[LLVMdev] Issue with paper http://llvm.org/devmtg/2008-08/Geoffray_VMKitProject.pdf and presentation http://hal.inria.fr/docs/00/35/45/77/PDF/RR-6799.pdf
kirill.kononenko at gmail.com
Tue Mar 3 02:16:05 PST 2009
Just in case that this information does not get corrupted I want to
point to the issue:
In http://llvm.org/devmtg/2008-08/Geoffray_VMKitProject.pdf it is on
page 21: "VMKit, Mono, PNet", Array bounds checks.
In http://hal.inria.fr/docs/00/35/45/77/PDF/RR-6799.pdf it is on page
17: "We compare N3 with Mono  version 1.2.6, and PNet version 0.7.4
, which uses libjit as dynamic compiler...."
This statement is false. PNet version 0.7.4 has not been using LibJIT.
It has been using an unroller.
"The PNet performance show that libjit has very few optimizations and
does not compete against robust compilers."
In fact any amateur in Just-In-Time compilers, and Virtual Machine
engines, by just looking at this benchmark on the next page "Figure 7:
PNetMark results (higher is better)", can tell that PNet engine used
to take these measures has not been a Just-In-Time compiler. However,
the authors of the paper insist.
This benchmark definitely does not tell anything about LibJIT, as
LibJIT has not been used. In fact, we started development of LibJIT as
an answer to such benchmarks issues.
I can suggest the authors of these papers to use a real industrial VM
engine to compare. For instance, check Microsoft .NET or if you cannot
do this "for legal reason" try Java HotSpot. But not a 5 years old
version, although even that should give you more objective results
that you can understand what these benchmark mean in fact.
2009/3/2 Chris Lattner <clattner at apple.com>:
> On Mar 2, 2009, at 8:25 AM, Kirill Kononenko wrote:
>> Hello Dear LLVM Developers and Users
> Hi Kirill,
> Please talk to Nicholas about this, thanks!
>> I would like to complain about scientific inaccuracy of the paper
>> http://hal.inria.fr/docs/00/35/45/77/PDF/RR-6799.pdf written by
>> Nicolas Geoffray, Gaël Thomas, Charles Clément, Bertil Folliot, Gilles
>> Muller and presentation done by Nicolas Geoffray on LLVM conference
>> supported by Google:
>> They provide comparison of their system that uses LLVM, with our
>> system Portable.NET that uses LibJIT for Just-In-Time compilation.
>> All measures in this paper have been done for Portable.NET 0.7.4,
>> which does not use LibJIT as its engine. However, the author provides
>> results of Portable.NET 0.7.4. as results of LibJIT. The Portable.NET
>> 0.7.4 uses a dynamic compiler, which has nothing in common with the
>> project and development of LibJIT. We are very upset with these papers
>> and that there are such "researchers" in your group. Portable.NET
>> 0.7.4 was released 3 years ago. LibJIT have been developed during the
>> last 3 years. We require that you remove this paper and this
>> presentation and make official excuses not only in this maillist but
>> also on the next conferences when you talk about LLVM.
>> LLVM Developers mailing list
>> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu
More information about the llvm-dev