llvm-hs segmentation fault with output of clang -g

[andrew-wja]

The EncodeM and DecodeM instances for debug metadata are currently quite fragile, and there are several places where a segmentation fault could be triggered by debug metadata structures which we don't handle, or don't handle correctly.

In particular, there are multiple places where we use maybe (return nullPtr) encodeM to pass a null pointer to the LLVM C/C++ API where we have Nothing on the Haskell side. An audit of the encoding needs to be carried out to make sure that we only pass null pointers to LLVM where the LLVM API explicitly accepts a null pointer as an indicator of the absence of something, and not anywhere else. We currently have at least one null pointer being passed to an LLVM API function which dereferences it, leading to the segmentation fault when attempting to load the IR produced by clang -g.

[aboeglin]

@andrew-wja
Can this explain the failing tests? I'd be happy to look into it but I'd need some guidance.

[aboeglin]

Nevermind, I dug into it last night and the failing tests are related to wrong types for DICompileUnit and DILocalVariable.