[MLton] benchmarks with -representation {normal,packed}

[Matthew Fluet]

> Now that the cast problem is fixed, all the regressions (except for
> fixed-integer) and benchmarks work with -representation packed.  Here
> are the benchmark results.  There was small speedup on some
> benchmarks, and only one with noticable slowdown.

Where's the speedup coming from?  Avoiding GCs?  Less memory traffic in
allocating heap objects?  I could imagine a slowdown due to extra
bit-fiddling to manipulate packed objects.

I would expect packed representations to pay off in programs that are
using a lot of memory and doing frequent GCs (so, I'm interested to hear
if there is a self-compile speedup), but I didn't think that applied to
many of the benchmarks.

> There was conistent
> small improvement on code size, up to a few percent.

Again, I would have thought the bit fiddling of packed representations
would have increased code size, rather than decrease.