Re: [PATCH][C++] Improve PR15855 (compile-time/memory hog)

[Daniel Berlin <dberlin at dberlin dot org>]

> > > The second big incremental update has:
> > >
> > > Number of virtual NEW -> OLD mappings: 5443467
> > > Number of real NEW -> OLD mappings:          0
> > > Number of total NEW -> OLD mappings:   5443467
> > >
> > > Number of virtual symbols: 21950
> >
> > Uh, why did we need 5 million mappings here?
> I haven't a clue.  Whatever's going on is certainly excessive,
> that's about all I know.  I suspect whomever figures this out
> will get a nice little speedup on 15855 :-)

BTW, i did look at changing the heuristics for SFT generation for this 
case.
I can't do anything that won't significantly affect the quality of the 
interprocedural call clobbering and alias analysis stuff i'm working on.
(In fact, the current pruning heuristics are already too conservative in 
that regard, and i'll rely on a bit of oracle work to make it work).
We need to be able to name most memory locations in structures to be able 
to say what is and isn't clobbered.


> Jeff