https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103571
--- Comment #18 from Hongtao.liu <crazylht at gmail dot com> ---
codegen for foo1/foo2 is suboptimal under -mavx2, i guess we can have
vec_setv16hf_0 and with vpblendw.
typedef _Float16 __v16hf __attribute__ ((__vector_size__ (32)));
typedef _Float16 __m256h __attribute__ ((__vector_size__ (32), __may_alias__));
__m256h
__attribute__ ((noinline, noclone))
foo1 (_Float16 x)
{
return __extension__ (__m256h)(__v16hf) { x, 0.0f, 0.0f, 0.0f,
0.0f, 0.0f, 0.0f, 0.0f,
0.0f, 0.0f, 0.0f, 0.0f,
0.0f, 0.0f, 0.0f, 0.0f };
}
__m256h
__attribute__ ((noinline, noclone))
foo2 (_Float16 *x)
{
return __extension__ (__m256h)(__v16hf) { *x, 0.0f, 0.0f, 0.0f,
0.0f, 0.0f, 0.0f, 0.0f,
0.0f, 0.0f, 0.0f, 0.0f,
0.0f, 0.0f, 0.0f, 0.0f };
}
foo1:
.LFB0:
.cfi_startproc
vpxor %xmm1, %xmm1, %xmm1
vpbroadcastw %xmm0, %ymm0
vpblendw $1, %ymm0, %ymm1, %ymm0
vpblendd $15, %ymm0, %ymm1, %ymm1
vmovdqa %ymm1, %ymm0
ret
.cfi_endproc
.LFE0:
.size foo1, .-foo1
.p2align 4
.globl foo2
.type foo2, @function
foo2:
.LFB1:
.cfi_startproc
vpbroadcastw (%rdi), %ymm1
vpxor %xmm0, %xmm0, %xmm0
vpblendw $1, %ymm1, %ymm0, %ymm1
vpblendd $15, %ymm1, %ymm0, %ymm0
ret
.cfi_endproc
