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 ir_mode *ir_get_low_unsigned_mode(void)
 308 {
 309         return env->low_unsigned;
 310 }
 311
 312 /**
 313  * Translate a Constant: create two.
 314  */
 315 static void lower_Const(ir_node *node, ir_mode *mode)
 316 {
 317         ir_graph  *irg      = get_irn_irg(node);
 318         dbg_info  *dbg      = get_irn_dbg_info(node);
 319         ir_mode   *low_mode = env->low_unsigned;
 320         ir_tarval *tv       = get_Const_tarval(node);
 321         ir_tarval *tv_l     = tarval_convert_to(tv, low_mode);
 322         ir_node   *res_low  = new_rd_Const(dbg, irg, tv_l);
 323         ir_tarval *tv_shrs  = tarval_shrs(tv, env->tv_mode_bits);
 324         ir_tarval *tv_h     = tarval_convert_to(tv_shrs, mode);
 325         ir_node   *res_high = new_rd_Const(dbg, irg, tv_h);
 326
 327         ir_set_dw_lowered(node, res_low, res_high);
 328 }
 329
 330 /**
 331  * Translate a Load: create two.
 332  */
 333 static void lower_Load(ir_node *node, ir_mode *mode)
 334 {
 335         ir_mode    *low_mode = env->low_unsigned;
 336         ir_graph   *irg = get_irn_irg(node);
 337         ir_node    *adr = get_Load_ptr(node);
 338         ir_node    *mem = get_Load_mem(node);
 339         ir_node    *low, *high, *proj;
 340         dbg_info   *dbg;
 341         ir_node    *block = get_nodes_block(node);
 342         ir_cons_flags volatility = get_Load_volatility(node) == volatility_is_volatile
 343                                  ? cons_volatile : cons_none;
 344         const ir_edge_t *edge;
 345         const ir_edge_t *next;
 346
 347         if (env->params->little_endian) {
 348                 low  = adr;
 349                 high = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
 350         } else {
 351                 low  = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
 352                 high = adr;
 353         }
 354
 355         /* create two loads */
 356         dbg  = get_irn_dbg_info(node);
 357         low  = new_rd_Load(dbg, block, mem,  low,  low_mode, volatility);
 358         proj = new_r_Proj(low, mode_M, pn_Load_M);
 359         high = new_rd_Load(dbg, block, proj, high, mode, volatility);
 360
 361         foreach_out_edge_safe(node, edge, next) {
 362                 ir_node *proj = get_edge_src_irn(edge);
 363                 if (!is_Proj(proj))
 364                         continue;
 365
 366                 switch (get_Proj_proj(proj)) {
 367                 case pn_Load_M:         /* Memory result. */
 368                         /* put it to the second one */
 369                         set_Proj_pred(proj, high);
 370                         break;
 371                 case pn_Load_X_except:  /* Execution result if exception occurred. */
 372                         /* put it to the first one */
 373                         set_Proj_pred(proj, low);
 374                         break;
 375                 case pn_Load_res: {       /* Result of load operation. */
 376                         ir_node *res_low  = new_r_Proj(low,  low_mode, pn_Load_res);
 377                         ir_node *res_high = new_r_Proj(high, mode,     pn_Load_res);
 378                         ir_set_dw_lowered(proj, res_low, res_high);
 379                         break;
 380                 }
 381                 default:
 382                         assert(0 && "unexpected Proj number");
 383                 }
 384                 /* mark this proj: we have handled it already, otherwise we might fall
 385                  * into out new nodes. */
 386                 mark_irn_visited(proj);
 387         }
 388 }
 389
 390 /**
 391  * Translate a Store: create two.
 392  */
 393 static void lower_Store(ir_node *node, ir_mode *mode)
 394 {
 395         ir_graph              *irg;
 396         ir_node               *block, *adr, *mem;
 397         ir_node               *low, *high, *proj;
 398         dbg_info              *dbg;
 399         ir_node               *value = get_Store_value(node);
 400         const lower64_entry_t *entry = get_node_entry(value);
 401         ir_cons_flags volatility = get_Store_volatility(node) == volatility_is_volatile
 402                                    ? cons_volatile : cons_none;
 403         const ir_edge_t *edge;
 404         const ir_edge_t *next;
 405         (void) mode;
 406
 407         assert(entry);
 408
 409         if (! entry->low_word) {
 410                 /* not ready yet, wait */
 411                 pdeq_putr(env->waitq, node);
 412                 return;
 413         }
 414
 415         irg = get_irn_irg(node);
 416         adr = get_Store_ptr(node);
 417         mem = get_Store_mem(node);
 418         block = get_nodes_block(node);
 419
 420         if (env->params->little_endian) {
 421                 low  = adr;
 422                 high = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
 423         } else {
 424                 low  = new_r_Add(block, adr, new_r_Const(irg, env->tv_mode_bytes), get_irn_mode(adr));
 425                 high = adr;
 426         }
 427
 428         /* create two Stores */
 429         dbg = get_irn_dbg_info(node);
 430         low  = new_rd_Store(dbg, block, mem, low,  entry->low_word, volatility);
 431         proj = new_r_Proj(low, mode_M, pn_Store_M);
 432         high = new_rd_Store(dbg, block, proj, high, entry->high_word, volatility);
 433
 434         foreach_out_edge_safe(node, edge, next) {
 435                 ir_node *proj = get_edge_src_irn(edge);
 436                 if (!is_Proj(proj))
 437                         continue;
 438
 439                 switch (get_Proj_proj(proj)) {
 440                 case pn_Store_M:         /* Memory result. */
 441                         /* put it to the second one */
 442                         set_Proj_pred(proj, high);
 443                         break;
 444                 case pn_Store_X_except:  /* Execution result if exception occurred. */
 445                         /* put it to the first one */
 446                         set_Proj_pred(proj, low);
 447                         break;
 448                 default:
 449                         assert(0 && "unexpected Proj number");
 450                 }
 451                 /* mark this proj: we have handled it already, otherwise we might fall into
 452                  * out new nodes. */
 453                 mark_irn_visited(proj);
 454         }
 455 }
 456
 457 /**
 458  * Return a node containing the address of the intrinsic emulation function.
 459  *
 460  * @param method  the method type of the emulation function
 461  * @param op      the emulated ir_op
 462  * @param imode   the input mode of the emulated opcode
 463  * @param omode   the output mode of the emulated opcode
 464  * @param env     the lower environment
 465  */
 466 static ir_node *get_intrinsic_address(ir_type *method, ir_op *op,
 467                                       ir_mode *imode, ir_mode *omode)
 468 {
 469         symconst_symbol sym;
 470         ir_entity *ent;
 471         op_mode_entry_t key, *entry;
 472
 473         key.op    = op;
 474         key.imode = imode;
 475         key.omode = omode;
 476         key.ent   = NULL;
 477
 478         entry = (op_mode_entry_t*)set_insert(intrinsic_fkt, &key, sizeof(key),
 479                                 HASH_PTR(op) ^ HASH_PTR(imode) ^ (HASH_PTR(omode) << 8));
 480         if (! entry->ent) {
 481                 /* create a new one */
 482                 ent = env->params->create_intrinsic(method, op, imode, omode, env->params->ctx);
 483
 484                 assert(ent && "Intrinsic creator must return an entity");
 485                 entry->ent = ent;
 486         } else {
 487                 ent = entry->ent;
 488         }
 489         sym.entity_p = ent;
 490         return new_r_SymConst(env->irg, mode_P_code, sym, symconst_addr_ent);
 491 }
 492
 493 /**
 494  * Translate a Div.
 495  *
 496  * Create an intrinsic Call.
 497  */
 498 static void lower_Div(ir_node *node, ir_mode *mode)
 499 {
 500         ir_node         *left   = get_Div_left(node);
 501         ir_node         *right  = get_Div_right(node);
 502         ir_node         *block  = get_nodes_block(node);
 503         dbg_info        *dbgi   = get_irn_dbg_info(node);
 504         ir_type         *mtp    = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
 505         ir_mode         *opmode = get_irn_op_mode(node);
 506         ir_node         *addr
 507             = get_intrinsic_address(mtp, get_irn_op(node), opmode, opmode);
 508         ir_node         *in[4];
 509         ir_node         *call;
 510         ir_node         *resproj;
 511         const ir_edge_t *edge;
 512         const ir_edge_t *next;
 513
 514         if (env->params->little_endian) {
 515                 in[0] = get_lowered_low(left);
 516                 in[1] = get_lowered_high(left);
 517                 in[2] = get_lowered_low(right);
 518                 in[3] = get_lowered_high(right);
 519         } else {
 520                 in[0] = get_lowered_high(left);
 521                 in[1] = get_lowered_low(left);
 522                 in[2] = get_lowered_high(right);
 523                 in[3] = get_lowered_low(right);
 524         }
 525         call    = new_rd_Call(dbgi, block, get_Div_mem(node), addr, 4, in, mtp);
 526         resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
 527         set_irn_pinned(call, get_irn_pinned(node));
 528
 529         foreach_out_edge_safe(node, edge, next) {
 530                 ir_node *proj = get_edge_src_irn(edge);
 531                 if (!is_Proj(proj))
 532                         continue;
 533
 534                 switch (get_Proj_proj(proj)) {
 535                 case pn_Div_M:         /* Memory result. */
 536                         /* reroute to the call */
 537                         set_Proj_pred(proj, call);
 538                         set_Proj_proj(proj, pn_Call_M);
 539                         break;
 540                 case pn_Div_X_regular:
 541                         set_Proj_pred(proj, call);
 542                         set_Proj_proj(proj, pn_Call_X_regular);
 543                         break;
 544                 case pn_Div_X_except:
 545                         set_Proj_pred(proj, call);
 546                         set_Proj_proj(proj, pn_Call_X_except);
 547                         break;
 548                 case pn_Div_res:
 549                         if (env->params->little_endian) {
 550                                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
 551                                 ir_node *res_high = new_r_Proj(resproj, mode,              1);
 552                                 ir_set_dw_lowered(proj, res_low, res_high);
 553                         } else {
 554                                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
 555                                 ir_node *res_high = new_r_Proj(resproj, mode,              0);
 556                                 ir_set_dw_lowered(proj, res_low, res_high);
 557                         }
 558                         break;
 559                 default:
 560                         assert(0 && "unexpected Proj number");
 561                 }
 562                 /* mark this proj: we have handled it already, otherwise we might fall into
 563                  * out new nodes. */
 564                 mark_irn_visited(proj);
 565         }
 566 }
 567
 568 /**
 569  * Translate a Mod.
 570  *
 571  * Create an intrinsic Call.
 572  */
 573 static void lower_Mod(ir_node *node, ir_mode *mode)
 574 {
 575         ir_node         *left   = get_Mod_left(node);
 576         ir_node         *right  = get_Mod_right(node);
 577         dbg_info        *dbgi   = get_irn_dbg_info(node);
 578         ir_node         *block  = get_nodes_block(node);
 579         ir_type         *mtp    = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
 580         ir_mode         *opmode = get_irn_op_mode(node);
 581         ir_node         *addr
 582             = get_intrinsic_address(mtp, get_irn_op(node), opmode, opmode);
 583         ir_node         *in[4];
 584         ir_node         *call;
 585         ir_node         *resproj;
 586         const ir_edge_t *edge;
 587         const ir_edge_t *next;
 588
 589         if (env->params->little_endian) {
 590                 in[0] = get_lowered_low(left);
 591                 in[1] = get_lowered_high(left);
 592                 in[2] = get_lowered_low(right);
 593                 in[3] = get_lowered_high(right);
 594         } else {
 595                 in[0] = get_lowered_high(left);
 596                 in[1] = get_lowered_low(left);
 597                 in[2] = get_lowered_high(right);
 598                 in[3] = get_lowered_low(right);
 599         }
 600         call    = new_rd_Call(dbgi, block, get_Mod_mem(node), addr, 4, in, mtp);
 601         resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
 602         set_irn_pinned(call, get_irn_pinned(node));
 603
 604         foreach_out_edge_safe(node, edge, next) {
 605                 ir_node *proj = get_edge_src_irn(edge);
 606                 if (!is_Proj(proj))
 607                         continue;
 608
 609                 switch (get_Proj_proj(proj)) {
 610                 case pn_Mod_M:         /* Memory result. */
 611                         /* reroute to the call */
 612                         set_Proj_pred(proj, call);
 613                         set_Proj_proj(proj, pn_Call_M);
 614                         break;
 615                 case pn_Div_X_regular:
 616                         set_Proj_pred(proj, call);
 617                         set_Proj_proj(proj, pn_Call_X_regular);
 618                         break;
 619                 case pn_Mod_X_except:
 620                         set_Proj_pred(proj, call);
 621                         set_Proj_proj(proj, pn_Call_X_except);
 622                         break;
 623                 case pn_Mod_res:
 624                         if (env->params->little_endian) {
 625                                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
 626                                 ir_node *res_high = new_r_Proj(resproj, mode,              1);
 627                                 ir_set_dw_lowered(proj, res_low, res_high);
 628                         } else {
 629                                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
 630                                 ir_node *res_high = new_r_Proj(resproj, mode,              0);
 631                                 ir_set_dw_lowered(proj, res_low, res_high);
 632                         }
 633                         break;
 634                 default:
 635                         assert(0 && "unexpected Proj number");
 636                 }
 637                 /* mark this proj: we have handled it already, otherwise we might fall
 638                  * into out new nodes. */
 639                 mark_irn_visited(proj);
 640         }
 641 }
 642
 643 /**
 644  * Translate a binop.
 645  *
 646  * Create an intrinsic Call.
 647  */
 648 static void lower_binop(ir_node *node, ir_mode *mode)
 649 {
 650         ir_node  *left  = get_binop_left(node);
 651         ir_node  *right = get_binop_right(node);
 652         dbg_info *dbgi  = get_irn_dbg_info(node);
 653         ir_node  *block = get_nodes_block(node);
 654         ir_graph *irg   = get_irn_irg(block);
 655         ir_type  *mtp   = mode_is_signed(mode) ? binop_tp_s : binop_tp_u;
 656         ir_node  *addr  = get_intrinsic_address(mtp, get_irn_op(node), mode, mode);
 657         ir_node  *in[4];
 658         ir_node  *call;
 659         ir_node  *resproj;
 660
 661         if (env->params->little_endian) {
 662                 in[0] = get_lowered_low(left);
 663                 in[1] = get_lowered_high(left);
 664                 in[2] = get_lowered_low(right);
 665                 in[3] = get_lowered_high(right);
 666         } else {
 667                 in[0] = get_lowered_high(left);
 668                 in[1] = get_lowered_low(left);
 669                 in[2] = get_lowered_high(right);
 670                 in[3] = get_lowered_low(right);
 671         }
 672         call    = new_rd_Call(dbgi, block, get_irg_no_mem(irg), addr, 4, in, mtp);
 673         resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
 674         set_irn_pinned(call, get_irn_pinned(node));
 675
 676         if (env->params->little_endian) {
 677                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
 678                 ir_node *res_high = new_r_Proj(resproj, mode,              1);
 679                 ir_set_dw_lowered(node, res_low, res_high);
 680         } else {
 681                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
 682                 ir_node *res_high = new_r_Proj(resproj, mode,              0);
 683                 ir_set_dw_lowered(node, res_low, res_high);
 684         }
 685 }
 686
 687 static ir_node *create_conv(ir_node *block, ir_node *node, ir_mode *dest_mode)
 688 {
 689         if (get_irn_mode(node) == dest_mode)
 690                 return node;
 691         return new_r_Conv(block, node, dest_mode);
 692 }
 693
 694 /**
 695  * Moves node and all predecessors of node from from_bl to to_bl.
 696  * Does not move predecessors of Phi nodes (or block nodes).
 697  */
 698 static void move(ir_node *node, ir_node *from_bl, ir_node *to_bl)
 699 {
 700         int i, arity;
 701
 702         /* move this node */
 703         set_nodes_block(node, to_bl);
 704
 705         /* move its Projs */
 706         if (get_irn_mode(node) == mode_T) {
 707                 const ir_edge_t *edge;
 708                 foreach_out_edge(node, edge) {
 709                         ir_node *proj = get_edge_src_irn(edge);
 710                         if (!is_Proj(proj))
 711                                 continue;
 712                         move(proj, from_bl, to_bl);
 713                 }
 714         }
 715
 716         /* We must not move predecessors of Phi nodes, even if they are in
 717          * from_bl. (because these are values from an earlier loop iteration
 718          * which are not predecessors of node here)
 719          */
 720         if (is_Phi(node))
 721                 return;
 722
 723         /* recursion ... */
 724         arity = get_irn_arity(node);
 725         for (i = 0; i < arity; i++) {
 726                 ir_node *pred      = get_irn_n(node, i);
 727                 ir_mode *pred_mode = get_irn_mode(pred);
 728                 if (get_nodes_block(pred) == from_bl)
 729                         move(pred, from_bl, to_bl);
 730                 if (pred_mode == env->high_signed || pred_mode == env->high_unsigned) {
 731                         ir_node *pred_low  = get_lowered_low(pred);
 732                         ir_node *pred_high = get_lowered_high(pred);
 733                         if (get_nodes_block(pred_low) == from_bl)
 734                                 move(pred_low, from_bl, to_bl);
 735                         if (pred_high != NULL && get_nodes_block(pred_high) == from_bl)
 736                                 move(pred_high, from_bl, to_bl);
 737                 }
 738         }
 739 }
 740
 741 /**
 742  * We need a custom version of part_block_edges because during transformation
 743  * not all data-dependencies are explicit yet if a lowered nodes users are not
 744  * lowered yet.
 745  * We can fix this by modifying move to look for such implicit dependencies.
 746  * Additionally we have to keep the proj_2_block map updated
 747  */
 748 static ir_node *part_block_dw(ir_node *node)
 749 {
 750         ir_graph *irg        = get_irn_irg(node);
 751         ir_node  *old_block  = get_nodes_block(node);
 752         int       n_cfgpreds = get_Block_n_cfgpreds(old_block);
 753         ir_node **cfgpreds   = get_Block_cfgpred_arr(old_block);
 754         ir_node  *new_block  = new_r_Block(irg, n_cfgpreds, cfgpreds);
 755         const ir_edge_t *edge;
 756         const ir_edge_t *next;
 757
 758         /* old_block has no predecessors anymore for now */
 759         set_irn_in(old_block, 0, NULL);
 760
 761         /* move node and its predecessors to new_block */
 762         move(node, old_block, new_block);
 763
 764         /* move Phi nodes to new_block */
 765         foreach_out_edge_safe(old_block, edge, next) {
 766                 ir_node *phi = get_edge_src_irn(edge);
 767                 if (!is_Phi(phi))
 768                         continue;
 769                 set_nodes_block(phi, new_block);
 770         }
 771         return old_block;
 772 }
 773
 774 typedef ir_node* (*new_rd_shr_func)(dbg_info *dbgi, ir_node *block,
 775                                     ir_node *left, ir_node *right,
 776                                     ir_mode *mode);
 777
 778 static void lower_shr_helper(ir_node *node, ir_mode *mode,
 779                              new_rd_shr_func new_rd_shrs)
 780 {
 781         ir_node  *right         = get_binop_right(node);
 782         ir_node  *left          = get_binop_left(node);
 783         ir_mode  *shr_mode      = get_irn_mode(node);
 784         unsigned  modulo_shift  = get_mode_modulo_shift(shr_mode);
 785         ir_mode  *low_unsigned  = env->low_unsigned;
 786         unsigned  modulo_shift2 = get_mode_modulo_shift(mode);
 787         ir_graph *irg           = get_irn_irg(node);
 788         ir_node  *left_low      = get_lowered_low(left);
 789         ir_node  *left_high     = get_lowered_high(left);
 790         dbg_info *dbgi          = get_irn_dbg_info(node);
 791         ir_node  *lower_block;
 792         ir_node  *block;
 793         ir_node  *cnst;
 794         ir_node  *and;
 795         ir_node  *cmp;
 796         ir_node  *cond;
 797         ir_node  *proj_true;
 798         ir_node  *proj_false;
 799         ir_node  *phi_low;
 800         ir_node  *phi_high;
 801         ir_node  *lower_in[2];
 802         ir_node  *phi_low_in[2];
 803         ir_node  *phi_high_in[2];
 804
 805         /* this version is optimized for modulo shift architectures
 806          * (and can't handle anything else) */
 807         if (modulo_shift != get_mode_size_bits(shr_mode)
 808                         || modulo_shift2<<1 != modulo_shift) {
 809                 panic("Shr lowering only implemented for modulo shift shr operations");
 810         }
 811         if (!is_po2(modulo_shift) || !is_po2(modulo_shift2)) {
 812                 panic("Shr lowering only implemented for power-of-2 modes");
 813         }
 814         /* without 2-complement the -x instead of (bit_width-x) trick won't work */
 815         if (get_mode_arithmetic(shr_mode) != irma_twos_complement) {
 816                 panic("Shr lowering only implemented for two-complement modes");
 817         }
 818
 819         /* if the right operand is a 64bit value, we're only interested in the
 820          * lower word */
 821         if (get_irn_mode(right) == env->high_unsigned) {
 822                 right = get_lowered_low(right);
 823         } else {
 824                 /* shift should never have signed mode on the right */
 825                 ir_node *block = get_nodes_block(node);
 826                 assert(get_irn_mode(right) != env->high_signed);
 827                 right = create_conv(block, right, low_unsigned);
 828         }
 829
 830         lower_block = part_block_dw(node);
 831         env->flags |= CF_CHANGED;
 832         block = get_nodes_block(node);
 833
 834         /* add a Cmp to test if highest bit is set <=> whether we shift more
 835          * than half the word width */
 836         cnst       = new_r_Const_long(irg, low_unsigned, modulo_shift2);
 837         and        = new_r_And(block, right, cnst, low_unsigned);
 838         cnst       = new_r_Const(irg, get_mode_null(low_unsigned));
 839         cmp        = new_rd_Cmp(dbgi, block, and, cnst, ir_relation_equal);
 840         cond       = new_rd_Cond(dbgi, block, cmp);
 841         proj_true  = new_r_Proj(cond, mode_X, pn_Cond_true);
 842         proj_false = new_r_Proj(cond, mode_X, pn_Cond_false);
 843
 844         /* the true block => shift_width < 1word */
 845         {
 846                 /* In theory the low value (for 64bit shifts) is:
 847                  *    Or(High << (32-x)), Low >> x)
 848                  * In practice High << 32-x will fail when x is zero (since we have
 849                  * modulo shift and 32 will be 0). So instead we use:
 850                  *    Or(High<<1<<~x, Low >> x)
 851                  */
 852                 ir_node *in[1]        = { proj_true };
 853                 ir_node *block_true   = new_r_Block(irg, ARRAY_SIZE(in), in);
 854                 ir_node *res_high     = new_rd_shrs(dbgi, block_true, left_high,
 855                                                     right, mode);
 856                 ir_node *shift_low    = new_rd_Shr(dbgi, block_true, left_low, right,
 857                                                    low_unsigned);
 858                 ir_node *not_shiftval = new_rd_Not(dbgi, block_true, right,
 859                                                    low_unsigned);
 860                 ir_node *conv         = create_conv(block_true, left_high,
 861                                                     low_unsigned);
 862                 ir_node *one          = new_r_Const(irg, get_mode_one(low_unsigned));
 863                 ir_node *carry0       = new_rd_Shl(dbgi, block_true, conv, one,
 864                                                    low_unsigned);
 865                 ir_node *carry1       = new_rd_Shl(dbgi, block_true, carry0,
 866                                                    not_shiftval, low_unsigned);
 867                 ir_node *res_low      = new_rd_Or(dbgi, block_true, shift_low, carry1,
 868                                                   low_unsigned);
 869                 lower_in[0]           = new_r_Jmp(block_true);
 870                 phi_low_in[0]         = res_low;
 871                 phi_high_in[0]        = res_high;
 872         }
 873
 874         /* false block => shift_width > 1word */
 875         {
 876                 ir_node *in[1]       = { proj_false };
 877                 ir_node *block_false = new_r_Block(irg, ARRAY_SIZE(in), in);
 878                 ir_node *conv        = create_conv(block_false, left_high, low_unsigned);
 879                 ir_node *res_low     = new_rd_shrs(dbgi, block_false, conv, right,
 880                                                    low_unsigned);
 881                 int      cnsti       = modulo_shift2-1;
 882                 ir_node *cnst        = new_r_Const_long(irg, low_unsigned, cnsti);
 883                 ir_node *res_high;
 884                 if (new_rd_shrs == new_rd_Shrs) {
 885                         res_high = new_rd_shrs(dbgi, block_false, left_high, cnst, mode);
 886                 } else {
 887                         res_high = new_r_Const(irg, get_mode_null(mode));
 888                 }
 889                 lower_in[1]          = new_r_Jmp(block_false);
 890                 phi_low_in[1]        = res_low;
 891                 phi_high_in[1]       = res_high;
 892         }
 893
 894         /* patch lower block */
 895         set_irn_in(lower_block, ARRAY_SIZE(lower_in), lower_in);
 896         phi_low  = new_r_Phi(lower_block, ARRAY_SIZE(phi_low_in), phi_low_in,
 897                              low_unsigned);
 898         phi_high = new_r_Phi(lower_block, ARRAY_SIZE(phi_high_in), phi_high_in,
 899                              mode);
 900         ir_set_dw_lowered(node, phi_low, phi_high);
 901 }
 902
 903 static void lower_Shr(ir_node *node, ir_mode *mode)
 904 {
 905         lower_shr_helper(node, mode, new_rd_Shr);
 906 }
 907
 908 static void lower_Shrs(ir_node *node, ir_mode *mode)
 909 {
 910         lower_shr_helper(node, mode, new_rd_Shrs);
 911 }
 912
 913 static void lower_Shl(ir_node *node, ir_mode *mode)
 914 {
 915         ir_node  *right         = get_binop_right(node);
 916         ir_node  *left          = get_binop_left(node);
 917         ir_mode  *shr_mode      = get_irn_mode(node);
 918         unsigned  modulo_shift  = get_mode_modulo_shift(shr_mode);
 919         ir_mode  *low_unsigned  = env->low_unsigned;
 920         unsigned  modulo_shift2 = get_mode_modulo_shift(mode);
 921         ir_graph *irg           = get_irn_irg(node);
 922         ir_node  *left_low      = get_lowered_low(left);
 923         ir_node  *left_high     = get_lowered_high(left);
 924         dbg_info *dbgi          = get_irn_dbg_info(node);
 925         ir_node  *lower_block   = get_nodes_block(node);
 926         ir_node  *block;
 927         ir_node  *cnst;
 928         ir_node  *and;
 929         ir_node  *cmp;
 930         ir_node  *cond;
 931         ir_node  *proj_true;
 932         ir_node  *proj_false;
 933         ir_node  *phi_low;
 934         ir_node  *phi_high;
 935         ir_node  *lower_in[2];
 936         ir_node  *phi_low_in[2];
 937         ir_node  *phi_high_in[2];
 938
 939         /* this version is optimized for modulo shift architectures
 940          * (and can't handle anything else) */
 941         if (modulo_shift != get_mode_size_bits(shr_mode)
 942                         || modulo_shift2<<1 != modulo_shift) {
 943                 panic("Shr lowering only implemented for modulo shift shr operations");
 944         }
 945         if (!is_po2(modulo_shift) || !is_po2(modulo_shift2)) {
 946                 panic("Shr lowering only implemented for power-of-2 modes");
 947         }
 948         /* without 2-complement the -x instead of (bit_width-x) trick won't work */
 949         if (get_mode_arithmetic(shr_mode) != irma_twos_complement) {
 950                 panic("Shr lowering only implemented for two-complement modes");
 951         }
 952
 953         /* if the right operand is a 64bit value, we're only interested in the
 954          * lower word */
 955         if (get_irn_mode(right) == env->high_unsigned) {
 956                 right = get_lowered_low(right);
 957         } else {
 958                 /* shift should never have signed mode on the right */
 959                 assert(get_irn_mode(right) != env->high_signed);
 960                 right = create_conv(lower_block, right, low_unsigned);
 961         }
 962
 963         part_block_dw(node);
 964         env->flags |= CF_CHANGED;
 965         block = get_nodes_block(node);
 966
 967         /* add a Cmp to test if highest bit is set <=> whether we shift more
 968          * than half the word width */
 969         cnst       = new_r_Const_long(irg, low_unsigned, modulo_shift2);
 970         and        = new_r_And(block, right, cnst, low_unsigned);
 971         cnst       = new_r_Const(irg, get_mode_null(low_unsigned));
 972         cmp        = new_rd_Cmp(dbgi, block, and, cnst, ir_relation_equal);
 973         cond       = new_rd_Cond(dbgi, block, cmp);
 974         proj_true  = new_r_Proj(cond, mode_X, pn_Cond_true);
 975         proj_false = new_r_Proj(cond, mode_X, pn_Cond_false);
 976
 977         /* the true block => shift_width < 1word */
 978         {
 979                 ir_node *in[1]        = { proj_true };
 980                 ir_node *block_true   = new_r_Block(irg, ARRAY_SIZE(in), in);
 981
 982                 ir_node *res_low      = new_rd_Shl(dbgi, block_true, left_low,
 983                                                    right, low_unsigned);
 984                 ir_node *shift_high   = new_rd_Shl(dbgi, block_true, left_high, right,
 985                                                    mode);
 986                 ir_node *not_shiftval = new_rd_Not(dbgi, block_true, right,
 987                                                    low_unsigned);
 988                 ir_node *conv         = create_conv(block_true, left_low, mode);
 989                 ir_node *one          = new_r_Const(irg, get_mode_one(low_unsigned));
 990                 ir_node *carry0       = new_rd_Shr(dbgi, block_true, conv, one, mode);
 991                 ir_node *carry1       = new_rd_Shr(dbgi, block_true, carry0,
 992                                                    not_shiftval, mode);
 993                 ir_node *res_high     = new_rd_Or(dbgi, block_true, shift_high, carry1,
 994                                                   mode);
 995                 lower_in[0]           = new_r_Jmp(block_true);
 996                 phi_low_in[0]         = res_low;
 997                 phi_high_in[0]        = res_high;
 998         }
 999
1000         /* false block => shift_width > 1word */
1001         {
1002                 ir_node *in[1]       = { proj_false };
1003                 ir_node *block_false = new_r_Block(irg, ARRAY_SIZE(in), in);
1004                 ir_node *res_low     = new_r_Const(irg, get_mode_null(low_unsigned));
1005                 ir_node *conv        = create_conv(block_false, left_low, mode);
1006                 ir_node *res_high    = new_rd_Shl(dbgi, block_false, conv, right, mode);
1007                 lower_in[1]          = new_r_Jmp(block_false);
1008                 phi_low_in[1]        = res_low;
1009                 phi_high_in[1]       = res_high;
1010         }
1011
1012         /* patch lower block */
1013         set_irn_in(lower_block, ARRAY_SIZE(lower_in), lower_in);
1014         phi_low  = new_r_Phi(lower_block, ARRAY_SIZE(phi_low_in), phi_low_in,
1015                              low_unsigned);
1016         phi_high = new_r_Phi(lower_block, ARRAY_SIZE(phi_high_in), phi_high_in,
1017                              mode);
1018         ir_set_dw_lowered(node, phi_low, phi_high);
1019 }
1020
1021 /**
1022  * Rebuild Rotl nodes into Or(Shl, Shr) and prepare all nodes.
1023  */
1024 static void prepare_links_and_handle_rotl(ir_node *node, void *data)
1025 {
1026         (void) data;
1027         if (is_Rotl(node)) {
1028                 ir_mode  *mode = get_irn_op_mode(node);
1029                 ir_node  *right;
1030                 ir_node  *left, *shl, *shr, *ornode, *block, *sub, *c;
1031                 ir_mode  *omode, *rmode;
1032                 ir_graph *irg;
1033                 dbg_info *dbg;
1034                 optimization_state_t state;
1035
1036                 if (mode != env->high_signed && mode != env->high_unsigned) {
1037                         prepare_links(node);
1038                         return;
1039                 }
1040
1041                 /* replace the Rotl(x,y) by an Or(Shl(x,y), Shr(x,64-y)) */
1042                 right = get_Rotl_right(node);
1043                 irg   = get_irn_irg(node);
1044                 dbg   = get_irn_dbg_info(node);
1045                 omode = get_irn_mode(node);
1046                 left  = get_Rotl_left(node);
1047                 block = get_nodes_block(node);
1048                 shl   = new_rd_Shl(dbg, block, left, right, omode);
1049                 rmode = get_irn_mode(right);
1050                 c     = new_r_Const_long(irg, rmode, get_mode_size_bits(omode));
1051                 sub   = new_rd_Sub(dbg, block, c, right, rmode);
1052                 shr   = new_rd_Shr(dbg, block, left, sub, omode);
1053
1054                 /* switch optimization off here, or we will get the Rotl back */
1055                 save_optimization_state(&state);
1056                 set_opt_algebraic_simplification(0);
1057                 ornode = new_rd_Or(dbg, block, shl, shr, omode);
1058                 restore_optimization_state(&state);
1059
1060                 exchange(node, ornode);
1061
1062                 /* do lowering on the new nodes */
1063                 prepare_links(shl);
1064                 prepare_links(c);
1065                 prepare_links(sub);
1066                 prepare_links(shr);
1067                 prepare_links(ornode);
1068                 return;
1069         }
1070
1071         prepare_links(node);
1072 }
1073
1074 /**
1075  * Translate an Unop.
1076  *
1077  * Create an intrinsic Call.
1078  */
1079 static void lower_unop(ir_node *node, ir_mode *mode)
1080 {
1081         ir_node  *op       = get_unop_op(node);
1082         dbg_info *dbgi     = get_irn_dbg_info(node);
1083         ir_node  *block    = get_nodes_block(node);
1084         ir_graph *irg      = get_irn_irg(block);
1085         ir_type  *mtp      = mode_is_signed(mode) ? unop_tp_s : unop_tp_u;
1086         ir_op    *irop     = get_irn_op(node);
1087         ir_node  *addr     = get_intrinsic_address(mtp, irop, mode, mode);
1088         ir_node  *nomem    = get_irg_no_mem(irg);
1089         ir_node  *in[2];
1090         ir_node  *call;
1091         ir_node  *resproj;
1092
1093         if (env->params->little_endian) {
1094                 in[0] = get_lowered_low(op);
1095                 in[1] = get_lowered_high(op);
1096         } else {
1097                 in[0] = get_lowered_high(op);
1098                 in[1] = get_lowered_low(op);
1099         }
1100         call    = new_rd_Call(dbgi, block, nomem, addr, 2, in, mtp);
1101         resproj = new_r_Proj(call, mode_T, pn_Call_T_result);
1102         set_irn_pinned(call, get_irn_pinned(node));
1103
1104         if (env->params->little_endian) {
1105                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 0);
1106                 ir_node *res_high = new_r_Proj(resproj, mode,              1);
1107                 ir_set_dw_lowered(node, res_low, res_high);
1108         } else {
1109                 ir_node *res_low  = new_r_Proj(resproj, env->low_unsigned, 1);
1110                 ir_node *res_high = new_r_Proj(resproj, mode,              0);
1111                 ir_set_dw_lowered(node, res_low, res_high);
1112         }
1113 }
1114
1115 /**
1116  * Translate a logical binop.
1117  *
1118  * Create two logical binops.
1119  */
1120 static void lower_binop_logical(ir_node *node, ir_mode *mode,
1121                                                                 ir_node *(*constr_rd)(dbg_info *db, ir_node *block, ir_node *op1, ir_node *op2, ir_mode *mode) )
1122 {
1123         ir_node               *left        = get_binop_left(node);
1124         ir_node               *right       = get_binop_right(node);
1125         const lower64_entry_t *left_entry  = get_node_entry(left);
1126         const lower64_entry_t *right_entry = get_node_entry(right);
1127         dbg_info              *dbgi        = get_irn_dbg_info(node);
1128         ir_node               *block       = get_nodes_block(node);
1129         ir_node               *res_low
1130                 = constr_rd(dbgi, block, left_entry->low_word, right_entry->low_word,
1131                             env->low_unsigned);
1132         ir_node               *res_high
1133                 = constr_rd(dbgi, block, left_entry->high_word, right_entry->high_word,
1134                             mode);
1135         ir_set_dw_lowered(node, res_low, res_high);
1136 }
1137
1138 static void lower_And(ir_node *node, ir_mode *mode)
1139 {
1140         lower_binop_logical(node, mode, new_rd_And);
1141 }
1142
1143 static void lower_Or(ir_node *node, ir_mode *mode)
1144 {
1145         lower_binop_logical(node, mode, new_rd_Or);
1146 }
1147
1148 static void lower_Eor(ir_node *node, ir_mode *mode)
1149 {
1150         lower_binop_logical(node, mode, new_rd_Eor);
1151 }
1152
1153 /**
1154  * Translate a Not.
1155  *
1156  * Create two logical Nots.
1157  */
1158 static void lower_Not(ir_node *node, ir_mode *mode)
1159 {
1160         ir_node               *op       = get_Not_op(node);
1161         const lower64_entry_t *op_entry = get_node_entry(op);
1162         dbg_info              *dbgi     = get_irn_dbg_info(node);
1163         ir_node               *block    = get_nodes_block(node);
1164         ir_node               *res_low
1165                 = new_rd_Not(dbgi, block, op_entry->low_word, env->low_unsigned);
1166         ir_node               *res_high
1167                 = new_rd_Not(dbgi, block, op_entry->high_word, mode);
1168         ir_set_dw_lowered(node, res_low, res_high);
1169 }
1170
1171 static bool is_equality_cmp_0(const ir_node *node)
1172 {
1173         ir_relation relation = get_Cmp_relation(node);
1174         ir_node    *left     = get_Cmp_left(node);
1175         ir_node    *right    = get_Cmp_right(node);
1176         ir_mode    *mode     = get_irn_mode(left);
1177
1178         /* this probably makes no sense if unordered is involved */
1179         assert(!mode_is_float(mode));
1180
1181         if (!is_Const(right) || !is_Const_null(right))
1182                 return false;
1183         if (relation == ir_relation_equal)
1184                 return true;
1185         if (mode_is_signed(mode)) {
1186                 return relation == ir_relation_less_greater;
1187         } else {
1188                 return relation == ir_relation_greater;
1189         }
1190 }
1191
1192 static ir_node *get_cfop_destination(const ir_node *cfop)
1193 {
1194         const ir_edge_t *first = get_irn_out_edge_first(cfop);
1195         /* we should only have 1 destination */
1196         assert(get_irn_n_edges(cfop) == 1);
1197         return get_edge_src_irn(first);
1198 }
1199
1200 /**
1201  * Translate a Cond.
1202  */
1203 static void lower_Cond(ir_node *node, ir_mode *mode)
1204 {
1205         ir_node *left, *right, *block;
1206         ir_node *sel = get_Cond_selector(node);
1207         ir_mode *m = get_irn_mode(sel);
1208         ir_mode *cmp_mode;
1209         const lower64_entry_t *lentry, *rentry;
1210         ir_node  *proj, *projT = NULL, *projF = NULL;
1211         ir_node  *new_bl, *irn;
1212         ir_node  *projHF, *projHT;
1213         ir_node  *dst_blk;
1214         ir_relation relation;
1215         ir_graph *irg;
1216         dbg_info *dbg;
1217         const ir_edge_t *edge;
1218         const ir_edge_t *next;
1219
1220         (void) mode;
1221
1222         if (m != mode_b) {
1223                 if (m == env->high_signed || m == env->high_unsigned) {
1224                         /* bad we can't really handle Switch with 64bit offsets */
1225                         panic("Cond with 64bit jumptable not supported");
1226                 }
1227                 lower_node(sel);
1228                 return;
1229         }
1230
1231         if (!is_Cmp(sel)) {
1232                 lower_node(sel);
1233                 return;
1234         }
1235
1236         left     = get_Cmp_left(sel);
1237         cmp_mode = get_irn_mode(left);
1238         if (cmp_mode != env->high_signed && cmp_mode != env->high_unsigned) {
1239                 lower_node(sel);
1240                 return;
1241         }
1242
1243         right  = get_Cmp_right(sel);
1244         lower_node(left);
1245         lower_node(right);
1246         lentry = get_node_entry(left);
1247         rentry = get_node_entry(right);
1248
1249         /* all right, build the code */
1250         foreach_out_edge_safe(node, edge, next) {
1251                 ir_node *proj    = get_edge_src_irn(edge);
1252                 long     proj_nr;
1253                 if (!is_Proj(proj))
1254                         continue;
1255                 proj_nr = get_Proj_proj(proj);
1256
1257                 if (proj_nr == pn_Cond_true) {
1258                         assert(projT == NULL && "more than one Proj(true)");
1259                         projT = proj;
1260                 } else {
1261                         assert(proj_nr == pn_Cond_false);
1262                         assert(projF == NULL && "more than one Proj(false)");
1263                         projF = proj;
1264                 }
1265                 mark_irn_visited(proj);
1266         }
1267         assert(projT && projF);
1268
1269         /* create a new high compare */
1270         block    = get_nodes_block(node);
1271         irg      = get_Block_irg(block);
1272         dbg      = get_irn_dbg_info(sel);
1273         relation = get_Cmp_relation(sel);
1274
1275         if (is_equality_cmp_0(sel)) {
1276                 /* x ==/!= 0 ==> or(low,high) ==/!= 0 */
1277                 ir_mode *mode   = env->low_unsigned;
1278                 ir_node *low    = new_r_Conv(block, lentry->low_word, mode);
1279                 ir_node *high   = new_r_Conv(block, lentry->high_word, mode);
1280                 ir_node *ornode = new_rd_Or(dbg, block, low, high, mode);
1281                 ir_node *cmp    = new_rd_Cmp(dbg, block, ornode, new_r_Const_long(irg, mode, 0), relation);
1282                 set_Cond_selector(node, cmp);
1283                 return;
1284         }
1285
1286         if (relation == ir_relation_equal) {
1287                 /* simple case:a == b <==> a_h == b_h && a_l == b_l */
1288                 dst_blk = get_cfop_destination(projF);
1289
1290                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1291                                  ir_relation_equal);
1292                 dbg = get_irn_dbg_info(node);
1293                 irn = new_rd_Cond(dbg, block, irn);
1294
1295                 projHF = new_r_Proj(irn, mode_X, pn_Cond_false);
1296                 mark_irn_visited(projHF);
1297                 exchange(projF, projHF);
1298
1299                 projHT = new_r_Proj(irn, mode_X, pn_Cond_true);
1300                 mark_irn_visited(projHT);
1301
1302                 new_bl = new_r_Block(irg, 1, &projHT);
1303
1304                 dbg = get_irn_dbg_info(sel);
1305                 irn = new_rd_Cmp(dbg, new_bl, lentry->low_word, rentry->low_word,
1306                                   ir_relation_equal);
1307                 dbg = get_irn_dbg_info(node);
1308                 irn = new_rd_Cond(dbg, new_bl, irn);
1309
1310                 proj = new_r_Proj(irn, mode_X, pn_Cond_false);
1311                 mark_irn_visited(proj);
1312                 add_block_cf_input(dst_blk, projHF, proj);
1313
1314                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1315                 mark_irn_visited(proj);
1316                 exchange(projT, proj);
1317         } else if (relation == ir_relation_less_greater) {
1318                 /* simple case:a != b <==> a_h != b_h || a_l != b_l */
1319                 dst_blk = get_cfop_destination(projT);
1320
1321                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1322                                  ir_relation_less_greater);
1323                 dbg = get_irn_dbg_info(node);
1324                 irn = new_rd_Cond(dbg, block, irn);
1325
1326                 projHT = new_r_Proj(irn, mode_X, pn_Cond_true);
1327                 mark_irn_visited(projHT);
1328                 exchange(projT, projHT);
1329
1330                 projHF = new_r_Proj(irn, mode_X, pn_Cond_false);
1331                 mark_irn_visited(projHF);
1332
1333                 new_bl = new_r_Block(irg, 1, &projHF);
1334
1335                 dbg = get_irn_dbg_info(sel);
1336                 irn = new_rd_Cmp(dbg, new_bl, lentry->low_word, rentry->low_word,
1337                                  ir_relation_less_greater);
1338                 dbg = get_irn_dbg_info(node);
1339                 irn = new_rd_Cond(dbg, new_bl, irn);
1340
1341                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1342                 mark_irn_visited(proj);
1343                 add_block_cf_input(dst_blk, projHT, proj);
1344
1345                 proj = new_r_Proj(irn, mode_X, pn_Cond_false);
1346                 mark_irn_visited(proj);
1347                 exchange(projF, proj);
1348         } else {
1349                 /* a rel b <==> a_h REL b_h || (a_h == b_h && a_l rel b_l) */
1350                 ir_node *dstT, *dstF, *newbl_eq, *newbl_l;
1351                 ir_node *projEqF;
1352
1353                 dstT = get_cfop_destination(projT);
1354                 dstF = get_cfop_destination(projF);
1355
1356                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1357                                  relation & ~ir_relation_equal);
1358                 dbg = get_irn_dbg_info(node);
1359                 irn = new_rd_Cond(dbg, block, irn);
1360
1361                 projHT = new_r_Proj(irn, mode_X, pn_Cond_true);
1362                 mark_irn_visited(projHT);
1363
1364                 projHF = new_r_Proj(irn, mode_X, pn_Cond_false);
1365                 mark_irn_visited(projHF);
1366
1367                 newbl_eq = new_r_Block(irg, 1, &projHF);
1368
1369                 irn = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1370                                  ir_relation_equal);
1371                 irn = new_rd_Cond(dbg, newbl_eq, irn);
1372
1373                 projEqF = new_r_Proj(irn, mode_X, pn_Cond_false);
1374                 mark_irn_visited(projEqF);
1375
1376                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1377                 mark_irn_visited(proj);
1378
1379                 newbl_l = new_r_Block(irg, 1, &proj);
1380
1381                 dbg = get_irn_dbg_info(sel);
1382                 irn = new_rd_Cmp(dbg, newbl_l, lentry->low_word, rentry->low_word,
1383                                  relation);
1384                 dbg = get_irn_dbg_info(node);
1385                 irn = new_rd_Cond(dbg, newbl_l, irn);
1386
1387                 proj = new_r_Proj(irn, mode_X, pn_Cond_true);
1388                 mark_irn_visited(proj);
1389                 add_block_cf_input(dstT, projT, proj);
1390
1391                 proj = new_r_Proj(irn, mode_X, pn_Cond_false);
1392                 mark_irn_visited(proj);
1393                 add_block_cf_input(dstF, projF, proj);
1394
1395                 exchange(projT, projHT);
1396                 exchange(projF, projEqF);
1397         }
1398
1399         /* we have changed the control flow */
1400         env->flags |= CF_CHANGED;
1401 }
1402
1403 /**
1404  * Translate a Conv to higher_signed
1405  */
1406 static void lower_Conv_to_Ll(ir_node *node)
1407 {
1408         ir_mode  *omode        = get_irn_mode(node);
1409         ir_node  *op           = get_Conv_op(node);
1410         ir_mode  *imode        = get_irn_mode(op);
1411         ir_graph *irg          = get_irn_irg(node);
1412         ir_node  *block        = get_nodes_block(node);
1413         dbg_info *dbg          = get_irn_dbg_info(node);
1414         ir_node  *res_low;
1415         ir_node  *res_high;
1416
1417         ir_mode  *low_unsigned = env->low_unsigned;
1418         ir_mode  *low_signed
1419                 = mode_is_signed(omode) ? env->low_signed : low_unsigned;
1420
1421         if (mode_is_int(imode) || mode_is_reference(imode)) {
1422                 if (imode == env->high_signed || imode == env->high_unsigned) {
1423                         /* a Conv from Lu to Ls or Ls to Lu */
1424                         const lower64_entry_t *op_entry = get_node_entry(op);
1425                         res_low  = op_entry->low_word;
1426                         res_high = new_rd_Conv(dbg, block, op_entry->high_word, low_signed);
1427                 } else {
1428                         /* simple case: create a high word */
1429                         if (imode != low_unsigned)
1430                                 op = new_rd_Conv(dbg, block, op, low_unsigned);
1431
1432                         res_low = op;
1433
1434                         if (mode_is_signed(imode)) {
1435                                 int      c       = get_mode_size_bits(low_signed) - 1;
1436                                 ir_node *cnst    = new_r_Const_long(irg, low_unsigned, c);
1437                                 if (get_irn_mode(op) != low_signed)
1438                                         op = new_rd_Conv(dbg, block, op, low_signed);
1439                                 res_high = new_rd_Shrs(dbg, block, op, cnst, low_signed);
1440                         } else {
1441                                 res_high = new_r_Const(irg, get_mode_null(low_signed));
1442                         }
1443                 }
1444         } else if (imode == mode_b) {
1445                 res_low  = new_rd_Conv(dbg, block, op, low_unsigned);
1446                 res_high = new_r_Const(irg, get_mode_null(low_signed));
1447         } else {
1448                 ir_node *irn, *call;
1449                 ir_type *mtp = get_conv_type(imode, omode);
1450
1451                 irn = get_intrinsic_address(mtp, get_irn_op(node), imode, omode);
1452                 call = new_rd_Call(dbg, block, get_irg_no_mem(irg), irn, 1, &op, mtp);
1453                 set_irn_pinned(call, get_irn_pinned(node));
1454                 irn = new_r_Proj(call, mode_T, pn_Call_T_result);
1455
1456                 res_low  = new_r_Proj(irn, low_unsigned, 0);
1457                 res_high = new_r_Proj(irn, low_signed, 1);
1458         }
1459         ir_set_dw_lowered(node, res_low, res_high);
1460 }
1461
1462 /**
1463  * Translate a Conv from higher_unsigned
1464  */
1465 static void lower_Conv_from_Ll(ir_node *node)
1466 {
1467         ir_node               *op    = get_Conv_op(node);
1468         ir_mode               *omode = get_irn_mode(node);
1469         ir_node               *block = get_nodes_block(node);
1470         dbg_info              *dbg   = get_irn_dbg_info(node);
1471         ir_graph              *irg   = get_irn_irg(node);
1472         const lower64_entry_t *entry = get_node_entry(op);
1473
1474         if (mode_is_int(omode) || mode_is_reference(omode)) {
1475                 op = entry->low_word;
1476
1477                 /* simple case: create a high word */
1478                 if (omode != env->low_unsigned)
1479                         op = new_rd_Conv(dbg, block, op, omode);
1480
1481                 set_Conv_op(node, op);
1482         } else if (omode == mode_b) {
1483                 /* llu ? true : false  <=> (low|high) ? true : false */
1484                 ir_mode *mode   = env->low_unsigned;
1485                 ir_node *ornode = new_rd_Or(dbg, block, entry->low_word,
1486                                             entry->high_word, mode);
1487                 set_Conv_op(node, ornode);
1488         } else {
1489                 ir_node *irn, *call, *in[2];
1490                 ir_mode *imode = get_irn_mode(op);
1491                 ir_type *mtp   = get_conv_type(imode, omode);
1492                 ir_node *res;
1493
1494                 irn   = get_intrinsic_address(mtp, get_irn_op(node), imode, omode);
1495                 in[0] = entry->low_word;
1496                 in[1] = entry->high_word;
1497
1498                 call = new_rd_Call(dbg, block, get_irg_no_mem(irg), irn, 2, in, mtp);
1499                 set_irn_pinned(call, get_irn_pinned(node));
1500                 irn = new_r_Proj(call, mode_T, pn_Call_T_result);
1501                 res = new_r_Proj(irn, omode, 0);
1502
1503                 exchange(node, res);
1504         }
1505 }
1506
1507 /**
1508  * lower Cmp
1509  */
1510 static void lower_Cmp(ir_node *cmp, ir_mode *m)
1511 {
1512         ir_node  *l    = get_Cmp_left(cmp);
1513         ir_mode  *mode = get_irn_mode(l);
1514         ir_node  *r, *low, *high, *t, *res;
1515         ir_relation relation;
1516         ir_node  *block;
1517         dbg_info *dbg;
1518         const lower64_entry_t *lentry;
1519         const lower64_entry_t *rentry;
1520         (void) m;
1521
1522         if (mode != env->high_signed && mode != env->high_unsigned)
1523                 return;
1524
1525         r        = get_Cmp_right(cmp);
1526         lentry   = get_node_entry(l);
1527         rentry   = get_node_entry(r);
1528         relation = get_Cmp_relation(cmp);
1529         block    = get_nodes_block(cmp);
1530         dbg      = get_irn_dbg_info(cmp);
1531
1532         /* easy case for x ==/!= 0 (see lower_Cond for details) */
1533         if (is_equality_cmp_0(cmp)) {
1534                 ir_graph *irg     = get_irn_irg(cmp);
1535                 ir_mode  *mode    = env->low_unsigned;
1536                 ir_node  *low     = new_r_Conv(block, lentry->low_word, mode);
1537                 ir_node  *high    = new_r_Conv(block, lentry->high_word, mode);
1538                 ir_node  *ornode  = new_rd_Or(dbg, block, low, high, mode);
1539                 ir_node  *new_cmp = new_rd_Cmp(dbg, block, ornode, new_r_Const_long(irg, mode, 0), relation);
1540                 exchange(cmp, new_cmp);
1541                 return;
1542         }
1543
1544         if (relation == ir_relation_equal) {
1545                 /* simple case:a == b <==> a_h == b_h && a_l == b_l */
1546                 low  = new_rd_Cmp(dbg, block, lentry->low_word, rentry->low_word,
1547                                   relation);
1548                 high = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1549                                   relation);
1550                 res  = new_rd_And(dbg, block, low, high, mode_b);
1551         } else if (relation == ir_relation_less_greater) {
1552                 /* simple case:a != b <==> a_h != b_h || a_l != b_l */
1553                 low  = new_rd_Cmp(dbg, block, lentry->low_word, rentry->low_word,
1554                                   relation);
1555                 high = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1556                                   relation);
1557                 res = new_rd_Or(dbg, block, low, high, mode_b);
1558         } else {
1559                 /* a rel b <==> a_h REL b_h || (a_h == b_h && a_l rel b_l) */
1560                 ir_node *high1 = new_rd_Cmp(dbg, block, lentry->high_word,
1561                         rentry->high_word, relation & ~ir_relation_equal);
1562                 low  = new_rd_Cmp(dbg, block, lentry->low_word, rentry->low_word,
1563                                   relation);
1564                 high = new_rd_Cmp(dbg, block, lentry->high_word, rentry->high_word,
1565                                   ir_relation_equal);
1566                 t = new_rd_And(dbg, block, low, high, mode_b);
1567                 res = new_rd_Or(dbg, block, high1, t, mode_b);
1568         }
1569         exchange(cmp, res);
1570 }
1571
1572 /**
1573  * Translate a Conv.
1574  */
1575 static void lower_Conv(ir_node *node, ir_mode *mode)
1576 {
1577         mode = get_irn_mode(node);
1578
1579         if (mode == env->high_signed || mode == env->high_unsigned) {
1580                 lower_Conv_to_Ll(node);
1581         } else {
1582                 ir_mode *op_mode = get_irn_mode(get_Conv_op(node));
1583
1584                 if (op_mode == env->high_signed || op_mode == env->high_unsigned) {
1585                         lower_Conv_from_Ll(node);
1586                 }
1587         }
1588 }
1589
1590 /**
1591  * Remember the new argument index of this value type entity in the lowered
1592  * method type.
1593  *
1594  * @param ent  the entity
1595  * @param pos  the argument index of this entity
1596  */
1597 static inline void set_entity_arg_idx(ir_entity *ent, size_t pos)
1598 {
1599         set_entity_link(ent, INT_TO_PTR(pos));
1600 }
1601
1602 /**
1603  * Retrieve the argument index of a value type entity.
1604  *
1605  * @param ent  the entity
1606  */
1607 static size_t get_entity_arg_idx(const ir_entity *ent) {
1608         return PTR_TO_INT(get_entity_link(ent));
1609 }
1610
1611 /**
1612  * Lower the method type.
1613  *
1614  * @param env  the lower environment
1615  * @param mtp  the method type to lower
1616  *
1617  * @return the lowered type
1618  */
1619 static ir_type *lower_mtp(ir_type *mtp)
1620 {
1621         pmap_entry *entry;
1622         ir_type    *res, *value_type;
1623
1624         entry = pmap_find(lowered_type, mtp);
1625         if (! entry) {
1626                 size_t i, orig_n_params, orig_n_res, n_param, n_res;
1627
1628                 /* count new number of params */
1629                 n_param = orig_n_params = get_method_n_params(mtp);
1630                 for (i = orig_n_params; i > 0;) {
1631                         ir_type *tp = get_method_param_type(mtp, --i);
1632
1633                         if (is_Primitive_type(tp)) {
1634                                 ir_mode *mode = get_type_mode(tp);
1635
1636                                 if (mode == env->high_signed || mode == env->high_unsigned)
1637                                         ++n_param;
1638                         }
1639                 }
1640
1641                 /* count new number of results */
1642                 n_res = orig_n_res = get_method_n_ress(mtp);
1643                 for (i = orig_n_res; i > 0;) {
1644                         ir_type *tp = get_method_res_type(mtp, --i);
1645
1646                         if (is_Primitive_type(tp)) {
1647                                 ir_mode *mode = get_type_mode(tp);
1648
1649                                 if (mode == env->high_signed || mode == env->high_unsigned)
1650                                         ++n_res;
1651                         }
1652                 }
1653
1654                 res = new_type_method(n_param, n_res);
1655
1656                 /* set param types and result types */
1657                 for (i = n_param = 0; i < orig_n_params; ++i) {
1658                         ir_type *tp = get_method_param_type(mtp, i);
1659
1660                         if (is_Primitive_type(tp)) {
1661                                 ir_mode *mode = get_type_mode(tp);
1662
1663                                 if (mode == env->high_signed) {
1664                                         if (env->params->little_endian) {
1665                                                 set_method_param_type(res, n_param++, tp_u);
1666                                                 set_method_param_type(res, n_param++, tp_s);
1667                                         } else {
1668                                                 set_method_param_type(res, n_param++, tp_s);
1669                                                 set_method_param_type(res, n_param++, tp_u);
1670                                         }
1671                                 } else if (mode == env->high_unsigned) {
1672                                         set_method_param_type(res, n_param++, tp_u);
1673                                         set_method_param_type(res, n_param++, tp_u);
1674                                 } else {
1675                                         set_method_param_type(res, n_param++, tp);
1676                                 }
1677                         } else {
1678                                 set_method_param_type(res, n_param++, tp);
1679                         }
1680                 }
1681                 for (i = n_res = 0; i < orig_n_res; ++i) {
1682                         ir_type *tp = get_method_res_type(mtp, i);
1683
1684                         if (is_Primitive_type(tp)) {
1685                                 ir_mode *mode = get_type_mode(tp);
1686
1687                                 if (mode == env->high_signed) {
1688                                         if (env->params->little_endian) {
1689                                                 set_method_res_type(res, n_res++, tp_u);
1690                                                 set_method_res_type(res, n_res++, tp_s);
1691                                         } else {
1692                                                 set_method_res_type(res, n_res++, tp_s);
1693                                                 set_method_res_type(res, n_res++, tp_u);
1694                                         }
1695                                 } else if (mode == env->high_unsigned) {
1696                                         set_method_res_type(res, n_res++, tp_u);
1697                                         set_method_res_type(res, n_res++, tp_u);
1698                                 } else {
1699                                         set_method_res_type(res, n_res++, tp);
1700                                 }
1701                         } else {
1702                                 set_method_res_type(res, n_res++, tp);
1703                         }
1704                 }
1705                 set_lowered_type(mtp, res);
1706                 pmap_insert(lowered_type, mtp, res);
1707
1708                 value_type = get_method_value_param_type(mtp);
1709                 if (value_type != NULL) {
1710                         /* this creates a new value parameter type */
1711                         (void)get_method_value_param_ent(res, 0);
1712
1713                         /* set new param positions for all entities of the value type */
1714                         for (i = n_param = 0; i < orig_n_params; ++i) {
1715                                 ir_type   *tp  = get_method_param_type(mtp, i);
1716                                 ir_entity *ent = get_method_value_param_ent(mtp, i);
1717
1718                                 set_entity_arg_idx(ent, n_param);
1719                                 if (is_Primitive_type(tp)) {
1720                                         ir_mode *mode = get_type_mode(tp);
1721
1722                                         if (mode == env->high_signed
1723                                             || mode == env->high_unsigned) {
1724                                                 n_param += 2;
1725                                                 continue;
1726                                         }
1727                                 }
1728                                 ++n_param;
1729                         }
1730
1731                         set_lowered_type(value_type, get_method_value_param_type(res));
1732                 }
1733         } else {
1734                 res = (ir_type*)entry->value;
1735         }
1736         return res;
1737 }
1738
1739 /**
1740  * Translate a Return.
1741  */
1742 static void lower_Return(ir_node *node, ir_mode *mode)
1743 {
1744         ir_graph  *irg = get_irn_irg(node);
1745         ir_entity *ent = get_irg_entity(irg);
1746         ir_type   *mtp = get_entity_type(ent);
1747         ir_node  **in;
1748         size_t     i, j, n;
1749         int        need_conv = 0;
1750         (void) mode;
1751
1752         /* check if this return must be lowered */
1753         for (i = 0, n = get_Return_n_ress(node); i < n; ++i) {
1754                 ir_node *pred = get_Return_res(node, i);
1755                 ir_mode *mode = get_irn_op_mode(pred);
1756
1757                 if (mode == env->high_signed || mode == env->high_unsigned)
1758                         need_conv = 1;
1759         }
1760         if (! need_conv)
1761                 return;
1762
1763         ent = get_irg_entity(irg);
1764         mtp = get_entity_type(ent);
1765
1766         mtp = lower_mtp(mtp);
1767         set_entity_type(ent, mtp);
1768
1769         /* create a new in array */
1770         NEW_ARR_A(ir_node *, in, get_method_n_ress(mtp) + 1);
1771         in[0] = get_Return_mem(node);
1772
1773         for (j = i = 0, n = get_Return_n_ress(node); i < n; ++i) {
1774                 ir_node *pred      = get_Return_res(node, i);
1775                 ir_mode *pred_mode = get_irn_mode(pred);
1776
1777                 if (pred_mode == env->high_signed || pred_mode == env->high_unsigned) {
1778                         const lower64_entry_t *entry = get_node_entry(pred);
1779                         if (env->params->little_endian) {
1780                                 in[++j] = entry->low_word;
1781                                 in[++j] = entry->high_word;
1782                         } else {
1783                                 in[++j] = entry->high_word;
1784                                 in[++j] = entry->low_word;
1785                         }
1786                 } else {
1787                         in[++j] = pred;
1788                 }
1789         }
1790
1791         set_irn_in(node, j+1, in);
1792 }
1793
1794 /**
1795  * Translate the parameters.
1796  */
1797 static void lower_Start(ir_node *node, ir_mode *mode)
1798 {
1799         ir_graph  *irg = get_irn_irg(node);
1800         ir_entity *ent = get_irg_entity(irg);
1801         ir_type   *tp  = get_entity_type(ent);
1802         ir_node   *args;
1803         long      *new_projs;
1804         size_t    i, j, n_params;
1805         const ir_edge_t *edge;
1806         const ir_edge_t *next;
1807         (void) mode;
1808
1809         if (!mtp_must_be_lowered(tp))
1810                 return;
1811
1812         n_params = get_method_n_params(tp);
1813
1814         NEW_ARR_A(long, new_projs, n_params);
1815
1816         /* Calculate mapping of proj numbers in new_projs */
1817         for (i = j = 0; i < n_params; ++i, ++j) {
1818                 ir_type *ptp = get_method_param_type(tp, i);
1819
1820                 new_projs[i] = j;
1821                 if (is_Primitive_type(ptp)) {
1822                         ir_mode *mode = get_type_mode(ptp);
1823
1824                         if (mode == env->high_signed || mode == env->high_unsigned)
1825                                 ++j;
1826                 }
1827         }
1828
1829         /* lower method type */
1830         tp = lower_mtp(tp);
1831         set_entity_type(ent, tp);
1832
1833         args = NULL;
1834         foreach_out_edge(node, edge) {
1835                 ir_node *proj = get_edge_src_irn(edge);
1836                 if (!is_Proj(proj))
1837                         continue;
1838                 if (get_Proj_proj(proj) == pn_Start_T_args) {
1839                         args = proj;
1840                         break;
1841                 }
1842         }
1843         if (args == NULL)
1844                 return;
1845
1846         /* fix all Proj's and create new ones */
1847         foreach_out_edge_safe(args, edge, next) {
1848                 ir_node *proj   = get_edge_src_irn(edge);
1849                 ir_mode *mode   = get_irn_mode(proj);
1850                 ir_mode *mode_l = env->low_unsigned;
1851                 ir_node *pred;
1852                 long     proj_nr;
1853                 ir_mode *mode_h;
1854                 ir_node *res_low;
1855                 ir_node *res_high;
1856                 dbg_info *dbg;
1857
1858                 if (!is_Proj(proj))
1859                         continue;
1860                 pred    = get_Proj_pred(proj);
1861                 proj_nr = get_Proj_proj(proj);
1862
1863                 if (mode == env->high_signed) {
1864                         mode_h = env->low_signed;
1865                 } else if (mode == env->high_unsigned) {
1866                         mode_h = env->low_unsigned;
1867                 } else {
1868                         long new_pn = new_projs[proj_nr];
1869                         set_Proj_proj(proj, new_pn);
1870                         continue;
1871                 }
1872
1873                 dbg = get_irn_dbg_info(proj);
1874                 if (env->params->little_endian) {
1875                         res_low  = new_rd_Proj(dbg, pred, mode_l, new_projs[proj_nr]);
1876                         res_high = new_rd_Proj(dbg, pred, mode_h, new_projs[proj_nr] + 1);
1877                 } else {
1878                         res_high = new_rd_Proj(dbg, pred, mode_h, new_projs[proj_nr]);
1879                         res_low  = new_rd_Proj(dbg, pred, mode_l, new_projs[proj_nr] + 1);
1880                 }
1881                 ir_set_dw_lowered(proj, res_low, res_high);
1882         }
1883 }
1884
1885 /**
1886  * Translate a Call.
1887  */
1888 static void lower_Call(ir_node *node, ir_mode *mode)
1889 {
1890         ir_type  *tp = get_Call_type(node);
1891         ir_node  **in;
1892         size_t   n_params, n_res;
1893         bool     need_lower = false;
1894         size_t   i, j;
1895         size_t   p;
1896         long     *res_numbers = NULL;
1897         ir_node  *resproj;
1898         const ir_edge_t *edge;
1899         const ir_edge_t *next;
1900         (void) mode;
1901
1902         n_params = get_method_n_params(tp);
1903         for (p = 0; p < n_params; ++p) {
1904                 ir_type *ptp = get_method_param_type(tp, p);
1905
1906                 if (is_Primitive_type(ptp)) {
1907                         ir_mode *mode = get_type_mode(ptp);
1908
1909                         if (mode == env->high_signed || mode == env->high_unsigned) {
1910                                 need_lower = true;
1911                                 break;
1912                         }
1913                 }
1914         }
1915         n_res = get_method_n_ress(tp);
1916         if (n_res > 0) {
1917                 NEW_ARR_A(long, res_numbers, n_res);
1918
1919                 for (i = j = 0; i < n_res; ++i, ++j) {
1920                         ir_type *ptp = get_method_res_type(tp, i);
1921
1922                         res_numbers[i] = j;
1923                         if (is_Primitive_type(ptp)) {
1924                                 ir_mode *mode = get_type_mode(ptp);
1925
1926                                 if (mode == env->high_signed || mode == env->high_unsigned) {
1927                                         need_lower = true;
1928                                         ++j;
1929                                 }
1930                         }
1931                 }
1932         }
1933
1934         if (! need_lower)
1935                 return;
1936
1937         /* let's lower it */
1938         tp = lower_mtp(tp);
1939         set_Call_type(node, tp);
1940
1941         NEW_ARR_A(ir_node *, in, get_method_n_params(tp) + 2);
1942
1943         in[0] = get_Call_mem(node);
1944         in[1] = get_Call_ptr(node);
1945
1946         for (j = 2, i = 0; i < n_params; ++i) {
1947                 ir_node *pred      = get_Call_param(node, i);
1948                 ir_mode *pred_mode = get_irn_mode(pred);
1949
1950                 if (pred_mode == env->high_signed || pred_mode == env->high_unsigned) {
1951                         const lower64_entry_t *pred_entry = get_node_entry(pred);
1952                         if (env->params->little_endian) {
1953                                 in[j++] = pred_entry->low_word;
1954                                 in[j++] = pred_entry->high_word;
1955                         } else {
1956                                 in[j++] = pred_entry->high_word;
1957                                 in[j++] = pred_entry->low_word;
1958                         }
1959                 } else {
1960                         in[j++] = pred;
1961                 }
1962         }
1963
1964         set_irn_in(node, j, in);
1965
1966         /* find results T */
1967         resproj = NULL;
1968         foreach_out_edge(node, edge) {
1969                 ir_node *proj = get_edge_src_irn(edge);
1970                 if (!is_Proj(proj))
1971                         continue;
1972                 if (get_Proj_proj(proj) == pn_Call_T_result) {
1973                         resproj = proj;
1974                         break;
1975                 }
1976         }
1977         if (resproj == NULL)
1978                 return;
1979
1980         /* fix the results */
1981         foreach_out_edge_safe(resproj, edge, next) {
1982                 ir_node  *proj      = get_edge_src_irn(edge);
1983                 ir_mode  *proj_mode = get_irn_mode(proj);
1984                 ir_mode  *mode_l    = env->low_unsigned;
1985                 ir_node  *pred;
1986                 long      proj_nr;
1987                 ir_mode  *mode_h;
1988                 ir_node  *res_low;
1989                 ir_node  *res_high;
1990                 dbg_info *dbg;
1991
1992                 if (!is_Proj(proj))
1993                         continue;
1994                 pred    = get_Proj_pred(proj);
1995                 proj_nr = get_Proj_proj(proj);
1996
1997                 if (proj_mode == env->high_signed) {
1998                         mode_h = env->low_signed;
1999                 } else if (proj_mode == env->high_unsigned) {
2000                         mode_h = env->low_unsigned;
2001                 } else {
2002                         long new_nr = res_numbers[proj_nr];
2003                         set_Proj_proj(proj, new_nr);
2004                         continue;
2005                 }
2006
2007                 dbg = get_irn_dbg_info(proj);
2008                 if (env->params->little_endian) {
2009                         res_low  = new_rd_Proj(dbg, pred, mode_l, res_numbers[proj_nr]);
2010                         res_high = new_rd_Proj(dbg, pred, mode_h, res_numbers[proj_nr] + 1);
2011                 } else {
2012                         res_high = new_rd_Proj(dbg, pred, mode_h, res_numbers[proj_nr]);
2013                         res_low  = new_rd_Proj(dbg, pred, mode_l, res_numbers[proj_nr] + 1);
2014                 }
2015                 ir_set_dw_lowered(proj, res_low, res_high);
2016         }
2017 }
2018
2019 /**
2020  * Translate an Unknown into two.
2021  */
2022 static void lower_Unknown(ir_node *node, ir_mode *mode)
2023 {
2024         ir_mode  *low_mode = env->low_unsigned;
2025         ir_graph *irg      = get_irn_irg(node);
2026         ir_node  *res_low  = new_r_Unknown(irg, low_mode);
2027         ir_node  *res_high = new_r_Unknown(irg, mode);
2028         ir_set_dw_lowered(node, res_low, res_high);
2029 }
2030
2031 /**
2032  * Translate a Bad into two.
2033  */
2034 static void lower_Bad(ir_node *node, ir_mode *mode)
2035 {
2036         ir_mode  *low_mode = env->low_unsigned;
2037         ir_graph *irg      = get_irn_irg(node);
2038         ir_node  *res_low  = new_r_Bad(irg, low_mode);
2039         ir_node  *res_high = new_r_Bad(irg, mode);
2040         ir_set_dw_lowered(node, res_low, res_high);
2041 }
2042
2043 /**
2044  * Translate a Phi.
2045  *
2046  * First step: just create two templates
2047  */
2048 static void lower_Phi(ir_node *phi)
2049 {
2050         ir_mode  *mode = get_irn_mode(phi);
2051         int       i;
2052         int       arity;
2053         ir_node **in_l;
2054         ir_node **in_h;
2055         ir_node  *unk_l;
2056         ir_node  *unk_h;
2057         ir_node  *phi_l;
2058         ir_node  *phi_h;
2059         dbg_info *dbg;
2060         ir_node  *block;
2061         ir_graph *irg;
2062         ir_mode  *mode_l;
2063         ir_mode  *mode_h;
2064
2065         /* enqueue predecessors */
2066         arity = get_Phi_n_preds(phi);
2067         for (i = 0; i < arity; ++i) {
2068                 ir_node *pred = get_Phi_pred(phi, i);
2069                 pdeq_putr(env->waitq, pred);
2070         }
2071
2072         if (mode != env->high_signed && mode != env->high_unsigned)
2073                 return;
2074
2075         /* first create a new in array */
2076         NEW_ARR_A(ir_node *, in_l, arity);
2077         NEW_ARR_A(ir_node *, in_h, arity);
2078         irg    = get_irn_irg(phi);
2079         mode_l = env->low_unsigned;
2080         mode_h = mode == env->high_signed ? env->low_signed : env->low_unsigned;
2081         unk_l  = new_r_Dummy(irg, mode_l);
2082         unk_h  = new_r_Dummy(irg, mode_h);
2083         for (i = 0; i < arity; ++i) {
2084                 in_l[i] = unk_l;
2085                 in_h[i] = unk_h;
2086         }
2087
2088         dbg   = get_irn_dbg_info(phi);
2089         block = get_nodes_block(phi);
2090         phi_l = new_rd_Phi(dbg, block, arity, in_l, mode_l);
2091         phi_h = new_rd_Phi(dbg, block, arity, in_h, mode_h);
2092
2093         ir_set_dw_lowered(phi, phi_l, phi_h);
2094
2095         /* remember that we need to fixup the predecessors later */
2096         ARR_APP1(ir_node*, env->lowered_phis, phi);
2097 }
2098
2099 static void fixup_phi(ir_node *phi)
2100 {
2101         const lower64_entry_t *entry = get_node_entry(phi);
2102         ir_node               *phi_l = entry->low_word;
2103         ir_node               *phi_h = entry->high_word;
2104         int                    arity = get_Phi_n_preds(phi);
2105         int                    i;
2106
2107         /* exchange phi predecessors which are lowered by now */
2108         for (i = 0; i < arity; ++i) {
2109                 ir_node               *pred       = get_Phi_pred(phi, i);
2110                 const lower64_entry_t *pred_entry = get_node_entry(pred);
2111
2112                 set_Phi_pred(phi_l, i, pred_entry->low_word);
2113                 set_Phi_pred(phi_h, i, pred_entry->high_word);
2114         }
2115 }
2116
2117 /**
2118  * Translate a Mux.
2119  */
2120 static void lower_Mux(ir_node *mux, ir_mode *mode)
2121 {
2122         ir_node               *truen       = get_Mux_true(mux);
2123         ir_node               *falsen      = get_Mux_false(mux);
2124         ir_node               *sel         = get_Mux_sel(mux);
2125         const lower64_entry_t *true_entry  = get_node_entry(truen);
2126         const lower64_entry_t *false_entry = get_node_entry(falsen);
2127         ir_node               *true_l      = true_entry->low_word;
2128         ir_node               *true_h      = true_entry->high_word;
2129         ir_node               *false_l     = false_entry->low_word;
2130         ir_node               *false_h     = false_entry->high_word;
2131         dbg_info              *dbgi        = get_irn_dbg_info(mux);
2132         ir_node               *block       = get_nodes_block(mux);
2133         ir_node               *res_low
2134                 = new_rd_Mux(dbgi, block, sel, false_l, true_l, env->low_unsigned);
2135         ir_node               *res_high
2136                 = new_rd_Mux(dbgi, block, sel, false_h, true_h, mode);
2137         ir_set_dw_lowered(mux, res_low, res_high);
2138 }
2139
2140 /**
2141  * Translate an ASM node.
2142  */
2143 static void lower_ASM(ir_node *asmn, ir_mode *mode)
2144 {
2145         ir_mode           *high_signed        = env->high_signed;
2146         ir_mode           *high_unsigned      = env->high_unsigned;
2147         int                n_outs             = get_ASM_n_output_constraints(asmn);
2148         ir_asm_constraint *output_constraints = get_ASM_output_constraints(asmn);
2149         ir_asm_constraint *input_constraints  = get_ASM_input_constraints(asmn);
2150         unsigned           n_64bit_outs       = 0;
2151         int      i;
2152
2153         (void)mode;
2154
2155         for (i = get_irn_arity(asmn) - 1; i >= 0; --i) {
2156                 ir_node *op      = get_irn_n(asmn, i);
2157                 ir_mode *op_mode = get_irn_mode(op);
2158                 if (op_mode == high_signed || op_mode == high_unsigned) {
2159                         panic("lowering ASM 64bit input unimplemented");
2160                 }
2161         }
2162
2163         for (i = 0; i < n_outs; ++i) {
2164                 const ir_asm_constraint *constraint = &output_constraints[i];
2165                 if (constraint->mode == high_signed || constraint->mode == high_unsigned) {
2166                         const char *constr = get_id_str(constraint->constraint);
2167                         ++n_64bit_outs;
2168                         /* TODO: How to do this architecture neutral? This is very
2169                          * i386 specific... */
2170                         if (constr[0] != '=' || constr[1] != 'A') {
2171                                 panic("lowering ASM 64bit output only supports '=A' currently");
2172                         }
2173                 }
2174         }
2175
2176         if (n_64bit_outs == 0)
2177                 return;
2178
2179         {
2180                 dbg_info          *dbgi       = get_irn_dbg_info(asmn);
2181                 ir_node           *block      = get_nodes_block(asmn);
2182                 int                arity      = get_irn_arity(asmn);
2183                 ir_node          **in         = get_irn_in(asmn) + 1;
2184                 int                n_outs     = get_ASM_n_output_constraints(asmn);
2185                 int                new_n_outs = 0;
2186                 int                n_clobber  = get_ASM_n_clobbers(asmn);
2187                 long              *proj_map   = ALLOCAN(long, n_outs);
2188                 ident            **clobbers   = get_ASM_clobbers(asmn);
2189                 ident             *asm_text   = get_ASM_text(asmn);
2190                 ir_asm_constraint *new_outputs
2191                         = ALLOCAN(ir_asm_constraint, n_outs+n_64bit_outs);
2192                 ir_node           *new_asm;
2193                 const ir_edge_t   *edge;
2194                 const ir_edge_t   *next;
2195
2196                 for (i = 0; i < n_outs; ++i) {
2197                         const ir_asm_constraint *constraint = &output_constraints[i];
2198                         if (constraint->mode == high_signed || constraint->mode == high_unsigned) {
2199                                 new_outputs[new_n_outs].pos        = constraint->pos;
2200                                 new_outputs[new_n_outs].constraint = new_id_from_str("=a");
2201                                 new_outputs[new_n_outs].mode       = env->low_unsigned;
2202                                 proj_map[i] = new_n_outs;
2203                                 ++new_n_outs;
2204                                 new_outputs[new_n_outs].pos        = constraint->pos;
2205                                 new_outputs[new_n_outs].constraint = new_id_from_str("=d");
2206                                 if (constraint->mode == high_signed)
2207                                         new_outputs[new_n_outs].mode = env->low_signed;
2208                                 else
2209                                         new_outputs[new_n_outs].mode = env->low_unsigned;
2210                                 ++new_n_outs;
2211                         } else {
2212                                 new_outputs[new_n_outs] = *constraint;
2213                                 proj_map[i] = new_n_outs;
2214                                 ++new_n_outs;
2215                         }
2216                 }
2217                 assert(new_n_outs == n_outs+(int)n_64bit_outs);
2218
2219                 new_asm = new_rd_ASM(dbgi, block, arity, in, input_constraints,
2220                                      new_n_outs, new_outputs, n_clobber, clobbers,
2221                                      asm_text);
2222
2223                 foreach_out_edge_safe(asmn, edge, next) {
2224                         ir_node *proj      = get_edge_src_irn(edge);
2225                         ir_mode *proj_mode = get_irn_mode(proj);
2226                         long     pn;
2227
2228                         if (!is_Proj(proj))
2229                                 continue;
2230                         pn = get_Proj_proj(proj);
2231
2232                         if (pn < n_outs)
2233                                 pn = proj_map[pn];
2234                         else
2235                                 pn = new_n_outs + pn - n_outs;
2236
2237                         if (proj_mode == high_signed || proj_mode == high_unsigned) {
2238                                 ir_mode *high_mode
2239                                         = proj_mode == high_signed ? env->low_signed : env->low_unsigned;
2240                                 ir_node *np_low  = new_r_Proj(new_asm, env->low_unsigned, pn);
2241                                 ir_node *np_high = new_r_Proj(new_asm, high_mode, pn+1);
2242                                 ir_set_dw_lowered(proj, np_low, np_high);
2243                         } else {
2244                                 ir_node *np = new_r_Proj(new_asm, proj_mode, pn);
2245                                 exchange(proj, np);
2246                         }
2247                 }
2248         }
2249 }
2250
2251 /**
2252  * Translate a Sel node.
2253  */
2254 static void lower_Sel(ir_node *sel, ir_mode *mode)
2255 {
2256         (void) mode;
2257
2258         /* we must only lower value parameter Sels if we change the
2259            value parameter type. */
2260         if (env->value_param_tp != NULL) {
2261                 ir_entity *ent = get_Sel_entity(sel);
2262             if (get_entity_owner(ent) == env->value_param_tp) {
2263                         size_t pos = get_entity_arg_idx(ent);
2264
2265                         ent = get_method_value_param_ent(env->l_mtp, pos);
2266                         set_Sel_entity(sel, ent);
2267                 }
2268         }
2269 }
2270
2271 /**
2272  * check for opcodes that must always be lowered.
2273  */
2274 static bool always_lower(unsigned code)
2275 {
2276         switch (code) {
2277         case iro_ASM:
2278         case iro_Proj:
2279         case iro_Start:
2280         case iro_Call:
2281         case iro_Return:
2282         case iro_Cond:
2283         case iro_Conv:
2284         case iro_Sel:
2285                 return true;
2286         default:
2287                 return false;
2288         }
2289 }
2290
2291 /**
2292  * Compare two op_mode_entry_t's.
2293  */
2294 static int cmp_op_mode(const void *elt, const void *key, size_t size)
2295 {
2296         const op_mode_entry_t *e1 = (const op_mode_entry_t*)elt;
2297         const op_mode_entry_t *e2 = (const op_mode_entry_t*)key;
2298         (void) size;
2299
2300         return (e1->op != e2->op) | (e1->imode != e2->imode) | (e1->omode != e2->omode);
2301 }
2302
2303 /**
2304  * Compare two conv_tp_entry_t's.
2305  */
2306 static int cmp_conv_tp(const void *elt, const void *key, size_t size)
2307 {
2308         const conv_tp_entry_t *e1 = (const conv_tp_entry_t*)elt;
2309         const conv_tp_entry_t *e2 = (const conv_tp_entry_t*)key;
2310         (void) size;
2311
2312         return (e1->imode != e2->imode) | (e1->omode != e2->omode);
2313 }
2314
2315 /**
2316  * Enter a lowering function into an ir_op.
2317  */
2318 void ir_register_dw_lower_function(ir_op *op, lower_dw_func func)
2319 {
2320         op->ops.generic = (op_func)func;
2321 }
2322
2323 /**
2324  * Returns non-zero if a method type must be lowered.
2325  *
2326  * @param mtp  the method type
2327  */
2328 static bool mtp_must_be_lowered(ir_type *mtp)
2329 {
2330         size_t i, n_params = get_method_n_params(mtp);
2331
2332         /* first check if we have parameters that must be fixed */
2333         for (i = 0; i < n_params; ++i) {
2334                 ir_type *tp = get_method_param_type(mtp, i);
2335
2336                 if (is_Primitive_type(tp)) {
2337                         ir_mode *mode = get_type_mode(tp);
2338
2339                         if (mode == env->high_signed || mode == env->high_unsigned)
2340                                 return true;
2341                 }
2342         }
2343         return false;
2344 }
2345
2346 /* Determine which modes need to be lowered */
2347 static void setup_modes(void)
2348 {
2349         unsigned           size_bits           = env->params->doubleword_size;
2350         ir_mode           *doubleword_signed   = NULL;
2351         ir_mode           *doubleword_unsigned = NULL;
2352         size_t             n_modes             = get_irp_n_modes();
2353         ir_mode_arithmetic arithmetic;
2354         unsigned           modulo_shift;
2355         size_t             i;
2356
2357         /* search for doubleword modes... */
2358         for (i = 0; i < n_modes; ++i) {
2359                 ir_mode *mode = get_irp_mode(i);
2360                 if (!mode_is_int(mode))
2361                         continue;
2362                 if (get_mode_size_bits(mode) != size_bits)
2363                         continue;
2364                 if (mode_is_signed(mode)) {
2365                         if (doubleword_signed != NULL) {
2366                                 /* sigh - the lowerer should really just lower all mode with
2367                                  * size_bits it finds. Unfortunately this required a bigger
2368                                  * rewrite. */
2369                                 panic("multiple double word signed modes found");
2370                         }
2371                         doubleword_signed = mode;
2372                 } else {
2373                         if (doubleword_unsigned != NULL) {
2374                                 /* sigh - the lowerer should really just lower all mode with
2375                                  * size_bits it finds. Unfortunately this required a bigger
2376                                  * rewrite. */
2377                                 panic("multiple double word unsigned modes found");
2378                         }
2379                         doubleword_unsigned = mode;
2380                 }
2381         }
2382         if (doubleword_signed == NULL || doubleword_unsigned == NULL) {
2383                 panic("Couldn't find doubleword modes");
2384         }
2385
2386         arithmetic   = get_mode_arithmetic(doubleword_signed);
2387         modulo_shift = get_mode_modulo_shift(doubleword_signed);
2388
2389         assert(get_mode_size_bits(doubleword_unsigned) == size_bits);
2390         assert(size_bits % 2 == 0);
2391         assert(get_mode_sign(doubleword_signed) == 1);
2392         assert(get_mode_sign(doubleword_unsigned) == 0);
2393         assert(get_mode_sort(doubleword_signed) == irms_int_number);
2394         assert(get_mode_sort(doubleword_unsigned) == irms_int_number);
2395         assert(get_mode_arithmetic(doubleword_unsigned) == arithmetic);
2396         assert(get_mode_modulo_shift(doubleword_unsigned) == modulo_shift);
2397
2398         /* try to guess a sensible modulo shift for the new mode.
2399          * (This is IMO another indication that this should really be a node
2400          *  attribute instead of a mode thing) */
2401         if (modulo_shift == size_bits) {
2402                 modulo_shift = modulo_shift / 2;
2403         } else if (modulo_shift == 0) {
2404                 /* fine */
2405         } else {
2406                 panic("Don't know what new modulo shift to use for lowered doubleword mode");
2407         }
2408         size_bits /= 2;
2409
2410         /* produce lowered modes */
2411         env->high_signed   = doubleword_signed;
2412         env->high_unsigned = doubleword_unsigned;
2413         env->low_signed    = new_ir_mode("WS", irms_int_number, size_bits, 1,
2414                                          arithmetic, modulo_shift);
2415         env->low_unsigned  = new_ir_mode("WU", irms_int_number, size_bits, 0,
2416                                          arithmetic, modulo_shift);
2417 }
2418
2419 static void enqueue_preds(ir_node *node)
2420 {
2421         int arity = get_irn_arity(node);
2422         int i;
2423
2424         for (i = 0; i < arity; ++i) {
2425                 ir_node *pred = get_irn_n(node, i);
2426                 pdeq_putr(env->waitq, pred);
2427         }
2428 }
2429
2430 static void lower_node(ir_node *node)
2431 {
2432         int              arity;
2433         int              i;
2434         lower_dw_func    func;
2435         ir_op           *op;
2436         ir_mode         *mode;
2437         unsigned         idx;
2438         lower64_entry_t *entry;
2439
2440         if (irn_visited_else_mark(node))
2441                 return;
2442
2443         /* cycles are always broken at Phi and Block nodes. So we don't need special
2444          * magic in all the other lower functions */
2445         if (is_Block(node)) {
2446                 enqueue_preds(node);
2447                 return;
2448         } else if (is_Phi(node)) {
2449                 lower_Phi(node);
2450                 return;
2451         }
2452
2453         /* depth-first: descend into operands */
2454         if (!is_Block(node)) {
2455                 ir_node *block = get_nodes_block(node);
2456                 lower_node(block);
2457         }
2458
2459         if (!is_Cond(node)) {
2460                 arity = get_irn_arity(node);
2461                 for (i = 0; i < arity; ++i) {
2462                         ir_node *pred = get_irn_n(node, i);
2463                         lower_node(pred);
2464                 }
2465         }
2466
2467         op   = get_irn_op(node);
2468         func = (lower_dw_func) op->ops.generic;
2469         if (func == NULL)
2470                 return;
2471
2472         idx   = get_irn_idx(node);
2473         entry = idx < env->n_entries ? env->entries[idx] : NULL;
2474         if (entry != NULL || always_lower(get_irn_opcode(node))) {
2475                 mode = get_irn_op_mode(node);
2476                 if (mode == env->high_signed) {
2477                         mode = env->low_signed;
2478                 } else {
2479                         mode = env->low_unsigned;
2480                 }
2481                 DB((dbg, LEVEL_1, "  %+F\n", node));
2482                 func(node, mode);
2483         }
2484 }
2485
2486 static void clear_node_and_phi_links(ir_node *node, void *data)
2487 {
2488         (void) data;
2489         if (get_irn_mode(node) == mode_T) {
2490                 set_irn_link(node, node);
2491         } else {
2492                 set_irn_link(node, NULL);
2493         }
2494         if (is_Block(node))
2495                 set_Block_phis(node, NULL);
2496         else if (is_Phi(node))
2497                 set_Phi_next(node, NULL);
2498 }
2499
2500 static void lower_irg(ir_graph *irg)
2501 {
2502         ir_entity *ent;
2503         ir_type   *mtp;
2504         unsigned   n_idx;
2505
2506         obstack_init(&env->obst);
2507
2508         /* just here for debugging */
2509         current_ir_graph = irg;
2510         edges_assure(irg);
2511
2512         n_idx = get_irg_last_idx(irg);
2513         n_idx = n_idx + (n_idx >> 2);  /* add 25% */
2514         env->n_entries = n_idx;
2515         env->entries   = NEW_ARR_F(lower64_entry_t*, n_idx);
2516         memset(env->entries, 0, sizeof(env->entries[0]) * n_idx);
2517
2518         env->irg            = irg;
2519         env->l_mtp          = NULL;
2520         env->flags          = 0;
2521         env->value_param_tp = NULL;
2522
2523         ent = get_irg_entity(irg);
2524         mtp = get_entity_type(ent);
2525
2526         if (mtp_must_be_lowered(mtp)) {
2527                 ir_type *ltp = lower_mtp(mtp);
2528                 /* Do not update the entity type yet, this will be done by lower_Start! */
2529                 env->flags |= MUST_BE_LOWERED;
2530                 env->l_mtp = ltp;
2531                 env->value_param_tp = get_method_value_param_type(mtp);
2532         }
2533
2534         /* first step: link all nodes and allocate data */
2535         ir_reserve_resources(irg, IR_RESOURCE_PHI_LIST | IR_RESOURCE_IRN_LINK);
2536         visit_all_identities(irg, clear_node_and_phi_links, NULL);
2537         irg_walk_graph(irg, NULL, prepare_links_and_handle_rotl, env);
2538
2539         if (env->flags & MUST_BE_LOWERED) {
2540                 size_t i;
2541                 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
2542                 inc_irg_visited(irg);
2543
2544                 assert(pdeq_empty(env->waitq));
2545                 pdeq_putr(env->waitq, get_irg_end(irg));
2546
2547                 env->lowered_phis = NEW_ARR_F(ir_node*, 0);
2548                 while (!pdeq_empty(env->waitq)) {
2549                         ir_node *node = (ir_node*)pdeq_getl(env->waitq);
2550                         lower_node(node);
2551                 }
2552
2553                 /* we need to fixup phis */
2554                 for (i = 0; i < ARR_LEN(env->lowered_phis); ++i) {
2555                         ir_node *phi = env->lowered_phis[i];
2556                         fixup_phi(phi);
2557                 }
2558                 DEL_ARR_F(env->lowered_phis);
2559
2560
2561                 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
2562
2563                 if (env->flags & CF_CHANGED) {
2564                         /* control flow changed, dominance info is invalid */
2565                         set_irg_doms_inconsistent(irg);
2566                         set_irg_extblk_inconsistent(irg);
2567                 }
2568         }
2569
2570         ir_free_resources(irg, IR_RESOURCE_PHI_LIST | IR_RESOURCE_IRN_LINK);
2571
2572         DEL_ARR_F(env->entries);
2573         obstack_free(&env->obst, NULL);
2574 }
2575
2576 static const lwrdw_param_t *param;
2577
2578 void ir_prepare_dw_lowering(const lwrdw_param_t *new_param)
2579 {
2580         assert(new_param != NULL);
2581         FIRM_DBG_REGISTER(dbg, "firm.lower.dw");
2582
2583         param = new_param;
2584
2585         clear_irp_opcodes_generic_func();
2586         ir_register_dw_lower_function(op_ASM,     lower_ASM);
2587         ir_register_dw_lower_function(op_Add,     lower_binop);
2588         ir_register_dw_lower_function(op_And,     lower_And);
2589         ir_register_dw_lower_function(op_Bad,     lower_Bad);
2590         ir_register_dw_lower_function(op_Call,    lower_Call);
2591         ir_register_dw_lower_function(op_Cmp,     lower_Cmp);
2592         ir_register_dw_lower_function(op_Cond,    lower_Cond);
2593         ir_register_dw_lower_function(op_Const,   lower_Const);
2594         ir_register_dw_lower_function(op_Conv,    lower_Conv);
2595         ir_register_dw_lower_function(op_Div,     lower_Div);
2596         ir_register_dw_lower_function(op_Eor,     lower_Eor);
2597         ir_register_dw_lower_function(op_Load,    lower_Load);
2598         ir_register_dw_lower_function(op_Minus,   lower_unop);
2599         ir_register_dw_lower_function(op_Mod,     lower_Mod);
2600         ir_register_dw_lower_function(op_Mul,     lower_binop);
2601         ir_register_dw_lower_function(op_Mux,     lower_Mux);
2602         ir_register_dw_lower_function(op_Not,     lower_Not);
2603         ir_register_dw_lower_function(op_Or,      lower_Or);
2604         ir_register_dw_lower_function(op_Return,  lower_Return);
2605         ir_register_dw_lower_function(op_Sel,     lower_Sel);
2606         ir_register_dw_lower_function(op_Shl,     lower_Shl);
2607         ir_register_dw_lower_function(op_Shr,     lower_Shr);
2608         ir_register_dw_lower_function(op_Shrs,    lower_Shrs);
2609         ir_register_dw_lower_function(op_Start,   lower_Start);
2610         ir_register_dw_lower_function(op_Store,   lower_Store);
2611         ir_register_dw_lower_function(op_Sub,     lower_binop);
2612         ir_register_dw_lower_function(op_Unknown, lower_Unknown);
2613 }
2614
2615 /*
2616  * Do the lowering.
2617  */
2618 void ir_lower_dw_ops(void)
2619 {
2620         lower_dw_env_t lenv;
2621         size_t      i, n;
2622
2623         memset(&lenv, 0, sizeof(lenv));
2624         lenv.params = param;
2625         env = &lenv;
2626
2627         setup_modes();
2628
2629         /* create the necessary maps */
2630         if (! intrinsic_fkt)
2631                 intrinsic_fkt = new_set(cmp_op_mode, iro_Last + 1);
2632         if (! conv_types)
2633                 conv_types = new_set(cmp_conv_tp, 16);
2634         if (! lowered_type)
2635                 lowered_type = pmap_create();
2636
2637         /* create a primitive unsigned and signed type */
2638         if (! tp_u)
2639                 tp_u = get_type_for_mode(lenv.low_unsigned);
2640         if (! tp_s)
2641                 tp_s = get_type_for_mode(lenv.low_signed);
2642
2643         /* create method types for the created binop calls */
2644         if (! binop_tp_u) {
2645                 binop_tp_u = new_type_method(4, 2);
2646                 set_method_param_type(binop_tp_u, 0, tp_u);
2647                 set_method_param_type(binop_tp_u, 1, tp_u);
2648                 set_method_param_type(binop_tp_u, 2, tp_u);
2649                 set_method_param_type(binop_tp_u, 3, tp_u);
2650                 set_method_res_type(binop_tp_u, 0, tp_u);
2651                 set_method_res_type(binop_tp_u, 1, tp_u);
2652         }
2653         if (! binop_tp_s) {
2654                 binop_tp_s = new_type_method(4, 2);
2655                 if (env->params->little_endian) {
2656                         set_method_param_type(binop_tp_s, 0, tp_u);
2657                         set_method_param_type(binop_tp_s, 1, tp_s);
2658                         set_method_param_type(binop_tp_s, 2, tp_u);
2659                         set_method_param_type(binop_tp_s, 3, tp_s);
2660                         set_method_res_type(binop_tp_s, 0, tp_u);
2661                         set_method_res_type(binop_tp_s, 1, tp_s);
2662                 } else {
2663                         set_method_param_type(binop_tp_s, 0, tp_s);
2664                         set_method_param_type(binop_tp_s, 1, tp_u);
2665                         set_method_param_type(binop_tp_s, 2, tp_s);
2666                         set_method_param_type(binop_tp_s, 3, tp_u);
2667                         set_method_res_type(binop_tp_s, 0, tp_s);
2668                         set_method_res_type(binop_tp_s, 1, tp_u);
2669                 }
2670         }
2671         if (! unop_tp_u) {
2672                 unop_tp_u = new_type_method(2, 2);
2673                 set_method_param_type(unop_tp_u, 0, tp_u);
2674                 set_method_param_type(unop_tp_u, 1, tp_u);
2675                 set_method_res_type(unop_tp_u, 0, tp_u);
2676                 set_method_res_type(unop_tp_u, 1, tp_u);
2677         }
2678         if (! unop_tp_s) {
2679                 unop_tp_s = new_type_method(2, 2);
2680                 if (env->params->little_endian) {
2681                         set_method_param_type(unop_tp_s, 0, tp_u);
2682                         set_method_param_type(unop_tp_s, 1, tp_s);
2683                         set_method_res_type(unop_tp_s, 0, tp_u);
2684                         set_method_res_type(unop_tp_s, 1, tp_s);
2685                 } else {
2686                         set_method_param_type(unop_tp_s, 0, tp_s);
2687                         set_method_param_type(unop_tp_s, 1, tp_u);
2688                         set_method_res_type(unop_tp_s, 0, tp_s);
2689                         set_method_res_type(unop_tp_s, 1, tp_u);
2690                 }
2691         }
2692
2693         lenv.tv_mode_bytes = new_tarval_from_long(param->doubleword_size/(2*8), lenv.low_unsigned);
2694         lenv.tv_mode_bits  = new_tarval_from_long(param->doubleword_size/2, lenv.low_unsigned);
2695         lenv.waitq         = new_pdeq();
2696         lenv.first_id      = new_id_from_chars(param->little_endian ? ".l" : ".h", 2);
2697         lenv.next_id       = new_id_from_chars(param->little_endian ? ".h" : ".l", 2);
2698
2699         /* transform all graphs */
2700         for (i = 0, n = get_irp_n_irgs(); i < n; ++i) {
2701                 ir_graph *irg = get_irp_irg(i);
2702                 lower_irg(irg);
2703         }
2704         del_pdeq(lenv.waitq);
2705
2706         env = NULL;
2707 }
2708
2709 /* Default implementation. */
2710 ir_entity *def_create_intrinsic_fkt(ir_type *method, const ir_op *op,
2711                                     const ir_mode *imode, const ir_mode *omode,
2712                                     void *context)
2713 {
2714         char buf[64];
2715         ident *id;
2716         ir_entity *ent;
2717         (void) context;
2718
2719         if (imode == omode) {
2720                 snprintf(buf, sizeof(buf), "__l%s%s", get_op_name(op), get_mode_name(imode));
2721         } else {
2722                 snprintf(buf, sizeof(buf), "__l%s%s%s", get_op_name(op),
2723                         get_mode_name(imode), get_mode_name(omode));
2724         }
2725         id = new_id_from_str(buf);
2726
2727         ent = new_entity(get_glob_type(), id, method);
2728         set_entity_ld_ident(ent, get_entity_ident(ent));
2729         set_entity_visibility(ent, ir_visibility_external);
2730         return ent;
2731 }