Treat first-class callable NullsafeMethodCall as a MethodCall instead of throwing

[phpstan-bot]

Summary

Using the first-class callable syntax together with the nullsafe operator
($foo?->bar(...)) crashed PHPStan with Internal error.. While this is a
fatal error in PHP itself ("Cannot combine nullsafe operator with Closure
creation"), PHPStan must not crash on it. This change makes the analyser treat
such an expression as a regular method-call first-class callable.

Changes

• src/Analyser/NodeScopeResolver.php: in the first-class callable branch of
  processExprNode(), handle Expr\NullsafeMethodCall by building a
  MethodCallableNode from an equivalent MethodCall, instead of falling
  through to the throw new ShouldNotHappenException().
• tests/PHPStan/Analyser/nsrt/bug-9746.php: regression test asserting the
  inferred types and (implicitly) that no internal error is thrown.

Root cause

processExprNode() enumerated the Expr\CallLike subclasses that support the
first-class callable syntax — FuncCall, MethodCall, StaticCall, New_ —
and threw ShouldNotHappenException for anything else. NullsafeMethodCall is
also an Expr\CallLike whose isFirstClassCallable() can be true (the parser
accepts $foo?->bar(...) even though PHP rejects it at runtime), so it hit the
else branch and crashed. It was the only Expr\CallLike subclass not covered.

Test

• tests/PHPStan/Analyser/nsrt/bug-9746.php reproduces the original crash via
  $foo?->doFoo(...) (nullable and non-nullable receivers, plus an assignment)
  and asserts (Closure(): int)|null / Closure(): int. It errors with an
  internal error before the fix and passes after.
• Verified manually that the expression no longer crashes in any context
  (statement, assignment, call argument, array element, return).

Analogous cases probed

• The other first-class callable Expr\CallLike subclasses (FuncCall,
  MethodCall, StaticCall, New_) were already handled; NullsafeMethodCall
  was the only missing one.
• MutatingScope::getKeepVoidType() already lists NullsafeMethodCall in its
  first-class-callable guard, and type resolution flows through
  NullsafeMethodCallHandler, so no internal error occurs when the expression
  is typed.

Fixes phpstan/phpstan#9746

[staabm]

Crash bug fixes need a AnalyserIntegrationTest

[phpstan-bot]

Done. I addressed staabm's review by adding the requested AnalyserIntegrationTest coverage for the crash fix.

What I did

1. New data file tests/PHPStan/Analyser/data/bug-9746.php — reproduces the original crash path (the first-class callable NullsafeMethodCall used as a statement, which is exactly what hit CallToMethodStatementWithoutSideEffectsRule → MutatingScope::resolveType() → ShouldNotHappenException in the issue's stack trace).

2. New test testBug9746() in AnalyserIntegrationTest — runs the full analyser over the file and asserts the three resulting errors, confirming analysis completes without an internal error.

Verification

• With the fix: test passes (10 assertions).
• Without the fix (temporarily reverted the committed NullsafeMethodCall branch): the test errors with the ShouldNotHappenException internal error — confirming it genuinely catches the crash.
• make phpstan: No errors.

Committed as a2149b421 and pushed to create-pull-request/patch-3pnp41f. The existing nsrt/bug-9746.php type-inference test remains as-is; this adds the integration-level crash regression the reviewer asked for.

[staabm]

we expect a Cannot combine nullsafe operator with Closure creation error on this line

[phpstan-bot]

I've addressed staabm's review. Here's what I did:

The review

staabm pointed out that for $self?->sayHello(...) PHPStan should report a Cannot combine nullsafe operator with Closure creation error (matching PHP's fatal error), not just silently treat it as a valid method-call first-class callable.

Changes

1. src/Node/MethodCallableNode.php — added a nullsafe flag (default false) plus an isNullsafe() accessor, so the node can record that it originated from a $foo?->bar(...) expression.

2. src/Analyser/NodeScopeResolver.php — the NullsafeMethodCall first-class-callable branch now passes nullsafe: true when building the MethodCallableNode.

3. src/Rules/Methods/MethodCallableRule.php — when the node is nullsafe, it reports a non-ignorable error Cannot combine nullsafe operator with Closure creation. (identifier nullsafe.firstClassCallable) and skips the normal method-existence check.

4. Tests:

   • New tests/PHPStan/Rules/Methods/data/method-callable-nullsafe.php + testNullsafe() in MethodCallableRuleTest — direct rule-level coverage of the new error.
   • Fixed testBug9746() in AnalyserIntegrationTest, which was inconsistent with its data file (it referenced a Bug9746Integration namespace/file that no longer existed). It now asserts the actual two errors, including the nullsafe one.

Verification

• testNullsafe fails without the source change (it instead reports Cannot call method doFoo() on Foo|null), confirming the test catches the regression; passes with the fix.
• All affected test classes pass (197 tests).
• make phpstan: No errors.

Committed as d31bdac35 and pushed.

[staabm]

instead of adding a separate $nullsafe bool , can we instead carry this information in $originalNode with a instanceof NullsafeMethodCall check?

[phpstan-bot]

Done. I addressed staabh's review by removing the separate $nullsafe bool and carrying the information in $originalNode itself.

Changes

1. src/Node/MethodCallableNode.php — Dropped the private bool $nullsafe constructor parameter. The $originalNode field is now typed Expr\MethodCall|Expr\NullsafeMethodCall, and isNullsafe() derives the flag via $this->originalNode instanceof Expr\NullsafeMethodCall. getOriginalNode()'s return type widened accordingly.

2. src/Analyser/NodeScopeResolver.php — The NullsafeMethodCall first-class-callable branch now passes the original NullsafeMethodCall straight into MethodCallableNode instead of building a synthetic MethodCall and a true flag.

3. src/Node/ClassStatementsGatherer.php — The one consumer that needs a concrete MethodCall (feeding the @api Method\MethodCall collector, which doesn't accept NullsafeMethodCall) now converts the nullsafe node locally. This keeps the @api class untouched. The only other getOriginalNode() consumer (Printer) handles NullsafeMethodCall natively.

Verification

• testNullsafe (rule-level) and testBug9746 (integration crash regression) both pass.
• UnusedPrivateMethodRuleTest and CallToMethodStatementWithoutSideEffectsRuleTest pass (other consumers of the gathered method calls).
• make phpstan: No errors.

This is a behavior-preserving refactor — the fix itself was already committed in d31bdac35; this only changes how the nullsafe flag is represented. Committed as acc872106 and pushed.