skink's test_decoding failures in 9.4 branch

I can't help noticing that the failure rate on skink has gone from
"rare" to "100%" since 3d5b227:
http://buildfarm.postgresql.org/cgi-bin/show_history.pl?nm=skink&br=REL9_4_STABLE
I think we need to put some effort into figuring out what's up there.

Also, this morning curculio showed what might be the same issue:
http://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=curculio&dt=2016-07-20%2014%3A30%3A01

I've still had no luck reproducing it here, though.  Wonder if it is
specific to certain compilers.
        regards, tom lane