You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Streamline some elements of variadic types support (#15924)
Fixes#13981Fixes#15241Fixes#15495Fixes#15633Fixes#15667Fixes#15897Fixes#15929
OK, I started following the plan outlined in
#15879. In this PR I focused mostly
on "kinematics". Here are some notes (in random order):
* I decided to normalize `TupleType` and `Instance` items in
`semanal_typeargs.py` (not in the type constructors, like for unions).
It looks like a simpler way to normalize for now. After this, we can
rely on the fact that only non-trivial (more below on what is trivial)
variadic items in a type list is either `*Ts` or `*tuple[X, ...]`. A
single top-level `TupleType` can appear in an unpack only as type of
`*args`.
* Callables turned out to be tricky. There is certain tight coupling
between `FuncDef.type` and `FuncDef.arguments` that makes it hard to
normalize prefix to be expressed as individual arguments _at
definition_. I faced exactly the same problem when I implemented `**`
unpacking for TypedDicts. So we have two choices: either handle prefixes
everywhere, or use normalization helper in relevant code. I propose to
go with the latter (it worked well for `**` unpacking).
* I decided to switch `Unpack` to be disallowed by default in
`typeanal.py`, only very specific potions are allowed now. Although this
required plumbing `allow_unpack` all the way from `semanal.py`,
conceptually it is simple. This is similar to how `ParamSpec` is
handled.
* This PR fixes all currently open crash issues (some intentionally,
some accidentally) plus a bunch of TODOs I found in the tests (but not
all).
* I decided to simplify `TypeAliasExpr` (and made it simple reference to
the `SymbolNode`, like e.g. `TypedDictExpr` and `NamedTupleExpr`). This
is not strictly necessary for this PR, but it makes some parts of it a
bit simpler, and I wanted to do it for long time.
Here is a more detailed plan of what I am leaving for future PRs (in
rough order of priority):
* Close non-crash open issues (it looks like there are only three, and
all seem to be straightforward)
* Handle trivial items in `UnpackType` gracefully. These are `<nothing>`
and `Any` (and also potentially `object`). They can appear e.g. after a
user error. Currently they can cause assert crashes. (Not sure what is
the best way to do this).
* Go over current places where `Unpack` is handled, and verify both
possible variadic items are handled.
* Audit variadic `Instance` constrains and subtyping (the latter is
probably OK, but the former may be broken).
* Audit `Callable` and `Tuple` subtyping for variadic-related edge cases
(constraints seem OK for these).
* Figure out story about `map_instance_to_supertype()` (if no changes
are needed, add tests for subclassing).
* Clear most remaining TODOs.
* Go once more over the large scale picture and check whether we have
some important parts missing (or unhandled interactions between those).
* Verify various "advanced" typing features work well with
`TypeVarTuple`s (and add some support if missing but looks important).
* Enable this feature by default.
I hope to finish these in next few weeks.
0 commit comments