Problems with C and C++ Separate Compilation

11 11 2008

After graduation, a couple months of watching television, driving cross-country (if you get the chance you should drive across northern Wyoming), settling in at Microsoft and living in Seattle, I’m back.  And I’m annoyed at C.

C is a fantastic language in many ways.  It is essentially an abstract assembly language.  Almost any general-purpose operation which can be done in assembly can be done in C, and it makes building large, relatively portable systems much easier.  The only things which can’t be done directly in C are operations on specific registers (and it’s easy enough to link in short assembly routines when that’s necessary.

Most of my early interest in programming languages and most of my problems when I first started doing systems work were related to basic typing issues: the ugliness of casting things to void pointers and back, the conversions between various integer types, and other relatively mundane C errors which are easy to make and hard to debug.  I came to believe that additional features other than type system and memory safety improvements in other languages, while extremely useful, were mostly great conveniences rather than fundamental improvements.

But the past several months have changed my mind.  While the ease of turning a pointer to one type of object into a pointer to another type in the same place is certainly a bane as often as it is a boon, a reasonably experienced C programmer begins to recognize common symptoms for such problems.  A more serious, though less frequently encountered problem has to do with type identity and versioning.

Consider the case where you write an application (in C) to use an external library.  Your application interfaces with this library through two means: #include-ing its public header, and being linked to the library’s object file.  Initially these two interfaces will probably be fine (if, for example, you just installed this library).  Now move forward a couple months.  Update your library.  Did your update include the object file and the header file?  If not then any of sizes or layout changes to the library’s data types might cause non-obvious errors; your application will happily compile and link, but the results you get back from the library may not be what you expect.

What if it’s your library, or just an object file in your project?  These tend to have a fair amount of turnover.  Most moderately-sized projects use separate compilation to separate code changes and avoid recompiling the same code repeatedly if it doesn’t change.  But when tying these object files together, there are no checks to ensure that data structures exchanged between object files are consistent; the C compilation model assumes that your data structure definitions are stable, or you recompile from scratch every time.  It also makes the reasonable assumption that the same compiler is used for every object file.  On the off chance you violate that expectation (perhaps with a compiler update), memory layouts of the same structure definition may differ between object files.

It’s possible to work around this problem with a build system if you track every header file dependency explicitly.  For large projects, this can be difficult.  Especially with fast-moving projects, it’s easy to add an include to a .c file without remembering to add the dependency to the build system configuration.  Once this missing dependency goes unnoticed for some time it becomes considerably more difficult to track down, and developers end up either spending their time debugging the build system or resorting to rebuilding from scratch every time in favor of the broken incremental build.

Another permutation of the same problem is that of unrelated structures with the same name.  It’s easy to imagine a large system with two subsystems defining structures named CALLBACK_ARGS.  What happens when one section of code needs to interact with both of these systems?  If all appropriate headers are included, then the name collision will be detected.  If only one of the conflicting headers is included, then depending on how the headers are organized it becomes trivially easy to pass the wrong structure to a function.  Especially when working on a new system, it usually seems reasonable to assume that structures of the same name are the same semantic (and in-memory) structure.

Namespaces can help alleviate the same-name problem: including only one structure’s header and trying to pass that to another function will result in an error complaining about passing an argument of type Subsystem1::CALLBACK_ARGS* to a function expecting a Subsystem2::CALLBACK_ARGS*.  This doesn’t actually prevent you from declaring two structures of the same name in the same namespace in separate header files, but if namespaces are used judiciously to separate subsystems then the likelihood of doing so accidentally is greatly reduced.

The versioning problem is a direct result of how #include works in C.  Rather than being a direct part of the language, #include is a preprocessor directive equivalent to “take the text of the specified file and pretend I typed it in place right here, then pass that result to the actual compiler.”  At its core most C compilers only handle single files at a time, so they don’t actually know anything about other object files (or at least, they don’t directly use information about other object files).  That’s the linker’s job, and the linker knows nothing about structures per se – only matching symbolic references.

One solution is to store all structure layout information in object files, and generate code for accessing those structures once at link time.  This slows the linking process, but prevents the mismatched definition problem; all code for accessing the structure is generated at the same time from the same definition.  This blurs the distinction between compiler and linker, but adds great value.

Doing this at compile time for static linking is relatively cheap and straightforward.  Doing this at load-link time is a bit trickier.  While compilers and static linkers can play any tricks they want for code which only interacts directly with itself, dynamically linked executable formats must be defined in standard ways, limiting what can be done.  I don’t know of any major executable formats which support this (most were designed in the heyday of C and C++, when they were still the best languages around), but that is a matter of format standards rather than a technical limitation.  This would be more expensive than current dynamic linking, but doable.   A compiler could choose to use a richer format for its own object files and then resort to standard formats when asked to generate a standard library or executable.  OCaml does this; for a Test.cmx and Mod.cmx compiled to objects using differing interface files for a Test module data structure:

Yggdrasil:caml colin$ ocamlopt Test.cmx Mod.cmx
Files Mod.cmx and Test.cmx make inconsistent assumptions over interface Test
Yggdrasil:caml colin$ 

Unfortunately C and C++ have a compilation and linking model which is now so well-established that I suspect any proposal to fix this in the standards for those languages would likely meet with significant resistance.  Though at the same time, I can’t think of any desired C\C++ semantics that this would break, so maybe it could happen.