[Bug middle-end/66881] New: Possibly inefficient std::atomic<int> codegen on x86 for simple arithmetic

["tkoeppe at google dot com" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66881

            Bug ID: 66881
           Summary: Possibly inefficient std::atomic<int> codegen on x86
                    for simple arithmetic
           Product: gcc
           Version: 4.9.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: tkoeppe at google dot com
  Target Milestone: ---

Consider these two simple versions of addition:

  #include <atomic>

  std::atomic<int> x;
  int y;

  void f(int a) {
    x.store(x.load(std::memory_order_relaxed) + a, std::memory_order_relaxed);
  }

  void g(int a) {
    y += a;
  }

GCC generates the following assembly:

  f(int):
        mov     eax, DWORD PTR x[rip]
        add     edi, eax
        mov     DWORD PTR x[rip], edi
        ret

  g(int):
        add     DWORD PTR y[rip], edi
        ret

Now, it is clear to me that the correct atomic codegen for store() and load()
is "mov", as it appears here, but why aren't the two consecutive operations not
folded into a single add? Aren't the semantics and the memory ordering the
same? x86 says that (most) "reads" and "writes" are strongly ordered; doesn't
that apply to the read and write produced by "add", too?

(My original motivation came from a variant of this with floats, where the
non-atomic code executed noticeably faster, even though I would have expected
the two to produce the same machine code.)