splitting generated code up into multiple files that can be separately compiled, to check that the behavior remains the same when whole program optimization is used.
We did this, though I'm not sure the functionality is currently working or even exposed to external users.
It found some early bugs in GCC's LTO but they tended to be relatively shallow bugs that were going to get fixed anyway, so I never spent much time doing this sort of testing.
In GCC at least, the optimizations that fire are link time are the same ones that fire at module compilation time, so the opportunity for finding extra bugs isn't that huge. LLVM operates the same way. If a compiler had a totally separate whole-program mode, this would be more useful.
John