[llvm] [RISCV] Extract spread(2,4,8) shuffle lowering from interleave(2) (PR #118822)

[Craig Topper via llvm-commits]

================
@@ -4816,12 +4816,46 @@ static SDValue lowerVECTOR_SHUFFLEAsVSlide1(const SDLoc &DL, MVT VT,
   return convertFromScalableVector(VT, Vec, DAG, Subtarget);
 }
 
+// Given a vector a, b, c, d return a vector Factor times longer
+// with Factor-1 undef's between elements. Ex:
+//   a, undef, b, undef, c, undef, d, undef (Factor=2, Index=0)
+//   undef, a, undef, b, undef, c, undef, d (Factor=2, Index=1)
+static SDValue getWideningSpread(SDValue V, unsigned Factor, unsigned Index,
+                                 const SDLoc &DL, SelectionDAG &DAG) {
+
+  MVT VT = V.getSimpleValueType();
+  unsigned EltBits = VT.getScalarSizeInBits();
+  ElementCount EC = VT.getVectorElementCount();
+  V = DAG.getBitcast(VT.changeTypeToInteger(), V);
+
+  MVT WideVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Factor), EC);
+
+  SDValue Result = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT, V);
+  // TODO: On rv32, the constant becomes a splat_vector_parts which does not
+  // allow the SHL to fold away if Index is 0.
+  if (Index != 0)
+    Result = DAG.getNode(ISD::SHL, DL, WideVT, Result,
+                         DAG.getConstant(EltBits * Index, DL, WideVT));
+  // Make sure to use original element type
+  MVT ResultVT = MVT::getVectorVT(VT.getVectorElementType(),
+                                  EC.multiplyCoefficientBy(Factor));
+  return DAG.getBitcast(ResultVT, Result);
+}
+
 // Given two input vectors of <[vscale x ]n x ty>, use vwaddu.vv and vwmaccu.vx
 // to create an interleaved vector of <[vscale x] n*2 x ty>.
 // This requires that the size of ty is less than the subtarget's maximum ELEN.
 static SDValue getWideningInterleave(SDValue EvenV, SDValue OddV,
                                      const SDLoc &DL, SelectionDAG &DAG,
                                      const RISCVSubtarget &Subtarget) {
+
+  // FIXME: Not only does this optimize the code, it fixes some correctness
+  // issues because MIR does not have freeze.
+  if (EvenV.isUndef())
+    return getWideningSpread(OddV, 2, 1, DL, DAG);
+  else if (OddV.isUndef())
----------------
topperc wrote:

Drop else

https://github.com/llvm/llvm-project/pull/118822