Of course there have been regressions, but the semantics of the code don't typically change. A "dead heap allocation", or removing an unused call to malloc, is an interesting example but I think you'd be hard pressed to find a case where it caused a massive performance regression.

I think there may be some edge cases (constant folding, dead code removal) of course, but a change to LLVM cannot cause the 3 sorts of changes I identified in Go, Haskell, and JavaScript.