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] PR/7344, O(n^2) branch prediction


> 
> > > > > > I tried to make bootstrap, but it didn't finish due to unrelated issues.
> > > > > > Would appreciate if you could run it.
> > > > > 
> > > > > OK, it's in progress.  Results in a few hours.
> > > > > 
> > > > > Brad
> > > > > 
> > > > 
> > > > Also, it would be good if someone more familiar with branch prediction
> > > > checked the patch, and make sure I'm testing for the right condition...
> > > 
> > > !   for (e = BASIC_BLOCK (y)->succ, count = 0; e; e = e->succ_next)
> > > !     {
> > > !       if (++count >= 2)
> > > !         break;
> > > !       if (e->dest->index >= 0
> > > !           && dominated_by_p (post_dominators, e->dest, bb))
> > > !         predict_edge_def (e, pred, taken);
> > > !     }
> > > 
> > > I believe there should be "if (count++ >= 2)"; the way it is here it
> > > will break too soon.
> > 
> > No it won't, it'll break if the branch is multiway. Multiway is  > 1, not 
> > > 2.
> > 
> > In fact, the whole loop is pointless, since it will only execute once. You 
> > might as well just say
> > if (BASIC_BLOCK (y)->succ)
> > {
> > 	...
> > }	
> 
> But this is wrong -- we want it to work for ordinary conditional jumps,
> i.e. we really want to break only if >2; alternatively
> we could test for anycondjump_p.

Well, the patch as it was originally written was tested at

http://gcc.gnu.org/ml/gcc-testresults/2002-09/msg00163.html

which has many fewer g++ and g77 failures than without the patch:

http://gcc.gnu.org/ml/gcc-testresults/2002-09/msg00149.html

Perhaps if you guys can figure out what the right condition should be,
it can be tested again.

Brad


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