SLP/Loop vectorizer pass ordering

[Zinovy Nis]

Moving Loop Rotation pass from its current location to just before
Loop Vectorizer helped in my case!
Other benchmarks from our internal suit did not degrade also.


2014-10-10 12:38 GMT+04:00 Renato Golin <renato.golin at linaro.org>:
> On 26 July 2014 03:05, Chandler Carruth <chandlerc at google.com> wrote:
>> It would be nice for the inliner to
>> have visibility into how the vectorizer will be able to pack the code of a
>> function into vectors so it can more accurately estimate the size cost it
>> trades off against.
>
> A bit late to the party, and going on a tangent, but could this be
> done via exposing the vectorization validation/cost models to external
> passes?
>
> When we were discussing about the interaction between Polly and the
> loop vectorizer, the idea was that either would use the other's
> information on legality and costs at different vector lengths
> (somewhat similar to safelen metadata). Another idea was to expose the
> canVectorize() method to other passes, so that you could infer some
> knowledge about the end result of that being inlined or not.
>
> I realise that this is a gamble, and inlined code can be more (or
> less) easily vectorized, but at least it gives you an idea on how to
> gamble. For instance, if the current vectorized cost is greater, but
> very close to the scalar version, and the code reduction is 4x, it may
> be profitable to inline. I'm not a big fan of gambling with
> optimizations, but inlining is always a gambling, so, maybe that fits
> into the model. Happy to be wrong, though.
>
> cheers,
> --renato
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits