I have found the term that was in my mind, it is Old Norse svilar, and here is the quote from Hans Henrich Hock: Principles of historical linguistics, p. 836
For instance, there
is just one Germanic language, Old Norse, which has the word svilar
referring to ‘brothers-in-law whose wives are sisters of each other’.
Occam’s Razor would prevent us from reconstructing this form for
Proto-Germanic. If further reconstruction had to proceed from
Proto-Germanic, without consideration of its individual daughter
languages, we would miss the connection of svilar with the dialectal
Greek aelioi, a word with the same meaning. We would therefore fail to
reconstruct the PIE word *swelio- from which both the Norse and the
dialectal Greek words can be derived. Such an approach, however, would
be patently wrong, for the kinship term in question is of such highly
specialized reference that the correspondence is not likely to reflect
borrowing (or chance). And since Greek and Germanic otherwise are not
closely related members of the Indo-European family, common innovation
is unlikely. Under the circumstances, reconstruction seems to be the
only acceptable solution.
It's other attestation is from Greek, not Iranian, and the dialectal part is on the Greek side, not the Germanic (Old Norse) side; this part of my memory was wrong.