Re: musl libc, memcpy

[Rich Felker <dalias@aerifal.cx>]

On Tue, Jul 31, 2012 at 12:19:13AM +0200, Kim Walisch wrote:
> > I'd like to know what block sizes you were looking at, because for
> > memcpy that makes all the difference in the world:
> 
> I copied blocks of 16 kilobytes.

OK, that sounds (off-hand) like a good size for testing.

> > I don't think this is necessary or useful. If we want better
> > performance on these archs, a tiny asm file that does almost nothing
> > but "rep movsd" is known to be the fastest solution on 32-bit x86, and
> > is at least the second-fastest on 64-bit, with the faster solutions
> > not being available on all cpus. On pretty much all other archs,
> > unaligned access is illegal.
> 
> My point is that your code uses byte (char) copying for unaligned data
> but on x86 this is not necessary. Using a simple macro in your memcpy
> implementation that always uses the size_t copying path for x86 speeds
> up your memcpy implementation by about 500% for unaligned data on my
> PC (Intel i5-670 3.46GHz, gcc-4.7, SL Linux 6.2 x86_64). You can also
> use a separate asm file with "rep movsd" for x86, I guess it will run
> at the same speed as my macro solution.

I'm attaching a (possibly buggy; not heavily tested) rep-movsd-based
version. I'd be interested in hearing how it performs.

> Another interesting thing to mention is that gcc-4.5 vectorizes the 3
> copying loops of your memcpy implementation if it is compiled with the
> -ftree-vectorize flag (add -ftree-vectorizer-verbose=1 for
> vectorization report) but not if simply compiled with -O2 or -O3. With

Odd, the gcc manual claims -ftree-vectorize is included in -O3:

http://gcc.gnu.org/onlinedocs/gcc/Optimize-Options.html

> $ gcc -O2 -ftree-vectorize -ftree-vectorizer-verbose=1 memcpy.c main.c -o memcpy
> 
> memcpy.c:25: note: created 1 versioning for alias checks.
> memcpy.c:25: note: LOOP VECTORIZED.
> memcpy.c:21: note: created 1 versioning for alias checks.
> memcpy.c:21: note: LOOP VECTORIZED.
> memcpy.c:9: note: vectorized 2 loops in function.

From the sound of those notes, I suspect duplicate code (and wasteful
conditional branches) are getting generated to handle the possibility
that the source and destination pointers might alias. I think this
means it would be a good idea to add proper use of "restrict" pointers
(per C99 requirements) in musl sooner rather than later; it might both
reduce code size and improve performance.

Rich