During code generation for a CmmSwitch, if we do not have a default branch to take (switchTargetsDefault = Nothing), generate the "morally correct" LLVM IR code, where by default we branch to an unreachable basic block.
I benchmarked this (possibly incorrectly) using nofib, and saw no increase or decrease in performance.
However, I still believe this should be upstreamed since it informs the LLVM optimizer about the possible values that will be branched upon.
Feedback on making the code better is greatly appreciated :)