Cheaper file cache, type memo bookkeeping and expression-key lookups

[ondrejmirtes]

Three performance changes and one bug fix the second of them depends on.

Full self-analysis (build/phpstan.neon, parallel, result cache cleared, disk caches warm), 6 interleaved pairs against 2.3.x: user CPU −3.4 % (6/6, t = −5.8), wall −3.2 %, output identical.

Store file cache entries serialized instead of as var_export()ed PHP

FileCacheStorage wrote every entry as a PHP file returning a var_export()ed CacheItem and loaded it with include. Compiling such a file costs more than unserializing the same data, every object in it is hydrated through a userland __set_state() call, and the compiled code stays in memory for the rest of the process. Entries are now serialize() payloads in *.dat files; the separate extension keeps an older PHPStan sharing the directory from including a payload that is not PHP. clearUnusedFiles() (cache version v3-serialized) deletes the *.php entries of the former formats once.

Single process, src/Analyser src/Rules src/Type, 8 pairs: −2.86 % user CPU, peak RSS 589 MB → 526 MB (−11 %). PHP only - applies without the extension as well.

Keep a type's structural hash and memo keys in a header in front of the object (turbo)

TypeCombinatorCache kept the structural hash of every hashed type and the memo keys each result answers in weak maps - an EG(weakrefs) registration per object and a weak-reference notification when it dies. The shadowing classes it hashes structurally now allocate a 32-byte header in front of the zend_object holding both; PHP subclasses inherit the allocator, free_obj tombstones the object's memo slots and clone_obj gives the clone a fresh header.

Single process, 8 pairs: −1.82 % user CPU.

Do not let a node built from another node's attributes inherit its printed key (bug fix)

ExprPrinter caches an expression's printed key in the node's phpstan_cache_printer attribute, and every node constructed from another node's attributes copied it along - the BooleanAnd/BooleanOr and callable virtual nodes, the single-subject Isset_ nodes of a multi-variable isset(), the plain MethodCall read for an argument-less nullsafe call, the __invoke() call of an invokable callee, the clone-with property fetches, the foreach value assignment and the $this property fetches ClassStatementsGatherer records. Such a node then printed as the node it was made from whenever that one had been printed first, and whatever was tracked for it answered for the original expression: the keep-void type of $foo && $i++; in a loop read mixed instead of bool. The cached key is now dropped from the copied attributes, in PHP and in the native mirrors. InheritedExpressionKeyRuleTest fails without the fix (with and without the extension).

Answer hasExpressionType() without printing a key no tracked expression can match (turbo)

ScopeOps::hasExpressionType() printed the expression's key before looking it up - for a node never printed before that is a trip through the PHP pretty printer, most of the printer misses of an analysis, and most such lookups answer No. A printed key determines the kind of node it was printed from (PHP syntax its node class, every virtual node its own __phpstan marker; a short list shares the array's kind), so when none of the tracked expressions shares the node's kind, the native implementation answers No without printing. Exact since the previous commit.

Single process, 8 pairs: −1.08 % user CPU, walk-trace identical.

Verification

Before the rebase onto the current 2.3.x: make tests with and without the extension (22,237 tests), make phpstan, side-by-side, signature parity, walk-trace identical, make lint-turbo.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QSGnoYacTfJv3taAhZkj5o