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

   1 /*
   2  * Copyright (C) 1995-2007 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief   Load/Store optimizations.
  23  * @author  Michael Beck
  24  * @version $Id$
  25  */
  26 #ifdef HAVE_CONFIG_H
  27 # include "config.h"
  28 #endif
  29
  30 #include <string.h>
  31
  32 #include "iroptimize.h"
  33 #include "irnode_t.h"
  34 #include "irgraph_t.h"
  35 #include "irmode_t.h"
  36 #include "iropt_t.h"
  37 #include "ircons_t.h"
  38 #include "irgmod.h"
  39 #include "irgwalk.h"
  40 #include "irvrfy.h"
  41 #include "tv_t.h"
  42 #include "dbginfo_t.h"
  43 #include "iropt_dbg.h"
  44 #include "irflag_t.h"
  45 #include "array.h"
  46 #include "irhooks.h"
  47 #include "iredges.h"
  48 #include "irtools.h"
  49 #include "opt_polymorphy.h"
  50 #include "irmemory.h"
  51 #include "xmalloc.h"
  52
  53 #ifdef DO_CACHEOPT
  54 #include "cacheopt/cachesim.h"
  55 #endif
  56
  57 #undef IMAX
  58 #define IMAX(a,b)       ((a) > (b) ? (a) : (b))
  59
  60 #define MAX_PROJ        IMAX(pn_Load_max, pn_Store_max)
  61
  62 enum changes_t {
  63         DF_CHANGED = 1,       /**< data flow changed */
  64         CF_CHANGED = 2,       /**< control flow changed */
  65 };
  66
  67 /**
  68  * walker environment
  69  */
  70 typedef struct _walk_env_t {
  71         struct obstack obst;          /**< list of all stores */
  72         unsigned changes;             /**< a bitmask of graph changes */
  73 } walk_env_t;
  74
  75 /**
  76  * flags for Load/Store
  77  */
  78 enum ldst_flags_t {
  79         LDST_VISITED = 1              /**< if set, this Load/Store is already visited */
  80 };
  81
  82 /** A Load/Store info. */
  83 typedef struct _ldst_info_t {
  84         ir_node  *projs[MAX_PROJ];    /**< list of Proj's of this node */
  85         ir_node  *exc_block;          /**< the exception block if available */
  86         int      exc_idx;             /**< predecessor index in the exception block */
  87         unsigned flags;               /**< flags */
  88         unsigned visited;             /**< visited counter for breaking loops */
  89 } ldst_info_t;
  90
  91 /**
  92  * flags for control flow.
  93  */
  94 enum block_flags_t {
  95         BLOCK_HAS_COND = 1,      /**< Block has conditional control flow */
  96         BLOCK_HAS_EXC  = 2       /**< Block has exceptional control flow */
  97 };
  98
  99 /**
 100  * a Block info.
 101  */
 102 typedef struct _block_info_t {
 103         unsigned flags;               /**< flags for the block */
 104 } block_info_t;
 105
 106 /** the master visited flag for loop detection. */
 107 static unsigned master_visited = 0;
 108
 109 #define INC_MASTER()       ++master_visited
 110 #define MARK_NODE(info)    (info)->visited = master_visited
 111 #define NODE_VISITED(info) (info)->visited >= master_visited
 112
 113 /**
 114  * get the Load/Store info of a node
 115  */
 116 static ldst_info_t *get_ldst_info(ir_node *node, walk_env_t *env) {
 117         ldst_info_t *info = get_irn_link(node);
 118
 119         if (! info) {
 120                 info = obstack_alloc(&env->obst, sizeof(*info));
 121                 memset(info, 0, sizeof(*info));
 122                 set_irn_link(node, info);
 123         }
 124         return info;
 125 }  /* get_ldst_info */
 126
 127 /**
 128  * get the Block info of a node
 129  */
 130 static block_info_t *get_block_info(ir_node *node, walk_env_t *env) {
 131         block_info_t *info = get_irn_link(node);
 132
 133         if (! info) {
 134                 info = obstack_alloc(&env->obst, sizeof(*info));
 135                 memset(info, 0, sizeof(*info));
 136                 set_irn_link(node, info);
 137         }
 138         return info;
 139 }  /* get_block_info */
 140
 141 /**
 142  * update the projection info for a Load/Store
 143  */
 144 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
 145 {
 146         long nr = get_Proj_proj(proj);
 147
 148         assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
 149
 150         if (info->projs[nr]) {
 151                 /* there is already one, do CSE */
 152                 exchange(proj, info->projs[nr]);
 153                 return DF_CHANGED;
 154         }
 155         else {
 156                 info->projs[nr] = proj;
 157                 return 0;
 158         }
 159 }  /* update_projs */
 160
 161 /**
 162  * update the exception block info for a Load/Store node.
 163  *
 164  * @param info   the load/store info struct
 165  * @param block  the exception handler block for this load/store
 166  * @param pos    the control flow input of the block
 167  */
 168 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
 169 {
 170         assert(info->exc_block == NULL && "more than one exception block found");
 171
 172         info->exc_block = block;
 173         info->exc_idx   = pos;
 174         return 0;
 175 }  /* update_exc */
 176
 177 /** Return the number of uses of an address node */
 178 #define get_irn_n_uses(adr)     get_irn_n_edges(adr)
 179
 180 /**
 181  * walker, collects all Load/Store/Proj nodes
 182  *
 183  * walks from Start -> End
 184  */
 185 static void collect_nodes(ir_node *node, void *env)
 186 {
 187         ir_op       *op = get_irn_op(node);
 188         ir_node     *pred, *blk, *pred_blk;
 189         ldst_info_t *ldst_info;
 190         walk_env_t  *wenv = env;
 191
 192         if (op == op_Proj) {
 193                 ir_node *adr;
 194                 ir_op *op;
 195
 196                 pred = get_Proj_pred(node);
 197                 op   = get_irn_op(pred);
 198
 199                 if (op == op_Load) {
 200                         ldst_info = get_ldst_info(pred, wenv);
 201
 202                         wenv->changes |= update_projs(ldst_info, node);
 203
 204                         if ((ldst_info->flags & LDST_VISITED) == 0) {
 205                                 adr = get_Load_ptr(pred);
 206                                 ldst_info->flags |= LDST_VISITED;
 207                         }
 208
 209                         /*
 210                          * Place the Proj's to the same block as the
 211                          * predecessor Load. This is always ok and prevents
 212                          * "non-SSA" form after optimizations if the Proj
 213                          * is in a wrong block.
 214                          */
 215                         blk      = get_nodes_block(node);
 216                         pred_blk = get_nodes_block(pred);
 217                         if (blk != pred_blk) {
 218                                 wenv->changes |= DF_CHANGED;
 219                                 set_nodes_block(node, pred_blk);
 220                         }
 221                 } else if (op == op_Store) {
 222                         ldst_info = get_ldst_info(pred, wenv);
 223
 224                         wenv->changes |= update_projs(ldst_info, node);
 225
 226                         if ((ldst_info->flags & LDST_VISITED) == 0) {
 227                                 adr = get_Store_ptr(pred);
 228                                 ldst_info->flags |= LDST_VISITED;
 229                         }
 230
 231                         /*
 232                         * Place the Proj's to the same block as the
 233                         * predecessor Store. This is always ok and prevents
 234                         * "non-SSA" form after optimizations if the Proj
 235                         * is in a wrong block.
 236                         */
 237                         blk      = get_nodes_block(node);
 238                         pred_blk = get_nodes_block(pred);
 239                         if (blk != pred_blk) {
 240                                 wenv->changes |= DF_CHANGED;
 241                                 set_nodes_block(node, pred_blk);
 242                         }
 243                 }
 244         } else if (op == op_Block) {
 245                 int i;
 246
 247                 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
 248                         ir_node      *pred_block, *proj;
 249                         block_info_t *bl_info;
 250                         int          is_exc = 0;
 251
 252                         pred = proj = get_Block_cfgpred(node, i);
 253
 254                         if (is_Proj(proj)) {
 255                                 pred   = get_Proj_pred(proj);
 256                                 is_exc = get_Proj_proj(proj) == pn_Generic_X_except;
 257                         }
 258
 259                         /* ignore Bad predecessors, they will be removed later */
 260                         if (is_Bad(pred))
 261                                 continue;
 262
 263                         pred_block = get_nodes_block(pred);
 264                         bl_info    = get_block_info(pred_block, wenv);
 265
 266                         if (is_fragile_op(pred) && is_exc)
 267                                 bl_info->flags |= BLOCK_HAS_EXC;
 268                         else if (is_irn_forking(pred))
 269                                 bl_info->flags |= BLOCK_HAS_COND;
 270
 271                         if (is_exc && (get_irn_op(pred) == op_Load || get_irn_op(pred) == op_Store)) {
 272                                 ldst_info = get_ldst_info(pred, wenv);
 273
 274                                 wenv->changes |= update_exc(ldst_info, node, i);
 275                         }
 276                 }
 277         }
 278 }  /* collect_nodes */
 279
 280 /**
 281  * Returns an entity if the address ptr points to a constant one.
 282  *
 283  * @param ptr  the address
 284  *
 285  * @return an entity or NULL
 286  */
 287 static ir_entity *find_constant_entity(ir_node *ptr)
 288 {
 289         for (;;) {
 290                 ir_op *op = get_irn_op(ptr);
 291
 292                 if (op == op_SymConst && (get_SymConst_kind(ptr) == symconst_addr_ent)) {
 293                         ir_entity *ent = get_SymConst_entity(ptr);
 294                         if (variability_constant == get_entity_variability(ent))
 295                                 return ent;
 296                         return NULL;
 297                 } else if (op == op_Sel) {
 298                         ir_entity *ent = get_Sel_entity(ptr);
 299                         ir_type   *tp  = get_entity_owner(ent);
 300
 301                         /* Do not fiddle with polymorphism. */
 302                         if (is_Class_type(get_entity_owner(ent)) &&
 303                                 ((get_entity_n_overwrites(ent)    != 0) ||
 304                                 (get_entity_n_overwrittenby(ent) != 0)   ) )
 305                                 return NULL;
 306
 307                         if (is_Array_type(tp)) {
 308                                 /* check bounds */
 309                                 int i, n;
 310
 311                                 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
 312                                         ir_node *bound;
 313                                         tarval *tlower, *tupper;
 314                                         ir_node *index = get_Sel_index(ptr, i);
 315                                         tarval *tv     = computed_value(index);
 316
 317                                         /* check if the index is constant */
 318                                         if (tv == tarval_bad)
 319                                                 return NULL;
 320
 321                                         bound  = get_array_lower_bound(tp, i);
 322                                         tlower = computed_value(bound);
 323                                         bound  = get_array_upper_bound(tp, i);
 324                                         tupper = computed_value(bound);
 325
 326                                         if (tlower == tarval_bad || tupper == tarval_bad)
 327                                                 return NULL;
 328
 329                                         if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
 330                                                 return NULL;
 331                                         if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
 332                                                 return NULL;
 333
 334                                         /* ok, bounds check finished */
 335                                 }
 336                         }
 337
 338                         if (variability_constant == get_entity_variability(ent))
 339                                 return ent;
 340
 341                         /* try next */
 342                         ptr = get_Sel_ptr(ptr);
 343                 } else
 344                         return NULL;
 345         }
 346 }  /* find_constant_entity */
 347
 348 /**
 349  * Return the Selection index of a Sel node from dimension n
 350  */
 351 static long get_Sel_array_index_long(ir_node *n, int dim) {
 352         ir_node *index = get_Sel_index(n, dim);
 353         assert(get_irn_op(index) == op_Const);
 354         return get_tarval_long(get_Const_tarval(index));
 355 }  /* get_Sel_array_index_long */
 356
 357 /**
 358  * Returns the accessed component graph path for an
 359  * node computing an address.
 360  *
 361  * @param ptr    the node computing the address
 362  * @param depth  current depth in steps upward from the root
 363  *               of the address
 364  */
 365 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
 366         compound_graph_path *res = NULL;
 367         ir_entity           *root, *field;
 368         int                 path_len, pos;
 369
 370         if (get_irn_op(ptr) == op_SymConst) {
 371                 /* a SymConst. If the depth is 0, this is an access to a global
 372                  * entity and we don't need a component path, else we know
 373                  * at least it's length.
 374                  */
 375                 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
 376                 root = get_SymConst_entity(ptr);
 377                 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
 378         } else {
 379                 assert(get_irn_op(ptr) == op_Sel);
 380                 /* it's a Sel, go up until we find the root */
 381                 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
 382
 383                 /* fill up the step in the path at the current position */
 384                 field    = get_Sel_entity(ptr);
 385                 path_len = get_compound_graph_path_length(res);
 386                 pos      = path_len - depth - 1;
 387                 set_compound_graph_path_node(res, pos, field);
 388
 389                 if (is_Array_type(get_entity_owner(field))) {
 390                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 391                         set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
 392                 }
 393         }
 394         return res;
 395 }  /* rec_get_accessed_path */
 396
 397 /** Returns an access path or NULL.  The access path is only
 398  *  valid, if the graph is in phase_high and _no_ address computation is used.
 399  */
 400 static compound_graph_path *get_accessed_path(ir_node *ptr) {
 401         return rec_get_accessed_path(ptr, 0);
 402 }  /* get_accessed_path */
 403
 404 /* forward */
 405 static void reduce_adr_usage(ir_node *ptr);
 406
 407 /**
 408  * Update a Load that may lost it's usage.
 409  */
 410 static void handle_load_update(ir_node *load) {
 411         ldst_info_t *info = get_irn_link(load);
 412
 413         /* do NOT touch volatile loads for now */
 414         if (get_Load_volatility(load) == volatility_is_volatile)
 415                 return;
 416
 417         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 418                 ir_node *ptr = get_Load_ptr(load);
 419                 ir_node *mem = get_Load_mem(load);
 420
 421                 /* a Load which value is neither used nor exception checked, remove it */
 422                 exchange(info->projs[pn_Load_M], mem);
 423                 exchange(load, new_Bad());
 424                 reduce_adr_usage(ptr);
 425         }
 426 }  /* handle_load_update */
 427
 428 /**
 429  * A Use of an address node is vanished. Check if this was a Proj
 430  * node and update the counters.
 431  */
 432 static void reduce_adr_usage(ir_node *ptr) {
 433         if (is_Proj(ptr)) {
 434                 if (get_irn_n_edges(ptr) <= 0) {
 435                         /* this Proj is dead now */
 436                         ir_node *pred = get_Proj_pred(ptr);
 437
 438                         if (is_Load(pred)) {
 439                                 ldst_info_t *info = get_irn_link(pred);
 440                                 info->projs[get_Proj_proj(ptr)] = NULL;
 441
 442                                 /* this node lost it's result proj, handle that */
 443                                 handle_load_update(pred);
 444                         }
 445                 }
 446         }
 447 }  /* reduce_adr_usage */
 448
 449 /**
 450  * Check, if an already existing value of mode old_mode can be converted
 451  * into the needed one new_mode without loss.
 452  */
 453 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
 454         if (old_mode == new_mode)
 455                 return 1;
 456
 457         /* if both modes are two-complement ones, we can always convert the
 458            Stored value into the needed one. */
 459         if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
 460                   get_mode_arithmetic(old_mode) == irma_twos_complement &&
 461                   get_mode_arithmetic(new_mode) == irma_twos_complement)
 462                 return 1;
 463         return 0;
 464 }  /* can_use_stored_value */
 465
 466 /**
 467  * Follow the memory chain as long as there are only Loads
 468  * and alias free Stores and try to replace current Load or Store
 469  * by a previous ones.
 470  * Note that in unreachable loops it might happen that we reach
 471  * load again, as well as we can fall into a cycle.
 472  * We break such cycles using a special visited flag.
 473  *
 474  * INC_MASTER() must be called before dive into
 475  */
 476 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
 477         unsigned res = 0;
 478         ldst_info_t *info = get_irn_link(load);
 479         ir_node *pred;
 480         ir_node *ptr       = get_Load_ptr(load);
 481         ir_node *mem       = get_Load_mem(load);
 482         ir_mode *load_mode = get_Load_mode(load);
 483
 484         for (pred = curr; load != pred; ) {
 485                 ldst_info_t *pred_info = get_irn_link(pred);
 486
 487                 /*
 488                  * BEWARE: one might think that checking the modes is useless, because
 489                  * if the pointers are identical, they refer to the same object.
 490                  * This is only true in strong typed languages, not in C were the following
 491                  * is possible a = *(ir_type1 *)p; b = *(ir_type2 *)p ...
 492                  */
 493                 if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
 494                     can_use_stored_value(get_irn_mode(get_Store_value(pred)), load_mode)) {
 495                         /*
 496                          * a Load immediately after a Store -- a read after write.
 497                          * We may remove the Load, if both Load & Store does not have an exception handler
 498                          * OR they are in the same block. In the latter case the Load cannot
 499                          * throw an exception when the previous Store was quiet.
 500                          *
 501                          * Why we need to check for Store Exception? If the Store cannot
 502                          * be executed (ROM) the exception handler might simply jump into
 503                          * the load block :-(
 504                          * We could make it a little bit better if we would know that the exception
 505                          * handler of the Store jumps directly to the end...
 506                          */
 507                         if ((pred_info->projs[pn_Store_X_except] == NULL && info->projs[pn_Load_X_except] == NULL) ||
 508                             get_nodes_block(load) == get_nodes_block(pred)) {
 509                                 ir_node *value = get_Store_value(pred);
 510
 511                                 DBG_OPT_RAW(load, value);
 512
 513                                 /* add an convert if needed */
 514                                 if (get_irn_mode(get_Store_value(pred)) != load_mode) {
 515                                         value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
 516                                 }
 517
 518                                 if (info->projs[pn_Load_M])
 519                                         exchange(info->projs[pn_Load_M], mem);
 520
 521                                 /* no exception */
 522                                 if (info->projs[pn_Load_X_except]) {
 523                                         exchange( info->projs[pn_Load_X_except], new_Bad());
 524                                         res |= CF_CHANGED;
 525                                 }
 526
 527                                 if (info->projs[pn_Load_res])
 528                                         exchange(info->projs[pn_Load_res], value);
 529
 530                                 exchange(load, new_Bad());
 531                                 reduce_adr_usage(ptr);
 532                                 return res | DF_CHANGED;
 533                         }
 534                 } else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
 535                            can_use_stored_value(get_Load_mode(pred), load_mode)) {
 536                         /*
 537                          * a Load after a Load -- a read after read.
 538                          * We may remove the second Load, if it does not have an exception handler
 539                          * OR they are in the same block. In the later case the Load cannot
 540                          * throw an exception when the previous Load was quiet.
 541                          *
 542                          * Here, there is no need to check if the previous Load has an exception
 543                          * hander because they would have exact the same exception...
 544                          */
 545                         if (info->projs[pn_Load_X_except] == NULL || get_nodes_block(load) == get_nodes_block(pred)) {
 546                                 ir_node *value;
 547
 548                                 DBG_OPT_RAR(load, pred);
 549
 550                                 /* the result is used */
 551                                 if (info->projs[pn_Load_res]) {
 552                                         if (pred_info->projs[pn_Load_res] == NULL) {
 553                                                 /* create a new Proj again */
 554                                                 pred_info->projs[pn_Load_res] = new_r_Proj(current_ir_graph, get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res);
 555                                         }
 556                                         value = pred_info->projs[pn_Load_res];
 557
 558                                         /* add an convert if needed */
 559                                         if (get_Load_mode(pred) != load_mode) {
 560                                                 value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
 561                                         }
 562
 563                                         exchange(info->projs[pn_Load_res], value);
 564                                 }
 565
 566                                 if (info->projs[pn_Load_M])
 567                                         exchange(info->projs[pn_Load_M], mem);
 568
 569                                 /* no exception */
 570                                 if (info->projs[pn_Load_X_except]) {
 571                                         exchange(info->projs[pn_Load_X_except], new_Bad());
 572                                         res |= CF_CHANGED;
 573                                 }
 574
 575                                 exchange(load, new_Bad());
 576                                 reduce_adr_usage(ptr);
 577                                 return res |= DF_CHANGED;
 578                         }
 579                 }
 580
 581                 if (get_irn_op(pred) == op_Store) {
 582                         /* check if we can pass through this store */
 583                         ir_alias_relation rel = get_alias_relation(
 584                                 current_ir_graph,
 585                                 get_Store_ptr(pred),
 586                                 get_irn_mode(get_Store_value(pred)),
 587                                 ptr, load_mode);
 588                         /* if the might be an alias, we cannot pass this Store */
 589                         if (rel != no_alias)
 590                                 break;
 591                         pred = skip_Proj(get_Store_mem(pred));
 592                 } else if (get_irn_op(pred) == op_Load) {
 593                         pred = skip_Proj(get_Load_mem(pred));
 594                 } else {
 595                         /* follow only Load chains */
 596                         break;
 597                 }
 598
 599                 /* check for cycles */
 600                 if (NODE_VISITED(pred_info))
 601                         break;
 602                 MARK_NODE(pred_info);
 603         }
 604
 605         if (get_irn_op(pred) == op_Sync) {
 606                 int i;
 607
 608                 /* handle all Sync predecessors */
 609                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
 610                         res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
 611                         if (res)
 612                                 break;
 613                 }
 614         }
 615
 616         return res;
 617 }  /* follow_Mem_chain */
 618
 619 /**
 620  * optimize a Load
 621  *
 622  * @param load  the Load node
 623  */
 624 static unsigned optimize_load(ir_node *load)
 625 {
 626         ldst_info_t *info = get_irn_link(load);
 627         ir_node *mem, *ptr, *new_node;
 628         ir_entity *ent;
 629         unsigned res = 0;
 630
 631         /* do NOT touch volatile loads for now */
 632         if (get_Load_volatility(load) == volatility_is_volatile)
 633                 return 0;
 634
 635         /* the address of the load to be optimized */
 636         ptr = get_Load_ptr(load);
 637
 638         /*
 639          * Check if we can remove the exception from a Load:
 640          * This can be done, if the address is from an Sel(Alloc) and
 641          * the Sel type is a subtype of the allocated type.
 642          *
 643          * This optimizes some often used OO constructs,
 644          * like x = new O; x->t;
 645          */
 646         if (info->projs[pn_Load_X_except]) {
 647                 if (is_Sel(ptr)) {
 648                         ir_node *mem = get_Sel_mem(ptr);
 649
 650                         /* FIXME: works with the current FE, but better use the base */
 651                         if (get_irn_op(skip_Proj(mem)) == op_Alloc) {
 652                                 /* ok, check the types */
 653                                 ir_entity *ent    = get_Sel_entity(ptr);
 654                                 ir_type   *s_type = get_entity_type(ent);
 655                                 ir_type   *a_type = get_Alloc_type(mem);
 656
 657                                 if (is_SubClass_of(s_type, a_type)) {
 658                                         /* ok, condition met: there can't be an exception because
 659                                         * Alloc guarantees that enough memory was allocated */
 660
 661                                         exchange(info->projs[pn_Load_X_except], new_Bad());
 662                                         info->projs[pn_Load_X_except] = NULL;
 663                                         res |= CF_CHANGED;
 664                                 }
 665                         }
 666                 } else if ((get_irn_op(skip_Proj(ptr)) == op_Alloc) ||
 667                         ((get_irn_op(ptr) == op_Cast) && (get_irn_op(skip_Proj(get_Cast_op(ptr))) == op_Alloc))) {
 668                                 /* simple case: a direct load after an Alloc. Firm Alloc throw
 669                                  * an exception in case of out-of-memory. So, there is no way for an
 670                                  * exception in this load.
 671                                  * This code is constructed by the "exception lowering" in the Jack compiler.
 672                                  */
 673                                 exchange(info->projs[pn_Load_X_except], new_Bad());
 674                                 info->projs[pn_Load_X_except] = NULL;
 675                                 res |= CF_CHANGED;
 676                 }
 677         }
 678
 679         /* The mem of the Load. Must still be returned after optimization. */
 680         mem  = get_Load_mem(load);
 681
 682         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 683                 /* a Load which value is neither used nor exception checked, remove it */
 684                 exchange(info->projs[pn_Load_M], mem);
 685
 686                 exchange(load, new_Bad());
 687                 reduce_adr_usage(ptr);
 688                 return res | DF_CHANGED;
 689         }
 690
 691         /* Load from a constant polymorphic field, where we can resolve
 692            polymorphism. */
 693         new_node = transform_node_Load(load);
 694         if (new_node != load) {
 695                 if (info->projs[pn_Load_M]) {
 696                         exchange(info->projs[pn_Load_M], mem);
 697                         info->projs[pn_Load_M] = NULL;
 698                 }
 699                 if (info->projs[pn_Load_X_except]) {
 700                         exchange(info->projs[pn_Load_X_except], new_Bad());
 701                         info->projs[pn_Load_X_except] = NULL;
 702                 }
 703                 if (info->projs[pn_Load_res])
 704                         exchange(info->projs[pn_Load_res], new_node);
 705
 706                 exchange(load, new_Bad());
 707                 reduce_adr_usage(ptr);
 708                 return res | DF_CHANGED;
 709         }
 710
 711         /* check if we can determine the entity that will be loaded */
 712         ent = find_constant_entity(ptr);
 713         if (ent) {
 714                 if ((allocation_static == get_entity_allocation(ent)) &&
 715                         (visibility_external_allocated != get_entity_visibility(ent))) {
 716                         /* a static allocation that is not external: there should be NO exception
 717                          * when loading. */
 718
 719                         /* no exception, clear the info field as it might be checked later again */
 720                         if (info->projs[pn_Load_X_except]) {
 721                                 exchange(info->projs[pn_Load_X_except], new_Bad());
 722                                 info->projs[pn_Load_X_except] = NULL;
 723                                 res |= CF_CHANGED;
 724                         }
 725
 726                         if (variability_constant == get_entity_variability(ent)
 727                                 && is_atomic_entity(ent)) {
 728                                 /* Might not be atomic after
 729                                    lowering of Sels.  In this
 730                                    case we could also load, but
 731                                    it's more complicated. */
 732                                 /* more simpler case: we load the content of a constant value:
 733                                  * replace it by the constant itself
 734                                  */
 735
 736                                 /* no memory */
 737                                 if (info->projs[pn_Load_M]) {
 738                                         exchange(info->projs[pn_Load_M], mem);
 739                                         res |= DF_CHANGED;
 740                                 }
 741                                 /* no result :-) */
 742                                 if (info->projs[pn_Load_res]) {
 743                                         if (is_atomic_entity(ent)) {
 744                                                 ir_node *c = copy_const_value(get_irn_dbg_info(load), get_atomic_ent_value(ent));
 745
 746                                                 DBG_OPT_RC(load, c);
 747                                                 exchange(info->projs[pn_Load_res], c);
 748                                                 res |= DF_CHANGED;
 749                                         }
 750                                 }
 751                                 exchange(load, new_Bad());
 752                                 reduce_adr_usage(ptr);
 753                                 return res;
 754                         } else if (variability_constant == get_entity_variability(ent)) {
 755                                 compound_graph_path *path = get_accessed_path(ptr);
 756
 757                                 if (path) {
 758                                         ir_node *c;
 759
 760                                         assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
 761                                         /*
 762                                         {
 763                                                 int j;
 764                                                 for (j = 0; j < get_compound_graph_path_length(path); ++j) {
 765                                                         ir_entity *node = get_compound_graph_path_node(path, j);
 766                                                         fprintf(stdout, ".%s", get_entity_name(node));
 767                                                         if (is_Array_type(get_entity_owner(node)))
 768                                                                 fprintf(stdout, "[%d]", get_compound_graph_path_array_index(path, j));
 769                                                 }
 770                                                 printf("\n");
 771                                         }
 772                                         */
 773
 774                                         c = get_compound_ent_value_by_path(ent, path);
 775                                         free_compound_graph_path(path);
 776
 777                                         /* printf("  cons: "); DDMN(c); */
 778
 779                                         if (info->projs[pn_Load_M]) {
 780                                                 exchange(info->projs[pn_Load_M], mem);
 781                                                 res |= DF_CHANGED;
 782                                         }
 783                                         if (info->projs[pn_Load_res]) {
 784                                                 exchange(info->projs[pn_Load_res], copy_const_value(get_irn_dbg_info(load), c));
 785                                                 res |= DF_CHANGED;
 786                                         }
 787                                         exchange(load, new_Bad());
 788                                         reduce_adr_usage(ptr);
 789                                         return res;
 790                                 } else {
 791                                         /*  We can not determine a correct access path.  E.g., in jack, we load
 792                                         a byte from an object to generate an exception.   Happens in test program
 793                                         Reflectiontest.
 794                                         printf(">>>>>>>>>>>>> Found access to constant entity %s in function %s\n", get_entity_name(ent),
 795                                         get_entity_name(get_irg_entity(current_ir_graph)));
 796                                         printf("  load: "); DDMN(load);
 797                                         printf("  ptr:  "); DDMN(ptr);
 798                                         */
 799                                 }
 800                         }
 801                 }
 802         }
 803
 804         /* Check, if the address of this load is used more than once.
 805          * If not, this load cannot be removed in any case. */
 806         if (get_irn_n_uses(ptr) <= 1)
 807                 return res;
 808
 809         /*
 810          * follow the memory chain as long as there are only Loads
 811          * and try to replace current Load or Store by a previous one.
 812          * Note that in unreachable loops it might happen that we reach
 813          * load again, as well as we can fall into a cycle.
 814          * We break such cycles using a special visited flag.
 815          */
 816         INC_MASTER();
 817         res = follow_Mem_chain(load, skip_Proj(mem));
 818         return res;
 819 }  /* optimize_load */
 820
 821 /**
 822  * Check whether a value of mode new_mode would completely overwrite a value
 823  * of mode old_mode in memory.
 824  */
 825 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
 826 {
 827         return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
 828 }  /* is_completely_overwritten */
 829
 830 /**
 831  * follow the memory chain as long as there are only Loads and alias free Stores.
 832  *
 833  * INC_MASTER() must be called before dive into
 834  */
 835 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
 836         unsigned res = 0;
 837         ldst_info_t *info = get_irn_link(store);
 838         ir_node *pred;
 839         ir_node *ptr = get_Store_ptr(store);
 840         ir_node *mem = get_Store_mem(store);
 841         ir_node *value = get_Store_value(store);
 842         ir_mode *mode  = get_irn_mode(value);
 843         ir_node *block = get_nodes_block(store);
 844
 845         for (pred = curr; pred != store;) {
 846                 ldst_info_t *pred_info = get_irn_link(pred);
 847
 848                 /*
 849                  * BEWARE: one might think that checking the modes is useless, because
 850                  * if the pointers are identical, they refer to the same object.
 851                  * This is only true in strong typed languages, not is C were the following
 852                  * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
 853                  * However, if the mode that is written have a bigger  or equal size the the old
 854                  * one, the old value is completely overwritten and can be killed ...
 855                  */
 856                 if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
 857                     get_nodes_block(pred) == block &&
 858                     is_completely_overwritten(get_irn_mode(get_Store_value(pred)), mode)) {
 859                         /*
 860                          * a Store after a Store in the same block -- a write after write.
 861                          * We may remove the first Store, if it does not have an exception handler.
 862                          *
 863                          * TODO: What, if both have the same exception handler ???
 864                          */
 865                         if (get_Store_volatility(pred) != volatility_is_volatile && !pred_info->projs[pn_Store_X_except]) {
 866                                 DBG_OPT_WAW(pred, store);
 867                                 exchange( pred_info->projs[pn_Store_M], get_Store_mem(pred) );
 868                                 exchange(pred, new_Bad());
 869                                 reduce_adr_usage(ptr);
 870                                 return DF_CHANGED;
 871                         }
 872                 } else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
 873                            value == pred_info->projs[pn_Load_res]) {
 874                         /*
 875                          * a Store of a value after a Load -- a write after read.
 876                          * We may remove the second Store, if it does not have an exception handler.
 877                          */
 878                         if (! info->projs[pn_Store_X_except]) {
 879                                 DBG_OPT_WAR(store, pred);
 880                                 exchange( info->projs[pn_Store_M], mem );
 881                                 exchange(store, new_Bad());
 882                                 reduce_adr_usage(ptr);
 883                                 return DF_CHANGED;
 884                         }
 885                 }
 886
 887                 if (get_irn_op(pred) == op_Store) {
 888                         /* check if we can pass thru this store */
 889                         ir_alias_relation rel = get_alias_relation(
 890                                 current_ir_graph,
 891                                 get_Store_ptr(pred),
 892                                 get_irn_mode(get_Store_value(pred)),
 893                                 ptr, mode);
 894                         /* if the might be an alias, we cannot pass this Store */
 895                         if (rel != no_alias)
 896                                 break;
 897                         pred = skip_Proj(get_Store_mem(pred));
 898                 } else if (get_irn_op(pred) == op_Load) {
 899                         pred = skip_Proj(get_Load_mem(pred));
 900                 } else {
 901                         /* follow only Load chains */
 902                         break;
 903                 }
 904
 905                 /* check for cycles */
 906                 if (NODE_VISITED(pred_info))
 907                         break;
 908                 MARK_NODE(pred_info);
 909         }
 910
 911         if (get_irn_op(pred) == op_Sync) {
 912                 int i;
 913
 914                 /* handle all Sync predecessors */
 915                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
 916                         res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
 917                         if (res)
 918                                 break;
 919                 }
 920         }
 921         return res;
 922 }  /* follow_Mem_chain_for_Store */
 923
 924 /**
 925  * optimize a Store
 926  *
 927  * @param store  the Store node
 928  */
 929 static unsigned optimize_store(ir_node *store) {
 930         ir_node *ptr, *mem;
 931
 932         if (get_Store_volatility(store) == volatility_is_volatile)
 933                 return 0;
 934
 935         ptr = get_Store_ptr(store);
 936
 937         /* Check, if the address of this Store is used more than once.
 938          * If not, this Store cannot be removed in any case. */
 939         if (get_irn_n_uses(ptr) <= 1)
 940                 return 0;
 941
 942         mem = get_Store_mem(store);
 943
 944         /* follow the memory chain as long as there are only Loads */
 945         INC_MASTER();
 946         return follow_Mem_chain_for_Store(store, skip_Proj(mem));
 947 }  /* optimize_store */
 948
 949 /**
 950  * walker, optimizes Phi after Stores to identical places:
 951  * Does the following optimization:
 952  * @verbatim
 953  *
 954  *   val1   val2   val3          val1  val2  val3
 955  *    |      |      |               \    |    /
 956  *  Store  Store  Store              \   |   /
 957  *      \    |    /                   PhiData
 958  *       \   |   /                       |
 959  *        \  |  /                      Store
 960  *          PhiM
 961  *
 962  * @endverbatim
 963  * This reduces the number of stores and allows for predicated execution.
 964  * Moves Stores back to the end of a function which may be bad.
 965  *
 966  * This is only possible if the predecessor blocks have only one successor.
 967  */
 968 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
 969 {
 970         int i, n;
 971         ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
 972         ir_mode *mode;
 973         ir_node **inM, **inD, **projMs;
 974         int *idx;
 975         dbg_info *db = NULL;
 976         ldst_info_t *info;
 977         block_info_t *bl_info;
 978         unsigned res = 0;
 979
 980         /* Must be a memory Phi */
 981         if (get_irn_mode(phi) != mode_M)
 982                 return 0;
 983
 984         n = get_Phi_n_preds(phi);
 985         if (n <= 0)
 986                 return 0;
 987
 988         /* must be only one user */
 989         projM = get_Phi_pred(phi, 0);
 990         if (get_irn_n_edges(projM) != 1)
 991                 return 0;
 992
 993         store = skip_Proj(projM);
 994         old_store = store;
 995         if (get_irn_op(store) != op_Store)
 996                 return 0;
 997
 998         block = get_nodes_block(store);
 999
1000         /* abort on dead blocks */
1001         if (is_Block_dead(block))
1002                 return 0;
1003
1004         /* check if the block is post dominated by Phi-block
1005            and has no exception exit */
1006         bl_info = get_irn_link(block);
1007         if (bl_info->flags & BLOCK_HAS_EXC)
1008                 return 0;
1009
1010         phi_block = get_nodes_block(phi);
1011         if (! block_strictly_postdominates(phi_block, block))
1012                 return 0;
1013
1014         /* this is the address of the store */
1015         ptr  = get_Store_ptr(store);
1016         mode = get_irn_mode(get_Store_value(store));
1017         info = get_irn_link(store);
1018         exc  = info->exc_block;
1019
1020         for (i = 1; i < n; ++i) {
1021                 ir_node *pred = get_Phi_pred(phi, i);
1022
1023                 if (get_irn_n_edges(pred) != 1)
1024                         return 0;
1025
1026                 pred = skip_Proj(pred);
1027                 if (get_irn_op(pred) != op_Store)
1028                         return 0;
1029
1030                 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1031                         return 0;
1032
1033                 info = get_irn_link(pred);
1034
1035                 /* check, if all stores have the same exception flow */
1036                 if (exc != info->exc_block)
1037                         return 0;
1038
1039                 /* abort on dead blocks */
1040                 block = get_nodes_block(pred);
1041                 if (is_Block_dead(block))
1042                         return 0;
1043
1044                 /* check if the block is post dominated by Phi-block
1045                    and has no exception exit. Note that block must be different from
1046                    Phi-block, else we would move a Store from end End of a block to its
1047                    Start... */
1048                 bl_info = get_irn_link(block);
1049                 if (bl_info->flags & BLOCK_HAS_EXC)
1050                         return 0;
1051                 if (block == phi_block || ! block_postdominates(phi_block, block))
1052                         return 0;
1053         }
1054
1055         /*
1056          * ok, when we are here, we found all predecessors of a Phi that
1057          * are Stores to the same address and size. That means whatever
1058          * we do before we enter the block of the Phi, we do a Store.
1059          * So, we can move the Store to the current block:
1060          *
1061          *   val1    val2    val3          val1  val2  val3
1062          *    |       |       |               \    |    /
1063          * | Str | | Str | | Str |             \   |   /
1064          *      \     |     /                   PhiData
1065          *       \    |    /                       |
1066          *        \   |   /                       Str
1067          *           PhiM
1068          *
1069          * Is only allowed if the predecessor blocks have only one successor.
1070          */
1071
1072         NEW_ARR_A(ir_node *, projMs, n);
1073         NEW_ARR_A(ir_node *, inM, n);
1074         NEW_ARR_A(ir_node *, inD, n);
1075         NEW_ARR_A(int, idx, n);
1076
1077         /* Prepare: Collect all Store nodes.  We must do this
1078            first because we otherwise may loose a store when exchanging its
1079            memory Proj.
1080          */
1081         for (i = n - 1; i >= 0; --i) {
1082                 ir_node *store;
1083
1084                 projMs[i] = get_Phi_pred(phi, i);
1085                 assert(is_Proj(projMs[i]));
1086
1087                 store = get_Proj_pred(projMs[i]);
1088                 info  = get_irn_link(store);
1089
1090                 inM[i] = get_Store_mem(store);
1091                 inD[i] = get_Store_value(store);
1092                 idx[i] = info->exc_idx;
1093         }
1094         block = get_nodes_block(phi);
1095
1096         /* second step: create a new memory Phi */
1097         phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M);
1098
1099         /* third step: create a new data Phi */
1100         phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode);
1101
1102         /* rewire memory and kill the node */
1103         for (i = n - 1; i >= 0; --i) {
1104                 ir_node *proj  = projMs[i];
1105
1106                 if(is_Proj(proj)) {
1107                         ir_node *store = get_Proj_pred(proj);
1108                         exchange(proj, inM[i]);
1109                         kill_node(store);
1110                 }
1111         }
1112
1113         /* fourth step: create the Store */
1114         store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD);
1115 #ifdef DO_CACHEOPT
1116         co_set_irn_name(store, co_get_irn_ident(old_store));
1117 #endif
1118
1119         projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
1120
1121         info = get_ldst_info(store, wenv);
1122         info->projs[pn_Store_M] = projM;
1123
1124         /* fifths step: repair exception flow */
1125         if (exc) {
1126                 ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except);
1127
1128                 info->projs[pn_Store_X_except] = projX;
1129                 info->exc_block                = exc;
1130                 info->exc_idx                  = idx[0];
1131
1132                 for (i = 0; i < n; ++i) {
1133                         set_Block_cfgpred(exc, idx[i], projX);
1134                 }
1135
1136                 if (n > 1) {
1137                         /* the exception block should be optimized as some inputs are identical now */
1138                 }
1139
1140                 res |= CF_CHANGED;
1141         }
1142
1143         /* sixth step: replace old Phi */
1144         exchange(phi, projM);
1145
1146         return res | DF_CHANGED;
1147 }  /* optimize_phi */
1148
1149 /**
1150  * walker, do the optimizations
1151  */
1152 static void do_load_store_optimize(ir_node *n, void *env) {
1153         walk_env_t *wenv = env;
1154
1155         switch (get_irn_opcode(n)) {
1156
1157         case iro_Load:
1158                 wenv->changes |= optimize_load(n);
1159                 break;
1160
1161         case iro_Store:
1162                 wenv->changes |= optimize_store(n);
1163                 break;
1164
1165         case iro_Phi:
1166                 wenv->changes |= optimize_phi(n, wenv);
1167
1168         default:
1169                 ;
1170         }
1171 }  /* do_load_store_optimize */
1172
1173 /*
1174  * do the load store optimization
1175  */
1176 void optimize_load_store(ir_graph *irg) {
1177         walk_env_t env;
1178
1179         assert(get_irg_phase_state(irg) != phase_building);
1180         assert(get_irg_pinned(irg) != op_pin_state_floats &&
1181                 "LoadStore optimization needs pinned graph");
1182
1183         if (! get_opt_redundant_loadstore())
1184                 return;
1185
1186         edges_assure(irg);
1187
1188         /* for Phi optimization post-dominators are needed ... */
1189         assure_postdoms(irg);
1190
1191         if (get_opt_alias_analysis()) {
1192                 assure_irg_address_taken_computed(irg);
1193                 assure_irp_globals_address_taken_computed();
1194         }
1195
1196         obstack_init(&env.obst);
1197         env.changes = 0;
1198
1199         /* init the links, then collect Loads/Stores/Proj's in lists */
1200         master_visited = 0;
1201         irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
1202
1203         /* now we have collected enough information, optimize */
1204         irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
1205
1206         obstack_free(&env.obst, NULL);
1207
1208         /* Handle graph state */
1209         if (env.changes) {
1210                 set_irg_outs_inconsistent(irg);
1211         }
1212
1213         if (env.changes & CF_CHANGED) {
1214                 /* is this really needed: Yes, control flow changed, block might
1215                 have Bad() predecessors. */
1216                 set_irg_doms_inconsistent(irg);
1217         }
1218 }  /* optimize_load_store */