On 25 April 2017 at 10:16, Nikunj A Dadhania <nik...@linux.vnet.ibm.com> wrote: > Peter Maydell <peter.mayd...@linaro.org> writes: > >> On 25 April 2017 at 09:58, Nikunj A Dadhania <nik...@linux.vnet.ibm.com> >> wrote: >>> /tmp/atomic-1660e0.o: In function `main': >>> /tmp/atomic.c:(.text+0x28): undefined reference to `__atomic_load_8' >>> /tmp/atomic.c:(.text+0x40): undefined reference to `__atomic_store_8' >>> /tmp/atomic.c:(.text+0x69): undefined reference to >>> `__atomic_compare_exchange_8' >>> /tmp/atomic.c:(.text+0x7d): undefined reference to `__atomic_exchange_8' >>> /tmp/atomic.c:(.text+0x91): undefined reference to >>> `__atomic_fetch_add_8' >>> clang-3.9: error: linker command failed with exit code 1 (use -v to see >>> invocation) >>> >>> With -latomic, the linker succeeds in getting the binary. >> >> What host is this on ? > > $ uname -a > Linux abhimanyu 4.9.13-200.fc25.x86_64 #1 SMP Mon Feb 27 16:48:42 UTC 2017 > x86_64 x86_64 x86_64 GNU/Linux
Hmm, I can repro that, but there's a problem even if we do link against libatomic. If you disassemble the resulting binaries, gcc produces what we want: f0 48 01 45 e8 lock add %rax,-0x18(%rbp) and other inline atomic instructions. clang on the other hand produces e8 33 fe ff ff callq 400610 <__atomic_fetch_add_8@plt> which is an expensive call to an out of line subroutine. We need to figure out how to get clang to just emit the inline operations. thanks -- PMM
