[llvm-branch-commits] [llvm] [LoongArch] Use xvperm.w for cross-lane access within a single vector (PR #151634)

[via llvm-branch-commits]

================
@@ -1832,6 +1832,48 @@ static SDValue lowerVECTOR_SHUFFLE_XVSHUF4I(const SDLoc &DL, ArrayRef<int> Mask,
   return lowerVECTOR_SHUFFLE_VSHUF4I(DL, Mask, VT, V1, V2, DAG);
 }
 
+/// 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).
----------------
heiher wrote:

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

```llvm
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
}
```

https://github.com/llvm/llvm-project/pull/151634