I wrote: > ... or maybe not just yet. Andres' buildfarm critters seem to have > a different opinion than my machine about what the output of > collate.icu.utf8 ought to be. I wonder what the prevailing LANG > setting is for them, and which ICU version they're using.
Oh, I bet it's "C.utf8", because I can reproduce the failure with that. This crystallizes a nagging feeling I'd had that you were misdescribing the collate.icu.utf8 test as not being run under --no-locale. Actually, it's only skipped if the encoding isn't UTF8, not the same thing. I think we need to remove the default-collation cases from that test too. regards, tom lane