[westonruter]

The a, A, S, D, l, M, and F format characters of date(), date_format(), and DateTimeInterface::format() now produce unions of their possible values, alongside the numeric single-character formats added in #721. PHP formats these characters from hard-coded English tables in ext/date/php_date.c, so the values do not depend on the locale.

Format	Type
a	'am'|'pm'
A	'AM'|'PM'
S	'nd'|'rd'|'st'|'th'
D	'Fri'|'Mon'|'Sat'|'Sun'|'Thu'|'Tue'|'Wed'
l	'Friday'|'Monday'|'Saturday'|'Sunday'|'Thursday'|'Tuesday'|'Wednesday'
M	'Apr'|'Aug'|'Dec'|'Feb'|'Jan'|'Jul'|'Jun'|'Mar'|'May'|'Nov'|'Oct'|'Sep'
F	'April'|'August'|'December'|'February'|'January'|'July'|'June'|'March'|'May'|'November'|'October'|'September'

Removal of the non-empty-string widening

Adding the cases alone is not enough: getTypeFromFormatType() ended with

if ($type->isNumericString()->no() && $formatType->isNonEmptyString()->yes()) {
	$type = TypeCombinator::union($type, new IntersectionType([
		new StringType(), new AccessoryNonEmptyStringType(),
	]));
}

which widens any non-numeric constant result to non-empty-string, so the new unions came out as non-empty-string. The same happens today for formats without format characters, e.g. date(' ') is non-empty-string rather than ' '.

This block was added in #2888 for phpstan/phpstan#10468 (a non-empty format should give a non-empty result). That guarantee does not depend on it:

• A non-constant format is handled by the count($types) === 0 branch, which already returns non-empty-string for a non-empty-string format.
• For a constant non-empty format, every type buildReturnTypeFromFormat() returns is already non-empty: the explicit cases, ConstantStringType($date), numeric-string, and non-falsy-string.

So the block is removed rather than worked around. The bug-10468.php and bug-6613.php tests still pass.

Tests

• nsrt/bug-15537.php is the reproducer from the issue. It fails before this change (non-falsy-string for every assertion) and passes after.
• nsrt/date-format.php expected non-falsy-string for D; it now expects the weekday union.

Closes phpstan/phpstan#15537

🤖 Generated with Claude Code

[phpstan-bot]

This pull request has been marked as ready for review.

[staabm]

//cc @SanderMuller

[SanderMuller]

Looks good to me. I checked it on 6bf825f.

• The seven unions match what PHP returns. I collected every date() and format() value for a A S D l M F, hourly over two years plus years 1 and 9999. The sets are the same under the C, fr_FR, de_DE and ja_JP locales, on PHP 7.4 and 8.5.
• bug-15537.php and date-format.php fail with the base helper and pass with this one.
• Removing the non-empty-string widening only makes types narrower. With the base helper, date(' ') and date($b ? ' ' : '-') give non-empty-string. With this branch they give ' ' and ' '|'-'. A non-empty non-constant format still gives non-empty-string, and Y, d, u and 0 give the same types as before.
• make tests, make phpstan and phpcs pass. The class is not shadowed by Turbo.
• nextras/orm also fails on #6861 with the same two Nette DI errors, so it is not from this PR.

Performance: I timed a synthetic file with 200 functions and 4000 date format calls, 10 interleaved rounds, user CPU medians. When the calls are only evaluated, the analysis takes 2.07 s instead of 2.02 s. When the results are concatenated, it takes 2.53 s instead of 2.24 s. Most of that is the concat of the new unions, which already happens today for small numeric unions such as date('L') . date('N'). A real-world codebase of about 35000 files has 62 calls with these format characters, so I expect no measurable difference in practice.

[staabm]

I am not allowed to change the PR. @westonruter please pick the following commit onto the PR: staabm@613f7e6

-> found a way arround it

after that we are ready to go

[staabm]

thank you!