nsz Git - libfirm/blob - ir/be/ia32/ia32_common_transform.c

   1 /*
   2  * Copyright (C) 1995-2008 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       This file implements the common parts of IR transformation from
  23  *              firm into ia32-Firm.
  24  * @author      Matthias Braun, Sebastian Buchwald
  25  * @version     $Id: ia32_common_transform.c 21012 2008-08-06 13:35:17Z beck $
  26  */
  27 #include "config.h"
  28
  29 #include "error.h"
  30 #include "irargs_t.h"
  31 #include "ircons.h"
  32 #include "irprintf.h"
  33 #include "typerep.h"
  34
  35 #include "../betranshlp.h"
  36 #include "../beirg_t.h"
  37
  38 #include "ia32_architecture.h"
  39 #include "ia32_common_transform.h"
  40 #include "ia32_new_nodes.h"
  41
  42 #include "gen_ia32_new_nodes.h"
  43 #include "gen_ia32_regalloc_if.h"
  44
  45 /** hold the current code generator during transformation */
  46 ia32_code_gen_t *env_cg = NULL;
  47
  48 heights_t *heights = NULL;
  49
  50 static const arch_register_req_t no_register_req = {
  51         arch_register_req_type_none,
  52         NULL,                         /* regclass */
  53         NULL,                         /* limit bitset */
  54         0,                            /* same pos */
  55         0                             /* different pos */
  56 };
  57
  58 static int check_immediate_constraint(long val, char immediate_constraint_type)
  59 {
  60         switch (immediate_constraint_type) {
  61                 case 0:
  62                 case 'i': return 1;
  63
  64                 case 'I': return    0 <= val && val <=  31;
  65                 case 'J': return    0 <= val && val <=  63;
  66                 case 'K': return -128 <= val && val <= 127;
  67                 case 'L': return val == 0xff || val == 0xffff;
  68                 case 'M': return    0 <= val && val <=   3;
  69                 case 'N': return    0 <= val && val <= 255;
  70                 case 'O': return    0 <= val && val <= 127;
  71
  72                 default: panic("Invalid immediate constraint found");
  73         }
  74 }
  75
  76 /**
  77  * creates a unique ident by adding a number to a tag
  78  *
  79  * @param tag   the tag string, must contain a %d if a number
  80  *              should be added
  81  */
  82 static ident *unique_id(const char *tag)
  83 {
  84         static unsigned id = 0;
  85         char str[256];
  86
  87         snprintf(str, sizeof(str), tag, ++id);
  88         return new_id_from_str(str);
  89 }
  90
  91 /**
  92  * Get a primitive type for a mode.
  93  */
  94 static ir_type *ia32_get_prim_type(pmap *types, ir_mode *mode)
  95 {
  96         pmap_entry *e = pmap_find(types, mode);
  97         ir_type *res;
  98
  99         if (! e) {
 100                 char buf[64];
 101                 snprintf(buf, sizeof(buf), "prim_type_%s", get_mode_name(mode));
 102                 res = new_type_primitive(new_id_from_str(buf), mode);
 103                 set_type_alignment_bytes(res, 16);
 104                 pmap_insert(types, mode, res);
 105         }
 106         else
 107                 res = e->value;
 108         return res;
 109 }
 110
 111 ir_entity *create_float_const_entity(ir_node *cnst)
 112 {
 113         ia32_isa_t *isa = env_cg->isa;
 114         tarval *key     = get_Const_tarval(cnst);
 115         pmap_entry *e   = pmap_find(isa->tv_ent, key);
 116         ir_entity *res;
 117         ir_graph *rem;
 118
 119         if (e == NULL) {
 120                 tarval  *tv   = key;
 121                 ir_mode *mode = get_tarval_mode(tv);
 122                 ir_type *tp;
 123
 124                 if (! ia32_cg_config.use_sse2) {
 125                         /* try to reduce the mode to produce smaller sized entities */
 126                         if (mode != mode_F) {
 127                                 if (tarval_ieee754_can_conv_lossless(tv, mode_F)) {
 128                                         mode = mode_F;
 129                                         tv = tarval_convert_to(tv, mode);
 130                                 } else if (mode != mode_D) {
 131                                         if (tarval_ieee754_can_conv_lossless(tv, mode_D)) {
 132                                                 mode = mode_D;
 133                                                 tv = tarval_convert_to(tv, mode);
 134                                         }
 135                                 }
 136                         }
 137                 }
 138
 139                 if (mode == get_irn_mode(cnst)) {
 140                         /* mode was not changed */
 141                         tp = get_Const_type(cnst);
 142                         if (tp == firm_unknown_type)
 143                                 tp = ia32_get_prim_type(isa->types, mode);
 144                 } else
 145                         tp = ia32_get_prim_type(isa->types, mode);
 146
 147                 res = new_entity(get_glob_type(), unique_id(".LC%u"), tp);
 148
 149                 set_entity_ld_ident(res, get_entity_ident(res));
 150                 set_entity_visibility(res, visibility_local);
 151                 set_entity_variability(res, variability_constant);
 152                 set_entity_allocation(res, allocation_static);
 153
 154                  /* we create a new entity here: It's initialization must resist on the
 155                     const code irg */
 156                 rem = current_ir_graph;
 157                 current_ir_graph = get_const_code_irg();
 158                 set_atomic_ent_value(res, new_Const_type(tv, tp));
 159                 current_ir_graph = rem;
 160
 161                 pmap_insert(isa->tv_ent, key, res);
 162         } else {
 163                 res = e->value;
 164         }
 165
 166         return res;
 167 }
 168
 169 ir_node *create_Immediate(ir_entity *symconst, int symconst_sign, long val)
 170 {
 171         ir_graph *irg         = current_ir_graph;
 172         ir_node  *start_block = get_irg_start_block(irg);
 173         ir_node  *immediate   = new_rd_ia32_Immediate(NULL, irg, start_block,
 174                                                       symconst, symconst_sign, val);
 175         arch_set_irn_register(immediate, &ia32_gp_regs[REG_GP_NOREG]);
 176
 177         return immediate;
 178 }
 179
 180 const arch_register_t *ia32_get_clobber_register(const char *clobber)
 181 {
 182         const arch_register_t       *reg = NULL;
 183         int                          c;
 184         size_t                       r;
 185         const arch_register_class_t *cls;
 186
 187         /* TODO: construct a hashmap instead of doing linear search for clobber
 188          * register */
 189         for(c = 0; c < N_CLASSES; ++c) {
 190                 cls = & ia32_reg_classes[c];
 191                 for(r = 0; r < cls->n_regs; ++r) {
 192                         const arch_register_t *temp_reg = arch_register_for_index(cls, r);
 193                         if(strcmp(temp_reg->name, clobber) == 0
 194                                         || (c == CLASS_ia32_gp && strcmp(temp_reg->name+1, clobber) == 0)) {
 195                                 reg = temp_reg;
 196                                 break;
 197                         }
 198                 }
 199                 if(reg != NULL)
 200                         break;
 201         }
 202
 203         return reg;
 204 }
 205
 206 #ifndef NDEBUG
 207 const char *ia32_get_old_node_name(ia32_code_gen_t *cg, ir_node *irn) {
 208         const ia32_isa_t *isa = cg->isa;
 209
 210         lc_eoprintf(firm_get_arg_env(), isa->name_obst, "%+F", irn);
 211         obstack_1grow(isa->name_obst, 0);
 212         return obstack_finish(isa->name_obst);
 213 }
 214 #endif /* NDEBUG */
 215
 216 int ia32_mode_needs_gp_reg(ir_mode *mode) {
 217         if(mode == mode_fpcw)
 218                 return 0;
 219         if(get_mode_size_bits(mode) > 32)
 220                 return 0;
 221         return mode_is_int(mode) || mode_is_reference(mode) || mode == mode_b;
 222 }
 223
 224 static void parse_asm_constraints(constraint_t *constraint, const char *c,
 225                            int is_output)
 226 {
 227         char                         immediate_type     = '\0';
 228         unsigned                     limited            = 0;
 229         const arch_register_class_t *cls                = NULL;
 230         int                          memory_possible       = 0;
 231         int                          all_registers_allowed = 0;
 232         int                          p;
 233         int                          same_as = -1;
 234
 235         memset(constraint, 0, sizeof(constraint[0]));
 236         constraint->same_as = -1;
 237
 238         if(*c == 0) {
 239                 /* a memory constraint: no need to do anything in backend about it
 240                  * (the dependencies are already respected by the memory edge of
 241                  * the node) */
 242                 return;
 243         }
 244
 245         /* TODO: improve error messages with node and source info. (As users can
 246          * easily hit these) */
 247         while(*c != 0) {
 248                 switch(*c) {
 249                 case ' ':
 250                 case '\t':
 251                 case '\n':
 252                         break;
 253
 254                 /* Skip out/in-out marker */
 255                 case '=': break;
 256                 case '+': break;
 257
 258                 case '*':
 259                         ++c;
 260                         break;
 261                 case '#':
 262                         while(*c != 0 && *c != ',')
 263                                 ++c;
 264                         break;
 265
 266                 case 'a':
 267                         assert(cls == NULL || cls == &ia32_reg_classes[CLASS_ia32_gp]);
 268                         cls      = &ia32_reg_classes[CLASS_ia32_gp];
 269                         limited |= 1 << REG_EAX;
 270                         break;
 271                 case 'b':
 272                         assert(cls == NULL || cls == &ia32_reg_classes[CLASS_ia32_gp]);
 273                         cls      = &ia32_reg_classes[CLASS_ia32_gp];
 274                         limited |= 1 << REG_EBX;
 275                         break;
 276                 case 'c':
 277                         assert(cls == NULL || cls == &ia32_reg_classes[CLASS_ia32_gp]);
 278                         cls      = &ia32_reg_classes[CLASS_ia32_gp];
 279                         limited |= 1 << REG_ECX;
 280                         break;
 281                 case 'd':
 282                         assert(cls == NULL || cls == &ia32_reg_classes[CLASS_ia32_gp]);
 283                         cls      = &ia32_reg_classes[CLASS_ia32_gp];
 284                         limited |= 1 << REG_EDX;
 285                         break;
 286                 case 'D':
 287                         assert(cls == NULL || cls == &ia32_reg_classes[CLASS_ia32_gp]);
 288                         cls      = &ia32_reg_classes[CLASS_ia32_gp];
 289                         limited |= 1 << REG_EDI;
 290                         break;
 291                 case 'S':
 292                         assert(cls == NULL || cls == &ia32_reg_classes[CLASS_ia32_gp]);
 293                         cls      = &ia32_reg_classes[CLASS_ia32_gp];
 294                         limited |= 1 << REG_ESI;
 295                         break;
 296                 case 'Q':
 297                 case 'q':
 298                         /* q means lower part of the regs only, this makes no
 299                          * difference to Q for us (we only assign whole registers) */
 300                         assert(cls == NULL || cls == &ia32_reg_classes[CLASS_ia32_gp]);
 301                         cls      = &ia32_reg_classes[CLASS_ia32_gp];
 302                         limited |= 1 << REG_EAX | 1 << REG_EBX | 1 << REG_ECX |
 303                                    1 << REG_EDX;
 304                         break;
 305                 case 'A':
 306                         assert(cls == NULL || cls == &ia32_reg_classes[CLASS_ia32_gp]);
 307                         cls      = &ia32_reg_classes[CLASS_ia32_gp];
 308                         limited |= 1 << REG_EAX | 1 << REG_EDX;
 309                         break;
 310                 case 'l':
 311                         assert(cls == NULL || cls == &ia32_reg_classes[CLASS_ia32_gp]);
 312                         cls      = &ia32_reg_classes[CLASS_ia32_gp];
 313                         limited |= 1 << REG_EAX | 1 << REG_EBX | 1 << REG_ECX |
 314                                    1 << REG_EDX | 1 << REG_ESI | 1 << REG_EDI |
 315                                    1 << REG_EBP;
 316                         break;
 317
 318                 case 'R':
 319                 case 'r':
 320                 case 'p':
 321                         if (cls != NULL && cls != &ia32_reg_classes[CLASS_ia32_gp])
 322                                 panic("multiple register classes not supported");
 323                         cls                   = &ia32_reg_classes[CLASS_ia32_gp];
 324                         all_registers_allowed = 1;
 325                         break;
 326
 327                 case 'f':
 328                 case 't':
 329                 case 'u':
 330                         /* TODO: mark values so the x87 simulator knows about t and u */
 331                         if (cls != NULL && cls != &ia32_reg_classes[CLASS_ia32_vfp])
 332                                 panic("multiple register classes not supported");
 333                         cls                   = &ia32_reg_classes[CLASS_ia32_vfp];
 334                         all_registers_allowed = 1;
 335                         break;
 336
 337                 case 'Y':
 338                 case 'x':
 339                         if (cls != NULL && cls != &ia32_reg_classes[CLASS_ia32_xmm])
 340                                 panic("multiple register classes not supproted");
 341                         cls                   = &ia32_reg_classes[CLASS_ia32_xmm];
 342                         all_registers_allowed = 1;
 343                         break;
 344
 345                 case 'I':
 346                 case 'J':
 347                 case 'K':
 348                 case 'L':
 349                 case 'M':
 350                 case 'N':
 351                 case 'O':
 352                         if (cls != NULL && cls != &ia32_reg_classes[CLASS_ia32_gp])
 353                                 panic("multiple register classes not supported");
 354                         if (immediate_type != '\0')
 355                                 panic("multiple immediate types not supported");
 356                         cls            = &ia32_reg_classes[CLASS_ia32_gp];
 357                         immediate_type = *c;
 358                         break;
 359                 case 'n':
 360                 case 'i':
 361                         if (cls != NULL && cls != &ia32_reg_classes[CLASS_ia32_gp])
 362                                 panic("multiple register classes not supported");
 363                         if (immediate_type != '\0')
 364                                 panic("multiple immediate types not supported");
 365                         cls            = &ia32_reg_classes[CLASS_ia32_gp];
 366                         immediate_type = 'i';
 367                         break;
 368
 369                 case 'X':
 370                 case 'g':
 371                         if (cls != NULL && cls != &ia32_reg_classes[CLASS_ia32_gp])
 372                                 panic("multiple register classes not supported");
 373                         if (immediate_type != '\0')
 374                                 panic("multiple immediate types not supported");
 375                         immediate_type        = 'i';
 376                         cls                   = &ia32_reg_classes[CLASS_ia32_gp];
 377                         all_registers_allowed = 1;
 378                         memory_possible       = 1;
 379                         break;
 380
 381                 case '0':
 382                 case '1':
 383                 case '2':
 384                 case '3':
 385                 case '4':
 386                 case '5':
 387                 case '6':
 388                 case '7':
 389                 case '8':
 390                 case '9':
 391                         if (is_output)
 392                                 panic("can only specify same constraint on input");
 393
 394                         sscanf(c, "%d%n", &same_as, &p);
 395                         if(same_as >= 0) {
 396                                 c += p;
 397                                 continue;
 398                         }
 399                         break;
 400
 401                 case 'm':
 402                 case 'o':
 403                 case 'V':
 404                         /* memory constraint no need to do anything in backend about it
 405                          * (the dependencies are already respected by the memory edge of
 406                          * the node) */
 407                         memory_possible = 1;
 408                         break;
 409
 410                 case 'E': /* no float consts yet */
 411                 case 'F': /* no float consts yet */
 412                 case 's': /* makes no sense on x86 */
 413                 case '<': /* no autodecrement on x86 */
 414                 case '>': /* no autoincrement on x86 */
 415                 case 'C': /* sse constant not supported yet */
 416                 case 'G': /* 80387 constant not supported yet */
 417                 case 'y': /* we don't support mmx registers yet */
 418                 case 'Z': /* not available in 32 bit mode */
 419                 case 'e': /* not available in 32 bit mode */
 420                         panic("unsupported asm constraint '%c' found in (%+F)",
 421                               *c, current_ir_graph);
 422                         break;
 423                 default:
 424                         panic("unknown asm constraint '%c' found in (%+F)", *c,
 425                               current_ir_graph);
 426                         break;
 427                 }
 428                 ++c;
 429         }
 430
 431         if(same_as >= 0) {
 432                 if (cls != NULL)
 433                         panic("same as and register constraint not supported");
 434                 if (immediate_type != '\0')
 435                         panic("same as and immediate constraint not supported");
 436         }
 437
 438         if (cls == NULL && same_as < 0) {
 439                 if (!memory_possible)
 440                         panic("no constraint specified for assembler input");
 441         }
 442
 443         constraint->same_as               = same_as;
 444         constraint->cls                   = cls;
 445         constraint->allowed_registers     = limited;
 446         constraint->all_registers_allowed = all_registers_allowed;
 447         constraint->memory_possible       = memory_possible;
 448         constraint->immediate_type        = immediate_type;
 449 }
 450
 451 ir_node *gen_ASM(ir_node *node)
 452 {
 453         ir_graph                   *irg       = current_ir_graph;
 454         ir_node                    *block = NULL;
 455         ir_node                    *new_block = NULL;
 456         dbg_info                   *dbgi      = get_irn_dbg_info(node);
 457         int                         i, arity;
 458         int                         out_idx;
 459         ir_node                   **in;
 460         ir_node                    *new_node;
 461         int                         out_arity;
 462         int                         n_out_constraints;
 463         int                         n_clobbers;
 464         const arch_register_req_t **out_reg_reqs;
 465         const arch_register_req_t **in_reg_reqs;
 466         ia32_asm_reg_t             *register_map;
 467         unsigned                    reg_map_size = 0;
 468         struct obstack             *obst;
 469         const ir_asm_constraint    *in_constraints;
 470         const ir_asm_constraint    *out_constraints;
 471         ident                     **clobbers;
 472         int                         clobbers_flags = 0;
 473         unsigned                    clobber_bits[N_CLASSES];
 474
 475         memset(&clobber_bits, 0, sizeof(clobber_bits));
 476
 477         switch (be_transformer) {
 478         case TRANSFORMER_DEFAULT:
 479                 block     = get_nodes_block(node);
 480                 new_block = be_transform_node(block);
 481                 break;
 482
 483 #ifdef FIRM_GRGEN_BE
 484         case TRANSFORMER_PBQP:
 485         case TRANSFORMER_RAND:
 486                 new_block = get_nodes_block(node);
 487                 break;
 488 #endif
 489
 490         default:
 491                 panic("invalid transformer");
 492         }
 493
 494         /* workaround for lots of buggy code out there as most people think volatile
 495          * asm is enough for everything and forget the flags (linux kernel, etc.)
 496          */
 497         if (get_irn_pinned(node) == op_pin_state_pinned) {
 498                 clobbers_flags = 1;
 499         }
 500
 501         arity = get_irn_arity(node);
 502         in    = ALLOCANZ(ir_node*, arity);
 503
 504         clobbers   = get_ASM_clobbers(node);
 505         n_clobbers = 0;
 506         for (i = 0; i < get_ASM_n_clobbers(node); ++i) {
 507                 const arch_register_req_t *req;
 508                 const char                *c = get_id_str(clobbers[i]);
 509
 510                 if (strcmp(c, "memory") == 0)
 511                         continue;
 512                 if (strcmp(c, "cc") == 0) {
 513                         clobbers_flags = 1;
 514                         continue;
 515                 }
 516
 517                 req = parse_clobber(c);
 518                 clobber_bits[req->cls->index] |= *req->limited;
 519
 520                 n_clobbers++;
 521         }
 522         n_out_constraints = get_ASM_n_output_constraints(node);
 523         out_arity         = n_out_constraints + n_clobbers;
 524
 525         in_constraints  = get_ASM_input_constraints(node);
 526         out_constraints = get_ASM_output_constraints(node);
 527
 528         /* determine size of register_map */
 529         for (out_idx = 0; out_idx < n_out_constraints; ++out_idx) {
 530                 const ir_asm_constraint *constraint = &out_constraints[out_idx];
 531                 if (constraint->pos > reg_map_size)
 532                         reg_map_size = constraint->pos;
 533         }
 534         for (i = 0; i < arity; ++i) {
 535                 const ir_asm_constraint   *constraint = &in_constraints[i];
 536                 if(constraint->pos > reg_map_size)
 537                         reg_map_size = constraint->pos;
 538         }
 539         ++reg_map_size;
 540
 541         obst         = get_irg_obstack(irg);
 542         register_map = NEW_ARR_D(ia32_asm_reg_t, obst, reg_map_size);
 543         memset(register_map, 0, reg_map_size * sizeof(register_map[0]));
 544
 545         /* construct output constraints */
 546         out_reg_reqs = obstack_alloc(obst, out_arity * sizeof(out_reg_reqs[0]));
 547
 548         for (out_idx = 0; out_idx < n_out_constraints; ++out_idx) {
 549                 const ir_asm_constraint   *constraint = &out_constraints[out_idx];
 550                 const char                *c       = get_id_str(constraint->constraint);
 551                 unsigned                   pos        = constraint->pos;
 552                 constraint_t               parsed_constraint;
 553                 const arch_register_req_t *req;
 554
 555                 parse_asm_constraints(&parsed_constraint, c, 1);
 556                 req = make_register_req(&parsed_constraint, n_out_constraints,
 557                                         out_reg_reqs, out_idx);
 558                 out_reg_reqs[out_idx] = req;
 559
 560                 register_map[pos].use_input = 0;
 561                 register_map[pos].valid     = 1;
 562                 register_map[pos].memory    = 0;
 563                 register_map[pos].inout_pos = out_idx;
 564                 register_map[pos].mode      = constraint->mode;
 565         }
 566
 567         /* inputs + input constraints */
 568         in_reg_reqs = obstack_alloc(obst, arity * sizeof(in_reg_reqs[0]));
 569         for (i = 0; i < arity; ++i) {
 570                 ir_node                   *pred         = get_irn_n(node, i);
 571                 const ir_asm_constraint   *constraint   = &in_constraints[i];
 572                 ident                     *constr_id    = constraint->constraint;
 573                 const char                *c            = get_id_str(constr_id);
 574                 unsigned                   pos          = constraint->pos;
 575                 int                        is_memory_op = 0;
 576                 ir_node                   *input        = NULL;
 577                 unsigned                   r_clobber_bits;
 578                 constraint_t               parsed_constraint;
 579                 const arch_register_req_t *req;
 580
 581                 parse_asm_constraints(&parsed_constraint, c, 0);
 582                 if (parsed_constraint.cls != NULL) {
 583                         r_clobber_bits = clobber_bits[parsed_constraint.cls->index];
 584                         if (r_clobber_bits != 0) {
 585                                 if (parsed_constraint.all_registers_allowed) {
 586                                         parsed_constraint.all_registers_allowed = 0;
 587                                         be_abi_set_non_ignore_regs(env_cg->birg->abi,
 588                                                         parsed_constraint.cls,
 589                                                         &parsed_constraint.allowed_registers);
 590                                 }
 591                                 parsed_constraint.allowed_registers &= ~r_clobber_bits;
 592                         }
 593                 }
 594
 595                 req = make_register_req(&parsed_constraint, n_out_constraints,
 596                                         out_reg_reqs, i);
 597                 in_reg_reqs[i] = req;
 598
 599                 if (parsed_constraint.immediate_type != '\0') {
 600                         char imm_type = parsed_constraint.immediate_type;
 601                         input = try_create_Immediate(pred, imm_type);
 602                 }
 603
 604                 if (input == NULL) {
 605                         ir_node *pred = NULL;
 606                         switch (be_transformer) {
 607                         case TRANSFORMER_DEFAULT:
 608                                 pred  = get_irn_n(node, i);
 609                                 input = be_transform_node(pred);
 610                                 break;
 611
 612 #ifdef FIRM_GRGEN_BE
 613                         case TRANSFORMER_PBQP:
 614                         case TRANSFORMER_RAND:
 615                                 input = get_irn_n(node, i);
 616                                 break;
 617 #endif
 618
 619                         default: panic("invalid transformer");
 620                         }
 621
 622                         if (parsed_constraint.cls == NULL
 623                                         && parsed_constraint.same_as < 0) {
 624                                 is_memory_op = 1;
 625                         } else if(parsed_constraint.memory_possible) {
 626                                 /* TODO: match Load or Load/Store if memory possible is set */
 627                         }
 628                 }
 629                 in[i] = input;
 630
 631                 register_map[pos].use_input = 1;
 632                 register_map[pos].valid     = 1;
 633                 register_map[pos].memory    = is_memory_op;
 634                 register_map[pos].inout_pos = i;
 635                 register_map[pos].mode      = constraint->mode;
 636         }
 637
 638         /* parse clobbers */
 639         for (i = 0; i < get_ASM_n_clobbers(node); ++i) {
 640                 const char                *c = get_id_str(clobbers[i]);
 641                 const arch_register_req_t *req;
 642
 643                 if (strcmp(c, "memory") == 0 || strcmp(c, "cc") == 0)
 644                         continue;
 645
 646                 req = parse_clobber(c);
 647                 out_reg_reqs[out_idx] = req;
 648                 ++out_idx;
 649         }
 650
 651         new_node = new_rd_ia32_Asm(dbgi, irg, new_block, arity, in, out_arity,
 652                                    get_ASM_text(node), register_map);
 653
 654         if (arity == 0)
 655                 be_dep_on_frame(new_node);
 656
 657         set_ia32_out_req_all(new_node, out_reg_reqs);
 658         set_ia32_in_req_all(new_node, in_reg_reqs);
 659
 660         SET_IA32_ORIG_NODE(new_node, ia32_get_old_node_name(env_cg, node));
 661
 662         return new_node;
 663 }
 664
 665 ir_node *gen_CopyB(ir_node *node) {
 666         ir_node  *block    = NULL;
 667         ir_node  *src      = NULL;
 668         ir_node  *new_src  = NULL;
 669         ir_node  *dst      = NULL;
 670         ir_node  *new_dst  = NULL;
 671         ir_node  *mem      = NULL;
 672         ir_node  *new_mem  = NULL;
 673         ir_node  *res      = NULL;
 674         ir_graph *irg      = current_ir_graph;
 675         dbg_info *dbgi     = get_irn_dbg_info(node);
 676         int      size      = get_type_size_bytes(get_CopyB_type(node));
 677         int      rem;
 678
 679         switch (be_transformer) {
 680                 case TRANSFORMER_DEFAULT:
 681                         block    = be_transform_node(get_nodes_block(node));
 682                         src      = get_CopyB_src(node);
 683                         new_src  = be_transform_node(src);
 684                         dst      = get_CopyB_dst(node);
 685                         new_dst  = be_transform_node(dst);
 686                         mem      = get_CopyB_mem(node);
 687                         new_mem  = be_transform_node(mem);
 688                         break;
 689
 690 #ifdef FIRM_GRGEN_BE
 691                 case TRANSFORMER_PBQP:
 692                 case TRANSFORMER_RAND:
 693                         block    = get_nodes_block(node);
 694                         new_src  = get_CopyB_src(node);
 695                         new_dst  = get_CopyB_dst(node);
 696                         new_mem  = get_CopyB_mem(node);
 697                         break;
 698 #endif
 699
 700                 default: panic("invalid transformer");
 701         }
 702
 703         /* If we have to copy more than 32 bytes, we use REP MOVSx and */
 704         /* then we need the size explicitly in ECX.                    */
 705         if (size >= 32 * 4) {
 706                 rem = size & 0x3; /* size % 4 */
 707                 size >>= 2;
 708
 709                 res = new_rd_ia32_Const(dbgi, irg, block, NULL, 0, size);
 710                 be_dep_on_frame(res);
 711
 712                 res = new_rd_ia32_CopyB(dbgi, irg, block, new_dst, new_src, res, new_mem, rem);
 713         } else {
 714                 if(size == 0) {
 715                         ir_fprintf(stderr, "Optimization warning copyb %+F with size <4\n",
 716                                    node);
 717                 }
 718                 res = new_rd_ia32_CopyB_i(dbgi, irg, block, new_dst, new_src, new_mem, size);
 719         }
 720
 721         SET_IA32_ORIG_NODE(res, ia32_get_old_node_name(env_cg, node));
 722
 723         return res;
 724 }
 725
 726 ir_node *gen_Proj_tls(ir_node *node) {
 727         ir_node  *block = NULL;
 728         ir_graph *irg   = current_ir_graph;
 729         dbg_info *dbgi  = NULL;
 730         ir_node  *res   = NULL;
 731
 732         switch (be_transformer) {
 733                 case TRANSFORMER_DEFAULT:
 734                         block = be_transform_node(get_nodes_block(node));
 735                         break;
 736
 737 #ifdef FIRM_GRGEN_BE
 738                 case TRANSFORMER_PBQP:
 739                 case TRANSFORMER_RAND:
 740                         block = get_nodes_block(node);
 741                         break;
 742 #endif
 743
 744                 default: panic("invalid transformer");
 745         }
 746
 747         res   = new_rd_ia32_LdTls(dbgi, irg, block, mode_Iu);
 748
 749         return res;
 750 }
 751
 752 ir_node *gen_Unknown(ir_node *node)
 753 {
 754         ir_mode *mode = get_irn_mode(node);
 755
 756         if (mode_is_float(mode)) {
 757                 if (ia32_cg_config.use_sse2) {
 758                         return ia32_new_Unknown_xmm(env_cg);
 759                 } else {
 760                         /* Unknown nodes are buggy in x87 simulator, use zero for now... */
 761                         ir_graph *irg   = current_ir_graph;
 762                         dbg_info *dbgi  = get_irn_dbg_info(node);
 763                         ir_node  *block = get_irg_start_block(irg);
 764                         ir_node  *ret   = new_rd_ia32_vfldz(dbgi, irg, block);
 765
 766                         be_dep_on_frame(ret);
 767                         return ret;
 768                 }
 769         } else if (ia32_mode_needs_gp_reg(mode)) {
 770                 return ia32_new_Unknown_gp(env_cg);
 771         } else {
 772                 panic("unsupported Unknown-Mode");
 773         }
 774         return NULL;
 775 }
 776
 777 const arch_register_req_t *make_register_req(const constraint_t *constraint,
 778                 int n_outs, const arch_register_req_t **out_reqs, int pos)
 779 {
 780         struct obstack      *obst    = get_irg_obstack(current_ir_graph);
 781         int                  same_as = constraint->same_as;
 782         arch_register_req_t *req;
 783
 784         if (same_as >= 0) {
 785                 const arch_register_req_t *other_constr;
 786
 787                 if (same_as >= n_outs)
 788                         panic("invalid output number in same_as constraint");
 789
 790                 other_constr     = out_reqs[same_as];
 791
 792                 req              = obstack_alloc(obst, sizeof(req[0]));
 793                 *req             = *other_constr;
 794                 req->type       |= arch_register_req_type_should_be_same;
 795                 req->other_same  = 1U << pos;
 796
 797                 /* switch constraints. This is because in firm we have same_as
 798                  * constraints on the output constraints while in the gcc asm syntax
 799                  * they are specified on the input constraints */
 800                 out_reqs[same_as] = req;
 801                 return other_constr;
 802         }
 803
 804         /* pure memory ops */
 805         if (constraint->cls == NULL) {
 806                 return &no_register_req;
 807         }
 808
 809         if (constraint->allowed_registers != 0
 810                         && !constraint->all_registers_allowed) {
 811                 unsigned *limited_ptr;
 812
 813                 req         = obstack_alloc(obst, sizeof(req[0]) + sizeof(unsigned));
 814                 memset(req, 0, sizeof(req[0]));
 815                 limited_ptr = (unsigned*) (req+1);
 816
 817                 req->type    = arch_register_req_type_limited;
 818                 *limited_ptr = constraint->allowed_registers;
 819                 req->limited = limited_ptr;
 820         } else {
 821                 req       = obstack_alloc(obst, sizeof(req[0]));
 822                 memset(req, 0, sizeof(req[0]));
 823                 req->type = arch_register_req_type_normal;
 824         }
 825         req->cls = constraint->cls;
 826
 827         return req;
 828 }
 829
 830 const arch_register_req_t *parse_clobber(const char *clobber)
 831 {
 832         struct obstack        *obst = get_irg_obstack(current_ir_graph);
 833         const arch_register_t *reg  = ia32_get_clobber_register(clobber);
 834         arch_register_req_t   *req;
 835         unsigned              *limited;
 836
 837         if(reg == NULL) {
 838                 panic("Register '%s' mentioned in asm clobber is unknown", clobber);
 839         }
 840
 841         assert(reg->index < 32);
 842
 843         limited  = obstack_alloc(obst, sizeof(limited[0]));
 844         *limited = 1 << reg->index;
 845
 846         req          = obstack_alloc(obst, sizeof(req[0]));
 847         memset(req, 0, sizeof(req[0]));
 848         req->type    = arch_register_req_type_limited;
 849         req->cls     = arch_register_get_class(reg);
 850         req->limited = limited;
 851
 852         return req;
 853 }
 854
 855
 856 int prevents_AM(ir_node *const block, ir_node *const am_candidate,
 857                        ir_node *const other)
 858 {
 859         if (get_nodes_block(other) != block)
 860                 return 0;
 861
 862         if (is_Sync(other)) {
 863                 int i;
 864
 865                 for (i = get_Sync_n_preds(other) - 1; i >= 0; --i) {
 866                         ir_node *const pred = get_Sync_pred(other, i);
 867
 868                         if (get_nodes_block(pred) != block)
 869                                 continue;
 870
 871                         /* Do not block ourselves from getting eaten */
 872                         if (is_Proj(pred) && get_Proj_pred(pred) == am_candidate)
 873                                 continue;
 874
 875                         if (!heights_reachable_in_block(heights, pred, am_candidate))
 876                                 continue;
 877
 878                         return 1;
 879                 }
 880
 881                 return 0;
 882         } else {
 883                 /* Do not block ourselves from getting eaten */
 884                 if (is_Proj(other) && get_Proj_pred(other) == am_candidate)
 885                         return 0;
 886
 887                 if (!heights_reachable_in_block(heights, other, am_candidate))
 888                         return 0;
 889
 890                 return 1;
 891         }
 892 }
 893
 894 ir_node *try_create_Immediate(ir_node *node, char immediate_constraint_type)
 895 {
 896         int          minus         = 0;
 897         tarval      *offset        = NULL;
 898         int          offset_sign   = 0;
 899         long         val = 0;
 900         ir_entity   *symconst_ent  = NULL;
 901         int          symconst_sign = 0;
 902         ir_mode     *mode;
 903         ir_node     *cnst          = NULL;
 904         ir_node     *symconst      = NULL;
 905         ir_node     *new_node;
 906
 907         mode = get_irn_mode(node);
 908         if(!mode_is_int(mode) && !mode_is_reference(mode)) {
 909                 return NULL;
 910         }
 911
 912         if(is_Minus(node)) {
 913                 minus = 1;
 914                 node  = get_Minus_op(node);
 915         }
 916
 917         if(is_Const(node)) {
 918                 cnst        = node;
 919                 symconst    = NULL;
 920                 offset_sign = minus;
 921         } else if(is_SymConst(node)) {
 922                 cnst          = NULL;
 923                 symconst      = node;
 924                 symconst_sign = minus;
 925         } else if(is_Add(node)) {
 926                 ir_node *left  = get_Add_left(node);
 927                 ir_node *right = get_Add_right(node);
 928                 if(is_Const(left) && is_SymConst(right)) {
 929                         cnst          = left;
 930                         symconst      = right;
 931                         symconst_sign = minus;
 932                         offset_sign   = minus;
 933                 } else if(is_SymConst(left) && is_Const(right)) {
 934                         cnst          = right;
 935                         symconst      = left;
 936                         symconst_sign = minus;
 937                         offset_sign   = minus;
 938                 }
 939         } else if(is_Sub(node)) {
 940                 ir_node *left  = get_Sub_left(node);
 941                 ir_node *right = get_Sub_right(node);
 942                 if(is_Const(left) && is_SymConst(right)) {
 943                         cnst          = left;
 944                         symconst      = right;
 945                         symconst_sign = !minus;
 946                         offset_sign   = minus;
 947                 } else if(is_SymConst(left) && is_Const(right)) {
 948                         cnst          = right;
 949                         symconst      = left;
 950                         symconst_sign = minus;
 951                         offset_sign   = !minus;
 952                 }
 953         } else {
 954                 return NULL;
 955         }
 956
 957         if(cnst != NULL) {
 958                 offset = get_Const_tarval(cnst);
 959                 if(tarval_is_long(offset)) {
 960                         val = get_tarval_long(offset);
 961                 } else {
 962                         ir_fprintf(stderr, "Optimisation Warning: tarval from %+F is not a "
 963                                    "long?\n", cnst);
 964                         return NULL;
 965                 }
 966
 967                 if(!check_immediate_constraint(val, immediate_constraint_type))
 968                         return NULL;
 969         }
 970         if(symconst != NULL) {
 971                 if(immediate_constraint_type != 0) {
 972                         /* we need full 32bits for symconsts */
 973                         return NULL;
 974                 }
 975
 976                 /* unfortunately the assembler/linker doesn't support -symconst */
 977                 if(symconst_sign)
 978                         return NULL;
 979
 980                 if(get_SymConst_kind(symconst) != symconst_addr_ent)
 981                         return NULL;
 982                 symconst_ent = get_SymConst_entity(symconst);
 983         }
 984         if(cnst == NULL && symconst == NULL)
 985                 return NULL;
 986
 987         if(offset_sign && offset != NULL) {
 988                 offset = tarval_neg(offset);
 989         }
 990
 991         new_node = create_Immediate(symconst_ent, symconst_sign, val);
 992
 993         return new_node;
 994 }