RFC: Modeling horizontal vector reductions

Evan Cheng evan.cheng at apple.com
Thu Sep 12 08:49:31 PDT 2013



Sent from my iPad

On Sep 11, 2013, at 11:16 PM, Owen Anderson <resistor at mac.com> wrote:

> 
> On Sep 11, 2013, at 4:24 PM, Arnold Schwaighofer <aschwaighofer at apple.com> wrote:
> 
>> CC’ing Owen who I have been told might have started this at one point. Owen is my thinking that threading fast-math flags into ISel is a bigger undertaking a correct estimate or is it really just “creating and maintaining a BinOp->FastMathFlag map”? I am guessing the maintaining part might be hairy.
> 
> I would love to propagate fast-math flags through ISel and kill off the FastMathFlag, but it'll be a non-trivial amount of work to update every DAGCombine to respect them properly.

Fast math is not the only bit of information that's lost. There is also nsw and nuw which would have been useful for isel as well. That said I agree it's not worthwhile to add more complexity to sdisel. We should add them after we migrate to MI based isel. 

Evan

> 
> --Owen
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20130912/ae99004c/attachment.html>


More information about the llvm-commits mailing list