Re: Progress on fast path sorting, btree index creation time

On 5 January 2012 22:27, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> There is no compiler anywhere that implements "always inline", unless
> you are talking about a macro.  "inline" is a hint and nothing more,
> and if you think you can force it you are mistaken.  So this controversy
> is easily resolved: we do not need any such construct.

I'm slightly puzzled by your remarks here. GCC documentation on this
is sparse (although, as I've demonstrated, I can get better numbers
using the always_inline attribute on GCC 4.3), but take a look at
this:

http://msdn.microsoft.com/en-us/library/z8y1yy88.aspx

While it is not strictly true to say that pg_always_inline could
inline every function under every set of circumstances, it's pretty
close to the truth. I do accept that this facility could quite easily
be abused if its use isn't carefully measured. I also accept that
C99/GNU C inline functions are generally just requests to the compiler
that may be ignored (and indeed the compiler may inline without being
asked to).

It's short sighted to see this as a case of inlining itself making a
big difference, so much as it making a big difference as an enabling
transformation.

> The real question is whether we should accept a patch that is a
> performance loss when the compiler fails to inline some reasonably
> simple function.  I think that would depend on the actual numbers
> involved, so we'd need to see data before making a decision.

Who said anything about a performance loss? Since the raw improvement
to qsort_arg is so large, it's pretty difficult to imagine a
confluence of circumstances in which this results in a net loss. See
the figures at http://archives.postgresql.org/pgsql-hackers/2011-11/msg01316.php
, for example.

The pg_always_inline idea is relatively recent. It just serves to
provide additional encouragement to the compiler to inline, insofar as
that is possible on the platform in question. The compiler's
cost/benefit analysis cannot possibly appreciate how hot a code path
qsort_arg is, because it has a set of generic heuristics that are
quite fallible, and very probably are on quite conservative. We can
appreciate such things though.

--
Peter Geoghegan       http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training and Services