[Mlir-commits] [mlir] [mlir][scf] Refactor and improve ParallelLoopFusion (PR #179284)

[llvmlistbot at llvm.org]

================

----------------
fabrizio-indirli wrote:

Good point, in theory I'd agree with you; however it seems that lots of "innocuous" operations do not implement the `MemoryEffectOpInterface` yet (for example in the GPU dialect): if we enable this stricter check now, the pass won't trigger in lots of cases. I suspect this might disrupt lots of downstream uses of this pass, that are relying on the fusion to happen (this is also our case).
Note that the previous version of the pass wasn't checking this either - it considered only `memref.load` ops in the 2nd loop as potential data hazards. 
Of course ideally we should stop on unknown ops, but maybe for the time being we can ignore them to avoid disruption? In the code, I added a dedicated check where for now I just advance the walk when encountering unknown ops, but I added also a TODO to fix it in the future once more ops implement `MemoryEffectOpInterface`. Let me know if this could work

https://github.com/llvm/llvm-project/pull/179284