[llvmlinux] Heavy experiments

Behan Webster behanw at converseincode.com
Wed Jul 24 14:42:51 UTC 2013

Just to add a tiny bit to the conversation...

On 07/24/13 07:04, Renato Golin wrote:
> On 24 July 2013 02:08, Marcelo Sousa <marceloabsousa at gmail.com 
> <mailto:marceloabsousa at gmail.com>> wrote:
>      - Vectorization can explode the amount of IR beyond recognition
>     Are you referring to vectorization as a compiler transformation?
>     Please elaborate and if possible provide an example of this.
>     Nevertheless, I doubt that this happens in the particular case of
>     the Linux Kernel.
> Yes, and I agree, it's possible that the Kernel has nothing worth 
> vectorizing. It's just an example where your tool wouldn't work at all.
In general the various cpu vector extensions and the FPU aren't used in 
the kernel (though there are a few exceptions to this rule where it 
makes sense). The overhead in always saving/restoring the state of these 
parts of the CPU during system calls would be too high (context switches 
between user and kernel space are many and already relatively 
expensive). The FPU and vector extensions are saved for use in userspace 
where they tend to have the most impact.


Behan Webster
behanw at converseincode.com

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.linuxfoundation.org/pipermail/llvmlinux/attachments/20130724/a25e77dc/attachment.html>

More information about the LLVMLinux mailing list