Clang.File.getIncludedFile causes segfault on any code about libclang HOT 1 CLOSED

I'll write down the process of debugging this as I do it:

My first approach is to run the executable under lldb to see if the crash is happening in C/C++ code. If it is, this will give me a stack trace and usually the problem is immediately clear.

If the crash is happening in Haskell code, lldb is generally not as useful. In this case I usually add additional print statements to the crashing code to determine exactly where the crash happens. (I could also use ghci for this, but editing the actual source is convenient since you can run it over and over and verify whether the problem is fixed.)

Generally knowing where the crash is tells me the source of the problem immediately. In this case the crash happens in Clang.String.unpack, which means Clang.Cursor.getIncludedFile returned a null value and probably needs to return a Maybe (File s) instead of a File s.

At that point looking at the source of Clang.Cursor.getIncludedFile will tell me the low-level FFI function that it calls - in this case Clang.Internal.FFI.getIncludedFile, which is defined in FFI.gc. I can see from there that the C-level function being called is clang_getIncludedFile. (The mapping between the various names is obvious in this case but sometimes it's less straightforward).

To verify my guess that clang_getIncludedFile may fail and return null, I check its definition at the C level in clang/include/clang-c/Index.h. It seems pretty clear that it's only designed to be called on inclusion directives. At the Haskell level, then, we definitely need it to return a Maybe, since we have no type-level way of restricting the kind of cursor you can pass in.

To fix this, then, I just need to update the definition of getIncludedFile in FFI.gc and the functions that call it. The details of doing that require some knowledge of greencard but it's pretty easy to pick up. In this case we've needed this in several other places and there's already a maybe_file "dis" available, so it's a very simple change.

You can view the referenced commit to see exactly what I did. If you feel comfortable debugging issues like this, we'd love to have the help!

from libclang.