Fix array + array inference

[VincentLanglet]

Closes phpstan/phpstan#13561

[phpstan-bot]

This pull request has been marked as ready for review.

[staabm]

I am not sure why this only works on index 0.

in array_merge we iterate over all left arrays

[VincentLanglet]

This is something you could try for array_merge too I think.

Since we're looking only for certainty of hasOffsetValueType YES the key has to be in all the constant arrays, so checking only one constant array is enough.

• only keys in the first constant array can have YES certainty
• key which are not in the first contant array will at most have the MAYBE certainty

[staabm]

do you think it would make sense to adjust array_merge() and array_replace() with this PR, so we have it in sync over all implementations?

[VincentLanglet]

Honestly I'm not sure about it.

Those implementations seems more complicated, especially because there is an iteration

foreach ($argTypes as $argType) {

Also the logic seems to be different since there is some looking at code like

TrinaryLogic::createFromBoolean($argType->hasOffsetValueType($keyType)->yes());

or

if ($hasOffsetValue->yes()) {
					// the last string-keyed offset will overwrite previous values
					$hasOffsetType = new HasOffsetValueType(
						$keyType,
						$offsetType,
					);
				} elseif ($hasOffsetValue->maybe()) {
					$hasOffsetType = new HasOffsetType(
						$keyType,
					);
				} else {
					continue;
				}

which I don't have here.

So I would prefer avoid to delay with PR (with the risk of introducing a regression) and fixing the existing bug, and we could try a dedicated PR with the refacto for array_merge/replace ?

[staabm]

thank you