[ondrejmirtes]

Every PHPStan process, including each parallel worker, kept about 10 MB more than it needs from the first symbol lookup on. Bisecting pointed to three caches, and this PR changes how each one is stored. Each speedup stays:

• FileHelper's normalized path cache (from NodeDependencies: cache normalized filenames #4815). It kept an array per original path plus a second copy of the path, about 500 bytes per entry. FileFinder sends every found file through it, so on Sylius it held 6,653 paths (3.2 MB) before analysis started. It is now keyed by the directory separator first, and a path that normalizes to itself, which is nearly every path, is stored once as both key and value. Those 6,653 paths now take 0.3 MB, and cached lookups are slightly faster (0.46 ms vs 0.58 ms per pass over 6,653 paths).
• OptimizedDirectorySourceLocatorFactory's $checkedFiles. It kept every checked file's hash, stat signature and symbols for the life of the process: 5.4 MB for 3,810 files in every Sylius worker. A directory is scanned only once per process anyway (OptimizedDirectorySourceLocatorRepository), so it is now local to one scan. A locator whose files are all unchanged gets filled as soon as its cache entry is checked. On a warm run, that means only one directory's cache entry is in memory at a time instead of all of them. The files that were hashed or scanned are still shared by the locators of one scan. A file that two locators reach and that its stat signature vouches for costs one extra stat().
• DirectoryWalker's listings. Every listing ever walked was loaded on the first walk and kept: 2.3 MB for Sylius's 3,067 directories, in the main process and in every worker. Each walked directory now has its own listings file, loaded for that walk and released after it.

No caps or limits were added. None of the changed classes is turbo-shadowed. The C++ NodeScopeResolver only calls back into the PHP FileHelper::normalizePath(), so turbo-ext needs no change.

Measurements

Sylius (level 6, its own phpstan.neon.dist with phpstan-doctrine, phpstan-symfony and phpstan-webmozart-assert, extension versions that every PHAR below accepts), PHP 8.4, 16 cores, no turbo extension, warm caches, result cache cleared before each run. Best of 5. Peak memory is memory_get_peak_usage() recorded at shutdown of each process; max RSS and wall time come from /usr/bin/time. The 2.3.x-dev and this-PR PHARs were built with compiler/bin/prepare + box, as in phar.yml.

One-line file (function foo(mixed $c): mixed { strlen("a"); return $c; }, analysed with one spawned worker)

	2.1.34	2.3.0	2.3.x-dev	this PR
peak memory, worker	101.42 MB	74.09 MB	73.97 MB	63.42 MB
peak memory, main process	82.96 MB	55.97 MB	55.97 MB	51.19 MB
max RSS	134.6 MB	167.6 MB	168.4 MB	158.5 MB
wall time	1.96 s	2.14 s	2.15 s	2.14 s
peak memory, from source, --debug (one process)	24.87 MB	–	37.63 MB	27.01 MB

Full src/ analysis (16 workers)

	2.1.34	2.3.0	2.3.x-dev	this PR
wall time	8.59 s	6.28 s	6.95 s	6.18 s
peak memory, largest worker	213.40 MB	158.47 MB	162.79 MB	146.94 MB
peak memory, main process	131.56 MB	95.91 MB	95.91 MB	91.11 MB
max RSS	240.7 MB	265.1 MB	269.2 MB	252.6 MB

In PHAR mode, 2.1.34 uses more heap than 2.3 because its PHAR members have mtime 0. OPcache refuses to cache such members, so all the compiled code stays in the heap; since then resign.php stamps the members, and 2.3 enables OPcache in its workers. The source-mode row (with opcache.enable_cli=1) compares the analyser itself: this PR brings it back to within about 2 MB of 2.1.34.

Breakdown of what the one-line source-mode run keeps after the scan, on 2.3.x-dev: $checkedFiles 5.4 MB, FileHelper cache 3.2 MB (now 0.3 MB), directory listings 2.3 MB.