[Mlir-commits] [mlir] [mlir][vector] Add support for yielding loop bounds in `scf.for` distribution. (PR #163443)

[Charitha Saumya]

charithaintc wrote:

> Yeah, sure, I did not mean to hoist anything. I meant that scalar values coming from the outside may be referenced directly, not necessarily via the warpOp arguments. Meaning that some pass, similar to hoisting in terms of when it should be called, would clean up things beforehand, simplifying the work for the actual distribution patterns.

agreed. uniform scalars/vectors are hoisted before even we begin distribution. I guess this is to reduce compile time (no need of repeated application of WarpForward + WarpDead) . but unfortunately that does not help us here. 

https://github.com/llvm/llvm-project/pull/163443