ICFP presentation

[Matthew Fluet]

> I agree wrt (b).  The number of remaining functions as a bar chart using the
> same six benchmarks would be nice.  You could even normalize stuff to show how
> many original functions were there as well.  

I think that just a chart of the normalized function counts will be
enough.  It's going to get too complicated with both absolutes and
normalized, and I think its easy enough to see the trend on the normalized
graph.

>   * that A_call and A_dom consistently beat A_cont

count-graphs is anomalous in this respect

>   * the positive correlation between amount of contification and run time

as a general trend, yes.  For any individual benchmark, there are
anomalies.

>   * that A_dom almost always contifies more than A_cont -- this helps bring out
>    the point about anomalies 

the only exception in this respect is barnes-hut, which is not among the
six that are currently on the slides.  But, I went ahead and added it.
Actually, it's a very good correlation b/w functions contified and
run-time.

> As to (a), I think the <4% is enough.  And we're already hurting for time.

I'm thinking of dropping the Compile-time Performance slide altogether and
replacing it with a single Compile-time Statistics slide with the
normalized function counts and three bullets:

* contification runs three times in the optimizing cycle
* contification takes less than 4% of total compile time
* contification quiesces after two rounds

The fact that contification eliminates 1/2 the functions is now implicit
in the graph (not to mention that Acont actual only eliminates 40% of the
functions in mlton).