This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Compilation performance comparison of 3.5.0 and TreeSSA trees on MICO sources as requested in: [tree-ssa] Merge status 2004-05-03


> On Tue, 2004-05-04 at 16:50, Karel Gardas wrote:
> 
> > Conclusion: compile time regressions for both -O0 and -O2 compilation.
> > 
> Thanks.  -O0 is a known loser right now.  The gimplification process
> tends to create extremely bloated RTL.  That can only be fixed by
> introducing simplistic tree optimization passes at O0.  I see this as a
> post-merge exercise, as well.

I do little of that on tree-profilng (cleanup cfg basically) and it
compiles faster than -O0 than tree-ssa tought I don't remember exact
number.  The time spent by building the CFG pays back.  I meant to
implement equivalent of delete_trivially_dead_insns for trees but didn't
get into it yet.

Honza
> 
> As far as the -O2 regressions, do you have the -ftime-report output? 
> IMO, these C++ problems are better addressed once tree-ssa is in
> mainline.  If it's similar to the other C++ applications showing compile
> time regressions, this will need work in a bunch of places inside GCC
> (memory allocator, parser, trees, etc).
> 
> 
> Diego.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]