Re: [PATCH] x86_64/memset: use "small block" code for blocks up to 30 bytes long

[Rich Felker <dalias@libc.org>]

On Fri, Feb 13, 2015 at 05:39:49PM +0100, Denys Vlasenko wrote:
> Before this change, we were using it only for 15-byte blocks and smaller.
> Measurements on Sandy Bridge CPU show that "rep stosq" setup time
> is high enough to dominate speed of fills well above that size:

I just ran some tests with the latest three patches, including this
one, and aside from the significant improvement for sizes 16-30 in
this last patch, each patch makes various sizes 2-15 mildly slower. A
few of them go from 9 cycles to 11 cycles average; most just increase
by 1 cycle. And all sizes 8 and under are still slower than the C
code.

I don't think these small sizes are a typical deliberate usage case
for memset, but even just a couple cycles is a large relative
difference at such sizes, and they could arise in generic code (think
something like qsort which uses memcpy possibly with small sizes, but
something using memset instead) and make a significant performance
difference.

The main change whose value I really question is the conditional
widen_rax. If the value isn't used until a few cycles after the imul
instruction, doing it unconditionally is probably cheaper than testing
and branching even when the branch is predictable.

Rich