[Specialization] Alternatve Minimal implementation: only allow whole-impl specialization, haskell style

[lolbinarycat]

Many of the difficulties #31844 has encountered come from dealing with things that happen when some but not all of a trait's items are overridden. I would argue that this is overkill for the common usecase of making blanket impls not conflict with more specific impls. this would also address the most common reason for people asking for negative type constraints.

[workingjubilee]

Rust does not have Haskell's type system, it has a much more complicated one (unless you mean "Haskell with approximately 100 compiler features enabled", in which case it has a much simpler one), so, while I'm not sure if this is the best venue for such a proposal, in whatever venue you do take it to, you're going to want to actually explain what you mean.

[lolbinarycat]

in my proposal there would be two types of impl blocks: regular impl and default impl. it is legal for a default impl block to conflict with a regular impl block. for types with two matching implementations of a trait, the regular impl gets priority over the default impl. it is not legal for two impls of the same priority to overlap.

[SkiFire13]

Many of the difficulties #31844 has encountered come from dealing with things that happen when some but not all of a trait's items are overridden.

AFAIK the main soundness problem is with specialized implementation applying depending on some lifetimes. How are you proposing to solve that?

[lolbinarycat]

could you give an example of code that would be potentially unsound under this proposal?

[SkiFire13]

See #40582 for an example. This old blogpost has a nice explanation of what's the fundamental problem with specialization, along with a bunch of other tricky examples.