[PATCH] D69498: IR: Invert convergent attribute handling

[Mehdi AMINI via Phabricator via llvm-commits]

mehdi_amini added a comment.

What matter isn't "convergent" in itself, it is how it restricts transformations: if you know that all the control flow is always uniform, are there still restriction on any transformation in presence of convergent instructions?

If not, then the "target approach" seems nice: as @foad and @sameerds noted we just need to have a guarantee on these targets about the uniformity of the branches, it does not require to change the convergent semantics on a target basis I think.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D69498/new/

https://reviews.llvm.org/D69498