[Mlir-commits] [mlir] [mlir]introduce UnrollScopeInterface and apply it to funcOp and gpu.launch Op. (PR #123904)

[Oleksandr Alex Zinenko]

ftynse wrote:

> I claim that the PR you actually want is to change AffineLoopUnrollPass to work on an arbitrary operation, thus allowing you to -pass-pipeline='builtin.module(func.func(gpu.launch(affine-loop-unroll)))
>
> I can't think why the affine loop unroller needs to run on functions specifically

Such a change would still be welcome. The loop unroll pass was written before we turned functions into operations, that's why it is historically a function pass, but there is no reason for that apart from history. Then it can run either on `gpu.launch` or on `gpu.func` (it currently cannot!) after GPU outlining.



https://github.com/llvm/llvm-project/pull/123904