[Bug target/86693] inefficient atomic_fetch_xor

[Bug target/86693] inefficient atomic_fetch_xor

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86693

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement

[Bug target/86693] inefficient atomic_fetch_xor

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86693

--- Comment #4 from Hongtao.liu <crazylht at gmail dot com> ---
Change testcase a little bit, gcc now can generate lock btc


void func1();

void func(unsigned long *counter)
{
        if (__atomic_fetch_xor(counter, 1, __ATOMIC_ACQ_REL) & 1) {
                func1();
        }
}


func(unsigned long*):
        lock btc        QWORD PTR [rdi], 0
        jc      .L4
        ret
.L4:
        jmp     func1()

[Bug target/86693] inefficient atomic_fetch_xor

[hjl.tools at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86693

--- Comment #5 from H.J. Lu <hjl.tools at gmail dot com> ---
(In reply to Hongtao.liu from comment #4)
> Change testcase a little bit, gcc now can generate lock btc
> 
> 
> void func1();
> 
> void func(unsigned long *counter)
> {
>         if (__atomic_fetch_xor(counter, 1, __ATOMIC_ACQ_REL) & 1) {
>                 func1();
>         }
> }
> 
> 
> func(unsigned long*):
>         lock btc        QWORD PTR [rdi], 0
>         jc      .L4
>         ret
> .L4:
>         jmp     func1()

We should rewrite the original test to the canonical form, similar to r12-5102.
Hongtao, can you do that?

[Bug target/86693] inefficient atomic_fetch_xor

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86693

--- Comment #6 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to H.J. Lu from comment #5)
> (In reply to Hongtao.liu from comment #4)
> > Change testcase a little bit, gcc now can generate lock btc
> > 
> > 
> > void func1();
> > 
> > void func(unsigned long *counter)
> > {
> >         if (__atomic_fetch_xor(counter, 1, __ATOMIC_ACQ_REL) & 1) {
> >                 func1();
> >         }
> > }
> > 
> > 
> > func(unsigned long*):
> >         lock btc        QWORD PTR [rdi], 0
> >         jc      .L4
> >         ret
> > .L4:
> >         jmp     func1()
> 
> We should rewrite the original test to the canonical form, similar to
> r12-5102.
> Hongtao, can you do that?

The orginal testcase is not equal to btc, __atomic_fetch_xor(counter, 1,
__ATOMIC_ACQ_REL) == 1 require other bits of *counter is 0. And as #c1 said, we
don't have instructions to keep the old or new value of xor or ior in a
register.

I can't find a way to optimize off the exchangeloop.