[C++] Support requires-clause in more places

[jbms]

Feature or Bugfix

• Feature

Purpose

Previously a C++20 requires-clause was only supported on function
declarations. However, the C++ standard allows a require-clause on
class/union templates, alias templates, and variable templates, and
also allows a requires clause after each template parameter list, not
just the final one.

This moves the requiresClause to be a property of ASTTemplateParams
rather than ASTDeclaration to better match the C++ grammar and
allows requires clauses in many places that are supported by C++20 but
were not previously allowed by Sphinx, namely:

• On class/union templates, alias templates, and variable templates

• After each template parameter list, not just the last one.

• After the template parameter list in template template parameters.

Additionally:

• This adds support for template parameters on unions.

• This adds support for trailing requires clauses on functions without
  a template prefix. This is allowed by C++20 for non-template members
  of class templates.

When encoding the id, the requires clause of the last template
parameter list is treated specially in order to preserve compatibility
with existing v4 ids.

[jbms]

Note: I made some major changes to this PR to support a lot of additional things.

As it is currently implemented, it does not prohibit requires clauses on concepts even though those are disallowed by C++. But I don't think that is a problem.

Note: I noticed that there is support for non-standard concept introduction syntax, which I guess was part of an earlier concepts proposal. Should that be removed since it is not supported by C++?

[jakobandersen]

Thanks for the PR. There are a number of structural changes to the code, and you list multiple changes. Please split it up into multiple commits to make it easier to review.

[jakobandersen]

As it is currently implemented, it does not prohibit requires clauses on concepts even though those are disallowed by C++. But I don't think that is a problem.

It would be nice to not allow it, unless it is really cumbersome to disallow.

Note: I noticed that there is support for non-standard concept introduction syntax, which I guess was part of an earlier concepts proposal. Should that be removed since it is not supported by C++?

They were indeed implemented during the Concepts TS time-frame. If they are in the way of implementing new things we should deprecate them, otherwise I see no harm in letting them be.

[jbms]

Thanks for reviewing this change. I have split it up into multiple commits, and added a check to prevent a requires-clause on concepts.

[jbms]

@jakobandersen Please let me know if there is anything I can do to help make progress with this and my other C++ PRs. Thanks!

[jakobandersen]

When encoding the id, the requires clause of the last template
parameter list is treated specially in order to preserve compatibility
with existing v4 ids.

Is this really needed? As currently only the clause being ID-encoded is allowed, then all currently allowed declarations will have the same ID afterwards anyway. Or am I missing something?

[jbms]

When encoding the id, the requires clause of the last template
parameter list is treated specially in order to preserve compatibility
with existing v4 ids.

Is this really needed? As currently only the clause being ID-encoded is allowed, then all currently allowed declarations will have the same ID afterwards anyway. Or am I missing something?

Previously, the requires clause was encoded after the template parameter list, so with a requires clause after the last template parameter list we would have:

"I" + template_parameter_list_id + "E" + "IQ" + requires_clause_id + "E"

if we didn't special case requires clauses on the last template parameter list, it would instead be:

"I" + template_parameter_list_id + "IQ" + requires_clause_id + "E" + "E"

[jbms]

I didn't see anything in the Itanium ABI regarding requires clauses --- not sure if you just invented a convention for them or if you used a convention defined elsewhere. With this change, I just used a convention that seemed to make sense.

[AA-Turner]

@jakobandersen Should we move this to 5.x, or is it nearly ready to be merged for 5.1?

A

[jakobandersen]

@jakobandersen Should we move this to 5.x, or is it nearly ready to be merged for 5.1?

I won't have time to properly dig back into it before the weekend, so please go ahead with the release and move this to 5.2.

[jakobandersen]

if we didn't special case requires clauses on the last template parameter list ...

Indeed, good catch.

[jakobandersen]

I didn't see anything in the Itanium ABI regarding requires clauses --- not sure if you just invented a convention for them or if you used a convention defined elsewhere. With this change, I just used a convention that seemed to make sense.

Right, to my understanding it hasn't been updated officially yet, so this is just something I cooked up, and your changes are fine.

[jakobandersen]

I have rebased and updated this PR and merged it as #10703. Thanks again for your nice work.