They both compile to a memset so no compilation time in the first case leads to something useful.
Most unrolling should be fairly pointless for MMatrixes since LLVM should just unroll it by itself.
The different reduce functions (https://github.com/JuliaArrays/StaticArrays.jl/blob/master/src/mapreduce.jl#L71 & co) would also likely gain by just being rewritten as loops instead of forcing unrolling.