Currently Word32 and Int32 values are represented boxed. Some benchmarks (e.g., DLX) may benefit from using an unboxed representation of Word32 values. When GC is enabled, a possibility is to wrap Word32/Int32 operations in code that shifts values one bit. Thus the representation is [v] = (v<<1)+1.
Also, conversions like Word.fromLargeWord o Word32.toLargeWord are problematic as they involve boxing! Perhaps we can find a better way to deal with these conversions... Internally, we need operations like Word.fromWord32, which can be implemented without boxing also when GC is enabled.