llvm dev - Jan 2010 - [LLVMdev] Register Spilling and SSA

Hi

I just stumbled upon this paper. While i just skimmed over it it seems as if
the authors say that their algorithm is more efficient than the llvm 2.3 
algorithm? So i thought that might be interesting?
 
http://pp.info.uni-karlsruhe.de/uploads/publikationen/braun09cc.pdf

Disclaimer: I have no affiliation with the authors and stumbled in a slightly 
unrelated search over this paper.

ST

On Thursday 14 January 2010 15:56, ST wrote:
> Hi
>
> I just stumbled upon this paper. While i just skimmed over it it seems as
> if the authors say that their algorithm is more efficient than the llvm 2.3
> algorithm? So i thought that might be interesting?
>
> http://pp.info.uni-karlsruhe.de/uploads/publikationen/braun09cc.pdf
Don't trust it.  The abstract clearly states they're counting the number
of
dynamic spills.  That has almost nothing to do with performance.

Someone would have to reproduce their experiment to verify that performance
indeed improves.

And alas, our field is notoriously unscientific in this respect.  
Reproducability of experiments is nonexistant.

                              -Dave

Am Donnerstag, den 14.01.2010, 16:35 -0600 schrieb David
Greene:
> On Thursday 14 January 2010 15:56, ST wrote:
> > Hi
> >
> > I just stumbled upon this paper. While i just skimmed over it it seems
as
> > if the authors say that their algorithm is more efficient than the
llvm 2.3
> > algorithm? So i thought that might be interesting?
> >
> > http://pp.info.uni-karlsruhe.de/uploads/publikationen/braun09cc.pdf
As the author of this paper I have to defend it now (of course ;-)
> 
> Don't trust it.  The abstract clearly states they're counting the
number of
> dynamic spills.  That has almost nothing to do with performance.
So if not the number of dynamic spills/reloads and rematerialisations.
What else can you do to measure the quality of your spilling algorithm?
I admit that the impressive numbers in spill translate to smaller gains
in the range of 1-5% also it's hard to compare 2 different compilers.
Nonetheless we produce faster x86 code than llvm-2.3 for several spec
benchmarks (and produce way slower code for some others where I suspect
missing architecture neutral optimisations in firm).
> 
> Someone would have to reproduce their experiment to verify that performance
> indeed improves.
You can easily download libfirm-1.17.0 and cparser-0.9.9 from
http://www.libfirm.org and see for yourself.

If you're interested in the valgrind hacks to count the spills/reloads,
I uploaded them here:
pp.info.uni-karlsruhe.de/~matze/valgrind-3.5.0-countmem.tgz (there's a
new valgrind plugin called 'countmem' in it).

Greetings,
	Matthias Braun