Keep empty constant arrays out of the oversized merge in TypeCombinator::optimizeConstantArrays()

[phpstan-bot]

Summary

The bug-12671.php benchmark went from well under a second to ~26s. Bisecting pointed to e606ca2 ("Merge unions of constant arrays with many keys instead of degrading"). That commit exposed an existing weakness in TypeCombinator::optimizeConstantArrays(): a single array{} member kept the generalized oversized arrays from being merged into one. The union then compared hundreds of generalized arrays pairwise, each with a huge item-type union.

With this change, analysing the bench file takes ~1.7s (0.8s with the turbo extension) instead of ~18.5s locally.

Changes

• src/Type/TypeCombinator.php (optimizeConstantArrays()): empty constant arrays are set aside before the oversized generalization. The remaining members are merged as before ($eachIsOversized is no longer spoiled by array{}), and the empty arrays are put back next to the result.
• turbo-ext/src/TypeCombinator.cpp: same change in the native mirror. The regression test fails with the old native build and passes with the new one. Analysis output of the bench file is identical with and without the extension.
• Probed analogous paths:
  • The general-array branch of processArrayTypes() goes through the same function, so it is covered.
  • The inner value traversal in the same method already skips empty arrays.
  • Non-constant array members still prevent the merge. That is intended, since their item types are not generalized.

Root cause

The oversized generalization turns every non-empty ConstantArrayType into non-empty-array<K, V>&oversized-array. When all members were generalized, it merges them into one array. Empty arrays are deliberately not generalized (array{}&oversized-array would be contradictory), but that also left $isOversized false for them, so $eachIsOversized became false and the method returned every generalized member separately.

In the benchmark, $pr_states[$country] over a country→states map (many countries map to []) produced ~250 such members. Their item types are unions of list{code, name, localName}&oversized-array. TypeCombinator::union() then ran compareTypesInUnion() over them pairwise, and every comparison was a union-vs-union isSuperTypeOf(). That made about 557k nested isSuperTypeOf calls for a single offset fetch.

Before e606ca2, a hidden 62-distinct-key cap degraded such unions to a general array before reaching this code, which masked the problem.

The result is also more consistent now: array{}|X is exactly array{} plus what the union gives without the empty member. Previously the pairwise path relied on the intentionally lossy "oversized intersection is a subtype when maybe" rule in IntersectionType::isSubTypeOf(). For example, it absorbed Argentina's states into Spain's array and dropped them from the inferred type.

Test

tests/PHPStan/Analyser/nsrt/bug-15293.php builds a constant map of three entries with disjoint keys (together over ConstantArrayTypeBuilder::ARRAY_COUNT_LIMIT values) plus one [] entry. It asserts that an unknown-key fetch yields array{}|(non-empty-array<…, list{…}&oversized-array>&oversized-array), identical to the map without the empty entry. Without the fix, the first assertion fails with a union of separate list{n, n, n}&oversized-array members. The existing tests/bench/data/bug-12671.php covers the performance side.

Fixes phpstan/phpstan#15293

🤖 Generated with Claude Code