ShuriZma
/
xenia-canary
mirror of https://github.com/xenia-canary/xenia-canary.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Emulate some vector arithmetic opcodes (until we get a native implementation)

This commit is contained in:
Dr. Chat 2015-05-05 15:34:13 -05:00
parent 87811bbf2b
commit cb127ae9ba
1 changed files with 75 additions and 34 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
src/xenia/cpu/backend/x64/x64_sequences.cc
View File

@ -4726,7 +4726,18 @@ EMITTER(VECTOR_SHL_V128, MATCH(I<OPCODE_VECTOR_SHL, V128<>, V128<>, V128<>>)) {
e.CallNativeSafe(reinterpret_cast<void*>(EmulateVectorShlI16));
e.vmovaps(i.dest, e.xmm0);
}
static __m128i EmulateVectorShlI32(void*, __m128i src1, __m128i src2) {
alignas(16) uint32_t value[4];
alignas(16) uint32_t shamt[4];
_mm_store_si128(reinterpret_cast<__m128i*>(value), src1);
_mm_store_si128(reinterpret_cast<__m128i*>(shamt), src2);
for (size_t i = 0; i < 4; ++i) {
value[i] = value[i] << (shamt[i] & 0x1F);
}
return _mm_load_si128(reinterpret_cast<__m128i*>(value));
}
static void EmitInt32(X64Emitter& e, const EmitArgType& i) {
if (e.cpu()->has(Xbyak::util::Cpu::tAVX2)) {
if (i.src2.is_constant) {
const auto& shamt = i.src2.constant();
bool all_same = true;
@ -4755,6 +4766,18 @@ EMITTER(VECTOR_SHL_V128, MATCH(I<OPCODE_VECTOR_SHL, V128<>, V128<>, V128<>>)) {
e.vandps(e.xmm0, i.src2, e.GetXmmConstPtr(XMMShiftMaskPS));
e.vpsllvd(i.dest, i.src1, e.xmm0);
}
} else {
// TODO(benvanik): native version (with shift magic).
if (i.src2.is_constant) {
e.LoadConstantXmm(e.xmm0, i.src2.constant());
e.lea(e.r9, e.StashXmm(1, e.xmm0));
} else {
e.lea(e.r9, e.StashXmm(1, i.src2));
}
e.lea(e.r8, e.StashXmm(0, i.src1));
e.CallNativeSafe(reinterpret_cast<void*>(EmulateVectorShlI32));
e.vmovaps(i.dest, e.xmm0);
}
}
};
EMITTER_OPCODE_TABLE(
@ -5058,6 +5081,16 @@ EMITTER(VECTOR_ROTATE_LEFT_V128, MATCH(I<OPCODE_VECTOR_ROTATE_LEFT, V128<>, V128
}
return _mm_load_si128(reinterpret_cast<__m128i*>(value));
}
static __m128i EmulateVectorRotateLeftI32(void*, __m128i src1, __m128i src2) {
alignas(16) uint32_t value[4];
alignas(16) uint32_t shamt[4];
_mm_store_si128(reinterpret_cast<__m128i*>(value), src1);
_mm_store_si128(reinterpret_cast<__m128i*>(shamt), src2);
for (size_t i = 0; i < 4; ++i) {
value[i] = xe::rotate_left<uint32_t>(value[i], shamt[i] & 0x1F);
}
return _mm_load_si128(reinterpret_cast<__m128i*>(value));
}
static void Emit(X64Emitter& e, const EmitArgType& i) {
switch (i.instr->flags) {
case INT8_TYPE:
@ -5075,6 +5108,7 @@ EMITTER(VECTOR_ROTATE_LEFT_V128, MATCH(I<OPCODE_VECTOR_ROTATE_LEFT, V128<>, V128
e.vmovaps(i.dest, e.xmm0);
break;
case INT32_TYPE: {
if (e.cpu()->has(Xbyak::util::Cpu::tAVX2)) {
Xmm temp = i.dest;
if (i.dest == i.src1 || i.dest == i.src2) {
temp = e.xmm2;
@ -5088,6 +5122,13 @@ EMITTER(VECTOR_ROTATE_LEFT_V128, MATCH(I<OPCODE_VECTOR_ROTATE_LEFT, V128<>, V128
e.vpsrlvd(i.dest, i.src1, temp);
// Merge:
e.vpor(i.dest, e.xmm1);
} else {
// TODO: Non-AVX2 native version
e.lea(e.r8, e.StashXmm(0, i.src1));
e.lea(e.r9, e.StashXmm(1, i.src2));
e.CallNativeSafe(reinterpret_cast<void*>(EmulateVectorRotateLeftI32));
e.vmovaps(i.dest, e.xmm0);
}
break;
}
default: