llvm dev - Oct 2013 - [LLVMdev] loop vectorizer misses opportunity, exploit

On 31 October 2013 08:01, Frank Winter <fwinter at jlab.org> wrote:
>  With current trunk I tried the following on the above example:
>
> clang++ -emit-llvm -S loop_minimal.cc -std=c++11
> opt -O3 -vectorize-slp -S loop_minimal.ll
> opt -O3 -loop-vectorize -S loop_minimal.ll
> opt -O3 -bb-vectorize -S loop_minimal.ll
>
> All optimization passes miss the opportunity. It seems the SCEV AA pass
doesn't understand modulo arithmetic.
>
>
Hi Frank,

IIRC, opt -O3 will already pass all three, so you don't have to add it to
the command line.


How can the SCEV AA pass be extended to handle this type of
arithmetic?
>
>
I'm not an SCEV expert, so it might be possible that it does understand,
but the loop vectorizer is not understanding the evolution info, or
whatever.

What I recommend is to step through the loop vectorizer (canVectorize) and
see what the SCEV returns to the vectorizer. If the info is there, but not
accounted for, it's a vectorizer bug. If the SCEV gets lost, than you need
to add it into the SCEV logic.

cheers,
--renato
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20131031/0427103a/attachment.html>

----- Original Message -----
> 
> On 31 October 2013 08:01, Frank Winter < fwinter at jlab.org > wrote:
> 
> 
> 
> 
> 
> 
> With current trunk I tried the following on the above example:
> 
> clang++ -emit-llvm -S loop_minimal.cc -std=c++11
> opt -O3 -vectorize-slp -S loop_minimal.ll
> opt -O3 -loop-vectorize -S loop_minimal.ll
> opt -O3 -bb-vectorize -S loop_minimal.ll
> 
> All optimization passes miss the opportunity. It seems the SCEV AA
> pass doesn't understand modulo arithmetic.
> 
> 
> Hi Frank,
> 
> 
> IIRC, opt -O3 will already pass all three, so you don't have to add
> it to the command line.
-O3 will not automatically add the -bb-vectorize in clang or in opt. From Clang,
SLP and loop vectorization are enabled from -O3. From opt, I don't believe
that this is true (although we should probably fix that).
> 
> 
> 
> 
> 
> 
> 
> How can the SCEV AA pass be extended to handle this type of
> arithmetic?
To be clear, the loop vectorizer does not use the SCEV AA analysis to do
dependence checking. It uses SCEV in its own internal memory dependence checking
class.

 -Hal
> 
> 
> I'm not an SCEV expert, so it might be possible that it does
> understand, but the loop vectorizer is not understanding the
> evolution info, or whatever.
> 
> 
> What I recommend is to step through the loop vectorizer
> (canVectorize) and see what the SCEV returns to the vectorizer. If
> the info is there, but not accounted for, it's a vectorizer bug. If
> the SCEV gets lost, than you need to add it into the SCEV logic.
> 
> 
> cheers,
> --renato
-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory

On 31 October 2013 09:42, Hal Finkel <hfinkel at anl.gov> wrote:
> > How can the SCEV AA pass be extended to handle this type of
> > arithmetic?
>
> To be clear, the loop vectorizer does not use the SCEV AA analysis to do
> dependence checking. It uses SCEV in its own internal memory dependence
> checking class.
>
Hi Hal,

Thanks for the clarification! I have to say I haven't dug that deeply in
the AA recently, I assumed it was using it. Would it make sense to use it,
given that the induction could be a complex expression?

cheers,
--renato
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20131031/95be0cad/attachment.html>