[llvm] [RISCV] Optimize conditional branches that can be statically evaluated (PR #131684)

[Philip Reames via llvm-commits]

preames wrote:

Two related, but slightly off topic points.

1) The only use of MLI in PeepholeOpt is in the precondition for optimizeRecurrence.  Looking at that code, the primary heuristic seems oddly limited.  We could likely generalize this to work based on overlapping liveness (instead of recurrence cycles), and remove the MLI dependence.  Haven't fully explored this, but seems at least plausible.  

2) Have you looked at updating MachineLoopInfo instead of invalidating?  We do a bunch of this over IR Loops, but it should be doable on Machine Loops too.  

https://github.com/llvm/llvm-project/pull/131684