In a comment on the previous question, I claimed:
I have another benchmark to indicate ghc-7.4.1+llvm does unpacking of enumerations on strict data fields.
In fact, after some experimentation, I believe that, in at least some simplistic cases, using an enumeration is at least as fast as, and actually might be more memory efficient (hence faster in a more realistic application) than using a newtype of Word8 (or Int), even when used in a strict field of a datatype. As I said in the previous question, I have experienced the similar phenomenon in a more realistic (but still small) setting.
Can anyone point me to some relevant references as to what optimizations ghc/llvm does for enumerations? In particular, does it really unpack the internal tag of the enumeration on a strict data field? The assembly output and the result of profiling seem to suggest that that is the case, but to me there is no indication on the core level. Any insights will be greatly appreciated.
One more question: Are enumerations always at least as efficient as newtypes of the corresponding Integral, where using them makes sense? (Note that enumerations can behave like Integrals,too.) If not, what is the (hopefully realistically useful) exception? Daniel Fischer suggested in his answer that putting an enumeration on a strict field of a multiple-constructor datatype might prevent certain optimizations. However I have failed to verify this in the two-constructor case. Maybe there is a difference when putting them in a large multiple-constructor datatypes?
I'm also curious about what exactly is happening in the following benchmark. In all the four cases, roughly the same amount of bytes were allocated in