Re: clang build error on i686

[Peter Maydell]

On Sun, 11 Jul 2021 at 16:17, Richard Henderson
<richard.henderson@linaro.org> wrote:
>
> On 7/3/21 10:45 AM, Peter Maydell wrote:
> > On Sat, 3 Jul 2021 at 15:37, Cole Robinson <crobinso@redhat.com> wrote:
> >>
> >> Hi, I'm hitting build errors with clang on i686 userspace on x86_64
> >> kernel. Affects both qemu 6.0.0 and qemu.git, tested with fedora
> >> clang-12.0.1~rc3-1.fc35.i686.
> >>
> >> Full build log from the 6.0.0 build:
> >> https://gist.githubusercontent.com/crobinso/7b1206044eac7326490b2adce829e861/raw/9dddef968051fd6383ba7adb9e595081ad4f8fa4/gistfile1.txt
> >>
> >> Lots of errors like:
> >>
> >> /usr/bin/ld: libqemu-aarch64-softmmu.fa.p/accel_tcg_cputlb.c.o: in
> >> function `helper_atomic_cmpxchgq_le_mmu':
> >> /builddir/build/BUILD/qemu-6.0.0/accel/tcg/atomic_template.h:86:
> >> undefined reference to `__atomic_compare_exchange_8'
> >> /usr/bin/ld: libqemu-aarch64-softmmu.fa.p/accel_tcg_cputlb.c.o: in
> >> function `helper_atomic_xchgq_le_mmu':
> >> /builddir/build/BUILD/qemu-6.0.0/accel/tcg/atomic_template.h:134:
> >> undefined reference to `__atomic_exchange_8'
> >>
> >> Also warnings like:
> >>
> >> /builddir/build/BUILD/qemu-6.0.0/include/qemu/stats64.h:58:21: warning:
> >> misaligned atomic operation may incur significant performance penalty;
> >> the expected alignment (8 bytes) exceeds the actual alignment (4 bytes)
> >> [-Watomic-alignment]
> >>      uint64_t orig = qatomic_read__nocheck(&s->value);
> >>                      ^
> >> /builddir/build/BUILD/qemu-6.0.0/include/qemu/atomic.h:129:5: note:
> >> expanded from macro 'qatomic_read__nocheck'
> >>      __atomic_load_n(ptr, __ATOMIC_RELAXED)
> >
> > I think at least part of what is happening here is that this compiler/host
> > doesn't support native 64-bit atomics, but configure has selected
> > CONFIG_ATOMIC64 anyway.
>
> Not true.  The host certainly supports it.
>
> This is a new alignment warning in clang-12 wrt the alignment of the atomic 
> operation.
> Which may be complicated by the fact that the i386 abi does not normally 
> align structures
> beyond 4 bytes.
>
> We may need to disable this warning for i386 (but not x86_64).

The first part of the problem isn't just a warning, though -- it's clang
actually emitting calls to libatomic.

-- PMM