[llvm] [RISCV] Merge shuffle sources if lanes are disjoint (PR #119401)

[Philip Reames via llvm-commits]

================
@@ -5197,6 +5197,67 @@ static bool isCompressMask(ArrayRef<int> Mask) {
   return true;
 }
 
+/// Given a shuffle where the indices are disjoint between the two sources,
+/// e.g.:
+///
+/// t2:v4i8 = vector_shuffle t0:v4i8, t1:v4i8, <2, 7, 1, 4>
+///
+/// Merge the two sources into one and do a single source shuffle:
+///
+/// t2:v4i8 = vselect t1:v4i8, t0:v4i8, <0, 1, 0, 1>
+/// t3:v4i8 = vector_shuffle t2:v4i8, undef, <2, 3, 1, 0>
+///
+/// A vselect will either be merged into a masked instruction or be lowered as a
+/// vmerge.vvm, which is cheaper than a vrgather.vv.
+static SDValue lowerDisjointIndicesShuffle(ShuffleVectorSDNode *SVN,
+                                           SelectionDAG &DAG,
+                                           const RISCVSubtarget &Subtarget) {
+  MVT VT = SVN->getSimpleValueType(0);
+  MVT XLenVT = Subtarget.getXLenVT();
+  SDLoc DL(SVN);
+
+  const ArrayRef<int> Mask = SVN->getMask();
+
+  // Work out which source each lane will come from.
+  SmallVector<int, 16> Srcs(Mask.size(), -1);
+
+  for (int Idx : Mask) {
+    if (Idx == -1)
+      continue;
+    unsigned SrcIdx = Idx % Mask.size();
+    int Src = (uint32_t)Idx < Mask.size() ? 0 : 1;
+    if (Srcs[SrcIdx] == -1)
+      // Mark this source as using this lane.
+      Srcs[SrcIdx] = Src;
+    else if (Srcs[SrcIdx] != Src)
+      // The other source is using this lane: not disjoint.
+      return SDValue();
+  }
+
+  SmallVector<SDValue> SelectMaskVals;
+  for (int Lane : Srcs) {
----------------
preames wrote:

Can you reverse the select here so that it uses the same order as the generic fallthrough below?  From prior experience, our vmerge.vxm matching is oddly fragile.  I'd like to remove this unrelated change if possible.

https://github.com/llvm/llvm-project/pull/119401