The problem is that normal order reduction can be inefficient, so
compilers can change the reduction order provided this does not affect
the termination properties. They may use strictness analysis to decide
when this is safe, but this is not mandated nor is the algorithm used
standardised. As strictness analysis is AFAICS effectively the halting
problem in disguise, it has no general solution. Some compilers may well
do a better job than others, hence they won't all use same reduction
order or exhibit same space behaviour for the same source program. In
fact even with the same compiler you may get completely different space
behaviour for the same program, depending on optimisation settings.