An item is parametric over lifetimes is if
- every lifetime in it is different
- none of the lifetimes are 'static
- none of them are subtypes of each other (`'a: 'b`)
Conversation
Specialization is unsound because while we can check lifetimes at typeck time, by the time we generate code we don't have lifetimes anymore, so we might choose the less special impl.
2
2
there were problems that i do not recall in getting typeck to do that. AFAIK we are hoping to stabilize specialization without associated types so that typeck getting the wrong instance never matters
1
Ah, associated type specialization, that matches something I remember. Although you *would think* erasing lifetimes when doing specialization trait queries would "just work" - maybe it's non-trivial to detect due to caching? cc
1
might have to do with higher rank lifetimes but i mainly remember being confused by aaron and niko's conversation and the repeated invokation of the word 'skolemize'
1
1
just wanted to jump in the thread to voice support for any Rust feature that uses the word “skolemize” in its docs
1
7
Quote Tweet
Replying to @Gankra_
I am literally too excited to properly read this
I just whispered "fuck yeah skolemized lifetimes" to myself
1
3
AFAIK, rough explanation is that skolemization is a way of normalizing complex types with 'foralls' in them so that you can check that they are the same. In Rust this is important for higher rank lifetimes. Comes from logic, after 'Thoralf Skolem': en.wikipedia.org/wiki/Skolem_no
1
5
Yep! It basically treats "foralls" (i.e. generics) as if they were fresh concrete types that cannot be unified with any existing concrete types. That's how you make sure that your conclusions really do work *for all* concrete instantiations.
1
3
Ah neat TIL
1
Show replies