contification paper

[Stephen Weeks]

> Yeah, I'm trying to decide which are the right ones to include.  Two
> problems with my setup, though.  I couldn't compile mlkit with the
> snapshot that I took a few days ago; I ran out of memory with -max-heap
> 220m.  Which also clearly means I can't get G2 MLton to compile.  I know
> those are the two biggest benchmarks; I can get the function counts using
> a G0 MLton, but not compile times.

Yeah, you really need 400m for MLton and Kit compiles.  The right thing is
probably for me to rerun those two on a 512M machine here.  We just need compile 
times and runtimes, right?

> But, here are the one's I'm thinking of:
...
Fine.

> (Maybe just include self-compile static counts? 

I'd prefer to run 'em here.

> What do you use as a
> running time for the MLton benchmark?)

I don't remember, maybe compiling mlyacc.

> As another note, in thinking about the benchmarks section and talking
> briefly about the "real" CPS IL, I instrumented the transformation to
> count the number of functions it has to nest because of the mutual
> recursion restriction.  Useful statistic: over all benchmarks with the dom
> analysis, there are only 38 functions that need to be nested in this way,
> out of over 2700 contified functions.  (And the 38 might even be
> overcounting, because it will repeatedly count a function if the recursive
> calls to sccHeads keep nesting the same functions.)  So, I think we can
> reasonably say that it has negligible effect.

agreed.