steam-frame/FEX-Emu--FEX
1
0 Fork 0
mirror of https://github.com/FEX-Emu/FEX.git synced 2026-10-06 17:00:19 +02:00
Code Issues Packages Projects Releases 66 Wiki Activity

VectorOps: Remove unnecessary mov in VUShrNI2/VSQXTN2/VSQXTUN2

Browse Source 
We can move the initial move down by SPLICE, which not only lets us turn
it into a MOVPRFX, but also we can safely move into the final
destination register directly, since we can be sure there's no
potential dependencies at this point
This commit is contained in:
Lioncache committed 2023-02-15 17:14:26 -05:00
1 parent 81a89ab747
commit e78b573610
1 file changed
+19 -24
Download Patch File Download Diff File Expand all files Collapse all files
External/FEXCore/Source/Interface/Core/JIT/Arm64/VectorOps.cpp
Vendored
+19 -24
View File
@@ -2468,14 +2468,13 @@ DEF_OP(VUShrNI2) {
ElementSize == 4 ? ARMEmitter::SubRegSize::i32Bit : ARMEmitter::SubRegSize::i8Bit;
if (HostSupportsSVE && Is256Bit) {
mov(VTMP1.Z(), VectorLower.Z());
const auto Mask = PRED_TMP_16B;
shrnb(SubRegSize, VTMP2.Z(), VectorUpper.Z(), BitShift);
uzp1(SubRegSize, VTMP2.Z(), VTMP2.Z(), VTMP2.Z());
splice<ARMEmitter::OpType::Destructive>(SubRegSize, VTMP1.Z(), Mask, VTMP1.Z(), VTMP2.Z());
mov(Dst.Z(), VTMP1.Z());
movprfx(Dst.Z(), VectorLower.Z());
splice<ARMEmitter::OpType::Destructive>(SubRegSize, Dst.Z(), Mask, Dst.Z(), VTMP2.Z());
} else {
mov(VTMP1.Q(), VectorLower.Q());
shrn2(SubRegSize, VTMP1.Q(), VectorUpper.Q(), BitShift);
@@ -2651,12 +2650,6 @@ DEF_OP(VSQXTN2) {
ElementSize == 4 ? ARMEmitter::SubRegSize::i32Bit : ARMEmitter::SubRegSize::i8Bit;
if (HostSupportsSVE && Is256Bit) {
// Need to use the destructive variant of SPLICE, since
// the constructive variant requires a register list, and
// we can't guarantee VectorLower and VectorUpper will always
// have consecutive indexes with one another.
mov(VTMP1.Z(), VectorLower.Z());
// We use the 16 byte mask due to how SPLICE works. We only
// want to get at the first 16 bytes in the lower vector, so
// that SPLICE will then begin copying the first 16 bytes
@@ -2666,20 +2659,23 @@ DEF_OP(VSQXTN2) {
sqxtnb(SubRegSize, VTMP2.Z(), VectorUpper.Z());
uzp1(SubRegSize, VTMP2.Z(), VTMP2.Z(), VTMP2.Z());
splice<ARMEmitter::OpType::Destructive>(SubRegSize, VTMP1.Z(), Mask, VTMP1.Z(), VTMP2.Z());
mov(Dst.Z(), VTMP1.Z());
// Need to use the destructive variant of SPLICE, since
// the constructive variant requires a register list, and
// we can't guarantee VectorLower and VectorUpper will always
// have consecutive indexes with one another.
movprfx(Dst.Z(), VectorLower.Z());
splice<ARMEmitter::OpType::Destructive>(SubRegSize, Dst.Z(), Mask, Dst.Z(), VTMP2.Z());
} else {
mov(VTMP1.Q(), VectorLower.Q());
if (OpSize == 8) {
sqxtn(SubRegSize, VTMP2, VectorUpper);
ins(ARMEmitter::SubRegSize::i32Bit, VTMP1, 1, VTMP2, 0);
mov(Dst.Q(), VectorLower.Q());
ins(ARMEmitter::SubRegSize::i32Bit, Dst, 1, VTMP2, 0);
} else {
mov(VTMP1.Q(), VectorLower.Q());
sqxtn2(SubRegSize, VTMP1, VectorUpper);
mov(Dst.Q(), VTMP1.Q());
}
mov(Dst.Q(), VTMP1.Q());
}
}
@@ -2728,24 +2724,23 @@ DEF_OP(VSQXTUN2) {
// NOTE: See VSQXTN2 implementation for an in-depth explanation
// of everything going on here.
mov(VTMP1.Z(), VectorLower.Z());
const auto Mask = PRED_TMP_16B;
sqxtunb(SubRegSize, VTMP2.Z(), VectorUpper.Z());
uzp1(SubRegSize, VTMP2.Z(), VTMP2.Z(), VTMP2.Z());
splice<ARMEmitter::OpType::Destructive>(SubRegSize, VTMP1.Z(), Mask, VTMP1.Z(), VTMP2.Z());
mov(Dst.Z(), VTMP1.Z());
movprfx(Dst.Z(), VectorLower.Z());
splice<ARMEmitter::OpType::Destructive>(SubRegSize, Dst.Z(), Mask, Dst.Z(), VTMP2.Z());
} else {
mov(VTMP1.Q(), VectorLower.Q());
if (OpSize == 8) {
sqxtun(SubRegSize, VTMP2, VectorUpper);
ins(ARMEmitter::SubRegSize::i32Bit, VTMP1, 1, VTMP2, 0);
mov(Dst.Q(), VectorLower.Q());
ins(ARMEmitter::SubRegSize::i32Bit, Dst, 1, VTMP2, 0);
} else {
mov(VTMP1.Q(), VectorLower.Q());
sqxtun2(SubRegSize, VTMP1, VectorUpper);
mov(Dst.Q(), VTMP1.Q());
}
mov(Dst.Q(), VTMP1.Q());
}
}