[FFmpeg-devel] [PATCH 2/2] configure: Improve aarch64 feature detection on older, broken Clang versions

[Martin Storsjö]

On Tue, 24 Oct 2023, Martin Storsjö wrote:

> Clang versions before 17 (Xcode versions up to and including 15.0)
> had a very annoying bug in its behaviour of the ".arch" directive
> in assembly. If the directive only contained a level, such as
> ".arch armv8.2-a", it did validate the name of the level, but it
> didn't apply the level to what instructions are allowed. The level
> was applied if the directive contained an extra feature enabled,
> such as ".arch armv8.2-a+crc" though. It was also applied on the
> next ".arch_extension" directive.
>
> This bug, combined with the fact that the same versions of Clang
> didn't support the dotprod/i8mm extension names in either
> ".arch <level>+<feature>" or in ".arch_extension", could lead to
> unexepcted build failures.
>
> As the dotprod/i8mm extensions couldn't be enabled dynamically
> via the ".arch_extension" directive, someone building ffmpeg could
> try to enable them by configuring their build with
> --extra-cflags="-march=armv8.6-a".
>
> During configure, we test for support for the i8mm instructions
> like this:
>
>    # Built with -march=armv8.6-a
>    .arch armv8.2-a             # Has no visible effect here
>    #.arch_extension i8mm       # Omitted as the extension name isn't known
>    usdot v0.4s, v0.16b, v0.16b
>    # Successfully assembled as armv8.6-a is the effective level,
>    # and i8mm is enabled implicitly in armv8.6-a.
>
> Thus, we would enable assembling those instructions. However if
> we later check for another extension, such as sve (which those
> versions of Clang actually do support), we can later run into the
> following situation when building actual code:
>
>    # Built with -march=armv8.6-a
>    .arch armv8.2-a             # Has no visible effect here
>    #.arch_extension i8mm       # Omitted as the extension name isn't known
>    .arch_extension sve         # Included as "sve" is as supported extension name
>    # .arch_extension effectively activates the previous .arch directive,
>    # so the effective level is armv8.2-a+sve now.
>    usdot v0.4s, v0.16b, v0.16b
>    # Fails to build the instructions that require i8mm. Despite the
>    # configure check, the unrelated ".arch_extension sve" directive
>    # breaks the functionality of the i8mm feature.
>
> This patch avoids this situation:
> - By adding a dummy feature such as "+crc" on the .arch directive
>  (if supported), we make sure that it does get applied immediately,
>  avoiding it taking effect spuriously at a later unrelated
>  ".arch_extension" directive.
> - By checking for higher arch levels such as armv8.4-a and armv8.6-a,
>  we can assemble the dotprod and i8mm extensions without the user
>  needing to pass -march=armv8.6-a. This allows using the dotprod/i8mm
>  codepaths via runtime detection while keeping the binary runnable
>  on older versions. I.e. this enables the i8mm codepaths on Apple M2
>  machines while built with Xcode's Clang.
>
> TL;DR: Enable the I8MM extensions for Apple M2 without the user needing
> to do a custom configuration; avoid potential build breakage if a user
> does such a custom configuration.
>
> Once Xcode versions that have these issues fixed are prevalent, we
> can consider reverting this change.
> ---
> configure | 21 ++++++++++++++++++++-
> 1 file changed, 20 insertions(+), 1 deletion(-)

Will push now.

// Martin