[llvm] [RISCV][CG]Use processShuffleMasks for per-register shuffles (PR #120803)

[Philip Reames via llvm-commits]

================
@@ -5120,58 +5119,70 @@ static SDValue lowerShuffleViaVRegSplitting(ShuffleVectorSDNode *SVN,
 
   MVT ElemVT = VT.getVectorElementType();
   unsigned ElemsPerVReg = *VLen / ElemVT.getFixedSizeInBits();
-  unsigned VRegsPerSrc = NumElts / ElemsPerVReg;
-
-  SmallVector<std::pair<int, SmallVector<int>>>
-    OutMasks(VRegsPerSrc, {-1, {}});
-
-  // Check if our mask can be done as a 1-to-1 mapping from source
-  // to destination registers in the group without needing to
-  // write each destination more than once.
-  for (unsigned DstIdx = 0; DstIdx < Mask.size(); DstIdx++) {
-    int DstVecIdx = DstIdx / ElemsPerVReg;
-    int DstSubIdx = DstIdx % ElemsPerVReg;
-    int SrcIdx = Mask[DstIdx];
-    if (SrcIdx < 0 || (unsigned)SrcIdx >= 2 * NumElts)
-      continue;
-    int SrcVecIdx = SrcIdx / ElemsPerVReg;
-    int SrcSubIdx = SrcIdx % ElemsPerVReg;
-    if (OutMasks[DstVecIdx].first == -1)
-      OutMasks[DstVecIdx].first = SrcVecIdx;
-    if (OutMasks[DstVecIdx].first != SrcVecIdx)
-      // Note: This case could easily be handled by keeping track of a chain
----------------
preames wrote:

Possibly, yes.  In particular, using a quadratic number of shuffles (with distinct indices and masks) vs a single one is a *lot* of code size increase.  

I suspect this needs a bit of thought and investigation.  I am not proposing any particular heuristic, and am open to being convinced that the right heuristic is to just blindly expand.  I'd just like to see it explored and justified.  

Though 1-2 shuffles is way to low a threshold.  You definitely want something which allows at least the linear expansion of the code you replaced.

https://github.com/llvm/llvm-project/pull/120803