Implement declarative (macro_rules!) derive macros (RFC 3698)

[joshtriplett]

This is a draft for review, and should not be merged yet.

This is layered atop #145153 , and has
only two additional commits atop that. The first handles parsing and provides a
test for various parse errors. The second implements expansion and handles
application.

This implements RFC 3698, "Declarative (macro_rules!) derive macros".
Tracking issue: #143549

This has one remaining issue, which I could use some help debugging: in
tests/ui/macros/macro-rules-derive-error.rs, the diagnostics for
derive(fn_only) (for a fn_only with no derive rules) and
derive(ForwardReferencedDerive) both get emitted twice, as a duplicate
diagnostic.

From what I can tell via adding some debugging code,
unresolved_macro_suggestions is getting called twice from
finalize_macro_resolutions for each of them, because
self.single_segment_macro_resolutions has two entries for the macro, with two
different parent_scope values. I'm not clear on why that happened; it doesn't
happen with the equivalent code using attrs.

I'd welcome any suggestions for fixing this.

[joshtriplett]

r? @petrochenkov

[joshtriplett]

(Removing T-clippy and T-rustdoc, because the two commits specific to this PR don't touch them.)

[joshtriplett]

(Marking as "waiting on review" because I'm seeking help with the test failure here.)

[petrochenkov]

Blocked on #145153.
@rustbot blocked

[joshtriplett]

(Marking for review now that 145153 is queued.)

[petrochenkov]

(I'll wait until the rebase anyway.)

[joshtriplett]

@rustbot ready

[petrochenkov]

From what I can tell via adding some debugging code,
unresolved_macro_suggestions is getting called twice from
finalize_macro_resolutions for each of them, because
self.single_segment_macro_resolutions has two entries for the macro, with two
different parent_scope values. I'm not clear on why that happened; it doesn't
happen with the equivalent code using attrs.

Probably because of this FIXME in expand.rs

                                    // FIXME: Consider using the derive resolutions (`_exts`)
                                    // instead of enqueuing the derives to be resolved again later.

I don't think it's too important to fix, identically looking diagnostics are not shown to users anyway (UI tests suite specifically disables that deduplication).

It's suspicious that the parent_scope values are different though.

[joshtriplett]

From what I can tell via adding some debugging code,
unresolved_macro_suggestions is getting called twice from
finalize_macro_resolutions for each of them, because
self.single_segment_macro_resolutions has two entries for the macro, with two
different parent_scope values. I'm not clear on why that happened; it doesn't
happen with the equivalent code using attrs.

Probably because of this FIXME in expand.rs

                                    // FIXME: Consider using the derive resolutions (`_exts`)
                                    // instead of enqueuing the derives to be resolved again later.

Oh. Yeah, that would do it; thanks.

I don't think it's too important to fix, identically looking diagnostics are not shown to users anyway (UI tests suite specifically disables that deduplication).

Fair enough. It does seem worth fixing for performance reasons if nothing else, but then, the FIXME is there for a reason. I'll ignore it for now, thank you.

It's suspicious that the parent_scope values are different though.

That code appears to generate fresh expansions for the enqueuing.

I poked pretty extensively at test cases, though, and couldn't find any case where it happened more than twice. So, I think this is worst case 2N, not M*N or N^2.

[joshtriplett]

@rustbot ready

[petrochenkov]

@bors r+

[bors]

📌 Commit 354fcf2 has been approved by petrochenkov

It is now in the queue for this repository.