https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91546
--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Instruction count is not everything.
If instructions can be executed/issued together (I don't know x86 processors
that well), then GCC produces is better.
E.g.
clock 0
vmovd xmm2, edx
vmovd xmm3, edi
clock 1
vpinsrd xmm1, xmm2, ecx, 1
vpinsrd xmm0, xmm3, esi, 1
clock 2
vpunpcklqdq xmm0, xmm0, xmm1
clock 3
While clang/LLVM:
clock 0
vmovd xmm0, edi
clock 1
vpinsrd xmm0, xmm0, esi, 1
clock 2
vpinsrd xmm0, xmm0, edx, 2
clock 3
vpinsrd xmm0, xmm0, ecx, 3
clock 4
