[clang] [Clang] VectorExprEvaluator::VisitCallExpr / InterpretBuiltin - Allow shufps/pd shuffles intrinsics to be used in constexpr (PR #164078)

[Simon Pilgrim via cfe-commits]

================
@@ -3128,6 +3128,52 @@ static bool interp__builtin_ia32_vpconflict(InterpState &S, CodePtr OpPC,
   return true;
 }
 
+static bool interp__builtin_ia32_shuf(InterpState &S, CodePtr OpPC,
+                                      const CallExpr *Call) {
+  assert(Call->getNumArgs() == 3);
+
+  unsigned ShuffleMask = popToAPSInt(S, Call->getArg(2)).getZExtValue();
+  QualType Arg0Type = Call->getArg(0)->getType();
+  const auto *VecT = Arg0Type->castAs<VectorType>();
+  PrimType ElemT = *S.getContext().classify(VecT->getElementType());
+  unsigned NumElems = VecT->getNumElements();
+  unsigned LaneWidth = S.getContext().getBitWidth(VecT->getElementType());
----------------
RKSimon wrote:

LaneWidth -> ElemWidth (Lanes in x86 mean 128-bit subvectors)

https://github.com/llvm/llvm-project/pull/164078