i386-gen.c

   1 /*
   2  *  X86 code generator for TCC
   3  *
   4  *  Copyright (c) 2001-2004 Fabrice Bellard
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2 of the License, or (at your option) any later version.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  19  */
  20
  21 #ifdef TARGET_DEFS_ONLY
  22
  23 /* number of available registers */
  24 #define NB_REGS         4
  25 #define NB_ASM_REGS     8
  26
  27 /* a register can belong to several classes. The classes must be
  28    sorted from more general to more precise (see gv2() code which does
  29    assumptions on it). */
  30 #define RC_INT     0x0001 /* generic integer register */
  31 #define RC_FLOAT   0x0002 /* generic float register */
  32 #define RC_EAX     0x0004
  33 #define RC_ST0     0x0008
  34 #define RC_ECX     0x0010
  35 #define RC_EDX     0x0020
  36 #define RC_IRET    RC_EAX /* function return: integer register */
  37 #define RC_LRET    RC_EDX /* function return: second integer register */
  38 #define RC_FRET    RC_ST0 /* function return: float register */
  39
  40 /* pretty names for the registers */
  41 enum {
  42     TREG_EAX = 0,
  43     TREG_ECX,
  44     TREG_EDX,
  45     TREG_ST0,
  46     TREG_ESP = 4
  47 };
  48
  49 /* return registers for function */
  50 #define REG_IRET TREG_EAX /* single word int return register */
  51 #define REG_LRET TREG_EDX /* second word return register (for long long) */
  52 #define REG_FRET TREG_ST0 /* float return register */
  53
  54 /* defined if function parameters must be evaluated in reverse order */
  55 #define INVERT_FUNC_PARAMS
  56
  57 /* defined if structures are passed as pointers. Otherwise structures
  58    are directly pushed on stack. */
  59 /* #define FUNC_STRUCT_PARAM_AS_PTR */
  60
  61 /* pointer size, in bytes */
  62 #define PTR_SIZE 4
  63
  64 /* long double size and alignment, in bytes */
  65 #define LDOUBLE_SIZE  12
  66 #define LDOUBLE_ALIGN 4
  67 /* maximum alignment (for aligned attribute support) */
  68 #define MAX_ALIGN     8
  69
  70
  71 #define psym oad
  72
  73 /******************************************************/
  74 /* ELF defines */
  75
  76 #define EM_TCC_TARGET EM_386
  77
  78 /* relocation type for 32 bit data relocation */
  79 #define R_DATA_32   R_386_32
  80 #define R_DATA_PTR  R_386_32
  81 #define R_JMP_SLOT  R_386_JMP_SLOT
  82 #define R_COPY      R_386_COPY
  83
  84 #define ELF_START_ADDR 0x08048000
  85 #define ELF_PAGE_SIZE  0x1000
  86
  87 /******************************************************/
  88 #else /* ! TARGET_DEFS_ONLY */
  89 /******************************************************/
  90 #include "tcc.h"
  91
  92 ST_DATA const int reg_classes[NB_REGS] = {
  93     /* eax */ RC_INT | RC_EAX,
  94     /* ecx */ RC_INT | RC_ECX,
  95     /* edx */ RC_INT | RC_EDX,
  96     /* st0 */ RC_FLOAT | RC_ST0,
  97 };
  98
  99 static unsigned long func_sub_sp_offset;
 100 static int func_ret_sub;
 101 #ifdef CONFIG_TCC_BCHECK
 102 static unsigned long func_bound_offset;
 103 #endif
 104
 105 /* XXX: make it faster ? */
 106 ST_FUNC void g(int c)
 107 {
 108     int ind1;
 109     ind1 = ind + 1;
 110     if (ind1 > cur_text_section->data_allocated)
 111         section_realloc(cur_text_section, ind1);
 112     cur_text_section->data[ind] = c;
 113     ind = ind1;
 114 }
 115
 116 ST_FUNC void o(unsigned int c)
 117 {
 118     while (c) {
 119         g(c);
 120         c = c >> 8;
 121     }
 122 }
 123
 124 ST_FUNC void gen_le16(int v)
 125 {
 126     g(v);
 127     g(v >> 8);
 128 }
 129
 130 ST_FUNC void gen_le32(int c)
 131 {
 132     g(c);
 133     g(c >> 8);
 134     g(c >> 16);
 135     g(c >> 24);
 136 }
 137
 138 /* output a symbol and patch all calls to it */
 139 ST_FUNC void gsym_addr(int t, int a)
 140 {
 141     int n, *ptr;
 142     while (t) {
 143         ptr = (int *)(cur_text_section->data + t);
 144         n = *ptr; /* next value */
 145         *ptr = a - t - 4;
 146         t = n;
 147     }
 148 }
 149
 150 ST_FUNC void gsym(int t)
 151 {
 152     gsym_addr(t, ind);
 153 }
 154
 155 /* psym is used to put an instruction with a data field which is a
 156    reference to a symbol. It is in fact the same as oad ! */
 157 #define psym oad
 158
 159 /* instruction + 4 bytes data. Return the address of the data */
 160 ST_FUNC int oad(int c, int s)
 161 {
 162     int ind1;
 163
 164     o(c);
 165     ind1 = ind + 4;
 166     if (ind1 > cur_text_section->data_allocated)
 167         section_realloc(cur_text_section, ind1);
 168     *(int *)(cur_text_section->data + ind) = s;
 169     s = ind;
 170     ind = ind1;
 171     return s;
 172 }
 173
 174 /* output constant with relocation if 'r & VT_SYM' is true */
 175 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
 176 {
 177     if (r & VT_SYM)
 178         greloc(cur_text_section, sym, ind, R_386_32);
 179     gen_le32(c);
 180 }
 181
 182 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 183 {
 184     if (r & VT_SYM)
 185         greloc(cur_text_section, sym, ind, R_386_PC32);
 186     gen_le32(c - 4);
 187 }
 188
 189 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 190    opcode bits */
 191 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 192 {
 193     op_reg = op_reg << 3;
 194     if ((r & VT_VALMASK) == VT_CONST) {
 195         /* constant memory reference */
 196         o(0x05 | op_reg);
 197         gen_addr32(r, sym, c);
 198     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 199         /* currently, we use only ebp as base */
 200         if (c == (char)c) {
 201             /* short reference */
 202             o(0x45 | op_reg);
 203             g(c);
 204         } else {
 205             oad(0x85 | op_reg, c);
 206         }
 207     } else {
 208         g(0x00 | op_reg | (r & VT_VALMASK));
 209     }
 210 }
 211
 212 /* load 'r' from value 'sv' */
 213 ST_FUNC void load(int r, SValue *sv)
 214 {
 215     int v, t, ft, fc, fr;
 216     SValue v1;
 217
 218 #ifdef TCC_TARGET_PE
 219     SValue v2;
 220     sv = pe_getimport(sv, &v2);
 221 #endif
 222
 223     fr = sv->r;
 224     ft = sv->type.t;
 225     fc = sv->c.ul;
 226
 227     v = fr & VT_VALMASK;
 228     if (fr & VT_LVAL) {
 229         if (v == VT_LLOCAL) {
 230             v1.type.t = VT_INT;
 231             v1.r = VT_LOCAL | VT_LVAL;
 232             v1.c.ul = fc;
 233             fr = r;
 234             if (!(reg_classes[fr] & RC_INT))
 235                 fr = get_reg(RC_INT);
 236             load(fr, &v1);
 237         }
 238         if ((ft & VT_BTYPE) == VT_FLOAT) {
 239             o(0xd9); /* flds */
 240             r = 0;
 241         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 242             o(0xdd); /* fldl */
 243             r = 0;
 244         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 245             o(0xdb); /* fldt */
 246             r = 5;
 247         } else if ((ft & VT_TYPE) == VT_BYTE) {
 248             o(0xbe0f);   /* movsbl */
 249         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 250             o(0xb60f);   /* movzbl */
 251         } else if ((ft & VT_TYPE) == VT_SHORT) {
 252             o(0xbf0f);   /* movswl */
 253         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 254             o(0xb70f);   /* movzwl */
 255         } else {
 256             o(0x8b);     /* movl */
 257         }
 258         gen_modrm(r, fr, sv->sym, fc);
 259     } else {
 260         if (v == VT_CONST) {
 261             o(0xb8 + r); /* mov $xx, r */
 262             gen_addr32(fr, sv->sym, fc);
 263         } else if (v == VT_LOCAL) {
 264             if (fc) {
 265                 o(0x8d); /* lea xxx(%ebp), r */
 266                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 267             } else {
 268                 o(0x89);
 269                 o(0xe8 + r); /* mov %ebp, r */
 270             }
 271         } else if (v == VT_CMP) {
 272             oad(0xb8 + r, 0); /* mov $0, r */
 273             o(0x0f); /* setxx %br */
 274             o(fc);
 275             o(0xc0 + r);
 276         } else if (v == VT_JMP || v == VT_JMPI) {
 277             t = v & 1;
 278             oad(0xb8 + r, t); /* mov $1, r */
 279             o(0x05eb); /* jmp after */
 280             gsym(fc);
 281             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 282         } else if (v != r) {
 283             o(0x89);
 284             o(0xc0 + r + v * 8); /* mov v, r */
 285         }
 286     }
 287 }
 288
 289 /* store register 'r' in lvalue 'v' */
 290 ST_FUNC void store(int r, SValue *v)
 291 {
 292     int fr, bt, ft, fc;
 293
 294 #ifdef TCC_TARGET_PE
 295     SValue v2;
 296     v = pe_getimport(v, &v2);
 297 #endif
 298
 299     ft = v->type.t;
 300     fc = v->c.ul;
 301     fr = v->r & VT_VALMASK;
 302     bt = ft & VT_BTYPE;
 303     /* XXX: incorrect if float reg to reg */
 304     if (bt == VT_FLOAT) {
 305         o(0xd9); /* fsts */
 306         r = 2;
 307     } else if (bt == VT_DOUBLE) {
 308         o(0xdd); /* fstpl */
 309         r = 2;
 310     } else if (bt == VT_LDOUBLE) {
 311         o(0xc0d9); /* fld %st(0) */
 312         o(0xdb); /* fstpt */
 313         r = 7;
 314     } else {
 315         if (bt == VT_SHORT)
 316             o(0x66);
 317         if (bt == VT_BYTE || bt == VT_BOOL)
 318             o(0x88);
 319         else
 320             o(0x89);
 321     }
 322     if (fr == VT_CONST ||
 323         fr == VT_LOCAL ||
 324         (v->r & VT_LVAL)) {
 325         gen_modrm(r, v->r, v->sym, fc);
 326     } else if (fr != r) {
 327         o(0xc0 + fr + r * 8); /* mov r, fr */
 328     }
 329 }
 330
 331 static void gadd_sp(int val)
 332 {
 333     if (val == (char)val) {
 334         o(0xc483);
 335         g(val);
 336     } else {
 337         oad(0xc481, val); /* add $xxx, %esp */
 338     }
 339 }
 340
 341 /* 'is_jmp' is '1' if it is a jump */
 342 static void gcall_or_jmp(int is_jmp)
 343 {
 344     int r;
 345     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 346         /* constant case */
 347         if (vtop->r & VT_SYM) {
 348             /* relocation case */
 349             greloc(cur_text_section, vtop->sym,
 350                    ind + 1, R_386_PC32);
 351         } else {
 352             /* put an empty PC32 relocation */
 353             put_elf_reloc(symtab_section, cur_text_section,
 354                           ind + 1, R_386_PC32, 0);
 355         }
 356         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 357     } else {
 358         /* otherwise, indirect call */
 359         r = gv(RC_INT);
 360         o(0xff); /* call/jmp *r */
 361         o(0xd0 + r + (is_jmp << 4));
 362     }
 363 }
 364
 365 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 366 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 367
 368 /* Return 1 if this function returns via an sret pointer, 0 otherwise */
 369 ST_FUNC int gfunc_sret(CType *vt, CType *ret, int *ret_align)
 370 {
 371 #ifdef TCC_TARGET_PE
 372     int size, align;
 373
 374     *ret_align = 1; // Never have to re-align return values for x86
 375     size = type_size(vt, &align);
 376     if (size > 8) {
 377         return 1;
 378     } else if (size > 4) {
 379         ret->ref = NULL;
 380         ret->t = VT_LLONG;
 381         return 0;
 382     } else {
 383         ret->ref = NULL;
 384         ret->t = VT_INT;
 385         return 0;
 386     }
 387 #else
 388     *ret_align = 1; // Never have to re-align return values for x86
 389     return 1;
 390 #endif
 391 }
 392
 393 /* Generate function call. The function address is pushed first, then
 394    all the parameters in call order. This functions pops all the
 395    parameters and the function address. */
 396 ST_FUNC void gfunc_call(int nb_args)
 397 {
 398     int size, align, r, args_size, i, func_call;
 399     Sym *func_sym;
 400
 401     args_size = 0;
 402     for(i = 0;i < nb_args; i++) {
 403         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 404             size = type_size(&vtop->type, &align);
 405             /* align to stack align size */
 406             size = (size + 3) & ~3;
 407             /* allocate the necessary size on stack */
 408             oad(0xec81, size); /* sub $xxx, %esp */
 409             /* generate structure store */
 410             r = get_reg(RC_INT);
 411             o(0x89); /* mov %esp, r */
 412             o(0xe0 + r);
 413             vset(&vtop->type, r | VT_LVAL, 0);
 414             vswap();
 415             vstore();
 416             args_size += size;
 417         } else if (is_float(vtop->type.t)) {
 418             gv(RC_FLOAT); /* only one float register */
 419             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 420                 size = 4;
 421             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 422                 size = 8;
 423             else
 424                 size = 12;
 425             oad(0xec81, size); /* sub $xxx, %esp */
 426             if (size == 12)
 427                 o(0x7cdb);
 428             else
 429                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 430             g(0x24);
 431             g(0x00);
 432             args_size += size;
 433         } else {
 434             /* simple type (currently always same size) */
 435             /* XXX: implicit cast ? */
 436             r = gv(RC_INT);
 437             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 438                 size = 8;
 439                 o(0x50 + vtop->r2); /* push r */
 440             } else {
 441                 size = 4;
 442             }
 443             o(0x50 + r); /* push r */
 444             args_size += size;
 445         }
 446         vtop--;
 447     }
 448     save_regs(0); /* save used temporary registers */
 449     func_sym = vtop->type.ref;
 450     func_call = FUNC_CALL(func_sym->r);
 451     /* fast call case */
 452     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 453         func_call == FUNC_FASTCALLW) {
 454         int fastcall_nb_regs;
 455         uint8_t *fastcall_regs_ptr;
 456         if (func_call == FUNC_FASTCALLW) {
 457             fastcall_regs_ptr = fastcallw_regs;
 458             fastcall_nb_regs = 2;
 459         } else {
 460             fastcall_regs_ptr = fastcall_regs;
 461             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 462         }
 463         for(i = 0;i < fastcall_nb_regs; i++) {
 464             if (args_size <= 0)
 465                 break;
 466             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 467             /* XXX: incorrect for struct/floats */
 468             args_size -= 4;
 469         }
 470     }
 471 #ifndef TCC_TARGET_PE
 472     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 473         args_size -= 4;
 474 #endif
 475     gcall_or_jmp(0);
 476
 477     if (args_size && func_call != FUNC_STDCALL)
 478         gadd_sp(args_size);
 479     vtop--;
 480 }
 481
 482 #ifdef TCC_TARGET_PE
 483 #define FUNC_PROLOG_SIZE 10
 484 #else
 485 #define FUNC_PROLOG_SIZE 9
 486 #endif
 487
 488 /* generate function prolog of type 't' */
 489 ST_FUNC void gfunc_prolog(CType *func_type)
 490 {
 491     int addr, align, size, func_call, fastcall_nb_regs;
 492     int param_index, param_addr;
 493     uint8_t *fastcall_regs_ptr;
 494     Sym *sym;
 495     CType *type;
 496
 497     sym = func_type->ref;
 498     func_call = FUNC_CALL(sym->r);
 499     addr = 8;
 500     loc = 0;
 501     func_vc = 0;
 502
 503     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 504         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 505         fastcall_regs_ptr = fastcall_regs;
 506     } else if (func_call == FUNC_FASTCALLW) {
 507         fastcall_nb_regs = 2;
 508         fastcall_regs_ptr = fastcallw_regs;
 509     } else {
 510         fastcall_nb_regs = 0;
 511         fastcall_regs_ptr = NULL;
 512     }
 513     param_index = 0;
 514
 515     ind += FUNC_PROLOG_SIZE;
 516     func_sub_sp_offset = ind;
 517     /* if the function returns a structure, then add an
 518        implicit pointer parameter */
 519     func_vt = sym->type;
 520 #ifdef TCC_TARGET_PE
 521     size = type_size(&func_vt,&align);
 522     if (((func_vt.t & VT_BTYPE) == VT_STRUCT) && (size > 8)) {
 523 #else
 524     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 525 #endif
 526         /* XXX: fastcall case ? */
 527         func_vc = addr;
 528         addr += 4;
 529         param_index++;
 530     }
 531     /* define parameters */
 532     while ((sym = sym->next) != NULL) {
 533         type = &sym->type;
 534         size = type_size(type, &align);
 535         size = (size + 3) & ~3;
 536 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 537         /* structs are passed as pointer */
 538         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 539             size = 4;
 540         }
 541 #endif
 542         if (param_index < fastcall_nb_regs) {
 543             /* save FASTCALL register */
 544             loc -= 4;
 545             o(0x89);     /* movl */
 546             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 547             param_addr = loc;
 548         } else {
 549             param_addr = addr;
 550             addr += size;
 551         }
 552         sym_push(sym->v & ~SYM_FIELD, type,
 553                  VT_LOCAL | lvalue_type(type->t), param_addr);
 554         param_index++;
 555     }
 556     func_ret_sub = 0;
 557     /* pascal type call ? */
 558     if (func_call == FUNC_STDCALL)
 559         func_ret_sub = addr - 8;
 560 #ifndef TCC_TARGET_PE
 561     else if (func_vc)
 562         func_ret_sub = 4;
 563 #endif
 564
 565 #ifdef CONFIG_TCC_BCHECK
 566     /* leave some room for bound checking code */
 567     if (tcc_state->do_bounds_check) {
 568         oad(0xb8, 0); /* lbound section pointer */
 569         oad(0xb8, 0); /* call to function */
 570         func_bound_offset = lbounds_section->data_offset;
 571     }
 572 #endif
 573 }
 574
 575 /* generate function epilog */
 576 ST_FUNC void gfunc_epilog(void)
 577 {
 578     int v, saved_ind;
 579
 580 #ifdef CONFIG_TCC_BCHECK
 581     if (tcc_state->do_bounds_check
 582      && func_bound_offset != lbounds_section->data_offset) {
 583         int saved_ind;
 584         int *bounds_ptr;
 585         Sym *sym, *sym_data;
 586         /* add end of table info */
 587         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 588         *bounds_ptr = 0;
 589         /* generate bound local allocation */
 590         saved_ind = ind;
 591         ind = func_sub_sp_offset;
 592         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 593                                func_bound_offset, lbounds_section->data_offset);
 594         greloc(cur_text_section, sym_data,
 595                ind + 1, R_386_32);
 596         oad(0xb8, 0); /* mov %eax, xxx */
 597         sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
 598         greloc(cur_text_section, sym,
 599                ind + 1, R_386_PC32);
 600         oad(0xe8, -4);
 601         ind = saved_ind;
 602         /* generate bound check local freeing */
 603         o(0x5250); /* save returned value, if any */
 604         greloc(cur_text_section, sym_data,
 605                ind + 1, R_386_32);
 606         oad(0xb8, 0); /* mov %eax, xxx */
 607         sym = external_global_sym(TOK___bound_local_delete, &func_old_type, 0);
 608         greloc(cur_text_section, sym,
 609                ind + 1, R_386_PC32);
 610         oad(0xe8, -4);
 611         o(0x585a); /* restore returned value, if any */
 612     }
 613 #endif
 614     o(0xc9); /* leave */
 615     if (func_ret_sub == 0) {
 616         o(0xc3); /* ret */
 617     } else {
 618         o(0xc2); /* ret n */
 619         g(func_ret_sub);
 620         g(func_ret_sub >> 8);
 621     }
 622     /* align local size to word & save local variables */
 623
 624     v = (-loc + 3) & -4;
 625     saved_ind = ind;
 626     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 627 #ifdef TCC_TARGET_PE
 628     if (v >= 4096) {
 629         Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
 630         oad(0xb8, v); /* mov stacksize, %eax */
 631         oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
 632         greloc(cur_text_section, sym, ind-4, R_386_PC32);
 633     } else
 634 #endif
 635     {
 636         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 637         o(0xec81);  /* sub esp, stacksize */
 638         gen_le32(v);
 639 #if FUNC_PROLOG_SIZE == 10
 640         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 641 #endif
 642     }
 643     ind = saved_ind;
 644 }
 645
 646 /* generate a jump to a label */
 647 ST_FUNC int gjmp(int t)
 648 {
 649     return psym(0xe9, t);
 650 }
 651
 652 /* generate a jump to a fixed address */
 653 ST_FUNC void gjmp_addr(int a)
 654 {
 655     int r;
 656     r = a - ind - 2;
 657     if (r == (char)r) {
 658         g(0xeb);
 659         g(r);
 660     } else {
 661         oad(0xe9, a - ind - 5);
 662     }
 663 }
 664
 665 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 666 ST_FUNC int gtst(int inv, int t)
 667 {
 668     int v, *p;
 669
 670     v = vtop->r & VT_VALMASK;
 671     if (v == VT_CMP) {
 672         /* fast case : can jump directly since flags are set */
 673         g(0x0f);
 674         t = psym((vtop->c.i - 16) ^ inv, t);
 675     } else if (v == VT_JMP || v == VT_JMPI) {
 676         /* && or || optimization */
 677         if ((v & 1) == inv) {
 678             /* insert vtop->c jump list in t */
 679             p = &vtop->c.i;
 680             while (*p != 0)
 681                 p = (int *)(cur_text_section->data + *p);
 682             *p = t;
 683             t = vtop->c.i;
 684         } else {
 685             t = gjmp(t);
 686             gsym(vtop->c.i);
 687         }
 688     } else {
 689         if (is_float(vtop->type.t) ||
 690             (vtop->type.t & VT_BTYPE) == VT_LLONG) {
 691             vpushi(0);
 692             gen_op(TOK_NE);
 693         }
 694         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 695             /* constant jmp optimization */
 696             if ((vtop->c.i != 0) != inv)
 697                 t = gjmp(t);
 698         } else {
 699             v = gv(RC_INT);
 700             o(0x85);
 701             o(0xc0 + v * 9);
 702             g(0x0f);
 703             t = psym(0x85 ^ inv, t);
 704         }
 705     }
 706     vtop--;
 707     return t;
 708 }
 709
 710 /* generate an integer binary operation */
 711 ST_FUNC void gen_opi(int op)
 712 {
 713     int r, fr, opc, c;
 714
 715     switch(op) {
 716     case '+':
 717     case TOK_ADDC1: /* add with carry generation */
 718         opc = 0;
 719     gen_op8:
 720         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 721             /* constant case */
 722             vswap();
 723             r = gv(RC_INT);
 724             vswap();
 725             c = vtop->c.i;
 726             if (c == (char)c) {
 727                 /* generate inc and dec for smaller code */
 728                 if (c==1 && opc==0) {
 729                     o (0x40 | r); // inc
 730                 } else if (c==1 && opc==5) {
 731                     o (0x48 | r); // dec
 732                 } else {
 733                     o(0x83);
 734                     o(0xc0 | (opc << 3) | r);
 735                     g(c);
 736                 }
 737             } else {
 738                 o(0x81);
 739                 oad(0xc0 | (opc << 3) | r, c);
 740             }
 741         } else {
 742             gv2(RC_INT, RC_INT);
 743             r = vtop[-1].r;
 744             fr = vtop[0].r;
 745             o((opc << 3) | 0x01);
 746             o(0xc0 + r + fr * 8);
 747         }
 748         vtop--;
 749         if (op >= TOK_ULT && op <= TOK_GT) {
 750             vtop->r = VT_CMP;
 751             vtop->c.i = op;
 752         }
 753         break;
 754     case '-':
 755     case TOK_SUBC1: /* sub with carry generation */
 756         opc = 5;
 757         goto gen_op8;
 758     case TOK_ADDC2: /* add with carry use */
 759         opc = 2;
 760         goto gen_op8;
 761     case TOK_SUBC2: /* sub with carry use */
 762         opc = 3;
 763         goto gen_op8;
 764     case '&':
 765         opc = 4;
 766         goto gen_op8;
 767     case '^':
 768         opc = 6;
 769         goto gen_op8;
 770     case '|':
 771         opc = 1;
 772         goto gen_op8;
 773     case '*':
 774         gv2(RC_INT, RC_INT);
 775         r = vtop[-1].r;
 776         fr = vtop[0].r;
 777         vtop--;
 778         o(0xaf0f); /* imul fr, r */
 779         o(0xc0 + fr + r * 8);
 780         break;
 781     case TOK_SHL:
 782         opc = 4;
 783         goto gen_shift;
 784     case TOK_SHR:
 785         opc = 5;
 786         goto gen_shift;
 787     case TOK_SAR:
 788         opc = 7;
 789     gen_shift:
 790         opc = 0xc0 | (opc << 3);
 791         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 792             /* constant case */
 793             vswap();
 794             r = gv(RC_INT);
 795             vswap();
 796             c = vtop->c.i & 0x1f;
 797             o(0xc1); /* shl/shr/sar $xxx, r */
 798             o(opc | r);
 799             g(c);
 800         } else {
 801             /* we generate the shift in ecx */
 802             gv2(RC_INT, RC_ECX);
 803             r = vtop[-1].r;
 804             o(0xd3); /* shl/shr/sar %cl, r */
 805             o(opc | r);
 806         }
 807         vtop--;
 808         break;
 809     case '/':
 810     case TOK_UDIV:
 811     case TOK_PDIV:
 812     case '%':
 813     case TOK_UMOD:
 814     case TOK_UMULL:
 815         /* first operand must be in eax */
 816         /* XXX: need better constraint for second operand */
 817         gv2(RC_EAX, RC_ECX);
 818         r = vtop[-1].r;
 819         fr = vtop[0].r;
 820         vtop--;
 821         save_reg(TREG_EDX);
 822         if (op == TOK_UMULL) {
 823             o(0xf7); /* mul fr */
 824             o(0xe0 + fr);
 825             vtop->r2 = TREG_EDX;
 826             r = TREG_EAX;
 827         } else {
 828             if (op == TOK_UDIV || op == TOK_UMOD) {
 829                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 830                 o(0xf0 + fr);
 831             } else {
 832                 o(0xf799); /* cltd, idiv fr, %eax */
 833                 o(0xf8 + fr);
 834             }
 835             if (op == '%' || op == TOK_UMOD)
 836                 r = TREG_EDX;
 837             else
 838                 r = TREG_EAX;
 839         }
 840         vtop->r = r;
 841         break;
 842     default:
 843         opc = 7;
 844         goto gen_op8;
 845     }
 846 }
 847
 848 /* generate a floating point operation 'v = t1 op t2' instruction. The
 849    two operands are guaranted to have the same floating point type */
 850 /* XXX: need to use ST1 too */
 851 ST_FUNC void gen_opf(int op)
 852 {
 853     int a, ft, fc, swapped, r;
 854
 855     /* convert constants to memory references */
 856     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 857         vswap();
 858         gv(RC_FLOAT);
 859         vswap();
 860     }
 861     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 862         gv(RC_FLOAT);
 863
 864     /* must put at least one value in the floating point register */
 865     if ((vtop[-1].r & VT_LVAL) &&
 866         (vtop[0].r & VT_LVAL)) {
 867         vswap();
 868         gv(RC_FLOAT);
 869         vswap();
 870     }
 871     swapped = 0;
 872     /* swap the stack if needed so that t1 is the register and t2 is
 873        the memory reference */
 874     if (vtop[-1].r & VT_LVAL) {
 875         vswap();
 876         swapped = 1;
 877     }
 878     if (op >= TOK_ULT && op <= TOK_GT) {
 879         /* load on stack second operand */
 880         load(TREG_ST0, vtop);
 881         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 882         if (op == TOK_GE || op == TOK_GT)
 883             swapped = !swapped;
 884         else if (op == TOK_EQ || op == TOK_NE)
 885             swapped = 0;
 886         if (swapped)
 887             o(0xc9d9); /* fxch %st(1) */
 888         o(0xe9da); /* fucompp */
 889         o(0xe0df); /* fnstsw %ax */
 890         if (op == TOK_EQ) {
 891             o(0x45e480); /* and $0x45, %ah */
 892             o(0x40fC80); /* cmp $0x40, %ah */
 893         } else if (op == TOK_NE) {
 894             o(0x45e480); /* and $0x45, %ah */
 895             o(0x40f480); /* xor $0x40, %ah */
 896             op = TOK_NE;
 897         } else if (op == TOK_GE || op == TOK_LE) {
 898             o(0x05c4f6); /* test $0x05, %ah */
 899             op = TOK_EQ;
 900         } else {
 901             o(0x45c4f6); /* test $0x45, %ah */
 902             op = TOK_EQ;
 903         }
 904         vtop--;
 905         vtop->r = VT_CMP;
 906         vtop->c.i = op;
 907     } else {
 908         /* no memory reference possible for long double operations */
 909         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 910             load(TREG_ST0, vtop);
 911             swapped = !swapped;
 912         }
 913
 914         switch(op) {
 915         default:
 916         case '+':
 917             a = 0;
 918             break;
 919         case '-':
 920             a = 4;
 921             if (swapped)
 922                 a++;
 923             break;
 924         case '*':
 925             a = 1;
 926             break;
 927         case '/':
 928             a = 6;
 929             if (swapped)
 930                 a++;
 931             break;
 932         }
 933         ft = vtop->type.t;
 934         fc = vtop->c.ul;
 935         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 936             o(0xde); /* fxxxp %st, %st(1) */
 937             o(0xc1 + (a << 3));
 938         } else {
 939             /* if saved lvalue, then we must reload it */
 940             r = vtop->r;
 941             if ((r & VT_VALMASK) == VT_LLOCAL) {
 942                 SValue v1;
 943                 r = get_reg(RC_INT);
 944                 v1.type.t = VT_INT;
 945                 v1.r = VT_LOCAL | VT_LVAL;
 946                 v1.c.ul = fc;
 947                 load(r, &v1);
 948                 fc = 0;
 949             }
 950
 951             if ((ft & VT_BTYPE) == VT_DOUBLE)
 952                 o(0xdc);
 953             else
 954                 o(0xd8);
 955             gen_modrm(a, r, vtop->sym, fc);
 956         }
 957         vtop--;
 958     }
 959 }
 960
 961 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 962    and 'long long' cases. */
 963 ST_FUNC void gen_cvt_itof(int t)
 964 {
 965     save_reg(TREG_ST0);
 966     gv(RC_INT);
 967     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 968         /* signed long long to float/double/long double (unsigned case
 969            is handled generically) */
 970         o(0x50 + vtop->r2); /* push r2 */
 971         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 972         o(0x242cdf); /* fildll (%esp) */
 973         o(0x08c483); /* add $8, %esp */
 974     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 975                (VT_INT | VT_UNSIGNED)) {
 976         /* unsigned int to float/double/long double */
 977         o(0x6a); /* push $0 */
 978         g(0x00);
 979         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 980         o(0x242cdf); /* fildll (%esp) */
 981         o(0x08c483); /* add $8, %esp */
 982     } else {
 983         /* int to float/double/long double */
 984         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 985         o(0x2404db); /* fildl (%esp) */
 986         o(0x04c483); /* add $4, %esp */
 987     }
 988     vtop->r = TREG_ST0;
 989 }
 990
 991 /* convert fp to int 't' type */
 992 /* XXX: handle long long case */
 993 ST_FUNC void gen_cvt_ftoi(int t)
 994 {
 995     int r, r2, size;
 996     Sym *sym;
 997     CType ushort_type;
 998
 999     ushort_type.t = VT_SHORT | VT_UNSIGNED;
1000     ushort_type.ref = 0;
1001
1002     gv(RC_FLOAT);
1003     if (t != VT_INT)
1004         size = 8;
1005     else
1006         size = 4;
1007
1008     o(0x2dd9); /* ldcw xxx */
1009     sym = external_global_sym(TOK___tcc_int_fpu_control,
1010                               &ushort_type, VT_LVAL);
1011     greloc(cur_text_section, sym,
1012            ind, R_386_32);
1013     gen_le32(0);
1014
1015     oad(0xec81, size); /* sub $xxx, %esp */
1016     if (size == 4)
1017         o(0x1cdb); /* fistpl */
1018     else
1019         o(0x3cdf); /* fistpll */
1020     o(0x24);
1021     o(0x2dd9); /* ldcw xxx */
1022     sym = external_global_sym(TOK___tcc_fpu_control,
1023                               &ushort_type, VT_LVAL);
1024     greloc(cur_text_section, sym,
1025            ind, R_386_32);
1026     gen_le32(0);
1027
1028     r = get_reg(RC_INT);
1029     o(0x58 + r); /* pop r */
1030     if (size == 8) {
1031         if (t == VT_LLONG) {
1032             vtop->r = r; /* mark reg as used */
1033             r2 = get_reg(RC_INT);
1034             o(0x58 + r2); /* pop r2 */
1035             vtop->r2 = r2;
1036         } else {
1037             o(0x04c483); /* add $4, %esp */
1038         }
1039     }
1040     vtop->r = r;
1041 }
1042
1043 /* convert from one floating point type to another */
1044 ST_FUNC void gen_cvt_ftof(int t)
1045 {
1046     /* all we have to do on i386 is to put the float in a register */
1047     gv(RC_FLOAT);
1048 }
1049
1050 /* computed goto support */
1051 ST_FUNC void ggoto(void)
1052 {
1053     gcall_or_jmp(1);
1054     vtop--;
1055 }
1056
1057 /* bound check support functions */
1058 #ifdef CONFIG_TCC_BCHECK
1059
1060 /* generate a bounded pointer addition */
1061 ST_FUNC void gen_bounded_ptr_add(void)
1062 {
1063     Sym *sym;
1064
1065     /* prepare fast i386 function call (args in eax and edx) */
1066     gv2(RC_EAX, RC_EDX);
1067     /* save all temporary registers */
1068     vtop -= 2;
1069     save_regs(0);
1070     /* do a fast function call */
1071     sym = external_global_sym(TOK___bound_ptr_add, &func_old_type, 0);
1072     greloc(cur_text_section, sym,
1073            ind + 1, R_386_PC32);
1074     oad(0xe8, -4);
1075     /* returned pointer is in eax */
1076     vtop++;
1077     vtop->r = TREG_EAX | VT_BOUNDED;
1078     /* address of bounding function call point */
1079     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1080 }
1081
1082 /* patch pointer addition in vtop so that pointer dereferencing is
1083    also tested */
1084 ST_FUNC void gen_bounded_ptr_deref(void)
1085 {
1086     int func;
1087     int size, align;
1088     Elf32_Rel *rel;
1089     Sym *sym;
1090
1091     size = 0;
1092     /* XXX: put that code in generic part of tcc */
1093     if (!is_float(vtop->type.t)) {
1094         if (vtop->r & VT_LVAL_BYTE)
1095             size = 1;
1096         else if (vtop->r & VT_LVAL_SHORT)
1097             size = 2;
1098     }
1099     if (!size)
1100         size = type_size(&vtop->type, &align);
1101     switch(size) {
1102     case  1: func = TOK___bound_ptr_indir1; break;
1103     case  2: func = TOK___bound_ptr_indir2; break;
1104     case  4: func = TOK___bound_ptr_indir4; break;
1105     case  8: func = TOK___bound_ptr_indir8; break;
1106     case 12: func = TOK___bound_ptr_indir12; break;
1107     case 16: func = TOK___bound_ptr_indir16; break;
1108     default:
1109         tcc_error("unhandled size when dereferencing bounded pointer");
1110         func = 0;
1111         break;
1112     }
1113
1114     /* patch relocation */
1115     /* XXX: find a better solution ? */
1116     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1117     sym = external_global_sym(func, &func_old_type, 0);
1118     if (!sym->c)
1119         put_extern_sym(sym, NULL, 0, 0);
1120     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1121 }
1122 #endif
1123
1124 /* Save the stack pointer onto the stack */
1125 ST_FUNC void gen_vla_sp_save(int addr) {
1126     /* mov %esp,addr(%ebp)*/
1127     o(0x89);
1128     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1129 }
1130
1131 /* Restore the SP from a location on the stack */
1132 ST_FUNC void gen_vla_sp_restore(int addr) {
1133     o(0x8b);
1134     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1135 }
1136
1137 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1138 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1139 #ifdef TCC_TARGET_PE
1140     /* alloca does more than just adjust %rsp on Windows */
1141     vpush_global_sym(&func_old_type, TOK_alloca);
1142     vswap(); /* Move alloca ref past allocation size */
1143     gfunc_call(1);
1144     vset(type, REG_IRET, 0);
1145 #else
1146     int r;
1147     r = gv(RC_INT); /* allocation size */
1148     /* sub r,%rsp */
1149     o(0x2b);
1150     o(0xe0 | r);
1151     /* We align to 16 bytes rather than align */
1152     /* and ~15, %esp */
1153     o(0xf0e483);
1154     /* mov %esp, r */
1155     o(0x89);
1156     o(0xe0 | r);
1157     vpop();
1158     vset(type, r, 0);
1159 #endif
1160 }
1161
1162 /* end of X86 code generator */
1163 /*************************************************************/
1164 #endif
1165 /*************************************************************/