[Mlir-commits] [mlir] [mlir][gpu] Add innermost-first policy when mapping loops to GPU IDs (PR #160634)

[Fabian Mora]

fabianmcg wrote:

> But I feel that loop interchange is different semantically; as it is an actual IR semantic transformation and can have more involved side-effects as mentioned above while the other is just a binding problem.

The `scf.parallel` semantics already say it shouldn't have any side-effects:
```
    Semantically we require that the iteration space can be iterated in any
    order, and the loop body can be executed in parallel. If there are data
    races, the behavior is undefined.
```
If a reordering produced an undesired side-effect, then the op is already in violation of its assumptions.

https://github.com/llvm/llvm-project/pull/160634