Re: [PATCH] Aarch64: Add simd exp/expf functions

[Florian Weimer <fweimer at redhat dot com>]

* Szabolcs Nagy:

> On 07/03/2019 19:04, Joseph Myers wrote:
>> On Wed, 6 Mar 2019, Florian Weimer wrote:
>> 
>>> On the other hand, it could help to get libmvec out of the door more
>>> quickly.  I think it's not ideal that if you use an older compiler,
>>> you get only a subset of the glibc ABI.  We can get away with it here
>>> because it affects an entire soname.  Still it might be difficult to
>>> explain why applications are not portable.
>> 
>> On the whole I think I agree with Rich Felker's argument 
>> <https://sourceware.org/ml/libc-alpha/2015-11/msg00184.html> against 
>> having the presence of libmvec depend on the tools used for the build.  
>> (Note that the installed bits/math-vector.h file, which may be shared 
>> between multilibs, does not depend on the tools used, so if libmvec was 
>> disabled then the installed bits/math-vector.h is not actually correct and 
>> some programs will fail to build.)
>> 
>> This is an argument for removing the --disable-mathvec configure option as 
>> well as either having assembly wrappers or a requirement for new-enough 
>> tool versions for building libmvec functions on platforms where the oldest 
>> supported GCC / binutils aren't new enough.
>
> so is it acceptable to submit generated asm to the
> source tree together with the c source?

No, before we do that, I think we should just require GCC 9 and binutils
2.33 for building aarch64.

I had the hope that you could build a compatible ABI with just a few
assember trampolines, but that's not the case if the DSOs need markers
for disabling lazy binding in client code.  (But it is probably more
natural to disable lazy binding though function attributes in the header
file.)

Thanks,
Florian