Avoid assertion failure from SetFuncAllowAttrs when struct is empty

[tim-hoffman]

No description provided.

[iangneal]

I'm not sure that we should allow empty structs, since an empty struct can never be used---a call to @compute or @product is required to create a struct (since those are the only places where struct.new is allowed). Is there a use case for this?

[tim-hoffman]

I'm not sure that we should allow empty structs, since an empty struct can never be used---a call to @compute or @product is required to create a struct (since those are the only places where struct.new is allowed). Is there a use case for this?

That's a fair point. I ran into this in a unit test that uses the C API to create a StructDefOp and then calls the verifier (without generating anything else within the StructDefOp). It had an unhelpful assertion failure within getBody() so I avoided that with this PR.
Here's the weird part. The test uses the C API wrapper for mlir::verify() (in mlir/lib/IR/Verifier.cpp) and we do have the check you mention in StructDefOp::verifyRegions() but I guess the former does not call the latter. In fact, I'm just now noticing that it too calls getBody() which would fail in my test. Let me try a few things and I'll probably push more to this PR.

[tim-hoffman]

@iangneal I fixed the same issue in StructDefOp::verifyRegions() as mentioned earlier. The difference between mlir::verify() not checking regions and verifyRegions() doing so should be expected behavior I believe.

[iangneal]

LGTM.