Messages in this thread | | | Date | Sun, 8 Feb 2015 21:09:27 +0530 | Subject | Re: [PATCH] lib/int_sqrt.c: Optimize square root function | From | Anshul Garg <> |
| |
Dear Mr. linus,
Thanks for quick replies.
Yes performance numbers are not conclusive enough. So its better to discard this patch as of now.
I will try to explore more in this area.
Thanks & regards Anshul Garg
On Fri, Feb 6, 2015 at 1:07 AM, Linus Torvalds <torvalds@linux-foundation.org> wrote: > On Thu, Feb 5, 2015 at 10:43 AM, Anshul Garg <aksgarg1989@gmail.com> wrote: >> >> NOTE :: >> I have not used gcc optimizations while compilation. >> With O2 level optimization proposed solution is taking more time. > > The thing is, the kernel is compiled with -O2, so that's what matters. > > Also, for very tight loops like this, the major costs tend to be very > subtle microarchitectural details, particularly branch prediction. > Which in turn end up sometimes depending on just exactly where the > branches were placed, and even whether two conditional branches were > in the same 8-byte aligned region etc things (because the branch > prediction might be done ignoring the low bits of the EIP etc). So not > only does the exact microarchitecture matter, things that don't *seem* > like they should matter can change behavior a lot. > > My point is really that the performance numbers are very ambiguous. > The patch may well help in some situations, but hurt in others. > > Linus
| |