[Bug target/86019] [8/9 Regression] Unref implementation using atomic_thread_fence generates worse code on x86-64 in gcc 8.1 than 7.3

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86019

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
This used to be handled by peephole2, but that can't do anything with multiple
basic blocks and
(insn 12 11 13 3 (set (mem/v:BLK (scratch:DI) [0  A8])
        (unspec:BLK [
                (mem/v:BLK (scratch:DI) [0  A8])
            ] UNSPEC_MEMORY_BLOCKAGE))
"/usr/src/gcc-7/obj/x86_64-pc-linux-gnu/libstdc++-v3/include/bits/atomic_base.h":102:26
688 {*memory_blockag
e}
     (nil))
emitted in there.  And generic code doesn't know anything about the memory
blockage instruction that it would e.g. be able to move it from the conditional
to unconditional if it was the only thing (say let ifcvt pass do it?).