Doug Way <dway at riskmetrics.com> wrote:

> Type collecting is nice because it can run quickly, whereas a thorough type inferencer can be pretty slow.  The downside to type collecting is that sometimes there will be no instances of a given class in the image to collect from.  (Although this is often not a problem because as a general rule you're usually editing classes which you're actually running in the image.)  I was thinking about adding Dan Ingalls' simple type inferencer to Whisker as a backup method for when the type collector can't find any instances.  Also, I think Lex Spoon is working on a more thorough type inferencer... I'm not sure what the status of that is.  Anyway, this is sort of a digression from your post.
> 

Yeah.  The problem is that my inferencer gets bogged way down in too
many cases.

Another observation on the difference between what Whisker does and what
tools like mine do is that Whisker's is a *lower* bound on the types,
and mine gives *upper* bounds.

I believe Dan's just makes a complete guess, based on how the code seems
to use the variable, so it's not known to be an upper or lower bound.  
It sounds like the guesses are very good, though.  So, a good
combination for people who want results *right now*, would probably be a
Whisker-style query, followed by a Dan-style query if there are no
current instances of the class.

-Lex