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_type   *l_mtp;              /**< lowered method type of the current method */
 106         ir_tarval *tv_mode_bytes;      /**< a tarval containing the number of bytes in the lowered modes */
 107         ir_tarval *tv_mode_bits;       /**< a tarval containing the number of bits in the lowered modes */
 108         pdeq      *waitq;              /**< a wait queue of all nodes that must be handled later */
 109         ir_node  **lowered_phis;       /**< list of lowered phis */
 110         ir_mode   *high_signed;        /**< doubleword signed type */
 111         ir_mode   *high_unsigned;      /**< doubleword unsigned type */
 112         ir_mode   *low_signed;         /**< word signed type */
 113         ir_mode   *low_unsigned;       /**< word unsigned type */
 114         ident     *first_id;           /**< .l for little and .h for big endian */
 115         ident     *next_id;            /**< .h for little and .l for big endian */
 116         const lwrdw_param_t *params;   /**< transformation parameter */
 117         unsigned flags;                /**< some flags */
 118         unsigned n_entries;            /**< number of entries */
 119         ir_type  *value_param_tp;      /**< the old value param type */
 120 } lower_dw_env_t;
 121
 122 static lower_dw_env_t *env;
 123
 124 static void lower_node(ir_node *node);
 125 static bool mtp_must_be_lowered(ir_type *mtp);
 126
 127 /**
 128  * Create a method type for a Conv emulation from imode to omode.
 129  */
 130 static ir_type *get_conv_type(ir_mode *imode, ir_mode *omode)
 131 {
 132         conv_tp_entry_t key, *entry;
 133         ir_type *mtd;
 134
 135         key.imode = imode;
 136         key.omode = omode;
 137         key.mtd   = NULL;
 138
 139         entry = (conv_tp_entry_t*)set_insert(conv_types, &key, sizeof(key), HASH_PTR(imode) ^ HASH_PTR(omode));
 140         if (! entry->mtd) {
 141                 int n_param = 1, n_res = 1;
 142
 143                 if (imode == env->high_signed || imode == env->high_unsigned)
 144                         n_param = 2;
 145                 if (omode == env->high_signed || omode == env->high_unsigned)
 146                         n_res = 2;
 147
 148                 /* create a new one */
 149                 mtd = new_type_method(n_param, n_res);
 150
 151                 /* set param types and result types */
 152                 n_param = 0;
 153                 if (imode == env->high_signed) {
 154                         set_method_param_type(mtd, n_param++, tp_u);
 155                         set_method_param_type(mtd, n_param++, tp_s);
 156                 } else if (imode == env->high_unsigned) {
 157                         set_method_param_type(mtd, n_param++, tp_u);
 158                         set_method_param_type(mtd, n_param++, tp_u);
 159                 } else {
 160                         ir_type *tp = get_type_for_mode(imode);
 161                         set_method_param_type(mtd, n_param++, tp);
 162                 }
 163
 164                 n_res = 0;
 165                 if (omode == env->high_signed) {
 166                         set_method_res_type(mtd, n_res++, tp_u);
 167                         set_method_res_type(mtd, n_res++, tp_s);
 168                 } else if (omode == env->high_unsigned) {
 169                         set_method_res_type(mtd, n_res++, tp_u);
 170                         set_method_res_type(mtd, n_res++, tp_u);
 171                 } else {
 172                         ir_type *tp = get_type_for_mode(omode);
 173                         set_method_res_type(mtd, n_res++, tp);
 174                 }
 175                 entry->mtd = mtd;
 176         } else {
 177                 mtd = entry->mtd;
 178         }
 179         return mtd;
 180 }
 181
 182 /**
 183  * Add an additional control flow input to a block.
 184  * Patch all Phi nodes. The new Phi inputs are copied from
 185  * old input number nr.
 186  */
 187 static void add_block_cf_input_nr(ir_node *block, int nr, ir_node *cf)
 188 {
 189         int i, arity = get_irn_arity(block);
 190         ir_node **in;
 191         const ir_edge_t *edge;
 192
 193         assert(nr < arity);
 194
 195         NEW_ARR_A(ir_node *, in, arity + 1);
 196         for (i = 0; i < arity; ++i)
 197                 in[i] = get_irn_n(block, i);
 198         in[i] = cf;
 199
 200         set_irn_in(block, i + 1, in);
 201
 202         foreach_out_edge(block, edge) {
 203                 ir_node *phi = get_edge_src_irn(edge);
 204                 if (!is_Phi(phi))
 205                         continue;
 206
 207                 for (i = 0; i < arity; ++i)
 208                         in[i] = get_irn_n(phi, i);
 209                 in[i] = in[nr];
 210                 set_irn_in(phi, i + 1, in);
 211         }
 212 }
 213
 214 /**
 215  * Add an additional control flow input to a block.
 216  * Patch all Phi nodes. The new Phi inputs are copied from
 217  * old input from cf tmpl.
 218  */
 219 static void add_block_cf_input(ir_node *block, ir_node *tmpl, ir_node *cf)
 220 {
 221         int i, arity = get_irn_arity(block);
 222         int nr = 0;
 223
 224         for (i = 0; i < arity; ++i) {
 225                 if (get_irn_n(block, i) == tmpl) {
 226                         nr = i;
 227                         break;
 228                 }
 229         }
 230         assert(i < arity);
 231         add_block_cf_input_nr(block, nr, cf);
 232 }
 233
 234 /**
 235  * Return the "operational" mode of a Firm node.
 236  */
 237 static ir_mode *get_irn_op_mode(ir_node *node)
 238 {
 239         switch (get_irn_opcode(node)) {
 240         case iro_Load:
 241                 return get_Load_mode(node);
 242         case iro_Store:
 243                 return get_irn_mode(get_Store_value(node));
 244         case iro_Div:
 245                 return get_irn_mode(get_Div_left(node));
 246         case iro_Mod:
 247                 return get_irn_mode(get_Mod_left(node));
 248         case iro_Cmp:
 249                 return get_irn_mode(get_Cmp_left(node));
 250         default:
 251                 return get_irn_mode(node);
 252         }
 253 }
 254
 255 /**
 256  * Walker, prepare the node links and determine which nodes need to be lowered
 257  * at all.
 258  */
 259 static void prepare_links(ir_node *node)
 260 {
 261         ir_mode         *mode = get_irn_op_mode(node);
 262         lower64_entry_t *link;
 263
 264         if (mode == env->high_signed || mode == env->high_unsigned) {
 265                 unsigned idx = get_irn_idx(node);
 266                 /* ok, found a node that will be lowered */
 267                 link = OALLOCZ(&env->obst, lower64_entry_t);
 268
 269                 if (idx >= env->n_entries) {
 270                         /* enlarge: this happens only for Rotl nodes which is RARELY */
 271                         unsigned old   = env->n_entries;
 272                         unsigned n_idx = idx + (idx >> 3);
 273
 274                         ARR_RESIZE(lower64_entry_t *, env->entries, n_idx);
 275                         memset(&env->entries[old], 0, (n_idx - old) * sizeof(env->entries[0]));
 276                         env->n_entries = n_idx;
 277                 }
 278                 env->entries[idx] = link;
 279                 env->flags |= MUST_BE_LOWERED;
 280         } else if (is_Conv(node)) {
 281                 /* Conv nodes have two modes */
 282                 ir_node *pred = get_Conv_op(node);
 283                 mode = get_irn_mode(pred);
 284
 285                 if (mode == env->high_signed || mode == env->high_unsigned) {
 286                         /* must lower this node either but don't need a link */
 287                         env->flags |= MUST_BE_LOWERED;
 288                 }
 289                 return;
 290         }
 291 }
 292
 293 lower64_entry_t *get_node_entry(ir_node *node)
 294 {
 295         unsigned idx = get_irn_idx(node);
 296         assert(idx < env->n_entries);
 297         return env->entries[idx];
 298 }
 299
 300 void ir_set_dw_lowered(ir_node *old, ir_node *new_low, ir_node *new_high)
 301 {
 302         lower64_entry_t *entry = get_node_entry(old);
 303         entry->low_word  = new_low;
 304         entry->high_word = new_high;
 305 }
 306
 307 /**
 308  * Translate a Constant: create two.
 309  */
 310 static void lower_Const(ir_node *node, ir_mode *mode)
 311 {
 312         ir_graph  *irg      = get_irn_irg(node);
 313         dbg_info  *dbg      = get_irn_dbg_info(node);
 314         ir_mode   *low_mode = env->low_unsigned;
 315         ir_tarval *tv       = get_Const_tarval(node);
 316         ir_tarval *tv_l     = tarval_convert_to(tv, low_mode);
 317         ir_node   *res_low  = new_rd_Const(dbg, irg, tv_l);
 318         ir_tarval *tv_shrs  = tarval_shrs(tv, env->tv_mode_bits);
 319         ir_tarval *tv_h     = tarval_convert_to(tv_shrs, mode);
 320         ir_node   *res_high = new_rd_Const(dbg, irg, tv_h);
 321
 322         ir_set_dw_lowered(node, res_low, res_high);
 323 }
 324
 325 /**
 326  * Translate a Load: create two.
 327  */
 328 static void lower_Load(ir_node *node, ir_mode *mode)
 329 {
 330         ir_mode    *low_mode = env->low_unsigned;
 331         ir_graph   *irg = get_irn_irg(node);
 332         ir_node    *adr = get_Load_ptr(node);
 333         ir_node    *mem = get_Load_mem(node);
 334         ir_node    *low, *high, *proj;
 335         dbg_info   *dbg;
 336         ir_node    *block = get_nodes_block(node);
 337         ir_cons_flags volatility = get_Load_volatility(node) == volatility_is_volatile
 338                                  ? cons_volatile : cons_none;
 339         const ir_edge_t *edge;
 340         const ir_edge_t *next;
 341
 342         if (env->params->little_endian) {
 343                 low  = adr;
 344                 high = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
 345         } else {
 346                 low  = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
 347                 high = adr;
 348         }
 349
 350         /* create two loads */
 351         dbg  = get_irn_dbg_info(node);
 352         low  = new_rd_Load(dbg, block, mem,  low,  low_mode, volatility);
 353         proj = new_r_Proj(low, mode_M, pn_Load_M);
 354         high = new_rd_Load(dbg, block, proj, high, mode, volatility);
 355
 356         foreach_out_edge_safe(node, edge, next) {
 357                 ir_node *proj = get_edge_src_irn(edge);
 358                 if (!is_Proj(proj))
 359                         continue;
 360
 361                 switch (get_Proj_proj(proj)) {
 362                 case pn_Load_M:         /* Memory result. */
 363                         /* put it to the second one */
 364                         set_Proj_pred(proj, high);
 365                         break;
 366                 case pn_Load_X_except:  /* Execution result if exception occurred. */
 367                         /* put it to the first one */
 368                         set_Proj_pred(proj, low);
 369                         break;
 370                 case pn_Load_res: {       /* Result of load operation. */
 371                         ir_node *res_low  = new_r_Proj(low,  low_mode, pn_Load_res);
 372                         ir_node *res_high = new_r_Proj(high, mode,     pn_Load_res);
 373                         ir_set_dw_lowered(proj, res_low, res_high);
 374                         break;
 375                 }
 376                 default:
 377                         assert(0 && "unexpected Proj number");
 378                 }
 379                 /* mark this proj: we have handled it already, otherwise we might fall
 380                  * into out new nodes. */
 381                 mark_irn_visited(proj);
 382         }
 383 }
 384
 385 /**
 386  * Translate a Store: create two.
 387  */
 388 static void lower_Store(ir_node *node, ir_mode *mode)
 389 {
 390         ir_graph              *irg;
 391         ir_node               *block, *adr, *mem;
 392         ir_node               *low, *high, *proj;
 393         dbg_info              *dbg;
 394         ir_node               *value = get_Store_value(node);
 395         const lower64_entry_t *entry = get_node_entry(value);
 396         ir_cons_flags volatility = get_Store_volatility(node) == volatility_is_volatile
 397                                    ? cons_volatile : cons_none;
 398         const ir_edge_t *edge;
 399         const ir_edge_t *next;
 400         (void) mode;
 401
 402         assert(entry);
 403
 404         if (! entry->low_word) {
 405                 /* not ready yet, wait */
 406                 pdeq_putr(env->waitq, node);
 407                 return;
 408         }
 409
 410         irg = get_irn_irg(node);
 411         adr = get_Store_ptr(node);
 412         mem = get_Store_mem(node);
 413         block = get_nodes_block(node);
 414
 415         if (env->params->little_endian) {
 416                 low  = adr;
 417                 high = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
 418         } else {
 419                 low  = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
 420                 high = adr;
 421         }
 422
 423         /* create two Stores */
 424         dbg = get_irn_dbg_info(node);
 425         low  = new_rd_Store(dbg, block, mem, low,  entry->low_word, volatility);
 426         proj = new_r_Proj(low, mode_M, pn_Store_M);
 427         high = new_rd_Store(dbg, block, proj, high, entry->high_word, volatility);
 428
 429         foreach_out_edge_safe(node, edge, next) {
 430                 ir_node *proj = get_edge_src_irn(edge);
 431                 if (!is_Proj(proj))
 432                         continue;
 433
 434                 switch (get_Proj_proj(proj)) {
 435                 case pn_Store_M:         /* Memory result. */
 436                         /* put it to the second one */
 437                         set_Proj_pred(proj, high);
 438                         break;
 439                 case pn_Store_X_except:  /* Execution result if exception occurred. */
 440                         /* put it to the first one */
 441                         set_Proj_pred(proj, low);
 442                         break;
 443                 default:
 444                         assert(0 && "unexpected Proj number");
 445                 }
 446                 /* mark this proj: we have handled it already, otherwise we might fall into
 447                  * out new nodes. */
 448                 mark_irn_visited(proj);
 449         }
 450 }
 451
 452 /**
 453  * Return a node containing the address of the intrinsic emulation function.
 454  *
 455  * @param method  the method type of the emulation function
 456  * @param op      the emulated ir_op
 457  * @param imode   the input mode of the emulated opcode
 458  * @param omode   the output mode of the emulated opcode
 459  * @param env     the lower environment
 460  */
 461 static ir_node *get_intrinsic_address(ir_type *method, ir_op *op,
 462                                       ir_mode *imode, ir_mode *omode)
 463 {
 464         symconst_symbol sym;
 465         ir_entity *ent;
 466         op_mode_entry_t key, *entry;
 467
 468         key.op    = op;
 469         key.imode = imode;
 470         key.omode = omode;
 471         key.ent   = NULL;
 472
 473         entry = (op_mode_entry_t*)set_insert(intrinsic_fkt, &key, sizeof(key),
 474                                 HASH_PTR(op) ^ HASH_PTR(imode) ^ (HASH_PTR(omode) << 8));
 475         if (! entry->ent) {
 476                 /* create a new one */
 477                 ent = env->params->create_intrinsic(method, op, imode, omode, env->params->ctx);
 478
 479                 assert(ent && "Intrinsic creator must return an entity");
 480                 entry->ent = ent;
 481         } else {
 482                 ent = entry->ent;
 483         }
 484         sym.entity_p = ent;
 485         return new_r_SymConst(env->irg, mode_P_code, sym, symconst_addr_ent);
 486 }
 487
 488 /**
 489  * Translate a Div.
 490  *
 491  * Create an intrinsic Call.
 492  */
 493 static void lower_Div(ir_node *node, ir_mode *mode)
 494 {
 495         ir_node         *left   = get_Div_left(node);
 496         ir_node         *right  = get_Div_right(node);
 497         ir_node         *block  = get_nodes_block(node);
 498         dbg_info        *dbgi   = get_irn_dbg_info(node);
 499         ir_type         *mtp    = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
 500         ir_mode         *opmode = get_irn_op_mode(node);
 501         ir_node         *addr
 502             = get_intrinsic_address(mtp, get_irn_op(node), opmode, opmode);
 503         ir_node         *in[4];
 504         ir_node         *call;
 505         ir_node         *resproj;
 506         const ir_edge_t *edge;
 507         const ir_edge_t *next;
 508
 509         if (env->params->little_endian) {
 510                 in[0] = get_lowered_low(left);
 511                 in[1] = get_lowered_high(left);
 512                 in[2] = get_lowered_low(right);
 513                 in[3] = get_lowered_high(right);
 514         } else {
 515                 in[0] = get_lowered_high(left);
 516                 in[1] = get_lowered_low(left);
 517                 in[2] = get_lowered_high(right);
 518                 in[3] = get_lowered_low(right);
 519         }
 520         call    = new_rd_Call(dbgi, block, get_Div_mem(node), addr, 4, in, mtp);
 521         resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
 522         set_irn_pinned(call, get_irn_pinned(node));
 523
 524         foreach_out_edge_safe(node, edge, next) {
 525                 ir_node *proj = get_edge_src_irn(edge);
 526                 if (!is_Proj(proj))
 527                         continue;
 528
 529                 switch (get_Proj_proj(proj)) {
 530                 case pn_Div_M:         /* Memory result. */
 531                         /* reroute to the call */
 532                         set_Proj_pred(proj, call);
 533                         set_Proj_proj(proj, pn_Call_M);
 534                         break;
 535                 case pn_Div_X_except:  /* Execution result if exception occurred. */
 536                         /* reroute to the call */
 537                         set_Proj_pred(proj, call);
 538                         set_Proj_proj(proj, pn_Call_X_except);
 539                         break;
 540                 case pn_Div_res:
 541                         if (env->params->little_endian) {
 542                                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
 543                                 ir_node *res_high = new_r_Proj(resproj, mode,              1);
 544                                 ir_set_dw_lowered(proj, res_low, res_high);
 545                         } else {
 546                                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
 547                                 ir_node *res_high = new_r_Proj(resproj, mode,              0);
 548                                 ir_set_dw_lowered(proj, res_low, res_high);
 549                         }
 550                         break;
 551                 default:
 552                         assert(0 && "unexpected Proj number");
 553                 }
 554                 /* mark this proj: we have handled it already, otherwise we might fall into
 555                  * out new nodes. */
 556                 mark_irn_visited(proj);
 557         }
 558 }
 559
 560 /**
 561  * Translate a Mod.
 562  *
 563  * Create an intrinsic Call.
 564  */
 565 static void lower_Mod(ir_node *node, ir_mode *mode)
 566 {
 567         ir_node         *left   = get_Mod_left(node);
 568         ir_node         *right  = get_Mod_right(node);
 569         dbg_info        *dbgi   = get_irn_dbg_info(node);
 570         ir_node         *block  = get_nodes_block(node);
 571         ir_type         *mtp    = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
 572         ir_mode         *opmode = get_irn_op_mode(node);
 573         ir_node         *addr
 574             = get_intrinsic_address(mtp, get_irn_op(node), opmode, opmode);
 575         ir_node         *in[4];
 576         ir_node         *call;
 577         ir_node         *resproj;
 578         const ir_edge_t *edge;
 579         const ir_edge_t *next;
 580
 581         if (env->params->little_endian) {
 582                 in[0] = get_lowered_low(left);
 583                 in[1] = get_lowered_high(left);
 584                 in[2] = get_lowered_low(right);
 585                 in[3] = get_lowered_high(right);
 586         } else {
 587                 in[0] = get_lowered_high(left);
 588                 in[1] = get_lowered_low(left);
 589                 in[2] = get_lowered_high(right);
 590                 in[3] = get_lowered_low(right);
 591         }
 592         call    = new_rd_Call(dbgi, block, get_Mod_mem(node), addr, 4, in, mtp);
 593         resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
 594         set_irn_pinned(call, get_irn_pinned(node));
 595
 596         foreach_out_edge_safe(node, edge, next) {
 597                 ir_node *proj = get_edge_src_irn(edge);
 598                 if (!is_Proj(proj))
 599                         continue;
 600
 601                 switch (get_Proj_proj(proj)) {
 602                 case pn_Mod_M:         /* Memory result. */
 603                         /* reroute to the call */
 604                         set_Proj_pred(proj, call);
 605                         set_Proj_proj(proj, pn_Call_M);
 606                         break;
 607                 case pn_Mod_X_except:  /* Execution result if exception occurred. */
 608                         /* reroute to the call */
 609                         set_Proj_pred(proj, call);
 610                         set_Proj_proj(proj, pn_Call_X_except);
 611                         break;
 612                 case pn_Mod_res:
 613                         if (env->params->little_endian) {
 614                                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
 615                                 ir_node *res_high = new_r_Proj(resproj, mode,              1);
 616                                 ir_set_dw_lowered(proj, res_low, res_high);
 617                         } else {
 618                                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
 619                                 ir_node *res_high = new_r_Proj(resproj, mode,              0);
 620                                 ir_set_dw_lowered(proj, res_low, res_high);
 621                         }
 622                         break;
 623                 default:
 624                         assert(0 && "unexpected Proj number");
 625                 }
 626                 /* mark this proj: we have handled it already, otherwise we might fall
 627                  * into out new nodes. */
 628                 mark_irn_visited(proj);
 629         }
 630 }
 631
 632 /**
 633  * Translate a binop.
 634  *
 635  * Create an intrinsic Call.
 636  */
 637 static void lower_binop(ir_node *node, ir_mode *mode)
 638 {
 639         ir_node  *left  = get_binop_left(node);
 640         ir_node  *right = get_binop_right(node);
 641         dbg_info *dbgi  = get_irn_dbg_info(node);
 642         ir_node  *block = get_nodes_block(node);
 643         ir_graph *irg   = get_irn_irg(block);
 644         ir_type  *mtp   = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
 645         ir_node  *addr  = get_intrinsic_address(mtp, get_irn_op(node), mode, mode);
 646         ir_node  *in[4];
 647         ir_node  *call;
 648         ir_node  *resproj;
 649
 650         if (env->params->little_endian) {
 651                 in[0] = get_lowered_low(left);
 652                 in[1] = get_lowered_high(left);
 653                 in[2] = get_lowered_low(right);
 654                 in[3] = get_lowered_high(right);
 655         } else {
 656                 in[0] = get_lowered_high(left);
 657                 in[1] = get_lowered_low(left);
 658                 in[2] = get_lowered_high(right);
 659                 in[3] = get_lowered_low(right);
 660         }
 661         call    = new_rd_Call(dbgi, block, get_irg_no_mem(irg), addr, 4, in, mtp);
 662         resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
 663         set_irn_pinned(call, get_irn_pinned(node));
 664
 665         if (env->params->little_endian) {
 666                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
 667                 ir_node *res_high = new_r_Proj(resproj, mode,              1);
 668                 ir_set_dw_lowered(node, res_low, res_high);
 669         } else {
 670                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
 671                 ir_node *res_high = new_r_Proj(resproj, mode,              0);
 672                 ir_set_dw_lowered(node, res_low, res_high);
 673         }
 674 }
 675
 676 static ir_node *create_conv(ir_node *block, ir_node *node, ir_mode *dest_mode)
 677 {
 678         if (get_irn_mode(node) == dest_mode)
 679                 return node;
 680         return new_r_Conv(block, node, dest_mode);
 681 }
 682
 683 /**
 684  * Moves node and all predecessors of node from from_bl to to_bl.
 685  * Does not move predecessors of Phi nodes (or block nodes).
 686  */
 687 static void move(ir_node *node, ir_node *from_bl, ir_node *to_bl)
 688 {
 689         int i, arity;
 690
 691         /* move this node */
 692         set_nodes_block(node, to_bl);
 693
 694         /* move its Projs */
 695         if (get_irn_mode(node) == mode_T) {
 696                 const ir_edge_t *edge;
 697                 foreach_out_edge(node, edge) {
 698                         ir_node *proj = get_edge_src_irn(edge);
 699                         if (!is_Proj(proj))
 700                                 continue;
 701                         move(proj, from_bl, to_bl);
 702                 }
 703         }
 704
 705         /* We must not move predecessors of Phi nodes, even if they are in
 706          * from_bl. (because these are values from an earlier loop iteration
 707          * which are not predecessors of node here)
 708          */
 709         if (is_Phi(node))
 710                 return;
 711
 712         /* recursion ... */
 713         arity = get_irn_arity(node);
 714         for (i = 0; i < arity; i++) {
 715                 ir_node *pred      = get_irn_n(node, i);
 716                 ir_mode *pred_mode = get_irn_mode(pred);
 717                 if (get_nodes_block(pred) == from_bl)
 718                         move(pred, from_bl, to_bl);
 719                 if (pred_mode == env->high_signed || pred_mode == env->high_unsigned) {
 720                         ir_node *pred_low  = get_lowered_low(pred);
 721                         ir_node *pred_high = get_lowered_high(pred);
 722                         if (get_nodes_block(pred_low) == from_bl)
 723                                 move(pred_low, from_bl, to_bl);
 724                         if (pred_high != NULL && get_nodes_block(pred_high) == from_bl)
 725                                 move(pred_high, from_bl, to_bl);
 726                 }
 727         }
 728 }
 729
 730 /**
 731  * We need a custom version of part_block_edges because during transformation
 732  * not all data-dependencies are explicit yet if a lowered nodes users are not
 733  * lowered yet.
 734  * We can fix this by modifying move to look for such implicit dependencies.
 735  * Additionally we have to keep the proj_2_block map updated
 736  */
 737 static ir_node *part_block_dw(ir_node *node)
 738 {
 739         ir_graph *irg        = get_irn_irg(node);
 740         ir_node  *old_block  = get_nodes_block(node);
 741         int       n_cfgpreds = get_Block_n_cfgpreds(old_block);
 742         ir_node **cfgpreds   = get_Block_cfgpred_arr(old_block);
 743         ir_node  *new_block  = new_r_Block(irg, n_cfgpreds, cfgpreds);
 744         const ir_edge_t *edge;
 745         const ir_edge_t *next;
 746
 747         /* old_block has no predecessors anymore for now */
 748         set_irn_in(old_block, 0, NULL);
 749
 750         /* move node and its predecessors to new_block */
 751         move(node, old_block, new_block);
 752
 753         /* move Phi nodes to new_block */
 754         foreach_out_edge_safe(old_block, edge, next) {
 755                 ir_node *phi = get_edge_src_irn(edge);
 756                 if (!is_Phi(phi))
 757                         continue;
 758                 set_nodes_block(phi, new_block);
 759         }
 760         return old_block;
 761 }
 762
 763 typedef ir_node* (*new_rd_shr_func)(dbg_info *dbgi, ir_node *block,
 764                                     ir_node *left, ir_node *right,
 765                                     ir_mode *mode);
 766
 767 static void lower_shr_helper(ir_node *node, ir_mode *mode,
 768                              new_rd_shr_func new_rd_shrs)
 769 {
 770         ir_node  *right         = get_binop_right(node);
 771         ir_node  *left          = get_binop_left(node);
 772         ir_mode  *shr_mode      = get_irn_mode(node);
 773         unsigned  modulo_shift  = get_mode_modulo_shift(shr_mode);
 774         ir_mode  *low_unsigned  = env->low_unsigned;
 775         unsigned  modulo_shift2 = get_mode_modulo_shift(mode);
 776         ir_graph *irg           = get_irn_irg(node);
 777         ir_node  *left_low      = get_lowered_low(left);
 778         ir_node  *left_high     = get_lowered_high(left);
 779         dbg_info *dbgi          = get_irn_dbg_info(node);
 780         ir_node  *lower_block;
 781         ir_node  *block;
 782         ir_node  *cnst;
 783         ir_node  *and;
 784         ir_node  *cmp;
 785         ir_node  *cond;
 786         ir_node  *proj_true;
 787         ir_node  *proj_false;
 788         ir_node  *phi_low;
 789         ir_node  *phi_high;
 790         ir_node  *lower_in[2];
 791         ir_node  *phi_low_in[2];
 792         ir_node  *phi_high_in[2];
 793
 794         /* this version is optimized for modulo shift architectures
 795          * (and can't handle anything else) */
 796         if (modulo_shift != get_mode_size_bits(shr_mode)
 797                         || modulo_shift2<<1 != modulo_shift) {
 798                 panic("Shr lowering only implemented for modulo shift shr operations");
 799         }
 800         if (!is_po2(modulo_shift) || !is_po2(modulo_shift2)) {
 801                 panic("Shr lowering only implemented for power-of-2 modes");
 802         }
 803         /* without 2-complement the -x instead of (bit_width-x) trick won't work */
 804         if (get_mode_arithmetic(shr_mode) != irma_twos_complement) {
 805                 panic("Shr lowering only implemented for two-complement modes");
 806         }
 807
 808         /* if the right operand is a 64bit value, we're only interested in the
 809          * lower word */
 810         if (get_irn_mode(right) == env->high_unsigned) {
 811                 right = get_lowered_low(right);
 812         } else {
 813                 /* shift should never have signed mode on the right */
 814                 assert(get_irn_mode(right) != env->high_signed);
 815                 ir_node *block = get_nodes_block(node);
 816                 right = create_conv(block, right, low_unsigned);
 817         }
 818
 819         lower_block = part_block_dw(node);
 820         env->flags |= CF_CHANGED;
 821         block = get_nodes_block(node);
 822
 823         /* add a Cmp to test if highest bit is set <=> wether we shift more
 824          * than half the word width */
 825         cnst       = new_r_Const_long(irg, low_unsigned, modulo_shift2);
 826         and        = new_r_And(block, right, cnst, low_unsigned);
 827         cnst       = new_r_Const(irg, get_mode_null(low_unsigned));
 828         cmp        = new_rd_Cmp(dbgi, block, and, cnst, ir_relation_equal);
 829         cond       = new_rd_Cond(dbgi, block, cmp);
 830         proj_true  = new_r_Proj(cond, mode_X, pn_Cond_true);
 831         proj_false = new_r_Proj(cond, mode_X, pn_Cond_false);
 832
 833         /* the true block => shift_width < 1word */
 834         {
 835                 /* In theory the low value (for 64bit shifts) is:
 836                  *    Or(High << (32-x)), Low >> x)
 837                  * In practice High << 32-x will fail when x is zero (since we have
 838                  * modulo shift and 32 will be 0). So instead we use:
 839                  *    Or(High<<1<<~x, Low >> x)
 840                  */
 841                 ir_node *in[1]        = { proj_true };
 842                 ir_node *block_true   = new_r_Block(irg, ARRAY_SIZE(in), in);
 843                 ir_node *res_high     = new_rd_shrs(dbgi, block_true, left_high,
 844                                                     right, mode);
 845                 ir_node *shift_low    = new_rd_Shr(dbgi, block_true, left_low, right,
 846                                                    low_unsigned);
 847                 ir_node *not_shiftval = new_rd_Not(dbgi, block_true, right,
 848                                                    low_unsigned);
 849                 ir_node *conv         = create_conv(block_true, left_high,
 850                                                     low_unsigned);
 851                 ir_node *one          = new_r_Const(irg, get_mode_one(low_unsigned));
 852                 ir_node *carry0       = new_rd_Shl(dbgi, block_true, conv, one,
 853                                                    low_unsigned);
 854                 ir_node *carry1       = new_rd_Shl(dbgi, block_true, carry0,
 855                                                    not_shiftval, low_unsigned);
 856                 ir_node *res_low      = new_rd_Or(dbgi, block_true, shift_low, carry1,
 857                                                   low_unsigned);
 858                 lower_in[0]           = new_r_Jmp(block_true);
 859                 phi_low_in[0]         = res_low;
 860                 phi_high_in[0]        = res_high;
 861         }
 862
 863         /* false block => shift_width > 1word */
 864         {
 865                 ir_node *in[1]       = { proj_false };
 866                 ir_node *block_false = new_r_Block(irg, ARRAY_SIZE(in), in);
 867                 ir_node *conv        = create_conv(block_false, left_high, low_unsigned);
 868                 ir_node *res_low     = new_rd_shrs(dbgi, block_false, conv, right,
 869                                                    low_unsigned);
 870                 int      cnsti       = modulo_shift2-1;
 871                 ir_node *cnst        = new_r_Const_long(irg, low_unsigned, cnsti);
 872                 ir_node *res_high;
 873                 if (new_rd_shrs == new_rd_Shrs) {
 874                         res_high = new_rd_shrs(dbgi, block_false, left_high, cnst, mode);
 875                 } else {
 876                         res_high = new_r_Const(irg, get_mode_null(mode));
 877                 }
 878                 lower_in[1]          = new_r_Jmp(block_false);
 879                 phi_low_in[1]        = res_low;
 880                 phi_high_in[1]       = res_high;
 881         }
 882
 883         /* patch lower block */
 884         set_irn_in(lower_block, ARRAY_SIZE(lower_in), lower_in);
 885         phi_low  = new_r_Phi(lower_block, ARRAY_SIZE(phi_low_in), phi_low_in,
 886                              low_unsigned);
 887         phi_high = new_r_Phi(lower_block, ARRAY_SIZE(phi_high_in), phi_high_in,
 888                              mode);
 889         ir_set_dw_lowered(node, phi_low, phi_high);
 890 }
 891
 892 static void lower_Shr(ir_node *node, ir_mode *mode)
 893 {
 894         lower_shr_helper(node, mode, new_rd_Shr);
 895 }
 896
 897 static void lower_Shrs(ir_node *node, ir_mode *mode)
 898 {
 899         lower_shr_helper(node, mode, new_rd_Shrs);
 900 }
 901
 902 static void lower_Shl(ir_node *node, ir_mode *mode)
 903 {
 904         ir_node  *right         = get_binop_right(node);
 905         ir_node  *left          = get_binop_left(node);
 906         ir_mode  *shr_mode      = get_irn_mode(node);
 907         unsigned  modulo_shift  = get_mode_modulo_shift(shr_mode);
 908         ir_mode  *low_unsigned  = env->low_unsigned;
 909         unsigned  modulo_shift2 = get_mode_modulo_shift(mode);
 910         ir_graph *irg           = get_irn_irg(node);
 911         ir_node  *left_low      = get_lowered_low(left);
 912         ir_node  *left_high     = get_lowered_high(left);
 913         dbg_info *dbgi          = get_irn_dbg_info(node);
 914         ir_node  *lower_block   = get_nodes_block(node);
 915         ir_node  *block;
 916         ir_node  *cnst;
 917         ir_node  *and;
 918         ir_node  *cmp;
 919         ir_node  *cond;
 920         ir_node  *proj_true;
 921         ir_node  *proj_false;
 922         ir_node  *phi_low;
 923         ir_node  *phi_high;
 924         ir_node  *lower_in[2];
 925         ir_node  *phi_low_in[2];
 926         ir_node  *phi_high_in[2];
 927
 928         /* this version is optimized for modulo shift architectures
 929          * (and can't handle anything else) */
 930         if (modulo_shift != get_mode_size_bits(shr_mode)
 931                         || modulo_shift2<<1 != modulo_shift) {
 932                 panic("Shr lowering only implemented for modulo shift shr operations");
 933         }
 934         if (!is_po2(modulo_shift) || !is_po2(modulo_shift2)) {
 935                 panic("Shr lowering only implemented for power-of-2 modes");
 936         }
 937         /* without 2-complement the -x instead of (bit_width-x) trick won't work */
 938         if (get_mode_arithmetic(shr_mode) != irma_twos_complement) {
 939                 panic("Shr lowering only implemented for two-complement modes");
 940         }
 941
 942         /* if the right operand is a 64bit value, we're only interested in the
 943          * lower word */
 944         if (get_irn_mode(right) == env->high_unsigned) {
 945                 right = get_lowered_low(right);
 946         } else {
 947                 /* shift should never have signed mode on the right */
 948                 assert(get_irn_mode(right) != env->high_signed);
 949                 right = create_conv(lower_block, right, low_unsigned);
 950         }
 951
 952         part_block_dw(node);
 953         env->flags |= CF_CHANGED;
 954         block = get_nodes_block(node);
 955
 956         /* add a Cmp to test if highest bit is set <=> wether we shift more
 957          * than half the word width */
 958         cnst       = new_r_Const_long(irg, low_unsigned, modulo_shift2);
 959         and        = new_r_And(block, right, cnst, low_unsigned);
 960         cnst       = new_r_Const(irg, get_mode_null(low_unsigned));
 961         cmp        = new_rd_Cmp(dbgi, block, and, cnst, ir_relation_equal);
 962         cond       = new_rd_Cond(dbgi, block, cmp);
 963         proj_true  = new_r_Proj(cond, mode_X, pn_Cond_true);
 964         proj_false = new_r_Proj(cond, mode_X, pn_Cond_false);
 965
 966         /* the true block => shift_width < 1word */
 967         {
 968                 ir_node *in[1]        = { proj_true };
 969                 ir_node *block_true   = new_r_Block(irg, ARRAY_SIZE(in), in);
 970
 971                 ir_node *res_low      = new_rd_Shl(dbgi, block_true, left_low,
 972                                                    right, low_unsigned);
 973                 ir_node *shift_high   = new_rd_Shl(dbgi, block_true, left_high, right,
 974                                                    mode);
 975                 ir_node *not_shiftval = new_rd_Not(dbgi, block_true, right,
 976                                                    low_unsigned);
 977                 ir_node *conv         = create_conv(block_true, left_low, mode);
 978                 ir_node *one          = new_r_Const(irg, get_mode_one(mode));
 979                 ir_node *carry0       = new_rd_Shr(dbgi, block_true, conv, one, mode);
 980                 ir_node *carry1       = new_rd_Shr(dbgi, block_true, carry0,
 981                                                    not_shiftval, mode);
 982                 ir_node *res_high     = new_rd_Or(dbgi, block_true, shift_high, carry1,
 983                                                   mode);
 984                 lower_in[0]           = new_r_Jmp(block_true);
 985                 phi_low_in[0]         = res_low;
 986                 phi_high_in[0]        = res_high;
 987         }
 988
 989         /* false block => shift_width > 1word */
 990         {
 991                 ir_node *in[1]       = { proj_false };
 992                 ir_node *block_false = new_r_Block(irg, ARRAY_SIZE(in), in);
 993                 ir_node *res_low     = new_r_Const(irg, get_mode_null(low_unsigned));
 994                 ir_node *conv        = create_conv(block_false, left_low, mode);
 995                 ir_node *res_high    = new_rd_Shl(dbgi, block_false, conv, right, mode);
 996                 lower_in[1]          = new_r_Jmp(block_false);
 997                 phi_low_in[1]        = res_low;
 998                 phi_high_in[1]       = res_high;
 999         }
1000
1001         /* patch lower block */
1002         set_irn_in(lower_block, ARRAY_SIZE(lower_in), lower_in);
1003         phi_low  = new_r_Phi(lower_block, ARRAY_SIZE(phi_low_in), phi_low_in,
1004                              low_unsigned);
1005         phi_high = new_r_Phi(lower_block, ARRAY_SIZE(phi_high_in), phi_high_in,
1006                              mode);
1007         ir_set_dw_lowered(node, phi_low, phi_high);
1008 }
1009
1010 /**
1011  * Rebuild Rotl nodes into Or(Shl, Shr) and prepare all nodes.
1012  */
1013 static void prepare_links_and_handle_rotl(ir_node *node, void *data)
1014 {
1015         (void) data;
1016         if (is_Rotl(node)) {
1017                 ir_mode  *mode = get_irn_op_mode(node);
1018                 ir_node  *right;
1019                 ir_node  *left, *shl, *shr, *ornode, *block, *sub, *c;
1020                 ir_mode  *omode, *rmode;
1021                 ir_graph *irg;
1022                 dbg_info *dbg;
1023                 optimization_state_t state;
1024
1025                 if (mode != env->high_signed && mode != env->high_unsigned) {
1026                         prepare_links(node);
1027                         return;
1028                 }
1029
1030                 /* replace the Rotl(x,y) by an Or(Shl(x,y), Shr(x,64-y)) */
1031                 right = get_Rotl_right(node);
1032                 irg   = get_irn_irg(node);
1033                 dbg   = get_irn_dbg_info(node);
1034                 omode = get_irn_mode(node);
1035                 left  = get_Rotl_left(node);
1036                 block = get_nodes_block(node);
1037                 shl   = new_rd_Shl(dbg, block, left, right, omode);
1038                 rmode = get_irn_mode(right);
1039                 c     = new_r_Const_long(irg, rmode, get_mode_size_bits(omode));
1040                 sub   = new_rd_Sub(dbg, block, c, right, rmode);
1041                 shr   = new_rd_Shr(dbg, block, left, sub, omode);
1042
1043                 /* switch optimization off here, or we will get the Rotl back */
1044                 save_optimization_state(&state);
1045                 set_opt_algebraic_simplification(0);
1046                 ornode = new_rd_Or(dbg, block, shl, shr, omode);
1047                 restore_optimization_state(&state);
1048
1049                 exchange(node, ornode);
1050
1051                 /* do lowering on the new nodes */
1052                 prepare_links(shl);
1053                 prepare_links(c);
1054                 prepare_links(sub);
1055                 prepare_links(shr);
1056                 prepare_links(ornode);
1057                 return;
1058         }
1059
1060         prepare_links(node);
1061 }
1062
1063 /**
1064  * Translate an Unop.
1065  *
1066  * Create an intrinsic Call.
1067  */
1068 static void lower_unop(ir_node *node, ir_mode *mode)
1069 {
1070         ir_node  *op       = get_unop_op(node);
1071         dbg_info *dbgi     = get_irn_dbg_info(node);
1072         ir_node  *block    = get_nodes_block(node);
1073         ir_graph *irg      = get_irn_irg(block);
1074         ir_type  *mtp      = mode_is_signed(mode) ? unop_tp_s : unop_tp_u;
1075         ir_op    *irop     = get_irn_op(node);
1076         ir_node  *addr     = get_intrinsic_address(mtp, irop, mode, mode);
1077         ir_node  *nomem    = get_irg_no_mem(irg);
1078         ir_node  *in[2];
1079         ir_node  *call;
1080         ir_node  *resproj;
1081
1082         if (env->params->little_endian) {
1083                 in[0] = get_lowered_low(op);
1084                 in[1] = get_lowered_high(op);
1085         } else {
1086                 in[0] = get_lowered_high(op);
1087                 in[1] = get_lowered_low(op);
1088         }
1089         call    = new_rd_Call(dbgi, block, nomem, addr, 2, in, mtp);
1090         resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
1091         set_irn_pinned(call, get_irn_pinned(node));
1092
1093         if (env->params->little_endian) {
1094                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
1095                 ir_node *res_high = new_r_Proj(resproj, mode,              1);
1096                 ir_set_dw_lowered(node, res_low, res_high);
1097         } else {
1098                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
1099                 ir_node *res_high = new_r_Proj(resproj, mode,              0);
1100                 ir_set_dw_lowered(node, res_low, res_high);
1101         }
1102 }
1103
1104 /**
1105  * Translate a logical binop.
1106  *
1107  * Create two logical binops.
1108  */
1109 static void lower_binop_logical(ir_node *node, ir_mode *mode,
1110                                                                 ir_node *(*constr_rd)(dbg_info *db, ir_node *block, ir_node *op1, ir_node *op2, ir_mode *mode) )
1111 {
1112         ir_node               *left        = get_binop_left(node);
1113         ir_node               *right       = get_binop_right(node);
1114         const lower64_entry_t *left_entry  = get_node_entry(left);
1115         const lower64_entry_t *right_entry = get_node_entry(right);
1116         dbg_info              *dbgi        = get_irn_dbg_info(node);
1117         ir_node               *block       = get_nodes_block(node);
1118         ir_node               *res_low
1119                 = constr_rd(dbgi, block, left_entry->low_word, right_entry->low_word,
1120                             env->low_unsigned);
1121         ir_node               *res_high
1122                 = constr_rd(dbgi, block, left_entry->high_word, right_entry->high_word,
1123                             mode);
1124         ir_set_dw_lowered(node, res_low, res_high);
1125 }
1126
1127 static void lower_And(ir_node *node, ir_mode *mode)
1128 {
1129         lower_binop_logical(node, mode, new_rd_And);
1130 }
1131
1132 static void lower_Or(ir_node *node, ir_mode *mode)
1133 {
1134         lower_binop_logical(node, mode, new_rd_Or);
1135 }
1136
1137 static void lower_Eor(ir_node *node, ir_mode *mode)
1138 {
1139         lower_binop_logical(node, mode, new_rd_Eor);
1140 }
1141
1142 /**
1143  * Translate a Not.
1144  *
1145  * Create two logical Nots.
1146  */
1147 static void lower_Not(ir_node *node, ir_mode *mode)
1148 {
1149         ir_node               *op       = get_Not_op(node);
1150         const lower64_entry_t *op_entry = get_node_entry(op);
1151         dbg_info              *dbgi     = get_irn_dbg_info(node);
1152         ir_node               *block    = get_nodes_block(node);
1153         ir_node               *res_low
1154                 = new_rd_Not(dbgi, block, op_entry->low_word, env->low_unsigned);
1155         ir_node               *res_high
1156                 = new_rd_Not(dbgi, block, op_entry->high_word, mode);
1157         ir_set_dw_lowered(node, res_low, res_high);
1158 }
1159
1160 static bool is_equality_cmp_0(const ir_node *node)
1161 {
1162         ir_relation relation = get_Cmp_relation(node);
1163         ir_node    *left     = get_Cmp_left(node);
1164         ir_node    *right    = get_Cmp_right(node);
1165         ir_mode    *mode     = get_irn_mode(left);
1166
1167         /* this probably makes no sense if unordered is involved */
1168         assert(!mode_is_float(mode));
1169
1170         if (!is_Const(right) || !is_Const_null(right))
1171                 return false;
1172         if (relation == ir_relation_equal)
1173                 return true;
1174         if (mode_is_signed(mode)) {
1175                 return relation == ir_relation_less_greater;
1176         } else {
1177                 return relation == ir_relation_greater;
1178         }
1179 }
1180
1181 static ir_node *get_cfop_destination(const ir_node *cfop)
1182 {
1183         const ir_edge_t *first = get_irn_out_edge_first(cfop);
1184         /* we should only have 1 destination */
1185         assert(get_irn_n_edges(cfop) == 1);
1186         return get_edge_src_irn(first);
1187 }
1188
1189 /**
1190  * Translate a Cond.
1191  */
1192 static void lower_Cond(ir_node *node, ir_mode *mode)
1193 {
1194         ir_node *left, *right, *block;
1195         ir_node *sel = get_Cond_selector(node);
1196         ir_mode *m = get_irn_mode(sel);
1197         ir_mode *cmp_mode;
1198         const lower64_entry_t *lentry, *rentry;
1199         ir_node  *proj, *projT = NULL, *projF = NULL;
1200         ir_node  *new_bl, *irn;
1201         ir_node  *projHF, *projHT;
1202         ir_node  *dst_blk;
1203         ir_relation relation;
1204         ir_graph *irg;
1205         dbg_info *dbg;
1206         const ir_edge_t *edge;
1207         const ir_edge_t *next;
1208
1209         (void) mode;
1210
1211         if (m != mode_b) {
1212                 if (m == env->high_signed || m == env->high_unsigned) {
1213                         /* bad we can't really handle Switch with 64bit offsets */
1214                         panic("Cond with 64bit jumptable not supported");
1215                 }
1216                 lower_node(sel);
1217                 return;
1218         }
1219
1220         if (!is_Cmp(sel)) {
1221                 lower_node(sel);
1222                 return;
1223         }
1224
1225         left     = get_Cmp_left(sel);
1226         cmp_mode = get_irn_mode(left);
1227         if (cmp_mode != env->high_signed && cmp_mode != env->high_unsigned) {
1228                 lower_node(sel);
1229                 return;
1230         }
1231
1232         right  = get_Cmp_right(sel);
1233         lower_node(left);
1234         lower_node(right);
1235         lentry = get_node_entry(left);
1236         rentry = get_node_entry(right);
1237
1238         /* all right, build the code */
1239         foreach_out_edge_safe(node, edge, next) {
1240                 ir_node *proj    = get_edge_src_irn(edge);
1241                 long     proj_nr;
1242                 if (!is_Proj(proj))
1243                         continue;
1244                 proj_nr = get_Proj_proj(proj);
1245
1246                 if (proj_nr == pn_Cond_true) {
1247                         assert(projT == NULL && "more than one Proj(true)");
1248                         projT = proj;
1249                 } else {
1250                         assert(proj_nr == pn_Cond_false);
1251                         assert(projF == NULL && "more than one Proj(false)");
1252                         projF = proj;
1253                 }
1254                 mark_irn_visited(proj);
1255         }
1256         assert(projT && projF);
1257
1258         /* create a new high compare */
1259         block    = get_nodes_block(node);
1260         irg      = get_Block_irg(block);
1261         dbg      = get_irn_dbg_info(sel);
1262         relation = get_Cmp_relation(sel);
1263
1264         if (is_equality_cmp_0(sel)) {
1265                 /* x ==/!= 0 ==> or(low,high) ==/!= 0 */
1266                 ir_mode *mode   = env->low_unsigned;
1267                 ir_node *low    = new_r_Conv(block, lentry->low_word, mode);
1268                 ir_node *high   = new_r_Conv(block, lentry->high_word, mode);
1269                 ir_node *ornode = new_rd_Or(dbg, block, low, high, mode);
1270                 ir_node *cmp    = new_rd_Cmp(dbg, block, ornode, new_r_Const_long(irg, mode, 0), relation);
1271                 set_Cond_selector(node, cmp);
1272                 return;
1273         }
1274
1275         if (relation == ir_relation_equal) {
1276                 /* simple case:a == b <==> a_h == b_h && a_l == b_l */
1277                 dst_blk = get_cfop_destination(projF);
1278
1279                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1280                                  ir_relation_equal);
1281                 dbg = get_irn_dbg_info(node);
1282                 irn = new_rd_Cond(dbg, block, irn);
1283
1284                 projHF = new_r_Proj(irn, mode_X, pn_Cond_false);
1285                 mark_irn_visited(projHF);
1286                 exchange(projF, projHF);
1287
1288                 projHT = new_r_Proj(irn, mode_X, pn_Cond_true);
1289                 mark_irn_visited(projHT);
1290
1291                 new_bl = new_r_Block(irg, 1, &projHT);
1292
1293                 dbg = get_irn_dbg_info(sel);
1294                 irn = new_rd_Cmp(dbg, new_bl, lentry->low_word, rentry->low_word,
1295                                   ir_relation_equal);
1296                 dbg = get_irn_dbg_info(node);
1297                 irn = new_rd_Cond(dbg, new_bl, irn);
1298
1299                 proj = new_r_Proj(irn, mode_X, pn_Cond_false);
1300                 mark_irn_visited(proj);
1301                 add_block_cf_input(dst_blk, projHF, proj);
1302
1303                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1304                 mark_irn_visited(proj);
1305                 exchange(projT, proj);
1306         } else if (relation == ir_relation_less_greater) {
1307                 /* simple case:a != b <==> a_h != b_h || a_l != b_l */
1308                 dst_blk = get_cfop_destination(projT);
1309
1310                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1311                                  ir_relation_less_greater);
1312                 dbg = get_irn_dbg_info(node);
1313                 irn = new_rd_Cond(dbg, block, irn);
1314
1315                 projHT = new_r_Proj(irn, mode_X, pn_Cond_true);
1316                 mark_irn_visited(projHT);
1317                 exchange(projT, projHT);
1318
1319                 projHF = new_r_Proj(irn, mode_X, pn_Cond_false);
1320                 mark_irn_visited(projHF);
1321
1322                 new_bl = new_r_Block(irg, 1, &projHF);
1323
1324                 dbg = get_irn_dbg_info(sel);
1325                 irn = new_rd_Cmp(dbg, new_bl, lentry->low_word, rentry->low_word,
1326                                  ir_relation_less_greater);
1327                 dbg = get_irn_dbg_info(node);
1328                 irn = new_rd_Cond(dbg, new_bl, irn);
1329
1330                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1331                 mark_irn_visited(proj);
1332                 add_block_cf_input(dst_blk, projHT, proj);
1333
1334                 proj = new_r_Proj(irn, mode_X, pn_Cond_false);
1335                 mark_irn_visited(proj);
1336                 exchange(projF, proj);
1337         } else {
1338                 /* a rel b <==> a_h REL b_h || (a_h == b_h && a_l rel b_l) */
1339                 ir_node *dstT, *dstF, *newbl_eq, *newbl_l;
1340                 ir_node *projEqF;
1341
1342                 dstT = get_cfop_destination(projT);
1343                 dstF = get_cfop_destination(projF);
1344
1345                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1346                                  relation & ~ir_relation_equal);
1347                 dbg = get_irn_dbg_info(node);
1348                 irn = new_rd_Cond(dbg, block, irn);
1349
1350                 projHT = new_r_Proj(irn, mode_X, pn_Cond_true);
1351                 mark_irn_visited(projHT);
1352
1353                 projHF = new_r_Proj(irn, mode_X, pn_Cond_false);
1354                 mark_irn_visited(projHF);
1355
1356                 newbl_eq = new_r_Block(irg, 1, &projHF);
1357
1358                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1359                                  ir_relation_equal);
1360                 irn = new_rd_Cond(dbg, newbl_eq, irn);
1361
1362                 projEqF = new_r_Proj(irn, mode_X, pn_Cond_false);
1363                 mark_irn_visited(projEqF);
1364
1365                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1366                 mark_irn_visited(proj);
1367
1368                 newbl_l = new_r_Block(irg, 1, &proj);
1369
1370                 dbg = get_irn_dbg_info(sel);
1371                 irn = new_rd_Cmp(dbg, newbl_l, lentry->low_word, rentry->low_word,
1372                                  relation);
1373                 dbg = get_irn_dbg_info(node);
1374                 irn = new_rd_Cond(dbg, newbl_l, irn);
1375
1376                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1377                 mark_irn_visited(proj);
1378                 add_block_cf_input(dstT, projT, proj);
1379
1380                 proj = new_r_Proj(irn, mode_X, pn_Cond_false);
1381                 mark_irn_visited(proj);
1382                 add_block_cf_input(dstF, projF, proj);
1383
1384                 exchange(projT, projHT);
1385                 exchange(projF, projEqF);
1386         }
1387
1388         /* we have changed the control flow */
1389         env->flags |= CF_CHANGED;
1390 }
1391
1392 /**
1393  * Translate a Conv to higher_signed
1394  */
1395 static void lower_Conv_to_Ll(ir_node *node)
1396 {
1397         ir_mode  *omode        = get_irn_mode(node);
1398         ir_node  *op           = get_Conv_op(node);
1399         ir_mode  *imode        = get_irn_mode(op);
1400         ir_graph *irg          = get_irn_irg(node);
1401         ir_node  *block        = get_nodes_block(node);
1402         dbg_info *dbg          = get_irn_dbg_info(node);
1403         ir_node  *res_low;
1404         ir_node  *res_high;
1405
1406         ir_mode  *low_unsigned = env->low_unsigned;
1407         ir_mode  *low_signed
1408                 = mode_is_signed(omode) ? env->low_signed : low_unsigned;
1409
1410         if (mode_is_int(imode) || mode_is_reference(imode)) {
1411                 if (imode == env->high_signed || imode == env->high_unsigned) {
1412                         /* a Conv from Lu to Ls or Ls to Lu */
1413                         const lower64_entry_t *op_entry = get_node_entry(op);
1414                         res_low  = op_entry->low_word;
1415                         res_high = new_rd_Conv(dbg, block, op_entry->high_word, low_signed);
1416                 } else {
1417                         /* simple case: create a high word */
1418                         if (imode != low_unsigned)
1419                                 op = new_rd_Conv(dbg, block, op, low_unsigned);
1420
1421                         res_low = op;
1422
1423                         if (mode_is_signed(imode)) {
1424                                 int      c       = get_mode_size_bits(low_signed) - 1;
1425                                 ir_node *cnst    = new_r_Const_long(irg, low_unsigned, c);
1426                                 if (get_irn_mode(op) != low_signed)
1427                                         op = new_rd_Conv(dbg, block, op, low_signed);
1428                                 res_high = new_rd_Shrs(dbg, block, op, cnst, low_signed);
1429                         } else {
1430                                 res_high = new_r_Const(irg, get_mode_null(low_signed));
1431                         }
1432                 }
1433         } else if (imode == mode_b) {
1434                 res_low  = new_rd_Conv(dbg, block, op, low_unsigned);
1435                 res_high = new_r_Const(irg, get_mode_null(low_signed));
1436         } else {
1437                 ir_node *irn, *call;
1438                 ir_type *mtp = get_conv_type(imode, omode);
1439
1440                 irn = get_intrinsic_address(mtp, get_irn_op(node), imode, omode);
1441                 call = new_rd_Call(dbg, block, get_irg_no_mem(irg), irn, 1, &op, mtp);
1442                 set_irn_pinned(call, get_irn_pinned(node));
1443                 irn = new_r_Proj(call, mode_T, pn_Call_T_result);
1444
1445                 res_low  = new_r_Proj(irn, low_unsigned, 0);
1446                 res_high = new_r_Proj(irn, low_signed, 1);
1447         }
1448         ir_set_dw_lowered(node, res_low, res_high);
1449 }
1450
1451 /**
1452  * Translate a Conv from higher_unsigned
1453  */
1454 static void lower_Conv_from_Ll(ir_node *node)
1455 {
1456         ir_node               *op    = get_Conv_op(node);
1457         ir_mode               *omode = get_irn_mode(node);
1458         ir_node               *block = get_nodes_block(node);
1459         dbg_info              *dbg   = get_irn_dbg_info(node);
1460         ir_graph              *irg   = get_irn_irg(node);
1461         const lower64_entry_t *entry = get_node_entry(op);
1462
1463         if (mode_is_int(omode) || mode_is_reference(omode)) {
1464                 op = entry->low_word;
1465
1466                 /* simple case: create a high word */
1467                 if (omode != env->low_unsigned)
1468                         op = new_rd_Conv(dbg, block, op, omode);
1469
1470                 set_Conv_op(node, op);
1471         } else if (omode == mode_b) {
1472                 /* llu ? true : false  <=> (low|high) ? true : false */
1473                 ir_mode *mode   = env->low_unsigned;
1474                 ir_node *ornode = new_rd_Or(dbg, block, entry->low_word,
1475                                             entry->high_word, mode);
1476                 set_Conv_op(node, ornode);
1477         } else {
1478                 ir_node *irn, *call, *in[2];
1479                 ir_mode *imode = get_irn_mode(op);
1480                 ir_type *mtp   = get_conv_type(imode, omode);
1481                 ir_node *res;
1482
1483                 irn   = get_intrinsic_address(mtp, get_irn_op(node), imode, omode);
1484                 in[0] = entry->low_word;
1485                 in[1] = entry->high_word;
1486
1487                 call = new_rd_Call(dbg, block, get_irg_no_mem(irg), irn, 2, in, mtp);
1488                 set_irn_pinned(call, get_irn_pinned(node));
1489                 irn = new_r_Proj(call, mode_T, pn_Call_T_result);
1490                 res = new_r_Proj(irn, omode, 0);
1491
1492                 exchange(node, res);
1493         }
1494 }
1495
1496 /**
1497  * lower Cmp
1498  */
1499 static void lower_Cmp(ir_node *cmp, ir_mode *m)
1500 {
1501         ir_node  *l    = get_Cmp_left(cmp);
1502         ir_mode  *mode = get_irn_mode(l);
1503         ir_node  *r, *low, *high, *t, *res;
1504         ir_relation relation;
1505         ir_node  *block;
1506         dbg_info *dbg;
1507         const lower64_entry_t *lentry;
1508         const lower64_entry_t *rentry;
1509         (void) m;
1510
1511         if (mode != env->high_signed && mode != env->high_unsigned)
1512                 return;
1513
1514         r        = get_Cmp_right(cmp);
1515         lentry   = get_node_entry(l);
1516         rentry   = get_node_entry(r);
1517         relation = get_Cmp_relation(cmp);
1518         block    = get_nodes_block(cmp);
1519         dbg      = get_irn_dbg_info(cmp);
1520
1521         /* easy case for x ==/!= 0 (see lower_Cond for details) */
1522         if (is_equality_cmp_0(cmp)) {
1523                 ir_graph *irg     = get_irn_irg(cmp);
1524                 ir_mode  *mode    = env->low_unsigned;
1525                 ir_node  *low     = new_r_Conv(block, lentry->low_word, mode);
1526                 ir_node  *high    = new_r_Conv(block, lentry->high_word, mode);
1527                 ir_node  *ornode  = new_rd_Or(dbg, block, low, high, mode);
1528                 ir_node  *new_cmp = new_rd_Cmp(dbg, block, ornode, new_r_Const_long(irg, mode, 0), relation);
1529                 exchange(cmp, new_cmp);
1530                 return;
1531         }
1532
1533         if (relation == ir_relation_equal) {
1534                 /* simple case:a == b <==> a_h == b_h && a_l == b_l */
1535                 low  = new_rd_Cmp(dbg, block, lentry->low_word, rentry->low_word,
1536                                   relation);
1537                 high = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1538                                   relation);
1539                 res  = new_rd_And(dbg, block, low, high, mode_b);
1540         } else if (relation == ir_relation_less_greater) {
1541                 /* simple case:a != b <==> a_h != b_h || a_l != b_l */
1542                 low  = new_rd_Cmp(dbg, block, lentry->low_word, rentry->low_word,
1543                                   relation);
1544                 high = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1545                                   relation);
1546                 res = new_rd_Or(dbg, block, low, high, mode_b);
1547         } else {
1548                 /* a rel b <==> a_h REL b_h || (a_h == b_h && a_l rel b_l) */
1549                 ir_node *high1 = new_rd_Cmp(dbg, block, lentry->high_word,
1550                         rentry->high_word, relation & ~ir_relation_equal);
1551                 low  = new_rd_Cmp(dbg, block, lentry->low_word, rentry->low_word,
1552                                   relation);
1553                 high = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1554                                   ir_relation_equal);
1555                 t = new_rd_And(dbg, block, low, high, mode_b);
1556                 res = new_rd_Or(dbg, block, high1, t, mode_b);
1557         }
1558         exchange(cmp, res);
1559 }
1560
1561 /**
1562  * Translate a Conv.
1563  */
1564 static void lower_Conv(ir_node *node, ir_mode *mode)
1565 {
1566         mode = get_irn_mode(node);
1567
1568         if (mode == env->high_signed || mode == env->high_unsigned) {
1569                 lower_Conv_to_Ll(node);
1570         } else {
1571                 ir_mode *op_mode = get_irn_mode(get_Conv_op(node));
1572
1573                 if (op_mode == env->high_signed || op_mode == env->high_unsigned) {
1574                         lower_Conv_from_Ll(node);
1575                 }
1576         }
1577 }
1578
1579 /**
1580  * Remember the new argument index of this value type entity in the lowered
1581  * method type.
1582  *
1583  * @param ent  the entity
1584  * @param pos  the argument index of this entity
1585  */
1586 static inline void set_entity_arg_idx(ir_entity *ent, size_t pos)
1587 {
1588         set_entity_link(ent, INT_TO_PTR(pos));
1589 }
1590
1591 /**
1592  * Retrieve the argument index of a value type entity.
1593  *
1594  * @param ent  the entity
1595  */
1596 static size_t get_entity_arg_idx(const ir_entity *ent) {
1597         return PTR_TO_INT(get_entity_link(ent));
1598 }
1599
1600 /**
1601  * Lower the method type.
1602  *
1603  * @param env  the lower environment
1604  * @param mtp  the method type to lower
1605  *
1606  * @return the lowered type
1607  */
1608 static ir_type *lower_mtp(ir_type *mtp)
1609 {
1610         pmap_entry *entry;
1611         ir_type    *res, *value_type;
1612
1613         entry = pmap_find(lowered_type, mtp);
1614         if (! entry) {
1615                 size_t i, orig_n_params, orig_n_res, n_param, n_res;
1616
1617                 /* count new number of params */
1618                 n_param = orig_n_params = get_method_n_params(mtp);
1619                 for (i = orig_n_params; i > 0;) {
1620                         ir_type *tp = get_method_param_type(mtp, --i);
1621
1622                         if (is_Primitive_type(tp)) {
1623                                 ir_mode *mode = get_type_mode(tp);
1624
1625                                 if (mode == env->high_signed || mode == env->high_unsigned)
1626                                         ++n_param;
1627                         }
1628                 }
1629
1630                 /* count new number of results */
1631                 n_res = orig_n_res = get_method_n_ress(mtp);
1632                 for (i = orig_n_res; i > 0;) {
1633                         ir_type *tp = get_method_res_type(mtp, --i);
1634
1635                         if (is_Primitive_type(tp)) {
1636                                 ir_mode *mode = get_type_mode(tp);
1637
1638                                 if (mode == env->high_signed || mode == env->high_unsigned)
1639                                         ++n_res;
1640                         }
1641                 }
1642
1643                 res = new_type_method(n_param, n_res);
1644
1645                 /* set param types and result types */
1646                 for (i = n_param = 0; i < orig_n_params; ++i) {
1647                         ir_type *tp = get_method_param_type(mtp, i);
1648
1649                         if (is_Primitive_type(tp)) {
1650                                 ir_mode *mode = get_type_mode(tp);
1651
1652                                 if (mode == env->high_signed) {
1653                                         if (env->params->little_endian) {
1654                                                 set_method_param_type(res, n_param++, tp_u);
1655                                                 set_method_param_type(res, n_param++, tp_s);
1656                                         } else {
1657                                                 set_method_param_type(res, n_param++, tp_s);
1658                                                 set_method_param_type(res, n_param++, tp_u);
1659                                         }
1660                                 } else if (mode == env->high_unsigned) {
1661                                         set_method_param_type(res, n_param++, tp_u);
1662                                         set_method_param_type(res, n_param++, tp_u);
1663                                 } else {
1664                                         set_method_param_type(res, n_param++, tp);
1665                                 }
1666                         } else {
1667                                 set_method_param_type(res, n_param++, tp);
1668                         }
1669                 }
1670                 for (i = n_res = 0; i < orig_n_res; ++i) {
1671                         ir_type *tp = get_method_res_type(mtp, i);
1672
1673                         if (is_Primitive_type(tp)) {
1674                                 ir_mode *mode = get_type_mode(tp);
1675
1676                                 if (mode == env->high_signed) {
1677                                         if (env->params->little_endian) {
1678                                                 set_method_res_type(res, n_res++, tp_u);
1679                                                 set_method_res_type(res, n_res++, tp_s);
1680                                         } else {
1681                                                 set_method_res_type(res, n_res++, tp_s);
1682                                                 set_method_res_type(res, n_res++, tp_u);
1683                                         }
1684                                 } else if (mode == env->high_unsigned) {
1685                                         set_method_res_type(res, n_res++, tp_u);
1686                                         set_method_res_type(res, n_res++, tp_u);
1687                                 } else {
1688                                         set_method_res_type(res, n_res++, tp);
1689                                 }
1690                         } else {
1691                                 set_method_res_type(res, n_res++, tp);
1692                         }
1693                 }
1694                 set_lowered_type(mtp, res);
1695                 pmap_insert(lowered_type, mtp, res);
1696
1697                 value_type = get_method_value_param_type(mtp);
1698                 if (value_type != NULL) {
1699                         /* this creates a new value parameter type */
1700                         (void)get_method_value_param_ent(res, 0);
1701
1702                         /* set new param positions for all entities of the value type */
1703                         for (i = n_param = 0; i < orig_n_params; ++i) {
1704                                 ir_type   *tp  = get_method_param_type(mtp, i);
1705                                 ir_entity *ent = get_method_value_param_ent(mtp, i);
1706
1707                                 set_entity_arg_idx(ent, n_param);
1708                                 if (is_Primitive_type(tp)) {
1709                                         ir_mode *mode = get_type_mode(tp);
1710
1711                                         if (mode == env->high_signed
1712                                             || mode == env->high_unsigned) {
1713                                                 n_param += 2;
1714                                                 continue;
1715                                         }
1716                                 }
1717                                 ++n_param;
1718                         }
1719
1720                         set_lowered_type(value_type, get_method_value_param_type(res));
1721                 }
1722         } else {
1723                 res = (ir_type*)entry->value;
1724         }
1725         return res;
1726 }
1727
1728 /**
1729  * Translate a Return.
1730  */
1731 static void lower_Return(ir_node *node, ir_mode *mode)
1732 {
1733         ir_graph  *irg = get_irn_irg(node);
1734         ir_entity *ent = get_irg_entity(irg);
1735         ir_type   *mtp = get_entity_type(ent);
1736         ir_node  **in;
1737         size_t     i, j, n;
1738         int        need_conv = 0;
1739         (void) mode;
1740
1741         /* check if this return must be lowered */
1742         for (i = 0, n = get_Return_n_ress(node); i < n; ++i) {
1743                 ir_node *pred = get_Return_res(node, i);
1744                 ir_mode *mode = get_irn_op_mode(pred);
1745
1746                 if (mode == env->high_signed || mode == env->high_unsigned)
1747                         need_conv = 1;
1748         }
1749         if (! need_conv)
1750                 return;
1751
1752         ent = get_irg_entity(irg);
1753         mtp = get_entity_type(ent);
1754
1755         mtp = lower_mtp(mtp);
1756         set_entity_type(ent, mtp);
1757
1758         /* create a new in array */
1759         NEW_ARR_A(ir_node *, in, get_method_n_ress(mtp) + 1);
1760         in[0] = get_Return_mem(node);
1761
1762         for (j = i = 0, n = get_Return_n_ress(node); i < n; ++i) {
1763                 ir_node *pred      = get_Return_res(node, i);
1764                 ir_mode *pred_mode = get_irn_mode(pred);
1765
1766                 if (pred_mode == env->high_signed || pred_mode == env->high_unsigned) {
1767                         const lower64_entry_t *entry = get_node_entry(pred);
1768                         if (env->params->little_endian) {
1769                                 in[++j] = entry->low_word;
1770                                 in[++j] = entry->high_word;
1771                         } else {
1772                                 in[++j] = entry->high_word;
1773                                 in[++j] = entry->low_word;
1774                         }
1775                 } else {
1776                         in[++j] = pred;
1777                 }
1778         }
1779
1780         set_irn_in(node, j+1, in);
1781 }
1782
1783 /**
1784  * Translate the parameters.
1785  */
1786 static void lower_Start(ir_node *node, ir_mode *mode)
1787 {
1788         ir_graph  *irg = get_irn_irg(node);
1789         ir_entity *ent = get_irg_entity(irg);
1790         ir_type   *tp  = get_entity_type(ent);
1791         ir_node   *args;
1792         long      *new_projs;
1793         size_t    i, j, n_params;
1794         const ir_edge_t *edge;
1795         const ir_edge_t *next;
1796         (void) mode;
1797
1798         if (!mtp_must_be_lowered(tp))
1799                 return;
1800
1801         n_params = get_method_n_params(tp);
1802
1803         NEW_ARR_A(long, new_projs, n_params);
1804
1805         /* Calculate mapping of proj numbers in new_projs */
1806         for (i = j = 0; i < n_params; ++i, ++j) {
1807                 ir_type *ptp = get_method_param_type(tp, i);
1808
1809                 new_projs[i] = j;
1810                 if (is_Primitive_type(ptp)) {
1811                         ir_mode *mode = get_type_mode(ptp);
1812
1813                         if (mode == env->high_signed || mode == env->high_unsigned)
1814                                 ++j;
1815                 }
1816         }
1817
1818         /* lower method type */
1819         tp = lower_mtp(tp);
1820         set_entity_type(ent, tp);
1821
1822         args = NULL;
1823         foreach_out_edge(node, edge) {
1824                 ir_node *proj = get_edge_src_irn(edge);
1825                 if (!is_Proj(proj))
1826                         continue;
1827                 if (get_Proj_proj(proj) == pn_Start_T_args) {
1828                         args = proj;
1829                         break;
1830                 }
1831         }
1832         if (args == NULL)
1833                 return;
1834
1835         /* fix all Proj's and create new ones */
1836         foreach_out_edge_safe(args, edge, next) {
1837                 ir_node *proj   = get_edge_src_irn(edge);
1838                 ir_mode *mode   = get_irn_mode(proj);
1839                 ir_mode *mode_l = env->low_unsigned;
1840                 ir_node *pred;
1841                 long     proj_nr;
1842                 ir_mode *mode_h;
1843                 ir_node *res_low;
1844                 ir_node *res_high;
1845                 dbg_info *dbg;
1846
1847                 if (!is_Proj(proj))
1848                         continue;
1849                 pred    = get_Proj_pred(proj);
1850                 proj_nr = get_Proj_proj(proj);
1851
1852                 if (mode == env->high_signed) {
1853                         mode_h = env->low_signed;
1854                 } else if (mode == env->high_unsigned) {
1855                         mode_h = env->low_unsigned;
1856                 } else {
1857                         long new_pn = new_projs[proj_nr];
1858                         if (new_pn != proj_nr) {
1859                                 ir_node *new_proj = new_r_Proj(pred, mode, new_pn);
1860                                 exchange(proj, new_proj);
1861                         }
1862                         continue;
1863                 }
1864
1865                 dbg = get_irn_dbg_info(proj);
1866                 if (env->params->little_endian) {
1867                         res_low  = new_rd_Proj(dbg, pred, mode_l, new_projs[proj_nr]);
1868                         res_high = new_rd_Proj(dbg, pred, mode_h, new_projs[proj_nr] + 1);
1869                 } else {
1870                         res_high = new_rd_Proj(dbg, pred, mode_h, new_projs[proj_nr]);
1871                         res_low  = new_rd_Proj(dbg, pred, mode_l, new_projs[proj_nr] + 1);
1872                 }
1873                 ir_set_dw_lowered(proj, res_low, res_high);
1874         }
1875 }
1876
1877 /**
1878  * Translate a Call.
1879  */
1880 static void lower_Call(ir_node *node, ir_mode *mode)
1881 {
1882         ir_type  *tp = get_Call_type(node);
1883         ir_node  **in;
1884         size_t   n_params, n_res;
1885         bool     need_lower = false;
1886         size_t   i, j;
1887         size_t   p;
1888         long     *res_numbers = NULL;
1889         ir_node  *resproj;
1890         const ir_edge_t *edge;
1891         const ir_edge_t *next;
1892         (void) mode;
1893
1894         n_params = get_method_n_params(tp);
1895         for (p = 0; p < n_params; ++p) {
1896                 ir_type *ptp = get_method_param_type(tp, p);
1897
1898                 if (is_Primitive_type(ptp)) {
1899                         ir_mode *mode = get_type_mode(ptp);
1900
1901                         if (mode == env->high_signed || mode == env->high_unsigned) {
1902                                 need_lower = true;
1903                                 break;
1904                         }
1905                 }
1906         }
1907         n_res = get_method_n_ress(tp);
1908         if (n_res > 0) {
1909                 NEW_ARR_A(long, res_numbers, n_res);
1910
1911                 for (i = j = 0; i < n_res; ++i, ++j) {
1912                         ir_type *ptp = get_method_res_type(tp, i);
1913
1914                         res_numbers[i] = j;
1915                         if (is_Primitive_type(ptp)) {
1916                                 ir_mode *mode = get_type_mode(ptp);
1917
1918                                 if (mode == env->high_signed || mode == env->high_unsigned) {
1919                                         need_lower = true;
1920                                         ++j;
1921                                 }
1922                         }
1923                 }
1924         }
1925
1926         if (! need_lower)
1927                 return;
1928
1929         /* let's lower it */
1930         tp = lower_mtp(tp);
1931         set_Call_type(node, tp);
1932
1933         NEW_ARR_A(ir_node *, in, get_method_n_params(tp) + 2);
1934
1935         in[0] = get_Call_mem(node);
1936         in[1] = get_Call_ptr(node);
1937
1938         for (j = 2, i = 0; i < n_params; ++i) {
1939                 ir_node *pred      = get_Call_param(node, i);
1940                 ir_mode *pred_mode = get_irn_mode(pred);
1941
1942                 if (pred_mode == env->high_signed || pred_mode == env->high_unsigned) {
1943                         const lower64_entry_t *pred_entry = get_node_entry(pred);
1944                         if (env->params->little_endian) {
1945                                 in[j++] = pred_entry->low_word;
1946                                 in[j++] = pred_entry->high_word;
1947                         } else {
1948                                 in[j++] = pred_entry->high_word;
1949                                 in[j++] = pred_entry->low_word;
1950                         }
1951                 } else {
1952                         in[j++] = pred;
1953                 }
1954         }
1955
1956         set_irn_in(node, j, in);
1957
1958         /* find results T */
1959         resproj = NULL;
1960         foreach_out_edge(node, edge) {
1961                 ir_node *proj = get_edge_src_irn(edge);
1962                 if (!is_Proj(proj))
1963                         continue;
1964                 if (get_Proj_proj(proj) == pn_Call_T_result) {
1965                         resproj = proj;
1966                         break;
1967                 }
1968         }
1969         if (resproj == NULL)
1970                 return;
1971
1972         /* fix the results */
1973         foreach_out_edge_safe(resproj, edge, next) {
1974                 ir_node  *proj      = get_edge_src_irn(edge);
1975                 ir_mode  *proj_mode = get_irn_mode(proj);
1976                 ir_mode  *mode_l    = env->low_unsigned;
1977                 ir_node  *pred;
1978                 long      proj_nr;
1979                 ir_mode  *mode_h;
1980                 ir_node  *res_low;
1981                 ir_node  *res_high;
1982                 dbg_info *dbg;
1983
1984                 if (!is_Proj(proj))
1985                         continue;
1986                 pred    = get_Proj_pred(proj);
1987                 proj_nr = get_Proj_proj(proj);
1988
1989                 if (proj_mode == env->high_signed) {
1990                         mode_h = env->low_signed;
1991                 } else if (proj_mode == env->high_unsigned) {
1992                         mode_h = env->low_unsigned;
1993                 } else {
1994                         long new_nr = res_numbers[proj_nr];
1995                         if (proj_nr != new_nr) {
1996                                 ir_node *new_proj = new_r_Proj(pred, proj_mode, new_nr);
1997                                 exchange(proj, new_proj);
1998                         }
1999                         continue;
2000                 }
2001
2002                 dbg = get_irn_dbg_info(proj);
2003                 if (env->params->little_endian) {
2004                         res_low  = new_rd_Proj(dbg, pred, mode_l, res_numbers[proj_nr]);
2005                         res_high = new_rd_Proj(dbg, pred, mode_h, res_numbers[proj_nr] + 1);
2006                 } else {
2007                         res_high = new_rd_Proj(dbg, pred, mode_h, res_numbers[proj_nr]);
2008                         res_low  = new_rd_Proj(dbg, pred, mode_l, res_numbers[proj_nr] + 1);
2009                 }
2010                 ir_set_dw_lowered(proj, res_low, res_high);
2011         }
2012 }
2013
2014 /**
2015  * Translate an Unknown into two.
2016  */
2017 static void lower_Unknown(ir_node *node, ir_mode *mode)
2018 {
2019         ir_mode  *low_mode = env->low_unsigned;
2020         ir_graph *irg      = get_irn_irg(node);
2021         ir_node  *res_low  = new_r_Unknown(irg, low_mode);
2022         ir_node  *res_high = new_r_Unknown(irg, mode);
2023         ir_set_dw_lowered(node, res_low, res_high);
2024 }
2025
2026 /**
2027  * Translate a Bad into two.
2028  */
2029 static void lower_Bad(ir_node *node, ir_mode *mode)
2030 {
2031         ir_mode  *low_mode = env->low_unsigned;
2032         ir_graph *irg      = get_irn_irg(node);
2033         ir_node  *res_low  = new_r_Bad(irg, low_mode);
2034         ir_node  *res_high = new_r_Bad(irg, mode);
2035         ir_set_dw_lowered(node, res_low, res_high);
2036 }
2037
2038 /**
2039  * Translate a Phi.
2040  *
2041  * First step: just create two templates
2042  */
2043 static void lower_Phi(ir_node *phi)
2044 {
2045         ir_mode  *mode = get_irn_mode(phi);
2046         int       i;
2047         int       arity;
2048         ir_node **in_l;
2049         ir_node **in_h;
2050         ir_node  *unk_l;
2051         ir_node  *unk_h;
2052         ir_node  *phi_l;
2053         ir_node  *phi_h;
2054         dbg_info *dbg;
2055         ir_node  *block;
2056         ir_graph *irg;
2057         ir_mode  *mode_l;
2058         ir_mode  *mode_h;
2059
2060         /* enqueue predecessors */
2061         arity = get_Phi_n_preds(phi);
2062         for (i = 0; i < arity; ++i) {
2063                 ir_node *pred = get_Phi_pred(phi, i);
2064                 pdeq_putr(env->waitq, pred);
2065         }
2066
2067         if (mode != env->high_signed && mode != env->high_unsigned)
2068                 return;
2069
2070         /* first create a new in array */
2071         NEW_ARR_A(ir_node *, in_l, arity);
2072         NEW_ARR_A(ir_node *, in_h, arity);
2073         irg    = get_irn_irg(phi);
2074         mode_l = env->low_unsigned;
2075         mode_h = mode == env->high_signed ? env->low_signed : env->low_unsigned;
2076         unk_l  = new_r_Dummy(irg, mode_l);
2077         unk_h  = new_r_Dummy(irg, mode_h);
2078         for (i = 0; i < arity; ++i) {
2079                 in_l[i] = unk_l;
2080                 in_h[i] = unk_h;
2081         }
2082
2083         dbg   = get_irn_dbg_info(phi);
2084         block = get_nodes_block(phi);
2085         phi_l = new_rd_Phi(dbg, block, arity, in_l, mode_l);
2086         phi_h = new_rd_Phi(dbg, block, arity, in_h, mode_h);
2087
2088         ir_set_dw_lowered(phi, phi_l, phi_h);
2089
2090         /* remember that we need to fixup the predecessors later */
2091         ARR_APP1(ir_node*, env->lowered_phis, phi);
2092 }
2093
2094 static void fixup_phi(ir_node *phi)
2095 {
2096         const lower64_entry_t *entry = get_node_entry(phi);
2097         ir_node               *phi_l = entry->low_word;
2098         ir_node               *phi_h = entry->high_word;
2099         int                    arity = get_Phi_n_preds(phi);
2100         int                    i;
2101
2102         /* exchange phi predecessors which are lowered by now */
2103         for (i = 0; i < arity; ++i) {
2104                 ir_node               *pred       = get_Phi_pred(phi, i);
2105                 const lower64_entry_t *pred_entry = get_node_entry(pred);
2106
2107                 set_Phi_pred(phi_l, i, pred_entry->low_word);
2108                 set_Phi_pred(phi_h, i, pred_entry->high_word);
2109         }
2110 }
2111
2112 /**
2113  * Translate a Mux.
2114  */
2115 static void lower_Mux(ir_node *mux, ir_mode *mode)
2116 {
2117         ir_node               *truen       = get_Mux_true(mux);
2118         ir_node               *falsen      = get_Mux_false(mux);
2119         ir_node               *sel         = get_Mux_sel(mux);
2120         const lower64_entry_t *true_entry  = get_node_entry(truen);
2121         const lower64_entry_t *false_entry = get_node_entry(falsen);
2122         ir_node               *true_l      = true_entry->low_word;
2123         ir_node               *true_h      = true_entry->high_word;
2124         ir_node               *false_l     = false_entry->low_word;
2125         ir_node               *false_h     = false_entry->high_word;
2126         dbg_info              *dbgi        = get_irn_dbg_info(mux);
2127         ir_node               *block       = get_nodes_block(mux);
2128         ir_node               *res_low
2129                 = new_rd_Mux(dbgi, block, sel, false_l, true_l, env->low_unsigned);
2130         ir_node               *res_high
2131                 = new_rd_Mux(dbgi, block, sel, false_h, true_h, mode);
2132         ir_set_dw_lowered(mux, res_low, res_high);
2133 }
2134
2135 /**
2136  * Translate an ASM node.
2137  */
2138 static void lower_ASM(ir_node *asmn, ir_mode *mode)
2139 {
2140         ir_mode           *high_signed        = env->high_signed;
2141         ir_mode           *high_unsigned      = env->high_unsigned;
2142         int                n_outs             = get_ASM_n_output_constraints(asmn);
2143         ir_asm_constraint *output_constraints = get_ASM_output_constraints(asmn);
2144         ir_asm_constraint *input_constraints  = get_ASM_input_constraints(asmn);
2145         unsigned           n_64bit_outs       = 0;
2146         int      i;
2147
2148         (void)mode;
2149
2150         for (i = get_irn_arity(asmn) - 1; i >= 0; --i) {
2151                 ir_node *op      = get_irn_n(asmn, i);
2152                 ir_mode *op_mode = get_irn_mode(op);
2153                 if (op_mode == high_signed || op_mode == high_unsigned) {
2154                         panic("lowering ASM 64bit input unimplemented");
2155                 }
2156         }
2157
2158         for (i = 0; i < n_outs; ++i) {
2159                 const ir_asm_constraint *constraint = &output_constraints[i];
2160                 if (constraint->mode == high_signed || constraint->mode == high_unsigned) {
2161                         const char *constr = get_id_str(constraint->constraint);
2162                         ++n_64bit_outs;
2163                         /* TODO: How to do this architecture neutral? This is very
2164                          * i386 specific... */
2165                         if (constr[0] != '=' || constr[1] != 'A') {
2166                                 panic("lowering ASM 64bit output only supports '=A' currently");
2167                         }
2168                 }
2169         }
2170
2171         if (n_64bit_outs == 0)
2172                 return;
2173
2174         {
2175                 dbg_info          *dbgi       = get_irn_dbg_info(asmn);
2176                 ir_node           *block      = get_nodes_block(asmn);
2177                 int                arity      = get_irn_arity(asmn);
2178                 ir_node          **in         = get_irn_in(asmn) + 1;
2179                 int                n_outs     = get_ASM_n_output_constraints(asmn);
2180                 int                new_n_outs = 0;
2181                 int                n_clobber  = get_ASM_n_clobbers(asmn);
2182                 long              *proj_map   = ALLOCAN(long, n_outs);
2183                 ident            **clobbers   = get_ASM_clobbers(asmn);
2184                 ident             *asm_text   = get_ASM_text(asmn);
2185                 ir_asm_constraint *new_outputs
2186                         = ALLOCAN(ir_asm_constraint, n_outs+n_64bit_outs);
2187                 ir_node           *new_asm;
2188                 const ir_edge_t   *edge;
2189                 const ir_edge_t   *next;
2190
2191                 for (i = 0; i < n_outs; ++i) {
2192                         const ir_asm_constraint *constraint = &output_constraints[i];
2193                         if (constraint->mode == high_signed || constraint->mode == high_unsigned) {
2194                                 new_outputs[new_n_outs].pos        = constraint->pos;
2195                                 new_outputs[new_n_outs].constraint = new_id_from_str("=a");
2196                                 new_outputs[new_n_outs].mode       = env->low_unsigned;
2197                                 proj_map[i] = new_n_outs;
2198                                 ++new_n_outs;
2199                                 new_outputs[new_n_outs].pos        = constraint->pos;
2200                                 new_outputs[new_n_outs].constraint = new_id_from_str("=d");
2201                                 if (constraint->mode == high_signed)
2202                                         new_outputs[new_n_outs].mode = env->low_signed;
2203                                 else
2204                                         new_outputs[new_n_outs].mode = env->low_unsigned;
2205                                 ++new_n_outs;
2206                         } else {
2207                                 new_outputs[new_n_outs] = *constraint;
2208                                 proj_map[i] = new_n_outs;
2209                                 ++new_n_outs;
2210                         }
2211                 }
2212                 assert(new_n_outs == n_outs+(int)n_64bit_outs);
2213
2214                 new_asm = new_rd_ASM(dbgi, block, arity, in, input_constraints,
2215                                      new_n_outs, new_outputs, n_clobber, clobbers,
2216                                      asm_text);
2217
2218                 foreach_out_edge_safe(asmn, edge, next) {
2219                         ir_node *proj      = get_edge_src_irn(edge);
2220                         ir_mode *proj_mode = get_irn_mode(proj);
2221                         long     pn;
2222
2223                         if (!is_Proj(proj))
2224                                 continue;
2225                         pn = get_Proj_proj(proj);
2226
2227                         if (pn < n_outs)
2228                                 pn = proj_map[pn];
2229                         else
2230                                 pn = new_n_outs + pn - n_outs;
2231
2232                         if (proj_mode == high_signed || proj_mode == high_unsigned) {
2233                                 ir_mode *high_mode
2234                                         = proj_mode == high_signed ? env->low_signed : env->low_unsigned;
2235                                 ir_node *np_low  = new_r_Proj(new_asm, env->low_unsigned, pn);
2236                                 ir_node *np_high = new_r_Proj(new_asm, high_mode, pn+1);
2237                                 ir_set_dw_lowered(proj, np_low, np_high);
2238                         } else {
2239                                 ir_node *np = new_r_Proj(new_asm, proj_mode, pn);
2240                                 exchange(proj, np);
2241                         }
2242                 }
2243         }
2244 }
2245
2246 /**
2247  * Translate a Sel node.
2248  */
2249 static void lower_Sel(ir_node *sel, ir_mode *mode)
2250 {
2251         (void) mode;
2252
2253         /* we must only lower value parameter Sels if we change the
2254            value parameter type. */
2255         if (env->value_param_tp != NULL) {
2256                 ir_entity *ent = get_Sel_entity(sel);
2257             if (get_entity_owner(ent) == env->value_param_tp) {
2258                         size_t pos = get_entity_arg_idx(ent);
2259
2260                         ent = get_method_value_param_ent(env->l_mtp, pos);
2261                         set_Sel_entity(sel, ent);
2262                 }
2263         }
2264 }
2265
2266 /**
2267  * check for opcodes that must always be lowered.
2268  */
2269 static bool always_lower(unsigned code)
2270 {
2271         switch (code) {
2272         case iro_ASM:
2273         case iro_Proj:
2274         case iro_Start:
2275         case iro_Call:
2276         case iro_Return:
2277         case iro_Cond:
2278         case iro_Conv:
2279         case iro_Sel:
2280                 return true;
2281         default:
2282                 return false;
2283         }
2284 }
2285
2286 /**
2287  * Compare two op_mode_entry_t's.
2288  */
2289 static int cmp_op_mode(const void *elt, const void *key, size_t size)
2290 {
2291         const op_mode_entry_t *e1 = (const op_mode_entry_t*)elt;
2292         const op_mode_entry_t *e2 = (const op_mode_entry_t*)key;
2293         (void) size;
2294
2295         return (e1->op != e2->op) | (e1->imode != e2->imode) | (e1->omode != e2->omode);
2296 }
2297
2298 /**
2299  * Compare two conv_tp_entry_t's.
2300  */
2301 static int cmp_conv_tp(const void *elt, const void *key, size_t size)
2302 {
2303         const conv_tp_entry_t *e1 = (const conv_tp_entry_t*)elt;
2304         const conv_tp_entry_t *e2 = (const conv_tp_entry_t*)key;
2305         (void) size;
2306
2307         return (e1->imode != e2->imode) | (e1->omode != e2->omode);
2308 }
2309
2310 /**
2311  * Enter a lowering function into an ir_op.
2312  */
2313 void ir_register_dw_lower_function(ir_op *op, lower_dw_func func)
2314 {
2315         op->ops.generic = (op_func)func;
2316 }
2317
2318 /**
2319  * Returns non-zero if a method type must be lowered.
2320  *
2321  * @param mtp  the method type
2322  */
2323 static bool mtp_must_be_lowered(ir_type *mtp)
2324 {
2325         size_t i, n_params = get_method_n_params(mtp);
2326
2327         /* first check if we have parameters that must be fixed */
2328         for (i = 0; i < n_params; ++i) {
2329                 ir_type *tp = get_method_param_type(mtp, i);
2330
2331                 if (is_Primitive_type(tp)) {
2332                         ir_mode *mode = get_type_mode(tp);
2333
2334                         if (mode == env->high_signed || mode == env->high_unsigned)
2335                                 return true;
2336                 }
2337         }
2338         return false;
2339 }
2340
2341 /* Determine which modes need to be lowered */
2342 static void setup_modes(void)
2343 {
2344         unsigned           size_bits           = env->params->doubleword_size;
2345         ir_mode           *doubleword_signed   = NULL;
2346         ir_mode           *doubleword_unsigned = NULL;
2347         size_t             n_modes             = get_irp_n_modes();
2348         ir_mode_arithmetic arithmetic;
2349         unsigned           modulo_shift;
2350         size_t             i;
2351
2352         /* search for doubleword modes... */
2353         for (i = 0; i < n_modes; ++i) {
2354                 ir_mode *mode = get_irp_mode(i);
2355                 if (!mode_is_int(mode))
2356                         continue;
2357                 if (get_mode_size_bits(mode) != size_bits)
2358                         continue;
2359                 if (mode_is_signed(mode)) {
2360                         if (doubleword_signed != NULL) {
2361                                 /* sigh - the lowerer should really just lower all mode with
2362                                  * size_bits it finds. Unfortunately this required a bigger
2363                                  * rewrite. */
2364                                 panic("multiple double word signed modes found");
2365                         }
2366                         doubleword_signed = mode;
2367                 } else {
2368                         if (doubleword_unsigned != NULL) {
2369                                 /* sigh - the lowerer should really just lower all mode with
2370                                  * size_bits it finds. Unfortunately this required a bigger
2371                                  * rewrite. */
2372                                 panic("multiple double word unsigned modes found");
2373                         }
2374                         doubleword_unsigned = mode;
2375                 }
2376         }
2377         if (doubleword_signed == NULL || doubleword_unsigned == NULL) {
2378                 panic("Couldn't find doubleword modes");
2379         }
2380
2381         arithmetic   = get_mode_arithmetic(doubleword_signed);
2382         modulo_shift = get_mode_modulo_shift(doubleword_signed);
2383
2384         assert(get_mode_size_bits(doubleword_unsigned) == size_bits);
2385         assert(size_bits % 2 == 0);
2386         assert(get_mode_sign(doubleword_signed) == 1);
2387         assert(get_mode_sign(doubleword_unsigned) == 0);
2388         assert(get_mode_sort(doubleword_signed) == irms_int_number);
2389         assert(get_mode_sort(doubleword_unsigned) == irms_int_number);
2390         assert(get_mode_arithmetic(doubleword_unsigned) == arithmetic);
2391         assert(get_mode_modulo_shift(doubleword_unsigned) == modulo_shift);
2392
2393         /* try to guess a sensible modulo shift for the new mode.
2394          * (This is IMO another indication that this should really be a node
2395          *  attribute instead of a mode thing) */
2396         if (modulo_shift == size_bits) {
2397                 modulo_shift = modulo_shift / 2;
2398         } else if (modulo_shift == 0) {
2399                 /* fine */
2400         } else {
2401                 panic("Don't know what new modulo shift to use for lowered doubleword mode");
2402         }
2403         size_bits /= 2;
2404
2405         /* produce lowered modes */
2406         env->high_signed   = doubleword_signed;
2407         env->high_unsigned = doubleword_unsigned;
2408         env->low_signed    = new_ir_mode("WS", irms_int_number, size_bits, 1,
2409                                          arithmetic, modulo_shift);
2410         env->low_unsigned  = new_ir_mode("WU", irms_int_number, size_bits, 0,
2411                                          arithmetic, modulo_shift);
2412 }
2413
2414 static void enqueue_preds(ir_node *node)
2415 {
2416         int arity = get_irn_arity(node);
2417         int i;
2418
2419         for (i = 0; i < arity; ++i) {
2420                 ir_node *pred = get_irn_n(node, i);
2421                 pdeq_putr(env->waitq, pred);
2422         }
2423 }
2424
2425 static void lower_node(ir_node *node)
2426 {
2427         int              arity;
2428         int              i;
2429         lower_dw_func    func;
2430         ir_op           *op;
2431         ir_mode         *mode;
2432         unsigned         idx;
2433         lower64_entry_t *entry;
2434
2435         if (irn_visited_else_mark(node))
2436                 return;
2437
2438         /* cycles are always broken at Phi and Block nodes. So we don't need special
2439          * magic in all the other lower functions */
2440         if (is_Block(node)) {
2441                 enqueue_preds(node);
2442                 return;
2443         } else if (is_Phi(node)) {
2444                 lower_Phi(node);
2445                 return;
2446         }
2447
2448         /* depth-first: descend into operands */
2449         if (!is_Block(node)) {
2450                 ir_node *block = get_nodes_block(node);
2451                 lower_node(block);
2452         }
2453
2454         if (!is_Cond(node)) {
2455                 arity = get_irn_arity(node);
2456                 for (i = 0; i < arity; ++i) {
2457                         ir_node *pred = get_irn_n(node, i);
2458                         lower_node(pred);
2459                 }
2460         }
2461
2462         op   = get_irn_op(node);
2463         func = (lower_dw_func) op->ops.generic;
2464         if (func == NULL)
2465                 return;
2466
2467         idx   = get_irn_idx(node);
2468         entry = idx < env->n_entries ? env->entries[idx] : NULL;
2469         if (entry != NULL || always_lower(get_irn_opcode(node))) {
2470                 mode = get_irn_op_mode(node);
2471                 if (mode == env->high_signed) {
2472                         mode = env->low_signed;
2473                 } else {
2474                         mode = env->low_unsigned;
2475                 }
2476                 DB((dbg, LEVEL_1, "  %+F\n", node));
2477                 func(node, mode);
2478         }
2479 }
2480
2481 static void clear_node_and_phi_links(ir_node *node, void *data)
2482 {
2483         (void) data;
2484         if (get_irn_mode(node) == mode_T) {
2485                 set_irn_link(node, node);
2486         } else {
2487                 set_irn_link(node, NULL);
2488         }
2489         if (is_Block(node))
2490                 set_Block_phis(node, NULL);
2491         else if (is_Phi(node))
2492                 set_Phi_next(node, NULL);
2493 }
2494
2495 static void lower_irg(ir_graph *irg)
2496 {
2497         ir_entity *ent;
2498         ir_type   *mtp;
2499         unsigned   n_idx;
2500
2501         obstack_init(&env->obst);
2502
2503         /* just here for debugging */
2504         current_ir_graph = irg;
2505         edges_assure(irg);
2506
2507         n_idx = get_irg_last_idx(irg);
2508         n_idx = n_idx + (n_idx >> 2);  /* add 25% */
2509         env->n_entries = n_idx;
2510         env->entries   = NEW_ARR_F(lower64_entry_t*, n_idx);
2511         memset(env->entries, 0, sizeof(env->entries[0]) * n_idx);
2512
2513         env->irg            = irg;
2514         env->l_mtp          = NULL;
2515         env->flags          = 0;
2516         env->value_param_tp = NULL;
2517
2518         ent = get_irg_entity(irg);
2519         mtp = get_entity_type(ent);
2520
2521         if (mtp_must_be_lowered(mtp)) {
2522                 ir_type *ltp = lower_mtp(mtp);
2523                 /* Do not update the entity type yet, this will be done by lower_Start! */
2524                 env->flags |= MUST_BE_LOWERED;
2525                 env->l_mtp = ltp;
2526                 env->value_param_tp = get_method_value_param_type(mtp);
2527         }
2528
2529         /* first step: link all nodes and allocate data */
2530         ir_reserve_resources(irg, IR_RESOURCE_PHI_LIST | IR_RESOURCE_IRN_LINK);
2531         visit_all_identities(irg, clear_node_and_phi_links, NULL);
2532         irg_walk_graph(irg, NULL, prepare_links_and_handle_rotl, env);
2533
2534         if (env->flags & MUST_BE_LOWERED) {
2535                 size_t i;
2536                 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
2537                 inc_irg_visited(irg);
2538
2539                 assert(pdeq_empty(env->waitq));
2540                 pdeq_putr(env->waitq, get_irg_end(irg));
2541
2542                 env->lowered_phis = NEW_ARR_F(ir_node*, 0);
2543                 while (!pdeq_empty(env->waitq)) {
2544                         ir_node *node = (ir_node*)pdeq_getl(env->waitq);
2545                         lower_node(node);
2546                 }
2547
2548                 /* we need to fixup phis */
2549                 for (i = 0; i < ARR_LEN(env->lowered_phis); ++i) {
2550                         ir_node *phi = env->lowered_phis[i];
2551                         fixup_phi(phi);
2552                 }
2553                 DEL_ARR_F(env->lowered_phis);
2554
2555
2556                 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
2557
2558                 if (env->flags & CF_CHANGED) {
2559                         /* control flow changed, dominance info is invalid */
2560                         set_irg_doms_inconsistent(irg);
2561                         set_irg_extblk_inconsistent(irg);
2562                 }
2563         }
2564
2565         ir_free_resources(irg, IR_RESOURCE_PHI_LIST | IR_RESOURCE_IRN_LINK);
2566
2567         DEL_ARR_F(env->entries);
2568         obstack_free(&env->obst, NULL);
2569 }
2570
2571 static const lwrdw_param_t *param;
2572
2573 void ir_prepare_dw_lowering(const lwrdw_param_t *new_param)
2574 {
2575         assert(new_param != NULL);
2576         FIRM_DBG_REGISTER(dbg, "firm.lower.dw");
2577
2578         param = new_param;
2579
2580         clear_irp_opcodes_generic_func();
2581         ir_register_dw_lower_function(op_ASM,     lower_ASM);
2582         ir_register_dw_lower_function(op_Add,     lower_binop);
2583         ir_register_dw_lower_function(op_And,     lower_And);
2584         ir_register_dw_lower_function(op_Bad,     lower_Bad);
2585         ir_register_dw_lower_function(op_Call,    lower_Call);
2586         ir_register_dw_lower_function(op_Cmp,     lower_Cmp);
2587         ir_register_dw_lower_function(op_Cond,    lower_Cond);
2588         ir_register_dw_lower_function(op_Const,   lower_Const);
2589         ir_register_dw_lower_function(op_Conv,    lower_Conv);
2590         ir_register_dw_lower_function(op_Div,     lower_Div);
2591         ir_register_dw_lower_function(op_Eor,     lower_Eor);
2592         ir_register_dw_lower_function(op_Load,    lower_Load);
2593         ir_register_dw_lower_function(op_Minus,   lower_unop);
2594         ir_register_dw_lower_function(op_Mod,     lower_Mod);
2595         ir_register_dw_lower_function(op_Mul,     lower_binop);
2596         ir_register_dw_lower_function(op_Mux,     lower_Mux);
2597         ir_register_dw_lower_function(op_Not,     lower_Not);
2598         ir_register_dw_lower_function(op_Or,      lower_Or);
2599         ir_register_dw_lower_function(op_Return,  lower_Return);
2600         ir_register_dw_lower_function(op_Sel,     lower_Sel);
2601         ir_register_dw_lower_function(op_Shl,     lower_Shl);
2602         ir_register_dw_lower_function(op_Shr,     lower_Shr);
2603         ir_register_dw_lower_function(op_Shrs,    lower_Shrs);
2604         ir_register_dw_lower_function(op_Start,   lower_Start);
2605         ir_register_dw_lower_function(op_Store,   lower_Store);
2606         ir_register_dw_lower_function(op_Sub,     lower_binop);
2607         ir_register_dw_lower_function(op_Unknown, lower_Unknown);
2608 }
2609
2610 /*
2611  * Do the lowering.
2612  */
2613 void ir_lower_dw_ops(void)
2614 {
2615         lower_dw_env_t lenv;
2616         size_t      i, n;
2617
2618         memset(&lenv, 0, sizeof(lenv));
2619         lenv.params = param;
2620         env = &lenv;
2621
2622         setup_modes();
2623
2624         /* create the necessary maps */
2625         if (! intrinsic_fkt)
2626                 intrinsic_fkt = new_set(cmp_op_mode, iro_Last + 1);
2627         if (! conv_types)
2628                 conv_types = new_set(cmp_conv_tp, 16);
2629         if (! lowered_type)
2630                 lowered_type = pmap_create();
2631
2632         /* create a primitive unsigned and signed type */
2633         if (! tp_u)
2634                 tp_u = get_type_for_mode(lenv.low_unsigned);
2635         if (! tp_s)
2636                 tp_s = get_type_for_mode(lenv.low_signed);
2637
2638         /* create method types for the created binop calls */
2639         if (! binop_tp_u) {
2640                 binop_tp_u = new_type_method(4, 2);
2641                 set_method_param_type(binop_tp_u, 0, tp_u);
2642                 set_method_param_type(binop_tp_u, 1, tp_u);
2643                 set_method_param_type(binop_tp_u, 2, tp_u);
2644                 set_method_param_type(binop_tp_u, 3, tp_u);
2645                 set_method_res_type(binop_tp_u, 0, tp_u);
2646                 set_method_res_type(binop_tp_u, 1, tp_u);
2647         }
2648         if (! binop_tp_s) {
2649                 binop_tp_s = new_type_method(4, 2);
2650                 if (env->params->little_endian) {
2651                         set_method_param_type(binop_tp_s, 0, tp_u);
2652                         set_method_param_type(binop_tp_s, 1, tp_s);
2653                         set_method_param_type(binop_tp_s, 2, tp_u);
2654                         set_method_param_type(binop_tp_s, 3, tp_s);
2655                         set_method_res_type(binop_tp_s, 0, tp_u);
2656                         set_method_res_type(binop_tp_s, 1, tp_s);
2657                 } else {
2658                         set_method_param_type(binop_tp_s, 0, tp_s);
2659                         set_method_param_type(binop_tp_s, 1, tp_u);
2660                         set_method_param_type(binop_tp_s, 2, tp_s);
2661                         set_method_param_type(binop_tp_s, 3, tp_u);
2662                         set_method_res_type(binop_tp_s, 0, tp_s);
2663                         set_method_res_type(binop_tp_s, 1, tp_u);
2664                 }
2665         }
2666         if (! unop_tp_u) {
2667                 unop_tp_u = new_type_method(2, 2);
2668                 set_method_param_type(unop_tp_u, 0, tp_u);
2669                 set_method_param_type(unop_tp_u, 1, tp_u);
2670                 set_method_res_type(unop_tp_u, 0, tp_u);
2671                 set_method_res_type(unop_tp_u, 1, tp_u);
2672         }
2673         if (! unop_tp_s) {
2674                 unop_tp_s = new_type_method(2, 2);
2675                 if (env->params->little_endian) {
2676                         set_method_param_type(unop_tp_s, 0, tp_u);
2677                         set_method_param_type(unop_tp_s, 1, tp_s);
2678                         set_method_res_type(unop_tp_s, 0, tp_u);
2679                         set_method_res_type(unop_tp_s, 1, tp_s);
2680                 } else {
2681                         set_method_param_type(unop_tp_s, 0, tp_s);
2682                         set_method_param_type(unop_tp_s, 1, tp_u);
2683                         set_method_res_type(unop_tp_s, 0, tp_s);
2684                         set_method_res_type(unop_tp_s, 1, tp_u);
2685                 }
2686         }
2687
2688         lenv.tv_mode_bytes = new_tarval_from_long(param->doubleword_size/(2*8), lenv.low_unsigned);
2689         lenv.tv_mode_bits  = new_tarval_from_long(param->doubleword_size/2, lenv.low_unsigned);
2690         lenv.waitq         = new_pdeq();
2691         lenv.first_id      = new_id_from_chars(param->little_endian ? ".l" : ".h", 2);
2692         lenv.next_id       = new_id_from_chars(param->little_endian ? ".h" : ".l", 2);
2693
2694         /* transform all graphs */
2695         for (i = 0, n = get_irp_n_irgs(); i < n; ++i) {
2696                 ir_graph *irg = get_irp_irg(i);
2697                 lower_irg(irg);
2698         }
2699         del_pdeq(lenv.waitq);
2700
2701         env = NULL;
2702 }
2703
2704 /* Default implementation. */
2705 ir_entity *def_create_intrinsic_fkt(ir_type *method, const ir_op *op,
2706                                     const ir_mode *imode, const ir_mode *omode,
2707                                     void *context)
2708 {
2709         char buf[64];
2710         ident *id;
2711         ir_entity *ent;
2712         (void) context;
2713
2714         if (imode == omode) {
2715                 snprintf(buf, sizeof(buf), "__l%s%s", get_op_name(op), get_mode_name(imode));
2716         } else {
2717                 snprintf(buf, sizeof(buf), "__l%s%s%s", get_op_name(op),
2718                         get_mode_name(imode), get_mode_name(omode));
2719         }
2720         id = new_id_from_str(buf);
2721
2722         ent = new_entity(get_glob_type(), id, method);
2723         set_entity_ld_ident(ent, get_entity_ident(ent));
2724         set_entity_visibility(ent, ir_visibility_external);
2725         return ent;
2726 }