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 #ifdef HAVE_STRING_H
  31 # include <string.h>
  32 #endif
  33
  34 #include "irnode_t.h"
  35 #include "irgraph_t.h"
  36 #include "irmode_t.h"
  37 #include "iropt_t.h"
  38 #include "ircons_t.h"
  39 #include "irgmod.h"
  40 #include "irgwalk.h"
  41 #include "irvrfy.h"
  42 #include "tv_t.h"
  43 #include "dbginfo_t.h"
  44 #include "iropt_dbg.h"
  45 #include "irflag_t.h"
  46 #include "array.h"
  47 #include "irhooks.h"
  48 #include "iredges.h"
  49 #include "irtools.h"
  50 #include "opt_polymorphy.h"
  51 #include "irmemory.h"
  52 #include "xmalloc.h"
  53
  54 #ifdef DO_CACHEOPT
  55 #include "cacheopt/cachesim.h"
  56 #endif
  57
  58 #undef IMAX
  59 #define IMAX(a,b)       ((a) > (b) ? (a) : (b))
  60
  61 #define MAX_PROJ        IMAX(pn_Load_max, pn_Store_max)
  62
  63 enum changes_t {
  64         DF_CHANGED = 1,       /**< data flow changed */
  65         CF_CHANGED = 2,       /**< control flow changed */
  66 };
  67
  68 /**
  69  * walker environment
  70  */
  71 typedef struct _walk_env_t {
  72         struct obstack obst;          /**< list of all stores */
  73         unsigned changes;             /**< a bitmask of graph changes */
  74 } walk_env_t;
  75
  76 /**
  77  * flags for Load/Store
  78  */
  79 enum ldst_flags_t {
  80         LDST_VISITED = 1              /**< if set, this Load/Store is already visited */
  81 };
  82
  83 /** A Load/Store info. */
  84 typedef struct _ldst_info_t {
  85         ir_node  *projs[MAX_PROJ];    /**< list of Proj's of this node */
  86         ir_node  *exc_block;          /**< the exception block if available */
  87         int      exc_idx;             /**< predecessor index in the exception block */
  88         unsigned flags;               /**< flags */
  89         unsigned visited;             /**< visited counter for breaking loops */
  90 } ldst_info_t;
  91
  92 /**
  93  * flags for control flow.
  94  */
  95 enum block_flags_t {
  96         BLOCK_HAS_COND = 1,      /**< Block has conditional control flow */
  97         BLOCK_HAS_EXC  = 2       /**< Block has exceptional control flow */
  98 };
  99
 100 /**
 101  * a Block info.
 102  */
 103 typedef struct _block_info_t {
 104         unsigned flags;               /**< flags for the block */
 105 } block_info_t;
 106
 107 /** the master visited flag for loop detection. */
 108 static unsigned master_visited = 0;
 109
 110 #define INC_MASTER()       ++master_visited
 111 #define MARK_NODE(info)    (info)->visited = master_visited
 112 #define NODE_VISITED(info) (info)->visited >= master_visited
 113
 114 /**
 115  * get the Load/Store info of a node
 116  */
 117 static ldst_info_t *get_ldst_info(ir_node *node, walk_env_t *env) {
 118         ldst_info_t *info = get_irn_link(node);
 119
 120         if (! info) {
 121                 info = obstack_alloc(&env->obst, sizeof(*info));
 122                 memset(info, 0, sizeof(*info));
 123                 set_irn_link(node, info);
 124         }
 125         return info;
 126 }  /* get_ldst_info */
 127
 128 /**
 129  * get the Block info of a node
 130  */
 131 static block_info_t *get_block_info(ir_node *node, walk_env_t *env) {
 132         block_info_t *info = get_irn_link(node);
 133
 134         if (! info) {
 135                 info = obstack_alloc(&env->obst, sizeof(*info));
 136                 memset(info, 0, sizeof(*info));
 137                 set_irn_link(node, info);
 138         }
 139         return info;
 140 }  /* get_block_info */
 141
 142 /**
 143  * update the projection info for a Load/Store
 144  */
 145 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
 146 {
 147         long nr = get_Proj_proj(proj);
 148
 149         assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
 150
 151         if (info->projs[nr]) {
 152                 /* there is already one, do CSE */
 153                 exchange(proj, info->projs[nr]);
 154                 return DF_CHANGED;
 155         }
 156         else {
 157                 info->projs[nr] = proj;
 158                 return 0;
 159         }
 160 }  /* update_projs */
 161
 162 /**
 163  * update the exception block info for a Load/Store node.
 164  *
 165  * @param info   the load/store info struct
 166  * @param block  the exception handler block for this load/store
 167  * @param pos    the control flow input of the block
 168  */
 169 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
 170 {
 171         assert(info->exc_block == NULL && "more than one exception block found");
 172
 173         info->exc_block = block;
 174         info->exc_idx   = pos;
 175         return 0;
 176 }  /* update_exc */
 177
 178 /** Return the number of uses of an address node */
 179 #define get_irn_n_uses(adr)     get_irn_n_edges(adr)
 180
 181 /**
 182  * walker, collects all Load/Store/Proj nodes
 183  *
 184  * walks from Start -> End
 185  */
 186 static void collect_nodes(ir_node *node, void *env)
 187 {
 188         ir_op       *op = get_irn_op(node);
 189         ir_node     *pred, *blk, *pred_blk;
 190         ldst_info_t *ldst_info;
 191         walk_env_t  *wenv = env;
 192
 193         if (op == op_Proj) {
 194                 ir_node *adr;
 195                 ir_op *op;
 196
 197                 pred = get_Proj_pred(node);
 198                 op   = get_irn_op(pred);
 199
 200                 if (op == op_Load) {
 201                         ldst_info = get_ldst_info(pred, wenv);
 202
 203                         wenv->changes |= update_projs(ldst_info, node);
 204
 205                         if ((ldst_info->flags & LDST_VISITED) == 0) {
 206                                 adr = get_Load_ptr(pred);
 207                                 ldst_info->flags |= LDST_VISITED;
 208                         }
 209
 210                         /*
 211                         * Place the Proj's to the same block as the
 212                         * predecessor Load. This is always ok and prevents
 213                         * "non-SSA" form after optimizations if the Proj
 214                         * is in a wrong block.
 215                         */
 216                         blk      = get_nodes_block(node);
 217                         pred_blk = get_nodes_block(pred);
 218                         if (blk != pred_blk) {
 219                                 wenv->changes |= DF_CHANGED;
 220                                 set_nodes_block(node, pred_blk);
 221                         }
 222                 } else if (op == op_Store) {
 223                         ldst_info = get_ldst_info(pred, wenv);
 224
 225                         wenv->changes |= update_projs(ldst_info, node);
 226
 227                         if ((ldst_info->flags & LDST_VISITED) == 0) {
 228                                 adr = get_Store_ptr(pred);
 229                                 ldst_info->flags |= LDST_VISITED;
 230                         }
 231
 232                         /*
 233                         * Place the Proj's to the same block as the
 234                         * predecessor Store. This is always ok and prevents
 235                         * "non-SSA" form after optimizations if the Proj
 236                         * is in a wrong block.
 237                         */
 238                         blk      = get_nodes_block(node);
 239                         pred_blk = get_nodes_block(pred);
 240                         if (blk != pred_blk) {
 241                                 wenv->changes |= DF_CHANGED;
 242                                 set_nodes_block(node, pred_blk);
 243                         }
 244                 }
 245         } else if (op == op_Block) {
 246                 int i;
 247
 248                 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
 249                         ir_node      *pred_block;
 250                         block_info_t *bl_info;
 251
 252                         pred = skip_Proj(get_Block_cfgpred(node, i));
 253
 254                         /* ignore Bad predecessors, they will be removed later */
 255                         if (is_Bad(pred))
 256                                 continue;
 257
 258                         pred_block = get_nodes_block(pred);
 259                         bl_info    = get_block_info(pred_block, wenv);
 260
 261                         if (is_fragile_op(pred))
 262                                 bl_info->flags |= BLOCK_HAS_EXC;
 263                         else if (is_irn_forking(pred))
 264                                 bl_info->flags |= BLOCK_HAS_COND;
 265
 266                         if (get_irn_op(pred) == op_Load || get_irn_op(pred) == op_Store) {
 267                                 ldst_info = get_ldst_info(pred, wenv);
 268
 269                                 wenv->changes |= update_exc(ldst_info, node, i);
 270                         }
 271                 }
 272         }
 273 }  /* collect_nodes */
 274
 275 /**
 276  * Returns an entity if the address ptr points to a constant one.
 277  *
 278  * @param ptr  the address
 279  *
 280  * @return an entity or NULL
 281  */
 282 static ir_entity *find_constant_entity(ir_node *ptr)
 283 {
 284         for (;;) {
 285                 ir_op *op = get_irn_op(ptr);
 286
 287                 if (op == op_SymConst && (get_SymConst_kind(ptr) == symconst_addr_ent)) {
 288                         ir_entity *ent = get_SymConst_entity(ptr);
 289                         if (variability_constant == get_entity_variability(ent))
 290                                 return ent;
 291                         return NULL;
 292                 } else if (op == op_Sel) {
 293                         ir_entity *ent = get_Sel_entity(ptr);
 294                         ir_type   *tp  = get_entity_owner(ent);
 295
 296                         /* Do not fiddle with polymorphism. */
 297                         if (is_Class_type(get_entity_owner(ent)) &&
 298                                 ((get_entity_n_overwrites(ent)    != 0) ||
 299                                 (get_entity_n_overwrittenby(ent) != 0)   ) )
 300                                 return NULL;
 301
 302                         if (is_Array_type(tp)) {
 303                                 /* check bounds */
 304                                 int i, n;
 305
 306                                 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
 307                                         ir_node *bound;
 308                                         tarval *tlower, *tupper;
 309                                         ir_node *index = get_Sel_index(ptr, i);
 310                                         tarval *tv     = computed_value(index);
 311
 312                                         /* check if the index is constant */
 313                                         if (tv == tarval_bad)
 314                                                 return NULL;
 315
 316                                         bound  = get_array_lower_bound(tp, i);
 317                                         tlower = computed_value(bound);
 318                                         bound  = get_array_upper_bound(tp, i);
 319                                         tupper = computed_value(bound);
 320
 321                                         if (tlower == tarval_bad || tupper == tarval_bad)
 322                                                 return NULL;
 323
 324                                         if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
 325                                                 return NULL;
 326                                         if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
 327                                                 return NULL;
 328
 329                                         /* ok, bounds check finished */
 330                                 }
 331                         }
 332
 333                         if (variability_constant == get_entity_variability(ent))
 334                                 return ent;
 335
 336                         /* try next */
 337                         ptr = get_Sel_ptr(ptr);
 338                 } else
 339                         return NULL;
 340         }
 341 }  /* find_constant_entity */
 342
 343 /**
 344  * Return the Selection index of a Sel node from dimension n
 345  */
 346 static long get_Sel_array_index_long(ir_node *n, int dim) {
 347         ir_node *index = get_Sel_index(n, dim);
 348         assert(get_irn_op(index) == op_Const);
 349         return get_tarval_long(get_Const_tarval(index));
 350 }  /* get_Sel_array_index_long */
 351
 352 /**
 353  * Returns the accessed component graph path for an
 354  * node computing an address.
 355  *
 356  * @param ptr    the node computing the address
 357  * @param depth  current depth in steps upward from the root
 358  *               of the address
 359  */
 360 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
 361         compound_graph_path *res = NULL;
 362         ir_entity           *root, *field;
 363         int                 path_len, pos;
 364
 365         if (get_irn_op(ptr) == op_SymConst) {
 366                 /* a SymConst. If the depth is 0, this is an access to a global
 367                  * entity and we don't need a component path, else we know
 368                  * at least it's length.
 369                  */
 370                 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
 371                 root = get_SymConst_entity(ptr);
 372                 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
 373         } else {
 374                 assert(get_irn_op(ptr) == op_Sel);
 375                 /* it's a Sel, go up until we find the root */
 376                 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
 377
 378                 /* fill up the step in the path at the current position */
 379                 field    = get_Sel_entity(ptr);
 380                 path_len = get_compound_graph_path_length(res);
 381                 pos      = path_len - depth - 1;
 382                 set_compound_graph_path_node(res, pos, field);
 383
 384                 if (is_Array_type(get_entity_owner(field))) {
 385                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 386                         set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
 387                 }
 388         }
 389         return res;
 390 }  /* rec_get_accessed_path */
 391
 392 /** Returns an access path or NULL.  The access path is only
 393  *  valid, if the graph is in phase_high and _no_ address computation is used.
 394  */
 395 static compound_graph_path *get_accessed_path(ir_node *ptr) {
 396         return rec_get_accessed_path(ptr, 0);
 397 }  /* get_accessed_path */
 398
 399 /* forward */
 400 static void reduce_adr_usage(ir_node *ptr);
 401
 402 /**
 403  * Update a Load that may lost it's usage.
 404  */
 405 static void handle_load_update(ir_node *load) {
 406         ldst_info_t *info = get_irn_link(load);
 407
 408         /* do NOT touch volatile loads for now */
 409         if (get_Load_volatility(load) == volatility_is_volatile)
 410                 return;
 411
 412         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 413                 ir_node *ptr = get_Load_ptr(load);
 414                 ir_node *mem = get_Load_mem(load);
 415
 416                 /* a Load which value is neither used nor exception checked, remove it */
 417                 exchange(info->projs[pn_Load_M], mem);
 418                 exchange(load, new_Bad());
 419                 reduce_adr_usage(ptr);
 420         }
 421 }  /* handle_load_update */
 422
 423 /**
 424  * A Use of an address node is vanished. Check if this was a Proj
 425  * node and update the counters.
 426  */
 427 static void reduce_adr_usage(ir_node *ptr) {
 428         if (is_Proj(ptr)) {
 429                 if (get_irn_n_edges(ptr) <= 0) {
 430                         /* this Proj is dead now */
 431                         ir_node *pred = get_Proj_pred(ptr);
 432
 433                         if (is_Load(pred)) {
 434                                 ldst_info_t *info = get_irn_link(pred);
 435                                 info->projs[get_Proj_proj(ptr)] = NULL;
 436
 437                                 /* this node lost it's result proj, handle that */
 438                                 handle_load_update(pred);
 439                         }
 440                 }
 441         }
 442 }  /* reduce_adr_usage */
 443
 444 /**
 445  * Check, if an already existing value of mode old_mode can be converted
 446  * into the needed one new_mode without loss.
 447  */
 448 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
 449         if (old_mode == new_mode)
 450                 return 1;
 451
 452         /* if both modes are two-complement ones, we can always convert the
 453            Stored value into the needed one. */
 454         if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
 455                   get_mode_arithmetic(old_mode) == irma_twos_complement &&
 456                   get_mode_arithmetic(new_mode) == irma_twos_complement)
 457                 return 1;
 458         return 0;
 459 }  /* can_use_stored_value */
 460
 461 /**
 462  * Follow the memory chain as long as there are only Loads
 463  * and alias free Stores and try to replace current Load or Store
 464  * by a previous ones.
 465  * Note that in unreachable loops it might happen that we reach
 466  * load again, as well as we can fall into a cycle.
 467  * We break such cycles using a special visited flag.
 468  *
 469  * INC_MASTER() must be called before dive into
 470  */
 471 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
 472         unsigned res = 0;
 473         ldst_info_t *info = get_irn_link(load);
 474         ir_node *pred;
 475         ir_node *ptr       = get_Load_ptr(load);
 476         ir_node *mem       = get_Load_mem(load);
 477         ir_mode *load_mode = get_Load_mode(load);
 478
 479         for (pred = curr; load != pred; ) {
 480                 ldst_info_t *pred_info = get_irn_link(pred);
 481
 482                 /*
 483                  * BEWARE: one might think that checking the modes is useless, because
 484                  * if the pointers are identical, they refer to the same object.
 485                  * This is only true in strong typed languages, not in C were the following
 486                  * is possible a = *(ir_type1 *)p; b = *(ir_type2 *)p ...
 487                  */
 488                 if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
 489                     can_use_stored_value(get_irn_mode(get_Store_value(pred)), load_mode)) {
 490                         /*
 491                          * a Load immediately after a Store -- a read after write.
 492                          * We may remove the Load, if both Load & Store does not have an exception handler
 493                          * OR they are in the same block. In the latter case the Load cannot
 494                          * throw an exception when the previous Store was quiet.
 495                          *
 496                          * Why we need to check for Store Exception? If the Store cannot
 497                          * be executed (ROM) the exception handler might simply jump into
 498                          * the load block :-(
 499                          * We could make it a little bit better if we would know that the exception
 500                          * handler of the Store jumps directly to the end...
 501                          */
 502                         if ((pred_info->projs[pn_Store_X_except] == NULL && info->projs[pn_Load_X_except] == NULL) ||
 503                             get_nodes_block(load) == get_nodes_block(pred)) {
 504                                 ir_node *value = get_Store_value(pred);
 505
 506                                 DBG_OPT_RAW(load, value);
 507
 508                                 /* add an convert if needed */
 509                                 if (get_irn_mode(get_Store_value(pred)) != load_mode) {
 510                                         value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
 511                                 }
 512
 513                                 if (info->projs[pn_Load_M])
 514                                         exchange(info->projs[pn_Load_M], mem);
 515
 516                                 /* no exception */
 517                                 if (info->projs[pn_Load_X_except]) {
 518                                         exchange( info->projs[pn_Load_X_except], new_Bad());
 519                                         res |= CF_CHANGED;
 520                                 }
 521
 522                                 if (info->projs[pn_Load_res])
 523                                         exchange(info->projs[pn_Load_res], value);
 524
 525                                 exchange(load, new_Bad());
 526                                 reduce_adr_usage(ptr);
 527                                 return res | DF_CHANGED;
 528                         }
 529                 } else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
 530                            can_use_stored_value(get_Load_mode(pred), load_mode)) {
 531                         /*
 532                          * a Load after a Load -- a read after read.
 533                          * We may remove the second Load, if it does not have an exception handler
 534                          * OR they are in the same block. In the later case the Load cannot
 535                          * throw an exception when the previous Load was quiet.
 536                          *
 537                          * Here, there is no need to check if the previous Load has an exception
 538                          * hander because they would have exact the same exception...
 539                          */
 540                         if (info->projs[pn_Load_X_except] == NULL || get_nodes_block(load) == get_nodes_block(pred)) {
 541                                 ir_node *value;
 542
 543                                 DBG_OPT_RAR(load, pred);
 544
 545                                 /* the result is used */
 546                                 if (info->projs[pn_Load_res]) {
 547                                         if (pred_info->projs[pn_Load_res] == NULL) {
 548                                                 /* create a new Proj again */
 549                                                 pred_info->projs[pn_Load_res] = new_r_Proj(current_ir_graph, get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res);
 550                                         }
 551                                         value = pred_info->projs[pn_Load_res];
 552
 553                                         /* add an convert if needed */
 554                                         if (get_Load_mode(pred) != load_mode) {
 555                                                 value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
 556                                         }
 557
 558                                         exchange(info->projs[pn_Load_res], value);
 559                                 }
 560
 561                                 if (info->projs[pn_Load_M])
 562                                         exchange(info->projs[pn_Load_M], mem);
 563
 564                                 /* no exception */
 565                                 if (info->projs[pn_Load_X_except]) {
 566                                         exchange(info->projs[pn_Load_X_except], new_Bad());
 567                                         res |= CF_CHANGED;
 568                                 }
 569
 570                                 exchange(load, new_Bad());
 571                                 reduce_adr_usage(ptr);
 572                                 return res |= DF_CHANGED;
 573                         }
 574                 }
 575
 576                 if (get_irn_op(pred) == op_Store) {
 577                         /* check if we can pass through this store */
 578                         ir_alias_relation rel = get_alias_relation(
 579                                 current_ir_graph,
 580                                 get_Store_ptr(pred),
 581                                 get_irn_mode(get_Store_value(pred)),
 582                                 ptr, load_mode);
 583                         /* if the might be an alias, we cannot pass this Store */
 584                         if (rel != no_alias)
 585                                 break;
 586                         pred = skip_Proj(get_Store_mem(pred));
 587                 } else if (get_irn_op(pred) == op_Load) {
 588                         pred = skip_Proj(get_Load_mem(pred));
 589                 } else {
 590                         /* follow only Load chains */
 591                         break;
 592                 }
 593
 594                 /* check for cycles */
 595                 if (NODE_VISITED(pred_info))
 596                         break;
 597                 MARK_NODE(pred_info);
 598         }
 599
 600         if (get_irn_op(pred) == op_Sync) {
 601                 int i;
 602
 603                 /* handle all Sync predecessors */
 604                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
 605                         res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
 606                         if (res)
 607                                 break;
 608                 }
 609         }
 610
 611         return res;
 612 }  /* follow_Mem_chain */
 613
 614 /**
 615  * optimize a Load
 616  *
 617  * @param load  the Load node
 618  */
 619 static unsigned optimize_load(ir_node *load)
 620 {
 621         ldst_info_t *info = get_irn_link(load);
 622         ir_node *mem, *ptr, *new_node;
 623         ir_entity *ent;
 624         unsigned res = 0;
 625
 626         /* do NOT touch volatile loads for now */
 627         if (get_Load_volatility(load) == volatility_is_volatile)
 628                 return 0;
 629
 630         /* the address of the load to be optimized */
 631         ptr = get_Load_ptr(load);
 632
 633         /*
 634          * Check if we can remove the exception from a Load:
 635          * This can be done, if the address is from an Sel(Alloc) and
 636          * the Sel type is a subtype of the allocated type.
 637          *
 638          * This optimizes some often used OO constructs,
 639          * like x = new O; x->t;
 640          */
 641         if (info->projs[pn_Load_X_except]) {
 642                 if (is_Sel(ptr)) {
 643                         ir_node *mem = get_Sel_mem(ptr);
 644
 645                         /* FIXME: works with the current FE, but better use the base */
 646                         if (get_irn_op(skip_Proj(mem)) == op_Alloc) {
 647                                 /* ok, check the types */
 648                                 ir_entity *ent    = get_Sel_entity(ptr);
 649                                 ir_type   *s_type = get_entity_type(ent);
 650                                 ir_type   *a_type = get_Alloc_type(mem);
 651
 652                                 if (is_SubClass_of(s_type, a_type)) {
 653                                         /* ok, condition met: there can't be an exception because
 654                                         * Alloc guarantees that enough memory was allocated */
 655
 656                                         exchange(info->projs[pn_Load_X_except], new_Bad());
 657                                         info->projs[pn_Load_X_except] = NULL;
 658                                         res |= CF_CHANGED;
 659                                 }
 660                         }
 661                 } else if ((get_irn_op(skip_Proj(ptr)) == op_Alloc) ||
 662                         ((get_irn_op(ptr) == op_Cast) && (get_irn_op(skip_Proj(get_Cast_op(ptr))) == op_Alloc))) {
 663                                 /* simple case: a direct load after an Alloc. Firm Alloc throw
 664                                  * an exception in case of out-of-memory. So, there is no way for an
 665                                  * exception in this load.
 666                                  * This code is constructed by the "exception lowering" in the Jack compiler.
 667                                  */
 668                                 exchange(info->projs[pn_Load_X_except], new_Bad());
 669                                 info->projs[pn_Load_X_except] = NULL;
 670                                 res |= CF_CHANGED;
 671                 }
 672         }
 673
 674         /* The mem of the Load. Must still be returned after optimization. */
 675         mem  = get_Load_mem(load);
 676
 677         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 678                 /* a Load which value is neither used nor exception checked, remove it */
 679                 exchange(info->projs[pn_Load_M], mem);
 680
 681                 exchange(load, new_Bad());
 682                 reduce_adr_usage(ptr);
 683                 return res | DF_CHANGED;
 684         }
 685
 686         /* Load from a constant polymorphic field, where we can resolve
 687            polymorphism. */
 688         new_node = transform_node_Load(load);
 689         if (new_node != load) {
 690                 if (info->projs[pn_Load_M]) {
 691                         exchange(info->projs[pn_Load_M], mem);
 692                         info->projs[pn_Load_M] = NULL;
 693                 }
 694                 if (info->projs[pn_Load_X_except]) {
 695                         exchange(info->projs[pn_Load_X_except], new_Bad());
 696                         info->projs[pn_Load_X_except] = NULL;
 697                 }
 698                 if (info->projs[pn_Load_res])
 699                         exchange(info->projs[pn_Load_res], new_node);
 700
 701                 exchange(load, new_Bad());
 702                 reduce_adr_usage(ptr);
 703                 return res | DF_CHANGED;
 704         }
 705
 706         /* check if we can determine the entity that will be loaded */
 707         ent = find_constant_entity(ptr);
 708         if (ent) {
 709                 if ((allocation_static == get_entity_allocation(ent)) &&
 710                         (visibility_external_allocated != get_entity_visibility(ent))) {
 711                         /* a static allocation that is not external: there should be NO exception
 712                          * when loading. */
 713
 714                         /* no exception, clear the info field as it might be checked later again */
 715                         if (info->projs[pn_Load_X_except]) {
 716                                 exchange(info->projs[pn_Load_X_except], new_Bad());
 717                                 info->projs[pn_Load_X_except] = NULL;
 718                                 res |= CF_CHANGED;
 719                         }
 720
 721                         if (variability_constant == get_entity_variability(ent)
 722                                 && is_atomic_entity(ent)) {
 723                                 /* Might not be atomic after
 724                                    lowering of Sels.  In this
 725                                    case we could also load, but
 726                                    it's more complicated. */
 727                                 /* more simpler case: we load the content of a constant value:
 728                                  * replace it by the constant itself
 729                                  */
 730
 731                                 /* no memory */
 732                                 if (info->projs[pn_Load_M]) {
 733                                         exchange(info->projs[pn_Load_M], mem);
 734                                         res |= DF_CHANGED;
 735                                 }
 736                                 /* no result :-) */
 737                                 if (info->projs[pn_Load_res]) {
 738                                         if (is_atomic_entity(ent)) {
 739                                                 ir_node *c = copy_const_value(get_irn_dbg_info(load), get_atomic_ent_value(ent));
 740
 741                                                 DBG_OPT_RC(load, c);
 742                                                 exchange(info->projs[pn_Load_res], c);
 743                                                 res |= DF_CHANGED;
 744                                         }
 745                                 }
 746                                 exchange(load, new_Bad());
 747                                 reduce_adr_usage(ptr);
 748                                 return res;
 749                         } else if (variability_constant == get_entity_variability(ent)) {
 750                                 compound_graph_path *path = get_accessed_path(ptr);
 751
 752                                 if (path) {
 753                                         ir_node *c;
 754
 755                                         assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
 756                                         /*
 757                                         {
 758                                                 int j;
 759                                                 for (j = 0; j < get_compound_graph_path_length(path); ++j) {
 760                                                         ir_entity *node = get_compound_graph_path_node(path, j);
 761                                                         fprintf(stdout, ".%s", get_entity_name(node));
 762                                                         if (is_Array_type(get_entity_owner(node)))
 763                                                                 fprintf(stdout, "[%d]", get_compound_graph_path_array_index(path, j));
 764                                                 }
 765                                                 printf("\n");
 766                                         }
 767                                         */
 768
 769                                         c = get_compound_ent_value_by_path(ent, path);
 770                                         free_compound_graph_path(path);
 771
 772                                         /* printf("  cons: "); DDMN(c); */
 773
 774                                         if (info->projs[pn_Load_M]) {
 775                                                 exchange(info->projs[pn_Load_M], mem);
 776                                                 res |= DF_CHANGED;
 777                                         }
 778                                         if (info->projs[pn_Load_res]) {
 779                                                 exchange(info->projs[pn_Load_res], copy_const_value(get_irn_dbg_info(load), c));
 780                                                 res |= DF_CHANGED;
 781                                         }
 782                                         exchange(load, new_Bad());
 783                                         reduce_adr_usage(ptr);
 784                                         return res;
 785                                 } else {
 786                                         /*  We can not determine a correct access path.  E.g., in jack, we load
 787                                         a byte from an object to generate an exception.   Happens in test program
 788                                         Reflectiontest.
 789                                         printf(">>>>>>>>>>>>> Found access to constant entity %s in function %s\n", get_entity_name(ent),
 790                                         get_entity_name(get_irg_entity(current_ir_graph)));
 791                                         printf("  load: "); DDMN(load);
 792                                         printf("  ptr:  "); DDMN(ptr);
 793                                         */
 794                                 }
 795                         }
 796                 }
 797         }
 798
 799         /* Check, if the address of this load is used more than once.
 800          * If not, this load cannot be removed in any case. */
 801         if (get_irn_n_uses(ptr) <= 1)
 802                 return res;
 803
 804         /*
 805          * follow the memory chain as long as there are only Loads
 806          * and try to replace current Load or Store by a previous one.
 807          * Note that in unreachable loops it might happen that we reach
 808          * load again, as well as we can fall into a cycle.
 809          * We break such cycles using a special visited flag.
 810          */
 811         INC_MASTER();
 812         res = follow_Mem_chain(load, skip_Proj(mem));
 813         return res;
 814 }  /* optimize_load */
 815
 816 /**
 817  * Check whether a value of mode new_mode would completely overwrite a value
 818  * of mode old_mode in memory.
 819  */
 820 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
 821 {
 822         return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
 823 }  /* is_completely_overwritten */
 824
 825 /**
 826  * follow the memory chain as long as there are only Loads and alias free Stores.
 827  *
 828  * INC_MASTER() must be called before dive into
 829  */
 830 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
 831         unsigned res = 0;
 832         ldst_info_t *info = get_irn_link(store);
 833         ir_node *pred;
 834         ir_node *ptr = get_Store_ptr(store);
 835         ir_node *mem = get_Store_mem(store);
 836         ir_node *value = get_Store_value(store);
 837         ir_mode *mode  = get_irn_mode(value);
 838         ir_node *block = get_nodes_block(store);
 839
 840         for (pred = curr; pred != store;) {
 841                 ldst_info_t *pred_info = get_irn_link(pred);
 842
 843                 /*
 844                  * BEWARE: one might think that checking the modes is useless, because
 845                  * if the pointers are identical, they refer to the same object.
 846                  * This is only true in strong typed languages, not is C were the following
 847                  * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
 848                  * However, if the mode that is written have a bigger  or equal size the the old
 849                  * one, the old value is completely overwritten and can be killed ...
 850                  */
 851                 if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
 852                     get_nodes_block(pred) == block &&
 853                     is_completely_overwritten(get_irn_mode(get_Store_value(pred)), mode)) {
 854                         /*
 855                          * a Store after a Store in the same block -- a write after write.
 856                          * We may remove the first Store, if it does not have an exception handler.
 857                          *
 858                          * TODO: What, if both have the same exception handler ???
 859                          */
 860                         if (get_Store_volatility(pred) != volatility_is_volatile && !pred_info->projs[pn_Store_X_except]) {
 861                                 DBG_OPT_WAW(pred, store);
 862                                 exchange( pred_info->projs[pn_Store_M], get_Store_mem(pred) );
 863                                 exchange(pred, new_Bad());
 864                                 reduce_adr_usage(ptr);
 865                                 return DF_CHANGED;
 866                         }
 867                 } else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
 868                            value == pred_info->projs[pn_Load_res]) {
 869                         /*
 870                          * a Store of a value after a Load -- a write after read.
 871                          * We may remove the second Store, if it does not have an exception handler.
 872                          */
 873                         if (! info->projs[pn_Store_X_except]) {
 874                                 DBG_OPT_WAR(store, pred);
 875                                 exchange( info->projs[pn_Store_M], mem );
 876                                 exchange(store, new_Bad());
 877                                 reduce_adr_usage(ptr);
 878                                 return DF_CHANGED;
 879                         }
 880                 }
 881
 882                 if (get_irn_op(pred) == op_Store) {
 883                         /* check if we can pass thru this store */
 884                         ir_alias_relation rel = get_alias_relation(
 885                                 current_ir_graph,
 886                                 get_Store_ptr(pred),
 887                                 get_irn_mode(get_Store_value(pred)),
 888                                 ptr, mode);
 889                         /* if the might be an alias, we cannot pass this Store */
 890                         if (rel != no_alias)
 891                                 break;
 892                         pred = skip_Proj(get_Store_mem(pred));
 893                 } else if (get_irn_op(pred) == op_Load) {
 894                         pred = skip_Proj(get_Load_mem(pred));
 895                 } else {
 896                         /* follow only Load chains */
 897                         break;
 898                 }
 899
 900                 /* check for cycles */
 901                 if (NODE_VISITED(pred_info))
 902                         break;
 903                 MARK_NODE(pred_info);
 904         }
 905
 906         if (get_irn_op(pred) == op_Sync) {
 907                 int i;
 908
 909                 /* handle all Sync predecessors */
 910                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
 911                         res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
 912                         if (res)
 913                                 break;
 914                 }
 915         }
 916         return res;
 917 }  /* follow_Mem_chain_for_Store */
 918
 919 /**
 920  * optimize a Store
 921  *
 922  * @param store  the Store node
 923  */
 924 static unsigned optimize_store(ir_node *store) {
 925         ir_node *ptr, *mem;
 926
 927         if (get_Store_volatility(store) == volatility_is_volatile)
 928                 return 0;
 929
 930         ptr = get_Store_ptr(store);
 931
 932         /* Check, if the address of this Store is used more than once.
 933          * If not, this Store cannot be removed in any case. */
 934         if (get_irn_n_uses(ptr) <= 1)
 935                 return 0;
 936
 937         mem = get_Store_mem(store);
 938
 939         /* follow the memory chain as long as there are only Loads */
 940         INC_MASTER();
 941         return follow_Mem_chain_for_Store(store, skip_Proj(mem));
 942 }  /* optimize_store */
 943
 944 /**
 945  * walker, optimizes Phi after Stores to identical places:
 946  * Does the following optimization:
 947  * @verbatim
 948  *
 949  *   val1   val2   val3          val1  val2  val3
 950  *    |      |      |               \    |    /
 951  *  Store  Store  Store              \   |   /
 952  *      \    |    /                   PhiData
 953  *       \   |   /                       |
 954  *        \  |  /                      Store
 955  *          PhiM
 956  *
 957  * @endverbatim
 958  * This reduces the number of stores and allows for predicated execution.
 959  * Moves Stores back to the end of a function which may be bad.
 960  *
 961  * This is only possible if the predecessor blocks have only one successor.
 962  */
 963 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
 964 {
 965         int i, n;
 966         ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
 967         ir_mode *mode;
 968         ir_node **inM, **inD, **stores;
 969         int *idx;
 970         dbg_info *db = NULL;
 971         ldst_info_t *info;
 972         block_info_t *bl_info;
 973         unsigned res = 0;
 974
 975         /* Must be a memory Phi */
 976         if (get_irn_mode(phi) != mode_M)
 977                 return 0;
 978
 979         n = get_Phi_n_preds(phi);
 980         if (n <= 0)
 981                 return 0;
 982
 983         store = skip_Proj(get_Phi_pred(phi, 0));
 984         old_store = store;
 985         if (get_irn_op(store) != op_Store)
 986                 return 0;
 987
 988         block = get_nodes_block(store);
 989
 990         /* abort on dead blocks */
 991         if (is_Block_dead(block))
 992                 return 0;
 993
 994         /* check if the block is post dominated by Phi-block
 995            and has no exception exit */
 996         bl_info = get_irn_link(block);
 997         if (bl_info->flags & BLOCK_HAS_EXC)
 998                 return 0;
 999
1000         phi_block = get_nodes_block(phi);
1001         if (! block_postdominates(phi_block, block))
1002                 return 0;
1003
1004         /* this is the address of the store */
1005         ptr  = get_Store_ptr(store);
1006         mode = get_irn_mode(get_Store_value(store));
1007         info = get_irn_link(store);
1008         exc  = info->exc_block;
1009
1010         for (i = 1; i < n; ++i) {
1011                 ir_node *pred = skip_Proj(get_Phi_pred(phi, i));
1012
1013                 if (get_irn_op(pred) != op_Store)
1014                         return 0;
1015
1016                 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1017                         return 0;
1018
1019                 info = get_irn_link(pred);
1020
1021                 /* check, if all stores have the same exception flow */
1022                 if (exc != info->exc_block)
1023                         return 0;
1024
1025                 /* abort on dead blocks */
1026                 block = get_nodes_block(pred);
1027                 if (is_Block_dead(block))
1028                         return 0;
1029
1030                 /* check if the block is post dominated by Phi-block
1031                    and has no exception exit. Note that block must be different from
1032                    Phi-block, else we would move a Store from end End of a block to its
1033                    Start... */
1034                 bl_info = get_irn_link(block);
1035                 if (bl_info->flags & BLOCK_HAS_EXC)
1036                         return 0;
1037                 if (block == phi_block || ! block_postdominates(phi_block, block))
1038                         return 0;
1039         }
1040
1041         /*
1042          * ok, when we are here, we found all predecessors of a Phi that
1043          * are Stores to the same address and size. That means whatever
1044          * we do before we enter the block of the Phi, we do a Store.
1045          * So, we can move the Store to the current block:
1046          *
1047          *   val1    val2    val3          val1  val2  val3
1048          *    |       |       |               \    |    /
1049          * | Str | | Str | | Str |             \   |   /
1050          *      \     |     /                   PhiData
1051          *       \    |    /                       |
1052          *        \   |   /                       Str
1053          *           PhiM
1054          *
1055          * Is only allowed if the predecessor blocks have only one successor.
1056          */
1057
1058         NEW_ARR_A(ir_node *, stores, n);
1059         NEW_ARR_A(ir_node *, inM, n);
1060         NEW_ARR_A(ir_node *, inD, n);
1061         NEW_ARR_A(int, idx, n);
1062
1063         /* Prepare: Collect all Store nodes.  We must do this
1064            first because we otherwise may loose a store when exchanging its
1065            memory Proj.
1066          */
1067         for (i = 0; i < n; ++i)
1068                 stores[i] = skip_Proj(get_Phi_pred(phi, i));
1069
1070         /* Prepare: Skip the memory Proj: we need this in the case some stores
1071            are cascaded.
1072            Beware: One Store might be included more than once in the stores[]
1073            list, so we must prevent to do the exchange more than once.
1074          */
1075         for (i = 0; i < n; ++i) {
1076                 ir_node *store = stores[i];
1077                 ir_node *proj_m;
1078
1079                 info = get_irn_link(store);
1080                 proj_m = info->projs[pn_Store_M];
1081
1082                 if (is_Proj(proj_m) && get_Proj_pred(proj_m) == store)
1083                         exchange(proj_m, get_Store_mem(store));
1084         }
1085
1086         /* first step: collect all inputs */
1087         for (i = 0; i < n; ++i) {
1088                 ir_node *store = stores[i];
1089                 info = get_irn_link(store);
1090
1091                 inM[i] = get_Store_mem(store);
1092                 inD[i] = get_Store_value(store);
1093                 idx[i] = info->exc_idx;
1094         }
1095         block = get_nodes_block(phi);
1096
1097         /* second step: create a new memory Phi */
1098         phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M);
1099
1100         /* third step: create a new data Phi */
1101         phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode);
1102
1103         /* fourth step: create the Store */
1104         store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD);
1105 #ifdef DO_CACHEOPT
1106         co_set_irn_name(store, co_get_irn_ident(old_store));
1107 #endif
1108
1109         projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
1110
1111         info = get_ldst_info(store, wenv);
1112         info->projs[pn_Store_M] = projM;
1113
1114         /* fifths step: repair exception flow */
1115         if (exc) {
1116                 ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except);
1117
1118                 info->projs[pn_Store_X_except] = projX;
1119                 info->exc_block                = exc;
1120                 info->exc_idx                  = idx[0];
1121
1122                 for (i = 0; i < n; ++i) {
1123                         set_Block_cfgpred(exc, idx[i], projX);
1124                 }
1125
1126                 if (n > 1) {
1127                         /* the exception block should be optimized as some inputs are identical now */
1128                 }
1129
1130                 res |= CF_CHANGED;
1131         }
1132
1133         /* sixth step: replace old Phi */
1134         exchange(phi, projM);
1135
1136         return res | DF_CHANGED;
1137 }  /* optimize_phi */
1138
1139 /**
1140  * walker, do the optimizations
1141  */
1142 static void do_load_store_optimize(ir_node *n, void *env) {
1143         walk_env_t *wenv = env;
1144
1145         switch (get_irn_opcode(n)) {
1146
1147         case iro_Load:
1148                 wenv->changes |= optimize_load(n);
1149                 break;
1150
1151         case iro_Store:
1152                 wenv->changes |= optimize_store(n);
1153                 break;
1154
1155         case iro_Phi:
1156                 wenv->changes |= optimize_phi(n, wenv);
1157
1158         default:
1159                 ;
1160         }
1161 }  /* do_load_store_optimize */
1162
1163 /*
1164  * do the load store optimization
1165  */
1166 void optimize_load_store(ir_graph *irg) {
1167         walk_env_t env;
1168
1169         assert(get_irg_phase_state(irg) != phase_building);
1170         assert(get_irg_pinned(irg) != op_pin_state_floats &&
1171                 "LoadStore optimization needs pinned graph");
1172
1173         if (! get_opt_redundant_loadstore())
1174                 return;
1175
1176         edges_assure(irg);
1177
1178         /* for Phi optimization post-dominators are needed ... */
1179         assure_postdoms(irg);
1180
1181         if (get_opt_alias_analysis()) {
1182                 assure_irg_address_taken_computed(irg);
1183                 assure_irp_globals_address_taken_computed();
1184         }
1185
1186         obstack_init(&env.obst);
1187         env.changes = 0;
1188
1189         /* init the links, then collect Loads/Stores/Proj's in lists */
1190         master_visited = 0;
1191         irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
1192
1193         /* now we have collected enough information, optimize */
1194         irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
1195
1196         obstack_free(&env.obst, NULL);
1197
1198         /* Handle graph state */
1199         if (env.changes) {
1200                 if (get_irg_outs_state(irg) == outs_consistent)
1201                         set_irg_outs_inconsistent(irg);
1202         }
1203
1204         if (env.changes & CF_CHANGED) {
1205                 /* is this really needed: Yes, control flow changed, block might
1206                 have Bad() predecessors. */
1207                 set_irg_doms_inconsistent(irg);
1208         }
1209 }  /* optimize_load_store */