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         5
  25 #define NB_ASM_REGS     8
  26 #define CONFIG_TCC_ASM
  27
  28 /* a register can belong to several classes. The classes must be
  29    sorted from more general to more precise (see gv2() code which does
  30    assumptions on it). */
  31 #define RC_INT     0x0001 /* generic integer register */
  32 #define RC_FLOAT   0x0002 /* generic float register */
  33 #define RC_EAX     0x0004
  34 #define RC_ST0     0x0008
  35 #define RC_ECX     0x0010
  36 #define RC_EDX     0x0020
  37 #define RC_EBX     0x0040
  38
  39 #define RC_IRET    RC_EAX /* function return: integer register */
  40 #define RC_IRE2    RC_EDX /* function return: second integer register */
  41 #define RC_FRET    RC_ST0 /* function return: float register */
  42
  43 /* pretty names for the registers */
  44 enum {
  45     TREG_EAX = 0,
  46     TREG_ECX,
  47     TREG_EDX,
  48     TREG_EBX,
  49     TREG_ST0,
  50     TREG_ESP = 4
  51 };
  52
  53 /* return registers for function */
  54 #define REG_IRET TREG_EAX /* single word int return register */
  55 #define REG_IRE2 TREG_EDX /* second word return register (for long long) */
  56 #define REG_FRET TREG_ST0 /* float return register */
  57
  58 /* defined if function parameters must be evaluated in reverse order */
  59 #define INVERT_FUNC_PARAMS
  60
  61 /* defined if structures are passed as pointers. Otherwise structures
  62    are directly pushed on stack. */
  63 /* #define FUNC_STRUCT_PARAM_AS_PTR */
  64
  65 /* pointer size, in bytes */
  66 #define PTR_SIZE 4
  67
  68 /* long double size and alignment, in bytes */
  69 #define LDOUBLE_SIZE  12
  70 #define LDOUBLE_ALIGN 4
  71 /* maximum alignment (for aligned attribute support) */
  72 #define MAX_ALIGN     8
  73
  74 /* define if return values need to be extended explicitely
  75    at caller side (for interfacing with non-TCC compilers) */
  76 #define PROMOTE_RET
  77
  78 /******************************************************/
  79 #else /* ! TARGET_DEFS_ONLY */
  80 /******************************************************/
  81 #define USING_GLOBALS
  82 #include "tcc.h"
  83
  84 /* define to 1/0 to [not] have EBX as 4th register */
  85 #define USE_EBX 0
  86
  87 ST_DATA const int reg_classes[NB_REGS] = {
  88     /* eax */ RC_INT | RC_EAX,
  89     /* ecx */ RC_INT | RC_ECX,
  90     /* edx */ RC_INT | RC_EDX,
  91     /* ebx */ (RC_INT | RC_EBX) * USE_EBX,
  92     /* st0 */ RC_FLOAT | RC_ST0,
  93 };
  94
  95 static unsigned long func_sub_sp_offset;
  96 static int func_ret_sub;
  97 #ifdef CONFIG_TCC_BCHECK
  98 static addr_t func_bound_offset;
  99 static unsigned long func_bound_ind;
 100 static int func_bound_alloca_used;
 101 static void gen_bounds_prolog(void);
 102 static void gen_bounds_epilog(void);
 103 #endif
 104
 105 /* XXX: make it faster ? */
 106 ST_FUNC void g(int c)
 107 {
 108     int ind1;
 109     if (nocode_wanted)
 110         return;
 111     ind1 = ind + 1;
 112     if (ind1 > cur_text_section->data_allocated)
 113         section_realloc(cur_text_section, ind1);
 114     cur_text_section->data[ind] = c;
 115     ind = ind1;
 116 }
 117
 118 ST_FUNC void o(unsigned int c)
 119 {
 120     while (c) {
 121         g(c);
 122         c = c >> 8;
 123     }
 124 }
 125
 126 ST_FUNC void gen_le16(int v)
 127 {
 128     g(v);
 129     g(v >> 8);
 130 }
 131
 132 ST_FUNC void gen_le32(int c)
 133 {
 134     g(c);
 135     g(c >> 8);
 136     g(c >> 16);
 137     g(c >> 24);
 138 }
 139
 140 /* output a symbol and patch all calls to it */
 141 ST_FUNC void gsym_addr(int t, int a)
 142 {
 143     while (t) {
 144         unsigned char *ptr = cur_text_section->data + t;
 145         uint32_t n = read32le(ptr); /* next value */
 146         write32le(ptr, a - t - 4);
 147         t = n;
 148     }
 149 }
 150
 151 /* instruction + 4 bytes data. Return the address of the data */
 152 static int oad(int c, int s)
 153 {
 154     int t;
 155     if (nocode_wanted)
 156         return s;
 157     o(c);
 158     t = ind;
 159     gen_le32(s);
 160     return t;
 161 }
 162
 163 ST_FUNC void gen_fill_nops(int bytes)
 164 {
 165     while (bytes--)
 166       g(0x90);
 167 }
 168
 169 /* generate jmp to a label */
 170 #define gjmp2(instr,lbl) oad(instr,lbl)
 171
 172 /* output constant with relocation if 'r & VT_SYM' is true */
 173 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
 174 {
 175     if (r & VT_SYM)
 176         greloc(cur_text_section, sym, ind, R_386_32);
 177     gen_le32(c);
 178 }
 179
 180 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 181 {
 182     if (r & VT_SYM)
 183         greloc(cur_text_section, sym, ind, R_386_PC32);
 184     gen_le32(c - 4);
 185 }
 186
 187 /* generate a modrm reference. 'op_reg' contains the additional 3
 188    opcode bits */
 189 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 190 {
 191     op_reg = op_reg << 3;
 192     if ((r & VT_VALMASK) == VT_CONST) {
 193         /* constant memory reference */
 194         o(0x05 | op_reg);
 195         gen_addr32(r, sym, c);
 196     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 197         /* currently, we use only ebp as base */
 198         if (c == (char)c) {
 199             /* short reference */
 200             o(0x45 | op_reg);
 201             g(c);
 202         } else {
 203             oad(0x85 | op_reg, c);
 204         }
 205     } else {
 206         g(0x00 | op_reg | (r & VT_VALMASK));
 207     }
 208 }
 209
 210 /* load 'r' from value 'sv' */
 211 ST_FUNC void load(int r, SValue *sv)
 212 {
 213     int v, t, ft, fc, fr;
 214     SValue v1;
 215
 216 #ifdef TCC_TARGET_PE
 217     SValue v2;
 218     sv = pe_getimport(sv, &v2);
 219 #endif
 220
 221     fr = sv->r;
 222     ft = sv->type.t & ~VT_DEFSIGN;
 223     fc = sv->c.i;
 224
 225     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 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.i = 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 || (ft & VT_TYPE) == VT_BOOL) {
 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             o(0x0f); /* setxx %br */
 273             o(fc);
 274             o(0xc0 + r);
 275             o(0xc0b60f + r * 0x90000); /* movzbl %al, %eax */
 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.i;
 301     fr = v->r & VT_VALMASK;
 302     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 303     bt = ft & VT_BTYPE;
 304     /* XXX: incorrect if float reg to reg */
 305     if (bt == VT_FLOAT) {
 306         o(0xd9); /* fsts */
 307         r = 2;
 308     } else if (bt == VT_DOUBLE) {
 309         o(0xdd); /* fstpl */
 310         r = 2;
 311     } else if (bt == VT_LDOUBLE) {
 312         o(0xc0d9); /* fld %st(0) */
 313         o(0xdb); /* fstpt */
 314         r = 7;
 315     } else {
 316         if (bt == VT_SHORT)
 317             o(0x66);
 318         if (bt == VT_BYTE || bt == VT_BOOL)
 319             o(0x88);
 320         else
 321             o(0x89);
 322     }
 323     if (fr == VT_CONST ||
 324         fr == VT_LOCAL ||
 325         (v->r & VT_LVAL)) {
 326         gen_modrm(r, v->r, v->sym, fc);
 327     } else if (fr != r) {
 328         o(0xc0 + fr + r * 8); /* mov r, fr */
 329     }
 330 }
 331
 332 static void gadd_sp(int val)
 333 {
 334     if (val == (char)val) {
 335         o(0xc483);
 336         g(val);
 337     } else {
 338         oad(0xc481, val); /* add $xxx, %esp */
 339     }
 340 }
 341
 342 #if defined CONFIG_TCC_BCHECK || defined TCC_TARGET_PE
 343 static void gen_static_call(int v)
 344 {
 345     Sym *sym;
 346
 347     sym = external_global_sym(v, &func_old_type);
 348     oad(0xe8, -4);
 349     greloc(cur_text_section, sym, ind-4, R_386_PC32);
 350 }
 351 #endif
 352
 353 /* 'is_jmp' is '1' if it is a jump */
 354 static void gcall_or_jmp(int is_jmp)
 355 {
 356     int r;
 357     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST && (vtop->r & VT_SYM)) {
 358         /* constant and relocation case */
 359         greloc(cur_text_section, vtop->sym, ind + 1, R_386_PC32);
 360         oad(0xe8 + is_jmp, vtop->c.i - 4); /* call/jmp im */
 361 #ifdef CONFIG_TCC_BCHECK
 362         if (tcc_state->do_bounds_check && vtop->sym->v == TOK_alloca)
 363             func_bound_alloca_used = 1;
 364 #endif
 365     } else {
 366         /* otherwise, indirect call */
 367         r = gv(RC_INT);
 368         o(0xff); /* call/jmp *r */
 369         o(0xd0 + r + (is_jmp << 4));
 370     }
 371 }
 372
 373 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 374 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 375
 376 /* Return the number of registers needed to return the struct, or 0 if
 377    returning via struct pointer. */
 378 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
 379 {
 380 #ifdef TCC_TARGET_PE
 381     int size, align;
 382     *ret_align = 1; // Never have to re-align return values for x86
 383     *regsize = 4;
 384     size = type_size(vt, &align);
 385     if (size > 8 || (size & (size - 1)))
 386         return 0;
 387     if (size == 8)
 388         ret->t = VT_LLONG;
 389     else if (size == 4)
 390         ret->t = VT_INT;
 391     else if (size == 2)
 392         ret->t = VT_SHORT;
 393     else
 394         ret->t = VT_BYTE;
 395     ret->ref = NULL;
 396     return 1;
 397 #else
 398     *ret_align = 1; // Never have to re-align return values for x86
 399     return 0;
 400 #endif
 401 }
 402
 403 /* Generate function call. The function address is pushed first, then
 404    all the parameters in call order. This functions pops all the
 405    parameters and the function address. */
 406 ST_FUNC void gfunc_call(int nb_args)
 407 {
 408     int size, align, r, args_size, i, func_call;
 409     Sym *func_sym;
 410
 411 #ifdef CONFIG_TCC_BCHECK
 412     if (tcc_state->do_bounds_check)
 413         gbound_args(nb_args);
 414 #endif
 415
 416     args_size = 0;
 417     for(i = 0;i < nb_args; i++) {
 418         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 419             size = type_size(&vtop->type, &align);
 420             /* align to stack align size */
 421             size = (size + 3) & ~3;
 422             /* allocate the necessary size on stack */
 423             oad(0xec81, size); /* sub $xxx, %esp */
 424             /* generate structure store */
 425             r = get_reg(RC_INT);
 426             o(0x89); /* mov %esp, r */
 427             o(0xe0 + r);
 428             vset(&vtop->type, r | VT_LVAL, 0);
 429             vswap();
 430             vstore();
 431             args_size += size;
 432         } else if (is_float(vtop->type.t)) {
 433             gv(RC_FLOAT); /* only one float register */
 434             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 435                 size = 4;
 436             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 437                 size = 8;
 438             else
 439                 size = 12;
 440             oad(0xec81, size); /* sub $xxx, %esp */
 441             if (size == 12)
 442                 o(0x7cdb);
 443             else
 444                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 445             g(0x24);
 446             g(0x00);
 447             args_size += size;
 448         } else {
 449             /* simple type (currently always same size) */
 450             /* XXX: implicit cast ? */
 451             r = gv(RC_INT);
 452             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 453                 size = 8;
 454                 o(0x50 + vtop->r2); /* push r */
 455             } else {
 456                 size = 4;
 457             }
 458             o(0x50 + r); /* push r */
 459             args_size += size;
 460         }
 461         vtop--;
 462     }
 463     save_regs(0); /* save used temporary registers */
 464     func_sym = vtop->type.ref;
 465     func_call = func_sym->f.func_call;
 466     /* fast call case */
 467     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 468         func_call == FUNC_FASTCALLW) {
 469         int fastcall_nb_regs;
 470         uint8_t *fastcall_regs_ptr;
 471         if (func_call == FUNC_FASTCALLW) {
 472             fastcall_regs_ptr = fastcallw_regs;
 473             fastcall_nb_regs = 2;
 474         } else {
 475             fastcall_regs_ptr = fastcall_regs;
 476             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 477         }
 478         for(i = 0;i < fastcall_nb_regs; i++) {
 479             if (args_size <= 0)
 480                 break;
 481             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 482             /* XXX: incorrect for struct/floats */
 483             args_size -= 4;
 484         }
 485     }
 486 #ifndef TCC_TARGET_PE
 487     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 488         args_size -= 4;
 489 #endif
 490
 491     gcall_or_jmp(0);
 492
 493     if (args_size && func_call != FUNC_STDCALL && func_call != FUNC_FASTCALLW)
 494         gadd_sp(args_size);
 495     vtop--;
 496 }
 497
 498 #ifdef TCC_TARGET_PE
 499 #define FUNC_PROLOG_SIZE (10 + USE_EBX)
 500 #else
 501 #define FUNC_PROLOG_SIZE (9 + USE_EBX)
 502 #endif
 503
 504 /* generate function prolog of type 't' */
 505 ST_FUNC void gfunc_prolog(Sym *func_sym)
 506 {
 507     CType *func_type = &func_sym->type;
 508     int addr, align, size, func_call, fastcall_nb_regs;
 509     int param_index, param_addr;
 510     uint8_t *fastcall_regs_ptr;
 511     Sym *sym;
 512     CType *type;
 513
 514     sym = func_type->ref;
 515     func_call = sym->f.func_call;
 516     addr = 8;
 517     loc = 0;
 518     func_vc = 0;
 519
 520     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 521         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 522         fastcall_regs_ptr = fastcall_regs;
 523     } else if (func_call == FUNC_FASTCALLW) {
 524         fastcall_nb_regs = 2;
 525         fastcall_regs_ptr = fastcallw_regs;
 526     } else {
 527         fastcall_nb_regs = 0;
 528         fastcall_regs_ptr = NULL;
 529     }
 530     param_index = 0;
 531
 532     ind += FUNC_PROLOG_SIZE;
 533     func_sub_sp_offset = ind;
 534     /* if the function returns a structure, then add an
 535        implicit pointer parameter */
 536 #ifdef TCC_TARGET_PE
 537     size = type_size(&func_vt,&align);
 538     if (((func_vt.t & VT_BTYPE) == VT_STRUCT)
 539         && (size > 8 || (size & (size - 1)))) {
 540 #else
 541     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 542 #endif
 543         /* XXX: fastcall case ? */
 544         func_vc = addr;
 545         addr += 4;
 546         param_index++;
 547     }
 548     /* define parameters */
 549     while ((sym = sym->next) != NULL) {
 550         type = &sym->type;
 551         size = type_size(type, &align);
 552         size = (size + 3) & ~3;
 553 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 554         /* structs are passed as pointer */
 555         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 556             size = 4;
 557         }
 558 #endif
 559         if (param_index < fastcall_nb_regs) {
 560             /* save FASTCALL register */
 561             loc -= 4;
 562             o(0x89);     /* movl */
 563             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 564             param_addr = loc;
 565         } else {
 566             param_addr = addr;
 567             addr += size;
 568         }
 569         sym_push(sym->v & ~SYM_FIELD, type,
 570                  VT_LOCAL | VT_LVAL, param_addr);
 571         param_index++;
 572     }
 573     func_ret_sub = 0;
 574     /* pascal type call or fastcall ? */
 575     if (func_call == FUNC_STDCALL || func_call == FUNC_FASTCALLW)
 576         func_ret_sub = addr - 8;
 577 #ifndef TCC_TARGET_PE
 578     else if (func_vc)
 579         func_ret_sub = 4;
 580 #endif
 581
 582 #ifdef CONFIG_TCC_BCHECK
 583     if (tcc_state->do_bounds_check)
 584         gen_bounds_prolog();
 585 #endif
 586 }
 587
 588 /* generate function epilog */
 589 ST_FUNC void gfunc_epilog(void)
 590 {
 591     addr_t v, saved_ind;
 592
 593 #ifdef CONFIG_TCC_BCHECK
 594     if (tcc_state->do_bounds_check)
 595         gen_bounds_epilog();
 596 #endif
 597
 598     /* align local size to word & save local variables */
 599     v = (-loc + 3) & -4;
 600
 601 #if USE_EBX
 602     o(0x8b);
 603     gen_modrm(TREG_EBX, VT_LOCAL, NULL, -(v+4));
 604 #endif
 605
 606     o(0xc9); /* leave */
 607     if (func_ret_sub == 0) {
 608         o(0xc3); /* ret */
 609     } else {
 610         o(0xc2); /* ret n */
 611         g(func_ret_sub);
 612         g(func_ret_sub >> 8);
 613     }
 614     saved_ind = ind;
 615     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 616 #ifdef TCC_TARGET_PE
 617     if (v >= 4096) {
 618         oad(0xb8, v); /* mov stacksize, %eax */
 619         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 620     } else
 621 #endif
 622     {
 623         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 624         o(0xec81);  /* sub esp, stacksize */
 625         gen_le32(v);
 626 #ifdef TCC_TARGET_PE
 627         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 628 #endif
 629     }
 630     o(0x53 * USE_EBX); /* push ebx */
 631     ind = saved_ind;
 632 }
 633
 634 /* generate a jump to a label */
 635 ST_FUNC int gjmp(int t)
 636 {
 637     return gjmp2(0xe9, t);
 638 }
 639
 640 /* generate a jump to a fixed address */
 641 ST_FUNC void gjmp_addr(int a)
 642 {
 643     int r;
 644     r = a - ind - 2;
 645     if (r == (char)r) {
 646         g(0xeb);
 647         g(r);
 648     } else {
 649         oad(0xe9, a - ind - 5);
 650     }
 651 }
 652
 653 #if 0
 654 /* generate a jump to a fixed address */
 655 ST_FUNC void gjmp_cond_addr(int a, int op)
 656 {
 657     int r = a - ind - 2;
 658     if (r == (char)r)
 659         g(op - 32), g(r);
 660     else
 661         g(0x0f), gjmp2(op - 16, r - 4);
 662 }
 663 #endif
 664
 665 ST_FUNC int gjmp_append(int n, int t)
 666 {
 667     void *p;
 668     /* insert vtop->c jump list in t */
 669     if (n) {
 670         uint32_t n1 = n, n2;
 671         while ((n2 = read32le(p = cur_text_section->data + n1)))
 672             n1 = n2;
 673         write32le(p, t);
 674         t = n;
 675     }
 676     return t;
 677 }
 678
 679 ST_FUNC int gjmp_cond(int op, int t)
 680 {
 681     g(0x0f);
 682     t = gjmp2(op - 16, t);
 683     return t;
 684 }
 685
 686 ST_FUNC void gen_opi(int op)
 687 {
 688     int r, fr, opc, c;
 689
 690     switch(op) {
 691     case '+':
 692     case TOK_ADDC1: /* add with carry generation */
 693         opc = 0;
 694     gen_op8:
 695         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 696             /* constant case */
 697             vswap();
 698             r = gv(RC_INT);
 699             vswap();
 700             c = vtop->c.i;
 701             if (c == (char)c) {
 702                 /* generate inc and dec for smaller code */
 703                 if ((c == 1 || c == -1) && (op == '+' || op == '-')) {
 704                     opc = (c == 1) ^ (op == '+');
 705                     o (0x40 | (opc << 3) | r); // inc,dec
 706                 } else {
 707                     o(0x83);
 708                     o(0xc0 | (opc << 3) | r);
 709                     g(c);
 710                 }
 711             } else {
 712                 o(0x81);
 713                 oad(0xc0 | (opc << 3) | r, c);
 714             }
 715         } else {
 716             gv2(RC_INT, RC_INT);
 717             r = vtop[-1].r;
 718             fr = vtop[0].r;
 719             o((opc << 3) | 0x01);
 720             o(0xc0 + r + fr * 8);
 721         }
 722         vtop--;
 723         if (op >= TOK_ULT && op <= TOK_GT)
 724             vset_VT_CMP(op);
 725         break;
 726     case '-':
 727     case TOK_SUBC1: /* sub with carry generation */
 728         opc = 5;
 729         goto gen_op8;
 730     case TOK_ADDC2: /* add with carry use */
 731         opc = 2;
 732         goto gen_op8;
 733     case TOK_SUBC2: /* sub with carry use */
 734         opc = 3;
 735         goto gen_op8;
 736     case '&':
 737         opc = 4;
 738         goto gen_op8;
 739     case '^':
 740         opc = 6;
 741         goto gen_op8;
 742     case '|':
 743         opc = 1;
 744         goto gen_op8;
 745     case '*':
 746         gv2(RC_INT, RC_INT);
 747         r = vtop[-1].r;
 748         fr = vtop[0].r;
 749         vtop--;
 750         o(0xaf0f); /* imul fr, r */
 751         o(0xc0 + fr + r * 8);
 752         break;
 753     case TOK_SHL:
 754         opc = 4;
 755         goto gen_shift;
 756     case TOK_SHR:
 757         opc = 5;
 758         goto gen_shift;
 759     case TOK_SAR:
 760         opc = 7;
 761     gen_shift:
 762         opc = 0xc0 | (opc << 3);
 763         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 764             /* constant case */
 765             vswap();
 766             r = gv(RC_INT);
 767             vswap();
 768             c = vtop->c.i & 0x1f;
 769             o(0xc1); /* shl/shr/sar $xxx, r */
 770             o(opc | r);
 771             g(c);
 772         } else {
 773             /* we generate the shift in ecx */
 774             gv2(RC_INT, RC_ECX);
 775             r = vtop[-1].r;
 776             o(0xd3); /* shl/shr/sar %cl, r */
 777             o(opc | r);
 778         }
 779         vtop--;
 780         break;
 781     case '/':
 782     case TOK_UDIV:
 783     case TOK_PDIV:
 784     case '%':
 785     case TOK_UMOD:
 786     case TOK_UMULL:
 787         /* first operand must be in eax */
 788         /* XXX: need better constraint for second operand */
 789         gv2(RC_EAX, RC_ECX);
 790         r = vtop[-1].r;
 791         fr = vtop[0].r;
 792         vtop--;
 793         save_reg(TREG_EDX);
 794         /* save EAX too if used otherwise */
 795         save_reg_upstack(TREG_EAX, 1);
 796         if (op == TOK_UMULL) {
 797             o(0xf7); /* mul fr */
 798             o(0xe0 + fr);
 799             vtop->r2 = TREG_EDX;
 800             r = TREG_EAX;
 801         } else {
 802             if (op == TOK_UDIV || op == TOK_UMOD) {
 803                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 804                 o(0xf0 + fr);
 805             } else {
 806                 o(0xf799); /* cltd, idiv fr, %eax */
 807                 o(0xf8 + fr);
 808             }
 809             if (op == '%' || op == TOK_UMOD)
 810                 r = TREG_EDX;
 811             else
 812                 r = TREG_EAX;
 813         }
 814         vtop->r = r;
 815         break;
 816     default:
 817         opc = 7;
 818         goto gen_op8;
 819     }
 820 }
 821
 822 /* generate a floating point operation 'v = t1 op t2' instruction. The
 823    two operands are guaranteed to have the same floating point type */
 824 /* XXX: need to use ST1 too */
 825 ST_FUNC void gen_opf(int op)
 826 {
 827     int a, ft, fc, swapped, r;
 828
 829     /* convert constants to memory references */
 830     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 831         vswap();
 832         gv(RC_FLOAT);
 833         vswap();
 834     }
 835     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 836         gv(RC_FLOAT);
 837
 838     /* must put at least one value in the floating point register */
 839     if ((vtop[-1].r & VT_LVAL) &&
 840         (vtop[0].r & VT_LVAL)) {
 841         vswap();
 842         gv(RC_FLOAT);
 843         vswap();
 844     }
 845     swapped = 0;
 846     /* swap the stack if needed so that t1 is the register and t2 is
 847        the memory reference */
 848     if (vtop[-1].r & VT_LVAL) {
 849         vswap();
 850         swapped = 1;
 851     }
 852     if (op >= TOK_ULT && op <= TOK_GT) {
 853         /* load on stack second operand */
 854         load(TREG_ST0, vtop);
 855         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 856         if (op == TOK_GE || op == TOK_GT)
 857             swapped = !swapped;
 858         else if (op == TOK_EQ || op == TOK_NE)
 859             swapped = 0;
 860         if (swapped)
 861             o(0xc9d9); /* fxch %st(1) */
 862         if (op == TOK_EQ || op == TOK_NE)
 863             o(0xe9da); /* fucompp */
 864         else
 865             o(0xd9de); /* fcompp */
 866         o(0xe0df); /* fnstsw %ax */
 867         if (op == TOK_EQ) {
 868             o(0x45e480); /* and $0x45, %ah */
 869             o(0x40fC80); /* cmp $0x40, %ah */
 870         } else if (op == TOK_NE) {
 871             o(0x45e480); /* and $0x45, %ah */
 872             o(0x40f480); /* xor $0x40, %ah */
 873             op = TOK_NE;
 874         } else if (op == TOK_GE || op == TOK_LE) {
 875             o(0x05c4f6); /* test $0x05, %ah */
 876             op = TOK_EQ;
 877         } else {
 878             o(0x45c4f6); /* test $0x45, %ah */
 879             op = TOK_EQ;
 880         }
 881         vtop--;
 882         vset_VT_CMP(op);
 883     } else {
 884         /* no memory reference possible for long double operations */
 885         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 886             load(TREG_ST0, vtop);
 887             swapped = !swapped;
 888         }
 889
 890         switch(op) {
 891         default:
 892         case '+':
 893             a = 0;
 894             break;
 895         case '-':
 896             a = 4;
 897             if (swapped)
 898                 a++;
 899             break;
 900         case '*':
 901             a = 1;
 902             break;
 903         case '/':
 904             a = 6;
 905             if (swapped)
 906                 a++;
 907             break;
 908         }
 909         ft = vtop->type.t;
 910         fc = vtop->c.i;
 911         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 912             o(0xde); /* fxxxp %st, %st(1) */
 913             o(0xc1 + (a << 3));
 914         } else {
 915             /* if saved lvalue, then we must reload it */
 916             r = vtop->r;
 917             if ((r & VT_VALMASK) == VT_LLOCAL) {
 918                 SValue v1;
 919                 r = get_reg(RC_INT);
 920                 v1.type.t = VT_INT;
 921                 v1.r = VT_LOCAL | VT_LVAL;
 922                 v1.c.i = fc;
 923                 load(r, &v1);
 924                 fc = 0;
 925             }
 926
 927             if ((ft & VT_BTYPE) == VT_DOUBLE)
 928                 o(0xdc);
 929             else
 930                 o(0xd8);
 931             gen_modrm(a, r, vtop->sym, fc);
 932         }
 933         vtop--;
 934     }
 935 }
 936
 937 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 938    and 'long long' cases. */
 939 ST_FUNC void gen_cvt_itof(int t)
 940 {
 941     save_reg(TREG_ST0);
 942     gv(RC_INT);
 943     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 944         /* signed long long to float/double/long double (unsigned case
 945            is handled generically) */
 946         o(0x50 + vtop->r2); /* push r2 */
 947         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 948         o(0x242cdf); /* fildll (%esp) */
 949         o(0x08c483); /* add $8, %esp */
 950         vtop->r2 = VT_CONST;
 951     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 952                (VT_INT | VT_UNSIGNED)) {
 953         /* unsigned int to float/double/long double */
 954         o(0x6a); /* push $0 */
 955         g(0x00);
 956         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 957         o(0x242cdf); /* fildll (%esp) */
 958         o(0x08c483); /* add $8, %esp */
 959     } else {
 960         /* int to float/double/long double */
 961         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 962         o(0x2404db); /* fildl (%esp) */
 963         o(0x04c483); /* add $4, %esp */
 964     }
 965     vtop->r2 = VT_CONST;
 966     vtop->r = TREG_ST0;
 967 }
 968
 969 /* convert fp to int 't' type */
 970 ST_FUNC void gen_cvt_ftoi(int t)
 971 {
 972     int bt = vtop->type.t & VT_BTYPE;
 973     if (bt == VT_FLOAT)
 974         vpush_global_sym(&func_old_type, TOK___fixsfdi);
 975     else if (bt == VT_LDOUBLE)
 976         vpush_global_sym(&func_old_type, TOK___fixxfdi);
 977     else
 978         vpush_global_sym(&func_old_type, TOK___fixdfdi);
 979     vswap();
 980     gfunc_call(1);
 981     vpushi(0);
 982     vtop->r = REG_IRET;
 983     if ((t & VT_BTYPE) == VT_LLONG)
 984         vtop->r2 = REG_IRE2;
 985 }
 986
 987 /* convert from one floating point type to another */
 988 ST_FUNC void gen_cvt_ftof(int t)
 989 {
 990     /* all we have to do on i386 is to put the float in a register */
 991     gv(RC_FLOAT);
 992 }
 993
 994 /* char/short to int conversion */
 995 ST_FUNC void gen_cvt_csti(int t)
 996 {
 997     int r, sz, xl;
 998     r = gv(RC_INT);
 999     sz = !(t & VT_UNSIGNED);
1000     xl = (t & VT_BTYPE) == VT_SHORT;
1001     o(0xc0b60f /* mov[sz] %a[xl], %eax */
1002         | (sz << 3 | xl) << 8
1003         | (r << 3 | r) << 16
1004         );
1005 }
1006
1007 /* computed goto support */
1008 ST_FUNC void ggoto(void)
1009 {
1010     gcall_or_jmp(1);
1011     vtop--;
1012 }
1013
1014 /* bound check support functions */
1015 #ifdef CONFIG_TCC_BCHECK
1016 /* generate a bounded pointer addition */
1017 ST_FUNC void gen_bounded_ptr_add(void)
1018 {
1019     vpush_global_sym(&func_old_type, TOK___bound_ptr_add);
1020     vrott(3);
1021     gfunc_call(2);
1022     vpushi(0);
1023     /* returned pointer is in eax */
1024     vtop->r = TREG_EAX | VT_BOUNDED;
1025     if (nocode_wanted)
1026         return;
1027     /* relocation offset of the bounding function call point */
1028     vtop->c.i = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1029 }
1030
1031 /* patch pointer addition in vtop so that pointer dereferencing is
1032    also tested */
1033 ST_FUNC void gen_bounded_ptr_deref(void)
1034 {
1035     addr_t func;
1036     int  size, align;
1037     Elf32_Rel *rel;
1038     Sym *sym;
1039
1040     if (nocode_wanted)
1041         return;
1042
1043     size = type_size(&vtop->type, &align);
1044     switch(size) {
1045     case  1: func = TOK___bound_ptr_indir1; break;
1046     case  2: func = TOK___bound_ptr_indir2; break;
1047     case  4: func = TOK___bound_ptr_indir4; break;
1048     case  8: func = TOK___bound_ptr_indir8; break;
1049     case 12: func = TOK___bound_ptr_indir12; break;
1050     case 16: func = TOK___bound_ptr_indir16; break;
1051     default:
1052         /* may happen with struct member access */
1053         return;
1054         //tcc_error("unhandled size when dereferencing bounded pointer");
1055         //func = 0;
1056         //break;
1057     }
1058     sym = external_global_sym(func, &func_old_type);
1059     if (!sym->c)
1060         put_extern_sym(sym, NULL, 0, 0);
1061     /* patch relocation */
1062     /* XXX: find a better solution ? */
1063     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.i);
1064     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1065 }
1066
1067 static void gen_bounds_prolog(void)
1068 {
1069     /* leave some room for bound checking code */
1070     func_bound_offset = lbounds_section->data_offset;
1071     func_bound_ind = ind;
1072     func_bound_alloca_used = 0;
1073     oad(0xb8, 0); /* lbound section pointer */
1074     oad(0xb8, 0); /* call to function */
1075 }
1076
1077 static void gen_bounds_epilog(void)
1078 {
1079     addr_t saved_ind;
1080     addr_t *bounds_ptr;
1081     Sym *sym_data;
1082
1083     if (func_bound_offset == lbounds_section->data_offset && !func_bound_alloca_used)
1084         return;
1085
1086     /* add end of table info */
1087     bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
1088     *bounds_ptr = 0;
1089
1090     /* generate bound local allocation */
1091     saved_ind = ind;
1092     ind = func_bound_ind;
1093     sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
1094                            func_bound_offset, lbounds_section->data_offset);
1095     greloc(cur_text_section, sym_data, ind + 1, R_386_32);
1096     ind = ind + 5;
1097     gen_static_call(TOK___bound_local_new);
1098     ind = saved_ind;
1099
1100     /* generate bound check local freeing */
1101     o(0x5250); /* save returned value, if any */
1102     greloc(cur_text_section, sym_data, ind + 1, R_386_32);
1103     oad(0xb8, 0); /* mov %eax, xxx */
1104     gen_static_call(TOK___bound_local_delete);
1105     o(0x585a); /* restore returned value, if any */
1106 }
1107 #endif
1108
1109 /* Save the stack pointer onto the stack */
1110 ST_FUNC void gen_vla_sp_save(int addr) {
1111     /* mov %esp,addr(%ebp)*/
1112     o(0x89);
1113     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1114 }
1115
1116 /* Restore the SP from a location on the stack */
1117 ST_FUNC void gen_vla_sp_restore(int addr) {
1118     o(0x8b);
1119     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1120 }
1121
1122 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1123 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1124     int use_call = 0;
1125
1126 #if defined(CONFIG_TCC_BCHECK)
1127     use_call = tcc_state->do_bounds_check;
1128 #endif
1129 #ifdef TCC_TARGET_PE    /* alloca does more than just adjust %rsp on Windows */
1130     use_call = 1;
1131 #endif
1132     if (use_call)
1133     {
1134         vpush_global_sym(&func_old_type, TOK_alloca);
1135         vswap(); /* Move alloca ref past allocation size */
1136         gfunc_call(1);
1137     }
1138     else {
1139         int r;
1140         r = gv(RC_INT); /* allocation size */
1141         /* sub r,%rsp */
1142         o(0x2b);
1143         o(0xe0 | r);
1144         /* We align to 16 bytes rather than align */
1145         /* and ~15, %esp */
1146         o(0xf0e483);
1147         vpop();
1148     }
1149 }
1150
1151 /* end of X86 code generator */
1152 /*************************************************************/
1153 #endif
1154 /*************************************************************/