Skip to content

[LoongArch] Use xvperm.w for cross-lane access within a single vector #151634

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Sep 4, 2025
Merged

[LoongArch] Use xvperm.w for cross-lane access within a single vector #151634

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
aa379f6
[LoongArch] Use xvperm.w for cross-lane access within a single vector
zhaoqi5 Aug 1, 2025
b9b2cc4
opt code style
zhaoqi5 Aug 1, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
44 changes: 44 additions & 0 deletions llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1990,6 +1990,48 @@ lowerVECTOR_SHUFFLE_XVSHUF4I(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
return lowerVECTOR_SHUFFLE_VSHUF4I(DL, Mask, VT, V1, V2, DAG, Subtarget);
}

/// Lower VECTOR_SHUFFLE into XVPERM (if possible).
static SDValue lowerVECTOR_SHUFFLE_XVPERM(const SDLoc &DL, ArrayRef<int> Mask,
MVT VT, SDValue V1, SDValue V2,
SelectionDAG &DAG) {
// LoongArch LASX only have XVPERM_W.
if (Mask.size() != 8 || (VT != MVT::v8i32 && VT != MVT::v8f32))
return SDValue();

unsigned NumElts = VT.getVectorNumElements();
unsigned HalfSize = NumElts / 2;
bool FrontLo = true, FrontHi = true;
bool BackLo = true, BackHi = true;

auto inRange = [](int val, int low, int high) {
return (val == -1) || (val >= low && val < high);
};

for (unsigned i = 0; i < HalfSize; ++i) {
int Fronti = Mask[i];
int Backi = Mask[i + HalfSize];

FrontLo &= inRange(Fronti, 0, HalfSize);
FrontHi &= inRange(Fronti, HalfSize, NumElts);
BackLo &= inRange(Backi, 0, HalfSize);
BackHi &= inRange(Backi, HalfSize, NumElts);
}

// If both the lower and upper 128-bit parts access only one half of the
// vector (either lower or upper), avoid using xvperm.w. The latency of
// xvperm.w(3) is higher than using xvshuf(1) and xvori(1).
Copy link
Member

@heiher heiher Sep 2, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

For a shuffle that swaps the upper and lower 128-bit halves, xvperm.w alone should be enough and likely faster.

define <8 x i32> @shuffle_v8i32(<8 x i32> %a) {
  %shuffle = shufflevector <8 x i32> %a, <8 x i32> poison, <8 x i32> <i32 5, i32 4, i32 6, i32 7, i32 3, i32 2, i32 0, i32 1>
  ret <8 x i32> %shuffle
}

Copy link
Contributor Author

@zhaoqi5 zhaoqi5 Sep 2, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, xvperm.w is actually enough for this case.

But when legalizing vector_shuffle, canonicalizeShuffleVectorByLane() will firstly generate xvpermi.d to avoid cross-lane access (similar to discussion in #151633 (comment)).
When comming to here, the masks has already changed to be 1,0,2,3,3,2,1,0, the already generated xvpermi.d is unable to be removed anyway.
I haven't come up with a good solution about this yet.

However, based solely on the masks obtained here, it is necessary to avoid conversion here.

Copy link
Member

@heiher heiher Sep 2, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If it’s not something that can be done quickly, feel free to handle it in a separate PR. LGTM

Copy link
Contributor Author

@zhaoqi5 zhaoqi5 Sep 2, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

OK. Thanks for your review.

if ((FrontLo || FrontHi) && (BackLo || BackHi))
return SDValue();

SmallVector<SDValue, 8> Masks;
for (unsigned i = 0; i < NumElts; ++i)
Masks.push_back(Mask[i] == -1 ? DAG.getUNDEF(MVT::i64)
: DAG.getConstant(Mask[i], DL, MVT::i64));
SDValue MaskVec = DAG.getBuildVector(MVT::v8i32, DL, Masks);

return DAG.getNode(LoongArchISD::XVPERM, DL, VT, V1, MaskVec);
}

/// Lower VECTOR_SHUFFLE into XVPACKEV (if possible).
static SDValue lowerVECTOR_SHUFFLE_XVPACKEV(const SDLoc &DL, ArrayRef<int> Mask,
MVT VT, SDValue V1, SDValue V2,
Expand Down Expand Up @@ -2396,6 +2438,8 @@ static SDValue lower256BitShuffle(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
if ((Result = lowerVECTOR_SHUFFLE_XVSHUF4I(DL, NewMask, VT, V1, V2, DAG,
Subtarget)))
return Result;
if ((Result = lowerVECTOR_SHUFFLE_XVPERM(DL, NewMask, VT, V1, V2, DAG)))
return Result;
if ((Result = lowerVECTOR_SHUFFLEAsLanePermuteAndShuffle(DL, NewMask, VT,
V1, V2, DAG)))
return Result;
Expand Down
18 changes: 4 additions & 14 deletions llvm/test/CodeGen/LoongArch/lasx/shuffle-as-permute-and-shuffle.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,13 +61,8 @@ define <8 x i32> @shuffle_v8i32(<8 x i32> %a) {
; CHECK-LABEL: shuffle_v8i32:
; CHECK: # %bb.0:
; CHECK-NEXT: pcalau12i $a0, %pc_hi20(.LCPI4_0)
; CHECK-NEXT: xvld $xr2, $a0, %pc_lo12(.LCPI4_0)
; CHECK-NEXT: pcalau12i $a0, %pc_hi20(.LCPI4_1)
; CHECK-NEXT: xvld $xr1, $a0, %pc_lo12(.LCPI4_1)
; CHECK-NEXT: xvpermi.d $xr3, $xr0, 78
; CHECK-NEXT: xvshuf.d $xr2, $xr0, $xr3
; CHECK-NEXT: xvshuf.d $xr1, $xr2, $xr0
; CHECK-NEXT: xvori.b $xr0, $xr1, 0
; CHECK-NEXT: xvld $xr1, $a0, %pc_lo12(.LCPI4_0)
; CHECK-NEXT: xvperm.w $xr0, $xr0, $xr1
; CHECK-NEXT: ret
%shuffle = shufflevector <8 x i32> %a, <8 x i32> poison, <8 x i32> <i32 4, i32 5, i32 0, i32 1, i32 4, i32 5, i32 6, i32 7>
ret <8 x i32> %shuffle
Expand Down Expand Up @@ -117,13 +112,8 @@ define <8 x float> @shuffle_v8f32(<8 x float> %a) {
; CHECK-LABEL: shuffle_v8f32:
; CHECK: # %bb.0:
; CHECK-NEXT: pcalau12i $a0, %pc_hi20(.LCPI8_0)
; CHECK-NEXT: xvld $xr2, $a0, %pc_lo12(.LCPI8_0)
; CHECK-NEXT: pcalau12i $a0, %pc_hi20(.LCPI8_1)
; CHECK-NEXT: xvld $xr1, $a0, %pc_lo12(.LCPI8_1)
; CHECK-NEXT: xvpermi.d $xr3, $xr0, 78
; CHECK-NEXT: xvshuf.d $xr2, $xr0, $xr3
; CHECK-NEXT: xvshuf.d $xr1, $xr2, $xr0
; CHECK-NEXT: xvori.b $xr0, $xr1, 0
; CHECK-NEXT: xvld $xr1, $a0, %pc_lo12(.LCPI8_0)
; CHECK-NEXT: xvperm.w $xr0, $xr0, $xr1
; CHECK-NEXT: ret
%shuffle = shufflevector <8 x float> %a, <8 x float> poison, <8 x i32> <i32 4, i32 5, i32 0, i32 1, i32 4, i32 5, i32 6, i32 7>
ret <8 x float> %shuffle
Expand Down