https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115833
Hongtao Liu <liuhongt at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |liuhongt at gcc dot gnu.org
--- Comment #3 from Hongtao Liu <liuhongt at gcc dot gnu.org> ---
> It seems the very bad code generation is mostly from constructing the
> V4HImode vectors going via GPRs with shifts and ORs. Possibly
> constructing a V4SImode vector and then packing to V4HImode would be
> better?
void v4hi_contruct(signed short *t, signed short tt, short tt1)
{
t[0] = tt;
t[1] = tt1;
t[2] = tt1;
t[3] = tt1;
}
void v4si_contruct(int *t, int tt, int tt2)
{
t[0] = tt;
t[1] = tt2;
t[2] = tt2;
t[3] = tt2;
}
v4hi_contruct(short*, short, short):
movzx eax, dx
movzx esi, si
mov rdx, rax
sal rdx, 16
or rdx, rax
sal rdx, 16
or rdx, rax
sal rdx, 16
or rdx, rsi
mov QWORD PTR [rdi], rdx
ret
v4si_contruct(int*, int, int):
vmovd xmm2, edx
vmovd xmm3, esi
vpinsrd xmm1, xmm2, edx, 1
vpinsrd xmm0, xmm3, edx, 1
vpunpcklqdq xmm0, xmm0, xmm1
vmovdqu XMMWORD PTR [rdi], xmm0
ret
both vmovd and vpinsrd is expensive, and v4hi_contruct is not necessary worse
than v4si_construct, but v4hi_construct can be optimized to be a little more
parallel via GPRs.
