https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122382
Uroš Bizjak <ubizjak at gmail dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|NEW |ASSIGNED
Assignee|unassigned at gcc dot gnu.org |ubizjak at gmail dot com
--- Comment #6 from Uroš Bizjak <ubizjak at gmail dot com> ---
Created attachment 65984
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=65984&action=edit
Patch in testing
Following test:
--cut here--
void
foo (_Float16 *__restrict h, const float *__restrict f)
{
for (int n = 0; n < 128; ++n)
h[n] = f[n];
}
void
bar (float *__restrict f, const _Float16 *__restrict h)
{
for (int n = 0; n < 128; ++n)
f[n] = h[n];
}
--cut here--
compiles with "-O2 -mf16c -mno-avx512f" to:
foo:
xorl %eax, %eax
.p2align 6
.p2align 4
.p2align 3
.L2:
vmovups (%rsi,%rax,2), %ymm0
vmovups 32(%rsi,%rax,2), %ymm1
vcvtps2ph $4, %ymm0, %xmm0
vcvtps2ph $4, %ymm1, %xmm1
vinsertf128 $0x1, %xmm1, %ymm0, %ymm0
vmovdqu %ymm0, (%rdi,%rax)
addq $32, %rax
cmpq $256, %rax
jne .L2
vzeroupper
ret
bar:
xorl %eax, %eax
.p2align 6
.p2align 4
.p2align 3
.L6:
vmovdqu (%rsi,%rax), %ymm0
vcvtph2ps %xmm0, %ymm1
vextractf128 $0x1, %ymm0, %xmm0
vmovups %ymm1, (%rdi,%rax,2)
vcvtph2ps %xmm0, %ymm0
vmovups %ymm0, 32(%rdi,%rax,2)
addq $32, %rax
cmpq $256, %rax
jne .L6
vzeroupper
ret
