[Bug target/80885] New: Do not use the opaque _mm256_cmp_ps to produce -1

[glisse at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80885

            Bug ID: 80885
           Summary: Do not use the opaque _mm256_cmp_ps to produce -1
           Product: gcc
           Version: 8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: glisse at gcc dot gnu.org
  Target Milestone: ---
            Target: x86_64-*-*

Hello,

in avx2intrin.h, I see a lot of

  __v8sf __src = _mm256_setzero_ps ();
  __v8sf __mask = _mm256_cmp_ps (__src, __src, _CMP_EQ_OQ);

(and variants with double)
Where gcc knows nothing about _mm256_cmp_ps and blindly generates the insn. It
seems to me that it would be better to use some way to let gcc know the
constant (for instance generate it as __v8si then cast), even if in the end it
might decide to generate the constant this way.

Macros (for the -O0 case) have the rather suspicious
 (__v4df)_mm256_set1_pd((double)(long long int) -1)