[Mlir-commits] [mlir] [mlir] introduce cf.assume (PR #220095)

[Mehdi Amini]

https://github.com/joker-eph requested changes to this pull request.

I think such an op is very useful, however there is an important design question that can on whether to make it work by "side effect" (it affects an existing value) vs making it inserted in the dataflow (it could return a new value with the assumption).  In general I prefer the dataflow version, because it is independent of the placement, can be hoisted and manipulated like a pure operation in general)

There is also an ambiguity in your current description about whether the assumption attached to an SSA value only takes place after the assume is "executed" (path/flow sensitivity).

https://github.com/llvm/llvm-project/pull/220095