Remove Pod bounds on traits

[tamird]

Because Pod has a 'static bound, its present on many traits means it is
impossible to implement those traits on any type that has a non-static
lifetime. Removing the Pod bounds requires granular bounds elsewhere, which are
added here.

[bjorn3]

An 'static bound on a type means that values of this type can life forever. Not that they actually do. In practice this means that any references contained inside this type are 'static. Pod types can't contain references so the 'static bound doesn't restrict anything.

[bjorn3]

To elaborate on why Pod types can't contain references, the safety condition of Pod requires that "A type that is Pod must have no invalid byte values". References have many invalid byte values, null pointers being the simplest of them. Enums also violate this safety condition. Any discriminant value other than those of a valid variant is an invalid byte value.

[tamird]

So the alternative is to remove the Pod bound from various traits e.g. FileHeader. In that case you'd move the Pod bound to methods like Parse which effectively cast a byte slice into a FileHeader. Would that be better?

[bjorn3]

I think moving the Pod bound makes sense.

[tamird]

After much toil: done.

[philipc]

After much toil: done.

I wish you had waited for more discussion before doing this much work. My initial reaction is that we don't want this change.

Can you show what this allows you to accomplish? I don't understand how you can implement FileHeader for an enum even with this change, or why you would want to. FileHeader probably should have been a sealed trait.

[philipc]

I'll discuss further in #549

[tamird]

Separately from the discussion in #549 I think it might be good to land this from a separation-of-concerns perspective. These traits don't care about being Pod apart from when they provide ingress from a byte slice. Let me know if this resonates with you.

[philipc]

Yes, technically the Pod bound could be moved to the parse, but in practice this trait should only be implemented for Pod types anyway, and the loss of ergonomics is not worth the technical purity. The Debug bound is similar. It shouldn't be needed at all, but it was added to make the types much more pleasant to work with.

The amount of boilerplate that you had to add in readobj indicates to me that moving the Pod bound is not the right choice.