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

   1 /*
   2  * Project:     libFIRM
   3  * File name:   ir/opt/ldstopt.c
   4  * Purpose:     load store optimizations
   5  * Author:      Michael Beck
   6  * Created:
   7  * CVS-ID:      $Id$
   8  * Copyright:   (c) 1998-2007 Universität Karlsruhe
   9  * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
  10  */
  11 #ifdef HAVE_CONFIG_H
  12 # include "config.h"
  13 #endif
  14
  15 #ifdef HAVE_STRING_H
  16 # include <string.h>
  17 #endif
  18
  19 #include "irnode_t.h"
  20 #include "irgraph_t.h"
  21 #include "irmode_t.h"
  22 #include "iropt_t.h"
  23 #include "ircons_t.h"
  24 #include "irgmod.h"
  25 #include "irgwalk.h"
  26 #include "irvrfy.h"
  27 #include "tv_t.h"
  28 #include "dbginfo_t.h"
  29 #include "iropt_dbg.h"
  30 #include "irflag_t.h"
  31 #include "array.h"
  32 #include "irhooks.h"
  33 #include "iredges.h"
  34 #include "irtools.h"
  35 #include "opt_polymorphy.h"
  36 #include "irmemory.h"
  37 #include "xmalloc.h"
  38
  39 #ifdef DO_CACHEOPT
  40 #include "cacheopt/cachesim.h"
  41 #endif
  42
  43 #undef IMAX
  44 #define IMAX(a,b)       ((a) > (b) ? (a) : (b))
  45
  46 #define MAX_PROJ        IMAX(pn_Load_max, pn_Store_max)
  47
  48 enum changes_t {
  49         DF_CHANGED = 1,       /**< data flow changed */
  50         CF_CHANGED = 2,       /**< control flow changed */
  51 };
  52
  53 /**
  54  * walker environment
  55  */
  56 typedef struct _walk_env_t {
  57         struct obstack obst;          /**< list of all stores */
  58         unsigned changes;             /**< a bitmask of graph changes */
  59 } walk_env_t;
  60
  61 /**
  62  * flags for Load/Store
  63  */
  64 enum ldst_flags_t {
  65         LDST_VISITED = 1              /**< if set, this Load/Store is already visited */
  66 };
  67
  68 /** A Load/Store info. */
  69 typedef struct _ldst_info_t {
  70         ir_node  *projs[MAX_PROJ];    /**< list of Proj's of this node */
  71         ir_node  *exc_block;          /**< the exception block if available */
  72         int      exc_idx;             /**< predecessor index in the exception block */
  73         unsigned flags;               /**< flags */
  74         unsigned visited;             /**< visited counter for breaking loops */
  75 } ldst_info_t;
  76
  77 /**
  78  * flags for control flow.
  79  */
  80 enum block_flags_t {
  81         BLOCK_HAS_COND = 1,      /**< Block has conditional control flow */
  82         BLOCK_HAS_EXC  = 2       /**< Block has exceptional control flow */
  83 };
  84
  85 /**
  86  * a Block info.
  87  */
  88 typedef struct _block_info_t {
  89         unsigned flags;               /**< flags for the block */
  90 } block_info_t;
  91
  92 /** the master visited flag for loop detection. */
  93 static unsigned master_visited = 0;
  94
  95 #define INC_MASTER()       ++master_visited
  96 #define MARK_NODE(info)    (info)->visited = master_visited
  97 #define NODE_VISITED(info) (info)->visited >= master_visited
  98
  99 /**
 100  * get the Load/Store info of a node
 101  */
 102 static ldst_info_t *get_ldst_info(ir_node *node, walk_env_t *env) {
 103         ldst_info_t *info = get_irn_link(node);
 104
 105         if (! info) {
 106                 info = obstack_alloc(&env->obst, sizeof(*info));
 107                 memset(info, 0, sizeof(*info));
 108                 set_irn_link(node, info);
 109         }
 110         return info;
 111 }  /* get_ldst_info */
 112
 113 /**
 114  * get the Block info of a node
 115  */
 116 static block_info_t *get_block_info(ir_node *node, walk_env_t *env) {
 117         block_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_block_info */
 126
 127 /**
 128  * update the projection info for a Load/Store
 129  */
 130 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
 131 {
 132         long nr = get_Proj_proj(proj);
 133
 134         assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
 135
 136         if (info->projs[nr]) {
 137                 /* there is already one, do CSE */
 138                 exchange(proj, info->projs[nr]);
 139                 return DF_CHANGED;
 140         }
 141         else {
 142                 info->projs[nr] = proj;
 143                 return 0;
 144         }
 145 }  /* update_projs */
 146
 147 /**
 148  * update the exception block info for a Load/Store node.
 149  *
 150  * @param info   the load/store info struct
 151  * @param block  the exception handler block for this load/store
 152  * @param pos    the control flow input of the block
 153  */
 154 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
 155 {
 156         assert(info->exc_block == NULL && "more than one exception block found");
 157
 158         info->exc_block = block;
 159         info->exc_idx   = pos;
 160         return 0;
 161 }  /* update_exc */
 162
 163 /** Return the number of uses of an address node */
 164 #define get_irn_n_uses(adr)     get_irn_n_edges(adr)
 165
 166 /**
 167  * walker, collects all Load/Store/Proj nodes
 168  *
 169  * walks from Start -> End
 170  */
 171 static void collect_nodes(ir_node *node, void *env)
 172 {
 173         ir_op       *op = get_irn_op(node);
 174         ir_node     *pred, *blk, *pred_blk;
 175         ldst_info_t *ldst_info;
 176         walk_env_t  *wenv = env;
 177
 178         if (op == op_Proj) {
 179                 ir_node *adr;
 180                 ir_op *op;
 181
 182                 pred = get_Proj_pred(node);
 183                 op   = get_irn_op(pred);
 184
 185                 if (op == op_Load) {
 186                         ldst_info = get_ldst_info(pred, wenv);
 187
 188                         wenv->changes |= update_projs(ldst_info, node);
 189
 190                         if ((ldst_info->flags & LDST_VISITED) == 0) {
 191                                 adr = get_Load_ptr(pred);
 192                                 ldst_info->flags |= LDST_VISITED;
 193                         }
 194
 195                         /*
 196                         * Place the Proj's to the same block as the
 197                         * predecessor Load. This is always ok and prevents
 198                         * "non-SSA" form after optimizations if the Proj
 199                         * is in a wrong block.
 200                         */
 201                         blk      = get_nodes_block(node);
 202                         pred_blk = get_nodes_block(pred);
 203                         if (blk != pred_blk) {
 204                                 wenv->changes |= DF_CHANGED;
 205                                 set_nodes_block(node, pred_blk);
 206                         }
 207                 } else if (op == op_Store) {
 208                         ldst_info = get_ldst_info(pred, wenv);
 209
 210                         wenv->changes |= update_projs(ldst_info, node);
 211
 212                         if ((ldst_info->flags & LDST_VISITED) == 0) {
 213                                 adr = get_Store_ptr(pred);
 214                                 ldst_info->flags |= LDST_VISITED;
 215                         }
 216
 217                         /*
 218                         * Place the Proj's to the same block as the
 219                         * predecessor Store. This is always ok and prevents
 220                         * "non-SSA" form after optimizations if the Proj
 221                         * is in a wrong block.
 222                         */
 223                         blk      = get_nodes_block(node);
 224                         pred_blk = get_nodes_block(pred);
 225                         if (blk != pred_blk) {
 226                                 wenv->changes |= DF_CHANGED;
 227                                 set_nodes_block(node, pred_blk);
 228                         }
 229                 }
 230         } else if (op == op_Block) {
 231                 int i;
 232
 233                 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
 234                         ir_node      *pred_block;
 235                         block_info_t *bl_info;
 236
 237                         pred = skip_Proj(get_Block_cfgpred(node, i));
 238
 239                         /* ignore Bad predecessors, they will be removed later */
 240                         if (is_Bad(pred))
 241                                 continue;
 242
 243                         pred_block = get_nodes_block(pred);
 244                         bl_info    = get_block_info(pred_block, wenv);
 245
 246                         if (is_fragile_op(pred))
 247                                 bl_info->flags |= BLOCK_HAS_EXC;
 248                         else if (is_irn_forking(pred))
 249                                 bl_info->flags |= BLOCK_HAS_COND;
 250
 251                         if (get_irn_op(pred) == op_Load || get_irn_op(pred) == op_Store) {
 252                                 ldst_info = get_ldst_info(pred, wenv);
 253
 254                                 wenv->changes |= update_exc(ldst_info, node, i);
 255                         }
 256                 }
 257         }
 258 }  /* collect_nodes */
 259
 260 /**
 261  * Returns an entity if the address ptr points to a constant one.
 262  *
 263  * @param ptr  the address
 264  *
 265  * @return an entity or NULL
 266  */
 267 static ir_entity *find_constant_entity(ir_node *ptr)
 268 {
 269         for (;;) {
 270                 ir_op *op = get_irn_op(ptr);
 271
 272                 if (op == op_SymConst && (get_SymConst_kind(ptr) == symconst_addr_ent)) {
 273                         ir_entity *ent = get_SymConst_entity(ptr);
 274                         if (variability_constant == get_entity_variability(ent))
 275                                 return ent;
 276                         return NULL;
 277                 } else if (op == op_Sel) {
 278                         ir_entity *ent = get_Sel_entity(ptr);
 279                         ir_type   *tp  = get_entity_owner(ent);
 280
 281                         /* Do not fiddle with polymorphism. */
 282                         if (is_Class_type(get_entity_owner(ent)) &&
 283                                 ((get_entity_n_overwrites(ent)    != 0) ||
 284                                 (get_entity_n_overwrittenby(ent) != 0)   ) )
 285                                 return NULL;
 286
 287                         if (is_Array_type(tp)) {
 288                                 /* check bounds */
 289                                 int i, n;
 290
 291                                 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
 292                                         ir_node *bound;
 293                                         tarval *tlower, *tupper;
 294                                         ir_node *index = get_Sel_index(ptr, i);
 295                                         tarval *tv     = computed_value(index);
 296
 297                                         /* check if the index is constant */
 298                                         if (tv == tarval_bad)
 299                                                 return NULL;
 300
 301                                         bound  = get_array_lower_bound(tp, i);
 302                                         tlower = computed_value(bound);
 303                                         bound  = get_array_upper_bound(tp, i);
 304                                         tupper = computed_value(bound);
 305
 306                                         if (tlower == tarval_bad || tupper == tarval_bad)
 307                                                 return NULL;
 308
 309                                         if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
 310                                                 return NULL;
 311                                         if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
 312                                                 return NULL;
 313
 314                                         /* ok, bounds check finished */
 315                                 }
 316                         }
 317
 318                         if (variability_constant == get_entity_variability(ent))
 319                                 return ent;
 320
 321                         /* try next */
 322                         ptr = get_Sel_ptr(ptr);
 323                 } else
 324                         return NULL;
 325         }
 326 }  /* find_constant_entity */
 327
 328 /**
 329  * Return the Selection index of a Sel node from dimension n
 330  */
 331 static long get_Sel_array_index_long(ir_node *n, int dim) {
 332         ir_node *index = get_Sel_index(n, dim);
 333         assert(get_irn_op(index) == op_Const);
 334         return get_tarval_long(get_Const_tarval(index));
 335 }  /* get_Sel_array_index_long */
 336
 337 /**
 338  * Returns the accessed component graph path for an
 339  * node computing an address.
 340  *
 341  * @param ptr    the node computing the address
 342  * @param depth  current depth in steps upward from the root
 343  *               of the address
 344  */
 345 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
 346         compound_graph_path *res = NULL;
 347         ir_entity           *root, *field;
 348         int                 path_len, pos;
 349
 350         if (get_irn_op(ptr) == op_SymConst) {
 351                 /* a SymConst. If the depth is 0, this is an access to a global
 352                  * entity and we don't need a component path, else we know
 353                  * at least it's length.
 354                  */
 355                 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
 356                 root = get_SymConst_entity(ptr);
 357                 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
 358         } else {
 359                 assert(get_irn_op(ptr) == op_Sel);
 360                 /* it's a Sel, go up until we find the root */
 361                 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
 362
 363                 /* fill up the step in the path at the current position */
 364                 field    = get_Sel_entity(ptr);
 365                 path_len = get_compound_graph_path_length(res);
 366                 pos      = path_len - depth - 1;
 367                 set_compound_graph_path_node(res, pos, field);
 368
 369                 if (is_Array_type(get_entity_owner(field))) {
 370                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 371                         set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
 372                 }
 373         }
 374         return res;
 375 }  /* rec_get_accessed_path */
 376
 377 /** Returns an access path or NULL.  The access path is only
 378  *  valid, if the graph is in phase_high and _no_ address computation is used.
 379  */
 380 static compound_graph_path *get_accessed_path(ir_node *ptr) {
 381         return rec_get_accessed_path(ptr, 0);
 382 }  /* get_accessed_path */
 383
 384 /* forward */
 385 static void reduce_adr_usage(ir_node *ptr);
 386
 387 /**
 388  * Update a Load that may lost it's usage.
 389  */
 390 static void handle_load_update(ir_node *load) {
 391         ldst_info_t *info = get_irn_link(load);
 392
 393         /* do NOT touch volatile loads for now */
 394         if (get_Load_volatility(load) == volatility_is_volatile)
 395                 return;
 396
 397         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 398                 ir_node *ptr = get_Load_ptr(load);
 399                 ir_node *mem = get_Load_mem(load);
 400
 401                 /* a Load which value is neither used nor exception checked, remove it */
 402                 exchange(info->projs[pn_Load_M], mem);
 403                 exchange(load, new_Bad());
 404                 reduce_adr_usage(ptr);
 405         }
 406 }  /* handle_load_update */
 407
 408 /**
 409  * A Use of an address node is vanished. Check if this was a Proj
 410  * node and update the counters.
 411  */
 412 static void reduce_adr_usage(ir_node *ptr) {
 413         if (is_Proj(ptr)) {
 414                 if (get_irn_n_edges(ptr) <= 0) {
 415                         /* this Proj is dead now */
 416                         ir_node *pred = get_Proj_pred(ptr);
 417
 418                         if (is_Load(pred)) {
 419                                 ldst_info_t *info = get_irn_link(pred);
 420                                 info->projs[get_Proj_proj(ptr)] = NULL;
 421
 422                                 /* this node lost it's result proj, handle that */
 423                                 handle_load_update(pred);
 424                         }
 425                 }
 426         }
 427 }  /* reduce_adr_usage */
 428
 429 /**
 430  * Check, if an already existing value of mode old_mode can be converted
 431  * into the needed one new_mode without loss.
 432  */
 433 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
 434         if (old_mode == new_mode)
 435                 return 1;
 436
 437         /* if both modes are two-complement ones, we can always convert the
 438            Stored value into the needed one. */
 439         if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
 440                   get_mode_arithmetic(old_mode) == irma_twos_complement &&
 441                   get_mode_arithmetic(new_mode) == irma_twos_complement)
 442                 return 1;
 443         return 0;
 444 }  /* can_use_stored_value */
 445
 446 /**
 447  * Follow the memory chain as long as there are only Loads
 448  * and alias free Stores and try to replace current Load or Store
 449  * by a previous ones.
 450  * Note that in unreachable loops it might happen that we reach
 451  * load again, as well as we can fall into a cycle.
 452  * We break such cycles using a special visited flag.
 453  *
 454  * INC_MASTER() must be called before dive into
 455  */
 456 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
 457         unsigned res = 0;
 458         ldst_info_t *info = get_irn_link(load);
 459         ir_node *pred;
 460         ir_node *ptr       = get_Load_ptr(load);
 461         ir_node *mem       = get_Load_mem(load);
 462         ir_mode *load_mode = get_Load_mode(load);
 463
 464         for (pred = curr; load != pred; ) {
 465                 ldst_info_t *pred_info = get_irn_link(pred);
 466
 467                 /*
 468                  * BEWARE: one might think that checking the modes is useless, because
 469                  * if the pointers are identical, they refer to the same object.
 470                  * This is only true in strong typed languages, not in C were the following
 471                  * is possible a = *(ir_type1 *)p; b = *(ir_type2 *)p ...
 472                  */
 473                 if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
 474                     can_use_stored_value(get_irn_mode(get_Store_value(pred)), load_mode)) {
 475                         /*
 476                          * a Load immediately after a Store -- a read after write.
 477                          * We may remove the Load, if both Load & Store does not have an exception handler
 478                          * OR they are in the same block. In the latter case the Load cannot
 479                          * throw an exception when the previous Store was quiet.
 480                          *
 481                          * Why we need to check for Store Exception? If the Store cannot
 482                          * be executed (ROM) the exception handler might simply jump into
 483                          * the load block :-(
 484                          * We could make it a little bit better if we would know that the exception
 485                          * handler of the Store jumps directly to the end...
 486                          */
 487                         if ((pred_info->projs[pn_Store_X_except] == NULL && info->projs[pn_Load_X_except] == NULL) ||
 488                             get_nodes_block(load) == get_nodes_block(pred)) {
 489                                 ir_node *value = get_Store_value(pred);
 490
 491                                 DBG_OPT_RAW(load, value);
 492
 493                                 /* add an convert if needed */
 494                                 if (get_irn_mode(get_Store_value(pred)) != load_mode) {
 495                                         value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
 496                                 }
 497
 498                                 if (info->projs[pn_Load_M])
 499                                         exchange(info->projs[pn_Load_M], mem);
 500
 501                                 /* no exception */
 502                                 if (info->projs[pn_Load_X_except]) {
 503                                         exchange( info->projs[pn_Load_X_except], new_Bad());
 504                                         res |= CF_CHANGED;
 505                                 }
 506
 507                                 if (info->projs[pn_Load_res])
 508                                         exchange(info->projs[pn_Load_res], value);
 509
 510                                 exchange(load, new_Bad());
 511                                 reduce_adr_usage(ptr);
 512                                 return res | DF_CHANGED;
 513                         }
 514                 } else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
 515                            can_use_stored_value(get_Load_mode(pred), load_mode)) {
 516                         /*
 517                          * a Load after a Load -- a read after read.
 518                          * We may remove the second Load, if it does not have an exception handler
 519                          * OR they are in the same block. In the later case the Load cannot
 520                          * throw an exception when the previous Load was quiet.
 521                          *
 522                          * Here, there is no need to check if the previous Load has an exception
 523                          * hander because they would have exact the same exception...
 524                          */
 525                         if (info->projs[pn_Load_X_except] == NULL || get_nodes_block(load) == get_nodes_block(pred)) {
 526                                 ir_node *value;
 527
 528                                 DBG_OPT_RAR(load, pred);
 529
 530                                 /* the result is used */
 531                                 if (info->projs[pn_Load_res]) {
 532                                         if (pred_info->projs[pn_Load_res] == NULL) {
 533                                                 /* create a new Proj again */
 534                                                 pred_info->projs[pn_Load_res] = new_r_Proj(current_ir_graph, get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res);
 535                                         }
 536                                         value = pred_info->projs[pn_Load_res];
 537
 538                                         /* add an convert if needed */
 539                                         if (get_Load_mode(pred) != load_mode) {
 540                                                 value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
 541                                         }
 542
 543                                         exchange(info->projs[pn_Load_res], value);
 544                                 }
 545
 546                                 if (info->projs[pn_Load_M])
 547                                         exchange(info->projs[pn_Load_M], mem);
 548
 549                                 /* no exception */
 550                                 if (info->projs[pn_Load_X_except]) {
 551                                         exchange(info->projs[pn_Load_X_except], new_Bad());
 552                                         res |= CF_CHANGED;
 553                                 }
 554
 555                                 exchange(load, new_Bad());
 556                                 reduce_adr_usage(ptr);
 557                                 return res |= DF_CHANGED;
 558                         }
 559                 }
 560
 561                 if (get_irn_op(pred) == op_Store) {
 562                         /* check if we can pass through this store */
 563                         ir_alias_relation rel = get_alias_relation(
 564                                 current_ir_graph,
 565                                 get_Store_ptr(pred),
 566                                 get_irn_mode(get_Store_value(pred)),
 567                                 ptr, load_mode);
 568                         /* if the might be an alias, we cannot pass this Store */
 569                         if (rel != no_alias)
 570                                 break;
 571                         pred = skip_Proj(get_Store_mem(pred));
 572                 } else if (get_irn_op(pred) == op_Load) {
 573                         pred = skip_Proj(get_Load_mem(pred));
 574                 } else {
 575                         /* follow only Load chains */
 576                         break;
 577                 }
 578
 579                 /* check for cycles */
 580                 if (NODE_VISITED(pred_info))
 581                         break;
 582                 MARK_NODE(pred_info);
 583         }
 584
 585         if (get_irn_op(pred) == op_Sync) {
 586                 int i;
 587
 588                 /* handle all Sync predecessors */
 589                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
 590                         res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
 591                         if (res)
 592                                 break;
 593                 }
 594         }
 595
 596         return res;
 597 }  /* follow_Mem_chain */
 598
 599 /**
 600  * optimize a Load
 601  *
 602  * @param load  the Load node
 603  */
 604 static unsigned optimize_load(ir_node *load)
 605 {
 606         ldst_info_t *info = get_irn_link(load);
 607         ir_node *mem, *ptr, *new_node;
 608         ir_entity *ent;
 609         unsigned res = 0;
 610
 611         /* do NOT touch volatile loads for now */
 612         if (get_Load_volatility(load) == volatility_is_volatile)
 613                 return 0;
 614
 615         /* the address of the load to be optimized */
 616         ptr = get_Load_ptr(load);
 617
 618         /*
 619          * Check if we can remove the exception from a Load:
 620          * This can be done, if the address is from an Sel(Alloc) and
 621          * the Sel type is a subtype of the allocated type.
 622          *
 623          * This optimizes some often used OO constructs,
 624          * like x = new O; x->t;
 625          */
 626         if (info->projs[pn_Load_X_except]) {
 627                 if (is_Sel(ptr)) {
 628                         ir_node *mem = get_Sel_mem(ptr);
 629
 630                         /* FIXME: works with the current FE, but better use the base */
 631                         if (get_irn_op(skip_Proj(mem)) == op_Alloc) {
 632                                 /* ok, check the types */
 633                                 ir_entity *ent    = get_Sel_entity(ptr);
 634                                 ir_type   *s_type = get_entity_type(ent);
 635                                 ir_type   *a_type = get_Alloc_type(mem);
 636
 637                                 if (is_SubClass_of(s_type, a_type)) {
 638                                         /* ok, condition met: there can't be an exception because
 639                                         * Alloc guarantees that enough memory was allocated */
 640
 641                                         exchange(info->projs[pn_Load_X_except], new_Bad());
 642                                         info->projs[pn_Load_X_except] = NULL;
 643                                         res |= CF_CHANGED;
 644                                 }
 645                         }
 646                 } else if ((get_irn_op(skip_Proj(ptr)) == op_Alloc) ||
 647                         ((get_irn_op(ptr) == op_Cast) && (get_irn_op(skip_Proj(get_Cast_op(ptr))) == op_Alloc))) {
 648                                 /* simple case: a direct load after an Alloc. Firm Alloc throw
 649                                  * an exception in case of out-of-memory. So, there is no way for an
 650                                  * exception in this load.
 651                                  * This code is constructed by the "exception lowering" in the Jack compiler.
 652                                  */
 653                                 exchange(info->projs[pn_Load_X_except], new_Bad());
 654                                 info->projs[pn_Load_X_except] = NULL;
 655                                 res |= CF_CHANGED;
 656                 }
 657         }
 658
 659         /* The mem of the Load. Must still be returned after optimization. */
 660         mem  = get_Load_mem(load);
 661
 662         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 663                 /* a Load which value is neither used nor exception checked, remove it */
 664                 exchange(info->projs[pn_Load_M], mem);
 665
 666                 exchange(load, new_Bad());
 667                 reduce_adr_usage(ptr);
 668                 return res | DF_CHANGED;
 669         }
 670
 671         /* Load from a constant polymorphic field, where we can resolve
 672            polymorphism. */
 673         new_node = transform_node_Load(load);
 674         if (new_node != load) {
 675                 if (info->projs[pn_Load_M]) {
 676                         exchange(info->projs[pn_Load_M], mem);
 677                         info->projs[pn_Load_M] = NULL;
 678                 }
 679                 if (info->projs[pn_Load_X_except]) {
 680                         exchange(info->projs[pn_Load_X_except], new_Bad());
 681                         info->projs[pn_Load_X_except] = NULL;
 682                 }
 683                 if (info->projs[pn_Load_res])
 684                         exchange(info->projs[pn_Load_res], new_node);
 685
 686                 exchange(load, new_Bad());
 687                 reduce_adr_usage(ptr);
 688                 return res | DF_CHANGED;
 689         }
 690
 691         /* check if we can determine the entity that will be loaded */
 692         ent = find_constant_entity(ptr);
 693         if (ent) {
 694                 if ((allocation_static == get_entity_allocation(ent)) &&
 695                         (visibility_external_allocated != get_entity_visibility(ent))) {
 696                         /* a static allocation that is not external: there should be NO exception
 697                          * when loading. */
 698
 699                         /* no exception, clear the info field as it might be checked later again */
 700                         if (info->projs[pn_Load_X_except]) {
 701                                 exchange(info->projs[pn_Load_X_except], new_Bad());
 702                                 info->projs[pn_Load_X_except] = NULL;
 703                                 res |= CF_CHANGED;
 704                         }
 705
 706                         if (variability_constant == get_entity_variability(ent)
 707                                 && is_atomic_entity(ent)) {
 708                                 /* Might not be atomic after
 709                                    lowering of Sels.  In this
 710                                    case we could also load, but
 711                                    it's more complicated. */
 712                                 /* more simpler case: we load the content of a constant value:
 713                                  * replace it by the constant itself
 714                                  */
 715
 716                                 /* no memory */
 717                                 if (info->projs[pn_Load_M]) {
 718                                         exchange(info->projs[pn_Load_M], mem);
 719                                         res |= DF_CHANGED;
 720                                 }
 721                                 /* no result :-) */
 722                                 if (info->projs[pn_Load_res]) {
 723                                         if (is_atomic_entity(ent)) {
 724                                                 ir_node *c = copy_const_value(get_irn_dbg_info(load), get_atomic_ent_value(ent));
 725
 726                                                 DBG_OPT_RC(load, c);
 727                                                 exchange(info->projs[pn_Load_res], c);
 728                                                 res |= DF_CHANGED;
 729                                         }
 730                                 }
 731                                 exchange(load, new_Bad());
 732                                 reduce_adr_usage(ptr);
 733                                 return res;
 734                         } else if (variability_constant == get_entity_variability(ent)) {
 735                                 compound_graph_path *path = get_accessed_path(ptr);
 736
 737                                 if (path) {
 738                                         ir_node *c;
 739
 740                                         assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
 741                                         /*
 742                                         {
 743                                                 int j;
 744                                                 for (j = 0; j < get_compound_graph_path_length(path); ++j) {
 745                                                         ir_entity *node = get_compound_graph_path_node(path, j);
 746                                                         fprintf(stdout, ".%s", get_entity_name(node));
 747                                                         if (is_Array_type(get_entity_owner(node)))
 748                                                                 fprintf(stdout, "[%d]", get_compound_graph_path_array_index(path, j));
 749                                                 }
 750                                                 printf("\n");
 751                                         }
 752                                         */
 753
 754                                         c = get_compound_ent_value_by_path(ent, path);
 755                                         free_compound_graph_path(path);
 756
 757                                         /* printf("  cons: "); DDMN(c); */
 758
 759                                         if (info->projs[pn_Load_M]) {
 760                                                 exchange(info->projs[pn_Load_M], mem);
 761                                                 res |= DF_CHANGED;
 762                                         }
 763                                         if (info->projs[pn_Load_res]) {
 764                                                 exchange(info->projs[pn_Load_res], copy_const_value(get_irn_dbg_info(load), c));
 765                                                 res |= DF_CHANGED;
 766                                         }
 767                                         exchange(load, new_Bad());
 768                                         reduce_adr_usage(ptr);
 769                                         return res;
 770                                 } else {
 771                                         /*  We can not determine a correct access path.  E.g., in jack, we load
 772                                         a byte from an object to generate an exception.   Happens in test program
 773                                         Reflectiontest.
 774                                         printf(">>>>>>>>>>>>> Found access to constant entity %s in function %s\n", get_entity_name(ent),
 775                                         get_entity_name(get_irg_entity(current_ir_graph)));
 776                                         printf("  load: "); DDMN(load);
 777                                         printf("  ptr:  "); DDMN(ptr);
 778                                         */
 779                                 }
 780                         }
 781                 }
 782         }
 783
 784         /* Check, if the address of this load is used more than once.
 785          * If not, this load cannot be removed in any case. */
 786         if (get_irn_n_uses(ptr) <= 1)
 787                 return res;
 788
 789         /*
 790          * follow the memory chain as long as there are only Loads
 791          * and try to replace current Load or Store by a previous one.
 792          * Note that in unreachable loops it might happen that we reach
 793          * load again, as well as we can fall into a cycle.
 794          * We break such cycles using a special visited flag.
 795          */
 796         INC_MASTER();
 797         res = follow_Mem_chain(load, skip_Proj(mem));
 798         return res;
 799 }  /* optimize_load */
 800
 801 /**
 802  * Check whether a value of mode new_mode would completely overwrite a value
 803  * of mode old_mode in memory.
 804  */
 805 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
 806 {
 807         return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
 808 }  /* is_completely_overwritten */
 809
 810 /**
 811  * follow the memory chain as long as there are only Loads and alias free Stores.
 812  *
 813  * INC_MASTER() must be called before dive into
 814  */
 815 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
 816         unsigned res = 0;
 817         ldst_info_t *info = get_irn_link(store);
 818         ir_node *pred;
 819         ir_node *ptr = get_Store_ptr(store);
 820         ir_node *mem = get_Store_mem(store);
 821         ir_node *value = get_Store_value(store);
 822         ir_mode *mode  = get_irn_mode(value);
 823         ir_node *block = get_nodes_block(store);
 824
 825         for (pred = curr; pred != store;) {
 826                 ldst_info_t *pred_info = get_irn_link(pred);
 827
 828                 /*
 829                  * BEWARE: one might think that checking the modes is useless, because
 830                  * if the pointers are identical, they refer to the same object.
 831                  * This is only true in strong typed languages, not is C were the following
 832                  * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
 833                  * However, if the mode that is written have a bigger  or equal size the the old
 834                  * one, the old value is completely overwritten and can be killed ...
 835                  */
 836                 if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
 837                     get_nodes_block(pred) == block &&
 838                     is_completely_overwritten(get_irn_mode(get_Store_value(pred)), mode)) {
 839                         /*
 840                          * a Store after a Store in the same block -- a write after write.
 841                          * We may remove the first Store, if it does not have an exception handler.
 842                          *
 843                          * TODO: What, if both have the same exception handler ???
 844                          */
 845                         if (get_Store_volatility(pred) != volatility_is_volatile && !pred_info->projs[pn_Store_X_except]) {
 846                                 DBG_OPT_WAW(pred, store);
 847                                 exchange( pred_info->projs[pn_Store_M], get_Store_mem(pred) );
 848                                 exchange(pred, new_Bad());
 849                                 reduce_adr_usage(ptr);
 850                                 return DF_CHANGED;
 851                         }
 852                 } else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
 853                            value == pred_info->projs[pn_Load_res]) {
 854                         /*
 855                          * a Store of a value after a Load -- a write after read.
 856                          * We may remove the second Store, if it does not have an exception handler.
 857                          */
 858                         if (! info->projs[pn_Store_X_except]) {
 859                                 DBG_OPT_WAR(store, pred);
 860                                 exchange( info->projs[pn_Store_M], mem );
 861                                 exchange(store, new_Bad());
 862                                 reduce_adr_usage(ptr);
 863                                 return DF_CHANGED;
 864                         }
 865                 }
 866
 867                 if (get_irn_op(pred) == op_Store) {
 868                         /* check if we can pass thru this store */
 869                         ir_alias_relation rel = get_alias_relation(
 870                                 current_ir_graph,
 871                                 get_Store_ptr(pred),
 872                                 get_irn_mode(get_Store_value(pred)),
 873                                 ptr, mode);
 874                         /* if the might be an alias, we cannot pass this Store */
 875                         if (rel != no_alias)
 876                                 break;
 877                         pred = skip_Proj(get_Store_mem(pred));
 878                 } else if (get_irn_op(pred) == op_Load) {
 879                         pred = skip_Proj(get_Load_mem(pred));
 880                 } else {
 881                         /* follow only Load chains */
 882                         break;
 883                 }
 884
 885                 /* check for cycles */
 886                 if (NODE_VISITED(pred_info))
 887                         break;
 888                 MARK_NODE(pred_info);
 889         }
 890
 891         if (get_irn_op(pred) == op_Sync) {
 892                 int i;
 893
 894                 /* handle all Sync predecessors */
 895                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
 896                         res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
 897                         if (res)
 898                                 break;
 899                 }
 900         }
 901         return res;
 902 }  /* follow_Mem_chain_for_Store */
 903
 904 /**
 905  * optimize a Store
 906  *
 907  * @param store  the Store node
 908  */
 909 static unsigned optimize_store(ir_node *store) {
 910         ir_node *ptr, *mem;
 911
 912         if (get_Store_volatility(store) == volatility_is_volatile)
 913                 return 0;
 914
 915         ptr = get_Store_ptr(store);
 916
 917         /* Check, if the address of this Store is used more than once.
 918          * If not, this Store cannot be removed in any case. */
 919         if (get_irn_n_uses(ptr) <= 1)
 920                 return 0;
 921
 922         mem = get_Store_mem(store);
 923
 924         /* follow the memory chain as long as there are only Loads */
 925         INC_MASTER();
 926         return follow_Mem_chain_for_Store(store, skip_Proj(mem));
 927 }  /* optimize_store */
 928
 929 /**
 930  * walker, optimizes Phi after Stores to identical places:
 931  * Does the following optimization:
 932  * @verbatim
 933  *
 934  *   val1   val2   val3          val1  val2  val3
 935  *    |      |      |               \    |    /
 936  *  Store  Store  Store              \   |   /
 937  *      \    |    /                   PhiData
 938  *       \   |   /                       |
 939  *        \  |  /                      Store
 940  *          PhiM
 941  *
 942  * @endverbatim
 943  * This reduces the number of stores and allows for predicated execution.
 944  * Moves Stores back to the end of a function which may be bad.
 945  *
 946  * This is only possible if the predecessor blocks have only one successor.
 947  */
 948 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
 949 {
 950         int i, n;
 951         ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
 952         ir_mode *mode;
 953         ir_node **inM, **inD, **stores;
 954         int *idx;
 955         dbg_info *db = NULL;
 956         ldst_info_t *info;
 957         block_info_t *bl_info;
 958         unsigned res = 0;
 959
 960         /* Must be a memory Phi */
 961         if (get_irn_mode(phi) != mode_M)
 962                 return 0;
 963
 964         n = get_Phi_n_preds(phi);
 965         if (n <= 0)
 966                 return 0;
 967
 968         store = skip_Proj(get_Phi_pred(phi, 0));
 969         old_store = store;
 970         if (get_irn_op(store) != op_Store)
 971                 return 0;
 972
 973         block = get_nodes_block(store);
 974
 975         /* abort on dead blocks */
 976         if (is_Block_dead(block))
 977                 return 0;
 978
 979         /* check if the block is post dominated by Phi-block
 980            and has no exception exit */
 981         bl_info = get_irn_link(block);
 982         if (bl_info->flags & BLOCK_HAS_EXC)
 983                 return 0;
 984
 985         phi_block = get_nodes_block(phi);
 986         if (! block_postdominates(phi_block, block))
 987                 return 0;
 988
 989         /* this is the address of the store */
 990         ptr  = get_Store_ptr(store);
 991         mode = get_irn_mode(get_Store_value(store));
 992         info = get_irn_link(store);
 993         exc  = info->exc_block;
 994
 995         for (i = 1; i < n; ++i) {
 996                 ir_node *pred = skip_Proj(get_Phi_pred(phi, i));
 997
 998                 if (get_irn_op(pred) != op_Store)
 999                         return 0;
1000
1001                 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1002                         return 0;
1003
1004                 info = get_irn_link(pred);
1005
1006                 /* check, if all stores have the same exception flow */
1007                 if (exc != info->exc_block)
1008                         return 0;
1009
1010                 /* abort on dead blocks */
1011                 block = get_nodes_block(pred);
1012                 if (is_Block_dead(block))
1013                         return 0;
1014
1015                 /* check if the block is post dominated by Phi-block
1016                    and has no exception exit. Note that block must be different from
1017                    Phi-block, else we would move a Store from end End of a block to its
1018                    Start... */
1019                 bl_info = get_irn_link(block);
1020                 if (bl_info->flags & BLOCK_HAS_EXC)
1021                         return 0;
1022                 if (block == phi_block || ! block_postdominates(phi_block, block))
1023                         return 0;
1024         }
1025
1026         /*
1027          * ok, when we are here, we found all predecessors of a Phi that
1028          * are Stores to the same address and size. That means whatever
1029          * we do before we enter the block of the Phi, we do a Store.
1030          * So, we can move the Store to the current block:
1031          *
1032          *   val1    val2    val3          val1  val2  val3
1033          *    |       |       |               \    |    /
1034          * | Str | | Str | | Str |             \   |   /
1035          *      \     |     /                   PhiData
1036          *       \    |    /                       |
1037          *        \   |   /                       Str
1038          *           PhiM
1039          *
1040          * Is only allowed if the predecessor blocks have only one successor.
1041          */
1042
1043         NEW_ARR_A(ir_node *, stores, n);
1044         NEW_ARR_A(ir_node *, inM, n);
1045         NEW_ARR_A(ir_node *, inD, n);
1046         NEW_ARR_A(int, idx, n);
1047
1048         /* Prepare: Collect all Store nodes.  We must do this
1049            first because we otherwise may loose a store when exchanging its
1050            memory Proj.
1051          */
1052         for (i = 0; i < n; ++i)
1053                 stores[i] = skip_Proj(get_Phi_pred(phi, i));
1054
1055         /* Prepare: Skip the memory Proj: we need this in the case some stores
1056            are cascaded.
1057            Beware: One Store might be included more than once in the stores[]
1058            list, so we must prevent to do the exchange more than once.
1059          */
1060         for (i = 0; i < n; ++i) {
1061                 ir_node *store = stores[i];
1062                 ir_node *proj_m;
1063
1064                 info = get_irn_link(store);
1065                 proj_m = info->projs[pn_Store_M];
1066
1067                 if (is_Proj(proj_m) && get_Proj_pred(proj_m) == store)
1068                         exchange(proj_m, get_Store_mem(store));
1069         }
1070
1071         /* first step: collect all inputs */
1072         for (i = 0; i < n; ++i) {
1073                 ir_node *store = stores[i];
1074                 info = get_irn_link(store);
1075
1076                 inM[i] = get_Store_mem(store);
1077                 inD[i] = get_Store_value(store);
1078                 idx[i] = info->exc_idx;
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 */