Account for subclasses in get_parent_class() of a $this/static value by zonuexe · Pull Request #6048 · phpstan/phpstan-src

zonuexe · 2026-07-14T03:23:13Z

get_parent_class($this) inside a non-final class returned the declared parent (or false when there was none), treating $this as exactly the current class.

Through late static binding, $this can be an instance of a subclass, so for a direct child the parent is the class itself, and for a deeper descendant it is a subclass. The return type now also includes class-string<Class> in that case — which removes a false-positive identical.alwaysFalse on get_parent_class($this) === self::class (the reported bug).

Final classes and non-$this/static arguments (typed parameters, new X()) keep the exact parent type, so get_parent_class($typedParam) and get_parent_class(new B()) are unchanged.

Behaviour

Inside a non-final class A (no parent), with class B extends A:

get_parent_class($this) in A → class-string<A>|false (was false).
get_parent_class($this) in B → 'A'|class-string<B> (was 'A').
get_parent_class($this) in a final class → unchanged (exact parent / false).
get_parent_class($typedParam), get_parent_class(new B()), get_parent_class(SomeClass::class) → unchanged.

The two updated assertions in get-parent-class.php reflect this corrected behaviour.

Note

While checking whether phpstan/phpstan#14951 would interfere with the self/parent/static scope and member-visibility work in #4081 (Closure::bind()), it turned out the two are independent and this one could be fixed on its own — so it is submitted here as a separate PR.

staabm

needs a rule test showing the rule error in the origin reproducer no longer occurs

staabm · 2026-07-14T07:40:04Z

do we have similar problems with other functions, e.g. get_called_class?

zonuexe · 2026-07-14T07:50:46Z

Added tests asserting get_called_class()/get_class() keep the subclass possibility (class-string<static(...)>), so they don't share this issue — it was specific to get_parent_class() pinning the parent to the argument's static class.

I also looked at unifying the three: get_class() already goes through Type::toGetClassResultType(), but get_parent_class() intentionally keeps typed params precise (exact parent, see bug-4302b) whereas get_class() widens them to class-string<Sub>, so a shared abstraction would either regress that precision or add indirection without removing real duplication. Kept them separate.

staabm · 2026-07-14T08:11:56Z

+		assertType('\'ParentClass\\\\Foo\'', get_parent_class(self::class));
+		assertType('\'ParentClass\\\\Foo\'|class-string<ParentClass\Bar>', get_parent_class(static::class));


not sure it makes sense that inside the trait we assume we know the class/all classes this trait might be used by. @VincentLanglet wdyt?

That's the classic behavior of the trait, everything is computed in the context of the class which use the trait. I'm not sure it's easy to change

ok, was not sure tbh. I though this is true from within the class using the trait, but not the trait itself. I might missremember.

in case you think it is fine, we move ahead.

I think you're right, looking at
https://phpstan.org/r/e81f1d9e-19da-42b4-b576-d52cd54be556

We should have a similar behavior than

assertType('class-string|false', get_parent_class($this));

(which already work for get_parent_class(static::class))

VincentLanglet

It should be class-string|false in a trait.

VincentLanglet · 2026-07-14T08:57:29Z

+		assertType('\'ParentClass\\\\Foo\'', get_parent_class(self::class));
+		assertType('\'ParentClass\\\\Foo\'|class-string<ParentClass\Bar>', get_parent_class(static::class));


I think you're right, looking at
https://phpstan.org/r/e81f1d9e-19da-42b4-b576-d52cd54be556

We should have a similar behavior than

assertType('class-string|false', get_parent_class($this));

(which already work for get_parent_class(static::class))

VincentLanglet · 2026-07-14T09:06:35Z

 		assertType('class-string|false', get_parent_class());
 		assertType('class-string|false', get_parent_class($this));
-		// self::class / static::class resolve against the using class (Bar).
-		assertType('\'ParentClass\\\\Foo\'', get_parent_class(self::class));


What about self::class ?

self::class isn't late static bound — inside the trait it's already the using class as a constant class-string ('ParentClass\Bar'), which is indistinguishable from writing Bar::class. So the extension sees the very same ConstantString in both cases and there's nothing to be conservative about; making it class-string|false would mean special-casing the syntax and would contradict get_parent_class(Bar::class) returning 'ParentClass\Foo'.

The guard applies to $this / static::class because those carry the subclass uncertainty. I added assertions covering all three (plus Bar::class side by side) to make this explicit.

VincentLanglet

Isn't weird having a different behavior in trait for

get_parent_class(static::class);

and

get_parent_class(self::class)

@zonuexe @staabm ?

I would expect class-string|false

zonuexe · 2026-07-14T09:49:59Z

I don't think it's the same kind of thing. self::class is not late static bound: inside the trait it is already resolved to the using class as a plain constant ('ParentClass\Bar'), exactly like new self(), self::C and parent::class are.

Two consequences:

A guard on self::class would be purely syntactic and trivially bypassed:

$c = self::class;               // 'ParentClass\Bar'
get_parent_class($c);           // 'ParentClass\Foo' — nothing to guard against

parent::class inside the trait already gives 'ParentClass\Foo', so making get_parent_class(self::class) return class-string|false would hide nothing.

Making it conservative would mean not resolving self::class against the using class at all, which is the core trait behavior you mentioned isn't easy to change.

$this / static::class are different: they carry the subclass uncertainty (the runtime class can be a subclass of the using class), which is why they stay class-string|false — and that mirrors what happens outside a trait, where self::class is exact while static::class widens to class-string<A>|false.

VincentLanglet

I think we could add a test with @final phpdoc ? Which should be considered as final too.

get_parent_class($this) inside a non-final class returned the declared parent (or false when there was none), treating $this as exactly the current class. Through late static binding $this can be an instance of a subclass, so for a direct child the parent is the class itself and for a deeper descendant it is a subclass. The return type now also includes class-string<Class> in that case, which removes a false-positive identical.alwaysFalse on get_parent_class($this) === self::class. Final classes and non-$this/static arguments (typed parameters, new X()) keep the exact parent type. Closes phpstan/phpstan#14951

…lways false Covers the reported identical.alwaysFalse false positive at the rule level, in addition to the type-inference assertions.

…union Addresses review: avoid calling TypeCombinator::union() inside the loop and use a plain foreach for readability.

They already return class-string<static(...)>/class-string<$this(...)>, so they keep the subclass possibility instead of pinning to the current class.

A class-string argument names an exact class, so it is not subject to the subclass widening: get_parent_class(self::class) stays false / the exact parent.

static::class names the late-static-bound class, so it carries the same subclass possibility as $this. Unwrap the class-string<static(...)> and reuse the object handling, giving class-string<Class>|<parent> instead of the wide class-string|false. self::class / X::class name an exact class and keep the exact parent.

…t method $this stays deferred to the general type via the existing in-trait guard; self::class and static::class resolve against the using class the same way as outside a trait.

… a trait The in-trait guard already bails out for $this. Apply it to static::class as well, so a trait does not resolve a late-static-bound argument against one particular using class. Outside a trait static::class keeps the narrowed type.

self::class is not late static bound: inside a trait it is already the using class as a constant class-string, indistinguishable from writing Bar::class, so it keeps the exact parent while $this / static::class stay conservative.

ClassReflection::isFinal() already honours the @Final PHPDoc, so an @Final class gets no subclass widening, same as the final keyword.

getClassStringObjectType() returns an ErrorType for non-class-string types, so the isClassString()->yes() guard was redundant. Dropping it also removes an escaped TrinaryLogic mutant reported by mutation testing.

staabm · 2026-07-14T12:25:52Z

thank you @zonuexe

staabm requested changes Jul 14, 2026

View reviewed changes

zonuexe force-pushed the fix/get-parent-class-non-final-this branch from c4258ce to d814e38 Compare July 14, 2026 05:42

staabm requested changes Jul 14, 2026

View reviewed changes

Comment thread src/Type/Php/GetParentClassDynamicFunctionReturnTypeExtension.php Outdated

Comment thread src/Type/Php/GetParentClassDynamicFunctionReturnTypeExtension.php Outdated

staabm reviewed Jul 14, 2026

View reviewed changes

Comment thread src/Type/Php/GetParentClassDynamicFunctionReturnTypeExtension.php

staabm reviewed Jul 14, 2026

View reviewed changes

Comment thread tests/PHPStan/Rules/Comparison/data/bug-14951.php

staabm reviewed Jul 14, 2026

View reviewed changes

staabm requested a review from VincentLanglet July 14, 2026 08:12

staabm approved these changes Jul 14, 2026

View reviewed changes

VincentLanglet requested changes Jul 14, 2026

View reviewed changes

VincentLanglet reviewed Jul 14, 2026

View reviewed changes

VincentLanglet approved these changes Jul 14, 2026

View reviewed changes

staabm approved these changes Jul 14, 2026

View reviewed changes

zonuexe added 11 commits July 14, 2026 13:55

Add rule test for get_parent_class($this) === self::class not being a…

15e73fc

…lways false Covers the reported identical.alwaysFalse false positive at the rule level, in addition to the type-inference assertions.

Build the get_parent_class() return type with a foreach and a single …

fabdaf6

…union Addresses review: avoid calling TypeCombinator::union() inside the loop and use a plain foreach for readability.

Assert get_called_class()/get_class() are unaffected by the same problem

2903e1e

They already return class-string<static(...)>/class-string<$this(...)>, so they keep the subclass possibility instead of pinning to the current class.

Cover get_parent_class(self::class) keeping the exact parent

8d0091a

A class-string argument names an exact class, so it is not subject to the subclass widening: get_parent_class(self::class) stays false / the exact parent.

Cover get_parent_class() with self::class/static::class inside a trai…

b342212

…t method $this stays deferred to the general type via the existing in-trait guard; self::class and static::class resolve against the using class the same way as outside a trait.

Cover get_parent_class(self::class) inside a trait

fc781b9

self::class is not late static bound: inside a trait it is already the using class as a constant class-string, indistinguishable from writing Bar::class, so it keeps the exact parent while $this / static::class stay conservative.

Cover @Final classes in get_parent_class()

2e4c4af

ClassReflection::isFinal() already honours the @Final PHPDoc, so an @Final class gets no subclass widening, same as the final keyword.

Resolve the class-string object type without a TrinaryLogic check

035e89f

getClassStringObjectType() returns an ErrorType for non-class-string types, so the isClassString()->yes() guard was redundant. Dropping it also removes an escaped TrinaryLogic mutant reported by mutation testing.

staabm force-pushed the fix/get-parent-class-non-final-this branch from 7393c1e to 035e89f Compare July 14, 2026 11:55

staabm merged commit 01262e4 into phpstan:2.2.x Jul 14, 2026
679 of 684 checks passed

zonuexe deleted the fix/get-parent-class-non-final-this branch July 14, 2026 12:26

		assertType('\'ParentClass\\\\Foo\'', get_parent_class(self::class));
		assertType('\'ParentClass\\\\Foo\'\|class-string<ParentClass\Bar>', get_parent_class(static::class));

Uh oh!

Conversation

zonuexe commented Jul 14, 2026

Behaviour

Note

Uh oh!

staabm left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

staabm commented Jul 14, 2026

Uh oh!

Uh oh!

zonuexe commented Jul 14, 2026

Uh oh!

Uh oh!

staabm Jul 14, 2026

Choose a reason for hiding this comment

Uh oh!

VincentLanglet Jul 14, 2026

Choose a reason for hiding this comment

Uh oh!

staabm Jul 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

VincentLanglet Jul 14, 2026

Choose a reason for hiding this comment

Uh oh!

VincentLanglet left a comment

Choose a reason for hiding this comment

Uh oh!

VincentLanglet Jul 14, 2026

Choose a reason for hiding this comment

Uh oh!

VincentLanglet Jul 14, 2026

Choose a reason for hiding this comment

Uh oh!

zonuexe Jul 14, 2026

Choose a reason for hiding this comment

Uh oh!

VincentLanglet left a comment

Choose a reason for hiding this comment

Uh oh!

zonuexe commented Jul 14, 2026

Uh oh!

VincentLanglet left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

staabm commented Jul 14, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

staabm Jul 14, 2026 •

edited

Loading