As we already model external (escaping uses) as explicit memory accesses and
the block builder generates scalar stores for all memory accesses, all scalar
values inside a basic block should already be initialized when generating
escape values. Hence, isNew will always be false and the corresponding code
path is never executed. This change passes both 'make check-polly' as well as
LNT.
Johannes, do you remember why you introduced this code path? Did I miss
something? If so, could you provide a test case that would fail with this
patch applied?