nsz Git - libfirm/blob - ir/lower/lower_dw.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   Lower double word operations, i.e. 64bit -> 32bit, 32bit -> 16bit etc.
  23  * @date    8.10.2004
  24  * @author  Michael Beck
  25  */
  26 #include "config.h"
  27
  28 #include <string.h>
  29 #include <stdlib.h>
  30 #include <stdbool.h>
  31 #include <assert.h>
  32
  33 #include "be.h"
  34 #include "error.h"
  35 #include "lowering.h"
  36 #include "irnode_t.h"
  37 #include "irnodeset.h"
  38 #include "irgraph_t.h"
  39 #include "irmode_t.h"
  40 #include "iropt_t.h"
  41 #include "irgmod.h"
  42 #include "tv_t.h"
  43 #include "dbginfo_t.h"
  44 #include "iropt_dbg.h"
  45 #include "irflag_t.h"
  46 #include "firmstat.h"
  47 #include "irgwalk.h"
  48 #include "ircons.h"
  49 #include "irflag.h"
  50 #include "iroptimize.h"
  51 #include "irtools.h"
  52 #include "debug.h"
  53 #include "set.h"
  54 #include "pmap.h"
  55 #include "pdeq.h"
  56 #include "irdump.h"
  57 #include "array_t.h"
  58 #include "irpass_t.h"
  59 #include "lower_dw.h"
  60
  61 /** A map from (op, imode, omode) to Intrinsic functions entities. */
  62 static set *intrinsic_fkt;
  63
  64 /** A map from (imode, omode) to conv function types. */
  65 static set *conv_types;
  66
  67 /** A map from a method type to its lowered type. */
  68 static pmap *lowered_type;
  69
  70 /** A map from a builtin type to its lower and higher type. */
  71 static pmap *lowered_builtin_type_high;
  72 static pmap *lowered_builtin_type_low;
  73
  74 /** The types for the binop and unop intrinsics. */
  75 static ir_type *binop_tp_u, *binop_tp_s, *unop_tp_u, *unop_tp_s, *tp_s, *tp_u;
  76
  77 static ir_nodeset_t created_mux_nodes;
  78
  79 /** the debug handle */
  80 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  81
  82 /**
  83  * An entry in the (op, imode, omode) -> entity map.
  84  */
  85 typedef struct op_mode_entry {
  86         const ir_op   *op;    /**< the op */
  87         const ir_mode *imode; /**< the input mode */
  88         const ir_mode *omode; /**< the output mode */
  89         ir_entity     *ent;   /**< the associated entity of this (op, imode, omode) triple */
  90 } op_mode_entry_t;
  91
  92 /**
  93  * An entry in the (imode, omode) -> tp map.
  94  */
  95 typedef struct conv_tp_entry {
  96         const ir_mode *imode; /**< the input mode */
  97         const ir_mode *omode; /**< the output mode */
  98         ir_type       *mtd;   /**< the associated method type of this (imode, omode) pair */
  99 } conv_tp_entry_t;
 100
 101 enum lower_flags {
 102         MUST_BE_LOWERED = 1,  /**< graph must be lowered */
 103         CF_CHANGED      = 2,  /**< control flow was changed */
 104 };
 105
 106 /**
 107  * The lower environment.
 108  */
 109 typedef struct lower_dw_env_t {
 110         lower64_entry_t **entries;     /**< entries per node */
 111         ir_graph      *irg;
 112         struct obstack obst;           /**< an obstack holding the temporary data */
 113         ir_tarval *tv_mode_bytes;      /**< a tarval containing the number of bytes in the lowered modes */
 114         ir_tarval *tv_mode_bits;       /**< a tarval containing the number of bits in the lowered modes */
 115         pdeq      *waitq;              /**< a wait queue of all nodes that must be handled later */
 116         ir_node  **lowered_phis;       /**< list of lowered phis */
 117         ir_mode   *high_signed;        /**< doubleword signed type */
 118         ir_mode   *high_unsigned;      /**< doubleword unsigned type */
 119         ir_mode   *low_signed;         /**< word signed type */
 120         ir_mode   *low_unsigned;       /**< word unsigned type */
 121         ident     *first_id;           /**< .l for little and .h for big endian */
 122         ident     *next_id;            /**< .h for little and .l for big endian */
 123         const lwrdw_param_t *params;   /**< transformation parameter */
 124         unsigned flags;                /**< some flags */
 125         unsigned n_entries;            /**< number of entries */
 126 } lower_dw_env_t;
 127
 128 static lower_dw_env_t *env;
 129
 130 static void lower_node(ir_node *node);
 131
 132 /**
 133  * Create a method type for a Conv emulation from imode to omode.
 134  */
 135 static ir_type *get_conv_type(ir_mode *imode, ir_mode *omode)
 136 {
 137         conv_tp_entry_t key, *entry;
 138         ir_type *mtd;
 139
 140         key.imode = imode;
 141         key.omode = omode;
 142         key.mtd   = NULL;
 143
 144         entry = (conv_tp_entry_t*)set_insert(conv_types, &key, sizeof(key), hash_ptr(imode) ^ hash_ptr(omode));
 145         if (! entry->mtd) {
 146                 int n_param = 1, n_res = 1;
 147
 148                 if (imode == env->high_signed || imode == env->high_unsigned)
 149                         n_param = 2;
 150                 if (omode == env->high_signed || omode == env->high_unsigned)
 151                         n_res = 2;
 152
 153                 /* create a new one */
 154                 mtd = new_type_method(n_param, n_res);
 155
 156                 /* set param types and result types */
 157                 n_param = 0;
 158                 if (imode == env->high_signed) {
 159                         if (env->params->little_endian) {
 160                                 set_method_param_type(mtd, n_param++, tp_u);
 161                                 set_method_param_type(mtd, n_param++, tp_s);
 162                         } else {
 163                                 set_method_param_type(mtd, n_param++, tp_s);
 164                                 set_method_param_type(mtd, n_param++, tp_u);
 165                         }
 166                 } else if (imode == env->high_unsigned) {
 167                         set_method_param_type(mtd, n_param++, tp_u);
 168                         set_method_param_type(mtd, n_param++, tp_u);
 169                 } else {
 170                         ir_type *tp = get_type_for_mode(imode);
 171                         set_method_param_type(mtd, n_param++, tp);
 172                 }
 173
 174                 n_res = 0;
 175                 if (omode == env->high_signed) {
 176                         if (env->params->little_endian) {
 177                                 set_method_res_type(mtd, n_res++, tp_u);
 178                                 set_method_res_type(mtd, n_res++, tp_s);
 179                         } else {
 180                                 set_method_res_type(mtd, n_res++, tp_s);
 181                                 set_method_res_type(mtd, n_res++, tp_u);
 182                         }
 183                 } else if (omode == env->high_unsigned) {
 184                         set_method_res_type(mtd, n_res++, tp_u);
 185                         set_method_res_type(mtd, n_res++, tp_u);
 186                 } else {
 187                         ir_type *tp = get_type_for_mode(omode);
 188                         set_method_res_type(mtd, n_res++, tp);
 189                 }
 190                 entry->mtd = mtd;
 191         } else {
 192                 mtd = entry->mtd;
 193         }
 194         return mtd;
 195 }
 196
 197 /**
 198  * Add an additional control flow input to a block.
 199  * Patch all Phi nodes. The new Phi inputs are copied from
 200  * old input number nr.
 201  */
 202 static void add_block_cf_input_nr(ir_node *block, int nr, ir_node *cf)
 203 {
 204         int i, arity = get_irn_arity(block);
 205         ir_node **in;
 206         const ir_edge_t *edge;
 207
 208         assert(nr < arity);
 209
 210         NEW_ARR_A(ir_node *, in, arity + 1);
 211         for (i = 0; i < arity; ++i)
 212                 in[i] = get_irn_n(block, i);
 213         in[i] = cf;
 214
 215         set_irn_in(block, i + 1, in);
 216
 217         foreach_out_edge(block, edge) {
 218                 ir_node *phi = get_edge_src_irn(edge);
 219                 if (!is_Phi(phi))
 220                         continue;
 221
 222                 for (i = 0; i < arity; ++i)
 223                         in[i] = get_irn_n(phi, i);
 224                 in[i] = in[nr];
 225                 set_irn_in(phi, i + 1, in);
 226         }
 227 }
 228
 229 /**
 230  * Add an additional control flow input to a block.
 231  * Patch all Phi nodes. The new Phi inputs are copied from
 232  * old input from cf tmpl.
 233  */
 234 static void add_block_cf_input(ir_node *block, ir_node *tmpl, ir_node *cf)
 235 {
 236         int i, arity = get_irn_arity(block);
 237         int nr = 0;
 238
 239         for (i = 0; i < arity; ++i) {
 240                 if (get_irn_n(block, i) == tmpl) {
 241                         nr = i;
 242                         break;
 243                 }
 244         }
 245         assert(i < arity);
 246         add_block_cf_input_nr(block, nr, cf);
 247 }
 248
 249 /**
 250  * Return the "operational" mode of a Firm node.
 251  */
 252 static ir_mode *get_irn_op_mode(ir_node *node)
 253 {
 254         switch (get_irn_opcode(node)) {
 255         case iro_Load:
 256                 return get_Load_mode(node);
 257         case iro_Store:
 258                 return get_irn_mode(get_Store_value(node));
 259         case iro_Div:
 260                 return get_irn_mode(get_Div_left(node));
 261         case iro_Mod:
 262                 return get_irn_mode(get_Mod_left(node));
 263         case iro_Cmp:
 264                 return get_irn_mode(get_Cmp_left(node));
 265         default:
 266                 return get_irn_mode(node);
 267         }
 268 }
 269
 270 /**
 271  * Walker, prepare the node links and determine which nodes need to be lowered
 272  * at all.
 273  */
 274 static void prepare_links(ir_node *node)
 275 {
 276         ir_mode         *mode = get_irn_op_mode(node);
 277         lower64_entry_t *link;
 278
 279         if (mode == env->high_signed || mode == env->high_unsigned) {
 280                 unsigned idx = get_irn_idx(node);
 281                 /* ok, found a node that will be lowered */
 282                 link = OALLOCZ(&env->obst, lower64_entry_t);
 283
 284                 if (idx >= env->n_entries) {
 285                         /* enlarge: this happens only for Rotl nodes which is RARELY */
 286                         unsigned old   = env->n_entries;
 287                         unsigned n_idx = idx + (idx >> 3);
 288
 289                         ARR_RESIZE(lower64_entry_t *, env->entries, n_idx);
 290                         memset(&env->entries[old], 0, (n_idx - old) * sizeof(env->entries[0]));
 291                         env->n_entries = n_idx;
 292                 }
 293                 env->entries[idx] = link;
 294                 env->flags |= MUST_BE_LOWERED;
 295         } else if (is_Conv(node)) {
 296                 /* Conv nodes have two modes */
 297                 ir_node *pred = get_Conv_op(node);
 298                 mode = get_irn_mode(pred);
 299
 300                 if (mode == env->high_signed || mode == env->high_unsigned) {
 301                         /* must lower this node either but don't need a link */
 302                         env->flags |= MUST_BE_LOWERED;
 303                 }
 304                 return;
 305         } else if (is_Call(node)) {
 306                 /* Special case:  If the result of the Call is never used, we won't
 307                  * find a Proj with a mode that potentially triggers MUST_BE_LOWERED
 308                  * to be set.  Thus, if we see a call, we check its result types and
 309                  * decide whether MUST_BE_LOWERED has to be set.
 310                  */
 311                 ir_type *tp = get_Call_type(node);
 312                 size_t   n_res, i;
 313
 314                 n_res = get_method_n_ress(tp);
 315                 for (i = 0; i < n_res; ++i) {
 316                         ir_type *rtp = get_method_res_type(tp, i);
 317
 318                         if (is_Primitive_type(rtp)) {
 319                                 ir_mode *rmode = get_type_mode(rtp);
 320
 321                                 if (rmode == env->high_signed || rmode == env->high_unsigned) {
 322                                         env->flags |= MUST_BE_LOWERED;
 323                                 }
 324                         }
 325                 }
 326         }
 327 }
 328
 329 lower64_entry_t *get_node_entry(ir_node *node)
 330 {
 331         unsigned idx = get_irn_idx(node);
 332         assert(idx < env->n_entries);
 333         return env->entries[idx];
 334 }
 335
 336 void ir_set_dw_lowered(ir_node *old, ir_node *new_low, ir_node *new_high)
 337 {
 338         lower64_entry_t *entry = get_node_entry(old);
 339         entry->low_word  = new_low;
 340         entry->high_word = new_high;
 341 }
 342
 343 ir_mode *ir_get_low_unsigned_mode(void)
 344 {
 345         return env->low_unsigned;
 346 }
 347
 348 /**
 349  * Translate a Constant: create two.
 350  */
 351 static void lower_Const(ir_node *node, ir_mode *mode)
 352 {
 353         ir_graph  *irg      = get_irn_irg(node);
 354         dbg_info  *dbg      = get_irn_dbg_info(node);
 355         ir_mode   *low_mode = env->low_unsigned;
 356         ir_tarval *tv       = get_Const_tarval(node);
 357         ir_tarval *tv_l     = tarval_convert_to(tv, low_mode);
 358         ir_node   *res_low  = new_rd_Const(dbg, irg, tv_l);
 359         ir_tarval *tv_shrs  = tarval_shrs(tv, env->tv_mode_bits);
 360         ir_tarval *tv_h     = tarval_convert_to(tv_shrs, mode);
 361         ir_node   *res_high = new_rd_Const(dbg, irg, tv_h);
 362
 363         ir_set_dw_lowered(node, res_low, res_high);
 364 }
 365
 366 /**
 367  * Translate a Load: create two.
 368  */
 369 static void lower_Load(ir_node *node, ir_mode *mode)
 370 {
 371         ir_mode    *low_mode = env->low_unsigned;
 372         ir_graph   *irg = get_irn_irg(node);
 373         ir_node    *adr = get_Load_ptr(node);
 374         ir_node    *mem = get_Load_mem(node);
 375         ir_node    *low;
 376         ir_node    *high;
 377         ir_node    *proj_m;
 378         dbg_info   *dbg;
 379         ir_node    *block = get_nodes_block(node);
 380         ir_cons_flags volatility = get_Load_volatility(node) == volatility_is_volatile
 381                                  ? cons_volatile : cons_none;
 382         const ir_edge_t *edge;
 383         const ir_edge_t *next;
 384
 385         if (env->params->little_endian) {
 386                 low  = adr;
 387                 high = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
 388         } else {
 389                 low  = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
 390                 high = adr;
 391         }
 392
 393         /* create two loads */
 394         dbg    = get_irn_dbg_info(node);
 395         low    = new_rd_Load(dbg, block, mem,  low,  low_mode, volatility);
 396         proj_m = new_r_Proj(low, mode_M, pn_Load_M);
 397         high   = new_rd_Load(dbg, block, proj_m, high, mode, volatility);
 398
 399         foreach_out_edge_safe(node, edge, next) {
 400                 ir_node *proj = get_edge_src_irn(edge);
 401                 if (!is_Proj(proj))
 402                         continue;
 403
 404                 switch (get_Proj_proj(proj)) {
 405                 case pn_Load_M:         /* Memory result. */
 406                         /* put it to the second one */
 407                         set_Proj_pred(proj, high);
 408                         break;
 409                 case pn_Load_X_except:  /* Execution result if exception occurred. */
 410                         /* put it to the first one */
 411                         set_Proj_pred(proj, low);
 412                         break;
 413                 case pn_Load_res: {       /* Result of load operation. */
 414                         ir_node *res_low  = new_r_Proj(low,  low_mode, pn_Load_res);
 415                         ir_node *res_high = new_r_Proj(high, mode,     pn_Load_res);
 416                         ir_set_dw_lowered(proj, res_low, res_high);
 417                         break;
 418                 }
 419                 default:
 420                         assert(0 && "unexpected Proj number");
 421                 }
 422                 /* mark this proj: we have handled it already, otherwise we might fall
 423                  * into out new nodes. */
 424                 mark_irn_visited(proj);
 425         }
 426 }
 427
 428 /**
 429  * Translate a Store: create two.
 430  */
 431 static void lower_Store(ir_node *node, ir_mode *mode)
 432 {
 433         ir_graph              *irg;
 434         ir_node               *block, *adr, *mem;
 435         ir_node               *low, *high, *proj_m;
 436         dbg_info              *dbg;
 437         ir_node               *value = get_Store_value(node);
 438         const lower64_entry_t *entry = get_node_entry(value);
 439         ir_cons_flags volatility = get_Store_volatility(node) == volatility_is_volatile
 440                                    ? cons_volatile : cons_none;
 441         const ir_edge_t *edge;
 442         const ir_edge_t *next;
 443         (void) mode;
 444
 445         assert(entry);
 446
 447         if (! entry->low_word) {
 448                 /* not ready yet, wait */
 449                 pdeq_putr(env->waitq, node);
 450                 return;
 451         }
 452
 453         irg = get_irn_irg(node);
 454         adr = get_Store_ptr(node);
 455         mem = get_Store_mem(node);
 456         block = get_nodes_block(node);
 457
 458         if (env->params->little_endian) {
 459                 low  = adr;
 460                 high = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
 461         } else {
 462                 low  = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
 463                 high = adr;
 464         }
 465
 466         /* create two Stores */
 467         dbg    = get_irn_dbg_info(node);
 468         low    = new_rd_Store(dbg, block, mem, low,  entry->low_word, volatility);
 469         proj_m = new_r_Proj(low, mode_M, pn_Store_M);
 470         high   = new_rd_Store(dbg, block, proj_m, high, entry->high_word, volatility);
 471
 472         foreach_out_edge_safe(node, edge, next) {
 473                 ir_node *proj = get_edge_src_irn(edge);
 474                 if (!is_Proj(proj))
 475                         continue;
 476
 477                 switch (get_Proj_proj(proj)) {
 478                 case pn_Store_M:         /* Memory result. */
 479                         /* put it to the second one */
 480                         set_Proj_pred(proj, high);
 481                         break;
 482                 case pn_Store_X_except:  /* Execution result if exception occurred. */
 483                         /* put it to the first one */
 484                         set_Proj_pred(proj, low);
 485                         break;
 486                 default:
 487                         assert(0 && "unexpected Proj number");
 488                 }
 489                 /* mark this proj: we have handled it already, otherwise we might fall into
 490                  * out new nodes. */
 491                 mark_irn_visited(proj);
 492         }
 493 }
 494
 495 /**
 496  * Return a node containing the address of the intrinsic emulation function.
 497  *
 498  * @param method  the method type of the emulation function
 499  * @param op      the emulated ir_op
 500  * @param imode   the input mode of the emulated opcode
 501  * @param omode   the output mode of the emulated opcode
 502  * @param env     the lower environment
 503  */
 504 static ir_node *get_intrinsic_address(ir_type *method, ir_op *op,
 505                                       ir_mode *imode, ir_mode *omode)
 506 {
 507         symconst_symbol sym;
 508         ir_entity *ent;
 509         op_mode_entry_t key, *entry;
 510
 511         key.op    = op;
 512         key.imode = imode;
 513         key.omode = omode;
 514         key.ent   = NULL;
 515
 516         entry = (op_mode_entry_t*)set_insert(intrinsic_fkt, &key, sizeof(key),
 517                                 hash_ptr(op) ^ hash_ptr(imode) ^ (hash_ptr(omode) << 8));
 518         if (! entry->ent) {
 519                 /* create a new one */
 520                 ent = env->params->create_intrinsic(method, op, imode, omode, env->params->ctx);
 521
 522                 assert(ent && "Intrinsic creator must return an entity");
 523                 entry->ent = ent;
 524         } else {
 525                 ent = entry->ent;
 526         }
 527         sym.entity_p = ent;
 528         return new_r_SymConst(env->irg, mode_P_code, sym, symconst_addr_ent);
 529 }
 530
 531 /**
 532  * Translate a Div.
 533  *
 534  * Create an intrinsic Call.
 535  */
 536 static void lower_Div(ir_node *node, ir_mode *mode)
 537 {
 538         ir_node         *left   = get_Div_left(node);
 539         ir_node         *right  = get_Div_right(node);
 540         ir_node         *block  = get_nodes_block(node);
 541         dbg_info        *dbgi   = get_irn_dbg_info(node);
 542         ir_type         *mtp    = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
 543         ir_mode         *opmode = get_irn_op_mode(node);
 544         ir_node         *addr
 545             = get_intrinsic_address(mtp, get_irn_op(node), opmode, opmode);
 546         ir_node         *in[4];
 547         ir_node         *call;
 548         ir_node         *resproj;
 549         const ir_edge_t *edge;
 550         const ir_edge_t *next;
 551
 552         if (env->params->little_endian) {
 553                 in[0] = get_lowered_low(left);
 554                 in[1] = get_lowered_high(left);
 555                 in[2] = get_lowered_low(right);
 556                 in[3] = get_lowered_high(right);
 557         } else {
 558                 in[0] = get_lowered_high(left);
 559                 in[1] = get_lowered_low(left);
 560                 in[2] = get_lowered_high(right);
 561                 in[3] = get_lowered_low(right);
 562         }
 563         call    = new_rd_Call(dbgi, block, get_Div_mem(node), addr, 4, in, mtp);
 564         resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
 565         set_irn_pinned(call, get_irn_pinned(node));
 566
 567         foreach_out_edge_safe(node, edge, next) {
 568                 ir_node *proj = get_edge_src_irn(edge);
 569                 if (!is_Proj(proj))
 570                         continue;
 571
 572                 switch (get_Proj_proj(proj)) {
 573                 case pn_Div_M:         /* Memory result. */
 574                         /* reroute to the call */
 575                         set_Proj_pred(proj, call);
 576                         set_Proj_proj(proj, pn_Call_M);
 577                         break;
 578                 case pn_Div_X_regular:
 579                         set_Proj_pred(proj, call);
 580                         set_Proj_proj(proj, pn_Call_X_regular);
 581                         break;
 582                 case pn_Div_X_except:
 583                         set_Proj_pred(proj, call);
 584                         set_Proj_proj(proj, pn_Call_X_except);
 585                         break;
 586                 case pn_Div_res:
 587                         if (env->params->little_endian) {
 588                                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
 589                                 ir_node *res_high = new_r_Proj(resproj, mode,              1);
 590                                 ir_set_dw_lowered(proj, res_low, res_high);
 591                         } else {
 592                                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
 593                                 ir_node *res_high = new_r_Proj(resproj, mode,              0);
 594                                 ir_set_dw_lowered(proj, res_low, res_high);
 595                         }
 596                         break;
 597                 default:
 598                         assert(0 && "unexpected Proj number");
 599                 }
 600                 /* mark this proj: we have handled it already, otherwise we might fall into
 601                  * out new nodes. */
 602                 mark_irn_visited(proj);
 603         }
 604 }
 605
 606 /**
 607  * Translate a Mod.
 608  *
 609  * Create an intrinsic Call.
 610  */
 611 static void lower_Mod(ir_node *node, ir_mode *mode)
 612 {
 613         ir_node         *left   = get_Mod_left(node);
 614         ir_node         *right  = get_Mod_right(node);
 615         dbg_info        *dbgi   = get_irn_dbg_info(node);
 616         ir_node         *block  = get_nodes_block(node);
 617         ir_type         *mtp    = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
 618         ir_mode         *opmode = get_irn_op_mode(node);
 619         ir_node         *addr
 620             = get_intrinsic_address(mtp, get_irn_op(node), opmode, opmode);
 621         ir_node         *in[4];
 622         ir_node         *call;
 623         ir_node         *resproj;
 624         const ir_edge_t *edge;
 625         const ir_edge_t *next;
 626
 627         if (env->params->little_endian) {
 628                 in[0] = get_lowered_low(left);
 629                 in[1] = get_lowered_high(left);
 630                 in[2] = get_lowered_low(right);
 631                 in[3] = get_lowered_high(right);
 632         } else {
 633                 in[0] = get_lowered_high(left);
 634                 in[1] = get_lowered_low(left);
 635                 in[2] = get_lowered_high(right);
 636                 in[3] = get_lowered_low(right);
 637         }
 638         call    = new_rd_Call(dbgi, block, get_Mod_mem(node), addr, 4, in, mtp);
 639         resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
 640         set_irn_pinned(call, get_irn_pinned(node));
 641
 642         foreach_out_edge_safe(node, edge, next) {
 643                 ir_node *proj = get_edge_src_irn(edge);
 644                 if (!is_Proj(proj))
 645                         continue;
 646
 647                 switch (get_Proj_proj(proj)) {
 648                 case pn_Mod_M:         /* Memory result. */
 649                         /* reroute to the call */
 650                         set_Proj_pred(proj, call);
 651                         set_Proj_proj(proj, pn_Call_M);
 652                         break;
 653                 case pn_Div_X_regular:
 654                         set_Proj_pred(proj, call);
 655                         set_Proj_proj(proj, pn_Call_X_regular);
 656                         break;
 657                 case pn_Mod_X_except:
 658                         set_Proj_pred(proj, call);
 659                         set_Proj_proj(proj, pn_Call_X_except);
 660                         break;
 661                 case pn_Mod_res:
 662                         if (env->params->little_endian) {
 663                                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
 664                                 ir_node *res_high = new_r_Proj(resproj, mode,              1);
 665                                 ir_set_dw_lowered(proj, res_low, res_high);
 666                         } else {
 667                                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
 668                                 ir_node *res_high = new_r_Proj(resproj, mode,              0);
 669                                 ir_set_dw_lowered(proj, res_low, res_high);
 670                         }
 671                         break;
 672                 default:
 673                         assert(0 && "unexpected Proj number");
 674                 }
 675                 /* mark this proj: we have handled it already, otherwise we might fall
 676                  * into out new nodes. */
 677                 mark_irn_visited(proj);
 678         }
 679 }
 680
 681 /**
 682  * Translate a binop.
 683  *
 684  * Create an intrinsic Call.
 685  */
 686 static void lower_binop(ir_node *node, ir_mode *mode)
 687 {
 688         ir_node  *left  = get_binop_left(node);
 689         ir_node  *right = get_binop_right(node);
 690         dbg_info *dbgi  = get_irn_dbg_info(node);
 691         ir_node  *block = get_nodes_block(node);
 692         ir_graph *irg   = get_irn_irg(block);
 693         ir_type  *mtp   = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
 694         ir_node  *addr  = get_intrinsic_address(mtp, get_irn_op(node), mode, mode);
 695         ir_node  *in[4];
 696         ir_node  *call;
 697         ir_node  *resproj;
 698
 699         if (env->params->little_endian) {
 700                 in[0] = get_lowered_low(left);
 701                 in[1] = get_lowered_high(left);
 702                 in[2] = get_lowered_low(right);
 703                 in[3] = get_lowered_high(right);
 704         } else {
 705                 in[0] = get_lowered_high(left);
 706                 in[1] = get_lowered_low(left);
 707                 in[2] = get_lowered_high(right);
 708                 in[3] = get_lowered_low(right);
 709         }
 710         call    = new_rd_Call(dbgi, block, get_irg_no_mem(irg), addr, 4, in, mtp);
 711         resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
 712         set_irn_pinned(call, get_irn_pinned(node));
 713
 714         if (env->params->little_endian) {
 715                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
 716                 ir_node *res_high = new_r_Proj(resproj, mode,              1);
 717                 ir_set_dw_lowered(node, res_low, res_high);
 718         } else {
 719                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
 720                 ir_node *res_high = new_r_Proj(resproj, mode,              0);
 721                 ir_set_dw_lowered(node, res_low, res_high);
 722         }
 723 }
 724
 725 static ir_node *create_conv(ir_node *block, ir_node *node, ir_mode *dest_mode)
 726 {
 727         if (get_irn_mode(node) == dest_mode)
 728                 return node;
 729         return new_r_Conv(block, node, dest_mode);
 730 }
 731
 732 /**
 733  * Moves node and all predecessors of node from from_bl to to_bl.
 734  * Does not move predecessors of Phi nodes (or block nodes).
 735  */
 736 static void move(ir_node *node, ir_node *from_bl, ir_node *to_bl)
 737 {
 738         int i, arity;
 739
 740         /* move this node */
 741         set_nodes_block(node, to_bl);
 742
 743         /* move its Projs */
 744         if (get_irn_mode(node) == mode_T) {
 745                 const ir_edge_t *edge;
 746                 foreach_out_edge(node, edge) {
 747                         ir_node *proj = get_edge_src_irn(edge);
 748                         if (!is_Proj(proj))
 749                                 continue;
 750                         move(proj, from_bl, to_bl);
 751                 }
 752         }
 753
 754         /* We must not move predecessors of Phi nodes, even if they are in
 755          * from_bl. (because these are values from an earlier loop iteration
 756          * which are not predecessors of node here)
 757          */
 758         if (is_Phi(node))
 759                 return;
 760
 761         /* recursion ... */
 762         arity = get_irn_arity(node);
 763         for (i = 0; i < arity; i++) {
 764                 ir_node *pred      = get_irn_n(node, i);
 765                 ir_mode *pred_mode = get_irn_mode(pred);
 766                 if (get_nodes_block(pred) == from_bl)
 767                         move(pred, from_bl, to_bl);
 768                 if (pred_mode == env->high_signed || pred_mode == env->high_unsigned) {
 769                         ir_node *pred_low  = get_lowered_low(pred);
 770                         ir_node *pred_high = get_lowered_high(pred);
 771                         if (get_nodes_block(pred_low) == from_bl)
 772                                 move(pred_low, from_bl, to_bl);
 773                         if (pred_high != NULL && get_nodes_block(pred_high) == from_bl)
 774                                 move(pred_high, from_bl, to_bl);
 775                 }
 776         }
 777 }
 778
 779 /**
 780  * We need a custom version of part_block_edges because during transformation
 781  * not all data-dependencies are explicit yet if a lowered nodes users are not
 782  * lowered yet.
 783  * We can fix this by modifying move to look for such implicit dependencies.
 784  * Additionally we have to keep the proj_2_block map updated
 785  */
 786 static ir_node *part_block_dw(ir_node *node)
 787 {
 788         ir_graph *irg        = get_irn_irg(node);
 789         ir_node  *old_block  = get_nodes_block(node);
 790         int       n_cfgpreds = get_Block_n_cfgpreds(old_block);
 791         ir_node **cfgpreds   = get_Block_cfgpred_arr(old_block);
 792         ir_node  *new_block  = new_r_Block(irg, n_cfgpreds, cfgpreds);
 793         const ir_edge_t *edge;
 794         const ir_edge_t *next;
 795
 796         /* old_block has no predecessors anymore for now */
 797         set_irn_in(old_block, 0, NULL);
 798
 799         /* move node and its predecessors to new_block */
 800         move(node, old_block, new_block);
 801
 802         /* move Phi nodes to new_block */
 803         foreach_out_edge_safe(old_block, edge, next) {
 804                 ir_node *phi = get_edge_src_irn(edge);
 805                 if (!is_Phi(phi))
 806                         continue;
 807                 set_nodes_block(phi, new_block);
 808         }
 809         return old_block;
 810 }
 811
 812 typedef ir_node* (*new_rd_shr_func)(dbg_info *dbgi, ir_node *block,
 813                                     ir_node *left, ir_node *right,
 814                                     ir_mode *mode);
 815
 816 static void lower_shr_helper(ir_node *node, ir_mode *mode,
 817                              new_rd_shr_func new_rd_shrs)
 818 {
 819         ir_node  *right         = get_binop_right(node);
 820         ir_node  *left          = get_binop_left(node);
 821         ir_mode  *shr_mode      = get_irn_mode(node);
 822         unsigned  modulo_shift  = get_mode_modulo_shift(shr_mode);
 823         ir_mode  *low_unsigned  = env->low_unsigned;
 824         unsigned  modulo_shift2 = get_mode_modulo_shift(mode);
 825         ir_graph *irg           = get_irn_irg(node);
 826         ir_node  *left_low      = get_lowered_low(left);
 827         ir_node  *left_high     = get_lowered_high(left);
 828         dbg_info *dbgi          = get_irn_dbg_info(node);
 829         ir_node  *lower_block;
 830         ir_node  *block;
 831         ir_node  *cnst;
 832         ir_node  *andn;
 833         ir_node  *cmp;
 834         ir_node  *cond;
 835         ir_node  *proj_true;
 836         ir_node  *proj_false;
 837         ir_node  *phi_low;
 838         ir_node  *phi_high;
 839         ir_node  *lower_in[2];
 840         ir_node  *phi_low_in[2];
 841         ir_node  *phi_high_in[2];
 842
 843         /* this version is optimized for modulo shift architectures
 844          * (and can't handle anything else) */
 845         if (modulo_shift != get_mode_size_bits(shr_mode)
 846                         || modulo_shift2<<1 != modulo_shift) {
 847                 panic("Shr lowering only implemented for modulo shift shr operations");
 848         }
 849         if (!is_po2(modulo_shift) || !is_po2(modulo_shift2)) {
 850                 panic("Shr lowering only implemented for power-of-2 modes");
 851         }
 852         /* without 2-complement the -x instead of (bit_width-x) trick won't work */
 853         if (get_mode_arithmetic(shr_mode) != irma_twos_complement) {
 854                 panic("Shr lowering only implemented for two-complement modes");
 855         }
 856
 857         block = get_nodes_block(node);
 858
 859         /* if the right operand is a 64bit value, we're only interested in the
 860          * lower word */
 861         if (get_irn_mode(right) == env->high_unsigned) {
 862                 right = get_lowered_low(right);
 863         } else {
 864                 /* shift should never have signed mode on the right */
 865                 assert(get_irn_mode(right) != env->high_signed);
 866                 right = create_conv(block, right, low_unsigned);
 867         }
 868
 869         lower_block = part_block_dw(node);
 870         env->flags |= CF_CHANGED;
 871         block = get_nodes_block(node);
 872
 873         /* add a Cmp to test if highest bit is set <=> whether we shift more
 874          * than half the word width */
 875         cnst       = new_r_Const_long(irg, low_unsigned, modulo_shift2);
 876         andn       = new_r_And(block, right, cnst, low_unsigned);
 877         cnst       = new_r_Const(irg, get_mode_null(low_unsigned));
 878         cmp        = new_rd_Cmp(dbgi, block, andn, cnst, ir_relation_equal);
 879         cond       = new_rd_Cond(dbgi, block, cmp);
 880         proj_true  = new_r_Proj(cond, mode_X, pn_Cond_true);
 881         proj_false = new_r_Proj(cond, mode_X, pn_Cond_false);
 882
 883         /* the true block => shift_width < 1word */
 884         {
 885                 /* In theory the low value (for 64bit shifts) is:
 886                  *    Or(High << (32-x)), Low >> x)
 887                  * In practice High << 32-x will fail when x is zero (since we have
 888                  * modulo shift and 32 will be 0). So instead we use:
 889                  *    Or(High<<1<<~x, Low >> x)
 890                  */
 891                 ir_node *in[1]        = { proj_true };
 892                 ir_node *block_true   = new_r_Block(irg, ARRAY_SIZE(in), in);
 893                 ir_node *res_high     = new_rd_shrs(dbgi, block_true, left_high,
 894                                                     right, mode);
 895                 ir_node *shift_low    = new_rd_Shr(dbgi, block_true, left_low, right,
 896                                                    low_unsigned);
 897                 ir_node *not_shiftval = new_rd_Not(dbgi, block_true, right,
 898                                                    low_unsigned);
 899                 ir_node *conv         = create_conv(block_true, left_high,
 900                                                     low_unsigned);
 901                 ir_node *one          = new_r_Const(irg, get_mode_one(low_unsigned));
 902                 ir_node *carry0       = new_rd_Shl(dbgi, block_true, conv, one,
 903                                                    low_unsigned);
 904                 ir_node *carry1       = new_rd_Shl(dbgi, block_true, carry0,
 905                                                    not_shiftval, low_unsigned);
 906                 ir_node *res_low      = new_rd_Or(dbgi, block_true, shift_low, carry1,
 907                                                   low_unsigned);
 908                 lower_in[0]           = new_r_Jmp(block_true);
 909                 phi_low_in[0]         = res_low;
 910                 phi_high_in[0]        = res_high;
 911         }
 912
 913         /* false block => shift_width > 1word */
 914         {
 915                 ir_node *in[1]       = { proj_false };
 916                 ir_node *block_false = new_r_Block(irg, ARRAY_SIZE(in), in);
 917                 ir_node *conv        = create_conv(block_false, left_high, low_unsigned);
 918                 ir_node *res_low     = new_rd_shrs(dbgi, block_false, conv, right,
 919                                                    low_unsigned);
 920                 int      cnsti       = modulo_shift2-1;
 921                 ir_node *cnst2       = new_r_Const_long(irg, low_unsigned, cnsti);
 922                 ir_node *res_high;
 923                 if (new_rd_shrs == new_rd_Shrs) {
 924                         res_high = new_rd_shrs(dbgi, block_false, left_high, cnst2, mode);
 925                 } else {
 926                         res_high = new_r_Const(irg, get_mode_null(mode));
 927                 }
 928                 lower_in[1]          = new_r_Jmp(block_false);
 929                 phi_low_in[1]        = res_low;
 930                 phi_high_in[1]       = res_high;
 931         }
 932
 933         /* patch lower block */
 934         set_irn_in(lower_block, ARRAY_SIZE(lower_in), lower_in);
 935         phi_low  = new_r_Phi(lower_block, ARRAY_SIZE(phi_low_in), phi_low_in,
 936                              low_unsigned);
 937         phi_high = new_r_Phi(lower_block, ARRAY_SIZE(phi_high_in), phi_high_in,
 938                              mode);
 939         ir_set_dw_lowered(node, phi_low, phi_high);
 940 }
 941
 942 static void lower_Shr(ir_node *node, ir_mode *mode)
 943 {
 944         lower_shr_helper(node, mode, new_rd_Shr);
 945 }
 946
 947 static void lower_Shrs(ir_node *node, ir_mode *mode)
 948 {
 949         lower_shr_helper(node, mode, new_rd_Shrs);
 950 }
 951
 952 static void lower_Shl(ir_node *node, ir_mode *mode)
 953 {
 954         ir_node  *right         = get_binop_right(node);
 955         ir_node  *left          = get_binop_left(node);
 956         ir_mode  *shr_mode      = get_irn_mode(node);
 957         unsigned  modulo_shift  = get_mode_modulo_shift(shr_mode);
 958         ir_mode  *low_unsigned  = env->low_unsigned;
 959         unsigned  modulo_shift2 = get_mode_modulo_shift(mode);
 960         ir_graph *irg           = get_irn_irg(node);
 961         ir_node  *left_low      = get_lowered_low(left);
 962         ir_node  *left_high     = get_lowered_high(left);
 963         dbg_info *dbgi          = get_irn_dbg_info(node);
 964         ir_node  *lower_block   = get_nodes_block(node);
 965         ir_node  *block;
 966         ir_node  *cnst;
 967         ir_node  *andn;
 968         ir_node  *cmp;
 969         ir_node  *cond;
 970         ir_node  *proj_true;
 971         ir_node  *proj_false;
 972         ir_node  *phi_low;
 973         ir_node  *phi_high;
 974         ir_node  *lower_in[2];
 975         ir_node  *phi_low_in[2];
 976         ir_node  *phi_high_in[2];
 977
 978         /* this version is optimized for modulo shift architectures
 979          * (and can't handle anything else) */
 980         if (modulo_shift != get_mode_size_bits(shr_mode)
 981                         || modulo_shift2<<1 != modulo_shift) {
 982                 panic("Shl lowering only implemented for modulo shift shl operations");
 983         }
 984         if (!is_po2(modulo_shift) || !is_po2(modulo_shift2)) {
 985                 panic("Shl lowering only implemented for power-of-2 modes");
 986         }
 987         /* without 2-complement the -x instead of (bit_width-x) trick won't work */
 988         if (get_mode_arithmetic(shr_mode) != irma_twos_complement) {
 989                 panic("Shl lowering only implemented for two-complement modes");
 990         }
 991
 992         /* if the right operand is a 64bit value, we're only interested in the
 993          * lower word */
 994         if (get_irn_mode(right) == env->high_unsigned) {
 995                 right = get_lowered_low(right);
 996         } else {
 997                 /* shift should never have signed mode on the right */
 998                 assert(get_irn_mode(right) != env->high_signed);
 999                 right = create_conv(lower_block, right, low_unsigned);
1000         }
1001
1002         part_block_dw(node);
1003         env->flags |= CF_CHANGED;
1004         block = get_nodes_block(node);
1005
1006         /* add a Cmp to test if highest bit is set <=> whether we shift more
1007          * than half the word width */
1008         cnst       = new_r_Const_long(irg, low_unsigned, modulo_shift2);
1009         andn       = new_r_And(block, right, cnst, low_unsigned);
1010         cnst       = new_r_Const(irg, get_mode_null(low_unsigned));
1011         cmp        = new_rd_Cmp(dbgi, block, andn, cnst, ir_relation_equal);
1012         cond       = new_rd_Cond(dbgi, block, cmp);
1013         proj_true  = new_r_Proj(cond, mode_X, pn_Cond_true);
1014         proj_false = new_r_Proj(cond, mode_X, pn_Cond_false);
1015
1016         /* the true block => shift_width < 1word */
1017         {
1018                 ir_node *in[1]        = { proj_true };
1019                 ir_node *block_true   = new_r_Block(irg, ARRAY_SIZE(in), in);
1020
1021                 ir_node *res_low      = new_rd_Shl(dbgi, block_true, left_low,
1022                                                    right, low_unsigned);
1023                 ir_node *shift_high   = new_rd_Shl(dbgi, block_true, left_high, right,
1024                                                    mode);
1025                 ir_node *not_shiftval = new_rd_Not(dbgi, block_true, right,
1026                                                    low_unsigned);
1027                 ir_node *conv         = create_conv(block_true, left_low, mode);
1028                 ir_node *one          = new_r_Const(irg, get_mode_one(low_unsigned));
1029                 ir_node *carry0       = new_rd_Shr(dbgi, block_true, conv, one, mode);
1030                 ir_node *carry1       = new_rd_Shr(dbgi, block_true, carry0,
1031                                                    not_shiftval, mode);
1032                 ir_node *res_high     = new_rd_Or(dbgi, block_true, shift_high, carry1,
1033                                                   mode);
1034                 lower_in[0]           = new_r_Jmp(block_true);
1035                 phi_low_in[0]         = res_low;
1036                 phi_high_in[0]        = res_high;
1037         }
1038
1039         /* false block => shift_width > 1word */
1040         {
1041                 ir_node *in[1]       = { proj_false };
1042                 ir_node *block_false = new_r_Block(irg, ARRAY_SIZE(in), in);
1043                 ir_node *res_low     = new_r_Const(irg, get_mode_null(low_unsigned));
1044                 ir_node *conv        = create_conv(block_false, left_low, mode);
1045                 ir_node *res_high    = new_rd_Shl(dbgi, block_false, conv, right, mode);
1046                 lower_in[1]          = new_r_Jmp(block_false);
1047                 phi_low_in[1]        = res_low;
1048                 phi_high_in[1]       = res_high;
1049         }
1050
1051         /* patch lower block */
1052         set_irn_in(lower_block, ARRAY_SIZE(lower_in), lower_in);
1053         phi_low  = new_r_Phi(lower_block, ARRAY_SIZE(phi_low_in), phi_low_in,
1054                              low_unsigned);
1055         phi_high = new_r_Phi(lower_block, ARRAY_SIZE(phi_high_in), phi_high_in,
1056                              mode);
1057         ir_set_dw_lowered(node, phi_low, phi_high);
1058 }
1059
1060 /**
1061  * Rebuild Rotl nodes into Or(Shl, Shr) and prepare all nodes.
1062  */
1063 static void prepare_links_and_handle_rotl(ir_node *node, void *data)
1064 {
1065         (void) data;
1066         if (is_Rotl(node)) {
1067                 ir_mode  *mode = get_irn_op_mode(node);
1068                 ir_node  *right;
1069                 ir_node  *left, *shl, *shr, *ornode, *block, *sub, *c;
1070                 ir_mode  *omode, *rmode;
1071                 ir_graph *irg;
1072                 dbg_info *dbg;
1073                 optimization_state_t state;
1074
1075                 if (mode != env->high_signed && mode != env->high_unsigned) {
1076                         prepare_links(node);
1077                         return;
1078                 }
1079
1080                 /* replace the Rotl(x,y) by an Or(Shl(x,y), Shr(x,64-y)) */
1081                 right = get_Rotl_right(node);
1082                 irg   = get_irn_irg(node);
1083                 dbg   = get_irn_dbg_info(node);
1084                 omode = get_irn_mode(node);
1085                 left  = get_Rotl_left(node);
1086                 block = get_nodes_block(node);
1087                 shl   = new_rd_Shl(dbg, block, left, right, omode);
1088                 rmode = get_irn_mode(right);
1089                 c     = new_r_Const_long(irg, rmode, get_mode_size_bits(omode));
1090                 sub   = new_rd_Sub(dbg, block, c, right, rmode);
1091                 shr   = new_rd_Shr(dbg, block, left, sub, omode);
1092
1093                 /* switch optimization off here, or we will get the Rotl back */
1094                 save_optimization_state(&state);
1095                 set_opt_algebraic_simplification(0);
1096                 ornode = new_rd_Or(dbg, block, shl, shr, omode);
1097                 restore_optimization_state(&state);
1098
1099                 exchange(node, ornode);
1100
1101                 /* do lowering on the new nodes */
1102                 prepare_links(shl);
1103                 prepare_links(c);
1104                 prepare_links(sub);
1105                 prepare_links(shr);
1106                 prepare_links(ornode);
1107                 return;
1108         }
1109
1110         prepare_links(node);
1111 }
1112
1113 /**
1114  * Translate an Unop.
1115  *
1116  * Create an intrinsic Call.
1117  */
1118 static void lower_unop(ir_node *node, ir_mode *mode)
1119 {
1120         ir_node  *op       = get_unop_op(node);
1121         dbg_info *dbgi     = get_irn_dbg_info(node);
1122         ir_node  *block    = get_nodes_block(node);
1123         ir_graph *irg      = get_irn_irg(block);
1124         ir_type  *mtp      = mode_is_signed(mode) ? unop_tp_s : unop_tp_u;
1125         ir_op    *irop     = get_irn_op(node);
1126         ir_node  *addr     = get_intrinsic_address(mtp, irop, mode, mode);
1127         ir_node  *nomem    = get_irg_no_mem(irg);
1128         ir_node  *in[2];
1129         ir_node  *call;
1130         ir_node  *resproj;
1131
1132         if (env->params->little_endian) {
1133                 in[0] = get_lowered_low(op);
1134                 in[1] = get_lowered_high(op);
1135         } else {
1136                 in[0] = get_lowered_high(op);
1137                 in[1] = get_lowered_low(op);
1138         }
1139         call    = new_rd_Call(dbgi, block, nomem, addr, 2, in, mtp);
1140         resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
1141         set_irn_pinned(call, get_irn_pinned(node));
1142
1143         if (env->params->little_endian) {
1144                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
1145                 ir_node *res_high = new_r_Proj(resproj, mode,              1);
1146                 ir_set_dw_lowered(node, res_low, res_high);
1147         } else {
1148                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
1149                 ir_node *res_high = new_r_Proj(resproj, mode,              0);
1150                 ir_set_dw_lowered(node, res_low, res_high);
1151         }
1152 }
1153
1154 /**
1155  * Translate a logical binop.
1156  *
1157  * Create two logical binops.
1158  */
1159 static void lower_binop_logical(ir_node *node, ir_mode *mode,
1160                                                                 ir_node *(*constr_rd)(dbg_info *db, ir_node *block, ir_node *op1, ir_node *op2, ir_mode *mode) )
1161 {
1162         ir_node               *left        = get_binop_left(node);
1163         ir_node               *right       = get_binop_right(node);
1164         const lower64_entry_t *left_entry  = get_node_entry(left);
1165         const lower64_entry_t *right_entry = get_node_entry(right);
1166         dbg_info              *dbgi        = get_irn_dbg_info(node);
1167         ir_node               *block       = get_nodes_block(node);
1168         ir_node               *res_low
1169                 = constr_rd(dbgi, block, left_entry->low_word, right_entry->low_word,
1170                             env->low_unsigned);
1171         ir_node               *res_high
1172                 = constr_rd(dbgi, block, left_entry->high_word, right_entry->high_word,
1173                             mode);
1174         ir_set_dw_lowered(node, res_low, res_high);
1175 }
1176
1177 static void lower_And(ir_node *node, ir_mode *mode)
1178 {
1179         lower_binop_logical(node, mode, new_rd_And);
1180 }
1181
1182 static void lower_Or(ir_node *node, ir_mode *mode)
1183 {
1184         lower_binop_logical(node, mode, new_rd_Or);
1185 }
1186
1187 static void lower_Eor(ir_node *node, ir_mode *mode)
1188 {
1189         lower_binop_logical(node, mode, new_rd_Eor);
1190 }
1191
1192 /**
1193  * Translate a Not.
1194  *
1195  * Create two logical Nots.
1196  */
1197 static void lower_Not(ir_node *node, ir_mode *mode)
1198 {
1199         ir_node               *op       = get_Not_op(node);
1200         const lower64_entry_t *op_entry = get_node_entry(op);
1201         dbg_info              *dbgi     = get_irn_dbg_info(node);
1202         ir_node               *block    = get_nodes_block(node);
1203         ir_node               *res_low
1204                 = new_rd_Not(dbgi, block, op_entry->low_word, env->low_unsigned);
1205         ir_node               *res_high
1206                 = new_rd_Not(dbgi, block, op_entry->high_word, mode);
1207         ir_set_dw_lowered(node, res_low, res_high);
1208 }
1209
1210 static void lower_Proj(ir_node *node, ir_mode *op_mode)
1211 {
1212         ir_mode *mode = get_irn_mode(node);
1213         ir_node *pred;
1214         (void)op_mode;
1215         if (mode != env->high_signed && mode != env->high_unsigned)
1216                 return;
1217         /* skip tuples */
1218         pred = get_Proj_pred(node);
1219         if (is_Tuple(pred)) {
1220                 long                   pn    = get_Proj_proj(node);
1221                 ir_node               *op    = get_irn_n(pred, pn);
1222                 const lower64_entry_t *entry = get_node_entry(op);
1223                 ir_set_dw_lowered(node, entry->low_word, entry->high_word);
1224         }
1225 }
1226
1227 static bool is_equality_cmp(const ir_node *node)
1228 {
1229         ir_relation relation = get_Cmp_relation(node);
1230         ir_node    *left     = get_Cmp_left(node);
1231         ir_node    *right    = get_Cmp_right(node);
1232         ir_mode    *mode     = get_irn_mode(left);
1233
1234         /* this probably makes no sense if unordered is involved */
1235         assert(!mode_is_float(mode));
1236
1237         if (relation == ir_relation_equal || relation == ir_relation_less_greater)
1238                 return true;
1239
1240         if (!is_Const(right) || !is_Const_null(right))
1241                 return false;
1242         if (mode_is_signed(mode)) {
1243                 return relation == ir_relation_less_greater;
1244         } else {
1245                 return relation == ir_relation_greater;
1246         }
1247 }
1248
1249 static ir_node *get_cfop_destination(const ir_node *cfop)
1250 {
1251         const ir_edge_t *first = get_irn_out_edge_first(cfop);
1252         /* we should only have 1 destination */
1253         assert(get_irn_n_edges(cfop) == 1);
1254         return get_edge_src_irn(first);
1255 }
1256
1257 static void lower_Switch(ir_node *node, ir_mode *high_mode)
1258 {
1259         ir_node *selector = get_Switch_selector(node);
1260         ir_mode *mode     = get_irn_mode(selector);
1261         (void)high_mode;
1262         if (mode == env->high_signed || mode == env->high_unsigned) {
1263                 /* we can't really handle Switch with 64bit offsets */
1264                 panic("Switch with 64bit jumptable not supported");
1265         }
1266         lower_node(selector);
1267 }
1268
1269 /**
1270  * Translate a Cond.
1271  */
1272 static void lower_Cond(ir_node *node, ir_mode *high_mode)
1273 {
1274         ir_node *left, *right, *block;
1275         ir_node *sel = get_Cond_selector(node);
1276         ir_mode *cmp_mode;
1277         const lower64_entry_t *lentry, *rentry;
1278         ir_node  *projT = NULL, *projF = NULL;
1279         ir_node  *new_bl, *irn;
1280         ir_node  *projHF, *projHT;
1281         ir_node  *dst_blk;
1282         ir_relation relation;
1283         ir_graph *irg;
1284         dbg_info *dbg;
1285         const ir_edge_t *edge;
1286         const ir_edge_t *next;
1287
1288         (void) high_mode;
1289
1290         if (!is_Cmp(sel)) {
1291                 lower_node(sel);
1292                 return;
1293         }
1294
1295         left     = get_Cmp_left(sel);
1296         cmp_mode = get_irn_mode(left);
1297         if (cmp_mode != env->high_signed && cmp_mode != env->high_unsigned) {
1298                 lower_node(sel);
1299                 return;
1300         }
1301
1302         right  = get_Cmp_right(sel);
1303         lower_node(left);
1304         lower_node(right);
1305         lentry = get_node_entry(left);
1306         rentry = get_node_entry(right);
1307
1308         /* all right, build the code */
1309         foreach_out_edge_safe(node, edge, next) {
1310                 ir_node *proj    = get_edge_src_irn(edge);
1311                 long     proj_nr;
1312                 if (!is_Proj(proj))
1313                         continue;
1314                 proj_nr = get_Proj_proj(proj);
1315
1316                 if (proj_nr == pn_Cond_true) {
1317                         assert(projT == NULL && "more than one Proj(true)");
1318                         projT = proj;
1319                 } else {
1320                         assert(proj_nr == pn_Cond_false);
1321                         assert(projF == NULL && "more than one Proj(false)");
1322                         projF = proj;
1323                 }
1324                 mark_irn_visited(proj);
1325         }
1326         assert(projT && projF);
1327
1328         /* create a new high compare */
1329         block    = get_nodes_block(node);
1330         irg      = get_Block_irg(block);
1331         dbg      = get_irn_dbg_info(sel);
1332         relation = get_Cmp_relation(sel);
1333
1334         if (is_equality_cmp(sel)) {
1335                 /* x ==/!= y ==> or(x_low^y_low,x_high^y_high) ==/!= 0 */
1336                 ir_mode *mode       = env->low_unsigned;
1337                 ir_node *low_left   = new_rd_Conv(dbg, block, lentry->low_word, mode);
1338                 ir_node *high_left  = new_rd_Conv(dbg, block, lentry->high_word, mode);
1339                 ir_node *low_right  = new_rd_Conv(dbg, block, rentry->low_word, mode);
1340                 ir_node *high_right = new_rd_Conv(dbg, block, rentry->high_word, mode);
1341                 ir_node *xor_low    = new_rd_Eor(dbg, block, low_left, low_right, mode);
1342                 ir_node *xor_high   = new_rd_Eor(dbg, block, high_left, high_right, mode);
1343                 ir_node *ornode = new_rd_Or(dbg, block, xor_low, xor_high, mode);
1344                 ir_node *cmp    = new_rd_Cmp(dbg, block, ornode, new_r_Const(irg, get_mode_null(mode)), relation);
1345                 set_Cond_selector(node, cmp);
1346                 return;
1347         }
1348
1349         if (relation == ir_relation_equal) {
1350                 ir_node *proj;
1351                 /* simple case:a == b <==> a_h == b_h && a_l == b_l */
1352                 dst_blk = get_cfop_destination(projF);
1353
1354                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1355                                  ir_relation_equal);
1356                 dbg = get_irn_dbg_info(node);
1357                 irn = new_rd_Cond(dbg, block, irn);
1358
1359                 projHF = new_r_Proj(irn, mode_X, pn_Cond_false);
1360                 mark_irn_visited(projHF);
1361                 exchange(projF, projHF);
1362
1363                 projHT = new_r_Proj(irn, mode_X, pn_Cond_true);
1364                 mark_irn_visited(projHT);
1365
1366                 new_bl = new_r_Block(irg, 1, &projHT);
1367
1368                 dbg = get_irn_dbg_info(sel);
1369                 irn = new_rd_Cmp(dbg, new_bl, lentry->low_word, rentry->low_word,
1370                                   ir_relation_equal);
1371                 dbg = get_irn_dbg_info(node);
1372                 irn = new_rd_Cond(dbg, new_bl, irn);
1373
1374                 proj = new_r_Proj(irn, mode_X, pn_Cond_false);
1375                 mark_irn_visited(proj);
1376                 add_block_cf_input(dst_blk, projHF, proj);
1377
1378                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1379                 mark_irn_visited(proj);
1380                 exchange(projT, proj);
1381         } else if (relation == ir_relation_less_greater) {
1382                 ir_node *proj;
1383                 /* simple case:a != b <==> a_h != b_h || a_l != b_l */
1384                 dst_blk = get_cfop_destination(projT);
1385
1386                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1387                                  ir_relation_less_greater);
1388                 dbg = get_irn_dbg_info(node);
1389                 irn = new_rd_Cond(dbg, block, irn);
1390
1391                 projHT = new_r_Proj(irn, mode_X, pn_Cond_true);
1392                 mark_irn_visited(projHT);
1393                 exchange(projT, projHT);
1394
1395                 projHF = new_r_Proj(irn, mode_X, pn_Cond_false);
1396                 mark_irn_visited(projHF);
1397
1398                 new_bl = new_r_Block(irg, 1, &projHF);
1399
1400                 dbg = get_irn_dbg_info(sel);
1401                 irn = new_rd_Cmp(dbg, new_bl, lentry->low_word, rentry->low_word,
1402                                  ir_relation_less_greater);
1403                 dbg = get_irn_dbg_info(node);
1404                 irn = new_rd_Cond(dbg, new_bl, irn);
1405
1406                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1407                 mark_irn_visited(proj);
1408                 add_block_cf_input(dst_blk, projHT, proj);
1409
1410                 proj = new_r_Proj(irn, mode_X, pn_Cond_false);
1411                 mark_irn_visited(proj);
1412                 exchange(projF, proj);
1413         } else {
1414                 ir_node *proj;
1415                 /* a rel b <==> a_h REL b_h || (a_h == b_h && a_l rel b_l) */
1416                 ir_node *dstT, *dstF, *newbl_eq, *newbl_l;
1417                 ir_node *projEqF;
1418
1419                 dstT = get_cfop_destination(projT);
1420                 dstF = get_cfop_destination(projF);
1421
1422                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1423                                  relation & ~ir_relation_equal);
1424                 dbg = get_irn_dbg_info(node);
1425                 irn = new_rd_Cond(dbg, block, irn);
1426
1427                 projHT = new_r_Proj(irn, mode_X, pn_Cond_true);
1428                 mark_irn_visited(projHT);
1429
1430                 projHF = new_r_Proj(irn, mode_X, pn_Cond_false);
1431                 mark_irn_visited(projHF);
1432
1433                 newbl_eq = new_r_Block(irg, 1, &projHF);
1434
1435                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1436                                  ir_relation_equal);
1437                 irn = new_rd_Cond(dbg, newbl_eq, irn);
1438
1439                 projEqF = new_r_Proj(irn, mode_X, pn_Cond_false);
1440                 mark_irn_visited(projEqF);
1441
1442                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1443                 mark_irn_visited(proj);
1444
1445                 newbl_l = new_r_Block(irg, 1, &proj);
1446
1447                 dbg = get_irn_dbg_info(sel);
1448                 irn = new_rd_Cmp(dbg, newbl_l, lentry->low_word, rentry->low_word,
1449                                  relation);
1450                 dbg = get_irn_dbg_info(node);
1451                 irn = new_rd_Cond(dbg, newbl_l, irn);
1452
1453                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1454                 mark_irn_visited(proj);
1455                 add_block_cf_input(dstT, projT, proj);
1456
1457                 proj = new_r_Proj(irn, mode_X, pn_Cond_false);
1458                 mark_irn_visited(proj);
1459                 add_block_cf_input(dstF, projF, proj);
1460
1461                 exchange(projT, projHT);
1462                 exchange(projF, projEqF);
1463         }
1464
1465         /* we have changed the control flow */
1466         env->flags |= CF_CHANGED;
1467 }
1468
1469 /**
1470  * Translate a Conv to higher_signed
1471  */
1472 static void lower_Conv_to_Ll(ir_node *node)
1473 {
1474         ir_mode  *omode        = get_irn_mode(node);
1475         ir_node  *op           = get_Conv_op(node);
1476         ir_mode  *imode        = get_irn_mode(op);
1477         ir_graph *irg          = get_irn_irg(node);
1478         ir_node  *block        = get_nodes_block(node);
1479         dbg_info *dbg          = get_irn_dbg_info(node);
1480         ir_node  *res_low;
1481         ir_node  *res_high;
1482
1483         ir_mode  *low_unsigned = env->low_unsigned;
1484         ir_mode  *low_signed
1485                 = mode_is_signed(omode) ? env->low_signed : low_unsigned;
1486
1487         if (mode_is_int(imode) || mode_is_reference(imode)) {
1488                 if (imode == env->high_signed || imode == env->high_unsigned) {
1489                         /* a Conv from Lu to Ls or Ls to Lu */
1490                         const lower64_entry_t *op_entry = get_node_entry(op);
1491                         res_low  = op_entry->low_word;
1492                         res_high = new_rd_Conv(dbg, block, op_entry->high_word, low_signed);
1493                 } else {
1494                         /* simple case: create a high word */
1495                         if (imode != low_unsigned)
1496                                 op = new_rd_Conv(dbg, block, op, low_unsigned);
1497
1498                         res_low = op;
1499
1500                         if (mode_is_signed(imode)) {
1501                                 int      c       = get_mode_size_bits(low_signed) - 1;
1502                                 ir_node *cnst    = new_r_Const_long(irg, low_unsigned, c);
1503                                 if (get_irn_mode(op) != low_signed)
1504                                         op = new_rd_Conv(dbg, block, op, low_signed);
1505                                 res_high = new_rd_Shrs(dbg, block, op, cnst, low_signed);
1506                         } else {
1507                                 res_high = new_r_Const(irg, get_mode_null(low_signed));
1508                         }
1509                 }
1510         } else if (imode == mode_b) {
1511                 res_low  = new_rd_Conv(dbg, block, op, low_unsigned);
1512                 res_high = new_r_Const(irg, get_mode_null(low_signed));
1513         } else {
1514                 ir_node *irn, *call;
1515                 ir_type *mtp = get_conv_type(imode, omode);
1516
1517                 irn = get_intrinsic_address(mtp, get_irn_op(node), imode, omode);
1518                 call = new_rd_Call(dbg, block, get_irg_no_mem(irg), irn, 1, &op, mtp);
1519                 set_irn_pinned(call, get_irn_pinned(node));
1520                 irn = new_r_Proj(call, mode_T, pn_Call_T_result);
1521
1522                 if (env->params->little_endian) {
1523                         res_low  = new_r_Proj(irn, low_unsigned, 0);
1524                         res_high = new_r_Proj(irn, low_signed, 1);
1525                 } else {
1526                         res_low  = new_r_Proj(irn, low_unsigned, 1);
1527                         res_high = new_r_Proj(irn, low_signed,   0);
1528                 }
1529         }
1530         ir_set_dw_lowered(node, res_low, res_high);
1531 }
1532
1533 /**
1534  * Translate a Conv from higher_unsigned
1535  */
1536 static void lower_Conv_from_Ll(ir_node *node)
1537 {
1538         ir_node               *op    = get_Conv_op(node);
1539         ir_mode               *omode = get_irn_mode(node);
1540         ir_node               *block = get_nodes_block(node);
1541         dbg_info              *dbg   = get_irn_dbg_info(node);
1542         ir_graph              *irg   = get_irn_irg(node);
1543         const lower64_entry_t *entry = get_node_entry(op);
1544
1545         if (mode_is_int(omode) || mode_is_reference(omode)) {
1546                 op = entry->low_word;
1547
1548                 /* simple case: create a high word */
1549                 if (omode != env->low_unsigned)
1550                         op = new_rd_Conv(dbg, block, op, omode);
1551
1552                 set_Conv_op(node, op);
1553         } else if (omode == mode_b) {
1554                 /* llu ? true : false  <=> (low|high) ? true : false */
1555                 ir_mode *mode   = env->low_unsigned;
1556                 ir_node *ornode = new_rd_Or(dbg, block, entry->low_word,
1557                                             entry->high_word, mode);
1558                 set_Conv_op(node, ornode);
1559         } else {
1560                 ir_node *irn, *call, *in[2];
1561                 ir_mode *imode = get_irn_mode(op);
1562                 ir_type *mtp   = get_conv_type(imode, omode);
1563                 ir_node *res;
1564
1565                 irn   = get_intrinsic_address(mtp, get_irn_op(node), imode, omode);
1566                 if (env->params->little_endian) {
1567                         in[0] = entry->low_word;
1568                         in[1] = entry->high_word;
1569                 } else {
1570                         in[0] = entry->high_word;
1571                         in[1] = entry->low_word;
1572                 }
1573
1574                 call = new_rd_Call(dbg, block, get_irg_no_mem(irg), irn, 2, in, mtp);
1575                 set_irn_pinned(call, get_irn_pinned(node));
1576                 irn = new_r_Proj(call, mode_T, pn_Call_T_result);
1577                 res = new_r_Proj(irn, omode, 0);
1578
1579                 exchange(node, res);
1580         }
1581 }
1582
1583 /**
1584  * lower Cmp
1585  */
1586 static void lower_Cmp(ir_node *cmp, ir_mode *m)
1587 {
1588         ir_node  *l        = get_Cmp_left(cmp);
1589         ir_mode  *cmp_mode = get_irn_mode(l);
1590         ir_node  *r, *low, *high, *t, *res;
1591         ir_relation relation;
1592         ir_node  *block;
1593         dbg_info *dbg;
1594         const lower64_entry_t *lentry;
1595         const lower64_entry_t *rentry;
1596         (void) m;
1597
1598         if (cmp_mode != env->high_signed && cmp_mode != env->high_unsigned)
1599                 return;
1600
1601         r        = get_Cmp_right(cmp);
1602         lentry   = get_node_entry(l);
1603         rentry   = get_node_entry(r);
1604         relation = get_Cmp_relation(cmp);
1605         block    = get_nodes_block(cmp);
1606         dbg      = get_irn_dbg_info(cmp);
1607
1608         /* easy case for x ==/!= 0 (see lower_Cond for details) */
1609         if (is_equality_cmp(cmp)) {
1610                 ir_graph *irg        = get_irn_irg(cmp);
1611                 ir_mode  *mode       = env->low_unsigned;
1612                 ir_node  *low_left   = new_rd_Conv(dbg, block, lentry->low_word, mode);
1613                 ir_node  *high_left  = new_rd_Conv(dbg, block, lentry->high_word, mode);
1614                 ir_node  *low_right  = new_rd_Conv(dbg, block, rentry->low_word, mode);
1615                 ir_node  *high_right = new_rd_Conv(dbg, block, rentry->high_word, mode);
1616                 ir_node  *xor_low    = new_rd_Eor(dbg, block, low_left, low_right, mode);
1617                 ir_node  *xor_high   = new_rd_Eor(dbg, block, high_left, high_right, mode);
1618                 ir_node  *ornode     = new_rd_Or(dbg, block, xor_low, xor_high, mode);
1619                 ir_node  *new_cmp    = new_rd_Cmp(dbg, block, ornode, new_r_Const(irg, get_mode_null(mode)), relation);
1620                 exchange(cmp, new_cmp);
1621                 return;
1622         }
1623
1624         if (relation == ir_relation_equal) {
1625                 /* simple case:a == b <==> a_h == b_h && a_l == b_l */
1626                 low  = new_rd_Cmp(dbg, block, lentry->low_word, rentry->low_word,
1627                                   relation);
1628                 high = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1629                                   relation);
1630                 res  = new_rd_And(dbg, block, low, high, mode_b);
1631         } else if (relation == ir_relation_less_greater) {
1632                 /* simple case:a != b <==> a_h != b_h || a_l != b_l */
1633                 low  = new_rd_Cmp(dbg, block, lentry->low_word, rentry->low_word,
1634                                   relation);
1635                 high = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1636                                   relation);
1637                 res = new_rd_Or(dbg, block, low, high, mode_b);
1638         } else {
1639                 /* a rel b <==> a_h REL b_h || (a_h == b_h && a_l rel b_l) */
1640                 ir_node *high1 = new_rd_Cmp(dbg, block, lentry->high_word,
1641                         rentry->high_word, relation & ~ir_relation_equal);
1642                 low  = new_rd_Cmp(dbg, block, lentry->low_word, rentry->low_word,
1643                                   relation);
1644                 high = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1645                                   ir_relation_equal);
1646                 t = new_rd_And(dbg, block, low, high, mode_b);
1647                 res = new_rd_Or(dbg, block, high1, t, mode_b);
1648         }
1649         exchange(cmp, res);
1650 }
1651
1652 /**
1653  * Translate a Conv.
1654  */
1655 static void lower_Conv(ir_node *node, ir_mode *mode)
1656 {
1657         mode = get_irn_mode(node);
1658
1659         if (mode == env->high_signed || mode == env->high_unsigned) {
1660                 lower_Conv_to_Ll(node);
1661         } else {
1662                 ir_mode *op_mode = get_irn_mode(get_Conv_op(node));
1663
1664                 if (op_mode == env->high_signed || op_mode == env->high_unsigned) {
1665                         lower_Conv_from_Ll(node);
1666                 }
1667         }
1668 }
1669
1670 static void fix_parameter_entities(ir_graph *irg, ir_type *orig_mtp)
1671 {
1672         size_t      orig_n_params      = get_method_n_params(orig_mtp);
1673         ir_entity **parameter_entities;
1674
1675         parameter_entities = ALLOCANZ(ir_entity*, orig_n_params);
1676
1677         ir_type *frame_type = get_irg_frame_type(irg);
1678         size_t   n          = get_compound_n_members(frame_type);
1679         size_t   i;
1680         size_t   n_param;
1681
1682         /* collect parameter entities */
1683         for (i = 0; i < n; ++i) {
1684                 ir_entity *entity = get_compound_member(frame_type, i);
1685                 size_t     p;
1686                 if (!is_parameter_entity(entity))
1687                         continue;
1688                 p = get_entity_parameter_number(entity);
1689                 if (p == IR_VA_START_PARAMETER_NUMBER)
1690                         continue;
1691                 assert(p < orig_n_params);
1692                 assert(parameter_entities[p] == NULL);
1693                 parameter_entities[p] = entity;
1694         }
1695
1696         /* adjust indices */
1697         n_param = 0;
1698         for (i = 0; i < orig_n_params; ++i, ++n_param) {
1699                 ir_entity *entity = parameter_entities[i];
1700                 ir_type   *tp;
1701
1702                 if (entity != NULL)
1703                         set_entity_parameter_number(entity, n_param);
1704
1705                 tp = get_method_param_type(orig_mtp, i);
1706                 if (is_Primitive_type(tp)) {
1707                         ir_mode *mode = get_type_mode(tp);
1708                         if (mode == env->high_signed || mode == env->high_unsigned) {
1709                                 ++n_param;
1710                                 /* note that we do not change the type of the parameter
1711                                  * entities, as calling convention fixup later still needs to
1712                                  * know which is/was a lowered doubleword.
1713                                  * So we just mark/remember it for later */
1714                                 if (entity != NULL) {
1715                                         assert(entity->attr.parameter.doubleword_low_mode == NULL);
1716                                         entity->attr.parameter.doubleword_low_mode
1717                                                 = env->low_unsigned;
1718                                 }
1719                         }
1720                 }
1721         }
1722 }
1723
1724 /**
1725  * Lower the method type.
1726  *
1727  * @param env  the lower environment
1728  * @param mtp  the method type to lower
1729  *
1730  * @return the lowered type
1731  */
1732 static ir_type *lower_mtp(ir_type *mtp)
1733 {
1734         ir_type *res;
1735         size_t   i;
1736         size_t   orig_n_params;
1737         size_t   orig_n_res;
1738         size_t   n_param;
1739         size_t   n_res;
1740         bool     must_be_lowered;
1741
1742         res = pmap_get(ir_type, lowered_type, mtp);
1743         if (res != NULL)
1744                 return res;
1745         if (type_visited(mtp))
1746                 return mtp;
1747         mark_type_visited(mtp);
1748
1749         orig_n_params   = get_method_n_params(mtp);
1750         orig_n_res      = get_method_n_ress(mtp);
1751         n_param         = orig_n_params;
1752         n_res           = orig_n_res;
1753         must_be_lowered = false;
1754
1755         /* count new number of params */
1756         for (i = orig_n_params; i > 0;) {
1757                 ir_type *tp = get_method_param_type(mtp, --i);
1758
1759                 if (is_Primitive_type(tp)) {
1760                         ir_mode *mode = get_type_mode(tp);
1761
1762                         if (mode == env->high_signed || mode == env->high_unsigned) {
1763                                 ++n_param;
1764                                 must_be_lowered = true;
1765                         }
1766                 }
1767         }
1768
1769         /* count new number of results */
1770         for (i = orig_n_res; i > 0;) {
1771                 ir_type *tp = get_method_res_type(mtp, --i);
1772
1773                 if (is_Primitive_type(tp)) {
1774                         ir_mode *mode = get_type_mode(tp);
1775
1776                         if (mode == env->high_signed || mode == env->high_unsigned) {
1777                                 ++n_res;
1778                                 must_be_lowered = true;
1779                         }
1780                 }
1781         }
1782         if (!must_be_lowered) {
1783                 set_type_link(mtp, NULL);
1784                 return mtp;
1785         }
1786
1787         res = new_d_type_method(n_param, n_res, get_type_dbg_info(mtp));
1788
1789         /* set param types and result types */
1790         for (i = n_param = 0; i < orig_n_params; ++i) {
1791                 ir_type *tp = get_method_param_type(mtp, i);
1792
1793                 if (is_Primitive_type(tp)) {
1794                         ir_mode *mode = get_type_mode(tp);
1795
1796                         if (mode == env->high_signed) {
1797                                 if (env->params->little_endian) {
1798                                         set_method_param_type(res, n_param++, tp_u);
1799                                         set_method_param_type(res, n_param++, tp_s);
1800                                 } else {
1801                                         set_method_param_type(res, n_param++, tp_s);
1802                                         set_method_param_type(res, n_param++, tp_u);
1803                                 }
1804                         } else if (mode == env->high_unsigned) {
1805                                 set_method_param_type(res, n_param++, tp_u);
1806                                 set_method_param_type(res, n_param++, tp_u);
1807                         } else {
1808                                 set_method_param_type(res, n_param, tp);
1809                                 ++n_param;
1810                         }
1811                 } else {
1812                         set_method_param_type(res, n_param, tp);
1813                         ++n_param;
1814                 }
1815         }
1816         for (i = n_res = 0; i < orig_n_res; ++i) {
1817                 ir_type *tp = get_method_res_type(mtp, i);
1818
1819                 if (is_Primitive_type(tp)) {
1820                         ir_mode *mode = get_type_mode(tp);
1821
1822                         if (mode == env->high_signed) {
1823                                 if (env->params->little_endian) {
1824                                         set_method_res_type(res, n_res++, tp_u);
1825                                         set_method_res_type(res, n_res++, tp_s);
1826                                 } else {
1827                                         set_method_res_type(res, n_res++, tp_s);
1828                                         set_method_res_type(res, n_res++, tp_u);
1829                                 }
1830                         } else if (mode == env->high_unsigned) {
1831                                 set_method_res_type(res, n_res++, tp_u);
1832                                 set_method_res_type(res, n_res++, tp_u);
1833                         } else {
1834                                 set_method_res_type(res, n_res++, tp);
1835                         }
1836                 } else {
1837                         set_method_res_type(res, n_res++, tp);
1838                 }
1839         }
1840
1841         set_method_variadicity(res, get_method_variadicity(mtp));
1842         set_method_calling_convention(res, get_method_calling_convention(mtp));
1843         set_method_additional_properties(res, get_method_additional_properties(mtp));
1844
1845         set_higher_type(res, mtp);
1846         set_type_link(res, mtp);
1847
1848         mark_type_visited(res);
1849         pmap_insert(lowered_type, mtp, res);
1850         return res;
1851 }
1852
1853 /**
1854  * Translate a Return.
1855  */
1856 static void lower_Return(ir_node *node, ir_mode *mode)
1857 {
1858         ir_graph  *irg = get_irn_irg(node);
1859         ir_entity *ent = get_irg_entity(irg);
1860         ir_type   *mtp = get_entity_type(ent);
1861         ir_node  **in;
1862         size_t     i, j, n;
1863         int        need_conv = 0;
1864         (void) mode;
1865
1866         /* check if this return must be lowered */
1867         for (i = 0, n = get_Return_n_ress(node); i < n; ++i) {
1868                 ir_node *pred  = get_Return_res(node, i);
1869                 ir_mode *rmode = get_irn_op_mode(pred);
1870
1871                 if (rmode == env->high_signed || rmode == env->high_unsigned)
1872                         need_conv = 1;
1873         }
1874         if (! need_conv)
1875                 return;
1876
1877         ent = get_irg_entity(irg);
1878         mtp = get_entity_type(ent);
1879
1880         /* create a new in array */
1881         NEW_ARR_A(ir_node *, in, get_method_n_ress(mtp) + 1);
1882         j = 0;
1883         in[j++] = get_Return_mem(node);
1884
1885         for (i = 0, n = get_Return_n_ress(node); i < n; ++i) {
1886                 ir_node *pred      = get_Return_res(node, i);
1887                 ir_mode *pred_mode = get_irn_mode(pred);
1888
1889                 if (pred_mode == env->high_signed || pred_mode == env->high_unsigned) {
1890                         const lower64_entry_t *entry = get_node_entry(pred);
1891                         if (env->params->little_endian) {
1892                                 in[j++] = entry->low_word;
1893                                 in[j++] = entry->high_word;
1894                         } else {
1895                                 in[j++] = entry->high_word;
1896                                 in[j++] = entry->low_word;
1897                         }
1898                 } else {
1899                         in[j++] = pred;
1900                 }
1901         }
1902         assert(j == get_method_n_ress(mtp)+1);
1903
1904         set_irn_in(node, j, in);
1905 }
1906
1907 /**
1908  * Translate the parameters.
1909  */
1910 static void lower_Start(ir_node *node, ir_mode *high_mode)
1911 {
1912         ir_graph  *irg      = get_irn_irg(node);
1913         ir_entity *ent      = get_irg_entity(irg);
1914         ir_type   *mtp      = get_entity_type(ent);
1915         ir_type   *orig_mtp = (ir_type*)get_type_link(mtp);
1916         ir_node   *args;
1917         long      *new_projs;
1918         size_t    i, j, n_params;
1919         const ir_edge_t *edge;
1920         const ir_edge_t *next;
1921         (void) high_mode;
1922
1923         /* if type link is NULL then the type was not lowered, hence no changes
1924          * at Start necessary */
1925         if (orig_mtp == NULL)
1926                 return;
1927
1928         n_params = get_method_n_params(orig_mtp);
1929
1930         NEW_ARR_A(long, new_projs, n_params);
1931
1932         /* Calculate mapping of proj numbers in new_projs */
1933         for (i = j = 0; i < n_params; ++i, ++j) {
1934                 ir_type *ptp = get_method_param_type(orig_mtp, i);
1935
1936                 new_projs[i] = j;
1937                 if (is_Primitive_type(ptp)) {
1938                         ir_mode *amode = get_type_mode(ptp);
1939                         if (amode == env->high_signed || amode == env->high_unsigned)
1940                                 ++j;
1941                 }
1942         }
1943
1944         /* find args Proj */
1945         args = NULL;
1946         foreach_out_edge(node, edge) {
1947                 ir_node *proj = get_edge_src_irn(edge);
1948                 if (!is_Proj(proj))
1949                         continue;
1950                 if (get_Proj_proj(proj) == pn_Start_T_args) {
1951                         args = proj;
1952                         break;
1953                 }
1954         }
1955         if (args == NULL)
1956                 return;
1957
1958         /* fix all Proj's and create new ones */
1959         foreach_out_edge_safe(args, edge, next) {
1960                 ir_node *proj   = get_edge_src_irn(edge);
1961                 ir_mode *mode   = get_irn_mode(proj);
1962                 ir_mode *mode_l = env->low_unsigned;
1963                 ir_node *pred;
1964                 long     proj_nr;
1965                 ir_mode *mode_h;
1966                 ir_node *res_low;
1967                 ir_node *res_high;
1968                 int      old_cse;
1969                 dbg_info *dbg;
1970
1971                 if (!is_Proj(proj))
1972                         continue;
1973                 pred    = get_Proj_pred(proj);
1974                 proj_nr = get_Proj_proj(proj);
1975
1976                 if (mode == env->high_signed) {
1977                         mode_h = env->low_signed;
1978                 } else if (mode == env->high_unsigned) {
1979                         mode_h = env->low_unsigned;
1980                 } else {
1981                         long new_pn = new_projs[proj_nr];
1982                         set_Proj_proj(proj, new_pn);
1983                         continue;
1984                 }
1985
1986                 /* Switch off CSE or we might get an already existing Proj. */
1987                 old_cse = get_opt_cse();
1988                 set_opt_cse(0);
1989                 dbg = get_irn_dbg_info(proj);
1990                 if (env->params->little_endian) {
1991                         res_low  = new_rd_Proj(dbg, pred, mode_l, new_projs[proj_nr]);
1992                         res_high = new_rd_Proj(dbg, pred, mode_h, new_projs[proj_nr] + 1);
1993                 } else {
1994                         res_high = new_rd_Proj(dbg, pred, mode_h, new_projs[proj_nr]);
1995                         res_low  = new_rd_Proj(dbg, pred, mode_l, new_projs[proj_nr] + 1);
1996                 }
1997                 set_opt_cse(old_cse);
1998                 ir_set_dw_lowered(proj, res_low, res_high);
1999         }
2000 }
2001
2002 /**
2003  * Translate a Call.
2004  */
2005 static void lower_Call(ir_node *node, ir_mode *mode)
2006 {
2007         ir_type  *tp = get_Call_type(node);
2008         ir_node  **in;
2009         size_t   n_params, n_res;
2010         bool     need_lower = false;
2011         size_t   i, j;
2012         size_t   p;
2013         long     *res_numbers = NULL;
2014         ir_node  *resproj;
2015         const ir_edge_t *edge;
2016         const ir_edge_t *next;
2017         (void) mode;
2018
2019         n_params = get_method_n_params(tp);
2020         for (p = 0; p < n_params; ++p) {
2021                 ir_type *ptp = get_method_param_type(tp, p);
2022
2023                 if (is_Primitive_type(ptp)) {
2024                         ir_mode *pmode = get_type_mode(ptp);
2025                         if (pmode == env->high_signed || pmode == env->high_unsigned) {
2026                                 need_lower = true;
2027                                 break;
2028                         }
2029                 }
2030         }
2031         n_res = get_method_n_ress(tp);
2032         if (n_res > 0) {
2033                 NEW_ARR_A(long, res_numbers, n_res);
2034
2035                 for (i = j = 0; i < n_res; ++i, ++j) {
2036                         ir_type *ptp = get_method_res_type(tp, i);
2037
2038                         res_numbers[i] = j;
2039                         if (is_Primitive_type(ptp)) {
2040                                 ir_mode *rmode = get_type_mode(ptp);
2041                                 if (rmode == env->high_signed || rmode == env->high_unsigned) {
2042                                         need_lower = true;
2043                                         ++j;
2044                                 }
2045                         }
2046                 }
2047         }
2048
2049         if (! need_lower)
2050                 return;
2051
2052         /* let's lower it */
2053         tp = lower_mtp(tp);
2054         set_Call_type(node, tp);
2055
2056         NEW_ARR_A(ir_node *, in, get_method_n_params(tp) + 2);
2057
2058         in[0] = get_Call_mem(node);
2059         in[1] = get_Call_ptr(node);
2060
2061         for (j = 2, i = 0; i < n_params; ++i) {
2062                 ir_node *pred      = get_Call_param(node, i);
2063                 ir_mode *pred_mode = get_irn_mode(pred);
2064
2065                 if (pred_mode == env->high_signed || pred_mode == env->high_unsigned) {
2066                         const lower64_entry_t *pred_entry = get_node_entry(pred);
2067                         if (env->params->little_endian) {
2068                                 in[j++] = pred_entry->low_word;
2069                                 in[j++] = pred_entry->high_word;
2070                         } else {
2071                                 in[j++] = pred_entry->high_word;
2072                                 in[j++] = pred_entry->low_word;
2073                         }
2074                 } else {
2075                         in[j++] = pred;
2076                 }
2077         }
2078
2079         set_irn_in(node, j, in);
2080
2081         /* find results T */
2082         resproj = NULL;
2083         foreach_out_edge(node, edge) {
2084                 ir_node *proj = get_edge_src_irn(edge);
2085                 if (!is_Proj(proj))
2086                         continue;
2087                 if (get_Proj_proj(proj) == pn_Call_T_result) {
2088                         resproj = proj;
2089                         break;
2090                 }
2091         }
2092         if (resproj == NULL)
2093                 return;
2094
2095         /* fix the results */
2096         foreach_out_edge_safe(resproj, edge, next) {
2097                 ir_node  *proj      = get_edge_src_irn(edge);
2098                 ir_mode  *proj_mode = get_irn_mode(proj);
2099                 ir_mode  *mode_l    = env->low_unsigned;
2100                 ir_node  *pred;
2101                 long      proj_nr;
2102                 ir_mode  *mode_h;
2103                 ir_node  *res_low;
2104                 ir_node  *res_high;
2105                 dbg_info *dbg;
2106
2107                 if (!is_Proj(proj))
2108                         continue;
2109                 pred    = get_Proj_pred(proj);
2110                 proj_nr = get_Proj_proj(proj);
2111
2112                 if (proj_mode == env->high_signed) {
2113                         mode_h = env->low_signed;
2114                 } else if (proj_mode == env->high_unsigned) {
2115                         mode_h = env->low_unsigned;
2116                 } else {
2117                         long new_nr = res_numbers[proj_nr];
2118                         set_Proj_proj(proj, new_nr);
2119                         continue;
2120                 }
2121
2122                 dbg = get_irn_dbg_info(proj);
2123                 if (env->params->little_endian) {
2124                         res_low  = new_rd_Proj(dbg, pred, mode_l, res_numbers[proj_nr]);
2125                         res_high = new_rd_Proj(dbg, pred, mode_h, res_numbers[proj_nr] + 1);
2126                 } else {
2127                         res_high = new_rd_Proj(dbg, pred, mode_h, res_numbers[proj_nr]);
2128                         res_low  = new_rd_Proj(dbg, pred, mode_l, res_numbers[proj_nr] + 1);
2129                 }
2130                 ir_set_dw_lowered(proj, res_low, res_high);
2131         }
2132 }
2133
2134 /**
2135  * Translate an Unknown into two.
2136  */
2137 static void lower_Unknown(ir_node *node, ir_mode *mode)
2138 {
2139         ir_mode  *low_mode = env->low_unsigned;
2140         ir_graph *irg      = get_irn_irg(node);
2141         ir_node  *res_low  = new_r_Unknown(irg, low_mode);
2142         ir_node  *res_high = new_r_Unknown(irg, mode);
2143         ir_set_dw_lowered(node, res_low, res_high);
2144 }
2145
2146 /**
2147  * Translate a Bad into two.
2148  */
2149 static void lower_Bad(ir_node *node, ir_mode *mode)
2150 {
2151         ir_mode  *low_mode = env->low_unsigned;
2152         ir_graph *irg      = get_irn_irg(node);
2153         ir_node  *res_low  = new_r_Bad(irg, low_mode);
2154         ir_node  *res_high = new_r_Bad(irg, mode);
2155         ir_set_dw_lowered(node, res_low, res_high);
2156 }
2157
2158 /**
2159  * Translate a Phi.
2160  *
2161  * First step: just create two templates
2162  */
2163 static void lower_Phi(ir_node *phi)
2164 {
2165         ir_mode  *mode = get_irn_mode(phi);
2166         int       i;
2167         int       arity;
2168         ir_node **in_l;
2169         ir_node **in_h;
2170         ir_node  *unk_l;
2171         ir_node  *unk_h;
2172         ir_node  *phi_l;
2173         ir_node  *phi_h;
2174         dbg_info *dbg;
2175         ir_node  *block;
2176         ir_graph *irg;
2177         ir_mode  *mode_l;
2178         ir_mode  *mode_h;
2179
2180         /* enqueue predecessors */
2181         arity = get_Phi_n_preds(phi);
2182         for (i = 0; i < arity; ++i) {
2183                 ir_node *pred = get_Phi_pred(phi, i);
2184                 pdeq_putr(env->waitq, pred);
2185         }
2186
2187         if (mode != env->high_signed && mode != env->high_unsigned)
2188                 return;
2189
2190         /* first create a new in array */
2191         NEW_ARR_A(ir_node *, in_l, arity);
2192         NEW_ARR_A(ir_node *, in_h, arity);
2193         irg    = get_irn_irg(phi);
2194         mode_l = env->low_unsigned;
2195         mode_h = mode == env->high_signed ? env->low_signed : env->low_unsigned;
2196         unk_l  = new_r_Dummy(irg, mode_l);
2197         unk_h  = new_r_Dummy(irg, mode_h);
2198         for (i = 0; i < arity; ++i) {
2199                 in_l[i] = unk_l;
2200                 in_h[i] = unk_h;
2201         }
2202
2203         dbg   = get_irn_dbg_info(phi);
2204         block = get_nodes_block(phi);
2205         phi_l = new_rd_Phi(dbg, block, arity, in_l, mode_l);
2206         phi_h = new_rd_Phi(dbg, block, arity, in_h, mode_h);
2207
2208         ir_set_dw_lowered(phi, phi_l, phi_h);
2209
2210         /* remember that we need to fixup the predecessors later */
2211         ARR_APP1(ir_node*, env->lowered_phis, phi);
2212 }
2213
2214 static void fixup_phi(ir_node *phi)
2215 {
2216         const lower64_entry_t *entry = get_node_entry(phi);
2217         ir_node               *phi_l = entry->low_word;
2218         ir_node               *phi_h = entry->high_word;
2219         int                    arity = get_Phi_n_preds(phi);
2220         int                    i;
2221
2222         /* exchange phi predecessors which are lowered by now */
2223         for (i = 0; i < arity; ++i) {
2224                 ir_node               *pred       = get_Phi_pred(phi, i);
2225                 const lower64_entry_t *pred_entry = get_node_entry(pred);
2226
2227                 set_Phi_pred(phi_l, i, pred_entry->low_word);
2228                 set_Phi_pred(phi_h, i, pred_entry->high_word);
2229         }
2230 }
2231
2232 /**
2233  * Translate a Mux.
2234  */
2235 static void lower_Mux(ir_node *mux, ir_mode *mode)
2236 {
2237         ir_node               *truen       = get_Mux_true(mux);
2238         ir_node               *falsen      = get_Mux_false(mux);
2239         ir_node               *sel         = get_Mux_sel(mux);
2240         const lower64_entry_t *true_entry  = get_node_entry(truen);
2241         const lower64_entry_t *false_entry = get_node_entry(falsen);
2242         ir_node               *true_l      = true_entry->low_word;
2243         ir_node               *true_h      = true_entry->high_word;
2244         ir_node               *false_l     = false_entry->low_word;
2245         ir_node               *false_h     = false_entry->high_word;
2246         dbg_info              *dbgi        = get_irn_dbg_info(mux);
2247         ir_node               *block       = get_nodes_block(mux);
2248         ir_node               *res_low
2249                 = new_rd_Mux(dbgi, block, sel, false_l, true_l, env->low_unsigned);
2250         ir_node               *res_high
2251                 = new_rd_Mux(dbgi, block, sel, false_h, true_h, mode);
2252         ir_set_dw_lowered(mux, res_low, res_high);
2253 }
2254
2255 /**
2256  * Translate an ASM node.
2257  */
2258 static void lower_ASM(ir_node *asmn, ir_mode *mode)
2259 {
2260         ir_mode           *high_signed        = env->high_signed;
2261         ir_mode           *high_unsigned      = env->high_unsigned;
2262         int                n_outs             = get_ASM_n_output_constraints(asmn);
2263         ir_asm_constraint *output_constraints = get_ASM_output_constraints(asmn);
2264         ir_asm_constraint *input_constraints  = get_ASM_input_constraints(asmn);
2265         unsigned           n_64bit_outs       = 0;
2266         int      i;
2267
2268         (void)mode;
2269
2270         for (i = get_irn_arity(asmn) - 1; i >= 0; --i) {
2271                 ir_node *op      = get_irn_n(asmn, i);
2272                 ir_mode *op_mode = get_irn_mode(op);
2273                 if (op_mode == high_signed || op_mode == high_unsigned) {
2274                         panic("lowering ASM 64bit input unimplemented");
2275                 }
2276         }
2277
2278         for (i = 0; i < n_outs; ++i) {
2279                 const ir_asm_constraint *constraint = &output_constraints[i];
2280                 if (constraint->mode == high_signed || constraint->mode == high_unsigned) {
2281                         const char *constr = get_id_str(constraint->constraint);
2282                         ++n_64bit_outs;
2283                         /* TODO: How to do this architecture neutral? This is very
2284                          * i386 specific... */
2285                         if (constr[0] != '=' || constr[1] != 'A') {
2286                                 panic("lowering ASM 64bit output only supports '=A' currently");
2287                         }
2288                 }
2289         }
2290
2291         if (n_64bit_outs == 0)
2292                 return;
2293
2294         {
2295                 dbg_info          *dbgi       = get_irn_dbg_info(asmn);
2296                 ir_node           *block      = get_nodes_block(asmn);
2297                 int                arity      = get_irn_arity(asmn);
2298                 ir_node          **in         = get_irn_in(asmn) + 1;
2299                 int                new_n_outs = 0;
2300                 int                n_clobber  = get_ASM_n_clobbers(asmn);
2301                 long              *proj_map   = ALLOCAN(long, n_outs);
2302                 ident            **clobbers   = get_ASM_clobbers(asmn);
2303                 ident             *asm_text   = get_ASM_text(asmn);
2304                 ir_asm_constraint *new_outputs
2305                         = ALLOCAN(ir_asm_constraint, n_outs+n_64bit_outs);
2306                 ir_node           *new_asm;
2307                 const ir_edge_t   *edge;
2308                 const ir_edge_t   *next;
2309
2310                 for (i = 0; i < n_outs; ++i) {
2311                         const ir_asm_constraint *constraint = &output_constraints[i];
2312                         if (constraint->mode == high_signed || constraint->mode == high_unsigned) {
2313                                 new_outputs[new_n_outs].pos        = constraint->pos;
2314                                 new_outputs[new_n_outs].constraint = new_id_from_str("=a");
2315                                 new_outputs[new_n_outs].mode       = env->low_unsigned;
2316                                 proj_map[i] = new_n_outs;
2317                                 ++new_n_outs;
2318                                 new_outputs[new_n_outs].pos        = constraint->pos;
2319                                 new_outputs[new_n_outs].constraint = new_id_from_str("=d");
2320                                 if (constraint->mode == high_signed)
2321                                         new_outputs[new_n_outs].mode = env->low_signed;
2322                                 else
2323                                         new_outputs[new_n_outs].mode = env->low_unsigned;
2324                                 ++new_n_outs;
2325                         } else {
2326                                 new_outputs[new_n_outs] = *constraint;
2327                                 proj_map[i] = new_n_outs;
2328                                 ++new_n_outs;
2329                         }
2330                 }
2331                 assert(new_n_outs == n_outs+(int)n_64bit_outs);
2332
2333                 new_asm = new_rd_ASM(dbgi, block, arity, in, input_constraints,
2334                                      new_n_outs, new_outputs, n_clobber, clobbers,
2335                                      asm_text);
2336
2337                 foreach_out_edge_safe(asmn, edge, next) {
2338                         ir_node *proj      = get_edge_src_irn(edge);
2339                         ir_mode *proj_mode = get_irn_mode(proj);
2340                         long     pn;
2341
2342                         if (!is_Proj(proj))
2343                                 continue;
2344                         pn = get_Proj_proj(proj);
2345
2346                         if (pn < n_outs)
2347                                 pn = proj_map[pn];
2348                         else
2349                                 pn = new_n_outs + pn - n_outs;
2350
2351                         if (proj_mode == high_signed || proj_mode == high_unsigned) {
2352                                 ir_mode *high_mode
2353                                         = proj_mode == high_signed ? env->low_signed : env->low_unsigned;
2354                                 ir_node *np_low  = new_r_Proj(new_asm, env->low_unsigned, pn);
2355                                 ir_node *np_high = new_r_Proj(new_asm, high_mode, pn+1);
2356                                 ir_set_dw_lowered(proj, np_low, np_high);
2357                         } else {
2358                                 ir_node *np = new_r_Proj(new_asm, proj_mode, pn);
2359                                 exchange(proj, np);
2360                         }
2361                 }
2362         }
2363 }
2364
2365 /**
2366  * Lower the builtin type to its higher part.
2367  *
2368  * @param mtp  the builtin type to lower
2369  *
2370  * @return the lowered type
2371  */
2372 static ir_type *lower_Builtin_type_high(ir_type *mtp)
2373 {
2374         ir_type *res;
2375         size_t   i;
2376         size_t   n_params;
2377         size_t   n_results;
2378         bool     must_be_lowered;
2379
2380         res = pmap_get(ir_type, lowered_builtin_type_high, mtp);
2381         if (res != NULL)
2382                 return res;
2383
2384         n_params        = get_method_n_params(mtp);
2385         n_results       = get_method_n_ress(mtp);
2386         must_be_lowered = false;
2387
2388         /* check for double word parameter */
2389         for (i = n_params; i > 0;) {
2390                 ir_type *tp = get_method_param_type(mtp, --i);
2391
2392                 if (is_Primitive_type(tp)) {
2393                         ir_mode *mode = get_type_mode(tp);
2394
2395                         if (mode == env->high_signed || mode == env->high_unsigned) {
2396                                 must_be_lowered = true;
2397                                 break;
2398                         }
2399                 }
2400         }
2401
2402         if (!must_be_lowered) {
2403                 set_type_link(mtp, NULL);
2404                 return mtp;
2405         }
2406
2407         res = new_d_type_method(n_params, n_results, get_type_dbg_info(mtp));
2408
2409         /* set param types and result types */
2410         for (i = 0; i < n_params; ++i) {
2411                 ir_type *tp = get_method_param_type(mtp, i);
2412
2413                 if (is_Primitive_type(tp)) {
2414                         ir_mode *mode = get_type_mode(tp);
2415
2416                         if (mode == env->high_signed) {
2417                                 if (env->params->little_endian) {
2418                                         set_method_param_type(res, i, tp_u);
2419                                 } else {
2420                                         set_method_param_type(res, i, tp_s);
2421                                 }
2422                         } else if (mode == env->high_unsigned) {
2423                                 set_method_param_type(res, i, tp_u);
2424                         } else {
2425                                 set_method_param_type(res, i, tp);
2426                         }
2427                 } else {
2428                         set_method_param_type(res, i, tp);
2429                 }
2430         }
2431         for (i = n_results = 0; i < n_results; ++i) {
2432                 ir_type *tp = get_method_res_type(mtp, i);
2433
2434                 if (is_Primitive_type(tp)) {
2435                         ir_mode *mode = get_type_mode(tp);
2436
2437                         if (mode == env->high_signed) {
2438                                 if (env->params->little_endian) {
2439                                         set_method_res_type(res, i, tp_u);
2440                                 } else {
2441                                         set_method_res_type(res, i, tp_s);
2442                                 }
2443                         } else if (mode == env->high_unsigned) {
2444                                 set_method_res_type(res, i, tp_u);
2445                         } else {
2446                                 set_method_res_type(res, i, tp);
2447                         }
2448                 } else {
2449                         set_method_res_type(res, i, tp);
2450                 }
2451         }
2452
2453         set_method_variadicity(res, get_method_variadicity(mtp));
2454         set_method_calling_convention(res, get_method_calling_convention(mtp));
2455         set_method_additional_properties(res, get_method_additional_properties(mtp));
2456
2457         pmap_insert(lowered_builtin_type_high, mtp, res);
2458         return res;
2459 }
2460
2461 /**
2462  * Lower the builtin type to its lower part.
2463  *
2464  * @param mtp  the builtin type to lower
2465  *
2466  * @return the lowered type
2467  */
2468 static ir_type *lower_Builtin_type_low(ir_type *mtp)
2469 {
2470         ir_type *res;
2471         size_t   i;
2472         size_t   n_params;
2473         size_t   n_results;
2474         bool     must_be_lowered;
2475
2476         res = pmap_get(ir_type, lowered_builtin_type_low, mtp);
2477         if (res != NULL)
2478                 return res;
2479
2480         n_params        = get_method_n_params(mtp);
2481         n_results       = get_method_n_ress(mtp);
2482         must_be_lowered = false;
2483
2484         /* check for double word parameter */
2485         for (i = n_params; i > 0;) {
2486                 ir_type *tp = get_method_param_type(mtp, --i);
2487
2488                 if (is_Primitive_type(tp)) {
2489                         ir_mode *mode = get_type_mode(tp);
2490
2491                         if (mode == env->high_signed || mode == env->high_unsigned) {
2492                                 must_be_lowered = true;
2493                                 break;
2494                         }
2495                 }
2496         }
2497
2498         if (!must_be_lowered) {
2499                 set_type_link(mtp, NULL);
2500                 return mtp;
2501         }
2502
2503         res = new_d_type_method(n_params, n_results, get_type_dbg_info(mtp));
2504
2505         /* set param types and result types */
2506         for (i = 0; i < n_params; ++i) {
2507                 ir_type *tp = get_method_param_type(mtp, i);
2508
2509                 if (is_Primitive_type(tp)) {
2510                         ir_mode *mode = get_type_mode(tp);
2511
2512                         if (mode == env->high_signed) {
2513                                 if (env->params->little_endian) {
2514                                         set_method_param_type(res, i, tp_s);
2515                                 } else {
2516                                         set_method_param_type(res, i, tp_u);
2517                                 }
2518                         } else if (mode == env->high_unsigned) {
2519                                 set_method_param_type(res, i, tp_u);
2520                         } else {
2521                                 set_method_param_type(res, i, tp);
2522                         }
2523                 } else {
2524                         set_method_param_type(res, i, tp);
2525                 }
2526         }
2527         for (i = 0; i < n_results; ++i) {
2528                 ir_type *tp = get_method_res_type(mtp, i);
2529
2530                 if (is_Primitive_type(tp)) {
2531                         ir_mode *mode = get_type_mode(tp);
2532
2533                         if (mode == env->high_signed) {
2534                                 if (env->params->little_endian) {
2535                                         set_method_res_type(res, i, tp_s);
2536                                 } else {
2537                                         set_method_res_type(res, i, tp_u);
2538                                 }
2539                         } else if (mode == env->high_unsigned) {
2540                                 set_method_res_type(res, i, tp_u);
2541                         } else {
2542                                 set_method_res_type(res, i, tp);
2543                         }
2544                 } else {
2545                         set_method_res_type(res, i, tp);
2546                 }
2547         }
2548
2549         set_method_variadicity(res, get_method_variadicity(mtp));
2550         set_method_calling_convention(res, get_method_calling_convention(mtp));
2551         set_method_additional_properties(res, get_method_additional_properties(mtp));
2552
2553         pmap_insert(lowered_builtin_type_low, mtp, res);
2554         return res;
2555 }
2556
2557 /**
2558  * lowers a builtin which reduces a 64bit value to a simple summary value
2559  * (popcount, ffs, ...)
2560  */
2561 static void lower_reduce_builtin(ir_node *builtin, ir_mode *mode)
2562 {
2563         ir_builtin_kind  kind         = get_Builtin_kind(builtin);
2564         ir_node         *operand      = get_Builtin_param(builtin, 0);
2565         ir_mode         *operand_mode = get_irn_mode(operand);
2566         if (operand_mode != env->high_signed && operand_mode != env->high_unsigned)
2567                 return;
2568
2569         {
2570         arch_allow_ifconv_func  allow_ifconv      = be_get_backend_param()->allow_ifconv;
2571         int                     arity             = get_irn_arity(builtin);
2572         dbg_info               *dbgi              = get_irn_dbg_info(builtin);
2573         ir_graph               *irg               = get_irn_irg(builtin);
2574         ir_type                *type              = get_Builtin_type(builtin);
2575         ir_type                *lowered_type_high = lower_Builtin_type_high(type);
2576         ir_type                *lowered_type_low  = lower_Builtin_type_low(type);
2577         ir_type                *result_type       = get_method_res_type(lowered_type_low, 0);
2578         ir_mode                *result_mode       = get_type_mode(result_type);
2579         ir_node                *block             = get_nodes_block(builtin);
2580         ir_node                *mem               = get_Builtin_mem(builtin);
2581         const lower64_entry_t  *entry             = get_node_entry(operand);
2582         ir_mode                *high_mode         = get_irn_mode(entry->high_word);
2583         ir_node                *in_high[1]        = {entry->high_word};
2584         ir_node                *in_low[1]         = {entry->low_word};
2585         ir_node                *res;
2586
2587         assert(is_NoMem(mem));
2588         assert(arity == 2);
2589
2590         switch (kind) {
2591         case ir_bk_ffs: {
2592                 ir_node               *number_of_bits = new_r_Const_long(irg, result_mode, get_mode_size_bits(env->low_unsigned));
2593                 ir_node               *zero_high      = new_rd_Const(dbgi, irg, get_mode_null(high_mode));
2594                 ir_node               *zero_unsigned  = new_rd_Const(dbgi, irg, get_mode_null(env->low_unsigned));
2595                 ir_node               *zero_result    = new_rd_Const(dbgi, irg, get_mode_null(result_mode));
2596                 ir_node               *cmp_low        = new_rd_Cmp(dbgi, block, entry->low_word, zero_unsigned, ir_relation_equal);
2597                 ir_node               *cmp_high       = new_rd_Cmp(dbgi, block, entry->high_word, zero_high, ir_relation_equal);
2598                 ir_node               *ffs_high       = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2599                 ir_node               *high_proj      = new_r_Proj(ffs_high, result_mode, pn_Builtin_max+1);
2600                 ir_node               *high           = new_rd_Add(dbgi, block, high_proj, number_of_bits, result_mode);
2601                 ir_node               *ffs_low        = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2602                 ir_node               *low            = new_r_Proj(ffs_low, result_mode, pn_Builtin_max+1);
2603                 ir_node               *mux_high       = new_rd_Mux(dbgi, block, cmp_high, high, zero_result, result_mode);
2604
2605                 if (! allow_ifconv(cmp_high, high, zero_result))
2606                         ir_nodeset_insert(&created_mux_nodes, mux_high);
2607
2608                 res = new_rd_Mux(dbgi, block, cmp_low, low, mux_high, result_mode);
2609
2610                 if (! allow_ifconv(cmp_low, low, mux_high))
2611                         ir_nodeset_insert(&created_mux_nodes, res);
2612                 break;
2613         }
2614         case ir_bk_clz: {
2615                 ir_node               *zero           = new_rd_Const(dbgi, irg, get_mode_null(high_mode));
2616                 ir_node               *cmp_high       = new_rd_Cmp(dbgi, block, entry->high_word, zero, ir_relation_equal);
2617                 ir_node               *clz_high       = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2618                 ir_node               *high           = new_r_Proj(clz_high, result_mode, pn_Builtin_max+1);
2619                 ir_node               *clz_low        = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2620                 ir_node               *low_proj       = new_r_Proj(clz_low, result_mode, pn_Builtin_max+1);
2621                 ir_node               *number_of_bits = new_r_Const_long(irg, result_mode, get_mode_size_bits(mode));
2622                 ir_node               *low            = new_rd_Add(dbgi, block, low_proj, number_of_bits, result_mode);
2623
2624                 res = new_rd_Mux(dbgi, block, cmp_high, high, low, result_mode);
2625
2626                 if (! allow_ifconv(cmp_high, high, low))
2627                         ir_nodeset_insert(&created_mux_nodes, res);
2628                 break;
2629         }
2630         case ir_bk_ctz: {
2631                 ir_node               *zero_unsigned  = new_rd_Const(dbgi, irg, get_mode_null(env->low_unsigned));
2632                 ir_node               *cmp_low        = new_rd_Cmp(dbgi, block, entry->low_word, zero_unsigned, ir_relation_equal);
2633                 ir_node               *ffs_high       = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2634                 ir_node               *high_proj      = new_r_Proj(ffs_high, result_mode, pn_Builtin_max+1);
2635                 ir_node               *number_of_bits = new_r_Const_long(irg, result_mode, get_mode_size_bits(env->low_unsigned));
2636                 ir_node               *high           = new_rd_Add(dbgi, block, high_proj, number_of_bits, result_mode);
2637                 ir_node               *ffs_low        = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2638                 ir_node               *low            = new_r_Proj(ffs_low, result_mode, pn_Builtin_max+1);
2639
2640                 res = new_rd_Mux(dbgi, block, cmp_low, low, high, result_mode);
2641
2642                 if (! allow_ifconv(cmp_low, low, high))
2643                         ir_nodeset_insert(&created_mux_nodes, res);
2644                 break;
2645         }
2646         case ir_bk_popcount: {
2647                 ir_node               *popcount_high = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2648                 ir_node               *popcount_low  = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2649                 ir_node               *high          = new_r_Proj(popcount_high, result_mode, pn_Builtin_max+1);
2650                 ir_node               *low           = new_r_Proj(popcount_low, result_mode, pn_Builtin_max+1);
2651
2652                 res = new_rd_Add(dbgi, block, high, low, result_mode);
2653                 break;
2654         }
2655         case ir_bk_parity: {
2656                 ir_node  *parity_high;
2657                 ir_node  *parity_low;
2658                 ir_node  *high;
2659                 ir_node  *low;
2660
2661                 assert(arity == 2);
2662
2663                 parity_high = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2664                 high        = new_r_Proj(parity_high, result_mode, pn_Builtin_max+1);
2665                 parity_low  = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2666                 low         = new_r_Proj(parity_low, result_mode, pn_Builtin_max+1);
2667                 res         = new_rd_Eor(dbgi, block, high, low, result_mode);
2668                 break;
2669         }
2670         default:
2671                 panic("unexpected builtin");
2672         }
2673
2674         turn_into_tuple(builtin, 2);
2675         set_irn_n(builtin, pn_Builtin_M, mem);
2676         set_irn_n(builtin, pn_Builtin_max+1, res);
2677         }
2678 }
2679
2680 /**
2681  * lowers builtins performing arithmetic (bswap)
2682  */
2683 static void lower_arithmetic_builtin(ir_node *builtin, ir_mode *mode)
2684 {
2685         ir_builtin_kind  kind         = get_Builtin_kind(builtin);
2686         ir_node         *operand      = get_Builtin_param(builtin, 0);
2687         ir_mode         *operand_mode = get_irn_mode(operand);
2688         (void) mode;
2689         if (operand_mode != env->high_signed && operand_mode != env->high_unsigned)
2690                 return;
2691
2692         {
2693         dbg_info              *dbgi              = get_irn_dbg_info(builtin);
2694         ir_type               *type              = get_Builtin_type(builtin);
2695         ir_type               *lowered_type_high = lower_Builtin_type_high(type);
2696         ir_type               *lowered_type_low  = lower_Builtin_type_low(type);
2697         ir_node               *block             = get_nodes_block(builtin);
2698         ir_node               *mem               = get_Builtin_mem(builtin);
2699         const lower64_entry_t *entry             = get_node_entry(operand);
2700         ir_mode               *mode_high         = get_irn_mode(entry->high_word);
2701         const ir_edge_t       *edge;
2702         const ir_edge_t       *next;
2703         ir_node               *res_high;
2704         ir_node               *res_low;
2705
2706         switch (kind) {
2707         case ir_bk_bswap: {
2708                 ir_node               *in_high[1] = { entry->high_word };
2709                 ir_node               *in_low[1]  = { entry->low_word };
2710                 ir_node               *swap_high  = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2711                 ir_node               *swap_low   = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2712                 ir_node               *high       = new_r_Proj(swap_high, mode_high, pn_Builtin_max+1);
2713                 ir_node               *low        = new_r_Proj(swap_low, env->low_unsigned, pn_Builtin_max+1);
2714                 if (mode_high == env->low_signed) {
2715                         res_high = new_rd_Conv(dbgi, block, low, env->low_signed);
2716                         res_low  = new_rd_Conv(dbgi, block, high, env->low_unsigned);
2717                 } else {
2718                         res_high = low;
2719                         res_low  = high;
2720                 }
2721                 break;
2722         }
2723         default:
2724                 panic("unexpected builtin");
2725         }
2726
2727         /* search result Proj */
2728         foreach_out_edge_safe(builtin, edge, next) {
2729                 ir_node *proj = get_edge_src_irn(edge);
2730                 if (!is_Proj(proj))
2731                         continue;
2732
2733                 if (get_Proj_proj(proj) == pn_Builtin_max+1) {
2734                         ir_set_dw_lowered(proj, res_low, res_high);
2735                 }
2736         }
2737         }
2738 }
2739
2740 /**
2741  * Lower double word builtins.
2742  */
2743 static void lower_Builtin(ir_node *builtin, ir_mode *mode)
2744 {
2745         ir_builtin_kind kind = get_Builtin_kind(builtin);
2746
2747         switch (kind) {
2748         case ir_bk_trap:
2749         case ir_bk_debugbreak:
2750         case ir_bk_return_address:
2751         case ir_bk_frame_address:
2752         case ir_bk_prefetch:
2753         case ir_bk_inport:
2754         case ir_bk_outport:
2755         case ir_bk_inner_trampoline:
2756                 /* Nothing to do. */
2757                 return;
2758         case ir_bk_bswap:
2759                 lower_arithmetic_builtin(builtin, mode);
2760                 return;
2761         case ir_bk_ffs:
2762         case ir_bk_clz:
2763         case ir_bk_ctz:
2764         case ir_bk_popcount:
2765         case ir_bk_parity:
2766                 lower_reduce_builtin(builtin, mode);
2767                 return;
2768         }
2769         panic("unknown builtin");
2770 }
2771
2772 /**
2773  * check for opcodes that must always be lowered.
2774  */
2775 static bool always_lower(unsigned code)
2776 {
2777         switch (code) {
2778         case iro_ASM:
2779         case iro_Builtin:
2780         case iro_Proj:
2781         case iro_Start:
2782         case iro_Call:
2783         case iro_Return:
2784         case iro_Cond:
2785         case iro_Conv:
2786         case iro_Sel:
2787                 return true;
2788         default:
2789                 return false;
2790         }
2791 }
2792
2793 /**
2794  * Compare two op_mode_entry_t's.
2795  */
2796 static int cmp_op_mode(const void *elt, const void *key, size_t size)
2797 {
2798         const op_mode_entry_t *e1 = (const op_mode_entry_t*)elt;
2799         const op_mode_entry_t *e2 = (const op_mode_entry_t*)key;
2800         (void) size;
2801
2802         return (e1->op != e2->op) | (e1->imode != e2->imode) | (e1->omode != e2->omode);
2803 }
2804
2805 /**
2806  * Compare two conv_tp_entry_t's.
2807  */
2808 static int cmp_conv_tp(const void *elt, const void *key, size_t size)
2809 {
2810         const conv_tp_entry_t *e1 = (const conv_tp_entry_t*)elt;
2811         const conv_tp_entry_t *e2 = (const conv_tp_entry_t*)key;
2812         (void) size;
2813
2814         return (e1->imode != e2->imode) | (e1->omode != e2->omode);
2815 }
2816
2817 /**
2818  * Enter a lowering function into an ir_op.
2819  */
2820 void ir_register_dw_lower_function(ir_op *op, lower_dw_func func)
2821 {
2822         op->ops.generic = (op_func)func;
2823 }
2824
2825 /* Determine which modes need to be lowered */
2826 static void setup_modes(void)
2827 {
2828         unsigned           size_bits           = env->params->doubleword_size;
2829         ir_mode           *doubleword_signed   = NULL;
2830         ir_mode           *doubleword_unsigned = NULL;
2831         size_t             n_modes             = ir_get_n_modes();
2832         ir_mode_arithmetic arithmetic;
2833         unsigned           modulo_shift;
2834         size_t             i;
2835
2836         /* search for doubleword modes... */
2837         for (i = 0; i < n_modes; ++i) {
2838                 ir_mode *mode = ir_get_mode(i);
2839                 if (!mode_is_int(mode))
2840                         continue;
2841                 if (get_mode_size_bits(mode) != size_bits)
2842                         continue;
2843                 if (mode_is_signed(mode)) {
2844                         if (doubleword_signed != NULL) {
2845                                 /* sigh - the lowerer should really just lower all mode with
2846                                  * size_bits it finds. Unfortunately this required a bigger
2847                                  * rewrite. */
2848                                 panic("multiple double word signed modes found");
2849                         }
2850                         doubleword_signed = mode;
2851                 } else {
2852                         if (doubleword_unsigned != NULL) {
2853                                 /* sigh - the lowerer should really just lower all mode with
2854                                  * size_bits it finds. Unfortunately this required a bigger
2855                                  * rewrite. */
2856                                 panic("multiple double word unsigned modes found");
2857                         }
2858                         doubleword_unsigned = mode;
2859                 }
2860         }
2861         if (doubleword_signed == NULL || doubleword_unsigned == NULL) {
2862                 panic("Couldn't find doubleword modes");
2863         }
2864
2865         arithmetic   = get_mode_arithmetic(doubleword_signed);
2866         modulo_shift = get_mode_modulo_shift(doubleword_signed);
2867
2868         assert(get_mode_size_bits(doubleword_unsigned) == size_bits);
2869         assert(size_bits % 2 == 0);
2870         assert(get_mode_sign(doubleword_signed) == 1);
2871         assert(get_mode_sign(doubleword_unsigned) == 0);
2872         assert(get_mode_sort(doubleword_signed) == irms_int_number);
2873         assert(get_mode_sort(doubleword_unsigned) == irms_int_number);
2874         assert(get_mode_arithmetic(doubleword_unsigned) == arithmetic);
2875         assert(get_mode_modulo_shift(doubleword_unsigned) == modulo_shift);
2876
2877         /* try to guess a sensible modulo shift for the new mode.
2878          * (This is IMO another indication that this should really be a node
2879          *  attribute instead of a mode thing) */
2880         if (modulo_shift == size_bits) {
2881                 modulo_shift = modulo_shift / 2;
2882         } else if (modulo_shift == 0) {
2883                 /* fine */
2884         } else {
2885                 panic("Don't know what new modulo shift to use for lowered doubleword mode");
2886         }
2887         size_bits /= 2;
2888
2889         /* produce lowered modes */
2890         env->high_signed   = doubleword_signed;
2891         env->high_unsigned = doubleword_unsigned;
2892         env->low_signed    = new_int_mode("WS", arithmetic, size_bits, 1,
2893                                           modulo_shift);
2894         env->low_unsigned  = new_int_mode("WU", arithmetic, size_bits, 0,
2895                                           modulo_shift);
2896 }
2897
2898 static void enqueue_preds(ir_node *node)
2899 {
2900         int arity = get_irn_arity(node);
2901         int i;
2902
2903         for (i = 0; i < arity; ++i) {
2904                 ir_node *pred = get_irn_n(node, i);
2905                 pdeq_putr(env->waitq, pred);
2906         }
2907 }
2908
2909 static void lower_node(ir_node *node)
2910 {
2911         int              arity;
2912         int              i;
2913         lower_dw_func    func;
2914         ir_op           *op;
2915         ir_mode         *mode;
2916         unsigned         idx;
2917         lower64_entry_t *entry;
2918
2919         if (irn_visited_else_mark(node))
2920                 return;
2921
2922         /* cycles are always broken at Phi and Block nodes. So we don't need special
2923          * magic in all the other lower functions */
2924         if (is_Block(node)) {
2925                 enqueue_preds(node);
2926                 return;
2927         } else if (is_Phi(node)) {
2928                 lower_Phi(node);
2929                 return;
2930         }
2931
2932         /* depth-first: descend into operands */
2933         if (!is_Block(node)) {
2934                 ir_node *block = get_nodes_block(node);
2935                 lower_node(block);
2936         }
2937
2938         if (!is_Cond(node)) {
2939                 arity = get_irn_arity(node);
2940                 for (i = 0; i < arity; ++i) {
2941                         ir_node *pred = get_irn_n(node, i);
2942                         lower_node(pred);
2943                 }
2944         }
2945
2946         op   = get_irn_op(node);
2947         func = (lower_dw_func) op->ops.generic;
2948         if (func == NULL)
2949                 return;
2950
2951         idx   = get_irn_idx(node);
2952         entry = idx < env->n_entries ? env->entries[idx] : NULL;
2953         if (entry != NULL || always_lower(get_irn_opcode(node))) {
2954                 mode = get_irn_op_mode(node);
2955                 if (mode == env->high_signed) {
2956                         mode = env->low_signed;
2957                 } else {
2958                         mode = env->low_unsigned;
2959                 }
2960                 DB((dbg, LEVEL_1, "  %+F\n", node));
2961                 func(node, mode);
2962         }
2963 }
2964
2965 static void clear_node_and_phi_links(ir_node *node, void *data)
2966 {
2967         (void) data;
2968         if (get_irn_mode(node) == mode_T) {
2969                 set_irn_link(node, node);
2970         } else {
2971                 set_irn_link(node, NULL);
2972         }
2973         if (is_Block(node))
2974                 set_Block_phis(node, NULL);
2975         else if (is_Phi(node))
2976                 set_Phi_next(node, NULL);
2977 }
2978
2979 static void lower_irg(ir_graph *irg)
2980 {
2981         ir_entity *ent;
2982         ir_type   *mtp;
2983         ir_type   *lowered_mtp;
2984         unsigned   n_idx;
2985
2986         obstack_init(&env->obst);
2987
2988         /* just here for debugging */
2989         current_ir_graph = irg;
2990         assure_edges(irg);
2991
2992         n_idx = get_irg_last_idx(irg);
2993         n_idx = n_idx + (n_idx >> 2);  /* add 25% */
2994         env->n_entries = n_idx;
2995         env->entries   = NEW_ARR_F(lower64_entry_t*, n_idx);
2996         memset(env->entries, 0, sizeof(env->entries[0]) * n_idx);
2997
2998         env->irg            = irg;
2999         env->flags          = 0;
3000
3001         ent = get_irg_entity(irg);
3002         mtp = get_entity_type(ent);
3003         lowered_mtp = lower_mtp(mtp);
3004
3005         if (lowered_mtp != mtp) {
3006                 set_entity_type(ent, lowered_mtp);
3007                 env->flags |= MUST_BE_LOWERED;
3008
3009                 fix_parameter_entities(irg, mtp);
3010         }
3011
3012         /* first step: link all nodes and allocate data */
3013         ir_reserve_resources(irg, IR_RESOURCE_PHI_LIST | IR_RESOURCE_IRN_LINK);
3014         visit_all_identities(irg, clear_node_and_phi_links, NULL);
3015         irg_walk_graph(irg, NULL, prepare_links_and_handle_rotl, env);
3016
3017         if (env->flags & MUST_BE_LOWERED) {
3018                 size_t i;
3019                 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
3020                 inc_irg_visited(irg);
3021
3022                 assert(pdeq_empty(env->waitq));
3023                 pdeq_putr(env->waitq, get_irg_end(irg));
3024
3025                 env->lowered_phis = NEW_ARR_F(ir_node*, 0);
3026                 while (!pdeq_empty(env->waitq)) {
3027                         ir_node *node = (ir_node*)pdeq_getl(env->waitq);
3028                         lower_node(node);
3029                 }
3030
3031                 /* we need to fixup phis */
3032                 for (i = 0; i < ARR_LEN(env->lowered_phis); ++i) {
3033                         ir_node *phi = env->lowered_phis[i];
3034                         fixup_phi(phi);
3035                 }
3036                 DEL_ARR_F(env->lowered_phis);
3037
3038
3039                 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
3040
3041                 if (env->flags & CF_CHANGED) {
3042                         /* control flow changed, dominance info is invalid */
3043                         clear_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE);
3044                 }
3045                 edges_deactivate(irg);
3046         }
3047
3048         ir_free_resources(irg, IR_RESOURCE_PHI_LIST | IR_RESOURCE_IRN_LINK);
3049
3050         DEL_ARR_F(env->entries);
3051         obstack_free(&env->obst, NULL);
3052 }
3053
3054 static const lwrdw_param_t *param;
3055
3056 void ir_prepare_dw_lowering(const lwrdw_param_t *new_param)
3057 {
3058         assert(new_param != NULL);
3059         FIRM_DBG_REGISTER(dbg, "firm.lower.dw");
3060
3061         param = new_param;
3062
3063         ir_clear_opcodes_generic_func();
3064         ir_register_dw_lower_function(op_ASM,     lower_ASM);
3065         ir_register_dw_lower_function(op_Add,     lower_binop);
3066         ir_register_dw_lower_function(op_And,     lower_And);
3067         ir_register_dw_lower_function(op_Bad,     lower_Bad);
3068         ir_register_dw_lower_function(op_Builtin, lower_Builtin);
3069         ir_register_dw_lower_function(op_Call,    lower_Call);
3070         ir_register_dw_lower_function(op_Cmp,     lower_Cmp);
3071         ir_register_dw_lower_function(op_Cond,    lower_Cond);
3072         ir_register_dw_lower_function(op_Const,   lower_Const);
3073         ir_register_dw_lower_function(op_Conv,    lower_Conv);
3074         ir_register_dw_lower_function(op_Div,     lower_Div);
3075         ir_register_dw_lower_function(op_Eor,     lower_Eor);
3076         ir_register_dw_lower_function(op_Load,    lower_Load);
3077         ir_register_dw_lower_function(op_Minus,   lower_unop);
3078         ir_register_dw_lower_function(op_Mod,     lower_Mod);
3079         ir_register_dw_lower_function(op_Mul,     lower_binop);
3080         ir_register_dw_lower_function(op_Mux,     lower_Mux);
3081         ir_register_dw_lower_function(op_Not,     lower_Not);
3082         ir_register_dw_lower_function(op_Or,      lower_Or);
3083         ir_register_dw_lower_function(op_Proj,    lower_Proj);
3084         ir_register_dw_lower_function(op_Return,  lower_Return);
3085         ir_register_dw_lower_function(op_Shl,     lower_Shl);
3086         ir_register_dw_lower_function(op_Shr,     lower_Shr);
3087         ir_register_dw_lower_function(op_Shrs,    lower_Shrs);
3088         ir_register_dw_lower_function(op_Start,   lower_Start);
3089         ir_register_dw_lower_function(op_Store,   lower_Store);
3090         ir_register_dw_lower_function(op_Sub,     lower_binop);
3091         ir_register_dw_lower_function(op_Switch,  lower_Switch);
3092         ir_register_dw_lower_function(op_Unknown, lower_Unknown);
3093 }
3094
3095 /**
3096  * Callback to lower only the Mux nodes we created.
3097  */
3098 static int lower_mux_cb(ir_node *mux)
3099 {
3100         return ir_nodeset_contains(&created_mux_nodes, mux);
3101 }
3102
3103 /*
3104  * Do the lowering.
3105  */
3106 void ir_lower_dw_ops(void)
3107 {
3108         lower_dw_env_t lenv;
3109         size_t      i, n;
3110
3111         memset(&lenv, 0, sizeof(lenv));
3112         lenv.params = param;
3113         env = &lenv;
3114
3115         setup_modes();
3116
3117         /* create the necessary maps */
3118         if (! intrinsic_fkt)
3119                 intrinsic_fkt = new_set(cmp_op_mode, iro_Last + 1);
3120         if (! conv_types)
3121                 conv_types = new_set(cmp_conv_tp, 16);
3122         if (! lowered_type)
3123                 lowered_type = pmap_create();
3124         if (! lowered_builtin_type_low)
3125                 lowered_builtin_type_low = pmap_create();
3126         if (! lowered_builtin_type_high)
3127                 lowered_builtin_type_high = pmap_create();
3128
3129         /* create a primitive unsigned and signed type */
3130         if (! tp_u)
3131                 tp_u = get_type_for_mode(lenv.low_unsigned);
3132         if (! tp_s)
3133                 tp_s = get_type_for_mode(lenv.low_signed);
3134
3135         /* create method types for the created binop calls */
3136         if (! binop_tp_u) {
3137                 binop_tp_u = new_type_method(4, 2);
3138                 set_method_param_type(binop_tp_u, 0, tp_u);
3139                 set_method_param_type(binop_tp_u, 1, tp_u);
3140                 set_method_param_type(binop_tp_u, 2, tp_u);
3141                 set_method_param_type(binop_tp_u, 3, tp_u);
3142                 set_method_res_type(binop_tp_u, 0, tp_u);
3143                 set_method_res_type(binop_tp_u, 1, tp_u);
3144         }
3145         if (! binop_tp_s) {
3146                 binop_tp_s = new_type_method(4, 2);
3147                 if (env->params->little_endian) {
3148                         set_method_param_type(binop_tp_s, 0, tp_u);
3149                         set_method_param_type(binop_tp_s, 1, tp_s);
3150                         set_method_param_type(binop_tp_s, 2, tp_u);
3151                         set_method_param_type(binop_tp_s, 3, tp_s);
3152                         set_method_res_type(binop_tp_s, 0, tp_u);
3153                         set_method_res_type(binop_tp_s, 1, tp_s);
3154                 } else {
3155                         set_method_param_type(binop_tp_s, 0, tp_s);
3156                         set_method_param_type(binop_tp_s, 1, tp_u);
3157                         set_method_param_type(binop_tp_s, 2, tp_s);
3158                         set_method_param_type(binop_tp_s, 3, tp_u);
3159                         set_method_res_type(binop_tp_s, 0, tp_s);
3160                         set_method_res_type(binop_tp_s, 1, tp_u);
3161                 }
3162         }
3163         if (! unop_tp_u) {
3164                 unop_tp_u = new_type_method(2, 2);
3165                 set_method_param_type(unop_tp_u, 0, tp_u);
3166                 set_method_param_type(unop_tp_u, 1, tp_u);
3167                 set_method_res_type(unop_tp_u, 0, tp_u);
3168                 set_method_res_type(unop_tp_u, 1, tp_u);
3169         }
3170         if (! unop_tp_s) {
3171                 unop_tp_s = new_type_method(2, 2);
3172                 if (env->params->little_endian) {
3173                         set_method_param_type(unop_tp_s, 0, tp_u);
3174                         set_method_param_type(unop_tp_s, 1, tp_s);
3175                         set_method_res_type(unop_tp_s, 0, tp_u);
3176                         set_method_res_type(unop_tp_s, 1, tp_s);
3177                 } else {
3178                         set_method_param_type(unop_tp_s, 0, tp_s);
3179                         set_method_param_type(unop_tp_s, 1, tp_u);
3180                         set_method_res_type(unop_tp_s, 0, tp_s);
3181                         set_method_res_type(unop_tp_s, 1, tp_u);
3182                 }
3183         }
3184
3185         lenv.tv_mode_bytes = new_tarval_from_long(param->doubleword_size/(2*8), lenv.low_unsigned);
3186         lenv.tv_mode_bits  = new_tarval_from_long(param->doubleword_size/2, lenv.low_unsigned);
3187         lenv.waitq         = new_pdeq();
3188         lenv.first_id      = new_id_from_chars(param->little_endian ? ".l" : ".h", 2);
3189         lenv.next_id       = new_id_from_chars(param->little_endian ? ".h" : ".l", 2);
3190
3191         irp_reserve_resources(irp, IRP_RESOURCE_TYPE_LINK | IRP_RESOURCE_TYPE_VISITED);
3192         inc_master_type_visited();
3193         /* transform all graphs */
3194         for (i = 0, n = get_irp_n_irgs(); i < n; ++i) {
3195                 ir_graph *irg = get_irp_irg(i);
3196
3197                 ir_nodeset_init(&created_mux_nodes);
3198
3199                 lower_irg(irg);
3200
3201                 if (ir_nodeset_size(&created_mux_nodes) > 0)
3202                         lower_mux(irg, lower_mux_cb);
3203
3204                 ir_nodeset_destroy(&created_mux_nodes);
3205         }
3206         irp_free_resources(irp, IRP_RESOURCE_TYPE_LINK | IRP_RESOURCE_TYPE_VISITED);
3207         del_pdeq(lenv.waitq);
3208
3209         env = NULL;
3210 }
3211
3212 /* Default implementation. */
3213 ir_entity *def_create_intrinsic_fkt(ir_type *method, const ir_op *op,
3214                                     const ir_mode *imode, const ir_mode *omode,
3215                                     void *context)
3216 {
3217         char buf[64];
3218         ident *id;
3219         ir_entity *ent;
3220         (void) context;
3221
3222         if (imode == omode) {
3223                 snprintf(buf, sizeof(buf), "__l%s%s", get_op_name(op), get_mode_name(imode));
3224         } else {
3225                 snprintf(buf, sizeof(buf), "__l%s%s%s", get_op_name(op),
3226                         get_mode_name(imode), get_mode_name(omode));
3227         }
3228         id = new_id_from_str(buf);
3229
3230         ent = new_entity(get_glob_type(), id, method);
3231         set_entity_ld_ident(ent, get_entity_ident(ent));
3232         set_entity_visibility(ent, ir_visibility_external);
3233         return ent;
3234 }