[LLVMdev] SCEV ordering

[Chris Lattner]

On Mon, 23 Apr 2007, Dan Gohman wrote:
> I think what I was missing is that the reassociate pass itself usually
> saves the day. With the -std-compile-opts passes, reassociate is run before
> the SCEV-using passes, so most chains of adds are already in a canonical
> order. There still are some cases where having ScalarEvolution do its own
> sorting does reduce the total number of SCEVs allocated, but I now see that
> it's pretty rare in practice.

Ok!  If you find a specific example, please do find a bugzilla.  This is 
the sort of thing that is very difficult to notice while studying .s 
files, yet can have a big impact. :)

-Chris

-- 
http://nondot.org/sabre/
http://llvm.org/