I'm thinking the layer framework needs to allocate the layer
transformer instance in a thread-local manner so that if it
constructs mutable state that will be thread local also
One downside to this is layer transformers that are thread-safe
unnecessarily incur some amount of overhead, both in
In the CRC32 layer, the "optimization" was really just laziness. The use
case we have only needs constant known explicit lengths, so we wanted to
hard wire that, and not bother to support nor test any other cases.
But, had that worked, I would bet we would have made the mistake and
pre-allocated
I don't remember why NoCacheEvluatable was used, but you're I suspect
you're right that it was done so implementations are less likely to mess
things up. Note that NoCacheEvaluatable is also used for
ImplicitLengthEv. I'm not sure why it's used there either. Maybe
LayerLengthEv just copied
