src/lj_target_x86.h

   1 /*
   2 ** Definitions for x86 and x64 CPUs.
   3 ** Copyright (C) 2005-2012 Mike Pall. See Copyright Notice in luajit.h
   4 */
   5
   6 #ifndef _LJ_TARGET_X86_H
   7 #define _LJ_TARGET_X86_H
   8
   9 /* -- Registers IDs ------------------------------------------------------- */
  10
  11 #if LJ_64
  12 #define GPRDEF(_) \
  13   _(EAX) _(ECX) _(EDX) _(EBX) _(ESP) _(EBP) _(ESI) _(EDI) \
  14   _(R8D) _(R9D) _(R10D) _(R11D) _(R12D) _(R13D) _(R14D) _(R15D)
  15 #define FPRDEF(_) \
  16   _(XMM0) _(XMM1) _(XMM2) _(XMM3) _(XMM4) _(XMM5) _(XMM6) _(XMM7) \
  17   _(XMM8) _(XMM9) _(XMM10) _(XMM11) _(XMM12) _(XMM13) _(XMM14) _(XMM15)
  18 #else
  19 #define GPRDEF(_) \
  20   _(EAX) _(ECX) _(EDX) _(EBX) _(ESP) _(EBP) _(ESI) _(EDI)
  21 #define FPRDEF(_) \
  22   _(XMM0) _(XMM1) _(XMM2) _(XMM3) _(XMM4) _(XMM5) _(XMM6) _(XMM7)
  23 #endif
  24 #define VRIDDEF(_) \
  25   _(MRM)
  26
  27 #define RIDENUM(name)   RID_##name,
  28
  29 enum {
  30   GPRDEF(RIDENUM)               /* General-purpose registers (GPRs). */
  31   FPRDEF(RIDENUM)               /* Floating-point registers (FPRs). */
  32   RID_MAX,
  33   RID_MRM = RID_MAX,            /* Pseudo-id for ModRM operand. */
  34
  35   /* Calling conventions. */
  36   RID_RET = RID_EAX,
  37 #if LJ_64
  38   RID_FPRET = RID_XMM0,
  39 #else
  40   RID_RETLO = RID_EAX,
  41   RID_RETHI = RID_EDX,
  42 #endif
  43
  44   /* These definitions must match with the *.dasc file(s): */
  45   RID_BASE = RID_EDX,           /* Interpreter BASE. */
  46 #if LJ_64 && !LJ_ABI_WIN
  47   RID_LPC = RID_EBX,            /* Interpreter PC. */
  48   RID_DISPATCH = RID_R14D,      /* Interpreter DISPATCH table. */
  49 #else
  50   RID_LPC = RID_ESI,            /* Interpreter PC. */
  51   RID_DISPATCH = RID_EBX,       /* Interpreter DISPATCH table. */
  52 #endif
  53
  54   /* Register ranges [min, max) and number of registers. */
  55   RID_MIN_GPR = RID_EAX,
  56   RID_MIN_FPR = RID_XMM0,
  57   RID_MAX_GPR = RID_MIN_FPR,
  58   RID_MAX_FPR = RID_MAX,
  59   RID_NUM_GPR = RID_MAX_GPR - RID_MIN_GPR,
  60   RID_NUM_FPR = RID_MAX_FPR - RID_MIN_FPR,
  61 };
  62
  63 /* -- Register sets ------------------------------------------------------- */
  64
  65 /* Make use of all registers, except the stack pointer. */
  66 #define RSET_GPR        (RSET_RANGE(RID_MIN_GPR, RID_MAX_GPR)-RID2RSET(RID_ESP))
  67 #define RSET_FPR        (RSET_RANGE(RID_MIN_FPR, RID_MAX_FPR))
  68 #define RSET_ALL        (RSET_GPR|RSET_FPR)
  69 #define RSET_INIT       RSET_ALL
  70
  71 #if LJ_64
  72 /* Note: this requires the use of FORCE_REX! */
  73 #define RSET_GPR8       RSET_GPR
  74 #else
  75 #define RSET_GPR8       (RSET_RANGE(RID_EAX, RID_EBX+1))
  76 #endif
  77
  78 /* ABI-specific register sets. */
  79 #define RSET_ACD        (RID2RSET(RID_EAX)|RID2RSET(RID_ECX)|RID2RSET(RID_EDX))
  80 #if LJ_64
  81 #if LJ_ABI_WIN
  82 /* Windows x64 ABI. */
  83 #define RSET_SCRATCH \
  84   (RSET_ACD|RSET_RANGE(RID_R8D, RID_R11D+1)|RSET_RANGE(RID_XMM0, RID_XMM5+1))
  85 #define REGARG_GPRS \
  86   (RID_ECX|((RID_EDX|((RID_R8D|(RID_R9D<<5))<<5))<<5))
  87 #define REGARG_NUMGPR   4
  88 #define REGARG_NUMFPR   4
  89 #define REGARG_FIRSTFPR RID_XMM0
  90 #define REGARG_LASTFPR  RID_XMM3
  91 #define STACKARG_OFS    (4*8)
  92 #else
  93 /* The rest of the civilized x64 world has a common ABI. */
  94 #define RSET_SCRATCH \
  95   (RSET_ACD|RSET_RANGE(RID_ESI, RID_R11D+1)|RSET_FPR)
  96 #define REGARG_GPRS \
  97   (RID_EDI|((RID_ESI|((RID_EDX|((RID_ECX|((RID_R8D|(RID_R9D \
  98    <<5))<<5))<<5))<<5))<<5))
  99 #define REGARG_NUMGPR   6
 100 #define REGARG_NUMFPR   8
 101 #define REGARG_FIRSTFPR RID_XMM0
 102 #define REGARG_LASTFPR  RID_XMM7
 103 #define STACKARG_OFS    0
 104 #endif
 105 #else
 106 /* Common x86 ABI. */
 107 #define RSET_SCRATCH    (RSET_ACD|RSET_FPR)
 108 #define REGARG_GPRS     (RID_ECX|(RID_EDX<<5))  /* Fastcall only. */
 109 #define REGARG_NUMGPR   2  /* Fastcall only. */
 110 #define REGARG_NUMFPR   0
 111 #define STACKARG_OFS    0
 112 #endif
 113
 114 #if LJ_64
 115 /* Prefer the low 8 regs of each type to reduce REX prefixes. */
 116 #undef rset_picktop
 117 #define rset_picktop(rs)        (lj_fls(lj_bswap(rs)) ^ 0x18)
 118 #endif
 119
 120 /* -- Spill slots --------------------------------------------------------- */
 121
 122 /* Spill slots are 32 bit wide. An even/odd pair is used for FPRs.
 123 **
 124 ** SPS_FIXED: Available fixed spill slots in interpreter frame.
 125 ** This definition must match with the *.dasc file(s).
 126 **
 127 ** SPS_FIRST: First spill slot for general use. Reserve min. two 32 bit slots.
 128 */
 129 #if LJ_64
 130 #if LJ_ABI_WIN
 131 #define SPS_FIXED       (4*2)
 132 #define SPS_FIRST       (4*2)   /* Don't use callee register save area. */
 133 #else
 134 #define SPS_FIXED       4
 135 #define SPS_FIRST       2
 136 #endif
 137 #else
 138 #define SPS_FIXED       6
 139 #define SPS_FIRST       2
 140 #endif
 141
 142 #define SPOFS_TMP       0
 143
 144 #define sps_scale(slot)         (4 * (int32_t)(slot))
 145 #define sps_align(slot)         (((slot) - SPS_FIXED + 3) & ~3)
 146
 147 /* -- Exit state ---------------------------------------------------------- */
 148
 149 /* This definition must match with the *.dasc file(s). */
 150 typedef struct {
 151   lua_Number fpr[RID_NUM_FPR];  /* Floating-point registers. */
 152   intptr_t gpr[RID_NUM_GPR];    /* General-purpose registers. */
 153   int32_t spill[256];           /* Spill slots. */
 154 } ExitState;
 155
 156 /* Limited by the range of a short fwd jump (127): (2+2)*(32-1)-2 = 122. */
 157 #define EXITSTUB_SPACING        (2+2)
 158 #define EXITSTUBS_PER_GROUP     32
 159
 160 /* -- x86 ModRM operand encoding ------------------------------------------ */
 161
 162 typedef enum {
 163   XM_OFS0 = 0x00, XM_OFS8 = 0x40, XM_OFS32 = 0x80, XM_REG = 0xc0,
 164   XM_SCALE1 = 0x00, XM_SCALE2 = 0x40, XM_SCALE4 = 0x80, XM_SCALE8 = 0xc0,
 165   XM_MASK = 0xc0
 166 } x86Mode;
 167
 168 /* Structure to hold variable ModRM operand. */
 169 typedef struct {
 170   int32_t ofs;          /* Offset. */
 171   uint8_t base;         /* Base register or RID_NONE. */
 172   uint8_t idx;          /* Index register or RID_NONE. */
 173   uint8_t scale;        /* Index scale (XM_SCALE1 .. XM_SCALE8). */
 174 } x86ModRM;
 175
 176 /* -- Opcodes ------------------------------------------------------------- */
 177
 178 /* Macros to construct variable-length x86 opcodes. -(len+1) is in LSB. */
 179 #define XO_(o)          ((uint32_t)(0x0000fe + (0x##o<<24)))
 180 #define XO_FPU(a,b)     ((uint32_t)(0x00fd + (0x##a<<16)+(0x##b<<24)))
 181 #define XO_0f(o)        ((uint32_t)(0x0f00fd + (0x##o<<24)))
 182 #define XO_66(o)        ((uint32_t)(0x6600fd + (0x##o<<24)))
 183 #define XO_660f(o)      ((uint32_t)(0x0f66fc + (0x##o<<24)))
 184 #define XO_f20f(o)      ((uint32_t)(0x0ff2fc + (0x##o<<24)))
 185 #define XO_f30f(o)      ((uint32_t)(0x0ff3fc + (0x##o<<24)))
 186
 187 /* This list of x86 opcodes is not intended to be complete. Opcodes are only
 188 ** included when needed. Take a look at DynASM or jit.dis_x86 to see the
 189 ** whole mess.
 190 */
 191 typedef enum {
 192   /* Fixed length opcodes. XI_* prefix. */
 193   XI_NOP =      0x90,
 194   XI_XCHGa =    0x90,
 195   XI_CALL =     0xe8,
 196   XI_JMP =      0xe9,
 197   XI_JMPs =     0xeb,
 198   XI_PUSH =     0x50, /* Really 50+r. */
 199   XI_JCCs =     0x70, /* Really 7x. */
 200   XI_JCCn =     0x80, /* Really 0f8x. */
 201   XI_LEA =      0x8d,
 202   XI_MOVrib =   0xb0, /* Really b0+r. */
 203   XI_MOVri =    0xb8, /* Really b8+r. */
 204   XI_ARITHib =  0x80,
 205   XI_ARITHi =   0x81,
 206   XI_ARITHi8 =  0x83,
 207   XI_PUSHi8 =   0x6a,
 208   XI_TEST =     0x85,
 209   XI_MOVmi =    0xc7,
 210   XI_GROUP5 =   0xff,
 211
 212   /* Note: little-endian byte-order! */
 213   XI_FLDZ =     0xeed9,
 214   XI_FLD1 =     0xe8d9,
 215   XI_FLDLG2 =   0xecd9,
 216   XI_FLDLN2 =   0xedd9,
 217   XI_FDUP =     0xc0d9,  /* Really fld st0. */
 218   XI_FPOP =     0xd8dd,  /* Really fstp st0. */
 219   XI_FPOP1 =    0xd9dd,  /* Really fstp st1. */
 220   XI_FRNDINT =  0xfcd9,
 221   XI_FSIN =     0xfed9,
 222   XI_FCOS =     0xffd9,
 223   XI_FPTAN =    0xf2d9,
 224   XI_FPATAN =   0xf3d9,
 225   XI_FSCALE =   0xfdd9,
 226   XI_FYL2X =    0xf1d9,
 227
 228   /* Variable-length opcodes. XO_* prefix. */
 229   XO_MOV =      XO_(8b),
 230   XO_MOVto =    XO_(89),
 231   XO_MOVtow =   XO_66(89),
 232   XO_MOVtob =   XO_(88),
 233   XO_MOVmi =    XO_(c7),
 234   XO_MOVmib =   XO_(c6),
 235   XO_LEA =      XO_(8d),
 236   XO_ARITHib =  XO_(80),
 237   XO_ARITHi =   XO_(81),
 238   XO_ARITHi8 =  XO_(83),
 239   XO_ARITHiw8 = XO_66(83),
 240   XO_SHIFTi =   XO_(c1),
 241   XO_SHIFT1 =   XO_(d1),
 242   XO_SHIFTcl =  XO_(d3),
 243   XO_IMUL =     XO_0f(af),
 244   XO_IMULi =    XO_(69),
 245   XO_IMULi8 =   XO_(6b),
 246   XO_CMP =      XO_(3b),
 247   XO_TEST =     XO_(85),
 248   XO_GROUP3b =  XO_(f6),
 249   XO_GROUP3 =   XO_(f7),
 250   XO_GROUP5b =  XO_(fe),
 251   XO_GROUP5 =   XO_(ff),
 252   XO_MOVZXb =   XO_0f(b6),
 253   XO_MOVZXw =   XO_0f(b7),
 254   XO_MOVSXb =   XO_0f(be),
 255   XO_MOVSXw =   XO_0f(bf),
 256   XO_MOVSXd =   XO_(63),
 257   XO_BSWAP =    XO_0f(c8),
 258   XO_CMOV =     XO_0f(40),
 259
 260   XO_MOVSD =    XO_f20f(10),
 261   XO_MOVSDto =  XO_f20f(11),
 262   XO_MOVSS =    XO_f30f(10),
 263   XO_MOVSSto =  XO_f30f(11),
 264   XO_MOVLPD =   XO_660f(12),
 265   XO_MOVAPS =   XO_0f(28),
 266   XO_XORPS =    XO_0f(57),
 267   XO_ANDPS =    XO_0f(54),
 268   XO_ADDSD =    XO_f20f(58),
 269   XO_SUBSD =    XO_f20f(5c),
 270   XO_MULSD =    XO_f20f(59),
 271   XO_DIVSD =    XO_f20f(5e),
 272   XO_SQRTSD =   XO_f20f(51),
 273   XO_MINSD =    XO_f20f(5d),
 274   XO_MAXSD =    XO_f20f(5f),
 275   XO_ROUNDSD =  0x0b3a0ffc,  /* Really 66 0f 3a 0b. See asm_fpmath. */
 276   XO_UCOMISD =  XO_660f(2e),
 277   XO_CVTSI2SD = XO_f20f(2a),
 278   XO_CVTSD2SI = XO_f20f(2d),
 279   XO_CVTTSD2SI= XO_f20f(2c),
 280   XO_CVTSI2SS = XO_f30f(2a),
 281   XO_CVTSS2SI = XO_f30f(2d),
 282   XO_CVTTSS2SI= XO_f30f(2c),
 283   XO_CVTSS2SD = XO_f30f(5a),
 284   XO_CVTSD2SS = XO_f20f(5a),
 285   XO_ADDSS =    XO_f30f(58),
 286   XO_MOVD =     XO_660f(6e),
 287   XO_MOVDto =   XO_660f(7e),
 288
 289   XO_FLDd =     XO_(d9), XOg_FLDd = 0,
 290   XO_FLDq =     XO_(dd), XOg_FLDq = 0,
 291   XO_FILDd =    XO_(db), XOg_FILDd = 0,
 292   XO_FILDq =    XO_(df), XOg_FILDq = 5,
 293   XO_FSTPd =    XO_(d9), XOg_FSTPd = 3,
 294   XO_FSTPq =    XO_(dd), XOg_FSTPq = 3,
 295   XO_FISTPq =   XO_(df), XOg_FISTPq = 7,
 296   XO_FISTTPq =  XO_(dd), XOg_FISTTPq = 1,
 297   XO_FADDq =    XO_(dc), XOg_FADDq = 0,
 298   XO_FLDCW =    XO_(d9), XOg_FLDCW = 5,
 299   XO_FNSTCW =   XO_(d9), XOg_FNSTCW = 7
 300 } x86Op;
 301
 302 /* x86 opcode groups. */
 303 typedef uint32_t x86Group;
 304
 305 #define XG_(i8, i, g)   ((x86Group)(((i8) << 16) + ((i) << 8) + (g)))
 306 #define XG_ARITHi(g)    XG_(XI_ARITHi8, XI_ARITHi, g)
 307 #define XG_TOXOi(xg)    ((x86Op)(0x000000fe + (((xg)<<16) & 0xff000000)))
 308 #define XG_TOXOi8(xg)   ((x86Op)(0x000000fe + (((xg)<<8) & 0xff000000)))
 309
 310 #define XO_ARITH(a)     ((x86Op)(0x030000fe + ((a)<<27)))
 311 #define XO_ARITHw(a)    ((x86Op)(0x036600fd + ((a)<<27)))
 312
 313 typedef enum {
 314   XOg_ADD, XOg_OR, XOg_ADC, XOg_SBB, XOg_AND, XOg_SUB, XOg_XOR, XOg_CMP,
 315   XOg_X_IMUL
 316 } x86Arith;
 317
 318 typedef enum {
 319   XOg_ROL, XOg_ROR, XOg_RCL, XOg_RCR, XOg_SHL, XOg_SHR, XOg_SAL, XOg_SAR
 320 } x86Shift;
 321
 322 typedef enum {
 323   XOg_TEST, XOg_TEST_, XOg_NOT, XOg_NEG, XOg_MUL, XOg_IMUL, XOg_DIV, XOg_IDIV
 324 } x86Group3;
 325
 326 typedef enum {
 327   XOg_INC, XOg_DEC, XOg_CALL, XOg_CALLfar, XOg_JMP, XOg_JMPfar, XOg_PUSH
 328 } x86Group5;
 329
 330 /* x86 condition codes. */
 331 typedef enum {
 332   CC_O, CC_NO, CC_B, CC_NB, CC_E, CC_NE, CC_BE, CC_NBE,
 333   CC_S, CC_NS, CC_P, CC_NP, CC_L, CC_NL, CC_LE, CC_NLE,
 334   CC_C = CC_B, CC_NAE = CC_C, CC_NC = CC_NB, CC_AE = CC_NB,
 335   CC_Z = CC_E, CC_NZ = CC_NE, CC_NA = CC_BE, CC_A = CC_NBE,
 336   CC_PE = CC_P, CC_PO = CC_NP, CC_NGE = CC_L, CC_GE = CC_NL,
 337   CC_NG = CC_LE, CC_G = CC_NLE
 338 } x86CC;
 339
 340 #endif