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 ST_DATA int func_bound_add_epilog;
 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             v1.sym = NULL;
 234             fr = r;
 235             if (!(reg_classes[fr] & RC_INT))
 236                 fr = get_reg(RC_INT);
 237             load(fr, &v1);
 238         }
 239         if ((ft & VT_BTYPE) == VT_FLOAT) {
 240             o(0xd9); /* flds */
 241             r = 0;
 242         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 243             o(0xdd); /* fldl */
 244             r = 0;
 245         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 246             o(0xdb); /* fldt */
 247             r = 5;
 248         } else if ((ft & VT_TYPE) == VT_BYTE || (ft & VT_TYPE) == VT_BOOL) {
 249             o(0xbe0f);   /* movsbl */
 250         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 251             o(0xb60f);   /* movzbl */
 252         } else if ((ft & VT_TYPE) == VT_SHORT) {
 253             o(0xbf0f);   /* movswl */
 254         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 255             o(0xb70f);   /* movzwl */
 256         } else {
 257             o(0x8b);     /* movl */
 258         }
 259         gen_modrm(r, fr, sv->sym, fc);
 260     } else {
 261         if (v == VT_CONST) {
 262             o(0xb8 + r); /* mov $xx, r */
 263             gen_addr32(fr, sv->sym, fc);
 264         } else if (v == VT_LOCAL) {
 265             if (fc) {
 266                 o(0x8d); /* lea xxx(%ebp), r */
 267                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 268             } else {
 269                 o(0x89);
 270                 o(0xe8 + r); /* mov %ebp, r */
 271             }
 272         } else if (v == VT_CMP) {
 273             o(0x0f); /* setxx %br */
 274             o(fc);
 275             o(0xc0 + r);
 276             o(0xc0b60f + r * 0x90000); /* movzbl %al, %eax */
 277         } else if (v == VT_JMP || v == VT_JMPI) {
 278             t = v & 1;
 279             oad(0xb8 + r, t); /* mov $1, r */
 280             o(0x05eb); /* jmp after */
 281             gsym(fc);
 282             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 283         } else if (v != r) {
 284             o(0x89);
 285             o(0xc0 + r + v * 8); /* mov v, r */
 286         }
 287     }
 288 }
 289
 290 /* store register 'r' in lvalue 'v' */
 291 ST_FUNC void store(int r, SValue *v)
 292 {
 293     int fr, bt, ft, fc;
 294
 295 #ifdef TCC_TARGET_PE
 296     SValue v2;
 297     v = pe_getimport(v, &v2);
 298 #endif
 299
 300     ft = v->type.t;
 301     fc = v->c.i;
 302     fr = v->r & VT_VALMASK;
 303     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 304     bt = ft & VT_BTYPE;
 305     /* XXX: incorrect if float reg to reg */
 306     if (bt == VT_FLOAT) {
 307         o(0xd9); /* fsts */
 308         r = 2;
 309     } else if (bt == VT_DOUBLE) {
 310         o(0xdd); /* fstpl */
 311         r = 2;
 312     } else if (bt == VT_LDOUBLE) {
 313         o(0xc0d9); /* fld %st(0) */
 314         o(0xdb); /* fstpt */
 315         r = 7;
 316     } else {
 317         if (bt == VT_SHORT)
 318             o(0x66);
 319         if (bt == VT_BYTE || bt == VT_BOOL)
 320             o(0x88);
 321         else
 322             o(0x89);
 323     }
 324     if (fr == VT_CONST ||
 325         fr == VT_LOCAL ||
 326         (v->r & VT_LVAL)) {
 327         gen_modrm(r, v->r, v->sym, fc);
 328     } else if (fr != r) {
 329         o(0xc0 + fr + r * 8); /* mov r, fr */
 330     }
 331 }
 332
 333 static void gadd_sp(int val)
 334 {
 335     if (val == (char)val) {
 336         o(0xc483);
 337         g(val);
 338     } else {
 339         oad(0xc481, val); /* add $xxx, %esp */
 340     }
 341 }
 342
 343 #if defined CONFIG_TCC_BCHECK || defined TCC_TARGET_PE
 344 static void gen_static_call(int v)
 345 {
 346     Sym *sym;
 347
 348     sym = external_helper_sym(v);
 349     oad(0xe8, -4);
 350     greloc(cur_text_section, sym, ind-4, R_386_PC32);
 351 }
 352 #endif
 353
 354 /* 'is_jmp' is '1' if it is a jump */
 355 static void gcall_or_jmp(int is_jmp)
 356 {
 357     int r;
 358     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST && (vtop->r & VT_SYM)) {
 359         /* constant and relocation case */
 360         greloc(cur_text_section, vtop->sym, ind + 1, R_386_PC32);
 361         oad(0xe8 + is_jmp, vtop->c.i - 4); /* call/jmp im */
 362     } else {
 363         /* otherwise, indirect call */
 364         r = gv(RC_INT);
 365         o(0xff); /* call/jmp *r */
 366         o(0xd0 + r + (is_jmp << 4));
 367     }
 368 }
 369
 370 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 371 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 372
 373 /* Return the number of registers needed to return the struct, or 0 if
 374    returning via struct pointer. */
 375 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
 376 {
 377 #ifdef TCC_TARGET_PE
 378     int size, align;
 379     *ret_align = 1; // Never have to re-align return values for x86
 380     *regsize = 4;
 381     size = type_size(vt, &align);
 382     if (size > 8 || (size & (size - 1)))
 383         return 0;
 384     if (size == 8)
 385         ret->t = VT_LLONG;
 386     else if (size == 4)
 387         ret->t = VT_INT;
 388     else if (size == 2)
 389         ret->t = VT_SHORT;
 390     else
 391         ret->t = VT_BYTE;
 392     ret->ref = NULL;
 393     return 1;
 394 #else
 395     *ret_align = 1; // Never have to re-align return values for x86
 396     return 0;
 397 #endif
 398 }
 399
 400 /* Generate function call. The function address is pushed first, then
 401    all the parameters in call order. This functions pops all the
 402    parameters and the function address. */
 403 ST_FUNC void gfunc_call(int nb_args)
 404 {
 405     int size, align, r, args_size, i, func_call;
 406     Sym *func_sym;
 407
 408 #ifdef CONFIG_TCC_BCHECK
 409     if (tcc_state->do_bounds_check)
 410         gbound_args(nb_args);
 411 #endif
 412
 413     args_size = 0;
 414     for(i = 0;i < nb_args; i++) {
 415         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 416             size = type_size(&vtop->type, &align);
 417             /* align to stack align size */
 418             size = (size + 3) & ~3;
 419             /* allocate the necessary size on stack */
 420 #ifdef TCC_TARGET_PE
 421             if (size >= 0x4096) {
 422                 /* cannot call alloca with bound checking. Do stack probing. */
 423                 o(0x50);               // push %eax
 424                 oad(0xb8, size - 4);   // mov size-4,%eax
 425                 oad(0x3d, 4096);       // p1: cmp $4096,%eax
 426                 o(0x1476);             // jbe <p2>
 427                 oad(0x248485,-4096);   // test %eax,-4096(%esp)
 428                 oad(0xec81, 4096);     // sub $4096,%esp
 429                 oad(0x2d, 4096);       // sub $4096,%eax
 430                 o(0xe5eb);             // jmp <p1>
 431                 o(0xc429);             // p2: sub %eax,%esp
 432                 oad(0xc481, size - 4); // add size-4,%esp
 433                 o(0x58);               // pop %eax
 434             }
 435 #endif
 436             oad(0xec81, size); /* sub $xxx, %esp */
 437             /* generate structure store */
 438             r = get_reg(RC_INT);
 439             o(0x89); /* mov %esp, r */
 440             o(0xe0 + r);
 441             vset(&vtop->type, r | VT_LVAL, 0);
 442             vswap();
 443             vstore();
 444             args_size += size;
 445         } else if (is_float(vtop->type.t)) {
 446             gv(RC_FLOAT); /* only one float register */
 447             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 448                 size = 4;
 449             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 450                 size = 8;
 451             else
 452                 size = 12;
 453             oad(0xec81, size); /* sub $xxx, %esp */
 454             if (size == 12)
 455                 o(0x7cdb);
 456             else
 457                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 458             g(0x24);
 459             g(0x00);
 460             args_size += size;
 461         } else {
 462             /* simple type (currently always same size) */
 463             /* XXX: implicit cast ? */
 464             r = gv(RC_INT);
 465             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 466                 size = 8;
 467                 o(0x50 + vtop->r2); /* push r */
 468             } else {
 469                 size = 4;
 470             }
 471             o(0x50 + r); /* push r */
 472             args_size += size;
 473         }
 474         vtop--;
 475     }
 476     save_regs(0); /* save used temporary registers */
 477     func_sym = vtop->type.ref;
 478     func_call = func_sym->f.func_call;
 479     /* fast call case */
 480     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 481         func_call == FUNC_FASTCALLW) {
 482         int fastcall_nb_regs;
 483         uint8_t *fastcall_regs_ptr;
 484         if (func_call == FUNC_FASTCALLW) {
 485             fastcall_regs_ptr = fastcallw_regs;
 486             fastcall_nb_regs = 2;
 487         } else {
 488             fastcall_regs_ptr = fastcall_regs;
 489             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 490         }
 491         for(i = 0;i < fastcall_nb_regs; i++) {
 492             if (args_size <= 0)
 493                 break;
 494             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 495             /* XXX: incorrect for struct/floats */
 496             args_size -= 4;
 497         }
 498     }
 499 #ifndef TCC_TARGET_PE
 500     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 501         args_size -= 4;
 502 #endif
 503
 504     gcall_or_jmp(0);
 505
 506     if (args_size && func_call != FUNC_STDCALL && func_call != FUNC_FASTCALLW)
 507         gadd_sp(args_size);
 508     vtop--;
 509 }
 510
 511 #ifdef TCC_TARGET_PE
 512 #define FUNC_PROLOG_SIZE (10 + USE_EBX)
 513 #else
 514 #define FUNC_PROLOG_SIZE (9 + USE_EBX)
 515 #endif
 516
 517 /* generate function prolog of type 't' */
 518 ST_FUNC void gfunc_prolog(Sym *func_sym)
 519 {
 520     CType *func_type = &func_sym->type;
 521     int addr, align, size, func_call, fastcall_nb_regs;
 522     int param_index, param_addr;
 523     uint8_t *fastcall_regs_ptr;
 524     Sym *sym;
 525     CType *type;
 526
 527     sym = func_type->ref;
 528     func_call = sym->f.func_call;
 529     addr = 8;
 530     loc = 0;
 531     func_vc = 0;
 532
 533     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 534         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 535         fastcall_regs_ptr = fastcall_regs;
 536     } else if (func_call == FUNC_FASTCALLW) {
 537         fastcall_nb_regs = 2;
 538         fastcall_regs_ptr = fastcallw_regs;
 539     } else {
 540         fastcall_nb_regs = 0;
 541         fastcall_regs_ptr = NULL;
 542     }
 543     param_index = 0;
 544
 545     ind += FUNC_PROLOG_SIZE;
 546     func_sub_sp_offset = ind;
 547     /* if the function returns a structure, then add an
 548        implicit pointer parameter */
 549 #ifdef TCC_TARGET_PE
 550     size = type_size(&func_vt,&align);
 551     if (((func_vt.t & VT_BTYPE) == VT_STRUCT)
 552         && (size > 8 || (size & (size - 1)))) {
 553 #else
 554     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 555 #endif
 556         /* XXX: fastcall case ? */
 557         func_vc = addr;
 558         addr += 4;
 559         param_index++;
 560     }
 561     /* define parameters */
 562     while ((sym = sym->next) != NULL) {
 563         type = &sym->type;
 564         size = type_size(type, &align);
 565         size = (size + 3) & ~3;
 566 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 567         /* structs are passed as pointer */
 568         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 569             size = 4;
 570         }
 571 #endif
 572         if (param_index < fastcall_nb_regs) {
 573             /* save FASTCALL register */
 574             loc -= 4;
 575             o(0x89);     /* movl */
 576             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 577             param_addr = loc;
 578         } else {
 579             param_addr = addr;
 580             addr += size;
 581         }
 582         sym_push(sym->v & ~SYM_FIELD, type,
 583                  VT_LOCAL | VT_LVAL, param_addr);
 584         param_index++;
 585     }
 586     func_ret_sub = 0;
 587     /* pascal type call or fastcall ? */
 588     if (func_call == FUNC_STDCALL || func_call == FUNC_FASTCALLW)
 589         func_ret_sub = addr - 8;
 590 #ifndef TCC_TARGET_PE
 591     else if (func_vc)
 592         func_ret_sub = 4;
 593 #endif
 594
 595 #ifdef CONFIG_TCC_BCHECK
 596     if (tcc_state->do_bounds_check)
 597         gen_bounds_prolog();
 598 #endif
 599 }
 600
 601 /* generate function epilog */
 602 ST_FUNC void gfunc_epilog(void)
 603 {
 604     addr_t v, saved_ind;
 605
 606 #ifdef CONFIG_TCC_BCHECK
 607     if (tcc_state->do_bounds_check)
 608         gen_bounds_epilog();
 609 #endif
 610
 611     /* align local size to word & save local variables */
 612     v = (-loc + 3) & -4;
 613
 614 #if USE_EBX
 615     o(0x8b);
 616     gen_modrm(TREG_EBX, VT_LOCAL, NULL, -(v+4));
 617 #endif
 618
 619     o(0xc9); /* leave */
 620     if (func_ret_sub == 0) {
 621         o(0xc3); /* ret */
 622     } else {
 623         o(0xc2); /* ret n */
 624         g(func_ret_sub);
 625         g(func_ret_sub >> 8);
 626     }
 627     saved_ind = ind;
 628     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 629 #ifdef TCC_TARGET_PE
 630     if (v >= 4096) {
 631         oad(0xb8, v); /* mov stacksize, %eax */
 632         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 633     } else
 634 #endif
 635     {
 636         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 637         o(0xec81);  /* sub esp, stacksize */
 638         gen_le32(v);
 639 #ifdef TCC_TARGET_PE
 640         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 641 #endif
 642     }
 643     o(0x53 * USE_EBX); /* push ebx */
 644     ind = saved_ind;
 645 }
 646
 647 /* generate a jump to a label */
 648 ST_FUNC int gjmp(int t)
 649 {
 650     return gjmp2(0xe9, t);
 651 }
 652
 653 /* generate a jump to a fixed address */
 654 ST_FUNC void gjmp_addr(int a)
 655 {
 656     int r;
 657     r = a - ind - 2;
 658     if (r == (char)r) {
 659         g(0xeb);
 660         g(r);
 661     } else {
 662         oad(0xe9, a - ind - 5);
 663     }
 664 }
 665
 666 #if 0
 667 /* generate a jump to a fixed address */
 668 ST_FUNC void gjmp_cond_addr(int a, int op)
 669 {
 670     int r = a - ind - 2;
 671     if (r == (char)r)
 672         g(op - 32), g(r);
 673     else
 674         g(0x0f), gjmp2(op - 16, r - 4);
 675 }
 676 #endif
 677
 678 ST_FUNC int gjmp_append(int n, int t)
 679 {
 680     void *p;
 681     /* insert vtop->c jump list in t */
 682     if (n) {
 683         uint32_t n1 = n, n2;
 684         while ((n2 = read32le(p = cur_text_section->data + n1)))
 685             n1 = n2;
 686         write32le(p, t);
 687         t = n;
 688     }
 689     return t;
 690 }
 691
 692 ST_FUNC int gjmp_cond(int op, int t)
 693 {
 694     g(0x0f);
 695     t = gjmp2(op - 16, t);
 696     return t;
 697 }
 698
 699 ST_FUNC void gen_opi(int op)
 700 {
 701     int r, fr, opc, c;
 702
 703     switch(op) {
 704     case '+':
 705     case TOK_ADDC1: /* add with carry generation */
 706         opc = 0;
 707     gen_op8:
 708         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 709             /* constant case */
 710             vswap();
 711             r = gv(RC_INT);
 712             vswap();
 713             c = vtop->c.i;
 714             if (c == (char)c) {
 715                 /* generate inc and dec for smaller code */
 716                 if ((c == 1 || c == -1) && (op == '+' || op == '-')) {
 717                     opc = (c == 1) ^ (op == '+');
 718                     o (0x40 | (opc << 3) | r); // inc,dec
 719                 } else {
 720                     o(0x83);
 721                     o(0xc0 | (opc << 3) | r);
 722                     g(c);
 723                 }
 724             } else {
 725                 o(0x81);
 726                 oad(0xc0 | (opc << 3) | r, c);
 727             }
 728         } else {
 729             gv2(RC_INT, RC_INT);
 730             r = vtop[-1].r;
 731             fr = vtop[0].r;
 732             o((opc << 3) | 0x01);
 733             o(0xc0 + r + fr * 8);
 734         }
 735         vtop--;
 736         if (op >= TOK_ULT && op <= TOK_GT)
 737             vset_VT_CMP(op);
 738         break;
 739     case '-':
 740     case TOK_SUBC1: /* sub with carry generation */
 741         opc = 5;
 742         goto gen_op8;
 743     case TOK_ADDC2: /* add with carry use */
 744         opc = 2;
 745         goto gen_op8;
 746     case TOK_SUBC2: /* sub with carry use */
 747         opc = 3;
 748         goto gen_op8;
 749     case '&':
 750         opc = 4;
 751         goto gen_op8;
 752     case '^':
 753         opc = 6;
 754         goto gen_op8;
 755     case '|':
 756         opc = 1;
 757         goto gen_op8;
 758     case '*':
 759         gv2(RC_INT, RC_INT);
 760         r = vtop[-1].r;
 761         fr = vtop[0].r;
 762         vtop--;
 763         o(0xaf0f); /* imul fr, r */
 764         o(0xc0 + fr + r * 8);
 765         break;
 766     case TOK_SHL:
 767         opc = 4;
 768         goto gen_shift;
 769     case TOK_SHR:
 770         opc = 5;
 771         goto gen_shift;
 772     case TOK_SAR:
 773         opc = 7;
 774     gen_shift:
 775         opc = 0xc0 | (opc << 3);
 776         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 777             /* constant case */
 778             vswap();
 779             r = gv(RC_INT);
 780             vswap();
 781             c = vtop->c.i & 0x1f;
 782             o(0xc1); /* shl/shr/sar $xxx, r */
 783             o(opc | r);
 784             g(c);
 785         } else {
 786             /* we generate the shift in ecx */
 787             gv2(RC_INT, RC_ECX);
 788             r = vtop[-1].r;
 789             o(0xd3); /* shl/shr/sar %cl, r */
 790             o(opc | r);
 791         }
 792         vtop--;
 793         break;
 794     case '/':
 795     case TOK_UDIV:
 796     case TOK_PDIV:
 797     case '%':
 798     case TOK_UMOD:
 799     case TOK_UMULL:
 800         /* first operand must be in eax */
 801         /* XXX: need better constraint for second operand */
 802         gv2(RC_EAX, RC_ECX);
 803         r = vtop[-1].r;
 804         fr = vtop[0].r;
 805         vtop--;
 806         save_reg(TREG_EDX);
 807         /* save EAX too if used otherwise */
 808         save_reg_upstack(TREG_EAX, 1);
 809         if (op == TOK_UMULL) {
 810             o(0xf7); /* mul fr */
 811             o(0xe0 + fr);
 812             vtop->r2 = TREG_EDX;
 813             r = TREG_EAX;
 814         } else {
 815             if (op == TOK_UDIV || op == TOK_UMOD) {
 816                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 817                 o(0xf0 + fr);
 818             } else {
 819                 o(0xf799); /* cltd, idiv fr, %eax */
 820                 o(0xf8 + fr);
 821             }
 822             if (op == '%' || op == TOK_UMOD)
 823                 r = TREG_EDX;
 824             else
 825                 r = TREG_EAX;
 826         }
 827         vtop->r = r;
 828         break;
 829     default:
 830         opc = 7;
 831         goto gen_op8;
 832     }
 833 }
 834
 835 /* generate a floating point operation 'v = t1 op t2' instruction. The
 836    two operands are guaranteed to have the same floating point type */
 837 /* XXX: need to use ST1 too */
 838 ST_FUNC void gen_opf(int op)
 839 {
 840     int a, ft, fc, swapped, r;
 841
 842     /* convert constants to memory references */
 843     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 844         vswap();
 845         gv(RC_FLOAT);
 846         vswap();
 847     }
 848     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 849         gv(RC_FLOAT);
 850
 851     /* must put at least one value in the floating point register */
 852     if ((vtop[-1].r & VT_LVAL) &&
 853         (vtop[0].r & VT_LVAL)) {
 854         vswap();
 855         gv(RC_FLOAT);
 856         vswap();
 857     }
 858     swapped = 0;
 859     /* swap the stack if needed so that t1 is the register and t2 is
 860        the memory reference */
 861     if (vtop[-1].r & VT_LVAL) {
 862         vswap();
 863         swapped = 1;
 864     }
 865     if (op >= TOK_ULT && op <= TOK_GT) {
 866         /* load on stack second operand */
 867         load(TREG_ST0, vtop);
 868         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 869         if (op == TOK_GE || op == TOK_GT)
 870             swapped = !swapped;
 871         else if (op == TOK_EQ || op == TOK_NE)
 872             swapped = 0;
 873         if (swapped)
 874             o(0xc9d9); /* fxch %st(1) */
 875         if (op == TOK_EQ || op == TOK_NE)
 876             o(0xe9da); /* fucompp */
 877         else
 878             o(0xd9de); /* fcompp */
 879         o(0xe0df); /* fnstsw %ax */
 880         if (op == TOK_EQ) {
 881             o(0x45e480); /* and $0x45, %ah */
 882             o(0x40fC80); /* cmp $0x40, %ah */
 883         } else if (op == TOK_NE) {
 884             o(0x45e480); /* and $0x45, %ah */
 885             o(0x40f480); /* xor $0x40, %ah */
 886             op = TOK_NE;
 887         } else if (op == TOK_GE || op == TOK_LE) {
 888             o(0x05c4f6); /* test $0x05, %ah */
 889             op = TOK_EQ;
 890         } else {
 891             o(0x45c4f6); /* test $0x45, %ah */
 892             op = TOK_EQ;
 893         }
 894         vtop--;
 895         vset_VT_CMP(op);
 896     } else {
 897         /* no memory reference possible for long double operations */
 898         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 899             load(TREG_ST0, vtop);
 900             swapped = !swapped;
 901         }
 902
 903         switch(op) {
 904         default:
 905         case '+':
 906             a = 0;
 907             break;
 908         case '-':
 909             a = 4;
 910             if (swapped)
 911                 a++;
 912             break;
 913         case '*':
 914             a = 1;
 915             break;
 916         case '/':
 917             a = 6;
 918             if (swapped)
 919                 a++;
 920             break;
 921         }
 922         ft = vtop->type.t;
 923         fc = vtop->c.i;
 924         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 925             o(0xde); /* fxxxp %st, %st(1) */
 926             o(0xc1 + (a << 3));
 927         } else {
 928             /* if saved lvalue, then we must reload it */
 929             r = vtop->r;
 930             if ((r & VT_VALMASK) == VT_LLOCAL) {
 931                 SValue v1;
 932                 r = get_reg(RC_INT);
 933                 v1.type.t = VT_INT;
 934                 v1.r = VT_LOCAL | VT_LVAL;
 935                 v1.c.i = fc;
 936                 v1.sym = NULL;
 937                 load(r, &v1);
 938                 fc = 0;
 939             }
 940
 941             if ((ft & VT_BTYPE) == VT_DOUBLE)
 942                 o(0xdc);
 943             else
 944                 o(0xd8);
 945             gen_modrm(a, r, vtop->sym, fc);
 946         }
 947         vtop--;
 948     }
 949 }
 950
 951 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 952    and 'long long' cases. */
 953 ST_FUNC void gen_cvt_itof(int t)
 954 {
 955     save_reg(TREG_ST0);
 956     gv(RC_INT);
 957     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 958         /* signed long long to float/double/long double (unsigned case
 959            is handled generically) */
 960         o(0x50 + vtop->r2); /* push r2 */
 961         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 962         o(0x242cdf); /* fildll (%esp) */
 963         o(0x08c483); /* add $8, %esp */
 964         vtop->r2 = VT_CONST;
 965     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 966                (VT_INT | VT_UNSIGNED)) {
 967         /* unsigned int to float/double/long double */
 968         o(0x6a); /* push $0 */
 969         g(0x00);
 970         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 971         o(0x242cdf); /* fildll (%esp) */
 972         o(0x08c483); /* add $8, %esp */
 973     } else {
 974         /* int to float/double/long double */
 975         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 976         o(0x2404db); /* fildl (%esp) */
 977         o(0x04c483); /* add $4, %esp */
 978     }
 979     vtop->r2 = VT_CONST;
 980     vtop->r = TREG_ST0;
 981 }
 982
 983 /* convert fp to int 't' type */
 984 ST_FUNC void gen_cvt_ftoi(int t)
 985 {
 986     int bt = vtop->type.t & VT_BTYPE;
 987     if (bt == VT_FLOAT)
 988         vpush_helper_func(TOK___fixsfdi);
 989     else if (bt == VT_LDOUBLE)
 990         vpush_helper_func(TOK___fixxfdi);
 991     else
 992         vpush_helper_func(TOK___fixdfdi);
 993     vswap();
 994     gfunc_call(1);
 995     vpushi(0);
 996     vtop->r = REG_IRET;
 997     if ((t & VT_BTYPE) == VT_LLONG)
 998         vtop->r2 = REG_IRE2;
 999 }
1000
1001 /* convert from one floating point type to another */
1002 ST_FUNC void gen_cvt_ftof(int t)
1003 {
1004     /* all we have to do on i386 is to put the float in a register */
1005     gv(RC_FLOAT);
1006 }
1007
1008 /* char/short to int conversion */
1009 ST_FUNC void gen_cvt_csti(int t)
1010 {
1011     int r, sz, xl;
1012     r = gv(RC_INT);
1013     sz = !(t & VT_UNSIGNED);
1014     xl = (t & VT_BTYPE) == VT_SHORT;
1015     o(0xc0b60f /* mov[sz] %a[xl], %eax */
1016         | (sz << 3 | xl) << 8
1017         | (r << 3 | r) << 16
1018         );
1019 }
1020
1021 /* computed goto support */
1022 ST_FUNC void ggoto(void)
1023 {
1024     gcall_or_jmp(1);
1025     vtop--;
1026 }
1027
1028 /* bound check support functions */
1029 #ifdef CONFIG_TCC_BCHECK
1030
1031 static void gen_bounds_prolog(void)
1032 {
1033     /* leave some room for bound checking code */
1034     func_bound_offset = lbounds_section->data_offset;
1035     func_bound_ind = ind;
1036     func_bound_add_epilog = 0;
1037     oad(0xb8, 0); /* lbound section pointer */
1038     oad(0xb8, 0); /* call to function */
1039 }
1040
1041 static void gen_bounds_epilog(void)
1042 {
1043     addr_t saved_ind;
1044     addr_t *bounds_ptr;
1045     Sym *sym_data;
1046     int offset_modified = func_bound_offset != lbounds_section->data_offset;
1047
1048     if (!offset_modified && !func_bound_add_epilog)
1049         return;
1050
1051     /* add end of table info */
1052     bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
1053     *bounds_ptr = 0;
1054
1055     sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
1056                            func_bound_offset, lbounds_section->data_offset);
1057
1058     /* generate bound local allocation */
1059     if (offset_modified) {
1060         saved_ind = ind;
1061         ind = func_bound_ind;
1062         greloc(cur_text_section, sym_data, ind + 1, R_386_32);
1063         ind = ind + 5;
1064         gen_static_call(TOK___bound_local_new);
1065         ind = saved_ind;
1066     }
1067
1068     /* generate bound check local freeing */
1069     o(0x5250); /* save returned value, if any */
1070     greloc(cur_text_section, sym_data, ind + 1, R_386_32);
1071     oad(0xb8, 0); /* mov %eax, xxx */
1072     gen_static_call(TOK___bound_local_delete);
1073     o(0x585a); /* restore returned value, if any */
1074 }
1075 #endif
1076
1077 /* Save the stack pointer onto the stack */
1078 ST_FUNC void gen_vla_sp_save(int addr) {
1079     /* mov %esp,addr(%ebp)*/
1080     o(0x89);
1081     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1082 }
1083
1084 /* Restore the SP from a location on the stack */
1085 ST_FUNC void gen_vla_sp_restore(int addr) {
1086     o(0x8b);
1087     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1088 }
1089
1090 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1091 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1092     int use_call = 0;
1093
1094 #if defined(CONFIG_TCC_BCHECK)
1095     use_call = tcc_state->do_bounds_check;
1096 #endif
1097 #ifdef TCC_TARGET_PE    /* alloca does more than just adjust %rsp on Windows */
1098     use_call = 1;
1099 #endif
1100     if (use_call)
1101     {
1102         vpush_helper_func(TOK_alloca);
1103         vswap(); /* Move alloca ref past allocation size */
1104         gfunc_call(1);
1105     }
1106     else {
1107         int r;
1108         r = gv(RC_INT); /* allocation size */
1109         /* sub r,%rsp */
1110         o(0x2b);
1111         o(0xe0 | r);
1112         /* We align to 16 bytes rather than align */
1113         /* and ~15, %esp */
1114         o(0xf0e483);
1115         vpop();
1116     }
1117 }
1118
1119 /* end of X86 code generator */
1120 /*************************************************************/
1121 #endif
1122 /*************************************************************/