Re: Curing plpgsql's memory leaks for statement-lifespan values

Attached is a draft patch for creating statement-level temporary memory
contexts in plpgsql. It creates such contexts only as needed, and there
are a lot of simpler plpgsql statements that don't need them, so I think
the performance impact should be pretty minimal --- though I've not tried
to measure it, since I don't have a credible benchmark for overall plpgsql
performance.

This fixes the originally complained-of leak and several others besides,
including at least one path that leaked function-lifespan memory even
without an error :-(.

In addition to the patch proper, I attach for amusement's sake some
additional hacking I did for testing purposes, to make sure I'd found and
accounted for all the places that were allocating memory in the SPI Proc
context. There's a glibc-dependent hack in aset.c that reports any
plpgsql-driven palloc or pfree against a context named "SPI Proc", as
well as changes in pl_comp.c so that transient junk created during initial
parsing of a plpgsql function body doesn't end up in the SPI Proc context.
(I did that just to cut the amount of noise I had to chase down. I suppose
in large functions it might be worth adopting such a change so that that
junk can be released immediately after parsing; but I suspect for most
functions it'd just be an extra context without much gain.)

Although this is in principle a bug fix, it's invasive enough that I'd
be pretty hesitant to back-patch it, or even to stick it into HEAD
post-beta. I'm inclined to sign it up for the next commitfest instead.

regards, tom lane