On 06/22/2012 03:02 PM, Michael Matz wrote:
Hi,
On Fri, 22 Jun 2012, Tobias Grosser wrote:
There are two areas that do not yet have an isl counterpart. The
openscop import/export and the loop interchange heuristic. I think you
can safely crop the openscop import/export, as it is currently unused
and can later be added back if needed. In respect of the loop
interchange. I am personally not interested in a dedicated loop
interchange pass, but the isl scheduling optimizer does not yet provide
a valuable alternative. This is one of the other reasons I did not push
the patches through, as I was afraid of the discuss that the loop
interchange removal may cause (and I was too lazy to rewrite and fix the
buggy loop interchange heuristic).
I assume you'd be able to help with answering questions when somebody
isl/ppl-illiterate tries to fiddle with that, though?
If s.o. is interested to work on the loop interchange heuristic, I will
happily answer questions to the best of my knowledge (I did not write
the loop interchange heuristic (Sebastian did it)).
I personally hope we can remove the specific transformations (like
interchange, blocking, ...) and move to a more uniform optimizer.
However, this is touching the area of research and it may be valuable to
have the existing stuff around to compare.
One reason, I would prefer to remove them is that they enforce the
maintenance of the loop statement tree (an imperative data structure)
next to the abstract model we have. Being able to work just within our
abstract model seems a lot nicer both conceptually, but also in terms of
maintenance work needed.
Cheers
Tobi
