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_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_global_sym(v, &func_old_type);
 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 #ifdef CONFIG_TCC_BCHECK
 363         if (tcc_state->do_bounds_check &&
 364             (vtop->sym->v == TOK_alloca ||
 365              vtop->sym->v == TOK_setjmp ||
 366              vtop->sym->v == TOK__setjmp
 367 #ifndef TCC_TARGET_PE
 368              || vtop->sym->v == TOK_sigsetjmp
 369              || vtop->sym->v == TOK___sigsetjmp
 370 #endif
 371             ))
 372             func_bound_add_epilog = 1;
 373 #endif
 374     } else {
 375         /* otherwise, indirect call */
 376         r = gv(RC_INT);
 377         o(0xff); /* call/jmp *r */
 378         o(0xd0 + r + (is_jmp << 4));
 379     }
 380 }
 381
 382 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 383 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 384
 385 /* Return the number of registers needed to return the struct, or 0 if
 386    returning via struct pointer. */
 387 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
 388 {
 389 #ifdef TCC_TARGET_PE
 390     int size, align;
 391     *ret_align = 1; // Never have to re-align return values for x86
 392     *regsize = 4;
 393     size = type_size(vt, &align);
 394     if (size > 8 || (size & (size - 1)))
 395         return 0;
 396     if (size == 8)
 397         ret->t = VT_LLONG;
 398     else if (size == 4)
 399         ret->t = VT_INT;
 400     else if (size == 2)
 401         ret->t = VT_SHORT;
 402     else
 403         ret->t = VT_BYTE;
 404     ret->ref = NULL;
 405     return 1;
 406 #else
 407     *ret_align = 1; // Never have to re-align return values for x86
 408     return 0;
 409 #endif
 410 }
 411
 412 /* Generate function call. The function address is pushed first, then
 413    all the parameters in call order. This functions pops all the
 414    parameters and the function address. */
 415 ST_FUNC void gfunc_call(int nb_args)
 416 {
 417     int size, align, r, args_size, i, func_call;
 418     Sym *func_sym;
 419
 420 #ifdef CONFIG_TCC_BCHECK
 421     if (tcc_state->do_bounds_check)
 422         gbound_args(nb_args);
 423 #endif
 424
 425     args_size = 0;
 426     for(i = 0;i < nb_args; i++) {
 427         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 428             size = type_size(&vtop->type, &align);
 429             /* align to stack align size */
 430             size = (size + 3) & ~3;
 431             /* allocate the necessary size on stack */
 432             oad(0xec81, size); /* sub $xxx, %esp */
 433             /* generate structure store */
 434             r = get_reg(RC_INT);
 435             o(0x89); /* mov %esp, r */
 436             o(0xe0 + r);
 437             vset(&vtop->type, r | VT_LVAL, 0);
 438             vswap();
 439             vstore();
 440             args_size += size;
 441         } else if (is_float(vtop->type.t)) {
 442             gv(RC_FLOAT); /* only one float register */
 443             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 444                 size = 4;
 445             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 446                 size = 8;
 447             else
 448                 size = 12;
 449             oad(0xec81, size); /* sub $xxx, %esp */
 450             if (size == 12)
 451                 o(0x7cdb);
 452             else
 453                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 454             g(0x24);
 455             g(0x00);
 456             args_size += size;
 457         } else {
 458             /* simple type (currently always same size) */
 459             /* XXX: implicit cast ? */
 460             r = gv(RC_INT);
 461             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 462                 size = 8;
 463                 o(0x50 + vtop->r2); /* push r */
 464             } else {
 465                 size = 4;
 466             }
 467             o(0x50 + r); /* push r */
 468             args_size += size;
 469         }
 470         vtop--;
 471     }
 472     save_regs(0); /* save used temporary registers */
 473     func_sym = vtop->type.ref;
 474     func_call = func_sym->f.func_call;
 475     /* fast call case */
 476     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 477         func_call == FUNC_FASTCALLW) {
 478         int fastcall_nb_regs;
 479         uint8_t *fastcall_regs_ptr;
 480         if (func_call == FUNC_FASTCALLW) {
 481             fastcall_regs_ptr = fastcallw_regs;
 482             fastcall_nb_regs = 2;
 483         } else {
 484             fastcall_regs_ptr = fastcall_regs;
 485             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 486         }
 487         for(i = 0;i < fastcall_nb_regs; i++) {
 488             if (args_size <= 0)
 489                 break;
 490             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 491             /* XXX: incorrect for struct/floats */
 492             args_size -= 4;
 493         }
 494     }
 495 #ifndef TCC_TARGET_PE
 496     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 497         args_size -= 4;
 498 #endif
 499
 500     gcall_or_jmp(0);
 501
 502     if (args_size && func_call != FUNC_STDCALL && func_call != FUNC_FASTCALLW)
 503         gadd_sp(args_size);
 504     vtop--;
 505 }
 506
 507 #ifdef TCC_TARGET_PE
 508 #define FUNC_PROLOG_SIZE (10 + USE_EBX)
 509 #else
 510 #define FUNC_PROLOG_SIZE (9 + USE_EBX)
 511 #endif
 512
 513 /* generate function prolog of type 't' */
 514 ST_FUNC void gfunc_prolog(Sym *func_sym)
 515 {
 516     CType *func_type = &func_sym->type;
 517     int addr, align, size, func_call, fastcall_nb_regs;
 518     int param_index, param_addr;
 519     uint8_t *fastcall_regs_ptr;
 520     Sym *sym;
 521     CType *type;
 522
 523     sym = func_type->ref;
 524     func_call = sym->f.func_call;
 525     addr = 8;
 526     loc = 0;
 527     func_vc = 0;
 528
 529     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 530         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 531         fastcall_regs_ptr = fastcall_regs;
 532     } else if (func_call == FUNC_FASTCALLW) {
 533         fastcall_nb_regs = 2;
 534         fastcall_regs_ptr = fastcallw_regs;
 535     } else {
 536         fastcall_nb_regs = 0;
 537         fastcall_regs_ptr = NULL;
 538     }
 539     param_index = 0;
 540
 541     ind += FUNC_PROLOG_SIZE;
 542     func_sub_sp_offset = ind;
 543     /* if the function returns a structure, then add an
 544        implicit pointer parameter */
 545 #ifdef TCC_TARGET_PE
 546     size = type_size(&func_vt,&align);
 547     if (((func_vt.t & VT_BTYPE) == VT_STRUCT)
 548         && (size > 8 || (size & (size - 1)))) {
 549 #else
 550     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 551 #endif
 552         /* XXX: fastcall case ? */
 553         func_vc = addr;
 554         addr += 4;
 555         param_index++;
 556     }
 557     /* define parameters */
 558     while ((sym = sym->next) != NULL) {
 559         type = &sym->type;
 560         size = type_size(type, &align);
 561         size = (size + 3) & ~3;
 562 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 563         /* structs are passed as pointer */
 564         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 565             size = 4;
 566         }
 567 #endif
 568         if (param_index < fastcall_nb_regs) {
 569             /* save FASTCALL register */
 570             loc -= 4;
 571             o(0x89);     /* movl */
 572             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 573             param_addr = loc;
 574         } else {
 575             param_addr = addr;
 576             addr += size;
 577         }
 578         sym_push(sym->v & ~SYM_FIELD, type,
 579                  VT_LOCAL | VT_LVAL, param_addr);
 580         param_index++;
 581     }
 582     func_ret_sub = 0;
 583     /* pascal type call or fastcall ? */
 584     if (func_call == FUNC_STDCALL || func_call == FUNC_FASTCALLW)
 585         func_ret_sub = addr - 8;
 586 #ifndef TCC_TARGET_PE
 587     else if (func_vc)
 588         func_ret_sub = 4;
 589 #endif
 590
 591 #ifdef CONFIG_TCC_BCHECK
 592     if (tcc_state->do_bounds_check)
 593         gen_bounds_prolog();
 594 #endif
 595 }
 596
 597 /* generate function epilog */
 598 ST_FUNC void gfunc_epilog(void)
 599 {
 600     addr_t v, saved_ind;
 601
 602 #ifdef CONFIG_TCC_BCHECK
 603     if (tcc_state->do_bounds_check)
 604         gen_bounds_epilog();
 605 #endif
 606
 607     /* align local size to word & save local variables */
 608     v = (-loc + 3) & -4;
 609
 610 #if USE_EBX
 611     o(0x8b);
 612     gen_modrm(TREG_EBX, VT_LOCAL, NULL, -(v+4));
 613 #endif
 614
 615     o(0xc9); /* leave */
 616     if (func_ret_sub == 0) {
 617         o(0xc3); /* ret */
 618     } else {
 619         o(0xc2); /* ret n */
 620         g(func_ret_sub);
 621         g(func_ret_sub >> 8);
 622     }
 623     saved_ind = ind;
 624     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 625 #ifdef TCC_TARGET_PE
 626     if (v >= 4096) {
 627         oad(0xb8, v); /* mov stacksize, %eax */
 628         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 629     } else
 630 #endif
 631     {
 632         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 633         o(0xec81);  /* sub esp, stacksize */
 634         gen_le32(v);
 635 #ifdef TCC_TARGET_PE
 636         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 637 #endif
 638     }
 639     o(0x53 * USE_EBX); /* push ebx */
 640     ind = saved_ind;
 641 }
 642
 643 /* generate a jump to a label */
 644 ST_FUNC int gjmp(int t)
 645 {
 646     return gjmp2(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 #if 0
 663 /* generate a jump to a fixed address */
 664 ST_FUNC void gjmp_cond_addr(int a, int op)
 665 {
 666     int r = a - ind - 2;
 667     if (r == (char)r)
 668         g(op - 32), g(r);
 669     else
 670         g(0x0f), gjmp2(op - 16, r - 4);
 671 }
 672 #endif
 673
 674 ST_FUNC int gjmp_append(int n, int t)
 675 {
 676     void *p;
 677     /* insert vtop->c jump list in t */
 678     if (n) {
 679         uint32_t n1 = n, n2;
 680         while ((n2 = read32le(p = cur_text_section->data + n1)))
 681             n1 = n2;
 682         write32le(p, t);
 683         t = n;
 684     }
 685     return t;
 686 }
 687
 688 ST_FUNC int gjmp_cond(int op, int t)
 689 {
 690     g(0x0f);
 691     t = gjmp2(op - 16, t);
 692     return t;
 693 }
 694
 695 ST_FUNC void gen_opi(int op)
 696 {
 697     int r, fr, opc, c;
 698
 699     switch(op) {
 700     case '+':
 701     case TOK_ADDC1: /* add with carry generation */
 702         opc = 0;
 703     gen_op8:
 704         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 705             /* constant case */
 706             vswap();
 707             r = gv(RC_INT);
 708             vswap();
 709             c = vtop->c.i;
 710             if (c == (char)c) {
 711                 /* generate inc and dec for smaller code */
 712                 if ((c == 1 || c == -1) && (op == '+' || op == '-')) {
 713                     opc = (c == 1) ^ (op == '+');
 714                     o (0x40 | (opc << 3) | r); // inc,dec
 715                 } else {
 716                     o(0x83);
 717                     o(0xc0 | (opc << 3) | r);
 718                     g(c);
 719                 }
 720             } else {
 721                 o(0x81);
 722                 oad(0xc0 | (opc << 3) | r, c);
 723             }
 724         } else {
 725             gv2(RC_INT, RC_INT);
 726             r = vtop[-1].r;
 727             fr = vtop[0].r;
 728             o((opc << 3) | 0x01);
 729             o(0xc0 + r + fr * 8);
 730         }
 731         vtop--;
 732         if (op >= TOK_ULT && op <= TOK_GT)
 733             vset_VT_CMP(op);
 734         break;
 735     case '-':
 736     case TOK_SUBC1: /* sub with carry generation */
 737         opc = 5;
 738         goto gen_op8;
 739     case TOK_ADDC2: /* add with carry use */
 740         opc = 2;
 741         goto gen_op8;
 742     case TOK_SUBC2: /* sub with carry use */
 743         opc = 3;
 744         goto gen_op8;
 745     case '&':
 746         opc = 4;
 747         goto gen_op8;
 748     case '^':
 749         opc = 6;
 750         goto gen_op8;
 751     case '|':
 752         opc = 1;
 753         goto gen_op8;
 754     case '*':
 755         gv2(RC_INT, RC_INT);
 756         r = vtop[-1].r;
 757         fr = vtop[0].r;
 758         vtop--;
 759         o(0xaf0f); /* imul fr, r */
 760         o(0xc0 + fr + r * 8);
 761         break;
 762     case TOK_SHL:
 763         opc = 4;
 764         goto gen_shift;
 765     case TOK_SHR:
 766         opc = 5;
 767         goto gen_shift;
 768     case TOK_SAR:
 769         opc = 7;
 770     gen_shift:
 771         opc = 0xc0 | (opc << 3);
 772         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 773             /* constant case */
 774             vswap();
 775             r = gv(RC_INT);
 776             vswap();
 777             c = vtop->c.i & 0x1f;
 778             o(0xc1); /* shl/shr/sar $xxx, r */
 779             o(opc | r);
 780             g(c);
 781         } else {
 782             /* we generate the shift in ecx */
 783             gv2(RC_INT, RC_ECX);
 784             r = vtop[-1].r;
 785             o(0xd3); /* shl/shr/sar %cl, r */
 786             o(opc | r);
 787         }
 788         vtop--;
 789         break;
 790     case '/':
 791     case TOK_UDIV:
 792     case TOK_PDIV:
 793     case '%':
 794     case TOK_UMOD:
 795     case TOK_UMULL:
 796         /* first operand must be in eax */
 797         /* XXX: need better constraint for second operand */
 798         gv2(RC_EAX, RC_ECX);
 799         r = vtop[-1].r;
 800         fr = vtop[0].r;
 801         vtop--;
 802         save_reg(TREG_EDX);
 803         /* save EAX too if used otherwise */
 804         save_reg_upstack(TREG_EAX, 1);
 805         if (op == TOK_UMULL) {
 806             o(0xf7); /* mul fr */
 807             o(0xe0 + fr);
 808             vtop->r2 = TREG_EDX;
 809             r = TREG_EAX;
 810         } else {
 811             if (op == TOK_UDIV || op == TOK_UMOD) {
 812                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 813                 o(0xf0 + fr);
 814             } else {
 815                 o(0xf799); /* cltd, idiv fr, %eax */
 816                 o(0xf8 + fr);
 817             }
 818             if (op == '%' || op == TOK_UMOD)
 819                 r = TREG_EDX;
 820             else
 821                 r = TREG_EAX;
 822         }
 823         vtop->r = r;
 824         break;
 825     default:
 826         opc = 7;
 827         goto gen_op8;
 828     }
 829 }
 830
 831 /* generate a floating point operation 'v = t1 op t2' instruction. The
 832    two operands are guaranteed to have the same floating point type */
 833 /* XXX: need to use ST1 too */
 834 ST_FUNC void gen_opf(int op)
 835 {
 836     int a, ft, fc, swapped, r;
 837
 838     /* convert constants to memory references */
 839     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 840         vswap();
 841         gv(RC_FLOAT);
 842         vswap();
 843     }
 844     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 845         gv(RC_FLOAT);
 846
 847     /* must put at least one value in the floating point register */
 848     if ((vtop[-1].r & VT_LVAL) &&
 849         (vtop[0].r & VT_LVAL)) {
 850         vswap();
 851         gv(RC_FLOAT);
 852         vswap();
 853     }
 854     swapped = 0;
 855     /* swap the stack if needed so that t1 is the register and t2 is
 856        the memory reference */
 857     if (vtop[-1].r & VT_LVAL) {
 858         vswap();
 859         swapped = 1;
 860     }
 861     if (op >= TOK_ULT && op <= TOK_GT) {
 862         /* load on stack second operand */
 863         load(TREG_ST0, vtop);
 864         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 865         if (op == TOK_GE || op == TOK_GT)
 866             swapped = !swapped;
 867         else if (op == TOK_EQ || op == TOK_NE)
 868             swapped = 0;
 869         if (swapped)
 870             o(0xc9d9); /* fxch %st(1) */
 871         if (op == TOK_EQ || op == TOK_NE)
 872             o(0xe9da); /* fucompp */
 873         else
 874             o(0xd9de); /* fcompp */
 875         o(0xe0df); /* fnstsw %ax */
 876         if (op == TOK_EQ) {
 877             o(0x45e480); /* and $0x45, %ah */
 878             o(0x40fC80); /* cmp $0x40, %ah */
 879         } else if (op == TOK_NE) {
 880             o(0x45e480); /* and $0x45, %ah */
 881             o(0x40f480); /* xor $0x40, %ah */
 882             op = TOK_NE;
 883         } else if (op == TOK_GE || op == TOK_LE) {
 884             o(0x05c4f6); /* test $0x05, %ah */
 885             op = TOK_EQ;
 886         } else {
 887             o(0x45c4f6); /* test $0x45, %ah */
 888             op = TOK_EQ;
 889         }
 890         vtop--;
 891         vset_VT_CMP(op);
 892     } else {
 893         /* no memory reference possible for long double operations */
 894         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 895             load(TREG_ST0, vtop);
 896             swapped = !swapped;
 897         }
 898
 899         switch(op) {
 900         default:
 901         case '+':
 902             a = 0;
 903             break;
 904         case '-':
 905             a = 4;
 906             if (swapped)
 907                 a++;
 908             break;
 909         case '*':
 910             a = 1;
 911             break;
 912         case '/':
 913             a = 6;
 914             if (swapped)
 915                 a++;
 916             break;
 917         }
 918         ft = vtop->type.t;
 919         fc = vtop->c.i;
 920         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 921             o(0xde); /* fxxxp %st, %st(1) */
 922             o(0xc1 + (a << 3));
 923         } else {
 924             /* if saved lvalue, then we must reload it */
 925             r = vtop->r;
 926             if ((r & VT_VALMASK) == VT_LLOCAL) {
 927                 SValue v1;
 928                 r = get_reg(RC_INT);
 929                 v1.type.t = VT_INT;
 930                 v1.r = VT_LOCAL | VT_LVAL;
 931                 v1.c.i = fc;
 932                 v1.sym = NULL;
 933                 load(r, &v1);
 934                 fc = 0;
 935             }
 936
 937             if ((ft & VT_BTYPE) == VT_DOUBLE)
 938                 o(0xdc);
 939             else
 940                 o(0xd8);
 941             gen_modrm(a, r, vtop->sym, fc);
 942         }
 943         vtop--;
 944     }
 945 }
 946
 947 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 948    and 'long long' cases. */
 949 ST_FUNC void gen_cvt_itof(int t)
 950 {
 951     save_reg(TREG_ST0);
 952     gv(RC_INT);
 953     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 954         /* signed long long to float/double/long double (unsigned case
 955            is handled generically) */
 956         o(0x50 + vtop->r2); /* push r2 */
 957         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 958         o(0x242cdf); /* fildll (%esp) */
 959         o(0x08c483); /* add $8, %esp */
 960         vtop->r2 = VT_CONST;
 961     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 962                (VT_INT | VT_UNSIGNED)) {
 963         /* unsigned int to float/double/long double */
 964         o(0x6a); /* push $0 */
 965         g(0x00);
 966         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 967         o(0x242cdf); /* fildll (%esp) */
 968         o(0x08c483); /* add $8, %esp */
 969     } else {
 970         /* int to float/double/long double */
 971         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 972         o(0x2404db); /* fildl (%esp) */
 973         o(0x04c483); /* add $4, %esp */
 974     }
 975     vtop->r2 = VT_CONST;
 976     vtop->r = TREG_ST0;
 977 }
 978
 979 /* convert fp to int 't' type */
 980 ST_FUNC void gen_cvt_ftoi(int t)
 981 {
 982     int bt = vtop->type.t & VT_BTYPE;
 983     if (bt == VT_FLOAT)
 984         vpush_global_sym(&func_old_type, TOK___fixsfdi);
 985     else if (bt == VT_LDOUBLE)
 986         vpush_global_sym(&func_old_type, TOK___fixxfdi);
 987     else
 988         vpush_global_sym(&func_old_type, TOK___fixdfdi);
 989     vswap();
 990     gfunc_call(1);
 991     vpushi(0);
 992     vtop->r = REG_IRET;
 993     if ((t & VT_BTYPE) == VT_LLONG)
 994         vtop->r2 = REG_IRE2;
 995 }
 996
 997 /* convert from one floating point type to another */
 998 ST_FUNC void gen_cvt_ftof(int t)
 999 {
1000     /* all we have to do on i386 is to put the float in a register */
1001     gv(RC_FLOAT);
1002 }
1003
1004 /* char/short to int conversion */
1005 ST_FUNC void gen_cvt_csti(int t)
1006 {
1007     int r, sz, xl;
1008     r = gv(RC_INT);
1009     sz = !(t & VT_UNSIGNED);
1010     xl = (t & VT_BTYPE) == VT_SHORT;
1011     o(0xc0b60f /* mov[sz] %a[xl], %eax */
1012         | (sz << 3 | xl) << 8
1013         | (r << 3 | r) << 16
1014         );
1015 }
1016
1017 /* computed goto support */
1018 ST_FUNC void ggoto(void)
1019 {
1020     gcall_or_jmp(1);
1021     vtop--;
1022 }
1023
1024 /* bound check support functions */
1025 #ifdef CONFIG_TCC_BCHECK
1026 /* generate a bounded pointer addition */
1027 ST_FUNC void gen_bounded_ptr_add(void)
1028 {
1029     vpush_global_sym(&func_old_type, TOK___bound_ptr_add);
1030     vrott(3);
1031     gfunc_call(2);
1032     vpushi(0);
1033     /* returned pointer is in eax */
1034     vtop->r = TREG_EAX | VT_BOUNDED;
1035     if (nocode_wanted)
1036         return;
1037     /* relocation offset of the bounding function call point */
1038     vtop->c.i = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1039 }
1040
1041 /* patch pointer addition in vtop so that pointer dereferencing is
1042    also tested */
1043 ST_FUNC void gen_bounded_ptr_deref(void)
1044 {
1045     addr_t func;
1046     int  size, align;
1047     Elf32_Rel *rel;
1048     Sym *sym;
1049
1050     if (nocode_wanted)
1051         return;
1052
1053     size = type_size(&vtop->type, &align);
1054     switch(size) {
1055     case  1: func = TOK___bound_ptr_indir1; break;
1056     case  2: func = TOK___bound_ptr_indir2; break;
1057     case  4: func = TOK___bound_ptr_indir4; break;
1058     case  8: func = TOK___bound_ptr_indir8; break;
1059     case 12: func = TOK___bound_ptr_indir12; break;
1060     case 16: func = TOK___bound_ptr_indir16; break;
1061     default:
1062         /* may happen with struct member access */
1063         return;
1064         //tcc_error("unhandled size when dereferencing bounded pointer");
1065         //func = 0;
1066         //break;
1067     }
1068     sym = external_global_sym(func, &func_old_type);
1069     if (!sym->c)
1070         put_extern_sym(sym, NULL, 0, 0);
1071     /* patch relocation */
1072     /* XXX: find a better solution ? */
1073     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.i);
1074     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1075 }
1076
1077 static void gen_bounds_prolog(void)
1078 {
1079     /* leave some room for bound checking code */
1080     func_bound_offset = lbounds_section->data_offset;
1081     func_bound_ind = ind;
1082     func_bound_add_epilog = 0;
1083     oad(0xb8, 0); /* lbound section pointer */
1084     oad(0xb8, 0); /* call to function */
1085 }
1086
1087 static void gen_bounds_epilog(void)
1088 {
1089     addr_t saved_ind;
1090     addr_t *bounds_ptr;
1091     Sym *sym_data;
1092     int offset_modified = func_bound_offset != lbounds_section->data_offset;
1093
1094     if (!offset_modified && !func_bound_add_epilog)
1095         return;
1096
1097     /* add end of table info */
1098     bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
1099     *bounds_ptr = 0;
1100
1101     sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
1102                            func_bound_offset, lbounds_section->data_offset);
1103
1104     /* generate bound local allocation */
1105     if (offset_modified) {
1106         saved_ind = ind;
1107         ind = func_bound_ind;
1108         greloc(cur_text_section, sym_data, ind + 1, R_386_32);
1109         ind = ind + 5;
1110         gen_static_call(TOK___bound_local_new);
1111         ind = saved_ind;
1112     }
1113
1114     /* generate bound check local freeing */
1115     o(0x5250); /* save returned value, if any */
1116     greloc(cur_text_section, sym_data, ind + 1, R_386_32);
1117     oad(0xb8, 0); /* mov %eax, xxx */
1118     gen_static_call(TOK___bound_local_delete);
1119     o(0x585a); /* restore returned value, if any */
1120 }
1121 #endif
1122
1123 /* Save the stack pointer onto the stack */
1124 ST_FUNC void gen_vla_sp_save(int addr) {
1125     /* mov %esp,addr(%ebp)*/
1126     o(0x89);
1127     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1128 }
1129
1130 /* Restore the SP from a location on the stack */
1131 ST_FUNC void gen_vla_sp_restore(int addr) {
1132     o(0x8b);
1133     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1134 }
1135
1136 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1137 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1138     int use_call = 0;
1139
1140 #if defined(CONFIG_TCC_BCHECK)
1141     use_call = tcc_state->do_bounds_check;
1142 #endif
1143 #ifdef TCC_TARGET_PE    /* alloca does more than just adjust %rsp on Windows */
1144     use_call = 1;
1145 #endif
1146     if (use_call)
1147     {
1148         vpush_global_sym(&func_old_type, TOK_alloca);
1149         vswap(); /* Move alloca ref past allocation size */
1150         gfunc_call(1);
1151     }
1152     else {
1153         int r;
1154         r = gv(RC_INT); /* allocation size */
1155         /* sub r,%rsp */
1156         o(0x2b);
1157         o(0xe0 | r);
1158         /* We align to 16 bytes rather than align */
1159         /* and ~15, %esp */
1160         o(0xf0e483);
1161         vpop();
1162     }
1163 }
1164
1165 /* end of X86 code generator */
1166 /*************************************************************/
1167 #endif
1168 /*************************************************************/