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