llvm dev - May 2008 - [LLVMdev] [PATCH] Split LoopUnroll pass into mechanism and policy

Hi,

the attached patch splits the loop unroll pass into a LoopUnroll superclass
that implements the unrolling mechanism, and a SimpleLoopUnroll subclass
implementing the current policy. This split is modeled after the split between
Inliner and SimpleInliner.

The superclass currently still finds out the TripCount and TripMultiple, and
passes those, together with the Loop in question, to a policy method.
Currently, TripMultiple is not yet used in the SimpleLoopUnroll, but I can
imagine that this might change, so I included it already.

Currently, there is only the SimpleLoopUnroll subclass, so there is no direct
gain. However, this opens up the way for other policies and in particular I
need to implement a platform specific policy so I require this change.

Comments are welcome, I hope this change can be committed.

Gr.

Matthijs
-------------- next part --------------
A non-text attachment was scrubbed...
Name: split-unroll.diff
Type: text/x-diff
Size: 14632 bytes
Desc: not available
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20080506/a747bafc/attachment.diff>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20080506/a747bafc/attachment.sig>

Hi,

please find an updated patch attached that incorporates the (trivial) changes
introduced by r50696 and which makes this patch apply cleanly again.

Gr.

Matthijs
-------------- next part --------------
A non-text attachment was scrubbed...
Name: split-unroll.diff
Type: text/x-diff
Size: 14660 bytes
Desc: not available
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20080506/73a12eb3/attachment.diff>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20080506/73a12eb3/attachment.sig>

Hello Matthijs,

Separating mechanism from policy is a good thing for the LoopUnroll
pass. Instead of moving the policy to a subclass though, I think it'd
be better to move the mechanism, the unrollLoop function, out to be a
standalone utility function, with the LoopInfo object passed in
explicitly. FoldBlockIntoPredecessor would also be good to make into
a standalone utility function, since it isn't specific to unrolling.

This should still make it easy to write new unrolling passes with
custom heuristics, but it would also be more flexible for passes to
do unrolling in combination with other transformations.

What do you think?

Dan

On May 6, 2008, at 3:11 AM, Matthijs Kooijman wrote:
> Hi,
>
> please find an updated patch attached that incorporates the  
> (trivial) changes
> introduced by r50696 and which makes this patch apply cleanly again.
>
> Gr.
>
> Matthijs
> <split-unroll.diff>_______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev