[SanderMuller]

BetterReflection 6.73.0.6 (ondrejmirtes/BetterReflection#47) no longer throws CircularReference for an inheritance cycle. A class-like on a cycle has no supertypes of its kind instead. Since 613b1b1, 2.3.x therefore reports nothing for such a cycle. This is the phpstan-src part of the change, with a rule that reports the cycle.

The changes:

• CircularInheritanceRule, at level 0, reports the cycle on each of its members. The identifiers are class.circularExtends, interface.circularExtends and traitUse.circular, and the error cannot be ignored. Examples: Class Foo extends itself. and Class Foo extends Bar, which extends Foo.
• ClassReflection::getParentClass() returns null for a class that extends itself, instead of throwing CircularReference. A class that extends a class on a cycle keeps that parent.
• Inside a trait, TraitUseHandler processes only the traits that reflection says the trait uses. A trait on a trait cycle uses none.
• ClassReflection.cpp and TraitUseHandler.cpp get the same changes.
• When the rule finds a cycle, it tracks a dependency on the other members of the cycle. For a trait declaration, it also tracks the traits it uses. A trait declaration is not analysed on its own, so its file does not depend on the traits it uses otherwise. Without these dependencies, the result cache keeps a stale error after a cycle is broken or closed elsewhere.

The TraitUseHandler change also fixes a crash on 2.3.x. Take a class that uses a trait on a cycle, where the members of one trait use the members of the other one. Analysing it stops with an internal error, MissingPropertyFromReflectionException. The new AnalyserIntegrationTest::testTraitCycleMembers() covers this case. A trait on a cycle uses no traits, so the class gets the members of the first trait only. The test therefore also expects Access to an undefined property.

#6729 removed the part of the result-cache-parse-error e2e step that expected a reflection error for a class that extends itself. The new result-cache-circular-inheritance step checks instead that the result cache is saved when the rule reports a cycle.

Verification:

• CircularInheritanceRuleTest covers:
  • classes on cycles of one, two and three members, and interfaces and traits on cycles of one and two,
  • a class that extends itself with a name in different case,
  • diamonds of interfaces and traits without a cycle,
  • class-likes that extend, implement or use a member of a cycle, which are not reported.
• testTraitCycleMembers() fails on 2.3.x with the internal error.
• The new e2e test result-cache-circular-inheritance breaks a class cycle and closes a trait cycle between two runs, and checks that the cached errors follow. It also checks that the first run saves the result cache.
• make tests passes with and without the Turbo extension (22,636 tests, 74 skipped). make phpstan, phpcs and make lint report no errors.
• Turbo: the strict build passes, smoke.php reports ALL OK, and side-by-side.php, signature-parity.php and walk-trace.php pass. On the rule's fixtures and the cycle fixtures of AnalyserIntegrationTest, the raw output is identical with the extension loaded and not loaded.

Performance on wordpress-develop at level 9, cold runs from source without Turbo, on an M4 Pro. Base is cb3b7e994. Each value is the median of 3 alternating rounds, with the range in brackets. The output is identical, because wordpress-develop has no inheritance cycles.

	Base	This PR
Wall time	34.04 s (33.40 to 34.11)	34.54 s (33.84 to 34.60)
User CPU	309.1 s (308.1 to 309.4)	308.5 s (307.9 to 311.2)

🤖 Generated with Claude Code

[staabm]

when did the phpstan-laravel and the tempest-framework CI check broke? I saw them failling also in others PRs but did not yet saw the root-cause PR (unrelated to this PR - just asking)

[staabm]

did you verify the results of newly added tests in this PR against the state we had before 613b1b1 ?

[SanderMuller]

when did the phpstan-laravel and the tempest-framework CI check broke? I saw them failling also in others PRs but did not yet saw the root-cause PR (unrelated to this PR - just asking)

calebdw/phpstan-laravel#50 is the reason for phpstan-laravel I think

[SanderMuller]

did you verify the results of newly added tests in this PR against the state we had before 613b1b1 ?

Yes. I analysed the PR's cycle fixtures one file at a time, at level 8, on three trees:

• 28f54de, the commit before 613b1b1, with BetterReflection 6.73.0.5. Each file reported Reflection error: Circular reference to class "..." without a line, for the first cycle only. For example, circular-inheritance.php reported only ExtendsSelf.
• 2.3.x at bb2b515. The class and trait cycles report nothing, and trait-cycle-members.php crashes with MissingPropertyFromReflectionException.
• This PR. Every member of every cycle gets the rule error on its line. That is 13 errors in circular-inheritance.php, and one for each class of the e2e fixture. trait-cycle-members.php does not crash.

So each cycle that was reported before 613b1b1 is reported again, now as a rule error with a line.