I find this very strange.  Gcc accepted some parallelized STL
implementation too in recent years, and I always questioned it.  For
example, 25.1.1/1 in c++98 (for_each def'n):

 Effects: Applies f to the result of dereferencing every iterator in
          the range [first, last), starting from first and proceeding
          to last - 1.

How could you possibly parallelize that?  It's defined to operate in
series.  I guess if the compiler could prove that `f' had no side
effects, and there were no volatiles involved, it might be possible (is
that strong enough to prove that that execution order is irrelevant?
hrm..)  It seems like the former would run up against solving the
aliasing problem, though.  What am I missing here?

(Granted, I can't recall the last time I wrote a for_each that *relied*
on that execution order.  Further, I could probably be convinced that
requiring a particular ordering here was a specification mistake.
Still, a standard is a standard.)

-tom