Shadow the parser visitors, ExprPrinter and ClassStatementsGatherer natively

[ondrejmirtes]

Shadows twenty parser visitors, ExprPrinter and ClassStatementsGatherer natively. Apart from the #[ShadowedByTurboExtension] / #[ReferencedByTurboExtension] attributes, the PHP twins are unchanged.

Parser visitors

The node traverser calls every registered visitor once per node. Visitor enterNode()/leaveNode() bodies accounted for 18.1M of the 19.26M native→PHP crossings under PhpParser\NodeTraverser in a self-analysis run, which is 46.5% of all crossings. Most of these bodies only type-check the node and return null.

• Twenty visitors in src/Parser/ are declared under their real names at activation, like any other shadowed class, so a direct $visitor->enterNode($node) still works.
• Each one also registers a pt_native_visitor entry. The native traverser's per-traverse visitor plan resolves it by class entry and calls the C++ function pointer instead of going through the engine. A visitor without an entry takes the engine path as before.
• Only visitors whose enterNode()/leaveNode() always return null and whose beforeTraverse() only resets state qualify, so the fast path doesn't have to duplicate the traverser's return-value protocol. GotoLabelVisitor, AnonymousClassVisitor, PipeTransformerVisitor, StandaloneThrowExprVisitor and UseAliasVisitor stay in PHP.
• tests/parser-visitors.php traverses crafted snippets and real source files four ways (PHP and native visitors, each under the PHP and the native traverser) and requires byte-identical serialized ASTs.

ExprPrinter

The Variable fast path, the read of the node's attribute cache and the cache write happen in C++. Only a genuine cache miss crosses into PHP, straight to Printer::prettyPrintExpr(). The node-key path in support.cpp used to take both fast paths and then call the twin's printExpr(), which re-took them; it now calls the miss half directly, which removes that PHP frame for every expression key the analyser prints.

ClassStatementsGatherer

ClassLikeHandler hands a gatherer to every class body walk as its node callback. The native gatherer forwards each node to the wrapped callback and collects natively. It queries the scope through new MutatingScope direct entries (isInExpressionAssign, isInTrait, isInAnonymousFunction, getFunction) and the existing ScopeContext and ClassReflection ones, so the ~2.7M scope and reflection calls a self-analysis makes from here stay native. The collected value objects are still the PHP classes, built through the class map (16 new keys).

tests/class-statements-gatherer.php replays the node stream of a real walk into both gatherers and compares everything they collect. Self-analysis A/B, 8 interleaved pairs: −1.76% user time, with byte-identical output.

🤖 Generated with Claude Code

https://claude.ai/code/session_017MvPby652L7wUqGAHEiEcN