[llvm] [PowerPC] Optimize bitcast(truncate) patterns using vbpermq/vbpermd (PR #181233)

[Nemanja Ivanovic via llvm-commits]

================

----------------
nemanjai wrote:

I agree 100%. The build vector vs. insert element issue is completely orthogonal to this, but I would imagine it comes up in real code and should probably be addressed in a separate patch.
This shows the issue quite clearly: https://godbolt.org/z/x4M4c7oaP

The inlined version does `li + rldimi + rldimi + mtvsrdd` whereas the non-inlined version simply does `xxlxor + vinsw` (of course the stack setup code isn't relevant).

https://github.com/llvm/llvm-project/pull/181233