Fixes to tuple fallbacks #6442
Fixes to tuple fallbacks #6442
Conversation
This introduces a patch-up pass after the main semantic analysis pass.
| @@ -2785,7 +2787,7 @@ def visit_tuple_expr(self, e: TupleExpr) -> Type: | |||
| tt = self.accept(item, type_context_items[j]) | |||
| j += 1 | |||
| items.append(tt) | |||
| fallback_item = join.join_type_list(items) | |||
| fallback_item = AnyType(TypeOfAny.special_form) | |||
There was a problem hiding this comment.
Maybe add a comment above reminding that this is just a partial fallback, and a real fallback is always calculated on-demand.
mypy/newsemanal/semanal_shared.py
Outdated
| @@ -19,12 +20,8 @@ | |||
| # Priorities for ordering of patches within the final "patch" phase of semantic analysis | |||
| # (after pass 3): | |||
There was a problem hiding this comment.
There is no pass three anymore, so this comment can also be updated.
|
|
||
| This must be called only after the main semantic analysis pass, since joins | ||
| aren't available before that. | ||
| """ |
There was a problem hiding this comment.
I would mention the chicken and egg problem about tuple fallbacks and checking upper bounds for instances here, and/or maybe in a comment in semanal_main() where the two are called.
test-data/unit/check-newsemanal.test
Outdated
| class C(A): pass | ||
|
|
||
| class T(Tuple[B, C]): | ||
| pass |
There was a problem hiding this comment.
Maybe put the classes in reverse order (just to check that forward refs also work, although I am quite they should).
|
|
||
| class NTStr(NamedTuple): | ||
| x: str | ||
| y: str |
There was a problem hiding this comment.
I would add couple non-trivial upper bound checks, just be sure they behave reasonably with the current ordering (fallbacks first, then bounds).
| for nx in nt: | ||
| reveal_type(nx) # E: Revealed type is 'Union[builtins.int*, builtins.str*]' | ||
|
|
||
| t: Union[Tuple[int, int], Tuple[str, str]] | ||
| for x in t: | ||
| # TODO(Ivan): This will be OK when tuple fallback patches are added (like above) | ||
| reveal_type(x) # E: Revealed type is 'Any' |
mypy/newsemanal/semanal_main.py
Outdated
| @@ -204,6 +217,11 @@ def semantic_analyze_target(target: str, | |||
| return [], analyzer.incomplete | |||
|
|
|||
|
|
|||
| def process_patches(graph: 'Graph', scc: List[str]) -> None: | |||
| for module in scc: | |||
| graph[module].semantic_analysis_apply_patches() | |||
There was a problem hiding this comment.
I would add a motivational docstring explaining why we do mini-passes for one things, and patches for other. My understanding that there will be quite few tuple types in a typical program, but quite a lot instances, so we a make a traverser for the latter.
Also there is an option to combine these two in a single type fixing traverser pass, and then repeat it until we solved the chicken and egg problem (this would require to somehow mark types as ready). I however didn't think much about this, maybe this will not help.
This changes tuple fallbacks to be calculated on demand during
type checking. This fixes some issues with fallbacks being imprecise.
In the new semantic analyzer, this introduces a new pass just after
the main semantic analysis pass to calculate precise item types for
tuplebase classes. These can't be calculated on demand sincethe base class is a variable-length tuple type (
Instance) insteadof
TupleType. We can't calculate these during the main semanticanalysis pass since base classes can be incomplete.
Fixes #6400.