Butchered `std::any::type_name` output for trait object types with lifetime params

[fmease]

Reproducer:

fn main() {
    trait Trait<'a, T> {}
    dbg!(std::any::type_name::<dyn Trait<'_, ()>>()); // yields `"dyn playground::main::Trait<, ()>"`
                                                      // notice the "bad comma"
}

Marked T-libs for obvious reasons but this is actually something that needs to be fixed in the compiler.

Earlier Zulip discussion: #t-compiler > `type_name` omits lifetimes?.

[fmease]

The linked Zulip discussion also concerns itself with the way higher-ranked / late-bound lifetimes should be rendered which is only tangentially related. E.g., how we render some distinct types the same like for<'a> fn(for<'b> fn(S<'a>, S<'b>)) and for<'a> fn(for<'b> fn(S<'b>, S<'a>)) for which we both output "fn(fn(playground::main::S, playground::main::S))". Of course, we don't guarantee that property for type_name, it would just be nice.

[fmease]

The cause for this bug is the fact that we don't/can't reuse path_generic_args in pretty_print_dyn_existential which would've filtered out lifetime args entirely in the PrettyPrinter impl for type_name::AbsolutePathPrinter. Instead we just don't print any lifetimes (leaving odd holes and thus "bad commas") because should_print_regions is still false. And we can't use path_generic_args there because we need to insert associated item bindings (equality constraints) into the angle-bracketed list.

[fmease]

The most minimal fix might just be to adjust path_generic_args's signature to allow for a "callback" that's to be run just before inserting the final >, so pretty_print_dyn_existential gets the chance to inject assoc item bindings that way. However, that doesn't work as is because we want to omit the delimiters if the args are empty (or the constraints). So I dunno.

[fmease]

I guess the easiest solution would be to just inspect should_print_regions inside pretty_print_dyn_existential. That's a bit hacky tho since it's a bit spaghetti / entangling.