Warn if compartment call returns void

[rmn30]

It would be good if clang would warn if the return type of a compartment call is void. Compartment calls can fail unexpectedly and return -1 (e.g. due to trusted stack exhaustion). Callers should be aware of and check for this, therefore it is never appropriate for a compartment call to return void. For the same reason we should also consider automatically adding the nodiscard attribute to compartment calls.

[v01dXYZ]

Hi, I'm working on this. How to deal with C++ constructors with a cheri_compartment attr ?

[v01dXYZ]

I think it is worth to take the c++ cases into account (non static method, constructor/destructor, lambda ?, etc.). But it would be better to write another ticket for that.

[davidchisnall]

I'm not sure what it would even mean to expose a constructor across a compartment boundary. A factory method, maybe, but what would the security model be for a constructor? Non-static methods can't be compartment exports, because they need to in the vtable and the ABI does not allow that. It would be possible to make a lambda's call operator a cross-compartment call, but there's no way of exposing that in the source language.

[v01dXYZ]

I'll open a ticket about generating an error when non static methods are used as compartment entry points. (Or you can if you have other edge cases in mind).

There are more things to mention as I progress:

• in Attr.td (the tablegen file defining attributes), cheri_compartment has Subjects = [Function]. Contrary to FunctionLike, it does not include function pointers. It seems strange to me but there is maybe a reason.
• cheri_compartment is a Calling Convention attribute. Clang considers it as a Function type attribute. Thus, when the function declaration is annotated with such an attribute, clang will also try to add this attribute to the function type. But there is a problem, later when trying to handle the attributes, it will iterate first on the different parts of the type and then on the declaration itself (at each iteration, it essentially takes the attribute list and call the same processing function on it). Which means the warning will be shown twice. I could disable that but it will also disable showing multiple warnings when there are redundant cheri_compartment attributes. Which path should I take ?

[davidchisnall]

To the first, yes this makes no sense for function pointers. Function pointers must use the Cheri callback convention. They are not (and cannot be) explicitly tied to a compartment.

I think showing the warning twice to start with is fine for a first cut.

[v01dXYZ]

I wrote a first version for cheri_compartment. I reported the duplicate attribute problem caused by GetTypeForDeclarator which is the root cause of the previous problem (https://github.com/llvm/llvm-project/issues/116543).

There are still to support cheri_ccall and cheri_ccallback. But this is a little bit more complicated than expected as the attribute is added to the AST in a different part where it's more difficult to also add an additional WarnUnusedResult attr. Emitting a diagnosis about void return type could easily be done though.