Performance regression: switch-heavy array-shape code ~30% slower since the type "reasons" commits (bench bug-7581.php) #14918

Bug report

Performance regression on the 2.2.x dev branch, found by auditing the bench.yml Test job history: analysing switch-heavy code that builds array shapes key by key (bench file tests/bench/data/bug-7581.php) got ~30 % slower on 2026-06-09, and this is the largest single unexplained share of why bug-7581.php currently fails the bench job in every run (currently +63–75 % over the committed baseline on slow runners, +27–40 % even on the fastest runner class).

When it started

The regression window is exactly the "reasons" batch merged between the 11:59 and 12:23 bench runs on 2026-06-09:

• 57ea87c835 Give reason for unresolvable PHPDoc types
• a01080ece4 Explain unresolvable and impossible types with reasons
• e5caec56fa Fold reasons tip into addTip closure

Evidence from the Test job (mode vs. the committed baseline.xml of 2026-05-27; both runs on same-speed runners — the unrelated borderline bench bug-9690.php reads +29.28 % / +28.99 % in the two runs, so runner speed is equal):

No other file moved between those two runs.

Bimodality, then pinned

Right after the batch the effect was bimodal: two fastest-class runners on effectively identical src gave +30.06 % (2026-06-09 19:32, Mo 511.8ms vs. Mo 393.5ms) and −0.05 % (2026-06-10 09:17) — there were no src commits between those runs. That looks like the file sits right at some cliff (union-size / degradation threshold) that run-to-run noise could still tip either way. After ad34ca7395 (phpstan/phpstan-src#5848, tracked separately) landed on 2026-06-11, it has been pinned on the slow side: 44 consecutive failing runs, never once passing on any runner class since.

Local confirmation that this share is separate from phpstan/phpstan-src#5848/phpstan/phpstan-src#5876

In-process bench (median of 5 steady-state iterations, same setup as PHPStan\Benchmark\RegressionBench, PHP 8.5.5): with the later MutatingScope changes from phpstan/phpstan-src#5848 + phpstan/phpstan-src#5876 reverted on current 2.2.x HEAD, bug-7581.php improves 354 ms → 285 ms (−19 %), but is still ~+30 % above the scaled pre-June baseline — matching the share this batch introduced on CI before phpstan/phpstan-src#5848 existed.

The sample produces no errors; all the time goes into NodeScopeResolver/scope merging while $data grows a new constant-array-shape key per case, inside a foreach convergence loop over mixed. A plausible suspect is reason-string construction (AcceptsResult/IsSuperTypeOfResult reasons, the "impossible types" explanation path) now happening in hot paths even when the reasons are never displayed.

Reproduce

# in phpstan-src, 2.2.x
cp tests/bench/data/bug-7581.php test.php
bin/phpstan analyse -l 8 test.php --debug
# or the bench itself:
tests/vendor/bin/phpbench run --file=tests/bench/storage/baseline.xml --report=my-report

<?php

namespace Bug7581;

/** @param mixed[] $parsed */
function parse(array $parsed): void
{
	$data = [];

	foreach ($parsed as $tag => $val):
		switch ($tag):
			case 'A0':
				$data['A0'] = '';
				break;
			case 'A1':
				$data['A1'] = '';
				break;
			// … identical cases A2 through A31 …
			case 'A32':
				$data['A32'] = '';
				break;
		endswitch;
	endforeach;

	echo 'test';
}

Code snippet that reproduces the problem

https://github.com/phpstan/phpstan-src/blob/2.2.x/tests/bench/data/bug-7581.php

Expected output

Analysis time of bug-7581.php back near the committed CI baseline (Mo ≈ 394 ms on GitHub runners, i.e. within the ±25 % bench tolerance) with no change in reported errors. The reasons/tips machinery should be lazy or otherwise free when no error message ends up being rendered.

Related bench regressions keeping the Test job red since 2026-06-11 (this file was later pinned permanently red by the first one): #14920 (phpstan-src#5848 guard check), #14921 (phpstan-src#5876 conditional targets), #14919 (ExprHandler dispatch refactor, same week).