I wrote:
> Hmm, TimestampTzGetDatum is not a no-op on 32-bit machines. If you're
> correct about this, why are our 32-bit BF animals not crashing? Are
> we failing to test this code?
Oh, I had the polarity backwards: this error doesn't result in trying
to dereference something that's not a pointer, but rather in
constructing an extra indirection layer, with the end result being
that the timestamp displayed in the pg_stat_io view is garbage
(I saw output like "1999-12-31 19:11:45.880208-05" while testing in
a 32-bit VM). It's not so surprising that our regression tests are
insensitive to the values being displayed there.
I confirm that this fixes the problem. Will push shortly.
regards, tom lane