Peter Eisentraut wrote:
> Neil Conway writes:
>
> > MemSet(): ~9.6
> > memset(): ~19.5
> > __builtin_memset(): ~10.00
>
> I did my own tests with this code and the results vary wildly between
> platforms. (I do not list __builtin_memset() because the results were
> invariably equal to memset().)
>
> Platform buffer memset() MemSet()
>
> freebsd 32 5.3 4.9
> freebsd 256 23.3 24.2
> mingw 32 0.5 2.0
> mingw 256 6.6 10.5
> unixware 256 15.2 10.3
> unixware 1024 29.2 34.1
> cygwin 256 6.7 15.8
>
> "freebsd" is i386-unknown-freebsd4.7 with GCC 2.95.4.
> "mingw" is i686-pc-mingw32 with GCC 3.2.
> "unixware" is i586-sco-sysv5uw7.1.3 with vendor compiler version 4.1.
> "cygwin" is i686-pc-cygwin with GCC 2.95.3.
>
> GCC was run as 'gcc -O2 -Wall'. (I also tried 'gcc -O3 -finline' but
> that gave only minimally better results.) The SCO compiler was run as
> 'cc -O -Kinine -v'.
>
> Make of those results what you will, but the current cross-over point of
> 1024 seems very wrong.
No question 1024 looks wrong on a lot of platforms. On Sparc, we were
seeing a crossover even higher, and on many like BSD/OS, the crossover
is much lower. Without some platform-specific test, I don't know how we
are going to set this correctly. Should we develop such a test? Do we
have enough MemSet usage in the 256-4k range that people would see a
difference between different MEMSET_LOOP_LIMIT values?
--
Bruce Momjian | http://candle.pha.pa.us
pgman@candle.pha.pa.us | (610) 359-1001
+ If your life is a hard drive, | 13 Roberts Road
+ Christ can be your backup. | Newtown Square, Pennsylvania 19073