rustdoc doesn't consider blanket(!) impls of auto(!) traits

[fmease]

Background: User-written trait impls of a given type disqualify potential built-in auto trait impls (modulo simplify_type) ignoring some extra details.

Consider the following code:

#![feature(auto_traits)]

pub auto trait Marker {}
pub struct Subject;

impl<T> Marker for T {}

Here, the existence of the blanket impl of Marker disqualifies any auto trait candidates. Meaning Subject impls Marker not due to a built-in auto trait impl but due to the blanket impl.

Therefore I expect to find impl<T> Marker for T under section Blanket Implementations on Subject's page. However, there is no such impl. It does correctly determine it shouldn't synthesize an auto trait impl of Marker for Subject.

If you delete the auto qualifier of trait Marker, then rustdoc will correctly detect and render impl<T> Marker for T.

[fmease]

A "more exciting" example:

#![feature(auto_traits)]

pub auto trait Marker {}

#[derive(Clone, Copy)]
pub struct Subject;

impl<T: Copy> Marker for T {}

Same story, it should place an impl<T> Marker for T where T: Copy under Blanket Implementations but it doesn't.

[fmease]

Very likely due to

rust/src/librustdoc/clean/auto_trait.rs

Lines 74 to 77 in e1a2ec6

	if !cx.generated_synthetics.insert((ty, trait_def_id)) {
	debug!("already generated, aborting");
	return None;
	}

and

rust/src/librustdoc/clean/blanket_impl.rs

Lines 27 to 31 in e1a2ec6

	if !cx.cache.effective_visibilities.is_reachable(tcx, trait_def_id)
	|| cx.generated_synthetics.contains(&(ty.skip_binder(), trait_def_id))
	{
	continue;
	}

We mark the auto trait impl "request" as visited even if we don't end up synthesizing an impl (due to a user-written impl existing). Then, when computing blanket impls, we skip the "same" request due to being "visited".

We should split cx.generated_synthetics into two, one set for auto trait impls and one for blanket impls.

[fmease]

Alternatively we could synthesize blanket impls before auto trait impls since blanket impls should "always" trump auto trait impls. Not sure if there are edge cases where that doesn't hold.