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

[Charitha Saumya]

charithaintc wrote:



> The case is valid, although generally I'd expect scalar usages inside warpOp to be "forwarded" to outer definitions beforehand, similar to how scalar definitions are hoisted.

I get your point. But above not always true. In flash attention loop bounds comes from kernel args. then they can not be hoisted like scalar constants. hence compilation fails in flash attention. Even the test case I have in this PR also fails (much simpler). 

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