suyu/dynarmic
26
1
Fork 1
Code Issues Pull requests Projects Releases Packages Wiki Activity

RegAlloc: Allow allocation of XMM registers

Browse source
This commit is contained in:
MerryMage 2016-08-02 13:46:12 +01:00
parent 864081d1a0
commit 4414ec5bc8
5 changed files with 218 additions and 131 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

136
src/backend_x64/emit_x64.cpp
View file

@ -89,12 +89,12 @@ CodePtr EmitX64::Emit(const Arm::LocationDescriptor descriptor, Dynarmic::IR::Bl
void EmitX64::EmitIdentity(IR::Block& block, IR::Inst* inst) {
// TODO: Possible unnecessary mov here.
reg_alloc.UseDefRegister(inst->GetArg(0), inst);
reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
}
void EmitX64::EmitGetRegister(IR::Block&, IR::Inst* inst) {
Arm::Reg reg = inst->GetArg(0).GetRegRef();
X64Reg result = reg_alloc.DefRegister(inst);
X64Reg result = reg_alloc.DefRegister(inst, any_gpr);
code->MOV(32, R(result), MJitStateReg(reg));
}
@ -104,13 +104,13 @@ void EmitX64::EmitSetRegister(IR::Block&, IR::Inst* inst) {
if (arg.IsImmediate()) {
code->MOV(32, MJitStateReg(reg), Imm32(arg.GetU32()));
} else {
X64Reg to_store = reg_alloc.UseRegister(arg.GetInst());
X64Reg to_store = reg_alloc.UseRegister(arg.GetInst(), any_gpr);
code->MOV(32, MJitStateReg(reg), R(to_store));
}
}
void EmitX64::EmitGetNFlag(IR::Block&, IR::Inst* inst) {
X64Reg result = reg_alloc.DefRegister(inst);
X64Reg result = reg_alloc.DefRegister(inst, any_gpr);
code->MOV(32, R(result), MJitStateCpsr());
code->SHR(32, R(result), Imm8(31));
}
@ -126,7 +126,7 @@ void EmitX64::EmitSetNFlag(IR::Block&, IR::Inst* inst) {
code->AND(32, MJitStateCpsr(), Imm32(~flag_mask));
}
} else {
X64Reg to_store = reg_alloc.UseScratchRegister(arg.GetInst());
X64Reg to_store = reg_alloc.UseScratchRegister(arg.GetInst(), any_gpr);
code->SHL(32, R(to_store), Imm8(flag_bit));
code->AND(32, MJitStateCpsr(), Imm32(~flag_mask));
@ -135,7 +135,7 @@ void EmitX64::EmitSetNFlag(IR::Block&, IR::Inst* inst) {
}
void EmitX64::EmitGetZFlag(IR::Block&, IR::Inst* inst) {
X64Reg result = reg_alloc.DefRegister(inst);
X64Reg result = reg_alloc.DefRegister(inst, any_gpr);
code->MOV(32, R(result), MJitStateCpsr());
code->SHR(32, R(result), Imm8(30));
code->AND(32, R(result), Imm32(1));
@ -152,7 +152,7 @@ void EmitX64::EmitSetZFlag(IR::Block&, IR::Inst* inst) {
code->AND(32, MJitStateCpsr(), Imm32(~flag_mask));
}
} else {
X64Reg to_store = reg_alloc.UseScratchRegister(arg.GetInst());
X64Reg to_store = reg_alloc.UseScratchRegister(arg.GetInst(), any_gpr);
code->SHL(32, R(to_store), Imm8(flag_bit));
code->AND(32, MJitStateCpsr(), Imm32(~flag_mask));
@ -161,7 +161,7 @@ void EmitX64::EmitSetZFlag(IR::Block&, IR::Inst* inst) {
}
void EmitX64::EmitGetCFlag(IR::Block&, IR::Inst* inst) {
X64Reg result = reg_alloc.DefRegister(inst);
X64Reg result = reg_alloc.DefRegister(inst, any_gpr);
code->MOV(32, R(result), MJitStateCpsr());
code->SHR(32, R(result), Imm8(29));
code->AND(32, R(result), Imm32(1));
@ -178,7 +178,7 @@ void EmitX64::EmitSetCFlag(IR::Block&, IR::Inst* inst) {
code->AND(32, MJitStateCpsr(), Imm32(~flag_mask));
}
} else {
X64Reg to_store = reg_alloc.UseScratchRegister(arg.GetInst());
X64Reg to_store = reg_alloc.UseScratchRegister(arg.GetInst(), any_gpr);
code->SHL(32, R(to_store), Imm8(flag_bit));
code->AND(32, MJitStateCpsr(), Imm32(~flag_mask));
@ -187,7 +187,7 @@ void EmitX64::EmitSetCFlag(IR::Block&, IR::Inst* inst) {
}
void EmitX64::EmitGetVFlag(IR::Block&, IR::Inst* inst) {
X64Reg result = reg_alloc.DefRegister(inst);
X64Reg result = reg_alloc.DefRegister(inst, any_gpr);
code->MOV(32, R(result), MJitStateCpsr());
code->SHR(32, R(result), Imm8(28));
code->AND(32, R(result), Imm32(1));
@ -204,7 +204,7 @@ void EmitX64::EmitSetVFlag(IR::Block&, IR::Inst* inst) {
code->AND(32, MJitStateCpsr(), Imm32(~flag_mask));
}
} else {
X64Reg to_store = reg_alloc.UseScratchRegister(arg.GetInst());
X64Reg to_store = reg_alloc.UseScratchRegister(arg.GetInst(), any_gpr);
code->SHL(32, R(to_store), Imm8(flag_bit));
code->AND(32, MJitStateCpsr(), Imm32(~flag_mask));
@ -237,9 +237,9 @@ void EmitX64::EmitBXWritePC(IR::Block&, IR::Inst* inst) {
code->AND(32, MJitStateCpsr(), Imm32(~T_bit));
}
} else {
X64Reg new_pc = reg_alloc.UseScratchRegister(arg.GetInst());
X64Reg tmp1 = reg_alloc.ScratchRegister();
X64Reg tmp2 = reg_alloc.ScratchRegister();
X64Reg new_pc = reg_alloc.UseScratchRegister(arg.GetInst(), any_gpr);
X64Reg tmp1 = reg_alloc.ScratchRegister(any_gpr);
X64Reg tmp2 = reg_alloc.ScratchRegister(any_gpr);
code->MOV(32, R(tmp1), MJitStateCpsr());
code->MOV(32, R(tmp2), R(tmp1));
@ -274,17 +274,17 @@ void EmitX64::EmitGetOverflowFromOp(IR::Block&, IR::Inst*) {
void EmitX64::EmitLeastSignificantHalf(IR::Block&, IR::Inst* inst) {
// TODO: Optimize
reg_alloc.UseDefRegister(inst->GetArg(0), inst);
reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
}
void EmitX64::EmitLeastSignificantByte(IR::Block&, IR::Inst* inst) {
// TODO: Optimize
reg_alloc.UseDefRegister(inst->GetArg(0), inst);
reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
}
void EmitX64::EmitMostSignificantBit(IR::Block&, IR::Inst* inst) {
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
// TODO: Flag optimization
@ -292,7 +292,7 @@ void EmitX64::EmitMostSignificantBit(IR::Block&, IR::Inst* inst) {
}
void EmitX64::EmitIsZero(IR::Block&, IR::Inst* inst) {
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
// TODO: Flag optimization
@ -315,7 +315,7 @@ void EmitX64::EmitLogicalShiftLeft(IR::Block& block, IR::Inst* inst) {
auto shift_arg = inst->GetArg(1);
if (shift_arg.IsImmediate()) {
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
u8 shift = shift_arg.GetU8();
if (shift <= 31) {
@ -325,8 +325,8 @@ void EmitX64::EmitLogicalShiftLeft(IR::Block& block, IR::Inst* inst) {
}
} else {
X64Reg shift = reg_alloc.UseRegister(shift_arg.GetInst(), {HostLoc::RCX});
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg zero = reg_alloc.ScratchRegister();
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
X64Reg zero = reg_alloc.ScratchRegister(any_gpr);
// The 32-bit x64 SHL instruction masks the shift count by 0x1F before performing the shift.
// ARM differs from the behaviour: It does not mask the count, so shifts above 31 result in zeros.
@ -344,8 +344,8 @@ void EmitX64::EmitLogicalShiftLeft(IR::Block& block, IR::Inst* inst) {
if (shift_arg.IsImmediate()) {
u8 shift = shift_arg.GetU8();
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg carry = reg_alloc.UseDefRegister(inst->GetArg(2), carry_inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
X64Reg carry = reg_alloc.UseDefRegister(inst->GetArg(2), carry_inst, any_gpr);
if (shift == 0) {
// There is nothing more to do.
@ -363,8 +363,8 @@ void EmitX64::EmitLogicalShiftLeft(IR::Block& block, IR::Inst* inst) {
}
} else {
X64Reg shift = reg_alloc.UseRegister(shift_arg.GetInst(), {HostLoc::RCX});
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg carry = reg_alloc.UseDefRegister(inst->GetArg(2), carry_inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
X64Reg carry = reg_alloc.UseDefRegister(inst->GetArg(2), carry_inst, any_gpr);
// TODO: Optimize this.
@ -405,7 +405,7 @@ void EmitX64::EmitLogicalShiftRight(IR::Block& block, IR::Inst* inst) {
auto shift_arg = inst->GetArg(1);
if (shift_arg.IsImmediate()) {
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
u8 shift = shift_arg.GetU8();
if (shift <= 31) {
@ -415,8 +415,8 @@ void EmitX64::EmitLogicalShiftRight(IR::Block& block, IR::Inst* inst) {
}
} else {
X64Reg shift = reg_alloc.UseRegister(shift_arg.GetInst(), {HostLoc::RCX});
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg zero = reg_alloc.ScratchRegister();
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
X64Reg zero = reg_alloc.ScratchRegister(any_gpr);
// The 32-bit x64 SHR instruction masks the shift count by 0x1F before performing the shift.
// ARM differs from the behaviour: It does not mask the count, so shifts above 31 result in zeros.
@ -434,8 +434,8 @@ void EmitX64::EmitLogicalShiftRight(IR::Block& block, IR::Inst* inst) {
if (shift_arg.IsImmediate()) {
u8 shift = shift_arg.GetU8();
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg carry = reg_alloc.UseDefRegister(inst->GetArg(2), carry_inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
X64Reg carry = reg_alloc.UseDefRegister(inst->GetArg(2), carry_inst, any_gpr);
if (shift == 0) {
// There is nothing more to do.
@ -452,8 +452,8 @@ void EmitX64::EmitLogicalShiftRight(IR::Block& block, IR::Inst* inst) {
}
} else {
X64Reg shift = reg_alloc.UseRegister(shift_arg.GetInst(), {HostLoc::RCX});
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg carry = reg_alloc.UseDefRegister(inst->GetArg(2), carry_inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
X64Reg carry = reg_alloc.UseDefRegister(inst->GetArg(2), carry_inst, any_gpr);
// TODO: Optimize this.
@ -498,13 +498,13 @@ void EmitX64::EmitArithmeticShiftRight(IR::Block& block, IR::Inst* inst) {
if (shift_arg.IsImmediate()) {
u8 shift = shift_arg.GetU8();
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
code->SAR(32, R(result), Imm8(shift < 31 ? shift : 31));
} else {
X64Reg shift = reg_alloc.UseScratchRegister(shift_arg.GetInst(), {HostLoc::RCX});
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg const31 = reg_alloc.ScratchRegister();
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
X64Reg const31 = reg_alloc.ScratchRegister(any_gpr);
// The 32-bit x64 SAR instruction masks the shift count by 0x1F before performing the shift.
// ARM differs from the behaviour: It does not mask the count.
@ -524,8 +524,8 @@ void EmitX64::EmitArithmeticShiftRight(IR::Block& block, IR::Inst* inst) {
if (shift_arg.IsImmediate()) {
u8 shift = shift_arg.GetU8();
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg carry = reg_alloc.UseDefRegister(inst->GetArg(2), carry_inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
X64Reg carry = reg_alloc.UseDefRegister(inst->GetArg(2), carry_inst, any_gpr);
if (shift == 0) {
// There is nothing more to do.
@ -539,8 +539,8 @@ void EmitX64::EmitArithmeticShiftRight(IR::Block& block, IR::Inst* inst) {
}
} else {
X64Reg shift = reg_alloc.UseRegister(shift_arg.GetInst(), {HostLoc::RCX});
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg carry = reg_alloc.UseDefRegister(inst->GetArg(2), carry_inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
X64Reg carry = reg_alloc.UseDefRegister(inst->GetArg(2), carry_inst, any_gpr);
// TODO: Optimize this.
@ -578,12 +578,12 @@ void EmitX64::EmitRotateRight(IR::Block& block, IR::Inst* inst) {
if (shift_arg.IsImmediate()) {
u8 shift = shift_arg.GetU8();
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
code->ROR(32, R(result), Imm8(shift & 0x1F));
} else {
X64Reg shift = reg_alloc.UseRegister(shift_arg.GetInst(), {HostLoc::RCX});
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
// x64 ROR instruction does (shift & 0x1F) for us.
code->ROR(32, R(result), R(shift));
@ -596,8 +596,8 @@ void EmitX64::EmitRotateRight(IR::Block& block, IR::Inst* inst) {
if (shift_arg.IsImmediate()) {
u8 shift = shift_arg.GetU8();
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg carry = reg_alloc.UseDefRegister(inst->GetArg(2), carry_inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
X64Reg carry = reg_alloc.UseDefRegister(inst->GetArg(2), carry_inst, any_gpr);
if (shift == 0) {
// There is nothing more to do.
@ -610,8 +610,8 @@ void EmitX64::EmitRotateRight(IR::Block& block, IR::Inst* inst) {
}
} else {
X64Reg shift = reg_alloc.UseScratchRegister(shift_arg.GetInst(), {HostLoc::RCX});
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg carry = reg_alloc.UseDefRegister(inst->GetArg(2), carry_inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
X64Reg carry = reg_alloc.UseDefRegister(inst->GetArg(2), carry_inst, any_gpr);
// TODO: Optimize
@ -638,10 +638,10 @@ void EmitX64::EmitRotateRight(IR::Block& block, IR::Inst* inst) {
static X64Reg DoCarry(RegAlloc& reg_alloc, const IR::Value& carry_in, IR::Inst* carry_out) {
if (carry_in.IsImmediate()) {
return carry_out ? reg_alloc.DefRegister(carry_out) : INVALID_REG;
return carry_out ? reg_alloc.DefRegister(carry_out, any_gpr) : INVALID_REG;
} else {
IR::Inst* in = carry_in.GetInst();
return carry_out ? reg_alloc.UseDefRegister(in, carry_out) : reg_alloc.UseRegister(in);
return carry_out ? reg_alloc.UseDefRegister(in, carry_out, any_gpr) : reg_alloc.UseRegister(in, any_gpr);
}
}
@ -653,15 +653,15 @@ void EmitX64::EmitAddWithCarry(IR::Block& block, IR::Inst* inst) {
IR::Value b = inst->GetArg(1);
IR::Value carry_in = inst->GetArg(2);
X64Reg result = reg_alloc.UseDefRegister(a, inst);
X64Reg result = reg_alloc.UseDefRegister(a, inst, any_gpr);
X64Reg carry = DoCarry(reg_alloc, carry_in, carry_inst);
X64Reg overflow = overflow_inst ? reg_alloc.DefRegister(overflow_inst) : INVALID_REG;
X64Reg overflow = overflow_inst ? reg_alloc.DefRegister(overflow_inst, any_gpr) : INVALID_REG;
// TODO: Consider using LEA.
OpArg op_arg = b.IsImmediate()
? Imm32(b.GetU32())
: R(reg_alloc.UseRegister(b.GetInst()));
: R(reg_alloc.UseRegister(b.GetInst(), any_gpr));
if (carry_in.IsImmediate()) {
if (carry_in.GetU1()) {
@ -695,9 +695,9 @@ void EmitX64::EmitSubWithCarry(IR::Block& block, IR::Inst* inst) {
IR::Value b = inst->GetArg(1);
IR::Value carry_in = inst->GetArg(2);
X64Reg result = reg_alloc.UseDefRegister(a, inst);
X64Reg result = reg_alloc.UseDefRegister(a, inst, any_gpr);
X64Reg carry = DoCarry(reg_alloc, carry_in, carry_inst);
X64Reg overflow = overflow_inst ? reg_alloc.DefRegister(overflow_inst) : INVALID_REG;
X64Reg overflow = overflow_inst ? reg_alloc.DefRegister(overflow_inst, any_gpr) : INVALID_REG;
// TODO: Consider using LEA.
// TODO: Optimize CMP case.
@ -705,7 +705,7 @@ void EmitX64::EmitSubWithCarry(IR::Block& block, IR::Inst* inst) {
OpArg op_arg = b.IsImmediate()
? Imm32(b.GetU32())
: R(reg_alloc.UseRegister(b.GetInst()));
: R(reg_alloc.UseRegister(b.GetInst(), any_gpr));
if (carry_in.IsImmediate()) {
if (carry_in.GetU1()) {
@ -736,10 +736,10 @@ void EmitX64::EmitAnd(IR::Block&, IR::Inst* inst) {
IR::Value a = inst->GetArg(0);
IR::Value b = inst->GetArg(1);
X64Reg result = reg_alloc.UseDefRegister(a, inst);
X64Reg result = reg_alloc.UseDefRegister(a, inst, any_gpr);
OpArg op_arg = b.IsImmediate()
? Imm32(b.GetU32())
: R(reg_alloc.UseRegister(b.GetInst()));
: R(reg_alloc.UseRegister(b.GetInst(), any_gpr));
code->AND(32, R(result), op_arg);
}
@ -748,10 +748,10 @@ void EmitX64::EmitEor(IR::Block&, IR::Inst* inst) {
IR::Value a = inst->GetArg(0);
IR::Value b = inst->GetArg(1);
X64Reg result = reg_alloc.UseDefRegister(a, inst);
X64Reg result = reg_alloc.UseDefRegister(a, inst, any_gpr);
OpArg op_arg = b.IsImmediate()
? Imm32(b.GetU32())
: R(reg_alloc.UseRegister(b.GetInst()));
: R(reg_alloc.UseRegister(b.GetInst(), any_gpr));
code->XOR(32, R(result), op_arg);
}
@ -760,10 +760,10 @@ void EmitX64::EmitOr(IR::Block&, IR::Inst* inst) {
IR::Value a = inst->GetArg(0);
IR::Value b = inst->GetArg(1);
X64Reg result = reg_alloc.UseDefRegister(a, inst);
X64Reg result = reg_alloc.UseDefRegister(a, inst, any_gpr);
OpArg op_arg = b.IsImmediate()
? Imm32(b.GetU32())
: R(reg_alloc.UseRegister(b.GetInst()));
: R(reg_alloc.UseRegister(b.GetInst(), any_gpr));
code->OR(32, R(result), op_arg);
}
@ -772,11 +772,11 @@ void EmitX64::EmitNot(IR::Block&, IR::Inst* inst) {
IR::Value a = inst->GetArg(0);
if (a.IsImmediate()) {
X64Reg result = reg_alloc.DefRegister(inst);
X64Reg result = reg_alloc.DefRegister(inst, any_gpr);
code->MOV(32, R(result), Imm32(~a.GetU32()));
} else {
X64Reg result = reg_alloc.UseDefRegister(a.GetInst(), inst);
X64Reg result = reg_alloc.UseDefRegister(a.GetInst(), inst, any_gpr);
code->NOT(32, R(result));
}
@ -784,46 +784,46 @@ void EmitX64::EmitNot(IR::Block&, IR::Inst* inst) {
void EmitX64::EmitSignExtendHalfToWord(IR::Block&, IR::Inst* inst) {
// TODO: Remove unnecessary mov that may occur here
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
code->MOVSX(32, 16, result, R(result));
}
void EmitX64::EmitSignExtendByteToWord(IR::Block&, IR::Inst* inst) {
// TODO: Remove unnecessary mov that may occur here
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
code->MOVSX(32, 8, result, R(result));
}
void EmitX64::EmitZeroExtendHalfToWord(IR::Block&, IR::Inst* inst) {
// TODO: Remove unnecessary mov that may occur here
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
code->MOVZX(32, 16, result, R(result));
}
void EmitX64::EmitZeroExtendByteToWord(IR::Block&, IR::Inst* inst) {
// TODO: Remove unnecessary mov that may occur here
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
code->MOVZX(32, 8, result, R(result));
}
void EmitX64::EmitByteReverseWord(IR::Block&, IR::Inst* inst) {
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
code->BSWAP(32, result);
}
void EmitX64::EmitByteReverseHalf(IR::Block&, IR::Inst* inst) {
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
code->ROL(16, R(result), Imm8(8));
}
void EmitX64::EmitByteReverseDual(IR::Block&, IR::Inst* inst) {
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst);
X64Reg result = reg_alloc.UseDefRegister(inst->GetArg(0), inst, any_gpr);
code->BSWAP(64, result);
}

2
src/backend_x64/jitstate.h
View file

@ -21,7 +21,7 @@ struct JitState {
u32 Fpscr = 0;
std::array<u32, SpillCount> Spill{}; // Spill.
std::array<u64, SpillCount> Spill{}; // Spill.
// For internal use (See: Routines::RunCode)
u64 save_host_RSP = 0;

135
src/backend_x64/reg_alloc.cpp
View file

@ -21,13 +21,14 @@ static Gen::X64Reg HostLocToX64(HostLoc loc) {
}
static Gen::OpArg SpillToOpArg(HostLoc loc) {
static_assert(std::is_same<decltype(JitState{}.Spill[0]), u64&>::value, "Spill must be u64");
DEBUG_ASSERT(HostLocIsSpill(loc));
size_t i = static_cast<size_t>(loc) - static_cast<size_t>(HostLoc::FirstSpill);
return Gen::MDisp(Gen::R15, static_cast<int>(offsetof(JitState, Spill) + i * sizeof(u32)));
return Gen::MDisp(Gen::R15, static_cast<int>(offsetof(JitState, Spill) + i * sizeof(u64)));
}
Gen::X64Reg RegAlloc::DefRegister(IR::Inst* def_inst, std::initializer_list<HostLoc> desired_locations) {
Gen::X64Reg RegAlloc::DefRegister(IR::Inst* def_inst, HostLocList desired_locations) {
DEBUG_ASSERT(std::all_of(desired_locations.begin(), desired_locations.end(), HostLocIsRegister));
DEBUG_ASSERT_MSG(ValueLocations(def_inst).empty(), "def_inst has already been defined");
@ -38,13 +39,11 @@ Gen::X64Reg RegAlloc::DefRegister(IR::Inst* def_inst, std::initializer_list<Host
}
// Update state
hostloc_state[static_cast<size_t>(location)] = HostLocState::Def;
hostloc_to_inst[static_cast<size_t>(location)] = def_inst;
LocInfo(location) = {def_inst, HostLocState::Def};
return HostLocToX64(location);
}
Gen::X64Reg RegAlloc::UseDefRegister(IR::Value use_value, IR::Inst* def_inst, std::initializer_list<HostLoc> desired_locations) {
Gen::X64Reg RegAlloc::UseDefRegister(IR::Value use_value, IR::Inst* def_inst, HostLocList desired_locations) {
if (!use_value.IsImmediate()) {
return UseDefRegister(use_value.GetInst(), def_inst, desired_locations);
}
@ -52,19 +51,19 @@ Gen::X64Reg RegAlloc::UseDefRegister(IR::Value use_value, IR::Inst* def_inst, st
return LoadImmediateIntoRegister(use_value, DefRegister(def_inst, desired_locations));
}
Gen::X64Reg RegAlloc::UseDefRegister(IR::Inst* use_inst, IR::Inst* def_inst, std::initializer_list<HostLoc> desired_locations) {
Gen::X64Reg RegAlloc::UseDefRegister(IR::Inst* use_inst, IR::Inst* def_inst, HostLocList desired_locations) {
DEBUG_ASSERT(std::all_of(desired_locations.begin(), desired_locations.end(), HostLocIsRegister));
DEBUG_ASSERT_MSG(ValueLocations(def_inst).empty(), "def_inst has already been defined");
DEBUG_ASSERT_MSG(!ValueLocations(use_inst).empty(), "use_inst has not been defined");
// TODO: Optimize the case when this is the last use_inst use.
Gen::X64Reg use_reg = UseRegister(use_inst);
Gen::X64Reg use_reg = UseRegister(use_inst, any_gpr);
Gen::X64Reg def_reg = DefRegister(def_inst, desired_locations);
code->MOV(32, Gen::R(def_reg), Gen::R(use_reg));
code->MOV(64, Gen::R(def_reg), Gen::R(use_reg));
return def_reg;
}
Gen::X64Reg RegAlloc::UseRegister(IR::Value use_value, std::initializer_list<HostLoc> desired_locations) {
Gen::X64Reg RegAlloc::UseRegister(IR::Value use_value, HostLocList desired_locations) {
if (!use_value.IsImmediate()) {
return UseRegister(use_value.GetInst(), desired_locations);
}
@ -72,17 +71,18 @@ Gen::X64Reg RegAlloc::UseRegister(IR::Value use_value, std::initializer_list<Hos
return LoadImmediateIntoRegister(use_value, ScratchRegister(desired_locations));
}
Gen::X64Reg RegAlloc::UseRegister(IR::Inst* use_inst, std::initializer_list<HostLoc> desired_locations) {
Gen::X64Reg RegAlloc::UseRegister(IR::Inst* use_inst, HostLocList desired_locations) {
DEBUG_ASSERT(std::all_of(desired_locations.begin(), desired_locations.end(), HostLocIsRegister));
DEBUG_ASSERT_MSG(!ValueLocations(use_inst).empty(), "use_inst has not been defined");
HostLoc current_location = ValueLocations(use_inst).front();
auto iter = std::find(desired_locations.begin(), desired_locations.end(), current_location);
if (iter != desired_locations.end()) {
ASSERT(hostloc_state[static_cast<size_t>(current_location)] == HostLocState::Idle || hostloc_state[static_cast<size_t>(current_location)] == HostLocState::Use);
ASSERT(LocInfo(current_location).state == HostLocState::Idle ||
LocInfo(current_location).state == HostLocState::Use);
// Update state
hostloc_state[static_cast<size_t>(current_location)] = HostLocState::Use;
LocInfo(current_location).state = HostLocState::Use;
DecrementRemainingUses(use_inst);
return HostLocToX64(current_location);
@ -95,18 +95,19 @@ Gen::X64Reg RegAlloc::UseRegister(IR::Inst* use_inst, std::initializer_list<Host
SpillRegister(new_location);
}
code->MOV(32, Gen::R(HostLocToX64(new_location)), SpillToOpArg(current_location));
EmitMove(new_location, current_location);
hostloc_state[static_cast<size_t>(new_location)] = HostLocState::Use;
std::swap(hostloc_to_inst[static_cast<size_t>(new_location)], hostloc_to_inst[static_cast<size_t>(current_location)]);
LocInfo(new_location) = LocInfo(current_location);
LocInfo(new_location).state = HostLocState::Use;
LocInfo(current_location) = {};
DecrementRemainingUses(use_inst);
} else if (HostLocIsRegister(current_location)) {
ASSERT(hostloc_state[static_cast<size_t>(current_location)] == HostLocState::Idle);
ASSERT(LocInfo(current_location).state == HostLocState::Idle);
code->XCHG(32, Gen::R(HostLocToX64(new_location)), Gen::R(HostLocToX64(current_location)));
EmitExchange(new_location, current_location);
hostloc_state[static_cast<size_t>(new_location)] = HostLocState::Use;
std::swap(hostloc_to_inst[static_cast<size_t>(new_location)], hostloc_to_inst[static_cast<size_t>(current_location)]);
std::swap(LocInfo(new_location), LocInfo(current_location));
LocInfo(new_location).state = HostLocState::Use;
DecrementRemainingUses(use_inst);
} else {
ASSERT_MSG(0, "Invalid current_location");
@ -115,7 +116,7 @@ Gen::X64Reg RegAlloc::UseRegister(IR::Inst* use_inst, std::initializer_list<Host
return HostLocToX64(new_location);
}
Gen::X64Reg RegAlloc::UseScratchRegister(IR::Value use_value, std::initializer_list<HostLoc> desired_locations) {
Gen::X64Reg RegAlloc::UseScratchRegister(IR::Value use_value, HostLocList desired_locations) {
if (!use_value.IsImmediate()) {
return UseScratchRegister(use_value.GetInst(), desired_locations);
}
@ -123,7 +124,7 @@ Gen::X64Reg RegAlloc::UseScratchRegister(IR::Value use_value, std::initializer_l
return LoadImmediateIntoRegister(use_value, ScratchRegister(desired_locations));
}
Gen::X64Reg RegAlloc::UseScratchRegister(IR::Inst* use_inst, std::initializer_list<HostLoc> desired_locations) {
Gen::X64Reg RegAlloc::UseScratchRegister(IR::Inst* use_inst, HostLocList desired_locations) {
DEBUG_ASSERT(std::all_of(desired_locations.begin(), desired_locations.end(), HostLocIsRegister));
DEBUG_ASSERT_MSG(!ValueLocations(use_inst).empty(), "use_inst has not been defined");
ASSERT_MSG(use_inst->use_count != 0, "use_inst ran out of uses. (Use-d an IR::Inst* too many times)");
@ -136,23 +137,23 @@ Gen::X64Reg RegAlloc::UseScratchRegister(IR::Inst* use_inst, std::initializer_li
SpillRegister(new_location);
}
code->MOV(32, Gen::R(HostLocToX64(new_location)), SpillToOpArg(current_location));
EmitMove(new_location, current_location);
hostloc_state[static_cast<size_t>(new_location)] = HostLocState::Scratch;
LocInfo(new_location).state = HostLocState::Scratch;
DecrementRemainingUses(use_inst);
} else if (HostLocIsRegister(current_location)) {
ASSERT(hostloc_state[static_cast<size_t>(current_location)] == HostLocState::Idle);
ASSERT(LocInfo(current_location).state == HostLocState::Idle);
if (IsRegisterOccupied(new_location)) {
SpillRegister(new_location);
if (current_location != new_location) {
code->MOV(32, Gen::R(HostLocToX64(new_location)), Gen::R(HostLocToX64(current_location)));
EmitMove(new_location, current_location);
}
} else {
code->MOV(32, Gen::R(HostLocToX64(new_location)), Gen::R(HostLocToX64(current_location)));
EmitMove(new_location, current_location);
}
hostloc_state[static_cast<size_t>(new_location)] = HostLocState::Scratch;
LocInfo(new_location).state = HostLocState::Scratch;
DecrementRemainingUses(use_inst);
} else {
ASSERT_MSG(0, "Invalid current_location");
@ -161,8 +162,7 @@ Gen::X64Reg RegAlloc::UseScratchRegister(IR::Inst* use_inst, std::initializer_li
return HostLocToX64(new_location);
}
Gen::X64Reg RegAlloc::ScratchRegister(std::initializer_list<HostLoc> desired_locations) {
Gen::X64Reg RegAlloc::ScratchRegister(HostLocList desired_locations) {
DEBUG_ASSERT(std::all_of(desired_locations.begin(), desired_locations.end(), HostLocIsRegister));
HostLoc location = SelectARegister(desired_locations);
@ -172,7 +172,7 @@ Gen::X64Reg RegAlloc::ScratchRegister(std::initializer_list<HostLoc> desired_loc
}
// Update state
hostloc_state[static_cast<size_t>(location)] = HostLocState::Scratch;
LocInfo(location).state = HostLocState::Scratch;
return HostLocToX64(location);
}
@ -235,7 +235,7 @@ void RegAlloc::HostCall(IR::Inst* result_def, IR::Value arg0_use, IR::Value arg1
code->MOV(64, Gen::R(Gen::RSP), Gen::MDisp(Gen::R15, offsetof(JitState, save_host_RSP)));
}
HostLoc RegAlloc::SelectARegister(std::initializer_list<HostLoc> desired_locations) const {
HostLoc RegAlloc::SelectARegister(HostLocList desired_locations) const {
std::vector<HostLoc> candidates = desired_locations;
// Find all locations that have not been allocated..
@ -259,32 +259,32 @@ std::vector<HostLoc> RegAlloc::ValueLocations(IR::Inst* value) const {
std::vector<HostLoc> locations;
for (size_t i = 0; i < HostLocCount; i++)
if (hostloc_to_inst[i] == value)
if (hostloc_info[i].value == value)
locations.emplace_back(static_cast<HostLoc>(i));
return locations;
}
bool RegAlloc::IsRegisterOccupied(HostLoc loc) const {
return hostloc_to_inst.at(static_cast<size_t>(loc)) != nullptr;
return GetLocInfo(loc).value != nullptr;
}
bool RegAlloc::IsRegisterAllocated(HostLoc loc) const {
return hostloc_state.at(static_cast<size_t>(loc)) != HostLocState::Idle;
return GetLocInfo(loc).state != HostLocState::Idle;
}
void RegAlloc::SpillRegister(HostLoc loc) {
ASSERT_MSG(HostLocIsRegister(loc), "Only registers can be spilled");
ASSERT_MSG(hostloc_state[static_cast<size_t>(loc)] == HostLocState::Idle, "Allocated registers cannot be spilled");
ASSERT_MSG(LocInfo(loc).state == HostLocState::Idle, "Allocated registers cannot be spilled");
ASSERT_MSG(IsRegisterOccupied(loc), "There is no need to spill unoccupied registers");
ASSERT_MSG(!IsRegisterAllocated(loc), "Registers that have been allocated must not be spilt");
HostLoc new_loc = FindFreeSpill();
code->MOV(32, SpillToOpArg(new_loc), Gen::R(HostLocToX64(loc)));
EmitMove(new_loc, loc);
hostloc_to_inst[static_cast<size_t>(new_loc)] = hostloc_to_inst[static_cast<size_t>(loc)];
hostloc_to_inst[static_cast<size_t>(loc)] = nullptr;
LocInfo(new_loc).value = LocInfo(loc).value;
LocInfo(loc).value = nullptr;
}
HostLoc RegAlloc::FindFreeSpill() const {
@ -296,11 +296,11 @@ HostLoc RegAlloc::FindFreeSpill() const {
}
void RegAlloc::EndOfAllocScope() {
hostloc_state.fill(HostLocState::Idle);
for (auto& iter : hostloc_to_inst)
if (iter && iter->use_count == 0)
iter = nullptr;
for (auto& iter : hostloc_info) {
iter.state = HostLocState::Idle;
if (iter.value && iter.value->use_count == 0)
iter.value = nullptr;
}
}
void RegAlloc::DecrementRemainingUses(IR::Inst* value) {
@ -309,12 +309,53 @@ void RegAlloc::DecrementRemainingUses(IR::Inst* value) {
}
void RegAlloc::AssertNoMoreUses() {
ASSERT(std::all_of(hostloc_to_inst.begin(), hostloc_to_inst.end(), [](const auto& inst){ return !inst; }));
ASSERT(std::all_of(hostloc_info.begin(), hostloc_info.end(), [](const auto& i){ return !i.value; }));
}
void RegAlloc::Reset() {
hostloc_to_inst.fill(nullptr);
hostloc_state.fill(HostLocState::Idle);
hostloc_info.fill({});
}
void RegAlloc::EmitMove(HostLoc to, HostLoc from) {
const auto& from_info = LocInfo(from);
if (HostLocIsXMM(to) && HostLocIsSpill(from)) {
if (from_info.GetType() == IR::Type::F64) {
code->MOVSD(HostLocToX64(to), SpillToOpArg(from));
} else if (from_info.GetType() == IR::Type::F32) {
code->MOVSS(HostLocToX64(to), SpillToOpArg(from));
} else {
ASSERT_MSG(false, "Tried to move a non-fp value into an XMM register");
}
} else if (HostLocIsSpill(to) && HostLocIsXMM(from)) {
if (from_info.GetType() == IR::Type::F64) {
code->MOVSD(SpillToOpArg(to), HostLocToX64(from));
} else if (from_info.GetType() == IR::Type::F32) {
code->MOVSS(SpillToOpArg(to), HostLocToX64(from));
} else {
ASSERT_MSG(false, "Tried to move a non-fp value into an XMM register");
}
} else if (HostLocIsXMM(to) && HostLocIsXMM(from)) {
code->MOVAPS(HostLocToX64(to), Gen::R(HostLocToX64(from)));
} else if (HostLocIsGPR(to) && HostLocIsSpill(from)) {
code->MOV(64, Gen::R(HostLocToX64(to)), SpillToOpArg(from));
} else if (HostLocIsSpill(to) && HostLocIsGPR(from)) {
code->MOV(64, SpillToOpArg(to), Gen::R(HostLocToX64(from)));
} else if (HostLocIsGPR(to) && HostLocIsGPR(from)){
code->MOV(64, Gen::R(HostLocToX64(to)), Gen::R(HostLocToX64(from)));
} else {
ASSERT_MSG(false, "Invalid RegAlloc::EmitMove");
}
}
void RegAlloc::EmitExchange(HostLoc a, HostLoc b) {
if (HostLocIsGPR(a) && HostLocIsGPR(b)) {
code->XCHG(64, Gen::R(HostLocToX64(a)), Gen::R(HostLocToX64(b)));
} else if (HostLocIsXMM(a) && HostLocIsXMM(b)) {
ASSERT_MSG(false, "Exchange is unnecessary for XMM registers");
} else {
ASSERT_MSG(false, "Invalid RegAlloc::EmitExchange");
}
}
} // namespace BackendX64

74
src/backend_x64/reg_alloc.h
View file

@ -19,6 +19,8 @@ namespace BackendX64 {
enum class HostLoc {
// Ordering of the registers is intentional. See also: HostLocToX64.
RAX, RCX, RDX, RBX, RSP, RBP, RSI, RDI, R8, R9, R10, R11, R12, R13, R14,
XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, XMM15,
CF, PF, AF, ZF, SF, OF,
FirstSpill,
};
@ -29,10 +31,18 @@ enum class HostLocState {
Idle, Def, Use, Scratch
};
inline bool HostLocIsRegister(HostLoc reg) {
inline bool HostLocIsGPR(HostLoc reg) {
return reg >= HostLoc::RAX && reg <= HostLoc::R14;
}
inline bool HostLocIsXMM(HostLoc reg) {
return reg >= HostLoc::XMM0 && reg <= HostLoc::XMM15;
}
inline bool HostLocIsRegister(HostLoc reg) {
return HostLocIsGPR(reg) || HostLocIsXMM(reg);
}
inline bool HostLocIsFlag(HostLoc reg) {
return reg >= HostLoc::CF && reg <= HostLoc::OF;
}
@ -46,7 +56,9 @@ inline bool HostLocIsSpill(HostLoc reg) {
return reg >= HostLoc::FirstSpill && reg <= HostLocSpill(SpillCount - 1);
}
const std::initializer_list<HostLoc> hostloc_any_register = {
using HostLocList = std::initializer_list<HostLoc>;
const HostLocList any_gpr = {
HostLoc::RAX,
HostLoc::RBX,
HostLoc::RCX,
@ -61,7 +73,26 @@ const std::initializer_list<HostLoc> hostloc_any_register = {
HostLoc::R11,
HostLoc::R12,
HostLoc::R13,
HostLoc::R14
HostLoc::R14,
};
const HostLocList any_xmm = {
HostLoc::XMM0,
HostLoc::XMM1,
HostLoc::XMM2,
HostLoc::XMM3,
HostLoc::XMM4,
HostLoc::XMM5,
HostLoc::XMM6,
HostLoc::XMM7,
HostLoc::XMM8,
HostLoc::XMM9,
HostLoc::XMM10,
HostLoc::XMM11,
HostLoc::XMM12,
HostLoc::XMM13,
HostLoc::XMM14,
HostLoc::XMM15,
};
class RegAlloc final {
@ -69,18 +100,18 @@ public:
RegAlloc(Gen::XEmitter* code) : code(code) {}
/// Late-def
Gen::X64Reg DefRegister(IR::Inst* def_inst, std::initializer_list<HostLoc> desired_locations = hostloc_any_register);
Gen::X64Reg DefRegister(IR::Inst* def_inst, HostLocList desired_locations);
/// Early-use, Late-def
Gen::X64Reg UseDefRegister(IR::Value use_value, IR::Inst* def_inst, std::initializer_list<HostLoc> desired_locations = hostloc_any_register);
Gen::X64Reg UseDefRegister(IR::Inst* use_inst, IR::Inst* def_inst, std::initializer_list<HostLoc> desired_locations = hostloc_any_register);
Gen::X64Reg UseDefRegister(IR::Value use_value, IR::Inst* def_inst, HostLocList desired_locations);
Gen::X64Reg UseDefRegister(IR::Inst* use_inst, IR::Inst* def_inst, HostLocList desired_locations);
/// Early-use
Gen::X64Reg UseRegister(IR::Value use_value, std::initializer_list<HostLoc> desired_locations = hostloc_any_register);
Gen::X64Reg UseRegister(IR::Inst* use_inst, std::initializer_list<HostLoc> desired_locations = hostloc_any_register);
Gen::X64Reg UseRegister(IR::Value use_value, HostLocList desired_locations);
Gen::X64Reg UseRegister(IR::Inst* use_inst, HostLocList desired_locations);
/// Early-use, Destroyed
Gen::X64Reg UseScratchRegister(IR::Value use_value, std::initializer_list<HostLoc> desired_locations = hostloc_any_register);
Gen::X64Reg UseScratchRegister(IR::Inst* use_inst, std::initializer_list<HostLoc> desired_locations = hostloc_any_register);
Gen::X64Reg UseScratchRegister(IR::Value use_value, HostLocList desired_locations);
Gen::X64Reg UseScratchRegister(IR::Inst* use_inst, HostLocList desired_locations);
/// Early-def, Late-use, single-use
Gen::X64Reg ScratchRegister(std::initializer_list<HostLoc> desired_locations = hostloc_any_register);
Gen::X64Reg ScratchRegister(HostLocList desired_locations);
Gen::X64Reg LoadImmediateIntoRegister(IR::Value imm, Gen::X64Reg reg);
/// Late-def for result register, Early-use for all arguments, Each value is placed into registers according to host ABI.
@ -97,19 +128,32 @@ public:
void Reset();
private:
HostLoc SelectARegister(std::initializer_list<HostLoc> desired_locations) const;
HostLoc SelectARegister(HostLocList desired_locations) const;
std::vector<HostLoc> ValueLocations(IR::Inst* value) const;
bool IsRegisterOccupied(HostLoc loc) const;
bool IsRegisterAllocated(HostLoc loc) const;
void EmitMove(HostLoc to, HostLoc from);
void EmitExchange(HostLoc a, HostLoc b);
void SpillRegister(HostLoc loc);
HostLoc FindFreeSpill() const;
Gen::XEmitter* code = nullptr;
using mapping_map_t = std::array<IR::Inst*, HostLocCount>;
mapping_map_t hostloc_to_inst;
std::array<HostLocState, HostLocCount> hostloc_state;
struct HostLocInfo {
IR::Inst* value = nullptr;
HostLocState state = HostLocState::Idle;
IR::Type GetType() const {
return value ? value->GetType() : IR::Type::Void;
}
};
std::array<HostLocInfo, HostLocCount> hostloc_info;
HostLocInfo& LocInfo(HostLoc loc) {
return hostloc_info[static_cast<size_t>(loc)];
}
HostLocInfo GetLocInfo(HostLoc loc) const {
return hostloc_info[static_cast<size_t>(loc)];
}
};
} // namespace BackendX64

2
src/frontend/ir/ir.h
View file

@ -41,6 +41,8 @@ enum class Type {
U16 = 1 << 5,
U32 = 1 << 6,
U64 = 1 << 7,
F32 = 1 << 8,
F64 = 1 << 9,
};
Type GetTypeOf(Opcode op);