Re: [PATCHES] Eliminate more detoast copies for packed varlenas

Gregory Stark <stark@enterprisedb.com> writes:
> This may have been a misdiagnosis of the buildfarm failures but it looks like
> a correct bug fix to me. That is, it looks like regexp_split_to_array() was
> capable of passing a packed varlena to setup_regexp_matches which wasn't
> expecting it. But this I don't understand why it wouldn't cause regression
> failures and indeed when I tested it with my build it didn't cause any
> problems.

The particular regression tests we have for those functions seem to pass
constants to them, not table columns, and so they don't see packed inputs.

(It might be interesting to make textin produce a packed result when
possible, just to see what breaks; but I would be afraid to try to do
that for production...)

> This all brings up the question of what other files should be considered for
> fixing.

I'm very much against such a wholesale edit as you seem to have in mind
here.  We already had some destabilization from the limited patch that
went in; now when we're trying to get to beta is not the time for more.
Maybe at the beginning of 8.4 devel cycle would be a reasonable time
to consider touching a lot of files.
        regards, tom lane