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

   1 /*
   2  * Copyright (C) 1995-2007 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief   Load/Store optimizations.
  23  * @author  Michael Beck
  24  * @version $Id$
  25  */
  26 #ifdef HAVE_CONFIG_H
  27 # include "config.h"
  28 #endif
  29
  30 #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 #include "irphase_t.h"
  53 #include "irgopt.h"
  54 #include "debug.h"
  55
  56 /** The debug handle. */
  57 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
  58
  59 #ifdef DO_CACHEOPT
  60 #include "cacheopt/cachesim.h"
  61 #endif
  62
  63 #undef IMAX
  64 #define IMAX(a,b)       ((a) > (b) ? (a) : (b))
  65
  66 #define MAX_PROJ        IMAX(pn_Load_max, pn_Store_max)
  67
  68 enum changes_t {
  69         DF_CHANGED = 1,       /**< data flow changed */
  70         CF_CHANGED = 2,       /**< control flow changed */
  71 };
  72
  73 /**
  74  * walker environment
  75  */
  76 typedef struct _walk_env_t {
  77         struct obstack obst;          /**< list of all stores */
  78         unsigned changes;             /**< a bitmask of graph changes */
  79 } walk_env_t;
  80
  81 /**
  82  * flags for Load/Store
  83  */
  84 enum ldst_flags_t {
  85         LDST_VISITED = 1              /**< if set, this Load/Store is already visited */
  86 };
  87
  88 /** A Load/Store info. */
  89 typedef struct _ldst_info_t {
  90         ir_node  *projs[MAX_PROJ];    /**< list of Proj's of this node */
  91         ir_node  *exc_block;          /**< the exception block if available */
  92         int      exc_idx;             /**< predecessor index in the exception block */
  93         unsigned flags;               /**< flags */
  94         unsigned visited;             /**< visited counter for breaking loops */
  95 } ldst_info_t;
  96
  97 /**
  98  * flags for control flow.
  99  */
 100 enum block_flags_t {
 101         BLOCK_HAS_COND = 1,      /**< Block has conditional control flow */
 102         BLOCK_HAS_EXC  = 2       /**< Block has exceptional control flow */
 103 };
 104
 105 /**
 106  * a Block info.
 107  */
 108 typedef struct _block_info_t {
 109         unsigned flags;               /**< flags for the block */
 110 } block_info_t;
 111
 112 /** the master visited flag for loop detection. */
 113 static unsigned master_visited = 0;
 114
 115 #define INC_MASTER()       ++master_visited
 116 #define MARK_NODE(info)    (info)->visited = master_visited
 117 #define NODE_VISITED(info) (info)->visited >= master_visited
 118
 119 /**
 120  * get the Load/Store info of a node
 121  */
 122 static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst) {
 123         ldst_info_t *info = get_irn_link(node);
 124
 125         if (! info) {
 126                 info = obstack_alloc(obst, sizeof(*info));
 127                 memset(info, 0, sizeof(*info));
 128                 set_irn_link(node, info);
 129         }
 130         return info;
 131 }  /* get_ldst_info */
 132
 133 /**
 134  * get the Block info of a node
 135  */
 136 static block_info_t *get_block_info(ir_node *node, struct obstack *obst) {
 137         block_info_t *info = get_irn_link(node);
 138
 139         if (! info) {
 140                 info = obstack_alloc(obst, sizeof(*info));
 141                 memset(info, 0, sizeof(*info));
 142                 set_irn_link(node, info);
 143         }
 144         return info;
 145 }  /* get_block_info */
 146
 147 /**
 148  * update the projection info for a Load/Store
 149  */
 150 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
 151 {
 152         long nr = get_Proj_proj(proj);
 153
 154         assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
 155
 156         if (info->projs[nr]) {
 157                 /* there is already one, do CSE */
 158                 exchange(proj, info->projs[nr]);
 159                 return DF_CHANGED;
 160         }
 161         else {
 162                 info->projs[nr] = proj;
 163                 return 0;
 164         }
 165 }  /* update_projs */
 166
 167 /**
 168  * update the exception block info for a Load/Store node.
 169  *
 170  * @param info   the load/store info struct
 171  * @param block  the exception handler block for this load/store
 172  * @param pos    the control flow input of the block
 173  */
 174 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
 175 {
 176         assert(info->exc_block == NULL && "more than one exception block found");
 177
 178         info->exc_block = block;
 179         info->exc_idx   = pos;
 180         return 0;
 181 }  /* update_exc */
 182
 183 /** Return the number of uses of an address node */
 184 #define get_irn_n_uses(adr)     get_irn_n_edges(adr)
 185
 186 /**
 187  * walker, collects all Load/Store/Proj nodes
 188  *
 189  * walks from Start -> End
 190  */
 191 static void collect_nodes(ir_node *node, void *env)
 192 {
 193         ir_op       *op = get_irn_op(node);
 194         ir_node     *pred, *blk, *pred_blk;
 195         ldst_info_t *ldst_info;
 196         walk_env_t  *wenv = env;
 197
 198         if (op == op_Proj) {
 199                 ir_node *adr;
 200                 ir_op *op;
 201
 202                 pred = get_Proj_pred(node);
 203                 op   = get_irn_op(pred);
 204
 205                 if (op == op_Load) {
 206                         ldst_info = get_ldst_info(pred, &wenv->obst);
 207
 208                         wenv->changes |= update_projs(ldst_info, node);
 209
 210                         if ((ldst_info->flags & LDST_VISITED) == 0) {
 211                                 adr = get_Load_ptr(pred);
 212                                 ldst_info->flags |= LDST_VISITED;
 213                         }
 214
 215                         /*
 216                          * Place the Proj's to the same block as the
 217                          * predecessor Load. This is always ok and prevents
 218                          * "non-SSA" form after optimizations if the Proj
 219                          * is in a wrong block.
 220                          */
 221                         blk      = get_nodes_block(node);
 222                         pred_blk = get_nodes_block(pred);
 223                         if (blk != pred_blk) {
 224                                 wenv->changes |= DF_CHANGED;
 225                                 set_nodes_block(node, pred_blk);
 226                         }
 227                 } else if (op == op_Store) {
 228                         ldst_info = get_ldst_info(pred, &wenv->obst);
 229
 230                         wenv->changes |= update_projs(ldst_info, node);
 231
 232                         if ((ldst_info->flags & LDST_VISITED) == 0) {
 233                                 adr = get_Store_ptr(pred);
 234                                 ldst_info->flags |= LDST_VISITED;
 235                         }
 236
 237                         /*
 238                         * Place the Proj's to the same block as the
 239                         * predecessor Store. This is always ok and prevents
 240                         * "non-SSA" form after optimizations if the Proj
 241                         * is in a wrong block.
 242                         */
 243                         blk      = get_nodes_block(node);
 244                         pred_blk = get_nodes_block(pred);
 245                         if (blk != pred_blk) {
 246                                 wenv->changes |= DF_CHANGED;
 247                                 set_nodes_block(node, pred_blk);
 248                         }
 249                 }
 250         } else if (op == op_Block) {
 251                 int i;
 252
 253                 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
 254                         ir_node      *pred_block, *proj;
 255                         block_info_t *bl_info;
 256                         int          is_exc = 0;
 257
 258                         pred = proj = get_Block_cfgpred(node, i);
 259
 260                         if (is_Proj(proj)) {
 261                                 pred   = get_Proj_pred(proj);
 262                                 is_exc = get_Proj_proj(proj) == pn_Generic_X_except;
 263                         }
 264
 265                         /* ignore Bad predecessors, they will be removed later */
 266                         if (is_Bad(pred))
 267                                 continue;
 268
 269                         pred_block = get_nodes_block(pred);
 270                         bl_info    = get_block_info(pred_block, &wenv->obst);
 271
 272                         if (is_fragile_op(pred) && is_exc)
 273                                 bl_info->flags |= BLOCK_HAS_EXC;
 274                         else if (is_irn_forking(pred))
 275                                 bl_info->flags |= BLOCK_HAS_COND;
 276
 277                         if (is_exc && (get_irn_op(pred) == op_Load || get_irn_op(pred) == op_Store)) {
 278                                 ldst_info = get_ldst_info(pred, &wenv->obst);
 279
 280                                 wenv->changes |= update_exc(ldst_info, node, i);
 281                         }
 282                 }
 283         }
 284 }  /* collect_nodes */
 285
 286 /**
 287  * Returns an entity if the address ptr points to a constant one.
 288  *
 289  * @param ptr  the address
 290  *
 291  * @return an entity or NULL
 292  */
 293 static ir_entity *find_constant_entity(ir_node *ptr)
 294 {
 295         for (;;) {
 296                 ir_op *op = get_irn_op(ptr);
 297
 298                 if (op == op_SymConst && (get_SymConst_kind(ptr) == symconst_addr_ent)) {
 299                         ir_entity *ent = get_SymConst_entity(ptr);
 300                         if (variability_constant == get_entity_variability(ent))
 301                                 return ent;
 302                         return NULL;
 303                 } else if (op == op_Sel) {
 304                         ir_entity *ent = get_Sel_entity(ptr);
 305                         ir_type   *tp  = get_entity_owner(ent);
 306
 307                         /* Do not fiddle with polymorphism. */
 308                         if (is_Class_type(get_entity_owner(ent)) &&
 309                                 ((get_entity_n_overwrites(ent)    != 0) ||
 310                                 (get_entity_n_overwrittenby(ent) != 0)   ) )
 311                                 return NULL;
 312
 313                         if (is_Array_type(tp)) {
 314                                 /* check bounds */
 315                                 int i, n;
 316
 317                                 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
 318                                         ir_node *bound;
 319                                         tarval *tlower, *tupper;
 320                                         ir_node *index = get_Sel_index(ptr, i);
 321                                         tarval *tv     = computed_value(index);
 322
 323                                         /* check if the index is constant */
 324                                         if (tv == tarval_bad)
 325                                                 return NULL;
 326
 327                                         bound  = get_array_lower_bound(tp, i);
 328                                         tlower = computed_value(bound);
 329                                         bound  = get_array_upper_bound(tp, i);
 330                                         tupper = computed_value(bound);
 331
 332                                         if (tlower == tarval_bad || tupper == tarval_bad)
 333                                                 return NULL;
 334
 335                                         if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
 336                                                 return NULL;
 337                                         if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
 338                                                 return NULL;
 339
 340                                         /* ok, bounds check finished */
 341                                 }
 342                         }
 343
 344                         if (variability_constant == get_entity_variability(ent))
 345                                 return ent;
 346
 347                         /* try next */
 348                         ptr = get_Sel_ptr(ptr);
 349                 } else
 350                         return NULL;
 351         }
 352 }  /* find_constant_entity */
 353
 354 /**
 355  * Return the Selection index of a Sel node from dimension n
 356  */
 357 static long get_Sel_array_index_long(ir_node *n, int dim) {
 358         ir_node *index = get_Sel_index(n, dim);
 359         assert(get_irn_op(index) == op_Const);
 360         return get_tarval_long(get_Const_tarval(index));
 361 }  /* get_Sel_array_index_long */
 362
 363 /**
 364  * Returns the accessed component graph path for an
 365  * node computing an address.
 366  *
 367  * @param ptr    the node computing the address
 368  * @param depth  current depth in steps upward from the root
 369  *               of the address
 370  */
 371 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
 372         compound_graph_path *res = NULL;
 373         ir_entity           *root, *field;
 374         int                 path_len, pos;
 375
 376         if (get_irn_op(ptr) == op_SymConst) {
 377                 /* a SymConst. If the depth is 0, this is an access to a global
 378                  * entity and we don't need a component path, else we know
 379                  * at least it's length.
 380                  */
 381                 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
 382                 root = get_SymConst_entity(ptr);
 383                 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
 384         } else {
 385                 assert(get_irn_op(ptr) == op_Sel);
 386                 /* it's a Sel, go up until we find the root */
 387                 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
 388
 389                 /* fill up the step in the path at the current position */
 390                 field    = get_Sel_entity(ptr);
 391                 path_len = get_compound_graph_path_length(res);
 392                 pos      = path_len - depth - 1;
 393                 set_compound_graph_path_node(res, pos, field);
 394
 395                 if (is_Array_type(get_entity_owner(field))) {
 396                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 397                         set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
 398                 }
 399         }
 400         return res;
 401 }  /* rec_get_accessed_path */
 402
 403 /** Returns an access path or NULL.  The access path is only
 404  *  valid, if the graph is in phase_high and _no_ address computation is used.
 405  */
 406 static compound_graph_path *get_accessed_path(ir_node *ptr) {
 407         return rec_get_accessed_path(ptr, 0);
 408 }  /* get_accessed_path */
 409
 410 /* forward */
 411 static void reduce_adr_usage(ir_node *ptr);
 412
 413 /**
 414  * Update a Load that may lost it's usage.
 415  */
 416 static void handle_load_update(ir_node *load) {
 417         ldst_info_t *info = get_irn_link(load);
 418
 419         /* do NOT touch volatile loads for now */
 420         if (get_Load_volatility(load) == volatility_is_volatile)
 421                 return;
 422
 423         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 424                 ir_node *ptr = get_Load_ptr(load);
 425                 ir_node *mem = get_Load_mem(load);
 426
 427                 /* a Load which value is neither used nor exception checked, remove it */
 428                 exchange(info->projs[pn_Load_M], mem);
 429                 if (info->projs[pn_Load_X_regular])
 430                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
 431                 exchange(load, new_Bad());
 432                 reduce_adr_usage(ptr);
 433         }
 434 }  /* handle_load_update */
 435
 436 /**
 437  * A Use of an address node is vanished. Check if this was a Proj
 438  * node and update the counters.
 439  */
 440 static void reduce_adr_usage(ir_node *ptr) {
 441         if (is_Proj(ptr)) {
 442                 if (get_irn_n_edges(ptr) <= 0) {
 443                         /* this Proj is dead now */
 444                         ir_node *pred = get_Proj_pred(ptr);
 445
 446                         if (is_Load(pred)) {
 447                                 ldst_info_t *info = get_irn_link(pred);
 448                                 info->projs[get_Proj_proj(ptr)] = NULL;
 449
 450                                 /* this node lost it's result proj, handle that */
 451                                 handle_load_update(pred);
 452                         }
 453                 }
 454         }
 455 }  /* reduce_adr_usage */
 456
 457 /**
 458  * Check, if an already existing value of mode old_mode can be converted
 459  * into the needed one new_mode without loss.
 460  */
 461 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
 462         if (old_mode == new_mode)
 463                 return 1;
 464
 465         /* if both modes are two-complement ones, we can always convert the
 466            Stored value into the needed one. */
 467         if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
 468                   get_mode_arithmetic(old_mode) == irma_twos_complement &&
 469                   get_mode_arithmetic(new_mode) == irma_twos_complement)
 470                 return 1;
 471         return 0;
 472 }  /* can_use_stored_value */
 473
 474 /**
 475  * Follow the memory chain as long as there are only Loads
 476  * and alias free Stores and try to replace current Load or Store
 477  * by a previous ones.
 478  * Note that in unreachable loops it might happen that we reach
 479  * load again, as well as we can fall into a cycle.
 480  * We break such cycles using a special visited flag.
 481  *
 482  * INC_MASTER() must be called before dive into
 483  */
 484 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
 485         unsigned res = 0;
 486         ldst_info_t *info = get_irn_link(load);
 487         ir_node *pred;
 488         ir_node *ptr       = get_Load_ptr(load);
 489         ir_node *mem       = get_Load_mem(load);
 490         ir_mode *load_mode = get_Load_mode(load);
 491
 492         for (pred = curr; load != pred; ) {
 493                 ldst_info_t *pred_info = get_irn_link(pred);
 494
 495                 /*
 496                  * BEWARE: one might think that checking the modes is useless, because
 497                  * if the pointers are identical, they refer to the same object.
 498                  * This is only true in strong typed languages, not in C were the following
 499                  * is possible a = *(ir_type1 *)p; b = *(ir_type2 *)p ...
 500                  */
 501                 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
 502                     can_use_stored_value(get_irn_mode(get_Store_value(pred)), load_mode)) {
 503                         /*
 504                          * a Load immediately after a Store -- a read after write.
 505                          * We may remove the Load, if both Load & Store does not have an exception handler
 506                          * OR they are in the same MacroBlock. In the latter case the Load cannot
 507                          * throw an exception when the previous Store was quiet.
 508                          *
 509                          * Why we need to check for Store Exception? If the Store cannot
 510                          * be executed (ROM) the exception handler might simply jump into
 511                          * the load MacroBlock :-(
 512                          * We could make it a little bit better if we would know that the exception
 513                          * handler of the Store jumps directly to the end...
 514                          */
 515                         if ((pred_info->projs[pn_Store_X_except] == NULL && info->projs[pn_Load_X_except] == NULL) ||
 516                             get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) {
 517                                 ir_node *value = get_Store_value(pred);
 518
 519                                 DBG_OPT_RAW(load, value);
 520
 521                                 /* add an convert if needed */
 522                                 if (get_irn_mode(get_Store_value(pred)) != load_mode) {
 523                                         value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
 524                                 }
 525
 526                                 if (info->projs[pn_Load_M])
 527                                         exchange(info->projs[pn_Load_M], mem);
 528
 529                                 /* no exception */
 530                                 if (info->projs[pn_Load_X_except]) {
 531                                         exchange( info->projs[pn_Load_X_except], new_Bad());
 532                                         res |= CF_CHANGED;
 533                                 }
 534                                 if (info->projs[pn_Load_X_regular]) {
 535                                         exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
 536                                         res |= CF_CHANGED;
 537                                 }
 538
 539                                 if (info->projs[pn_Load_res])
 540                                         exchange(info->projs[pn_Load_res], value);
 541
 542                                 exchange(load, new_Bad());
 543                                 reduce_adr_usage(ptr);
 544                                 return res | DF_CHANGED;
 545                         }
 546                 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
 547                            can_use_stored_value(get_Load_mode(pred), load_mode)) {
 548                         /*
 549                          * a Load after a Load -- a read after read.
 550                          * We may remove the second Load, if it does not have an exception handler
 551                          * OR they are in the same MacroBlock. In the later case the Load cannot
 552                          * throw an exception when the previous Load was quiet.
 553                          *
 554                          * Here, there is no need to check if the previous Load has an exception
 555                          * hander because they would have exact the same exception...
 556                          */
 557                         if (info->projs[pn_Load_X_except] == NULL || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) {
 558                                 ir_node *value;
 559
 560                                 DBG_OPT_RAR(load, pred);
 561
 562                                 /* the result is used */
 563                                 if (info->projs[pn_Load_res]) {
 564                                         if (pred_info->projs[pn_Load_res] == NULL) {
 565                                                 /* create a new Proj again */
 566                                                 pred_info->projs[pn_Load_res] = new_r_Proj(current_ir_graph, get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res);
 567                                         }
 568                                         value = pred_info->projs[pn_Load_res];
 569
 570                                         /* add an convert if needed */
 571                                         if (get_Load_mode(pred) != load_mode) {
 572                                                 value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
 573                                         }
 574
 575                                         exchange(info->projs[pn_Load_res], value);
 576                                 }
 577
 578                                 if (info->projs[pn_Load_M])
 579                                         exchange(info->projs[pn_Load_M], mem);
 580
 581                                 /* no exception */
 582                                 if (info->projs[pn_Load_X_except]) {
 583                                         exchange(info->projs[pn_Load_X_except], new_Bad());
 584                                         res |= CF_CHANGED;
 585                                 }
 586                                 if (info->projs[pn_Load_X_regular]) {
 587                                         exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
 588                                         res |= CF_CHANGED;
 589                                 }
 590
 591                                 exchange(load, new_Bad());
 592                                 reduce_adr_usage(ptr);
 593                                 return res |= DF_CHANGED;
 594                         }
 595                 }
 596
 597                 if (is_Store(pred)) {
 598                         /* check if we can pass through this store */
 599                         ir_alias_relation rel = get_alias_relation(
 600                                 current_ir_graph,
 601                                 get_Store_ptr(pred),
 602                                 get_irn_mode(get_Store_value(pred)),
 603                                 ptr, load_mode);
 604                         /* if the might be an alias, we cannot pass this Store */
 605                         if (rel != no_alias)
 606                                 break;
 607                         pred = skip_Proj(get_Store_mem(pred));
 608                 } else if (get_irn_op(pred) == op_Load) {
 609                         pred = skip_Proj(get_Load_mem(pred));
 610                 } else {
 611                         /* follow only Load chains */
 612                         break;
 613                 }
 614
 615                 /* check for cycles */
 616                 if (NODE_VISITED(pred_info))
 617                         break;
 618                 MARK_NODE(pred_info);
 619         }
 620
 621         if (is_Sync(pred)) {
 622                 int i;
 623
 624                 /* handle all Sync predecessors */
 625                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
 626                         res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
 627                         if (res)
 628                                 break;
 629                 }
 630         }
 631
 632         return res;
 633 }  /* follow_Mem_chain */
 634
 635 /**
 636  * optimize a Load
 637  *
 638  * @param load  the Load node
 639  */
 640 static unsigned optimize_load(ir_node *load)
 641 {
 642         ldst_info_t *info = get_irn_link(load);
 643         ir_node *mem, *ptr, *new_node;
 644         ir_entity *ent;
 645         unsigned res = 0;
 646
 647         /* do NOT touch volatile loads for now */
 648         if (get_Load_volatility(load) == volatility_is_volatile)
 649                 return 0;
 650
 651         /* the address of the load to be optimized */
 652         ptr = get_Load_ptr(load);
 653
 654         /*
 655          * Check if we can remove the exception from a Load:
 656          * This can be done, if the address is from an Sel(Alloc) and
 657          * the Sel type is a subtype of the allocated type.
 658          *
 659          * This optimizes some often used OO constructs,
 660          * like x = new O; x->t;
 661          */
 662         if (info->projs[pn_Load_X_except]) {
 663                 if (is_Sel(ptr)) {
 664                         ir_node *mem = get_Sel_mem(ptr);
 665
 666                         /* FIXME: works with the current FE, but better use the base */
 667                         if (is_Alloc(skip_Proj(mem))) {
 668                                 /* ok, check the types */
 669                                 ir_entity *ent    = get_Sel_entity(ptr);
 670                                 ir_type   *s_type = get_entity_type(ent);
 671                                 ir_type   *a_type = get_Alloc_type(mem);
 672
 673                                 if (is_SubClass_of(s_type, a_type)) {
 674                                         /* ok, condition met: there can't be an exception because
 675                                         * Alloc guarantees that enough memory was allocated */
 676
 677                                         exchange(info->projs[pn_Load_X_except], new_Bad());
 678                                         info->projs[pn_Load_X_except] = NULL;
 679                                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
 680                                         info->projs[pn_Load_X_regular] = NULL;
 681                                         res |= CF_CHANGED;
 682                                 }
 683                         }
 684                 } else if (is_Alloc(skip_Proj(skip_Cast(ptr)))) {
 685                                 /* simple case: a direct load after an Alloc. Firm Alloc throw
 686                                  * an exception in case of out-of-memory. So, there is no way for an
 687                                  * exception in this load.
 688                                  * This code is constructed by the "exception lowering" in the Jack compiler.
 689                                  */
 690                                 exchange(info->projs[pn_Load_X_except], new_Bad());
 691                                 info->projs[pn_Load_X_except] = NULL;
 692                                 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
 693                                 info->projs[pn_Load_X_regular] = NULL;
 694                                 res |= CF_CHANGED;
 695                 }
 696         }
 697
 698         /* The mem of the Load. Must still be returned after optimization. */
 699         mem  = get_Load_mem(load);
 700
 701         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 702                 /* a Load which value is neither used nor exception checked, remove it */
 703                 exchange(info->projs[pn_Load_M], mem);
 704
 705                 if (info->projs[pn_Load_X_regular]) {
 706                         /* should not happen, but if it does, remove it */
 707                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
 708                         res |= CF_CHANGED;
 709                 }
 710                 exchange(load, new_Bad());
 711                 reduce_adr_usage(ptr);
 712                 return res | DF_CHANGED;
 713         }
 714
 715         /* Load from a constant polymorphic field, where we can resolve
 716            polymorphism. */
 717         new_node = transform_node_Load(load);
 718         if (new_node != load) {
 719                 if (info->projs[pn_Load_M]) {
 720                         exchange(info->projs[pn_Load_M], mem);
 721                         info->projs[pn_Load_M] = NULL;
 722                 }
 723                 if (info->projs[pn_Load_X_except]) {
 724                         exchange(info->projs[pn_Load_X_except], new_Bad());
 725                         info->projs[pn_Load_X_except] = NULL;
 726                         res |= CF_CHANGED;
 727                 }
 728                 if (info->projs[pn_Load_X_regular]) {
 729                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
 730                         info->projs[pn_Load_X_regular] = NULL;
 731                         res |= CF_CHANGED;
 732                 }
 733                 if (info->projs[pn_Load_res])
 734                         exchange(info->projs[pn_Load_res], new_node);
 735
 736                 exchange(load, new_Bad());
 737                 reduce_adr_usage(ptr);
 738                 return res | DF_CHANGED;
 739         }
 740
 741         /* check if we can determine the entity that will be loaded */
 742         ent = find_constant_entity(ptr);
 743         if (ent) {
 744                 if ((allocation_static == get_entity_allocation(ent)) &&
 745                         (visibility_external_allocated != get_entity_visibility(ent))) {
 746                         /* a static allocation that is not external: there should be NO exception
 747                          * when loading. */
 748
 749                         /* no exception, clear the info field as it might be checked later again */
 750                         if (info->projs[pn_Load_X_except]) {
 751                                 exchange(info->projs[pn_Load_X_except], new_Bad());
 752                                 info->projs[pn_Load_X_except] = NULL;
 753                                 res |= CF_CHANGED;
 754                         }
 755                         if (info->projs[pn_Load_X_regular]) {
 756                                 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
 757                                 info->projs[pn_Load_X_regular] = NULL;
 758                                 res |= CF_CHANGED;
 759                         }
 760
 761                         if (variability_constant == get_entity_variability(ent)
 762                                 && is_atomic_entity(ent)) {
 763                                 /* Might not be atomic after
 764                                    lowering of Sels.  In this
 765                                    case we could also load, but
 766                                    it's more complicated. */
 767                                 /* more simpler case: we load the content of a constant value:
 768                                  * replace it by the constant itself
 769                                  */
 770
 771                                 /* no memory */
 772                                 if (info->projs[pn_Load_M]) {
 773                                         exchange(info->projs[pn_Load_M], mem);
 774                                         res |= DF_CHANGED;
 775                                 }
 776                                 /* no result :-) */
 777                                 if (info->projs[pn_Load_res]) {
 778                                         if (is_atomic_entity(ent)) {
 779                                                 ir_node *c = copy_const_value(get_irn_dbg_info(load), get_atomic_ent_value(ent));
 780
 781                                                 DBG_OPT_RC(load, c);
 782                                                 exchange(info->projs[pn_Load_res], c);
 783                                                 res |= DF_CHANGED;
 784                                         }
 785                                 }
 786                                 exchange(load, new_Bad());
 787                                 reduce_adr_usage(ptr);
 788                                 return res;
 789                         } else if (variability_constant == get_entity_variability(ent)) {
 790                                 compound_graph_path *path = get_accessed_path(ptr);
 791
 792                                 if (path) {
 793                                         ir_node *c;
 794
 795                                         assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
 796                                         /*
 797                                         {
 798                                                 int j;
 799                                                 for (j = 0; j < get_compound_graph_path_length(path); ++j) {
 800                                                         ir_entity *node = get_compound_graph_path_node(path, j);
 801                                                         fprintf(stdout, ".%s", get_entity_name(node));
 802                                                         if (is_Array_type(get_entity_owner(node)))
 803                                                                 fprintf(stdout, "[%d]", get_compound_graph_path_array_index(path, j));
 804                                                 }
 805                                                 printf("\n");
 806                                         }
 807                                         */
 808
 809                                         c = get_compound_ent_value_by_path(ent, path);
 810                                         free_compound_graph_path(path);
 811
 812                                         /* printf("  cons: "); DDMN(c); */
 813
 814                                         if (info->projs[pn_Load_M]) {
 815                                                 exchange(info->projs[pn_Load_M], mem);
 816                                                 res |= DF_CHANGED;
 817                                         }
 818                                         if (info->projs[pn_Load_res]) {
 819                                                 exchange(info->projs[pn_Load_res], copy_const_value(get_irn_dbg_info(load), c));
 820                                                 res |= DF_CHANGED;
 821                                         }
 822                                         exchange(load, new_Bad());
 823                                         reduce_adr_usage(ptr);
 824                                         return res;
 825                                 } else {
 826                                         /*  We can not determine a correct access path.  E.g., in jack, we load
 827                                         a byte from an object to generate an exception.   Happens in test program
 828                                         Reflectiontest.
 829                                         printf(">>>>>>>>>>>>> Found access to constant entity %s in function %s\n", get_entity_name(ent),
 830                                         get_entity_name(get_irg_entity(current_ir_graph)));
 831                                         printf("  load: "); DDMN(load);
 832                                         printf("  ptr:  "); DDMN(ptr);
 833                                         */
 834                                 }
 835                         }
 836                 }
 837         }
 838
 839         /* Check, if the address of this load is used more than once.
 840          * If not, this load cannot be removed in any case. */
 841         if (get_irn_n_uses(ptr) <= 1)
 842                 return res;
 843
 844         /*
 845          * follow the memory chain as long as there are only Loads
 846          * and try to replace current Load or Store by a previous one.
 847          * Note that in unreachable loops it might happen that we reach
 848          * load again, as well as we can fall into a cycle.
 849          * We break such cycles using a special visited flag.
 850          */
 851         INC_MASTER();
 852         res = follow_Mem_chain(load, skip_Proj(mem));
 853         return res;
 854 }  /* optimize_load */
 855
 856 /**
 857  * Check whether a value of mode new_mode would completely overwrite a value
 858  * of mode old_mode in memory.
 859  */
 860 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
 861 {
 862         return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
 863 }  /* is_completely_overwritten */
 864
 865 /**
 866  * follow the memory chain as long as there are only Loads and alias free Stores.
 867  *
 868  * INC_MASTER() must be called before dive into
 869  */
 870 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
 871         unsigned res = 0;
 872         ldst_info_t *info = get_irn_link(store);
 873         ir_node *pred;
 874         ir_node *ptr = get_Store_ptr(store);
 875         ir_node *mem = get_Store_mem(store);
 876         ir_node *value = get_Store_value(store);
 877         ir_mode *mode  = get_irn_mode(value);
 878         ir_node *block = get_nodes_block(store);
 879         ir_node *mblk  = get_Block_MacroBlock(block);
 880
 881         for (pred = curr; pred != store;) {
 882                 ldst_info_t *pred_info = get_irn_link(pred);
 883
 884                 /*
 885                  * BEWARE: one might think that checking the modes is useless, because
 886                  * if the pointers are identical, they refer to the same object.
 887                  * This is only true in strong typed languages, not is C were the following
 888                  * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
 889                  * However, if the mode that is written have a bigger  or equal size the the old
 890                  * one, the old value is completely overwritten and can be killed ...
 891                  */
 892                 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
 893                     get_nodes_MacroBlock(pred) == mblk &&
 894                     is_completely_overwritten(get_irn_mode(get_Store_value(pred)), mode)) {
 895                         /*
 896                          * a Store after a Store in the same block -- a write after write.
 897                          * We may remove the first Store, if it does not have an exception handler.
 898                          *
 899                          * TODO: What, if both have the same exception handler ???
 900                          */
 901                         if (get_Store_volatility(pred) != volatility_is_volatile && !pred_info->projs[pn_Store_X_except]) {
 902                                 DBG_OPT_WAW(pred, store);
 903                                 exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
 904                                 exchange(pred, new_Bad());
 905                                 reduce_adr_usage(ptr);
 906                                 return DF_CHANGED;
 907                         }
 908                 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
 909                            value == pred_info->projs[pn_Load_res]) {
 910                         /*
 911                          * a Store of a value after a Load -- a write after read.
 912                          * We may remove the second Store, if it does not have an exception handler.
 913                          */
 914                         if (! info->projs[pn_Store_X_except]) {
 915                                 DBG_OPT_WAR(store, pred);
 916                                 exchange(info->projs[pn_Store_M], mem);
 917                                 exchange(store, new_Bad());
 918                                 reduce_adr_usage(ptr);
 919                                 return DF_CHANGED;
 920                         }
 921                 }
 922
 923                 if (is_Store(pred)) {
 924                         /* check if we can pass thru this store */
 925                         ir_alias_relation rel = get_alias_relation(
 926                                 current_ir_graph,
 927                                 get_Store_ptr(pred),
 928                                 get_irn_mode(get_Store_value(pred)),
 929                                 ptr, mode);
 930                         /* if the might be an alias, we cannot pass this Store */
 931                         if (rel != no_alias)
 932                                 break;
 933                         pred = skip_Proj(get_Store_mem(pred));
 934                 } else if (get_irn_op(pred) == op_Load) {
 935                         pred = skip_Proj(get_Load_mem(pred));
 936                 } else {
 937                         /* follow only Load chains */
 938                         break;
 939                 }
 940
 941                 /* check for cycles */
 942                 if (NODE_VISITED(pred_info))
 943                         break;
 944                 MARK_NODE(pred_info);
 945         }
 946
 947         if (is_Sync(pred)) {
 948                 int i;
 949
 950                 /* handle all Sync predecessors */
 951                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
 952                         res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
 953                         if (res)
 954                                 break;
 955                 }
 956         }
 957         return res;
 958 }  /* follow_Mem_chain_for_Store */
 959
 960 /**
 961  * optimize a Store
 962  *
 963  * @param store  the Store node
 964  */
 965 static unsigned optimize_store(ir_node *store) {
 966         ir_node *ptr, *mem;
 967
 968         if (get_Store_volatility(store) == volatility_is_volatile)
 969                 return 0;
 970
 971         ptr = get_Store_ptr(store);
 972
 973         /* Check, if the address of this Store is used more than once.
 974          * If not, this Store cannot be removed in any case. */
 975         if (get_irn_n_uses(ptr) <= 1)
 976                 return 0;
 977
 978         mem = get_Store_mem(store);
 979
 980         /* follow the memory chain as long as there are only Loads */
 981         INC_MASTER();
 982         return follow_Mem_chain_for_Store(store, skip_Proj(mem));
 983 }  /* optimize_store */
 984
 985 /**
 986  * walker, optimizes Phi after Stores to identical places:
 987  * Does the following optimization:
 988  * @verbatim
 989  *
 990  *   val1   val2   val3          val1  val2  val3
 991  *    |      |      |               \    |    /
 992  *  Store  Store  Store              \   |   /
 993  *      \    |    /                   PhiData
 994  *       \   |   /                       |
 995  *        \  |  /                      Store
 996  *          PhiM
 997  *
 998  * @endverbatim
 999  * This reduces the number of stores and allows for predicated execution.
1000  * Moves Stores back to the end of a function which may be bad.
1001  *
1002  * This is only possible if the predecessor blocks have only one successor.
1003  */
1004 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
1005 {
1006         int i, n;
1007         ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
1008         ir_mode *mode;
1009         ir_node **inM, **inD, **projMs;
1010         int *idx;
1011         dbg_info *db = NULL;
1012         ldst_info_t *info;
1013         block_info_t *bl_info;
1014         unsigned res = 0;
1015
1016         /* Must be a memory Phi */
1017         if (get_irn_mode(phi) != mode_M)
1018                 return 0;
1019
1020         n = get_Phi_n_preds(phi);
1021         if (n <= 0)
1022                 return 0;
1023
1024         /* must be only one user */
1025         projM = get_Phi_pred(phi, 0);
1026         if (get_irn_n_edges(projM) != 1)
1027                 return 0;
1028
1029         store = skip_Proj(projM);
1030         old_store = store;
1031         if (get_irn_op(store) != op_Store)
1032                 return 0;
1033
1034         block = get_nodes_block(store);
1035
1036         /* abort on dead blocks */
1037         if (is_Block_dead(block))
1038                 return 0;
1039
1040         /* check if the block is post dominated by Phi-block
1041            and has no exception exit */
1042         bl_info = get_irn_link(block);
1043         if (bl_info->flags & BLOCK_HAS_EXC)
1044                 return 0;
1045
1046         phi_block = get_nodes_block(phi);
1047         if (! block_strictly_postdominates(phi_block, block))
1048                 return 0;
1049
1050         /* this is the address of the store */
1051         ptr  = get_Store_ptr(store);
1052         mode = get_irn_mode(get_Store_value(store));
1053         info = get_irn_link(store);
1054         exc  = info->exc_block;
1055
1056         for (i = 1; i < n; ++i) {
1057                 ir_node *pred = get_Phi_pred(phi, i);
1058
1059                 if (get_irn_n_edges(pred) != 1)
1060                         return 0;
1061
1062                 pred = skip_Proj(pred);
1063                 if (!is_Store(pred))
1064                         return 0;
1065
1066                 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1067                         return 0;
1068
1069                 info = get_irn_link(pred);
1070
1071                 /* check, if all stores have the same exception flow */
1072                 if (exc != info->exc_block)
1073                         return 0;
1074
1075                 /* abort on dead blocks */
1076                 block = get_nodes_block(pred);
1077                 if (is_Block_dead(block))
1078                         return 0;
1079
1080                 /* check if the block is post dominated by Phi-block
1081                    and has no exception exit. Note that block must be different from
1082                    Phi-block, else we would move a Store from end End of a block to its
1083                    Start... */
1084                 bl_info = get_irn_link(block);
1085                 if (bl_info->flags & BLOCK_HAS_EXC)
1086                         return 0;
1087                 if (block == phi_block || ! block_postdominates(phi_block, block))
1088                         return 0;
1089         }
1090
1091         /*
1092          * ok, when we are here, we found all predecessors of a Phi that
1093          * are Stores to the same address and size. That means whatever
1094          * we do before we enter the block of the Phi, we do a Store.
1095          * So, we can move the Store to the current block:
1096          *
1097          *   val1    val2    val3          val1  val2  val3
1098          *    |       |       |               \    |    /
1099          * | Str | | Str | | Str |             \   |   /
1100          *      \     |     /                   PhiData
1101          *       \    |    /                       |
1102          *        \   |   /                       Str
1103          *           PhiM
1104          *
1105          * Is only allowed if the predecessor blocks have only one successor.
1106          */
1107
1108         NEW_ARR_A(ir_node *, projMs, n);
1109         NEW_ARR_A(ir_node *, inM, n);
1110         NEW_ARR_A(ir_node *, inD, n);
1111         NEW_ARR_A(int, idx, n);
1112
1113         /* Prepare: Collect all Store nodes.  We must do this
1114            first because we otherwise may loose a store when exchanging its
1115            memory Proj.
1116          */
1117         for (i = n - 1; i >= 0; --i) {
1118                 ir_node *store;
1119
1120                 projMs[i] = get_Phi_pred(phi, i);
1121                 assert(is_Proj(projMs[i]));
1122
1123                 store = get_Proj_pred(projMs[i]);
1124                 info  = get_irn_link(store);
1125
1126                 inM[i] = get_Store_mem(store);
1127                 inD[i] = get_Store_value(store);
1128                 idx[i] = info->exc_idx;
1129         }
1130         block = get_nodes_block(phi);
1131
1132         /* second step: create a new memory Phi */
1133         phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M);
1134
1135         /* third step: create a new data Phi */
1136         phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode);
1137
1138         /* rewire memory and kill the node */
1139         for (i = n - 1; i >= 0; --i) {
1140                 ir_node *proj  = projMs[i];
1141
1142                 if(is_Proj(proj)) {
1143                         ir_node *store = get_Proj_pred(proj);
1144                         exchange(proj, inM[i]);
1145                         kill_node(store);
1146                 }
1147         }
1148
1149         /* fourth step: create the Store */
1150         store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD);
1151 #ifdef DO_CACHEOPT
1152         co_set_irn_name(store, co_get_irn_ident(old_store));
1153 #endif
1154
1155         projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
1156
1157         info = get_ldst_info(store, &wenv->obst);
1158         info->projs[pn_Store_M] = projM;
1159
1160         /* fifths step: repair exception flow */
1161         if (exc) {
1162                 ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except);
1163
1164                 info->projs[pn_Store_X_except] = projX;
1165                 info->exc_block                = exc;
1166                 info->exc_idx                  = idx[0];
1167
1168                 for (i = 0; i < n; ++i) {
1169                         set_Block_cfgpred(exc, idx[i], projX);
1170                 }
1171
1172                 if (n > 1) {
1173                         /* the exception block should be optimized as some inputs are identical now */
1174                 }
1175
1176                 res |= CF_CHANGED;
1177         }
1178
1179         /* sixth step: replace old Phi */
1180         exchange(phi, projM);
1181
1182         return res | DF_CHANGED;
1183 }  /* optimize_phi */
1184
1185 /**
1186  * walker, do the optimizations
1187  */
1188 static void do_load_store_optimize(ir_node *n, void *env) {
1189         walk_env_t *wenv = env;
1190
1191         switch (get_irn_opcode(n)) {
1192
1193         case iro_Load:
1194                 wenv->changes |= optimize_load(n);
1195                 break;
1196
1197         case iro_Store:
1198                 wenv->changes |= optimize_store(n);
1199                 break;
1200
1201         case iro_Phi:
1202                 wenv->changes |= optimize_phi(n, wenv);
1203
1204         default:
1205                 ;
1206         }
1207 }  /* do_load_store_optimize */
1208
1209 /** A scc. */
1210 typedef struct scc {
1211         ir_node *head;          /**< the head of the list */
1212 } scc;
1213
1214 /** A node entry. */
1215 typedef struct node_entry {
1216         unsigned DFSnum;    /**< the DFS number of this node */
1217         unsigned low;       /**< the low number of this node */
1218         ir_node  *header;   /**< the header of this node */
1219         int      in_stack;  /**< flag, set if the node is on the stack */
1220         ir_node  *next;     /**< link to the next node the the same scc */
1221         scc      *pscc;     /**< the scc of this node */
1222         unsigned POnum;     /**< the post order number for blocks */
1223 } node_entry;
1224
1225 /** A loop entry. */
1226 typedef struct loop_env {
1227         ir_phase ph;           /**< the phase object */
1228         ir_node  **stack;      /**< the node stack */
1229         int      tos;          /**< tos index */
1230         unsigned nextDFSnum;   /**< the current DFS number */
1231         unsigned POnum;        /**< current post order number */
1232
1233         unsigned changes;      /**< a bitmask of graph changes */
1234 } loop_env;
1235
1236 /**
1237 * Gets the node_entry of a node
1238 */
1239 static node_entry *get_irn_ne(ir_node *irn, loop_env *env) {
1240         ir_phase   *ph = &env->ph;
1241         node_entry *e  = phase_get_irn_data(&env->ph, irn);
1242
1243         if (! e) {
1244                 e = phase_alloc(ph, sizeof(*e));
1245                 memset(e, 0, sizeof(*e));
1246                 phase_set_irn_data(ph, irn, e);
1247         }
1248         return e;
1249 }  /* get_irn_ne */
1250
1251 /**
1252  * Push a node onto the stack.
1253  *
1254  * @param env   the loop environment
1255  * @param n     the node to push
1256  */
1257 static void push(loop_env *env, ir_node *n) {
1258         node_entry *e;
1259
1260         if (env->tos == ARR_LEN(env->stack)) {
1261                 int nlen = ARR_LEN(env->stack) * 2;
1262                 ARR_RESIZE(ir_node *, env->stack, nlen);
1263         }
1264         env->stack[env->tos++] = n;
1265         e = get_irn_ne(n, env);
1266         e->in_stack = 1;
1267 }  /* push */
1268
1269 /**
1270  * pop a node from the stack
1271  *
1272  * @param env   the loop environment
1273  *
1274  * @return  The topmost node
1275  */
1276 static ir_node *pop(loop_env *env) {
1277         ir_node *n = env->stack[--env->tos];
1278         node_entry *e = get_irn_ne(n, env);
1279
1280         e->in_stack = 0;
1281         return n;
1282 }  /* pop */
1283
1284 /**
1285  * Check if irn is a region constant.
1286  * The block or irn must strictly dominate the header block.
1287  *
1288  * @param irn           the node to check
1289  * @param header_block  the header block of the induction variable
1290  */
1291 static int is_rc(ir_node *irn, ir_node *header_block) {
1292         ir_node *block = get_nodes_block(irn);
1293
1294         return (block != header_block) && block_dominates(block, header_block);
1295 }  /* is_rc */
1296
1297 typedef struct phi_entry phi_entry;
1298 struct phi_entry {
1299         ir_node   *phi;    /**< A phi with a region const memory. */
1300         int       pos;     /**< The position of the region const memory */
1301         ir_node   *load;   /**< the newly created load for this phi */
1302         phi_entry *next;
1303 };
1304
1305 /**
1306  * Move loops out of loops if possible
1307  */
1308 static void move_loads_in_loops(scc *pscc, loop_env *env) {
1309         ir_node   *phi, *load, *next, *other, *next_other;
1310         ir_entity *ent;
1311         int       j;
1312         phi_entry *phi_list = NULL;
1313
1314         /* collect all outer memories */
1315         for (phi = pscc->head; phi != NULL; phi = next) {
1316                 node_entry *ne = get_irn_ne(phi, env);
1317                 next = ne->next;
1318
1319                 /* check all memory Phi's */
1320                 if (! is_Phi(phi))
1321                         continue;
1322
1323                 assert(get_irn_mode(phi) == mode_M && "DFS geturn non-memory Phi");
1324
1325                 for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
1326                         ir_node    *pred = get_irn_n(phi, j);
1327                         node_entry *pe   = get_irn_ne(pred, env);
1328
1329                         if (pe->pscc != ne->pscc) {
1330                                 /* not in the same SCC, is region const */
1331                                 phi_entry *pe = phase_alloc(&env->ph, sizeof(*pe));
1332
1333                                 pe->phi  = phi;
1334                                 pe->pos  = j;
1335                                 pe->next = phi_list;
1336                                 phi_list = pe;
1337                         }
1338                 }
1339         }
1340         /* no Phis no fun */
1341         assert(phi_list != NULL && "DFS found a loop without Phi");
1342
1343         for (load = pscc->head; load; load = next) {
1344                 ir_mode *load_mode;
1345                 node_entry *ne = get_irn_ne(load, env);
1346                 next = ne->next;
1347
1348                 if (is_Load(load)) {
1349                         ldst_info_t *info = get_irn_link(load);
1350                         ir_node     *ptr = get_Load_ptr(load);
1351
1352                         /* for now, we cannot handle Loads with exceptions */
1353                         if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
1354                                 continue;
1355
1356                         /* for now, we can only handle Load(SymConst) */
1357                         if (! is_SymConst(ptr) || get_SymConst_kind(ptr) != symconst_addr_ent)
1358                                 continue;
1359                         ent = get_SymConst_entity(ptr);
1360                         load_mode = get_Load_mode(load);
1361                         for (other = pscc->head; other != NULL; other = next_other) {
1362                                 node_entry *ne = get_irn_ne(other, env);
1363                                 next_other = ne->next;
1364
1365                                 if (is_Store(other)) {
1366                                         ir_alias_relation rel = get_alias_relation(
1367                                                 current_ir_graph,
1368                                                 get_Store_ptr(other),
1369                                                 get_irn_mode(get_Store_value(other)),
1370                                                 ptr, load_mode);
1371                                         /* if the might be an alias, we cannot pass this Store */
1372                                         if (rel != no_alias)
1373                                                 break;
1374                                 }
1375                         }
1376                         if (other == NULL) {
1377                                 ldst_info_t *ninfo;
1378                                 phi_entry   *pe;
1379                                 dbg_info    *db;
1380
1381                                 /* for now, we cannot handle more than one input */
1382                                 if (phi_list->next != NULL)
1383                                         return;
1384
1385                                 /* yep, no aliasing Store found, Load can be moved */
1386                                 DB((dbg, LEVEL_1, "  Found a Load that could be moved: %+F\n", load));
1387
1388                                 db   = get_irn_dbg_info(load);
1389                                 for (pe = phi_list; pe != NULL; pe = pe->next) {
1390                                         int     pos   = pe->pos;
1391                                         ir_node *phi  = pe->phi;
1392                                         ir_node *blk  = get_nodes_block(phi);
1393                                         ir_node *pred = get_Block_cfgpred_block(blk, pos);
1394                                         ir_node *irn, *mem;
1395
1396                                         pe->load = irn = new_rd_Load(db, current_ir_graph, pred, get_Phi_pred(phi, pos), ptr, load_mode);
1397                                         ninfo = get_ldst_info(irn, phase_obst(&env->ph));
1398
1399                                         ninfo->projs[pn_Load_M] = mem = new_r_Proj(current_ir_graph, pred, irn, mode_M, pn_Load_M);
1400                                         set_Phi_pred(phi, pos, mem);
1401
1402                                         ninfo->projs[pn_Load_res] = new_r_Proj(current_ir_graph, pred, irn, load_mode, pn_Load_res);
1403
1404                                         DB((dbg, LEVEL_1, "  Created %+F in %+F\n", irn, pred));
1405                                 }
1406
1407                                 /* now kill the old Load */
1408                                 exchange(info->projs[pn_Load_M], get_Load_mem(load));
1409                                 exchange(info->projs[pn_Load_res], ninfo->projs[pn_Load_res]);
1410
1411                                 env->changes |= DF_CHANGED;
1412                         }
1413                 }
1414         }
1415 }  /* move_loads_in_loops */
1416
1417 /**
1418  * Process a loop SCC.
1419  *
1420  * @param pscc  the SCC
1421  * @param env   the loop environment
1422  */
1423 static void process_loop(scc *pscc, loop_env *env) {
1424         ir_node *irn, *next, *header = NULL;
1425         node_entry *b, *h = NULL;
1426         int j, only_phi, num_outside, process = 0;
1427         ir_node *out_rc;
1428
1429         /* find the header block for this scc */
1430         for (irn = pscc->head; irn; irn = next) {
1431                 node_entry *e = get_irn_ne(irn, env);
1432                 ir_node *block = get_nodes_block(irn);
1433
1434                 next = e->next;
1435                 b = get_irn_ne(block, env);
1436
1437                 if (header) {
1438                         if (h->POnum < b->POnum) {
1439                                 header = block;
1440                                 h      = b;
1441                         }
1442                 }
1443                 else {
1444                         header = block;
1445                         h      = b;
1446                 }
1447         }
1448
1449         /* check if this scc contains only Phi, Loads or Stores nodes */
1450         only_phi    = 1;
1451         num_outside = 0;
1452         out_rc      = NULL;
1453         for (irn = pscc->head; irn; irn = next) {
1454                 node_entry *e = get_irn_ne(irn, env);
1455
1456                 next = e->next;
1457                 switch (get_irn_opcode(irn)) {
1458                 case iro_Call:
1459                 case iro_CopyB:
1460                         /* cannot handle Calls or CopyB yet */
1461                         goto fail;
1462                 case iro_Load:
1463                         process = 1;
1464                         if (get_Load_volatility(irn) == volatility_is_volatile) {
1465                                 /* cannot handle loops with volatile Loads */
1466                                 goto fail;
1467                         }
1468                         only_phi = 0;
1469                         break;
1470                 case iro_Store:
1471                         if (get_Store_volatility(irn) == volatility_is_volatile) {
1472                                 /* cannot handle loops with volatile Stores */
1473                                 goto fail;
1474                         }
1475                         only_phi = 0;
1476                         break;
1477                 default:
1478                         only_phi = 0;
1479                         break;
1480                 case iro_Phi:
1481                         for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
1482                                 ir_node *pred  = get_irn_n(irn, j);
1483                                 node_entry *pe = get_irn_ne(pred, env);
1484
1485                                 if (pe->pscc != e->pscc) {
1486                                         /* not in the same SCC, must be a region const */
1487                                         if (! is_rc(pred, header)) {
1488                                                 /* not a memory loop */
1489                                                 goto fail;
1490                                         }
1491                                         if (! out_rc) {
1492                                                 out_rc = pred;
1493                                                 ++num_outside;
1494                                         } else if (out_rc != pred) {
1495                                                 ++num_outside;
1496                                         }
1497                                 }
1498                         }
1499                         break;
1500                 }
1501         }
1502         if (! process)
1503                 goto fail;
1504
1505         /* found a memory loop */
1506         DB((dbg, LEVEL_2, "  Found a memory loop:\n  "));
1507         if (only_phi && num_outside == 1) {
1508                 /* a phi cycle with only one real predecessor can be collapsed */
1509                 DB((dbg, LEVEL_2, "  Found an USELESS Phi cycle:\n  "));
1510
1511                 for (irn = pscc->head; irn; irn = next) {
1512                         node_entry *e = get_irn_ne(irn, env);
1513                         next = e->next;
1514                         e->header = NULL;
1515                         exchange(irn, out_rc);
1516                 }
1517                 env->changes |= DF_CHANGED;
1518                 return;
1519         }
1520
1521         /* set the header for every node in this scc */
1522         for (irn = pscc->head; irn; irn = next) {
1523                 node_entry *e = get_irn_ne(irn, env);
1524                 e->header = header;
1525                 next = e->next;
1526                 DB((dbg, LEVEL_2, " %+F,", irn));
1527         }
1528         DB((dbg, LEVEL_2, "\n"));
1529
1530         move_loads_in_loops(pscc, env);
1531
1532 fail:
1533         ;
1534 }  /* process_loop */
1535
1536 /**
1537  * Process a SCC.
1538  *
1539  * @param pscc  the SCC
1540  * @param env   the loop environment
1541  */
1542 static void process_scc(scc *pscc, loop_env *env) {
1543         ir_node *head = pscc->head;
1544         node_entry *e = get_irn_ne(head, env);
1545
1546 #ifdef DEBUG_libfirm
1547         {
1548                 ir_node *irn, *next;
1549
1550                 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
1551                 for (irn = pscc->head; irn; irn = next) {
1552                         node_entry *e = get_irn_ne(irn, env);
1553
1554                         next = e->next;
1555
1556                         DB((dbg, LEVEL_4, " %+F,", irn));
1557                 }
1558                 DB((dbg, LEVEL_4, "\n"));
1559         }
1560 #endif
1561
1562         if (e->next != NULL) {
1563                 /* this SCC has more than one member */
1564                 process_loop(pscc, env);
1565         }
1566 }  /* process_scc */
1567
1568 /**
1569  * Do Tarjan's SCC algorithm and drive load/store optimization.
1570  *
1571  * @param irn  start at this node
1572  * @param env  the loop environment
1573  */
1574 static void dfs(ir_node *irn, loop_env *env)
1575 {
1576         int i, n;
1577         node_entry *node = get_irn_ne(irn, env);
1578
1579         mark_irn_visited(irn);
1580
1581         node->DFSnum = env->nextDFSnum++;
1582         node->low    = node->DFSnum;
1583         push(env, irn);
1584
1585         /* handle preds */
1586         if (is_Phi(irn) || is_Sync(irn)) {
1587                 n = get_irn_arity(irn);
1588                 for (i = 0; i < n; ++i) {
1589                         ir_node *pred = get_irn_n(irn, i);
1590                         node_entry *o = get_irn_ne(pred, env);
1591
1592                         if (irn_not_visited(pred)) {
1593                                 dfs(pred, env);
1594                                 node->low = MIN(node->low, o->low);
1595                         }
1596                         if (o->DFSnum < node->DFSnum && o->in_stack)
1597                                 node->low = MIN(o->DFSnum, node->low);
1598                 }
1599         } else if (is_fragile_op(irn)) {
1600                 ir_node *pred = get_fragile_op_mem(irn);
1601                 node_entry *o = get_irn_ne(pred, env);
1602
1603                 if (irn_not_visited(pred)) {
1604                         dfs(pred, env);
1605                         node->low = MIN(node->low, o->low);
1606                 }
1607                 if (o->DFSnum < node->DFSnum && o->in_stack)
1608                         node->low = MIN(o->DFSnum, node->low);
1609         } else if (is_Proj(irn)) {
1610                 ir_node *pred = get_Proj_pred(irn);
1611                 node_entry *o = get_irn_ne(pred, env);
1612
1613                 if (irn_not_visited(pred)) {
1614                         dfs(pred, env);
1615                         node->low = MIN(node->low, o->low);
1616                 }
1617                 if (o->DFSnum < node->DFSnum && o->in_stack)
1618                         node->low = MIN(o->DFSnum, node->low);
1619         }
1620         else {
1621                  /* IGNORE predecessors */
1622         }
1623
1624         if (node->low == node->DFSnum) {
1625                 scc *pscc = phase_alloc(&env->ph, sizeof(*pscc));
1626                 ir_node *x;
1627
1628                 pscc->head = NULL;
1629                 do {
1630                         node_entry *e;
1631
1632                         x = pop(env);
1633                         e = get_irn_ne(x, env);
1634                         e->pscc    = pscc;
1635                         e->next    = pscc->head;
1636                         pscc->head = x;
1637                 } while (x != irn);
1638
1639                 process_scc(pscc, env);
1640         }
1641 }  /* dfs */
1642
1643 /**
1644  * Do the DFS on the memory edges a graph.
1645  *
1646  * @param irg  the graph to process
1647  * @param env  the loop environment
1648  */
1649 static void do_dfs(ir_graph *irg, loop_env *env) {
1650         ir_graph *rem = current_ir_graph;
1651         ir_node  *endblk, *end;
1652         int      i;
1653
1654         current_ir_graph = irg;
1655         inc_irg_visited(irg);
1656
1657         /* visit all memory nodes */
1658         endblk = get_irg_end_block(irg);
1659         for (i = get_Block_n_cfgpreds(endblk) - 1; i >= 0; --i) {
1660                 ir_node *pred = get_Block_cfgpred(endblk, i);
1661
1662                 if (is_Return(pred))
1663                         dfs(get_Return_mem(pred), env);
1664                 else if (is_Raise(pred))
1665                         dfs(get_Raise_mem(pred), env);
1666                 else if (is_fragile_op(pred))
1667                         dfs(get_fragile_op_mem(pred), env);
1668                 else {
1669                         assert(0 && "Unknown EndBlock predecessor");
1670                 }
1671         }
1672
1673         /* visit the keep-alives */
1674         end = get_irg_end(irg);
1675         for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
1676                 ir_node *ka = get_End_keepalive(end, i);
1677
1678                 if (is_Phi(ka) && irn_not_visited(ka))
1679                         dfs(ka, env);
1680         }
1681         current_ir_graph = rem;
1682 }  /* do_dfs */
1683
1684 /**
1685  * Initialize new phase data. We do this always explicit, so return NULL here
1686  */
1687 static void *init_loop_data(ir_phase *ph, ir_node *irn, void *data) {
1688         (void)ph;
1689         (void)irn;
1690         (void)data;
1691         return NULL;
1692 }  /* init_loop_data */
1693
1694 /**
1695  * Optimize Loads/Stores in loops.
1696  *
1697  * @param irg  the graph
1698  */
1699 static int optimize_loops(ir_graph *irg) {
1700         loop_env env;
1701
1702         env.stack         = NEW_ARR_F(ir_node *, 128);
1703         env.tos           = 0;
1704         env.nextDFSnum    = 0;
1705         env.POnum         = 0;
1706         env.changes       = 0;
1707         phase_init(&env.ph, "ldstopt", irg, PHASE_DEFAULT_GROWTH, init_loop_data, NULL);
1708
1709         /* calculate the SCC's and drive loop optimization. */
1710         do_dfs(irg, &env);
1711
1712         DEL_ARR_F(env.stack);
1713         phase_free(&env.ph);
1714
1715         return env.changes;
1716 }  /* optimize_loops */
1717
1718 /*
1719  * do the load store optimization
1720  */
1721 void optimize_load_store(ir_graph *irg) {
1722         walk_env_t env;
1723
1724         FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
1725         firm_dbg_set_mask(dbg, SET_LEVEL_4);
1726
1727         assert(get_irg_phase_state(irg) != phase_building);
1728         assert(get_irg_pinned(irg) != op_pin_state_floats &&
1729                 "LoadStore optimization needs pinned graph");
1730
1731         if (! get_opt_redundant_loadstore())
1732                 return;
1733
1734         /* we need landing pads */
1735         remove_critical_cf_edges(irg);
1736
1737         edges_assure(irg);
1738
1739         /* loop optimizations need dominators ... */
1740         assure_doms(irg);
1741
1742         /* for Phi optimization post-dominators are needed ... */
1743         assure_postdoms(irg);
1744
1745         if (get_opt_alias_analysis()) {
1746                 assure_irg_address_taken_computed(irg);
1747                 assure_irp_globals_address_taken_computed();
1748         }
1749
1750         obstack_init(&env.obst);
1751         env.changes = 0;
1752
1753         /* init the links, then collect Loads/Stores/Proj's in lists */
1754         master_visited = 0;
1755         irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
1756
1757         /* now we have collected enough information, optimize */
1758         irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
1759
1760         env.changes |= optimize_loops(irg);
1761
1762         obstack_free(&env.obst, NULL);
1763
1764         /* Handle graph state */
1765         if (env.changes) {
1766                 set_irg_outs_inconsistent(irg);
1767         }
1768
1769         if (env.changes & CF_CHANGED) {
1770                 /* is this really needed: Yes, control flow changed, block might
1771                 have Bad() predecessors. */
1772                 set_irg_doms_inconsistent(irg);
1773         }
1774 }  /* optimize_load_store */