Currently our AArch64 backend only supports MulReduction for 64/128 bits vectors by using scalar instructions instead, as it's missing corresponding vector instruction. That's caused a performance gap.