Loop optimizer issues

Zdenek Dvorak rakdver@atrey.karlin.mff.cuni.cz
Tue Jul 29 19:15:00 GMT 2003


Hello,

> >>I've been thinking about this proposal and I am not convinced that it
> >>would be a good idea to merge the two branches.
> >...
> >
> >You give some good reasons from the point of view of the people
> >working on the tree-ssa branch.
> >
> >Unfortunately, I think the larger picture says something different:
> >
> >    * rtlopt is not getting merged into mainline for 3.4 unless we
> >      make a special exception to the rules.
> >    * tree-ssa is slated to be merged very early in 3.5 stage 1.
> >    * At that point tree-ssa may well include the total removal of
> >      the RTL loop optimizer, which would leave rtlopt with nothing
> >      to merge into.
> >    * But tree-ssa can usefully incorporate ideas from rtlopt if
> >      they're made visible to it now.
> 
> This is an assumption.
> The RTL loop optimizer is not SSA based, and we want most, if not all, 
> of our loop opts to be SSA based.  This is a large strike against it 
> IMHO.

This gives other possibility.  I have the tree-ssa with my loop
optimizer merged in my private tree.  In about a week of work I may make
the optimizer work also on tree level.  Would you then agree with merge?
It would of course mean merging more code to the branch  at a time,
so I personally disagree with this approach, but perhaps it would give
more weight to my opinion that sharing the code is possible and good
thing to do.

> It may be nice to share code with the RTL loop optimizer, but i'd 
> rather see nice SSA based loop optimizations that are easy to maintain 
> (by themselves), fast, and good at what they do.

I may of course write all the code I have already written for rtl again
for trees from the scratch, but I am not really sure whether such a stupid
activity is fun enough to spend my time on it :-(

> >To me, that says that there are effectively three choices:
> >
> >    * Merge rtlopt into tree-ssa now and deal with the pain.
> >    * Merge rtlopt into mainline now and deal with the pain.
> >    * Throw rtlopt away now.
> >
> >Given what Zdenek said about the state of play on his branch, I am
> >inclined toward option 1.
> 
> I would be inclined toward option 2, particularly since there is no 
> evidence that that rtlopt furthers the ssa part of tree-ssa, or ever 
> will.

I of course also find this opinion appealing, although I fear it is not
feasible.  But anything that does not force me to 3) is fine with me.

Zdenek



More information about the Gcc mailing list