On Tue, Dec 7, 2021 at 12:28 PM Peter Zijlstra [off-list ref] wrote:

Argh.. __atomic_add_fetch() != __atomic_fetch_add(); much confusion for
GCC having both. With the right primitive it becomes:

        movl    $1, %eax
        lock xaddl      %eax, (%rdi)
        testl   %eax, %eax
        je      .L5
        js      .L6

Which makes a whole lot more sense.

Note that the above misses the case where the old value was MAX_INT
and the result now became negative.

That isn't a _problem_, of course. I think it's fine. But if you cared
about it, you'd have to do something like

        movl    $1, %eax
        lock xaddl      %eax, (%rdi)
        jl      .L6
        testl   %eax, %eax
        je      .L5

instead (I might have gotten that "jl" wrong, needs more testing.

But if you don't care about the MAX_INT overflow and make the overflow
boundary be the next increment, then just make it be one error case:

        movl    $1, %eax
        lock xaddl      %eax, (%rdi)
        testl   %eax, %eax
        jle      .L5

and then (if you absolutely have to distinguish them) you can test eax
again in the slow path.

                     Linus