[llvm-commits] Fwd: Patch for review: Speeding up ScheduleDAG computations

[Dan Gohman]

Hi Roman,

Just one comment on this patch from me; in this change:

 >      SUnit *pop() {
 >        if (empty()) return NULL;
 > -      SUnit *V = Queue.top();
 > -      Queue.pop();
 > +      SUnit *V = *Queue.rbegin();
 > +      Queue.erase(V);
 >        return V;
 >      }

The erase call here uses the "key" form of erase, so it incurs a
lookup.  This can can be avoided by using the iterator form,
something like this:

   set::iterator i = prior(Queue.end());
   SUnit *V = *i;
   Queue.erase(i);

Dan

On Mar 3, 2008, at 3:06 AM, Roman Levenstein wrote:

> Hi,
>
>
> 2008/3/1, Evan Cheng <evan.cheng at apple.com>:
>
>> Let's try to solve one problem at a time.
>>
>> 1. Roman's depth / height patch calculation is obviously goodness.
>> Please commit that first.
>
>
> Here is the patch for ScheduleDAGList.cpp: computing priorities using
> a linear algorithm. It uses the same idea as the height/depth
> computation, i.e. dynamic programming. I took into account all
> comments from Dan & Evan from last review.
>
> Additionally, I use now std::set instead of a priority queue. The
> results are exactly the same as with the current SVN version - I
> really compared it side-by-side during execution. Using std::set makes
> operations such as removal from the middle of the queue much faster
> and removes a bottleneck. Scheduling of very bigs BBs is now up to 3-4
> times faster.
>
> Please, review this new version, if it is OK for submission.
>
> And there is one more patch to come, for the ScheduleDAGRRList.cpp,
> which would also use std::set instead of priority queues, introduce
> strict ordering and contain queue update changes as David suggested.
>
> -Roman
> <ScheduleDAGList.patch>_______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits