i386-gen.c

   1 /*
   2  *  X86 code generator for TCC
   3  *
   4  *  Copyright (c) 2001-2004 Fabrice Bellard
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2 of the License, or (at your option) any later version.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  19  */
  20
  21 #ifdef TARGET_DEFS_ONLY
  22
  23 /* number of available registers */
  24 #define NB_REGS         4
  25 #define NB_ASM_REGS     8
  26
  27 typedef int RegArgs;
  28
  29 /* a register can belong to several classes. The classes must be
  30    sorted from more general to more precise (see gv2() code which does
  31    assumptions on it). */
  32 #define RC_INT     0x0001 /* generic integer register */
  33 #define RC_FLOAT   0x0002 /* generic float register */
  34 #define RC_EAX     0x0004
  35 #define RC_ST0     0x0008
  36 #define RC_ECX     0x0010
  37 #define RC_EDX     0x0020
  38 #define RC_IRET    RC_EAX /* function return: integer register */
  39 #define RC_LRET    RC_EDX /* function return: second integer register */
  40 #define RC_FRET    RC_ST0 /* function return: float register */
  41
  42 /* pretty names for the registers */
  43 enum {
  44     TREG_EAX = 0,
  45     TREG_ECX,
  46     TREG_EDX,
  47     TREG_ST0,
  48     TREG_ESP = 4
  49 };
  50
  51 /* return registers for function */
  52 #define REG_IRET TREG_EAX /* single word int return register */
  53 #define REG_LRET TREG_EDX /* second word return register (for long long) */
  54 #define REG_FRET TREG_ST0 /* float return register */
  55
  56 /* defined if function parameters must be evaluated in reverse order */
  57 #define INVERT_FUNC_PARAMS
  58
  59 /* defined if structures are passed as pointers. Otherwise structures
  60    are directly pushed on stack. */
  61 /* #define FUNC_STRUCT_PARAM_AS_PTR */
  62
  63 /* pointer size, in bytes */
  64 #define PTR_SIZE 4
  65
  66 /* long double size and alignment, in bytes */
  67 #define LDOUBLE_SIZE  12
  68 #define LDOUBLE_ALIGN 4
  69 /* maximum alignment (for aligned attribute support) */
  70 #define MAX_ALIGN     8
  71
  72
  73 #define psym oad
  74
  75 /******************************************************/
  76 /* ELF defines */
  77
  78 #define EM_TCC_TARGET EM_386
  79
  80 /* relocation type for 32 bit data relocation */
  81 #define R_DATA_32   R_386_32
  82 #define R_DATA_PTR  R_386_32
  83 #define R_JMP_SLOT  R_386_JMP_SLOT
  84 #define R_COPY      R_386_COPY
  85
  86 #define ELF_START_ADDR 0x08048000
  87 #define ELF_PAGE_SIZE  0x1000
  88
  89 /******************************************************/
  90 #else /* ! TARGET_DEFS_ONLY */
  91 /******************************************************/
  92 #include "tcc.h"
  93
  94 ST_DATA const int reg_classes[NB_REGS] = {
  95     /* eax */ RC_INT | RC_EAX,
  96     /* ecx */ RC_INT | RC_ECX,
  97     /* edx */ RC_INT | RC_EDX,
  98     /* st0 */ RC_FLOAT | RC_ST0,
  99 };
 100
 101 static unsigned long func_sub_sp_offset;
 102 static int func_ret_sub;
 103 #ifdef CONFIG_TCC_BCHECK
 104 static addr_t func_bound_offset;
 105 #endif
 106
 107 /* XXX: make it faster ? */
 108 ST_FUNC void g(int c)
 109 {
 110     int ind1;
 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     int n, *ptr;
 144     while (t) {
 145         ptr = (int *)(cur_text_section->data + t);
 146         n = *ptr; /* next value */
 147         *ptr = a - t - 4;
 148         t = n;
 149     }
 150 }
 151
 152 ST_FUNC void gsym(int t)
 153 {
 154     gsym_addr(t, ind);
 155 }
 156
 157 /* psym is used to put an instruction with a data field which is a
 158    reference to a symbol. It is in fact the same as oad ! */
 159 #define psym oad
 160
 161 /* instruction + 4 bytes data. Return the address of the data */
 162 ST_FUNC int oad(int c, int s)
 163 {
 164     int ind1;
 165
 166     o(c);
 167     ind1 = ind + 4;
 168     if (ind1 > cur_text_section->data_allocated)
 169         section_realloc(cur_text_section, ind1);
 170     *(int *)(cur_text_section->data + ind) = s;
 171     s = ind;
 172     ind = ind1;
 173     return s;
 174 }
 175
 176 /* output constant with relocation if 'r & VT_SYM' is true */
 177 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
 178 {
 179     if (r & VT_SYM)
 180         greloc(cur_text_section, sym, ind, R_386_32);
 181     gen_le32(c);
 182 }
 183
 184 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 185 {
 186     if (r & VT_SYM)
 187         greloc(cur_text_section, sym, ind, R_386_PC32);
 188     gen_le32(c - 4);
 189 }
 190
 191 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 192    opcode bits */
 193 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 194 {
 195     op_reg = op_reg << 3;
 196     if ((r & VT_VALMASK) == VT_CONST) {
 197         /* constant memory reference */
 198         o(0x05 | op_reg);
 199         gen_addr32(r, sym, c);
 200     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 201         /* currently, we use only ebp as base */
 202         if (c == (char)c) {
 203             /* short reference */
 204             o(0x45 | op_reg);
 205             g(c);
 206         } else {
 207             oad(0x85 | op_reg, c);
 208         }
 209     } else {
 210         g(0x00 | op_reg | (r & VT_VALMASK));
 211     }
 212 }
 213
 214 /* load 'r' from value 'sv' */
 215 ST_FUNC void load(int r, SValue *sv)
 216 {
 217     int v, t, ft, fc, fr;
 218     SValue v1;
 219
 220 #ifdef TCC_TARGET_PE
 221     SValue v2;
 222     sv = pe_getimport(sv, &v2);
 223 #endif
 224
 225     fr = sv->r;
 226     ft = sv->type.t;
 227     fc = sv->c.ul;
 228
 229     v = fr & VT_VALMASK;
 230     if (fr & VT_LVAL) {
 231         if (v == VT_LLOCAL) {
 232             v1.type.t = VT_INT;
 233             v1.r = VT_LOCAL | VT_LVAL;
 234             v1.c.ul = fc;
 235             fr = r;
 236             if (!(reg_classes[fr] & RC_INT))
 237                 fr = get_reg(RC_INT);
 238             load(fr, &v1);
 239         }
 240         if ((ft & VT_BTYPE) == VT_FLOAT) {
 241             o(0xd9); /* flds */
 242             r = 0;
 243         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 244             o(0xdd); /* fldl */
 245             r = 0;
 246         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 247             o(0xdb); /* fldt */
 248             r = 5;
 249         } else if ((ft & VT_TYPE) == VT_BYTE || (ft & VT_TYPE) == VT_BOOL) {
 250             o(0xbe0f);   /* movsbl */
 251         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 252             o(0xb60f);   /* movzbl */
 253         } else if ((ft & VT_TYPE) == VT_SHORT) {
 254             o(0xbf0f);   /* movswl */
 255         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 256             o(0xb70f);   /* movzwl */
 257         } else {
 258             o(0x8b);     /* movl */
 259         }
 260         gen_modrm(r, fr, sv->sym, fc);
 261     } else {
 262         if (v == VT_CONST) {
 263             o(0xb8 + r); /* mov $xx, r */
 264             gen_addr32(fr, sv->sym, fc);
 265         } else if (v == VT_LOCAL) {
 266             if (fc) {
 267                 o(0x8d); /* lea xxx(%ebp), r */
 268                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 269             } else {
 270                 o(0x89);
 271                 o(0xe8 + r); /* mov %ebp, r */
 272             }
 273         } else if (v == VT_CMP) {
 274             oad(0xb8 + r, 0); /* mov $0, r */
 275             o(0x0f); /* setxx %br */
 276             o(fc);
 277             o(0xc0 + r);
 278         } else if (v == VT_JMP || v == VT_JMPI) {
 279             t = v & 1;
 280             oad(0xb8 + r, t); /* mov $1, r */
 281             o(0x05eb); /* jmp after */
 282             gsym(fc);
 283             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 284         } else if (v != r) {
 285             o(0x89);
 286             o(0xc0 + r + v * 8); /* mov v, r */
 287         }
 288     }
 289 }
 290
 291 /* store register 'r' in lvalue 'v' */
 292 ST_FUNC void store(int r, SValue *v)
 293 {
 294     int fr, bt, ft, fc;
 295
 296 #ifdef TCC_TARGET_PE
 297     SValue v2;
 298     v = pe_getimport(v, &v2);
 299 #endif
 300
 301     ft = v->type.t;
 302     fc = v->c.ul;
 303     fr = v->r & VT_VALMASK;
 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 static void gen_static_call(int v)
 344 {
 345     Sym *sym;
 346
 347     sym = external_global_sym(v, &func_old_type, 0);
 348     oad(0xe8, -4);
 349     greloc(cur_text_section, sym, ind-4, R_386_PC32);
 350 }
 351
 352 /* 'is_jmp' is '1' if it is a jump */
 353 static void gcall_or_jmp(int is_jmp)
 354 {
 355     int r;
 356     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 357         /* constant case */
 358         if (vtop->r & VT_SYM) {
 359             /* relocation case */
 360             greloc(cur_text_section, vtop->sym,
 361                    ind + 1, R_386_PC32);
 362         } else {
 363             /* put an empty PC32 relocation */
 364             put_elf_reloc(symtab_section, cur_text_section,
 365                           ind + 1, R_386_PC32, 0);
 366         }
 367         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 368     } else {
 369         /* otherwise, indirect call */
 370         r = gv(RC_INT);
 371         o(0xff); /* call/jmp *r */
 372         o(0xd0 + r + (is_jmp << 4));
 373     }
 374 }
 375
 376 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 377 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 378
 379 ST_FUNC int regargs_nregs(RegArgs *args)
 380 {
 381     return *args;
 382 }
 383
 384 /* Return the number of registers needed to return the struct, or 0 if
 385    returning via struct pointer. */
 386 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize, RegArgs *args)
 387 {
 388 #ifdef TCC_TARGET_PE
 389     int size, align;
 390
 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) {
 395         *args = 0;
 396     } else if (size > 4) {
 397         ret->ref = NULL;
 398         ret->t = VT_LLONG;
 399         *args = 1;
 400     } else {
 401         ret->ref = NULL;
 402         ret->t = VT_INT;
 403         *args = 1;
 404     }
 405 #else
 406     *ret_align = 1; // Never have to re-align return values for x86
 407     *args = 0;
 408 #endif
 409
 410     return *args != 0;
 411 }
 412
 413 /* Generate function call. The function address is pushed first, then
 414    all the parameters in call order. This functions pops all the
 415    parameters and the function address. */
 416 ST_FUNC void gfunc_call(int nb_args)
 417 {
 418     int size, align, r, args_size, i, func_call;
 419     Sym *func_sym;
 420
 421     args_size = 0;
 422     for(i = 0;i < nb_args; i++) {
 423         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 424             size = type_size(&vtop->type, &align);
 425             /* align to stack align size */
 426             size = (size + 3) & ~3;
 427             /* allocate the necessary size on stack */
 428             oad(0xec81, size); /* sub $xxx, %esp */
 429             /* generate structure store */
 430             r = get_reg(RC_INT);
 431             o(0x89); /* mov %esp, r */
 432             o(0xe0 + r);
 433             vset(&vtop->type, r | VT_LVAL, 0);
 434             vswap();
 435             vstore();
 436             args_size += size;
 437         } else if (is_float(vtop->type.t)) {
 438             gv(RC_FLOAT); /* only one float register */
 439             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 440                 size = 4;
 441             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 442                 size = 8;
 443             else
 444                 size = 12;
 445             oad(0xec81, size); /* sub $xxx, %esp */
 446             if (size == 12)
 447                 o(0x7cdb);
 448             else
 449                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 450             g(0x24);
 451             g(0x00);
 452             args_size += size;
 453         } else {
 454             /* simple type (currently always same size) */
 455             /* XXX: implicit cast ? */
 456             r = gv(RC_INT);
 457             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 458                 size = 8;
 459                 o(0x50 + vtop->r2); /* push r */
 460             } else {
 461                 size = 4;
 462             }
 463             o(0x50 + r); /* push r */
 464             args_size += size;
 465         }
 466         vtop--;
 467     }
 468     save_regs(0); /* save used temporary registers */
 469     func_sym = vtop->type.ref;
 470     func_call = func_sym->a.func_call;
 471     /* fast call case */
 472     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 473         func_call == FUNC_FASTCALLW) {
 474         int fastcall_nb_regs;
 475         uint8_t *fastcall_regs_ptr;
 476         if (func_call == FUNC_FASTCALLW) {
 477             fastcall_regs_ptr = fastcallw_regs;
 478             fastcall_nb_regs = 2;
 479         } else {
 480             fastcall_regs_ptr = fastcall_regs;
 481             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 482         }
 483         for(i = 0;i < fastcall_nb_regs; i++) {
 484             if (args_size <= 0)
 485                 break;
 486             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 487             /* XXX: incorrect for struct/floats */
 488             args_size -= 4;
 489         }
 490     }
 491 #ifndef TCC_TARGET_PE
 492     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 493         args_size -= 4;
 494 #endif
 495     gcall_or_jmp(0);
 496
 497     if (args_size && func_call != FUNC_STDCALL)
 498         gadd_sp(args_size);
 499     vtop--;
 500 }
 501
 502 #ifdef TCC_TARGET_PE
 503 #define FUNC_PROLOG_SIZE 10
 504 #else
 505 #define FUNC_PROLOG_SIZE 9
 506 #endif
 507
 508 /* generate function prolog of type 't' */
 509 ST_FUNC void gfunc_prolog(CType *func_type)
 510 {
 511     int addr, align, size, func_call, fastcall_nb_regs;
 512     int param_index, param_addr;
 513     uint8_t *fastcall_regs_ptr;
 514     Sym *sym;
 515     CType *type;
 516
 517     sym = func_type->ref;
 518     func_call = sym->a.func_call;
 519     addr = 8;
 520     loc = 0;
 521     func_vc = 0;
 522
 523     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 524         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 525         fastcall_regs_ptr = fastcall_regs;
 526     } else if (func_call == FUNC_FASTCALLW) {
 527         fastcall_nb_regs = 2;
 528         fastcall_regs_ptr = fastcallw_regs;
 529     } else {
 530         fastcall_nb_regs = 0;
 531         fastcall_regs_ptr = NULL;
 532     }
 533     param_index = 0;
 534
 535     ind += FUNC_PROLOG_SIZE;
 536     func_sub_sp_offset = ind;
 537     /* if the function returns a structure, then add an
 538        implicit pointer parameter */
 539     func_vt = sym->type;
 540     func_var = (sym->c == FUNC_ELLIPSIS);
 541 #ifdef TCC_TARGET_PE
 542     size = type_size(&func_vt,&align);
 543     if (((func_vt.t & VT_BTYPE) == VT_STRUCT) && (size > 8)) {
 544 #else
 545     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 546 #endif
 547         /* XXX: fastcall case ? */
 548         func_vc = addr;
 549         addr += 4;
 550         param_index++;
 551     }
 552     /* define parameters */
 553     while ((sym = sym->next) != NULL) {
 554         type = &sym->type;
 555         size = type_size(type, &align);
 556         size = (size + 3) & ~3;
 557 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 558         /* structs are passed as pointer */
 559         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 560             size = 4;
 561         }
 562 #endif
 563         if (param_index < fastcall_nb_regs) {
 564             /* save FASTCALL register */
 565             loc -= 4;
 566             o(0x89);     /* movl */
 567             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 568             param_addr = loc;
 569         } else {
 570             param_addr = addr;
 571             addr += size;
 572         }
 573         sym_push(sym->v & ~SYM_FIELD, type,
 574                  VT_LOCAL | lvalue_type(type->t), param_addr);
 575         param_index++;
 576     }
 577     func_ret_sub = 0;
 578     /* pascal type call ? */
 579     if (func_call == FUNC_STDCALL)
 580         func_ret_sub = addr - 8;
 581 #ifndef TCC_TARGET_PE
 582     else if (func_vc)
 583         func_ret_sub = 4;
 584 #endif
 585
 586 #ifdef CONFIG_TCC_BCHECK
 587     /* leave some room for bound checking code */
 588     if (tcc_state->do_bounds_check) {
 589         oad(0xb8, 0); /* lbound section pointer */
 590         oad(0xb8, 0); /* call to function */
 591         func_bound_offset = lbounds_section->data_offset;
 592     }
 593 #endif
 594 }
 595
 596 /* generate function epilog */
 597 ST_FUNC void gfunc_epilog(void)
 598 {
 599     addr_t v, saved_ind;
 600
 601 #ifdef CONFIG_TCC_BCHECK
 602     if (tcc_state->do_bounds_check
 603      && func_bound_offset != lbounds_section->data_offset) {
 604         addr_t saved_ind;
 605         addr_t *bounds_ptr;
 606         Sym *sym_data;
 607         /* add end of table info */
 608         bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
 609         *bounds_ptr = 0;
 610         /* generate bound local allocation */
 611         saved_ind = ind;
 612         ind = func_sub_sp_offset;
 613         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 614                                func_bound_offset, lbounds_section->data_offset);
 615         greloc(cur_text_section, sym_data,
 616                ind + 1, R_386_32);
 617         oad(0xb8, 0); /* mov %eax, xxx */
 618         gen_static_call(TOK___bound_local_new);
 619
 620         ind = saved_ind;
 621         /* generate bound check local freeing */
 622         o(0x5250); /* save returned value, if any */
 623         greloc(cur_text_section, sym_data,
 624                ind + 1, R_386_32);
 625         oad(0xb8, 0); /* mov %eax, xxx */
 626         gen_static_call(TOK___bound_local_delete);
 627
 628         o(0x585a); /* restore returned value, if any */
 629     }
 630 #endif
 631     o(0xc9); /* leave */
 632     if (func_ret_sub == 0) {
 633         o(0xc3); /* ret */
 634     } else {
 635         o(0xc2); /* ret n */
 636         g(func_ret_sub);
 637         g(func_ret_sub >> 8);
 638     }
 639     /* align local size to word & save local variables */
 640
 641     v = (-loc + 3) & -4;
 642     saved_ind = ind;
 643     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 644 #ifdef TCC_TARGET_PE
 645     if (v >= 4096) {
 646         oad(0xb8, v); /* mov stacksize, %eax */
 647         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 648     } else
 649 #endif
 650     {
 651         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 652         o(0xec81);  /* sub esp, stacksize */
 653         gen_le32(v);
 654 #if FUNC_PROLOG_SIZE == 10
 655         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 656 #endif
 657     }
 658     ind = saved_ind;
 659 }
 660
 661 /* generate a jump to a label */
 662 ST_FUNC int gjmp(int t)
 663 {
 664     return psym(0xe9, t);
 665 }
 666
 667 /* generate a jump to a fixed address */
 668 ST_FUNC void gjmp_addr(int a)
 669 {
 670     int r;
 671     r = a - ind - 2;
 672     if (r == (char)r) {
 673         g(0xeb);
 674         g(r);
 675     } else {
 676         oad(0xe9, a - ind - 5);
 677     }
 678 }
 679
 680 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 681 ST_FUNC int gtst(int inv, int t)
 682 {
 683     int v, *p;
 684
 685     v = vtop->r & VT_VALMASK;
 686     if (v == VT_CMP) {
 687         /* fast case : can jump directly since flags are set */
 688         g(0x0f);
 689         t = psym((vtop->c.i - 16) ^ inv, t);
 690     } else if (v == VT_JMP || v == VT_JMPI) {
 691         /* && or || optimization */
 692         if ((v & 1) == inv) {
 693             /* insert vtop->c jump list in t */
 694             p = &vtop->c.i;
 695             while (*p != 0)
 696                 p = (int *)(cur_text_section->data + *p);
 697             *p = t;
 698             t = vtop->c.i;
 699         } else {
 700             t = gjmp(t);
 701             gsym(vtop->c.i);
 702         }
 703     }
 704     vtop--;
 705     return t;
 706 }
 707
 708 /* generate an integer binary operation */
 709 ST_FUNC void gen_opi(int op)
 710 {
 711     int r, fr, opc, c;
 712
 713     switch(op) {
 714     case '+':
 715     case TOK_ADDC1: /* add with carry generation */
 716         opc = 0;
 717     gen_op8:
 718         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 719             /* constant case */
 720             vswap();
 721             r = gv(RC_INT);
 722             vswap();
 723             c = vtop->c.i;
 724             if (c == (char)c) {
 725                 /* generate inc and dec for smaller code */
 726                 if (c==1 && opc==0) {
 727                     o (0x40 | r); // inc
 728                 } else if (c==1 && opc==5) {
 729                     o (0x48 | r); // dec
 730                 } else {
 731                     o(0x83);
 732                     o(0xc0 | (opc << 3) | r);
 733                     g(c);
 734                 }
 735             } else {
 736                 o(0x81);
 737                 oad(0xc0 | (opc << 3) | r, c);
 738             }
 739         } else {
 740             gv2(RC_INT, RC_INT);
 741             r = vtop[-1].r;
 742             fr = vtop[0].r;
 743             o((opc << 3) | 0x01);
 744             o(0xc0 + r + fr * 8);
 745         }
 746         vtop--;
 747         if (op >= TOK_ULT && op <= TOK_GT) {
 748             vtop->r = VT_CMP;
 749             vtop->c.i = op;
 750         }
 751         break;
 752     case '-':
 753     case TOK_SUBC1: /* sub with carry generation */
 754         opc = 5;
 755         goto gen_op8;
 756     case TOK_ADDC2: /* add with carry use */
 757         opc = 2;
 758         goto gen_op8;
 759     case TOK_SUBC2: /* sub with carry use */
 760         opc = 3;
 761         goto gen_op8;
 762     case '&':
 763         opc = 4;
 764         goto gen_op8;
 765     case '^':
 766         opc = 6;
 767         goto gen_op8;
 768     case '|':
 769         opc = 1;
 770         goto gen_op8;
 771     case '*':
 772         gv2(RC_INT, RC_INT);
 773         r = vtop[-1].r;
 774         fr = vtop[0].r;
 775         vtop--;
 776         o(0xaf0f); /* imul fr, r */
 777         o(0xc0 + fr + r * 8);
 778         break;
 779     case TOK_SHL:
 780         opc = 4;
 781         goto gen_shift;
 782     case TOK_SHR:
 783         opc = 5;
 784         goto gen_shift;
 785     case TOK_SAR:
 786         opc = 7;
 787     gen_shift:
 788         opc = 0xc0 | (opc << 3);
 789         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 790             /* constant case */
 791             vswap();
 792             r = gv(RC_INT);
 793             vswap();
 794             c = vtop->c.i & 0x1f;
 795             o(0xc1); /* shl/shr/sar $xxx, r */
 796             o(opc | r);
 797             g(c);
 798         } else {
 799             /* we generate the shift in ecx */
 800             gv2(RC_INT, RC_ECX);
 801             r = vtop[-1].r;
 802             o(0xd3); /* shl/shr/sar %cl, r */
 803             o(opc | r);
 804         }
 805         vtop--;
 806         break;
 807     case '/':
 808     case TOK_UDIV:
 809     case TOK_PDIV:
 810     case '%':
 811     case TOK_UMOD:
 812     case TOK_UMULL:
 813         /* first operand must be in eax */
 814         /* XXX: need better constraint for second operand */
 815         gv2(RC_EAX, RC_ECX);
 816         r = vtop[-1].r;
 817         fr = vtop[0].r;
 818         vtop--;
 819         save_reg(TREG_EDX);
 820         if (op == TOK_UMULL) {
 821             o(0xf7); /* mul fr */
 822             o(0xe0 + fr);
 823             vtop->r2 = TREG_EDX;
 824             r = TREG_EAX;
 825         } else {
 826             if (op == TOK_UDIV || op == TOK_UMOD) {
 827                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 828                 o(0xf0 + fr);
 829             } else {
 830                 o(0xf799); /* cltd, idiv fr, %eax */
 831                 o(0xf8 + fr);
 832             }
 833             if (op == '%' || op == TOK_UMOD)
 834                 r = TREG_EDX;
 835             else
 836                 r = TREG_EAX;
 837         }
 838         vtop->r = r;
 839         break;
 840     default:
 841         opc = 7;
 842         goto gen_op8;
 843     }
 844 }
 845
 846 /* generate a floating point operation 'v = t1 op t2' instruction. The
 847    two operands are guaranted to have the same floating point type */
 848 /* XXX: need to use ST1 too */
 849 ST_FUNC void gen_opf(int op)
 850 {
 851     int a, ft, fc, swapped, r;
 852
 853     /* convert constants to memory references */
 854     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 855         vswap();
 856         gv(RC_FLOAT);
 857         vswap();
 858     }
 859     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 860         gv(RC_FLOAT);
 861
 862     /* must put at least one value in the floating point register */
 863     if ((vtop[-1].r & VT_LVAL) &&
 864         (vtop[0].r & VT_LVAL)) {
 865         vswap();
 866         gv(RC_FLOAT);
 867         vswap();
 868     }
 869     swapped = 0;
 870     /* swap the stack if needed so that t1 is the register and t2 is
 871        the memory reference */
 872     if (vtop[-1].r & VT_LVAL) {
 873         vswap();
 874         swapped = 1;
 875     }
 876     if (op >= TOK_ULT && op <= TOK_GT) {
 877         /* load on stack second operand */
 878         load(TREG_ST0, vtop);
 879         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 880         if (op == TOK_GE || op == TOK_GT)
 881             swapped = !swapped;
 882         else if (op == TOK_EQ || op == TOK_NE)
 883             swapped = 0;
 884         if (swapped)
 885             o(0xc9d9); /* fxch %st(1) */
 886         if (op == TOK_EQ || op == TOK_NE)
 887             o(0xe9da); /* fucompp */
 888         else
 889             o(0xd9de); /* fcompp */
 890         o(0xe0df); /* fnstsw %ax */
 891         if (op == TOK_EQ) {
 892             o(0x45e480); /* and $0x45, %ah */
 893             o(0x40fC80); /* cmp $0x40, %ah */
 894         } else if (op == TOK_NE) {
 895             o(0x45e480); /* and $0x45, %ah */
 896             o(0x40f480); /* xor $0x40, %ah */
 897             op = TOK_NE;
 898         } else if (op == TOK_GE || op == TOK_LE) {
 899             o(0x05c4f6); /* test $0x05, %ah */
 900             op = TOK_EQ;
 901         } else {
 902             o(0x45c4f6); /* test $0x45, %ah */
 903             op = TOK_EQ;
 904         }
 905         vtop--;
 906         vtop->r = VT_CMP;
 907         vtop->c.i = op;
 908     } else {
 909         /* no memory reference possible for long double operations */
 910         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 911             load(TREG_ST0, vtop);
 912             swapped = !swapped;
 913         }
 914
 915         switch(op) {
 916         default:
 917         case '+':
 918             a = 0;
 919             break;
 920         case '-':
 921             a = 4;
 922             if (swapped)
 923                 a++;
 924             break;
 925         case '*':
 926             a = 1;
 927             break;
 928         case '/':
 929             a = 6;
 930             if (swapped)
 931                 a++;
 932             break;
 933         }
 934         ft = vtop->type.t;
 935         fc = vtop->c.ul;
 936         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 937             o(0xde); /* fxxxp %st, %st(1) */
 938             o(0xc1 + (a << 3));
 939         } else {
 940             /* if saved lvalue, then we must reload it */
 941             r = vtop->r;
 942             if ((r & VT_VALMASK) == VT_LLOCAL) {
 943                 SValue v1;
 944                 r = get_reg(RC_INT);
 945                 v1.type.t = VT_INT;
 946                 v1.r = VT_LOCAL | VT_LVAL;
 947                 v1.c.ul = fc;
 948                 load(r, &v1);
 949                 fc = 0;
 950             }
 951
 952             if ((ft & VT_BTYPE) == VT_DOUBLE)
 953                 o(0xdc);
 954             else
 955                 o(0xd8);
 956             gen_modrm(a, r, vtop->sym, fc);
 957         }
 958         vtop--;
 959     }
 960 }
 961
 962 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 963    and 'long long' cases. */
 964 ST_FUNC void gen_cvt_itof(int t)
 965 {
 966     save_reg(TREG_ST0);
 967     gv(RC_INT);
 968     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 969         /* signed long long to float/double/long double (unsigned case
 970            is handled generically) */
 971         o(0x50 + vtop->r2); /* push r2 */
 972         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 973         o(0x242cdf); /* fildll (%esp) */
 974         o(0x08c483); /* add $8, %esp */
 975     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 976                (VT_INT | VT_UNSIGNED)) {
 977         /* unsigned int to float/double/long double */
 978         o(0x6a); /* push $0 */
 979         g(0x00);
 980         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 981         o(0x242cdf); /* fildll (%esp) */
 982         o(0x08c483); /* add $8, %esp */
 983     } else {
 984         /* int to float/double/long double */
 985         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 986         o(0x2404db); /* fildl (%esp) */
 987         o(0x04c483); /* add $4, %esp */
 988     }
 989     vtop->r = TREG_ST0;
 990 }
 991
 992 /* convert fp to int 't' type */
 993 ST_FUNC void gen_cvt_ftoi(int t)
 994 {
 995  #ifndef COMMIT_4ad186c5ef61_IS_FIXED
 996     /* a good version but it takes a more time to execute */
 997     gv(RC_FLOAT);
 998     save_reg(TREG_EAX);
 999     save_reg(TREG_EDX);
1000     gen_static_call(TOK___tcc_cvt_ftol);
1001     vtop->r = TREG_EAX; /* mark reg as used */
1002     if (t == VT_LLONG)
1003         vtop->r2 = TREG_EDX;
1004  #else
1005     /* a new version with a bug: t2a = 44100312 */
1006     /*
1007     #include<stdio.h>
1008     int main() {
1009         int t1 = 176401255;
1010         float f = 0.25f;
1011         int t2a = (int)(t1 * f); // must be 44100313
1012         int t2b = (int)(t1 * (float)0.25f);
1013         printf("t2a=%d t2b=%d \n",t2a,t2b);
1014         return 0;
1015     }
1016     */
1017     int bt = vtop->type.t & VT_BTYPE;
1018     if (bt == VT_FLOAT)
1019         vpush_global_sym(&func_old_type, TOK___fixsfdi);
1020     else if (bt == VT_LDOUBLE)
1021         vpush_global_sym(&func_old_type, TOK___fixxfdi);
1022     else
1023         vpush_global_sym(&func_old_type, TOK___fixdfdi);
1024     vswap();
1025     gfunc_call(1);
1026     vpushi(0);
1027     vtop->r = REG_IRET;
1028     vtop->r2 = REG_LRET;
1029  #endif
1030 }
1031
1032 /* convert from one floating point type to another */
1033 ST_FUNC void gen_cvt_ftof(int t)
1034 {
1035     /* all we have to do on i386 is to put the float in a register */
1036     gv(RC_FLOAT);
1037 }
1038
1039 /* computed goto support */
1040 ST_FUNC void ggoto(void)
1041 {
1042     gcall_or_jmp(1);
1043     vtop--;
1044 }
1045
1046 /* bound check support functions */
1047 #ifdef CONFIG_TCC_BCHECK
1048
1049 /* generate a bounded pointer addition */
1050 ST_FUNC void gen_bounded_ptr_add(void)
1051 {
1052     /* prepare fast i386 function call (args in eax and edx) */
1053     gv2(RC_EAX, RC_EDX);
1054     /* save all temporary registers */
1055     vtop -= 2;
1056     save_regs(0);
1057     /* do a fast function call */
1058     gen_static_call(TOK___bound_ptr_add);
1059     /* returned pointer is in eax */
1060     vtop++;
1061     vtop->r = TREG_EAX | VT_BOUNDED;
1062     /* address of bounding function call point */
1063     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1064 }
1065
1066 /* patch pointer addition in vtop so that pointer dereferencing is
1067    also tested */
1068 ST_FUNC void gen_bounded_ptr_deref(void)
1069 {
1070     addr_t func;
1071     addr_t size, align;
1072     Elf32_Rel *rel;
1073     Sym *sym;
1074
1075     size = 0;
1076     /* XXX: put that code in generic part of tcc */
1077     if (!is_float(vtop->type.t)) {
1078         if (vtop->r & VT_LVAL_BYTE)
1079             size = 1;
1080         else if (vtop->r & VT_LVAL_SHORT)
1081             size = 2;
1082     }
1083     if (!size)
1084         size = type_size(&vtop->type, &align);
1085     switch(size) {
1086     case  1: func = TOK___bound_ptr_indir1; break;
1087     case  2: func = TOK___bound_ptr_indir2; break;
1088     case  4: func = TOK___bound_ptr_indir4; break;
1089     case  8: func = TOK___bound_ptr_indir8; break;
1090     case 12: func = TOK___bound_ptr_indir12; break;
1091     case 16: func = TOK___bound_ptr_indir16; break;
1092     default:
1093         tcc_error("unhandled size when dereferencing bounded pointer");
1094         func = 0;
1095         break;
1096     }
1097
1098     /* patch relocation */
1099     /* XXX: find a better solution ? */
1100     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1101     sym = external_global_sym(func, &func_old_type, 0);
1102     if (!sym->c)
1103         put_extern_sym(sym, NULL, 0, 0);
1104     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1105 }
1106 #endif
1107
1108 /* Save the stack pointer onto the stack */
1109 ST_FUNC void gen_vla_sp_save(int addr) {
1110     /* mov %esp,addr(%ebp)*/
1111     o(0x89);
1112     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1113 }
1114
1115 /* Restore the SP from a location on the stack */
1116 ST_FUNC void gen_vla_sp_restore(int addr) {
1117     o(0x8b);
1118     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1119 }
1120
1121 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1122 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1123 #ifdef TCC_TARGET_PE
1124     /* alloca does more than just adjust %rsp on Windows */
1125     vpush_global_sym(&func_old_type, TOK_alloca);
1126     vswap(); /* Move alloca ref past allocation size */
1127     gfunc_call(1);
1128 #else
1129     int r;
1130     r = gv(RC_INT); /* allocation size */
1131     /* sub r,%rsp */
1132     o(0x2b);
1133     o(0xe0 | r);
1134     /* We align to 16 bytes rather than align */
1135     /* and ~15, %esp */
1136     o(0xf0e483);
1137     vpop();
1138 #endif
1139 }
1140
1141 /* end of X86 code generator */
1142 /*************************************************************/
1143 #endif
1144 /*************************************************************/