[vector] Some fine-tunes for AArch64 reduce_add operations

[Qi Feng (Arm Technology China)]

Hi all,

Could you please help me review this patch?

In this patch, I make some fine-tunes for AArch64 reduce_add operations.
As reduce_add operation can be applied to signed values, we should use
sign-extension when possible. And using faddp instead of fadd can reduce the
instructions from 3 to 2.

http://cr.openjdk.java.net/~yzhang/vectorapi/vectorapi.finetune/webrev.00/

Thanks,
Qi