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  *and;
 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         and        = 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, and, 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  *and;
 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         and        = 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, and, 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 bool is_equality_cmp(const ir_node *node)
1211 {
1212         ir_relation relation = get_Cmp_relation(node);
1213         ir_node    *left     = get_Cmp_left(node);
1214         ir_node    *right    = get_Cmp_right(node);
1215         ir_mode    *mode     = get_irn_mode(left);
1216
1217         /* this probably makes no sense if unordered is involved */
1218         assert(!mode_is_float(mode));
1219
1220         if (relation == ir_relation_equal || relation == ir_relation_less_greater)
1221                 return true;
1222
1223         if (!is_Const(right) || !is_Const_null(right))
1224                 return false;
1225         if (mode_is_signed(mode)) {
1226                 return relation == ir_relation_less_greater;
1227         } else {
1228                 return relation == ir_relation_greater;
1229         }
1230 }
1231
1232 static ir_node *get_cfop_destination(const ir_node *cfop)
1233 {
1234         const ir_edge_t *first = get_irn_out_edge_first(cfop);
1235         /* we should only have 1 destination */
1236         assert(get_irn_n_edges(cfop) == 1);
1237         return get_edge_src_irn(first);
1238 }
1239
1240 static void lower_Switch(ir_node *node, ir_mode *high_mode)
1241 {
1242         ir_node *selector = get_Switch_selector(node);
1243         ir_mode *mode     = get_irn_mode(selector);
1244         (void)high_mode;
1245         if (mode == env->high_signed || mode == env->high_unsigned) {
1246                 /* we can't really handle Switch with 64bit offsets */
1247                 panic("Switch with 64bit jumptable not supported");
1248         }
1249         lower_node(selector);
1250 }
1251
1252 /**
1253  * Translate a Cond.
1254  */
1255 static void lower_Cond(ir_node *node, ir_mode *high_mode)
1256 {
1257         ir_node *left, *right, *block;
1258         ir_node *sel = get_Cond_selector(node);
1259         ir_mode *cmp_mode;
1260         const lower64_entry_t *lentry, *rentry;
1261         ir_node  *projT = NULL, *projF = NULL;
1262         ir_node  *new_bl, *irn;
1263         ir_node  *projHF, *projHT;
1264         ir_node  *dst_blk;
1265         ir_relation relation;
1266         ir_graph *irg;
1267         dbg_info *dbg;
1268         const ir_edge_t *edge;
1269         const ir_edge_t *next;
1270
1271         (void) high_mode;
1272
1273         if (!is_Cmp(sel)) {
1274                 lower_node(sel);
1275                 return;
1276         }
1277
1278         left     = get_Cmp_left(sel);
1279         cmp_mode = get_irn_mode(left);
1280         if (cmp_mode != env->high_signed && cmp_mode != env->high_unsigned) {
1281                 lower_node(sel);
1282                 return;
1283         }
1284
1285         right  = get_Cmp_right(sel);
1286         lower_node(left);
1287         lower_node(right);
1288         lentry = get_node_entry(left);
1289         rentry = get_node_entry(right);
1290
1291         /* all right, build the code */
1292         foreach_out_edge_safe(node, edge, next) {
1293                 ir_node *proj    = get_edge_src_irn(edge);
1294                 long     proj_nr;
1295                 if (!is_Proj(proj))
1296                         continue;
1297                 proj_nr = get_Proj_proj(proj);
1298
1299                 if (proj_nr == pn_Cond_true) {
1300                         assert(projT == NULL && "more than one Proj(true)");
1301                         projT = proj;
1302                 } else {
1303                         assert(proj_nr == pn_Cond_false);
1304                         assert(projF == NULL && "more than one Proj(false)");
1305                         projF = proj;
1306                 }
1307                 mark_irn_visited(proj);
1308         }
1309         assert(projT && projF);
1310
1311         /* create a new high compare */
1312         block    = get_nodes_block(node);
1313         irg      = get_Block_irg(block);
1314         dbg      = get_irn_dbg_info(sel);
1315         relation = get_Cmp_relation(sel);
1316
1317         if (is_equality_cmp(sel)) {
1318                 /* x ==/!= y ==> or(x_low^y_low,x_high^y_high) ==/!= 0 */
1319                 ir_mode *mode       = env->low_unsigned;
1320                 ir_node *low_left   = new_rd_Conv(dbg, block, lentry->low_word, mode);
1321                 ir_node *high_left  = new_rd_Conv(dbg, block, lentry->high_word, mode);
1322                 ir_node *low_right  = new_rd_Conv(dbg, block, rentry->low_word, mode);
1323                 ir_node *high_right = new_rd_Conv(dbg, block, rentry->high_word, mode);
1324                 ir_node *xor_low    = new_rd_Eor(dbg, block, low_left, low_right, mode);
1325                 ir_node *xor_high   = new_rd_Eor(dbg, block, high_left, high_right, mode);
1326                 ir_node *ornode = new_rd_Or(dbg, block, xor_low, xor_high, mode);
1327                 ir_node *cmp    = new_rd_Cmp(dbg, block, ornode, new_r_Const(irg, get_mode_null(mode)), relation);
1328                 set_Cond_selector(node, cmp);
1329                 return;
1330         }
1331
1332         if (relation == ir_relation_equal) {
1333                 ir_node *proj;
1334                 /* simple case:a == b <==> a_h == b_h && a_l == b_l */
1335                 dst_blk = get_cfop_destination(projF);
1336
1337                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1338                                  ir_relation_equal);
1339                 dbg = get_irn_dbg_info(node);
1340                 irn = new_rd_Cond(dbg, block, irn);
1341
1342                 projHF = new_r_Proj(irn, mode_X, pn_Cond_false);
1343                 mark_irn_visited(projHF);
1344                 exchange(projF, projHF);
1345
1346                 projHT = new_r_Proj(irn, mode_X, pn_Cond_true);
1347                 mark_irn_visited(projHT);
1348
1349                 new_bl = new_r_Block(irg, 1, &projHT);
1350
1351                 dbg = get_irn_dbg_info(sel);
1352                 irn = new_rd_Cmp(dbg, new_bl, lentry->low_word, rentry->low_word,
1353                                   ir_relation_equal);
1354                 dbg = get_irn_dbg_info(node);
1355                 irn = new_rd_Cond(dbg, new_bl, irn);
1356
1357                 proj = new_r_Proj(irn, mode_X, pn_Cond_false);
1358                 mark_irn_visited(proj);
1359                 add_block_cf_input(dst_blk, projHF, proj);
1360
1361                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1362                 mark_irn_visited(proj);
1363                 exchange(projT, proj);
1364         } else if (relation == ir_relation_less_greater) {
1365                 ir_node *proj;
1366                 /* simple case:a != b <==> a_h != b_h || a_l != b_l */
1367                 dst_blk = get_cfop_destination(projT);
1368
1369                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1370                                  ir_relation_less_greater);
1371                 dbg = get_irn_dbg_info(node);
1372                 irn = new_rd_Cond(dbg, block, irn);
1373
1374                 projHT = new_r_Proj(irn, mode_X, pn_Cond_true);
1375                 mark_irn_visited(projHT);
1376                 exchange(projT, projHT);
1377
1378                 projHF = new_r_Proj(irn, mode_X, pn_Cond_false);
1379                 mark_irn_visited(projHF);
1380
1381                 new_bl = new_r_Block(irg, 1, &projHF);
1382
1383                 dbg = get_irn_dbg_info(sel);
1384                 irn = new_rd_Cmp(dbg, new_bl, lentry->low_word, rentry->low_word,
1385                                  ir_relation_less_greater);
1386                 dbg = get_irn_dbg_info(node);
1387                 irn = new_rd_Cond(dbg, new_bl, irn);
1388
1389                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1390                 mark_irn_visited(proj);
1391                 add_block_cf_input(dst_blk, projHT, proj);
1392
1393                 proj = new_r_Proj(irn, mode_X, pn_Cond_false);
1394                 mark_irn_visited(proj);
1395                 exchange(projF, proj);
1396         } else {
1397                 ir_node *proj;
1398                 /* a rel b <==> a_h REL b_h || (a_h == b_h && a_l rel b_l) */
1399                 ir_node *dstT, *dstF, *newbl_eq, *newbl_l;
1400                 ir_node *projEqF;
1401
1402                 dstT = get_cfop_destination(projT);
1403                 dstF = get_cfop_destination(projF);
1404
1405                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1406                                  relation & ~ir_relation_equal);
1407                 dbg = get_irn_dbg_info(node);
1408                 irn = new_rd_Cond(dbg, block, irn);
1409
1410                 projHT = new_r_Proj(irn, mode_X, pn_Cond_true);
1411                 mark_irn_visited(projHT);
1412
1413                 projHF = new_r_Proj(irn, mode_X, pn_Cond_false);
1414                 mark_irn_visited(projHF);
1415
1416                 newbl_eq = new_r_Block(irg, 1, &projHF);
1417
1418                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1419                                  ir_relation_equal);
1420                 irn = new_rd_Cond(dbg, newbl_eq, irn);
1421
1422                 projEqF = new_r_Proj(irn, mode_X, pn_Cond_false);
1423                 mark_irn_visited(projEqF);
1424
1425                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1426                 mark_irn_visited(proj);
1427
1428                 newbl_l = new_r_Block(irg, 1, &proj);
1429
1430                 dbg = get_irn_dbg_info(sel);
1431                 irn = new_rd_Cmp(dbg, newbl_l, lentry->low_word, rentry->low_word,
1432                                  relation);
1433                 dbg = get_irn_dbg_info(node);
1434                 irn = new_rd_Cond(dbg, newbl_l, irn);
1435
1436                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1437                 mark_irn_visited(proj);
1438                 add_block_cf_input(dstT, projT, proj);
1439
1440                 proj = new_r_Proj(irn, mode_X, pn_Cond_false);
1441                 mark_irn_visited(proj);
1442                 add_block_cf_input(dstF, projF, proj);
1443
1444                 exchange(projT, projHT);
1445                 exchange(projF, projEqF);
1446         }
1447
1448         /* we have changed the control flow */
1449         env->flags |= CF_CHANGED;
1450 }
1451
1452 /**
1453  * Translate a Conv to higher_signed
1454  */
1455 static void lower_Conv_to_Ll(ir_node *node)
1456 {
1457         ir_mode  *omode        = get_irn_mode(node);
1458         ir_node  *op           = get_Conv_op(node);
1459         ir_mode  *imode        = get_irn_mode(op);
1460         ir_graph *irg          = get_irn_irg(node);
1461         ir_node  *block        = get_nodes_block(node);
1462         dbg_info *dbg          = get_irn_dbg_info(node);
1463         ir_node  *res_low;
1464         ir_node  *res_high;
1465
1466         ir_mode  *low_unsigned = env->low_unsigned;
1467         ir_mode  *low_signed
1468                 = mode_is_signed(omode) ? env->low_signed : low_unsigned;
1469
1470         if (mode_is_int(imode) || mode_is_reference(imode)) {
1471                 if (imode == env->high_signed || imode == env->high_unsigned) {
1472                         /* a Conv from Lu to Ls or Ls to Lu */
1473                         const lower64_entry_t *op_entry = get_node_entry(op);
1474                         res_low  = op_entry->low_word;
1475                         res_high = new_rd_Conv(dbg, block, op_entry->high_word, low_signed);
1476                 } else {
1477                         /* simple case: create a high word */
1478                         if (imode != low_unsigned)
1479                                 op = new_rd_Conv(dbg, block, op, low_unsigned);
1480
1481                         res_low = op;
1482
1483                         if (mode_is_signed(imode)) {
1484                                 int      c       = get_mode_size_bits(low_signed) - 1;
1485                                 ir_node *cnst    = new_r_Const_long(irg, low_unsigned, c);
1486                                 if (get_irn_mode(op) != low_signed)
1487                                         op = new_rd_Conv(dbg, block, op, low_signed);
1488                                 res_high = new_rd_Shrs(dbg, block, op, cnst, low_signed);
1489                         } else {
1490                                 res_high = new_r_Const(irg, get_mode_null(low_signed));
1491                         }
1492                 }
1493         } else if (imode == mode_b) {
1494                 res_low  = new_rd_Conv(dbg, block, op, low_unsigned);
1495                 res_high = new_r_Const(irg, get_mode_null(low_signed));
1496         } else {
1497                 ir_node *irn, *call;
1498                 ir_type *mtp = get_conv_type(imode, omode);
1499
1500                 irn = get_intrinsic_address(mtp, get_irn_op(node), imode, omode);
1501                 call = new_rd_Call(dbg, block, get_irg_no_mem(irg), irn, 1, &op, mtp);
1502                 set_irn_pinned(call, get_irn_pinned(node));
1503                 irn = new_r_Proj(call, mode_T, pn_Call_T_result);
1504
1505                 if (env->params->little_endian) {
1506                         res_low  = new_r_Proj(irn, low_unsigned, 0);
1507                         res_high = new_r_Proj(irn, low_signed, 1);
1508                 } else {
1509                         res_low  = new_r_Proj(irn, low_unsigned, 1);
1510                         res_high = new_r_Proj(irn, low_signed,   0);
1511                 }
1512         }
1513         ir_set_dw_lowered(node, res_low, res_high);
1514 }
1515
1516 /**
1517  * Translate a Conv from higher_unsigned
1518  */
1519 static void lower_Conv_from_Ll(ir_node *node)
1520 {
1521         ir_node               *op    = get_Conv_op(node);
1522         ir_mode               *omode = get_irn_mode(node);
1523         ir_node               *block = get_nodes_block(node);
1524         dbg_info              *dbg   = get_irn_dbg_info(node);
1525         ir_graph              *irg   = get_irn_irg(node);
1526         const lower64_entry_t *entry = get_node_entry(op);
1527
1528         if (mode_is_int(omode) || mode_is_reference(omode)) {
1529                 op = entry->low_word;
1530
1531                 /* simple case: create a high word */
1532                 if (omode != env->low_unsigned)
1533                         op = new_rd_Conv(dbg, block, op, omode);
1534
1535                 set_Conv_op(node, op);
1536         } else if (omode == mode_b) {
1537                 /* llu ? true : false  <=> (low|high) ? true : false */
1538                 ir_mode *mode   = env->low_unsigned;
1539                 ir_node *ornode = new_rd_Or(dbg, block, entry->low_word,
1540                                             entry->high_word, mode);
1541                 set_Conv_op(node, ornode);
1542         } else {
1543                 ir_node *irn, *call, *in[2];
1544                 ir_mode *imode = get_irn_mode(op);
1545                 ir_type *mtp   = get_conv_type(imode, omode);
1546                 ir_node *res;
1547
1548                 irn   = get_intrinsic_address(mtp, get_irn_op(node), imode, omode);
1549                 if (env->params->little_endian) {
1550                         in[0] = entry->low_word;
1551                         in[1] = entry->high_word;
1552                 } else {
1553                         in[0] = entry->high_word;
1554                         in[1] = entry->low_word;
1555                 }
1556
1557                 call = new_rd_Call(dbg, block, get_irg_no_mem(irg), irn, 2, in, mtp);
1558                 set_irn_pinned(call, get_irn_pinned(node));
1559                 irn = new_r_Proj(call, mode_T, pn_Call_T_result);
1560                 res = new_r_Proj(irn, omode, 0);
1561
1562                 exchange(node, res);
1563         }
1564 }
1565
1566 /**
1567  * lower Cmp
1568  */
1569 static void lower_Cmp(ir_node *cmp, ir_mode *m)
1570 {
1571         ir_node  *l        = get_Cmp_left(cmp);
1572         ir_mode  *cmp_mode = get_irn_mode(l);
1573         ir_node  *r, *low, *high, *t, *res;
1574         ir_relation relation;
1575         ir_node  *block;
1576         dbg_info *dbg;
1577         const lower64_entry_t *lentry;
1578         const lower64_entry_t *rentry;
1579         (void) m;
1580
1581         if (cmp_mode != env->high_signed && cmp_mode != env->high_unsigned)
1582                 return;
1583
1584         r        = get_Cmp_right(cmp);
1585         lentry   = get_node_entry(l);
1586         rentry   = get_node_entry(r);
1587         relation = get_Cmp_relation(cmp);
1588         block    = get_nodes_block(cmp);
1589         dbg      = get_irn_dbg_info(cmp);
1590
1591         /* easy case for x ==/!= 0 (see lower_Cond for details) */
1592         if (is_equality_cmp(cmp)) {
1593                 ir_graph *irg        = get_irn_irg(cmp);
1594                 ir_mode  *mode       = env->low_unsigned;
1595                 ir_node  *low_left   = new_rd_Conv(dbg, block, lentry->low_word, mode);
1596                 ir_node  *high_left  = new_rd_Conv(dbg, block, lentry->high_word, mode);
1597                 ir_node  *low_right  = new_rd_Conv(dbg, block, rentry->low_word, mode);
1598                 ir_node  *high_right = new_rd_Conv(dbg, block, rentry->high_word, mode);
1599                 ir_node  *xor_low    = new_rd_Eor(dbg, block, low_left, low_right, mode);
1600                 ir_node  *xor_high   = new_rd_Eor(dbg, block, high_left, high_right, mode);
1601                 ir_node  *ornode     = new_rd_Or(dbg, block, xor_low, xor_high, mode);
1602                 ir_node  *new_cmp    = new_rd_Cmp(dbg, block, ornode, new_r_Const(irg, get_mode_null(mode)), relation);
1603                 exchange(cmp, new_cmp);
1604                 return;
1605         }
1606
1607         if (relation == ir_relation_equal) {
1608                 /* simple case:a == b <==> a_h == b_h && a_l == b_l */
1609                 low  = new_rd_Cmp(dbg, block, lentry->low_word, rentry->low_word,
1610                                   relation);
1611                 high = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1612                                   relation);
1613                 res  = new_rd_And(dbg, block, low, high, mode_b);
1614         } else if (relation == ir_relation_less_greater) {
1615                 /* simple case:a != b <==> a_h != b_h || a_l != b_l */
1616                 low  = new_rd_Cmp(dbg, block, lentry->low_word, rentry->low_word,
1617                                   relation);
1618                 high = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1619                                   relation);
1620                 res = new_rd_Or(dbg, block, low, high, mode_b);
1621         } else {
1622                 /* a rel b <==> a_h REL b_h || (a_h == b_h && a_l rel b_l) */
1623                 ir_node *high1 = new_rd_Cmp(dbg, block, lentry->high_word,
1624                         rentry->high_word, relation & ~ir_relation_equal);
1625                 low  = new_rd_Cmp(dbg, block, lentry->low_word, rentry->low_word,
1626                                   relation);
1627                 high = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1628                                   ir_relation_equal);
1629                 t = new_rd_And(dbg, block, low, high, mode_b);
1630                 res = new_rd_Or(dbg, block, high1, t, mode_b);
1631         }
1632         exchange(cmp, res);
1633 }
1634
1635 /**
1636  * Translate a Conv.
1637  */
1638 static void lower_Conv(ir_node *node, ir_mode *mode)
1639 {
1640         mode = get_irn_mode(node);
1641
1642         if (mode == env->high_signed || mode == env->high_unsigned) {
1643                 lower_Conv_to_Ll(node);
1644         } else {
1645                 ir_mode *op_mode = get_irn_mode(get_Conv_op(node));
1646
1647                 if (op_mode == env->high_signed || op_mode == env->high_unsigned) {
1648                         lower_Conv_from_Ll(node);
1649                 }
1650         }
1651 }
1652
1653 static void fix_parameter_entities(ir_graph *irg)
1654 {
1655         ir_entity *entity   = get_irg_entity(irg);
1656         ir_type   *mtp      = get_entity_type(entity);
1657         ir_type   *orig_mtp = get_type_link(mtp);
1658
1659         size_t      orig_n_params      = get_method_n_params(orig_mtp);
1660         ir_entity **parameter_entities;
1661
1662         parameter_entities = ALLOCANZ(ir_entity*, orig_n_params);
1663
1664         ir_type *frame_type = get_irg_frame_type(irg);
1665         size_t   n          = get_compound_n_members(frame_type);
1666         size_t   i;
1667         size_t   n_param;
1668
1669         /* collect parameter entities */
1670         for (i = 0; i < n; ++i) {
1671                 ir_entity *entity = get_compound_member(frame_type, i);
1672                 size_t     p;
1673                 if (!is_parameter_entity(entity))
1674                         continue;
1675                 p = get_entity_parameter_number(entity);
1676                 if (p == IR_VA_START_PARAMETER_NUMBER)
1677                         continue;
1678                 assert(p < orig_n_params);
1679                 assert(parameter_entities[p] == NULL);
1680                 parameter_entities[p] = entity;
1681         }
1682
1683         /* adjust indices */
1684         n_param = 0;
1685         for (i = 0; i < orig_n_params; ++i, ++n_param) {
1686                 ir_entity *entity = parameter_entities[i];
1687                 ir_type   *tp;
1688
1689                 if (entity != NULL)
1690                         set_entity_parameter_number(entity, n_param);
1691
1692                 tp = get_method_param_type(orig_mtp, i);
1693                 if (is_Primitive_type(tp)) {
1694                         ir_mode *mode = get_type_mode(tp);
1695                         if (mode == env->high_signed || mode == env->high_unsigned) {
1696                                 ++n_param;
1697                                 /* note that we do not change the type of the parameter
1698                                  * entities, as calling convention fixup later still needs to
1699                                  * know which is/was a lowered doubleword.
1700                                  * So we just mark/remember it for later */
1701                                 if (entity != NULL) {
1702                                         assert(entity->attr.parameter.doubleword_low_mode == NULL);
1703                                         entity->attr.parameter.doubleword_low_mode
1704                                                 = env->low_unsigned;
1705                                 }
1706                         }
1707                 }
1708         }
1709 }
1710
1711 /**
1712  * Lower the method type.
1713  *
1714  * @param env  the lower environment
1715  * @param mtp  the method type to lower
1716  *
1717  * @return the lowered type
1718  */
1719 static ir_type *lower_mtp(ir_type *mtp)
1720 {
1721         ir_type *res;
1722         size_t   i;
1723         size_t   orig_n_params;
1724         size_t   orig_n_res;
1725         size_t   n_param;
1726         size_t   n_res;
1727         bool     must_be_lowered;
1728
1729         res = (ir_type*)pmap_get(lowered_type, mtp);
1730         if (res != NULL)
1731                 return res;
1732
1733         orig_n_params   = get_method_n_params(mtp);
1734         orig_n_res      = get_method_n_ress(mtp);
1735         n_param         = orig_n_params;
1736         n_res           = orig_n_res;
1737         must_be_lowered = false;
1738
1739         /* count new number of params */
1740         for (i = orig_n_params; i > 0;) {
1741                 ir_type *tp = get_method_param_type(mtp, --i);
1742
1743                 if (is_Primitive_type(tp)) {
1744                         ir_mode *mode = get_type_mode(tp);
1745
1746                         if (mode == env->high_signed || mode == env->high_unsigned) {
1747                                 ++n_param;
1748                                 must_be_lowered = true;
1749                         }
1750                 }
1751         }
1752
1753         /* count new number of results */
1754         for (i = orig_n_res; i > 0;) {
1755                 ir_type *tp = get_method_res_type(mtp, --i);
1756
1757                 if (is_Primitive_type(tp)) {
1758                         ir_mode *mode = get_type_mode(tp);
1759
1760                         if (mode == env->high_signed || mode == env->high_unsigned) {
1761                                 ++n_res;
1762                                 must_be_lowered = true;
1763                         }
1764                 }
1765         }
1766         if (!must_be_lowered) {
1767                 set_type_link(mtp, NULL);
1768                 return mtp;
1769         }
1770
1771         res = new_d_type_method(n_param, n_res, get_type_dbg_info(mtp));
1772
1773         /* set param types and result types */
1774         for (i = n_param = 0; i < orig_n_params; ++i) {
1775                 ir_type *tp = get_method_param_type(mtp, i);
1776
1777                 if (is_Primitive_type(tp)) {
1778                         ir_mode *mode = get_type_mode(tp);
1779
1780                         if (mode == env->high_signed) {
1781                                 if (env->params->little_endian) {
1782                                         set_method_param_type(res, n_param++, tp_u);
1783                                         set_method_param_type(res, n_param++, tp_s);
1784                                 } else {
1785                                         set_method_param_type(res, n_param++, tp_s);
1786                                         set_method_param_type(res, n_param++, tp_u);
1787                                 }
1788                         } else if (mode == env->high_unsigned) {
1789                                 set_method_param_type(res, n_param++, tp_u);
1790                                 set_method_param_type(res, n_param++, tp_u);
1791                         } else {
1792                                 set_method_param_type(res, n_param, tp);
1793                                 ++n_param;
1794                         }
1795                 } else {
1796                         set_method_param_type(res, n_param, tp);
1797                         ++n_param;
1798                 }
1799         }
1800         for (i = n_res = 0; i < orig_n_res; ++i) {
1801                 ir_type *tp = get_method_res_type(mtp, i);
1802
1803                 if (is_Primitive_type(tp)) {
1804                         ir_mode *mode = get_type_mode(tp);
1805
1806                         if (mode == env->high_signed) {
1807                                 if (env->params->little_endian) {
1808                                         set_method_res_type(res, n_res++, tp_u);
1809                                         set_method_res_type(res, n_res++, tp_s);
1810                                 } else {
1811                                         set_method_res_type(res, n_res++, tp_s);
1812                                         set_method_res_type(res, n_res++, tp_u);
1813                                 }
1814                         } else if (mode == env->high_unsigned) {
1815                                 set_method_res_type(res, n_res++, tp_u);
1816                                 set_method_res_type(res, n_res++, tp_u);
1817                         } else {
1818                                 set_method_res_type(res, n_res++, tp);
1819                         }
1820                 } else {
1821                         set_method_res_type(res, n_res++, tp);
1822                 }
1823         }
1824
1825         set_method_variadicity(res, get_method_variadicity(mtp));
1826         set_method_calling_convention(res, get_method_calling_convention(mtp));
1827         set_method_additional_properties(res, get_method_additional_properties(mtp));
1828
1829         set_higher_type(res, mtp);
1830         set_type_link(res, mtp);
1831
1832         pmap_insert(lowered_type, mtp, res);
1833         return res;
1834 }
1835
1836 /**
1837  * Translate a Return.
1838  */
1839 static void lower_Return(ir_node *node, ir_mode *mode)
1840 {
1841         ir_graph  *irg = get_irn_irg(node);
1842         ir_entity *ent = get_irg_entity(irg);
1843         ir_type   *mtp = get_entity_type(ent);
1844         ir_node  **in;
1845         size_t     i, j, n;
1846         int        need_conv = 0;
1847         (void) mode;
1848
1849         /* check if this return must be lowered */
1850         for (i = 0, n = get_Return_n_ress(node); i < n; ++i) {
1851                 ir_node *pred  = get_Return_res(node, i);
1852                 ir_mode *rmode = get_irn_op_mode(pred);
1853
1854                 if (rmode == env->high_signed || rmode == env->high_unsigned)
1855                         need_conv = 1;
1856         }
1857         if (! need_conv)
1858                 return;
1859
1860         ent = get_irg_entity(irg);
1861         mtp = get_entity_type(ent);
1862
1863         /* create a new in array */
1864         NEW_ARR_A(ir_node *, in, get_method_n_ress(mtp) + 1);
1865         j = 0;
1866         in[j++] = get_Return_mem(node);
1867
1868         for (i = 0, n = get_Return_n_ress(node); i < n; ++i) {
1869                 ir_node *pred      = get_Return_res(node, i);
1870                 ir_mode *pred_mode = get_irn_mode(pred);
1871
1872                 if (pred_mode == env->high_signed || pred_mode == env->high_unsigned) {
1873                         const lower64_entry_t *entry = get_node_entry(pred);
1874                         if (env->params->little_endian) {
1875                                 in[j++] = entry->low_word;
1876                                 in[j++] = entry->high_word;
1877                         } else {
1878                                 in[j++] = entry->high_word;
1879                                 in[j++] = entry->low_word;
1880                         }
1881                 } else {
1882                         in[j++] = pred;
1883                 }
1884         }
1885         assert(j == get_method_n_ress(mtp)+1);
1886
1887         set_irn_in(node, j, in);
1888 }
1889
1890 /**
1891  * Translate the parameters.
1892  */
1893 static void lower_Start(ir_node *node, ir_mode *high_mode)
1894 {
1895         ir_graph  *irg      = get_irn_irg(node);
1896         ir_entity *ent      = get_irg_entity(irg);
1897         ir_type   *mtp      = get_entity_type(ent);
1898         ir_type   *orig_mtp = get_type_link(mtp);
1899         ir_node   *args;
1900         long      *new_projs;
1901         size_t    i, j, n_params;
1902         const ir_edge_t *edge;
1903         const ir_edge_t *next;
1904         (void) high_mode;
1905
1906         /* if type link is NULL then the type was not lowered, hence no changes
1907          * at Start necessary */
1908         if (orig_mtp == NULL)
1909                 return;
1910
1911         n_params = get_method_n_params(orig_mtp);
1912
1913         NEW_ARR_A(long, new_projs, n_params);
1914
1915         /* Calculate mapping of proj numbers in new_projs */
1916         for (i = j = 0; i < n_params; ++i, ++j) {
1917                 ir_type *ptp = get_method_param_type(orig_mtp, i);
1918
1919                 new_projs[i] = j;
1920                 if (is_Primitive_type(ptp)) {
1921                         ir_mode *amode = get_type_mode(ptp);
1922                         if (amode == env->high_signed || amode == env->high_unsigned)
1923                                 ++j;
1924                 }
1925         }
1926
1927         /* lower method type */
1928         args = NULL;
1929         foreach_out_edge(node, edge) {
1930                 ir_node *proj = get_edge_src_irn(edge);
1931                 if (!is_Proj(proj))
1932                         continue;
1933                 if (get_Proj_proj(proj) == pn_Start_T_args) {
1934                         args = proj;
1935                         break;
1936                 }
1937         }
1938         if (args == NULL)
1939                 return;
1940
1941         /* fix all Proj's and create new ones */
1942         foreach_out_edge_safe(args, edge, next) {
1943                 ir_node *proj   = get_edge_src_irn(edge);
1944                 ir_mode *mode   = get_irn_mode(proj);
1945                 ir_mode *mode_l = env->low_unsigned;
1946                 ir_node *pred;
1947                 long     proj_nr;
1948                 ir_mode *mode_h;
1949                 ir_node *res_low;
1950                 ir_node *res_high;
1951                 int      old_cse;
1952                 dbg_info *dbg;
1953
1954                 if (!is_Proj(proj))
1955                         continue;
1956                 pred    = get_Proj_pred(proj);
1957                 proj_nr = get_Proj_proj(proj);
1958
1959                 if (mode == env->high_signed) {
1960                         mode_h = env->low_signed;
1961                 } else if (mode == env->high_unsigned) {
1962                         mode_h = env->low_unsigned;
1963                 } else {
1964                         long new_pn = new_projs[proj_nr];
1965                         set_Proj_proj(proj, new_pn);
1966                         continue;
1967                 }
1968
1969                 /* Switch off CSE or we might get an already existing Proj. */
1970                 old_cse = get_opt_cse();
1971                 set_opt_cse(0);
1972                 dbg = get_irn_dbg_info(proj);
1973                 if (env->params->little_endian) {
1974                         res_low  = new_rd_Proj(dbg, pred, mode_l, new_projs[proj_nr]);
1975                         res_high = new_rd_Proj(dbg, pred, mode_h, new_projs[proj_nr] + 1);
1976                 } else {
1977                         res_high = new_rd_Proj(dbg, pred, mode_h, new_projs[proj_nr]);
1978                         res_low  = new_rd_Proj(dbg, pred, mode_l, new_projs[proj_nr] + 1);
1979                 }
1980                 set_opt_cse(old_cse);
1981                 ir_set_dw_lowered(proj, res_low, res_high);
1982         }
1983 }
1984
1985 /**
1986  * Translate a Call.
1987  */
1988 static void lower_Call(ir_node *node, ir_mode *mode)
1989 {
1990         ir_type  *tp = get_Call_type(node);
1991         ir_node  **in;
1992         size_t   n_params, n_res;
1993         bool     need_lower = false;
1994         size_t   i, j;
1995         size_t   p;
1996         long     *res_numbers = NULL;
1997         ir_node  *resproj;
1998         const ir_edge_t *edge;
1999         const ir_edge_t *next;
2000         (void) mode;
2001
2002         n_params = get_method_n_params(tp);
2003         for (p = 0; p < n_params; ++p) {
2004                 ir_type *ptp = get_method_param_type(tp, p);
2005
2006                 if (is_Primitive_type(ptp)) {
2007                         ir_mode *pmode = get_type_mode(ptp);
2008                         if (pmode == env->high_signed || pmode == env->high_unsigned) {
2009                                 need_lower = true;
2010                                 break;
2011                         }
2012                 }
2013         }
2014         n_res = get_method_n_ress(tp);
2015         if (n_res > 0) {
2016                 NEW_ARR_A(long, res_numbers, n_res);
2017
2018                 for (i = j = 0; i < n_res; ++i, ++j) {
2019                         ir_type *ptp = get_method_res_type(tp, i);
2020
2021                         res_numbers[i] = j;
2022                         if (is_Primitive_type(ptp)) {
2023                                 ir_mode *rmode = get_type_mode(ptp);
2024                                 if (rmode == env->high_signed || rmode == env->high_unsigned) {
2025                                         need_lower = true;
2026                                         ++j;
2027                                 }
2028                         }
2029                 }
2030         }
2031
2032         if (! need_lower)
2033                 return;
2034
2035         /* let's lower it */
2036         tp = lower_mtp(tp);
2037         set_Call_type(node, tp);
2038
2039         NEW_ARR_A(ir_node *, in, get_method_n_params(tp) + 2);
2040
2041         in[0] = get_Call_mem(node);
2042         in[1] = get_Call_ptr(node);
2043
2044         for (j = 2, i = 0; i < n_params; ++i) {
2045                 ir_node *pred      = get_Call_param(node, i);
2046                 ir_mode *pred_mode = get_irn_mode(pred);
2047
2048                 if (pred_mode == env->high_signed || pred_mode == env->high_unsigned) {
2049                         const lower64_entry_t *pred_entry = get_node_entry(pred);
2050                         if (env->params->little_endian) {
2051                                 in[j++] = pred_entry->low_word;
2052                                 in[j++] = pred_entry->high_word;
2053                         } else {
2054                                 in[j++] = pred_entry->high_word;
2055                                 in[j++] = pred_entry->low_word;
2056                         }
2057                 } else {
2058                         in[j++] = pred;
2059                 }
2060         }
2061
2062         set_irn_in(node, j, in);
2063
2064         /* find results T */
2065         resproj = NULL;
2066         foreach_out_edge(node, edge) {
2067                 ir_node *proj = get_edge_src_irn(edge);
2068                 if (!is_Proj(proj))
2069                         continue;
2070                 if (get_Proj_proj(proj) == pn_Call_T_result) {
2071                         resproj = proj;
2072                         break;
2073                 }
2074         }
2075         if (resproj == NULL)
2076                 return;
2077
2078         /* fix the results */
2079         foreach_out_edge_safe(resproj, edge, next) {
2080                 ir_node  *proj      = get_edge_src_irn(edge);
2081                 ir_mode  *proj_mode = get_irn_mode(proj);
2082                 ir_mode  *mode_l    = env->low_unsigned;
2083                 ir_node  *pred;
2084                 long      proj_nr;
2085                 ir_mode  *mode_h;
2086                 ir_node  *res_low;
2087                 ir_node  *res_high;
2088                 dbg_info *dbg;
2089
2090                 if (!is_Proj(proj))
2091                         continue;
2092                 pred    = get_Proj_pred(proj);
2093                 proj_nr = get_Proj_proj(proj);
2094
2095                 if (proj_mode == env->high_signed) {
2096                         mode_h = env->low_signed;
2097                 } else if (proj_mode == env->high_unsigned) {
2098                         mode_h = env->low_unsigned;
2099                 } else {
2100                         long new_nr = res_numbers[proj_nr];
2101                         set_Proj_proj(proj, new_nr);
2102                         continue;
2103                 }
2104
2105                 dbg = get_irn_dbg_info(proj);
2106                 if (env->params->little_endian) {
2107                         res_low  = new_rd_Proj(dbg, pred, mode_l, res_numbers[proj_nr]);
2108                         res_high = new_rd_Proj(dbg, pred, mode_h, res_numbers[proj_nr] + 1);
2109                 } else {
2110                         res_high = new_rd_Proj(dbg, pred, mode_h, res_numbers[proj_nr]);
2111                         res_low  = new_rd_Proj(dbg, pred, mode_l, res_numbers[proj_nr] + 1);
2112                 }
2113                 ir_set_dw_lowered(proj, res_low, res_high);
2114         }
2115 }
2116
2117 /**
2118  * Translate an Unknown into two.
2119  */
2120 static void lower_Unknown(ir_node *node, ir_mode *mode)
2121 {
2122         ir_mode  *low_mode = env->low_unsigned;
2123         ir_graph *irg      = get_irn_irg(node);
2124         ir_node  *res_low  = new_r_Unknown(irg, low_mode);
2125         ir_node  *res_high = new_r_Unknown(irg, mode);
2126         ir_set_dw_lowered(node, res_low, res_high);
2127 }
2128
2129 /**
2130  * Translate a Bad into two.
2131  */
2132 static void lower_Bad(ir_node *node, ir_mode *mode)
2133 {
2134         ir_mode  *low_mode = env->low_unsigned;
2135         ir_graph *irg      = get_irn_irg(node);
2136         ir_node  *res_low  = new_r_Bad(irg, low_mode);
2137         ir_node  *res_high = new_r_Bad(irg, mode);
2138         ir_set_dw_lowered(node, res_low, res_high);
2139 }
2140
2141 /**
2142  * Translate a Phi.
2143  *
2144  * First step: just create two templates
2145  */
2146 static void lower_Phi(ir_node *phi)
2147 {
2148         ir_mode  *mode = get_irn_mode(phi);
2149         int       i;
2150         int       arity;
2151         ir_node **in_l;
2152         ir_node **in_h;
2153         ir_node  *unk_l;
2154         ir_node  *unk_h;
2155         ir_node  *phi_l;
2156         ir_node  *phi_h;
2157         dbg_info *dbg;
2158         ir_node  *block;
2159         ir_graph *irg;
2160         ir_mode  *mode_l;
2161         ir_mode  *mode_h;
2162
2163         /* enqueue predecessors */
2164         arity = get_Phi_n_preds(phi);
2165         for (i = 0; i < arity; ++i) {
2166                 ir_node *pred = get_Phi_pred(phi, i);
2167                 pdeq_putr(env->waitq, pred);
2168         }
2169
2170         if (mode != env->high_signed && mode != env->high_unsigned)
2171                 return;
2172
2173         /* first create a new in array */
2174         NEW_ARR_A(ir_node *, in_l, arity);
2175         NEW_ARR_A(ir_node *, in_h, arity);
2176         irg    = get_irn_irg(phi);
2177         mode_l = env->low_unsigned;
2178         mode_h = mode == env->high_signed ? env->low_signed : env->low_unsigned;
2179         unk_l  = new_r_Dummy(irg, mode_l);
2180         unk_h  = new_r_Dummy(irg, mode_h);
2181         for (i = 0; i < arity; ++i) {
2182                 in_l[i] = unk_l;
2183                 in_h[i] = unk_h;
2184         }
2185
2186         dbg   = get_irn_dbg_info(phi);
2187         block = get_nodes_block(phi);
2188         phi_l = new_rd_Phi(dbg, block, arity, in_l, mode_l);
2189         phi_h = new_rd_Phi(dbg, block, arity, in_h, mode_h);
2190
2191         ir_set_dw_lowered(phi, phi_l, phi_h);
2192
2193         /* remember that we need to fixup the predecessors later */
2194         ARR_APP1(ir_node*, env->lowered_phis, phi);
2195 }
2196
2197 static void fixup_phi(ir_node *phi)
2198 {
2199         const lower64_entry_t *entry = get_node_entry(phi);
2200         ir_node               *phi_l = entry->low_word;
2201         ir_node               *phi_h = entry->high_word;
2202         int                    arity = get_Phi_n_preds(phi);
2203         int                    i;
2204
2205         /* exchange phi predecessors which are lowered by now */
2206         for (i = 0; i < arity; ++i) {
2207                 ir_node               *pred       = get_Phi_pred(phi, i);
2208                 const lower64_entry_t *pred_entry = get_node_entry(pred);
2209
2210                 set_Phi_pred(phi_l, i, pred_entry->low_word);
2211                 set_Phi_pred(phi_h, i, pred_entry->high_word);
2212         }
2213 }
2214
2215 /**
2216  * Translate a Mux.
2217  */
2218 static void lower_Mux(ir_node *mux, ir_mode *mode)
2219 {
2220         ir_node               *truen       = get_Mux_true(mux);
2221         ir_node               *falsen      = get_Mux_false(mux);
2222         ir_node               *sel         = get_Mux_sel(mux);
2223         const lower64_entry_t *true_entry  = get_node_entry(truen);
2224         const lower64_entry_t *false_entry = get_node_entry(falsen);
2225         ir_node               *true_l      = true_entry->low_word;
2226         ir_node               *true_h      = true_entry->high_word;
2227         ir_node               *false_l     = false_entry->low_word;
2228         ir_node               *false_h     = false_entry->high_word;
2229         dbg_info              *dbgi        = get_irn_dbg_info(mux);
2230         ir_node               *block       = get_nodes_block(mux);
2231         ir_node               *res_low
2232                 = new_rd_Mux(dbgi, block, sel, false_l, true_l, env->low_unsigned);
2233         ir_node               *res_high
2234                 = new_rd_Mux(dbgi, block, sel, false_h, true_h, mode);
2235         ir_set_dw_lowered(mux, res_low, res_high);
2236 }
2237
2238 /**
2239  * Translate an ASM node.
2240  */
2241 static void lower_ASM(ir_node *asmn, ir_mode *mode)
2242 {
2243         ir_mode           *high_signed        = env->high_signed;
2244         ir_mode           *high_unsigned      = env->high_unsigned;
2245         int                n_outs             = get_ASM_n_output_constraints(asmn);
2246         ir_asm_constraint *output_constraints = get_ASM_output_constraints(asmn);
2247         ir_asm_constraint *input_constraints  = get_ASM_input_constraints(asmn);
2248         unsigned           n_64bit_outs       = 0;
2249         int      i;
2250
2251         (void)mode;
2252
2253         for (i = get_irn_arity(asmn) - 1; i >= 0; --i) {
2254                 ir_node *op      = get_irn_n(asmn, i);
2255                 ir_mode *op_mode = get_irn_mode(op);
2256                 if (op_mode == high_signed || op_mode == high_unsigned) {
2257                         panic("lowering ASM 64bit input unimplemented");
2258                 }
2259         }
2260
2261         for (i = 0; i < n_outs; ++i) {
2262                 const ir_asm_constraint *constraint = &output_constraints[i];
2263                 if (constraint->mode == high_signed || constraint->mode == high_unsigned) {
2264                         const char *constr = get_id_str(constraint->constraint);
2265                         ++n_64bit_outs;
2266                         /* TODO: How to do this architecture neutral? This is very
2267                          * i386 specific... */
2268                         if (constr[0] != '=' || constr[1] != 'A') {
2269                                 panic("lowering ASM 64bit output only supports '=A' currently");
2270                         }
2271                 }
2272         }
2273
2274         if (n_64bit_outs == 0)
2275                 return;
2276
2277         {
2278                 dbg_info          *dbgi       = get_irn_dbg_info(asmn);
2279                 ir_node           *block      = get_nodes_block(asmn);
2280                 int                arity      = get_irn_arity(asmn);
2281                 ir_node          **in         = get_irn_in(asmn) + 1;
2282                 int                new_n_outs = 0;
2283                 int                n_clobber  = get_ASM_n_clobbers(asmn);
2284                 long              *proj_map   = ALLOCAN(long, n_outs);
2285                 ident            **clobbers   = get_ASM_clobbers(asmn);
2286                 ident             *asm_text   = get_ASM_text(asmn);
2287                 ir_asm_constraint *new_outputs
2288                         = ALLOCAN(ir_asm_constraint, n_outs+n_64bit_outs);
2289                 ir_node           *new_asm;
2290                 const ir_edge_t   *edge;
2291                 const ir_edge_t   *next;
2292
2293                 for (i = 0; i < n_outs; ++i) {
2294                         const ir_asm_constraint *constraint = &output_constraints[i];
2295                         if (constraint->mode == high_signed || constraint->mode == high_unsigned) {
2296                                 new_outputs[new_n_outs].pos        = constraint->pos;
2297                                 new_outputs[new_n_outs].constraint = new_id_from_str("=a");
2298                                 new_outputs[new_n_outs].mode       = env->low_unsigned;
2299                                 proj_map[i] = new_n_outs;
2300                                 ++new_n_outs;
2301                                 new_outputs[new_n_outs].pos        = constraint->pos;
2302                                 new_outputs[new_n_outs].constraint = new_id_from_str("=d");
2303                                 if (constraint->mode == high_signed)
2304                                         new_outputs[new_n_outs].mode = env->low_signed;
2305                                 else
2306                                         new_outputs[new_n_outs].mode = env->low_unsigned;
2307                                 ++new_n_outs;
2308                         } else {
2309                                 new_outputs[new_n_outs] = *constraint;
2310                                 proj_map[i] = new_n_outs;
2311                                 ++new_n_outs;
2312                         }
2313                 }
2314                 assert(new_n_outs == n_outs+(int)n_64bit_outs);
2315
2316                 new_asm = new_rd_ASM(dbgi, block, arity, in, input_constraints,
2317                                      new_n_outs, new_outputs, n_clobber, clobbers,
2318                                      asm_text);
2319
2320                 foreach_out_edge_safe(asmn, edge, next) {
2321                         ir_node *proj      = get_edge_src_irn(edge);
2322                         ir_mode *proj_mode = get_irn_mode(proj);
2323                         long     pn;
2324
2325                         if (!is_Proj(proj))
2326                                 continue;
2327                         pn = get_Proj_proj(proj);
2328
2329                         if (pn < n_outs)
2330                                 pn = proj_map[pn];
2331                         else
2332                                 pn = new_n_outs + pn - n_outs;
2333
2334                         if (proj_mode == high_signed || proj_mode == high_unsigned) {
2335                                 ir_mode *high_mode
2336                                         = proj_mode == high_signed ? env->low_signed : env->low_unsigned;
2337                                 ir_node *np_low  = new_r_Proj(new_asm, env->low_unsigned, pn);
2338                                 ir_node *np_high = new_r_Proj(new_asm, high_mode, pn+1);
2339                                 ir_set_dw_lowered(proj, np_low, np_high);
2340                         } else {
2341                                 ir_node *np = new_r_Proj(new_asm, proj_mode, pn);
2342                                 exchange(proj, np);
2343                         }
2344                 }
2345         }
2346 }
2347
2348 /**
2349  * Lower the builtin type to its higher part.
2350  *
2351  * @param mtp  the builtin type to lower
2352  *
2353  * @return the lowered type
2354  */
2355 static ir_type *lower_Builtin_type_high(ir_type *mtp)
2356 {
2357         ir_type *res;
2358         size_t   i;
2359         size_t   n_params;
2360         size_t   n_results;
2361         bool     must_be_lowered;
2362
2363         res = (ir_type*)pmap_get(lowered_builtin_type_high, mtp);
2364         if (res != NULL)
2365                 return res;
2366
2367         n_params        = get_method_n_params(mtp);
2368         n_results       = get_method_n_ress(mtp);
2369         must_be_lowered = false;
2370
2371         /* check for double word parameter */
2372         for (i = n_params; i > 0;) {
2373                 ir_type *tp = get_method_param_type(mtp, --i);
2374
2375                 if (is_Primitive_type(tp)) {
2376                         ir_mode *mode = get_type_mode(tp);
2377
2378                         if (mode == env->high_signed || mode == env->high_unsigned) {
2379                                 must_be_lowered = true;
2380                                 break;
2381                         }
2382                 }
2383         }
2384
2385         if (!must_be_lowered) {
2386                 set_type_link(mtp, NULL);
2387                 return mtp;
2388         }
2389
2390         res = new_d_type_method(n_params, n_results, get_type_dbg_info(mtp));
2391
2392         /* set param types and result types */
2393         for (i = 0; i < n_params; ++i) {
2394                 ir_type *tp = get_method_param_type(mtp, i);
2395
2396                 if (is_Primitive_type(tp)) {
2397                         ir_mode *mode = get_type_mode(tp);
2398
2399                         if (mode == env->high_signed) {
2400                                 if (env->params->little_endian) {
2401                                         set_method_param_type(res, i, tp_u);
2402                                 } else {
2403                                         set_method_param_type(res, i, tp_s);
2404                                 }
2405                         } else if (mode == env->high_unsigned) {
2406                                 set_method_param_type(res, i, tp_u);
2407                         } else {
2408                                 set_method_param_type(res, i, tp);
2409                         }
2410                 } else {
2411                         set_method_param_type(res, i, tp);
2412                 }
2413         }
2414         for (i = n_results = 0; i < n_results; ++i) {
2415                 ir_type *tp = get_method_res_type(mtp, i);
2416
2417                 if (is_Primitive_type(tp)) {
2418                         ir_mode *mode = get_type_mode(tp);
2419
2420                         if (mode == env->high_signed) {
2421                                 if (env->params->little_endian) {
2422                                         set_method_res_type(res, i, tp_u);
2423                                 } else {
2424                                         set_method_res_type(res, i, tp_s);
2425                                 }
2426                         } else if (mode == env->high_unsigned) {
2427                                 set_method_res_type(res, i, tp_u);
2428                         } else {
2429                                 set_method_res_type(res, i, tp);
2430                         }
2431                 } else {
2432                         set_method_res_type(res, i, tp);
2433                 }
2434         }
2435
2436         set_method_variadicity(res, get_method_variadicity(mtp));
2437         set_method_calling_convention(res, get_method_calling_convention(mtp));
2438         set_method_additional_properties(res, get_method_additional_properties(mtp));
2439
2440         pmap_insert(lowered_builtin_type_high, mtp, res);
2441         return res;
2442 }
2443
2444 /**
2445  * Lower the builtin type to its lower part.
2446  *
2447  * @param mtp  the builtin type to lower
2448  *
2449  * @return the lowered type
2450  */
2451 static ir_type *lower_Builtin_type_low(ir_type *mtp)
2452 {
2453         ir_type *res;
2454         size_t   i;
2455         size_t   n_params;
2456         size_t   n_results;
2457         bool     must_be_lowered;
2458
2459         res = (ir_type*)pmap_get(lowered_builtin_type_low, mtp);
2460         if (res != NULL)
2461                 return res;
2462
2463         n_params        = get_method_n_params(mtp);
2464         n_results       = get_method_n_ress(mtp);
2465         must_be_lowered = false;
2466
2467         /* check for double word parameter */
2468         for (i = n_params; i > 0;) {
2469                 ir_type *tp = get_method_param_type(mtp, --i);
2470
2471                 if (is_Primitive_type(tp)) {
2472                         ir_mode *mode = get_type_mode(tp);
2473
2474                         if (mode == env->high_signed || mode == env->high_unsigned) {
2475                                 must_be_lowered = true;
2476                                 break;
2477                         }
2478                 }
2479         }
2480
2481         if (!must_be_lowered) {
2482                 set_type_link(mtp, NULL);
2483                 return mtp;
2484         }
2485
2486         res = new_d_type_method(n_params, n_results, get_type_dbg_info(mtp));
2487
2488         /* set param types and result types */
2489         for (i = 0; i < n_params; ++i) {
2490                 ir_type *tp = get_method_param_type(mtp, i);
2491
2492                 if (is_Primitive_type(tp)) {
2493                         ir_mode *mode = get_type_mode(tp);
2494
2495                         if (mode == env->high_signed) {
2496                                 if (env->params->little_endian) {
2497                                         set_method_param_type(res, i, tp_s);
2498                                 } else {
2499                                         set_method_param_type(res, i, tp_u);
2500                                 }
2501                         } else if (mode == env->high_unsigned) {
2502                                 set_method_param_type(res, i, tp_u);
2503                         } else {
2504                                 set_method_param_type(res, i, tp);
2505                         }
2506                 } else {
2507                         set_method_param_type(res, i, tp);
2508                 }
2509         }
2510         for (i = 0; i < n_results; ++i) {
2511                 ir_type *tp = get_method_res_type(mtp, i);
2512
2513                 if (is_Primitive_type(tp)) {
2514                         ir_mode *mode = get_type_mode(tp);
2515
2516                         if (mode == env->high_signed) {
2517                                 if (env->params->little_endian) {
2518                                         set_method_res_type(res, i, tp_s);
2519                                 } else {
2520                                         set_method_res_type(res, i, tp_u);
2521                                 }
2522                         } else if (mode == env->high_unsigned) {
2523                                 set_method_res_type(res, i, tp_u);
2524                         } else {
2525                                 set_method_res_type(res, i, tp);
2526                         }
2527                 } else {
2528                         set_method_res_type(res, i, tp);
2529                 }
2530         }
2531
2532         set_method_variadicity(res, get_method_variadicity(mtp));
2533         set_method_calling_convention(res, get_method_calling_convention(mtp));
2534         set_method_additional_properties(res, get_method_additional_properties(mtp));
2535
2536         pmap_insert(lowered_builtin_type_low, mtp, res);
2537         return res;
2538 }
2539
2540 /**
2541  * lowers a builtin which reduces a 64bit value to a simple summary value
2542  * (popcount, ffs, ...)
2543  */
2544 static void lower_reduce_builtin(ir_node *builtin, ir_mode *mode)
2545 {
2546         ir_builtin_kind  kind         = get_Builtin_kind(builtin);
2547         ir_node         *operand      = get_Builtin_param(builtin, 0);
2548         ir_mode         *operand_mode = get_irn_mode(operand);
2549         if (operand_mode != env->high_signed && operand_mode != env->high_unsigned)
2550                 return;
2551
2552         {
2553         arch_allow_ifconv_func  allow_ifconv      = be_get_backend_param()->allow_ifconv;
2554         int                     arity             = get_irn_arity(builtin);
2555         dbg_info               *dbgi              = get_irn_dbg_info(builtin);
2556         ir_graph               *irg               = get_irn_irg(builtin);
2557         ir_type                *type              = get_Builtin_type(builtin);
2558         ir_type                *lowered_type_high = lower_Builtin_type_high(type);
2559         ir_type                *lowered_type_low  = lower_Builtin_type_low(type);
2560         ir_type                *result_type       = get_method_res_type(lowered_type_low, 0);
2561         ir_mode                *result_mode       = get_type_mode(result_type);
2562         ir_node                *block             = get_nodes_block(builtin);
2563         ir_node                *mem               = get_Builtin_mem(builtin);
2564         const lower64_entry_t  *entry             = get_node_entry(operand);
2565         ir_mode                *high_mode         = get_irn_mode(entry->high_word);
2566         ir_node                *in_high[1]        = {entry->high_word};
2567         ir_node                *in_low[1]         = {entry->low_word};
2568         ir_node                *res;
2569
2570         assert(is_NoMem(mem));
2571         assert(arity == 2);
2572
2573         switch (kind) {
2574         case ir_bk_ffs: {
2575                 ir_node               *number_of_bits = new_r_Const_long(irg, result_mode, get_mode_size_bits(env->low_unsigned));
2576                 ir_node               *zero_high      = new_rd_Const(dbgi, irg, get_mode_null(high_mode));
2577                 ir_node               *zero_unsigned  = new_rd_Const(dbgi, irg, get_mode_null(env->low_unsigned));
2578                 ir_node               *zero_result    = new_rd_Const(dbgi, irg, get_mode_null(result_mode));
2579                 ir_node               *cmp_low        = new_rd_Cmp(dbgi, block, entry->low_word, zero_unsigned, ir_relation_equal);
2580                 ir_node               *cmp_high       = new_rd_Cmp(dbgi, block, entry->high_word, zero_high, ir_relation_equal);
2581                 ir_node               *ffs_high       = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2582                 ir_node               *high_proj      = new_r_Proj(ffs_high, result_mode, pn_Builtin_max+1);
2583                 ir_node               *high           = new_rd_Add(dbgi, block, high_proj, number_of_bits, result_mode);
2584                 ir_node               *ffs_low        = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2585                 ir_node               *low            = new_r_Proj(ffs_low, result_mode, pn_Builtin_max+1);
2586                 ir_node               *mux_high       = new_rd_Mux(dbgi, block, cmp_high, high, zero_result, result_mode);
2587
2588                 if (! allow_ifconv(cmp_high, high, zero_result))
2589                         ir_nodeset_insert(&created_mux_nodes, mux_high);
2590
2591                 res = new_rd_Mux(dbgi, block, cmp_low, low, mux_high, result_mode);
2592
2593                 if (! allow_ifconv(cmp_low, low, mux_high))
2594                         ir_nodeset_insert(&created_mux_nodes, res);
2595                 break;
2596         }
2597         case ir_bk_clz: {
2598                 ir_node               *zero           = new_rd_Const(dbgi, irg, get_mode_null(high_mode));
2599                 ir_node               *cmp_high       = new_rd_Cmp(dbgi, block, entry->high_word, zero, ir_relation_equal);
2600                 ir_node               *clz_high       = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2601                 ir_node               *high           = new_r_Proj(clz_high, result_mode, pn_Builtin_max+1);
2602                 ir_node               *clz_low        = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2603                 ir_node               *low_proj       = new_r_Proj(clz_low, result_mode, pn_Builtin_max+1);
2604                 ir_node               *number_of_bits = new_r_Const_long(irg, result_mode, get_mode_size_bits(mode));
2605                 ir_node               *low            = new_rd_Add(dbgi, block, low_proj, number_of_bits, result_mode);
2606
2607                 res = new_rd_Mux(dbgi, block, cmp_high, high, low, result_mode);
2608
2609                 if (! allow_ifconv(cmp_high, high, low))
2610                         ir_nodeset_insert(&created_mux_nodes, res);
2611                 break;
2612         }
2613         case ir_bk_ctz: {
2614                 ir_node               *zero_unsigned  = new_rd_Const(dbgi, irg, get_mode_null(env->low_unsigned));
2615                 ir_node               *cmp_low        = new_rd_Cmp(dbgi, block, entry->low_word, zero_unsigned, ir_relation_equal);
2616                 ir_node               *ffs_high       = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2617                 ir_node               *high_proj      = new_r_Proj(ffs_high, result_mode, pn_Builtin_max+1);
2618                 ir_node               *number_of_bits = new_r_Const_long(irg, result_mode, get_mode_size_bits(env->low_unsigned));
2619                 ir_node               *high           = new_rd_Add(dbgi, block, high_proj, number_of_bits, result_mode);
2620                 ir_node               *ffs_low        = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2621                 ir_node               *low            = new_r_Proj(ffs_low, result_mode, pn_Builtin_max+1);
2622
2623                 res = new_rd_Mux(dbgi, block, cmp_low, low, high, result_mode);
2624
2625                 if (! allow_ifconv(cmp_low, low, high))
2626                         ir_nodeset_insert(&created_mux_nodes, res);
2627                 break;
2628         }
2629         case ir_bk_popcount: {
2630                 ir_node               *popcount_high = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2631                 ir_node               *popcount_low  = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2632                 ir_node               *high          = new_r_Proj(popcount_high, result_mode, pn_Builtin_max+1);
2633                 ir_node               *low           = new_r_Proj(popcount_low, result_mode, pn_Builtin_max+1);
2634
2635                 res = new_rd_Add(dbgi, block, high, low, result_mode);
2636                 break;
2637         }
2638         case ir_bk_parity: {
2639                 ir_node  *parity_high;
2640                 ir_node  *parity_low;
2641                 ir_node  *high;
2642                 ir_node  *low;
2643
2644                 assert(arity == 2);
2645
2646                 parity_high = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2647                 high        = new_r_Proj(parity_high, result_mode, pn_Builtin_max+1);
2648                 parity_low  = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2649                 low         = new_r_Proj(parity_low, result_mode, pn_Builtin_max+1);
2650                 res         = new_rd_Eor(dbgi, block, high, low, result_mode);
2651                 break;
2652         }
2653         default:
2654                 panic("unexpected builtin");
2655         }
2656
2657         turn_into_tuple(builtin, 2);
2658         set_irn_n(builtin, pn_Builtin_M, mem);
2659         set_irn_n(builtin, pn_Builtin_max+1, res);
2660         }
2661 }
2662
2663 /**
2664  * lowers builtins performing arithmetic (bswap)
2665  */
2666 static void lower_arithmetic_builtin(ir_node *builtin, ir_mode *mode)
2667 {
2668         ir_builtin_kind  kind         = get_Builtin_kind(builtin);
2669         ir_node         *operand      = get_Builtin_param(builtin, 0);
2670         ir_mode         *operand_mode = get_irn_mode(operand);
2671         (void) mode;
2672         if (operand_mode != env->high_signed && operand_mode != env->high_unsigned)
2673                 return;
2674
2675         {
2676         dbg_info              *dbgi              = get_irn_dbg_info(builtin);
2677         ir_type               *type              = get_Builtin_type(builtin);
2678         ir_type               *lowered_type_high = lower_Builtin_type_high(type);
2679         ir_type               *lowered_type_low  = lower_Builtin_type_low(type);
2680         ir_node               *block             = get_nodes_block(builtin);
2681         ir_node               *mem               = get_Builtin_mem(builtin);
2682         const lower64_entry_t *entry             = get_node_entry(operand);
2683         ir_mode               *mode_high         = get_irn_mode(entry->high_word);
2684         const ir_edge_t       *edge;
2685         const ir_edge_t       *next;
2686         ir_node               *res_high;
2687         ir_node               *res_low;
2688
2689         switch (kind) {
2690         case ir_bk_bswap: {
2691                 ir_node               *in_high[1] = { entry->high_word };
2692                 ir_node               *in_low[1]  = { entry->low_word };
2693                 ir_node               *swap_high  = new_rd_Builtin(dbgi, block, mem, 1, in_high, kind, lowered_type_high);
2694                 ir_node               *swap_low   = new_rd_Builtin(dbgi, block, mem, 1, in_low, kind, lowered_type_low);
2695                 ir_node               *high       = new_r_Proj(swap_high, mode_high, pn_Builtin_max+1);
2696                 ir_node               *low        = new_r_Proj(swap_low, env->low_unsigned, pn_Builtin_max+1);
2697                 if (mode_high == env->low_signed) {
2698                         res_high = new_rd_Conv(dbgi, block, low, env->low_signed);
2699                         res_low  = new_rd_Conv(dbgi, block, high, env->low_unsigned);
2700                 } else {
2701                         res_high = low;
2702                         res_low  = high;
2703                 }
2704                 break;
2705         }
2706         default:
2707                 panic("unexpected builtin");
2708         }
2709
2710         /* search result Proj */
2711         foreach_out_edge_safe(builtin, edge, next) {
2712                 ir_node *proj = get_edge_src_irn(edge);
2713                 if (!is_Proj(proj))
2714                         continue;
2715
2716                 if (get_Proj_proj(proj) == pn_Builtin_max+1) {
2717                         ir_set_dw_lowered(proj, res_low, res_high);
2718                 }
2719         }
2720         }
2721 }
2722
2723 /**
2724  * Lower double word builtins.
2725  */
2726 static void lower_Builtin(ir_node *builtin, ir_mode *mode)
2727 {
2728         ir_builtin_kind kind = get_Builtin_kind(builtin);
2729
2730         switch (kind) {
2731         case ir_bk_trap:
2732         case ir_bk_debugbreak:
2733         case ir_bk_return_address:
2734         case ir_bk_frame_address:
2735         case ir_bk_prefetch:
2736         case ir_bk_inport:
2737         case ir_bk_outport:
2738         case ir_bk_inner_trampoline:
2739                 /* Nothing to do. */
2740                 return;
2741         case ir_bk_bswap:
2742                 lower_arithmetic_builtin(builtin, mode);
2743                 return;
2744         case ir_bk_ffs:
2745         case ir_bk_clz:
2746         case ir_bk_ctz:
2747         case ir_bk_popcount:
2748         case ir_bk_parity:
2749                 lower_reduce_builtin(builtin, mode);
2750                 return;
2751         }
2752         panic("unknown builtin");
2753 }
2754
2755 /**
2756  * check for opcodes that must always be lowered.
2757  */
2758 static bool always_lower(unsigned code)
2759 {
2760         switch (code) {
2761         case iro_ASM:
2762         case iro_Builtin:
2763         case iro_Proj:
2764         case iro_Start:
2765         case iro_Call:
2766         case iro_Return:
2767         case iro_Cond:
2768         case iro_Conv:
2769         case iro_Sel:
2770                 return true;
2771         default:
2772                 return false;
2773         }
2774 }
2775
2776 /**
2777  * Compare two op_mode_entry_t's.
2778  */
2779 static int cmp_op_mode(const void *elt, const void *key, size_t size)
2780 {
2781         const op_mode_entry_t *e1 = (const op_mode_entry_t*)elt;
2782         const op_mode_entry_t *e2 = (const op_mode_entry_t*)key;
2783         (void) size;
2784
2785         return (e1->op != e2->op) | (e1->imode != e2->imode) | (e1->omode != e2->omode);
2786 }
2787
2788 /**
2789  * Compare two conv_tp_entry_t's.
2790  */
2791 static int cmp_conv_tp(const void *elt, const void *key, size_t size)
2792 {
2793         const conv_tp_entry_t *e1 = (const conv_tp_entry_t*)elt;
2794         const conv_tp_entry_t *e2 = (const conv_tp_entry_t*)key;
2795         (void) size;
2796
2797         return (e1->imode != e2->imode) | (e1->omode != e2->omode);
2798 }
2799
2800 /**
2801  * Enter a lowering function into an ir_op.
2802  */
2803 void ir_register_dw_lower_function(ir_op *op, lower_dw_func func)
2804 {
2805         op->ops.generic = (op_func)func;
2806 }
2807
2808 /* Determine which modes need to be lowered */
2809 static void setup_modes(void)
2810 {
2811         unsigned           size_bits           = env->params->doubleword_size;
2812         ir_mode           *doubleword_signed   = NULL;
2813         ir_mode           *doubleword_unsigned = NULL;
2814         size_t             n_modes             = ir_get_n_modes();
2815         ir_mode_arithmetic arithmetic;
2816         unsigned           modulo_shift;
2817         size_t             i;
2818
2819         /* search for doubleword modes... */
2820         for (i = 0; i < n_modes; ++i) {
2821                 ir_mode *mode = ir_get_mode(i);
2822                 if (!mode_is_int(mode))
2823                         continue;
2824                 if (get_mode_size_bits(mode) != size_bits)
2825                         continue;
2826                 if (mode_is_signed(mode)) {
2827                         if (doubleword_signed != NULL) {
2828                                 /* sigh - the lowerer should really just lower all mode with
2829                                  * size_bits it finds. Unfortunately this required a bigger
2830                                  * rewrite. */
2831                                 panic("multiple double word signed modes found");
2832                         }
2833                         doubleword_signed = mode;
2834                 } else {
2835                         if (doubleword_unsigned != NULL) {
2836                                 /* sigh - the lowerer should really just lower all mode with
2837                                  * size_bits it finds. Unfortunately this required a bigger
2838                                  * rewrite. */
2839                                 panic("multiple double word unsigned modes found");
2840                         }
2841                         doubleword_unsigned = mode;
2842                 }
2843         }
2844         if (doubleword_signed == NULL || doubleword_unsigned == NULL) {
2845                 panic("Couldn't find doubleword modes");
2846         }
2847
2848         arithmetic   = get_mode_arithmetic(doubleword_signed);
2849         modulo_shift = get_mode_modulo_shift(doubleword_signed);
2850
2851         assert(get_mode_size_bits(doubleword_unsigned) == size_bits);
2852         assert(size_bits % 2 == 0);
2853         assert(get_mode_sign(doubleword_signed) == 1);
2854         assert(get_mode_sign(doubleword_unsigned) == 0);
2855         assert(get_mode_sort(doubleword_signed) == irms_int_number);
2856         assert(get_mode_sort(doubleword_unsigned) == irms_int_number);
2857         assert(get_mode_arithmetic(doubleword_unsigned) == arithmetic);
2858         assert(get_mode_modulo_shift(doubleword_unsigned) == modulo_shift);
2859
2860         /* try to guess a sensible modulo shift for the new mode.
2861          * (This is IMO another indication that this should really be a node
2862          *  attribute instead of a mode thing) */
2863         if (modulo_shift == size_bits) {
2864                 modulo_shift = modulo_shift / 2;
2865         } else if (modulo_shift == 0) {
2866                 /* fine */
2867         } else {
2868                 panic("Don't know what new modulo shift to use for lowered doubleword mode");
2869         }
2870         size_bits /= 2;
2871
2872         /* produce lowered modes */
2873         env->high_signed   = doubleword_signed;
2874         env->high_unsigned = doubleword_unsigned;
2875         env->low_signed    = new_int_mode("WS", arithmetic, size_bits, 1,
2876                                           modulo_shift);
2877         env->low_unsigned  = new_int_mode("WU", arithmetic, size_bits, 0,
2878                                           modulo_shift);
2879 }
2880
2881 static void enqueue_preds(ir_node *node)
2882 {
2883         int arity = get_irn_arity(node);
2884         int i;
2885
2886         for (i = 0; i < arity; ++i) {
2887                 ir_node *pred = get_irn_n(node, i);
2888                 pdeq_putr(env->waitq, pred);
2889         }
2890 }
2891
2892 static void lower_node(ir_node *node)
2893 {
2894         int              arity;
2895         int              i;
2896         lower_dw_func    func;
2897         ir_op           *op;
2898         ir_mode         *mode;
2899         unsigned         idx;
2900         lower64_entry_t *entry;
2901
2902         if (irn_visited_else_mark(node))
2903                 return;
2904
2905         /* cycles are always broken at Phi and Block nodes. So we don't need special
2906          * magic in all the other lower functions */
2907         if (is_Block(node)) {
2908                 enqueue_preds(node);
2909                 return;
2910         } else if (is_Phi(node)) {
2911                 lower_Phi(node);
2912                 return;
2913         }
2914
2915         /* depth-first: descend into operands */
2916         if (!is_Block(node)) {
2917                 ir_node *block = get_nodes_block(node);
2918                 lower_node(block);
2919         }
2920
2921         if (!is_Cond(node)) {
2922                 arity = get_irn_arity(node);
2923                 for (i = 0; i < arity; ++i) {
2924                         ir_node *pred = get_irn_n(node, i);
2925                         lower_node(pred);
2926                 }
2927         }
2928
2929         op   = get_irn_op(node);
2930         func = (lower_dw_func) op->ops.generic;
2931         if (func == NULL)
2932                 return;
2933
2934         idx   = get_irn_idx(node);
2935         entry = idx < env->n_entries ? env->entries[idx] : NULL;
2936         if (entry != NULL || always_lower(get_irn_opcode(node))) {
2937                 mode = get_irn_op_mode(node);
2938                 if (mode == env->high_signed) {
2939                         mode = env->low_signed;
2940                 } else {
2941                         mode = env->low_unsigned;
2942                 }
2943                 DB((dbg, LEVEL_1, "  %+F\n", node));
2944                 func(node, mode);
2945         }
2946 }
2947
2948 static void clear_node_and_phi_links(ir_node *node, void *data)
2949 {
2950         (void) data;
2951         if (get_irn_mode(node) == mode_T) {
2952                 set_irn_link(node, node);
2953         } else {
2954                 set_irn_link(node, NULL);
2955         }
2956         if (is_Block(node))
2957                 set_Block_phis(node, NULL);
2958         else if (is_Phi(node))
2959                 set_Phi_next(node, NULL);
2960 }
2961
2962 static void lower_irg(ir_graph *irg)
2963 {
2964         ir_entity *ent;
2965         ir_type   *mtp;
2966         ir_type   *lowered_mtp;
2967         unsigned   n_idx;
2968
2969         obstack_init(&env->obst);
2970
2971         /* just here for debugging */
2972         current_ir_graph = irg;
2973         edges_assure(irg);
2974
2975         n_idx = get_irg_last_idx(irg);
2976         n_idx = n_idx + (n_idx >> 2);  /* add 25% */
2977         env->n_entries = n_idx;
2978         env->entries   = NEW_ARR_F(lower64_entry_t*, n_idx);
2979         memset(env->entries, 0, sizeof(env->entries[0]) * n_idx);
2980
2981         env->irg            = irg;
2982         env->flags          = 0;
2983
2984         ent = get_irg_entity(irg);
2985         mtp = get_entity_type(ent);
2986         lowered_mtp = lower_mtp(mtp);
2987
2988         if (lowered_mtp != mtp) {
2989                 set_entity_type(ent, lowered_mtp);
2990                 env->flags |= MUST_BE_LOWERED;
2991
2992                 fix_parameter_entities(irg);
2993         }
2994
2995         /* first step: link all nodes and allocate data */
2996         ir_reserve_resources(irg, IR_RESOURCE_PHI_LIST | IR_RESOURCE_IRN_LINK);
2997         visit_all_identities(irg, clear_node_and_phi_links, NULL);
2998         irg_walk_graph(irg, NULL, prepare_links_and_handle_rotl, env);
2999
3000         if (env->flags & MUST_BE_LOWERED) {
3001                 size_t i;
3002                 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
3003                 inc_irg_visited(irg);
3004
3005                 assert(pdeq_empty(env->waitq));
3006                 pdeq_putr(env->waitq, get_irg_end(irg));
3007
3008                 env->lowered_phis = NEW_ARR_F(ir_node*, 0);
3009                 while (!pdeq_empty(env->waitq)) {
3010                         ir_node *node = (ir_node*)pdeq_getl(env->waitq);
3011                         lower_node(node);
3012                 }
3013
3014                 /* we need to fixup phis */
3015                 for (i = 0; i < ARR_LEN(env->lowered_phis); ++i) {
3016                         ir_node *phi = env->lowered_phis[i];
3017                         fixup_phi(phi);
3018                 }
3019                 DEL_ARR_F(env->lowered_phis);
3020
3021
3022                 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
3023
3024                 if (env->flags & CF_CHANGED) {
3025                         /* control flow changed, dominance info is invalid */
3026                         clear_irg_state(irg, IR_GRAPH_STATE_CONSISTENT_DOMINANCE
3027                                            | IR_GRAPH_STATE_VALID_EXTENDED_BLOCKS);
3028                 }
3029                 edges_deactivate(irg);
3030         }
3031
3032         ir_free_resources(irg, IR_RESOURCE_PHI_LIST | IR_RESOURCE_IRN_LINK);
3033
3034         DEL_ARR_F(env->entries);
3035         obstack_free(&env->obst, NULL);
3036 }
3037
3038 static const lwrdw_param_t *param;
3039
3040 void ir_prepare_dw_lowering(const lwrdw_param_t *new_param)
3041 {
3042         assert(new_param != NULL);
3043         FIRM_DBG_REGISTER(dbg, "firm.lower.dw");
3044
3045         param = new_param;
3046
3047         ir_clear_opcodes_generic_func();
3048         ir_register_dw_lower_function(op_ASM,     lower_ASM);
3049         ir_register_dw_lower_function(op_Add,     lower_binop);
3050         ir_register_dw_lower_function(op_And,     lower_And);
3051         ir_register_dw_lower_function(op_Bad,     lower_Bad);
3052         ir_register_dw_lower_function(op_Builtin, lower_Builtin);
3053         ir_register_dw_lower_function(op_Call,    lower_Call);
3054         ir_register_dw_lower_function(op_Cmp,     lower_Cmp);
3055         ir_register_dw_lower_function(op_Cond,    lower_Cond);
3056         ir_register_dw_lower_function(op_Const,   lower_Const);
3057         ir_register_dw_lower_function(op_Conv,    lower_Conv);
3058         ir_register_dw_lower_function(op_Div,     lower_Div);
3059         ir_register_dw_lower_function(op_Eor,     lower_Eor);
3060         ir_register_dw_lower_function(op_Load,    lower_Load);
3061         ir_register_dw_lower_function(op_Minus,   lower_unop);
3062         ir_register_dw_lower_function(op_Mod,     lower_Mod);
3063         ir_register_dw_lower_function(op_Mul,     lower_binop);
3064         ir_register_dw_lower_function(op_Mux,     lower_Mux);
3065         ir_register_dw_lower_function(op_Not,     lower_Not);
3066         ir_register_dw_lower_function(op_Or,      lower_Or);
3067         ir_register_dw_lower_function(op_Return,  lower_Return);
3068         ir_register_dw_lower_function(op_Shl,     lower_Shl);
3069         ir_register_dw_lower_function(op_Shr,     lower_Shr);
3070         ir_register_dw_lower_function(op_Shrs,    lower_Shrs);
3071         ir_register_dw_lower_function(op_Start,   lower_Start);
3072         ir_register_dw_lower_function(op_Store,   lower_Store);
3073         ir_register_dw_lower_function(op_Sub,     lower_binop);
3074         ir_register_dw_lower_function(op_Switch,  lower_Switch);
3075         ir_register_dw_lower_function(op_Unknown, lower_Unknown);
3076 }
3077
3078 /**
3079  * Callback to lower only the Mux nodes we created.
3080  */
3081 static int lower_mux_cb(ir_node *mux)
3082 {
3083         return ir_nodeset_contains(&created_mux_nodes, mux);
3084 }
3085
3086 /*
3087  * Do the lowering.
3088  */
3089 void ir_lower_dw_ops(void)
3090 {
3091         lower_dw_env_t lenv;
3092         size_t      i, n;
3093
3094         memset(&lenv, 0, sizeof(lenv));
3095         lenv.params = param;
3096         env = &lenv;
3097
3098         setup_modes();
3099
3100         /* create the necessary maps */
3101         if (! intrinsic_fkt)
3102                 intrinsic_fkt = new_set(cmp_op_mode, iro_Last + 1);
3103         if (! conv_types)
3104                 conv_types = new_set(cmp_conv_tp, 16);
3105         if (! lowered_type)
3106                 lowered_type = pmap_create();
3107         if (! lowered_builtin_type_low)
3108                 lowered_builtin_type_low = pmap_create();
3109         if (! lowered_builtin_type_high)
3110                 lowered_builtin_type_high = pmap_create();
3111
3112         /* create a primitive unsigned and signed type */
3113         if (! tp_u)
3114                 tp_u = get_type_for_mode(lenv.low_unsigned);
3115         if (! tp_s)
3116                 tp_s = get_type_for_mode(lenv.low_signed);
3117
3118         /* create method types for the created binop calls */
3119         if (! binop_tp_u) {
3120                 binop_tp_u = new_type_method(4, 2);
3121                 set_method_param_type(binop_tp_u, 0, tp_u);
3122                 set_method_param_type(binop_tp_u, 1, tp_u);
3123                 set_method_param_type(binop_tp_u, 2, tp_u);
3124                 set_method_param_type(binop_tp_u, 3, tp_u);
3125                 set_method_res_type(binop_tp_u, 0, tp_u);
3126                 set_method_res_type(binop_tp_u, 1, tp_u);
3127         }
3128         if (! binop_tp_s) {
3129                 binop_tp_s = new_type_method(4, 2);
3130                 if (env->params->little_endian) {
3131                         set_method_param_type(binop_tp_s, 0, tp_u);
3132                         set_method_param_type(binop_tp_s, 1, tp_s);
3133                         set_method_param_type(binop_tp_s, 2, tp_u);
3134                         set_method_param_type(binop_tp_s, 3, tp_s);
3135                         set_method_res_type(binop_tp_s, 0, tp_u);
3136                         set_method_res_type(binop_tp_s, 1, tp_s);
3137                 } else {
3138                         set_method_param_type(binop_tp_s, 0, tp_s);
3139                         set_method_param_type(binop_tp_s, 1, tp_u);
3140                         set_method_param_type(binop_tp_s, 2, tp_s);
3141                         set_method_param_type(binop_tp_s, 3, tp_u);
3142                         set_method_res_type(binop_tp_s, 0, tp_s);
3143                         set_method_res_type(binop_tp_s, 1, tp_u);
3144                 }
3145         }
3146         if (! unop_tp_u) {
3147                 unop_tp_u = new_type_method(2, 2);
3148                 set_method_param_type(unop_tp_u, 0, tp_u);
3149                 set_method_param_type(unop_tp_u, 1, tp_u);
3150                 set_method_res_type(unop_tp_u, 0, tp_u);
3151                 set_method_res_type(unop_tp_u, 1, tp_u);
3152         }
3153         if (! unop_tp_s) {
3154                 unop_tp_s = new_type_method(2, 2);
3155                 if (env->params->little_endian) {
3156                         set_method_param_type(unop_tp_s, 0, tp_u);
3157                         set_method_param_type(unop_tp_s, 1, tp_s);
3158                         set_method_res_type(unop_tp_s, 0, tp_u);
3159                         set_method_res_type(unop_tp_s, 1, tp_s);
3160                 } else {
3161                         set_method_param_type(unop_tp_s, 0, tp_s);
3162                         set_method_param_type(unop_tp_s, 1, tp_u);
3163                         set_method_res_type(unop_tp_s, 0, tp_s);
3164                         set_method_res_type(unop_tp_s, 1, tp_u);
3165                 }
3166         }
3167
3168         lenv.tv_mode_bytes = new_tarval_from_long(param->doubleword_size/(2*8), lenv.low_unsigned);
3169         lenv.tv_mode_bits  = new_tarval_from_long(param->doubleword_size/2, lenv.low_unsigned);
3170         lenv.waitq         = new_pdeq();
3171         lenv.first_id      = new_id_from_chars(param->little_endian ? ".l" : ".h", 2);
3172         lenv.next_id       = new_id_from_chars(param->little_endian ? ".h" : ".l", 2);
3173
3174         irp_reserve_resources(irp, IRP_RESOURCE_TYPE_LINK);
3175         /* transform all graphs */
3176         for (i = 0, n = get_irp_n_irgs(); i < n; ++i) {
3177                 ir_graph *irg = get_irp_irg(i);
3178
3179                 ir_nodeset_init(&created_mux_nodes);
3180
3181                 lower_irg(irg);
3182
3183                 if (ir_nodeset_size(&created_mux_nodes) > 0)
3184                         lower_mux(irg, lower_mux_cb);
3185
3186                 ir_nodeset_destroy(&created_mux_nodes);
3187         }
3188         irp_free_resources(irp, IRP_RESOURCE_TYPE_LINK);
3189         del_pdeq(lenv.waitq);
3190
3191         env = NULL;
3192 }
3193
3194 /* Default implementation. */
3195 ir_entity *def_create_intrinsic_fkt(ir_type *method, const ir_op *op,
3196                                     const ir_mode *imode, const ir_mode *omode,
3197                                     void *context)
3198 {
3199         char buf[64];
3200         ident *id;
3201         ir_entity *ent;
3202         (void) context;
3203
3204         if (imode == omode) {
3205                 snprintf(buf, sizeof(buf), "__l%s%s", get_op_name(op), get_mode_name(imode));
3206         } else {
3207                 snprintf(buf, sizeof(buf), "__l%s%s%s", get_op_name(op),
3208                         get_mode_name(imode), get_mode_name(omode));
3209         }
3210         id = new_id_from_str(buf);
3211
3212         ent = new_entity(get_glob_type(), id, method);
3213         set_entity_ld_ident(ent, get_entity_ident(ent));
3214         set_entity_visibility(ent, ir_visibility_external);
3215         return ent;
3216 }