Resolve a node's file and package dependencies in one pass over unique reflections

[ondrejmirtes]

After every analysed node, FileAnalyserCallback resolves the node's dependencies to the files and Composer packages the result cache records. Profiling bin/phpstan analyse -l 8 src/Analyser/ src/Rules/ src/Type/ --debug showed three sources of avoidable work there:

• The reflections were walked twice. NodeDependencies::getFileDependencies() and getNonAnalysedDependencies() were called back to back on the same list. Both called getFileName() and normalizePath() on every reflection. They are now one method, getFileAndPackageDependencies(). It returns the analysed files, the non-analysed files and the packages in one pass, in the same order as before.
• 55% of the reflections walked were duplicates (917K walked, 415K unique per node). buildClassDependencies() added getInterfaces() again at every parent level, even though it already returns inherited interfaces, and the classes one node references share most of their ancestors. DependencyResolver now keys the reflections by spl_object_id() and merges them with += instead of array_merge(). Deduplicating this way costs nothing: resolveDependencies() takes the same time as before.
• The versioned-extension check (resolveVersionedExtensionPackage(), added in 677df31) called isBuiltin() first for every dependency, which goes through several layers of BetterReflection calls. Only ds and couchbase symbols can match. It now looks the name up in that small map first and calls isBuiltin() only on a hit.

Numbers

Measured on the command above, same session, OPcache on, without the turbo extension.

	Before	After
File/package resolution (both methods / the merged one)	0.72 s (1.39% of the run)	0.37 s (0.71%)
resolveDependencies()	1.18 s	1.18 s
Reflections walked	917K	415K

End to end, over 6 alternating base/head pairs, instructions retired dropped by 1.0% (432.7G → 428.3G). Every head run was below every base run (paired t ≈ −7.8). The wall-time difference (~0.4 s of ~52 s) is within the noise. The analysis output is identical to the base branch.

Verification

• make tests passes. make phpstan and phpcs are clean.
• All result-cache e2e scripts and the ext-ds-version scripts were run locally on the base branch and on this one. Their results match. The few scripts that fail locally fail the same way on the base branch because of the local environment.

🤖 Generated with Claude Code

https://claude.ai/code/session_0147ixdtEww7ewC6Jz6isghx