[PATCH] D33690: [PowerPC] Match vec_revb builtins to P9 instructions.

[Nemanja Ivanovic via Phabricator via llvm-commits]

nemanjai added a comment.

In https://reviews.llvm.org/D33690#768675, @inouehrs wrote:

> I agree to Nemanja's comment. Since we have so many special patterns of permutation (e.g. byte reverse, rotate, shift, merge, splat, pack, unpack etc etc), we need a robust framework to optimize these patterns. Maybe we can create a table that maps a shuffle pattern to a specific instruction.
>
> In complex case that can map onto a couple of instructions, there is a trade-off between consuming permutation pipeline resource and a vector register; e.g. "reverse-bytes-within-each-word-and-rotate-left-by-word" can be executed by one permute instruction using one additional vector register for the shuffle pattern. In hot loops, one permute approach may be a better choice than 2 instruction approach since we may be able to prepare shuffle pattern out side loop (if we have enough vector registers).


Of course, we have the "Perfect Shuffle" solution that works on BE and for v4i32 only. We could extend it to work on LE and to support the new instructions. However, I'm not convinced that's the best (or fastest) way forward.


https://reviews.llvm.org/D33690