[llvm] [ConstantFolding] Fold intrinsics of scalable vectors with splatted operands (PR #141845)

[Luke Lau via llvm-commits]

================
@@ -3780,7 +3780,35 @@ static Constant *ConstantFoldScalableVectorCall(
   default:
     break;
   }
-  return nullptr;
+
+  // If trivially vectorizable, try folding it via the scalar call if all
+  // operands are splats.
+
+  // TODO: ConstantFoldFixedVectorCall should probably check this too?
+  if (!isTriviallyVectorizable(IntrinsicID))
+    return nullptr;
+
+  SmallVector<Constant *, 4> SplatOps;
+  for (auto [I, Op] : enumerate(Operands)) {
+    if (isVectorIntrinsicWithScalarOpAtArg(IntrinsicID, I, /*TTI=*/nullptr)) {
+      SplatOps.push_back(Op);
+      continue;
+    }
+    // TODO: Should getSplatValue return a poison scalar for a poison vector?
----------------
lukel97 wrote:

Will do this in a follow up, seems to be NFC?

https://github.com/llvm/llvm-project/pull/141845