[PATCH] [X86][SSE] Avoid shuffles with zero by using pshufb to create zeros

[Simon Pilgrim]

Thanks Quentin.

A basic timing test of the pshufb vs 2xpshufb+por core loop gave a 30% improvement on my older Core2Duo machine (I guess due to throughput limitations), but this diminished to less than 5% on SandyBridge. However, its main use is the reduction in register pressure, as well as the obvious fact that it was pointlessly shuffling zero vectors.


REPOSITORY
  rL LLVM

http://reviews.llvm.org/D6878

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/