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] A new meta intrinsic header file for current and future x86 instrinsics.


On Thu, Nov 20, 2008 at 8:37 AM, Uros Bizjak <ubizjak@gmail.com> wrote:
> On Wed, Nov 19, 2008 at 9:30 PM, rajagopal, dwarak
> <dwarak.rajagopal@amd.com> wrote:
>>>
>>> If there are no further comments or objections from AMD people in the
>>> next 48 hours, and since this header file is intended for _Intel_
>>> intrinsics, it is OK for mainline.
>>>
>>> BTW: I guess there will be a header file (ammintrin.h ?) introduced to
>>> include AMD defined intrinsics, so gcc can still add x86intrin.h that
>>> will include both immintrin.h and eventual ammintrin.h. Hopefully,
>> these
>>> two intrinsics won't step onto each other toes.
>>>
>>> I hope this eventual solution is acceptable to both contributors.
>
>> This proposal is not acceptable to AMD for the following reasons:
>>
>> - GCC is an open source compiler that is vendor agnostic. We do not want
>> to see vendor specific intrinsic files, as the ISAs are architecture
>> specific primarily. This is how it has been till now (eg: emmintrin.h,
>> xmmintrin.h, etc). Also there are many examples when AMD adopted SSE
>> instructions after Intel created them, and Intel adopted AMD
>> instructions after AMD came up with them (latest example: popcnt).
>>
>> - We recommend that there is only 1 intrinsic file used for all x86
>> intrinsics. This seemed to have been agreed upon
>> (http://gcc.gnu.org/ml/gcc/2008-04/msg00081.html). We prefer that this
>> file be neutral and vendor agnostic, which immintrin.h is not.
>
> The majority of the patch is a rename from gmmintrin.h to avxintrin.h.
> This part is not problematic at all and since 4.4 is not yet released
> is OK to commit.
>
> Let me explain why I think that introducing immintrin.h as the step in
> the right direction for the "master plan" in a little more depth:
>
> immintrin.h collects all subheaders that are relevant for _intel_
> processors. It is just a header of headers, conditionally included on
> various __SSEx__ defines. Situation up to now was, that we had to
> include the latest Intel defined intrinsic, i.e. "gmmintrin.h" to get
> Intel's intrinsics (and this file included all earlier includes down
> to mmintrin.h) _AND_  latest AMD defined intrinsics, i.e.
> "bmmintrin.h" that includes all intrinsics relevant to AMD processors
> (but for some reason not mm3dnow.h).
>
> When new intrinsics file was introduced, we have to fix all users to
> include this new file. See for example log history for
> test-all-intrinsics tests gcc.target/i386/sse-{12,13,14}.h. This is
> simply not acceptable.
>
> However, there is a need for cross-compiler compatibility between
> various compilers, not only icc, FWIW. So, the proposal, to some
> extent harmonized between "other compilers" is to introduce two
> headers:
>
> -immintrin.h for Intel's intrinsics
> -<whatever>mmintrin.h for AMD's intrinsics, where I propose that
> <whatever> is "a" for some consistency.
>
> These two intrinsics will just conditionally switch on or off various
> includes, please look at proposed immintrin.h. I think that AMD
> defined will also include various SSEx headers (they are all protected
> against double inclusion), hopefully also mm3dnow.h.
>
> Finally, a top-level file will be introduced, that will include
> immintrin.h and <a>mmintrin.h. The name of this intrinsic is not yet
> fixed, but it has to be the same as for other compilers. IMO,
> "x86intrin.h" is a good name, I hope it will stay. So, users have the
> flexibility to include x86intrin.h, immintrin.h or <a>mmintrin.h,
> depending on their needs.
>
> By including "x86intrin.h", subheaders will be conditionally included
> by various -msseX -m<whatever> flags. This is the reason that a lot of
> work has been done on the infrastructure for "this -msseX option
> enables various other -msseX and disables some other -msseX".
>
> Now, the concrete example is gcc.target/i386/sse-12.c testcase. As you
> can see, it now includes <bmmintrin.h> for AMD intrinsics and
> <smmintrin.h> for Intel intrinsics. (Also it includes mm3dnow.h -
> forgotten by AMD?). With a new scheme, it will just include
> <x86intrin.h>, and everything will be controlled by various -msseX
> options. Try to compile this file without -msse5 for example, you will
> see where the problem lies.
>
> So I propose that AMD introduces its own intrinsic file, following
> immintrin.h example. There *will be* a vendor neutral x86intrin.h (or
> whatever the name will be agreed to), and AMDs file will be included
> through this vendor neutral header just in the same way as Intels
> file.

I fail to see a compelling reason to have vendor specific headers at all.

Richard.


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