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  * Project:     libFIRM
  22  * File name:   ir/opt/ldstopt.c
  23  * Purpose:     load store optimizations
  24  * Author:      Michael Beck
  25  * Created:
  26  * CVS-ID:      $Id$
  27  * Copyright:   (c) 1998-2007 Universität Karlsruhe
  28  */
  29 #ifdef HAVE_CONFIG_H
  30 # include "config.h"
  31 #endif
  32
  33 #ifdef HAVE_STRING_H
  34 # include <string.h>
  35 #endif
  36
  37 #include "irnode_t.h"
  38 #include "irgraph_t.h"
  39 #include "irmode_t.h"
  40 #include "iropt_t.h"
  41 #include "ircons_t.h"
  42 #include "irgmod.h"
  43 #include "irgwalk.h"
  44 #include "irvrfy.h"
  45 #include "tv_t.h"
  46 #include "dbginfo_t.h"
  47 #include "iropt_dbg.h"
  48 #include "irflag_t.h"
  49 #include "array.h"
  50 #include "irhooks.h"
  51 #include "iredges.h"
  52 #include "irtools.h"
  53 #include "opt_polymorphy.h"
  54 #include "irmemory.h"
  55 #include "xmalloc.h"
  56
  57 #ifdef DO_CACHEOPT
  58 #include "cacheopt/cachesim.h"
  59 #endif
  60
  61 #undef IMAX
  62 #define IMAX(a,b)       ((a) > (b) ? (a) : (b))
  63
  64 #define MAX_PROJ        IMAX(pn_Load_max, pn_Store_max)
  65
  66 enum changes_t {
  67         DF_CHANGED = 1,       /**< data flow changed */
  68         CF_CHANGED = 2,       /**< control flow changed */
  69 };
  70
  71 /**
  72  * walker environment
  73  */
  74 typedef struct _walk_env_t {
  75         struct obstack obst;          /**< list of all stores */
  76         unsigned changes;             /**< a bitmask of graph changes */
  77 } walk_env_t;
  78
  79 /**
  80  * flags for Load/Store
  81  */
  82 enum ldst_flags_t {
  83         LDST_VISITED = 1              /**< if set, this Load/Store is already visited */
  84 };
  85
  86 /** A Load/Store info. */
  87 typedef struct _ldst_info_t {
  88         ir_node  *projs[MAX_PROJ];    /**< list of Proj's of this node */
  89         ir_node  *exc_block;          /**< the exception block if available */
  90         int      exc_idx;             /**< predecessor index in the exception block */
  91         unsigned flags;               /**< flags */
  92         unsigned visited;             /**< visited counter for breaking loops */
  93 } ldst_info_t;
  94
  95 /**
  96  * flags for control flow.
  97  */
  98 enum block_flags_t {
  99         BLOCK_HAS_COND = 1,      /**< Block has conditional control flow */
 100         BLOCK_HAS_EXC  = 2       /**< Block has exceptional control flow */
 101 };
 102
 103 /**
 104  * a Block info.
 105  */
 106 typedef struct _block_info_t {
 107         unsigned flags;               /**< flags for the block */
 108 } block_info_t;
 109
 110 /** the master visited flag for loop detection. */
 111 static unsigned master_visited = 0;
 112
 113 #define INC_MASTER()       ++master_visited
 114 #define MARK_NODE(info)    (info)->visited = master_visited
 115 #define NODE_VISITED(info) (info)->visited >= master_visited
 116
 117 /**
 118  * get the Load/Store info of a node
 119  */
 120 static ldst_info_t *get_ldst_info(ir_node *node, walk_env_t *env) {
 121         ldst_info_t *info = get_irn_link(node);
 122
 123         if (! info) {
 124                 info = obstack_alloc(&env->obst, sizeof(*info));
 125                 memset(info, 0, sizeof(*info));
 126                 set_irn_link(node, info);
 127         }
 128         return info;
 129 }  /* get_ldst_info */
 130
 131 /**
 132  * get the Block info of a node
 133  */
 134 static block_info_t *get_block_info(ir_node *node, walk_env_t *env) {
 135         block_info_t *info = get_irn_link(node);
 136
 137         if (! info) {
 138                 info = obstack_alloc(&env->obst, sizeof(*info));
 139                 memset(info, 0, sizeof(*info));
 140                 set_irn_link(node, info);
 141         }
 142         return info;
 143 }  /* get_block_info */
 144
 145 /**
 146  * update the projection info for a Load/Store
 147  */
 148 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
 149 {
 150         long nr = get_Proj_proj(proj);
 151
 152         assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
 153
 154         if (info->projs[nr]) {
 155                 /* there is already one, do CSE */
 156                 exchange(proj, info->projs[nr]);
 157                 return DF_CHANGED;
 158         }
 159         else {
 160                 info->projs[nr] = proj;
 161                 return 0;
 162         }
 163 }  /* update_projs */
 164
 165 /**
 166  * update the exception block info for a Load/Store node.
 167  *
 168  * @param info   the load/store info struct
 169  * @param block  the exception handler block for this load/store
 170  * @param pos    the control flow input of the block
 171  */
 172 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
 173 {
 174         assert(info->exc_block == NULL && "more than one exception block found");
 175
 176         info->exc_block = block;
 177         info->exc_idx   = pos;
 178         return 0;
 179 }  /* update_exc */
 180
 181 /** Return the number of uses of an address node */
 182 #define get_irn_n_uses(adr)     get_irn_n_edges(adr)
 183
 184 /**
 185  * walker, collects all Load/Store/Proj nodes
 186  *
 187  * walks from Start -> End
 188  */
 189 static void collect_nodes(ir_node *node, void *env)
 190 {
 191         ir_op       *op = get_irn_op(node);
 192         ir_node     *pred, *blk, *pred_blk;
 193         ldst_info_t *ldst_info;
 194         walk_env_t  *wenv = env;
 195
 196         if (op == op_Proj) {
 197                 ir_node *adr;
 198                 ir_op *op;
 199
 200                 pred = get_Proj_pred(node);
 201                 op   = get_irn_op(pred);
 202
 203                 if (op == op_Load) {
 204                         ldst_info = get_ldst_info(pred, wenv);
 205
 206                         wenv->changes |= update_projs(ldst_info, node);
 207
 208                         if ((ldst_info->flags & LDST_VISITED) == 0) {
 209                                 adr = get_Load_ptr(pred);
 210                                 ldst_info->flags |= LDST_VISITED;
 211                         }
 212
 213                         /*
 214                         * Place the Proj's to the same block as the
 215                         * predecessor Load. This is always ok and prevents
 216                         * "non-SSA" form after optimizations if the Proj
 217                         * is in a wrong block.
 218                         */
 219                         blk      = get_nodes_block(node);
 220                         pred_blk = get_nodes_block(pred);
 221                         if (blk != pred_blk) {
 222                                 wenv->changes |= DF_CHANGED;
 223                                 set_nodes_block(node, pred_blk);
 224                         }
 225                 } else if (op == op_Store) {
 226                         ldst_info = get_ldst_info(pred, wenv);
 227
 228                         wenv->changes |= update_projs(ldst_info, node);
 229
 230                         if ((ldst_info->flags & LDST_VISITED) == 0) {
 231                                 adr = get_Store_ptr(pred);
 232                                 ldst_info->flags |= LDST_VISITED;
 233                         }
 234
 235                         /*
 236                         * Place the Proj's to the same block as the
 237                         * predecessor Store. This is always ok and prevents
 238                         * "non-SSA" form after optimizations if the Proj
 239                         * is in a wrong block.
 240                         */
 241                         blk      = get_nodes_block(node);
 242                         pred_blk = get_nodes_block(pred);
 243                         if (blk != pred_blk) {
 244                                 wenv->changes |= DF_CHANGED;
 245                                 set_nodes_block(node, pred_blk);
 246                         }
 247                 }
 248         } else if (op == op_Block) {
 249                 int i;
 250
 251                 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
 252                         ir_node      *pred_block;
 253                         block_info_t *bl_info;
 254
 255                         pred = skip_Proj(get_Block_cfgpred(node, i));
 256
 257                         /* ignore Bad predecessors, they will be removed later */
 258                         if (is_Bad(pred))
 259                                 continue;
 260
 261                         pred_block = get_nodes_block(pred);
 262                         bl_info    = get_block_info(pred_block, wenv);
 263
 264                         if (is_fragile_op(pred))
 265                                 bl_info->flags |= BLOCK_HAS_EXC;
 266                         else if (is_irn_forking(pred))
 267                                 bl_info->flags |= BLOCK_HAS_COND;
 268
 269                         if (get_irn_op(pred) == op_Load || get_irn_op(pred) == op_Store) {
 270                                 ldst_info = get_ldst_info(pred, wenv);
 271
 272                                 wenv->changes |= update_exc(ldst_info, node, i);
 273                         }
 274                 }
 275         }
 276 }  /* collect_nodes */
 277
 278 /**
 279  * Returns an entity if the address ptr points to a constant one.
 280  *
 281  * @param ptr  the address
 282  *
 283  * @return an entity or NULL
 284  */
 285 static ir_entity *find_constant_entity(ir_node *ptr)
 286 {
 287         for (;;) {
 288                 ir_op *op = get_irn_op(ptr);
 289
 290                 if (op == op_SymConst && (get_SymConst_kind(ptr) == symconst_addr_ent)) {
 291                         ir_entity *ent = get_SymConst_entity(ptr);
 292                         if (variability_constant == get_entity_variability(ent))
 293                                 return ent;
 294                         return NULL;
 295                 } else if (op == op_Sel) {
 296                         ir_entity *ent = get_Sel_entity(ptr);
 297                         ir_type   *tp  = get_entity_owner(ent);
 298
 299                         /* Do not fiddle with polymorphism. */
 300                         if (is_Class_type(get_entity_owner(ent)) &&
 301                                 ((get_entity_n_overwrites(ent)    != 0) ||
 302                                 (get_entity_n_overwrittenby(ent) != 0)   ) )
 303                                 return NULL;
 304
 305                         if (is_Array_type(tp)) {
 306                                 /* check bounds */
 307                                 int i, n;
 308
 309                                 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
 310                                         ir_node *bound;
 311                                         tarval *tlower, *tupper;
 312                                         ir_node *index = get_Sel_index(ptr, i);
 313                                         tarval *tv     = computed_value(index);
 314
 315                                         /* check if the index is constant */
 316                                         if (tv == tarval_bad)
 317                                                 return NULL;
 318
 319                                         bound  = get_array_lower_bound(tp, i);
 320                                         tlower = computed_value(bound);
 321                                         bound  = get_array_upper_bound(tp, i);
 322                                         tupper = computed_value(bound);
 323
 324                                         if (tlower == tarval_bad || tupper == tarval_bad)
 325                                                 return NULL;
 326
 327                                         if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
 328                                                 return NULL;
 329                                         if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
 330                                                 return NULL;
 331
 332                                         /* ok, bounds check finished */
 333                                 }
 334                         }
 335
 336                         if (variability_constant == get_entity_variability(ent))
 337                                 return ent;
 338
 339                         /* try next */
 340                         ptr = get_Sel_ptr(ptr);
 341                 } else
 342                         return NULL;
 343         }
 344 }  /* find_constant_entity */
 345
 346 /**
 347  * Return the Selection index of a Sel node from dimension n
 348  */
 349 static long get_Sel_array_index_long(ir_node *n, int dim) {
 350         ir_node *index = get_Sel_index(n, dim);
 351         assert(get_irn_op(index) == op_Const);
 352         return get_tarval_long(get_Const_tarval(index));
 353 }  /* get_Sel_array_index_long */
 354
 355 /**
 356  * Returns the accessed component graph path for an
 357  * node computing an address.
 358  *
 359  * @param ptr    the node computing the address
 360  * @param depth  current depth in steps upward from the root
 361  *               of the address
 362  */
 363 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
 364         compound_graph_path *res = NULL;
 365         ir_entity           *root, *field;
 366         int                 path_len, pos;
 367
 368         if (get_irn_op(ptr) == op_SymConst) {
 369                 /* a SymConst. If the depth is 0, this is an access to a global
 370                  * entity and we don't need a component path, else we know
 371                  * at least it's length.
 372                  */
 373                 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
 374                 root = get_SymConst_entity(ptr);
 375                 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
 376         } else {
 377                 assert(get_irn_op(ptr) == op_Sel);
 378                 /* it's a Sel, go up until we find the root */
 379                 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
 380
 381                 /* fill up the step in the path at the current position */
 382                 field    = get_Sel_entity(ptr);
 383                 path_len = get_compound_graph_path_length(res);
 384                 pos      = path_len - depth - 1;
 385                 set_compound_graph_path_node(res, pos, field);
 386
 387                 if (is_Array_type(get_entity_owner(field))) {
 388                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 389                         set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
 390                 }
 391         }
 392         return res;
 393 }  /* rec_get_accessed_path */
 394
 395 /** Returns an access path or NULL.  The access path is only
 396  *  valid, if the graph is in phase_high and _no_ address computation is used.
 397  */
 398 static compound_graph_path *get_accessed_path(ir_node *ptr) {
 399         return rec_get_accessed_path(ptr, 0);
 400 }  /* get_accessed_path */
 401
 402 /* forward */
 403 static void reduce_adr_usage(ir_node *ptr);
 404
 405 /**
 406  * Update a Load that may lost it's usage.
 407  */
 408 static void handle_load_update(ir_node *load) {
 409         ldst_info_t *info = get_irn_link(load);
 410
 411         /* do NOT touch volatile loads for now */
 412         if (get_Load_volatility(load) == volatility_is_volatile)
 413                 return;
 414
 415         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 416                 ir_node *ptr = get_Load_ptr(load);
 417                 ir_node *mem = get_Load_mem(load);
 418
 419                 /* a Load which value is neither used nor exception checked, remove it */
 420                 exchange(info->projs[pn_Load_M], mem);
 421                 exchange(load, new_Bad());
 422                 reduce_adr_usage(ptr);
 423         }
 424 }  /* handle_load_update */
 425
 426 /**
 427  * A Use of an address node is vanished. Check if this was a Proj
 428  * node and update the counters.
 429  */
 430 static void reduce_adr_usage(ir_node *ptr) {
 431         if (is_Proj(ptr)) {
 432                 if (get_irn_n_edges(ptr) <= 0) {
 433                         /* this Proj is dead now */
 434                         ir_node *pred = get_Proj_pred(ptr);
 435
 436                         if (is_Load(pred)) {
 437                                 ldst_info_t *info = get_irn_link(pred);
 438                                 info->projs[get_Proj_proj(ptr)] = NULL;
 439
 440                                 /* this node lost it's result proj, handle that */
 441                                 handle_load_update(pred);
 442                         }
 443                 }
 444         }
 445 }  /* reduce_adr_usage */
 446
 447 /**
 448  * Check, if an already existing value of mode old_mode can be converted
 449  * into the needed one new_mode without loss.
 450  */
 451 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
 452         if (old_mode == new_mode)
 453                 return 1;
 454
 455         /* if both modes are two-complement ones, we can always convert the
 456            Stored value into the needed one. */
 457         if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
 458                   get_mode_arithmetic(old_mode) == irma_twos_complement &&
 459                   get_mode_arithmetic(new_mode) == irma_twos_complement)
 460                 return 1;
 461         return 0;
 462 }  /* can_use_stored_value */
 463
 464 /**
 465  * Follow the memory chain as long as there are only Loads
 466  * and alias free Stores and try to replace current Load or Store
 467  * by a previous ones.
 468  * Note that in unreachable loops it might happen that we reach
 469  * load again, as well as we can fall into a cycle.
 470  * We break such cycles using a special visited flag.
 471  *
 472  * INC_MASTER() must be called before dive into
 473  */
 474 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
 475         unsigned res = 0;
 476         ldst_info_t *info = get_irn_link(load);
 477         ir_node *pred;
 478         ir_node *ptr       = get_Load_ptr(load);
 479         ir_node *mem       = get_Load_mem(load);
 480         ir_mode *load_mode = get_Load_mode(load);
 481
 482         for (pred = curr; load != pred; ) {
 483                 ldst_info_t *pred_info = get_irn_link(pred);
 484
 485                 /*
 486                  * BEWARE: one might think that checking the modes is useless, because
 487                  * if the pointers are identical, they refer to the same object.
 488                  * This is only true in strong typed languages, not in C were the following
 489                  * is possible a = *(ir_type1 *)p; b = *(ir_type2 *)p ...
 490                  */
 491                 if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
 492                     can_use_stored_value(get_irn_mode(get_Store_value(pred)), load_mode)) {
 493                         /*
 494                          * a Load immediately after a Store -- a read after write.
 495                          * We may remove the Load, if both Load & Store does not have an exception handler
 496                          * OR they are in the same block. In the latter case the Load cannot
 497                          * throw an exception when the previous Store was quiet.
 498                          *
 499                          * Why we need to check for Store Exception? If the Store cannot
 500                          * be executed (ROM) the exception handler might simply jump into
 501                          * the load block :-(
 502                          * We could make it a little bit better if we would know that the exception
 503                          * handler of the Store jumps directly to the end...
 504                          */
 505                         if ((pred_info->projs[pn_Store_X_except] == NULL && info->projs[pn_Load_X_except] == NULL) ||
 506                             get_nodes_block(load) == get_nodes_block(pred)) {
 507                                 ir_node *value = get_Store_value(pred);
 508
 509                                 DBG_OPT_RAW(load, value);
 510
 511                                 /* add an convert if needed */
 512                                 if (get_irn_mode(get_Store_value(pred)) != load_mode) {
 513                                         value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
 514                                 }
 515
 516                                 if (info->projs[pn_Load_M])
 517                                         exchange(info->projs[pn_Load_M], mem);
 518
 519                                 /* no exception */
 520                                 if (info->projs[pn_Load_X_except]) {
 521                                         exchange( info->projs[pn_Load_X_except], new_Bad());
 522                                         res |= CF_CHANGED;
 523                                 }
 524
 525                                 if (info->projs[pn_Load_res])
 526                                         exchange(info->projs[pn_Load_res], value);
 527
 528                                 exchange(load, new_Bad());
 529                                 reduce_adr_usage(ptr);
 530                                 return res | DF_CHANGED;
 531                         }
 532                 } else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
 533                            can_use_stored_value(get_Load_mode(pred), load_mode)) {
 534                         /*
 535                          * a Load after a Load -- a read after read.
 536                          * We may remove the second Load, if it does not have an exception handler
 537                          * OR they are in the same block. In the later case the Load cannot
 538                          * throw an exception when the previous Load was quiet.
 539                          *
 540                          * Here, there is no need to check if the previous Load has an exception
 541                          * hander because they would have exact the same exception...
 542                          */
 543                         if (info->projs[pn_Load_X_except] == NULL || get_nodes_block(load) == get_nodes_block(pred)) {
 544                                 ir_node *value;
 545
 546                                 DBG_OPT_RAR(load, pred);
 547
 548                                 /* the result is used */
 549                                 if (info->projs[pn_Load_res]) {
 550                                         if (pred_info->projs[pn_Load_res] == NULL) {
 551                                                 /* create a new Proj again */
 552                                                 pred_info->projs[pn_Load_res] = new_r_Proj(current_ir_graph, get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res);
 553                                         }
 554                                         value = pred_info->projs[pn_Load_res];
 555
 556                                         /* add an convert if needed */
 557                                         if (get_Load_mode(pred) != load_mode) {
 558                                                 value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
 559                                         }
 560
 561                                         exchange(info->projs[pn_Load_res], value);
 562                                 }
 563
 564                                 if (info->projs[pn_Load_M])
 565                                         exchange(info->projs[pn_Load_M], mem);
 566
 567                                 /* no exception */
 568                                 if (info->projs[pn_Load_X_except]) {
 569                                         exchange(info->projs[pn_Load_X_except], new_Bad());
 570                                         res |= CF_CHANGED;
 571                                 }
 572
 573                                 exchange(load, new_Bad());
 574                                 reduce_adr_usage(ptr);
 575                                 return res |= DF_CHANGED;
 576                         }
 577                 }
 578
 579                 if (get_irn_op(pred) == op_Store) {
 580                         /* check if we can pass through this store */
 581                         ir_alias_relation rel = get_alias_relation(
 582                                 current_ir_graph,
 583                                 get_Store_ptr(pred),
 584                                 get_irn_mode(get_Store_value(pred)),
 585                                 ptr, load_mode);
 586                         /* if the might be an alias, we cannot pass this Store */
 587                         if (rel != no_alias)
 588                                 break;
 589                         pred = skip_Proj(get_Store_mem(pred));
 590                 } else if (get_irn_op(pred) == op_Load) {
 591                         pred = skip_Proj(get_Load_mem(pred));
 592                 } else {
 593                         /* follow only Load chains */
 594                         break;
 595                 }
 596
 597                 /* check for cycles */
 598                 if (NODE_VISITED(pred_info))
 599                         break;
 600                 MARK_NODE(pred_info);
 601         }
 602
 603         if (get_irn_op(pred) == op_Sync) {
 604                 int i;
 605
 606                 /* handle all Sync predecessors */
 607                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
 608                         res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
 609                         if (res)
 610                                 break;
 611                 }
 612         }
 613
 614         return res;
 615 }  /* follow_Mem_chain */
 616
 617 /**
 618  * optimize a Load
 619  *
 620  * @param load  the Load node
 621  */
 622 static unsigned optimize_load(ir_node *load)
 623 {
 624         ldst_info_t *info = get_irn_link(load);
 625         ir_node *mem, *ptr, *new_node;
 626         ir_entity *ent;
 627         unsigned res = 0;
 628
 629         /* do NOT touch volatile loads for now */
 630         if (get_Load_volatility(load) == volatility_is_volatile)
 631                 return 0;
 632
 633         /* the address of the load to be optimized */
 634         ptr = get_Load_ptr(load);
 635
 636         /*
 637          * Check if we can remove the exception from a Load:
 638          * This can be done, if the address is from an Sel(Alloc) and
 639          * the Sel type is a subtype of the allocated type.
 640          *
 641          * This optimizes some often used OO constructs,
 642          * like x = new O; x->t;
 643          */
 644         if (info->projs[pn_Load_X_except]) {
 645                 if (is_Sel(ptr)) {
 646                         ir_node *mem = get_Sel_mem(ptr);
 647
 648                         /* FIXME: works with the current FE, but better use the base */
 649                         if (get_irn_op(skip_Proj(mem)) == op_Alloc) {
 650                                 /* ok, check the types */
 651                                 ir_entity *ent    = get_Sel_entity(ptr);
 652                                 ir_type   *s_type = get_entity_type(ent);
 653                                 ir_type   *a_type = get_Alloc_type(mem);
 654
 655                                 if (is_SubClass_of(s_type, a_type)) {
 656                                         /* ok, condition met: there can't be an exception because
 657                                         * Alloc guarantees that enough memory was allocated */
 658
 659                                         exchange(info->projs[pn_Load_X_except], new_Bad());
 660                                         info->projs[pn_Load_X_except] = NULL;
 661                                         res |= CF_CHANGED;
 662                                 }
 663                         }
 664                 } else if ((get_irn_op(skip_Proj(ptr)) == op_Alloc) ||
 665                         ((get_irn_op(ptr) == op_Cast) && (get_irn_op(skip_Proj(get_Cast_op(ptr))) == op_Alloc))) {
 666                                 /* simple case: a direct load after an Alloc. Firm Alloc throw
 667                                  * an exception in case of out-of-memory. So, there is no way for an
 668                                  * exception in this load.
 669                                  * This code is constructed by the "exception lowering" in the Jack compiler.
 670                                  */
 671                                 exchange(info->projs[pn_Load_X_except], new_Bad());
 672                                 info->projs[pn_Load_X_except] = NULL;
 673                                 res |= CF_CHANGED;
 674                 }
 675         }
 676
 677         /* The mem of the Load. Must still be returned after optimization. */
 678         mem  = get_Load_mem(load);
 679
 680         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 681                 /* a Load which value is neither used nor exception checked, remove it */
 682                 exchange(info->projs[pn_Load_M], mem);
 683
 684                 exchange(load, new_Bad());
 685                 reduce_adr_usage(ptr);
 686                 return res | DF_CHANGED;
 687         }
 688
 689         /* Load from a constant polymorphic field, where we can resolve
 690            polymorphism. */
 691         new_node = transform_node_Load(load);
 692         if (new_node != load) {
 693                 if (info->projs[pn_Load_M]) {
 694                         exchange(info->projs[pn_Load_M], mem);
 695                         info->projs[pn_Load_M] = NULL;
 696                 }
 697                 if (info->projs[pn_Load_X_except]) {
 698                         exchange(info->projs[pn_Load_X_except], new_Bad());
 699                         info->projs[pn_Load_X_except] = NULL;
 700                 }
 701                 if (info->projs[pn_Load_res])
 702                         exchange(info->projs[pn_Load_res], new_node);
 703
 704                 exchange(load, new_Bad());
 705                 reduce_adr_usage(ptr);
 706                 return res | DF_CHANGED;
 707         }
 708
 709         /* check if we can determine the entity that will be loaded */
 710         ent = find_constant_entity(ptr);
 711         if (ent) {
 712                 if ((allocation_static == get_entity_allocation(ent)) &&
 713                         (visibility_external_allocated != get_entity_visibility(ent))) {
 714                         /* a static allocation that is not external: there should be NO exception
 715                          * when loading. */
 716
 717                         /* no exception, clear the info field as it might be checked later again */
 718                         if (info->projs[pn_Load_X_except]) {
 719                                 exchange(info->projs[pn_Load_X_except], new_Bad());
 720                                 info->projs[pn_Load_X_except] = NULL;
 721                                 res |= CF_CHANGED;
 722                         }
 723
 724                         if (variability_constant == get_entity_variability(ent)
 725                                 && is_atomic_entity(ent)) {
 726                                 /* Might not be atomic after
 727                                    lowering of Sels.  In this
 728                                    case we could also load, but
 729                                    it's more complicated. */
 730                                 /* more simpler case: we load the content of a constant value:
 731                                  * replace it by the constant itself
 732                                  */
 733
 734                                 /* no memory */
 735                                 if (info->projs[pn_Load_M]) {
 736                                         exchange(info->projs[pn_Load_M], mem);
 737                                         res |= DF_CHANGED;
 738                                 }
 739                                 /* no result :-) */
 740                                 if (info->projs[pn_Load_res]) {
 741                                         if (is_atomic_entity(ent)) {
 742                                                 ir_node *c = copy_const_value(get_irn_dbg_info(load), get_atomic_ent_value(ent));
 743
 744                                                 DBG_OPT_RC(load, c);
 745                                                 exchange(info->projs[pn_Load_res], c);
 746                                                 res |= DF_CHANGED;
 747                                         }
 748                                 }
 749                                 exchange(load, new_Bad());
 750                                 reduce_adr_usage(ptr);
 751                                 return res;
 752                         } else if (variability_constant == get_entity_variability(ent)) {
 753                                 compound_graph_path *path = get_accessed_path(ptr);
 754
 755                                 if (path) {
 756                                         ir_node *c;
 757
 758                                         assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
 759                                         /*
 760                                         {
 761                                                 int j;
 762                                                 for (j = 0; j < get_compound_graph_path_length(path); ++j) {
 763                                                         ir_entity *node = get_compound_graph_path_node(path, j);
 764                                                         fprintf(stdout, ".%s", get_entity_name(node));
 765                                                         if (is_Array_type(get_entity_owner(node)))
 766                                                                 fprintf(stdout, "[%d]", get_compound_graph_path_array_index(path, j));
 767                                                 }
 768                                                 printf("\n");
 769                                         }
 770                                         */
 771
 772                                         c = get_compound_ent_value_by_path(ent, path);
 773                                         free_compound_graph_path(path);
 774
 775                                         /* printf("  cons: "); DDMN(c); */
 776
 777                                         if (info->projs[pn_Load_M]) {
 778                                                 exchange(info->projs[pn_Load_M], mem);
 779                                                 res |= DF_CHANGED;
 780                                         }
 781                                         if (info->projs[pn_Load_res]) {
 782                                                 exchange(info->projs[pn_Load_res], copy_const_value(get_irn_dbg_info(load), c));
 783                                                 res |= DF_CHANGED;
 784                                         }
 785                                         exchange(load, new_Bad());
 786                                         reduce_adr_usage(ptr);
 787                                         return res;
 788                                 } else {
 789                                         /*  We can not determine a correct access path.  E.g., in jack, we load
 790                                         a byte from an object to generate an exception.   Happens in test program
 791                                         Reflectiontest.
 792                                         printf(">>>>>>>>>>>>> Found access to constant entity %s in function %s\n", get_entity_name(ent),
 793                                         get_entity_name(get_irg_entity(current_ir_graph)));
 794                                         printf("  load: "); DDMN(load);
 795                                         printf("  ptr:  "); DDMN(ptr);
 796                                         */
 797                                 }
 798                         }
 799                 }
 800         }
 801
 802         /* Check, if the address of this load is used more than once.
 803          * If not, this load cannot be removed in any case. */
 804         if (get_irn_n_uses(ptr) <= 1)
 805                 return res;
 806
 807         /*
 808          * follow the memory chain as long as there are only Loads
 809          * and try to replace current Load or Store by a previous one.
 810          * Note that in unreachable loops it might happen that we reach
 811          * load again, as well as we can fall into a cycle.
 812          * We break such cycles using a special visited flag.
 813          */
 814         INC_MASTER();
 815         res = follow_Mem_chain(load, skip_Proj(mem));
 816         return res;
 817 }  /* optimize_load */
 818
 819 /**
 820  * Check whether a value of mode new_mode would completely overwrite a value
 821  * of mode old_mode in memory.
 822  */
 823 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
 824 {
 825         return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
 826 }  /* is_completely_overwritten */
 827
 828 /**
 829  * follow the memory chain as long as there are only Loads and alias free Stores.
 830  *
 831  * INC_MASTER() must be called before dive into
 832  */
 833 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
 834         unsigned res = 0;
 835         ldst_info_t *info = get_irn_link(store);
 836         ir_node *pred;
 837         ir_node *ptr = get_Store_ptr(store);
 838         ir_node *mem = get_Store_mem(store);
 839         ir_node *value = get_Store_value(store);
 840         ir_mode *mode  = get_irn_mode(value);
 841         ir_node *block = get_nodes_block(store);
 842
 843         for (pred = curr; pred != store;) {
 844                 ldst_info_t *pred_info = get_irn_link(pred);
 845
 846                 /*
 847                  * BEWARE: one might think that checking the modes is useless, because
 848                  * if the pointers are identical, they refer to the same object.
 849                  * This is only true in strong typed languages, not is C were the following
 850                  * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
 851                  * However, if the mode that is written have a bigger  or equal size the the old
 852                  * one, the old value is completely overwritten and can be killed ...
 853                  */
 854                 if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
 855                     get_nodes_block(pred) == block &&
 856                     is_completely_overwritten(get_irn_mode(get_Store_value(pred)), mode)) {
 857                         /*
 858                          * a Store after a Store in the same block -- a write after write.
 859                          * We may remove the first Store, if it does not have an exception handler.
 860                          *
 861                          * TODO: What, if both have the same exception handler ???
 862                          */
 863                         if (get_Store_volatility(pred) != volatility_is_volatile && !pred_info->projs[pn_Store_X_except]) {
 864                                 DBG_OPT_WAW(pred, store);
 865                                 exchange( pred_info->projs[pn_Store_M], get_Store_mem(pred) );
 866                                 exchange(pred, new_Bad());
 867                                 reduce_adr_usage(ptr);
 868                                 return DF_CHANGED;
 869                         }
 870                 } else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
 871                            value == pred_info->projs[pn_Load_res]) {
 872                         /*
 873                          * a Store of a value after a Load -- a write after read.
 874                          * We may remove the second Store, if it does not have an exception handler.
 875                          */
 876                         if (! info->projs[pn_Store_X_except]) {
 877                                 DBG_OPT_WAR(store, pred);
 878                                 exchange( info->projs[pn_Store_M], mem );
 879                                 exchange(store, new_Bad());
 880                                 reduce_adr_usage(ptr);
 881                                 return DF_CHANGED;
 882                         }
 883                 }
 884
 885                 if (get_irn_op(pred) == op_Store) {
 886                         /* check if we can pass thru this store */
 887                         ir_alias_relation rel = get_alias_relation(
 888                                 current_ir_graph,
 889                                 get_Store_ptr(pred),
 890                                 get_irn_mode(get_Store_value(pred)),
 891                                 ptr, mode);
 892                         /* if the might be an alias, we cannot pass this Store */
 893                         if (rel != no_alias)
 894                                 break;
 895                         pred = skip_Proj(get_Store_mem(pred));
 896                 } else if (get_irn_op(pred) == op_Load) {
 897                         pred = skip_Proj(get_Load_mem(pred));
 898                 } else {
 899                         /* follow only Load chains */
 900                         break;
 901                 }
 902
 903                 /* check for cycles */
 904                 if (NODE_VISITED(pred_info))
 905                         break;
 906                 MARK_NODE(pred_info);
 907         }
 908
 909         if (get_irn_op(pred) == op_Sync) {
 910                 int i;
 911
 912                 /* handle all Sync predecessors */
 913                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
 914                         res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
 915                         if (res)
 916                                 break;
 917                 }
 918         }
 919         return res;
 920 }  /* follow_Mem_chain_for_Store */
 921
 922 /**
 923  * optimize a Store
 924  *
 925  * @param store  the Store node
 926  */
 927 static unsigned optimize_store(ir_node *store) {
 928         ir_node *ptr, *mem;
 929
 930         if (get_Store_volatility(store) == volatility_is_volatile)
 931                 return 0;
 932
 933         ptr = get_Store_ptr(store);
 934
 935         /* Check, if the address of this Store is used more than once.
 936          * If not, this Store cannot be removed in any case. */
 937         if (get_irn_n_uses(ptr) <= 1)
 938                 return 0;
 939
 940         mem = get_Store_mem(store);
 941
 942         /* follow the memory chain as long as there are only Loads */
 943         INC_MASTER();
 944         return follow_Mem_chain_for_Store(store, skip_Proj(mem));
 945 }  /* optimize_store */
 946
 947 /**
 948  * walker, optimizes Phi after Stores to identical places:
 949  * Does the following optimization:
 950  * @verbatim
 951  *
 952  *   val1   val2   val3          val1  val2  val3
 953  *    |      |      |               \    |    /
 954  *  Store  Store  Store              \   |   /
 955  *      \    |    /                   PhiData
 956  *       \   |   /                       |
 957  *        \  |  /                      Store
 958  *          PhiM
 959  *
 960  * @endverbatim
 961  * This reduces the number of stores and allows for predicated execution.
 962  * Moves Stores back to the end of a function which may be bad.
 963  *
 964  * This is only possible if the predecessor blocks have only one successor.
 965  */
 966 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
 967 {
 968         int i, n;
 969         ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
 970         ir_mode *mode;
 971         ir_node **inM, **inD, **stores;
 972         int *idx;
 973         dbg_info *db = NULL;
 974         ldst_info_t *info;
 975         block_info_t *bl_info;
 976         unsigned res = 0;
 977
 978         /* Must be a memory Phi */
 979         if (get_irn_mode(phi) != mode_M)
 980                 return 0;
 981
 982         n = get_Phi_n_preds(phi);
 983         if (n <= 0)
 984                 return 0;
 985
 986         store = skip_Proj(get_Phi_pred(phi, 0));
 987         old_store = store;
 988         if (get_irn_op(store) != op_Store)
 989                 return 0;
 990
 991         block = get_nodes_block(store);
 992
 993         /* abort on dead blocks */
 994         if (is_Block_dead(block))
 995                 return 0;
 996
 997         /* check if the block is post dominated by Phi-block
 998            and has no exception exit */
 999         bl_info = get_irn_link(block);
1000         if (bl_info->flags & BLOCK_HAS_EXC)
1001                 return 0;
1002
1003         phi_block = get_nodes_block(phi);
1004         if (! block_postdominates(phi_block, block))
1005                 return 0;
1006
1007         /* this is the address of the store */
1008         ptr  = get_Store_ptr(store);
1009         mode = get_irn_mode(get_Store_value(store));
1010         info = get_irn_link(store);
1011         exc  = info->exc_block;
1012
1013         for (i = 1; i < n; ++i) {
1014                 ir_node *pred = skip_Proj(get_Phi_pred(phi, i));
1015
1016                 if (get_irn_op(pred) != op_Store)
1017                         return 0;
1018
1019                 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1020                         return 0;
1021
1022                 info = get_irn_link(pred);
1023
1024                 /* check, if all stores have the same exception flow */
1025                 if (exc != info->exc_block)
1026                         return 0;
1027
1028                 /* abort on dead blocks */
1029                 block = get_nodes_block(pred);
1030                 if (is_Block_dead(block))
1031                         return 0;
1032
1033                 /* check if the block is post dominated by Phi-block
1034                    and has no exception exit. Note that block must be different from
1035                    Phi-block, else we would move a Store from end End of a block to its
1036                    Start... */
1037                 bl_info = get_irn_link(block);
1038                 if (bl_info->flags & BLOCK_HAS_EXC)
1039                         return 0;
1040                 if (block == phi_block || ! block_postdominates(phi_block, block))
1041                         return 0;
1042         }
1043
1044         /*
1045          * ok, when we are here, we found all predecessors of a Phi that
1046          * are Stores to the same address and size. That means whatever
1047          * we do before we enter the block of the Phi, we do a Store.
1048          * So, we can move the Store to the current block:
1049          *
1050          *   val1    val2    val3          val1  val2  val3
1051          *    |       |       |               \    |    /
1052          * | Str | | Str | | Str |             \   |   /
1053          *      \     |     /                   PhiData
1054          *       \    |    /                       |
1055          *        \   |   /                       Str
1056          *           PhiM
1057          *
1058          * Is only allowed if the predecessor blocks have only one successor.
1059          */
1060
1061         NEW_ARR_A(ir_node *, stores, n);
1062         NEW_ARR_A(ir_node *, inM, n);
1063         NEW_ARR_A(ir_node *, inD, n);
1064         NEW_ARR_A(int, idx, n);
1065
1066         /* Prepare: Collect all Store nodes.  We must do this
1067            first because we otherwise may loose a store when exchanging its
1068            memory Proj.
1069          */
1070         for (i = 0; i < n; ++i)
1071                 stores[i] = skip_Proj(get_Phi_pred(phi, i));
1072
1073         /* Prepare: Skip the memory Proj: we need this in the case some stores
1074            are cascaded.
1075            Beware: One Store might be included more than once in the stores[]
1076            list, so we must prevent to do the exchange more than once.
1077          */
1078         for (i = 0; i < n; ++i) {
1079                 ir_node *store = stores[i];
1080                 ir_node *proj_m;
1081
1082                 info = get_irn_link(store);
1083                 proj_m = info->projs[pn_Store_M];
1084
1085                 if (is_Proj(proj_m) && get_Proj_pred(proj_m) == store)
1086                         exchange(proj_m, get_Store_mem(store));
1087         }
1088
1089         /* first step: collect all inputs */
1090         for (i = 0; i < n; ++i) {
1091                 ir_node *store = stores[i];
1092                 info = get_irn_link(store);
1093
1094                 inM[i] = get_Store_mem(store);
1095                 inD[i] = get_Store_value(store);
1096                 idx[i] = info->exc_idx;
1097         }
1098         block = get_nodes_block(phi);
1099
1100         /* second step: create a new memory Phi */
1101         phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M);
1102
1103         /* third step: create a new data Phi */
1104         phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode);
1105
1106         /* fourth step: create the Store */
1107         store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD);
1108 #ifdef DO_CACHEOPT
1109         co_set_irn_name(store, co_get_irn_ident(old_store));
1110 #endif
1111
1112         projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
1113
1114         info = get_ldst_info(store, wenv);
1115         info->projs[pn_Store_M] = projM;
1116
1117         /* fifths step: repair exception flow */
1118         if (exc) {
1119                 ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except);
1120
1121                 info->projs[pn_Store_X_except] = projX;
1122                 info->exc_block                = exc;
1123                 info->exc_idx                  = idx[0];
1124
1125                 for (i = 0; i < n; ++i) {
1126                         set_Block_cfgpred(exc, idx[i], projX);
1127                 }
1128
1129                 if (n > 1) {
1130                         /* the exception block should be optimized as some inputs are identical now */
1131                 }
1132
1133                 res |= CF_CHANGED;
1134         }
1135
1136         /* sixth step: replace old Phi */
1137         exchange(phi, projM);
1138
1139         return res | DF_CHANGED;
1140 }  /* optimize_phi */
1141
1142 /**
1143  * walker, do the optimizations
1144  */
1145 static void do_load_store_optimize(ir_node *n, void *env) {
1146         walk_env_t *wenv = env;
1147
1148         switch (get_irn_opcode(n)) {
1149
1150         case iro_Load:
1151                 wenv->changes |= optimize_load(n);
1152                 break;
1153
1154         case iro_Store:
1155                 wenv->changes |= optimize_store(n);
1156                 break;
1157
1158         case iro_Phi:
1159                 wenv->changes |= optimize_phi(n, wenv);
1160
1161         default:
1162                 ;
1163         }
1164 }  /* do_load_store_optimize */
1165
1166 /*
1167  * do the load store optimization
1168  */
1169 void optimize_load_store(ir_graph *irg) {
1170         walk_env_t env;
1171
1172         assert(get_irg_phase_state(irg) != phase_building);
1173         assert(get_irg_pinned(irg) != op_pin_state_floats &&
1174                 "LoadStore optimization needs pinned graph");
1175
1176         if (! get_opt_redundant_loadstore())
1177                 return;
1178
1179         edges_assure(irg);
1180
1181         /* for Phi optimization post-dominators are needed ... */
1182         assure_postdoms(irg);
1183
1184         if (get_opt_alias_analysis()) {
1185                 assure_irg_address_taken_computed(irg);
1186                 assure_irp_globals_address_taken_computed();
1187         }
1188
1189         obstack_init(&env.obst);
1190         env.changes = 0;
1191
1192         /* init the links, then collect Loads/Stores/Proj's in lists */
1193         master_visited = 0;
1194         irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
1195
1196         /* now we have collected enough information, optimize */
1197         irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
1198
1199         obstack_free(&env.obst, NULL);
1200
1201         /* Handle graph state */
1202         if (env.changes) {
1203                 if (get_irg_outs_state(irg) == outs_consistent)
1204                         set_irg_outs_inconsistent(irg);
1205         }
1206
1207         if (env.changes & CF_CHANGED) {
1208                 /* is this really needed: Yes, control flow changed, block might
1209                 have Bad() predecessors. */
1210                 set_irg_doms_inconsistent(irg);
1211         }
1212 }  /* optimize_load_store */