Support (package) in (mdx)

[emillon]

This adds a (package) field to (mdx) with the usual semantics.

The syntax is not ideal since (packages) is already supported as sugar
to specify dependencies.

Closes #3756

[emillon]

I had trouble implementing this while avoiding dependency cycles. the key problem is that the central check for "should we consider this stanza based on the current value of --only-packages?" is in Dune_file.t, and we can't refer to Mdx over there. This would make Dune_file.t depend on all the extensions and we don't want that.

To solve that I created a Package_proxy constructor that implements only package filtering for an arbitrary stanza. That works but it seems a bit heavy and could be error-prone. If we're happy with that, we could use it in more places. That could also be used to implement a generic (package ...) stanza that would wrap other stanzas.

An alternative is to do the same as in Test_rules and generate "fake" user rules (Dune_file.Rule.t). These are properly filtered by the existing code since they are defined in Dune_file.

The goal is to backport and include this feature in 2.9. It is also going to be added to the "new" version of the mdx stanza (0.2) in #3956, with more pleasing syntax. But this is a larger bit of work, and probably won't be backported.

Some remarks / TODO list:

• update documentation and note the syntax quirk
• find a solution to Package_proxy or become happy with it
• update changelog

Let me know what you think. Thanks!

[emillon]

Using user rules does not seem to be possible because of the way deps are encoded in a file. Another alternative seems to call Only_packages directly as done in Cram_rules. I'll try that.

[emillon]

That looks better now.

[emillon]

This is now ready for review.

[voodoos]

I did not see the previous version with Package_proxy so I cannot compare, but these changes looks good to me.

[ejgallego]

Hi folks, if we want this backported to 2.9 we should try to get it ready before Fri, cheers.

[ghost]

LGTM.

The package/packages discrepancy is unfortunate indeed. Perhaps we could replace (packages x y) by (deps (package x) (package y)). Or finally move to one package per directory and get rid of package fields altogether.