__ARM_ARCH incorrect for armv8.1-a and above

[llvmbot]

Bugzilla Link	49671
Version	trunk
OS	All
Reporter	LLVM Bugzilla Contributor
CC	@DavidSpickett

Extended Description

According to the ACLE Q3 2020 documentation
(https://developer.arm.com/documentation/101028/0012), section 5.4.1 ("A32/T32 instruction set architecture"), for 8.1+ __ARM_ARCH is supposed to be 801 for armv8.1-a, 802 for armv8.2-a, 803 for armv8.3-a, etc.

Clang currently just uses 8 for all versions of armv8. Testing is pretty straightforward: clang --target=aarch64-linux-gnu -march=armv8.6-a -dM -E - </dev/null | grep __ARM_ARCH.

This makes it difficult to determine support for features which are only available in later revisions but don't have an _ARM_FEATURE* macro, such as complex MLA (8.3).

Similar GCC bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99312

[DavidSpickett]

These macros are emitted in clang/lib/Basic/Targets/AArch64.cpp and we already have a switch block there that handles arch specific macros. So it should be easy enough to change.

As it is "8" is always used as you saw:

// ACLE predefines. Many can only have one possible value on v8 AArch64.
Builder.defineMacro("__ARM_ACLE", "200");
Builder.defineMacro("__ARM_ARCH", "8");

I'll try to find some time to fix this, probably next week.

[DavidSpickett]

So the basic fix is simple but the macro itself has some issues that need to be clarified. Both compilers just give "8" and that has presented an unofficial standard we need to be wary of. There are discussions ongoing around how to proceed.

One issue is something like: -march=armv8.1-a+nocrc
This would give you 801 with a basic fix. However from the ARMARM:
"All implementations of the Armv8.1 architecture are required to implement the CRC32* instructions."

I know from previous work this area of mix and matching extensions isn't straightforward. Clang is not very aware of what can and cannot be backported/removed etc. So the full fix would be more involved.

So I will not be working on this myself, at least until there's a new ACLE revision that clarifies what compilers should be doing. (especially in the corner cases)

(as I'm sure you saw elsewhere but I'll repeat anyway to keep this bug updated...)

The recommendation in general is that code should use per feature flags. If some are missing/not supported in clang we can certainly fix that.

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: None (llvmbot)

| | | | --- | --- | | Bugzilla Link | [49671](https://llvm.org/bz49671) | | Version | trunk | | OS | All | | Reporter | LLVM Bugzilla Contributor | | CC | @DavidSpickett |

Extended Description

According to the ACLE Q3 2020 documentation
(<https://developer.arm.com/documentation/101028/0012>), section 5.4.1 ("A32/T32 instruction set architecture"), for 8.1+ __ARM_ARCH is supposed to be 801 for armv8.1-a, 802 for armv8.2-a, 803 for armv8.3-a, etc.

Clang currently just uses 8 for all versions of armv8. Testing is pretty straightforward: clang --target=aarch64-linux-gnu -march=armv8.6-a -dM -E - </dev/null | grep __ARM_ARCH.

This makes it difficult to determine support for features which are only available in later revisions but don't have an _ARM_FEATURE* macro, such as complex MLA (8.3).

Similar GCC bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99312

[pinskia]

This has been fixed by #81493 .

[lenary]

This has been fixed by #81493 .

I think that was reverted: d022f32