[Bug target/122382] conversion to _Float16 does not vectorize on x86-64-v3

[ubizjak at gmail dot com via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122382

Uroš Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
             Status|ASSIGNED                    |RESOLVED

--- Comment #9 from Uroš Bizjak <ubizjak at gmail dot com> ---
The compiler now vectorizes conversions to and from _Float16 with -mf16c (part
of -march=x86-64-v3). This enables vectorization of e.g.:

void
foo (_Float16 *__restrict r, const _Float16 *__restrict x,
     const _Float16 *__restrict y, const _Float16 *__restrict z)
{
  for (int n = 0; n < 128; ++n)
    r[n] = (x[n] + y[n]) / z[n];
}

gcc -O2 -mf16c:

foo:
        xorl    %eax, %eax
        .p2align 4
        .p2align 3
.L2:
        vmovdqu (%rsi,%rax), %ymm3
        vmovdqu (%rdx,%rax), %ymm0
        vmovdqu (%rcx,%rax), %ymm2
        vcvtph2ps       %xmm3, %ymm4
        vcvtph2ps       %xmm0, %ymm1
        vextractf128    $0x1, %ymm3, %xmm3
        vaddps  %ymm4, %ymm1, %ymm1
        vextractf128    $0x1, %ymm0, %xmm0
        vcvtph2ps       %xmm3, %ymm3
        vcvtph2ps       %xmm2, %ymm4
        vextractf128    $0x1, %ymm2, %xmm2
        vcvtph2ps       %xmm0, %ymm0
        vaddps  %ymm3, %ymm0, %ymm0
        vcvtph2ps       %xmm2, %ymm2
        vdivps  %ymm4, %ymm1, %ymm1
        vdivps  %ymm2, %ymm0, %ymm0
        vcvtps2ph       $4, %ymm1, %xmm1
        vcvtps2ph       $4, %ymm0, %xmm0
        vinsertf128     $0x1, %xmm0, %ymm1, %ymm0
        vmovdqu %ymm0, (%rdi,%rax)
        addq    $32, %rax
        cmpq    $256, %rax
        jne     .L2
        vzeroupper
        ret

However, it still doesn't vectorize __builtin_sqrtf16, not even with
-ffast-math:

void
bar (_Float16 *__restrict r, const _Float16 *__restrict x,
     const _Float16 *__restrict y)
{
  for (int n = 0; n < 128; ++n)
    r[n] = __builtin_sqrtf16 (x[n] + y[n]);
}

I'll open another PR for that.

So, implemented for gcc-17.