[match-case] Fix narrowing of class pattern with union-argument.

[randolf-scholz]

Fixes #19468

I added factorization over UnionType inside the conditional_types function.
Then try_expanding_sum_type_to_union needed to move to the start of the function so that we expand current_type before attempting factorizating the union type.

I also reformatted the docstring for better legibility.

New Tests

• testMatchNarrowDownUnionUsingClassPattern (https://mypy-play.net/?mypy=1.17.0&python=3.12&gist=e9ec514f49903022bd32a82ae1774abd)

[randolf-scholz]

The default == current_type was necessary, otherwise quite a few tests break.
Adding or (default is None) and passing default=None in the recursive call also seems to break stuff. Not entirely sure what's going on there.

[sterliakov]

If I'm not mistaken, this only works by coincidence and is incorrect in general case.

The problem is that you try to union and restrict multiple results here. This is only reasonable when none of the ranges are upper bounds, otherwise this is unsound:

class Base: pass

class A1(Base): ...
class A2(Base): ...
class A3(Base): ...

ref: type[Base] = A1
x: Base | int

if isinstance(x, ref):
    reveal_type(x)  # Should be Base
else:
    reveal_type(x)  # Should be Base | int (e.g. `x = A2()`)

Applying the logic from this PR, you get Never | Base == Base as yes_type (correct), but then you get (Base | int) \ Base == int as no_type, which is wrong. Since this has nothing to do with default, likely this piece just happens to work on our testcases?

I can suggest only performing this expansion when not any(tr.is_upper_bound for tr in proposed_type_ranges), but maybe there's a better general solution similar to the logic below.

[randolf-scholz]

I ran your test example and it gives the expected results.

$ mypy tmp.py 
tmp.py:11: note: Revealed type is "tmp.Base"
tmp.py:13: note: Revealed type is "tmp.Base | builtins.int"

I believe it depends on from where this is called. conditional_types is called in a few places:

• refine_identity_comparison_expression without a default
• conditional_types_with_intersection with default, which itself is called in 7 places:
  1. TypeChecker.find_type_equals_check with current_type=self.lookup_type(expr) and default=None
  2. TypeChecker.find_isinstance_check_helper with current_type=self.lookup_type(expr) and default=None
  3. TypeChecker.infer_issubclass_maps with current_type=vartype and default=None
  4. PatternChecker.visit_as_pattern with current_type=current_typeanddefault=current_type`
  5. PatternChecker.visit_value_pattern with current_type=current_typeanddefault=get_proper_type(typ)`
  6. PatternChecker.visit_singleton_pattern with current_type=current_typeanddefault=current_type`
  7. PatternChecker.visit_sequence_pattern with current_type=inner_typeanddefault=inner_type`
  8. PatternChecker.visit_sequence_pattern with current_type=new_tuple_typeanddefault=new_tuple_type`
  9. PatternChecker.narrow_sequence_child with current_type=outer_typeanddefault=outer_type`
  10. PatternChecker.visist_class_pattern with current_type=current_typeanddefault=current_type`

So, we can see for the isinstance calls, default is None, so this branch is actually never taken currently, but may be taken for all the call-sites within PatternChecker except possibly PatternChecker.visit_value_pattern

[randolf-scholz]

If I try to replicate your example with match-case we get these results:

class Base: pass

class A1(Base): ...
class A2(Base): ...
class A3(Base): ...

ref: type[Base] = A1
x: Base | int

match x:
    case ref() as y: # E: expected type in class pattern; found "type[tmp.Base]"
        reveal_type(y)  # E: Statement is unreachable
    case other:
        reveal_type(other)  # Base | int

match x:
    case Base() as y:
        reveal_type(y)  # Base
    case other:
        reveal_type(other)  # int

match x:
    case A1() as y:
        reveal_type(y)  # A1
    case other:
        reveal_type(other)  # Base | int

[randolf-scholz]

@sterliakov I integrated your suggestion into a separate branch #19517, and was able to relax the condition to

    if (
        isinstance(current_type, UnionType)
        and not any(tr.is_upper_bound for tr in proposed_type_ranges)
        and (default in (current_type, None))
    ):

so that it also is applied with regular isinstance branches when default=None

[sterliakov]

Same question as in #19471, but this will be conflicted anyway:)

[sterliakov]

If I'm not mistaken, this only works by coincidence and is incorrect in general case.

The problem is that you try to union and restrict multiple results here. This is only reasonable when none of the ranges are upper bounds, otherwise this is unsound:

class Base: pass

class A1(Base): ...
class A2(Base): ...
class A3(Base): ...

ref: type[Base] = A1
x: Base | int

if isinstance(x, ref):
    reveal_type(x)  # Should be Base
else:
    reveal_type(x)  # Should be Base | int (e.g. `x = A2()`)

Applying the logic from this PR, you get Never | Base == Base as yes_type (correct), but then you get (Base | int) \ Base == int as no_type, which is wrong. Since this has nothing to do with default, likely this piece just happens to work on our testcases?

I can suggest only performing this expansion when not any(tr.is_upper_bound for tr in proposed_type_ranges), but maybe there's a better general solution similar to the logic below.

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅

[randolf-scholz]

Converting to draft in favor of sister PR #19517