[PATCH] D151982: [AArch64] Make .arch without extra features actually take effect

[Martin Storsjö via Phabricator via llvm-commits]

mstorsjo added a comment.

In D151982#4390604 <https://reviews.llvm.org/D151982#4390604>, @DavidSpickett wrote:

> - Is there any use case for the previous behaviour that would be broken by this?

Not a direct use case per se, but this does affect how things do work out in some cases.

https://patchwork.ffmpeg.org/project/ffmpeg/patch/20230530123043.52940-1-martin@martin.st/ is a patch for ffmpeg, that tries to allow assembling the dotprod/i8mm extensions (in handwritten assembly) even if the baseline target doesn't include those extensions. Support for those extensions will be detected at runtime, and functions that rely on those extensions will be used if supported, while the binaries in general can run on anything from ARMv8.0 upwards.

To achieve this, the patch detects whether `.arch armv8.2-a` is supported, and whether `.arch_extension i8mm` is supported. The `.arch` directive is required, since binutils refuses to enable those extensions unless the base architecture level is >= 8.2.

So for most cases, we're expected to end up with this:

  .arch armv8.2-a
  .arch_extension dotprod
  .arch_extension i8mm
  <dotprod/i8mm instructions>

However `dotprod` and `i8mm` aren't supported in `.arch_extension` right now (fixed by D151981 <https://reviews.llvm.org/D151981>), but `.arch armv8.2-a` is. Now if we're building with `-march=armv8.6-a`, we do support those extensions without any directives.

So with current released Clang, we get this, if building without `-march=armv8.6-a`:

  .arch armv8.2-a
  // No dotprod/i8mm instructions used, LLVM is unable to enable them via assembly source

If building with `-march=armv8.6-a`, we get this:

  .arch armv8.2-a
  // Dotprod/i8mm instructions used since the .arch doesn't have any effect

With this patch applied, we wouldn't be able to use those instructions here. But coupled with D151981 <https://reviews.llvm.org/D151981> we can, and we can also make use of them without `-march=armv8.6-a`.

But this patch also makes it possible to enable the use of those instructions with `.arch armv8.6-a` in the assembly source, which didn't work before.

So TL;DR, this patch _does_ change how things behave. But it does change things towards a more correct and expected behaviour - I don't see the previous behaviour as an expected behaviour that anyone would rely on, especially if they expect to assemble with binutils too.

> - Is it worth adding a release note for? (regardless of the answer to the above)

I guess so

> - Does gcc/gas actually behave in the way this patch implements?

Yes; this patch makes LLVM behave like binutils gas in all of those cases as mentioned above.

> - Do you know if other backends have this problem? (does not need to be included in this patch however)

Not sure offhand, dunno if `.arch` (or other similar directives) behaves in a similar way on other architectures.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D151982/new/

https://reviews.llvm.org/D151982