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

   1 /*
   2  * Copyright (C) 1995-2008 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(IMAX(pn_Load_max, pn_Store_max), pn_Call_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 /** A Load/Store info. */
  82 typedef struct _ldst_info_t {
  83         ir_node  *projs[MAX_PROJ];    /**< list of Proj's of this node */
  84         ir_node  *exc_block;          /**< the exception block if available */
  85         int      exc_idx;             /**< predecessor index in the exception block */
  86         unsigned visited;             /**< visited counter for breaking loops */
  87 } ldst_info_t;
  88
  89 /**
  90  * flags for control flow.
  91  */
  92 enum block_flags_t {
  93         BLOCK_HAS_COND = 1,      /**< Block has conditional control flow */
  94         BLOCK_HAS_EXC  = 2       /**< Block has exceptional control flow */
  95 };
  96
  97 /**
  98  * a Block info.
  99  */
 100 typedef struct _block_info_t {
 101         unsigned flags;               /**< flags for the block */
 102 } block_info_t;
 103
 104 /** the master visited flag for loop detection. */
 105 static unsigned master_visited = 0;
 106
 107 #define INC_MASTER()       ++master_visited
 108 #define MARK_NODE(info)    (info)->visited = master_visited
 109 #define NODE_VISITED(info) (info)->visited >= master_visited
 110
 111 /**
 112  * get the Load/Store info of a node
 113  */
 114 static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst) {
 115         ldst_info_t *info = get_irn_link(node);
 116
 117         if (! info) {
 118                 info = obstack_alloc(obst, sizeof(*info));
 119                 memset(info, 0, sizeof(*info));
 120                 set_irn_link(node, info);
 121         }
 122         return info;
 123 }  /* get_ldst_info */
 124
 125 /**
 126  * get the Block info of a node
 127  */
 128 static block_info_t *get_block_info(ir_node *node, struct obstack *obst) {
 129         block_info_t *info = get_irn_link(node);
 130
 131         if (! info) {
 132                 info = obstack_alloc(obst, sizeof(*info));
 133                 memset(info, 0, sizeof(*info));
 134                 set_irn_link(node, info);
 135         }
 136         return info;
 137 }  /* get_block_info */
 138
 139 /**
 140  * update the projection info for a Load/Store
 141  */
 142 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
 143 {
 144         long nr = get_Proj_proj(proj);
 145
 146         assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
 147
 148         if (info->projs[nr]) {
 149                 /* there is already one, do CSE */
 150                 exchange(proj, info->projs[nr]);
 151                 return DF_CHANGED;
 152         }
 153         else {
 154                 info->projs[nr] = proj;
 155                 return 0;
 156         }
 157 }  /* update_projs */
 158
 159 /**
 160  * update the exception block info for a Load/Store node.
 161  *
 162  * @param info   the load/store info struct
 163  * @param block  the exception handler block for this load/store
 164  * @param pos    the control flow input of the block
 165  */
 166 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
 167 {
 168         assert(info->exc_block == NULL && "more than one exception block found");
 169
 170         info->exc_block = block;
 171         info->exc_idx   = pos;
 172         return 0;
 173 }  /* update_exc */
 174
 175 /** Return the number of uses of an address node */
 176 #define get_irn_n_uses(adr)     get_irn_n_edges(adr)
 177
 178 /**
 179  * walker, collects all Load/Store/Proj nodes
 180  *
 181  * walks from Start -> End
 182  */
 183 static void collect_nodes(ir_node *node, void *env)
 184 {
 185         ir_opcode   opcode = get_irn_opcode(node);
 186         ir_node     *pred, *blk, *pred_blk;
 187         ldst_info_t *ldst_info;
 188         walk_env_t  *wenv = env;
 189
 190         if (opcode == iro_Proj) {
 191                 pred   = get_Proj_pred(node);
 192                 opcode = get_irn_opcode(pred);
 193
 194                 if (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call) {
 195                         ldst_info = get_ldst_info(pred, &wenv->obst);
 196
 197                         wenv->changes |= update_projs(ldst_info, node);
 198
 199                         /*
 200                          * Place the Proj's to the same block as the
 201                          * predecessor Load. This is always ok and prevents
 202                          * "non-SSA" form after optimizations if the Proj
 203                          * is in a wrong block.
 204                          */
 205                         blk      = get_nodes_block(node);
 206                         pred_blk = get_nodes_block(pred);
 207                         if (blk != pred_blk) {
 208                                 wenv->changes |= DF_CHANGED;
 209                                 set_nodes_block(node, pred_blk);
 210                         }
 211                 }
 212         } else if (opcode == iro_Block) {
 213                 int i;
 214
 215                 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
 216                         ir_node      *pred_block, *proj;
 217                         block_info_t *bl_info;
 218                         int          is_exc = 0;
 219
 220                         pred = proj = get_Block_cfgpred(node, i);
 221
 222                         if (is_Proj(proj)) {
 223                                 pred   = get_Proj_pred(proj);
 224                                 is_exc = get_Proj_proj(proj) == pn_Generic_X_except;
 225                         }
 226
 227                         /* ignore Bad predecessors, they will be removed later */
 228                         if (is_Bad(pred))
 229                                 continue;
 230
 231                         pred_block = get_nodes_block(pred);
 232                         bl_info    = get_block_info(pred_block, &wenv->obst);
 233
 234                         if (is_fragile_op(pred) && is_exc)
 235                                 bl_info->flags |= BLOCK_HAS_EXC;
 236                         else if (is_irn_forking(pred))
 237                                 bl_info->flags |= BLOCK_HAS_COND;
 238
 239                         opcode = get_irn_opcode(pred);
 240                         if (is_exc && (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call)) {
 241                                 ldst_info = get_ldst_info(pred, &wenv->obst);
 242
 243                                 wenv->changes |= update_exc(ldst_info, node, i);
 244                         }
 245                 }
 246         }
 247 }  /* collect_nodes */
 248
 249 /**
 250  * Returns an entity if the address ptr points to a constant one.
 251  *
 252  * @param ptr  the address
 253  *
 254  * @return an entity or NULL
 255  */
 256 static ir_entity *find_constant_entity(ir_node *ptr)
 257 {
 258         for (;;) {
 259                 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
 260                         return get_SymConst_entity(ptr);
 261                 } else if (is_Sel(ptr)) {
 262                         ir_entity *ent = get_Sel_entity(ptr);
 263                         ir_type   *tp  = get_entity_owner(ent);
 264
 265                         /* Do not fiddle with polymorphism. */
 266                         if (is_Class_type(get_entity_owner(ent)) &&
 267                                 ((get_entity_n_overwrites(ent)    != 0) ||
 268                                 (get_entity_n_overwrittenby(ent) != 0)   ) )
 269                                 return NULL;
 270
 271                         if (is_Array_type(tp)) {
 272                                 /* check bounds */
 273                                 int i, n;
 274
 275                                 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
 276                                         ir_node *bound;
 277                                         tarval *tlower, *tupper;
 278                                         ir_node *index = get_Sel_index(ptr, i);
 279                                         tarval *tv     = computed_value(index);
 280
 281                                         /* check if the index is constant */
 282                                         if (tv == tarval_bad)
 283                                                 return NULL;
 284
 285                                         bound  = get_array_lower_bound(tp, i);
 286                                         tlower = computed_value(bound);
 287                                         bound  = get_array_upper_bound(tp, i);
 288                                         tupper = computed_value(bound);
 289
 290                                         if (tlower == tarval_bad || tupper == tarval_bad)
 291                                                 return NULL;
 292
 293                                         if (tarval_cmp(tv, tlower) & pn_Cmp_Lt)
 294                                                 return NULL;
 295                                         if (tarval_cmp(tupper, tv) & pn_Cmp_Lt)
 296                                                 return NULL;
 297
 298                                         /* ok, bounds check finished */
 299                                 }
 300                         }
 301
 302                         if (variability_constant == get_entity_variability(ent))
 303                                 return ent;
 304
 305                         /* try next */
 306                         ptr = get_Sel_ptr(ptr);
 307                 } else if (is_Add(ptr)) {
 308                         ir_node *l = get_Add_left(ptr);
 309                         ir_node *r = get_Add_right(ptr);
 310
 311                         if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
 312                                 ptr = l;
 313                         else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
 314                                 ptr = r;
 315                         else
 316                                 return NULL;
 317
 318                         /* for now, we support only one addition, reassoc should fold all others */
 319                         if (! is_SymConst(ptr) && !is_Sel(ptr))
 320                                 return NULL;
 321                 } else if (is_Sub(ptr)) {
 322                         ir_node *l = get_Sub_left(ptr);
 323                         ir_node *r = get_Sub_right(ptr);
 324
 325                         if (get_irn_mode(l) == get_irn_mode(ptr) &&     is_Const(r))
 326                                 ptr = l;
 327                         else
 328                                 return NULL;
 329                         /* for now, we support only one substraction, reassoc should fold all others */
 330                         if (! is_SymConst(ptr) && !is_Sel(ptr))
 331                                 return NULL;
 332                 } else
 333                         return NULL;
 334         }
 335 }  /* find_constant_entity */
 336
 337 /**
 338  * Return the Selection index of a Sel node from dimension n
 339  */
 340 static long get_Sel_array_index_long(ir_node *n, int dim) {
 341         ir_node *index = get_Sel_index(n, dim);
 342         assert(is_Const(index));
 343         return get_tarval_long(get_Const_tarval(index));
 344 }  /* get_Sel_array_index_long */
 345
 346 /**
 347  * Returns the accessed component graph path for an
 348  * node computing an address.
 349  *
 350  * @param ptr    the node computing the address
 351  * @param depth  current depth in steps upward from the root
 352  *               of the address
 353  */
 354 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, int depth) {
 355         compound_graph_path *res = NULL;
 356         ir_entity           *root, *field, *ent;
 357         int                 path_len, pos, idx;
 358         tarval              *tv;
 359         ir_type             *tp;
 360
 361         if (is_SymConst(ptr)) {
 362                 /* a SymConst. If the depth is 0, this is an access to a global
 363                  * entity and we don't need a component path, else we know
 364                  * at least its length.
 365                  */
 366                 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
 367                 root = get_SymConst_entity(ptr);
 368                 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
 369         } else if (is_Sel(ptr)) {
 370                 /* it's a Sel, go up until we find the root */
 371                 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
 372                 if (res == NULL)
 373                         return NULL;
 374
 375                 /* fill up the step in the path at the current position */
 376                 field    = get_Sel_entity(ptr);
 377                 path_len = get_compound_graph_path_length(res);
 378                 pos      = path_len - depth - 1;
 379                 set_compound_graph_path_node(res, pos, field);
 380
 381                 if (is_Array_type(get_entity_owner(field))) {
 382                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 383                         set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
 384                 }
 385         } else if (is_Add(ptr)) {
 386                 ir_node *l    = get_Add_left(ptr);
 387                 ir_node *r    = get_Add_right(ptr);
 388                 ir_mode *mode = get_irn_mode(ptr);
 389                 tarval  *tmp;
 390
 391                 if (is_Const(r) && get_irn_mode(l) == mode) {
 392                         ptr = l;
 393                         tv  = get_Const_tarval(r);
 394                 } else {
 395                         ptr = r;
 396                         tv  = get_Const_tarval(l);
 397                 }
 398 ptr_arith:
 399                 mode = get_tarval_mode(tv);
 400                 tmp  = tv;
 401
 402                 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
 403                 if (is_Sel(ptr)) {
 404                         field = get_Sel_entity(ptr);
 405                 } else {
 406                         field = get_SymConst_entity(ptr);
 407                 }
 408                 idx = 0;
 409                 for (ent = field;;) {
 410                         unsigned size;
 411                         tarval   *sz, *tv_index, *tlower, *tupper;
 412                         ir_node  *bound;
 413
 414                         tp = get_entity_type(ent);
 415                         if (! is_Array_type(tp))
 416                                 break;
 417                         ent = get_array_element_entity(tp);
 418                         size = get_type_size_bytes(get_entity_type(ent));
 419                         sz   = new_tarval_from_long(size, mode);
 420
 421                         tv_index = tarval_div(tmp, sz);
 422                         tmp      = tarval_mod(tmp, sz);
 423
 424                         if (tv_index == tarval_bad || tmp == tarval_bad)
 425                                 return NULL;
 426
 427                         assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
 428                         bound  = get_array_lower_bound(tp, 0);
 429                         tlower = computed_value(bound);
 430                         bound  = get_array_upper_bound(tp, 0);
 431                         tupper = computed_value(bound);
 432
 433                         if (tlower == tarval_bad || tupper == tarval_bad)
 434                                 return NULL;
 435
 436                         if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
 437                                 return NULL;
 438                         if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
 439                                 return NULL;
 440
 441                         /* ok, bounds check finished */
 442                         ++idx;
 443                 }
 444                 if (! tarval_is_null(tmp)) {
 445                         /* access to some struct/union member */
 446                         return NULL;
 447                 }
 448
 449                 /* should be at least ONE array */
 450                 if (idx == 0)
 451                         return NULL;
 452
 453                 res = rec_get_accessed_path(ptr, depth + idx);
 454                 if (res == NULL)
 455                         return NULL;
 456
 457                 path_len = get_compound_graph_path_length(res);
 458                 pos      = path_len - depth - idx;
 459
 460                 for (ent = field;;) {
 461                         unsigned size;
 462                         tarval   *sz, *tv_index;
 463                         long     index;
 464
 465                         tp = get_entity_type(ent);
 466                         if (! is_Array_type(tp))
 467                                 break;
 468                         ent = get_array_element_entity(tp);
 469                         set_compound_graph_path_node(res, pos, ent);
 470
 471                         size = get_type_size_bytes(get_entity_type(ent));
 472                         sz   = new_tarval_from_long(size, mode);
 473
 474                         tv_index = tarval_div(tv, sz);
 475                         tv       = tarval_mod(tv, sz);
 476
 477                         /* worked above, should work again */
 478                         assert(tv_index != tarval_bad && tv != tarval_bad);
 479
 480                         /* bounds already checked above */
 481                         index = get_tarval_long(tv_index);
 482                         set_compound_graph_path_array_index(res, pos, index);
 483                         ++pos;
 484                 }
 485         } else if (is_Sub(ptr)) {
 486                 ir_node *l = get_Sub_left(ptr);
 487                 ir_node *r = get_Sub_right(ptr);
 488
 489                 ptr = l;
 490                 tv  = get_Const_tarval(r);
 491                 tv  = tarval_neg(tv);
 492                 goto ptr_arith;
 493         }
 494         return res;
 495 }  /* rec_get_accessed_path */
 496
 497 /**
 498  * Returns an access path or NULL.  The access path is only
 499  * valid, if the graph is in phase_high and _no_ address computation is used.
 500  */
 501 static compound_graph_path *get_accessed_path(ir_node *ptr) {
 502         compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
 503         return gr;
 504 }  /* get_accessed_path */
 505
 506 typedef struct path_entry {
 507         ir_entity         *ent;
 508         struct path_entry *next;
 509         long              index;
 510 } path_entry;
 511
 512 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next) {
 513         path_entry       entry, *p;
 514         ir_entity        *ent, *field;
 515         ir_initializer_t *initializer;
 516         tarval           *tv;
 517         ir_type          *tp;
 518         unsigned         n;
 519
 520         entry.next = next;
 521         if (is_SymConst(ptr)) {
 522                 /* found the root */
 523                 ent         = get_SymConst_entity(ptr);
 524                 initializer = get_entity_initializer(ent);
 525                 for (p = next; p != NULL;) {
 526                         if (initializer->kind != IR_INITIALIZER_COMPOUND)
 527                                 return NULL;
 528                         n  = get_initializer_compound_n_entries(initializer);
 529                         tp = get_entity_type(ent);
 530
 531                         if (is_Array_type(tp)) {
 532                                 ent = get_array_element_entity(tp);
 533                                 if (ent != p->ent) {
 534                                         /* a missing [0] */
 535                                         if (0 >= n)
 536                                                 return NULL;
 537                                         initializer = get_initializer_compound_value(initializer, 0);
 538                                         continue;
 539                                 }
 540                         }
 541                         if (p->index >= (int) n)
 542                                 return NULL;
 543                         initializer = get_initializer_compound_value(initializer, p->index);
 544
 545                         ent = p->ent;
 546                         p   = p->next;
 547                 }
 548                 tp = get_entity_type(ent);
 549                 while (is_Array_type(tp)) {
 550                         ent = get_array_element_entity(tp);
 551                         tp = get_entity_type(ent);
 552                         /* a missing [0] */
 553                         n  = get_initializer_compound_n_entries(initializer);
 554                         if (0 >= n)
 555                                 return NULL;
 556                         initializer = get_initializer_compound_value(initializer, 0);
 557                 }
 558
 559                 switch (initializer->kind) {
 560                 case IR_INITIALIZER_CONST:
 561                         return get_initializer_const_value(initializer);
 562                 case IR_INITIALIZER_TARVAL:
 563                 case IR_INITIALIZER_NULL:
 564                 default:
 565                         return NULL;
 566                 }
 567         } else if (is_Sel(ptr)) {
 568                 entry.ent = field = get_Sel_entity(ptr);
 569                 tp = get_entity_owner(field);
 570                 if (is_Array_type(tp)) {
 571                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 572                         entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
 573                 } else {
 574                         int i, n_members = get_compound_n_members(tp);
 575                         for (i = 0; i < n_members; ++i) {
 576                                 if (get_compound_member(tp, i) == field)
 577                                         break;
 578                         }
 579                         if (i >= n_members) {
 580                                 /* not found: should NOT happen */
 581                                 return NULL;
 582                         }
 583                         entry.index = i;
 584                 }
 585                 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
 586         }  else if (is_Add(ptr)) {
 587                 ir_node  *l = get_Add_left(ptr);
 588                 ir_node  *r = get_Add_right(ptr);
 589                 ir_mode  *mode;
 590                 unsigned pos;
 591
 592                 if (is_Const(r)) {
 593                         ptr = l;
 594                         tv  = get_Const_tarval(r);
 595                 } else {
 596                         ptr = r;
 597                         tv  = get_Const_tarval(l);
 598                 }
 599 ptr_arith:
 600                 mode = get_tarval_mode(tv);
 601
 602                 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
 603                 if (is_Sel(ptr)) {
 604                         field = get_Sel_entity(ptr);
 605                 } else {
 606                         field = get_SymConst_entity(ptr);
 607                 }
 608
 609                 /* count needed entries */
 610                 pos = 0;
 611                 for (ent = field;;) {
 612                         tp = get_entity_type(ent);
 613                         if (! is_Array_type(tp))
 614                                 break;
 615                         ent = get_array_element_entity(tp);
 616                         ++pos;
 617                 }
 618                 /* should be at least ONE entry */
 619                 if (pos == 0)
 620                         return NULL;
 621
 622                 /* allocate the right number of entries */
 623                 NEW_ARR_A(path_entry, p, pos);
 624
 625                 /* fill them up */
 626                 pos = 0;
 627                 for (ent = field;;) {
 628                         unsigned size;
 629                         tarval   *sz, *tv_index, *tlower, *tupper;
 630                         long     index;
 631                         ir_node  *bound;
 632
 633                         tp = get_entity_type(ent);
 634                         if (! is_Array_type(tp))
 635                                 break;
 636                         ent = get_array_element_entity(tp);
 637                         p[pos].ent  = ent;
 638                         p[pos].next = &p[pos + 1];
 639
 640                         size = get_type_size_bytes(get_entity_type(ent));
 641                         sz   = new_tarval_from_long(size, mode);
 642
 643                         tv_index = tarval_div(tv, sz);
 644                         tv       = tarval_mod(tv, sz);
 645
 646                         if (tv_index == tarval_bad || tv == tarval_bad)
 647                                 return NULL;
 648
 649                         assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
 650                         bound  = get_array_lower_bound(tp, 0);
 651                         tlower = computed_value(bound);
 652                         bound  = get_array_upper_bound(tp, 0);
 653                         tupper = computed_value(bound);
 654
 655                         if (tlower == tarval_bad || tupper == tarval_bad)
 656                                 return NULL;
 657
 658                         if (tarval_cmp(tv_index, tlower) & pn_Cmp_Lt)
 659                                 return NULL;
 660                         if (tarval_cmp(tupper, tv_index) & pn_Cmp_Lt)
 661                                 return NULL;
 662
 663                         /* ok, bounds check finished */
 664                         index = get_tarval_long(tv_index);
 665                         p[pos].index = index;
 666                         ++pos;
 667                 }
 668                 if (! tarval_is_null(tv)) {
 669                         /* hmm, wrong access */
 670                         return NULL;
 671                 }
 672                 p[pos - 1].next = next;
 673                 return rec_find_compound_ent_value(ptr, p);
 674         } else if (is_Sub(ptr)) {
 675                 ir_node *l = get_Sub_left(ptr);
 676                 ir_node *r = get_Sub_right(ptr);
 677
 678                 ptr = l;
 679                 tv  = get_Const_tarval(r);
 680                 tv  = tarval_neg(tv);
 681                 goto ptr_arith;
 682         }
 683         return NULL;
 684 }
 685
 686 static ir_node *find_compound_ent_value(ir_node *ptr) {
 687         return rec_find_compound_ent_value(ptr, NULL);
 688 }
 689
 690 /* forward */
 691 static void reduce_adr_usage(ir_node *ptr);
 692
 693 /**
 694  * Update a Load that may have lost its users.
 695  */
 696 static void handle_load_update(ir_node *load) {
 697         ldst_info_t *info = get_irn_link(load);
 698
 699         /* do NOT touch volatile loads for now */
 700         if (get_Load_volatility(load) == volatility_is_volatile)
 701                 return;
 702
 703         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 704                 ir_node *ptr = get_Load_ptr(load);
 705                 ir_node *mem = get_Load_mem(load);
 706
 707                 /* a Load whose value is neither used nor exception checked, remove it */
 708                 exchange(info->projs[pn_Load_M], mem);
 709                 if (info->projs[pn_Load_X_regular])
 710                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
 711                 kill_node(load);
 712                 reduce_adr_usage(ptr);
 713         }
 714 }  /* handle_load_update */
 715
 716 /**
 717  * A use of an address node has vanished. Check if this was a Proj
 718  * node and update the counters.
 719  */
 720 static void reduce_adr_usage(ir_node *ptr) {
 721         if (is_Proj(ptr)) {
 722                 if (get_irn_n_edges(ptr) <= 0) {
 723                         /* this Proj is dead now */
 724                         ir_node *pred = get_Proj_pred(ptr);
 725
 726                         if (is_Load(pred)) {
 727                                 ldst_info_t *info = get_irn_link(pred);
 728                                 info->projs[get_Proj_proj(ptr)] = NULL;
 729
 730                                 /* this node lost its result proj, handle that */
 731                                 handle_load_update(pred);
 732                         }
 733                 }
 734         }
 735 }  /* reduce_adr_usage */
 736
 737 /**
 738  * Check, if an already existing value of mode old_mode can be converted
 739  * into the needed one new_mode without loss.
 740  */
 741 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
 742         if (old_mode == new_mode)
 743                 return 1;
 744
 745         /* if both modes are two-complement ones, we can always convert the
 746            Stored value into the needed one. */
 747         if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
 748                   get_mode_arithmetic(old_mode) == irma_twos_complement &&
 749                   get_mode_arithmetic(new_mode) == irma_twos_complement)
 750                 return 1;
 751         return 0;
 752 }  /* can_use_stored_value */
 753
 754 /**
 755  * Check whether a Call is at least pure, ie. does only read memory.
 756  */
 757 static unsigned is_Call_pure(ir_node *call) {
 758         ir_type *call_tp = get_Call_type(call);
 759         unsigned prop = get_method_additional_properties(call_tp);
 760
 761         /* check first the call type */
 762         if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
 763                 /* try the called entity */
 764                 ir_node *ptr = get_Call_ptr(call);
 765
 766                 if (is_Global(ptr)) {
 767                         ir_entity *ent = get_Global_entity(ptr);
 768
 769                         prop = get_entity_additional_properties(ent);
 770                 }
 771         }
 772         return (prop & (mtp_property_const|mtp_property_pure)) != 0;
 773 }  /* is_Call_pure */
 774
 775 static ir_node *get_base_and_offset(ir_node *ptr, long *pOffset)
 776 {
 777         ir_mode *mode  = get_irn_mode(ptr);
 778         long    offset = 0;
 779
 780         /* TODO: long might not be enough, we should probably use some tarval thingy... */
 781         for (;;) {
 782                 if (is_Add(ptr)) {
 783                         ir_node *l = get_Add_left(ptr);
 784                         ir_node *r = get_Add_right(ptr);
 785
 786                         if (get_irn_mode(l) != mode || !is_Const(r))
 787                                 break;
 788
 789                         offset += get_tarval_long(get_Const_tarval(r));
 790                         ptr     = l;
 791                 } else if (is_Sub(ptr)) {
 792                         ir_node *l = get_Sub_left(ptr);
 793                         ir_node *r = get_Sub_right(ptr);
 794
 795                         if (get_irn_mode(l) != mode || !is_Const(r))
 796                                 break;
 797
 798                         offset -= get_tarval_long(get_Const_tarval(r));
 799                         ptr     = l;
 800                 } else if (is_Sel(ptr)) {
 801                         ir_entity *ent = get_Sel_entity(ptr);
 802                         ir_type   *tp  = get_entity_owner(ent);
 803
 804                         if (is_Array_type(tp)) {
 805                                 int     size;
 806                                 ir_node *index;
 807
 808                                 /* only one dimensional arrays yet */
 809                                 if (get_Sel_n_indexs(ptr) != 1)
 810                                         break;
 811                                 index = get_Sel_index(ptr, 0);
 812                                 if (! is_Const(index))
 813                                         break;
 814
 815                                 tp = get_entity_type(ent);
 816                                 if (get_type_state(tp) != layout_fixed)
 817                                         break;
 818
 819                                 size    = get_type_size_bytes(tp);
 820                                 offset += size * get_tarval_long(get_Const_tarval(index));
 821                         } else {
 822                                 if (get_type_state(tp) != layout_fixed)
 823                                         break;
 824                                 offset += get_entity_offset(ent);
 825                         }
 826                         ptr = get_Sel_ptr(ptr);
 827                 } else
 828                         break;
 829         }
 830
 831         *pOffset = offset;
 832         return ptr;
 833 }
 834
 835 static int try_load_after_store(ir_node *load,
 836                 ir_node *load_base_ptr, long load_offset, ir_node *store)
 837 {
 838         ldst_info_t *info;
 839         ir_node *store_ptr      = get_Store_ptr(store);
 840         long     store_offset;
 841         ir_node *store_base_ptr = get_base_and_offset(store_ptr, &store_offset);
 842         ir_node *store_value;
 843         ir_mode *store_mode;
 844         ir_node *load_ptr;
 845         ir_mode *load_mode;
 846         long     load_mode_len;
 847         long     store_mode_len;
 848         long     delta;
 849         int      res;
 850
 851         if (load_base_ptr != store_base_ptr)
 852                 return 0;
 853
 854         load_mode      = get_Load_mode(load);
 855         load_mode_len  = get_mode_size_bytes(load_mode);
 856         store_mode     = get_irn_mode(get_Store_value(store));
 857         store_mode_len = get_mode_size_bytes(store_mode);
 858         delta          = load_offset - store_offset;
 859         if (delta < 0 || delta + load_mode_len > store_mode_len)
 860                 return 0;
 861
 862         if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
 863             get_mode_arithmetic(load_mode)  != irma_twos_complement)
 864                 return 0;
 865
 866         store_value = get_Store_value(store);
 867
 868         /* produce a shift to adjust offset delta */
 869         if (delta > 0) {
 870                 ir_node *cnst;
 871
 872                 /* FIXME: only true for little endian */
 873                 cnst        = new_Const_long(mode_Iu, delta * 8);
 874                 store_value = new_r_Shr(current_ir_graph, get_nodes_block(load),
 875                                         store_value, cnst, store_mode);
 876         }
 877
 878         /* add an convert if needed */
 879         if (store_mode != load_mode) {
 880                 store_value = new_r_Conv(current_ir_graph, get_nodes_block(load),
 881                                          store_value, load_mode);
 882         }
 883
 884         DBG_OPT_RAW(load, store_value);
 885
 886         info = get_irn_link(load);
 887         if (info->projs[pn_Load_M])
 888                 exchange(info->projs[pn_Load_M], get_Load_mem(load));
 889
 890         res = 0;
 891         /* no exception */
 892         if (info->projs[pn_Load_X_except]) {
 893                 exchange( info->projs[pn_Load_X_except], new_Bad());
 894                 res |= CF_CHANGED;
 895         }
 896         if (info->projs[pn_Load_X_regular]) {
 897                 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
 898                 res |= CF_CHANGED;
 899         }
 900
 901         if (info->projs[pn_Load_res])
 902                 exchange(info->projs[pn_Load_res], store_value);
 903
 904         load_ptr = get_Load_ptr(load);
 905         kill_node(load);
 906         reduce_adr_usage(load_ptr);
 907         return res | DF_CHANGED;
 908 }
 909
 910 /**
 911  * Follow the memory chain as long as there are only Loads,
 912  * alias free Stores, and constant Calls and try to replace the
 913  * current Load by a previous ones.
 914  * Note that in unreachable loops it might happen that we reach
 915  * load again, as well as we can fall into a cycle.
 916  * We break such cycles using a special visited flag.
 917  *
 918  * INC_MASTER() must be called before dive into
 919  */
 920 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
 921         unsigned    res = 0;
 922         ldst_info_t *info = get_irn_link(load);
 923         ir_node     *pred;
 924         ir_node     *ptr       = get_Load_ptr(load);
 925         ir_node     *mem       = get_Load_mem(load);
 926         ir_mode     *load_mode = get_Load_mode(load);
 927
 928         for (pred = curr; load != pred; ) {
 929                 ldst_info_t *pred_info = get_irn_link(pred);
 930
 931                 /*
 932                  * a Load immediately after a Store -- a read after write.
 933                  * We may remove the Load, if both Load & Store does not have an
 934                  * exception handler OR they are in the same MacroBlock. In the latter
 935                  * case the Load cannot throw an exception when the previous Store was
 936                  * quiet.
 937                  *
 938                  * Why we need to check for Store Exception? If the Store cannot
 939                  * be executed (ROM) the exception handler might simply jump into
 940                  * the load MacroBlock :-(
 941                  * We could make it a little bit better if we would know that the
 942                  * exception handler of the Store jumps directly to the end...
 943                  */
 944                 if (is_Store(pred) && ((pred_info->projs[pn_Store_X_except] == NULL
 945                                 && info->projs[pn_Load_X_except] == NULL)
 946                                 || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)))
 947                 {
 948                         long    load_offset;
 949                         ir_node *base_ptr = get_base_and_offset(ptr, &load_offset);
 950                         int     changes   = try_load_after_store(load, base_ptr, load_offset, pred);
 951
 952                         if (changes != 0)
 953                                 return res | changes;
 954                 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
 955                            can_use_stored_value(get_Load_mode(pred), load_mode)) {
 956                         /*
 957                          * a Load after a Load -- a read after read.
 958                          * We may remove the second Load, if it does not have an exception handler
 959                          * OR they are in the same MacroBlock. In the later case the Load cannot
 960                          * throw an exception when the previous Load was quiet.
 961                          *
 962                          * Here, there is no need to check if the previous Load has an exception
 963                          * hander because they would have exact the same exception...
 964                          */
 965                         if (info->projs[pn_Load_X_except] == NULL || get_nodes_MacroBlock(load) == get_nodes_MacroBlock(pred)) {
 966                                 ir_node *value;
 967
 968                                 DBG_OPT_RAR(load, pred);
 969
 970                                 /* the result is used */
 971                                 if (info->projs[pn_Load_res]) {
 972                                         if (pred_info->projs[pn_Load_res] == NULL) {
 973                                                 /* create a new Proj again */
 974                                                 pred_info->projs[pn_Load_res] = new_r_Proj(current_ir_graph, get_nodes_block(pred), pred, get_Load_mode(pred), pn_Load_res);
 975                                         }
 976                                         value = pred_info->projs[pn_Load_res];
 977
 978                                         /* add an convert if needed */
 979                                         if (get_Load_mode(pred) != load_mode) {
 980                                                 value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
 981                                         }
 982
 983                                         exchange(info->projs[pn_Load_res], value);
 984                                 }
 985
 986                                 if (info->projs[pn_Load_M])
 987                                         exchange(info->projs[pn_Load_M], mem);
 988
 989                                 /* no exception */
 990                                 if (info->projs[pn_Load_X_except]) {
 991                                         exchange(info->projs[pn_Load_X_except], new_Bad());
 992                                         res |= CF_CHANGED;
 993                                 }
 994                                 if (info->projs[pn_Load_X_regular]) {
 995                                         exchange( info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
 996                                         res |= CF_CHANGED;
 997                                 }
 998
 999                                 kill_node(load);
1000                                 reduce_adr_usage(ptr);
1001                                 return res |= DF_CHANGED;
1002                         }
1003                 }
1004
1005                 if (is_Store(pred)) {
1006                         /* check if we can pass through this store */
1007                         ir_alias_relation rel = get_alias_relation(
1008                                 current_ir_graph,
1009                                 get_Store_ptr(pred),
1010                                 get_irn_mode(get_Store_value(pred)),
1011                                 ptr, load_mode);
1012                         /* if the might be an alias, we cannot pass this Store */
1013                         if (rel != ir_no_alias)
1014                                 break;
1015                         pred = skip_Proj(get_Store_mem(pred));
1016                 } else if (is_Load(pred)) {
1017                         pred = skip_Proj(get_Load_mem(pred));
1018                 } else if (is_Call(pred)) {
1019                         if (is_Call_pure(pred)) {
1020                                 /* The called graph is at least pure, so there are no Store's
1021                                    in it. We can handle it like a Load and skip it. */
1022                                 pred = skip_Proj(get_Call_mem(pred));
1023                         } else {
1024                                 /* there might be Store's in the graph, stop here */
1025                                 break;
1026                         }
1027                 } else {
1028                         /* follow only Load chains */
1029                         break;
1030                 }
1031
1032                 /* check for cycles */
1033                 if (NODE_VISITED(pred_info))
1034                         break;
1035                 MARK_NODE(pred_info);
1036         }
1037
1038         if (is_Sync(pred)) {
1039                 int i;
1040
1041                 /* handle all Sync predecessors */
1042                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1043                         res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
1044                         if (res)
1045                                 return res;
1046                 }
1047         }
1048
1049         return res;
1050 }  /* follow_Mem_chain */
1051
1052 /*
1053  * Check if we can replace the load by a given const from
1054  * the const code irg.
1055  */
1056 ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c) {
1057         ir_mode *c_mode = get_irn_mode(c);
1058         ir_mode *l_mode = get_Load_mode(load);
1059         ir_node *res    = NULL;
1060
1061         if (c_mode != l_mode) {
1062                 /* check, if the mode matches OR can be easily converted info */
1063                 if (is_reinterpret_cast(c_mode, l_mode)) {
1064                         /* we can safely cast */
1065                         dbg_info *dbg   = get_irn_dbg_info(load);
1066                         ir_node  *block = get_nodes_block(load);
1067
1068                         /* copy the value from the const code irg and cast it */
1069                         res = copy_const_value(dbg, c);
1070                         res = new_rd_Conv(dbg, current_ir_graph, block, res, l_mode);
1071                 }
1072         } else {
1073                 /* copy the value from the const code irg */
1074                 res = copy_const_value(get_irn_dbg_info(load), c);
1075         }
1076         return res;
1077 }  /* can_replace_load_by_const */
1078
1079 /**
1080  * optimize a Load
1081  *
1082  * @param load  the Load node
1083  */
1084 static unsigned optimize_load(ir_node *load)
1085 {
1086         ldst_info_t *info = get_irn_link(load);
1087         ir_node     *mem, *ptr, *value;
1088         ir_entity   *ent;
1089         long        dummy;
1090         unsigned    res = 0;
1091
1092         /* do NOT touch volatile loads for now */
1093         if (get_Load_volatility(load) == volatility_is_volatile)
1094                 return 0;
1095
1096         /* the address of the load to be optimized */
1097         ptr = get_Load_ptr(load);
1098
1099         /*
1100          * Check if we can remove the exception from a Load:
1101          * This can be done, if the address is from an Sel(Alloc) and
1102          * the Sel type is a subtype of the allocated type.
1103          *
1104          * This optimizes some often used OO constructs,
1105          * like x = new O; x->t;
1106          */
1107         if (info->projs[pn_Load_X_except]) {
1108                 ir_node *addr = ptr;
1109
1110                 /* find base address */
1111                 while (is_Sel(addr))
1112                         addr = get_Sel_ptr(addr);
1113                 if (is_Alloc(skip_Proj(skip_Cast(addr)))) {
1114                         /* simple case: a direct load after an Alloc. Firm Alloc throw
1115                          * an exception in case of out-of-memory. So, there is no way for an
1116                          * exception in this load.
1117                          * This code is constructed by the "exception lowering" in the Jack compiler.
1118                          */
1119                         exchange(info->projs[pn_Load_X_except], new_Bad());
1120                         info->projs[pn_Load_X_except] = NULL;
1121                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1122                         info->projs[pn_Load_X_regular] = NULL;
1123                         res |= CF_CHANGED;
1124                 }
1125         }
1126
1127         /* The mem of the Load. Must still be returned after optimization. */
1128         mem = get_Load_mem(load);
1129
1130         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
1131                 /* a Load which value is neither used nor exception checked, remove it */
1132                 exchange(info->projs[pn_Load_M], mem);
1133
1134                 if (info->projs[pn_Load_X_regular]) {
1135                         /* should not happen, but if it does, remove it */
1136                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1137                         res |= CF_CHANGED;
1138                 }
1139                 kill_node(load);
1140                 reduce_adr_usage(ptr);
1141                 return res | DF_CHANGED;
1142         }
1143
1144         /* Load from a constant polymorphic field, where we can resolve
1145            polymorphism. */
1146         value = transform_polymorph_Load(load);
1147         if (value == load) {
1148                 value = NULL;
1149                 /* check if we can determine the entity that will be loaded */
1150                 ent = find_constant_entity(ptr);
1151                 if (ent != NULL                                     &&
1152                     allocation_static == get_entity_allocation(ent) &&
1153                     visibility_external_allocated != get_entity_visibility(ent)) {
1154                         /* a static allocation that is not external: there should be NO exception
1155                          * when loading even if we cannot replace the load itself. */
1156
1157                         /* no exception, clear the info field as it might be checked later again */
1158                         if (info->projs[pn_Load_X_except]) {
1159                                 exchange(info->projs[pn_Load_X_except], new_Bad());
1160                                 info->projs[pn_Load_X_except] = NULL;
1161                                 res |= CF_CHANGED;
1162                         }
1163                         if (info->projs[pn_Load_X_regular]) {
1164                                 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1165                                 info->projs[pn_Load_X_regular] = NULL;
1166                                 res |= CF_CHANGED;
1167                         }
1168
1169                         if (variability_constant == get_entity_variability(ent)) {
1170                                 if (is_atomic_entity(ent)) {
1171                                         /* Might not be atomic after lowering of Sels.  In this case we
1172                                          * could also load, but it's more complicated. */
1173                                         /* more simpler case: we load the content of a constant value:
1174                                          * replace it by the constant itself */
1175                                         value = get_atomic_ent_value(ent);
1176                                 } else if (ent->has_initializer) {
1177                                         /* new style initializer */
1178                                         value = find_compound_ent_value(ptr);
1179                                 } else {
1180                                         /* old style initializer */
1181                                         compound_graph_path *path = get_accessed_path(ptr);
1182
1183                                         if (path != NULL) {
1184                                                 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1185
1186                                                 value = get_compound_ent_value_by_path(ent, path);
1187                                                 DB((dbg, LEVEL_1, "  Constant access at %F%F resulted in %+F\n", ent, path, value));
1188                                                 free_compound_graph_path(path);
1189                                         }
1190                                 }
1191                                 if (value != NULL)
1192                                         value = can_replace_load_by_const(load, value);
1193                         }
1194                 }
1195         }
1196         if (value != NULL) {
1197                 /* we completely replace the load by this value */
1198                 if (info->projs[pn_Load_X_except]) {
1199                         exchange(info->projs[pn_Load_X_except], new_Bad());
1200                         info->projs[pn_Load_X_except] = NULL;
1201                         res |= CF_CHANGED;
1202                 }
1203                 if (info->projs[pn_Load_X_regular]) {
1204                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(current_ir_graph, get_nodes_block(load)));
1205                         info->projs[pn_Load_X_regular] = NULL;
1206                         res |= CF_CHANGED;
1207                 }
1208                 if (info->projs[pn_Load_M]) {
1209                         exchange(info->projs[pn_Load_M], mem);
1210                         res |= DF_CHANGED;
1211                 }
1212                 if (info->projs[pn_Load_res]) {
1213                         exchange(info->projs[pn_Load_res], value);
1214                         res |= DF_CHANGED;
1215                 }
1216                 kill_node(load);
1217                 reduce_adr_usage(ptr);
1218                 return res;
1219         }
1220
1221         /* Check, if the address of this load is used more than once.
1222          * If not, more load cannot be removed in any case. */
1223         if (get_irn_n_uses(ptr) <= 1 && get_irn_n_uses(get_base_and_offset(ptr, &dummy)) <= 1)
1224                 return res;
1225
1226         /*
1227          * follow the memory chain as long as there are only Loads
1228          * and try to replace current Load or Store by a previous one.
1229          * Note that in unreachable loops it might happen that we reach
1230          * load again, as well as we can fall into a cycle.
1231          * We break such cycles using a special visited flag.
1232          */
1233         INC_MASTER();
1234         res = follow_Mem_chain(load, skip_Proj(mem));
1235         return res;
1236 }  /* optimize_load */
1237
1238 /**
1239  * Check whether a value of mode new_mode would completely overwrite a value
1240  * of mode old_mode in memory.
1241  */
1242 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
1243 {
1244         return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
1245 }  /* is_completely_overwritten */
1246
1247 /**
1248  * follow the memory chain as long as there are only Loads and alias free Stores.
1249  *
1250  * INC_MASTER() must be called before dive into
1251  */
1252 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
1253         unsigned res = 0;
1254         ldst_info_t *info = get_irn_link(store);
1255         ir_node *pred;
1256         ir_node *ptr = get_Store_ptr(store);
1257         ir_node *mem = get_Store_mem(store);
1258         ir_node *value = get_Store_value(store);
1259         ir_mode *mode  = get_irn_mode(value);
1260         ir_node *block = get_nodes_block(store);
1261         ir_node *mblk  = get_Block_MacroBlock(block);
1262
1263         for (pred = curr; pred != store;) {
1264                 ldst_info_t *pred_info = get_irn_link(pred);
1265
1266                 /*
1267                  * BEWARE: one might think that checking the modes is useless, because
1268                  * if the pointers are identical, they refer to the same object.
1269                  * This is only true in strong typed languages, not is C were the following
1270                  * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
1271                  * However, if the mode that is written have a bigger  or equal size the the old
1272                  * one, the old value is completely overwritten and can be killed ...
1273                  */
1274                 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
1275                     get_nodes_MacroBlock(pred) == mblk &&
1276                     is_completely_overwritten(get_irn_mode(get_Store_value(pred)), mode)) {
1277                         /*
1278                          * a Store after a Store in the same MacroBlock -- a write after write.
1279                          * We may remove the first Store, if it does not have an exception handler.
1280                          *
1281                          * TODO: What, if both have the same exception handler ???
1282                          */
1283                         if (get_Store_volatility(pred) != volatility_is_volatile && !pred_info->projs[pn_Store_X_except]) {
1284                                 DBG_OPT_WAW(pred, store);
1285                                 exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
1286                                 kill_node(pred);
1287                                 reduce_adr_usage(ptr);
1288                                 return DF_CHANGED;
1289                         }
1290                 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
1291                            value == pred_info->projs[pn_Load_res]) {
1292                         /*
1293                          * a Store of a value just loaded from the same address
1294                          * -- a write after read.
1295                          * We may remove the Store, if it does not have an exception
1296                          * handler.
1297                          */
1298                         if (! info->projs[pn_Store_X_except]) {
1299                                 DBG_OPT_WAR(store, pred);
1300                                 exchange(info->projs[pn_Store_M], mem);
1301                                 kill_node(store);
1302                                 reduce_adr_usage(ptr);
1303                                 return DF_CHANGED;
1304                         }
1305                 }
1306
1307                 if (is_Store(pred)) {
1308                         /* check if we can pass through this store */
1309                         ir_alias_relation rel = get_alias_relation(
1310                                 current_ir_graph,
1311                                 get_Store_ptr(pred),
1312                                 get_irn_mode(get_Store_value(pred)),
1313                                 ptr, mode);
1314                         /* if the might be an alias, we cannot pass this Store */
1315                         if (rel != ir_no_alias)
1316                                 break;
1317                         pred = skip_Proj(get_Store_mem(pred));
1318                 } else if (is_Load(pred)) {
1319                         ir_alias_relation rel = get_alias_relation(
1320                                 current_ir_graph, get_Load_ptr(pred), get_Load_mode(pred),
1321                                 ptr, mode);
1322                         if (rel != ir_no_alias)
1323                                 break;
1324
1325                         pred = skip_Proj(get_Load_mem(pred));
1326                 } else {
1327                         /* follow only Load chains */
1328                         break;
1329                 }
1330
1331                 /* check for cycles */
1332                 if (NODE_VISITED(pred_info))
1333                         break;
1334                 MARK_NODE(pred_info);
1335         }
1336
1337         if (is_Sync(pred)) {
1338                 int i;
1339
1340                 /* handle all Sync predecessors */
1341                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1342                         res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
1343                         if (res)
1344                                 break;
1345                 }
1346         }
1347         return res;
1348 }  /* follow_Mem_chain_for_Store */
1349
1350 /** find entity used as base for an address calculation */
1351 static ir_entity *find_entity(ir_node *ptr)
1352 {
1353         switch(get_irn_opcode(ptr)) {
1354         case iro_SymConst:
1355                 return get_SymConst_entity(ptr);
1356         case iro_Sel: {
1357                 ir_node *pred = get_Sel_ptr(ptr);
1358                 if (get_irg_frame(get_irn_irg(ptr)) == pred)
1359                         return get_Sel_entity(ptr);
1360
1361                 return find_entity(pred);
1362         }
1363         case iro_Sub:
1364         case iro_Add: {
1365                 ir_node *left = get_binop_left(ptr);
1366                 ir_node *right;
1367                 if (mode_is_reference(get_irn_mode(left)))
1368                         return find_entity(left);
1369                 right = get_binop_right(ptr);
1370                 if (mode_is_reference(get_irn_mode(right)))
1371                         return find_entity(right);
1372                 return NULL;
1373         }
1374         default:
1375                 return NULL;
1376         }
1377 }
1378
1379 /**
1380  * optimize a Store
1381  *
1382  * @param store  the Store node
1383  */
1384 static unsigned optimize_store(ir_node *store) {
1385         ir_node   *ptr;
1386         ir_node   *mem;
1387         ir_entity *entity;
1388
1389         if (get_Store_volatility(store) == volatility_is_volatile)
1390                 return 0;
1391
1392         ptr    = get_Store_ptr(store);
1393         entity = find_entity(ptr);
1394
1395         /* a store to an entity which is never read is unnecessary */
1396         if (entity != NULL && !(get_entity_usage(entity) & ir_usage_read)) {
1397                 ldst_info_t *info = get_irn_link(store);
1398                 if (info->projs[pn_Store_X_except] == NULL) {
1399                         exchange(info->projs[pn_Store_M], get_Store_mem(store));
1400                         kill_node(store);
1401                         reduce_adr_usage(ptr);
1402                         return DF_CHANGED;
1403                 }
1404         }
1405
1406         /* Check, if the address of this Store is used more than once.
1407          * If not, this Store cannot be removed in any case. */
1408         if (get_irn_n_uses(ptr) <= 1)
1409                 return 0;
1410
1411         mem = get_Store_mem(store);
1412
1413         /* follow the memory chain as long as there are only Loads */
1414         INC_MASTER();
1415
1416         return follow_Mem_chain_for_Store(store, skip_Proj(mem));
1417 }  /* optimize_store */
1418
1419 /**
1420  * walker, optimizes Phi after Stores to identical places:
1421  * Does the following optimization:
1422  * @verbatim
1423  *
1424  *   val1   val2   val3          val1  val2  val3
1425  *    |      |      |               \    |    /
1426  *  Store  Store  Store              \   |   /
1427  *      \    |    /                   PhiData
1428  *       \   |   /                       |
1429  *        \  |  /                      Store
1430  *          PhiM
1431  *
1432  * @endverbatim
1433  * This reduces the number of stores and allows for predicated execution.
1434  * Moves Stores back to the end of a function which may be bad.
1435  *
1436  * This is only possible if the predecessor blocks have only one successor.
1437  */
1438 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
1439 {
1440         int i, n;
1441         ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
1442         ir_mode *mode;
1443         ir_node **inM, **inD, **projMs;
1444         int *idx;
1445         dbg_info *db = NULL;
1446         ldst_info_t *info;
1447         block_info_t *bl_info;
1448         unsigned res = 0;
1449
1450         /* Must be a memory Phi */
1451         if (get_irn_mode(phi) != mode_M)
1452                 return 0;
1453
1454         n = get_Phi_n_preds(phi);
1455         if (n <= 0)
1456                 return 0;
1457
1458         /* must be only one user */
1459         projM = get_Phi_pred(phi, 0);
1460         if (get_irn_n_edges(projM) != 1)
1461                 return 0;
1462
1463         store = skip_Proj(projM);
1464         old_store = store;
1465         if (!is_Store(store))
1466                 return 0;
1467
1468         block = get_nodes_block(store);
1469
1470         /* abort on dead blocks */
1471         if (is_Block_dead(block))
1472                 return 0;
1473
1474         /* check if the block is post dominated by Phi-block
1475            and has no exception exit */
1476         bl_info = get_irn_link(block);
1477         if (bl_info->flags & BLOCK_HAS_EXC)
1478                 return 0;
1479
1480         phi_block = get_nodes_block(phi);
1481         if (! block_strictly_postdominates(phi_block, block))
1482                 return 0;
1483
1484         /* this is the address of the store */
1485         ptr  = get_Store_ptr(store);
1486         mode = get_irn_mode(get_Store_value(store));
1487         info = get_irn_link(store);
1488         exc  = info->exc_block;
1489
1490         for (i = 1; i < n; ++i) {
1491                 ir_node *pred = get_Phi_pred(phi, i);
1492
1493                 if (get_irn_n_edges(pred) != 1)
1494                         return 0;
1495
1496                 pred = skip_Proj(pred);
1497                 if (!is_Store(pred))
1498                         return 0;
1499
1500                 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1501                         return 0;
1502
1503                 info = get_irn_link(pred);
1504
1505                 /* check, if all stores have the same exception flow */
1506                 if (exc != info->exc_block)
1507                         return 0;
1508
1509                 /* abort on dead blocks */
1510                 block = get_nodes_block(pred);
1511                 if (is_Block_dead(block))
1512                         return 0;
1513
1514                 /* check if the block is post dominated by Phi-block
1515                    and has no exception exit. Note that block must be different from
1516                    Phi-block, else we would move a Store from end End of a block to its
1517                    Start... */
1518                 bl_info = get_irn_link(block);
1519                 if (bl_info->flags & BLOCK_HAS_EXC)
1520                         return 0;
1521                 if (block == phi_block || ! block_postdominates(phi_block, block))
1522                         return 0;
1523         }
1524
1525         /*
1526          * ok, when we are here, we found all predecessors of a Phi that
1527          * are Stores to the same address and size. That means whatever
1528          * we do before we enter the block of the Phi, we do a Store.
1529          * So, we can move the Store to the current block:
1530          *
1531          *   val1    val2    val3          val1  val2  val3
1532          *    |       |       |               \    |    /
1533          * | Str | | Str | | Str |             \   |   /
1534          *      \     |     /                   PhiData
1535          *       \    |    /                       |
1536          *        \   |   /                       Str
1537          *           PhiM
1538          *
1539          * Is only allowed if the predecessor blocks have only one successor.
1540          */
1541
1542         NEW_ARR_A(ir_node *, projMs, n);
1543         NEW_ARR_A(ir_node *, inM, n);
1544         NEW_ARR_A(ir_node *, inD, n);
1545         NEW_ARR_A(int, idx, n);
1546
1547         /* Prepare: Collect all Store nodes.  We must do this
1548            first because we otherwise may loose a store when exchanging its
1549            memory Proj.
1550          */
1551         for (i = n - 1; i >= 0; --i) {
1552                 ir_node *store;
1553
1554                 projMs[i] = get_Phi_pred(phi, i);
1555                 assert(is_Proj(projMs[i]));
1556
1557                 store = get_Proj_pred(projMs[i]);
1558                 info  = get_irn_link(store);
1559
1560                 inM[i] = get_Store_mem(store);
1561                 inD[i] = get_Store_value(store);
1562                 idx[i] = info->exc_idx;
1563         }
1564         block = get_nodes_block(phi);
1565
1566         /* second step: create a new memory Phi */
1567         phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M);
1568
1569         /* third step: create a new data Phi */
1570         phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode);
1571
1572         /* rewire memory and kill the node */
1573         for (i = n - 1; i >= 0; --i) {
1574                 ir_node *proj  = projMs[i];
1575
1576                 if(is_Proj(proj)) {
1577                         ir_node *store = get_Proj_pred(proj);
1578                         exchange(proj, inM[i]);
1579                         kill_node(store);
1580                 }
1581         }
1582
1583         /* fourth step: create the Store */
1584         store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD);
1585 #ifdef DO_CACHEOPT
1586         co_set_irn_name(store, co_get_irn_ident(old_store));
1587 #endif
1588
1589         projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
1590
1591         info = get_ldst_info(store, &wenv->obst);
1592         info->projs[pn_Store_M] = projM;
1593
1594         /* fifths step: repair exception flow */
1595         if (exc) {
1596                 ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except);
1597
1598                 info->projs[pn_Store_X_except] = projX;
1599                 info->exc_block                = exc;
1600                 info->exc_idx                  = idx[0];
1601
1602                 for (i = 0; i < n; ++i) {
1603                         set_Block_cfgpred(exc, idx[i], projX);
1604                 }
1605
1606                 if (n > 1) {
1607                         /* the exception block should be optimized as some inputs are identical now */
1608                 }
1609
1610                 res |= CF_CHANGED;
1611         }
1612
1613         /* sixth step: replace old Phi */
1614         exchange(phi, projM);
1615
1616         return res | DF_CHANGED;
1617 }  /* optimize_phi */
1618
1619 /**
1620  * walker, do the optimizations
1621  */
1622 static void do_load_store_optimize(ir_node *n, void *env) {
1623         walk_env_t *wenv = env;
1624
1625         switch (get_irn_opcode(n)) {
1626
1627         case iro_Load:
1628                 wenv->changes |= optimize_load(n);
1629                 break;
1630
1631         case iro_Store:
1632                 wenv->changes |= optimize_store(n);
1633                 break;
1634
1635         case iro_Phi:
1636                 wenv->changes |= optimize_phi(n, wenv);
1637                 break;
1638
1639         default:
1640                 ;
1641         }
1642 }  /* do_load_store_optimize */
1643
1644 /** A scc. */
1645 typedef struct scc {
1646         ir_node *head;          /**< the head of the list */
1647 } scc;
1648
1649 /** A node entry. */
1650 typedef struct node_entry {
1651         unsigned DFSnum;    /**< the DFS number of this node */
1652         unsigned low;       /**< the low number of this node */
1653         ir_node  *header;   /**< the header of this node */
1654         int      in_stack;  /**< flag, set if the node is on the stack */
1655         ir_node  *next;     /**< link to the next node the the same scc */
1656         scc      *pscc;     /**< the scc of this node */
1657         unsigned POnum;     /**< the post order number for blocks */
1658 } node_entry;
1659
1660 /** A loop entry. */
1661 typedef struct loop_env {
1662         ir_phase ph;           /**< the phase object */
1663         ir_node  **stack;      /**< the node stack */
1664         int      tos;          /**< tos index */
1665         unsigned nextDFSnum;   /**< the current DFS number */
1666         unsigned POnum;        /**< current post order number */
1667
1668         unsigned changes;      /**< a bitmask of graph changes */
1669 } loop_env;
1670
1671 /**
1672 * Gets the node_entry of a node
1673 */
1674 static node_entry *get_irn_ne(ir_node *irn, loop_env *env) {
1675         ir_phase   *ph = &env->ph;
1676         node_entry *e  = phase_get_irn_data(&env->ph, irn);
1677
1678         if (! e) {
1679                 e = phase_alloc(ph, sizeof(*e));
1680                 memset(e, 0, sizeof(*e));
1681                 phase_set_irn_data(ph, irn, e);
1682         }
1683         return e;
1684 }  /* get_irn_ne */
1685
1686 /**
1687  * Push a node onto the stack.
1688  *
1689  * @param env   the loop environment
1690  * @param n     the node to push
1691  */
1692 static void push(loop_env *env, ir_node *n) {
1693         node_entry *e;
1694
1695         if (env->tos == ARR_LEN(env->stack)) {
1696                 int nlen = ARR_LEN(env->stack) * 2;
1697                 ARR_RESIZE(ir_node *, env->stack, nlen);
1698         }
1699         env->stack[env->tos++] = n;
1700         e = get_irn_ne(n, env);
1701         e->in_stack = 1;
1702 }  /* push */
1703
1704 /**
1705  * pop a node from the stack
1706  *
1707  * @param env   the loop environment
1708  *
1709  * @return  The topmost node
1710  */
1711 static ir_node *pop(loop_env *env) {
1712         ir_node *n = env->stack[--env->tos];
1713         node_entry *e = get_irn_ne(n, env);
1714
1715         e->in_stack = 0;
1716         return n;
1717 }  /* pop */
1718
1719 /**
1720  * Check if irn is a region constant.
1721  * The block or irn must strictly dominate the header block.
1722  *
1723  * @param irn           the node to check
1724  * @param header_block  the header block of the induction variable
1725  */
1726 static int is_rc(ir_node *irn, ir_node *header_block) {
1727         ir_node *block = get_nodes_block(irn);
1728
1729         return (block != header_block) && block_dominates(block, header_block);
1730 }  /* is_rc */
1731
1732 typedef struct phi_entry phi_entry;
1733 struct phi_entry {
1734         ir_node   *phi;    /**< A phi with a region const memory. */
1735         int       pos;     /**< The position of the region const memory */
1736         ir_node   *load;   /**< the newly created load for this phi */
1737         phi_entry *next;
1738 };
1739
1740 /**
1741  * Move loops out of loops if possible.
1742  *
1743  * @param pscc   the loop described by an SCC
1744  * @param env    the loop environment
1745  */
1746 static void move_loads_out_of_loops(scc *pscc, loop_env *env) {
1747         ir_node   *phi, *load, *next, *other, *next_other;
1748         ir_entity *ent;
1749         int       j;
1750         phi_entry *phi_list = NULL;
1751
1752         /* collect all outer memories */
1753         for (phi = pscc->head; phi != NULL; phi = next) {
1754                 node_entry *ne = get_irn_ne(phi, env);
1755                 next = ne->next;
1756
1757                 /* check all memory Phi's */
1758                 if (! is_Phi(phi))
1759                         continue;
1760
1761                 assert(get_irn_mode(phi) == mode_M && "DFS geturn non-memory Phi");
1762
1763                 for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
1764                         ir_node    *pred = get_irn_n(phi, j);
1765                         node_entry *pe   = get_irn_ne(pred, env);
1766
1767                         if (pe->pscc != ne->pscc) {
1768                                 /* not in the same SCC, is region const */
1769                                 phi_entry *pe = phase_alloc(&env->ph, sizeof(*pe));
1770
1771                                 pe->phi  = phi;
1772                                 pe->pos  = j;
1773                                 pe->next = phi_list;
1774                                 phi_list = pe;
1775                         }
1776                 }
1777         }
1778         /* no Phis no fun */
1779         assert(phi_list != NULL && "DFS found a loop without Phi");
1780
1781         for (load = pscc->head; load; load = next) {
1782                 ir_mode *load_mode;
1783                 node_entry *ne = get_irn_ne(load, env);
1784                 next = ne->next;
1785
1786                 if (is_Load(load)) {
1787                         ldst_info_t *info = get_irn_link(load);
1788                         ir_node     *ptr = get_Load_ptr(load);
1789
1790                         /* for now, we cannot handle Loads with exceptions */
1791                         if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
1792                                 continue;
1793
1794                         /* for now, we can only handle Load(Global) */
1795                         if (! is_Global(ptr))
1796                                 continue;
1797                         ent = get_Global_entity(ptr);
1798                         load_mode = get_Load_mode(load);
1799                         for (other = pscc->head; other != NULL; other = next_other) {
1800                                 node_entry *ne = get_irn_ne(other, env);
1801                                 next_other = ne->next;
1802
1803                                 if (is_Store(other)) {
1804                                         ir_alias_relation rel = get_alias_relation(
1805                                                 current_ir_graph,
1806                                                 get_Store_ptr(other),
1807                                                 get_irn_mode(get_Store_value(other)),
1808                                                 ptr, load_mode);
1809                                         /* if the might be an alias, we cannot pass this Store */
1810                                         if (rel != ir_no_alias)
1811                                                 break;
1812                                 }
1813                                 /* only pure Calls are allowed here, so ignore them */
1814                         }
1815                         if (other == NULL) {
1816                                 ldst_info_t *ninfo;
1817                                 phi_entry   *pe;
1818                                 dbg_info    *db;
1819
1820                                 /* for now, we cannot handle more than one input */
1821                                 if (phi_list->next != NULL)
1822                                         return;
1823
1824                                 /* yep, no aliasing Store found, Load can be moved */
1825                                 DB((dbg, LEVEL_1, "  Found a Load that could be moved: %+F\n", load));
1826
1827                                 db   = get_irn_dbg_info(load);
1828                                 for (pe = phi_list; pe != NULL; pe = pe->next) {
1829                                         int     pos   = pe->pos;
1830                                         ir_node *phi  = pe->phi;
1831                                         ir_node *blk  = get_nodes_block(phi);
1832                                         ir_node *pred = get_Block_cfgpred_block(blk, pos);
1833                                         ir_node *irn, *mem;
1834
1835                                         pe->load = irn = new_rd_Load(db, current_ir_graph, pred, get_Phi_pred(phi, pos), ptr, load_mode);
1836                                         ninfo = get_ldst_info(irn, phase_obst(&env->ph));
1837
1838                                         ninfo->projs[pn_Load_M] = mem = new_r_Proj(current_ir_graph, pred, irn, mode_M, pn_Load_M);
1839                                         set_Phi_pred(phi, pos, mem);
1840
1841                                         ninfo->projs[pn_Load_res] = new_r_Proj(current_ir_graph, pred, irn, load_mode, pn_Load_res);
1842
1843                                         DB((dbg, LEVEL_1, "  Created %+F in %+F\n", irn, pred));
1844                                 }
1845
1846                                 /* now kill the old Load */
1847                                 exchange(info->projs[pn_Load_M], get_Load_mem(load));
1848                                 exchange(info->projs[pn_Load_res], ninfo->projs[pn_Load_res]);
1849
1850                                 env->changes |= DF_CHANGED;
1851                         }
1852                 }
1853         }
1854 }  /* move_loads_out_of_loops */
1855
1856 /**
1857  * Process a loop SCC.
1858  *
1859  * @param pscc  the SCC
1860  * @param env   the loop environment
1861  */
1862 static void process_loop(scc *pscc, loop_env *env) {
1863         ir_node *irn, *next, *header = NULL;
1864         node_entry *b, *h = NULL;
1865         int j, only_phi, num_outside, process = 0;
1866         ir_node *out_rc;
1867
1868         /* find the header block for this scc */
1869         for (irn = pscc->head; irn; irn = next) {
1870                 node_entry *e = get_irn_ne(irn, env);
1871                 ir_node *block = get_nodes_block(irn);
1872
1873                 next = e->next;
1874                 b = get_irn_ne(block, env);
1875
1876                 if (header) {
1877                         if (h->POnum < b->POnum) {
1878                                 header = block;
1879                                 h      = b;
1880                         }
1881                 }
1882                 else {
1883                         header = block;
1884                         h      = b;
1885                 }
1886         }
1887
1888         /* check if this scc contains only Phi, Loads or Stores nodes */
1889         only_phi    = 1;
1890         num_outside = 0;
1891         out_rc      = NULL;
1892         for (irn = pscc->head; irn; irn = next) {
1893                 node_entry *e = get_irn_ne(irn, env);
1894
1895                 next = e->next;
1896                 switch (get_irn_opcode(irn)) {
1897                 case iro_Call:
1898                         if (is_Call_pure(irn)) {
1899                                 /* pure calls can be treated like loads */
1900                                 only_phi = 0;
1901                                 break;
1902                         }
1903                         /* non-pure calls must be handle like may-alias Stores */
1904                         goto fail;
1905                 case iro_CopyB:
1906                         /* cannot handle CopyB yet */
1907                         goto fail;
1908                 case iro_Load:
1909                         process = 1;
1910                         if (get_Load_volatility(irn) == volatility_is_volatile) {
1911                                 /* cannot handle loops with volatile Loads */
1912                                 goto fail;
1913                         }
1914                         only_phi = 0;
1915                         break;
1916                 case iro_Store:
1917                         if (get_Store_volatility(irn) == volatility_is_volatile) {
1918                                 /* cannot handle loops with volatile Stores */
1919                                 goto fail;
1920                         }
1921                         only_phi = 0;
1922                         break;
1923                 default:
1924                         only_phi = 0;
1925                         break;
1926                 case iro_Phi:
1927                         for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
1928                                 ir_node *pred  = get_irn_n(irn, j);
1929                                 node_entry *pe = get_irn_ne(pred, env);
1930
1931                                 if (pe->pscc != e->pscc) {
1932                                         /* not in the same SCC, must be a region const */
1933                                         if (! is_rc(pred, header)) {
1934                                                 /* not a memory loop */
1935                                                 goto fail;
1936                                         }
1937                                         if (! out_rc) {
1938                                                 out_rc = pred;
1939                                                 ++num_outside;
1940                                         } else if (out_rc != pred) {
1941                                                 ++num_outside;
1942                                         }
1943                                 }
1944                         }
1945                         break;
1946                 }
1947         }
1948         if (! process)
1949                 goto fail;
1950
1951         /* found a memory loop */
1952         DB((dbg, LEVEL_2, "  Found a memory loop:\n  "));
1953         if (only_phi && num_outside == 1) {
1954                 /* a phi cycle with only one real predecessor can be collapsed */
1955                 DB((dbg, LEVEL_2, "  Found an USELESS Phi cycle:\n  "));
1956
1957                 for (irn = pscc->head; irn; irn = next) {
1958                         node_entry *e = get_irn_ne(irn, env);
1959                         next = e->next;
1960                         e->header = NULL;
1961                         exchange(irn, out_rc);
1962                 }
1963                 env->changes |= DF_CHANGED;
1964                 return;
1965         }
1966
1967         /* set the header for every node in this scc */
1968         for (irn = pscc->head; irn; irn = next) {
1969                 node_entry *e = get_irn_ne(irn, env);
1970                 e->header = header;
1971                 next = e->next;
1972                 DB((dbg, LEVEL_2, " %+F,", irn));
1973         }
1974         DB((dbg, LEVEL_2, "\n"));
1975
1976         move_loads_out_of_loops(pscc, env);
1977
1978 fail:
1979         ;
1980 }  /* process_loop */
1981
1982 /**
1983  * Process a SCC.
1984  *
1985  * @param pscc  the SCC
1986  * @param env   the loop environment
1987  */
1988 static void process_scc(scc *pscc, loop_env *env) {
1989         ir_node *head = pscc->head;
1990         node_entry *e = get_irn_ne(head, env);
1991
1992 #ifdef DEBUG_libfirm
1993         {
1994                 ir_node *irn, *next;
1995
1996                 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
1997                 for (irn = pscc->head; irn; irn = next) {
1998                         node_entry *e = get_irn_ne(irn, env);
1999
2000                         next = e->next;
2001
2002                         DB((dbg, LEVEL_4, " %+F,", irn));
2003                 }
2004                 DB((dbg, LEVEL_4, "\n"));
2005         }
2006 #endif
2007
2008         if (e->next != NULL) {
2009                 /* this SCC has more than one member */
2010                 process_loop(pscc, env);
2011         }
2012 }  /* process_scc */
2013
2014 /**
2015  * Do Tarjan's SCC algorithm and drive load/store optimization.
2016  *
2017  * @param irn  start at this node
2018  * @param env  the loop environment
2019  */
2020 static void dfs(ir_node *irn, loop_env *env)
2021 {
2022         int i, n;
2023         node_entry *node = get_irn_ne(irn, env);
2024
2025         mark_irn_visited(irn);
2026
2027         node->DFSnum = env->nextDFSnum++;
2028         node->low    = node->DFSnum;
2029         push(env, irn);
2030
2031         /* handle preds */
2032         if (is_Phi(irn) || is_Sync(irn)) {
2033                 n = get_irn_arity(irn);
2034                 for (i = 0; i < n; ++i) {
2035                         ir_node *pred = get_irn_n(irn, i);
2036                         node_entry *o = get_irn_ne(pred, env);
2037
2038                         if (irn_not_visited(pred)) {
2039                                 dfs(pred, env);
2040                                 node->low = MIN(node->low, o->low);
2041                         }
2042                         if (o->DFSnum < node->DFSnum && o->in_stack)
2043                                 node->low = MIN(o->DFSnum, node->low);
2044                 }
2045         } else if (is_fragile_op(irn)) {
2046                 ir_node *pred = get_fragile_op_mem(irn);
2047                 node_entry *o = get_irn_ne(pred, env);
2048
2049                 if (irn_not_visited(pred)) {
2050                         dfs(pred, env);
2051                         node->low = MIN(node->low, o->low);
2052                 }
2053                 if (o->DFSnum < node->DFSnum && o->in_stack)
2054                         node->low = MIN(o->DFSnum, node->low);
2055         } else if (is_Proj(irn)) {
2056                 ir_node *pred = get_Proj_pred(irn);
2057                 node_entry *o = get_irn_ne(pred, env);
2058
2059                 if (irn_not_visited(pred)) {
2060                         dfs(pred, env);
2061                         node->low = MIN(node->low, o->low);
2062                 }
2063                 if (o->DFSnum < node->DFSnum && o->in_stack)
2064                         node->low = MIN(o->DFSnum, node->low);
2065         }
2066         else {
2067                  /* IGNORE predecessors */
2068         }
2069
2070         if (node->low == node->DFSnum) {
2071                 scc *pscc = phase_alloc(&env->ph, sizeof(*pscc));
2072                 ir_node *x;
2073
2074                 pscc->head = NULL;
2075                 do {
2076                         node_entry *e;
2077
2078                         x = pop(env);
2079                         e = get_irn_ne(x, env);
2080                         e->pscc    = pscc;
2081                         e->next    = pscc->head;
2082                         pscc->head = x;
2083                 } while (x != irn);
2084
2085                 process_scc(pscc, env);
2086         }
2087 }  /* dfs */
2088
2089 /**
2090  * Do the DFS on the memory edges a graph.
2091  *
2092  * @param irg  the graph to process
2093  * @param env  the loop environment
2094  */
2095 static void do_dfs(ir_graph *irg, loop_env *env) {
2096         ir_graph *rem = current_ir_graph;
2097         ir_node  *endblk, *end;
2098         int      i;
2099
2100         current_ir_graph = irg;
2101         inc_irg_visited(irg);
2102
2103         /* visit all memory nodes */
2104         endblk = get_irg_end_block(irg);
2105         for (i = get_Block_n_cfgpreds(endblk) - 1; i >= 0; --i) {
2106                 ir_node *pred = get_Block_cfgpred(endblk, i);
2107
2108                 pred = skip_Proj(pred);
2109                 if (is_Return(pred))
2110                         dfs(get_Return_mem(pred), env);
2111                 else if (is_Raise(pred))
2112                         dfs(get_Raise_mem(pred), env);
2113                 else if (is_fragile_op(pred))
2114                         dfs(get_fragile_op_mem(pred), env);
2115                 else {
2116                         assert(0 && "Unknown EndBlock predecessor");
2117                 }
2118         }
2119
2120         /* visit the keep-alives */
2121         end = get_irg_end(irg);
2122         for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
2123                 ir_node *ka = get_End_keepalive(end, i);
2124
2125                 if (is_Phi(ka) && irn_not_visited(ka))
2126                         dfs(ka, env);
2127         }
2128         current_ir_graph = rem;
2129 }  /* do_dfs */
2130
2131 /**
2132  * Initialize new phase data. We do this always explicit, so return NULL here
2133  */
2134 static void *init_loop_data(ir_phase *ph, const ir_node *irn, void *data) {
2135         (void)ph;
2136         (void)irn;
2137         (void)data;
2138         return NULL;
2139 }  /* init_loop_data */
2140
2141 /**
2142  * Optimize Loads/Stores in loops.
2143  *
2144  * @param irg  the graph
2145  */
2146 static int optimize_loops(ir_graph *irg) {
2147         loop_env env;
2148
2149         env.stack         = NEW_ARR_F(ir_node *, 128);
2150         env.tos           = 0;
2151         env.nextDFSnum    = 0;
2152         env.POnum         = 0;
2153         env.changes       = 0;
2154         phase_init(&env.ph, "ldstopt", irg, PHASE_DEFAULT_GROWTH, init_loop_data, NULL);
2155
2156         /* calculate the SCC's and drive loop optimization. */
2157         do_dfs(irg, &env);
2158
2159         DEL_ARR_F(env.stack);
2160         phase_free(&env.ph);
2161
2162         return env.changes;
2163 }  /* optimize_loops */
2164
2165 /*
2166  * do the load store optimization
2167  */
2168 void optimize_load_store(ir_graph *irg) {
2169         walk_env_t env;
2170
2171         FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
2172
2173         assert(get_irg_phase_state(irg) != phase_building);
2174         assert(get_irg_pinned(irg) != op_pin_state_floats &&
2175                 "LoadStore optimization needs pinned graph");
2176
2177         /* we need landing pads */
2178         remove_critical_cf_edges(irg);
2179
2180         edges_assure(irg);
2181
2182         /* for Phi optimization post-dominators are needed ... */
2183         assure_postdoms(irg);
2184
2185         if (get_opt_alias_analysis()) {
2186                 assure_irg_entity_usage_computed(irg);
2187                 assure_irp_globals_entity_usage_computed();
2188         }
2189
2190         obstack_init(&env.obst);
2191         env.changes = 0;
2192
2193         /* init the links, then collect Loads/Stores/Proj's in lists */
2194         master_visited = 0;
2195         irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
2196
2197         /* now we have collected enough information, optimize */
2198         irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
2199
2200         env.changes |= optimize_loops(irg);
2201
2202         obstack_free(&env.obst, NULL);
2203
2204         /* Handle graph state */
2205         if (env.changes) {
2206                 set_irg_outs_inconsistent(irg);
2207                 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
2208         }
2209
2210         if (env.changes & CF_CHANGED) {
2211                 /* is this really needed: Yes, control flow changed, block might
2212                 have Bad() predecessors. */
2213                 set_irg_doms_inconsistent(irg);
2214         }
2215 }  /* optimize_load_store */