Re: Double sorting split patch

On Sat, Sep 17, 2011 at 9:09 PM, Peter Geoghegan <peter(at)2ndquadrant(dot)com> wrote:

> I find it curious that we go to the trouble of
> providing a custom qsort implementation in qsort.c, pg_qsort, but
> haven't gone one step further and considered inlining effects.

I think we provided the qsort implementation for the benefit of
platforms that didn't have a decent one and then ended up switching to
use it always for some reason I don't recall. It wasn't because we
thought we were better at writing qsort implementations than OS
vendors.

> I wondered in passing if compiler vendors had got around to figuring
> out a way of solving the "inlining function pointer" problem that I
> first read about years ago, so I ran this code, which benchmarks the
> macro-based qsort above:

You might need -fipa-pta or some other option. Or maybe LLVM would
have a better chance of pulling this off. Compilers are usually pretty
loath to generate specializations for every call site for fear of
bloating the code.

In any case I don't see how you're going to inline random database
functions. Those are the cases where large amounts of data are being
sorted. It's possible we sort small sets of data for internal reasons
very frequently but I don't recall it turning up at the top of the
profiles posted in recent days.

Now a JIT that actually inlined random database functions into qsort
and optimized the result would be pretty cool. But it doesn't seem
like it's going to happen tomorrow...

--
greg