[Mlir-commits] [mlir] [MLIR][OpenMP] Simplify OpenMP to LLVM dialect conversion (PR #132009)

[Sergio Afonso]

skatrak wrote:

One compromise I made for this initial implementation was adding the `SupportsMemRefOperand` template parameter to the `OpenMPOpConversion` class, which is only set to `false` for `AtomicUpdateOp`, `AtomicWriteOp`, `FlushOp`, `MapBoundsOp` and `ThreadprivateOp`.

This mirrors the previous behavior, where conversion would fail if an operand of `MemRefType` for the operation was found, and this was only checked for that (seemingly arbitrary) set of operations. I would like to propose not checking that in this conversion pass at all and simplify things further. If the types memrefs are converted to are an issue, it seems like we could trigger errors later during OpenMP to LLVM IR translation.

https://github.com/llvm/llvm-project/pull/132009