[0/2] x86: Optimize memchr() for x86-64

Message

Yu-Jen Chang May 28, 2022, 8:12 a.m. UTC

*** BLURB HERE ***
These patch series add an optimized "memchr()" for x86-64 and 
USER-MODE LINUX (UML).
 
There exists an assemebly implementation for x86-32. However, 
for x86-64, there isn't any optimized version. We implement word-wise 
comparison so that 8 characters can be compared at the same time on 
x86-64 CPU. The optimized “memchr()” is nearly 4x faster than the 
orginal implementation for long strings.

We test the optimized “memchr()” in UML and also recompile the 5.18 
Kernel with the optimized “memchr()”. They run correctly.

In this patch we add a new file "string_64.c", which only contains 
"memchr()". We can add more optimized string functions in it in the 
future.

Yu-Jen Chang (2):
  x86/lib: Optimize memchr()
  x86/um: Use x86_64-optimized memchr

 arch/x86/include/asm/string_64.h |  3 ++
 arch/x86/lib/Makefile            |  1 +
 arch/x86/lib/string_64.c         | 78 ++++++++++++++++++++++++++++++++
 arch/x86/um/Makefile             |  2 +-
 4 files changed, 83 insertions(+), 1 deletion(-)
 create mode 100644 arch/x86/lib/string_64.c

Comments

Andi Kleen May 29, 2022, 1:10 a.m. UTC | #1

On 5/28/2022 1:12 AM, Yu-Jen Chang wrote:
> *** BLURB HERE ***
> These patch series add an optimized "memchr()" for x86-64 and
> USER-MODE LINUX (UML).
>   
> There exists an assemebly implementation for x86-32. However,
> for x86-64, there isn't any optimized version. We implement word-wise
> comparison so that 8 characters can be compared at the same time on
> x86-64 CPU. The optimized “memchr()” is nearly 4x faster than the
> orginal implementation for long strings.
>
> We test the optimized “memchr()” in UML and also recompile the 5.18
> Kernel with the optimized “memchr()”. They run correctly.
>
> In this patch we add a new file "string_64.c", which only contains
> "memchr()". We can add more optimized string functions in it in the
> future.

Are there any workloads that care? From a quick grep I don't see any 
that look performance critical.

It would be good to describe what you optimized it for. For example 
optimization for small input strings is quite different than large 
strings. I don't know what is more common in the kernel.

I assume you ran it through some existing test suites for memchr (like 
glibc etc.) for correctness testing?

(bugs in optimized string functions are often subtle, it might be also 
worth trying some randomized testing comparing against a known reference)

-Andi