re new SML/NJ

[Matthew Fluet]

> As to moving the GC check out of the loop, doesn't that screw up the fact that
> it is also used for interrupt checking?

You are right.  Although I thought that there was a no-threads hack to
avoid inserting these limit checks.

> Mind you, if we have a lot of
> non-allocating loops, then we could also store the signal received flag in
> a separate global so you could just load it and branch if it is non-zero.
> Maybe this is no faster than the limit check any way.

It shouldn't be any faster.  I already special case the GC check for 0
additional bytes to just compare frontier (which is in a register) with
the GC.limit (which is a global), so I think that's just as fast as a
direct test of a global.