perf: replay cached dry-run diffs for unchanged files, 9-150x faster warm dry-runs

[SanderMuller]

Built on #8028 (its commit is first; review the last two here). Opening now so the numbers discussion has code attached — happy to rebase once #8028 lands.

Problem

Files with a pending diff are never marked clean in dry-run mode — correct, the diff must keep being reported — so every warm dry-run reprocesses them from scratch: parse, full PHPStan scope resolution, every rule. On real projects most warm time is exactly this. laravel/framework src/Illuminate with the prepared sets has 1,526 pending diffs: a warm dry-run costs the same as a cold one (220s vs 240s single process).

Change

Cache the produced FileDiff keyed on the file's own content hash, the parameter hash and one content hash per captured dependency (the capture from #8028). When everything still matches on the next run, replay the cached diff instead of reprocessing the file — skipping scope resolution entirely. Gist:

// ApplicationFileProcessor::processFile()
if ($useDiffCache) {
    $cachedFileProcessResult = $this->dryRunDiffCache->load($file, $configuration);
    if ($cachedFileProcessResult instanceof FileProcessResult) {
        return $cachedFileProcessResult;
    }
}

Dry-run only: write mode always computes fresh. Selective runs (--only, --only-suffix) bypass the cache entirely. --no-diffs results never cross into normal entries. The original hasChanged flag is replayed, since a rule can report line changes while printing identical content. A failed dependency capture means the file is never cached. The parameter hash is memoized per process, as computing it serializes the whole parameter bag and the cache key needs it per file.

Numbers

corpus	pending diffs	warm dry-run on main	with replay	speedup
laravel/framework src/Illuminate, single process	1,526	220-236s	1.5s	~150x
laravel/framework, parallel (14 cores)	1,526	31s	2.0s	~15x
laravel-queue-insights (public, 85 files)	38	4.6s	0.34s	~13x
hihaho/rector-rules (public, 25 files)	2	1.1s	0.28s	~4x
private 138-file corpus	62	2.7s	0.30s	~9x

Output byte-identical to a fresh run in every cache state, verified per measurement; numbers re-measured on the current minimal #8028 base. Cold cost is the #8028 capture (~7-8% interleaved in its minimal form); the replay itself adds nothing measurable on top. The warm gain scales with how many pending diffs the project has; a fully clean project sees no change.

Verification

Invalidation is covered end-to-end in tests: own-content change, dependency change (fresh-process simulation), --no-diffs cross-replay, the hasChanged flag round-trip. Replay works in parallel mode (workers save, workers replay).

[samsonasik]

Again, please don't do this.

Correctness is more important than speed. Doing this may cause inconsistent result for developers due to runtime check per available extensions in their local dev.

Use phpstan reflection instead, that's the way make it reliable.

[SanderMuller]

Applied — NativeFunctionCallAnalyzer is removed from the stack entirely (here and in #8028). Every node now goes through PHPStan's DependencyResolver, so native functions are resolved by PHPStan reflection like everything else. The skip-layer trade-off (~4% vs ~7-8% cold) can be a separate discussion if it ever comes back as its own PR.

[samsonasik]

Use $this->nodeNameResolver->getName($node) instead.

[SanderMuller]

Applied — the memo key now uses $this->nodeNameResolver->getName($node) (lowercased, as function names are case-insensitive). Measured cost: none (interleaved cold A/B within noise), and verified the namespace-fallback case end to end: two namespaces calling the same unqualified helper() with different resolutions record distinct dependency edges, and editing one helper re-reports only its caller, byte-identical to a fresh run.