This PR fixes an old bug in the interaction between [merge_constraint]
and [Typedecl.transl_with_constraint], where
variance (and now separability) are recomputed in an invalid type
environment. See #9624 and the new tests.
Because this changes the separability of standard library types, and
those separabilities are stored in the .cmi files, this commit changes
the .cmi files in the standard libraries in way that appear to require
a bootstrap (it looks like some part of the stdlib is built with
boot/ocamlc and others with ocamlc, and the two should produce/expect
the same .cmi exactly). The bootstrap will come as a separate commit.
improved error messages for fixed row polymorphic variants: keep track of the motivation behind a fixed row (e.g it was bound to an universal or existential type variable, or private) in the types themselves and use this explanation in error messages.
get_unboxed_type_representation is used in several other modules, and
split into its own Typedecl_unboxed unit.
(pair-programming with Gabriel Scherer)
The location can always be reconstructed from `decl.type_loc`. Note
however that variance-computation functions, in particular
`compute_variance_extension` and its callee (`compute_variance_gadt`,
etc.) need to be passed both the requirement and the location, despite
the fact that they also take a `decl` parameter: there are other
use-cases where the location is not the location of the type
declaration -- when processing an extension to an existing type
declaration.
Also included are a generic combinator, and a port of the variance
computation to the generic scheme -- without both those things, we
wouldn't know we have the right (internal) interface.
This change makes the code simpler and easier to extend. It also
avoids useless computation: the two fixpoints may take a different
number of iterations to converge, and the previous code would compute
both properties as long as one of them hadn't converged.
These are most likely a mistake for "type t = unit", but still valid, and lead
to quite confusing error messages afterwards because now `()` denotes two
different incompatible constructors.
Previously, not having a scope meant the type was used in every context,
now we set the scope to "Btype.lowest_level" to mean the same thing.
The equivalence was made obvious by the recent changes to identifiers
scoping.
- when including: the elements are rescoped to the current level (as
well as being given a fresh stamp, which was already the case)
- extension constructors idents cannot be local: they can be used as
types when the constructor's argument is an inlined record. They must be
given a scope
- in check_recmod_inclusion: create a new scope at each iteration
- when checking that type declarations inside recursive modules are well
founded, we now take a generic instance of the declaration (this is
reminiscent of what is done in Ctype.moregeneral)
- Ident.create now takes a scope as argument
- added Ident.create_var to use when the scope doesn't matter
- the current_time and the current_level are unrelated as of this
commit. But one has to remember to bump the level when creating new
scopes.
Florian Angeletti and myself ran into a problem when trying to use attributes
for ellision of parts of manual example. We wanted to be turn any ast-node
marked with the [@ellipsis] attribute into "..." in the rendering of the
corresponding code block, but for this we need the location of the
attributed node, and it turns out that some constructions supported
attributes without carrying a location:
- Rtag in row_field
- Otag in object_field
- type_exception record
- type_extension record
We added locations in all those positions, guaranteeing the invariant
that all nodes to which attributes can be attached have a precise
position.
Gabriel Scherer
Jacques Garrigue
Fourchaux
Leo White
Thomas Refis
Rodolphe Lepigre
Florian Angeletti
Jeremie Dimino
Rian Douglas
Alain Frisch
Florian Angeletti
Armaël Guéneau
Nicolás Ojeda Bär
Stefan Muenzel