[Mlir-commits] [mlir] [MLIR] Testing arith-to-emitc conversions using opaque types (PR #137936)

[Niklas Degener]

ndegener-amd wrote:

I'm gonna elaborate a bit on our use case. We had the issue, that our own opaque type conversions couldn't be tested in llvm and we had to do it locally. But for testing opaque type conversions, first of all a conversion is needed which can be tested. So the idea was to use some types, that are unsupported in emitc anyway, and convert them to opaque types. There is no specific reason for choosing f80 and i80 except them not being supported.
During testing, I noticed that a lot of the ArithToEmitC conversions don't support opaque types, so I went on and legalized them.

As you mentioned, this does introduce the issue that i80 and f80 are unconditionally legalized and no longer recognized as unsupported. But the ArithToEmitC pass is essentially only a testing pass, so I think it shouldn't matter.

https://github.com/llvm/llvm-project/pull/137936