[LLVMdev] some superoptimizer results

[Sean Silva]

On Wed, Jul 22, 2015 at 12:54 PM, Hal Finkel <hfinkel at anl.gov> wrote:

> One thing that is important to consider is where in the pipeline these
> kinds of optimizations fit. We normally try to put the IR into a canonical
> simplified form in the mid-level optimizer. This form is supposed to be
> whatever is most useful for exposing other optimizations, and for lowering,
> but only in a generic sense. We do have some optimizations near the end of
> our pipeline (vectorization, partial unrolling, etc.) that consider
> target-specific properties, but only because the alternative is doing those
> loop optimizations after instruction selection.
>
> Considering ILP and other pipeline-level costs are something we generally
> consider only in the SelectionDAG and after. If these are IR optimizations,
> then I'm not sure that considering ILP, etc. is the right metric -- so long
> as the transformations are sufficiently reversible to allow of efficient
> lowering afterward.
>

Agreed. It might just be that these initial results are from the "burn-in"
specifically targeting short simple sequences, but most of the
transformations in the link seem to be things that, if applicable, we would
want to do in the backend instead of in the middle-end.

-- Sean Silva


>
>  -Hal
>
> ----- Original Message -----
> > From: "Sean Silva" <chisophugis at gmail.com>
> > To: "John Regehr" <regehr at cs.utah.edu>
> > Cc: llvmdev at cs.uiuc.edu
> > Sent: Wednesday, July 22, 2015 2:35:51 PM
> > Subject: Re: [LLVMdev] some superoptimizer results
> >
> >
> >
> > Are you taking into account critical path length? Because e.g. for:
> >
> >
> >
> > %0:i64 = var
> > %1:i1 = slt 18446744073709551615:i64, %0
> > %2:i64 = subnsw 0:i64, %0
> > %3:i64 = select %1, %0, %2
> > infer %3
> > %4:i64 = ashr %0, 63:i64
> > %5:i64 = add %0, %4
> > %6:i64 = xor %5, %4
> > result %6
> >
> >
> > In the former case, the cmp and sub are independent, so can be
> > executed in parallel, while in the latter case all 3 instructions
> > are dependent. So the former case can execute in 2 cycles while the
> > latter takes 3. Modern OoO chips do in fact exploit this kind of
> > thing.
> >
> >
> > -- Sean Silva
> >
> >
> >
> >
> > On Wed, Jul 22, 2015 at 10:15 AM, John Regehr < regehr at cs.utah.edu >
> > wrote:
> >
> >
> > We (the folks working on Souper) would appreciate any feedback on
> > these IR-level superoptimizer results:
> >
> > http://blog.regehr.org/extra_files/souper-jul-15.html
> >
> > My impression is that while there's clearly plenty of material in
> > here that doesn't want to get implemented in an opt pass, there are
> > a number of gems hiding in there that are worth implementing.
> >
> > Blog post containing additional explanation and caveats is here:
> >
> > http://blog.regehr.org/archives/1252
> >
> > Thanks!
> >
> > John
> >
> > _______________________________________________
> > LLVM Developers mailing list
> > LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu
> > http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
> >
> >
> > _______________________________________________
> > LLVM Developers mailing list
> > LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> > http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
> >
>
> --
> Hal Finkel
> Assistant Computational Scientist
> Leadership Computing Facility
> Argonne National Laboratory
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150722/d540ee4d/attachment.html>