The use of self-compilation speed as a benchmark seems like a great way to target and then maintain a balance between complexity and the performance benefits of optimizations in a compiler. Complex optimizations where the cost of optimization in terms of code complexity, outweigh the benefits of faster binaries could be clearly recognised with such a metric.
An interesting idea for seeing the tradeoffs between compiler design/implementation complexity and speed.