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

A32: Implement ASIMD VRECPE

Browse source
This commit is contained in:
MerryMage 2020-06-20 15:07:06 +01:00
parent d3dc50d718
commit 6f59c2cd8e
8 changed files with 47 additions and 18 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

17
src/backend/x64/emit_x64_vector_floating_point.cpp
View file

@ -378,11 +378,12 @@ void EmitThreeOpVectorOperation(BlockOfCode& code, EmitContext& ctx, IR::Inst* i
ctx.reg_alloc.DefineValue(inst, result); ctx.reg_alloc.DefineValue(inst, result);
} }
template<typename Lambda> template<FpcrControlledArgument fcarg = FpcrControlledArgument::Absent, typename Lambda>
void EmitTwoOpFallback(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, Lambda lambda) { void EmitTwoOpFallback(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, Lambda lambda) {
const auto fn = static_cast<mp::equivalent_function_type<Lambda>*>(lambda); const auto fn = static_cast<mp::equivalent_function_type<Lambda>*>(lambda);
auto args = ctx.reg_alloc.GetArgumentInfo(inst); auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const bool fpcr_controlled = fcarg == FpcrControlledArgument::Absent || args[1].GetImmediateU1();
const Xbyak::Xmm arg1 = ctx.reg_alloc.UseXmm(args[0]); const Xbyak::Xmm arg1 = ctx.reg_alloc.UseXmm(args[0]);
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
ctx.reg_alloc.EndOfAllocScope(); ctx.reg_alloc.EndOfAllocScope();
@ -392,7 +393,7 @@ void EmitTwoOpFallback(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, Lamb
code.sub(rsp, stack_space + ABI_SHADOW_SPACE); code.sub(rsp, stack_space + ABI_SHADOW_SPACE);
code.lea(code.ABI_PARAM1, ptr[rsp + ABI_SHADOW_SPACE + 0 * 16]); code.lea(code.ABI_PARAM1, ptr[rsp + ABI_SHADOW_SPACE + 0 * 16]);
code.lea(code.ABI_PARAM2, ptr[rsp + ABI_SHADOW_SPACE + 1 * 16]); code.lea(code.ABI_PARAM2, ptr[rsp + ABI_SHADOW_SPACE + 1 * 16]);
code.mov(code.ABI_PARAM3.cvt32(), ctx.FPCR().Value()); code.mov(code.ABI_PARAM3.cvt32(), ctx.FPCR(fpcr_controlled).Value());
code.lea(code.ABI_PARAM4, code.ptr[code.r15 + code.GetJitStateInfo().offsetof_fpsr_exc]); code.lea(code.ABI_PARAM4, code.ptr[code.r15 + code.GetJitStateInfo().offsetof_fpsr_exc]);
code.movaps(xword[code.ABI_PARAM2], arg1); code.movaps(xword[code.ABI_PARAM2], arg1);
@ -1144,7 +1145,7 @@ void EmitX64::EmitFPVectorPairedAddLower64(EmitContext& ctx, IR::Inst* inst) {
template<typename FPT> template<typename FPT>
static void EmitRecipEstimate(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) { static void EmitRecipEstimate(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
EmitTwoOpFallback(code, ctx, inst, [](VectorArray<FPT>& result, const VectorArray<FPT>& operand, FP::FPCR fpcr, FP::FPSR& fpsr) { EmitTwoOpFallback<FpcrControlledArgument::Present>(code, ctx, inst, [](VectorArray<FPT>& result, const VectorArray<FPT>& operand, FP::FPCR fpcr, FP::FPSR& fpsr) {
for (size_t i = 0; i < result.size(); i++) { for (size_t i = 0; i < result.size(); i++) {
result[i] = FP::FPRecipEstimate<FPT>(operand[i], fpcr, fpsr); result[i] = FP::FPRecipEstimate<FPT>(operand[i], fpcr, fpsr);
} }
@ -1188,12 +1189,12 @@ static void EmitRecipStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
MaybeStandardFPSCRValue(code, ctx, fpcr_controlled, [&]{ MaybeStandardFPSCRValue(code, ctx, fpcr_controlled, [&]{
code.movaps(result, GetVectorOf<fsize, false, 0, 2>(code)); code.movaps(result, GetVectorOf<fsize, false, 0, 2>(code));
FCODE(vfnmadd231p)(result, operand1, operand2); FCODE(vfnmadd231p)(result, operand1, operand2);
});
FCODE(vcmpunordp)(tmp, result, result); FCODE(vcmpunordp)(tmp, result, result);
code.vptest(tmp, tmp); code.vptest(tmp, tmp);
code.jnz(fallback, code.T_NEAR); code.jnz(fallback, code.T_NEAR);
code.L(end); code.L(end);
});
code.SwitchToFarCode(); code.SwitchToFarCode();
code.L(fallback); code.L(fallback);

2
src/frontend/A32/decoder/asimd.inc
View file

@ -103,7 +103,7 @@ INST(asimd_VSWP, "VSWP", "111100111D110010dddd000
//INST(asimd_VQMOVN, "VQMOVN", "111100111-11--10----00101x-0----") // ASIMD //INST(asimd_VQMOVN, "VQMOVN", "111100111-11--10----00101x-0----") // ASIMD
//INST(asimd_VSHLL_max, "VSHLL_max", "111100111-11--10----001100-0----") // ASIMD //INST(asimd_VSHLL_max, "VSHLL_max", "111100111-11--10----001100-0----") // ASIMD
//INST(asimd_VCVT_half, "VCVT (half-precision)", "111100111-11--10----011x00-0----") // ASIMD //INST(asimd_VCVT_half, "VCVT (half-precision)", "111100111-11--10----011x00-0----") // ASIMD
//INST(asimd_VRECPE, "VRECPE", "111100111-11--11----010x0x-0----") // ASIMD INST(asimd_VRECPE, "VRECPE", "111100111D11zz11dddd010F0QM0mmmm") // ASIMD
//INST(asimd_VRSQRTE, "VRSQRTE", "111100111-11--11----010x1x-0----") // ASIMD //INST(asimd_VRSQRTE, "VRSQRTE", "111100111-11--11----010x1x-0----") // ASIMD
//INST(asimd_VCVT_integer, "VCVT (integer)", "111100111-11--11----011xxx-0----") // ASIMD //INST(asimd_VCVT_integer, "VCVT (integer)", "111100111-11--11----011xxx-0----") // ASIMD

27
src/frontend/A32/translate/impl/asimd_two_regs_misc.cpp
View file

@ -352,4 +352,31 @@ bool ArmTranslatorVisitor::asimd_VSWP(bool D, size_t Vd, bool Q, bool M, size_t
return true; return true;
} }
bool ArmTranslatorVisitor::asimd_VRECPE(bool D, size_t sz, size_t Vd, bool F, bool Q, bool M, size_t Vm) {
if (Q && (Common::Bit<0>(Vd) || Common::Bit<0>(Vm))) {
return UndefinedInstruction();
}
if (sz == 0b00 || sz == 0b11) {
return UndefinedInstruction();
}
if (!F && sz == 0b01) {
// TODO: Implement 16-bit VectorUnsignedRecipEstimate
return UndefinedInstruction();
}
const size_t esize = 8U << sz;
const auto d = ToVector(Q, Vd, D);
const auto m = ToVector(Q, Vm, M);
const auto reg_m = ir.GetVector(m);
const auto result = F ? ir.FPVectorRecipEstimate(esize, reg_m, false)
: ir.VectorUnsignedRecipEstimate(reg_m);
ir.SetVector(d, result);
return true;
}
} // namespace Dynarmic::A32 } // namespace Dynarmic::A32

1
src/frontend/A32/translate/impl/translate_arm.h
View file

@ -499,6 +499,7 @@ struct ArmTranslatorVisitor final {
bool asimd_VABS(bool D, size_t sz, size_t Vd, bool F, bool Q, bool M, size_t Vm); bool asimd_VABS(bool D, size_t sz, size_t Vd, bool F, bool Q, bool M, size_t Vm);
bool asimd_VNEG(bool D, size_t sz, size_t Vd, bool F, bool Q, bool M, size_t Vm); bool asimd_VNEG(bool D, size_t sz, size_t Vd, bool F, bool Q, bool M, size_t Vm);
bool asimd_VSWP(bool D, size_t Vd, bool Q, bool M, size_t Vm); bool asimd_VSWP(bool D, size_t Vd, bool Q, bool M, size_t Vm);
bool asimd_VRECPE(bool D, size_t sz, size_t Vd, bool F, bool Q, bool M, size_t Vm);
// Advanced SIMD load/store structures // Advanced SIMD load/store structures
bool v8_VST_multiple(bool D, Reg n, size_t Vd, Imm<4> type, size_t sz, size_t align, Reg m); bool v8_VST_multiple(bool D, Reg n, size_t Vd, Imm<4> type, size_t sz, size_t align, Reg m);

8
src/frontend/ir/ir_emitter.cpp
View file

@ -2440,14 +2440,14 @@ U128 IREmitter::FPVectorPairedAddLower(size_t esize, const U128& a, const U128&
UNREACHABLE(); UNREACHABLE();
} }
U128 IREmitter::FPVectorRecipEstimate(size_t esize, const U128& a) { U128 IREmitter::FPVectorRecipEstimate(size_t esize, const U128& a, bool fpcr_controlled) {
switch (esize) { switch (esize) {
case 16: case 16:
return Inst<U128>(Opcode::FPVectorRecipEstimate16, a); return Inst<U128>(Opcode::FPVectorRecipEstimate16, a, Imm1(fpcr_controlled));
case 32: case 32:
return Inst<U128>(Opcode::FPVectorRecipEstimate32, a); return Inst<U128>(Opcode::FPVectorRecipEstimate32, a, Imm1(fpcr_controlled));
case 64: case 64:
return Inst<U128>(Opcode::FPVectorRecipEstimate64, a); return Inst<U128>(Opcode::FPVectorRecipEstimate64, a, Imm1(fpcr_controlled));
} }
UNREACHABLE(); UNREACHABLE();
} }

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

@ -360,7 +360,7 @@ public:
U128 FPVectorNeg(size_t esize, const U128& a); U128 FPVectorNeg(size_t esize, const U128& a);
U128 FPVectorPairedAdd(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true); U128 FPVectorPairedAdd(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true);
U128 FPVectorPairedAddLower(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true); U128 FPVectorPairedAddLower(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true);
U128 FPVectorRecipEstimate(size_t esize, const U128& a); U128 FPVectorRecipEstimate(size_t esize, const U128& a, bool fpcr_controlled = true);
U128 FPVectorRecipStepFused(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true); U128 FPVectorRecipStepFused(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true);
U128 FPVectorRoundInt(size_t esize, const U128& operand, FP::RoundingMode rounding, bool exact); U128 FPVectorRoundInt(size_t esize, const U128& operand, FP::RoundingMode rounding, bool exact);
U128 FPVectorRSqrtEstimate(size_t esize, const U128& a); U128 FPVectorRSqrtEstimate(size_t esize, const U128& a);

6
src/frontend/ir/opcodes.inc
View file

@ -613,9 +613,9 @@ OPCODE(FPVectorPairedAdd32, U128, U128
OPCODE(FPVectorPairedAdd64, U128, U128, U128, U1 ) OPCODE(FPVectorPairedAdd64, U128, U128, U128, U1 )
OPCODE(FPVectorPairedAddLower32, U128, U128, U128, U1 ) OPCODE(FPVectorPairedAddLower32, U128, U128, U128, U1 )
OPCODE(FPVectorPairedAddLower64, U128, U128, U128, U1 ) OPCODE(FPVectorPairedAddLower64, U128, U128, U128, U1 )
OPCODE(FPVectorRecipEstimate16, U128, U128 ) OPCODE(FPVectorRecipEstimate16, U128, U128, U1 )
OPCODE(FPVectorRecipEstimate32, U128, U128 ) OPCODE(FPVectorRecipEstimate32, U128, U128, U1 )
OPCODE(FPVectorRecipEstimate64, U128, U128 ) OPCODE(FPVectorRecipEstimate64, U128, U128, U1 )
OPCODE(FPVectorRecipStepFused16, U128, U128, U128, U1 ) OPCODE(FPVectorRecipStepFused16, U128, U128, U128, U1 )
OPCODE(FPVectorRecipStepFused32, U128, U128, U128, U1 ) OPCODE(FPVectorRecipStepFused32, U128, U128, U128, U1 )
OPCODE(FPVectorRecipStepFused64, U128, U128, U128, U1 ) OPCODE(FPVectorRecipStepFused64, U128, U128, U128, U1 )

2
tests/A32/fuzz_arm.cpp
View file

@ -111,7 +111,7 @@ u32 GenRandomInst(u32 pc, bool is_last_inst) {
// FPSCR is inaccurate // FPSCR is inaccurate
"vfp_VMRS", "vfp_VMRS",
// Unimplemented in Unicorn // Unimplemented in Unicorn
"asimd_VPADD_float", "asimd_VPADD_float", "asimd_VRECPE",
// Incorrect Unicorn implementations // Incorrect Unicorn implementations
"asimd_VRECPS", // Unicorn does not fuse the multiply and subtraction, resulting in being off by 1ULP. "asimd_VRECPS", // Unicorn does not fuse the multiply and subtraction, resulting in being off by 1ULP.
}; };