Compiling the kernel with LLVM=1 fails for ARCH=arm CROSS_COMPILE=armv6l-unknown-linux-gnueabihf- kernels

Describe the bug

The issue seems to be that https://github.com/raspberrypi/linux/blob/rpi-5.15.y/arch/arm/lib/memcmp_rpi.S (and perhaps others) are not compatible with the clang assembler.

Steps to reproduce the behaviour

On a nix system with flakes enabled,

$ nix build github:eliasnaur/rpikernel-clang-issue
error: builder for '/nix/store/3zc9nww3crck85wir5fs78zfyfv1nm5q-Raspberry-Pi-Linux-kernel.drv' failed with exit code 2;
last 10 log lines:
>         ldrmib DAT5, [S_2]
>         ^
> <instantiation>:11:9: note: while in macro instantiation
>         memcmp_trailing_15bytes unaligned
>         ^
> arch/arm/lib/memcmp_rpi.S:265:6: note: while in macro instantiation
> 140: memcmp_short_inner_loop 1
>      ^
> make[1]: *** [scripts/Makefile.build:391: arch/arm/lib/memcmp_rpi.o] Error 1
> make: *** [Makefile:1900: arch/arm/lib] Error 2
For full logs, run 'nix log /nix/store/3zc9nww3crck85wir5fs78zfyfv1nm5q-Raspberry-Pi-Linux-kernel.drv'.

The relevant configuration and building scripts are in the repository. I'm using clang 14.

Device (s)

Raspberry Pi Zero, Raspberry Pi Zero W / WH

System

The issue is during building of custom kernel, not at runtime.

Logs

No response

Additional context

I'm building a custom kernel with LLVM because I have problems producing deterministic output from a gcc build.

The issue can be worked around by disabling BCM2835_FAST_MEMCPY in the kernel config.

Start with arch/arm/lib/memcmp_rpi.S and reproduce the failure using the linked Nix flake with LLVM=1, ARCH=arm, and the armv6l cross-compiler. Inspect the clang assembler diagnostics around the memcmp macros and compare the behavior with the existing GCC build. Done means the Raspberry Pi kernel builds with LLVM enabled without disabling BCM2835_FAST_MEMCPY.