Re: Compilation performance comparison of GCC 3.4.2 and GCC 4.0.0(041024) on MICO sources

[Kenneth Zadeck <zadeck at naturalbridge dot com>]

Jan Hubicka wrote:

> > On Mon, 25 Oct 2004, Jan Hubicka wrote:
> >
> >    
> >
> > > > > My bet: GC allocating lots and lots of bitmaps :-/
> > > > >          
> > > > Probably.
> > > >
> > > > Here's the interesting question:
> > > > If these bitmaps live over compilation, and we know exactly when they die
> > > > (cgraph_reset_static_var_maps), why the heck are they gc allocated?
> > > >        
> > > Ineed ;)
> > >
> > >      
Part of this is my lack of understanding of what is really happening in 
cgraph* and how compilation proceed if multiple files are put on the 
command line. 

The bitmaps are live over the time that you compile a single file.  No 
one has actually explained to me how multiple files are compiled.  I 
just assumed that it was my responcibility to clean up before the next 
file was compiled.  Since the cgraph node where these bitmaps lives is a 
ggc data structure, I just assumed that all of the substructures inside 
it also had to be ggced also.

> > > Also the current usage of these bitsmaps is disabled and Kenny already
> > > has patch that reduce good amount of them when inlining takes place.
> > > We probably need to decide what to do about mainline - either disable
> > > collecting the information too or try harder to fix it.
> > >      
> > If his patch to fix this stuff up for the mainline (which was submitted 
> > and is awaiting review from you and diego, i believe, since it touches 
> > cgraph and tree-ssa) isn't accepted, we should probably disable it until 
> > 4.1
> >    
>
> We apparently went here into trouble with lack of maintainer
> understanding both areas.  I already told Kenneth that the cgraph
> changes are good but obvioiusly never did format approval, sorry for
> that.
>
> The cgraph bits are OK with the cgraph-* file renamed to ipa-* file 
> (the cgraph prefix for ipa optimizers is bit confusing and we already
> have both with ipcp code).  Thinking about it we probably want cgraph
> prefix only for the cgraph itself use ip/ipa/ipo prefix for the
> optimizers and analysis and ipa is already taken.
>
> But my understanding is that it (except for Diego's approval) still need
> fixing in the global variable initializers area, right?  I plan to at
> least apply the patch into tree-profiling branch once we sort out the
> inliner issues so we can make progress there.
>
>  
I am about to submit a patch to remove the code from the mainline 
because of the not finding all of the names issue.

I will address these things directly as I get this stuff put on the 
profiling branch.  My first concern is to get the code to be correct.  
Until I can find ALL of the variables, this is my overriding problem.


> Honza
>  
>
> > > > Honza
> >