diff --git a/en_US.ISO8859-1/htdocs/news/status/report-2014-07-2014-09.xml b/en_US.ISO8859-1/htdocs/news/status/report-2014-07-2014-09.xml index de0f4223bd..58919fb42d 100644 --- a/en_US.ISO8859-1/htdocs/news/status/report-2014-07-2014-09.xml +++ b/en_US.ISO8859-1/htdocs/news/status/report-2014-07-2014-09.xml @@ -1019,4 +1019,47 @@ commit bits were taken in for safe keeping.

+ + + SSE variants of libc routines for amd64 + + + + + John + Baldwin + + jhb@FreeBSD.org + + + + + + + + +

I have written SSE/AVX-optimized versions of a few libc + routines for amd64. So far the list includes memcpy, + memset, and strlen. For each routine I have + written a simple regression test as well as performed some + simple microbenchmarks on various AMD and Intel CPUs.

+ +

The simplest routine is strlen which appears to be a + general win in microbenchmarks. memcpy and + memset have proven trickier as different variants can + behave quite differently on different CPUs.

+ +

At present, I do not yet have a patch relative to libc. Once + I do, this will be suitable for more testing. I'd like to see + some real-world benchmarks that show measurable improvement + before pushing any of this up into the tree.

+ + + + +

Create a branch that holds a modified libc and is suitable + for testing

+ + +