Re: Loop optimizer issues

[Zdenek Dvorak <rakdver at atrey dot karlin dot mff dot cuni dot cz>]

Hello,

> >> tree-ssa does regular merges with mainline, so once this was merged into
> >> mainline, it would become part of tree-ssa very shortly thereafter, and
> >> the code sharing could begin :-)
> >> 
> >> That way there are no multiple-experimental code base merge issues. 
> 
> Zdenek> but there is also no way how the merge you propose could be done in this
> Zdenek> stage of development.
> 
> 	I think it would be clearer if you proposed this as part of the
> Tree-SSA loop optimization work than RTL loop optimizations on the
> Tree-SSA branch.  The Tree-SSA branch is for Tree-SSA work, it is not
> gcc-3_5-basic_improvements_branch.

I agree; if I wanted to stay with just the rtl-level optimizer I have
now, there would be absolutely no reason for the merge.  My reason is
that I want to use parts of my work for both rtl and tree level
optimizers, so they would have to be merged in anyway; it would be very
impractical to merge them piece by piece.

The second reason is that before merging the changes from tree-ssa to
mainline we will have to make the rtl and tree level loop optimizer
to cooperate well anyway, so that neither the work is not unnecessarily
duplicated not some parts missing.  It could surely be handled even if
they are developed separately, but it seems clumsy to me.

And admitedly, my third reason is laziness.  It would be difficult to
merge this rtl-level work to mainline after tree-ssa merge, due to
divergence of branches (there might be some changes also on
rtl level as the tree-ssa approaches finishing, including reorganization
and disabling of some optimizations, perhaps some cleanups, etc.)
and due to the fact that all the decisions following from the previous
paragraph.  It would be much easier for me to have my code synchronized
with the changes due to its presence in the branch rather than doing
it somehow in my private copy.

> 	I do not know enough about your new loop optimization framework to
> advise whether you should convert the loop optimizations to Tree-SSA in
> your own private tree and only propose those patches or whether it is more
> effective for you to apply the rtl loop optimizations to the Tree-SSA
> branch and work on the conversion in place, which seems to be your intent.
> Maybe you can explain why you think it is better to have all of the rtl
> loop optimization changes in place before reusing it for the Tree-SSA loop
> optimization work.

What would be the purpose of making the changes in my private tree?  I think
it is better to expose my work to other people as well.

Zdenek