Rebuild GatheringNodeCallback layers in VirtualAssignNodeCallback instead of wrapping them

[phpstan-bot]

Summary

match ($x) { A => $this->counterA++, B => $this->counterB++ }; used as a statement was reported as Expression "match ($x) {…" on a separate line does not do anything. even though every arm mutates a property.

NoopRule skips the report when the statement contained an assignment — NodeScopeResolver tracks that with a $hasAssign gatherer that watches for PropertyAssignNode/VariableAssignNode. Inc/dec goes through NodeScopeResolver::processVirtualAssign(), which does emit a PropertyAssignNode, but the node arrived at the gatherer after $hasAssign had already been read.

Changes

• src/Analyser/VirtualAssignNodeCallback.php — added a create() factory that recursively rebuilds the callback chain: when the wrapped callback is a GatheringNodeCallback, a new GatheringNodeCallback is returned whose gatherer and inner callback are each filtered separately, instead of one filter wrapped around the whole chain. The constructor is now private.
• src/Analyser/NodeScopeResolver.php — processVirtualAssign() builds its callback with VirtualAssignNodeCallback::create().

Analogous cases fixed by the same change (all verified failing before the fix, passing after, and all covered by the regression test):

• match arms with $this->prop-- (post-decrement)
• match arms with $this->arr['k']++ / $this->arr['k']-- (array-dim virtual assign)
• nested match arms
• $cond && ($this->a++ > 0); → Unused result of "&&" operator.
• $cond || ($this->b-- > 0); → Unused result of "||" operator.
• [$this->a++]; → Expression "[$this->a++]" on a separate line does not do anything.

Probed and found already correct, so no test kept for them:

• and / or / xor statements and ternary statements — these are reported before the hasAssign() check in NoopRule by design (precedence tips), so they are unaffected.
• plain $this->a = 1; and $this->a += 1; inside match arms — AssignHandler/AssignOpHandler emit through the unwrapped node callback, so the gathering layer was never hidden.
• list/array destructuring targets (AssignHandler::prepareTarget() with AssignTargetWalkMode::virtualAssign()) — that path passes $nodeCallback straight through.
• NoopNodeCallback, the only other ShallowNodeCallback — it discards every node, so it has nothing to gather and cannot hide a gatherer.

Root cause

FiberNodeScopeResolver::callNodeCallback() peels GatheringNodeCallback layers off the top of the chain and runs their gatherers synchronously, precisely because "their arrays are read as soon as the enclosing body walk returns". Only the rule-facing remainder may be deferred into a fiber.

VirtualAssignNodeCallback broke that invariant: it wrapped the entire chain — gatherers included — in an opaque ShallowNodeCallback. callNodeCallback() therefore saw no gathering layer, put the whole chain into a fiber, and if the rule handling PropertyAssignNode asked for a type of an expression that had not been walked yet (TypesAssignedToPropertiesRule calls $scope->getType($node->getAssignedExpr())), the fiber parked. It was only resumed at the end of the function body by processPendingFibers() — long after NodeScopeResolver had read $hasAssign and emitted NoopExpressionNode($stmt->expr, false).

The pattern is "a node-callback wrapper that hides gathering layers from the fiber resolver". VirtualAssignNodeCallback was the only wrapper with that shape; the fix makes it preserve the chain structure rather than nest inside it. Because it is the single entry point for every virtual assign, all callers of processVirtualAssign() (inc/dec handlers, unset() of an offset, foreach targets, by-ref function-call writes such as preg_match/array_shift) are fixed at once.

Note that the symptom was rule-set dependent: with only NoopRule registered nothing parks and the bug does not appear, which is why the regression test runs the full rule set.

Test

tests/PHPStan/Analyser/data/bug-15038.php + AnalyserIntegrationTest::testBug15038(). The data file starts with the reporter's playground snippet verbatim, then adds a MoreVirtualAssigns class covering the analogous constructs listed above. The test runs the complete level-8 rule set (so a rule actually parks a fiber, as in the reported scenario) and asserts no errors.

Before the fix the test fails with 8 errors (lines 15, 41, 57, 65, 75, 76, 81); after the fix it passes. make tests, make phpstan and make cs-fix are green, and the affected tests also pass with PHPSTAN_FNSR=0 (fibers disabled).

Fixes phpstan/phpstan#15038

[ondrejmirtes]

Test should be in NoopRuleTest rather than AnalyserIntegrationTest.

[phpstan-bot]

The suite is still running; I'll report as soon as it lands.