Loop optimizer issues

[Zdenek Dvorak]

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.
> 
>       * The tree-ssa branch is primarily trying to replace RTL passes. 
>         This may include the loop optimizer.  So far, it is not clear
>         whether we will actually need to have a sophisticated loop
>         optimizer at the RTL level (if at all).  The only thing that
>         comes to mind that may needed from RTL are:
>               * instruction lengths/costs to guide the unroller
>               * prefetching capabilities and memory costs to guide
>                 software pipelining and memory hierarchy optimizations.

Exactly because this question is not clear, there is a need to
experiment.  This can hardly be done if the changes are kept separated.

>       * While there still isn't much loop-related work on the branch,
>         the loop discovery infrastructure that we inherited from
>         mainline seems adequate for the basic things.

There is a lot of more things I hope to reuse, especially the analysis
stuff I have written.  Yes, it is possible to take just the pieces I
need and I will do it if the other reasons won't persuade you.

>       * The branch is slowly becoming increasingly demanding wrt new
>         contributions and optimization work.  Nightly tests on many
>         platforms are very helpful.  I don't really want to add anything
>         that's not related to the tree optimization infrastructure and
>         may or may not be needed.

I propose the changes to be disabled by default (although I of course
would be grateful to people who would enable them for testing).

>       * The tree-ssa branch is already very difficult to merge with
>         mainline.  I don't want to make it more difficult.  I would
>         prefer to see the rtlopt branch merged into mainline separately.

Which will make the merge even more complicated, IMO.

Zdenek