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

[Mehdi Amini]

joker-eph wrote:

> I'm not necessarily a fan of having an additional vocabulary of assumptions (divisible-by, ranges, etc.) as attributes, this feels duplicative wrt operations. Hence the idea of potentially having a region instead. Note that this is orthogonal to the side-effect vs. value discussion.

Right, I'm not proposing we adopt the same thing to be clear. We can also look at the shape.assuming op where the computation is a region and the predicate is a "witness" (I think you're proposing the opposite right?), or something flat where we take witness/bool predicate and values that are returned and ensure a true predicate.

> What's your take on flow-sensitivity?

I don't like it too much :) 
It complicates analysis.

That said even without flow sensitivity, we're still "side-effecting" here right? Otherwise I'm not sure how it works with  assumes that would be path-sensitive. Basically I'm afraid of the composability aspect of this all. 

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