[PATCH] optimize vector fneg of bitcasted integer value

[Renato Golin]

Hi Sanjay,

The new code's "estimated" cost is so close the old one that I won't object. But I fear this might create some regressions, since the cost of using the vector pipeline in modern ARM cores is next to zero, and using both pipelines at the same time might actually be faster.

However, I tried to run them on an ARM OOO chip and made no noticeable difference, so go for it! Any relevant side effect should be picked up later and could be fixed on the ARM DAG legalizer, if needed.

cheers,
--renato

http://reviews.llvm.org/D4852