On Thu, 27 Feb 2025 13:05:29 -0500
Yury Norov <yury.norov@xxxxxxxxx> wrote:
> On Wed, Feb 26, 2025 at 10:29:11PM +0000, David Laight wrote:
> > On Mon, 24 Feb 2025 14:27:03 -0500
> > Yury Norov <yury.norov@xxxxxxxxx> wrote:
> > ....
> > > +#define parity(val) \
> > > +({ \
> > > + u64 __v = (val); \
> > > + int __ret; \
> > > + switch (BITS_PER_TYPE(val)) { \
> > > + case 64: \
> > > + __v ^= __v >> 32; \
> > > + fallthrough; \
> > > + case 32: \
> > > + __v ^= __v >> 16; \
> > > + fallthrough; \
> > > + case 16: \
> > > + __v ^= __v >> 8; \
> > > + fallthrough; \
> > > + case 8: \
> > > + __v ^= __v >> 4; \
> > > + __ret = (0x6996 >> (__v & 0xf)) & 1; \
> > > + break; \
> > > + default: \
> > > + BUILD_BUG(); \
> > > + } \
> > > + __ret; \
> > > +})
> > > +
> >
> > You really don't want to do that!
> > gcc makes a right hash of it for x86 (32bit).
> > See https://www.godbolt.org/z/jG8dv3cvs
>
> GCC fails to even understand this. Of course, the __v should be an
> __auto_type. But that way GCC fails to understand that case 64 is
> a dead code for all smaller type and throws a false-positive
> Wshift-count-overflow. This is a known issue, unfixed for 25 years!
Just do __v ^= __v >> 16 >> 16
>
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=4210
>
> > You do better using a __v32 after the 64bit xor.
>
> It should be an __auto_type. I already mentioned. So because of that,
> we can either do something like this:
>
> #define parity(val) \
> ({ \
> #ifdef CLANG \
> __auto_type __v = (val); \
> #else /* GCC; because of this and that */ \
> u64 __v = (val); \
> #endif \
> int __ret; \
>
> Or simply disable Wshift-count-overflow for GCC.
For 64bit values on 32bit it is probably better to do:
int p32(unsigned long long x)
{
unsigned int lo = x;
lo ^= x >> 32;
lo ^= lo >> 16;
lo ^= lo >> 8;
lo ^= lo >> 4;
return (0x6996 >> (lo & 0xf)) & 1;
}
That stops the compiler doing 64bit shifts (ok on x86, but probably not elsewhere).
It is likely to be reasonably optimal for most 64bit cpu as well.
(For x86-64 it probably removes a load of REX prefix.)
(It adds an extra instruction to arm because if its barrel shifter.)
>
> > Even the 64bit version is probably sub-optimal (both gcc and clang).
> > The whole lot ends up being a bit single register dependency chain.
> > You want to do:
>
> No, I don't. I want to have a sane compiler that does it for me.
>
> > mov %eax, %edx
> > shrl $n, %eax
> > xor %edx, %eax
> > so that the 'mov' and 'shrl' can happen in the same clock
> > (without relying on the register-register move being optimised out).
> >
> > I dropped in the arm64 for an example of where the magic shift of 6996
> > just adds an extra instruction.
>
> It's still unclear to me that this parity thing is used in hot paths.
> If that holds, it's unclear that your hand-made version is better than
> what's generated by GCC.
I wasn't seriously considering doing that optimisation.
Perhaps just hoping is might make a compiler person think :-)
David
>
> Do you have any perf test?
>
> Thanks,
> Yury
