[llvm-dev] RFC: Shrink wrapping vs SplitCSR

[Kit Barton via llvm-dev]

Hi all, 

We've seen several examples recently of performance opportunities on
POWER if we can improve the location of save/restore code for
callee-saved registers. Both Nemanja and myself have discussed this with
several people, and it seems that there are two possibilities for
improving this:

  1. Extend shrink wrapping to make the analysis of callee-saved
     registers more precise. 
  2. Focus on enabling and (possibly) improving SplitCSR. 

I would like opinions from people on the preferred way to proceed. 

I am leaning toward improving shrink wrapping, at least as a short-term
solution. However, I fully admit that this is because I am familiar with
the shrink wrapping code and completely naive about SplitCSR and what
work would be necessary to get this working well.

My proposal would be to implement the flow sensitive analysis described
by Fred Chow (PLDI '88) and make the necessary extensions in shrink
wrapping to handle multiple save/restore points. At that point we can do
an evaluation to understand the improvements it provides and the impact
on compile time. Once we have these results, we can look at the best way
to enable it (e.g., option, target opt-in, higher opts, etc.). 

Thoughts?

Kit