This is the mail archive of the gcc-patches@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: PATCH: tune ARM's rtx_costs function


> > OK, I see where this is coming from now.  The issue I'm concerned with is 
> > about generating extra pseudos for the intermediate constants.  These are 
> > generally unlikely to provide many CSE opportunities, and hence just add 
> > to the cost of simplifying things later on.  Perhaps the 
> > preserve_subexpressions_p test can be moved into arm_gen_constant and used 
> > to separate out those two cases.
> > 
> > I'd still like to see what the cost/benefit trade-offs are though.  After 
> > all, when using -O2 we always generate those extra pseudos.
> 
> For "time make all-target-libstdc++-v3" I get the following times.
> 
> Perhaps strangely, the compiles are slightly faster when always using a
> new pseudo.  Although the difference is very small (user ~0.07%).
> 
> Reference time with sometimes a new pseudo..
> real    7m6.692s
> user    5m2.740s
> sys     1m53.640s
> 
> Time with always a new pseudo..
> real    7m4.232s
> user    5m2.530s
> sys     1m52.470s
> 
> I will re-post the patches shortly with a few coding standards
> corrections.
> 

But unless you are forcing the compilation levels, that's going to use -O2 
everywhere, which isn't interesting, since we always did use a new pseudo 
at O2.  It's compilation times at O0 and O1 that are interesting.

R.


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