nsz Git - libfirm/blob - ir/be/ia32/ia32_optimize.c

   1 #ifdef HAVE_CONFIG_H
   2 #include "config.h"
   3 #endif
   4
   5 #include "irnode.h"
   6 #include "irprog_t.h"
   7 #include "ircons.h"
   8 #include "firm_types.h"
   9 #include "iredges.h"
  10 #include "tv.h"
  11 #include "irgmod.h"
  12
  13 #include "../be_t.h"
  14 #include "../beabi.h"
  15
  16 #include "ia32_new_nodes.h"
  17 #include "bearch_ia32_t.h"
  18 #include "gen_ia32_regalloc_if.h"     /* the generated interface (register type and class defenitions) */
  19
  20 #undef is_NoMem
  21 #define is_NoMem(irn) (get_irn_op(irn) == op_NoMem)
  22
  23 static int be_is_NoReg(be_abi_irg_t *babi, const ir_node *irn) {
  24         if (be_abi_get_callee_save_irn(babi, &ia32_gp_regs[REG_XXX]) == irn ||
  25                 be_abi_get_callee_save_irn(babi, &ia32_fp_regs[REG_XXXX]) == irn)
  26         {
  27                 return 1;
  28         }
  29
  30         return 0;
  31 }
  32
  33 /**
  34  * creates a unique ident by adding a number to a tag
  35  *
  36  * @param tag   the tag string, must contain a %d if a number
  37  *              should be added
  38  */
  39 static ident *unique_id(const char *tag)
  40 {
  41         static unsigned id = 0;
  42         char str[256];
  43
  44         snprintf(str, sizeof(str), tag, ++id);
  45         return new_id_from_str(str);
  46 }
  47
  48
  49
  50 /**
  51  * Transforms a SymConst.
  52  *
  53  * @param mod     the debug module
  54  * @param block   the block the new node should belong to
  55  * @param node    the ir SymConst node
  56  * @param mode    mode of the SymConst
  57  * @return the created ia32 Const node
  58  */
  59 static ir_node *gen_SymConst(ia32_transform_env_t *env) {
  60         ir_node  *cnst;
  61         dbg_info *dbg   = env->dbg;
  62         ir_mode  *mode  = env->mode;
  63         ir_graph *irg   = env->irg;
  64         ir_node  *block = env->block;
  65
  66         cnst = new_rd_ia32_Const(dbg, irg, block, mode);
  67         set_ia32_Const_attr(cnst, env->irn);
  68         return cnst;
  69 }
  70
  71 /**
  72  * Get a primitive type for a mode.
  73  */
  74 static ir_type *get_prim_type(pmap *types, ir_mode *mode)
  75 {
  76         pmap_entry *e = pmap_find(types, mode);
  77         ir_type *res;
  78
  79         if (! e) {
  80                 char buf[64];
  81                 snprintf(buf, sizeof(buf), "prim_type_%s", get_mode_name(mode));
  82                 res = new_type_primitive(new_id_from_str(buf), mode);
  83                 pmap_insert(types, mode, res);
  84         }
  85         else
  86                 res = e->value;
  87         return res;
  88 }
  89
  90 /**
  91  * Get an entity that is initialized with a tarval
  92  */
  93 static entity *get_entity_for_tv(ia32_code_gen_t *cg, ir_node *cnst)
  94 {
  95         tarval *tv    = get_Const_tarval(cnst);
  96         pmap_entry *e = pmap_find(cg->tv_ent, tv);
  97         entity *res;
  98         ir_graph *rem;
  99
 100         if (! e) {
 101                 ir_mode *mode = get_irn_mode(cnst);
 102                 ir_type *tp = get_Const_type(cnst);
 103                 if (tp == firm_unknown_type)
 104                         tp = get_prim_type(cg->types, mode);
 105
 106                 res = new_entity(get_glob_type(), unique_id("ia32FloatCnst_%u"), tp);
 107
 108                 set_entity_ld_ident(res, get_entity_ident(res));
 109                 set_entity_visibility(res, visibility_local);
 110                 set_entity_variability(res, variability_constant);
 111                 set_entity_allocation(res, allocation_static);
 112
 113                  /* we create a new entity here: It's initialization must resist on the
 114                     const code irg */
 115                 rem = current_ir_graph;
 116                 current_ir_graph = get_const_code_irg();
 117                 set_atomic_ent_value(res, new_Const_type(tv, tp));
 118                 current_ir_graph = rem;
 119         }
 120         else
 121                 res = e->value;
 122         return res;
 123 }
 124
 125 /**
 126  * Transforms a Const.
 127  *
 128  * @param mod     the debug module
 129  * @param block   the block the new node should belong to
 130  * @param node    the ir Const node
 131  * @param mode    mode of the Const
 132  * @return the created ia32 Const node
 133  */
 134 static ir_node *gen_Const(ia32_transform_env_t *env) {
 135         ir_node *cnst;
 136         symconst_symbol sym;
 137         ir_graph *irg   = env->irg;
 138         ir_node  *block = env->block;
 139         ir_node  *node  = env->irn;
 140         dbg_info *dbg   = env->dbg;
 141         ir_mode  *mode  = env->mode;
 142
 143         if (mode_is_float(mode)) {
 144                 sym.entity_p = get_entity_for_tv(env->cg, node);
 145
 146                 cnst = new_rd_SymConst(dbg, irg, block, sym, symconst_addr_ent);
 147                 env->irn = cnst;
 148                 cnst = gen_SymConst(env);
 149         }
 150         else {
 151                 cnst = new_rd_ia32_Const(dbg, irg, block, get_irn_mode(node));
 152                 set_ia32_Const_attr(cnst, node);
 153         }
 154         return cnst;
 155 }
 156
 157
 158
 159 /**
 160  * Transforms (all) Const's into ia32_Const and places them in the
 161  * block where they are used (or in the cfg-pred Block in case of Phi's)
 162  */
 163 void ia32_place_consts(ir_node *irn, void *env) {
 164         ia32_code_gen_t      *cg = env;
 165         ia32_transform_env_t  tenv;
 166         ir_mode              *mode;
 167         ir_node              *pred, *cnst;
 168         int                   i;
 169         opcode                opc;
 170
 171         if (is_Block(irn))
 172                 return;
 173
 174         mode = get_irn_mode(irn);
 175
 176         tenv.block    = get_nodes_block(irn);
 177         tenv.cg       = cg;
 178         tenv.irg      = cg->irg;
 179         tenv.mod      = cg->mod;
 180
 181         /* Loop over all predecessors and check for Sym/Const nodes */
 182         for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
 183                 pred      = get_irn_n(irn, i);
 184                 cnst      = NULL;
 185                 opc       = get_irn_opcode(pred);
 186                 tenv.irn  = pred;
 187                 tenv.mode = get_irn_mode(pred);
 188                 tenv.dbg  = get_irn_dbg_info(pred);
 189
 190                 /* If it's a Phi, then we need to create the */
 191                 /* new Const in it's predecessor block       */
 192                 if (is_Phi(irn)) {
 193                         tenv.block = get_Block_cfgpred_block(get_nodes_block(irn), i);
 194                 }
 195
 196                 switch (opc) {
 197                         case iro_Const:
 198                                 cnst = gen_Const(&tenv);
 199                                 break;
 200                         case iro_SymConst:
 201                                 cnst = gen_SymConst(&tenv);
 202                                 break;
 203                         default:
 204                                 break;
 205                 }
 206
 207                 /* if we found a const, then set it */
 208                 if (cnst) {
 209                         set_irn_n(irn, i, cnst);
 210                 }
 211         }
 212 }
 213
 214
 215 /******************************************************************
 216  *              _     _                   __  __           _
 217  *     /\      | |   | |                 |  \/  |         | |
 218  *    /  \   __| | __| |_ __ ___  ___ ___| \  / | ___   __| | ___
 219  *   / /\ \ / _` |/ _` | '__/ _ \/ __/ __| |\/| |/ _ \ / _` |/ _ \
 220  *  / ____ \ (_| | (_| | | |  __/\__ \__ \ |  | | (_) | (_| |  __/
 221  * /_/    \_\__,_|\__,_|_|  \___||___/___/_|  |_|\___/ \__,_|\___|
 222  *
 223  ******************************************************************/
 224
 225 static int node_is_comm(const ir_node *irn) {
 226         if (is_ia32_Add(irn)  ||
 227                 is_ia32_fAdd(irn) ||
 228                 is_ia32_Mul(irn)  ||
 229                 is_ia32_Mulh(irn) ||
 230                 is_ia32_fMul(irn) ||
 231                 is_ia32_And(irn)  ||
 232                 is_ia32_fAnd(irn) ||
 233                 is_ia32_Or(irn)   ||
 234                 is_ia32_fOr(irn)  ||
 235                 is_ia32_Eor(irn)  ||
 236                 is_ia32_fEor(irn) ||
 237                 is_ia32_Min(irn)  ||
 238                 is_ia32_fMin(irn) ||
 239                 is_ia32_Max(irn)  ||
 240                 is_ia32_fMax(irn))
 241         {
 242                 return 1;
 243         }
 244
 245         return 0;
 246 }
 247
 248 static int ia32_get_irn_n_edges(const ir_node *irn) {
 249         const ir_edge_t *edge;
 250         int cnt = 0;
 251
 252         foreach_out_edge(irn, edge) {
 253                 cnt++;
 254         }
 255
 256         return cnt;
 257 }
 258
 259 /**
 260  * Returns the first mode_M Proj connected to irn.
 261  */
 262 static ir_node *get_mem_proj(const ir_node *irn) {
 263         const ir_edge_t *edge;
 264         ir_node         *src;
 265
 266         assert(get_irn_mode(irn) == mode_T && "expected mode_T node");
 267
 268         foreach_out_edge(irn, edge) {
 269                 src = get_edge_src_irn(edge);
 270
 271                 assert(is_Proj(src) && "Proj expected");
 272
 273                 if (get_irn_mode(src) == mode_M)
 274                         return src;
 275         }
 276
 277         return NULL;
 278 }
 279
 280 /**
 281  * Returns the Proj with number 0 connected to irn.
 282  */
 283 static ir_node *get_res_proj(const ir_node *irn) {
 284         const ir_edge_t *edge;
 285         ir_node         *src;
 286
 287         assert(get_irn_mode(irn) == mode_T && "expected mode_T node");
 288
 289         foreach_out_edge(irn, edge) {
 290                 src = get_edge_src_irn(edge);
 291
 292                 assert(is_Proj(src) && "Proj expected");
 293
 294                 if (get_Proj_proj(src) == 0)
 295                         return src;
 296         }
 297
 298         return NULL;
 299 }
 300
 301 /**
 302  * Determines if pred is a Proj and if is_op_func returns true for it's predecessor.
 303  *
 304  * @param pred       The node to be checked
 305  * @param is_op_func The check-function
 306  * @return 1 if conditions are fulfilled, 0 otherwise
 307  */
 308 static int pred_is_specific_node(const ir_node *pred, int (*is_op_func)(const ir_node *n)) {
 309         if (is_Proj(pred) && is_op_func(get_Proj_pred(pred))) {
 310                 return 1;
 311         }
 312
 313         return 0;
 314 }
 315
 316 /**
 317  * Determines if pred is a Proj and if is_op_func returns true for it's predecessor
 318  * and if the predecessor is in block bl.
 319  *
 320  * @param bl         The block
 321  * @param pred       The node to be checked
 322  * @param is_op_func The check-function
 323  * @return 1 if conditions are fulfilled, 0 otherwise
 324  */
 325 static int pred_is_specific_nodeblock(const ir_node *bl, const ir_node *pred,
 326         int (*is_op_func)(const ir_node *n))
 327 {
 328         if (is_Proj(pred)) {
 329                 pred = get_Proj_pred(pred);
 330                 if ((bl == get_nodes_block(pred)) && is_op_func(pred)) {
 331                         return 1;
 332                 }
 333         }
 334
 335         return 0;
 336 }
 337
 338 /**
 339  * Folds Add or Sub to LEA if possible
 340  */
 341 static ir_node *fold_addr(be_abi_irg_t *babi, ir_node *irn, firm_dbg_module_t *mod, ir_node *noreg) {
 342         ir_graph *irg       = get_irn_irg(irn);
 343         ir_mode  *mode      = get_irn_mode(irn);
 344         dbg_info *dbg       = get_irn_dbg_info(irn);
 345         ir_node  *block     = get_nodes_block(irn);
 346         ir_node  *res       = irn;
 347         char     *offs      = NULL;
 348         char     *offs_cnst = NULL;
 349         char     *offs_lea  = NULL;
 350         int       scale     = 0;
 351         int       isadd     = 0;
 352         int       dolea     = 0;
 353         ir_node  *left, *right, *temp;
 354         ir_node  *base, *index;
 355         ia32_am_flavour_t am_flav;
 356
 357         if (is_ia32_Add(irn))
 358                 isadd = 1;
 359
 360         left  = get_irn_n(irn, 2);
 361         right = get_irn_n(irn, 3);
 362
 363         base    = left;
 364         index   = noreg;
 365         offs    = NULL;
 366         scale   = 0;
 367         am_flav = 0;
 368
 369         /* "normalize" arguments in case of add with two operands */
 370         if  (isadd && ! be_is_NoReg(babi, right)) {
 371                 /* put LEA == ia32_am_O as right operand */
 372                 if (is_ia32_Lea(left) && get_ia32_am_flavour(left) == ia32_am_O) {
 373                         set_irn_n(irn, 2, right);
 374                         set_irn_n(irn, 3, left);
 375                         temp  = left;
 376                         left  = right;
 377                         right = temp;
 378                 }
 379
 380                 /* put LEA != ia32_am_O as left operand */
 381                 if (is_ia32_Lea(right) && get_ia32_am_flavour(right) != ia32_am_O) {
 382                         set_irn_n(irn, 2, right);
 383                         set_irn_n(irn, 3, left);
 384                         temp  = left;
 385                         left  = right;
 386                         right = temp;
 387                 }
 388
 389                 /* put SHL as left operand iff left is NOT a LEA */
 390                 if (! is_ia32_Lea(left) && pred_is_specific_node(right, is_ia32_Shl)) {
 391                         set_irn_n(irn, 2, right);
 392                         set_irn_n(irn, 3, left);
 393                         temp  = left;
 394                         left  = right;
 395                         right = temp;
 396                 }
 397         }
 398
 399         /* check if operand is either const */
 400         if (get_ia32_cnst(irn)) {
 401                 DBG((mod, LEVEL_1, "\tfound op with imm"));
 402
 403                 offs_cnst = get_ia32_cnst(irn);
 404                 dolea     = 1;
 405         }
 406
 407         /* determine the operand which needs to be checked */
 408         if (be_is_NoReg(babi, right)) {
 409                 temp = left;
 410         }
 411         else {
 412                 temp = right;
 413         }
 414
 415         /* check if right operand is AMConst (LEA with ia32_am_O) */
 416         if (is_ia32_Lea(temp) && get_ia32_am_flavour(temp) == ia32_am_O) {
 417                 DBG((mod, LEVEL_1, "\tgot op with LEA am_O"));
 418
 419                 offs_lea = get_ia32_am_offs(temp);
 420                 dolea    = 1;
 421         }
 422
 423         if (isadd) {
 424                 /* default for add -> make right operand to index */
 425                 index = right;
 426                 dolea = 1;
 427
 428                 DBG((mod, LEVEL_1, "\tgot LEA candidate with index %+F\n", index));
 429
 430                 /* determine the operand which needs to be checked */
 431                 temp = left;
 432                 if (is_ia32_Lea(left)) {
 433                         temp = right;
 434                 }
 435
 436                 /* check for SHL 1,2,3 */
 437                 if (pred_is_specific_node(temp, is_ia32_Shl)) {
 438                         temp = get_Proj_pred(temp);
 439
 440                         if (get_ia32_Immop_tarval(temp)) {
 441                                 scale = get_tarval_long(get_ia32_Immop_tarval(temp));
 442
 443                                 if (scale <= 3) {
 444                                         scale = 1 << scale;
 445                                         index = get_irn_n(temp, 2);
 446
 447                                         DBG((mod, LEVEL_1, "\tgot scaled index %+F\n", index));
 448                                 }
 449                         }
 450                 }
 451
 452                 /* fix base */
 453                 if (! be_is_NoReg(babi, index)) {
 454                         /* if we have index, but left == right -> no base */
 455                         if (left == right) {
 456                                 base = noreg;
 457                         }
 458                         else if (! is_ia32_Lea(left) && (index != right)) {
 459                                 /* index != right -> we found a good Shl           */
 460                                 /* left  != LEA   -> this Shl was the left operand */
 461                                 /* -> base is right operand                        */
 462                                 base = right;
 463                         }
 464                 }
 465         }
 466
 467         /* Try to assimilate a LEA as left operand */
 468         if (is_ia32_Lea(left) && (get_ia32_am_flavour(left) != ia32_am_O)) {
 469                 am_flav = get_ia32_am_flavour(left);
 470
 471                 /* If we have an Add with a real right operand (not NoReg) and  */
 472                 /* the LEA contains already an index calculation then we create */
 473                 /* a new LEA.                                                   */
 474                 if (isadd && !be_is_NoReg(babi, index) && (am_flav & ia32_am_I)) {
 475                         DBG((mod, LEVEL_1, "\tleave old LEA, creating new one\n"));
 476                 }
 477                 else {
 478                         DBG((mod, LEVEL_1, "\tgot LEA as left operand ... assimilating\n"));
 479                         offs  = get_ia32_am_offs(left);
 480                         base  = get_irn_n(left, 0);
 481                         index = get_irn_n(left, 1);
 482                         scale = get_ia32_am_scale(left);
 483                 }
 484         }
 485
 486         /* ok, we can create a new LEA */
 487         if (dolea) {
 488                 res = new_rd_ia32_Lea(dbg, irg, block, base, index, mode_Is);
 489
 490                 /* add the old offset of a previous LEA */
 491                 if (offs) {
 492                         add_ia32_am_offs(res, offs);
 493                 }
 494
 495                 /* add the new offset */
 496                 if (isadd) {
 497                         if (offs_cnst) {
 498                                 add_ia32_am_offs(res, offs_cnst);
 499                         }
 500                         if (offs_lea) {
 501                                 add_ia32_am_offs(res, offs_lea);
 502                         }
 503                 }
 504                 else {
 505                         /* either lea_O-cnst, -cnst or -lea_O  */
 506                         if (offs_cnst) {
 507                                 if (offs_lea) {
 508                                         add_ia32_am_offs(res, offs_lea);
 509                                 }
 510
 511                                 sub_ia32_am_offs(res, offs_cnst);
 512                         }
 513                         else {
 514                                 sub_ia32_am_offs(res, offs_lea);
 515                         }
 516                 }
 517
 518                 /* set scale */
 519                 set_ia32_am_scale(res, scale);
 520
 521                 am_flav = ia32_am_N;
 522                 /* determine new am flavour */
 523                 if (offs || offs_cnst || offs_lea) {
 524                         am_flav |= ia32_O;
 525                 }
 526                 if (! be_is_NoReg(babi, base)) {
 527                         am_flav |= ia32_B;
 528                 }
 529                 if (! be_is_NoReg(babi, index)) {
 530                         am_flav |= ia32_I;
 531                 }
 532                 if (scale > 0) {
 533                         am_flav |= ia32_S;
 534                 }
 535                 set_ia32_am_flavour(res, am_flav);
 536
 537                 set_ia32_op_type(res, ia32_AddrModeS);
 538
 539                 DBG((mod, LEVEL_1, "\tLEA [%+F + %+F * %d + %s]\n", base, index, scale, get_ia32_am_offs(res)));
 540
 541                 /* get the result Proj of the Add/Sub */
 542                 irn = get_res_proj(irn);
 543
 544                 assert(irn && "Couldn't find result proj");
 545
 546                 /* exchange the old op with the new LEA */
 547                 exchange(irn, res);
 548         }
 549
 550         return res;
 551 }
 552
 553 /**
 554  * Optimizes a pattern around irn to address mode if possible.
 555  */
 556 void ia32_optimize_am(ir_node *irn, void *env) {
 557         ia32_code_gen_t   *cg   = env;
 558         ir_graph          *irg  = cg->irg;
 559         firm_dbg_module_t *mod  = cg->mod;
 560         ir_node           *res  = irn;
 561         be_abi_irg_t      *babi = cg->birg->abi;
 562         dbg_info          *dbg;
 563         ir_mode           *mode;
 564         ir_node           *block, *noreg_gp, *noreg_fp;
 565         ir_node           *left, *right, *temp;
 566         ir_node           *store, *load, *mem_proj;
 567         ir_node           *succ, *addr_b, *addr_i;
 568         int                check_am_src = 0;
 569
 570         if (! is_ia32_irn(irn))
 571                 return;
 572
 573         dbg      = get_irn_dbg_info(irn);
 574         mode     = get_irn_mode(irn);
 575         block    = get_nodes_block(irn);
 576         noreg_gp = ia32_new_NoReg_gp(cg);
 577         noreg_fp = ia32_new_NoReg_fp(cg);
 578
 579         DBG((mod, LEVEL_1, "checking for AM\n"));
 580
 581         /* 1st part: check for address calculations and transform the into Lea */
 582
 583         /* Following cases can occur:                                  */
 584         /* - Sub (l, imm) -> LEA [base - offset]                       */
 585         /* - Sub (l, r == LEA with ia32_am_O)   -> LEA [base - offset] */
 586         /* - Add (l, imm) -> LEA [base + offset]                       */
 587         /* - Add (l, r == LEA with ia32_am_O)  -> LEA [base + offset]  */
 588         /* - Add (l == LEA with ia32_am_O, r)  -> LEA [base + offset]  */
 589         /* - Add (l, r) -> LEA [base + index * scale]                  */
 590         /*              with scale > 1 iff l/r == shl (1,2,3)          */
 591
 592         if (is_ia32_Sub(irn) || is_ia32_Add(irn)) {
 593                 left  = get_irn_n(irn, 2);
 594                 right = get_irn_n(irn, 3);
 595
 596             /* Do not try to create a LEA if one of the operands is a Load. */
 597                 if (! pred_is_specific_nodeblock(block, left,  is_ia32_Load)  &&
 598                         ! pred_is_specific_nodeblock(block, right, is_ia32_Load))
 599                 {
 600                         res = fold_addr(babi, irn, mod, noreg_gp);
 601                 }
 602         }
 603
 604         /* 2nd part: fold following patterns:                                               */
 605         /* - Load  -> LEA into Load  } TODO: If the LEA is used by more than one Load/Store */
 606         /* - Store -> LEA into Store }       it might be better to keep the LEA             */
 607         /* - op -> Load into AMop with am_Source                                            */
 608         /*   conditions:                                                                    */
 609         /*     - op is am_Source capable AND                                                */
 610         /*     - the Load is only used by this op AND                                       */
 611         /*     - the Load is in the same block                                              */
 612         /* - Store -> op -> Load  into AMop with am_Dest                                    */
 613         /*   conditions:                                                                    */
 614         /*     - op is am_Dest capable AND                                                  */
 615         /*     - the Store uses the same address as the Load AND                            */
 616         /*     - the Load is only used by this op AND                                       */
 617         /*     - the Load and Store are in the same block AND                               */
 618         /*     - nobody else uses the result of the op                                      */
 619
 620         if ((res == irn) && (get_ia32_am_support(irn) != ia32_am_None) && !is_ia32_Lea(irn)) {
 621                 /* 1st: check for Load/Store -> LEA   */
 622                 if (is_ia32_Ld(irn) || is_ia32_St(irn)) {
 623                         left = get_irn_n(irn, 0);
 624
 625                         if (is_ia32_Lea(left)) {
 626                                 /* get the AM attributes from the LEA */
 627                                 add_ia32_am_offs(irn, get_ia32_am_offs(left));
 628                                 set_ia32_am_scale(irn, get_ia32_am_scale(left));
 629                                 set_ia32_am_flavour(irn, get_ia32_am_flavour(left));
 630                                 set_ia32_op_type(irn, get_ia32_op_type(left));
 631
 632                                 /* set base and index */
 633                                 set_irn_n(irn, 0, get_irn_n(left, 0));
 634                                 set_irn_n(irn, 1, get_irn_n(left, 1));
 635                         }
 636                 }
 637                 /* check if at least one operand is a Load */
 638                 else if (pred_is_specific_nodeblock(block, get_irn_n(irn, 2), is_ia32_Ld) ||
 639                                  pred_is_specific_nodeblock(block, get_irn_n(irn, 3), is_ia32_Ld))
 640                 {
 641                         left  = get_irn_n(irn, 2);
 642                         if (get_irn_arity(irn) == 4) {
 643                                 /* it's an "unary" operation */
 644                                 right = left;
 645                         }
 646                         else {
 647                                 right = get_irn_n(irn, 3);
 648                         }
 649
 650                         /* normalize commutative ops */
 651                         if (node_is_comm(irn)) {
 652                                 /* Assure that right operand is always a Load if there is one    */
 653                                 /* because non-commutative ops can only use Dest AM if the right */
 654                                 /* operand is a load, so we only need to check right operand.    */
 655                                 if (pred_is_specific_nodeblock(block, left, is_ia32_Ld))
 656                                 {
 657                                         set_irn_n(irn, 2, right);
 658                                         set_irn_n(irn, 3, left);
 659
 660                                         temp  = left;
 661                                         left  = right;
 662                                         right = temp;
 663                                 }
 664                         }
 665
 666                         /* check for Store -> op -> Load */
 667
 668                         /* Store -> op -> Load optimization is only possible if supported by op */
 669                         /* and if right operand is a Load                                       */
 670                         if ((get_ia32_am_support(irn) & ia32_am_Dest) &&
 671                                  pred_is_specific_nodeblock(block, right, is_ia32_Ld))
 672                         {
 673
 674                                 /* An address mode capable op always has a result Proj.                  */
 675                                 /* If this Proj is used by more than one other node, we don't need to    */
 676                                 /* check further, otherwise we check for Store and remember the address, */
 677                                 /* the Store points to. */
 678
 679                                 succ = get_res_proj(irn);
 680                                 assert(succ && "Couldn't find result proj");
 681
 682                                 addr_b = NULL;
 683                                 addr_i = NULL;
 684                                 store  = NULL;
 685
 686                                 /* now check for users and Store */
 687                                 if (ia32_get_irn_n_edges(succ) == 1) {
 688                                         succ = get_edge_src_irn(get_irn_out_edge_first(succ));
 689
 690                                         if (is_ia32_fStore(succ) || is_ia32_Store(succ)) {
 691                                                 store  = succ;
 692                                                 addr_b = get_irn_n(store, 0);
 693
 694                                                 /* Could be that the Store is connected to the address    */
 695                                                 /* calculating LEA while the Load is already transformed. */
 696                                                 if (is_ia32_Lea(addr_b)) {
 697                                                         succ   = addr_b;
 698                                                         addr_b = get_irn_n(succ, 0);
 699                                                         addr_i = get_irn_n(succ, 1);
 700                                                 }
 701                                                 else {
 702                                                         addr_i = noreg_gp;
 703                                                 }
 704                                         }
 705                                 }
 706
 707                                 if (store) {
 708                                         /* we found a Store as single user: Now check for Load */
 709
 710                                         /* Extra check for commutative ops with two Loads */
 711                                         /* -> put the interesting Load right              */
 712                                         if (node_is_comm(irn) &&
 713                                                 pred_is_specific_nodeblock(block, left, is_ia32_Ld))
 714                                         {
 715                                                 if ((addr_b == get_irn_n(get_Proj_pred(left), 0)) &&
 716                                                         (addr_i == get_irn_n(get_Proj_pred(left), 1)))
 717                                                 {
 718                                                         /* We exchange left and right, so it's easier to kill     */
 719                                                         /* the correct Load later and to handle unary operations. */
 720                                                         set_irn_n(irn, 2, right);
 721                                                         set_irn_n(irn, 3, left);
 722
 723                                                         temp  = left;
 724                                                         left  = right;
 725                                                         right = temp;
 726                                                 }
 727                                         }
 728
 729                                         /* skip the Proj for easier access */
 730                                         load = get_Proj_pred(right);
 731
 732                                         /* Compare Load and Store address */
 733                                         if ((addr_b == get_irn_n(load, 0)) && (addr_i == get_irn_n(load, 1)))
 734                                         {
 735                                                 /* Right Load is from same address, so we can */
 736                                                 /* disconnect the Load and Store here        */
 737
 738                                                 /* set new base, index and attributes */
 739                                                 set_irn_n(irn, 0, addr_b);
 740                                                 set_irn_n(irn, 1, addr_i);
 741                                                 add_ia32_am_offs(irn, get_ia32_am_offs(load));
 742                                                 set_ia32_am_scale(irn, get_ia32_am_scale(load));
 743                                                 set_ia32_am_flavour(irn, get_ia32_am_flavour(load));
 744                                                 set_ia32_op_type(irn, ia32_AddrModeD);
 745
 746                                                 /* connect to Load memory and disconnect Load */
 747                                                 if (get_irn_arity(irn) == 5) {
 748                                                         /* binary AMop */
 749                                                         set_irn_n(irn, 4, get_irn_n(load, 2));
 750                                                         set_irn_n(irn, 3, noreg_gp);
 751                                                 }
 752                                                 else {
 753                                                         /* unary AMop */
 754                                                         set_irn_n(irn, 3, get_irn_n(load, 2));
 755                                                         set_irn_n(irn, 2, noreg_gp);
 756                                                 }
 757
 758                                                 /* connect the memory Proj of the Store to the op */
 759                                                 mem_proj = get_mem_proj(store);
 760                                                 set_Proj_pred(mem_proj, irn);
 761                                                 set_Proj_proj(mem_proj, 1);
 762                                         }
 763                                 } /* if (store) */
 764                                 else if (get_ia32_am_support(irn) & ia32_am_Source) {
 765                                         /* There was no store, check if we still can optimize for source address mode */
 766                                         check_am_src = 1;
 767                                 }
 768                         } /* if (support AM Dest) */
 769                         else if (get_ia32_am_support(irn) & ia32_am_Source) {
 770                                 /* op doesn't support am AM Dest -> check for AM Source */
 771                                 check_am_src = 1;
 772                         }
 773
 774                         /* normalize commutative ops */
 775                         if (node_is_comm(irn)) {
 776                                 /* Assure that left operand is always a Load if there is one */
 777                                 /* because non-commutative ops can only use Source AM if the */
 778                                 /* left operand is a Load, so we only need to check the left */
 779                                 /* operand afterwards.                                       */
 780                                 if (pred_is_specific_nodeblock(block, right, is_ia32_Ld))       {
 781                                         set_irn_n(irn, 2, right);
 782                                         set_irn_n(irn, 3, left);
 783
 784                                         temp  = left;
 785                                         left  = right;
 786                                         right = temp;
 787                                 }
 788                         }
 789
 790                         /* optimize op -> Load iff Load is only used by this op   */
 791                         /* and left operand is a Load which only used by this irn */
 792                         if (check_am_src                                        &&
 793                                 pred_is_specific_nodeblock(block, left, is_ia32_Ld) &&
 794                                 (ia32_get_irn_n_edges(left) == 1))
 795                         {
 796                                 left = get_Proj_pred(left);
 797
 798                                 addr_b = get_irn_n(left, 0);
 799                                 addr_i = get_irn_n(left, 1);
 800
 801                                 /* set new base, index and attributes */
 802                                 set_irn_n(irn, 0, addr_b);
 803                                 set_irn_n(irn, 1, addr_i);
 804                                 add_ia32_am_offs(irn, get_ia32_am_offs(left));
 805                                 set_ia32_am_scale(irn, get_ia32_am_scale(left));
 806                                 set_ia32_am_flavour(irn, get_ia32_am_flavour(left));
 807                                 set_ia32_op_type(irn, ia32_AddrModeS);
 808
 809                                 /* connect to Load memory */
 810                                 if (get_irn_arity(irn) == 5) {
 811                                         /* binary AMop */
 812                                         set_irn_n(irn, 4, get_irn_n(left, 2));
 813                                 }
 814                                 else {
 815                                         /* unary AMop */
 816                                         set_irn_n(irn, 3, get_irn_n(left, 2));
 817                                 }
 818
 819                                 /* disconnect from Load */
 820                                 set_irn_n(irn, 2, noreg_gp);
 821
 822                                 /* If Load has a memory Proj, connect it to the op */
 823                                 mem_proj = get_mem_proj(left);
 824                                 if (mem_proj) {
 825                                         set_Proj_pred(mem_proj, irn);
 826                                         set_Proj_proj(mem_proj, 1);
 827                                 }
 828                         }
 829                 }
 830         }
 831 }