[llvm] [VPlan] Fold FirstOrderRecurrenceSplice(X, X) to X when X is uniform (PR #171772)

[Paul Walker via llvm-commits]

================
@@ -1449,6 +1449,14 @@ static void simplifyRecipe(VPSingleDefRecipe *Def, VPTypeAnalysis &TypeInfo) {
     return;
   }
 
+  // Replace splice(A, A, -1) with A if A is uniform.
+  if (match(Def, m_VPInstruction<VPInstruction::FirstOrderRecurrenceSplice>(
+                     m_VPValue(A), m_Deferred(A))) &&
+      vputils::isSingleScalar(A)) {
----------------
paulwalker-arm wrote:

I'm not familiar with VPlan corner cases so I apologise if this is nonsense but vputils::isSingleScalar says "because it produces the same value for all lanes or only has its first lane used".

For the second part of this classification when using vector operands (say <2 x i32>) we kind of have a bogus splice because the result of splice(A, A, -1) could be represented as `<poison, A[0]>` because `isSingleScalar` says we should only be using the first lane? Returning `splat(A[0])` here seems fine, but returning `A` seems wrong because then we're now saying Result[0] is valid but Result[1] is not, which is the opposite of what the splice returned (i.e. Result[1] is valid, but Result[0] is not).

https://github.com/llvm/llvm-project/pull/171772