Re: Popcount optimization using AVX512

On Sat, 6 Apr 2024 at 14:17, Nathan Bossart <nathandbossart(at)gmail(dot)com> wrote:
>
> On Sat, Apr 06, 2024 at 12:08:14PM +1300, David Rowley wrote:
> > Won't Valgrind complain about this?
> >
> > +pg_popcount_avx512(const char *buf, int bytes)
> >
> > + buf = (const char *) TYPEALIGN_DOWN(sizeof(__m512i), buf);
> >
> > + val = _mm512_maskz_loadu_epi8(mask, (const __m512i *) buf);
>
> I haven't been able to generate any complaints, at least with some simple
> tests. But I see your point. If this did cause such complaints, ISTM we'd
> just want to add it to the suppression file. Otherwise, I think we'd have
> to go back to the non-maskz approach (which I really wanted to avoid
> because of the weird function overhead juggling) or find another way to do
> a partial load into an __m512i.

[1] seems to think it's ok. If this is true then the following
shouldn't segfault:

The following seems to run without any issue and if I change the mask
to 1 it crashes, as you'd expect.

#include <immintrin.h>
#include <stdio.h>
int main(void)
{
__m512i val;
val = _mm512_maskz_loadu_epi8((__mmask64) 0, NULL);
printf("%llu\n", _mm512_reduce_add_epi64(val));
return 0;
}

gcc avx512.c -o avx512 -O0 -mavx512f -march=native

David

[1] https://stackoverflow.com/questions/54497141/when-using-a-mask-register-with-avx-512-load-and-stores-is-a-fault-raised-for-i