[staabm]

inspired by PHP-CS-Fixer/PHP-CS-Fixer#9924

benchmark strpbrk() with trim/preg_match():

➜  phpstan-src git:(pr/6954) ✗ php tests/bench/micro/fnmatch-pattern.php        
PHP 8.5.12; PCRE JIT: 1; 200000 iterations, median of 5 rounds
Input                 regex ns/op strpbrk ns/op    speedup
no wildcard                  42.3         35.5      1.19x
early wildcard               32.6         30.9      1.05x
late wildcard                39.5         36.3      1.09x
bracket pattern              37.5         33.0      1.14x
long no wildcard            108.0         86.4      1.25x
long late wildcard          106.2         91.0      1.17x

➜  phpstan-src git:(2.3.x) ✗ php tests/bench/micro/trim-digits.php   
PHP 8.5.12; PCRE JIT: 1; 200000 iterations, median of 5 rounds
Input                 regex ns/op strpbrk ns/op    speedup
empty                        32.8         26.7      1.23x
whitespace                   34.8         37.5      0.93x
no digit                     34.9         29.2      1.20x
early digit                  34.2         34.1      1.00x
late digit                   36.1         33.8      1.07x
trim range                   34.2         31.6      1.08x
long no digit                82.6         76.9      1.07x
long late digit              84.4         84.1      1.00x

[staabm]

//cc @SanderMuller

[SanderMuller]

Both replacements give the same result as the old regexes. I compared them on 201,025 strings: every byte value alone, doubled and next to a letter, and random strings of up to 10 characters. The random strings used digits, the 4 wildcard characters, \xB2, an Arabic-Indic digit and NUL. There was no difference, also with the UTF-8 ctype locale that the CLI sets.

On the PR head, FileExcluderTest, the trim tests in NodeScopeResolverTest, phpcs and make phpstan pass. The 3 red integration jobs are also red on #7007. Mutation Testing (8.4) did not run, because the download of infection got a 504 Gateway Time-out.

The run-time effect is very small. Self-analysis reaches the changed checks 1,591 times. With the largest difference in your tables (about 22 ns), that is less than 0.1 ms per run.

A separate problem that is older than this PR: trim() accepts a range like '/..:', which includes all digits without a digit character. So for a numeric-string $s, PHPStan types trim($s, '/..:') as numeric-string, but trim('12', '/..:') is ''. The old regex has the same gap. I can open an issue for it.

[staabm]

I can open an issue for it.

yes, please open a issue

[SanderMuller]

Opened phpstan/phpstan#15683. I also found that chop() on a constant string removes characters from the start, so I opened phpstan/phpstan#15684 for it.