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

   1 /*
   2  * Copyright (C) 1995-2011 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 #include "config.h"
  27
  28 #include <string.h>
  29
  30 #include "iroptimize.h"
  31 #include "irnode_t.h"
  32 #include "irgraph_t.h"
  33 #include "irmode_t.h"
  34 #include "iropt_t.h"
  35 #include "ircons_t.h"
  36 #include "irgmod.h"
  37 #include "irgwalk.h"
  38 #include "tv_t.h"
  39 #include "dbginfo_t.h"
  40 #include "iropt_dbg.h"
  41 #include "irflag_t.h"
  42 #include "array_t.h"
  43 #include "irhooks.h"
  44 #include "iredges.h"
  45 #include "irpass.h"
  46 #include "opt_polymorphy.h"
  47 #include "irmemory.h"
  48 #include "irphase_t.h"
  49 #include "irgopt.h"
  50 #include "set.h"
  51 #include "be.h"
  52 #include "debug.h"
  53
  54 /** The debug handle. */
  55 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
  56
  57 #undef IMAX
  58 #define IMAX(a,b)   ((a) > (b) ? (a) : (b))
  59
  60 #define MAX_PROJ    IMAX(IMAX((long)pn_Load_max, (long)pn_Store_max), (long)pn_Call_max)
  61
  62 enum changes_t {
  63         DF_CHANGED = 1,       /**< data flow changed */
  64         CF_CHANGED = 2,       /**< control flow changed */
  65 };
  66
  67 /**
  68  * walker environment
  69  */
  70 typedef struct walk_env_t {
  71         struct obstack obst;          /**< list of all stores */
  72         unsigned changes;             /**< a bitmask of graph changes */
  73 } walk_env_t;
  74
  75 /** A Load/Store info. */
  76 typedef struct ldst_info_t {
  77         ir_node  *projs[MAX_PROJ];    /**< list of Proj's of this node */
  78         ir_node  *exc_block;          /**< the exception block if available */
  79         int      exc_idx;             /**< predecessor index in the exception block */
  80         unsigned visited;             /**< visited counter for breaking loops */
  81 } ldst_info_t;
  82
  83 /**
  84  * flags for control flow.
  85  */
  86 enum block_flags_t {
  87         BLOCK_HAS_COND = 1,      /**< Block has conditional control flow */
  88         BLOCK_HAS_EXC  = 2       /**< Block has exceptional control flow */
  89 };
  90
  91 /**
  92  * a Block info.
  93  */
  94 typedef struct block_info_t {
  95         unsigned flags;               /**< flags for the block */
  96 } block_info_t;
  97
  98 /** the master visited flag for loop detection. */
  99 static unsigned master_visited = 0;
 100
 101 #define INC_MASTER()       ++master_visited
 102 #define MARK_NODE(info)    (info)->visited = master_visited
 103 #define NODE_VISITED(info) (info)->visited >= master_visited
 104
 105 /**
 106  * get the Load/Store info of a node
 107  */
 108 static ldst_info_t *get_ldst_info(ir_node *node, struct obstack *obst)
 109 {
 110         ldst_info_t *info = (ldst_info_t*)get_irn_link(node);
 111
 112         if (! info) {
 113                 info = OALLOCZ(obst, ldst_info_t);
 114                 set_irn_link(node, info);
 115         }
 116         return info;
 117 }  /* get_ldst_info */
 118
 119 /**
 120  * get the Block info of a node
 121  */
 122 static block_info_t *get_block_info(ir_node *node, struct obstack *obst)
 123 {
 124         block_info_t *info = (block_info_t*)get_irn_link(node);
 125
 126         if (! info) {
 127                 info = OALLOCZ(obst, block_info_t);
 128                 set_irn_link(node, info);
 129         }
 130         return info;
 131 }  /* get_block_info */
 132
 133 /**
 134  * update the projection info for a Load/Store
 135  */
 136 static unsigned update_projs(ldst_info_t *info, ir_node *proj)
 137 {
 138         long nr = get_Proj_proj(proj);
 139
 140         assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
 141
 142         if (info->projs[nr]) {
 143                 /* there is already one, do CSE */
 144                 exchange(proj, info->projs[nr]);
 145                 return DF_CHANGED;
 146         }
 147         else {
 148                 info->projs[nr] = proj;
 149                 return 0;
 150         }
 151 }  /* update_projs */
 152
 153 /**
 154  * update the exception block info for a Load/Store node.
 155  *
 156  * @param info   the load/store info struct
 157  * @param block  the exception handler block for this load/store
 158  * @param pos    the control flow input of the block
 159  */
 160 static unsigned update_exc(ldst_info_t *info, ir_node *block, int pos)
 161 {
 162         assert(info->exc_block == NULL && "more than one exception block found");
 163
 164         info->exc_block = block;
 165         info->exc_idx   = pos;
 166         return 0;
 167 }  /* update_exc */
 168
 169 /** Return the number of uses of an address node */
 170 #define get_irn_n_uses(adr)     get_irn_n_edges(adr)
 171
 172 /**
 173  * walker, collects all Load/Store/Proj nodes
 174  *
 175  * walks from Start -> End
 176  */
 177 static void collect_nodes(ir_node *node, void *env)
 178 {
 179         walk_env_t  *wenv   = (walk_env_t *)env;
 180         unsigned     opcode = get_irn_opcode(node);
 181         ir_node     *pred, *blk, *pred_blk;
 182         ldst_info_t *ldst_info;
 183
 184         if (opcode == iro_Proj) {
 185                 pred   = get_Proj_pred(node);
 186                 opcode = get_irn_opcode(pred);
 187
 188                 if (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call) {
 189                         ldst_info = get_ldst_info(pred, &wenv->obst);
 190
 191                         wenv->changes |= update_projs(ldst_info, node);
 192
 193                         /*
 194                          * Place the Proj's to the same block as the
 195                          * predecessor Load. This is always ok and prevents
 196                          * "non-SSA" form after optimizations if the Proj
 197                          * is in a wrong block.
 198                          */
 199                         blk      = get_nodes_block(node);
 200                         pred_blk = get_nodes_block(pred);
 201                         if (blk != pred_blk) {
 202                                 wenv->changes |= DF_CHANGED;
 203                                 set_nodes_block(node, pred_blk);
 204                         }
 205                 }
 206         } else if (opcode == iro_Block) {
 207                 int i;
 208
 209                 for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
 210                         ir_node      *pred_block, *proj;
 211                         block_info_t *bl_info;
 212                         int          is_exc = 0;
 213
 214                         pred = proj = get_Block_cfgpred(node, i);
 215
 216                         if (is_Proj(proj)) {
 217                                 pred   = get_Proj_pred(proj);
 218                                 is_exc = get_Proj_proj(proj) == pn_Generic_X_except;
 219                         }
 220
 221                         /* ignore Bad predecessors, they will be removed later */
 222                         if (is_Bad(pred))
 223                                 continue;
 224
 225                         pred_block = get_nodes_block(pred);
 226                         bl_info    = get_block_info(pred_block, &wenv->obst);
 227
 228                         if (is_fragile_op(pred) && is_exc)
 229                                 bl_info->flags |= BLOCK_HAS_EXC;
 230                         else if (is_irn_forking(pred))
 231                                 bl_info->flags |= BLOCK_HAS_COND;
 232
 233                         opcode = get_irn_opcode(pred);
 234                         if (is_exc && (opcode == iro_Load || opcode == iro_Store || opcode == iro_Call)) {
 235                                 ldst_info = get_ldst_info(pred, &wenv->obst);
 236
 237                                 wenv->changes |= update_exc(ldst_info, node, i);
 238                         }
 239                 }
 240         }
 241 }  /* collect_nodes */
 242
 243 /**
 244  * Returns an entity if the address ptr points to a constant one.
 245  *
 246  * @param ptr  the address
 247  *
 248  * @return an entity or NULL
 249  */
 250 static ir_entity *find_constant_entity(ir_node *ptr)
 251 {
 252         for (;;) {
 253                 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
 254                         return get_SymConst_entity(ptr);
 255                 } else if (is_Sel(ptr)) {
 256                         ir_entity *ent = get_Sel_entity(ptr);
 257                         ir_type   *tp  = get_entity_owner(ent);
 258
 259                         /* Do not fiddle with polymorphism. */
 260                         if (is_Class_type(get_entity_owner(ent)) &&
 261                                 ((get_entity_n_overwrites(ent)    != 0) ||
 262                                 (get_entity_n_overwrittenby(ent) != 0)   ) )
 263                                 return NULL;
 264
 265                         if (is_Array_type(tp)) {
 266                                 /* check bounds */
 267                                 int i, n;
 268
 269                                 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
 270                                         ir_node   *bound;
 271                                         ir_tarval *tlower, *tupper;
 272                                         ir_node   *index = get_Sel_index(ptr, i);
 273                                         ir_tarval *tv    = computed_value(index);
 274
 275                                         /* check if the index is constant */
 276                                         if (tv == tarval_bad)
 277                                                 return NULL;
 278
 279                                         bound  = get_array_lower_bound(tp, i);
 280                                         tlower = computed_value(bound);
 281                                         bound  = get_array_upper_bound(tp, i);
 282                                         tupper = computed_value(bound);
 283
 284                                         if (tlower == tarval_bad || tupper == tarval_bad)
 285                                                 return NULL;
 286
 287                                         if (tarval_cmp(tv, tlower) == ir_relation_less)
 288                                                 return NULL;
 289                                         if (tarval_cmp(tupper, tv) == ir_relation_less)
 290                                                 return NULL;
 291
 292                                         /* ok, bounds check finished */
 293                                 }
 294                         }
 295
 296                         if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT)
 297                                 return ent;
 298
 299                         /* try next */
 300                         ptr = get_Sel_ptr(ptr);
 301                 } else if (is_Add(ptr)) {
 302                         ir_node *l = get_Add_left(ptr);
 303                         ir_node *r = get_Add_right(ptr);
 304
 305                         if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
 306                                 ptr = l;
 307                         else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
 308                                 ptr = r;
 309                         else
 310                                 return NULL;
 311
 312                         /* for now, we support only one addition, reassoc should fold all others */
 313                         if (! is_SymConst(ptr) && !is_Sel(ptr))
 314                                 return NULL;
 315                 } else if (is_Sub(ptr)) {
 316                         ir_node *l = get_Sub_left(ptr);
 317                         ir_node *r = get_Sub_right(ptr);
 318
 319                         if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
 320                                 ptr = l;
 321                         else
 322                                 return NULL;
 323                         /* for now, we support only one substraction, reassoc should fold all others */
 324                         if (! is_SymConst(ptr) && !is_Sel(ptr))
 325                                 return NULL;
 326                 } else
 327                         return NULL;
 328         }
 329 }  /* find_constant_entity */
 330
 331 /**
 332  * Return the Selection index of a Sel node from dimension n
 333  */
 334 static long get_Sel_array_index_long(ir_node *n, int dim)
 335 {
 336         ir_node *index = get_Sel_index(n, dim);
 337         assert(is_Const(index));
 338         return get_tarval_long(get_Const_tarval(index));
 339 }  /* get_Sel_array_index_long */
 340
 341 /**
 342  * Returns the accessed component graph path for an
 343  * node computing an address.
 344  *
 345  * @param ptr    the node computing the address
 346  * @param depth  current depth in steps upward from the root
 347  *               of the address
 348  */
 349 static compound_graph_path *rec_get_accessed_path(ir_node *ptr, size_t depth)
 350 {
 351         compound_graph_path *res = NULL;
 352         ir_entity           *root, *field, *ent;
 353         size_t              path_len, pos, idx;
 354         ir_tarval           *tv;
 355         ir_type             *tp;
 356
 357         if (is_SymConst(ptr)) {
 358                 /* a SymConst. If the depth is 0, this is an access to a global
 359                  * entity and we don't need a component path, else we know
 360                  * at least its length.
 361                  */
 362                 assert(get_SymConst_kind(ptr) == symconst_addr_ent);
 363                 root = get_SymConst_entity(ptr);
 364                 res = (depth == 0) ? NULL : new_compound_graph_path(get_entity_type(root), depth);
 365         } else if (is_Sel(ptr)) {
 366                 /* it's a Sel, go up until we find the root */
 367                 res = rec_get_accessed_path(get_Sel_ptr(ptr), depth+1);
 368                 if (res == NULL)
 369                         return NULL;
 370
 371                 /* fill up the step in the path at the current position */
 372                 field    = get_Sel_entity(ptr);
 373                 path_len = get_compound_graph_path_length(res);
 374                 pos      = path_len - depth - 1;
 375                 set_compound_graph_path_node(res, pos, field);
 376
 377                 if (is_Array_type(get_entity_owner(field))) {
 378                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 379                         set_compound_graph_path_array_index(res, pos, get_Sel_array_index_long(ptr, 0));
 380                 }
 381         } else if (is_Add(ptr)) {
 382                 ir_mode   *mode;
 383                 ir_tarval *tmp;
 384
 385                 {
 386                         ir_node   *l    = get_Add_left(ptr);
 387                         ir_node   *r    = get_Add_right(ptr);
 388                         if (is_Const(r) && get_irn_mode(l) == get_irn_mode(ptr)) {
 389                                 ptr = l;
 390                                 tv  = get_Const_tarval(r);
 391                         } else {
 392                                 ptr = r;
 393                                 tv  = get_Const_tarval(l);
 394                         }
 395                 }
 396 ptr_arith:
 397                 mode = get_tarval_mode(tv);
 398                 tmp  = tv;
 399
 400                 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
 401                 if (is_Sel(ptr)) {
 402                         field = get_Sel_entity(ptr);
 403                 } else {
 404                         field = get_SymConst_entity(ptr);
 405                 }
 406                 idx = 0;
 407                 for (ent = field;;) {
 408                         unsigned   size;
 409                         ir_tarval *sz, *tv_index, *tlower, *tupper;
 410                         ir_node   *bound;
 411
 412                         tp = get_entity_type(ent);
 413                         if (! is_Array_type(tp))
 414                                 break;
 415                         ent = get_array_element_entity(tp);
 416                         size = get_type_size_bytes(get_entity_type(ent));
 417                         sz   = new_tarval_from_long(size, mode);
 418
 419                         tv_index = tarval_div(tmp, sz);
 420                         tmp      = tarval_mod(tmp, sz);
 421
 422                         if (tv_index == tarval_bad || tmp == tarval_bad)
 423                                 return NULL;
 424
 425                         assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
 426                         bound  = get_array_lower_bound(tp, 0);
 427                         tlower = computed_value(bound);
 428                         bound  = get_array_upper_bound(tp, 0);
 429                         tupper = computed_value(bound);
 430
 431                         if (tlower == tarval_bad || tupper == tarval_bad)
 432                                 return NULL;
 433
 434                         if (tarval_cmp(tv_index, tlower) == ir_relation_less)
 435                                 return NULL;
 436                         if (tarval_cmp(tupper, tv_index) == ir_relation_less)
 437                                 return NULL;
 438
 439                         /* ok, bounds check finished */
 440                         ++idx;
 441                 }
 442                 if (! tarval_is_null(tmp)) {
 443                         /* access to some struct/union member */
 444                         return NULL;
 445                 }
 446
 447                 /* should be at least ONE array */
 448                 if (idx == 0)
 449                         return NULL;
 450
 451                 res = rec_get_accessed_path(ptr, depth + idx);
 452                 if (res == NULL)
 453                         return NULL;
 454
 455                 path_len = get_compound_graph_path_length(res);
 456                 pos      = path_len - depth - idx;
 457
 458                 for (ent = field;;) {
 459                         unsigned   size;
 460                         ir_tarval *sz, *tv_index;
 461                         long       index;
 462
 463                         tp = get_entity_type(ent);
 464                         if (! is_Array_type(tp))
 465                                 break;
 466                         ent = get_array_element_entity(tp);
 467                         set_compound_graph_path_node(res, pos, ent);
 468
 469                         size = get_type_size_bytes(get_entity_type(ent));
 470                         sz   = new_tarval_from_long(size, mode);
 471
 472                         tv_index = tarval_div(tv, sz);
 473                         tv       = tarval_mod(tv, sz);
 474
 475                         /* worked above, should work again */
 476                         assert(tv_index != tarval_bad && tv != tarval_bad);
 477
 478                         /* bounds already checked above */
 479                         index = get_tarval_long(tv_index);
 480                         set_compound_graph_path_array_index(res, pos, index);
 481                         ++pos;
 482                 }
 483         } else if (is_Sub(ptr)) {
 484                 ir_node *l = get_Sub_left(ptr);
 485                 ir_node *r = get_Sub_right(ptr);
 486
 487                 ptr = l;
 488                 tv  = get_Const_tarval(r);
 489                 tv  = tarval_neg(tv);
 490                 goto ptr_arith;
 491         }
 492         return res;
 493 }  /* rec_get_accessed_path */
 494
 495 /**
 496  * Returns an access path or NULL.  The access path is only
 497  * valid, if the graph is in phase_high and _no_ address computation is used.
 498  */
 499 static compound_graph_path *get_accessed_path(ir_node *ptr)
 500 {
 501         compound_graph_path *gr = rec_get_accessed_path(ptr, 0);
 502         return gr;
 503 }  /* get_accessed_path */
 504
 505 typedef struct path_entry {
 506         ir_entity         *ent;
 507         struct path_entry *next;
 508         size_t            index;
 509 } path_entry;
 510
 511 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next)
 512 {
 513         path_entry       entry, *p;
 514         ir_entity        *ent, *field;
 515         ir_initializer_t *initializer;
 516         ir_tarval        *tv;
 517         ir_type          *tp;
 518         size_t           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 >= 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                         size_t 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_mode  *mode;
 588                 unsigned pos;
 589
 590                 {
 591                         ir_node *l = get_Add_left(ptr);
 592                         ir_node *r = get_Add_right(ptr);
 593                         if (is_Const(r)) {
 594                                 ptr = l;
 595                                 tv  = get_Const_tarval(r);
 596                         } else {
 597                                 ptr = r;
 598                                 tv  = get_Const_tarval(l);
 599                         }
 600                 }
 601 ptr_arith:
 602                 mode = get_tarval_mode(tv);
 603
 604                 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
 605                 if (is_Sel(ptr)) {
 606                         field = get_Sel_entity(ptr);
 607                 } else {
 608                         field = get_SymConst_entity(ptr);
 609                 }
 610
 611                 /* count needed entries */
 612                 pos = 0;
 613                 for (ent = field;;) {
 614                         tp = get_entity_type(ent);
 615                         if (! is_Array_type(tp))
 616                                 break;
 617                         ent = get_array_element_entity(tp);
 618                         ++pos;
 619                 }
 620                 /* should be at least ONE entry */
 621                 if (pos == 0)
 622                         return NULL;
 623
 624                 /* allocate the right number of entries */
 625                 NEW_ARR_A(path_entry, p, pos);
 626
 627                 /* fill them up */
 628                 pos = 0;
 629                 for (ent = field;;) {
 630                         unsigned   size;
 631                         ir_tarval *sz, *tv_index, *tlower, *tupper;
 632                         long       index;
 633                         ir_node   *bound;
 634
 635                         tp = get_entity_type(ent);
 636                         if (! is_Array_type(tp))
 637                                 break;
 638                         ent = get_array_element_entity(tp);
 639                         p[pos].ent  = ent;
 640                         p[pos].next = &p[pos + 1];
 641
 642                         size = get_type_size_bytes(get_entity_type(ent));
 643                         sz   = new_tarval_from_long(size, mode);
 644
 645                         tv_index = tarval_div(tv, sz);
 646                         tv       = tarval_mod(tv, sz);
 647
 648                         if (tv_index == tarval_bad || tv == tarval_bad)
 649                                 return NULL;
 650
 651                         assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
 652                         bound  = get_array_lower_bound(tp, 0);
 653                         tlower = computed_value(bound);
 654                         bound  = get_array_upper_bound(tp, 0);
 655                         tupper = computed_value(bound);
 656
 657                         if (tlower == tarval_bad || tupper == tarval_bad)
 658                                 return NULL;
 659
 660                         if (tarval_cmp(tv_index, tlower) == ir_relation_less)
 661                                 return NULL;
 662                         if (tarval_cmp(tupper, tv_index) == ir_relation_less)
 663                                 return NULL;
 664
 665                         /* ok, bounds check finished */
 666                         index = get_tarval_long(tv_index);
 667                         p[pos].index = index;
 668                         ++pos;
 669                 }
 670                 if (! tarval_is_null(tv)) {
 671                         /* hmm, wrong access */
 672                         return NULL;
 673                 }
 674                 p[pos - 1].next = next;
 675                 return rec_find_compound_ent_value(ptr, p);
 676         } else if (is_Sub(ptr)) {
 677                 ir_node *l = get_Sub_left(ptr);
 678                 ir_node *r = get_Sub_right(ptr);
 679
 680                 ptr = l;
 681                 tv  = get_Const_tarval(r);
 682                 tv  = tarval_neg(tv);
 683                 goto ptr_arith;
 684         }
 685         return NULL;
 686 }
 687
 688 static ir_node *find_compound_ent_value(ir_node *ptr)
 689 {
 690         return rec_find_compound_ent_value(ptr, NULL);
 691 }
 692
 693 /* forward */
 694 static void reduce_adr_usage(ir_node *ptr);
 695
 696 /**
 697  * Update a Load that may have lost its users.
 698  */
 699 static void handle_load_update(ir_node *load)
 700 {
 701         ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
 702
 703         /* do NOT touch volatile loads for now */
 704         if (get_Load_volatility(load) == volatility_is_volatile)
 705                 return;
 706
 707         if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
 708                 ir_node *ptr = get_Load_ptr(load);
 709                 ir_node *mem = get_Load_mem(load);
 710
 711                 /* a Load whose value is neither used nor exception checked, remove it */
 712                 if (info->projs[pn_Load_M])
 713                         exchange(info->projs[pn_Load_M], mem);
 714                 if (info->projs[pn_Load_X_regular])
 715                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
 716                 kill_node(load);
 717                 reduce_adr_usage(ptr);
 718         }
 719 }  /* handle_load_update */
 720
 721 /**
 722  * A use of an address node has vanished. Check if this was a Proj
 723  * node and update the counters.
 724  */
 725 static void reduce_adr_usage(ir_node *ptr)
 726 {
 727         ir_node *pred;
 728         if (!is_Proj(ptr))
 729                 return;
 730         if (get_irn_n_edges(ptr) > 0)
 731                 return;
 732
 733         /* this Proj is dead now */
 734         pred = get_Proj_pred(ptr);
 735         if (is_Load(pred)) {
 736                 ldst_info_t *info = (ldst_info_t*)get_irn_link(pred);
 737                 info->projs[get_Proj_proj(ptr)] = NULL;
 738
 739                 /* this node lost its result proj, handle that */
 740                 handle_load_update(pred);
 741         }
 742 }  /* reduce_adr_usage */
 743
 744 /**
 745  * Check, if an already existing value of mode old_mode can be converted
 746  * into the needed one new_mode without loss.
 747  */
 748 static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode)
 749 {
 750         unsigned old_size;
 751         unsigned new_size;
 752         if (old_mode == new_mode)
 753                 return true;
 754
 755         old_size = get_mode_size_bits(old_mode);
 756         new_size = get_mode_size_bits(new_mode);
 757
 758         /* if both modes are two-complement ones, we can always convert the
 759            Stored value into the needed one. (on big endian machines we currently
 760            only support this for modes of same size) */
 761         if (old_size >= new_size &&
 762                   get_mode_arithmetic(old_mode) == irma_twos_complement &&
 763                   get_mode_arithmetic(new_mode) == irma_twos_complement &&
 764                   (!be_get_backend_param()->byte_order_big_endian
 765                 || old_size == new_size)) {
 766                 return true;
 767         }
 768         return false;
 769 }
 770
 771 /**
 772  * Check whether a Call is at least pure, i.e. does only read memory.
 773  */
 774 static unsigned is_Call_pure(ir_node *call)
 775 {
 776         ir_type *call_tp = get_Call_type(call);
 777         unsigned prop = get_method_additional_properties(call_tp);
 778
 779         /* check first the call type */
 780         if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
 781                 /* try the called entity */
 782                 ir_node *ptr = get_Call_ptr(call);
 783
 784                 if (is_Global(ptr)) {
 785                         ir_entity *ent = get_Global_entity(ptr);
 786
 787                         prop = get_entity_additional_properties(ent);
 788                 }
 789         }
 790         return (prop & (mtp_property_const|mtp_property_pure)) != 0;
 791 }  /* is_Call_pure */
 792
 793 static ir_node *get_base_and_offset(ir_node *ptr, long *pOffset)
 794 {
 795         ir_mode *mode  = get_irn_mode(ptr);
 796         long    offset = 0;
 797
 798         /* TODO: long might not be enough, we should probably use some tarval thingy... */
 799         for (;;) {
 800                 if (is_Add(ptr)) {
 801                         ir_node *l = get_Add_left(ptr);
 802                         ir_node *r = get_Add_right(ptr);
 803
 804                         if (get_irn_mode(l) != mode || !is_Const(r))
 805                                 break;
 806
 807                         offset += get_tarval_long(get_Const_tarval(r));
 808                         ptr     = l;
 809                 } else if (is_Sub(ptr)) {
 810                         ir_node *l = get_Sub_left(ptr);
 811                         ir_node *r = get_Sub_right(ptr);
 812
 813                         if (get_irn_mode(l) != mode || !is_Const(r))
 814                                 break;
 815
 816                         offset -= get_tarval_long(get_Const_tarval(r));
 817                         ptr     = l;
 818                 } else if (is_Sel(ptr)) {
 819                         ir_entity *ent = get_Sel_entity(ptr);
 820                         ir_type   *tp  = get_entity_owner(ent);
 821
 822                         if (is_Array_type(tp)) {
 823                                 int     size;
 824                                 ir_node *index;
 825
 826                                 /* only one dimensional arrays yet */
 827                                 if (get_Sel_n_indexs(ptr) != 1)
 828                                         break;
 829                                 index = get_Sel_index(ptr, 0);
 830                                 if (! is_Const(index))
 831                                         break;
 832
 833                                 tp = get_entity_type(ent);
 834                                 if (get_type_state(tp) != layout_fixed)
 835                                         break;
 836
 837                                 size    = get_type_size_bytes(tp);
 838                                 offset += size * get_tarval_long(get_Const_tarval(index));
 839                         } else {
 840                                 if (get_type_state(tp) != layout_fixed)
 841                                         break;
 842                                 offset += get_entity_offset(ent);
 843                         }
 844                         ptr = get_Sel_ptr(ptr);
 845                 } else
 846                         break;
 847         }
 848
 849         *pOffset = offset;
 850         return ptr;
 851 }
 852
 853 static int try_load_after_store(ir_node *load,
 854                 ir_node *load_base_ptr, long load_offset, ir_node *store)
 855 {
 856         ldst_info_t *info;
 857         ir_node *store_ptr      = get_Store_ptr(store);
 858         long     store_offset;
 859         ir_node *store_base_ptr = get_base_and_offset(store_ptr, &store_offset);
 860         ir_node *store_value;
 861         ir_mode *store_mode;
 862         ir_node *load_ptr;
 863         ir_mode *load_mode;
 864         long     load_mode_len;
 865         long     store_mode_len;
 866         long     delta;
 867         int      res;
 868
 869         if (load_base_ptr != store_base_ptr)
 870                 return 0;
 871
 872         load_mode      = get_Load_mode(load);
 873         load_mode_len  = get_mode_size_bytes(load_mode);
 874         store_mode     = get_irn_mode(get_Store_value(store));
 875         store_mode_len = get_mode_size_bytes(store_mode);
 876         delta          = load_offset - store_offset;
 877         store_value    = get_Store_value(store);
 878
 879         if (delta != 0 || store_mode != load_mode) {
 880                 /* TODO: implement for big-endian */
 881                 if (delta < 0 || delta + load_mode_len > store_mode_len
 882                                 || (be_get_backend_param()->byte_order_big_endian
 883                                     && load_mode_len != store_mode_len))
 884                         return 0;
 885
 886                 if (get_mode_arithmetic(store_mode) != irma_twos_complement ||
 887                         get_mode_arithmetic(load_mode)  != irma_twos_complement)
 888                         return 0;
 889
 890
 891                 /* produce a shift to adjust offset delta */
 892                 if (delta > 0) {
 893                         ir_node *cnst;
 894                         ir_graph *irg = get_irn_irg(load);
 895
 896                         cnst        = new_r_Const_long(irg, mode_Iu, delta * 8);
 897                         store_value = new_r_Shr(get_nodes_block(load),
 898                                                                         store_value, cnst, store_mode);
 899                 }
 900
 901                 /* add an convert if needed */
 902                 if (store_mode != load_mode) {
 903                         store_value = new_r_Conv(get_nodes_block(load), store_value, load_mode);
 904                 }
 905         }
 906
 907         DBG_OPT_RAW(load, store_value);
 908
 909         info = (ldst_info_t*)get_irn_link(load);
 910         if (info->projs[pn_Load_M])
 911                 exchange(info->projs[pn_Load_M], get_Load_mem(load));
 912
 913         res = 0;
 914         /* no exception */
 915         if (info->projs[pn_Load_X_except]) {
 916                 ir_graph *irg = get_irn_irg(load);
 917                 exchange( info->projs[pn_Load_X_except], new_r_Bad(irg));
 918                 res |= CF_CHANGED;
 919         }
 920         if (info->projs[pn_Load_X_regular]) {
 921                 exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
 922                 res |= CF_CHANGED;
 923         }
 924
 925         if (info->projs[pn_Load_res])
 926                 exchange(info->projs[pn_Load_res], store_value);
 927
 928         load_ptr = get_Load_ptr(load);
 929         kill_node(load);
 930         reduce_adr_usage(load_ptr);
 931         return res | DF_CHANGED;
 932 }
 933
 934 /**
 935  * Follow the memory chain as long as there are only Loads,
 936  * alias free Stores, and constant Calls and try to replace the
 937  * current Load by a previous ones.
 938  * Note that in unreachable loops it might happen that we reach
 939  * load again, as well as we can fall into a cycle.
 940  * We break such cycles using a special visited flag.
 941  *
 942  * INC_MASTER() must be called before dive into
 943  */
 944 static unsigned follow_Mem_chain(ir_node *load, ir_node *curr)
 945 {
 946         unsigned    res = 0;
 947         ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
 948         ir_node     *pred;
 949         ir_node     *ptr       = get_Load_ptr(load);
 950         ir_node     *mem       = get_Load_mem(load);
 951         ir_mode     *load_mode = get_Load_mode(load);
 952
 953         for (pred = curr; load != pred; ) {
 954                 ldst_info_t *pred_info = (ldst_info_t*)get_irn_link(pred);
 955
 956                 /*
 957                  * a Load immediately after a Store -- a read after write.
 958                  * We may remove the Load, if both Load & Store does not have an
 959                  * exception handler OR they are in the same Block. In the latter
 960                  * case the Load cannot throw an exception when the previous Store was
 961                  * quiet.
 962                  *
 963                  * Why we need to check for Store Exception? If the Store cannot
 964                  * be executed (ROM) the exception handler might simply jump into
 965                  * the load Block :-(
 966                  * We could make it a little bit better if we would know that the
 967                  * exception handler of the Store jumps directly to the end...
 968                  */
 969                 if (is_Store(pred) && ((pred_info->projs[pn_Store_X_except] == NULL
 970                                 && info->projs[pn_Load_X_except] == NULL)
 971                                 || get_nodes_block(load) == get_nodes_block(pred)))
 972                 {
 973                         long    load_offset;
 974                         ir_node *base_ptr = get_base_and_offset(ptr, &load_offset);
 975                         int     changes   = try_load_after_store(load, base_ptr, load_offset, pred);
 976
 977                         if (changes != 0)
 978                                 return res | changes;
 979                 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
 980                            can_use_stored_value(get_Load_mode(pred), load_mode)) {
 981                         /*
 982                          * a Load after a Load -- a read after read.
 983                          * We may remove the second Load, if it does not have an exception
 984                          * handler OR they are in the same Block. In the later case
 985                          * the Load cannot throw an exception when the previous Load was
 986                          * quiet.
 987                          *
 988                          * Here, there is no need to check if the previous Load has an
 989                          * exception hander because they would have exact the same
 990                          * exception...
 991                          *
 992                          * TODO: implement load-after-load with different mode for big
 993                          *       endian
 994                          */
 995                         if (info->projs[pn_Load_X_except] == NULL
 996                                         || get_nodes_block(load) == get_nodes_block(pred)) {
 997                                 ir_node *value;
 998
 999                                 DBG_OPT_RAR(load, pred);
1000
1001                                 /* the result is used */
1002                                 if (info->projs[pn_Load_res]) {
1003                                         if (pred_info->projs[pn_Load_res] == NULL) {
1004                                                 /* create a new Proj again */
1005                                                 pred_info->projs[pn_Load_res] = new_r_Proj(pred, get_Load_mode(pred), pn_Load_res);
1006                                         }
1007                                         value = pred_info->projs[pn_Load_res];
1008
1009                                         /* add an convert if needed */
1010                                         if (get_Load_mode(pred) != load_mode) {
1011                                                 value = new_r_Conv(get_nodes_block(load), value, load_mode);
1012                                         }
1013
1014                                         exchange(info->projs[pn_Load_res], value);
1015                                 }
1016
1017                                 if (info->projs[pn_Load_M])
1018                                         exchange(info->projs[pn_Load_M], mem);
1019
1020                                 /* no exception */
1021                                 if (info->projs[pn_Load_X_except]) {
1022                                         ir_graph *irg = get_irn_irg(load);
1023                                         exchange(info->projs[pn_Load_X_except], new_r_Bad(irg));
1024                                         res |= CF_CHANGED;
1025                                 }
1026                                 if (info->projs[pn_Load_X_regular]) {
1027                                         exchange( info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1028                                         res |= CF_CHANGED;
1029                                 }
1030
1031                                 kill_node(load);
1032                                 reduce_adr_usage(ptr);
1033                                 return res |= DF_CHANGED;
1034                         }
1035                 }
1036
1037                 if (is_Store(pred)) {
1038                         /* check if we can pass through this store */
1039                         ir_alias_relation rel = get_alias_relation(
1040                                 get_Store_ptr(pred),
1041                                 get_irn_mode(get_Store_value(pred)),
1042                                 ptr, load_mode);
1043                         /* if the might be an alias, we cannot pass this Store */
1044                         if (rel != ir_no_alias)
1045                                 break;
1046                         pred = skip_Proj(get_Store_mem(pred));
1047                 } else if (is_Load(pred)) {
1048                         pred = skip_Proj(get_Load_mem(pred));
1049                 } else if (is_Call(pred)) {
1050                         if (is_Call_pure(pred)) {
1051                                 /* The called graph is at least pure, so there are no Store's
1052                                    in it. We can handle it like a Load and skip it. */
1053                                 pred = skip_Proj(get_Call_mem(pred));
1054                         } else {
1055                                 /* there might be Store's in the graph, stop here */
1056                                 break;
1057                         }
1058                 } else {
1059                         /* follow only Load chains */
1060                         break;
1061                 }
1062
1063                 /* check for cycles */
1064                 if (NODE_VISITED(pred_info))
1065                         break;
1066                 MARK_NODE(pred_info);
1067         }
1068
1069         if (is_Sync(pred)) {
1070                 int i;
1071
1072                 /* handle all Sync predecessors */
1073                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1074                         res |= follow_Mem_chain(load, skip_Proj(get_Sync_pred(pred, i)));
1075                         if (res)
1076                                 return res;
1077                 }
1078         }
1079
1080         return res;
1081 }  /* follow_Mem_chain */
1082
1083 /*
1084  * Check if we can replace the load by a given const from
1085  * the const code irg.
1086  */
1087 ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c)
1088 {
1089         ir_mode  *c_mode = get_irn_mode(c);
1090         ir_mode  *l_mode = get_Load_mode(load);
1091         ir_node  *block  = get_nodes_block(load);
1092         dbg_info *dbgi   = get_irn_dbg_info(load);
1093         ir_node  *res    = copy_const_value(dbgi, c, block);
1094
1095         if (c_mode != l_mode) {
1096                 /* check, if the mode matches OR can be easily converted info */
1097                 if (is_reinterpret_cast(c_mode, l_mode)) {
1098                         /* copy the value from the const code irg and cast it */
1099                         res = new_rd_Conv(dbgi, block, res, l_mode);
1100                 }
1101                 return NULL;
1102         }
1103         return res;
1104 }
1105
1106 /**
1107  * optimize a Load
1108  *
1109  * @param load  the Load node
1110  */
1111 static unsigned optimize_load(ir_node *load)
1112 {
1113         ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
1114         ir_node     *mem, *ptr, *value;
1115         ir_entity   *ent;
1116         long        dummy;
1117         unsigned    res = 0;
1118
1119         /* do NOT touch volatile loads for now */
1120         if (get_Load_volatility(load) == volatility_is_volatile)
1121                 return 0;
1122
1123         /* the address of the load to be optimized */
1124         ptr = get_Load_ptr(load);
1125
1126         /* The mem of the Load. Must still be returned after optimization. */
1127         mem = get_Load_mem(load);
1128
1129         if (info->projs[pn_Load_res] == NULL
1130                         && info->projs[pn_Load_X_except] == NULL) {
1131                 /* the value is never used and we don't care about exceptions, remove */
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(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                                 && get_entity_visibility(ent) != ir_visibility_external) {
1153                         /* a static allocation that is not external: there should be NO
1154                          * exception when loading even if we cannot replace the load itself.
1155                          */
1156
1157                         /* no exception, clear the info field as it might be checked later again */
1158                         if (info->projs[pn_Load_X_except]) {
1159                                 ir_graph *irg = get_irn_irg(load);
1160                                 exchange(info->projs[pn_Load_X_except], new_r_Bad(irg));
1161                                 info->projs[pn_Load_X_except] = NULL;
1162                                 res |= CF_CHANGED;
1163                         }
1164                         if (info->projs[pn_Load_X_regular]) {
1165                                 exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1166                                 info->projs[pn_Load_X_regular] = NULL;
1167                                 res |= CF_CHANGED;
1168                         }
1169
1170                         if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT) {
1171                                 if (ent->initializer != NULL) {
1172                                         /* new style initializer */
1173                                         value = find_compound_ent_value(ptr);
1174                                 } else if (entity_has_compound_ent_values(ent)) {
1175                                         /* old style initializer */
1176                                         compound_graph_path *path = get_accessed_path(ptr);
1177
1178                                         if (path != NULL) {
1179                                                 assert(is_proper_compound_graph_path(path, get_compound_graph_path_length(path)-1));
1180
1181                                                 value = get_compound_ent_value_by_path(ent, path);
1182                                                 DB((dbg, LEVEL_1, "  Constant access at %F%F resulted in %+F\n", ent, path, value));
1183                                                 free_compound_graph_path(path);
1184                                         }
1185                                 }
1186                                 if (value != NULL)
1187                                         value = can_replace_load_by_const(load, value);
1188                         }
1189                 }
1190         }
1191         if (value != NULL) {
1192                 /* we completely replace the load by this value */
1193                 if (info->projs[pn_Load_X_except]) {
1194                         ir_graph *irg = get_irn_irg(load);
1195                         exchange(info->projs[pn_Load_X_except], new_r_Bad(irg));
1196                         info->projs[pn_Load_X_except] = NULL;
1197                         res |= CF_CHANGED;
1198                 }
1199                 if (info->projs[pn_Load_X_regular]) {
1200                         exchange(info->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
1201                         info->projs[pn_Load_X_regular] = NULL;
1202                         res |= CF_CHANGED;
1203                 }
1204                 if (info->projs[pn_Load_M]) {
1205                         exchange(info->projs[pn_Load_M], mem);
1206                         res |= DF_CHANGED;
1207                 }
1208                 if (info->projs[pn_Load_res]) {
1209                         exchange(info->projs[pn_Load_res], value);
1210                         res |= DF_CHANGED;
1211                 }
1212                 kill_node(load);
1213                 reduce_adr_usage(ptr);
1214                 return res;
1215         }
1216
1217         /* Check, if the address of this load is used more than once.
1218          * If not, more load cannot be removed in any case. */
1219         if (get_irn_n_uses(ptr) <= 1 && get_irn_n_uses(get_base_and_offset(ptr, &dummy)) <= 1)
1220                 return res;
1221
1222         /*
1223          * follow the memory chain as long as there are only Loads
1224          * and try to replace current Load or Store by a previous one.
1225          * Note that in unreachable loops it might happen that we reach
1226          * load again, as well as we can fall into a cycle.
1227          * We break such cycles using a special visited flag.
1228          */
1229         INC_MASTER();
1230         res = follow_Mem_chain(load, skip_Proj(mem));
1231         return res;
1232 }  /* optimize_load */
1233
1234 /**
1235  * Check whether a value of mode new_mode would completely overwrite a value
1236  * of mode old_mode in memory.
1237  */
1238 static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
1239 {
1240         return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
1241 }  /* is_completely_overwritten */
1242
1243 /**
1244  * Check whether small is a part of large (starting at same address).
1245  */
1246 static int is_partially_same(ir_node *small, ir_node *large)
1247 {
1248         ir_mode *sm = get_irn_mode(small);
1249         ir_mode *lm = get_irn_mode(large);
1250
1251         /* FIXME: Check endianness */
1252         return is_Conv(small) && get_Conv_op(small) == large
1253             && get_mode_size_bytes(sm) < get_mode_size_bytes(lm)
1254             && get_mode_arithmetic(sm) == irma_twos_complement
1255             && get_mode_arithmetic(lm) == irma_twos_complement;
1256 }  /* is_partially_same */
1257
1258 /**
1259  * follow the memory chain as long as there are only Loads and alias free Stores.
1260  *
1261  * INC_MASTER() must be called before dive into
1262  */
1263 static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr)
1264 {
1265         unsigned res = 0;
1266         ldst_info_t *info = (ldst_info_t*)get_irn_link(store);
1267         ir_node *pred;
1268         ir_node *ptr = get_Store_ptr(store);
1269         ir_node *mem = get_Store_mem(store);
1270         ir_node *value = get_Store_value(store);
1271         ir_mode *mode  = get_irn_mode(value);
1272         ir_node *block = get_nodes_block(store);
1273
1274         for (pred = curr; pred != store;) {
1275                 ldst_info_t *pred_info = (ldst_info_t*)get_irn_link(pred);
1276
1277                 /*
1278                  * BEWARE: one might think that checking the modes is useless, because
1279                  * if the pointers are identical, they refer to the same object.
1280                  * This is only true in strong typed languages, not is C were the following
1281                  * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
1282                  * However, if the size of the mode that is written is bigger or equal the
1283                  * size of the old one, the old value is completely overwritten and can be
1284                  * killed ...
1285                  */
1286                 if (is_Store(pred) && get_Store_ptr(pred) == ptr &&
1287             get_nodes_block(pred) == block) {
1288                         /*
1289                          * a Store after a Store in the same Block -- a write after write.
1290                          */
1291
1292                         /*
1293                          * We may remove the first Store, if the old value is completely
1294                          * overwritten or the old value is a part of the new value,
1295                          * and if it does not have an exception handler.
1296                          *
1297                          * TODO: What, if both have the same exception handler ???
1298                          */
1299                         if (get_Store_volatility(pred) != volatility_is_volatile
1300                                 && !pred_info->projs[pn_Store_X_except]) {
1301                                 ir_node *predvalue = get_Store_value(pred);
1302                                 ir_mode *predmode  = get_irn_mode(predvalue);
1303
1304                                 if (is_completely_overwritten(predmode, mode)
1305                                         || is_partially_same(predvalue, value)) {
1306                                         DBG_OPT_WAW(pred, store);
1307                                         exchange(pred_info->projs[pn_Store_M], get_Store_mem(pred));
1308                                         kill_node(pred);
1309                                         reduce_adr_usage(ptr);
1310                                         return DF_CHANGED;
1311                                 }
1312                         }
1313
1314                         /*
1315                          * We may remove the Store, if the old value already contains
1316                          * the new value, and if it does not have an exception handler.
1317                          *
1318                          * TODO: What, if both have the same exception handler ???
1319                          */
1320                         if (get_Store_volatility(store) != volatility_is_volatile
1321                                 && !info->projs[pn_Store_X_except]) {
1322                                 ir_node *predvalue = get_Store_value(pred);
1323
1324                                 if (is_partially_same(value, predvalue)) {
1325                                         DBG_OPT_WAW(pred, store);
1326                                         exchange(info->projs[pn_Store_M], mem);
1327                                         kill_node(store);
1328                                         reduce_adr_usage(ptr);
1329                                         return DF_CHANGED;
1330                                 }
1331                         }
1332                 } else if (is_Load(pred) && get_Load_ptr(pred) == ptr &&
1333                            value == pred_info->projs[pn_Load_res]) {
1334                         /*
1335                          * a Store of a value just loaded from the same address
1336                          * -- a write after read.
1337                          * We may remove the Store, if it does not have an exception
1338                          * handler.
1339                          */
1340                         if (! info->projs[pn_Store_X_except]) {
1341                                 DBG_OPT_WAR(store, pred);
1342                                 exchange(info->projs[pn_Store_M], mem);
1343                                 kill_node(store);
1344                                 reduce_adr_usage(ptr);
1345                                 return DF_CHANGED;
1346                         }
1347                 }
1348
1349                 if (is_Store(pred)) {
1350                         /* check if we can pass through this store */
1351                         ir_alias_relation rel = get_alias_relation(
1352                                 get_Store_ptr(pred),
1353                                 get_irn_mode(get_Store_value(pred)),
1354                                 ptr, mode);
1355                         /* if the might be an alias, we cannot pass this Store */
1356                         if (rel != ir_no_alias)
1357                                 break;
1358                         pred = skip_Proj(get_Store_mem(pred));
1359                 } else if (is_Load(pred)) {
1360                         ir_alias_relation rel = get_alias_relation(
1361                                 get_Load_ptr(pred), get_Load_mode(pred),
1362                                 ptr, mode);
1363                         if (rel != ir_no_alias)
1364                                 break;
1365
1366                         pred = skip_Proj(get_Load_mem(pred));
1367                 } else {
1368                         /* follow only Load chains */
1369                         break;
1370                 }
1371
1372                 /* check for cycles */
1373                 if (NODE_VISITED(pred_info))
1374                         break;
1375                 MARK_NODE(pred_info);
1376         }
1377
1378         if (is_Sync(pred)) {
1379                 int i;
1380
1381                 /* handle all Sync predecessors */
1382                 for (i = get_Sync_n_preds(pred) - 1; i >= 0; --i) {
1383                         res |= follow_Mem_chain_for_Store(store, skip_Proj(get_Sync_pred(pred, i)));
1384                         if (res)
1385                                 break;
1386                 }
1387         }
1388         return res;
1389 }  /* follow_Mem_chain_for_Store */
1390
1391 /** find entity used as base for an address calculation */
1392 static ir_entity *find_entity(ir_node *ptr)
1393 {
1394         switch (get_irn_opcode(ptr)) {
1395         case iro_SymConst:
1396                 return get_SymConst_entity(ptr);
1397         case iro_Sel: {
1398                 ir_node *pred = get_Sel_ptr(ptr);
1399                 if (get_irg_frame(get_irn_irg(ptr)) == pred)
1400                         return get_Sel_entity(ptr);
1401
1402                 return find_entity(pred);
1403         }
1404         case iro_Sub:
1405         case iro_Add: {
1406                 ir_node *left = get_binop_left(ptr);
1407                 ir_node *right;
1408                 if (mode_is_reference(get_irn_mode(left)))
1409                         return find_entity(left);
1410                 right = get_binop_right(ptr);
1411                 if (mode_is_reference(get_irn_mode(right)))
1412                         return find_entity(right);
1413                 return NULL;
1414         }
1415         default:
1416                 return NULL;
1417         }
1418 }
1419
1420 /**
1421  * optimize a Store
1422  *
1423  * @param store  the Store node
1424  */
1425 static unsigned optimize_store(ir_node *store)
1426 {
1427         ir_node   *ptr;
1428         ir_node   *mem;
1429         ir_entity *entity;
1430
1431         if (get_Store_volatility(store) == volatility_is_volatile)
1432                 return 0;
1433
1434         ptr    = get_Store_ptr(store);
1435         entity = find_entity(ptr);
1436
1437         /* a store to an entity which is never read is unnecessary */
1438         if (entity != NULL && !(get_entity_usage(entity) & ir_usage_read)) {
1439                 ldst_info_t *info = (ldst_info_t*)get_irn_link(store);
1440                 if (info->projs[pn_Store_X_except] == NULL) {
1441                         DB((dbg, LEVEL_1, "  Killing useless %+F to never read entity %+F\n", store, entity));
1442                         exchange(info->projs[pn_Store_M], get_Store_mem(store));
1443                         kill_node(store);
1444                         reduce_adr_usage(ptr);
1445                         return DF_CHANGED;
1446                 }
1447         }
1448
1449         /* Check, if the address of this Store is used more than once.
1450          * If not, this Store cannot be removed in any case. */
1451         if (get_irn_n_uses(ptr) <= 1)
1452                 return 0;
1453
1454         mem = get_Store_mem(store);
1455
1456         /* follow the memory chain as long as there are only Loads */
1457         INC_MASTER();
1458
1459         return follow_Mem_chain_for_Store(store, skip_Proj(mem));
1460 }  /* optimize_store */
1461
1462 /**
1463  * walker, optimizes Phi after Stores to identical places:
1464  * Does the following optimization:
1465  * @verbatim
1466  *
1467  *   val1   val2   val3          val1  val2  val3
1468  *    |      |      |               \    |    /
1469  *  Store  Store  Store              \   |   /
1470  *      \    |    /                   PhiData
1471  *       \   |   /                       |
1472  *        \  |  /                      Store
1473  *          PhiM
1474  *
1475  * @endverbatim
1476  * This reduces the number of stores and allows for predicated execution.
1477  * Moves Stores back to the end of a function which may be bad.
1478  *
1479  * This is only possible if the predecessor blocks have only one successor.
1480  */
1481 static unsigned optimize_phi(ir_node *phi, walk_env_t *wenv)
1482 {
1483         int i, n;
1484         ir_node *store, *old_store, *ptr, *block, *phi_block, *phiM, *phiD, *exc, *projM;
1485         ir_mode *mode;
1486         ir_node **inM, **inD, **projMs;
1487         int *idx;
1488         dbg_info *db = NULL;
1489         ldst_info_t *info;
1490         block_info_t *bl_info;
1491         unsigned res = 0;
1492
1493         /* Must be a memory Phi */
1494         if (get_irn_mode(phi) != mode_M)
1495                 return 0;
1496
1497         n = get_Phi_n_preds(phi);
1498         if (n <= 0)
1499                 return 0;
1500
1501         /* must be only one user */
1502         projM = get_Phi_pred(phi, 0);
1503         if (get_irn_n_edges(projM) != 1)
1504                 return 0;
1505
1506         store = skip_Proj(projM);
1507         old_store = store;
1508         if (!is_Store(store))
1509                 return 0;
1510
1511         block = get_nodes_block(store);
1512
1513         /* abort on dead blocks */
1514         if (is_Block_dead(block))
1515                 return 0;
1516
1517         /* check if the block is post dominated by Phi-block
1518            and has no exception exit */
1519         bl_info = (block_info_t*)get_irn_link(block);
1520         if (bl_info->flags & BLOCK_HAS_EXC)
1521                 return 0;
1522
1523         phi_block = get_nodes_block(phi);
1524         if (! block_strictly_postdominates(phi_block, block))
1525                 return 0;
1526
1527         /* this is the address of the store */
1528         ptr  = get_Store_ptr(store);
1529         mode = get_irn_mode(get_Store_value(store));
1530         info = (ldst_info_t*)get_irn_link(store);
1531         exc  = info->exc_block;
1532
1533         for (i = 1; i < n; ++i) {
1534                 ir_node *pred = get_Phi_pred(phi, i);
1535
1536                 if (get_irn_n_edges(pred) != 1)
1537                         return 0;
1538
1539                 pred = skip_Proj(pred);
1540                 if (!is_Store(pred))
1541                         return 0;
1542
1543                 if (ptr != get_Store_ptr(pred) || mode != get_irn_mode(get_Store_value(pred)))
1544                         return 0;
1545
1546                 info = (ldst_info_t*)get_irn_link(pred);
1547
1548                 /* check, if all stores have the same exception flow */
1549                 if (exc != info->exc_block)
1550                         return 0;
1551
1552                 /* abort on dead blocks */
1553                 block = get_nodes_block(pred);
1554                 if (is_Block_dead(block))
1555                         return 0;
1556
1557                 /* check if the block is post dominated by Phi-block
1558                    and has no exception exit. Note that block must be different from
1559                    Phi-block, else we would move a Store from end End of a block to its
1560                    Start... */
1561                 bl_info = (block_info_t*)get_irn_link(block);
1562                 if (bl_info->flags & BLOCK_HAS_EXC)
1563                         return 0;
1564                 if (block == phi_block || ! block_postdominates(phi_block, block))
1565                         return 0;
1566         }
1567
1568         /*
1569          * ok, when we are here, we found all predecessors of a Phi that
1570          * are Stores to the same address and size. That means whatever
1571          * we do before we enter the block of the Phi, we do a Store.
1572          * So, we can move the Store to the current block:
1573          *
1574          *   val1    val2    val3          val1  val2  val3
1575          *    |       |       |               \    |    /
1576          * | Str | | Str | | Str |             \   |   /
1577          *      \     |     /                   PhiData
1578          *       \    |    /                       |
1579          *        \   |   /                       Str
1580          *           PhiM
1581          *
1582          * Is only allowed if the predecessor blocks have only one successor.
1583          */
1584
1585         NEW_ARR_A(ir_node *, projMs, n);
1586         NEW_ARR_A(ir_node *, inM, n);
1587         NEW_ARR_A(ir_node *, inD, n);
1588         NEW_ARR_A(int, idx, n);
1589
1590         /* Prepare: Collect all Store nodes.  We must do this
1591            first because we otherwise may loose a store when exchanging its
1592            memory Proj.
1593          */
1594         for (i = n - 1; i >= 0; --i) {
1595                 ir_node *store;
1596
1597                 projMs[i] = get_Phi_pred(phi, i);
1598                 assert(is_Proj(projMs[i]));
1599
1600                 store = get_Proj_pred(projMs[i]);
1601                 info  = (ldst_info_t*)get_irn_link(store);
1602
1603                 inM[i] = get_Store_mem(store);
1604                 inD[i] = get_Store_value(store);
1605                 idx[i] = info->exc_idx;
1606         }
1607         block = get_nodes_block(phi);
1608
1609         /* second step: create a new memory Phi */
1610         phiM = new_rd_Phi(get_irn_dbg_info(phi), block, n, inM, mode_M);
1611
1612         /* third step: create a new data Phi */
1613         phiD = new_rd_Phi(get_irn_dbg_info(phi), block, n, inD, mode);
1614
1615         /* rewire memory and kill the node */
1616         for (i = n - 1; i >= 0; --i) {
1617                 ir_node *proj  = projMs[i];
1618
1619                 if (is_Proj(proj)) {
1620                         ir_node *store = get_Proj_pred(proj);
1621                         exchange(proj, inM[i]);
1622                         kill_node(store);
1623                 }
1624         }
1625
1626         /* fourth step: create the Store */
1627         store = new_rd_Store(db, block, phiM, ptr, phiD, cons_none);
1628 #ifdef DO_CACHEOPT
1629         co_set_irn_name(store, co_get_irn_ident(old_store));
1630 #endif
1631
1632         projM = new_rd_Proj(NULL, store, mode_M, pn_Store_M);
1633
1634         info = get_ldst_info(store, &wenv->obst);
1635         info->projs[pn_Store_M] = projM;
1636
1637         /* fifths step: repair exception flow */
1638         if (exc) {
1639                 ir_node *projX = new_rd_Proj(NULL, store, mode_X, pn_Store_X_except);
1640
1641                 info->projs[pn_Store_X_except] = projX;
1642                 info->exc_block                = exc;
1643                 info->exc_idx                  = idx[0];
1644
1645                 for (i = 0; i < n; ++i) {
1646                         set_Block_cfgpred(exc, idx[i], projX);
1647                 }
1648
1649                 if (n > 1) {
1650                         /* the exception block should be optimized as some inputs are identical now */
1651                 }
1652
1653                 res |= CF_CHANGED;
1654         }
1655
1656         /* sixth step: replace old Phi */
1657         exchange(phi, projM);
1658
1659         return res | DF_CHANGED;
1660 }  /* optimize_phi */
1661
1662 /**
1663  * walker, do the optimizations
1664  */
1665 static void do_load_store_optimize(ir_node *n, void *env)
1666 {
1667         walk_env_t *wenv = (walk_env_t*)env;
1668
1669         switch (get_irn_opcode(n)) {
1670
1671         case iro_Load:
1672                 wenv->changes |= optimize_load(n);
1673                 break;
1674
1675         case iro_Store:
1676                 wenv->changes |= optimize_store(n);
1677                 break;
1678
1679         case iro_Phi:
1680                 wenv->changes |= optimize_phi(n, wenv);
1681                 break;
1682
1683         default:
1684                 ;
1685         }
1686 }  /* do_load_store_optimize */
1687
1688 /** A scc. */
1689 typedef struct scc {
1690         ir_node *head;      /**< the head of the list */
1691 } scc;
1692
1693 /** A node entry. */
1694 typedef struct node_entry {
1695         unsigned DFSnum;    /**< the DFS number of this node */
1696         unsigned low;       /**< the low number of this node */
1697         int      in_stack;  /**< flag, set if the node is on the stack */
1698         ir_node  *next;     /**< link to the next node the the same scc */
1699         scc      *pscc;     /**< the scc of this node */
1700         unsigned POnum;     /**< the post order number for blocks */
1701 } node_entry;
1702
1703 /** A loop entry. */
1704 typedef struct loop_env {
1705         ir_phase ph;           /**< the phase object */
1706         ir_node  **stack;      /**< the node stack */
1707         size_t   tos;          /**< tos index */
1708         unsigned nextDFSnum;   /**< the current DFS number */
1709         unsigned POnum;        /**< current post order number */
1710
1711         unsigned changes;      /**< a bitmask of graph changes */
1712 } loop_env;
1713
1714 /**
1715 * Gets the node_entry of a node
1716 */
1717 static node_entry *get_irn_ne(ir_node *irn, loop_env *env)
1718 {
1719         ir_phase   *ph = &env->ph;
1720         node_entry *e  = (node_entry*)phase_get_irn_data(&env->ph, irn);
1721
1722         if (! e) {
1723                 e = (node_entry*)phase_alloc(ph, sizeof(*e));
1724                 memset(e, 0, sizeof(*e));
1725                 phase_set_irn_data(ph, irn, e);
1726         }
1727         return e;
1728 }  /* get_irn_ne */
1729
1730 /**
1731  * Push a node onto the stack.
1732  *
1733  * @param env   the loop environment
1734  * @param n     the node to push
1735  */
1736 static void push(loop_env *env, ir_node *n)
1737 {
1738         node_entry *e;
1739
1740         if (env->tos == ARR_LEN(env->stack)) {
1741                 size_t nlen = ARR_LEN(env->stack) * 2;
1742                 ARR_RESIZE(ir_node *, env->stack, nlen);
1743         }
1744         env->stack[env->tos++] = n;
1745         e = get_irn_ne(n, env);
1746         e->in_stack = 1;
1747 }  /* push */
1748
1749 /**
1750  * pop a node from the stack
1751  *
1752  * @param env   the loop environment
1753  *
1754  * @return  The topmost node
1755  */
1756 static ir_node *pop(loop_env *env)
1757 {
1758         ir_node *n = env->stack[--env->tos];
1759         node_entry *e = get_irn_ne(n, env);
1760
1761         e->in_stack = 0;
1762         return n;
1763 }  /* pop */
1764
1765 /**
1766  * Check if irn is a region constant.
1767  * The block or irn must strictly dominate the header block.
1768  *
1769  * @param irn           the node to check
1770  * @param header_block  the header block of the induction variable
1771  */
1772 static int is_rc(ir_node *irn, ir_node *header_block)
1773 {
1774         ir_node *block = get_nodes_block(irn);
1775
1776         return (block != header_block) && block_dominates(block, header_block);
1777 }  /* is_rc */
1778
1779 typedef struct phi_entry phi_entry;
1780 struct phi_entry {
1781         ir_node   *phi;    /**< A phi with a region const memory. */
1782         int       pos;     /**< The position of the region const memory */
1783         ir_node   *load;   /**< the newly created load for this phi */
1784         phi_entry *next;
1785 };
1786
1787 /**
1788  * An entry in the avail set.
1789  */
1790 typedef struct avail_entry_t {
1791         ir_node *ptr;   /**< the address pointer */
1792         ir_mode *mode;  /**< the load mode */
1793         ir_node *load;  /**< the associated Load */
1794 } avail_entry_t;
1795
1796 /**
1797  * Compare two avail entries.
1798  */
1799 static int cmp_avail_entry(const void *elt, const void *key, size_t size)
1800 {
1801         const avail_entry_t *a = (const avail_entry_t*)elt;
1802         const avail_entry_t *b = (const avail_entry_t*)key;
1803         (void) size;
1804
1805         return a->ptr != b->ptr || a->mode != b->mode;
1806 }  /* cmp_avail_entry */
1807
1808 /**
1809  * Calculate the hash value of an avail entry.
1810  */
1811 static unsigned hash_cache_entry(const avail_entry_t *entry)
1812 {
1813         return get_irn_idx(entry->ptr) * 9 + HASH_PTR(entry->mode);
1814 }  /* hash_cache_entry */
1815
1816 /**
1817  * Move loops out of loops if possible.
1818  *
1819  * @param pscc   the loop described by an SCC
1820  * @param env    the loop environment
1821  */
1822 static void move_loads_out_of_loops(scc *pscc, loop_env *env)
1823 {
1824         ir_node   *phi, *load, *next, *other, *next_other;
1825         ir_entity *ent;
1826         int       j;
1827         phi_entry *phi_list = NULL;
1828         set       *avail;
1829
1830         avail = new_set(cmp_avail_entry, 8);
1831
1832         /* collect all outer memories */
1833         for (phi = pscc->head; phi != NULL; phi = next) {
1834                 node_entry *ne = get_irn_ne(phi, env);
1835                 next = ne->next;
1836
1837                 /* check all memory Phi's */
1838                 if (! is_Phi(phi))
1839                         continue;
1840
1841                 assert(get_irn_mode(phi) == mode_M && "DFS return non-memory Phi");
1842
1843                 for (j = get_irn_arity(phi) - 1; j >= 0; --j) {
1844                         ir_node    *pred = get_irn_n(phi, j);
1845                         node_entry *pe   = get_irn_ne(pred, env);
1846
1847                         if (pe->pscc != ne->pscc) {
1848                                 /* not in the same SCC, is region const */
1849                                 phi_entry *pe = (phi_entry*)phase_alloc(&env->ph, sizeof(*pe));
1850
1851                                 pe->phi  = phi;
1852                                 pe->pos  = j;
1853                                 pe->next = phi_list;
1854                                 phi_list = pe;
1855                         }
1856                 }
1857         }
1858         /* no Phis no fun */
1859         assert(phi_list != NULL && "DFS found a loop without Phi");
1860
1861         /* for now, we cannot handle more than one input (only reducible cf) */
1862         if (phi_list->next != NULL)
1863                 return;
1864
1865         for (load = pscc->head; load; load = next) {
1866                 ir_mode *load_mode;
1867                 node_entry *ne = get_irn_ne(load, env);
1868                 next = ne->next;
1869
1870                 if (is_Load(load)) {
1871                         ldst_info_t *info = (ldst_info_t*)get_irn_link(load);
1872                         ir_node     *ptr = get_Load_ptr(load);
1873
1874                         /* for now, we cannot handle Loads with exceptions */
1875                         if (info->projs[pn_Load_res] == NULL || info->projs[pn_Load_X_regular] != NULL || info->projs[pn_Load_X_except] != NULL)
1876                                 continue;
1877
1878                         /* for now, we can only move Load(Global) */
1879                         if (! is_Global(ptr))
1880                                 continue;
1881                         ent       = get_Global_entity(ptr);
1882                         load_mode = get_Load_mode(load);
1883                         for (other = pscc->head; other != NULL; other = next_other) {
1884                                 node_entry *ne = get_irn_ne(other, env);
1885                                 next_other = ne->next;
1886
1887                                 if (is_Store(other)) {
1888                                         ir_alias_relation rel = get_alias_relation(
1889                                                 get_Store_ptr(other),
1890                                                 get_irn_mode(get_Store_value(other)),
1891                                                 ptr, load_mode);
1892                                         /* if the might be an alias, we cannot pass this Store */
1893                                         if (rel != ir_no_alias)
1894                                                 break;
1895                                 }
1896                                 /* only Phis and pure Calls are allowed here, so ignore them */
1897                         }
1898                         if (other == NULL) {
1899                                 ldst_info_t *ninfo = NULL;
1900                                 phi_entry   *pe;
1901                                 dbg_info    *db;
1902
1903                                 /* yep, no aliasing Store found, Load can be moved */
1904                                 DB((dbg, LEVEL_1, "  Found a Load that could be moved: %+F\n", load));
1905
1906                                 db   = get_irn_dbg_info(load);
1907                                 for (pe = phi_list; pe != NULL; pe = pe->next) {
1908                                         int     pos   = pe->pos;
1909                                         ir_node *phi  = pe->phi;
1910                                         ir_node *blk  = get_nodes_block(phi);
1911                                         ir_node *pred = get_Block_cfgpred_block(blk, pos);
1912                                         ir_node *irn, *mem;
1913                                         avail_entry_t entry, *res;
1914
1915                                         entry.ptr  = ptr;
1916                                         entry.mode = load_mode;
1917                                         res = (avail_entry_t*)set_find(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1918                                         if (res != NULL) {
1919                                                 irn = res->load;
1920                                         } else {
1921                                                 irn = new_rd_Load(db, pred, get_Phi_pred(phi, pos), ptr, load_mode, cons_none);
1922                                                 entry.load = irn;
1923                                                 set_insert(avail, &entry, sizeof(entry), hash_cache_entry(&entry));
1924                                                 DB((dbg, LEVEL_1, "  Created %+F in %+F\n", irn, pred));
1925                                         }
1926                                         pe->load = irn;
1927                                         ninfo = get_ldst_info(irn, phase_obst(&env->ph));
1928
1929                                         ninfo->projs[pn_Load_M] = mem = new_r_Proj(irn, mode_M, pn_Load_M);
1930                                         set_Phi_pred(phi, pos, mem);
1931
1932                                         ninfo->projs[pn_Load_res] = new_r_Proj(irn, load_mode, pn_Load_res);
1933                                 }
1934
1935                                 /* now kill the old Load */
1936                                 exchange(info->projs[pn_Load_M], get_Load_mem(load));
1937                                 exchange(info->projs[pn_Load_res], ninfo->projs[pn_Load_res]);
1938
1939                                 env->changes |= DF_CHANGED;
1940                         }
1941                 }
1942         }
1943         del_set(avail);
1944 }  /* move_loads_out_of_loops */
1945
1946 /**
1947  * Process a loop SCC.
1948  *
1949  * @param pscc  the SCC
1950  * @param env   the loop environment
1951  */
1952 static void process_loop(scc *pscc, loop_env *env)
1953 {
1954         ir_node *irn, *next, *header = NULL;
1955         node_entry *b, *h = NULL;
1956         int j, only_phi, num_outside, process = 0;
1957         ir_node *out_rc;
1958
1959         /* find the header block for this scc */
1960         for (irn = pscc->head; irn; irn = next) {
1961                 node_entry *e = get_irn_ne(irn, env);
1962                 ir_node *block = get_nodes_block(irn);
1963
1964                 next = e->next;
1965                 b = get_irn_ne(block, env);
1966
1967                 if (header != NULL) {
1968                         if (h->POnum < b->POnum) {
1969                                 header = block;
1970                                 h      = b;
1971                         }
1972                 } else {
1973                         header = block;
1974                         h      = b;
1975                 }
1976         }
1977
1978         /* check if this scc contains only Phi, Loads or Stores nodes */
1979         only_phi    = 1;
1980         num_outside = 0;
1981         out_rc      = NULL;
1982         for (irn = pscc->head; irn; irn = next) {
1983                 node_entry *e = get_irn_ne(irn, env);
1984
1985                 next = e->next;
1986                 switch (get_irn_opcode(irn)) {
1987                 case iro_Call:
1988                         if (is_Call_pure(irn)) {
1989                                 /* pure calls can be treated like loads */
1990                                 only_phi = 0;
1991                                 break;
1992                         }
1993                         /* non-pure calls must be handle like may-alias Stores */
1994                         goto fail;
1995                 case iro_CopyB:
1996                         /* cannot handle CopyB yet */
1997                         goto fail;
1998                 case iro_Load:
1999                         process = 1;
2000                         if (get_Load_volatility(irn) == volatility_is_volatile) {
2001                                 /* cannot handle loops with volatile Loads */
2002                                 goto fail;
2003                         }
2004                         only_phi = 0;
2005                         break;
2006                 case iro_Store:
2007                         if (get_Store_volatility(irn) == volatility_is_volatile) {
2008                                 /* cannot handle loops with volatile Stores */
2009                                 goto fail;
2010                         }
2011                         only_phi = 0;
2012                         break;
2013                 default:
2014                         only_phi = 0;
2015                         break;
2016                 case iro_Phi:
2017                         for (j = get_irn_arity(irn) - 1; j >= 0; --j) {
2018                                 ir_node *pred  = get_irn_n(irn, j);
2019                                 node_entry *pe = get_irn_ne(pred, env);
2020
2021                                 if (pe->pscc != e->pscc) {
2022                                         /* not in the same SCC, must be a region const */
2023                                         if (! is_rc(pred, header)) {
2024                                                 /* not a memory loop */
2025                                                 goto fail;
2026                                         }
2027                                         if (out_rc == NULL) {
2028                                                 /* first region constant */
2029                                                 out_rc = pred;
2030                                                 ++num_outside;
2031                                         } else if (out_rc != pred) {
2032                                                 /* another region constant */
2033                                                 ++num_outside;
2034                                         }
2035                                 }
2036                         }
2037                         break;
2038                 }
2039         }
2040         if (! process)
2041                 goto fail;
2042
2043         /* found a memory loop */
2044         DB((dbg, LEVEL_2, "  Found a memory loop:\n  "));
2045         if (only_phi && num_outside == 1) {
2046                 /* a phi cycle with only one real predecessor can be collapsed */
2047                 DB((dbg, LEVEL_2, "  Found an USELESS Phi cycle:\n  "));
2048
2049                 for (irn = pscc->head; irn; irn = next) {
2050                         node_entry *e = get_irn_ne(irn, env);
2051                         next = e->next;
2052                         exchange(irn, out_rc);
2053                 }
2054                 env->changes |= DF_CHANGED;
2055                 return;
2056         }
2057
2058 #ifdef DEBUG_libfirm
2059         for (irn = pscc->head; irn; irn = next) {
2060                 node_entry *e = get_irn_ne(irn, env);
2061                 next = e->next;
2062                 DB((dbg, LEVEL_2, " %+F,", irn));
2063         }
2064         DB((dbg, LEVEL_2, "\n"));
2065 #endif
2066         move_loads_out_of_loops(pscc, env);
2067
2068 fail:
2069         ;
2070 }  /* process_loop */
2071
2072 /**
2073  * Process a SCC.
2074  *
2075  * @param pscc  the SCC
2076  * @param env   the loop environment
2077  */
2078 static void process_scc(scc *pscc, loop_env *env)
2079 {
2080         ir_node *head = pscc->head;
2081         node_entry *e = get_irn_ne(head, env);
2082
2083 #ifdef DEBUG_libfirm
2084         {
2085                 ir_node *irn, *next;
2086
2087                 DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc));
2088                 for (irn = pscc->head; irn; irn = next) {
2089                         node_entry *e = get_irn_ne(irn, env);
2090
2091                         next = e->next;
2092
2093                         DB((dbg, LEVEL_4, " %+F,", irn));
2094                 }
2095                 DB((dbg, LEVEL_4, "\n"));
2096         }
2097 #endif
2098
2099         if (e->next != NULL) {
2100                 /* this SCC has more than one member */
2101                 process_loop(pscc, env);
2102         }
2103 }  /* process_scc */
2104
2105 /**
2106  * Do Tarjan's SCC algorithm and drive load/store optimization.
2107  *
2108  * @param irn  start at this node
2109  * @param env  the loop environment
2110  */
2111 static void dfs(ir_node *irn, loop_env *env)
2112 {
2113         int i, n;
2114         node_entry *node = get_irn_ne(irn, env);
2115
2116         mark_irn_visited(irn);
2117
2118         node->DFSnum = env->nextDFSnum++;
2119         node->low    = node->DFSnum;
2120         push(env, irn);
2121
2122         /* handle preds */
2123         if (is_Phi(irn) || is_Sync(irn)) {
2124                 n = get_irn_arity(irn);
2125                 for (i = 0; i < n; ++i) {
2126                         ir_node *pred = get_irn_n(irn, i);
2127                         node_entry *o = get_irn_ne(pred, env);
2128
2129                         if (!irn_visited(pred)) {
2130                                 dfs(pred, env);
2131                                 node->low = MIN(node->low, o->low);
2132                         }
2133                         if (o->DFSnum < node->DFSnum && o->in_stack)
2134                                 node->low = MIN(o->DFSnum, node->low);
2135                 }
2136         } else if (is_fragile_op(irn)) {
2137                 ir_node *pred = get_fragile_op_mem(irn);
2138                 node_entry *o = get_irn_ne(pred, env);
2139
2140                 if (!irn_visited(pred)) {
2141                         dfs(pred, env);
2142                         node->low = MIN(node->low, o->low);
2143                 }
2144                 if (o->DFSnum < node->DFSnum && o->in_stack)
2145                         node->low = MIN(o->DFSnum, node->low);
2146         } else if (is_Proj(irn)) {
2147                 ir_node *pred = get_Proj_pred(irn);
2148                 node_entry *o = get_irn_ne(pred, env);
2149
2150                 if (!irn_visited(pred)) {
2151                         dfs(pred, env);
2152                         node->low = MIN(node->low, o->low);
2153                 }
2154                 if (o->DFSnum < node->DFSnum && o->in_stack)
2155                         node->low = MIN(o->DFSnum, node->low);
2156         }
2157         else {
2158                  /* IGNORE predecessors */
2159         }
2160
2161         if (node->low == node->DFSnum) {
2162                 scc *pscc = (scc*)phase_alloc(&env->ph, sizeof(*pscc));
2163                 ir_node *x;
2164
2165                 pscc->head = NULL;
2166                 do {
2167                         node_entry *e;
2168
2169                         x = pop(env);
2170                         e = get_irn_ne(x, env);
2171                         e->pscc    = pscc;
2172                         e->next    = pscc->head;
2173                         pscc->head = x;
2174                 } while (x != irn);
2175
2176                 process_scc(pscc, env);
2177         }
2178 }  /* dfs */
2179
2180 /**
2181  * Do the DFS on the memory edges a graph.
2182  *
2183  * @param irg  the graph to process
2184  * @param env  the loop environment
2185  */
2186 static void do_dfs(ir_graph *irg, loop_env *env)
2187 {
2188         ir_node  *endblk, *end;
2189         int      i;
2190
2191         inc_irg_visited(irg);
2192
2193         /* visit all memory nodes */
2194         endblk = get_irg_end_block(irg);
2195         for (i = get_Block_n_cfgpreds(endblk) - 1; i >= 0; --i) {
2196                 ir_node *pred = get_Block_cfgpred(endblk, i);
2197
2198                 pred = skip_Proj(pred);
2199                 if (is_Return(pred))
2200                         dfs(get_Return_mem(pred), env);
2201                 else if (is_Raise(pred))
2202                         dfs(get_Raise_mem(pred), env);
2203                 else if (is_fragile_op(pred))
2204                         dfs(get_fragile_op_mem(pred), env);
2205                 else if (is_Bad(pred))
2206                         /* ignore non-optimized block predecessor */;
2207                 else {
2208                         assert(0 && "Unknown EndBlock predecessor");
2209                 }
2210         }
2211
2212         /* visit the keep-alives */
2213         end = get_irg_end(irg);
2214         for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
2215                 ir_node *ka = get_End_keepalive(end, i);
2216
2217                 if (is_Phi(ka) && !irn_visited(ka))
2218                         dfs(ka, env);
2219         }
2220 }  /* do_dfs */
2221
2222 /**
2223  * Optimize Loads/Stores in loops.
2224  *
2225  * @param irg  the graph
2226  */
2227 static int optimize_loops(ir_graph *irg)
2228 {
2229         loop_env env;
2230
2231         env.stack         = NEW_ARR_F(ir_node *, 128);
2232         env.tos           = 0;
2233         env.nextDFSnum    = 0;
2234         env.POnum         = 0;
2235         env.changes       = 0;
2236         phase_init(&env.ph, irg, phase_irn_init_default);
2237
2238         /* calculate the SCC's and drive loop optimization. */
2239         do_dfs(irg, &env);
2240
2241         DEL_ARR_F(env.stack);
2242         phase_deinit(&env.ph);
2243
2244         return env.changes;
2245 }  /* optimize_loops */
2246
2247 /*
2248  * do the load store optimization
2249  */
2250 int optimize_load_store(ir_graph *irg)
2251 {
2252         walk_env_t env;
2253
2254         FIRM_DBG_REGISTER(dbg, "firm.opt.ldstopt");
2255
2256         assert(get_irg_phase_state(irg) != phase_building);
2257         assert(get_irg_pinned(irg) != op_pin_state_floats &&
2258                 "LoadStore optimization needs pinned graph");
2259
2260         /* we need landing pads */
2261         remove_critical_cf_edges(irg);
2262
2263         edges_assure(irg);
2264
2265         /* for Phi optimization post-dominators are needed ... */
2266         assure_postdoms(irg);
2267
2268         if (get_opt_alias_analysis()) {
2269                 assure_irg_entity_usage_computed(irg);
2270                 assure_irp_globals_entity_usage_computed();
2271         }
2272
2273         obstack_init(&env.obst);
2274         env.changes = 0;
2275
2276         /* init the links, then collect Loads/Stores/Proj's in lists */
2277         master_visited = 0;
2278         irg_walk_graph(irg, firm_clear_link, collect_nodes, &env);
2279
2280         /* now we have collected enough information, optimize */
2281         irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
2282
2283         env.changes |= optimize_loops(irg);
2284
2285         obstack_free(&env.obst, NULL);
2286
2287         /* Handle graph state */
2288         if (env.changes) {
2289                 set_irg_outs_inconsistent(irg);
2290                 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
2291         }
2292
2293         if (env.changes & CF_CHANGED) {
2294                 /* is this really needed: Yes, control flow changed, block might
2295                 have Bad() predecessors. */
2296                 set_irg_doms_inconsistent(irg);
2297         }
2298         return env.changes != 0;
2299 }  /* optimize_load_store */
2300
2301 ir_graph_pass_t *optimize_load_store_pass(const char *name)
2302 {
2303         return def_graph_pass_ret(name ? name : "ldst", optimize_load_store);
2304 }  /* optimize_load_store_pass */