nsz Git - libfirm/blob - ir/opt/opt_ldst.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   Dataflow driven Load/Store optimizations, uses some ideas from
  23  *          VanDrunen's LEPRE
  24  * @author  Michael Beck
  25  */
  26 #include "config.h"
  27
  28 #include "irnode_t.h"
  29 #include "irflag_t.h"
  30 #include "array_t.h"
  31 #include "ircons.h"
  32 #include "irdom.h"
  33 #include "irgmod.h"
  34 #include "irgwalk.h"
  35 #include "irouts.h"
  36 #include "irgraph.h"
  37 #include "irgopt.h"
  38 #include "iropt.h"
  39 #include "iroptimize.h"
  40 #include "irnodehashmap.h"
  41 #include "raw_bitset.h"
  42 #include "debug.h"
  43 #include "error.h"
  44 #include "irpass.h"
  45
  46 /* maximum number of output Proj's */
  47 #define MAX_PROJ ((long)pn_Load_max > (long)pn_Store_max ? (long)pn_Load_max : (long)pn_Store_max)
  48
  49 /**
  50  * Mapping an address to an dense ID.
  51  */
  52 typedef struct address_entry_t {
  53         unsigned id;          /**< The ID */
  54 } address_entry;
  55
  56 /**
  57  * Memop-flags.
  58  */
  59 enum memop_flags {
  60         FLAG_KILL_ALL    = 1, /**< KILL all addresses */
  61         FLAG_KILLED_NODE = 2, /**< this node was killed */
  62         FLAG_EXCEPTION   = 4, /**< this node has exception flow */
  63         FLAG_IGNORE      = 8, /**< ignore this node (volatile or other) */
  64 };
  65
  66 /**
  67  * A value: This represents a value stored at a given address in
  68  * memory. Do not confuse with values from value numbering.
  69  */
  70 typedef struct value_t value_t;
  71 struct value_t {
  72         ir_node  *address;    /**< the address of this value */
  73         ir_node  *value;      /**< the value itself */
  74         ir_mode  *mode;       /**< the mode of the value */
  75         unsigned id;          /**< address id */
  76 };
  77
  78 /**
  79  * A memop describes an memory-related operation.
  80  * These are Loads/Store and all other ops that might modify
  81  * memory (Calls, CopyB) or causing exceptions.
  82  */
  83 typedef struct memop_t memop_t;
  84 struct memop_t {
  85         value_t  value;      /**< the value of this memop: only defined for Load/Store */
  86         ir_node  *node;      /**< the memory op itself */
  87         ir_node  *mem;       /**< the memory FROM this node */
  88         ir_node  *replace;   /**< the replacement node if this memop is replaced */
  89         memop_t  *next;      /**< links to the next memory op in the block in forward order. */
  90         memop_t  *prev;      /**< links to the previous memory op in the block in forward order. */
  91         unsigned flags;      /**< memop flags */
  92         ir_node  *projs[MAX_PROJ+1]; /**< Projs of this memory op */
  93 };
  94
  95 /**
  96  * Additional data for every basic block.
  97  */
  98 typedef struct block_t block_t;
  99 struct block_t {
 100         memop_t  *memop_forward;     /**< topologically sorted list of memory ops in this block */
 101         memop_t  *memop_backward;    /**< last memop in the list */
 102         unsigned *avail_out;         /**< out-set of available addresses */
 103         memop_t  **id_2_memop_avail; /**< maps avail address ids to memops */
 104         unsigned *anticL_in;         /**< in-set of anticipated Load addresses */
 105         memop_t  **id_2_memop_antic; /**< maps anticipated address ids to memops */
 106         ir_node  *block;             /**< the associated block */
 107         block_t  *forward_next;      /**< next block entry for forward iteration */
 108         block_t  *backward_next;     /**< next block entry for backward iteration */
 109         memop_t  *avail;             /**< used locally for the avail map */
 110         memop_t  **trans_results;    /**< used to cached translated nodes due antic calculation. */
 111 };
 112
 113 /**
 114  * Metadata for this pass.
 115  */
 116 typedef struct ldst_env_t {
 117         struct obstack   obst;             /**< obstack for temporary data */
 118         ir_nodehashmap_t adr_map;          /**< Map addresses to */
 119         block_t         *forward;          /**< Inverse post-order list of all blocks Start->End */
 120         block_t         *backward;         /**< Inverse post-order list of all blocks End->Start */
 121         ir_node         *start_bl;         /**< start block of the current graph */
 122         ir_node         *end_bl;           /**< end block of the current graph */
 123         unsigned        *curr_set;         /**< current set of addresses */
 124         memop_t         **curr_id_2_memop; /**< current map of address ids to memops */
 125         unsigned        curr_adr_id;       /**< number for address mapping */
 126         unsigned        n_mem_ops;         /**< number of memory operations (Loads/Stores) */
 127         size_t          rbs_size;          /**< size of all bitsets in bytes */
 128         int             max_cfg_preds;     /**< maximum number of block cfg predecessors */
 129         int             changed;           /**< Flags for changed graph state */
 130 #ifdef DEBUG_libfirm
 131         ir_node         **id_2_address;    /**< maps an id to the used address */
 132 #endif
 133 } ldst_env;
 134
 135 /* the one and only environment */
 136 static ldst_env env;
 137
 138 #ifdef DEBUG_libfirm
 139
 140 static firm_dbg_module_t *dbg;
 141
 142 /**
 143  * Dumps the block list.
 144  *
 145  * @param ldst environment
 146  */
 147 static void dump_block_list(ldst_env *env)
 148 {
 149         block_t *entry;
 150         memop_t *op;
 151         int     i;
 152
 153         for (entry = env->forward; entry != NULL; entry = entry->forward_next) {
 154                 DB((dbg, LEVEL_2, "%+F {", entry->block));
 155
 156                 i = 0;
 157                 for (op = entry->memop_forward; op != NULL; op = op->next) {
 158                         if (i == 0) {
 159                                 DB((dbg, LEVEL_2, "\n\t"));
 160                         }
 161                         DB((dbg, LEVEL_2, "%+F", op->node));
 162                         if ((op->flags & FLAG_KILL_ALL) == FLAG_KILL_ALL)
 163                                 DB((dbg, LEVEL_2, "X"));
 164                         else if (op->flags & FLAG_KILL_ALL)
 165                                 DB((dbg, LEVEL_2, "K"));
 166                         DB((dbg, LEVEL_2, ", "));
 167
 168                         i = (i + 1) & 3;
 169                 }
 170                 DB((dbg, LEVEL_2, "\n}\n\n"));
 171         }
 172 }
 173
 174 /**
 175  * Dumps the current set.
 176  *
 177  * @param bl   current block
 178  * @param s    name of the set
 179  */
 180 static void dump_curr(block_t *bl, const char *s)
 181 {
 182         size_t end = env.rbs_size - 1;
 183         size_t pos;
 184         int    i;
 185
 186         DB((dbg, LEVEL_2, "%s[%+F] = {", s, bl->block));
 187         i = 0;
 188         for (pos = rbitset_next(env.curr_set, 0, 1); pos < end; pos = rbitset_next(env.curr_set, pos + 1, 1)) {
 189                 memop_t *op = env.curr_id_2_memop[pos];
 190
 191                 if (i == 0) {
 192                         DB((dbg, LEVEL_2, "\n\t"));
 193                 }
 194
 195                 DB((dbg, LEVEL_2, "<%+F, %+F>, ", op->value.address, op->value.value));
 196                 i = (i + 1) & 3;
 197         }
 198         DB((dbg, LEVEL_2, "\n}\n"));
 199 }
 200
 201 #else
 202 static void dump_block_list(ldst_env *env)
 203 {
 204         (void) env;
 205 }
 206 static void dump_curr(block_t *bl, const char *s)
 207 {
 208         (void) bl;
 209         (void) s;
 210 }
 211 #endif /* DEBUG_libfirm */
 212
 213 /** Get the block entry for a block node */
 214 static block_t *get_block_entry(const ir_node *block)
 215 {
 216         assert(is_Block(block));
 217
 218         return (block_t*)get_irn_link(block);
 219 }
 220
 221 /** Get the memop entry for a memory operation node */
 222 static memop_t *get_irn_memop(const ir_node *irn)
 223 {
 224         assert(! is_Block(irn));
 225         return (memop_t*)get_irn_link(irn);
 226 }
 227
 228 /**
 229  * Walk over the memory edges from definition to users.
 230  * This ensures, that even operation without memory output are found.
 231  *
 232  * @param irn   start node
 233  * @param pre   pre walker function
 234  * @param post  post walker function
 235  * @param ctx   context parameter for the walker functions
 236  */
 237 static void walk_memory(ir_node *irn, irg_walk_func *pre, irg_walk_func *post, void *ctx)
 238 {
 239         ir_mode *mode;
 240
 241         mark_irn_visited(irn);
 242
 243         if (pre)
 244                 pre(irn, ctx);
 245
 246         mode = get_irn_mode(irn);
 247         if (mode == mode_M) {
 248                 /* every successor uses memory */
 249                 for (unsigned i = get_irn_n_outs(irn); i-- > 0; ) {
 250                         ir_node *succ = get_irn_out(irn, i);
 251
 252                         if (! irn_visited(succ))
 253                                 walk_memory(succ, pre, post, ctx);
 254                 }
 255         } else if (mode == mode_T) {
 256                 /* only some Proj's uses memory */
 257                 for (unsigned i = get_irn_n_outs(irn); i-- > 0; ) {
 258                         ir_node *proj = get_irn_out(irn, i);
 259
 260                         if (get_irn_mode(proj) == mode_M && ! irn_visited(proj))
 261                                 walk_memory(proj, pre, post, ctx);
 262                 }
 263         }
 264         if (post)
 265                 post(irn, ctx);
 266 }
 267
 268 /**
 269  * Walks over all memory nodes of a graph.
 270  *
 271  * @param irg   a graph
 272  * @param pre   pre walker function
 273  * @param post  post walker function
 274  * @param ctx   context parameter for the walker functions
 275  */
 276 static void walk_memory_irg(ir_graph *irg, irg_walk_func pre, irg_walk_func post, void *ctx)
 277 {
 278         inc_irg_visited(irg);
 279
 280         ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
 281
 282         /*
 283          * there are two possible sources for memory: initial_mem and nomem
 284          * we ignore nomem as this should NOT change the memory
 285          */
 286         walk_memory(get_irg_initial_mem(irg), pre, post, ctx);
 287
 288         ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
 289 }
 290
 291 /**
 292  * Register an address and allocate a (sparse, 0..n) ID for it.
 293  *
 294  * @param adr  the IR-node representing the address
 295  *
 296  * @return the allocated id
 297  */
 298 static unsigned register_address(ir_node *adr)
 299 {
 300         address_entry *entry;
 301
 302         /* skip Confirms and Casts */
 303 restart:
 304         if (is_Confirm(adr)) {
 305                 adr = get_Confirm_value(adr);
 306                 goto restart;
 307         }
 308         if (is_Cast(adr)) {
 309                 adr = get_Cast_op(adr);
 310                 goto restart;
 311         }
 312
 313         entry = ir_nodehashmap_get(address_entry, &env.adr_map, adr);
 314
 315         if (entry == NULL) {
 316                 /* new address */
 317                 entry = OALLOC(&env.obst, address_entry);
 318
 319                 entry->id = env.curr_adr_id++;
 320                 ir_nodehashmap_insert(&env.adr_map, adr, entry);
 321
 322                 DB((dbg, LEVEL_3, "ADDRESS %+F has ID %u\n", adr, entry->id));
 323 #ifdef DEBUG_libfirm
 324                 ARR_APP1(ir_node *, env.id_2_address, adr);
 325 #endif
 326         }
 327         return entry->id;
 328 }
 329
 330
 331 /**
 332  * translate an address through a Phi node into a given predecessor
 333  * block.
 334  *
 335  * @param address  the address
 336  * @param block    the block
 337  * @param pos      the position of the predecessor in block
 338  */
 339 static ir_node *phi_translate(ir_node *address, const ir_node *block, int pos)
 340 {
 341         if (is_Phi(address) && get_nodes_block(address) == block)
 342                 address = get_Phi_pred(address, pos);
 343         return address;
 344 }
 345
 346 /**
 347  * Walker: allocate an block entry for every block
 348  * and register all potential addresses.
 349  */
 350 static void prepare_blocks(ir_node *irn, void *ctx)
 351 {
 352         (void)ctx;
 353
 354         if (is_Block(irn)) {
 355                 block_t *entry = OALLOC(&env.obst, block_t);
 356                 int     n;
 357
 358                 entry->memop_forward    = NULL;
 359                 entry->memop_backward   = NULL;
 360                 entry->avail_out        = NULL;
 361                 entry->id_2_memop_avail = NULL;
 362                 entry->anticL_in        = NULL;
 363                 entry->id_2_memop_antic = NULL;
 364                 entry->block            = irn;
 365                 entry->forward_next     = NULL;
 366                 entry->backward_next    = NULL;
 367                 entry->avail            = NULL;
 368                 entry->trans_results    = NULL;
 369                 set_irn_link(irn, entry);
 370
 371                 set_Block_phis(irn, NULL);
 372
 373                 /* use block marks to track unreachable blocks */
 374                 set_Block_mark(irn, 0);
 375
 376                 n = get_Block_n_cfgpreds(irn);
 377                 if (n > env.max_cfg_preds)
 378                         env.max_cfg_preds = n;
 379         } else {
 380                 ir_mode *mode = get_irn_mode(irn);
 381
 382                 if (mode_is_reference(mode)) {
 383                         /*
 384                          * Register ALL possible addresses: this is overkill yet but
 385                          * simpler then doing it for all possible translated addresses
 386                          * (which would be sufficient in the moment.
 387                          */
 388                         (void)register_address(irn);
 389                 }
 390         }
 391 }
 392
 393 /**
 394  * Post-Walker, link in all Phi's
 395  */
 396 static void link_phis(ir_node *irn, void *ctx)
 397 {
 398         (void)ctx;
 399
 400         if (is_Phi(irn)) {
 401                 ir_node *block = get_nodes_block(irn);
 402                 add_Block_phi(block, irn);
 403         }
 404 }
 405
 406 /**
 407  * Block walker: creates the inverse post-order list for the CFG.
 408  */
 409 static void inverse_post_order(ir_node *block, void *ctx)
 410 {
 411         block_t *entry = get_block_entry(block);
 412
 413         (void)ctx;
 414
 415         /* mark this block IS reachable from start */
 416         set_Block_mark(block, 1);
 417
 418         /* create the list in inverse order */
 419         entry->forward_next = env.forward;
 420         env.forward         = entry;
 421
 422         /* remember the first visited (last in list) entry, needed for later */
 423         if (env.backward == NULL)
 424                 env.backward = entry;
 425 }
 426
 427 /**
 428  * Block walker: create backward links for the memops of a block.
 429  */
 430 static void collect_backward(ir_node *block, void *ctx)
 431 {
 432         block_t *entry = get_block_entry(block);
 433         memop_t *last, *op;
 434
 435         (void)ctx;
 436
 437         /*
 438          * Do NOT link in the end block yet. We want it to be
 439          * the first in the list. This is NOT guaranteed by the walker
 440          * if we have endless loops.
 441          */
 442         if (block != env.end_bl) {
 443                 entry->backward_next = env.backward;
 444
 445                 /* create the list in inverse order */
 446                 env.backward = entry;
 447         }
 448
 449         /* create backward links for all memory ops */
 450         last = NULL;
 451         for (op = entry->memop_forward; op != NULL; op = op->next) {
 452                 op->prev = last;
 453                 last     = op;
 454         }
 455         entry->memop_backward = last;
 456 }
 457
 458 /**
 459  * Allocate a memop.
 460  *
 461  * @param irn  the IR-node representing the memop or NULL
 462  *             if this is a translated (virtual) memop
 463  *
 464  * @return the allocated memop
 465  */
 466 static memop_t *alloc_memop(ir_node *irn)
 467 {
 468         memop_t *m = OALLOC(&env.obst, memop_t);
 469
 470         m->value.address = NULL;
 471         m->value.value   = NULL;
 472         m->value.mode    = NULL;
 473
 474         m->node          = irn;
 475         m->mem           = NULL;
 476         m->replace       = NULL;
 477         m->next          = NULL;
 478         m->flags         = 0;
 479
 480         memset(m->projs, 0, sizeof(m->projs));
 481
 482         if (irn != NULL)
 483                 set_irn_link(irn, m);
 484         return m;
 485 }
 486
 487 /**
 488  * Create a memop for a Phi-replacement.
 489  *
 490  * @param op   the memop to clone
 491  * @param phi  the Phi-node representing the new value
 492  */
 493 static memop_t *clone_memop_phi(memop_t *op, ir_node *phi)
 494 {
 495         memop_t *m = OALLOC(&env.obst, memop_t);
 496
 497         m->value         = op->value;
 498         m->value.value   = phi;
 499
 500         m->node          = phi;
 501         m->replace       = NULL;
 502         m->next          = NULL;
 503         m->flags         = 0;
 504
 505         set_irn_link(phi, m);
 506         return m;
 507 }
 508
 509 /**
 510  * Return the memory properties of a call node.
 511  *
 512  * @param call  the call node
 513  *
 514  * return a bitset of mtp_property_const and mtp_property_pure
 515  */
 516 static unsigned get_Call_memory_properties(ir_node *call)
 517 {
 518         ir_type *call_tp = get_Call_type(call);
 519         unsigned prop = get_method_additional_properties(call_tp);
 520
 521         /* check first the call type */
 522         if ((prop & (mtp_property_const|mtp_property_pure)) == 0) {
 523                 /* try the called entity */
 524                 ir_node *ptr = get_Call_ptr(call);
 525
 526                 if (is_SymConst_addr_ent(ptr)) {
 527                         ir_entity *ent = get_SymConst_entity(ptr);
 528
 529                         prop = get_entity_additional_properties(ent);
 530                 }
 531         }
 532         return prop & (mtp_property_const|mtp_property_pure);
 533 }
 534
 535 /**
 536  * Returns an entity if the address ptr points to a constant one.
 537  *
 538  * @param ptr  the address
 539  *
 540  * @return an entity or NULL
 541  */
 542 static ir_entity *find_constant_entity(ir_node *ptr)
 543 {
 544         for (;;) {
 545                 if (is_SymConst(ptr) && get_SymConst_kind(ptr) == symconst_addr_ent) {
 546                         return get_SymConst_entity(ptr);
 547                 } else if (is_Sel(ptr)) {
 548                         ir_entity *ent = get_Sel_entity(ptr);
 549                         ir_type   *tp  = get_entity_owner(ent);
 550
 551                         /* Do not fiddle with polymorphism. */
 552                         if (is_Class_type(get_entity_owner(ent)) &&
 553                                 ((get_entity_n_overwrites(ent)    != 0) ||
 554                                 (get_entity_n_overwrittenby(ent) != 0)   ) )
 555                                 return NULL;
 556
 557                         if (is_Array_type(tp)) {
 558                                 /* check bounds */
 559                                 int i, n;
 560
 561                                 for (i = 0, n = get_Sel_n_indexs(ptr); i < n; ++i) {
 562                                         ir_node   *bound;
 563                                         ir_tarval *tlower, *tupper;
 564                                         ir_node   *index = get_Sel_index(ptr, i);
 565                                         ir_tarval *tv     = computed_value(index);
 566
 567                                         /* check if the index is constant */
 568                                         if (tv == tarval_bad)
 569                                                 return NULL;
 570
 571                                         bound  = get_array_lower_bound(tp, i);
 572                                         tlower = computed_value(bound);
 573                                         bound  = get_array_upper_bound(tp, i);
 574                                         tupper = computed_value(bound);
 575
 576                                         if (tlower == tarval_bad || tupper == tarval_bad)
 577                                                 return NULL;
 578
 579                                         if (tarval_cmp(tv, tlower) == ir_relation_less)
 580                                                 return NULL;
 581                                         if (tarval_cmp(tupper, tv) == ir_relation_less)
 582                                                 return NULL;
 583
 584                                         /* ok, bounds check finished */
 585                                 }
 586                         }
 587
 588                         if (get_entity_linkage(ent) == IR_LINKAGE_CONSTANT)
 589                                 return ent;
 590
 591                         /* try next */
 592                         ptr = get_Sel_ptr(ptr);
 593                 } else if (is_Add(ptr)) {
 594                         ir_node *l = get_Add_left(ptr);
 595                         ir_node *r = get_Add_right(ptr);
 596
 597                         if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
 598                                 ptr = l;
 599                         else if (get_irn_mode(r) == get_irn_mode(ptr) && is_Const(l))
 600                                 ptr = r;
 601                         else
 602                                 return NULL;
 603
 604                         /* for now, we support only one addition, reassoc should fold all others */
 605                         if (! is_SymConst(ptr) && !is_Sel(ptr))
 606                                 return NULL;
 607                 } else if (is_Sub(ptr)) {
 608                         ir_node *l = get_Sub_left(ptr);
 609                         ir_node *r = get_Sub_right(ptr);
 610
 611                         if (get_irn_mode(l) == get_irn_mode(ptr) && is_Const(r))
 612                                 ptr = l;
 613                         else
 614                                 return NULL;
 615                         /* for now, we support only one subtraction, reassoc should fold all others */
 616                         if (! is_SymConst(ptr) && !is_Sel(ptr))
 617                                 return NULL;
 618                 } else
 619                         return NULL;
 620         }
 621 }
 622
 623 /**
 624  * Return the Selection index of a Sel node from dimension n
 625  */
 626 static long get_Sel_array_index_long(ir_node *n, int dim)
 627 {
 628         ir_node *index = get_Sel_index(n, dim);
 629         return get_tarval_long(get_Const_tarval(index));
 630 }
 631
 632 typedef struct path_entry {
 633         ir_entity         *ent;
 634         struct path_entry *next;
 635         size_t            index;
 636 } path_entry;
 637
 638 static ir_node *rec_find_compound_ent_value(ir_node *ptr, path_entry *next)
 639 {
 640         path_entry       entry, *p;
 641         ir_entity        *ent, *field;
 642         ir_initializer_t *initializer;
 643         ir_tarval        *tv;
 644         ir_type          *tp;
 645         size_t           n;
 646
 647         entry.next = next;
 648         if (is_SymConst(ptr)) {
 649                 /* found the root */
 650                 ent         = get_SymConst_entity(ptr);
 651                 initializer = get_entity_initializer(ent);
 652                 for (p = next; p != NULL;) {
 653                         if (initializer->kind != IR_INITIALIZER_COMPOUND)
 654                                 return NULL;
 655                         n  = get_initializer_compound_n_entries(initializer);
 656                         tp = get_entity_type(ent);
 657
 658                         if (is_Array_type(tp)) {
 659                                 ent = get_array_element_entity(tp);
 660                                 if (ent != p->ent) {
 661                                         /* a missing [0] */
 662                                         if (0 >= n)
 663                                                 return NULL;
 664                                         initializer = get_initializer_compound_value(initializer, 0);
 665                                         continue;
 666                                 }
 667                         }
 668                         if (p->index >= n)
 669                                 return NULL;
 670                         initializer = get_initializer_compound_value(initializer, p->index);
 671
 672                         ent = p->ent;
 673                         p   = p->next;
 674                 }
 675                 tp = get_entity_type(ent);
 676                 while (is_Array_type(tp)) {
 677                         ent = get_array_element_entity(tp);
 678                         tp = get_entity_type(ent);
 679                         /* a missing [0] */
 680                         n  = get_initializer_compound_n_entries(initializer);
 681                         if (0 >= n)
 682                                 return NULL;
 683                         initializer = get_initializer_compound_value(initializer, 0);
 684                 }
 685
 686                 switch (initializer->kind) {
 687                 case IR_INITIALIZER_CONST:
 688                         return get_initializer_const_value(initializer);
 689                 case IR_INITIALIZER_TARVAL:
 690                 case IR_INITIALIZER_NULL:
 691                 default:
 692                         return NULL;
 693                 }
 694         } else if (is_Sel(ptr)) {
 695                 entry.ent = field = get_Sel_entity(ptr);
 696                 tp = get_entity_owner(field);
 697                 if (is_Array_type(tp)) {
 698                         assert(get_Sel_n_indexs(ptr) == 1 && "multi dim arrays not implemented");
 699                         entry.index = get_Sel_array_index_long(ptr, 0) - get_array_lower_bound_int(tp, 0);
 700                 } else {
 701                         size_t i, n_members = get_compound_n_members(tp);
 702                         for (i = 0; i < n_members; ++i) {
 703                                 if (get_compound_member(tp, i) == field)
 704                                         break;
 705                         }
 706                         if (i >= n_members) {
 707                                 /* not found: should NOT happen */
 708                                 return NULL;
 709                         }
 710                         entry.index = i;
 711                 }
 712                 return rec_find_compound_ent_value(get_Sel_ptr(ptr), &entry);
 713         }  else if (is_Add(ptr)) {
 714                 ir_mode *mode;
 715                 unsigned pos;
 716
 717                 {
 718                         ir_node *l = get_Add_left(ptr);
 719                         ir_node *r = get_Add_right(ptr);
 720                         if (is_Const(r)) {
 721                                 ptr = l;
 722                                 tv  = get_Const_tarval(r);
 723                         } else {
 724                                 ptr = r;
 725                                 tv  = get_Const_tarval(l);
 726                         }
 727                 }
 728 ptr_arith:
 729                 mode = get_tarval_mode(tv);
 730
 731                 /* ptr must be a Sel or a SymConst, this was checked in find_constant_entity() */
 732                 if (is_Sel(ptr)) {
 733                         field = get_Sel_entity(ptr);
 734                 } else {
 735                         field = get_SymConst_entity(ptr);
 736                 }
 737
 738                 /* count needed entries */
 739                 pos = 0;
 740                 for (ent = field;;) {
 741                         tp = get_entity_type(ent);
 742                         if (! is_Array_type(tp))
 743                                 break;
 744                         ent = get_array_element_entity(tp);
 745                         ++pos;
 746                 }
 747                 /* should be at least ONE entry */
 748                 if (pos == 0)
 749                         return NULL;
 750
 751                 /* allocate the right number of entries */
 752                 NEW_ARR_A(path_entry, p, pos);
 753
 754                 /* fill them up */
 755                 pos = 0;
 756                 for (ent = field;;) {
 757                         unsigned  size;
 758                         ir_tarval *sz, *tv_index, *tlower, *tupper;
 759                         size_t    index;
 760                         ir_node   *bound;
 761
 762                         tp = get_entity_type(ent);
 763                         if (! is_Array_type(tp))
 764                                 break;
 765                         ent = get_array_element_entity(tp);
 766                         p[pos].ent  = ent;
 767                         p[pos].next = &p[pos + 1];
 768
 769                         size = get_type_size_bytes(get_entity_type(ent));
 770                         sz   = new_tarval_from_long(size, mode);
 771
 772                         tv_index = tarval_div(tv, sz);
 773                         tv       = tarval_mod(tv, sz);
 774
 775                         if (tv_index == tarval_bad || tv == tarval_bad)
 776                                 return NULL;
 777
 778                         assert(get_array_n_dimensions(tp) == 1 && "multiarrays not implemented");
 779                         bound  = get_array_lower_bound(tp, 0);
 780                         tlower = computed_value(bound);
 781                         bound  = get_array_upper_bound(tp, 0);
 782                         tupper = computed_value(bound);
 783
 784                         if (tlower == tarval_bad || tupper == tarval_bad)
 785                                 return NULL;
 786
 787                         if (tarval_cmp(tv_index, tlower) == ir_relation_less)
 788                                 return NULL;
 789                         if (tarval_cmp(tupper, tv_index) == ir_relation_less)
 790                                 return NULL;
 791
 792                         /* ok, bounds check finished */
 793                         index = get_tarval_long(tv_index);
 794                         p[pos].index = index;
 795                         ++pos;
 796                 }
 797                 if (! tarval_is_null(tv)) {
 798                         /* hmm, wrong access */
 799                         return NULL;
 800                 }
 801                 p[pos - 1].next = next;
 802                 return rec_find_compound_ent_value(ptr, p);
 803         } else if (is_Sub(ptr)) {
 804                 ir_node *l = get_Sub_left(ptr);
 805                 ir_node *r = get_Sub_right(ptr);
 806
 807                 ptr = l;
 808                 tv  = get_Const_tarval(r);
 809                 tv  = tarval_neg(tv);
 810                 goto ptr_arith;
 811         }
 812         return NULL;
 813 }
 814
 815 static ir_node *find_compound_ent_value(ir_node *ptr)
 816 {
 817         return rec_find_compound_ent_value(ptr, NULL);
 818 }
 819
 820 /**
 821  * Mark a Load memop to be replace by a definition
 822  *
 823  * @param op  the Load memop
 824  */
 825 static void mark_replace_load(memop_t *op, ir_node *def)
 826 {
 827         op->replace = def;
 828         op->flags |= FLAG_KILLED_NODE;
 829         env.changed = 1;
 830 }
 831
 832 /**
 833  * Mark a Store memop to be removed.
 834  *
 835  * @param op  the Store memop
 836  */
 837 static void mark_remove_store(memop_t *op)
 838 {
 839         op->flags |= FLAG_KILLED_NODE;
 840         env.changed = 1;
 841 }
 842
 843 /**
 844  * Update a memop for a Load.
 845  *
 846  * @param m  the memop
 847  */
 848 static void update_Load_memop(memop_t *m)
 849 {
 850         ir_node   *load = m->node;
 851         ir_node   *ptr;
 852         ir_entity *ent;
 853
 854         if (get_Load_volatility(load) == volatility_is_volatile)
 855                 m->flags |= FLAG_IGNORE;
 856
 857         ptr = get_Load_ptr(load);
 858
 859         m->value.address = ptr;
 860
 861         for (unsigned i = get_irn_n_outs(load); i-- > 0; ) {
 862                 ir_node *proj = get_irn_out(load, i);
 863                 long    pn;
 864
 865                 /* beware of keep edges */
 866                 if (is_End(proj))
 867                         continue;
 868
 869                 pn = get_Proj_proj(proj);
 870                 m->projs[pn] = proj;
 871                 switch (pn) {
 872                 case pn_Load_res:
 873                         m->value.value = proj;
 874                         m->value.mode  = get_irn_mode(proj);
 875                         break;
 876                 case pn_Load_X_except:
 877                         m->flags |= FLAG_EXCEPTION;
 878                         break;
 879                 case pn_Load_M:
 880                         m->mem = proj;
 881                         break;
 882                 case pn_Load_X_regular:
 883                         break;
 884                 default:
 885                         panic("Unsupported Proj from Load %+F", proj);
 886                 }
 887         }
 888
 889         /* check if we can determine the entity that will be loaded */
 890         ent = find_constant_entity(ptr);
 891
 892         if (ent != NULL && get_entity_visibility(ent) != ir_visibility_external) {
 893                 /* a static allocation that is not external: there should be NO exception
 894                  * when loading even if we cannot replace the load itself. */
 895                 ir_node *value = NULL;
 896
 897                 /* no exception, clear the m fields as it might be checked later again */
 898                 if (m->projs[pn_Load_X_except]) {
 899                         ir_graph *irg = get_irn_irg(ptr);
 900                         exchange(m->projs[pn_Load_X_except], new_r_Bad(irg, mode_X));
 901                         m->projs[pn_Load_X_except] = NULL;
 902                         m->flags &= ~FLAG_EXCEPTION;
 903                         env.changed = 1;
 904                 }
 905                 if (m->projs[pn_Load_X_regular]) {
 906                         exchange(m->projs[pn_Load_X_regular], new_r_Jmp(get_nodes_block(load)));
 907                         m->projs[pn_Load_X_regular] = NULL;
 908                         env.changed = 1;
 909                 }
 910
 911                 if (get_entity_linkage(ent) & IR_LINKAGE_CONSTANT) {
 912                         if (ent->initializer) {
 913                                 /* new style initializer */
 914                                 value = find_compound_ent_value(ptr);
 915                         }
 916                         if (value != NULL)
 917                                 value = can_replace_load_by_const(load, value);
 918                 }
 919
 920                 if (value != NULL) {
 921                         /* we completely replace the load by this value */
 922                         DB((dbg, LEVEL_1, "Replacing Load %+F by constant %+F\n", m->node, value));
 923                         mark_replace_load(m, value);
 924                         return;
 925                 }
 926         }
 927
 928         if (m->value.value != NULL && !(m->flags & FLAG_IGNORE)) {
 929                 /* only create an address if this node is NOT killed immediately or ignored */
 930                 m->value.id = register_address(ptr);
 931                 ++env.n_mem_ops;
 932         } else {
 933                 /* no user, KILL it */
 934                 mark_replace_load(m, NULL);
 935         }
 936 }
 937
 938 /**
 939  * Update a memop for a Store.
 940  *
 941  * @param m  the memop
 942  */
 943 static void update_Store_memop(memop_t *m)
 944 {
 945         ir_node *store = m->node;
 946         ir_node *adr   = get_Store_ptr(store);
 947
 948         if (get_Store_volatility(store) == volatility_is_volatile) {
 949                 m->flags |= FLAG_IGNORE;
 950         } else {
 951                 /* only create an address if this node is NOT ignored */
 952                 m->value.id = register_address(adr);
 953                 ++env.n_mem_ops;
 954         }
 955
 956         m->value.address = adr;
 957
 958         for (unsigned i = get_irn_n_outs(store); i-- > 0; ) {
 959                 ir_node *proj = get_irn_out(store, i);
 960                 long    pn;
 961
 962                 /* beware of keep edges */
 963                 if (is_End(proj))
 964                         continue;
 965
 966                 pn = get_Proj_proj(proj);
 967                 m->projs[pn] = proj;
 968                 switch (pn) {
 969                 case pn_Store_X_except:
 970                         m->flags |= FLAG_EXCEPTION;
 971                         break;
 972                 case pn_Store_M:
 973                         m->mem = proj;
 974                         break;
 975                 case pn_Store_X_regular:
 976                         break;
 977                 default:
 978                         panic("Unsupported Proj from Store %+F", proj);
 979                 }
 980         }
 981         m->value.value = get_Store_value(store);
 982         m->value.mode  = get_irn_mode(m->value.value);
 983 }
 984
 985 /**
 986  * Update a memop for a Call.
 987  *
 988  * @param m  the memop
 989  */
 990 static void update_Call_memop(memop_t *m)
 991 {
 992         ir_node  *call = m->node;
 993         unsigned prop  = get_Call_memory_properties(call);
 994
 995         if (prop & mtp_property_const) {
 996                 /* A constant call did NOT use memory at all, we
 997                    can kick it from the list. */
 998         } else if (prop & mtp_property_pure) {
 999                 /* pure calls READ memory */
1000                 m->flags = 0;
1001         } else
1002                 m->flags = FLAG_KILL_ALL;
1003
1004         for (unsigned i = get_irn_n_outs(call); i-- > 0; ) {
1005                 ir_node *proj = get_irn_out(call, i);
1006
1007                 /* beware of keep edges */
1008                 if (is_End(proj))
1009                         continue;
1010
1011                 switch (get_Proj_proj(proj)) {
1012                 case pn_Call_X_except:
1013                         m->flags |= FLAG_EXCEPTION;
1014                         break;
1015                 case pn_Call_M:
1016                         m->mem = proj;
1017                         break;
1018                 }
1019         }
1020 }
1021
1022 /**
1023  * Update a memop for a Div/Mod.
1024  *
1025  * @param m  the memop
1026  */
1027 static void update_Div_memop(memop_t *m)
1028 {
1029         ir_node *div = m->node;
1030
1031         for (unsigned i = get_irn_n_outs(div); i-- > 0; ) {
1032                 ir_node *proj = get_irn_out(div, i);
1033
1034                 /* beware of keep edges */
1035                 if (is_End(proj))
1036                         continue;
1037
1038                 switch (get_Proj_proj(proj)) {
1039                 case pn_Div_X_except:
1040                         m->flags |= FLAG_EXCEPTION;
1041                         break;
1042                 case pn_Div_M:
1043                         m->mem = proj;
1044                         break;
1045                 }
1046         }
1047 }
1048
1049 static void update_Mod_memop(memop_t *m)
1050 {
1051         ir_node *div = m->node;
1052
1053         for (unsigned i = get_irn_n_outs(div); i-- > 0; ) {
1054                 ir_node *proj = get_irn_out(div, i);
1055
1056                 /* beware of keep edges */
1057                 if (is_End(proj))
1058                         continue;
1059
1060                 switch (get_Proj_proj(proj)) {
1061                 case pn_Mod_X_except:
1062                         m->flags |= FLAG_EXCEPTION;
1063                         break;
1064                 case pn_Mod_M:
1065                         m->mem = proj;
1066                         break;
1067                 }
1068         }
1069 }
1070
1071 /**
1072  * Update a memop for a Phi.
1073  *
1074  * @param m  the memop
1075  */
1076 static void update_Phi_memop(memop_t *m)
1077 {
1078         /* the Phi is its own mem */
1079         m->mem = m->node;
1080 }
1081
1082 /**
1083  * Memory walker: collect all memory ops and build topological lists.
1084  */
1085 static void collect_memops(ir_node *irn, void *ctx)
1086 {
1087         memop_t  *op;
1088         ir_node  *block;
1089         block_t  *entry;
1090
1091         (void) ctx;
1092         if (is_Proj(irn)) {
1093                 /* we can safely ignore ProjM's except the initial memory */
1094                 ir_graph *irg = get_irn_irg(irn);
1095                 if (irn != get_irg_initial_mem(irg))
1096                         return;
1097         }
1098
1099         op    = alloc_memop(irn);
1100         block = get_nodes_block(irn);
1101         entry = get_block_entry(block);
1102
1103         if (is_Phi(irn)) {
1104                 update_Phi_memop(op);
1105                 /* Phis must be always placed first */
1106                 op->next = entry->memop_forward;
1107                 entry->memop_forward = op;
1108                 if (entry->memop_backward == NULL)
1109                         entry->memop_backward = op;
1110         } else {
1111                 switch (get_irn_opcode(irn)) {
1112                 case iro_Load:
1113                         update_Load_memop(op);
1114                         break;
1115                 case iro_Store:
1116                         update_Store_memop(op);
1117                         break;
1118                 case iro_Call:
1119                         update_Call_memop(op);
1120                         break;
1121                 case iro_Sync:
1122                 case iro_Pin:
1123                         op->mem = irn;
1124                         break;
1125                 case iro_Proj:
1126                         /* initial memory */
1127                         op->mem = irn;
1128                         break;
1129                 case iro_Return:
1130                 case iro_End:
1131                         /* we can those to find the memory edge */
1132                         break;
1133                 case iro_Div:
1134                         update_Div_memop(op);
1135                         break;
1136                 case iro_Mod:
1137                         update_Mod_memop(op);
1138                         break;
1139
1140                 case iro_Builtin:
1141                         /* TODO: handle some builtins */
1142                 default:
1143                         /* unsupported operation */
1144                         op->flags = FLAG_KILL_ALL;
1145                 }
1146
1147
1148                 /* all other should be placed last */
1149                 if (entry->memop_backward == NULL) {
1150                         entry->memop_forward = entry->memop_backward = op;
1151                 } else {
1152                         entry->memop_backward->next = op;
1153                         entry->memop_backward       = op;
1154                 }
1155         }
1156 }
1157
1158 /**
1159  * Find an address in the current set.
1160  *
1161  * @param value  the value to be searched for
1162  *
1163  * @return a memop for the value or NULL if the value does
1164  *         not exists in the set or cannot be converted into
1165  *         the requested mode
1166  */
1167 static memop_t *find_address(const value_t *value)
1168 {
1169         if (rbitset_is_set(env.curr_set, value->id)) {
1170                 memop_t *res = env.curr_id_2_memop[value->id];
1171
1172                 if (res->value.mode == value->mode)
1173                         return res;
1174                 /* allow hidden casts */
1175                 if (get_mode_arithmetic(res->value.mode) == irma_twos_complement &&
1176                     get_mode_arithmetic(value->mode) == irma_twos_complement &&
1177                     get_mode_size_bits(res->value.mode) == get_mode_size_bits(value->mode))
1178                         return res;
1179         }
1180         return NULL;
1181 }
1182
1183 /**
1184  * Find an address in the avail_out set.
1185  *
1186  * @param bl     the block
1187  */
1188 static memop_t *find_address_avail(const block_t *bl, unsigned id, const ir_mode *mode)
1189 {
1190         if (rbitset_is_set(bl->avail_out, id)) {
1191                 memop_t *res = bl->id_2_memop_avail[id];
1192
1193                 if (res->value.mode == mode)
1194                         return res;
1195                 /* allow hidden casts */
1196                 if (get_mode_arithmetic(res->value.mode) == irma_twos_complement &&
1197                     get_mode_arithmetic(mode) == irma_twos_complement &&
1198                     get_mode_size_bits(res->value.mode) == get_mode_size_bits(mode))
1199                         return res;
1200         }
1201         return NULL;
1202 }
1203
1204 /**
1205  * Kill all addresses from the current set.
1206  */
1207 static void kill_all(void)
1208 {
1209         rbitset_clear_all(env.curr_set, env.rbs_size);
1210
1211         /* set sentinel */
1212         rbitset_set(env.curr_set, env.rbs_size - 1);
1213 }
1214
1215 /**
1216  * Kill memops that are not alias free due to a Store value from the current set.
1217  *
1218  * @param value  the Store value
1219  */
1220 static void kill_memops(const value_t *value)
1221 {
1222         size_t end = env.rbs_size - 1;
1223         size_t pos;
1224
1225         for (pos = rbitset_next(env.curr_set, 0, 1); pos < end; pos = rbitset_next(env.curr_set, pos + 1, 1)) {
1226                 memop_t *op = env.curr_id_2_memop[pos];
1227
1228                 if (ir_no_alias != get_alias_relation(value->address, value->mode,
1229                                                           op->value.address, op->value.mode)) {
1230                         rbitset_clear(env.curr_set, pos);
1231                         env.curr_id_2_memop[pos] = NULL;
1232                         DB((dbg, LEVEL_2, "KILLING %+F because of possible alias address %+F\n", op->node, value->address));
1233                 }
1234         }
1235 }
1236
1237 /**
1238  * Add the value of a memop to the current set.
1239  *
1240  * @param op  the memory op
1241  */
1242 static void add_memop(memop_t *op)
1243 {
1244         rbitset_set(env.curr_set, op->value.id);
1245         env.curr_id_2_memop[op->value.id] = op;
1246 }
1247
1248 /**
1249  * Add the value of a memop to the avail_out set.
1250  *
1251  * @param bl  the block
1252  * @param op  the memory op
1253  */
1254 static void add_memop_avail(block_t *bl, memop_t *op)
1255 {
1256         rbitset_set(bl->avail_out, op->value.id);
1257         bl->id_2_memop_avail[op->value.id] = op;
1258 }
1259
1260 /**
1261  * Check, if we can convert a value of one mode to another mode
1262  * without changing the representation of bits.
1263  *
1264  * @param from  the original mode
1265  * @param to    the destination mode
1266  */
1267 static int can_convert_to(const ir_mode *from, const ir_mode *to)
1268 {
1269         if (get_mode_arithmetic(from) == irma_twos_complement &&
1270             get_mode_arithmetic(to) == irma_twos_complement &&
1271             get_mode_size_bits(from) == get_mode_size_bits(to))
1272                 return 1;
1273         return 0;
1274 }
1275
1276 /**
1277  * Add a Conv to the requested mode if needed.
1278  *
1279  * @param irn   the IR-node to convert
1280  * @param mode  the destination mode
1281  *
1282  * @return the possible converted node or NULL
1283  *         if the conversion is not possible
1284  */
1285 static ir_node *conv_to(ir_node *irn, ir_mode *mode)
1286 {
1287         ir_mode *other = get_irn_mode(irn);
1288         if (other != mode) {
1289                 /* different modes: check if conversion is possible without changing the bits */
1290                 if (can_convert_to(other, mode)) {
1291                         ir_node *block = get_nodes_block(irn);
1292                         return new_r_Conv(block, irn, mode);
1293                 }
1294                 /* otherwise not possible ... yet */
1295                 return NULL;
1296         }
1297         return irn;
1298 }
1299
1300 /**
1301  * Update the address of an value if this address was a load result
1302  * and the load is killed now.
1303  *
1304  * @param value  the value whose address is updated
1305  */
1306 static void update_address(value_t *value)
1307 {
1308         if (is_Proj(value->address)) {
1309                 ir_node *load = get_Proj_pred(value->address);
1310
1311                 if (is_Load(load)) {
1312                         const memop_t *op = get_irn_memop(load);
1313
1314                         if (op->flags & FLAG_KILLED_NODE)
1315                                 value->address = op->replace;
1316                 }
1317         }
1318 }
1319
1320 /**
1321  * Do forward dataflow analysis on the given block and calculate the
1322  * GEN and KILL in the current (avail) set.
1323  *
1324  * @param bl  the block
1325  */
1326 static void calc_gen_kill_avail(block_t *bl)
1327 {
1328         memop_t *op;
1329         ir_node *def;
1330
1331         for (op = bl->memop_forward; op != NULL; op = op->next) {
1332                 switch (get_irn_opcode(op->node)) {
1333                 case iro_Phi:
1334                         /* meet */
1335                         break;
1336                 case iro_Sync:
1337                         /* join */
1338                         break;
1339                 case iro_Load:
1340                         if (! (op->flags & (FLAG_KILLED_NODE|FLAG_IGNORE))) {
1341                                 /* do we have this already? */
1342                                 memop_t *other;
1343
1344                                 update_address(&op->value);
1345                                 other = find_address(&op->value);
1346                                 if (other != NULL && other != op) {
1347                                         def = conv_to(other->value.value, op->value.mode);
1348                                         if (def != NULL) {
1349 #ifdef DEBUG_libfirm
1350                                                 if (is_Store(other->node)) {
1351                                                         /* RAW */
1352                                                         DB((dbg, LEVEL_1, "RAW %+F <- %+F(%+F)\n", op->node, def, other->node));
1353                                                 } else {
1354                                                         /* RAR */
1355                                                         DB((dbg, LEVEL_1, "RAR %+F <- %+F(%+F)\n", op->node, def, other->node));
1356                                                 }
1357 #endif
1358                                                 mark_replace_load(op, def);
1359                                                 /* do NOT change the memop table */
1360                                                 continue;
1361                                         }
1362                                 }
1363                                 /* add this value */
1364                                 add_memop(op);
1365                         }
1366                         break;
1367                 case iro_Store:
1368                         if (! (op->flags & FLAG_KILLED_NODE)) {
1369                                 /* do we have this store already */
1370                                 memop_t *other;
1371
1372                                 update_address(&op->value);
1373                                 other = find_address(&op->value);
1374                                 if (other != NULL) {
1375                                         if (is_Store(other->node)) {
1376                                                 if (op != other && !(other->flags & FLAG_IGNORE) &&
1377                                                     get_nodes_block(other->node) == get_nodes_block(op->node)) {
1378                                                         /*
1379                                                          * A WAW in the same block we can kick the first store.
1380                                                          * This is a shortcut: we know that the second Store will be anticipated
1381                                                          * then in an case.
1382                                                          */
1383                                                         DB((dbg, LEVEL_1, "WAW %+F <- %+F\n", other->node, op->node));
1384                                                         mark_remove_store(other);
1385                                                         /* FIXME: a Load might be get freed due to this killed store */
1386                                                 }
1387                                         } else if (other->value.value == op->value.value && !(op->flags & FLAG_IGNORE)) {
1388                                                 /* WAR */
1389                                                 DB((dbg, LEVEL_1, "WAR %+F <- %+F\n", op->node, other->node));
1390                                                 mark_remove_store(op);
1391                                                 /* do NOT change the memop table */
1392                                                 continue;
1393                                         }
1394                                 }
1395                                 /* KILL all possible aliases */
1396                                 kill_memops(&op->value);
1397                                 /* add this value */
1398                                 add_memop(op);
1399                         }
1400                         break;
1401                 default:
1402                         if (op->flags & FLAG_KILL_ALL)
1403                                 kill_all();
1404                 }
1405         }
1406 }
1407
1408 /**
1409  * Do forward dataflow analysis on a given block to calculate the avail_out set
1410  * for this block only.
1411  *
1412  * @param block  the block
1413  */
1414 static void forward_avail(block_t *bl)
1415 {
1416         /* fill the data from the current block */
1417         env.curr_id_2_memop = bl->id_2_memop_avail;
1418         env.curr_set        = bl->avail_out;
1419
1420         calc_gen_kill_avail(bl);
1421         dump_curr(bl, "Avail_out");
1422 }
1423
1424 /**
1425  * Do backward dataflow analysis on a given block to calculate the antic set
1426  * of Loaded addresses.
1427  *
1428  * @param bl  the block
1429  *
1430  * @return non-zero if the set has changed since last iteration
1431  */
1432 static int backward_antic(block_t *bl)
1433 {
1434         memop_t *op;
1435         ir_node *block = bl->block;
1436         int     n = get_Block_n_cfg_outs(block);
1437
1438         if (n == 1) {
1439                 ir_node  *succ    = get_Block_cfg_out(block, 0);
1440                 block_t  *succ_bl = get_block_entry(succ);
1441                 int      pred_pos = get_Block_cfgpred_pos(succ, block);
1442                 size_t   end      = env.rbs_size - 1;
1443                 size_t   pos;
1444
1445                 kill_all();
1446
1447                 if (bl->trans_results == NULL) {
1448                         /* allocate the translate cache */
1449                         bl->trans_results = OALLOCNZ(&env.obst, memop_t*, env.curr_adr_id);
1450                 }
1451
1452                 /* check for partly redundant values */
1453                 for (pos = rbitset_next(succ_bl->anticL_in, 0, 1);
1454                      pos < end;
1455                      pos = rbitset_next(succ_bl->anticL_in, pos + 1, 1)) {
1456                         /*
1457                          * do Phi-translation here: Note that at this point the nodes are
1458                          * not changed, so we can safely cache the results.
1459                          * However: Loads of Load results ARE bad, because we have no way
1460                           to translate them yet ...
1461                          */
1462                         memop_t *op = bl->trans_results[pos];
1463                         if (op == NULL) {
1464                                 /* not yet translated */
1465                                 ir_node *adr, *trans_adr;
1466
1467                                 op  = succ_bl->id_2_memop_antic[pos];
1468                                 adr = op->value.address;
1469
1470                                 trans_adr = phi_translate(adr, succ, pred_pos);
1471                                 if (trans_adr != adr) {
1472                                         /* create a new entry for the translated one */
1473                                         memop_t *new_op;
1474
1475                                         new_op = alloc_memop(NULL);
1476                                         new_op->value.address = trans_adr;
1477                                         new_op->value.id      = register_address(trans_adr);
1478                                         new_op->value.mode    = op->value.mode;
1479                                         new_op->node          = op->node; /* we need the node to decide if Load/Store */
1480                                         new_op->flags         = op->flags;
1481
1482                                         bl->trans_results[pos] = new_op;
1483                                         op = new_op;
1484                                 }
1485                         }
1486                         env.curr_id_2_memop[op->value.id] = op;
1487                         rbitset_set(env.curr_set, op->value.id);
1488                 }
1489         } else if (n > 1) {
1490                 ir_node *succ    = get_Block_cfg_out(block, 0);
1491                 block_t *succ_bl = get_block_entry(succ);
1492                 int i;
1493
1494                 rbitset_copy(env.curr_set, succ_bl->anticL_in, env.rbs_size);
1495                 memcpy(env.curr_id_2_memop, succ_bl->id_2_memop_antic, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
1496
1497                 /* Hmm: probably we want kill merges of Loads ans Stores here */
1498                 for (i = n - 1; i > 0; --i) {
1499                         ir_node *succ    = get_Block_cfg_out(bl->block, i);
1500                         block_t *succ_bl = get_block_entry(succ);
1501
1502                         rbitset_and(env.curr_set, succ_bl->anticL_in, env.rbs_size);
1503                 }
1504         } else {
1505                 /* block ends with a noreturn call */
1506                 kill_all();
1507         }
1508
1509         dump_curr(bl, "AnticL_out");
1510
1511         for (op = bl->memop_backward; op != NULL; op = op->prev) {
1512                 switch (get_irn_opcode(op->node)) {
1513                 case iro_Phi:
1514                         /* meet */
1515                         break;
1516                 case iro_Sync:
1517                         /* join */
1518                         break;
1519                 case iro_Load:
1520                         if (! (op->flags & (FLAG_KILLED_NODE|FLAG_IGNORE))) {
1521                                 /* always add it */
1522                                 add_memop(op);
1523                         }
1524                         break;
1525                 case iro_Store:
1526                         if (! (op->flags & FLAG_KILLED_NODE)) {
1527                                 /* a Store: check which memops must be killed */
1528                                 kill_memops(&op->value);
1529                         }
1530                         break;
1531                 default:
1532                         if (op->flags & FLAG_KILL_ALL)
1533                                 kill_all();
1534                 }
1535         }
1536
1537         memcpy(bl->id_2_memop_antic, env.curr_id_2_memop, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
1538         if (! rbitsets_equal(bl->anticL_in, env.curr_set, env.rbs_size)) {
1539                 /* changed */
1540                 rbitset_copy(bl->anticL_in, env.curr_set, env.rbs_size);
1541                 dump_curr(bl, "AnticL_in*");
1542                 return 1;
1543         }
1544         dump_curr(bl, "AnticL_in");
1545         return 0;
1546 }
1547
1548 /**
1549  * Replace a Load memop by a already known value.
1550  *
1551  * @param op  the Load memop
1552  */
1553 static void replace_load(memop_t *op)
1554 {
1555         ir_node *load = op->node;
1556         ir_node *def  = skip_Id(op->replace);
1557         ir_node *proj;
1558         ir_mode *mode;
1559
1560         if (def != NULL)
1561                 DB((dbg, LEVEL_1, "Replacing %+F by definition %+F\n", load, is_Proj(def) ? get_Proj_pred(def) : def));
1562         else {
1563                 if (op->flags & FLAG_EXCEPTION) {
1564                         /* bad: this node is unused and executed for exception only */
1565                         DB((dbg, LEVEL_1, "Unused %+F executed for exception only ...\n", load));
1566                         return;
1567                 }
1568                 DB((dbg, LEVEL_1, "Killing unused %+F\n", load));
1569         }
1570
1571         if (op->mem != NULL) {
1572                 /* in rare cases a Load might have NO memory */
1573                 exchange(op->mem, get_Load_mem(load));
1574         }
1575         proj = op->projs[pn_Load_res];
1576         if (proj != NULL) {
1577                 mode = get_irn_mode(proj);
1578                 if (get_irn_mode(def) != mode) {
1579                         /* a hidden cast */
1580                         dbg_info *db    = get_irn_dbg_info(load);
1581                         ir_node  *block = get_nodes_block(proj);
1582                         def = new_rd_Conv(db, block, def, mode);
1583                 }
1584                 exchange(proj, def);
1585         }
1586         proj = op->projs[pn_Load_X_except];
1587         if (proj != NULL) {
1588                 ir_graph *irg = get_irn_irg(load);
1589                 exchange(proj, new_r_Bad(irg, mode_X));
1590         }
1591         proj = op->projs[pn_Load_X_regular];
1592         if (proj != NULL) {
1593                 exchange(proj, new_r_Jmp(get_nodes_block(load)));
1594         }
1595 }
1596
1597 /**
1598  * Remove a Store memop.
1599  *
1600  * @param op  the Store memop
1601  */
1602 static void remove_store(memop_t *op)
1603 {
1604         ir_node *store = op->node;
1605         ir_node *proj;
1606
1607         DB((dbg, LEVEL_1, "Removing %+F\n", store));
1608
1609         if (op->mem != NULL) {
1610                 /* in rare cases a Store might have no memory */
1611                 exchange(op->mem, get_Store_mem(store));
1612         }
1613         proj = op->projs[pn_Store_X_except];
1614         if (proj != NULL) {
1615                 ir_graph *irg = get_irn_irg(store);
1616                 exchange(proj, new_r_Bad(irg, mode_X));
1617         }
1618         proj = op->projs[pn_Store_X_regular];
1619         if (proj != NULL) {
1620                 exchange(proj, new_r_Jmp(get_nodes_block(store)));
1621         }
1622 }
1623
1624
1625 /**
1626  * Do all necessary replacements for a given block.
1627  *
1628  * @param bl  the block
1629  */
1630 static void do_replacements(block_t *bl)
1631 {
1632         memop_t *op;
1633
1634         for (op = bl->memop_forward; op != NULL; op = op->next) {
1635                 if (op->flags & FLAG_KILLED_NODE) {
1636                         switch (get_irn_opcode(op->node)) {
1637                         case iro_Load:
1638                                 replace_load(op);
1639                                 break;
1640                         case iro_Store:
1641                                 remove_store(op);
1642                                 break;
1643                         }
1644                 }
1645         }
1646 }
1647
1648 /**
1649  * Calculate the Avail_out sets for all basic blocks.
1650  */
1651 static void calcAvail(void)
1652 {
1653         memop_t  **tmp_memop = env.curr_id_2_memop;
1654         unsigned *tmp_set    = env.curr_set;
1655         block_t  *bl;
1656
1657         /* calculate avail_out */
1658         DB((dbg, LEVEL_2, "Calculate Avail_out\n"));
1659
1660         /* iterate over all blocks in in any order, skip the start block */
1661         for (bl = env.forward->forward_next; bl != NULL; bl = bl->forward_next) {
1662                 forward_avail(bl);
1663         }
1664
1665         /* restore the current sets */
1666         env.curr_id_2_memop = tmp_memop;
1667         env.curr_set        = tmp_set;
1668 }
1669
1670 /**
1671  * Calculate the Antic_in sets for all basic blocks.
1672  */
1673 static void calcAntic(void)
1674 {
1675         int i, need_iter;
1676
1677         /* calculate antic_out */
1678         DB((dbg, LEVEL_2, "Calculate Antic_in\n"));
1679         i = 0;
1680         do {
1681                 block_t *bl;
1682
1683                 DB((dbg, LEVEL_2, "Iteration %d:\n=========\n", i));
1684
1685                 need_iter = 0;
1686
1687                 /* over all blocks in reverse post order */
1688                 for (bl = env.backward->backward_next; bl != NULL; bl = bl->backward_next) {
1689                         need_iter |= backward_antic(bl);
1690                 }
1691                 ++i;
1692         } while (need_iter);
1693         DB((dbg, LEVEL_2, "Get anticipated Load set after %d iterations\n", i));
1694 }
1695
1696 /**
1697  * Return the node representing the last memory in a block.
1698  *
1699  * @param bl  the block
1700  */
1701 static ir_node *find_last_memory(block_t *bl)
1702 {
1703         for (;;) {
1704                 if (bl->memop_backward != NULL) {
1705                         return bl->memop_backward->mem;
1706                 }
1707                 /* if there is NO memory in this block, go to the dominator */
1708                 bl = get_block_entry(get_Block_idom(bl->block));
1709         }
1710 }
1711
1712 /**
1713  * Reroute all memory users of old memory
1714  * to a new memory IR-node.
1715  *
1716  * @param omem  the old memory IR-node
1717  * @param nmem  the new memory IR-node
1718  */
1719 static void reroute_all_mem_users(ir_node *omem, ir_node *nmem)
1720 {
1721         for (unsigned i = get_irn_n_outs(omem); i-- > 0; ) {
1722                 int     n_pos;
1723                 ir_node *user = get_irn_out_ex(omem, i, &n_pos);
1724
1725                 set_irn_n(user, n_pos, nmem);
1726         }
1727
1728         /* all edges previously point to omem now point to nmem */
1729         nmem->o.out = omem->o.out;
1730 }
1731
1732 /**
1733  * Reroute memory users of old memory that are dominated by a given block
1734  * to a new memory IR-node.
1735  *
1736  * @param omem     the old memory IR-node
1737  * @param nmem     the new memory IR-node
1738  * @param pass_bl  the block the memory must pass
1739  */
1740 static void reroute_mem_through(ir_node *omem, ir_node *nmem, ir_node *pass_bl)
1741 {
1742         unsigned n = get_irn_n_outs(omem);
1743         ir_def_use_edges *new_out = OALLOCF(&env.obst, ir_def_use_edges, edges, n);
1744
1745         unsigned j = 0;
1746         for (unsigned i = 0; i < n; ++i) {
1747                 int     n_pos;
1748                 ir_node *user   = get_irn_out_ex(omem, i, &n_pos);
1749                 ir_node *use_bl = get_nodes_block(user);
1750
1751
1752                 if (is_Phi(user)) {
1753                         use_bl = get_Block_cfgpred_block(use_bl, n_pos);
1754                 }
1755                 if (block_dominates(pass_bl, use_bl)) {
1756                         /* found an user that is dominated */
1757                         new_out->edges[j].pos = n_pos;
1758                         new_out->edges[j].use = user;
1759                         ++j;
1760
1761                         set_irn_n(user, n_pos, nmem);
1762                 }
1763         }
1764         new_out->n_edges = j;
1765
1766         /* Modify the out structure: we create a new out edge array on our
1767            temporary obstack here. This should be no problem, as we invalidate the
1768            edges at the end either. */
1769         /* first entry is used for the length */
1770         nmem->o.out = new_out;
1771 }
1772
1773 /**
1774  * insert Loads, making partly redundant Loads fully redundant
1775  */
1776 static int insert_Load(block_t *bl)
1777 {
1778         ir_node  *block = bl->block;
1779         int      i, n = get_Block_n_cfgpreds(block);
1780         size_t   end = env.rbs_size - 1;
1781
1782         DB((dbg, LEVEL_3, "processing %+F\n", block));
1783
1784         if (n == 0) {
1785                 /* might still happen for an unreachable block (end for instance) */
1786                 return 0;
1787         }
1788
1789         if (n > 1) {
1790                 ir_node **ins;
1791                 size_t    pos;
1792
1793                 NEW_ARR_A(ir_node *, ins, n);
1794
1795                 rbitset_set_all(env.curr_set, env.rbs_size);
1796
1797                 /* More than one predecessors, calculate the join for all avail_outs ignoring unevaluated
1798                    Blocks. These put in Top anyway. */
1799                 for (i = n - 1; i >= 0; --i) {
1800                         ir_node *pred = skip_Proj(get_Block_cfgpred(block, i));
1801                         ir_node *blk  = get_nodes_block(pred);
1802                         block_t *pred_bl;
1803
1804                         pred_bl = get_block_entry(blk);
1805                         rbitset_and(env.curr_set, pred_bl->avail_out, env.rbs_size);
1806
1807                         if (is_Load(pred) || is_Store(pred)) {
1808                                 /* We reached this block by an exception from a Load or Store:
1809                                  * the memop creating the exception was NOT completed than, kill it
1810                                  */
1811                                 memop_t *exc_op = get_irn_memop(pred);
1812                                 rbitset_clear(env.curr_set, exc_op->value.id);
1813                         }
1814
1815                 }
1816                 /*
1817                  * Ensure that all values are in the map: build Phi's if necessary:
1818                  * Note: the last bit is the sentinel and ALWAYS set, so end with -2.
1819                  */
1820                 for (pos = 0; pos < env.rbs_size - 1; ++pos) {
1821                         if (! rbitset_is_set(env.curr_set, pos))
1822                                 env.curr_id_2_memop[pos] = NULL;
1823                         else {
1824                                 int      need_phi = 0;
1825                                 memop_t *first    = NULL;
1826                                 ir_mode *mode     = NULL;
1827
1828                                 for (i = 0; i < n; ++i) {
1829                                         ir_node *pred    = get_Block_cfgpred_block(bl->block, i);
1830                                         block_t *pred_bl = get_block_entry(pred);
1831
1832                                         memop_t *mop = pred_bl->id_2_memop_avail[pos];
1833                                         if (first == NULL) {
1834                                                 first = mop;
1835                                                 ins[0] = first->value.value;
1836                                                 mode = get_irn_mode(ins[0]);
1837
1838                                                 /* no Phi needed so far */
1839                                                 env.curr_id_2_memop[pos] = first;
1840                                         } else {
1841                                                 ins[i] = conv_to(mop->value.value, mode);
1842                                                 if (ins[i] != ins[0]) {
1843                                                         if (ins[i] == NULL) {
1844                                                                 /* conversion failed */
1845                                                                 env.curr_id_2_memop[pos] = NULL;
1846                                                                 rbitset_clear(env.curr_set, pos);
1847                                                                 break;
1848                                                         }
1849                                                         need_phi = 1;
1850                                                 }
1851                                         }
1852                                 }
1853                                 if (need_phi) {
1854                                         /* build a Phi  */
1855                                         ir_node *phi = new_r_Phi(bl->block, n, ins, mode);
1856                                         memop_t *phiop = alloc_memop(phi);
1857
1858                                         phiop->value = first->value;
1859                                         phiop->value.value = phi;
1860
1861                                         /* no need to link it in, as it is a DATA phi */
1862
1863                                         env.curr_id_2_memop[pos] = phiop;
1864
1865                                         DB((dbg, LEVEL_3, "Created new %+F on merging value for address %+F\n", phi, first->value.address));
1866                                 }
1867                         }
1868                 }
1869         } else {
1870                 /* only one predecessor, simply copy the map */
1871                 ir_node *pred    = get_Block_cfgpred_block(bl->block, 0);
1872                 block_t *pred_bl = get_block_entry(pred);
1873
1874                 rbitset_copy(env.curr_set, pred_bl->avail_out, env.rbs_size);
1875
1876                 memcpy(env.curr_id_2_memop, pred_bl->id_2_memop_avail, env.rbs_size * sizeof(bl->id_2_memop_avail[0]));
1877         }
1878
1879         if (n > 1) {
1880                 size_t pos;
1881
1882                 /* check for partly redundant values */
1883                 for (pos = rbitset_next(bl->anticL_in, 0, 1);
1884                      pos < end;
1885                      pos = rbitset_next(bl->anticL_in, pos + 1, 1)) {
1886                         memop_t *op = bl->id_2_memop_antic[pos];
1887                         int     have_some, all_same;
1888                         ir_node *first;
1889
1890                         if (rbitset_is_set(env.curr_set, pos)) {
1891                                 /* already avail */
1892                                 continue;
1893                         }
1894
1895                         assert(is_Load(op->node));
1896
1897                         DB((dbg, LEVEL_3, "anticipated %+F\n", op->node));
1898
1899                         have_some  = 0;
1900                         all_same   = 1;
1901                         first      = 0;
1902                         for (i = n - 1; i >= 0; --i) {
1903                                 ir_node *pred    = get_Block_cfgpred_block(block, i);
1904                                 block_t *pred_bl = get_block_entry(pred);
1905                                 ir_mode *mode    = op->value.mode;
1906                                 memop_t *e;
1907                                 ir_node *adr;
1908
1909                                 adr = phi_translate(op->value.address, block, i);
1910                                 DB((dbg, LEVEL_3, ".. using address %+F in pred %d\n", adr, i));
1911                                 e   = find_address_avail(pred_bl, register_address(adr), mode);
1912                                 if (e == NULL) {
1913                                         ir_node *ef_block = get_nodes_block(adr);
1914                                         if (! block_dominates(ef_block, pred)) {
1915                                                 /* cannot place a copy here */
1916                                                 have_some = 0;
1917                                                 DB((dbg, LEVEL_3, "%+F cannot be moved into predecessor %+F\n", op->node, pred));
1918                                                 break;
1919                                         }
1920                                         DB((dbg, LEVEL_3, "%+F is not available in predecessor %+F\n", op->node, pred));
1921                                         pred_bl->avail = NULL;
1922                                         all_same       = 0;
1923                                 } else {
1924                                         if (e->value.mode != mode && !can_convert_to(e->value.mode, mode)) {
1925                                                 /* cannot create a Phi due to different modes */
1926                                                 have_some = 0;
1927                                                 break;
1928                                         }
1929                                         pred_bl->avail = e;
1930                                         have_some      = 1;
1931                                         DB((dbg, LEVEL_3, "%+F is available for %+F in predecessor %+F\n", e->node, op->node, pred));
1932                                         if (first == NULL)
1933                                                 first = e->node;
1934                                         else if (first != e->node)
1935                                                 all_same = 0;
1936                                 }
1937                         }
1938                         if (have_some && !all_same) {
1939                                 ir_mode *mode = op->value.mode;
1940                                 ir_node **in, *phi;
1941                                 memop_t *phi_op;
1942
1943                                 NEW_ARR_A(ir_node *, in, n);
1944
1945                                 for (i = n - 1; i >= 0; --i) {
1946                                         ir_node *pred    = get_Block_cfgpred_block(block, i);
1947                                         block_t *pred_bl = get_block_entry(pred);
1948
1949                                         if (pred_bl->avail == NULL) {
1950                                                 /* create a new Load here and make to make it fully redundant */
1951                                                 dbg_info *db       = get_irn_dbg_info(op->node);
1952                                                 ir_node  *last_mem = find_last_memory(pred_bl);
1953                                                 ir_node  *load, *def, *adr;
1954                                                 memop_t  *new_op;
1955
1956                                                 assert(last_mem != NULL);
1957                                                 adr  = phi_translate(op->value.address, block, i);
1958                                                 load = new_rd_Load(db, pred, last_mem, adr, mode, cons_none);
1959                                                 def  = new_r_Proj(load, mode, pn_Load_res);
1960                                                 DB((dbg, LEVEL_1, "Created new %+F in %+F for party redundant %+F\n", load, pred, op->node));
1961
1962                                                 new_op                = alloc_memop(load);
1963                                                 new_op->mem           = new_r_Proj(load, mode_M, pn_Load_M);
1964                                                 new_op->value.address = adr;
1965                                                 new_op->value.id      = op->value.id;
1966                                                 new_op->value.mode    = mode;
1967                                                 new_op->value.value   = def;
1968
1969                                                 new_op->projs[pn_Load_M]   = new_op->mem;
1970                                                 new_op->projs[pn_Load_res] = def;
1971
1972                                                 new_op->prev = pred_bl->memop_backward;
1973                                                 if (pred_bl->memop_backward != NULL)
1974                                                         pred_bl->memop_backward->next = new_op;
1975
1976                                                 pred_bl->memop_backward = new_op;
1977
1978                                                 if (pred_bl->memop_forward == NULL)
1979                                                         pred_bl->memop_forward = new_op;
1980
1981                                                 if (get_nodes_block(last_mem) == pred) {
1982                                                         /* We have add a new last memory op in pred block.
1983                                                            If pred had already a last mem, reroute all memory
1984                                                            users. */
1985                                                         reroute_all_mem_users(last_mem, new_op->mem);
1986                                                 } else {
1987                                                         /* reroute only those memory going through the pre block */
1988                                                         reroute_mem_through(last_mem, new_op->mem, pred);
1989                                                 }
1990
1991                                                 /* we added this load at the end, so it will be avail anyway */
1992                                                 add_memop_avail(pred_bl, new_op);
1993                                                 pred_bl->avail = new_op;
1994                                         }
1995                                         in[i] = conv_to(pred_bl->avail->value.value, mode);
1996                                 }
1997                                 phi = new_r_Phi(block, n, in, mode);
1998                                 DB((dbg, LEVEL_1, "Created new %+F in %+F for now redundant %+F\n", phi, block, op->node));
1999
2000                                 phi_op = clone_memop_phi(op, phi);
2001                                 add_memop(phi_op);
2002                         }
2003                 }
2004         }
2005
2006         /* recalculate avail by gen and kill */
2007         calc_gen_kill_avail(bl);
2008
2009         /* always update the map after gen/kill, as values might have been changed due to RAR/WAR/WAW */
2010         memcpy(bl->id_2_memop_avail, env.curr_id_2_memop, env.rbs_size * sizeof(env.curr_id_2_memop[0]));
2011
2012         if (!rbitsets_equal(bl->avail_out, env.curr_set, env.rbs_size)) {
2013                 /* the avail set has changed */
2014                 rbitset_copy(bl->avail_out, env.curr_set, env.rbs_size);
2015                 dump_curr(bl, "Avail_out*");
2016                 return 1;
2017         }
2018         dump_curr(bl, "Avail_out");
2019         return 0;
2020 }
2021
2022 /**
2023  * Insert Loads upwards.
2024  */
2025 static void insert_Loads_upwards(void)
2026 {
2027         int i, need_iter;
2028         block_t *bl;
2029
2030         /* recalculate antic_out and insert Loads */
2031         DB((dbg, LEVEL_2, "Inserting Loads\n"));
2032
2033         i = 0;
2034         do {
2035                 DB((dbg, LEVEL_2, "Iteration %d:\n=========\n", i));
2036
2037                 need_iter = 0;
2038
2039                 /* over all blocks in reverse post order, skip the start block */
2040                 for (bl = env.forward->forward_next; bl != NULL; bl = bl->forward_next) {
2041                         need_iter |= insert_Load(bl);
2042                 }
2043                 ++i;
2044         } while (need_iter);
2045
2046         DB((dbg, LEVEL_2, "Finished Load inserting after %d iterations\n", i));
2047 }
2048
2049 void opt_ldst(ir_graph *irg)
2050 {
2051         block_t *bl;
2052
2053         FIRM_DBG_REGISTER(dbg, "firm.opt.ldst");
2054
2055         DB((dbg, LEVEL_1, "\nDoing Load/Store optimization on %+F\n", irg));
2056
2057         assure_irg_properties(irg,
2058                 IR_GRAPH_PROPERTY_NO_CRITICAL_EDGES /* we need landing pads */
2059                 | IR_GRAPH_PROPERTY_CONSISTENT_ENTITY_USAGE
2060                 | IR_GRAPH_PROPERTY_CONSISTENT_OUTS
2061                 | IR_GRAPH_PROPERTY_NO_UNREACHABLE_CODE
2062                 | IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE);
2063
2064         if (get_opt_alias_analysis()) {
2065                 assure_irp_globals_entity_usage_computed();
2066         }
2067
2068         obstack_init(&env.obst);
2069         ir_nodehashmap_init(&env.adr_map);
2070
2071         env.forward       = NULL;
2072         env.backward      = NULL;
2073         env.curr_adr_id   = 0;
2074         env.n_mem_ops     = 0;
2075         env.max_cfg_preds = 0;
2076         env.changed       = 0;
2077         env.start_bl      = get_irg_start_block(irg);
2078         env.end_bl        = get_irg_end_block(irg);
2079 #ifdef DEBUG_libfirm
2080         env.id_2_address  = NEW_ARR_F(ir_node *, 0);
2081 #endif
2082
2083         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_BLOCK_MARK);
2084
2085         /* first step: allocate block entries. Note that some blocks might be
2086            unreachable here. Using the normal walk ensures that ALL blocks are initialized. */
2087         irg_walk_graph(irg, prepare_blocks, link_phis, NULL);
2088
2089         /* produce an inverse post-order list for the CFG: this links only reachable
2090            blocks */
2091         irg_out_block_walk(get_irg_start_block(irg), NULL, inverse_post_order, NULL);
2092
2093         if (! get_Block_mark(env.end_bl)) {
2094                 /*
2095                  * The end block is NOT reachable due to endless loops
2096                  * or no_return calls.
2097                  * Place the end block last.
2098                  * env.backward points to the last block in the list for this purpose.
2099                  */
2100                 env.backward->forward_next = get_block_entry(env.end_bl);
2101
2102                 set_Block_mark(env.end_bl, 1);
2103         }
2104
2105         /* second step: find and sort all memory ops */
2106         walk_memory_irg(irg, collect_memops, NULL, NULL);
2107
2108 #ifdef DEBUG_libfirm
2109         /* check that the backward map is correct */
2110         assert((unsigned)ARR_LEN(env.id_2_address) == env.curr_adr_id);
2111 #endif
2112
2113         if (env.n_mem_ops == 0) {
2114                 /* no memory ops */
2115                 goto no_changes;
2116         }
2117
2118         /* create the backward links. */
2119         env.backward = NULL;
2120         irg_block_walk_graph(irg, NULL, collect_backward, NULL);
2121
2122         /* link the end block in */
2123         bl = get_block_entry(env.end_bl);
2124         bl->backward_next = env.backward;
2125         env.backward      = bl;
2126
2127         /* check that we really start with the start / end block */
2128         assert(env.forward->block  == env.start_bl);
2129         assert(env.backward->block == env.end_bl);
2130
2131         /* create address sets: for now, only the existing addresses are allowed plus one
2132            needed for the sentinel */
2133         env.rbs_size = env.curr_adr_id + 1;
2134
2135         /* create the current set */
2136         env.curr_set = rbitset_obstack_alloc(&env.obst, env.rbs_size);
2137         rbitset_set(env.curr_set, env.rbs_size - 1);
2138         env.curr_id_2_memop = NEW_ARR_DZ(memop_t*, &env.obst, env.rbs_size);
2139
2140         for (bl = env.forward; bl != NULL; bl = bl->forward_next) {
2141                 /* set sentinel bits */
2142                 bl->avail_out  = rbitset_obstack_alloc(&env.obst, env.rbs_size);
2143                 rbitset_set(bl->avail_out, env.rbs_size - 1);
2144
2145                 bl->id_2_memop_avail = NEW_ARR_DZ(memop_t*, &env.obst, env.rbs_size);
2146
2147                 bl->anticL_in  = rbitset_obstack_alloc(&env.obst, env.rbs_size);
2148                 rbitset_set(bl->anticL_in, env.rbs_size - 1);
2149
2150                 bl->id_2_memop_antic = NEW_ARR_DZ(memop_t*, &env.obst, env.rbs_size);
2151         }
2152
2153         (void) dump_block_list;
2154
2155         calcAvail();
2156         calcAntic();
2157
2158         insert_Loads_upwards();
2159
2160         if (env.changed) {
2161                 /* over all blocks in reverse post order */
2162                 for (bl = env.forward; bl != NULL; bl = bl->forward_next) {
2163                         do_replacements(bl);
2164                 }
2165
2166                 /* not only invalidate but free them. We might allocate new out arrays
2167                    on our obstack which will be deleted yet. */
2168                 confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_CONTROL_FLOW);
2169         } else {
2170 no_changes:
2171                 confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_ALL);
2172         }
2173
2174         ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_BLOCK_MARK);
2175         ir_nodehashmap_destroy(&env.adr_map);
2176         obstack_free(&env.obst, NULL);
2177
2178 #ifdef DEBUG_libfirm
2179         DEL_ARR_F(env.id_2_address);
2180 #endif
2181 }
2182
2183 ir_graph_pass_t *opt_ldst_pass(const char *name)
2184 {
2185         return def_graph_pass(name ? name : "ldst_df", opt_ldst);
2186 }