it's much simpler to handle down the line if they have a uniform structure;
empty tuples are easily converted into unit types when translating to OCaml.
Support for manipulating toplevel functions as values was buggy, because the
recursion after eta-expansion would fall into the pattern for a `let..in` and
not do the expected transformation.
The patch explicitely builds the closure in that case, avoiding such issues with
recursion.
some of the types (in particular, in hoisted closures) could not be
reconstructed afterwards. This properly propagates the types, including to
closure deconstruction time, giving additional insurance; and allowing
monomorphisation not to choke on the result.
NOTE: This is a temporary solution
A future approach could be to have Catala generate a module loader (with the
proper hash), relieving the user implementation from having to do the
registration.
This includes a few separate changes:
- pass visibility information of declarations (depending on wether the
declaration was in a ```catala-metadata block or not)
- add reasonable hash computation functions to discriminate the interfaces. In
particular:
* Uids have a `hash` function that depends on their string, but not on their
actual uid (which is not stable between runs of the compiler) ; the existing
`hash` function and its uses have been renamed to `id`.
* The `Hash` module provides the tools to properly combine hashes, etc. While
we rely on `Hashtbl.hash` for the atoms, we take care not to use it on any
recursive structure (it relies on a bounded traversal).
- insert the hashes in the artefacts, and properly check and report those (for
OCaml)
**Remains to do**:
- Record and check the hashes in the other backends
- Provide a way to get stable inline-test outputs in the presence of module
hashes
- Provide a way to write external modules that don't break at every Catala
update.
mostly reverting to the ones the interpreter was printing ; for the case of
divisions, we choose to point to the denominator instead of the operator as it's
where the only possible error (division by zero) comes from.
- Clearly distinguish Exceptions from Errors. The only catchable exception
available in our AST is `EmptyError`, so the corresponding nodes are made less
generic, and a node `FatalError` is added
- Runtime errors are defined as a specific type in the OCaml runtime, with a
carrier exception and printing functions. These are used throughout, and
consistently by the interpreter. They always carry a position, that can be
converted to be printed with the fancy compiler location printer, or in a
simpler way from the backends.
- All operators that might be subject to an error take a position as argument,
in order to print an informative message without relying on backtraces from
the backend
- This adds a `catala depends` command that recursively tracks module dependency.
It can then be used by Clerk for linking.
- Generation of cmo object files are added for OCaml (we only built native
objects, but jsoo requires bytecode).
- Some fixes to the generation of value embed/deembed shims (related to types
coming from different modules ; add support for options ; etc.)
Print to json directly rather than depend on yojson and a ppx.
Note: this should be tested with the website in order to validate that the Json
output is 1-to-1.
(a second step could be to simplify this output, now that it's manual)
so don't retype after monomorphisation, which is now possible as the pass itself
correctly preserves types.
In time the typer will need a special case to get knowledge of the new types and
modified operator types.
A little bit of effort enables us to propagate valid typing annotations, making
subsequent typing re-inference easier (and avoiding a traversal just to remove
type annotations)
since it doesn't correctly propagate and update types. (Ideally it would, but
otherwise it would be better to remove the type annotations on the fly instead
of introducing wrong ones then cleaning them up).
As part of making tuples first-class citizens, expliciting the arity upon
function application was needed (so that a function of two args can
transparently -- in the surface language -- be applied to either two arguments
or a pair).
It was decided to actually explicit the whole type of arguments because the cost
is the same, and this is consistent with lambda definitions.
A related change done here is the replacement of the `EOp` node for operators by
an "operator application" `EAppOp` node, enforcing a pervasive invariant that
operators are always directly applied. This makes matches terser, and highlights
the fact that the treatment of operator application is almost always different
from function application in practice.
