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     *ret_align = 1; // Never have to re-align return values for x86
 371 #ifdef TCC_TARGET_PE
 372     int size, align;
 373     size = type_size(vt, &align);
 374     if (size > 8) {
 375         return 1;
 376     } else if (size > 4) {
 377         ret->ref = NULL;
 378         ret->t = VT_LLONG;
 379         return 0;
 380     } else {
 381         ret->ref = NULL;
 382         ret->t = VT_INT;
 383         return 0;
 384     }
 385 #else
 386     return 1;
 387 #endif
 388 }
 389
 390 /* Generate function call. The function address is pushed first, then
 391    all the parameters in call order. This functions pops all the
 392    parameters and the function address. */
 393 ST_FUNC void gfunc_call(int nb_args)
 394 {
 395     int size, align, r, args_size, i, func_call;
 396     Sym *func_sym;
 397
 398     args_size = 0;
 399     for(i = 0;i < nb_args; i++) {
 400         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 401             size = type_size(&vtop->type, &align);
 402             /* align to stack align size */
 403             size = (size + 3) & ~3;
 404             /* allocate the necessary size on stack */
 405             oad(0xec81, size); /* sub $xxx, %esp */
 406             /* generate structure store */
 407             r = get_reg(RC_INT);
 408             o(0x89); /* mov %esp, r */
 409             o(0xe0 + r);
 410             vset(&vtop->type, r | VT_LVAL, 0);
 411             vswap();
 412             vstore();
 413             args_size += size;
 414         } else if (is_float(vtop->type.t)) {
 415             gv(RC_FLOAT); /* only one float register */
 416             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 417                 size = 4;
 418             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 419                 size = 8;
 420             else
 421                 size = 12;
 422             oad(0xec81, size); /* sub $xxx, %esp */
 423             if (size == 12)
 424                 o(0x7cdb);
 425             else
 426                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 427             g(0x24);
 428             g(0x00);
 429             args_size += size;
 430         } else {
 431             /* simple type (currently always same size) */
 432             /* XXX: implicit cast ? */
 433             r = gv(RC_INT);
 434             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 435                 size = 8;
 436                 o(0x50 + vtop->r2); /* push r */
 437             } else {
 438                 size = 4;
 439             }
 440             o(0x50 + r); /* push r */
 441             args_size += size;
 442         }
 443         vtop--;
 444     }
 445     save_regs(0); /* save used temporary registers */
 446     func_sym = vtop->type.ref;
 447     func_call = FUNC_CALL(func_sym->r);
 448     /* fast call case */
 449     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 450         func_call == FUNC_FASTCALLW) {
 451         int fastcall_nb_regs;
 452         uint8_t *fastcall_regs_ptr;
 453         if (func_call == FUNC_FASTCALLW) {
 454             fastcall_regs_ptr = fastcallw_regs;
 455             fastcall_nb_regs = 2;
 456         } else {
 457             fastcall_regs_ptr = fastcall_regs;
 458             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 459         }
 460         for(i = 0;i < fastcall_nb_regs; i++) {
 461             if (args_size <= 0)
 462                 break;
 463             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 464             /* XXX: incorrect for struct/floats */
 465             args_size -= 4;
 466         }
 467     }
 468 #ifndef TCC_TARGET_PE
 469     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 470         args_size -= 4;
 471 #endif
 472     gcall_or_jmp(0);
 473
 474     if (args_size && func_call != FUNC_STDCALL)
 475         gadd_sp(args_size);
 476     vtop--;
 477 }
 478
 479 #ifdef TCC_TARGET_PE
 480 #define FUNC_PROLOG_SIZE 10
 481 #else
 482 #define FUNC_PROLOG_SIZE 9
 483 #endif
 484
 485 /* generate function prolog of type 't' */
 486 ST_FUNC void gfunc_prolog(CType *func_type)
 487 {
 488     int addr, align, size, func_call, fastcall_nb_regs;
 489     int param_index, param_addr;
 490     uint8_t *fastcall_regs_ptr;
 491     Sym *sym;
 492     CType *type;
 493
 494     sym = func_type->ref;
 495     func_call = FUNC_CALL(sym->r);
 496     addr = 8;
 497     loc = 0;
 498     func_vc = 0;
 499
 500     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 501         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 502         fastcall_regs_ptr = fastcall_regs;
 503     } else if (func_call == FUNC_FASTCALLW) {
 504         fastcall_nb_regs = 2;
 505         fastcall_regs_ptr = fastcallw_regs;
 506     } else {
 507         fastcall_nb_regs = 0;
 508         fastcall_regs_ptr = NULL;
 509     }
 510     param_index = 0;
 511
 512     ind += FUNC_PROLOG_SIZE;
 513     func_sub_sp_offset = ind;
 514     /* if the function returns a structure, then add an
 515        implicit pointer parameter */
 516     func_vt = sym->type;
 517 #ifdef TCC_TARGET_PE
 518     size = type_size(&func_vt,&align);
 519     if (((func_vt.t & VT_BTYPE) == VT_STRUCT) && (size > 8)) {
 520 #else
 521     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 522 #endif
 523         /* XXX: fastcall case ? */
 524         func_vc = addr;
 525         addr += 4;
 526         param_index++;
 527     }
 528     /* define parameters */
 529     while ((sym = sym->next) != NULL) {
 530         type = &sym->type;
 531         size = type_size(type, &align);
 532         size = (size + 3) & ~3;
 533 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 534         /* structs are passed as pointer */
 535         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 536             size = 4;
 537         }
 538 #endif
 539         if (param_index < fastcall_nb_regs) {
 540             /* save FASTCALL register */
 541             loc -= 4;
 542             o(0x89);     /* movl */
 543             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 544             param_addr = loc;
 545         } else {
 546             param_addr = addr;
 547             addr += size;
 548         }
 549         sym_push(sym->v & ~SYM_FIELD, type,
 550                  VT_LOCAL | lvalue_type(type->t), param_addr);
 551         param_index++;
 552     }
 553     func_ret_sub = 0;
 554     /* pascal type call ? */
 555     if (func_call == FUNC_STDCALL)
 556         func_ret_sub = addr - 8;
 557 #ifndef TCC_TARGET_PE
 558     else if (func_vc)
 559         func_ret_sub = 4;
 560 #endif
 561
 562 #ifdef CONFIG_TCC_BCHECK
 563     /* leave some room for bound checking code */
 564     if (tcc_state->do_bounds_check) {
 565         oad(0xb8, 0); /* lbound section pointer */
 566         oad(0xb8, 0); /* call to function */
 567         func_bound_offset = lbounds_section->data_offset;
 568     }
 569 #endif
 570 }
 571
 572 /* generate function epilog */
 573 ST_FUNC void gfunc_epilog(void)
 574 {
 575     int v, saved_ind;
 576
 577 #ifdef CONFIG_TCC_BCHECK
 578     if (tcc_state->do_bounds_check
 579      && func_bound_offset != lbounds_section->data_offset) {
 580         int saved_ind;
 581         int *bounds_ptr;
 582         Sym *sym, *sym_data;
 583         /* add end of table info */
 584         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 585         *bounds_ptr = 0;
 586         /* generate bound local allocation */
 587         saved_ind = ind;
 588         ind = func_sub_sp_offset;
 589         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 590                                func_bound_offset, lbounds_section->data_offset);
 591         greloc(cur_text_section, sym_data,
 592                ind + 1, R_386_32);
 593         oad(0xb8, 0); /* mov %eax, xxx */
 594         sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
 595         greloc(cur_text_section, sym,
 596                ind + 1, R_386_PC32);
 597         oad(0xe8, -4);
 598         ind = saved_ind;
 599         /* generate bound check local freeing */
 600         o(0x5250); /* save returned value, if any */
 601         greloc(cur_text_section, sym_data,
 602                ind + 1, R_386_32);
 603         oad(0xb8, 0); /* mov %eax, xxx */
 604         sym = external_global_sym(TOK___bound_local_delete, &func_old_type, 0);
 605         greloc(cur_text_section, sym,
 606                ind + 1, R_386_PC32);
 607         oad(0xe8, -4);
 608         o(0x585a); /* restore returned value, if any */
 609     }
 610 #endif
 611     o(0xc9); /* leave */
 612     if (func_ret_sub == 0) {
 613         o(0xc3); /* ret */
 614     } else {
 615         o(0xc2); /* ret n */
 616         g(func_ret_sub);
 617         g(func_ret_sub >> 8);
 618     }
 619     /* align local size to word & save local variables */
 620
 621     v = (-loc + 3) & -4;
 622     saved_ind = ind;
 623     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 624 #ifdef TCC_TARGET_PE
 625     if (v >= 4096) {
 626         Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
 627         oad(0xb8, v); /* mov stacksize, %eax */
 628         oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
 629         greloc(cur_text_section, sym, ind-4, R_386_PC32);
 630     } else
 631 #endif
 632     {
 633         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 634         o(0xec81);  /* sub esp, stacksize */
 635         gen_le32(v);
 636 #if FUNC_PROLOG_SIZE == 10
 637         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 638 #endif
 639     }
 640     ind = saved_ind;
 641 }
 642
 643 /* generate a jump to a label */
 644 ST_FUNC int gjmp(int t)
 645 {
 646     return psym(0xe9, t);
 647 }
 648
 649 /* generate a jump to a fixed address */
 650 ST_FUNC void gjmp_addr(int a)
 651 {
 652     int r;
 653     r = a - ind - 2;
 654     if (r == (char)r) {
 655         g(0xeb);
 656         g(r);
 657     } else {
 658         oad(0xe9, a - ind - 5);
 659     }
 660 }
 661
 662 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 663 ST_FUNC int gtst(int inv, int t)
 664 {
 665     int v, *p;
 666
 667     v = vtop->r & VT_VALMASK;
 668     if (v == VT_CMP) {
 669         /* fast case : can jump directly since flags are set */
 670         g(0x0f);
 671         t = psym((vtop->c.i - 16) ^ inv, t);
 672     } else if (v == VT_JMP || v == VT_JMPI) {
 673         /* && or || optimization */
 674         if ((v & 1) == inv) {
 675             /* insert vtop->c jump list in t */
 676             p = &vtop->c.i;
 677             while (*p != 0)
 678                 p = (int *)(cur_text_section->data + *p);
 679             *p = t;
 680             t = vtop->c.i;
 681         } else {
 682             t = gjmp(t);
 683             gsym(vtop->c.i);
 684         }
 685     } else {
 686         if (is_float(vtop->type.t) ||
 687             (vtop->type.t & VT_BTYPE) == VT_LLONG) {
 688             vpushi(0);
 689             gen_op(TOK_NE);
 690         }
 691         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 692             /* constant jmp optimization */
 693             if ((vtop->c.i != 0) != inv)
 694                 t = gjmp(t);
 695         } else {
 696             v = gv(RC_INT);
 697             o(0x85);
 698             o(0xc0 + v * 9);
 699             g(0x0f);
 700             t = psym(0x85 ^ inv, t);
 701         }
 702     }
 703     vtop--;
 704     return t;
 705 }
 706
 707 /* generate an integer binary operation */
 708 ST_FUNC void gen_opi(int op)
 709 {
 710     int r, fr, opc, c;
 711
 712     switch(op) {
 713     case '+':
 714     case TOK_ADDC1: /* add with carry generation */
 715         opc = 0;
 716     gen_op8:
 717         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 718             /* constant case */
 719             vswap();
 720             r = gv(RC_INT);
 721             vswap();
 722             c = vtop->c.i;
 723             if (c == (char)c) {
 724                 /* generate inc and dec for smaller code */
 725                 if (c==1 && opc==0) {
 726                     o (0x40 | r); // inc
 727                 } else if (c==1 && opc==5) {
 728                     o (0x48 | r); // dec
 729                 } else {
 730                     o(0x83);
 731                     o(0xc0 | (opc << 3) | r);
 732                     g(c);
 733                 }
 734             } else {
 735                 o(0x81);
 736                 oad(0xc0 | (opc << 3) | r, c);
 737             }
 738         } else {
 739             gv2(RC_INT, RC_INT);
 740             r = vtop[-1].r;
 741             fr = vtop[0].r;
 742             o((opc << 3) | 0x01);
 743             o(0xc0 + r + fr * 8);
 744         }
 745         vtop--;
 746         if (op >= TOK_ULT && op <= TOK_GT) {
 747             vtop->r = VT_CMP;
 748             vtop->c.i = op;
 749         }
 750         break;
 751     case '-':
 752     case TOK_SUBC1: /* sub with carry generation */
 753         opc = 5;
 754         goto gen_op8;
 755     case TOK_ADDC2: /* add with carry use */
 756         opc = 2;
 757         goto gen_op8;
 758     case TOK_SUBC2: /* sub with carry use */
 759         opc = 3;
 760         goto gen_op8;
 761     case '&':
 762         opc = 4;
 763         goto gen_op8;
 764     case '^':
 765         opc = 6;
 766         goto gen_op8;
 767     case '|':
 768         opc = 1;
 769         goto gen_op8;
 770     case '*':
 771         gv2(RC_INT, RC_INT);
 772         r = vtop[-1].r;
 773         fr = vtop[0].r;
 774         vtop--;
 775         o(0xaf0f); /* imul fr, r */
 776         o(0xc0 + fr + r * 8);
 777         break;
 778     case TOK_SHL:
 779         opc = 4;
 780         goto gen_shift;
 781     case TOK_SHR:
 782         opc = 5;
 783         goto gen_shift;
 784     case TOK_SAR:
 785         opc = 7;
 786     gen_shift:
 787         opc = 0xc0 | (opc << 3);
 788         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 789             /* constant case */
 790             vswap();
 791             r = gv(RC_INT);
 792             vswap();
 793             c = vtop->c.i & 0x1f;
 794             o(0xc1); /* shl/shr/sar $xxx, r */
 795             o(opc | r);
 796             g(c);
 797         } else {
 798             /* we generate the shift in ecx */
 799             gv2(RC_INT, RC_ECX);
 800             r = vtop[-1].r;
 801             o(0xd3); /* shl/shr/sar %cl, r */
 802             o(opc | r);
 803         }
 804         vtop--;
 805         break;
 806     case '/':
 807     case TOK_UDIV:
 808     case TOK_PDIV:
 809     case '%':
 810     case TOK_UMOD:
 811     case TOK_UMULL:
 812         /* first operand must be in eax */
 813         /* XXX: need better constraint for second operand */
 814         gv2(RC_EAX, RC_ECX);
 815         r = vtop[-1].r;
 816         fr = vtop[0].r;
 817         vtop--;
 818         save_reg(TREG_EDX);
 819         if (op == TOK_UMULL) {
 820             o(0xf7); /* mul fr */
 821             o(0xe0 + fr);
 822             vtop->r2 = TREG_EDX;
 823             r = TREG_EAX;
 824         } else {
 825             if (op == TOK_UDIV || op == TOK_UMOD) {
 826                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 827                 o(0xf0 + fr);
 828             } else {
 829                 o(0xf799); /* cltd, idiv fr, %eax */
 830                 o(0xf8 + fr);
 831             }
 832             if (op == '%' || op == TOK_UMOD)
 833                 r = TREG_EDX;
 834             else
 835                 r = TREG_EAX;
 836         }
 837         vtop->r = r;
 838         break;
 839     default:
 840         opc = 7;
 841         goto gen_op8;
 842     }
 843 }
 844
 845 /* generate a floating point operation 'v = t1 op t2' instruction. The
 846    two operands are guaranted to have the same floating point type */
 847 /* XXX: need to use ST1 too */
 848 ST_FUNC void gen_opf(int op)
 849 {
 850     int a, ft, fc, swapped, r;
 851
 852     /* convert constants to memory references */
 853     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 854         vswap();
 855         gv(RC_FLOAT);
 856         vswap();
 857     }
 858     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 859         gv(RC_FLOAT);
 860
 861     /* must put at least one value in the floating point register */
 862     if ((vtop[-1].r & VT_LVAL) &&
 863         (vtop[0].r & VT_LVAL)) {
 864         vswap();
 865         gv(RC_FLOAT);
 866         vswap();
 867     }
 868     swapped = 0;
 869     /* swap the stack if needed so that t1 is the register and t2 is
 870        the memory reference */
 871     if (vtop[-1].r & VT_LVAL) {
 872         vswap();
 873         swapped = 1;
 874     }
 875     if (op >= TOK_ULT && op <= TOK_GT) {
 876         /* load on stack second operand */
 877         load(TREG_ST0, vtop);
 878         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 879         if (op == TOK_GE || op == TOK_GT)
 880             swapped = !swapped;
 881         else if (op == TOK_EQ || op == TOK_NE)
 882             swapped = 0;
 883         if (swapped)
 884             o(0xc9d9); /* fxch %st(1) */
 885         o(0xe9da); /* fucompp */
 886         o(0xe0df); /* fnstsw %ax */
 887         if (op == TOK_EQ) {
 888             o(0x45e480); /* and $0x45, %ah */
 889             o(0x40fC80); /* cmp $0x40, %ah */
 890         } else if (op == TOK_NE) {
 891             o(0x45e480); /* and $0x45, %ah */
 892             o(0x40f480); /* xor $0x40, %ah */
 893             op = TOK_NE;
 894         } else if (op == TOK_GE || op == TOK_LE) {
 895             o(0x05c4f6); /* test $0x05, %ah */
 896             op = TOK_EQ;
 897         } else {
 898             o(0x45c4f6); /* test $0x45, %ah */
 899             op = TOK_EQ;
 900         }
 901         vtop--;
 902         vtop->r = VT_CMP;
 903         vtop->c.i = op;
 904     } else {
 905         /* no memory reference possible for long double operations */
 906         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 907             load(TREG_ST0, vtop);
 908             swapped = !swapped;
 909         }
 910
 911         switch(op) {
 912         default:
 913         case '+':
 914             a = 0;
 915             break;
 916         case '-':
 917             a = 4;
 918             if (swapped)
 919                 a++;
 920             break;
 921         case '*':
 922             a = 1;
 923             break;
 924         case '/':
 925             a = 6;
 926             if (swapped)
 927                 a++;
 928             break;
 929         }
 930         ft = vtop->type.t;
 931         fc = vtop->c.ul;
 932         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 933             o(0xde); /* fxxxp %st, %st(1) */
 934             o(0xc1 + (a << 3));
 935         } else {
 936             /* if saved lvalue, then we must reload it */
 937             r = vtop->r;
 938             if ((r & VT_VALMASK) == VT_LLOCAL) {
 939                 SValue v1;
 940                 r = get_reg(RC_INT);
 941                 v1.type.t = VT_INT;
 942                 v1.r = VT_LOCAL | VT_LVAL;
 943                 v1.c.ul = fc;
 944                 load(r, &v1);
 945                 fc = 0;
 946             }
 947
 948             if ((ft & VT_BTYPE) == VT_DOUBLE)
 949                 o(0xdc);
 950             else
 951                 o(0xd8);
 952             gen_modrm(a, r, vtop->sym, fc);
 953         }
 954         vtop--;
 955     }
 956 }
 957
 958 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 959    and 'long long' cases. */
 960 ST_FUNC void gen_cvt_itof(int t)
 961 {
 962     save_reg(TREG_ST0);
 963     gv(RC_INT);
 964     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 965         /* signed long long to float/double/long double (unsigned case
 966            is handled generically) */
 967         o(0x50 + vtop->r2); /* push r2 */
 968         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 969         o(0x242cdf); /* fildll (%esp) */
 970         o(0x08c483); /* add $8, %esp */
 971     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 972                (VT_INT | VT_UNSIGNED)) {
 973         /* unsigned int to float/double/long double */
 974         o(0x6a); /* push $0 */
 975         g(0x00);
 976         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 977         o(0x242cdf); /* fildll (%esp) */
 978         o(0x08c483); /* add $8, %esp */
 979     } else {
 980         /* int to float/double/long double */
 981         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 982         o(0x2404db); /* fildl (%esp) */
 983         o(0x04c483); /* add $4, %esp */
 984     }
 985     vtop->r = TREG_ST0;
 986 }
 987
 988 /* convert fp to int 't' type */
 989 /* XXX: handle long long case */
 990 ST_FUNC void gen_cvt_ftoi(int t)
 991 {
 992     int r, r2, size;
 993     Sym *sym;
 994     CType ushort_type;
 995
 996     ushort_type.t = VT_SHORT | VT_UNSIGNED;
 997     ushort_type.ref = 0;
 998
 999     gv(RC_FLOAT);
1000     if (t != VT_INT)
1001         size = 8;
1002     else
1003         size = 4;
1004
1005     o(0x2dd9); /* ldcw xxx */
1006     sym = external_global_sym(TOK___tcc_int_fpu_control,
1007                               &ushort_type, VT_LVAL);
1008     greloc(cur_text_section, sym,
1009            ind, R_386_32);
1010     gen_le32(0);
1011
1012     oad(0xec81, size); /* sub $xxx, %esp */
1013     if (size == 4)
1014         o(0x1cdb); /* fistpl */
1015     else
1016         o(0x3cdf); /* fistpll */
1017     o(0x24);
1018     o(0x2dd9); /* ldcw xxx */
1019     sym = external_global_sym(TOK___tcc_fpu_control,
1020                               &ushort_type, VT_LVAL);
1021     greloc(cur_text_section, sym,
1022            ind, R_386_32);
1023     gen_le32(0);
1024
1025     r = get_reg(RC_INT);
1026     o(0x58 + r); /* pop r */
1027     if (size == 8) {
1028         if (t == VT_LLONG) {
1029             vtop->r = r; /* mark reg as used */
1030             r2 = get_reg(RC_INT);
1031             o(0x58 + r2); /* pop r2 */
1032             vtop->r2 = r2;
1033         } else {
1034             o(0x04c483); /* add $4, %esp */
1035         }
1036     }
1037     vtop->r = r;
1038 }
1039
1040 /* convert from one floating point type to another */
1041 ST_FUNC void gen_cvt_ftof(int t)
1042 {
1043     /* all we have to do on i386 is to put the float in a register */
1044     gv(RC_FLOAT);
1045 }
1046
1047 /* computed goto support */
1048 ST_FUNC void ggoto(void)
1049 {
1050     gcall_or_jmp(1);
1051     vtop--;
1052 }
1053
1054 /* bound check support functions */
1055 #ifdef CONFIG_TCC_BCHECK
1056
1057 /* generate a bounded pointer addition */
1058 ST_FUNC void gen_bounded_ptr_add(void)
1059 {
1060     Sym *sym;
1061
1062     /* prepare fast i386 function call (args in eax and edx) */
1063     gv2(RC_EAX, RC_EDX);
1064     /* save all temporary registers */
1065     vtop -= 2;
1066     save_regs(0);
1067     /* do a fast function call */
1068     sym = external_global_sym(TOK___bound_ptr_add, &func_old_type, 0);
1069     greloc(cur_text_section, sym,
1070            ind + 1, R_386_PC32);
1071     oad(0xe8, -4);
1072     /* returned pointer is in eax */
1073     vtop++;
1074     vtop->r = TREG_EAX | VT_BOUNDED;
1075     /* address of bounding function call point */
1076     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1077 }
1078
1079 /* patch pointer addition in vtop so that pointer dereferencing is
1080    also tested */
1081 ST_FUNC void gen_bounded_ptr_deref(void)
1082 {
1083     int func;
1084     int size, align;
1085     Elf32_Rel *rel;
1086     Sym *sym;
1087
1088     size = 0;
1089     /* XXX: put that code in generic part of tcc */
1090     if (!is_float(vtop->type.t)) {
1091         if (vtop->r & VT_LVAL_BYTE)
1092             size = 1;
1093         else if (vtop->r & VT_LVAL_SHORT)
1094             size = 2;
1095     }
1096     if (!size)
1097         size = type_size(&vtop->type, &align);
1098     switch(size) {
1099     case  1: func = TOK___bound_ptr_indir1; break;
1100     case  2: func = TOK___bound_ptr_indir2; break;
1101     case  4: func = TOK___bound_ptr_indir4; break;
1102     case  8: func = TOK___bound_ptr_indir8; break;
1103     case 12: func = TOK___bound_ptr_indir12; break;
1104     case 16: func = TOK___bound_ptr_indir16; break;
1105     default:
1106         tcc_error("unhandled size when dereferencing bounded pointer");
1107         func = 0;
1108         break;
1109     }
1110
1111     /* patch relocation */
1112     /* XXX: find a better solution ? */
1113     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1114     sym = external_global_sym(func, &func_old_type, 0);
1115     if (!sym->c)
1116         put_extern_sym(sym, NULL, 0, 0);
1117     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1118 }
1119 #endif
1120
1121 /* Save the stack pointer onto the stack */
1122 ST_FUNC void gen_vla_sp_save(int addr) {
1123     /* mov %esp,addr(%ebp)*/
1124     o(0x89);
1125     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1126 }
1127
1128 /* Restore the SP from a location on the stack */
1129 ST_FUNC void gen_vla_sp_restore(int addr) {
1130     o(0x8b);
1131     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1132 }
1133
1134 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1135 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1136 #ifdef TCC_TARGET_PE
1137     /* alloca does more than just adjust %rsp on Windows */
1138     vpush_global_sym(&func_old_type, TOK_alloca);
1139     vswap(); /* Move alloca ref past allocation size */
1140     gfunc_call(1);
1141     vset(type, REG_IRET, 0);
1142 #else
1143     int r;
1144     r = gv(RC_INT); /* allocation size */
1145     /* sub r,%rsp */
1146     o(0x2b);
1147     o(0xe0 | r);
1148     /* We align to 16 bytes rather than align */
1149     /* and ~15, %esp */
1150     o(0xf0e483);
1151     /* mov %esp, r */
1152     o(0x89);
1153     o(0xe0 | r);
1154     vpop();
1155     vset(type, r, 0);
1156 #endif
1157 }
1158
1159 /* end of X86 code generator */
1160 /*************************************************************/
1161 #endif
1162 /*************************************************************/