llvm dev - Jun 2015 - [LLVMdev] BasicAA unable to analyze recursive PHI nodes

----- Original Message -----
> From: "Tobias Edler von Koch" <tobias at codeaurora.org>
> To: "Daniel Berlin" <dberlin at dberlin.org>
> Cc: "LLVM Developers Mailing List" <llvmdev at cs.uiuc.edu>
> Sent: Thursday, June 11, 2015 10:02:37 AM
> Subject: Re: [LLVMdev] BasicAA unable to analyze recursive PHI nodes
> 
> Hi Daniel,
> 
> On Wed, 10 Jun 2015 17:11:02 -0700 Daniel Berlin
> <dberlin at dberlin.org>
> wrote:
> 
> > 
> > I didn't see numbers one way or the other. It probably has some
> > overhead, if it's the right solution, i'm sure it can be ade
> > faster.
> 
> As far as I remember, scev-aa is off because of limitations in the
> current pass manager that would mean it would have to be recalculated
> far too often.
That's correct. It also may need some caching, or other performance-related
work, to be sufficiently efficient to be on by default.
> 
> > 
> > >  I think this is a common enough case for it to be
> > > handled by BasicAA.
> > >
> > 
> > Can you provide some evidence?
> > 
> > IE hard numbers of how often it occurs on some set of benchmarks or
> > something, and what the compile time cost is?
> > 
> > It's really hard to say whether this is the right tradeoff without
> > any data.
> 
> I agree. We're seeing about 3% improvement on a number of internal
> benchmarks, but I'll give this a try on SPEC etc. and share the
> results
> here (including compile time).
Sounds good; please do.

 -Hal
> 
> Thanks for your comments!
> 
> Tobias
> 
> --
> Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a
> Linux Foundation Collaborative Project.
> _______________________________________________
> 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

Hi Daniel, Hal,

I've evaluated the patch on the SPEC CPU2006 C/C++ benchmarks. Here are
the results.

1. How many aliasPHI queries on recursive PHI nodes are there and what
are their results?

NoAlias       81317/10.6% (before) 118549/18.1% (after)
MayAlias     685637/89.4% (before) 534856/81.4% (after)
PartialAlias      0/ 0.0% (before)   3325/ 0.5% (after)
Total        766954       (before) 656730       (after)

As you can see, the share of NoAlias results for these queries goes up
noticeably. What's also interesting is that there is a 14.4% reduction
in total queries on recursive PHI nodes with the patch. 

2. What is the average number of incoming values on recursive PHI nodes?

There were 2 incoming values (one recursive, one 'real' value) in all
but 102 (0.013%) of aliasPHI queries on recursive PHI nodes before the
patch. With the patch, there were 103 queries with more than 2 incoming
values. So two incoming values is indeed the common case.

3. What is the compile time impact of the change?

I ran the SPEC2006 build 5 times, discarding the numbers for the first
build to warm up the fs cache. Without the patch, the average build
time over the remaining four builds is 5023.913s. With the patch, it is
5023.713s. Pretty much unchanged.

4. What is the impact of the change on the execution time of the
benchmarks?

This is with the ref data set, average of three runs.

(Speedup, higher is better)
400.perlbench   0.991404586
401.bzip2       0.999081718
403.gcc         1.006213196
429.mcf         0.990988836
445.gobmk       0.994521527
456.hmmer       0.996110677
458.sjeng       0.993703064
462.libquantum  1.014540594
464.h264ref     0.996489267
471.omnetpp     0.986799815
473.astar       0.987321483
483.xalancbmk   0.985122405
433.milc        1.023695308
444.namd        0.998587195
447.dealII      0.999585507
450.soplex      1.001369759
453.povray      0.978606911
470.lbm         1.018186381
482.sphinx3     0.998224144
Average         0.997923809

Not a huge change here although the numbers tend to be more on the side
of a slowdown. This is interesting - you'd think that better alias
analysis shouldn't have that effect. Any idea what could be causing
this? The numbers are from a real machine though so I'd expect
something like 1-2% noise, so it could be just that.

I hope this answers some of your questions!

Tobias

On Thu, 11 Jun 2015 10:20:02 -0500 Hal Finkel <hfinkel at anl.gov>
wrote:
> ----- Original Message -----
> > From: "Tobias Edler von Koch" <tobias at
codeaurora.org>
> > To: "Daniel Berlin" <dberlin at dberlin.org>
> > Cc: "LLVM Developers Mailing List" <llvmdev at
cs.uiuc.edu>
> > Sent: Thursday, June 11, 2015 10:02:37 AM
> > Subject: Re: [LLVMdev] BasicAA unable to analyze recursive PHI nodes
> > 
> > Hi Daniel,
> > 
> > On Wed, 10 Jun 2015 17:11:02 -0700 Daniel Berlin
> > <dberlin at dberlin.org>
> > wrote:
> > 
> > > 
> > > I didn't see numbers one way or the other. It probably has
some
> > > overhead, if it's the right solution, i'm sure it can be
ade
> > > faster.
> > 
> > As far as I remember, scev-aa is off because of limitations in the
> > current pass manager that would mean it would have to be recalculated
> > far too often.
> 
> That's correct. It also may need some caching, or other
performance-related work, to be sufficiently efficient to be on by default.
> 
> > 
> > > 
> > > >  I think this is a common enough case for it to be
> > > > handled by BasicAA.
> > > >
> > > 
> > > Can you provide some evidence?
> > > 
> > > IE hard numbers of how often it occurs on some set of benchmarks
or
> > > something, and what the compile time cost is?
> > > 
> > > It's really hard to say whether this is the right tradeoff
without
> > > any data.
> > 
> > I agree. We're seeing about 3% improvement on a number of internal
> > benchmarks, but I'll give this a try on SPEC etc. and share the
> > results
> > here (including compile time).
> 
> Sounds good; please do.
> 
>  -Hal
> 
> > 
> > Thanks for your comments!
> > 
> > Tobias
> > 
> > --
> > Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a
> > Linux Foundation Collaborative Project.
> > _______________________________________________
> > LLVM Developers mailing list
> > LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> > http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
> > 
> 


-- 
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a
Linux Foundation Collaborative Project.

> Not a huge change here although the numbers tend to be more on the side
> of a slowdown. This is interesting - you'd think that better alias
> analysis shouldn't have that effect. Any idea what could be causing
> this?

LLVM's alias analysis is good but not amazing.
If you improve it significantly (as you appear to have :P), you give
the compiler more freedom to do things.
What it chooses to do with that freedom  may not always be good :)

Basically, with any significant alias analysis change comes a lot of
tuning of heuristics/etc, or looking through regressions, to determine
what's going on now.  *Usually* it's "something used to not be able
to
do a ton of <whatever>, and now it can.  The cost model for
<whatever>
needs to be worked on so it doesn't go too crazy".