[PATCH] ARM: Do not select ARCH_HAS_KERNEL_FPU_SUPPORT

[Ard Biesheuvel]

On Wed, 15 May 2024 at 00:48, Samuel Holland <samuel.holland at sifive.com> wrote:
>
> Hello Ard,
>
> On 2024-05-09 2:39 AM, Ard Biesheuvel wrote:
> > On Thu, 9 May 2024 at 03:37, Samuel Holland <samuel.holland at sifive.com> wrote:
> >>
> >> On 32-bit ARM, conversions between `double` and `long long` require
> >> runtime library support. Since the kernel does not currently provide
> >> this library support, the amdgpu driver fails to build:
> >>
> >>   ERROR: modpost: "__aeabi_l2d" [drivers/gpu/drm/amd/amdgpu/amdgpu.ko] undefined!
> >>   ERROR: modpost: "__aeabi_d2ulz" [drivers/gpu/drm/amd/amdgpu/amdgpu.ko] undefined!
> >>
> >> As Arnd reports, there are likely no 32-bit ARM platforms which can use
> >> the amdgpu driver anyway, due to lack of features like 64-bit
> >> prefetchable BARs. Since amdgpu is currently the only real user of
> >> ARCH_HAS_KERNEL_FPU_SUPPORT, drop support for this option instead of
> >> bothering to implement the library functions.
> >>
> >> Fixes: 12624fe2d707 ("ARM: implement ARCH_HAS_KERNEL_FPU_SUPPORT")
> >
> > This commit is not in mainline yet. Could we just drop the original
> > patch instead?
>
> No, like I mentioned in the original thread, later patches in the series (for
> example bbce5cac4f5a ("lib/raid6: use CC_FLAGS_FPU for NEON CFLAGS")) depend on
> CC_FLAGS_FPU being defined for both arm and arm64.
>
> arm can't select ARCH_HAS_KERNEL_FPU_SUPPORT because the contract of that option
> is not fully implemented, but in my opinion it doesn't hurt to keep the part
> that is implemented.
>

Fair enough. But that still doesn't mean we have to add the 'select
ARCH_HAS_KERNEL_FPU_SUPPORT' only to back it out again.