Re: [PATCH v2 3/3] msvcrt: Add an SSE2 memset_aligned_32 implementation.

[Rémi Bernon]

On 9/14/21 1:10 PM, Piotr Caban wrote:
> On 9/14/21 12:55 PM, Rémi Bernon wrote:
> > On 9/14/21 12:15 PM, Marvin wrote:
> > - guard it with #ifdef __SSE2__, and in that case now that we import
> >
> > x86intrin.h we could use intel intrinsics instead of inline assembly,
>
> I didn't check that __SSE2__ is not defined when i686-w64-mingw is used. 
> I think that the SSE enabled version should be available in this case.

It may be defined, and it is for me when I build locally, but not on the 
testbot for some reason. You can have the same result with -mno-sse.
> > - or make it an external assembly function.
>
> We can also consider moving the ERMS version to separate function to 
> avoid using inline assembly.

That too. The call has an overhead though (but for large sizes that may 
not matter much).