llvm dev - Oct 2008 - [LLVMdev] A new project proposal for LLVM and calling help from a chinese student

Mingxing,

Your project sounds interesting and if it significantly improves over  
the live variable analysis that is in LLVM right now, I think it could  
be a useful contribution.  I'm copying the 'llvmdev at cs.uiuc.edu'
mailing list, which you should join.  Send all related messages to  
this list to get feedback on your goals and also to get help with any  
problems you face.  Good luck.

--Vikram
Associate Professor, Computer Science
University of Illinois at Urbana-Champaign
http://llvm.org/~vadve



On Oct 28, 2008, at 10:10 AM, 谭明星 wrote:
> Hello, Professor Adve:
> I’m a Ph.D student in Peking University of China. In this term，I  
> have a class “Advance Compiler Techniques”, which is about  
> Program Analysis and Compiler Construction. I’m very interested in  
> LLVM for its SSA-based IR and Lifelong Program Analysis. I have read  
> most of the LLVM documents onhttp://www.llvm.org/docs, some papers  
> on http://www.llvm.org/pubs/ and some source code in LLVM2.3. The  
> live variables analysis in LLVM seems not very efficient, and the  
> algorithm presented in Benoit Boissinot, Fast liveness checking for  
> ssa-form programs (the best paper of ACM SIGPLAN CGO’08) is much  
> better, especially for JIT compilation.  So, I want to implement  
> this algorithm in LLVM, which is also planed to be my final project  
> in “Advance Compiler Techniques” class.
> To my knowledge, there has been complete dominator tree  
> implementation in LLVM2.3, so what I need to do is only calculating  
> the back-edge target set(Tsure of this project; can you give me some  
> advices?
>
> 1.     Does it valuable to implement this new liveness checking  
> algorithm for LLVM?
> 2.     If this algorithm is useful for LLVM, is it possible to  
> integrate it to the next standard version LLVM? I really hope to  
> contribute my code to LLVM if possible. Surely I will try my best to  
> get it correct and give the evaluation.
> 3.     If I have problems in this project, how could I get help from  
> you?
>
> Thanks very much. Expect for your reply :)
>
> PS: The following are links about this paper:
> http://portal.acm.org/citation.cfm?id=1356064
>
http://www.if.insa-lyon.fr/chercheurs/jpbabau/emsoc/presentations/EmSoC07_Boissinot.pdf
>
> Yours   Mingxing Tan.
> 2008-10-28
>
> **************************************************
> * Microprocessor R&D Center, Peking University
> *   Beijing, P.R.China, 100871
> *     Tel:   8610-62765828 ext. 833
> *     Fax:   8610-62756231
> *     Email: tanmingxing at mprc.pku.edu.cn
> *            tanmx.star at gmail.com
> **************************************************
>

Hello,
> On Oct 28, 2008, at 10:10 AM, 谭明星 wrote:
>>
>> PS: The following are links about this paper:
>> http://portal.acm.org/citation.cfm?id=1356064
>>
http://www.if.insa-lyon.fr/chercheurs/jpbabau/emsoc/presentations/EmSoC07_Boissinot.pdf
>>
I've put the slides from CGO online:
http://perso.ens-lyon.fr/benoit.boissinot/upload/bboissin-liveness-cgo-slides.pdf
They should be much better than the one from my presentation at EmSoC.

regards,

Benoit

For what it's worth, it would be very useful to have fast liveness  
checking available at the MachineFunction level.  We could delay  
calculating full LiveIntervals until after PHI elimination that way,  
and it would make Strong PHI Elimination very, very happy.

--Owen

On Oct 29, 2008, at 7:05 AM, Benoit Boissinot wrote:
> Hello,
>
>> On Oct 28, 2008, at 10:10 AM, 谭明星 wrote:
>>>
>>> PS: The following are links about this paper:
>>> http://portal.acm.org/citation.cfm?id=1356064
>>>
http://www.if.insa-lyon.fr/chercheurs/jpbabau/emsoc/presentations/EmSoC07_Boissinot.pdf
>>>
>
> I've put the slides from CGO online:
>
http://perso.ens-lyon.fr/benoit.boissinot/upload/bboissin-liveness-cgo-slides.pdf
> They should be much better than the one from my presentation at EmSoC.
>
> regards,
>
> Benoit
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 2624 bytes
Desc: not available
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20081029/cdbd8ea9/attachment.bin>

Hi, Benoit,
Thanks very much for your advice.
You see the algorithm greatly improve the performance of liveness analysis.
However, it seems still not efficient.
First, it is inefficient in space. You have to pre-compute all Tq for every
Tq and save them, even though only the highest nodes of Tq are needed for a
given query(q,v); Second, it is inefficient in time. Given any query(q,v),
you have to  traverse all Tq to find the highest nodes. When the Tq is
large, it maybe will cost a lot. To conquer this problem, you first order
nodes according to dominance, and then the "highest node" will be the
first
node. However, when many nodes are not dominated by each other, you have to
traverse them.
In fact, I think the highest node proposed in your new slice is very similar
to the entry of SCC if the node is a loop entry. So, maybe I could use this
information to improve this algorithm, even though I don't know clearly how
to improve it now.
Thanks again, I will try to implement it in LLVM, and further more, try my
best to improve it.

Best wishes,
Star.

-----Original Message-----
From: Benoit Boissinot [mailto:bboissin at gmail.com] 
Sent: Wednesday, October 29, 2008 10:06 PM
To: LLVM Developers Mailing List
Cc: 谭明星
Subject: Re: [LLVMdev] A new project proposal for LLVM and calling help from
a chinese student

Hello,
> On Oct 28, 2008, at 10:10 AM, 谭明星 wrote:
>>
>> PS: The following are links about this paper:
>> http://portal.acm.org/citation.cfm?id=1356064
>>
http://www.if.insa-lyon.fr/chercheurs/jpbabau/emsoc/presentations/EmSoC07_Bo
issinot.pdf
>>
I've put the slides from CGO online:
http://perso.ens-lyon.fr/benoit.boissinot/upload/bboissin-liveness-cgo-slide
s.pdf
They should be much better than the one from my presentation at EmSoC.

regards,

Benoit