[LLVMdev] GSoC Proposal: Inter-Procedure Program Slicing in LLVM

Mingliang LIU liuml07 at mails.tsinghua.edu.cn
Wed May 1 21:36:31 PDT 2013

Hi all,

I had a second thought of the dynamic slicing, as well as the source code

Firstly, the dynamic slicing is very useful to software community (I'll
illustrate more in the refined proposal later), but it's already
implemented by Swarup and John Criswell from UIUC. The static slicing code
has been released as Giri project in LLVM, and they would kindly release
the dynamic slicing too if this proposal is accepted.

I'd like to study and enhance the Giri code to make it better. However, I
don't know the exact part to which I can contribute. I'll ask Swarup (I cc
this email to him) for help. I would be honored to be directed by him if
this proposal is selected.

Moreover, I don't know how the static and dynamic slicing fit together.
Does the static code of Giri need some improvements? For example, taking
into consideration the pointer aliases. Our group need the static slicing
to generate I/O benchmarks (see below please).

Secondly, to generate source code, which is human-readable and compilable
seems a bit ambitious. My previous scripts in python can generate the
source code of the original program by deleting useless line of code. It's
kind of tricky. We used dozens of regular expressions to match the boundary
of blocks, loops and functions to avoid deleting lines unexpectedly. I
thought employing clang was a better idea. I don't know whether you think
the source code genration is a good idea or not. I can release our simple
script and improve it later. So I have more time/energy to contribute to
the dynamic slicer.

I think the slicing pass can be loaded by the opt. There is no need to
change the front-end or other passes. I'm not sure whether the link
time optimization can be exploited. We borrowed the LLVMSlicer code
previously without LTO.

Lastly, I'd like to introduce myself briefly. I'm a three-years PhD
candidate student from Tsinghua University, China. My research area covers
performance analysis, compiler techniques for high performance
computing, and parallel computing (MPI/OpenMP).

One of my on-going work is to generate an I/O benchmark from the original
application. The base observation is that the computation and communication
statements can be deleted if they're irrelevant to the I/O pattern, e.g.
computing the buffer content to be written into a file. We take use of the
program slicing technique to find relevant/irrelevant statements. The
static slicer was borrowed from LLVMSlicer and I wrote code to make it work
for our application. We generated the line number of sliced code. There is
a very simple source code generation script using ugly and tricky regular
expressions to delete original source code according to the sliced line
number. We're going to submit the first version of the paper recently.

I also took part in one project in Open64 compiler several year ago, the
purpose of which is to fast collect the communication trace. We used
program slicing to delete the computation statements and kept the
communication related statements. We generated executable binaries instead
of source code from IR.

Now we plan to do a project that can automatically find manual
configuration errors in software deployment. The dynamic slicing can help a
lot since the input is the key factor to locate errors. We have not a
concrete plan for this project, but the dynamic slicing is heavily needed.


On Thu, May 2, 2013 at 8:00 AM, Sahoo, Swarup Kumar <ssahoo2 at illinois.edu>wrote:

>  Hi Mingliang,
>     We already implemented an usable version of Dynamic Backward Slicing
> as part of our ASPLOS paper. So, I think this will be a good idea to extend
> this project. I will also be interested in mentoring you for this project.
> Please let me know, if you need any help.
> Thanks,
> Swarup.
>  ------------------------------
> *From:* llvmdev-bounces at cs.uiuc.edu [llvmdev-bounces at cs.uiuc.edu] on
> behalf of Mingliang LIU [liuml07 at mails.tsinghua.edu.cn]
> *Sent:* Saturday, April 27, 2013 2:53 PM
> *To:* LLVM
> *Subject:* [LLVMdev] GSoC Proposal: Inter-Procedure Program Slicing in
>  Hi all,
>  This is a GSoC 2013 proposal for LLVM project. Please see the formatted
> version at here:
> http://pacman.cs.tsinghua.edu.cn/~liuml07/files/gsoc2013-proposal-program-slicing.pdf
>  Program slicing has been used in many applications, the criteria
> of which is a pair of statement and variables. I would like to write an
> inter-procedural program slicing pass in LLVM, which is able to calculate C
> program slices of source code effectively. There is no previous work
> implemented in LLVM, which considers both the dynamic program slicing and
> source code of the sliced program. Program slice contains all statements in
> a program that directly or indirectly act the value of a
> variable occurrence [5]. Program slicing has been used in many
> applications, e.g., program verification, testing, maintenance, automatic
> parallelization of program execution, automatic integration of program
> versions.
>  While it's straightforward to implement the slicer in the back-end of
> compiler using SSA form, the source code of the original program instead of
> intermediate
> representation is preferred in most cases. Moreover, we can further narrow
> the notion of slice, which contains statements that influence the value of a
> variable occurrence for speci c program inputs. This is referred as
> dynamic program slicing [1]. However, there is no previous work implemented
> in LLVM
> which solved the two problems.
>  There are two public projects which implement the backwards static
> slicing in LLVM.
>    - Giri Written by John Criswell from UIUC, a subproject of LLVM. The
>    Giri code contains the static backwards intra-procedure slicing passes, and
>    runs with an older version of LLVM. It also only backtracks until it hits a
>    load. Additional code must be written to backtrack further to
>    find potentially reaching stores.
>    - LLVMSlicer This implementation is a complete static backwards slicer
>    from Masaryk University. It works on the well de ned data and control flow
>    equations in a white paper by F. Tip [4]. However, this code was written
>    for special purpose, thus it's not general enough to be use by others. They
>    implemented the Andersen's alias algorithm [2], callgraph, and modi es
>    analysis to support the slicer, instead of using the LLVM APIs.
>  They eliminate the useless IR statements and keep the ones a ect the
> values of the criteria. However, neither of them generates the compilable
> source code
> slice, which is heavily needed in reality. There are several ways to do
> this. One is to generate the source code from sliced IR using llc tool. The
> issue is that
> the IR is not concerned with high-level semantic. The generated source
> code is different from the original program and not suitable for human
> reading. An-
> other approach is deleting the source code according to sliced IR, with
> line number information (in meta-data of each instruction). The naive
> script deleting
> sliced source code one by one fails to handle tricky cases. I think a
> better source code slicer is to take use of the AST info.
>  There are three main parts of this work.
>    1. First, borrow an implementation of static back-wards slicing from
>    Giri or LLVMSlicer, and use the LLVM callgraph, mod/ref and alias
>    interfaces as much as possible.
>    2. Second, implement the dynamic program slicing using the approach 3
>    in the paper [1].
>    3. Third, generate the source code of the sliced program. To make the
>    sliced source code compile directly, we need to employ clang front-end
>  The final result of this summer of code is to make this pass work e
> effectively and documented well. Further, I'll write test cases and behave
> as the active
> maintainer for this project. My long-term plan is to add more features,
> e.g. Objective-C/C++ support, thing slicing [3], to this project.
>  Any comment is highly appreciated.
>  [1] H AGRAWAL. Dynamic Program Slicing. In ACM SIGPLAN Conference on
> Programming Language Design and Implementation (PLDI),1990.
>  2] L.O. Andersen. Program analysis and specialization for the C
> programming language. PhD thesis, University of Cophenhagen, Germany, 1994.
> [3] M. Sridharan, S.J. Fink, and R. Bodik. Thin slicing. In PLDI'07, 2007.
> [4] F. Tip. A survey of program slicing techniques.Journal of Programming
> Languages, 1995.
> [5] M. Weiser. Program slicing. In Proceedings of the 5th International
> Conference on Software Engineering, ICSE, pages 439{449. IEEE, 1981.
>  --
> Mingliang LIU (刘明亮 in Chinese)
> PACMAN Group,  Dept. of Computer Science & Technology
> Tsinghua University, Beijing 100084, China
> Email: liuml07 at mails.tsinghua.edu.cn
> Homepage: http://pacman.cs.tsinghua.edu.cn/~liuml07/

Mingliang Liu (刘明亮 in Chinese)

PACMAN Group,  Dept. of Computer Science & Technology
Tsinghua University, Beijing 100084, China
Email: liuml07 at mails.tsinghua.edu.cn
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20130502/b78d927c/attachment.html>

More information about the llvm-dev mailing list