Good test suites for real languages are expensive to create and maintain. There's a reason that the Plum Hall test suite, which is industry standard for ANSI C, is so bloody expensive.
George Necula's translation validation is a brilliant idea but also quite expensive to implement.
The one thing that's cheap and easy is this: maintain a suite of regression tests, and every time you fix a bug in your compiler, put a suitable test into your regression suites. With compilers, it's unbelievable how easy it is to keep reintroducing the same bug over and over. Disciplined additions to your regression suite will prevent that, and they don't cost much.
