[FFmpeg-devel] [FFmpeg-cvslog] x86/intreadwrite: use intrinsics instead of inline asm for AV_COPY128

[Martin Storsjö]

On Thu, 11 Jul 2024, James Almer wrote:

> On 7/11/2024 10:54 AM, Martin Storsjö wrote:
>> On Thu, 11 Jul 2024, Martin Storsjö wrote:
>> 
>>> On Thu, 11 Jul 2024, James Almer wrote:
>>>
>>>> On 7/11/2024 10:08 AM, Martin Storsjö wrote:
>>>>> On Wed, 10 Jul 2024, James Almer wrote:
>>>>>
>>>>>> ffmpeg | branch: master | James Almer <jamrial at gmail.com> | Wed Jul 
>>>>>> 10 13:00:20 2024 -0300| [bd1bcb07e0f29c135103a402d71b343a09ad1690] 
>>>>>> | committer: James Almer
>>>>>>
>>>>>> x86/intreadwrite: use intrinsics instead of inline asm for AV_COPY128
>>>>>>
>>>>>> This has the benefit of removing any SSE -> AVX penalty that may 
>>>>>> happen when
>>>>>> the compiler emits VEX encoded instructions.
>>>>>>
>>>>>> Signed-off-by: James Almer <jamrial at gmail.com>
>>>>>>
>>>>>>>
>>>> 
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=bd1bcb07e0f29c135103a402d71b343a09ad1690
>>>>>> ---
>>>>>>
>>>>>> configure                    |  5 ++++-
>>>>>> libavutil/x86/intreadwrite.h | 20 +++++++-------------
>>>>>> 2 files changed, 11 insertions(+), 14 deletions(-)
>>>>>>
>>>>>> diff --git a/configure b/configure
>>>>>> index f84fefeaab..7151ed1de3 100755
>>>>>> --- a/configure
>>>>>> +++ b/configure
>>>>>> @@ -2314,6 +2314,7 @@ HEADERS_LIST="
>>>>>>
>>>>>> INTRINSICS_LIST="
>>>>>>     intrinsics_neon
>>>>>> +    intrinsics_sse
>>>>>>     intrinsics_sse2
>>>>>> "
>>>>>>
>>>>>> @@ -2744,7 +2745,8 @@ armv6t2_deps="arm"
>>>>>> armv8_deps="aarch64"
>>>>>> neon_deps_any="aarch64 arm"
>>>>>> intrinsics_neon_deps="neon"
>>>>>> -intrinsics_sse2_deps="sse2"
>>>>>> +intrinsics_sse_deps="sse"
>>>>>> +intrinsics_sse2_deps="sse2 intrinsics_sse"
>>>>>> vfp_deps="arm"
>>>>>> vfpv3_deps="vfp"
>>>>>> setend_deps="arm"
>>>>>> @@ -6446,6 +6448,7 @@ elif enabled loongarch; then
>>>>>> fi
>>>>>>
>>>>>> check_cc intrinsics_neon arm_neon.h "int16x8_t test = vdupq_n_s16(0)"
>>>>>> +check_cc intrinsics_sse immintrin.h "__m128 test = _mm_setzero_ps()"
>>>>>> check_cc intrinsics_sse2 emmintrin.h "__m128i test = 
>>>> _mm_setzero_si128()"
>>>>>>
>>>>>> check_ldflags -Wl,--as-needed
>>>>>> diff --git a/libavutil/x86/intreadwrite.h 
>>>> b/libavutil/x86/intreadwrite.h
>>>>>> index 9bbef00dba..6546eb016c 100644
>>>>>> --- a/libavutil/x86/intreadwrite.h
>>>>>> +++ b/libavutil/x86/intreadwrite.h
>>>>>> @@ -22,29 +22,25 @@
>>>>>> #define AVUTIL_X86_INTREADWRITE_H
>>>>>>
>>>>>> #include <stdint.h>
>>>>>> +#if HAVE_INTRINSICS_SSE
>>>>>> +#include <immintrin.h>
>>>>>> +#endif
>>>>>
>>>>> This change seems to have broken builds for x86 with Clang 16 or 
>>>>> newer. (Clang 15 and lower seems to be fine.)
>>>>>
>>>>> See e.g. 
>>>> 
> http://fate.ffmpeg.org/log.cgi?slot=i686-mingw32-clang-trunk&time=20240711035948&log=compile 
> for an example of the error. The issue is that a clang internal intrinsics 
> header contains "_mm_comige_sh(__m128h A,", i.e. a parameter with the name 
> "A" (which toolchain provided headers shouldn't use). This clashes with 
> libavcodec/huffuyv.h, which has a "#define A 3".
>>>>>
>>>>> This is obviously a Clang intrinsics header bug, but we can't fix 
>>>>> the existing Clang 16-18 releases that are out there, so I guess 
>>>>> what we 
>>>> can
>>>>> do is change our "define A" to something more elaborate. (IIRC there 
>>>> are
>>>>> some similar issues with names with ncurses and/or android headers 
>>>> too.)
>>>>>
>>>>> // Martin
>>>>
>>>> We also do "#define A AV_OPT_FLAG_AUDIO_PARAM" in options_table.h and 
>>>> probably other places, so changing huffyuv.h may not be enough.
>>>
>>> That's quite possible, but those cases may be including intreadwrite.h 
>>> before that, do it's possible it might not trigger there.
>>>
>>> I'll see how many places need to be changed here.
>> 
>> huffyuvenc.c seems to be the only file that runs into the issue; it 
>> includes put_bits.h (which brings in libavutil/intreadwrite.h) after 
>> huffyuv.h.
>
> Can we move the huffyuv.h include right after put_bits.h then? I 
> personally prefer that over changing the RGBA defines to workaround a 
> compiler bug.

That would probably work too.

However your other recent patch, which removes the include of immintrin.h, 
also avoids the issue altogether.

(FWIW, the upstream bug is fixed now in 
https://github.com/llvm/llvm-project/commit/6f04f46927cf54d19cc2a1470f47d5db4b3b96bb.)

// Martin