[PATCH] D32563: Add LiveRangeShrink pass to shrink live range within BB.

[Dehao Chen via llvm-commits]

On Fri, Apr 28, 2017 at 10:23 AM, David Li via Phabricator
<reviews at reviews.llvm.org> wrote:
> davidxl added a comment.
>
> A couple of comments:
>
> 1. Moving it out of the reassociation pass is probably the right direction to go.
> 2. having this pass allows future enhancement
> 3. This actually allows you to do more aggressive expression operand re-ordering (to enable more shrinking) -- but to void offsetting reassocation's effect (to enable cse,  constant folding, loop invariant code motion etc), the shrink pass should be pushed very late in the pipeline.

I've moved the pass to the end of the IR passes. Another thing that
might worth extending is to integrate BFI and DT into the pass and
make it allow instruction to be moved out of the BB if BFI shows
beneficial. But for the first version, I'd like to keep it simple. We
can add more aggressive optimization later when we find compelling use
cases.

Dehao

>
>
> https://reviews.llvm.org/D32563
>
>
>