[RISCV] Disable combineBinOpToReduce if the reduction AVL might be 0.

[craig.topper]

Summary

If the reduction AVL is 0, operand 0 of the reduction will be
returned rather than the scalar input.

To make the fold legal, we would need to fold the new scalar value
with whatever operand 0 is which may require a new scalar operation
before the reduction.

Block the combine if we can't prove AVL is non-zero.

[fakepaper56]

LGTM.

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

static SDValue combineBinOpToReduce(SDNode *N, SelectionDAG &DAG,
   // TODO: Deal with value other than neutral element.
   if (!isNeutralConstant(N->getOpcode(), N->getFlags(), ScalarV.getOperand(1),
                          0))
     return SDValue();
 
   if (!ScalarV.hasOneUse())
     return SDValue();
 
+  // If the AVL is zero, operand 0 will be returned. So it's not safe to fold.
+  // FIXME: We might be able to improve this if operand 0 is undef.
+  if (!isNonZeroAVL(Reduce.getOperand(5)))
+    return SDValue();
+
   SDValue NewStart = N->getOperand(1 - ReduceIdx);
 
   SDLoc DL(N);
   SDValue NewScalarV =
       lowerScalarInsert(NewStart, ScalarV.getOperand(2),
                         ScalarV.getSimpleValueType(), DL, DAG, Subtarget);
 
   // If we looked through an INSERT_SUBVECTOR we need to restore it.