Andres Freund <andres@2ndquadrant.com> writes:
> On 2013-11-11 12:31:55 -0500, Robert Haas wrote:
>> I seem to recall that our %m support involves rewriting the error
>> string twice, which I think is actually kind of expensive if, for
>> example, you've got a loop around a PL/pgsql EXCEPTION block.
> Yes, it does that. Is that actually where a significant amount of time
> is spent? I have a somewhat hard time believing that.
I don't see any double copying. There is *one* copy made by
expand_fmt_string. Like Andres, I'd want to see proof that
expand_fmt_string is a significant time sink before we jump through
these kinds of hoops to get rid of it. It looks like a pretty cheap
loop to me. (It might be less cheap if we made it smart enough to
identify 'z' flags, though :-()
>> I'd
>> actually like to find a way to get rid of the existing %m support,
>> maybe by having a flag that says "oh, and by the way append the system
>> error to my format string"; or by changing %m to %s and having the
>> caller pass system_error_string() or similar for that format position.
The first of these doesn't work unless you require translations to
assemble the string the same way the English version does. The second
would work, I guess, but it'd sure be a pain to convert everything.
regards, tom lane