Suyu/dynarmic
1
0
Fork 0
mirror of https://git.suyu.dev/suyu/dynarmic.git synced 2026-01-02 04:34:43 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions

A64: Implement FRSQRTS (vector), single/double variant

Browse source
This commit is contained in:
MerryMage 2018-07-23 22:58:52 +01:00
parent 45dc5f74f3
commit b2e4c16ef8
7 changed files with 123 additions and 34 deletions
Download patch file Download diff file Expand all files Collapse all files

61
src/backend_x64/emit_x64_vector_floating_point.cpp
View file

@ -250,6 +250,50 @@ inline void EmitOneArgumentFallback(BlockOfCode& code, EmitContext& ctx, IR::Ins
ctx.reg_alloc.DefineValue(inst, xmm0);
}
template <typename Lambda>
inline void EmitTwoArgumentFallback(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, Lambda lambda) {
const auto fn = static_cast<mp::equivalent_function_type_t<Lambda>*>(lambda);
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const Xbyak::Xmm arg1 = ctx.reg_alloc.UseXmm(args[0]);
const Xbyak::Xmm arg2 = ctx.reg_alloc.UseXmm(args[1]);
ctx.reg_alloc.EndOfAllocScope();
ctx.reg_alloc.HostCall(nullptr);
#ifdef _WIN32
constexpr u32 stack_space = 4 * 16;
code.sub(rsp, stack_space + ABI_SHADOW_SPACE);
code.lea(code.ABI_PARAM1, ptr[rsp + ABI_SHADOW_SPACE + 1 * 16]);
code.lea(code.ABI_PARAM2, ptr[rsp + ABI_SHADOW_SPACE + 2 * 16]);
code.lea(code.ABI_PARAM3, ptr[rsp + ABI_SHADOW_SPACE + 3 * 16]);
code.mov(code.ABI_PARAM4.cvt32(), ctx.FPCR());
code.lea(rax, code.ptr[code.r15 + code.GetJitStateInfo().offsetof_fpsr_exc]);
code.mov(qword[rsp + ABI_SHADOW_SPACE + 0 * 16], rax);
#else
constexpr u32 stack_space = 3 * 16;
code.sub(rsp, stack_space + ABI_SHADOW_SPACE);
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_PARAM3, ptr[rsp + ABI_SHADOW_SPACE + 2 * 16]);
code.mov(code.ABI_PARAM4.cvt32(), ctx.FPCR());
code.lea(code.ABI_PARAM5, code.ptr[code.r15 + code.GetJitStateInfo().offsetof_fpsr_exc]);
#endif
code.movaps(xword[code.ABI_PARAM2], arg1);
code.movaps(xword[code.ABI_PARAM3], arg2);
code.CallFunction(fn);
#ifdef _WIN32
code.movaps(xmm0, xword[rsp + ABI_SHADOW_SPACE + 1 * 16]);
#else
code.movaps(xmm0, xword[rsp + ABI_SHADOW_SPACE + 0 * 16]);
#endif
code.add(rsp, stack_space + ABI_SHADOW_SPACE);
ctx.reg_alloc.DefineValue(inst, xmm0);
}
void EmitX64::EmitFPVectorAbs16(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
@ -410,6 +454,23 @@ void EmitX64::EmitFPVectorRSqrtEstimate64(EmitContext& ctx, IR::Inst* inst) {
EmitRSqrtEstimate<u64>(code, ctx, inst);
}
template<typename FPT>
static void EmitRSqrtStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
EmitTwoArgumentFallback(code, ctx, inst, [](VectorArray<FPT>& result, const VectorArray<FPT>& op1, const VectorArray<FPT>& op2, FP::FPCR fpcr, FP::FPSR& fpsr) {
for (size_t i = 0; i < result.size(); i++) {
result[i] = FP::FPRSqrtStepFused<FPT>(op1[i], op2[i], fpcr, fpsr);
}
});
}
void EmitX64::EmitFPVectorRSqrtStepFused32(EmitContext& ctx, IR::Inst* inst) {
EmitRSqrtStepFused<u32>(code, ctx, inst);
}
void EmitX64::EmitFPVectorRSqrtStepFused64(EmitContext& ctx, IR::Inst* inst) {
EmitRSqrtStepFused<u64>(code, ctx, inst);
}
void EmitX64::EmitFPVectorS32ToSingle(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const Xbyak::Xmm xmm = ctx.reg_alloc.UseScratchXmm(args[0]);