[ondrejmirtes]

Since 9b24ee5 ("Do not use PHPDoc type of overridable class constants as their native type", first in 2.3.3), getClassConstFetchTypeByReflection() caches the native-types pass under a separate <Class>::<CONST>#native key. As a result, a fetch like self::X or Foo::X resolves and transforms the constant's initializer twice, once for each pass. For huge constant arrays this costs a lot. Bisecting with instruction counts showed that bug-12671.php, bug-12159.php and bug-10538.php in tests/bench/data got slower at this commit.

In this branch (a fetch on a class name, not on an object), the PHPDoc type is the only input that differs between the two passes. ConstantResolver::resolveClassConstantType() only uses it for a dynamicConstantNames list entry without a native type.

Change

• On a cache miss, the resolver now reads getConstantPhpDocType() once.
• When the constant has no PHPDoc type, the resolved type is stored under both keys, the plain one and #native, so the second pass hits the cache.
• Constants with a PHPDoc type keep separate entries as before, so the behaviour from 9b24ee5 is unchanged.
• A cache hit is still a single array lookup.
• The native mirror in turbo-ext/src/InitializerExprTypeResolver.cpp gets the identical change.

Instruction counts

Net instructions = cachegrind count of the file minus the count of an empty file at the same commit, so PHPStan's boot is excluded. Measured with PHP 8.4.27 at level 8 with bleedingEdge, without the turbo extension.

file	2.3.x (bcca2b0)	this PR (e178292)	change
bug-12671.php	14.97 G	10.24 G	−31.6 %
bug-12159.php	3.63 G	3.09 G	−14.9 %
bug-10538.php	20.43 G	17.19 G	−15.9 %

For reference, on 2f22669, the parent of 9b24ee5, the counts are 10.25 G, 2.83 G and 17.09 G. bug-12671 and bug-10538 are back to that level. The remaining ~0.26 G on bug-12159 comes from later commits, not from 9b24ee5.

Wall-clock times

These are medians of 9 runs after 2 warm-up runs (hyperfine) of bin/phpstan analyse -c <config> --debug --no-progress <file> on 2.3.x (bcca2b0) and on this branch. The repository has no tests/bench/phpstan.neon, so the config is the equivalent of tests/bench/BenchCase.php: it includes conf/bleedingEdge.neon and sets level: 8 and inferPrivatePropertyTypeFromConstructor: true.

Machine: Amp orb (E2B sandbox), Intel Xeon @ 2.60GHz, 16 vCPUs. PHP 8.4.27 NTS without the turbo extension.

file	2.3.x	this PR	change
bug-12671.php	2.457 s	1.988 s	−19.1 %
bug-12159.php	1.311 s	1.246 s	−4.9 %
bug-10538.php	3.294 s	2.929 s	−11.1 %

Tests

• make tests passes without the extension. The only failures were two DirectoryWalkerTest cases that failed only under parallel load; that test passes when run alone and is unrelated to this change.
• make tests passes with phpstan_turbo.so loaded.
• turbo-ext/tests/smoke.php prints ALL OK, and turbo-ext/tests/signature-parity.php passes.
• turbo-ext/bin/side-by-side.php pairs InitializerExprTypeResolver. Some generated headers are reported stale; that comes from the PHP 8.4 environment and also happens without this change.
• phpcs (build-cs) on src/Reflection/InitializerExprTypeResolver.php is clean, and self-analysis of that file with build/phpstan.neon finds no errors.
• For the three benchmark files, the errors are identical before and after the change, and identical with and without the extension.
• tests/PHPStan/Analyser/nsrt/class-constant-native-type.php from 9b24ee5 still passes.