https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90606
Bug ID: 90606
Summary: Replace mfence with faster xchg for
std::memory_order_seq_cst.
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: maxim.yegorushkin at gmail dot com
Target Milestone: ---
The following example:
#include <atomic>
std::atomic<int> a;
void foo_seq_cst(int b) { a = b; }
Compiles with `gcc-9.1 -O3 -std=c++17 -pthread` into
foo_seq_cst(int):
mov DWORD PTR a[rip], edi
mfence
ret
Whereas `clang++-9 -O3 -std=c++17 -pthread` compiles it into:
foo_seq_cst(int): # @foo_seq_cst(int)
xchg dword ptr [rip + a], edi
ret
xchg was benchmarked to be 2-3x faster than mfence and Linux kernel switched to
xchg were possible.
gcc should also switch to using xchg for std::memory_order_seq_cst.
See:
https://lore.kernel.org/lkml/20160112150032-mutt-send-email-...@redhat.com/
https://stackoverflow.com/questions/56205324/why-do-gcc-inserts-mfence-where-clang-dont-use-it/
