[PATCH] D47999: cl::opt ForceTrackRegLiveness for forcing liveness tracking on MIR (in MIR-Parser)

[Matthias Braun via Phabricator via llvm-commits]

MatzeB added a comment.

>From what I remember our use of this flag is strange:

My understanding of the flag is that it mostly indicates that the live-in lists are up-to-date. Which only really has a meaning after regalloc. Yet we always set the flag at the beginning when a MachineFunction is created and just happen to clear it for the targets that don't track liveness past regalloc (which should be close to noone nowadays). The flag keeps being very confusing to people IMO as it doesn't really tell you about other liveness tracking mechanisms like LiveRegister or LiveIntervals.

Either way: This patch is a strange solution, I'd much rather see us do some cleanup in the way MI handles it. You could try keeping the flag disabled by default and only set if in the regalloc pass for the post-ra passes...


Repository:
  rL LLVM

https://reviews.llvm.org/D47999