Re: Loop optimizer issues

["Dorit Naishlos" <DORIT at il dot ibm dot com>]

> However, I would like to see these bits in independent chunks, if
> possible.  One alternative that I've been thinking about is to create a
> sub-branch of tree-ssa to deal exclusively with loop optimizations and
> vectorization:
>
>      1. The tree-ssa branch would like to become 3.5 in the near
>         future.  It already contains a lot of experimental code that I
>         would very much like to stabilize.  Adding even more
>         experimental code at this stage seems a bit risky to me.
>
I understand.

>      2. Since work for loop optimizers and vectorization has just
>         started, it is unclear how much of it will be ready for 3.5.
>
Just so you know, depending on when 3.5 happens, we do hope to have some
(very) limited vectorization capabilites by then; i.e - to be able to
vectorize the simplest loops (for(i=0;i<N;i++) {a[i] = b[i]}...);
As a first step this will require very simple analysis. Gradually, as
other components become available (dependence analysis, aliasing,..) -
the range of vectorizable loops will broaden. Maybe it's too optimistic,
but we do hope to get something into 3.5.


>      3. A sub-branch of tree-ssa is no different than a mainline
>         branch.  Code that's ready for 3.5, can be merged at the same
>         time as tree-ssa.
>
> Thoughts?
>
My main concern, should we go with a sub-branch, is that it not become a
container for experimental components. We'd like this work to be merged
with mainline as soon as it's stable; we don't want to add any unnecessary
obstacles.

If it does end up in a separate sub-branch, then I'd guess we would want
certain building blocks (those vital for vectorization) to also be in the
same sub-branch. This would include IV/array/dependence/alias analysis.

thanks,
dorit