[LVI] Fix insertelement of constexpr

Bail out when evaluating an insertelement of a constant expression.
Unlike other ValueLattice kinds, these don't have implicit splat
semantics and we end up with type mismatches. If we actually wanted
to handle these, we should actually evaluate the insertion via
constant folding. I'm not bothering with that, as these should
get constant folded on construction already.

Author: nikic

llvm/lib/Analysis/LazyValueInfo.cpp

@@ -1005,6 +1005,13 @@ LazyValueInfoImpl::solveBlockValueInsertElement(InsertElementInst *IEI,
   if (!OptVecVal)
     return std::nullopt;
 
+  // Bail out if the inserted element is a constant expression. Unlike other
+  // ValueLattice types, these are not considered an implicit splat when a
+  // vector type is used.
+  // We could call ConstantFoldInsertElementInstruction here to handle these.
+  if (OptEltVal->isConstant())
+    return ValueLatticeElement::getOverdefined();
+
   Res.mergeIn(*OptVecVal);
   return Res;
 }

llvm/test/Transforms/CorrelatedValuePropagation/vectors.ll

@@ -352,3 +352,14 @@ define <2 x i1> @insertelement_fold2() {
   %icmp1 = icmp slt <2 x i32> %ie1, <i32 1024, i32 1024>
   ret <2 x i1> %icmp1
 }
+
+@g = external global i32
+
+define <2 x i16> @insertelement_constexpr() {
+; CHECK-LABEL: define <2 x i16> @insertelement_constexpr() {
+; CHECK-NEXT:    [[INS:%.*]] = insertelement <2 x i16> poison, i16 ptrtoint (ptr @g to i16), i32 0
+; CHECK-NEXT:    ret <2 x i16> [[INS]]
+;
+  %ins = insertelement <2 x i16> poison, i16 ptrtoint (ptr @g to i16), i32 0
+  ret <2 x i16> %ins
+}