Fix phpstan/phpstan#11054: [NAN] === mixed is always evaluated as false

[phpstan-bot]

Summary

PHPStan incorrectly reported Strict comparison using === between mixed and array{NAN} will always evaluate to false when a variable of type mixed was compared against multiple class constants containing [NAN]. After the first === CONST_A check, PHPStan narrowed the type by subtracting array{NAN}, causing subsequent comparisons against other [NAN] constants to be flagged as always false.

Changes

• Overrode getFiniteTypes() in src/Type/Constant/ConstantFloatType.php to return [] for NAN values
• Added regression test in tests/PHPStan/Rules/Comparison/data/bug-11054.php
• Added test method testBug11054 in tests/PHPStan/Rules/Comparison/StrictComparisonOfDifferentTypesRuleTest.php

Root cause

ConstantFloatType(NAN) inherited getFiniteTypes() from ConstantScalarTypeTrait which returned [$this]. This caused ConstantArrayType containing NAN to also report a single finite type. The TypeSpecifier uses getFiniteTypes() to decide whether to narrow (subtract) a type in the false branch of === checks. For NAN-containing arrays, this narrowing is incorrect because in PHP, two references to the same [NAN] array (e.g., class constants) ARE identical via === (same zval), even though structurally different [NAN] arrays are not. By returning [] from getFiniteTypes() for NAN, the false-branch narrowing is skipped, while the true-branch narrowing still works via the isConstantValue() fallback path.

Test

The regression test verifies that comparing a mixed parameter against multiple class constants containing [NAN] produces no false-positive "always false" errors.

Fixes phpstan/phpstan#11054

[staabm]

Add another test using INF instead of NAN

[phpstan-bot]

Both background tasks confirmed complete. All done — the commit with INF test cases has been pushed to the PR branch.

[VincentLanglet]

I find weird to have not consistency between getScalaraTypes, getScalarValues and getFiniteTypes.
But I made multiple tried and didn't see bad impact on this change. (while changing Scalar method would have)

[mvorisek]

Please add https://phpstan.org/r/45b081b0-e71e-43b2-99cd-a14590ea6687 from the linked issue. After this PR it seems still unfixed, I expect 4x always false reported (as cmp against NAN or array containing NAN is always false).

[VincentLanglet]

Please, as always requested multiples times, don't comment on closed PR and open dedicated issues.
Especially now because the phpstan-bot is mainly working with the snippet of the first message of the issue and solve problems one by one.

We fixed a bug, an wrong "Always false" message.
What you're asking is a feature request, an extra "Always false" message.