On Mon, 5 Jul 2021 at 17:25, Stefan Hajnoczi <stefa...@redhat.com> wrote: > > On Mon, Jul 05, 2021 at 12:31:01PM +0200, Philippe Mathieu-Daudé wrote: > > +Richard/Stefan for "atomic" error. > ... > > > [2363/9207] Linking target qemu-system-aarch64 > > > FAILED: qemu-system-aarch64 > > > clang++ @qemu-system-aarch64.rsp > > > libqemu-aarch64-softmmu.fa.p/accel_tcg_cputlb.c.o: In function > > > `helper_atomic_cmpxchgq_le_mmu': > > > /tmp/tmp.VkWONZ62bA/build/../accel/tcg/atomic_template.h:86: undefined > > > reference to `__atomic_compare_exchange_8' > > > libqemu-aarch64-softmmu.fa.p/accel_tcg_cputlb.c.o: In function > > > `helper_atomic_xchgq_le_mmu': > > > /tmp/tmp.VkWONZ62bA/build/../accel/tcg/atomic_template.h:134: undefined > > > reference to `__atomic_exchange_8' > > > libqemu-aarch64-softmmu.fa.p/accel_tcg_cputlb.c.o: In function > > > `helper_atomic_fetch_addq_le_mmu': > > According to docs/devel/atomics.rst: > > These operations are polymorphic; they operate on any type that is as > wide as a pointer or smaller. > > It looks like the compiler doesn't support 8-bit atomics here?
8 here means "8 bytes", not "8 bits". And indeed on i386 you can't do 8-byte atomics with simple insns. The compiler's answer to this is "emit a call to a helper in libatomic, which will emulate an atomic access by taking some kind of lock". We don't ever want to fall back to "take a lock" because sometimes our accesses to the atomic variables are from TCG generated code -- this is why we don't link against libatomic. The problem is that we have not correctly detected that this compiler can't do inline atomics for 64-bit values and avoided using them. But at least we have made this a compile failure rather than a silently-wrong-code bug :-) thanks -- PMM
