Support ::class magic constant in array shape keys and const types

[janedbal]

PHPStan didn't support the ::class magic constant as a value or key in PHPDoc array shapes.

class Test {
    /** @return array{Test::class: int} */
    static function foo(): array {
        $r = [self::class => 1];
        // inferred directly: array{X\Test: 1}  ✅
        return $r;
    }
}

\PHPStan\dumpType(Test::foo()); // was: non-empty-array<int>  ❌

The inferred type inside the method was correct (array{X\Test: 1}), but the declared @return resolved the Test::class key to ErrorType, collapsing the whole shape into non-empty-array<int>.

Cause: both PHPDoc resolution paths for Foo::CONST (resolveArrayShapeOffsetType() for keys, resolveConstTypeNode() for value types) look the constant up via $classReflection->hasConstant($constantName). But ::class is a magic pseudo-constant, not a declared one, so hasConstant('class') is false and resolution falls through to ErrorType. Regular class constants as keys (array{Foo::BAR: int}, already documented) were unaffected.

Fix: handle ::class in both paths, mirroring InitializerExprTypeResolver::getClassConstFetchTypeByReflection():

• static::class on a non-final class → class-string<static>
• everything else (self, parent, literal names, static on final) → a constant class-string

This also unblocks consumers that rely on exact Class::class keys surviving through array_merge, e.g. shipmonk-rnd/dead-code-detector#374 (shipmonk-rnd/dead-code-detector#374 (comment)).

Co-Authored-By: Claude Code

Docs: phpstan/phpstan#14837

[staabm]

this line seems to miss test-coverage

[janedbal]

You can place similar check to CI btw: https://github.com/shipmonk-rnd/coverage-guard ;) (even for MR-affected lines only)

[janedbal]

Anyway, coverage fixed

[staabm]

You can place similar check to CI btw: shipmonk-rnd/coverage-guard ;) (even for MR-affected lines only)

sounds like something we could try in a separate PR. might be useful, I am not sure yet.

could you send a PR?

[janedbal]

I can send a PR, but the rule decisions should be done by core maintainers. Or maybe you can use it to only observe what is tested in MRs (let fail all the time when coverage of new files < 100% or so) 🤷

[staabm]

thank you!

[staabm]

@janedbal what do you think about the magic "name" property and UnitEnum, EnumCase.

I think it has a similar problem?

https://phpstan.org/r/57dcadcd-0bba-428a-b8b1-319dc134e5bf

[janedbal]

Seems rare to me :) I didnt even know (forgot?) we have enum-string