Re: inline newNode()

Peter Eisentraut wrote:
> Tom Lane writes:
>
> > If you use memset() instead of MemSet(), I'm afraid you're going to blow
> > off most of the performance gain this was supposed to achieve.
>
> Can someone explain to me why memset() would ever be better than MemSet()?

I am surprised MemSet was ever faster than memset(). Remember, MemSet
was done only to prevent excessive function call overhead to memset().
I never anticipated that a simple while() loop would be faster than the
libc version, especially ones that have assembler memset versions, but,
for example on Sparc, this is true.

I looked at the Sparc assembler code and I can't see why MemSet would be
faster. Perhaps someone can send over the Sparc assembler output of
MemSet on their platform and I can compare it to the Solaris assembler I
see here. In fact, they can probably disassemble a memset() routine there
and see exactly what I see.

> Also, shouldn't GCC (at least 3.0 or later) inline memset() automatically?

Not sure, but yes, that may be true. I think it requires a high
optimizer level, perhaps higher than our default.

> What's the result of using -finline (or your favorite compiler's
> inlining flag)?

Yes, that would be it.

> And has someone wondered why the GEQO code needs so many new nodes?
> Perhaps a more lightweight data representation for internal use could be
> appropriate?

I assume the GEQO results he is seeing is only for a tests, and that the
macro version of newNode will help in all cases.

--
Bruce Momjian | http://candle.pha.pa.us
pgman(at)candle(dot)pha(dot)pa(dot)us | (610) 359-1001
+ If your life is a hard drive, | 13 Roberts Road
+ Christ can be your backup. | Newtown Square, Pennsylvania 19073