sneedmca/Lime3DS
1
0
Fork 0
mirror of https://github.com/Lime3DS/Lime3DS synced 2024-10-31 04:07:51 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 8

emitter: Pass OpArg by reference where possible

Browse source
This commit is contained in:
Lioncash 2015-08-20 04:14:51 -04:00
parent 0609b89f74
commit 55461125ba
2 changed files with 721 additions and 721 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

666
src/common/x64/emitter.cpp
View file

@ -374,7 +374,7 @@ void XEmitter::Rex(int w, int r, int x, int b)
Write8(rx);
}
void XEmitter::JMP(const u8 *addr, bool force5Bytes)
void XEmitter::JMP(const u8* addr, bool force5Bytes)
{
u64 fn = (u64)addr;
if (!force5Bytes)
@ -398,7 +398,7 @@ void XEmitter::JMP(const u8 *addr, bool force5Bytes)
}
}
void XEmitter::JMPptr(const OpArg &arg2)
void XEmitter::JMPptr(const OpArg& arg2)
{
OpArg arg = arg2;
if (arg.IsImm()) ASSERT_MSG(0, "JMPptr - Imm argument");
@ -425,7 +425,7 @@ void XEmitter::CALLptr(OpArg arg)
arg.WriteRest(this);
}
void XEmitter::CALL(const void *fnptr)
void XEmitter::CALL(const void* fnptr)
{
u64 distance = u64(fnptr) - (u64(code) + 5);
ASSERT_MSG(
@ -496,7 +496,7 @@ void XEmitter::J_CC(CCFlags conditionCode, const u8* addr, bool force5bytes)
}
}
void XEmitter::SetJumpTarget(const FixupBranch &branch)
void XEmitter::SetJumpTarget(const FixupBranch& branch)
{
if (branch.type == 0)
{
@ -667,7 +667,7 @@ void XEmitter::CBW(int bits)
void XEmitter::PUSH(X64Reg reg) {WriteSimple1Byte(32, 0x50, reg);}
void XEmitter::POP(X64Reg reg) {WriteSimple1Byte(32, 0x58, reg);}
void XEmitter::PUSH(int bits, const OpArg &reg)
void XEmitter::PUSH(int bits, const OpArg& reg)
{
if (reg.IsSimpleReg())
PUSH(reg.GetSimpleReg());
@ -703,7 +703,7 @@ void XEmitter::PUSH(int bits, const OpArg &reg)
}
}
void XEmitter::POP(int /*bits*/, const OpArg &reg)
void XEmitter::POP(int /*bits*/, const OpArg& reg)
{
if (reg.IsSimpleReg())
POP(reg.GetSimpleReg());
@ -791,12 +791,12 @@ void XEmitter::WriteMulDivType(int bits, OpArg src, int ext)
src.WriteRest(this);
}
void XEmitter::MUL(int bits, OpArg src) {WriteMulDivType(bits, src, 4);}
void XEmitter::DIV(int bits, OpArg src) {WriteMulDivType(bits, src, 6);}
void XEmitter::IMUL(int bits, OpArg src) {WriteMulDivType(bits, src, 5);}
void XEmitter::IDIV(int bits, OpArg src) {WriteMulDivType(bits, src, 7);}
void XEmitter::NEG(int bits, OpArg src) {WriteMulDivType(bits, src, 3);}
void XEmitter::NOT(int bits, OpArg src) {WriteMulDivType(bits, src, 2);}
void XEmitter::MUL(int bits, const OpArg& src) {WriteMulDivType(bits, src, 4);}
void XEmitter::DIV(int bits, const OpArg& src) {WriteMulDivType(bits, src, 6);}
void XEmitter::IMUL(int bits, const OpArg& src) {WriteMulDivType(bits, src, 5);}
void XEmitter::IDIV(int bits, const OpArg& src) {WriteMulDivType(bits, src, 7);}
void XEmitter::NEG(int bits, const OpArg& src) {WriteMulDivType(bits, src, 3);}
void XEmitter::NOT(int bits, const OpArg& src) {WriteMulDivType(bits, src, 2);}
void XEmitter::WriteBitSearchType(int bits, X64Reg dest, OpArg src, u8 byte2, bool rep)
{
@ -813,24 +813,24 @@ void XEmitter::WriteBitSearchType(int bits, X64Reg dest, OpArg src, u8 byte2, bo
src.WriteRest(this);
}
void XEmitter::MOVNTI(int bits, OpArg dest, X64Reg src)
void XEmitter::MOVNTI(int bits, const OpArg& dest, X64Reg src)
{
if (bits <= 16)
ASSERT_MSG(0, "MOVNTI - bits<=16");
WriteBitSearchType(bits, src, dest, 0xC3);
}
void XEmitter::BSF(int bits, X64Reg dest, OpArg src) {WriteBitSearchType(bits,dest,src,0xBC);} //bottom bit to top bit
void XEmitter::BSR(int bits, X64Reg dest, OpArg src) {WriteBitSearchType(bits,dest,src,0xBD);} //top bit to bottom bit
void XEmitter::BSF(int bits, X64Reg dest, const OpArg& src) {WriteBitSearchType(bits,dest,src,0xBC);} // Bottom bit to top bit
void XEmitter::BSR(int bits, X64Reg dest, const OpArg& src) {WriteBitSearchType(bits,dest,src,0xBD);} // Top bit to bottom bit
void XEmitter::TZCNT(int bits, X64Reg dest, OpArg src)
void XEmitter::TZCNT(int bits, X64Reg dest, const OpArg& src)
{
CheckFlags();
if (!Common::GetCPUCaps().bmi1)
ASSERT_MSG(0, "Trying to use BMI1 on a system that doesn't support it. Bad programmer.");
WriteBitSearchType(bits, dest, src, 0xBC, true);
}
void XEmitter::LZCNT(int bits, X64Reg dest, OpArg src)
void XEmitter::LZCNT(int bits, X64Reg dest, const OpArg& src)
{
CheckFlags();
if (!Common::GetCPUCaps().lzcnt)
@ -950,7 +950,7 @@ void XEmitter::LEA(int bits, X64Reg dest, OpArg src)
}
//shift can be either imm8 or cl
void XEmitter::WriteShift(int bits, OpArg dest, OpArg &shift, int ext)
void XEmitter::WriteShift(int bits, OpArg dest, const OpArg& shift, int ext)
{
CheckFlags();
bool writeImm = false;
@ -991,16 +991,16 @@ void XEmitter::WriteShift(int bits, OpArg dest, OpArg &shift, int ext)
// large rotates and shift are slower on intel than amd
// intel likes to rotate by 1, and the op is smaller too
void XEmitter::ROL(int bits, OpArg dest, OpArg shift) {WriteShift(bits, dest, shift, 0);}
void XEmitter::ROR(int bits, OpArg dest, OpArg shift) {WriteShift(bits, dest, shift, 1);}
void XEmitter::RCL(int bits, OpArg dest, OpArg shift) {WriteShift(bits, dest, shift, 2);}
void XEmitter::RCR(int bits, OpArg dest, OpArg shift) {WriteShift(bits, dest, shift, 3);}
void XEmitter::SHL(int bits, OpArg dest, OpArg shift) {WriteShift(bits, dest, shift, 4);}
void XEmitter::SHR(int bits, OpArg dest, OpArg shift) {WriteShift(bits, dest, shift, 5);}
void XEmitter::SAR(int bits, OpArg dest, OpArg shift) {WriteShift(bits, dest, shift, 7);}
void XEmitter::ROL(int bits, const OpArg& dest, const OpArg& shift) {WriteShift(bits, dest, shift, 0);}
void XEmitter::ROR(int bits, const OpArg& dest, const OpArg& shift) {WriteShift(bits, dest, shift, 1);}
void XEmitter::RCL(int bits, const OpArg& dest, const OpArg& shift) {WriteShift(bits, dest, shift, 2);}
void XEmitter::RCR(int bits, const OpArg& dest, const OpArg& shift) {WriteShift(bits, dest, shift, 3);}
void XEmitter::SHL(int bits, const OpArg& dest, const OpArg& shift) {WriteShift(bits, dest, shift, 4);}
void XEmitter::SHR(int bits, const OpArg& dest, const OpArg& shift) {WriteShift(bits, dest, shift, 5);}
void XEmitter::SAR(int bits, const OpArg& dest, const OpArg& shift) {WriteShift(bits, dest, shift, 7);}
// index can be either imm8 or register, don't use memory destination because it's slow
void XEmitter::WriteBitTest(int bits, OpArg &dest, OpArg &index, int ext)
void XEmitter::WriteBitTest(int bits, const OpArg& dest, const OpArg& index, int ext)
{
CheckFlags();
if (dest.IsImm())
@ -1029,13 +1029,13 @@ void XEmitter::WriteBitTest(int bits, OpArg &dest, OpArg &index, int ext)
}
}
void XEmitter::BT(int bits, OpArg dest, OpArg index) {WriteBitTest(bits, dest, index, 4);}
void XEmitter::BTS(int bits, OpArg dest, OpArg index) {WriteBitTest(bits, dest, index, 5);}
void XEmitter::BTR(int bits, OpArg dest, OpArg index) {WriteBitTest(bits, dest, index, 6);}
void XEmitter::BTC(int bits, OpArg dest, OpArg index) {WriteBitTest(bits, dest, index, 7);}
void XEmitter::BT(int bits, const OpArg& dest, const OpArg& index) {WriteBitTest(bits, dest, index, 4);}
void XEmitter::BTS(int bits, const OpArg& dest, const OpArg& index) {WriteBitTest(bits, dest, index, 5);}
void XEmitter::BTR(int bits, const OpArg& dest, const OpArg& index) {WriteBitTest(bits, dest, index, 6);}
void XEmitter::BTC(int bits, const OpArg& dest, const OpArg& index) {WriteBitTest(bits, dest, index, 7);}
//shift can be either imm8 or cl
void XEmitter::SHRD(int bits, OpArg dest, OpArg src, OpArg shift)
void XEmitter::SHRD(int bits, const OpArg& dest, const OpArg& src, const OpArg& shift)
{
CheckFlags();
if (dest.IsImm())
@ -1067,7 +1067,7 @@ void XEmitter::SHRD(int bits, OpArg dest, OpArg src, OpArg shift)
}
}
void XEmitter::SHLD(int bits, OpArg dest, OpArg src, OpArg shift)
void XEmitter::SHLD(int bits, const OpArg& dest, const OpArg& src, const OpArg& shift)
{
CheckFlags();
if (dest.IsImm())
@ -1111,7 +1111,7 @@ void OpArg::WriteSingleByteOp(XEmitter *emit, u8 op, X64Reg _operandReg, int bit
}
//operand can either be immediate or register
void OpArg::WriteNormalOp(XEmitter *emit, bool toRM, NormalOp op, const OpArg &operand, int bits) const
void OpArg::WriteNormalOp(XEmitter *emit, bool toRM, NormalOp op, const OpArg& operand, int bits) const
{
X64Reg _operandReg;
if (IsImm())
@ -1257,7 +1257,7 @@ void OpArg::WriteNormalOp(XEmitter *emit, bool toRM, NormalOp op, const OpArg &o
}
}
void XEmitter::WriteNormalOp(XEmitter *emit, int bits, NormalOp op, const OpArg &a1, const OpArg &a2)
void XEmitter::WriteNormalOp(XEmitter *emit, int bits, NormalOp op, const OpArg& a1, const OpArg& a2)
{
if (a1.IsImm())
{
@ -1283,24 +1283,24 @@ void XEmitter::WriteNormalOp(XEmitter *emit, int bits, NormalOp op, const OpArg
}
}
void XEmitter::ADD (int bits, const OpArg &a1, const OpArg &a2) {CheckFlags(); WriteNormalOp(this, bits, nrmADD, a1, a2);}
void XEmitter::ADC (int bits, const OpArg &a1, const OpArg &a2) {CheckFlags(); WriteNormalOp(this, bits, nrmADC, a1, a2);}
void XEmitter::SUB (int bits, const OpArg &a1, const OpArg &a2) {CheckFlags(); WriteNormalOp(this, bits, nrmSUB, a1, a2);}
void XEmitter::SBB (int bits, const OpArg &a1, const OpArg &a2) {CheckFlags(); WriteNormalOp(this, bits, nrmSBB, a1, a2);}
void XEmitter::AND (int bits, const OpArg &a1, const OpArg &a2) {CheckFlags(); WriteNormalOp(this, bits, nrmAND, a1, a2);}
void XEmitter::OR (int bits, const OpArg &a1, const OpArg &a2) {CheckFlags(); WriteNormalOp(this, bits, nrmOR , a1, a2);}
void XEmitter::XOR (int bits, const OpArg &a1, const OpArg &a2) {CheckFlags(); WriteNormalOp(this, bits, nrmXOR, a1, a2);}
void XEmitter::MOV (int bits, const OpArg &a1, const OpArg &a2)
void XEmitter::ADD (int bits, const OpArg& a1, const OpArg& a2) {CheckFlags(); WriteNormalOp(this, bits, nrmADD, a1, a2);}
void XEmitter::ADC (int bits, const OpArg& a1, const OpArg& a2) {CheckFlags(); WriteNormalOp(this, bits, nrmADC, a1, a2);}
void XEmitter::SUB (int bits, const OpArg& a1, const OpArg& a2) {CheckFlags(); WriteNormalOp(this, bits, nrmSUB, a1, a2);}
void XEmitter::SBB (int bits, const OpArg& a1, const OpArg& a2) {CheckFlags(); WriteNormalOp(this, bits, nrmSBB, a1, a2);}
void XEmitter::AND (int bits, const OpArg& a1, const OpArg& a2) {CheckFlags(); WriteNormalOp(this, bits, nrmAND, a1, a2);}
void XEmitter::OR (int bits, const OpArg& a1, const OpArg& a2) {CheckFlags(); WriteNormalOp(this, bits, nrmOR , a1, a2);}
void XEmitter::XOR (int bits, const OpArg& a1, const OpArg& a2) {CheckFlags(); WriteNormalOp(this, bits, nrmXOR, a1, a2);}
void XEmitter::MOV (int bits, const OpArg& a1, const OpArg& a2)
{
if (a1.IsSimpleReg() && a2.IsSimpleReg() && a1.GetSimpleReg() == a2.GetSimpleReg())
LOG_ERROR(Common, "Redundant MOV @ %p - bug in JIT?", code);
WriteNormalOp(this, bits, nrmMOV, a1, a2);
}
void XEmitter::TEST(int bits, const OpArg &a1, const OpArg &a2) {CheckFlags(); WriteNormalOp(this, bits, nrmTEST, a1, a2);}
void XEmitter::CMP (int bits, const OpArg &a1, const OpArg &a2) {CheckFlags(); WriteNormalOp(this, bits, nrmCMP, a1, a2);}
void XEmitter::XCHG(int bits, const OpArg &a1, const OpArg &a2) {WriteNormalOp(this, bits, nrmXCHG, a1, a2);}
void XEmitter::TEST(int bits, const OpArg& a1, const OpArg& a2) {CheckFlags(); WriteNormalOp(this, bits, nrmTEST, a1, a2);}
void XEmitter::CMP (int bits, const OpArg& a1, const OpArg& a2) {CheckFlags(); WriteNormalOp(this, bits, nrmCMP, a1, a2);}
void XEmitter::XCHG(int bits, const OpArg& a1, const OpArg& a2) {WriteNormalOp(this, bits, nrmXCHG, a1, a2);}
void XEmitter::IMUL(int bits, X64Reg regOp, OpArg a1, OpArg a2)
void XEmitter::IMUL(int bits, X64Reg regOp, const OpArg& a1, const OpArg& a2)
{
CheckFlags();
if (bits == 8)
@ -1353,7 +1353,7 @@ void XEmitter::IMUL(int bits, X64Reg regOp, OpArg a1, OpArg a2)
}
}
void XEmitter::IMUL(int bits, X64Reg regOp, OpArg a)
void XEmitter::IMUL(int bits, X64Reg regOp, const OpArg& a)
{
CheckFlags();
if (bits == 8)
@ -1390,7 +1390,7 @@ void XEmitter::WriteSSEOp(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extr
arg.WriteRest(this, extrabytes);
}
void XEmitter::WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extrabytes)
void XEmitter::WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp, const OpArg& arg, int extrabytes)
{
WriteAVXOp(opPrefix, op, regOp, INVALID_REG, arg, extrabytes);
}
@ -1418,7 +1418,7 @@ static int GetVEXpp(u8 opPrefix)
return 0;
}
void XEmitter::WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes)
void XEmitter::WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, const OpArg& arg, int extrabytes)
{
if (!Common::GetCPUCaps().avx)
ASSERT_MSG(0, "Trying to use AVX on a system that doesn't support it. Bad programmer.");
@ -1431,7 +1431,7 @@ void XEmitter::WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpA
}
// Like the above, but more general; covers GPR-based VEX operations, like BMI1/2
void XEmitter::WriteVEXOp(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes)
void XEmitter::WriteVEXOp(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, const OpArg& arg, int extrabytes)
{
if (size != 32 && size != 64)
ASSERT_MSG(0, "VEX GPR instructions only support 32-bit and 64-bit modes!");
@ -1442,7 +1442,7 @@ void XEmitter::WriteVEXOp(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg r
arg.WriteRest(this, extrabytes, regOp1);
}
void XEmitter::WriteBMI1Op(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes)
void XEmitter::WriteBMI1Op(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, const OpArg& arg, int extrabytes)
{
CheckFlags();
if (!Common::GetCPUCaps().bmi1)
@ -1450,7 +1450,7 @@ void XEmitter::WriteBMI1Op(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg
WriteVEXOp(size, opPrefix, op, regOp1, regOp2, arg, extrabytes);
}
void XEmitter::WriteBMI2Op(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes)
void XEmitter::WriteBMI2Op(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, const OpArg& arg, int extrabytes)
{
CheckFlags();
if (!Common::GetCPUCaps().bmi2)
@ -1517,135 +1517,135 @@ void XEmitter::WriteMXCSR(OpArg arg, int ext)
arg.WriteRest(this);
}
void XEmitter::STMXCSR(OpArg memloc) {WriteMXCSR(memloc, 3);}
void XEmitter::LDMXCSR(OpArg memloc) {WriteMXCSR(memloc, 2);}
void XEmitter::STMXCSR(const OpArg& memloc) {WriteMXCSR(memloc, 3);}
void XEmitter::LDMXCSR(const OpArg& memloc) {WriteMXCSR(memloc, 2);}
void XEmitter::MOVNTDQ(OpArg arg, X64Reg regOp) {WriteSSEOp(0x66, sseMOVNTDQ, regOp, arg);}
void XEmitter::MOVNTPS(OpArg arg, X64Reg regOp) {WriteSSEOp(0x00, sseMOVNTP, regOp, arg);}
void XEmitter::MOVNTPD(OpArg arg, X64Reg regOp) {WriteSSEOp(0x66, sseMOVNTP, regOp, arg);}
void XEmitter::MOVNTDQ(const OpArg& arg, X64Reg regOp) {WriteSSEOp(0x66, sseMOVNTDQ, regOp, arg);}
void XEmitter::MOVNTPS(const OpArg& arg, X64Reg regOp) {WriteSSEOp(0x00, sseMOVNTP, regOp, arg);}
void XEmitter::MOVNTPD(const OpArg& arg, X64Reg regOp) {WriteSSEOp(0x66, sseMOVNTP, regOp, arg);}
void XEmitter::ADDSS(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF3, sseADD, regOp, arg);}
void XEmitter::ADDSD(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF2, sseADD, regOp, arg);}
void XEmitter::SUBSS(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF3, sseSUB, regOp, arg);}
void XEmitter::SUBSD(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF2, sseSUB, regOp, arg);}
void XEmitter::CMPSS(X64Reg regOp, OpArg arg, u8 compare) {WriteSSEOp(0xF3, sseCMP, regOp, arg, 1); Write8(compare);}
void XEmitter::CMPSD(X64Reg regOp, OpArg arg, u8 compare) {WriteSSEOp(0xF2, sseCMP, regOp, arg, 1); Write8(compare);}
void XEmitter::MULSS(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF3, sseMUL, regOp, arg);}
void XEmitter::MULSD(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF2, sseMUL, regOp, arg);}
void XEmitter::DIVSS(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF3, sseDIV, regOp, arg);}
void XEmitter::DIVSD(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF2, sseDIV, regOp, arg);}
void XEmitter::MINSS(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF3, sseMIN, regOp, arg);}
void XEmitter::MINSD(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF2, sseMIN, regOp, arg);}
void XEmitter::MAXSS(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF3, sseMAX, regOp, arg);}
void XEmitter::MAXSD(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF2, sseMAX, regOp, arg);}
void XEmitter::SQRTSS(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF3, sseSQRT, regOp, arg);}
void XEmitter::SQRTSD(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF2, sseSQRT, regOp, arg);}
void XEmitter::RSQRTSS(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF3, sseRSQRT, regOp, arg);}
void XEmitter::ADDSS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF3, sseADD, regOp, arg);}
void XEmitter::ADDSD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF2, sseADD, regOp, arg);}
void XEmitter::SUBSS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF3, sseSUB, regOp, arg);}
void XEmitter::SUBSD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF2, sseSUB, regOp, arg);}
void XEmitter::CMPSS(X64Reg regOp, const OpArg& arg, u8 compare) {WriteSSEOp(0xF3, sseCMP, regOp, arg, 1); Write8(compare);}
void XEmitter::CMPSD(X64Reg regOp, const OpArg& arg, u8 compare) {WriteSSEOp(0xF2, sseCMP, regOp, arg, 1); Write8(compare);}
void XEmitter::MULSS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF3, sseMUL, regOp, arg);}
void XEmitter::MULSD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF2, sseMUL, regOp, arg);}
void XEmitter::DIVSS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF3, sseDIV, regOp, arg);}
void XEmitter::DIVSD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF2, sseDIV, regOp, arg);}
void XEmitter::MINSS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF3, sseMIN, regOp, arg);}
void XEmitter::MINSD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF2, sseMIN, regOp, arg);}
void XEmitter::MAXSS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF3, sseMAX, regOp, arg);}
void XEmitter::MAXSD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF2, sseMAX, regOp, arg);}
void XEmitter::SQRTSS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF3, sseSQRT, regOp, arg);}
void XEmitter::SQRTSD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF2, sseSQRT, regOp, arg);}
void XEmitter::RSQRTSS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF3, sseRSQRT, regOp, arg);}
void XEmitter::ADDPS(X64Reg regOp, OpArg arg) {WriteSSEOp(0x00, sseADD, regOp, arg);}
void XEmitter::ADDPD(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, sseADD, regOp, arg);}
void XEmitter::SUBPS(X64Reg regOp, OpArg arg) {WriteSSEOp(0x00, sseSUB, regOp, arg);}
void XEmitter::SUBPD(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, sseSUB, regOp, arg);}
void XEmitter::CMPPS(X64Reg regOp, OpArg arg, u8 compare) {WriteSSEOp(0x00, sseCMP, regOp, arg, 1); Write8(compare);}
void XEmitter::CMPPD(X64Reg regOp, OpArg arg, u8 compare) {WriteSSEOp(0x66, sseCMP, regOp, arg, 1); Write8(compare);}
void XEmitter::ANDPS(X64Reg regOp, OpArg arg) {WriteSSEOp(0x00, sseAND, regOp, arg);}
void XEmitter::ANDPD(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, sseAND, regOp, arg);}
void XEmitter::ANDNPS(X64Reg regOp, OpArg arg) {WriteSSEOp(0x00, sseANDN, regOp, arg);}
void XEmitter::ANDNPD(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, sseANDN, regOp, arg);}
void XEmitter::ORPS(X64Reg regOp, OpArg arg) {WriteSSEOp(0x00, sseOR, regOp, arg);}
void XEmitter::ORPD(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, sseOR, regOp, arg);}
void XEmitter::XORPS(X64Reg regOp, OpArg arg) {WriteSSEOp(0x00, sseXOR, regOp, arg);}
void XEmitter::XORPD(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, sseXOR, regOp, arg);}
void XEmitter::MULPS(X64Reg regOp, OpArg arg) {WriteSSEOp(0x00, sseMUL, regOp, arg);}
void XEmitter::MULPD(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, sseMUL, regOp, arg);}
void XEmitter::DIVPS(X64Reg regOp, OpArg arg) {WriteSSEOp(0x00, sseDIV, regOp, arg);}
void XEmitter::DIVPD(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, sseDIV, regOp, arg);}
void XEmitter::MINPS(X64Reg regOp, OpArg arg) {WriteSSEOp(0x00, sseMIN, regOp, arg);}
void XEmitter::MINPD(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, sseMIN, regOp, arg);}
void XEmitter::MAXPS(X64Reg regOp, OpArg arg) {WriteSSEOp(0x00, sseMAX, regOp, arg);}
void XEmitter::MAXPD(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, sseMAX, regOp, arg);}
void XEmitter::SQRTPS(X64Reg regOp, OpArg arg) {WriteSSEOp(0x00, sseSQRT, regOp, arg);}
void XEmitter::SQRTPD(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, sseSQRT, regOp, arg);}
void XEmitter::RCPPS(X64Reg regOp, OpArg arg) { WriteSSEOp(0x00, sseRCP, regOp, arg); }
void XEmitter::RSQRTPS(X64Reg regOp, OpArg arg) {WriteSSEOp(0x00, sseRSQRT, regOp, arg);}
void XEmitter::SHUFPS(X64Reg regOp, OpArg arg, u8 shuffle) {WriteSSEOp(0x00, sseSHUF, regOp, arg,1); Write8(shuffle);}
void XEmitter::SHUFPD(X64Reg regOp, OpArg arg, u8 shuffle) {WriteSSEOp(0x66, sseSHUF, regOp, arg,1); Write8(shuffle);}
void XEmitter::ADDPS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x00, sseADD, regOp, arg);}
void XEmitter::ADDPD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, sseADD, regOp, arg);}
void XEmitter::SUBPS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x00, sseSUB, regOp, arg);}
void XEmitter::SUBPD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, sseSUB, regOp, arg);}
void XEmitter::CMPPS(X64Reg regOp, const OpArg& arg, u8 compare) {WriteSSEOp(0x00, sseCMP, regOp, arg, 1); Write8(compare);}
void XEmitter::CMPPD(X64Reg regOp, const OpArg& arg, u8 compare) {WriteSSEOp(0x66, sseCMP, regOp, arg, 1); Write8(compare);}
void XEmitter::ANDPS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x00, sseAND, regOp, arg);}
void XEmitter::ANDPD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, sseAND, regOp, arg);}
void XEmitter::ANDNPS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x00, sseANDN, regOp, arg);}
void XEmitter::ANDNPD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, sseANDN, regOp, arg);}
void XEmitter::ORPS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x00, sseOR, regOp, arg);}
void XEmitter::ORPD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, sseOR, regOp, arg);}
void XEmitter::XORPS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x00, sseXOR, regOp, arg);}
void XEmitter::XORPD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, sseXOR, regOp, arg);}
void XEmitter::MULPS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x00, sseMUL, regOp, arg);}
void XEmitter::MULPD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, sseMUL, regOp, arg);}
void XEmitter::DIVPS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x00, sseDIV, regOp, arg);}
void XEmitter::DIVPD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, sseDIV, regOp, arg);}
void XEmitter::MINPS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x00, sseMIN, regOp, arg);}
void XEmitter::MINPD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, sseMIN, regOp, arg);}
void XEmitter::MAXPS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x00, sseMAX, regOp, arg);}
void XEmitter::MAXPD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, sseMAX, regOp, arg);}
void XEmitter::SQRTPS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x00, sseSQRT, regOp, arg);}
void XEmitter::SQRTPD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, sseSQRT, regOp, arg);}
void XEmitter::RCPPS(X64Reg regOp, const OpArg& arg) { WriteSSEOp(0x00, sseRCP, regOp, arg); }
void XEmitter::RSQRTPS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x00, sseRSQRT, regOp, arg);}
void XEmitter::SHUFPS(X64Reg regOp, const OpArg& arg, u8 shuffle) {WriteSSEOp(0x00, sseSHUF, regOp, arg,1); Write8(shuffle);}
void XEmitter::SHUFPD(X64Reg regOp, const OpArg& arg, u8 shuffle) {WriteSSEOp(0x66, sseSHUF, regOp, arg,1); Write8(shuffle);}
void XEmitter::HADDPS(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF2, sseHADD, regOp, arg);}
void XEmitter::HADDPS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF2, sseHADD, regOp, arg);}
void XEmitter::COMISS(X64Reg regOp, OpArg arg) {WriteSSEOp(0x00, sseCOMIS, regOp, arg);} //weird that these should be packed
void XEmitter::COMISD(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, sseCOMIS, regOp, arg);} //ordered
void XEmitter::UCOMISS(X64Reg regOp, OpArg arg) {WriteSSEOp(0x00, sseUCOMIS, regOp, arg);} //unordered
void XEmitter::UCOMISD(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, sseUCOMIS, regOp, arg);}
void XEmitter::COMISS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x00, sseCOMIS, regOp, arg);} //weird that these should be packed
void XEmitter::COMISD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, sseCOMIS, regOp, arg);} //ordered
void XEmitter::UCOMISS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x00, sseUCOMIS, regOp, arg);} //unordered
void XEmitter::UCOMISD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, sseUCOMIS, regOp, arg);}
void XEmitter::MOVAPS(X64Reg regOp, OpArg arg) {WriteSSEOp(0x00, sseMOVAPfromRM, regOp, arg);}
void XEmitter::MOVAPD(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, sseMOVAPfromRM, regOp, arg);}
void XEmitter::MOVAPS(OpArg arg, X64Reg regOp) {WriteSSEOp(0x00, sseMOVAPtoRM, regOp, arg);}
void XEmitter::MOVAPD(OpArg arg, X64Reg regOp) {WriteSSEOp(0x66, sseMOVAPtoRM, regOp, arg);}
void XEmitter::MOVAPS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x00, sseMOVAPfromRM, regOp, arg);}
void XEmitter::MOVAPD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, sseMOVAPfromRM, regOp, arg);}
void XEmitter::MOVAPS(const OpArg& arg, X64Reg regOp) {WriteSSEOp(0x00, sseMOVAPtoRM, regOp, arg);}
void XEmitter::MOVAPD(const OpArg& arg, X64Reg regOp) {WriteSSEOp(0x66, sseMOVAPtoRM, regOp, arg);}
void XEmitter::MOVUPS(X64Reg regOp, OpArg arg) {WriteSSEOp(0x00, sseMOVUPfromRM, regOp, arg);}
void XEmitter::MOVUPD(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, sseMOVUPfromRM, regOp, arg);}
void XEmitter::MOVUPS(OpArg arg, X64Reg regOp) {WriteSSEOp(0x00, sseMOVUPtoRM, regOp, arg);}
void XEmitter::MOVUPD(OpArg arg, X64Reg regOp) {WriteSSEOp(0x66, sseMOVUPtoRM, regOp, arg);}
void XEmitter::MOVUPS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x00, sseMOVUPfromRM, regOp, arg);}
void XEmitter::MOVUPD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, sseMOVUPfromRM, regOp, arg);}
void XEmitter::MOVUPS(const OpArg& arg, X64Reg regOp) {WriteSSEOp(0x00, sseMOVUPtoRM, regOp, arg);}
void XEmitter::MOVUPD(const OpArg& arg, X64Reg regOp) {WriteSSEOp(0x66, sseMOVUPtoRM, regOp, arg);}
void XEmitter::MOVDQA(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, sseMOVDQfromRM, regOp, arg);}
void XEmitter::MOVDQA(OpArg arg, X64Reg regOp) {WriteSSEOp(0x66, sseMOVDQtoRM, regOp, arg);}
void XEmitter::MOVDQU(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF3, sseMOVDQfromRM, regOp, arg);}
void XEmitter::MOVDQU(OpArg arg, X64Reg regOp) {WriteSSEOp(0xF3, sseMOVDQtoRM, regOp, arg);}
void XEmitter::MOVDQA(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, sseMOVDQfromRM, regOp, arg);}
void XEmitter::MOVDQA(const OpArg& arg, X64Reg regOp) {WriteSSEOp(0x66, sseMOVDQtoRM, regOp, arg);}
void XEmitter::MOVDQU(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF3, sseMOVDQfromRM, regOp, arg);}
void XEmitter::MOVDQU(const OpArg& arg, X64Reg regOp) {WriteSSEOp(0xF3, sseMOVDQtoRM, regOp, arg);}
void XEmitter::MOVSS(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF3, sseMOVUPfromRM, regOp, arg);}
void XEmitter::MOVSD(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF2, sseMOVUPfromRM, regOp, arg);}
void XEmitter::MOVSS(OpArg arg, X64Reg regOp) {WriteSSEOp(0xF3, sseMOVUPtoRM, regOp, arg);}
void XEmitter::MOVSD(OpArg arg, X64Reg regOp) {WriteSSEOp(0xF2, sseMOVUPtoRM, regOp, arg);}
void XEmitter::MOVSS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF3, sseMOVUPfromRM, regOp, arg);}
void XEmitter::MOVSD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF2, sseMOVUPfromRM, regOp, arg);}
void XEmitter::MOVSS(const OpArg& arg, X64Reg regOp) {WriteSSEOp(0xF3, sseMOVUPtoRM, regOp, arg);}
void XEmitter::MOVSD(const OpArg& arg, X64Reg regOp) {WriteSSEOp(0xF2, sseMOVUPtoRM, regOp, arg);}
void XEmitter::MOVLPS(X64Reg regOp, OpArg arg) { WriteSSEOp(0x00, sseMOVLPfromRM, regOp, arg); }
void XEmitter::MOVLPD(X64Reg regOp, OpArg arg) { WriteSSEOp(0x66, sseMOVLPfromRM, regOp, arg); }
void XEmitter::MOVLPS(OpArg arg, X64Reg regOp) { WriteSSEOp(0x00, sseMOVLPtoRM, regOp, arg); }
void XEmitter::MOVLPD(OpArg arg, X64Reg regOp) { WriteSSEOp(0x66, sseMOVLPtoRM, regOp, arg); }
void XEmitter::MOVLPS(X64Reg regOp, const OpArg& arg) { WriteSSEOp(0x00, sseMOVLPfromRM, regOp, arg); }
void XEmitter::MOVLPD(X64Reg regOp, const OpArg& arg) { WriteSSEOp(0x66, sseMOVLPfromRM, regOp, arg); }
void XEmitter::MOVLPS(const OpArg& arg, X64Reg regOp) { WriteSSEOp(0x00, sseMOVLPtoRM, regOp, arg); }
void XEmitter::MOVLPD(const OpArg& arg, X64Reg regOp) { WriteSSEOp(0x66, sseMOVLPtoRM, regOp, arg); }
void XEmitter::MOVHPS(X64Reg regOp, OpArg arg) { WriteSSEOp(0x00, sseMOVHPfromRM, regOp, arg); }
void XEmitter::MOVHPD(X64Reg regOp, OpArg arg) { WriteSSEOp(0x66, sseMOVHPfromRM, regOp, arg); }
void XEmitter::MOVHPS(OpArg arg, X64Reg regOp) { WriteSSEOp(0x00, sseMOVHPtoRM, regOp, arg); }
void XEmitter::MOVHPD(OpArg arg, X64Reg regOp) { WriteSSEOp(0x66, sseMOVHPtoRM, regOp, arg); }
void XEmitter::MOVHPS(X64Reg regOp, const OpArg& arg) { WriteSSEOp(0x00, sseMOVHPfromRM, regOp, arg); }
void XEmitter::MOVHPD(X64Reg regOp, const OpArg& arg) { WriteSSEOp(0x66, sseMOVHPfromRM, regOp, arg); }
void XEmitter::MOVHPS(const OpArg& arg, X64Reg regOp) { WriteSSEOp(0x00, sseMOVHPtoRM, regOp, arg); }
void XEmitter::MOVHPD(const OpArg& arg, X64Reg regOp) { WriteSSEOp(0x66, sseMOVHPtoRM, regOp, arg); }
void XEmitter::MOVHLPS(X64Reg regOp1, X64Reg regOp2) {WriteSSEOp(0x00, sseMOVHLPS, regOp1, R(regOp2));}
void XEmitter::MOVLHPS(X64Reg regOp1, X64Reg regOp2) {WriteSSEOp(0x00, sseMOVLHPS, regOp1, R(regOp2));}
void XEmitter::CVTPS2PD(X64Reg regOp, OpArg arg) {WriteSSEOp(0x00, 0x5A, regOp, arg);}
void XEmitter::CVTPD2PS(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, 0x5A, regOp, arg);}
void XEmitter::CVTPS2PD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x00, 0x5A, regOp, arg);}
void XEmitter::CVTPD2PS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, 0x5A, regOp, arg);}
void XEmitter::CVTSD2SS(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF2, 0x5A, regOp, arg);}
void XEmitter::CVTSS2SD(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF3, 0x5A, regOp, arg);}
void XEmitter::CVTSD2SI(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF2, 0x2D, regOp, arg);}
void XEmitter::CVTSS2SI(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF3, 0x2D, regOp, arg);}
void XEmitter::CVTSI2SD(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF2, 0x2A, regOp, arg);}
void XEmitter::CVTSI2SS(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF3, 0x2A, regOp, arg);}
void XEmitter::CVTSD2SS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF2, 0x5A, regOp, arg);}
void XEmitter::CVTSS2SD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF3, 0x5A, regOp, arg);}
void XEmitter::CVTSD2SI(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF2, 0x2D, regOp, arg);}
void XEmitter::CVTSS2SI(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF3, 0x2D, regOp, arg);}
void XEmitter::CVTSI2SD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF2, 0x2A, regOp, arg);}
void XEmitter::CVTSI2SS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF3, 0x2A, regOp, arg);}
void XEmitter::CVTDQ2PD(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF3, 0xE6, regOp, arg);}
void XEmitter::CVTDQ2PS(X64Reg regOp, OpArg arg) {WriteSSEOp(0x00, 0x5B, regOp, arg);}
void XEmitter::CVTPD2DQ(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF2, 0xE6, regOp, arg);}
void XEmitter::CVTPS2DQ(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, 0x5B, regOp, arg);}
void XEmitter::CVTDQ2PD(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF3, 0xE6, regOp, arg);}
void XEmitter::CVTDQ2PS(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x00, 0x5B, regOp, arg);}
void XEmitter::CVTPD2DQ(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF2, 0xE6, regOp, arg);}
void XEmitter::CVTPS2DQ(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, 0x5B, regOp, arg);}
void XEmitter::CVTTSD2SI(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF2, 0x2C, regOp, arg);}
void XEmitter::CVTTSS2SI(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF3, 0x2C, regOp, arg);}
void XEmitter::CVTTPS2DQ(X64Reg regOp, OpArg arg) {WriteSSEOp(0xF3, 0x5B, regOp, arg);}
void XEmitter::CVTTPD2DQ(X64Reg regOp, OpArg arg) {WriteSSEOp(0x66, 0xE6, regOp, arg);}
void XEmitter::CVTTSD2SI(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF2, 0x2C, regOp, arg);}
void XEmitter::CVTTSS2SI(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF3, 0x2C, regOp, arg);}
void XEmitter::CVTTPS2DQ(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0xF3, 0x5B, regOp, arg);}
void XEmitter::CVTTPD2DQ(X64Reg regOp, const OpArg& arg) {WriteSSEOp(0x66, 0xE6, regOp, arg);}
void XEmitter::MASKMOVDQU(X64Reg dest, X64Reg src) {WriteSSEOp(0x66, sseMASKMOVDQU, dest, R(src));}
void XEmitter::MOVMSKPS(X64Reg dest, OpArg arg) {WriteSSEOp(0x00, 0x50, dest, arg);}
void XEmitter::MOVMSKPD(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0x50, dest, arg);}
void XEmitter::MOVMSKPS(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x00, 0x50, dest, arg);}
void XEmitter::MOVMSKPD(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0x50, dest, arg);}
void XEmitter::LDDQU(X64Reg dest, OpArg arg) {WriteSSEOp(0xF2, sseLDDQU, dest, arg);} // For integer data only
void XEmitter::LDDQU(X64Reg dest, const OpArg& arg) {WriteSSEOp(0xF2, sseLDDQU, dest, arg);} // For integer data only
// THESE TWO ARE UNTESTED.
void XEmitter::UNPCKLPS(X64Reg dest, OpArg arg) {WriteSSEOp(0x00, 0x14, dest, arg);}
void XEmitter::UNPCKHPS(X64Reg dest, OpArg arg) {WriteSSEOp(0x00, 0x15, dest, arg);}
void XEmitter::UNPCKLPS(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x00, 0x14, dest, arg);}
void XEmitter::UNPCKHPS(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x00, 0x15, dest, arg);}
void XEmitter::UNPCKLPD(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0x14, dest, arg);}
void XEmitter::UNPCKHPD(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0x15, dest, arg);}
void XEmitter::UNPCKLPD(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0x14, dest, arg);}
void XEmitter::UNPCKHPD(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0x15, dest, arg);}
void XEmitter::MOVDDUP(X64Reg regOp, OpArg arg)
void XEmitter::MOVDDUP(X64Reg regOp, const OpArg& arg)
{
if (Common::GetCPUCaps().sse3)
{
@ -1663,9 +1663,9 @@ void XEmitter::MOVDDUP(X64Reg regOp, OpArg arg)
//There are a few more left
// Also some integer instructions are missing
void XEmitter::PACKSSDW(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0x6B, dest, arg);}
void XEmitter::PACKSSWB(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0x63, dest, arg);}
void XEmitter::PACKUSWB(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0x67, dest, arg);}
void XEmitter::PACKSSDW(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0x6B, dest, arg);}
void XEmitter::PACKSSWB(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0x63, dest, arg);}
void XEmitter::PACKUSWB(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0x67, dest, arg);}
void XEmitter::PUNPCKLBW(X64Reg dest, const OpArg &arg) {WriteSSEOp(0x66, 0x60, dest, arg);}
void XEmitter::PUNPCKLWD(X64Reg dest, const OpArg &arg) {WriteSSEOp(0x66, 0x61, dest, arg);}
@ -1690,7 +1690,7 @@ void XEmitter::PSRLQ(X64Reg reg, int shift)
Write8(shift);
}
void XEmitter::PSRLQ(X64Reg reg, OpArg arg)
void XEmitter::PSRLQ(X64Reg reg, const OpArg& arg)
{
WriteSSEOp(0x66, 0xd3, reg, arg);
}
@ -1735,212 +1735,212 @@ void XEmitter::PSRAD(X64Reg reg, int shift)
Write8(shift);
}
void XEmitter::WriteSSSE3Op(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extrabytes)
void XEmitter::WriteSSSE3Op(u8 opPrefix, u16 op, X64Reg regOp, const OpArg& arg, int extrabytes)
{
if (!Common::GetCPUCaps().ssse3)
ASSERT_MSG(0, "Trying to use SSSE3 on a system that doesn't support it. Bad programmer.");
WriteSSEOp(opPrefix, op, regOp, arg, extrabytes);
}
void XEmitter::WriteSSE41Op(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extrabytes)
void XEmitter::WriteSSE41Op(u8 opPrefix, u16 op, X64Reg regOp, const OpArg& arg, int extrabytes)
{
if (!Common::GetCPUCaps().sse4_1)
ASSERT_MSG(0, "Trying to use SSE4.1 on a system that doesn't support it. Bad programmer.");
WriteSSEOp(opPrefix, op, regOp, arg, extrabytes);
}
void XEmitter::PSHUFB(X64Reg dest, OpArg arg) {WriteSSSE3Op(0x66, 0x3800, dest, arg);}
void XEmitter::PTEST(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x3817, dest, arg);}
void XEmitter::PACKUSDW(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x382b, dest, arg);}
void XEmitter::DPPS(X64Reg dest, OpArg arg, u8 mask) {WriteSSE41Op(0x66, 0x3A40, dest, arg, 1); Write8(mask);}
void XEmitter::PSHUFB(X64Reg dest, const OpArg& arg) {WriteSSSE3Op(0x66, 0x3800, dest, arg);}
void XEmitter::PTEST(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x3817, dest, arg);}
void XEmitter::PACKUSDW(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x382b, dest, arg);}
void XEmitter::DPPS(X64Reg dest, const OpArg& arg, u8 mask) {WriteSSE41Op(0x66, 0x3A40, dest, arg, 1); Write8(mask);}
void XEmitter::PMINSB(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x3838, dest, arg);}
void XEmitter::PMINSD(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x3839, dest, arg);}
void XEmitter::PMINUW(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x383a, dest, arg);}
void XEmitter::PMINUD(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x383b, dest, arg);}
void XEmitter::PMAXSB(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x383c, dest, arg);}
void XEmitter::PMAXSD(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x383d, dest, arg);}
void XEmitter::PMAXUW(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x383e, dest, arg);}
void XEmitter::PMAXUD(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x383f, dest, arg);}
void XEmitter::PMINSB(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x3838, dest, arg);}
void XEmitter::PMINSD(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x3839, dest, arg);}
void XEmitter::PMINUW(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x383a, dest, arg);}
void XEmitter::PMINUD(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x383b, dest, arg);}
void XEmitter::PMAXSB(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x383c, dest, arg);}
void XEmitter::PMAXSD(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x383d, dest, arg);}
void XEmitter::PMAXUW(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x383e, dest, arg);}
void XEmitter::PMAXUD(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x383f, dest, arg);}
void XEmitter::PMOVSXBW(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x3820, dest, arg);}
void XEmitter::PMOVSXBD(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x3821, dest, arg);}
void XEmitter::PMOVSXBQ(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x3822, dest, arg);}
void XEmitter::PMOVSXWD(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x3823, dest, arg);}
void XEmitter::PMOVSXWQ(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x3824, dest, arg);}
void XEmitter::PMOVSXDQ(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x3825, dest, arg);}
void XEmitter::PMOVZXBW(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x3830, dest, arg);}
void XEmitter::PMOVZXBD(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x3831, dest, arg);}
void XEmitter::PMOVZXBQ(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x3832, dest, arg);}
void XEmitter::PMOVZXWD(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x3833, dest, arg);}
void XEmitter::PMOVZXWQ(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x3834, dest, arg);}
void XEmitter::PMOVZXDQ(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x3835, dest, arg);}
void XEmitter::PMOVSXBW(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x3820, dest, arg);}
void XEmitter::PMOVSXBD(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x3821, dest, arg);}
void XEmitter::PMOVSXBQ(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x3822, dest, arg);}
void XEmitter::PMOVSXWD(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x3823, dest, arg);}
void XEmitter::PMOVSXWQ(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x3824, dest, arg);}
void XEmitter::PMOVSXDQ(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x3825, dest, arg);}
void XEmitter::PMOVZXBW(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x3830, dest, arg);}
void XEmitter::PMOVZXBD(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x3831, dest, arg);}
void XEmitter::PMOVZXBQ(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x3832, dest, arg);}
void XEmitter::PMOVZXWD(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x3833, dest, arg);}
void XEmitter::PMOVZXWQ(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x3834, dest, arg);}
void XEmitter::PMOVZXDQ(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x3835, dest, arg);}
void XEmitter::PBLENDVB(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x3810, dest, arg);}
void XEmitter::BLENDVPS(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x3814, dest, arg);}
void XEmitter::BLENDVPD(X64Reg dest, OpArg arg) {WriteSSE41Op(0x66, 0x3815, dest, arg);}
void XEmitter::PBLENDVB(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x3810, dest, arg);}
void XEmitter::BLENDVPS(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x3814, dest, arg);}
void XEmitter::BLENDVPD(X64Reg dest, const OpArg& arg) {WriteSSE41Op(0x66, 0x3815, dest, arg);}
void XEmitter::BLENDPS(X64Reg dest, const OpArg& arg, u8 blend) { WriteSSE41Op(0x66, 0x3A0C, dest, arg, 1); Write8(blend); }
void XEmitter::BLENDPD(X64Reg dest, const OpArg& arg, u8 blend) { WriteSSE41Op(0x66, 0x3A0D, dest, arg, 1); Write8(blend); }
void XEmitter::ROUNDSS(X64Reg dest, OpArg arg, u8 mode) {WriteSSE41Op(0x66, 0x3A0A, dest, arg, 1); Write8(mode);}
void XEmitter::ROUNDSD(X64Reg dest, OpArg arg, u8 mode) {WriteSSE41Op(0x66, 0x3A0B, dest, arg, 1); Write8(mode);}
void XEmitter::ROUNDPS(X64Reg dest, OpArg arg, u8 mode) {WriteSSE41Op(0x66, 0x3A08, dest, arg, 1); Write8(mode);}
void XEmitter::ROUNDPD(X64Reg dest, OpArg arg, u8 mode) {WriteSSE41Op(0x66, 0x3A09, dest, arg, 1); Write8(mode);}
void XEmitter::ROUNDSS(X64Reg dest, const OpArg& arg, u8 mode) {WriteSSE41Op(0x66, 0x3A0A, dest, arg, 1); Write8(mode);}
void XEmitter::ROUNDSD(X64Reg dest, const OpArg& arg, u8 mode) {WriteSSE41Op(0x66, 0x3A0B, dest, arg, 1); Write8(mode);}
void XEmitter::ROUNDPS(X64Reg dest, const OpArg& arg, u8 mode) {WriteSSE41Op(0x66, 0x3A08, dest, arg, 1); Write8(mode);}
void XEmitter::ROUNDPD(X64Reg dest, const OpArg& arg, u8 mode) {WriteSSE41Op(0x66, 0x3A09, dest, arg, 1); Write8(mode);}
void XEmitter::PAND(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xDB, dest, arg);}
void XEmitter::PANDN(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xDF, dest, arg);}
void XEmitter::PXOR(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xEF, dest, arg);}
void XEmitter::POR(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xEB, dest, arg);}
void XEmitter::PAND(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xDB, dest, arg);}
void XEmitter::PANDN(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xDF, dest, arg);}
void XEmitter::PXOR(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xEF, dest, arg);}
void XEmitter::POR(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xEB, dest, arg);}
void XEmitter::PADDB(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xFC, dest, arg);}
void XEmitter::PADDW(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xFD, dest, arg);}
void XEmitter::PADDD(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xFE, dest, arg);}
void XEmitter::PADDQ(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xD4, dest, arg);}
void XEmitter::PADDB(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xFC, dest, arg);}
void XEmitter::PADDW(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xFD, dest, arg);}
void XEmitter::PADDD(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xFE, dest, arg);}
void XEmitter::PADDQ(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xD4, dest, arg);}
void XEmitter::PADDSB(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xEC, dest, arg);}
void XEmitter::PADDSW(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xED, dest, arg);}
void XEmitter::PADDUSB(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xDC, dest, arg);}
void XEmitter::PADDUSW(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xDD, dest, arg);}
void XEmitter::PADDSB(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xEC, dest, arg);}
void XEmitter::PADDSW(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xED, dest, arg);}
void XEmitter::PADDUSB(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xDC, dest, arg);}
void XEmitter::PADDUSW(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xDD, dest, arg);}
void XEmitter::PSUBB(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xF8, dest, arg);}
void XEmitter::PSUBW(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xF9, dest, arg);}
void XEmitter::PSUBD(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xFA, dest, arg);}
void XEmitter::PSUBQ(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xFB, dest, arg);}
void XEmitter::PSUBB(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xF8, dest, arg);}
void XEmitter::PSUBW(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xF9, dest, arg);}
void XEmitter::PSUBD(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xFA, dest, arg);}
void XEmitter::PSUBQ(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xFB, dest, arg);}
void XEmitter::PSUBSB(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xE8, dest, arg);}
void XEmitter::PSUBSW(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xE9, dest, arg);}
void XEmitter::PSUBUSB(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xD8, dest, arg);}
void XEmitter::PSUBUSW(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xD9, dest, arg);}
void XEmitter::PSUBSB(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xE8, dest, arg);}
void XEmitter::PSUBSW(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xE9, dest, arg);}
void XEmitter::PSUBUSB(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xD8, dest, arg);}
void XEmitter::PSUBUSW(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xD9, dest, arg);}
void XEmitter::PAVGB(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xE0, dest, arg);}
void XEmitter::PAVGW(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xE3, dest, arg);}
void XEmitter::PAVGB(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xE0, dest, arg);}
void XEmitter::PAVGW(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xE3, dest, arg);}
void XEmitter::PCMPEQB(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0x74, dest, arg);}
void XEmitter::PCMPEQW(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0x75, dest, arg);}
void XEmitter::PCMPEQD(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0x76, dest, arg);}
void XEmitter::PCMPEQB(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0x74, dest, arg);}
void XEmitter::PCMPEQW(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0x75, dest, arg);}
void XEmitter::PCMPEQD(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0x76, dest, arg);}
void XEmitter::PCMPGTB(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0x64, dest, arg);}
void XEmitter::PCMPGTW(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0x65, dest, arg);}
void XEmitter::PCMPGTD(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0x66, dest, arg);}
void XEmitter::PCMPGTB(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0x64, dest, arg);}
void XEmitter::PCMPGTW(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0x65, dest, arg);}
void XEmitter::PCMPGTD(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0x66, dest, arg);}
void XEmitter::PEXTRW(X64Reg dest, OpArg arg, u8 subreg) {WriteSSEOp(0x66, 0xC5, dest, arg, 1); Write8(subreg);}
void XEmitter::PINSRW(X64Reg dest, OpArg arg, u8 subreg) {WriteSSEOp(0x66, 0xC4, dest, arg, 1); Write8(subreg);}
void XEmitter::PEXTRW(X64Reg dest, const OpArg& arg, u8 subreg) {WriteSSEOp(0x66, 0xC5, dest, arg, 1); Write8(subreg);}
void XEmitter::PINSRW(X64Reg dest, const OpArg& arg, u8 subreg) {WriteSSEOp(0x66, 0xC4, dest, arg, 1); Write8(subreg);}
void XEmitter::PMADDWD(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xF5, dest, arg); }
void XEmitter::PSADBW(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xF6, dest, arg);}
void XEmitter::PMADDWD(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xF5, dest, arg); }
void XEmitter::PSADBW(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xF6, dest, arg);}
void XEmitter::PMAXSW(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xEE, dest, arg); }
void XEmitter::PMAXUB(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xDE, dest, arg); }
void XEmitter::PMINSW(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xEA, dest, arg); }
void XEmitter::PMINUB(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xDA, dest, arg); }
void XEmitter::PMAXSW(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xEE, dest, arg); }
void XEmitter::PMAXUB(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xDE, dest, arg); }
void XEmitter::PMINSW(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xEA, dest, arg); }
void XEmitter::PMINUB(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xDA, dest, arg); }
void XEmitter::PMOVMSKB(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0xD7, dest, arg); }
void XEmitter::PSHUFD(X64Reg regOp, OpArg arg, u8 shuffle) {WriteSSEOp(0x66, 0x70, regOp, arg, 1); Write8(shuffle);}
void XEmitter::PSHUFLW(X64Reg regOp, OpArg arg, u8 shuffle) {WriteSSEOp(0xF2, 0x70, regOp, arg, 1); Write8(shuffle);}
void XEmitter::PSHUFHW(X64Reg regOp, OpArg arg, u8 shuffle) {WriteSSEOp(0xF3, 0x70, regOp, arg, 1); Write8(shuffle);}
void XEmitter::PMOVMSKB(X64Reg dest, const OpArg& arg) {WriteSSEOp(0x66, 0xD7, dest, arg); }
void XEmitter::PSHUFD(X64Reg regOp, const OpArg& arg, u8 shuffle) {WriteSSEOp(0x66, 0x70, regOp, arg, 1); Write8(shuffle);}
void XEmitter::PSHUFLW(X64Reg regOp, const OpArg& arg, u8 shuffle) {WriteSSEOp(0xF2, 0x70, regOp, arg, 1); Write8(shuffle);}
void XEmitter::PSHUFHW(X64Reg regOp, const OpArg& arg, u8 shuffle) {WriteSSEOp(0xF3, 0x70, regOp, arg, 1); Write8(shuffle);}
// VEX
void XEmitter::VADDSD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0xF2, sseADD, regOp1, regOp2, arg);}
void XEmitter::VSUBSD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0xF2, sseSUB, regOp1, regOp2, arg);}
void XEmitter::VMULSD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0xF2, sseMUL, regOp1, regOp2, arg);}
void XEmitter::VDIVSD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0xF2, sseDIV, regOp1, regOp2, arg);}
void XEmitter::VADDPD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, sseADD, regOp1, regOp2, arg);}
void XEmitter::VSUBPD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, sseSUB, regOp1, regOp2, arg);}
void XEmitter::VMULPD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, sseMUL, regOp1, regOp2, arg);}
void XEmitter::VDIVPD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, sseDIV, regOp1, regOp2, arg);}
void XEmitter::VSQRTSD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0xF2, sseSQRT, regOp1, regOp2, arg);}
void XEmitter::VSHUFPD(X64Reg regOp1, X64Reg regOp2, OpArg arg, u8 shuffle) {WriteAVXOp(0x66, sseSHUF, regOp1, regOp2, arg, 1); Write8(shuffle);}
void XEmitter::VUNPCKLPD(X64Reg regOp1, X64Reg regOp2, OpArg arg){WriteAVXOp(0x66, 0x14, regOp1, regOp2, arg);}
void XEmitter::VUNPCKHPD(X64Reg regOp1, X64Reg regOp2, OpArg arg){WriteAVXOp(0x66, 0x15, regOp1, regOp2, arg);}
void XEmitter::VADDSD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) {WriteAVXOp(0xF2, sseADD, regOp1, regOp2, arg);}
void XEmitter::VSUBSD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) {WriteAVXOp(0xF2, sseSUB, regOp1, regOp2, arg);}
void XEmitter::VMULSD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) {WriteAVXOp(0xF2, sseMUL, regOp1, regOp2, arg);}
void XEmitter::VDIVSD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) {WriteAVXOp(0xF2, sseDIV, regOp1, regOp2, arg);}
void XEmitter::VADDPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) {WriteAVXOp(0x66, sseADD, regOp1, regOp2, arg);}
void XEmitter::VSUBPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) {WriteAVXOp(0x66, sseSUB, regOp1, regOp2, arg);}
void XEmitter::VMULPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) {WriteAVXOp(0x66, sseMUL, regOp1, regOp2, arg);}
void XEmitter::VDIVPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) {WriteAVXOp(0x66, sseDIV, regOp1, regOp2, arg);}
void XEmitter::VSQRTSD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) {WriteAVXOp(0xF2, sseSQRT, regOp1, regOp2, arg);}
void XEmitter::VSHUFPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg, u8 shuffle) {WriteAVXOp(0x66, sseSHUF, regOp1, regOp2, arg, 1); Write8(shuffle);}
void XEmitter::VUNPCKLPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg){WriteAVXOp(0x66, 0x14, regOp1, regOp2, arg);}
void XEmitter::VUNPCKHPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg){WriteAVXOp(0x66, 0x15, regOp1, regOp2, arg);}
void XEmitter::VANDPS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x00, sseAND, regOp1, regOp2, arg); }
void XEmitter::VANDPD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, sseAND, regOp1, regOp2, arg); }
void XEmitter::VANDNPS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x00, sseANDN, regOp1, regOp2, arg); }
void XEmitter::VANDNPD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, sseANDN, regOp1, regOp2, arg); }
void XEmitter::VORPS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x00, sseOR, regOp1, regOp2, arg); }
void XEmitter::VORPD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, sseOR, regOp1, regOp2, arg); }
void XEmitter::VXORPS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x00, sseXOR, regOp1, regOp2, arg); }
void XEmitter::VXORPD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, sseXOR, regOp1, regOp2, arg); }
void XEmitter::VANDPS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x00, sseAND, regOp1, regOp2, arg); }
void XEmitter::VANDPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, sseAND, regOp1, regOp2, arg); }
void XEmitter::VANDNPS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x00, sseANDN, regOp1, regOp2, arg); }
void XEmitter::VANDNPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, sseANDN, regOp1, regOp2, arg); }
void XEmitter::VORPS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x00, sseOR, regOp1, regOp2, arg); }
void XEmitter::VORPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, sseOR, regOp1, regOp2, arg); }
void XEmitter::VXORPS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x00, sseXOR, regOp1, regOp2, arg); }
void XEmitter::VXORPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, sseXOR, regOp1, regOp2, arg); }
void XEmitter::VPAND(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0xDB, regOp1, regOp2, arg); }
void XEmitter::VPANDN(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0xDF, regOp1, regOp2, arg); }
void XEmitter::VPOR(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0xEB, regOp1, regOp2, arg); }
void XEmitter::VPXOR(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0xEF, regOp1, regOp2, arg); }
void XEmitter::VPAND(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0xDB, regOp1, regOp2, arg); }
void XEmitter::VPANDN(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0xDF, regOp1, regOp2, arg); }
void XEmitter::VPOR(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0xEB, regOp1, regOp2, arg); }
void XEmitter::VPXOR(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0xEF, regOp1, regOp2, arg); }
void XEmitter::VFMADD132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x3898, regOp1, regOp2, arg); }
void XEmitter::VFMADD213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38A8, regOp1, regOp2, arg); }
void XEmitter::VFMADD231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38B8, regOp1, regOp2, arg); }
void XEmitter::VFMADD132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x3898, regOp1, regOp2, arg, 1); }
void XEmitter::VFMADD213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38A8, regOp1, regOp2, arg, 1); }
void XEmitter::VFMADD231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38B8, regOp1, regOp2, arg, 1); }
void XEmitter::VFMADD132SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x3899, regOp1, regOp2, arg); }
void XEmitter::VFMADD213SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38A9, regOp1, regOp2, arg); }
void XEmitter::VFMADD231SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38B9, regOp1, regOp2, arg); }
void XEmitter::VFMADD132SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x3899, regOp1, regOp2, arg, 1); }
void XEmitter::VFMADD213SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38A9, regOp1, regOp2, arg, 1); }
void XEmitter::VFMADD231SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38B9, regOp1, regOp2, arg, 1); }
void XEmitter::VFMSUB132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x389A, regOp1, regOp2, arg); }
void XEmitter::VFMSUB213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38AA, regOp1, regOp2, arg); }
void XEmitter::VFMSUB231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38BA, regOp1, regOp2, arg); }
void XEmitter::VFMSUB132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x389A, regOp1, regOp2, arg, 1); }
void XEmitter::VFMSUB213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38AA, regOp1, regOp2, arg, 1); }
void XEmitter::VFMSUB231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38BA, regOp1, regOp2, arg, 1); }
void XEmitter::VFMSUB132SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x389B, regOp1, regOp2, arg); }
void XEmitter::VFMSUB213SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38AB, regOp1, regOp2, arg); }
void XEmitter::VFMSUB231SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38BB, regOp1, regOp2, arg); }
void XEmitter::VFMSUB132SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x389B, regOp1, regOp2, arg, 1); }
void XEmitter::VFMSUB213SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38AB, regOp1, regOp2, arg, 1); }
void XEmitter::VFMSUB231SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38BB, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMADD132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x389C, regOp1, regOp2, arg); }
void XEmitter::VFNMADD213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38AC, regOp1, regOp2, arg); }
void XEmitter::VFNMADD231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38BC, regOp1, regOp2, arg); }
void XEmitter::VFNMADD132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x389C, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMADD213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38AC, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMADD231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38BC, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMADD132SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x389D, regOp1, regOp2, arg); }
void XEmitter::VFNMADD213SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38AD, regOp1, regOp2, arg); }
void XEmitter::VFNMADD231SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38BD, regOp1, regOp2, arg); }
void XEmitter::VFNMADD132SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x389D, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMADD213SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38AD, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMADD231SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38BD, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMSUB132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x389E, regOp1, regOp2, arg); }
void XEmitter::VFNMSUB213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38AE, regOp1, regOp2, arg); }
void XEmitter::VFNMSUB231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38BE, regOp1, regOp2, arg); }
void XEmitter::VFNMSUB132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x389E, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMSUB213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38AE, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMSUB231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38BE, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMSUB132SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x389F, regOp1, regOp2, arg); }
void XEmitter::VFNMSUB213SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38AF, regOp1, regOp2, arg); }
void XEmitter::VFNMSUB231SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38BF, regOp1, regOp2, arg); }
void XEmitter::VFNMSUB132SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x389F, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMSUB213SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38AF, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMSUB231SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38BF, regOp1, regOp2, arg, 1); }
void XEmitter::VFMADDSUB132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x3896, regOp1, regOp2, arg); }
void XEmitter::VFMADDSUB213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38A6, regOp1, regOp2, arg); }
void XEmitter::VFMADDSUB231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38B6, regOp1, regOp2, arg); }
void XEmitter::VFMADDSUB132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x3896, regOp1, regOp2, arg, 1); }
void XEmitter::VFMADDSUB213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38A6, regOp1, regOp2, arg, 1); }
void XEmitter::VFMADDSUB231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38B6, regOp1, regOp2, arg, 1); }
void XEmitter::VFMSUBADD132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x3897, regOp1, regOp2, arg); }
void XEmitter::VFMSUBADD213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38A7, regOp1, regOp2, arg); }
void XEmitter::VFMSUBADD231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38B7, regOp1, regOp2, arg); }
void XEmitter::VFMSUBADD132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x3897, regOp1, regOp2, arg, 1); }
void XEmitter::VFMSUBADD213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38A7, regOp1, regOp2, arg, 1); }
void XEmitter::VFMSUBADD231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) { WriteAVXOp(0x66, 0x38B7, regOp1, regOp2, arg, 1); }
void XEmitter::VFMADD132PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x3898, regOp1, regOp2, arg); }
void XEmitter::VFMADD213PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38A8, regOp1, regOp2, arg); }
void XEmitter::VFMADD231PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38B8, regOp1, regOp2, arg); }
void XEmitter::VFMADD132PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x3898, regOp1, regOp2, arg, 1); }
void XEmitter::VFMADD213PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38A8, regOp1, regOp2, arg, 1); }
void XEmitter::VFMADD231PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38B8, regOp1, regOp2, arg, 1); }
void XEmitter::VFMADD132SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x3899, regOp1, regOp2, arg); }
void XEmitter::VFMADD213SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38A9, regOp1, regOp2, arg); }
void XEmitter::VFMADD231SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38B9, regOp1, regOp2, arg); }
void XEmitter::VFMADD132SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x3899, regOp1, regOp2, arg, 1); }
void XEmitter::VFMADD213SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38A9, regOp1, regOp2, arg, 1); }
void XEmitter::VFMADD231SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38B9, regOp1, regOp2, arg, 1); }
void XEmitter::VFMSUB132PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x389A, regOp1, regOp2, arg); }
void XEmitter::VFMSUB213PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38AA, regOp1, regOp2, arg); }
void XEmitter::VFMSUB231PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38BA, regOp1, regOp2, arg); }
void XEmitter::VFMSUB132PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x389A, regOp1, regOp2, arg, 1); }
void XEmitter::VFMSUB213PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38AA, regOp1, regOp2, arg, 1); }
void XEmitter::VFMSUB231PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38BA, regOp1, regOp2, arg, 1); }
void XEmitter::VFMSUB132SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x389B, regOp1, regOp2, arg); }
void XEmitter::VFMSUB213SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38AB, regOp1, regOp2, arg); }
void XEmitter::VFMSUB231SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38BB, regOp1, regOp2, arg); }
void XEmitter::VFMSUB132SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x389B, regOp1, regOp2, arg, 1); }
void XEmitter::VFMSUB213SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38AB, regOp1, regOp2, arg, 1); }
void XEmitter::VFMSUB231SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38BB, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMADD132PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x389C, regOp1, regOp2, arg); }
void XEmitter::VFNMADD213PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38AC, regOp1, regOp2, arg); }
void XEmitter::VFNMADD231PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38BC, regOp1, regOp2, arg); }
void XEmitter::VFNMADD132PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x389C, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMADD213PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38AC, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMADD231PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38BC, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMADD132SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x389D, regOp1, regOp2, arg); }
void XEmitter::VFNMADD213SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38AD, regOp1, regOp2, arg); }
void XEmitter::VFNMADD231SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38BD, regOp1, regOp2, arg); }
void XEmitter::VFNMADD132SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x389D, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMADD213SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38AD, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMADD231SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38BD, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMSUB132PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x389E, regOp1, regOp2, arg); }
void XEmitter::VFNMSUB213PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38AE, regOp1, regOp2, arg); }
void XEmitter::VFNMSUB231PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38BE, regOp1, regOp2, arg); }
void XEmitter::VFNMSUB132PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x389E, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMSUB213PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38AE, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMSUB231PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38BE, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMSUB132SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x389F, regOp1, regOp2, arg); }
void XEmitter::VFNMSUB213SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38AF, regOp1, regOp2, arg); }
void XEmitter::VFNMSUB231SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38BF, regOp1, regOp2, arg); }
void XEmitter::VFNMSUB132SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x389F, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMSUB213SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38AF, regOp1, regOp2, arg, 1); }
void XEmitter::VFNMSUB231SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38BF, regOp1, regOp2, arg, 1); }
void XEmitter::VFMADDSUB132PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x3896, regOp1, regOp2, arg); }
void XEmitter::VFMADDSUB213PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38A6, regOp1, regOp2, arg); }
void XEmitter::VFMADDSUB231PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38B6, regOp1, regOp2, arg); }
void XEmitter::VFMADDSUB132PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x3896, regOp1, regOp2, arg, 1); }
void XEmitter::VFMADDSUB213PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38A6, regOp1, regOp2, arg, 1); }
void XEmitter::VFMADDSUB231PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38B6, regOp1, regOp2, arg, 1); }
void XEmitter::VFMSUBADD132PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x3897, regOp1, regOp2, arg); }
void XEmitter::VFMSUBADD213PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38A7, regOp1, regOp2, arg); }
void XEmitter::VFMSUBADD231PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38B7, regOp1, regOp2, arg); }
void XEmitter::VFMSUBADD132PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x3897, regOp1, regOp2, arg, 1); }
void XEmitter::VFMSUBADD213PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38A7, regOp1, regOp2, arg, 1); }
void XEmitter::VFMSUBADD231PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg) { WriteAVXOp(0x66, 0x38B7, regOp1, regOp2, arg, 1); }
void XEmitter::SARX(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2) {WriteBMI2Op(bits, 0xF3, 0x38F7, regOp1, regOp2, arg);}
void XEmitter::SHLX(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2) {WriteBMI2Op(bits, 0x66, 0x38F7, regOp1, regOp2, arg);}
void XEmitter::SHRX(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2) {WriteBMI2Op(bits, 0xF2, 0x38F7, regOp1, regOp2, arg);}
void XEmitter::RORX(int bits, X64Reg regOp, OpArg arg, u8 rotate) {WriteBMI2Op(bits, 0xF2, 0x3AF0, regOp, INVALID_REG, arg, 1); Write8(rotate);}
void XEmitter::PEXT(int bits, X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteBMI2Op(bits, 0xF3, 0x38F5, regOp1, regOp2, arg);}
void XEmitter::PDEP(int bits, X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteBMI2Op(bits, 0xF2, 0x38F5, regOp1, regOp2, arg);}
void XEmitter::MULX(int bits, X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteBMI2Op(bits, 0xF2, 0x38F6, regOp2, regOp1, arg);}
void XEmitter::BZHI(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2) {WriteBMI2Op(bits, 0x00, 0x38F5, regOp1, regOp2, arg);}
void XEmitter::BLSR(int bits, X64Reg regOp, OpArg arg) {WriteBMI1Op(bits, 0x00, 0x38F3, (X64Reg)0x1, regOp, arg);}
void XEmitter::BLSMSK(int bits, X64Reg regOp, OpArg arg) {WriteBMI1Op(bits, 0x00, 0x38F3, (X64Reg)0x2, regOp, arg);}
void XEmitter::BLSI(int bits, X64Reg regOp, OpArg arg) {WriteBMI1Op(bits, 0x00, 0x38F3, (X64Reg)0x3, regOp, arg);}
void XEmitter::BEXTR(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2){WriteBMI1Op(bits, 0x00, 0x38F7, regOp1, regOp2, arg);}
void XEmitter::ANDN(int bits, X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteBMI1Op(bits, 0x00, 0x38F2, regOp1, regOp2, arg);}
void XEmitter::SARX(int bits, X64Reg regOp1, const OpArg& arg, X64Reg regOp2) {WriteBMI2Op(bits, 0xF3, 0x38F7, regOp1, regOp2, arg);}
void XEmitter::SHLX(int bits, X64Reg regOp1, const OpArg& arg, X64Reg regOp2) {WriteBMI2Op(bits, 0x66, 0x38F7, regOp1, regOp2, arg);}
void XEmitter::SHRX(int bits, X64Reg regOp1, const OpArg& arg, X64Reg regOp2) {WriteBMI2Op(bits, 0xF2, 0x38F7, regOp1, regOp2, arg);}
void XEmitter::RORX(int bits, X64Reg regOp, const OpArg& arg, u8 rotate) {WriteBMI2Op(bits, 0xF2, 0x3AF0, regOp, INVALID_REG, arg, 1); Write8(rotate);}
void XEmitter::PEXT(int bits, X64Reg regOp1, X64Reg regOp2, const OpArg& arg) {WriteBMI2Op(bits, 0xF3, 0x38F5, regOp1, regOp2, arg);}
void XEmitter::PDEP(int bits, X64Reg regOp1, X64Reg regOp2, const OpArg& arg) {WriteBMI2Op(bits, 0xF2, 0x38F5, regOp1, regOp2, arg);}
void XEmitter::MULX(int bits, X64Reg regOp1, X64Reg regOp2, const OpArg& arg) {WriteBMI2Op(bits, 0xF2, 0x38F6, regOp2, regOp1, arg);}
void XEmitter::BZHI(int bits, X64Reg regOp1, const OpArg& arg, X64Reg regOp2) {WriteBMI2Op(bits, 0x00, 0x38F5, regOp1, regOp2, arg);}
void XEmitter::BLSR(int bits, X64Reg regOp, const OpArg& arg) {WriteBMI1Op(bits, 0x00, 0x38F3, (X64Reg)0x1, regOp, arg);}
void XEmitter::BLSMSK(int bits, X64Reg regOp, const OpArg& arg) {WriteBMI1Op(bits, 0x00, 0x38F3, (X64Reg)0x2, regOp, arg);}
void XEmitter::BLSI(int bits, X64Reg regOp, const OpArg& arg) {WriteBMI1Op(bits, 0x00, 0x38F3, (X64Reg)0x3, regOp, arg);}
void XEmitter::BEXTR(int bits, X64Reg regOp1, const OpArg& arg, X64Reg regOp2){WriteBMI1Op(bits, 0x00, 0x38F7, regOp1, regOp2, arg);}
void XEmitter::ANDN(int bits, X64Reg regOp1, X64Reg regOp2, const OpArg& arg) {WriteBMI1Op(bits, 0x00, 0x38F2, regOp1, regOp2, arg);}
// Prefixes
@ -1956,7 +1956,7 @@ void XEmitter::FWAIT()
}
// TODO: make this more generic
void XEmitter::WriteFloatLoadStore(int bits, FloatOp op, FloatOp op_80b, OpArg arg)
void XEmitter::WriteFloatLoadStore(int bits, FloatOp op, FloatOp op_80b, const OpArg& arg)
{
int mf = 0;
ASSERT_MSG(!(bits == 80 && op_80b == floatINVALID), "WriteFloatLoadStore: 80 bits not supported for this instruction");
@ -1974,9 +1974,9 @@ void XEmitter::WriteFloatLoadStore(int bits, FloatOp op, FloatOp op_80b, OpArg a
arg.WriteRest(this, 0, (X64Reg) op);
}
void XEmitter::FLD(int bits, OpArg src) {WriteFloatLoadStore(bits, floatLD, floatLD80, src);}
void XEmitter::FST(int bits, OpArg dest) {WriteFloatLoadStore(bits, floatST, floatINVALID, dest);}
void XEmitter::FSTP(int bits, OpArg dest) {WriteFloatLoadStore(bits, floatSTP, floatSTP80, dest);}
void XEmitter::FLD(int bits, const OpArg& src) {WriteFloatLoadStore(bits, floatLD, floatLD80, src);}
void XEmitter::FST(int bits, const OpArg& dest) {WriteFloatLoadStore(bits, floatST, floatINVALID, dest);}
void XEmitter::FSTP(int bits, const OpArg& dest) {WriteFloatLoadStore(bits, floatSTP, floatSTP80, dest);}
void XEmitter::FNSTSW_AX() { Write8(0xDF); Write8(0xE0); }
void XEmitter::RDTSC() { Write8(0x0F); Write8(0x31); }

776
src/common/x64/emitter.h
View file

@ -344,19 +344,19 @@ private:
void WriteSimple2Byte(int bits, u8 byte1, u8 byte2, X64Reg reg);
void WriteMulDivType(int bits, OpArg src, int ext);
void WriteBitSearchType(int bits, X64Reg dest, OpArg src, u8 byte2, bool rep = false);
void WriteShift(int bits, OpArg dest, OpArg &shift, int ext);
void WriteBitTest(int bits, OpArg &dest, OpArg &index, int ext);
void WriteShift(int bits, OpArg dest, const OpArg& shift, int ext);
void WriteBitTest(int bits, const OpArg& dest, const OpArg& index, int ext);
void WriteMXCSR(OpArg arg, int ext);
void WriteSSEOp(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extrabytes = 0);
void WriteSSSE3Op(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extrabytes = 0);
void WriteSSE41Op(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extrabytes = 0);
void WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extrabytes = 0);
void WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes = 0);
void WriteVEXOp(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes = 0);
void WriteBMI1Op(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes = 0);
void WriteBMI2Op(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes = 0);
void WriteFloatLoadStore(int bits, FloatOp op, FloatOp op_80b, OpArg arg);
void WriteNormalOp(XEmitter *emit, int bits, NormalOp op, const OpArg &a1, const OpArg &a2);
void WriteSSSE3Op(u8 opPrefix, u16 op, X64Reg regOp, const OpArg& arg, int extrabytes = 0);
void WriteSSE41Op(u8 opPrefix, u16 op, X64Reg regOp, const OpArg& arg, int extrabytes = 0);
void WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp, const OpArg& arg, int extrabytes = 0);
void WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, const OpArg& arg, int extrabytes = 0);
void WriteVEXOp(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, const OpArg& arg, int extrabytes = 0);
void WriteBMI1Op(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, const OpArg& arg, int extrabytes = 0);
void WriteBMI2Op(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, const OpArg& arg, int extrabytes = 0);
void WriteFloatLoadStore(int bits, FloatOp op, FloatOp op_80b, const OpArg& arg);
void WriteNormalOp(XEmitter *emit, int bits, NormalOp op, const OpArg& a1, const OpArg& a2);
void ABI_CalculateFrameSize(u32 mask, size_t rsp_alignment, size_t needed_frame_size, size_t* shadowp, size_t* subtractionp, size_t* xmm_offsetp);
@ -413,8 +413,8 @@ public:
// Stack control
void PUSH(X64Reg reg);
void POP(X64Reg reg);
void PUSH(int bits, const OpArg &reg);
void POP(int bits, const OpArg &reg);
void PUSH(int bits, const OpArg& reg);
void POP(int bits, const OpArg& reg);
void PUSHF();
void POPF();
@ -424,21 +424,21 @@ public:
void UD2();
FixupBranch J(bool force5bytes = false);
void JMP(const u8 * addr, bool force5Bytes = false);
void JMP(const u8* addr, bool force5Bytes = false);
void JMP(OpArg arg);
void JMPptr(const OpArg &arg);
void JMPptr(const OpArg& arg);
void JMPself(); //infinite loop!
#ifdef CALL
#undef CALL
#endif
void CALL(const void *fnptr);
void CALL(const void* fnptr);
void CALLptr(OpArg arg);
FixupBranch J_CC(CCFlags conditionCode, bool force5bytes = false);
//void J_CC(CCFlags conditionCode, JumpTarget target);
void J_CC(CCFlags conditionCode, const u8 * addr, bool force5Bytes = false);
void J_CC(CCFlags conditionCode, const u8* addr, bool force5Bytes = false);
void SetJumpTarget(const FixupBranch &branch);
void SetJumpTarget(const FixupBranch& branch);
void SETcc(CCFlags flag, OpArg dest);
// Note: CMOV brings small if any benefit on current cpus.
@ -450,8 +450,8 @@ public:
void SFENCE();
// Bit scan
void BSF(int bits, X64Reg dest, OpArg src); //bottom bit to top bit
void BSR(int bits, X64Reg dest, OpArg src); //top bit to bottom bit
void BSF(int bits, X64Reg dest, const OpArg& src); // Bottom bit to top bit
void BSR(int bits, X64Reg dest, const OpArg& src); // Top bit to bottom bit
// Cache control
enum PrefetchLevel
@ -462,37 +462,37 @@ public:
PF_T2, //Levels 3+ (aliased to T0 on AMD)
};
void PREFETCH(PrefetchLevel level, OpArg arg);
void MOVNTI(int bits, OpArg dest, X64Reg src);
void MOVNTDQ(OpArg arg, X64Reg regOp);
void MOVNTPS(OpArg arg, X64Reg regOp);
void MOVNTPD(OpArg arg, X64Reg regOp);
void MOVNTI(int bits, const OpArg& dest, X64Reg src);
void MOVNTDQ(const OpArg& arg, X64Reg regOp);
void MOVNTPS(const OpArg& arg, X64Reg regOp);
void MOVNTPD(const OpArg& arg, X64Reg regOp);
// Multiplication / division
void MUL(int bits, OpArg src); //UNSIGNED
void IMUL(int bits, OpArg src); //SIGNED
void IMUL(int bits, X64Reg regOp, OpArg src);
void IMUL(int bits, X64Reg regOp, OpArg src, OpArg imm);
void DIV(int bits, OpArg src);
void IDIV(int bits, OpArg src);
void MUL(int bits, const OpArg& src); //UNSIGNED
void IMUL(int bits, const OpArg& src); //SIGNED
void IMUL(int bits, X64Reg regOp, const OpArg& src);
void IMUL(int bits, X64Reg regOp, const OpArg& src, const OpArg& imm);
void DIV(int bits, const OpArg& src);
void IDIV(int bits, const OpArg& src);
// Shift
void ROL(int bits, OpArg dest, OpArg shift);
void ROR(int bits, OpArg dest, OpArg shift);
void RCL(int bits, OpArg dest, OpArg shift);
void RCR(int bits, OpArg dest, OpArg shift);
void SHL(int bits, OpArg dest, OpArg shift);
void SHR(int bits, OpArg dest, OpArg shift);
void SAR(int bits, OpArg dest, OpArg shift);
void ROL(int bits, const OpArg& dest, const OpArg& shift);
void ROR(int bits, const OpArg& dest, const OpArg& shift);
void RCL(int bits, const OpArg& dest, const OpArg& shift);
void RCR(int bits, const OpArg& dest, const OpArg& shift);
void SHL(int bits, const OpArg& dest, const OpArg& shift);
void SHR(int bits, const OpArg& dest, const OpArg& shift);
void SAR(int bits, const OpArg& dest, const OpArg& shift);
// Bit Test
void BT(int bits, OpArg dest, OpArg index);
void BTS(int bits, OpArg dest, OpArg index);
void BTR(int bits, OpArg dest, OpArg index);
void BTC(int bits, OpArg dest, OpArg index);
void BT(int bits, const OpArg& dest, const OpArg& index);
void BTS(int bits, const OpArg& dest, const OpArg& index);
void BTR(int bits, const OpArg& dest, const OpArg& index);
void BTC(int bits, const OpArg& dest, const OpArg& index);
// Double-Precision Shift
void SHRD(int bits, OpArg dest, OpArg src, OpArg shift);
void SHLD(int bits, OpArg dest, OpArg src, OpArg shift);
void SHRD(int bits, const OpArg& dest, const OpArg& src, const OpArg& shift);
void SHLD(int bits, const OpArg& dest, const OpArg& src, const OpArg& shift);
// Extend EAX into EDX in various ways
void CWD(int bits = 16);
@ -506,23 +506,23 @@ public:
void LEA(int bits, X64Reg dest, OpArg src);
// Integer arithmetic
void NEG (int bits, OpArg src);
void ADD (int bits, const OpArg &a1, const OpArg &a2);
void ADC (int bits, const OpArg &a1, const OpArg &a2);
void SUB (int bits, const OpArg &a1, const OpArg &a2);
void SBB (int bits, const OpArg &a1, const OpArg &a2);
void AND (int bits, const OpArg &a1, const OpArg &a2);
void CMP (int bits, const OpArg &a1, const OpArg &a2);
void NEG(int bits, const OpArg& src);
void ADD(int bits, const OpArg& a1, const OpArg& a2);
void ADC(int bits, const OpArg& a1, const OpArg& a2);
void SUB(int bits, const OpArg& a1, const OpArg& a2);
void SBB(int bits, const OpArg& a1, const OpArg& a2);
void AND(int bits, const OpArg& a1, const OpArg& a2);
void CMP(int bits, const OpArg& a1, const OpArg& a2);
// Bit operations
void NOT (int bits, OpArg src);
void OR (int bits, const OpArg &a1, const OpArg &a2);
void XOR (int bits, const OpArg &a1, const OpArg &a2);
void MOV (int bits, const OpArg &a1, const OpArg &a2);
void TEST(int bits, const OpArg &a1, const OpArg &a2);
void NOT (int bits, const OpArg& src);
void OR(int bits, const OpArg& a1, const OpArg& a2);
void XOR(int bits, const OpArg& a1, const OpArg& a2);
void MOV(int bits, const OpArg& a1, const OpArg& a2);
void TEST(int bits, const OpArg& a1, const OpArg& a2);
// Are these useful at all? Consider removing.
void XCHG(int bits, const OpArg &a1, const OpArg &a2);
void XCHG(int bits, const OpArg& a1, const OpArg& a2);
void XCHG_AHAL();
// Byte swapping (32 and 64-bit only).
@ -536,13 +536,13 @@ public:
void MOVBE(int dbits, const OpArg& dest, const OpArg& src);
// Available only on AMD >= Phenom or Intel >= Haswell
void LZCNT(int bits, X64Reg dest, OpArg src);
void LZCNT(int bits, X64Reg dest, const OpArg& src);
// Note: this one is actually part of BMI1
void TZCNT(int bits, X64Reg dest, OpArg src);
void TZCNT(int bits, X64Reg dest, const OpArg& src);
// WARNING - These two take 11-13 cycles and are VectorPath! (AMD64)
void STMXCSR(OpArg memloc);
void LDMXCSR(OpArg memloc);
void STMXCSR(const OpArg& memloc);
void LDMXCSR(const OpArg& memloc);
// Prefixes
void LOCK();
@ -569,259 +569,259 @@ public:
x87_FPUBusy = 0x8000,
};
void FLD(int bits, OpArg src);
void FST(int bits, OpArg dest);
void FSTP(int bits, OpArg dest);
void FLD(int bits, const OpArg& src);
void FST(int bits, const OpArg& dest);
void FSTP(int bits, const OpArg& dest);
void FNSTSW_AX();
void FWAIT();
// SSE/SSE2: Floating point arithmetic
void ADDSS(X64Reg regOp, OpArg arg);
void ADDSD(X64Reg regOp, OpArg arg);
void SUBSS(X64Reg regOp, OpArg arg);
void SUBSD(X64Reg regOp, OpArg arg);
void MULSS(X64Reg regOp, OpArg arg);
void MULSD(X64Reg regOp, OpArg arg);
void DIVSS(X64Reg regOp, OpArg arg);
void DIVSD(X64Reg regOp, OpArg arg);
void MINSS(X64Reg regOp, OpArg arg);
void MINSD(X64Reg regOp, OpArg arg);
void MAXSS(X64Reg regOp, OpArg arg);
void MAXSD(X64Reg regOp, OpArg arg);
void SQRTSS(X64Reg regOp, OpArg arg);
void SQRTSD(X64Reg regOp, OpArg arg);
void RSQRTSS(X64Reg regOp, OpArg arg);
void ADDSS(X64Reg regOp, const OpArg& arg);
void ADDSD(X64Reg regOp, const OpArg& arg);
void SUBSS(X64Reg regOp, const OpArg& arg);
void SUBSD(X64Reg regOp, const OpArg& arg);
void MULSS(X64Reg regOp, const OpArg& arg);
void MULSD(X64Reg regOp, const OpArg& arg);
void DIVSS(X64Reg regOp, const OpArg& arg);
void DIVSD(X64Reg regOp, const OpArg& arg);
void MINSS(X64Reg regOp, const OpArg& arg);
void MINSD(X64Reg regOp, const OpArg& arg);
void MAXSS(X64Reg regOp, const OpArg& arg);
void MAXSD(X64Reg regOp, const OpArg& arg);
void SQRTSS(X64Reg regOp, const OpArg& arg);
void SQRTSD(X64Reg regOp, const OpArg& arg);
void RSQRTSS(X64Reg regOp, const OpArg& arg);
// SSE/SSE2: Floating point bitwise (yes)
void CMPSS(X64Reg regOp, OpArg arg, u8 compare);
void CMPSD(X64Reg regOp, OpArg arg, u8 compare);
void CMPSS(X64Reg regOp, const OpArg& arg, u8 compare);
void CMPSD(X64Reg regOp, const OpArg& arg, u8 compare);
void CMPEQSS(X64Reg regOp, OpArg arg) { CMPSS(regOp, arg, CMP_EQ); }
void CMPLTSS(X64Reg regOp, OpArg arg) { CMPSS(regOp, arg, CMP_LT); }
void CMPLESS(X64Reg regOp, OpArg arg) { CMPSS(regOp, arg, CMP_LE); }
void CMPUNORDSS(X64Reg regOp, OpArg arg) { CMPSS(regOp, arg, CMP_UNORD); }
void CMPNEQSS(X64Reg regOp, OpArg arg) { CMPSS(regOp, arg, CMP_NEQ); }
void CMPNLTSS(X64Reg regOp, OpArg arg) { CMPSS(regOp, arg, CMP_NLT); }
void CMPORDSS(X64Reg regOp, OpArg arg) { CMPSS(regOp, arg, CMP_ORD); }
void CMPEQSS(X64Reg regOp, const OpArg& arg) { CMPSS(regOp, arg, CMP_EQ); }
void CMPLTSS(X64Reg regOp, const OpArg& arg) { CMPSS(regOp, arg, CMP_LT); }
void CMPLESS(X64Reg regOp, const OpArg& arg) { CMPSS(regOp, arg, CMP_LE); }
void CMPUNORDSS(X64Reg regOp, const OpArg& arg) { CMPSS(regOp, arg, CMP_UNORD); }
void CMPNEQSS(X64Reg regOp, const OpArg& arg) { CMPSS(regOp, arg, CMP_NEQ); }
void CMPNLTSS(X64Reg regOp, const OpArg& arg) { CMPSS(regOp, arg, CMP_NLT); }
void CMPORDSS(X64Reg regOp, const OpArg& arg) { CMPSS(regOp, arg, CMP_ORD); }
// SSE/SSE2: Floating point packed arithmetic (x4 for float, x2 for double)
void ADDPS(X64Reg regOp, OpArg arg);
void ADDPD(X64Reg regOp, OpArg arg);
void SUBPS(X64Reg regOp, OpArg arg);
void SUBPD(X64Reg regOp, OpArg arg);
void CMPPS(X64Reg regOp, OpArg arg, u8 compare);
void CMPPD(X64Reg regOp, OpArg arg, u8 compare);
void MULPS(X64Reg regOp, OpArg arg);
void MULPD(X64Reg regOp, OpArg arg);
void DIVPS(X64Reg regOp, OpArg arg);
void DIVPD(X64Reg regOp, OpArg arg);
void MINPS(X64Reg regOp, OpArg arg);
void MINPD(X64Reg regOp, OpArg arg);
void MAXPS(X64Reg regOp, OpArg arg);
void MAXPD(X64Reg regOp, OpArg arg);
void SQRTPS(X64Reg regOp, OpArg arg);
void SQRTPD(X64Reg regOp, OpArg arg);
void RCPPS(X64Reg regOp, OpArg arg);
void RSQRTPS(X64Reg regOp, OpArg arg);
void ADDPS(X64Reg regOp, const OpArg& arg);
void ADDPD(X64Reg regOp, const OpArg& arg);
void SUBPS(X64Reg regOp, const OpArg& arg);
void SUBPD(X64Reg regOp, const OpArg& arg);
void CMPPS(X64Reg regOp, const OpArg& arg, u8 compare);
void CMPPD(X64Reg regOp, const OpArg& arg, u8 compare);
void MULPS(X64Reg regOp, const OpArg& arg);
void MULPD(X64Reg regOp, const OpArg& arg);
void DIVPS(X64Reg regOp, const OpArg& arg);
void DIVPD(X64Reg regOp, const OpArg& arg);
void MINPS(X64Reg regOp, const OpArg& arg);
void MINPD(X64Reg regOp, const OpArg& arg);
void MAXPS(X64Reg regOp, const OpArg& arg);
void MAXPD(X64Reg regOp, const OpArg& arg);
void SQRTPS(X64Reg regOp, const OpArg& arg);
void SQRTPD(X64Reg regOp, const OpArg& arg);
void RCPPS(X64Reg regOp, const OpArg& arg);
void RSQRTPS(X64Reg regOp, const OpArg& arg);
// SSE/SSE2: Floating point packed bitwise (x4 for float, x2 for double)
void ANDPS(X64Reg regOp, OpArg arg);
void ANDPD(X64Reg regOp, OpArg arg);
void ANDNPS(X64Reg regOp, OpArg arg);
void ANDNPD(X64Reg regOp, OpArg arg);
void ORPS(X64Reg regOp, OpArg arg);
void ORPD(X64Reg regOp, OpArg arg);
void XORPS(X64Reg regOp, OpArg arg);
void XORPD(X64Reg regOp, OpArg arg);
void ANDPS(X64Reg regOp, const OpArg& arg);
void ANDPD(X64Reg regOp, const OpArg& arg);
void ANDNPS(X64Reg regOp, const OpArg& arg);
void ANDNPD(X64Reg regOp, const OpArg& arg);
void ORPS(X64Reg regOp, const OpArg& arg);
void ORPD(X64Reg regOp, const OpArg& arg);
void XORPS(X64Reg regOp, const OpArg& arg);
void XORPD(X64Reg regOp, const OpArg& arg);
// SSE/SSE2: Shuffle components. These are tricky - see Intel documentation.
void SHUFPS(X64Reg regOp, OpArg arg, u8 shuffle);
void SHUFPD(X64Reg regOp, OpArg arg, u8 shuffle);
void SHUFPS(X64Reg regOp, const OpArg& arg, u8 shuffle);
void SHUFPD(X64Reg regOp, const OpArg& arg, u8 shuffle);
// SSE/SSE2: Useful alternative to shuffle in some cases.
void MOVDDUP(X64Reg regOp, OpArg arg);
void MOVDDUP(X64Reg regOp, const OpArg& arg);
// TODO: Actually implement
#if 0
// SSE3: Horizontal operations in SIMD registers. Could be useful for various VFPU things like dot products...
void ADDSUBPS(X64Reg dest, OpArg src);
void ADDSUBPD(X64Reg dest, OpArg src);
void HADDPD(X64Reg dest, OpArg src);
void HSUBPS(X64Reg dest, OpArg src);
void HSUBPD(X64Reg dest, OpArg src);
void ADDSUBPS(X64Reg dest, const OpArg& src);
void ADDSUBPD(X64Reg dest, const OpArg& src);
void HADDPD(X64Reg dest, const OpArg& src);
void HSUBPS(X64Reg dest, const OpArg& src);
void HSUBPD(X64Reg dest, const OpArg& src);
// SSE4: Further horizontal operations - dot products. These are weirdly flexible, the arg contains both a read mask and a write "mask".
void DPPD(X64Reg dest, OpArg src, u8 arg);
void DPPD(X64Reg dest, const OpArg& src, u8 arg);
// These are probably useful for VFPU emulation.
void INSERTPS(X64Reg dest, OpArg src, u8 arg);
void EXTRACTPS(OpArg dest, X64Reg src, u8 arg);
void INSERTPS(X64Reg dest, const OpArg& src, u8 arg);
void EXTRACTPS(const OpArg& dest, X64Reg src, u8 arg);
#endif
// SSE3: Horizontal operations in SIMD registers. Very slow! shufps-based code beats it handily on Ivy.
void HADDPS(X64Reg dest, OpArg src);
void HADDPS(X64Reg dest, const OpArg& src);
// SSE4: Further horizontal operations - dot products. These are weirdly flexible, the arg contains both a read mask and a write "mask".
void DPPS(X64Reg dest, OpArg src, u8 arg);
void DPPS(X64Reg dest, const OpArg& src, u8 arg);
void UNPCKLPS(X64Reg dest, OpArg src);
void UNPCKHPS(X64Reg dest, OpArg src);
void UNPCKLPD(X64Reg dest, OpArg src);
void UNPCKHPD(X64Reg dest, OpArg src);
void UNPCKLPS(X64Reg dest, const OpArg& src);
void UNPCKHPS(X64Reg dest, const OpArg& src);
void UNPCKLPD(X64Reg dest, const OpArg& src);
void UNPCKHPD(X64Reg dest, const OpArg& src);
// SSE/SSE2: Compares.
void COMISS(X64Reg regOp, OpArg arg);
void COMISD(X64Reg regOp, OpArg arg);
void UCOMISS(X64Reg regOp, OpArg arg);
void UCOMISD(X64Reg regOp, OpArg arg);
void COMISS(X64Reg regOp, const OpArg& arg);
void COMISD(X64Reg regOp, const OpArg& arg);
void UCOMISS(X64Reg regOp, const OpArg& arg);
void UCOMISD(X64Reg regOp, const OpArg& arg);
// SSE/SSE2: Moves. Use the right data type for your data, in most cases.
void MOVAPS(X64Reg regOp, OpArg arg);
void MOVAPD(X64Reg regOp, OpArg arg);
void MOVAPS(OpArg arg, X64Reg regOp);
void MOVAPD(OpArg arg, X64Reg regOp);
void MOVAPS(X64Reg regOp, const OpArg& arg);
void MOVAPD(X64Reg regOp, const OpArg& arg);
void MOVAPS(const OpArg& arg, X64Reg regOp);
void MOVAPD(const OpArg& arg, X64Reg regOp);
void MOVUPS(X64Reg regOp, OpArg arg);
void MOVUPD(X64Reg regOp, OpArg arg);
void MOVUPS(OpArg arg, X64Reg regOp);
void MOVUPD(OpArg arg, X64Reg regOp);
void MOVUPS(X64Reg regOp, const OpArg& arg);
void MOVUPD(X64Reg regOp, const OpArg& arg);
void MOVUPS(const OpArg& arg, X64Reg regOp);
void MOVUPD(const OpArg& arg, X64Reg regOp);
void MOVDQA(X64Reg regOp, OpArg arg);
void MOVDQA(OpArg arg, X64Reg regOp);
void MOVDQU(X64Reg regOp, OpArg arg);
void MOVDQU(OpArg arg, X64Reg regOp);
void MOVDQA(X64Reg regOp, const OpArg& arg);
void MOVDQA(const OpArg& arg, X64Reg regOp);
void MOVDQU(X64Reg regOp, const OpArg& arg);
void MOVDQU(const OpArg& arg, X64Reg regOp);
void MOVSS(X64Reg regOp, OpArg arg);
void MOVSD(X64Reg regOp, OpArg arg);
void MOVSS(OpArg arg, X64Reg regOp);
void MOVSD(OpArg arg, X64Reg regOp);
void MOVSS(X64Reg regOp, const OpArg& arg);
void MOVSD(X64Reg regOp, const OpArg& arg);
void MOVSS(const OpArg& arg, X64Reg regOp);
void MOVSD(const OpArg& arg, X64Reg regOp);
void MOVLPS(X64Reg regOp, OpArg arg);
void MOVLPD(X64Reg regOp, OpArg arg);
void MOVLPS(OpArg arg, X64Reg regOp);
void MOVLPD(OpArg arg, X64Reg regOp);
void MOVLPS(X64Reg regOp, const OpArg& arg);
void MOVLPD(X64Reg regOp, const OpArg& arg);
void MOVLPS(const OpArg& arg, X64Reg regOp);
void MOVLPD(const OpArg& arg, X64Reg regOp);
void MOVHPS(X64Reg regOp, OpArg arg);
void MOVHPD(X64Reg regOp, OpArg arg);
void MOVHPS(OpArg arg, X64Reg regOp);
void MOVHPD(OpArg arg, X64Reg regOp);
void MOVHPS(X64Reg regOp, const OpArg& arg);
void MOVHPD(X64Reg regOp, const OpArg& arg);
void MOVHPS(const OpArg& arg, X64Reg regOp);
void MOVHPD(const OpArg& arg, X64Reg regOp);
void MOVHLPS(X64Reg regOp1, X64Reg regOp2);
void MOVLHPS(X64Reg regOp1, X64Reg regOp2);
void MOVD_xmm(X64Reg dest, const OpArg &arg);
void MOVD_xmm(X64Reg dest, const OpArg& arg);
void MOVQ_xmm(X64Reg dest, OpArg arg);
void MOVD_xmm(const OpArg &arg, X64Reg src);
void MOVD_xmm(const OpArg& arg, X64Reg src);
void MOVQ_xmm(OpArg arg, X64Reg src);
// SSE/SSE2: Generates a mask from the high bits of the components of the packed register in question.
void MOVMSKPS(X64Reg dest, OpArg arg);
void MOVMSKPD(X64Reg dest, OpArg arg);
void MOVMSKPS(X64Reg dest, const OpArg& arg);
void MOVMSKPD(X64Reg dest, const OpArg& arg);
// SSE2: Selective byte store, mask in src register. EDI/RDI specifies store address. This is a weird one.
void MASKMOVDQU(X64Reg dest, X64Reg src);
void LDDQU(X64Reg dest, OpArg src);
void LDDQU(X64Reg dest, const OpArg& src);
// SSE/SSE2: Data type conversions.
void CVTPS2PD(X64Reg dest, OpArg src);
void CVTPD2PS(X64Reg dest, OpArg src);
void CVTSS2SD(X64Reg dest, OpArg src);
void CVTSI2SS(X64Reg dest, OpArg src);
void CVTSD2SS(X64Reg dest, OpArg src);
void CVTSI2SD(X64Reg dest, OpArg src);
void CVTDQ2PD(X64Reg regOp, OpArg arg);
void CVTPD2DQ(X64Reg regOp, OpArg arg);
void CVTDQ2PS(X64Reg regOp, OpArg arg);
void CVTPS2DQ(X64Reg regOp, OpArg arg);
void CVTPS2PD(X64Reg dest, const OpArg& src);
void CVTPD2PS(X64Reg dest, const OpArg& src);
void CVTSS2SD(X64Reg dest, const OpArg& src);
void CVTSI2SS(X64Reg dest, const OpArg& src);
void CVTSD2SS(X64Reg dest, const OpArg& src);
void CVTSI2SD(X64Reg dest, const OpArg& src);
void CVTDQ2PD(X64Reg regOp, const OpArg& arg);
void CVTPD2DQ(X64Reg regOp, const OpArg& arg);
void CVTDQ2PS(X64Reg regOp, const OpArg& arg);
void CVTPS2DQ(X64Reg regOp, const OpArg& arg);
void CVTTPS2DQ(X64Reg regOp, OpArg arg);
void CVTTPD2DQ(X64Reg regOp, OpArg arg);
void CVTTPS2DQ(X64Reg regOp, const OpArg& arg);
void CVTTPD2DQ(X64Reg regOp, const OpArg& arg);
// Destinations are X64 regs (rax, rbx, ...) for these instructions.
void CVTSS2SI(X64Reg xregdest, OpArg src);
void CVTSD2SI(X64Reg xregdest, OpArg src);
void CVTTSS2SI(X64Reg xregdest, OpArg arg);
void CVTTSD2SI(X64Reg xregdest, OpArg arg);
void CVTSS2SI(X64Reg xregdest, const OpArg& src);
void CVTSD2SI(X64Reg xregdest, const OpArg& src);
void CVTTSS2SI(X64Reg xregdest, const OpArg& arg);
void CVTTSD2SI(X64Reg xregdest, const OpArg& arg);
// SSE2: Packed integer instructions
void PACKSSDW(X64Reg dest, OpArg arg);
void PACKSSWB(X64Reg dest, OpArg arg);
void PACKUSDW(X64Reg dest, OpArg arg);
void PACKUSWB(X64Reg dest, OpArg arg);
void PACKSSDW(X64Reg dest, const OpArg& arg);
void PACKSSWB(X64Reg dest, const OpArg& arg);
void PACKUSDW(X64Reg dest, const OpArg& arg);
void PACKUSWB(X64Reg dest, const OpArg& arg);
void PUNPCKLBW(X64Reg dest, const OpArg &arg);
void PUNPCKLWD(X64Reg dest, const OpArg &arg);
void PUNPCKLDQ(X64Reg dest, const OpArg &arg);
void PUNPCKLQDQ(X64Reg dest, const OpArg &arg);
void PTEST(X64Reg dest, OpArg arg);
void PAND(X64Reg dest, OpArg arg);
void PANDN(X64Reg dest, OpArg arg);
void PXOR(X64Reg dest, OpArg arg);
void POR(X64Reg dest, OpArg arg);
void PTEST(X64Reg dest, const OpArg& arg);
void PAND(X64Reg dest, const OpArg& arg);
void PANDN(X64Reg dest, const OpArg& arg);
void PXOR(X64Reg dest, const OpArg& arg);
void POR(X64Reg dest, const OpArg& arg);
void PADDB(X64Reg dest, OpArg arg);
void PADDW(X64Reg dest, OpArg arg);
void PADDD(X64Reg dest, OpArg arg);
void PADDQ(X64Reg dest, OpArg arg);
void PADDB(X64Reg dest, const OpArg& arg);
void PADDW(X64Reg dest, const OpArg& arg);
void PADDD(X64Reg dest, const OpArg& arg);
void PADDQ(X64Reg dest, const OpArg& arg);
void PADDSB(X64Reg dest, OpArg arg);
void PADDSW(X64Reg dest, OpArg arg);
void PADDUSB(X64Reg dest, OpArg arg);
void PADDUSW(X64Reg dest, OpArg arg);
void PADDSB(X64Reg dest, const OpArg& arg);
void PADDSW(X64Reg dest, const OpArg& arg);
void PADDUSB(X64Reg dest, const OpArg& arg);
void PADDUSW(X64Reg dest, const OpArg& arg);
void PSUBB(X64Reg dest, OpArg arg);
void PSUBW(X64Reg dest, OpArg arg);
void PSUBD(X64Reg dest, OpArg arg);
void PSUBQ(X64Reg dest, OpArg arg);
void PSUBB(X64Reg dest, const OpArg& arg);
void PSUBW(X64Reg dest, const OpArg& arg);
void PSUBD(X64Reg dest, const OpArg& arg);
void PSUBQ(X64Reg dest, const OpArg& arg);
void PSUBSB(X64Reg dest, OpArg arg);
void PSUBSW(X64Reg dest, OpArg arg);
void PSUBUSB(X64Reg dest, OpArg arg);
void PSUBUSW(X64Reg dest, OpArg arg);
void PSUBSB(X64Reg dest, const OpArg& arg);
void PSUBSW(X64Reg dest, const OpArg& arg);
void PSUBUSB(X64Reg dest, const OpArg& arg);
void PSUBUSW(X64Reg dest, const OpArg& arg);
void PAVGB(X64Reg dest, OpArg arg);
void PAVGW(X64Reg dest, OpArg arg);
void PAVGB(X64Reg dest, const OpArg& arg);
void PAVGW(X64Reg dest, const OpArg& arg);
void PCMPEQB(X64Reg dest, OpArg arg);
void PCMPEQW(X64Reg dest, OpArg arg);
void PCMPEQD(X64Reg dest, OpArg arg);
void PCMPEQB(X64Reg dest, const OpArg& arg);
void PCMPEQW(X64Reg dest, const OpArg& arg);
void PCMPEQD(X64Reg dest, const OpArg& arg);
void PCMPGTB(X64Reg dest, OpArg arg);
void PCMPGTW(X64Reg dest, OpArg arg);
void PCMPGTD(X64Reg dest, OpArg arg);
void PCMPGTB(X64Reg dest, const OpArg& arg);
void PCMPGTW(X64Reg dest, const OpArg& arg);
void PCMPGTD(X64Reg dest, const OpArg& arg);
void PEXTRW(X64Reg dest, OpArg arg, u8 subreg);
void PINSRW(X64Reg dest, OpArg arg, u8 subreg);
void PEXTRW(X64Reg dest, const OpArg& arg, u8 subreg);
void PINSRW(X64Reg dest, const OpArg& arg, u8 subreg);
void PMADDWD(X64Reg dest, OpArg arg);
void PSADBW(X64Reg dest, OpArg arg);
void PMADDWD(X64Reg dest, const OpArg& arg);
void PSADBW(X64Reg dest, const OpArg& arg);
void PMAXSW(X64Reg dest, OpArg arg);
void PMAXUB(X64Reg dest, OpArg arg);
void PMINSW(X64Reg dest, OpArg arg);
void PMINUB(X64Reg dest, OpArg arg);
void PMAXSW(X64Reg dest, const OpArg& arg);
void PMAXUB(X64Reg dest, const OpArg& arg);
void PMINSW(X64Reg dest, const OpArg& arg);
void PMINUB(X64Reg dest, const OpArg& arg);
// SSE4: More MAX/MIN instructions.
void PMINSB(X64Reg dest, OpArg arg);
void PMINSD(X64Reg dest, OpArg arg);
void PMINUW(X64Reg dest, OpArg arg);
void PMINUD(X64Reg dest, OpArg arg);
void PMAXSB(X64Reg dest, OpArg arg);
void PMAXSD(X64Reg dest, OpArg arg);
void PMAXUW(X64Reg dest, OpArg arg);
void PMAXUD(X64Reg dest, OpArg arg);
void PMINSB(X64Reg dest, const OpArg& arg);
void PMINSD(X64Reg dest, const OpArg& arg);
void PMINUW(X64Reg dest, const OpArg& arg);
void PMINUD(X64Reg dest, const OpArg& arg);
void PMAXSB(X64Reg dest, const OpArg& arg);
void PMAXSD(X64Reg dest, const OpArg& arg);
void PMAXUW(X64Reg dest, const OpArg& arg);
void PMAXUD(X64Reg dest, const OpArg& arg);
void PMOVMSKB(X64Reg dest, OpArg arg);
void PSHUFD(X64Reg dest, OpArg arg, u8 shuffle);
void PSHUFB(X64Reg dest, OpArg arg);
void PMOVMSKB(X64Reg dest, const OpArg& arg);
void PSHUFD(X64Reg dest, const OpArg& arg, u8 shuffle);
void PSHUFB(X64Reg dest, const OpArg& arg);
void PSHUFLW(X64Reg dest, OpArg arg, u8 shuffle);
void PSHUFHW(X64Reg dest, OpArg arg, u8 shuffle);
void PSHUFLW(X64Reg dest, const OpArg& arg, u8 shuffle);
void PSHUFHW(X64Reg dest, const OpArg& arg, u8 shuffle);
void PSRLW(X64Reg reg, int shift);
void PSRLD(X64Reg reg, int shift);
void PSRLQ(X64Reg reg, int shift);
void PSRLQ(X64Reg reg, OpArg arg);
void PSRLQ(X64Reg reg, const OpArg& arg);
void PSRLDQ(X64Reg reg, int shift);
void PSLLW(X64Reg reg, int shift);
@ -833,198 +833,198 @@ public:
void PSRAD(X64Reg reg, int shift);
// SSE4: data type conversions
void PMOVSXBW(X64Reg dest, OpArg arg);
void PMOVSXBD(X64Reg dest, OpArg arg);
void PMOVSXBQ(X64Reg dest, OpArg arg);
void PMOVSXWD(X64Reg dest, OpArg arg);
void PMOVSXWQ(X64Reg dest, OpArg arg);
void PMOVSXDQ(X64Reg dest, OpArg arg);
void PMOVZXBW(X64Reg dest, OpArg arg);
void PMOVZXBD(X64Reg dest, OpArg arg);
void PMOVZXBQ(X64Reg dest, OpArg arg);
void PMOVZXWD(X64Reg dest, OpArg arg);
void PMOVZXWQ(X64Reg dest, OpArg arg);
void PMOVZXDQ(X64Reg dest, OpArg arg);
void PMOVSXBW(X64Reg dest, const OpArg& arg);
void PMOVSXBD(X64Reg dest, const OpArg& arg);
void PMOVSXBQ(X64Reg dest, const OpArg& arg);
void PMOVSXWD(X64Reg dest, const OpArg& arg);
void PMOVSXWQ(X64Reg dest, const OpArg& arg);
void PMOVSXDQ(X64Reg dest, const OpArg& arg);
void PMOVZXBW(X64Reg dest, const OpArg& arg);
void PMOVZXBD(X64Reg dest, const OpArg& arg);
void PMOVZXBQ(X64Reg dest, const OpArg& arg);
void PMOVZXWD(X64Reg dest, const OpArg& arg);
void PMOVZXWQ(X64Reg dest, const OpArg& arg);
void PMOVZXDQ(X64Reg dest, const OpArg& arg);
// SSE4: variable blend instructions (xmm0 implicit argument)
void PBLENDVB(X64Reg dest, OpArg arg);
void BLENDVPS(X64Reg dest, OpArg arg);
void BLENDVPD(X64Reg dest, OpArg arg);
void PBLENDVB(X64Reg dest, const OpArg& arg);
void BLENDVPS(X64Reg dest, const OpArg& arg);
void BLENDVPD(X64Reg dest, const OpArg& arg);
void BLENDPS(X64Reg dest, const OpArg& arg, u8 blend);
void BLENDPD(X64Reg dest, const OpArg& arg, u8 blend);
// SSE4: rounding (see FloatRound for mode or use ROUNDNEARSS, etc. helpers.)
void ROUNDSS(X64Reg dest, OpArg arg, u8 mode);
void ROUNDSD(X64Reg dest, OpArg arg, u8 mode);
void ROUNDPS(X64Reg dest, OpArg arg, u8 mode);
void ROUNDPD(X64Reg dest, OpArg arg, u8 mode);
void ROUNDSS(X64Reg dest, const OpArg& arg, u8 mode);
void ROUNDSD(X64Reg dest, const OpArg& arg, u8 mode);
void ROUNDPS(X64Reg dest, const OpArg& arg, u8 mode);
void ROUNDPD(X64Reg dest, const OpArg& arg, u8 mode);
void ROUNDNEARSS(X64Reg dest, OpArg arg) { ROUNDSS(dest, arg, FROUND_NEAREST); }
void ROUNDFLOORSS(X64Reg dest, OpArg arg) { ROUNDSS(dest, arg, FROUND_FLOOR); }
void ROUNDCEILSS(X64Reg dest, OpArg arg) { ROUNDSS(dest, arg, FROUND_CEIL); }
void ROUNDZEROSS(X64Reg dest, OpArg arg) { ROUNDSS(dest, arg, FROUND_ZERO); }
void ROUNDNEARSS(X64Reg dest, const OpArg& arg) { ROUNDSS(dest, arg, FROUND_NEAREST); }
void ROUNDFLOORSS(X64Reg dest, const OpArg& arg) { ROUNDSS(dest, arg, FROUND_FLOOR); }
void ROUNDCEILSS(X64Reg dest, const OpArg& arg) { ROUNDSS(dest, arg, FROUND_CEIL); }
void ROUNDZEROSS(X64Reg dest, const OpArg& arg) { ROUNDSS(dest, arg, FROUND_ZERO); }
void ROUNDNEARSD(X64Reg dest, OpArg arg) { ROUNDSD(dest, arg, FROUND_NEAREST); }
void ROUNDFLOORSD(X64Reg dest, OpArg arg) { ROUNDSD(dest, arg, FROUND_FLOOR); }
void ROUNDCEILSD(X64Reg dest, OpArg arg) { ROUNDSD(dest, arg, FROUND_CEIL); }
void ROUNDZEROSD(X64Reg dest, OpArg arg) { ROUNDSD(dest, arg, FROUND_ZERO); }
void ROUNDNEARSD(X64Reg dest, const OpArg& arg) { ROUNDSD(dest, arg, FROUND_NEAREST); }
void ROUNDFLOORSD(X64Reg dest, const OpArg& arg) { ROUNDSD(dest, arg, FROUND_FLOOR); }
void ROUNDCEILSD(X64Reg dest, const OpArg& arg) { ROUNDSD(dest, arg, FROUND_CEIL); }
void ROUNDZEROSD(X64Reg dest, const OpArg& arg) { ROUNDSD(dest, arg, FROUND_ZERO); }
void ROUNDNEARPS(X64Reg dest, OpArg arg) { ROUNDPS(dest, arg, FROUND_NEAREST); }
void ROUNDFLOORPS(X64Reg dest, OpArg arg) { ROUNDPS(dest, arg, FROUND_FLOOR); }
void ROUNDCEILPS(X64Reg dest, OpArg arg) { ROUNDPS(dest, arg, FROUND_CEIL); }
void ROUNDZEROPS(X64Reg dest, OpArg arg) { ROUNDPS(dest, arg, FROUND_ZERO); }
void ROUNDNEARPS(X64Reg dest, const OpArg& arg) { ROUNDPS(dest, arg, FROUND_NEAREST); }
void ROUNDFLOORPS(X64Reg dest, const OpArg& arg) { ROUNDPS(dest, arg, FROUND_FLOOR); }
void ROUNDCEILPS(X64Reg dest, const OpArg& arg) { ROUNDPS(dest, arg, FROUND_CEIL); }
void ROUNDZEROPS(X64Reg dest, const OpArg& arg) { ROUNDPS(dest, arg, FROUND_ZERO); }
void ROUNDNEARPD(X64Reg dest, OpArg arg) { ROUNDPD(dest, arg, FROUND_NEAREST); }
void ROUNDFLOORPD(X64Reg dest, OpArg arg) { ROUNDPD(dest, arg, FROUND_FLOOR); }
void ROUNDCEILPD(X64Reg dest, OpArg arg) { ROUNDPD(dest, arg, FROUND_CEIL); }
void ROUNDZEROPD(X64Reg dest, OpArg arg) { ROUNDPD(dest, arg, FROUND_ZERO); }
void ROUNDNEARPD(X64Reg dest, const OpArg& arg) { ROUNDPD(dest, arg, FROUND_NEAREST); }
void ROUNDFLOORPD(X64Reg dest, const OpArg& arg) { ROUNDPD(dest, arg, FROUND_FLOOR); }
void ROUNDCEILPD(X64Reg dest, const OpArg& arg) { ROUNDPD(dest, arg, FROUND_CEIL); }
void ROUNDZEROPD(X64Reg dest, const OpArg& arg) { ROUNDPD(dest, arg, FROUND_ZERO); }
// AVX
void VADDSD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VSUBSD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VMULSD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VDIVSD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VADDPD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VSUBPD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VMULPD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VDIVPD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VSQRTSD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VSHUFPD(X64Reg regOp1, X64Reg regOp2, OpArg arg, u8 shuffle);
void VUNPCKLPD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VUNPCKHPD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VADDSD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VSUBSD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VMULSD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VDIVSD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VADDPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VSUBPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VMULPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VDIVPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VSQRTSD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VSHUFPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg, u8 shuffle);
void VUNPCKLPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VUNPCKHPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VANDPS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VANDPD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VANDNPS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VANDNPD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VORPS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VORPD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VXORPS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VXORPD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VANDPS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VANDPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VANDNPS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VANDNPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VORPS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VORPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VXORPS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VXORPD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VPAND(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VPANDN(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VPOR(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VPXOR(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VPAND(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VPANDN(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VPOR(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VPXOR(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
// FMA3
void VFMADD132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD132SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD213SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD231SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD132SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD213SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD231SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB132SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB213SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB231SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB132SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB213SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB231SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD132SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD213SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD231SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD132SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD213SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD231SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB132SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB213SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB231SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB132SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB213SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB231SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADDSUB132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADDSUB213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADDSUB231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADDSUB132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADDSUB213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADDSUB231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUBADD132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUBADD213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUBADD231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUBADD132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUBADD213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUBADD231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD132PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMADD213PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMADD231PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMADD132PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMADD213PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMADD231PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMADD132SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMADD213SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMADD231SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMADD132SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMADD213SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMADD231SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMSUB132PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMSUB213PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMSUB231PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMSUB132PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMSUB213PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMSUB231PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMSUB132SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMSUB213SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMSUB231SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMSUB132SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMSUB213SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMSUB231SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMADD132PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMADD213PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMADD231PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMADD132PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMADD213PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMADD231PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMADD132SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMADD213SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMADD231SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMADD132SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMADD213SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMADD231SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMSUB132PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMSUB213PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMSUB231PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMSUB132PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMSUB213PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMSUB231PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMSUB132SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMSUB213SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMSUB231SS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMSUB132SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMSUB213SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFNMSUB231SD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMADDSUB132PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMADDSUB213PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMADDSUB231PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMADDSUB132PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMADDSUB213PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMADDSUB231PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMSUBADD132PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMSUBADD213PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMSUBADD231PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMSUBADD132PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMSUBADD213PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMSUBADD231PD(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
// VEX GPR instructions
void SARX(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2);
void SHLX(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2);
void SHRX(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2);
void RORX(int bits, X64Reg regOp, OpArg arg, u8 rotate);
void PEXT(int bits, X64Reg regOp1, X64Reg regOp2, OpArg arg);
void PDEP(int bits, X64Reg regOp1, X64Reg regOp2, OpArg arg);
void MULX(int bits, X64Reg regOp1, X64Reg regOp2, OpArg arg);
void BZHI(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2);
void BLSR(int bits, X64Reg regOp, OpArg arg);
void BLSMSK(int bits, X64Reg regOp, OpArg arg);
void BLSI(int bits, X64Reg regOp, OpArg arg);
void BEXTR(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2);
void ANDN(int bits, X64Reg regOp1, X64Reg regOp2, OpArg arg);
void SARX(int bits, X64Reg regOp1, const OpArg& arg, X64Reg regOp2);
void SHLX(int bits, X64Reg regOp1, const OpArg& arg, X64Reg regOp2);
void SHRX(int bits, X64Reg regOp1, const OpArg& arg, X64Reg regOp2);
void RORX(int bits, X64Reg regOp, const OpArg& arg, u8 rotate);
void PEXT(int bits, X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void PDEP(int bits, X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void MULX(int bits, X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void BZHI(int bits, X64Reg regOp1, const OpArg& arg, X64Reg regOp2);
void BLSR(int bits, X64Reg regOp, const OpArg& arg);
void BLSMSK(int bits, X64Reg regOp, const OpArg& arg);
void BLSI(int bits, X64Reg regOp, const OpArg& arg);
void BEXTR(int bits, X64Reg regOp1, const OpArg& arg, X64Reg regOp2);
void ANDN(int bits, X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void RDTSC();
// Utility functions
// The difference between this and CALL is that this aligns the stack
// where appropriate.
void ABI_CallFunction(const void *func);
void ABI_CallFunction(const void* func);
template <typename T>
void ABI_CallFunction(T (*func)()) {
ABI_CallFunction((const void *)func);
ABI_CallFunction((const void*)func);
}
void ABI_CallFunction(const u8 *func) {
ABI_CallFunction((const void *)func);
void ABI_CallFunction(const u8* func) {
ABI_CallFunction((const void*)func);
}
void ABI_CallFunctionC16(const void *func, u16 param1);
void ABI_CallFunctionCC16(const void *func, u32 param1, u16 param2);
void ABI_CallFunctionC16(const void* func, u16 param1);
void ABI_CallFunctionCC16(const void* func, u32 param1, u16 param2);
// These only support u32 parameters, but that's enough for a lot of uses.
// These will destroy the 1 or 2 first "parameter regs".
void ABI_CallFunctionC(const void *func, u32 param1);
void ABI_CallFunctionCC(const void *func, u32 param1, u32 param2);
void ABI_CallFunctionCCC(const void *func, u32 param1, u32 param2, u32 param3);
void ABI_CallFunctionCCP(const void *func, u32 param1, u32 param2, void *param3);
void ABI_CallFunctionCCCP(const void *func, u32 param1, u32 param2, u32 param3, void *param4);
void ABI_CallFunctionP(const void *func, void *param1);
void ABI_CallFunctionPA(const void *func, void *param1, const Gen::OpArg &arg2);
void ABI_CallFunctionPAA(const void *func, void *param1, const Gen::OpArg &arg2, const Gen::OpArg &arg3);
void ABI_CallFunctionPPC(const void *func, void *param1, void *param2, u32 param3);
void ABI_CallFunctionAC(const void *func, const Gen::OpArg &arg1, u32 param2);
void ABI_CallFunctionACC(const void *func, const Gen::OpArg &arg1, u32 param2, u32 param3);
void ABI_CallFunctionA(const void *func, const Gen::OpArg &arg1);
void ABI_CallFunctionAA(const void *func, const Gen::OpArg &arg1, const Gen::OpArg &arg2);
void ABI_CallFunctionC(const void* func, u32 param1);
void ABI_CallFunctionCC(const void* func, u32 param1, u32 param2);
void ABI_CallFunctionCCC(const void* func, u32 param1, u32 param2, u32 param3);
void ABI_CallFunctionCCP(const void* func, u32 param1, u32 param2, void* param3);
void ABI_CallFunctionCCCP(const void* func, u32 param1, u32 param2, u32 param3, void* param4);
void ABI_CallFunctionP(const void* func, void* param1);
void ABI_CallFunctionPA(const void* func, void* param1, const OpArg& arg2);
void ABI_CallFunctionPAA(const void* func, void* param1, const OpArg& arg2, const OpArg& arg3);
void ABI_CallFunctionPPC(const void* func, void* param1, void* param2, u32 param3);
void ABI_CallFunctionAC(const void* func, const OpArg& arg1, u32 param2);
void ABI_CallFunctionACC(const void* func, const OpArg& arg1, u32 param2, u32 param3);
void ABI_CallFunctionA(const void* func, const OpArg& arg1);
void ABI_CallFunctionAA(const void* func, const OpArg& arg1, const OpArg& arg2);
// Pass a register as a parameter.
void ABI_CallFunctionR(const void *func, X64Reg reg1);
void ABI_CallFunctionRR(const void *func, X64Reg reg1, X64Reg reg2);
void ABI_CallFunctionR(const void* func, X64Reg reg1);
void ABI_CallFunctionRR(const void* func, X64Reg reg1, X64Reg reg2);
template <typename Tr, typename T1>
void ABI_CallFunctionC(Tr (*func)(T1), u32 param1) {
ABI_CallFunctionC((const void *)func, param1);
ABI_CallFunctionC((const void*)func, param1);
}
// A function that doesn't have any control over what it will do to regs,