If you are using -ftrapv, then you must implement the overflow handler. This can be used to throw an exception, abort, or some other behaviour (I use it in LanguageKit, for example, to box the result in an arbitrary-precision integer object, and we've also used it to implement some proposed C1x overflow handling drafts).
The signature for this function is:
long long overflow_handler(long long val, long long otherval, char op, char width);
If this function returns, then the result will be truncated and used in place of the real result. The op and width arguments represent the operation and the size of the arguments.
The function should be assigned to a function pointer declared like this, in some library that your code links against:
long long (*__overflow_handler)(long long a, long long b, char op, char width);
A sample implementation of this was part of the original patch that I sent implementing -ftrapv, but it was not added to svn because there was no clang runtime library. If you check the list archives, then you can find it.
You are correct that this performs run-time overflow checking. Compile-time overflow checking should be a default option, and not dependent on -ftrapv, but is only possible in a relatively limited number of cases. This option is intended for use in functions like calloc() that need to do some special handling if the result of multiplying two values together is 0. A trivial handler for this case might simply return 0. You would then implement a test-for-0 after the multiply and return NULL for 0-sized allocations.
One of the things that I never got around to adding was support for per-compilation-unit handler functions. Ideally, it would be nice to be able to do -ftrapv=my_handler and have my_handler() called on overflow.