No-compile Tests
No Compile Tests are used to verify that certain coding constructs will fail to compile. These were developed to test the base/functional/callback.h and base/functional/bind.h constructs which need to ban certain types of assignments in order to preserve type safety. The tests however can be used to enforce any sort of compile-time check, such as static_assert.
Why do we want to assert failure in compilation?
Most of the time you don't want to do this. However, with template code, you are playing with the type system. In particular with callbacks, it's possible to accidentally write code that allows unsafe type conversion leading to a code execution. Here's a specific example:
int func(double f) {
printf("woohoo! Called with %lf!\n", f);
}
Callback<int(void)> c = Bind(&func, 1);
If you look carefully at the types, Callback<int(void)> has no concept of how many arguments func originally needed, or what types they were. Somewhere between the call to Bind() and the assignment, the type of func:int(int), and the type of 1:int was captured and erased so that it could be assigned into a Callback that only knew the resulting function was int(void). To do this, there is some casting happening that can break type safety. In the Bind() implementation, the code has been constructed carefully to fail at compile time if you were to try and do something like:
Callback<void(void)> c = Bind(&func, 1);
However, the system is complicated and regressions can occur. In fact, they already snuck in once. To ensure we do not regress in behavior and allow a dangerous type safety violation, we need a set of "no compile tests" that assert that known dangerous constructs will not silently compile. This makes it much safer to modify the Bind() implementation.
Though this uses the Callback/Bind system as an example, any code that ends up needing to ensure a certain set of compile time constraints aren't violated can benefit from these kinds of tests.
How to write:
- Create a ".nc" source file.
- Import build/nocompile.gni
- Include the ".nc" source file into a sources section of a nocompile_test target (eg., base_nocompile_tests).
The nocompile.gni import adds a template target that will process the .nc files and produce a .cc file that represents the test results. On a failed no-compile test (aka success compilation), the generated .cc file will consist of a set of #error lines holding the compiler output. This will cause the unittest to fail to build. If all tests succeed, then the .cc file will contain a set of unittests representing each no-compile test which trivially succeed. This lets you know that the tests were at least run.
The comment at the top of build/nocompile.gni has most of the documentation for how to use this system.
The verification logic is built as part of clang. Full documentation is at “Clang” CFE Internals Manual > Specifying Diagnostics