This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH i386 AVX512] [81/n] Add new built-ins.


On Wed, Oct 22, 2014 at 03:34:22PM +0400, Kirill Yukhin wrote:
> > Can you test with -mavx512 (or whatever enables the builtins?)
> Done.
> I did:
> 	sync && time for i in `seq 1000` ; do ./build-x86_64-linux/gcc/xgcc -B./build-x86_64-linux/gcc -O0 -S test.c -mavx512vl ; done
> 
> Here're results.
> w/o the patch applied:
>     real    0m14.245s
>     user    0m10.753s
>     sys     0m2.150s
> 
> w/ the patch applied:
>     real    0m16.404s
>     user    0m12.935s
>     sys     0m2.577s
> 
> So, we have compilation 15% slowdown when -mavx512vl
> is provided and no difference when not.
> Is this change is acceptable?

As #include <x86intrin.h> enables everything these days,
it will affect also that.
I'd say the 15% slowdown for -mavx512vl and #include <x86intrin.h>
is not something that would preclude applying the patch, but definitely
something to work on later on.
Unfortunately, for the x86intrin.h case, lazy initialization of the builtins
supposedly doesn't help at all, because pretty much all the builtins will be
there.

	Jakub


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]