[alfredbez]

In RemoveUnusedPrivateMethodRector, IsClassMethodUsedAnalyzer::isClassMethodUsed() (which traverses the class AST) was invoked for private methods before checking whether the method is a PHPUnit data provider or target of a class-string static call (self::class::method()). In addition, ScopeFetcher::fetch($node) was called repeatedly on the same Class_ node for each private method.

Reproducer

use PHPUnit\Framework\Attributes\DataProvider;
use PHPUnit\Framework\TestCase;

final class Subject extends TestCase
{
    #[DataProvider('provideData')]
    public function testSomething(int $value): void
    {
        self::assertSame($value, self::class::computeExpected());
    }

    private static function provideData(): array
    {
        return [[1]];
    }

    private static function computeExpected(): int
    {
        return 1;
    }
}

Expected Behavior

RemoveUnusedPrivateMethodRector should check known-used method name lists (dataProviderMethodNames and classStringCallMethodNames) before running AST usage analysis, and fetch the class scope lazily at most once per class node.

Before / After

bench/remove_unused_private_method_bench.php (included in this PR): 150 TestCase classes, each with a private data provider, a private method called via self::class::method() and a locally used private method; 100 repeats = 15,000 refactor() calls, parsing and warmup excluded, CPU via getrusage().

	Before (e27ea90, parent)	After (a7b0485, this PR)
CPU, median of 3	7.26 s	1.76 s
CPU per call	484 us	117 us
Single runs	6.86 / 7.48 / 7.26 s	1.64 / 1.76 / 1.91 s
Process peak RSS	~120 MiB	~120 MiB

-75.7% CPU (~4.1x) for this case. Memory is unchanged.

Both revisions pass the same rule tests (49 tests, 51 assertions), including the new mixed_protected_and_unused_methods fixture. After each run, the script checked that all 150 classes were left unchanged.

Command, run in each checkout with the same vendor/ and the same bench file:

php -d memory_limit=512M -d opcache.enable_cli=0 bench/remove_unused_private_method_bench.php

Environment: PHP 8.4.27 CLI (NTS), Docker limited to 1 CPU / 768 MiB, AMD EPYC-Rome VPS. Fresh process per run, base/head order alternated.

The rule file, fixture and bench script in a7b0485 are identical to the current head after the rebase. A rerun on current main (e60b5ef) vs head (5dd7075) gave the same ratio: 6.86 s -> 1.66 s CPU (median of 5).

This measures direct rule calls on classes where the early name checks apply. It is not an end-to-end project runtime.

[TomasVotruba]

Reorder is behavior-neutral and the win looks real (name-list checks now gate isClassMethodUsed(), which traverses the whole class AST, plus scope fetched once per class instead of per method).

Could you add a quick before/after number though? This is the only one of the three perf PRs touching the hot traversal path, yet it ships no bench. A small bench/*.php micro-script (LazyContainerFactory + TestingParser, getrusage CPU, warmup excluded) on both branches would lock it in - same shape as the one in #8801.

[TomasVotruba]

Could you update PR descripotion with before/after bench time?

[alfredbez]

Could you update PR descripotion with before/after bench time?

Done ✅

[TomasVotruba]

Thanks 👍