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_ALLOCA_H
  16 #include <alloca.h>
  17 #endif
  18 #ifdef HAVE_MALLOC_H
  19 #include <malloc.h>
  20 #endif
  21 #ifdef HAVE_STRING_H
  22 # include <string.h>
  23 #endif
  24
  25 #include "irnode_t.h"
  26 #include "irgraph_t.h"
  27 #include "irmode_t.h"
  28 #include "iropt_t.h"
  29 #include "ircons_t.h"
  30 #include "irgmod.h"
  31 #include "irgwalk.h"
  32 #include "irvrfy.h"
  33 #include "tv_t.h"
  34 #include "dbginfo_t.h"
  35 #include "iropt_dbg.h"
  36 #include "irflag_t.h"
  37 #include "array.h"
  38 #include "irhooks.h"
  39 #include "iredges.h"
  40 #include "irtools.h"
  41 #include "opt_polymorphy.h"
  42 #include "irmemory.h"
  43
  44 #ifdef DO_CACHEOPT
  45 #include "cacheopt/cachesim.h"
  46 #endif
  47
  48 #undef IMAX
  49 #define IMAX(a,b)       ((a) > (b) ? (a) : (b))
  50
  51 #define MAX_PROJ        IMAX(pn_Load_max, pn_Store_max)
  52
  53 enum changes_t {
  54         DF_CHANGED = 1,       /**< data flow changed */
  55         CF_CHANGED = 2,       /**< control flow changed */
  56 };
  57
  58 /**
  59  * walker environment
  60  */
  61 typedef struct _walk_env_t {
  62         struct obstack obst;          /**< list of all stores */
  63         unsigned changes;             /**< a bitmask of graph changes */
  64 } walk_env_t;
  65
  66 /**
  67  * flags for Load/Store
  68  */
  69 enum ldst_flags_t {
  70         LDST_VISITED = 1              /**< if set, this Load/Store is already visited */
  71 };
  72
  73 /** A Load/Store info. */
  74 typedef struct _ldst_info_t {
  75         ir_node  *projs[MAX_PROJ];    /**< list of Proj's of this node */
  76         ir_node  *exc_block;          /**< the exception block if available */
  77         int      exc_idx;             /**< predecessor index in the exception block */
  78         unsigned flags;               /**< flags */
  79         unsigned visited;             /**< visited counter for breaking loops */
  80 } ldst_info_t;
  81
  82 /**
  83  * flags for control flow.
  84  */
  85 enum block_flags_t {
  86         BLOCK_HAS_COND = 1,      /**< Block has conditional control flow */
  87         BLOCK_HAS_EXC  = 2       /**< Block has exceptional control flow */
  88 };
  89
  90 /**
  91  * a Block info.
  92  */
  93 typedef struct _block_info_t {
  94         unsigned flags;               /**< flags for the block */
  95 } block_info_t;
  96
  97 /** the master visited flag for loop detection. */
  98 static unsigned master_visited = 0;
  99
 100 #define INC_MASTER()       ++master_visited
 101 #define MARK_NODE(info)    (info)->visited = master_visited
 102 #define NODE_VISITED(info) (info)->visited >= master_visited
 103
 104 /**
 105  * get the Load/Store info of a node
 106  */
 107 static ldst_info_t *get_ldst_info(ir_node *node, walk_env_t *env) {
 108         ldst_info_t *info = get_irn_link(node);
 109
 110         if (! info) {
 111                 info = obstack_alloc(&env->obst, sizeof(*info));
 112                 memset(info, 0, sizeof(*info));
 113                 set_irn_link(node, info);
 114         }
 115         return info;
 116 }  /* get_ldst_info */
 117
 118 /**
 119  * get the Block info of a node
 120  */
 121 static block_info_t *get_block_info(ir_node *node, walk_env_t *env) {
 122         block_info_t *info = get_irn_link(node);
 123
 124         if (! info) {
 125                 info = obstack_alloc(&env->obst, sizeof(*info));
 126                 memset(info, 0, sizeof(*info));
 127                 set_irn_link(node, info);
 128         }
 129         return info;
 130 }  /* get_block_info */
 131
 132 /**
 133  * update the projection info for a Load/Store
 134  */
 135 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
 136 {
 137         long nr = get_Proj_proj(proj);
 138
 139         assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
 140
 141         if (info->projs[nr]) {
 142                 /* there is already one, do CSE */
 143                 exchange(proj, info->projs[nr]);
 144                 return DF_CHANGED;
 145         }
 146         else {
 147                 info->projs[nr] = proj;
 148                 return 0;
 149         }
 150 }  /* update_projs */
 151
 152 /**
 153  * update the exception block info for a Load/Store node.
 154  *
 155  * @param info   the load/store info struct
 156  * @param block  the exception handler block for this load/store
 157  * @param pos    the control flow input of the block
 158  */
 159 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
 160 {
 161         assert(info->exc_block == NULL && "more than one exception block found");
 162
 163         info->exc_block = block;
 164         info->exc_idx   = pos;
 165         return 0;
 166 }  /* update_exc */
 167
 168 /** Return the number of uses of an address node */
 169 #define get_irn_n_uses(adr)     get_irn_n_edges(adr)
 170
 171 /**
 172  * walker, collects all Load/Store/Proj nodes
 173  *
 174  * walks from Start -> End
 175  */
 176 static void collect_nodes(ir_node *node, void *env)
 177 {
 178         ir_op       *op = get_irn_op(node);
 179         ir_node     *pred, *blk, *pred_blk;
 180         ldst_info_t *ldst_info;
 181         walk_env_t  *wenv = env;
 182
 183         if (op == op_Proj) {
 184                 ir_node *adr;
 185                 ir_op *op;
 186
 187                 pred = get_Proj_pred(node);
 188                 op   = get_irn_op(pred);
 189
 190                 if (op == op_Load) {
 191                         ldst_info = get_ldst_info(pred, wenv);
 192
 193                         wenv->changes |= update_projs(ldst_info, node);
 194
 195                         if ((ldst_info->flags & LDST_VISITED) == 0) {
 196                                 adr = get_Load_ptr(pred);
 197                                 ldst_info->flags |= LDST_VISITED;
 198                         }
 199
 200                         /*
 201                         * Place the Proj's to the same block as the
 202                         * predecessor Load. This is always ok and prevents
 203                         * "non-SSA" form after optimizations if the Proj
 204                         * is in a wrong block.
 205                         */
 206                         blk      = get_nodes_block(node);
 207                         pred_blk = get_nodes_block(pred);
 208                         if (blk != pred_blk) {
 209                                 wenv->changes |= DF_CHANGED;
 210                                 set_nodes_block(node, pred_blk);
 211                         }
 212                 } else if (op == op_Store) {
 213                         ldst_info = get_ldst_info(pred, wenv);
 214
 215                         wenv->changes |= update_projs(ldst_info, node);
 216
 217                         if ((ldst_info->flags & LDST_VISITED) == 0) {
 218                                 adr = get_Store_ptr(pred);
 219                                 ldst_info->flags |= LDST_VISITED;
 220                         }
 221
 222                         /*
 223                         * Place the Proj's to the same block as the
 224                         * predecessor Store. This is always ok and prevents
 225                         * "non-SSA" form after optimizations if the Proj
 226                         * is in a wrong block.
 227                         */
 228                         blk      = get_nodes_block(node);
 229                         pred_blk = get_nodes_block(pred);
 230                         if (blk != pred_blk) {
 231                                 wenv->changes |= DF_CHANGED;
 232                                 set_nodes_block(node, pred_blk);
 233                         }
 234                 }
 235         } else if (op == op_Block) {
 236                 int i;
 237
 238                 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
 239                         ir_node      *pred_block;
 240                         block_info_t *bl_info;
 241
 242                         pred = skip_Proj(get_Block_cfgpred(node, i));
 243
 244                         /* ignore Bad predecessors, they will be removed later */
 245                         if (is_Bad(pred))
 246                                 continue;
 247
 248                         pred_block = get_nodes_block(pred);
 249                         bl_info    = get_block_info(pred_block, wenv);
 250
 251                         if (is_fragile_op(pred))
 252                                 bl_info->flags |= BLOCK_HAS_EXC;
 253                         else if (is_irn_forking(pred))
 254                                 bl_info->flags |= BLOCK_HAS_COND;
 255
 256                         if (get_irn_op(pred) == op_Load || get_irn_op(pred) == op_Store) {
 257                                 ldst_info = get_ldst_info(pred, wenv);
 258
 259                                 wenv->changes |= update_exc(ldst_info, node, i);
 260                         }
 261                 }
 262         }
 263 }  /* collect_nodes */
 264
 265 /**
 266  * Returns an entity if the address ptr points to a constant one.
 267  *
 268  * @param ptr  the address
 269  *
 270  * @return an entity or NULL
 271  */
 272 static ir_entity *find_constant_entity(ir_node *ptr)
 273 {
 274         for (;;) {
 275                 ir_op *op = get_irn_op(ptr);
 276
 277                 if (op == op_SymConst && (get_SymConst_kind(ptr) == symconst_addr_ent)) {
 278                         ir_entity *ent = get_SymConst_entity(ptr);
 279                         if (variability_constant == get_entity_variability(ent))
 280                                 return ent;
 281                         return NULL;
 282                 } else if (op == op_Sel) {
 283                         ir_entity *ent = get_Sel_entity(ptr);
 284                         ir_type   *tp  = get_entity_owner(ent);
 285
 286                         /* Do not fiddle with polymorphism. */
 287                         if (is_Class_type(get_entity_owner(ent)) &&
 288                                 ((get_entity_n_overwrites(ent)    != 0) ||
 289                                 (get_entity_n_overwrittenby(ent) != 0)   ) )
 290                                 return NULL;
 291
 292                         if (is_Array_type(tp)) {
 293                                 /* check bounds */
 294                                 int i, n;
 295
 296                                 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
 297                                         ir_node *bound;
 298                                         tarval *tlower, *tupper;
 299                                         ir_node *index = get_Sel_index(ptr, i);
 300                                         tarval *tv     = computed_value(index);
 301
 302                                         /* check if the index is constant */
 303                                         if (tv == tarval_bad)
 304                                                 return NULL;
 305
 306                                         bound  = get_array_lower_bound(tp, i);
 307                                         tlower = computed_value(bound);
 308                                         bound  = get_array_upper_bound(tp, i);
 309                                         tupper = computed_value(bound);
 310
 311                                         if (tlower == tarval_bad || tupper == tarval_bad)
 312                                                 return NULL;
 313
 314                                         if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
 315                                                 return NULL;
 316                                         if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
 317                                                 return NULL;
 318
 319                                         /* ok, bounds check finished */
 320                                 }
 321                         }
 322
 323                         if (variability_constant == get_entity_variability(ent))
 324                                 return ent;
 325
 326                         /* try next */
 327                         ptr = get_Sel_ptr(ptr);
 328                 } else
 329                         return NULL;
 330         }
 331 }  /* find_constant_entity */
 332
 333 /**
 334  * Return the Selection index of a Sel node from dimension n
 335  */
 336 static long get_Sel_array_index_long(ir_node *n, int dim) {
 337         ir_node *index = get_Sel_index(n, dim);
 338         assert(get_irn_op(index) == op_Const);
 339         return get_tarval_long(get_Const_tarval(index));
 340 }  /* get_Sel_array_index_long */
 341
 342 /**
 343  * Returns the accessed component graph path for an
 344  * node computing an address.
 345  *
 346  * @param ptr    the node computing the address
 347  * @param depth  current depth in steps upward from the root
 348  *               of the address
 349  */
 350 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
 351         compound_graph_path *res = NULL;
 352         ir_entity           *root, *field;
 353         int                 path_len, pos;
 354
 355         if (get_irn_op(ptr) == op_SymConst) {
 356                 /* a SymConst. If the depth is 0, this is an access to a global
 357                  * entity and we don't need a component path, else we know
 358                  * at least it's length.
 359                  */
 360                 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
 361                 root = get_SymConst_entity(ptr);
 362                 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
 363         } else {
 364                 assert(get_irn_op(ptr) == op_Sel);
 365                 /* it's a Sel, go up until we find the root */
 366                 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
 367
 368                 /* fill up the step in the path at the current position */
 369                 field    = get_Sel_entity(ptr);
 370                 path_len = get_compound_graph_path_length(res);
 371                 pos      = path_len - depth - 1;
 372                 set_compound_graph_path_node(res, pos, field);
 373
 374                 if (is_Array_type(get_entity_owner(field))) {
 375                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 376                         set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
 377                 }
 378         }
 379         return res;
 380 }  /* rec_get_accessed_path */
 381
 382 /** Returns an access path or NULL.  The access path is only
 383  *  valid, if the graph is in phase_high and _no_ address computation is used.
 384  */
 385 static compound_graph_path *get_accessed_path(ir_node *ptr) {
 386         return rec_get_accessed_path(ptr, 0);
 387 }  /* get_accessed_path */
 388
 389 /* forward */
 390 static void reduce_adr_usage(ir_node *ptr);
 391
 392 /**
 393  * Update a Load that may lost it's usage.
 394  */
 395 static void handle_load_update(ir_node *load) {
 396         ldst_info_t *info = get_irn_link(load);
 397
 398         /* do NOT touch volatile loads for now */
 399         if (get_Load_volatility(load) == volatility_is_volatile)
 400                 return;
 401
 402         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 403                 ir_node *ptr = get_Load_ptr(load);
 404                 ir_node *mem = get_Load_mem(load);
 405
 406                 /* a Load which value is neither used nor exception checked, remove it */
 407                 exchange(info->projs[pn_Load_M], mem);
 408                 exchange(load, new_Bad());
 409                 reduce_adr_usage(ptr);
 410         }
 411 }  /* handle_load_update */
 412
 413 /**
 414  * A Use of an address node is vanished. Check if this was a Proj
 415  * node and update the counters.
 416  */
 417 static void reduce_adr_usage(ir_node *ptr) {
 418         if (is_Proj(ptr)) {
 419                 if (get_irn_n_edges(ptr) <= 0) {
 420                         /* this Proj is dead now */
 421                         ir_node *pred = get_Proj_pred(ptr);
 422
 423                         if (is_Load(pred)) {
 424                                 ldst_info_t *info = get_irn_link(pred);
 425                                 info->projs[get_Proj_proj(ptr)] = NULL;
 426
 427                                 /* this node lost it's result proj, handle that */
 428                                 handle_load_update(pred);
 429                         }
 430                 }
 431         }
 432 }  /* reduce_adr_usage */
 433
 434 /**
 435  * Check, if an already existing value of mode old_mode can be converted
 436  * into the needed one new_mode without loss.
 437  */
 438 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
 439         if (old_mode == new_mode)
 440                 return 1;
 441
 442         /* if both modes are two-complement ones, we can always convert the
 443            Stored value into the needed one. */
 444         if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
 445                   get_mode_arithmetic(old_mode) == irma_twos_complement &&
 446                   get_mode_arithmetic(new_mode) == irma_twos_complement)
 447                 return 1;
 448         return 0;
 449 }  /* can_use_stored_value */
 450
 451 /**
 452  * Follow the memory chain as long as there are only Loads
 453  * and alias free Stores and try to replace current Load or Store
 454  * by a previous ones.
 455  * Note that in unreachable loops it might happen that we reach
 456  * load again, as well as we can fall into a cycle.
 457  * We break such cycles using a special visited flag.
 458  *
 459  * INC_MASTER() must be called before dive into
 460  */
 461 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
 462         unsigned res = 0;
 463         ldst_info_t *info = get_irn_link(load);
 464         ir_node *pred;
 465         ir_node *ptr       = get_Load_ptr(load);
 466         ir_node *mem       = get_Load_mem(load);
 467         ir_mode *load_mode = get_Load_mode(load);
 468
 469         for (pred = curr; load != pred; ) {
 470                 ldst_info_t *pred_info = get_irn_link(pred);
 471
 472                 /*
 473                  * BEWARE: one might think that checking the modes is useless, because
 474                  * if the pointers are identical, they refer to the same object.
 475                  * This is only true in strong typed languages, not in C were the following
 476                  * is possible a = *(ir_type1 *)p; b = *(ir_type2 *)p ...
 477                  */
 478                 if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
 479                     can_use_stored_value(get_irn_mode(get_Store_value(pred)), load_mode)) {
 480                         /*
 481                          * a Load immediately after a Store -- a read after write.
 482                          * We may remove the Load, if both Load & Store does not have an exception handler
 483                          * OR they are in the same block. In the latter case the Load cannot
 484                          * throw an exception when the previous Store was quiet.
 485                          *
 486                          * Why we need to check for Store Exception? If the Store cannot
 487                          * be executed (ROM) the exception handler might simply jump into
 488                          * the load block :-(
 489                          * We could make it a little bit better if we would know that the exception
 490                          * handler of the Store jumps directly to the end...
 491                          */
 492                         if ((!pred_info->projs[pn_Store_X_except] && !info->projs[pn_Load_X_except]) ||
 493                             get_nodes_block(load) == get_nodes_block(pred)) {
 494                                 ir_node *value = get_Store_value(pred);
 495
 496                                 DBG_OPT_RAW(load, value);
 497
 498                                 /* add an convert if needed */
 499                                 if (get_irn_mode(get_Store_value(pred)) != load_mode) {
 500                                         value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
 501                                 }
 502
 503                                 if (info->projs[pn_Load_M])
 504                                         exchange(info->projs[pn_Load_M], mem);
 505
 506                                 /* no exception */
 507                                 if (info->projs[pn_Load_X_except]) {
 508                                         exchange( info->projs[pn_Load_X_except], new_Bad());
 509                                         res |= CF_CHANGED;
 510                                 }
 511
 512                                 if (info->projs[pn_Load_res])
 513                                         exchange(info->projs[pn_Load_res], value);
 514
 515                                 exchange(load, new_Bad());
 516                                 reduce_adr_usage(ptr);
 517                                 return res | DF_CHANGED;
 518                         }
 519                 } else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
 520                            can_use_stored_value(get_Load_mode(pred), load_mode)) {
 521                         /*
 522                          * a Load after a Load -- a read after read.
 523                          * We may remove the second Load, if it does not have an exception handler
 524                          * OR they are in the same block. In the later case the Load cannot
 525                          * throw an exception when the previous Load was quiet.
 526                          *
 527                          * Here, there is no need to check if the previous Load has an exception
 528                          * hander because they would have exact the same exception...
 529                          */
 530                         if (! info->projs[pn_Load_X_except] || get_nodes_block(load) == get_nodes_block(pred)) {
 531                                 DBG_OPT_RAR(load, pred);
 532
 533                                 if (pred_info->projs[pn_Load_res]) {
 534                                         ir_node *value = pred_info->projs[pn_Load_res];
 535
 536                                         /* add an convert if needed */
 537                                         if (get_Load_mode(pred) != load_mode) {
 538                                                 value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
 539                                         }
 540
 541                                         /* we need a data proj from the previous load for this optimization */
 542                                         if (info->projs[pn_Load_res])
 543                                                 exchange(info->projs[pn_Load_res], pred_info->projs[pn_Load_res]);
 544
 545                                         if (info->projs[pn_Load_M])
 546                                                 exchange(info->projs[pn_Load_M], mem);
 547                                 } else {
 548                                         if (info->projs[pn_Load_res]) {
 549                                                 set_Proj_pred(info->projs[pn_Load_res], pred);
 550                                                 set_nodes_block(info->projs[pn_Load_res], get_nodes_block(pred));
 551                                                 pred_info->projs[pn_Load_res] = info->projs[pn_Load_res];
 552                                         }
 553                                         if (info->projs[pn_Load_M]) {
 554                                                 /* Actually, this if should not be necessary.  Construct the Loads
 555                                                 properly!!! */
 556                                                 exchange(info->projs[pn_Load_M], mem);
 557                                         }
 558                                 }
 559
 560                                 /* no exception */
 561                                 if (info->projs[pn_Load_X_except]) {
 562                                         exchange(info->projs[pn_Load_X_except], new_Bad());
 563                                         res |= CF_CHANGED;
 564                                 }
 565
 566                                 exchange(load, new_Bad());
 567                                 reduce_adr_usage(ptr);
 568                                 return res |= DF_CHANGED;
 569                         }
 570                 }
 571
 572                 if (get_irn_op(pred) == op_Store) {
 573                         /* check if we can pass thru this store */
 574                         ir_alias_relation rel = get_alias_relation(
 575                                 current_ir_graph,
 576                                 get_Store_ptr(pred),
 577                                 get_irn_mode(get_Store_value(pred)),
 578                                 ptr, load_mode);
 579                         /* if the might be an alias, we cannot pass this Store */
 580                         if (rel != no_alias)
 581                                 break;
 582                         pred = skip_Proj(get_Store_mem(pred));
 583                 } else if (get_irn_op(pred) == op_Load) {
 584                         pred = skip_Proj(get_Load_mem(pred));
 585                 } else {
 586                         /* follow only Load chains */
 587                         break;
 588                 }
 589
 590                 /* check for cycles */
 591                 if (NODE_VISITED(pred_info))
 592                         break;
 593                 MARK_NODE(pred_info);
 594         }
 595
 596         if (get_irn_op(pred) == op_Sync) {
 597                 int i;
 598
 599                 /* handle all Sync predecessors */
 600                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
 601                         res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
 602                         if (res)
 603                                 break;
 604                 }
 605         }
 606
 607         return res;
 608 }  /* follow_Mem_chain */
 609
 610 /**
 611  * optimize a Load
 612  *
 613  * @param load  the Load node
 614  */
 615 static unsigned optimize_load(ir_node *load)
 616 {
 617         ldst_info_t *info = get_irn_link(load);
 618         ir_node *mem, *ptr, *new_node;
 619         ir_entity *ent;
 620         unsigned res = 0;
 621
 622         /* do NOT touch volatile loads for now */
 623         if (get_Load_volatility(load) == volatility_is_volatile)
 624                 return 0;
 625
 626         /* the address of the load to be optimized */
 627         ptr = get_Load_ptr(load);
 628
 629         /*
 630          * Check if we can remove the exception from a Load:
 631          * This can be done, if the address is from an Sel(Alloc) and
 632          * the Sel type is a subtype of the allocated type.
 633          *
 634          * This optimizes some often used OO constructs,
 635          * like x = new O; x->t;
 636          */
 637         if (info->projs[pn_Load_X_except]) {
 638                 if (is_Sel(ptr)) {
 639                         ir_node *mem = get_Sel_mem(ptr);
 640
 641                         /* FIXME: works with the current FE, but better use the base */
 642                         if (get_irn_op(skip_Proj(mem)) == op_Alloc) {
 643                                 /* ok, check the types */
 644                                 ir_entity *ent    = get_Sel_entity(ptr);
 645                                 ir_type   *s_type = get_entity_type(ent);
 646                                 ir_type   *a_type = get_Alloc_type(mem);
 647
 648                                 if (is_SubClass_of(s_type, a_type)) {
 649                                         /* ok, condition met: there can't be an exception because
 650                                         * Alloc guarantees that enough memory was allocated */
 651
 652                                         exchange(info->projs[pn_Load_X_except], new_Bad());
 653                                         info->projs[pn_Load_X_except] = NULL;
 654                                         res |= CF_CHANGED;
 655                                 }
 656                         }
 657                 } else if ((get_irn_op(skip_Proj(ptr)) == op_Alloc) ||
 658                         ((get_irn_op(ptr) == op_Cast) && (get_irn_op(skip_Proj(get_Cast_op(ptr))) == op_Alloc))) {
 659                                 /* simple case: a direct load after an Alloc. Firm Alloc throw
 660                                  * an exception in case of out-of-memory. So, there is no way for an
 661                                  * exception in this load.
 662                                  * This code is constructed by the "exception lowering" in the Jack compiler.
 663                                  */
 664                                 exchange(info->projs[pn_Load_X_except], new_Bad());
 665                                 info->projs[pn_Load_X_except] = NULL;
 666                                 res |= CF_CHANGED;
 667                 }
 668         }
 669
 670         /* The mem of the Load. Must still be returned after optimization. */
 671         mem  = get_Load_mem(load);
 672
 673         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 674                 /* a Load which value is neither used nor exception checked, remove it */
 675                 exchange(info->projs[pn_Load_M], mem);
 676
 677                 exchange(load, new_Bad());
 678                 reduce_adr_usage(ptr);
 679                 return res | DF_CHANGED;
 680         }
 681
 682         /* Load from a constant polymorphic field, where we can resolve
 683            polymorphism. */
 684         new_node = transform_node_Load(load);
 685         if (new_node != load) {
 686                 if (info->projs[pn_Load_M]) {
 687                         exchange(info->projs[pn_Load_M], mem);
 688                         info->projs[pn_Load_M] = NULL;
 689                 }
 690                 if (info->projs[pn_Load_X_except]) {
 691                         exchange(info->projs[pn_Load_X_except], new_Bad());
 692                         info->projs[pn_Load_X_except] = NULL;
 693                 }
 694                 if (info->projs[pn_Load_res])
 695                         exchange(info->projs[pn_Load_res], new_node);
 696
 697                 exchange(load, new_Bad());
 698                 reduce_adr_usage(ptr);
 699                 return res | DF_CHANGED;
 700         }
 701
 702         /* check if we can determine the entity that will be loaded */
 703         ent = find_constant_entity(ptr);
 704         if (ent) {
 705                 if ((allocation_static == get_entity_allocation(ent)) &&
 706                         (visibility_external_allocated != get_entity_visibility(ent))) {
 707                         /* a static allocation that is not external: there should be NO exception
 708                          * when loading. */
 709
 710                         /* no exception, clear the info field as it might be checked later again */
 711                         if (info->projs[pn_Load_X_except]) {
 712                                 exchange(info->projs[pn_Load_X_except], new_Bad());
 713                                 info->projs[pn_Load_X_except] = NULL;
 714                                 res |= CF_CHANGED;
 715                         }
 716
 717                         if (variability_constant == get_entity_variability(ent)
 718                                 && is_atomic_entity(ent)) {
 719                                 /* Might not be atomic after
 720                                    lowering of Sels.  In this
 721                                    case we could also load, but
 722                                    it's more complicated. */
 723                                 /* more simpler case: we load the content of a constant value:
 724                                  * replace it by the constant itself
 725                                  */
 726
 727                                 /* no memory */
 728                                 if (info->projs[pn_Load_M]) {
 729                                         exchange(info->projs[pn_Load_M], mem);
 730                                         res |= DF_CHANGED;
 731                                 }
 732                                 /* no result :-) */
 733                                 if (info->projs[pn_Load_res]) {
 734                                         if (is_atomic_entity(ent)) {
 735                                                 ir_node *c = copy_const_value(get_irn_dbg_info(load), get_atomic_ent_value(ent));
 736
 737                                                 DBG_OPT_RC(load, c);
 738                                                 exchange(info->projs[pn_Load_res], c);
 739                                                 res |= DF_CHANGED;
 740                                         }
 741                                 }
 742                                 exchange(load, new_Bad());
 743                                 reduce_adr_usage(ptr);
 744                                 return res;
 745                         } else if (variability_constant == get_entity_variability(ent)) {
 746                                 compound_graph_path *path = get_accessed_path(ptr);
 747
 748                                 if (path) {
 749                                         ir_node *c;
 750
 751                                         assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
 752                                         /*
 753                                         {
 754                                                 int j;
 755                                                 for (j = 0; j < get_compound_graph_path_length(path); ++j) {
 756                                                         ir_entity *node = get_compound_graph_path_node(path, j);
 757                                                         fprintf(stdout, ".%s", get_entity_name(node));
 758                                                         if (is_Array_type(get_entity_owner(node)))
 759                                                                 fprintf(stdout, "[%d]", get_compound_graph_path_array_index(path, j));
 760                                                 }
 761                                                 printf("\n");
 762                                         }
 763                                         */
 764
 765                                         c = get_compound_ent_value_by_path(ent, path);
 766                                         free_compound_graph_path(path);
 767
 768                                         /* printf("  cons: "); DDMN(c); */
 769
 770                                         if (info->projs[pn_Load_M]) {
 771                                                 exchange(info->projs[pn_Load_M], mem);
 772                                                 res |= DF_CHANGED;
 773                                         }
 774                                         if (info->projs[pn_Load_res]) {
 775                                                 exchange(info->projs[pn_Load_res], copy_const_value(get_irn_dbg_info(load), c));
 776                                                 res |= DF_CHANGED;
 777                                         }
 778                                         exchange(load, new_Bad());
 779                                         reduce_adr_usage(ptr);
 780                                         return res;
 781                                 } else {
 782                                         /*  We can not determine a correct access path.  E.g., in jack, we load
 783                                         a byte from an object to generate an exception.   Happens in test program
 784                                         Reflectiontest.
 785                                         printf(">>>>>>>>>>>>> Found access to constant entity %s in function %s\n", get_entity_name(ent),
 786                                         get_entity_name(get_irg_entity(current_ir_graph)));
 787                                         printf("  load: "); DDMN(load);
 788                                         printf("  ptr:  "); DDMN(ptr);
 789                                         */
 790                                 }
 791                         }
 792                 }
 793         }
 794
 795         /* Check, if the address of this load is used more than once.
 796          * If not, this load cannot be removed in any case. */
 797         if (get_irn_n_uses(ptr) <= 1)
 798                 return res;
 799
 800         /*
 801          * follow the memory chain as long as there are only Loads
 802          * and try to replace current Load or Store by a previous one.
 803          * Note that in unreachable loops it might happen that we reach
 804          * load again, as well as we can fall into a cycle.
 805          * We break such cycles using a special visited flag.
 806          */
 807         INC_MASTER();
 808         res = follow_Mem_chain(load, skip_Proj(mem));
 809         return res;
 810 }  /* optimize_load */
 811
 812 /**
 813  * Check whether a value of mode new_mode would completely overwrite a value
 814  * of mode old_mode in memory.
 815  */
 816 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
 817 {
 818         return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
 819 }  /* is_completely_overwritten */
 820
 821 /**
 822  * follow the memory chain as long as there are only Loads and alias free Stores.
 823  *
 824  * INC_MASTER() must be called before dive into
 825  */
 826 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
 827         unsigned res = 0;
 828         ldst_info_t *info = get_irn_link(store);
 829         ir_node *pred;
 830         ir_node *ptr = get_Store_ptr(store);
 831         ir_node *mem = get_Store_mem(store);
 832         ir_node *value = get_Store_value(store);
 833         ir_mode *mode  = get_irn_mode(value);
 834         ir_node *block = get_nodes_block(store);
 835
 836         for (pred = curr; pred != store;) {
 837                 ldst_info_t *pred_info = get_irn_link(pred);
 838
 839                 /*
 840                  * BEWARE: one might think that checking the modes is useless, because
 841                  * if the pointers are identical, they refer to the same object.
 842                  * This is only true in strong typed languages, not is C were the following
 843                  * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
 844                  * However, if the mode that is written have a bigger  or equal size the the old
 845                  * one, the old value is completely overwritten and can be killed ...
 846                  */
 847                 if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
 848                     get_nodes_block(pred) == block &&
 849                     is_completely_overwritten(get_irn_mode(get_Store_value(pred)), mode)) {
 850                         /*
 851                          * a Store after a Store in the same block -- a write after write.
 852                          * We may remove the first Store, if it does not have an exception handler.
 853                          *
 854                          * TODO: What, if both have the same exception handler ???
 855                          */
 856                         if (get_Store_volatility(pred) != volatility_is_volatile && !pred_info->projs[pn_Store_X_except]) {
 857                                 DBG_OPT_WAW(pred, store);
 858                                 exchange( pred_info->projs[pn_Store_M], get_Store_mem(pred) );
 859                                 exchange(pred, new_Bad());
 860                                 reduce_adr_usage(ptr);
 861                                 return DF_CHANGED;
 862                         }
 863                 } else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
 864                            value == pred_info->projs[pn_Load_res]) {
 865                         /*
 866                          * a Store of a value after a Load -- a write after read.
 867                          * We may remove the second Store, if it does not have an exception handler.
 868                          */
 869                         if (! info->projs[pn_Store_X_except]) {
 870                                 DBG_OPT_WAR(store, pred);
 871                                 exchange( info->projs[pn_Store_M], mem );
 872                                 exchange(store, new_Bad());
 873                                 reduce_adr_usage(ptr);
 874                                 return DF_CHANGED;
 875                         }
 876                 }
 877
 878                 if (get_irn_op(pred) == op_Store) {
 879                         /* check if we can pass thru this store */
 880                         ir_alias_relation rel = get_alias_relation(
 881                                 current_ir_graph,
 882                                 get_Store_ptr(pred),
 883                                 get_irn_mode(get_Store_value(pred)),
 884                                 ptr, mode);
 885                         /* if the might be an alias, we cannot pass this Store */
 886                         if (rel != no_alias)
 887                                 break;
 888                         pred = skip_Proj(get_Store_mem(pred));
 889                 } else if (get_irn_op(pred) == op_Load) {
 890                         pred = skip_Proj(get_Load_mem(pred));
 891                 } else {
 892                         /* follow only Load chains */
 893                         break;
 894                 }
 895
 896                 /* check for cycles */
 897                 if (NODE_VISITED(pred_info))
 898                         break;
 899                 MARK_NODE(pred_info);
 900         }
 901
 902         if (get_irn_op(pred) == op_Sync) {
 903                 int i;
 904
 905                 /* handle all Sync predecessors */
 906                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
 907                         res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
 908                         if (res)
 909                                 break;
 910                 }
 911         }
 912         return res;
 913 }  /* follow_Mem_chain_for_Store */
 914
 915 /**
 916  * optimize a Store
 917  *
 918  * @param store  the Store node
 919  */
 920 static unsigned optimize_store(ir_node *store) {
 921         ir_node *ptr, *mem;
 922
 923         if (get_Store_volatility(store) == volatility_is_volatile)
 924                 return 0;
 925
 926         ptr = get_Store_ptr(store);
 927
 928         /* Check, if the address of this Store is used more than once.
 929          * If not, this Store cannot be removed in any case. */
 930         if (get_irn_n_uses(ptr) <= 1)
 931                 return 0;
 932
 933         mem = get_Store_mem(store);
 934
 935         /* follow the memory chain as long as there are only Loads */
 936         INC_MASTER();
 937         return follow_Mem_chain_for_Store(store, skip_Proj(mem));
 938 }  /* optimize_store */
 939
 940 /**
 941  * walker, optimizes Phi after Stores to identical places:
 942  * Does the following optimization:
 943  * @verbatim
 944  *
 945  *   val1   val2   val3          val1  val2  val3
 946  *    |      |      |               \    |    /
 947  *  Store  Store  Store              \   |   /
 948  *      \    |    /                   PhiData
 949  *       \   |   /                       |
 950  *        \  |  /                      Store
 951  *          PhiM
 952  *
 953  * @endverbatim
 954  * This reduces the number of stores and allows for predicated execution.
 955  * Moves Stores back to the end of a function which may be bad.
 956  *
 957  * This is only possible if the predecessor blocks have only one successor.
 958  */
 959 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
 960 {
 961         int i, n;
 962         ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
 963         ir_mode *mode;
 964         ir_node **inM, **inD, **stores;
 965         int *idx;
 966         dbg_info *db = NULL;
 967         ldst_info_t *info;
 968         block_info_t *bl_info;
 969         unsigned res = 0;
 970
 971         /* Must be a memory Phi */
 972         if (get_irn_mode(phi) != mode_M)
 973                 return 0;
 974
 975         n = get_Phi_n_preds(phi);
 976         if (n <= 0)
 977                 return 0;
 978
 979         store = skip_Proj(get_Phi_pred(phi, 0));
 980         old_store = store;
 981         if (get_irn_op(store) != op_Store)
 982                 return 0;
 983
 984         block = get_nodes_block(store);
 985
 986         /* abort on dead blocks */
 987         if (is_Block_dead(block))
 988                 return 0;
 989
 990         /* check if the block is post dominated by Phi-block
 991            and has no exception exit */
 992         bl_info = get_irn_link(block);
 993         if (bl_info->flags & BLOCK_HAS_EXC)
 994                 return 0;
 995
 996         phi_block = get_nodes_block(phi);
 997         if (! block_postdominates(phi_block, block))
 998                 return 0;
 999
1000         /* this is the address of the store */
1001         ptr  = get_Store_ptr(store);
1002         mode = get_irn_mode(get_Store_value(store));
1003         info = get_irn_link(store);
1004         exc  = info->exc_block;
1005
1006         for (i = 1; i < n; ++i) {
1007                 ir_node *pred = skip_Proj(get_Phi_pred(phi, i));
1008
1009                 if (get_irn_op(pred) != op_Store)
1010                         return 0;
1011
1012                 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1013                         return 0;
1014
1015                 info = get_irn_link(pred);
1016
1017                 /* check, if all stores have the same exception flow */
1018                 if (exc != info->exc_block)
1019                         return 0;
1020
1021                 /* abort on dead blocks */
1022                 block = get_nodes_block(pred);
1023                 if (is_Block_dead(block))
1024                         return 0;
1025
1026                 /* check if the block is post dominated by Phi-block
1027                    and has no exception exit. Note that block must be different from
1028                    Phi-block, else we would move a Store from end End of a block to its
1029                    Start... */
1030                 bl_info = get_irn_link(block);
1031                 if (bl_info->flags & BLOCK_HAS_EXC)
1032                         return 0;
1033                 if (block == phi_block || ! block_postdominates(phi_block, block))
1034                         return 0;
1035         }
1036
1037         /*
1038          * ok, when we are here, we found all predecessors of a Phi that
1039          * are Stores to the same address and size. That means whatever
1040          * we do before we enter the block of the Phi, we do a Store.
1041          * So, we can move the Store to the current block:
1042          *
1043          *   val1    val2    val3          val1  val2  val3
1044          *    |       |       |               \    |    /
1045          * | Str | | Str | | Str |             \   |   /
1046          *      \     |     /                   PhiData
1047          *       \    |    /                       |
1048          *        \   |   /                       Str
1049          *           PhiM
1050          *
1051          * Is only allowed if the predecessor blocks have only one successor.
1052          */
1053
1054         NEW_ARR_A(ir_node *, stores, n);
1055         NEW_ARR_A(ir_node *, inM, n);
1056         NEW_ARR_A(ir_node *, inD, n);
1057         NEW_ARR_A(int, idx, n);
1058
1059         /* Prepare: Collect all Store nodes.  We must do this
1060            first because we otherwise may loose a store when exchanging its
1061            memory Proj.
1062          */
1063         for (i = 0; i < n; ++i)
1064                 stores[i] = skip_Proj(get_Phi_pred(phi, i));
1065
1066         /* Prepare: Skip the memory Proj: we need this in the case some stores
1067            are cascaded.
1068            Beware: One Store might be included more than once in the stores[]
1069            list, so we must prevent to do the exchange more than once.
1070          */
1071         for (i = 0; i < n; ++i) {
1072                 ir_node *store = stores[i];
1073                 ir_node *proj_m;
1074
1075                 info = get_irn_link(store);
1076                 proj_m = info->projs[pn_Store_M];
1077
1078                 if (is_Proj(proj_m) && get_Proj_pred(proj_m) == store)
1079                         exchange(proj_m, get_Store_mem(store));
1080         }
1081
1082         /* first step: collect all inputs */
1083         for (i = 0; i < n; ++i) {
1084                 ir_node *store = stores[i];
1085                 info = get_irn_link(store);
1086
1087                 inM[i] = get_Store_mem(store);
1088                 inD[i] = get_Store_value(store);
1089                 idx[i] = info->exc_idx;
1090         }
1091         block = get_nodes_block(phi);
1092
1093         /* second step: create a new memory Phi */
1094         phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M);
1095
1096         /* third step: create a new data Phi */
1097         phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode);
1098
1099         /* fourth step: create the Store */
1100         store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD);
1101 #ifdef DO_CACHEOPT
1102         co_set_irn_name(store, co_get_irn_ident(old_store));
1103 #endif
1104
1105         projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
1106
1107         info = get_ldst_info(store, wenv);
1108         info->projs[pn_Store_M] = projM;
1109
1110         /* fifths step: repair exception flow */
1111         if (exc) {
1112                 ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except);
1113
1114                 info->projs[pn_Store_X_except] = projX;
1115                 info->exc_block                = exc;
1116                 info->exc_idx                  = idx[0];
1117
1118                 for (i = 0; i < n; ++i) {
1119                         set_Block_cfgpred(exc, idx[i], projX);
1120                 }
1121
1122                 if (n > 1) {
1123                         /* the exception block should be optimized as some inputs are identical now */
1124                 }
1125
1126                 res |= CF_CHANGED;
1127         }
1128
1129         /* sixth step: replace old Phi */
1130         exchange(phi, projM);
1131
1132         return res | DF_CHANGED;
1133 }  /* optimize_phi */
1134
1135 /**
1136  * walker, do the optimizations
1137  */
1138 static void do_load_store_optimize(ir_node *n, void *env) {
1139         walk_env_t *wenv = env;
1140
1141         switch (get_irn_opcode(n)) {
1142
1143         case iro_Load:
1144                 wenv->changes |= optimize_load(n);
1145                 break;
1146
1147         case iro_Store:
1148                 wenv->changes |= optimize_store(n);
1149                 break;
1150
1151         case iro_Phi:
1152                 wenv->changes |= optimize_phi(n, wenv);
1153
1154         default:
1155                 ;
1156         }
1157 }  /* do_load_store_optimize */
1158
1159 /*
1160  * do the load store optimization
1161  */
1162 void optimize_load_store(ir_graph *irg) {
1163         walk_env_t env;
1164
1165         assert(get_irg_phase_state(irg) != phase_building);
1166         assert(get_irg_pinned(irg) != op_pin_state_floats &&
1167                 "LoadStore optimization needs pinned graph");
1168
1169         if (! get_opt_redundant_loadstore())
1170                 return;
1171
1172         edges_assure(irg);
1173
1174         /* for Phi optimization post-dominators are needed ... */
1175         assure_postdoms(irg);
1176
1177         if (get_opt_alias_analysis()) {
1178                 assure_irg_address_taken_computed(irg);
1179                 assure_irp_globals_address_taken_computed();
1180         }
1181
1182         obstack_init(&env.obst);
1183         env.changes = 0;
1184
1185         /* init the links, then collect Loads/Stores/Proj's in lists */
1186         master_visited = 0;
1187         irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
1188
1189         /* now we have collected enough information, optimize */
1190         irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
1191
1192         obstack_free(&env.obst, NULL);
1193
1194         /* Handle graph state */
1195         if (env.changes) {
1196                 if (get_irg_outs_state(irg) == outs_consistent)
1197                         set_irg_outs_inconsistent(irg);
1198         }
1199
1200         if (env.changes & CF_CHANGED) {
1201                 /* is this really needed: Yes, control flow changed, block might
1202                 have Bad() predecessors. */
1203                 set_irg_doms_inconsistent(irg);
1204         }
1205 }  /* optimize_load_store */