Wan Zhiyuan
2015-Apr-06 20:56 UTC
[LLVMdev] llvm DSA - reproduce the result in PLDI 07 paper
Dear all, I am trying to reproduce the "Percent May Alias" result described in PLDI 07's paper "Making Context-Sensitive Points-to Analysis with Heap Cloning Practical For The Real World" (http://llvm.org/pubs/2007-06-10-PLDI-DSA.html ). However, my "Percent May Alias" for all the benchmarks is much greater, especially "bzip2". The DSA code I use is from "https://github.com/smackers/smack". I have diff the code between smack and poolalloc_32. They are almost the same except the "#include" statements. I was wondering whether I need to do some configuration to make DSA work properly. Thank you! Zhiyuan -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150406/16f18f2d/attachment.html>
John Criswell
2015-Apr-06 21:22 UTC
[LLVMdev] llvm DSA - reproduce the result in PLDI 07 paper
Dear Zhiyuan, In order to reproduce the results from the paper, you'll need to replicate a system from that era. You'll need to use the same version of LLVM and DSA that the paper used. I think that was LLVM 1.9 (the release_19 branch of LLVM and poolalloc), but I'm not sure. You should check to see if the paper specifies the version. As you'll be using a very old version of LLVM, it may be worth setting up a VM with a corresponding old version of Linux. I suspect newer compilers will not compile a version of LLVM that is that old, so using an old version of Linux with an old version of GCC may be needed. I think Fedora Core 2 is the OS we were using at the time. To answer the question of why you can't use a modern version of LLVM and poolalloc, it's because LLVM has changed significantly. DSA relies upon the type annotations provided in the LLVM IR to "bootstrap" its type inference (bootstrap is not quite the right word, but it's the closest one of which I could think). As LLVM matured, transformations would ditch the type information (e.g., transforming typed GEPs into untyped GEPs into a byte array), making DSA's ability to do type-inference (and thereby improving field sensitivity) more difficult. Throw into the mix the fact that DSA is maintained by an academic research group, and the result is that modern DSA doesn't have the accuracy that the original DSA did. The good news is that I think DSA can be fixed by making its type-inferencing code smarter. The bad news is that it'd be a fair amount of work to do. So far, no one has had sufficient desire/motivation to design and implement the improvements. Regards, John Criswell On 4/6/15 4:56 PM, Wan Zhiyuan wrote:> Dear all, > I am trying to reproduce the "Percent May Alias" result described in > PLDI 07's paper "Making Context-Sensitive Points-to Analysis with Heap > Cloning Practical For The Real World" > (http://llvm.org/pubs/2007-06-10-PLDI-DSA.html). > > However, my "Percent May Alias" for all the benchmarks is much > greater, especially "bzip2". > > The DSA code I use is from "https://github.com/smackers/smack". I have > diff the code between smack and poolalloc_32. They are almost the same > except the "#include" statements. > > I was wondering whether I need to do some configuration to make DSA > work properly. > > Thank you! > > Zhiyuan > > > > > > _______________________________________________ > LLVM Developers mailing list > LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu > http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev-- John Criswell Assistant Professor Department of Computer Science, University of Rochester http://www.cs.rochester.edu/u/criswell -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150406/c69b71ab/attachment.html>
Wan Zhiyuan
2015-May-05 21:33 UTC
[LLVMdev] llvm DSA - reproduce the result in PLDI 07 paper
Dear John, I intend to implement the improvements on DSA. After running DSA on SPEC, I found DSA gives low precision for mcf and bzip2. I have checked the most imprecise c files in mcf an found that the code seems to be a mixture of "PHI" and "GEP" instructions. Could you please give me some hints about what the big picture of the improvement should be and how to start? Thank you! Regards, Zhiyuan On Mon, Apr 6, 2015 at 5:22 PM, John Criswell <jtcriswel at gmail.com> wrote:> Dear Zhiyuan, > > In order to reproduce the results from the paper, you'll need to replicate > a system from that era. You'll need to use the same version of LLVM and > DSA that the paper used. I think that was LLVM 1.9 (the release_19 branch > of LLVM and poolalloc), but I'm not sure. You should check to see if the > paper specifies the version. > > As you'll be using a very old version of LLVM, it may be worth setting up > a VM with a corresponding old version of Linux. I suspect newer compilers > will not compile a version of LLVM that is that old, so using an old > version of Linux with an old version of GCC may be needed. I think Fedora > Core 2 is the OS we were using at the time. > > To answer the question of why you can't use a modern version of LLVM and > poolalloc, it's because LLVM has changed significantly. DSA relies upon > the type annotations provided in the LLVM IR to "bootstrap" its type > inference (bootstrap is not quite the right word, but it's the closest one > of which I could think). As LLVM matured, transformations would ditch the > type information (e.g., transforming typed GEPs into untyped GEPs into a > byte array), making DSA's ability to do type-inference (and thereby > improving field sensitivity) more difficult. Throw into the mix the fact > that DSA is maintained by an academic research group, and the result is > that modern DSA doesn't have the accuracy that the original DSA did. > > The good news is that I think DSA can be fixed by making its > type-inferencing code smarter. The bad news is that it'd be a fair amount > of work to do. So far, no one has had sufficient desire/motivation to > design and implement the improvements. > > Regards, > > John Criswell > > > > On 4/6/15 4:56 PM, Wan Zhiyuan wrote: > > Dear all, > I am trying to reproduce the "Percent May Alias" result described in PLDI > 07's paper "Making Context-Sensitive Points-to Analysis with Heap Cloning > Practical For The Real World" ( > http://llvm.org/pubs/2007-06-10-PLDI-DSA.html). > > However, my "Percent May Alias" for all the benchmarks is much greater, > especially "bzip2". > > The DSA code I use is from "https://github.com/smackers/smack". I have > diff the code between smack and poolalloc_32. They are almost the same > except the "#include" statements. > > I was wondering whether I need to do some configuration to make DSA work > properly. > > Thank you! > > Zhiyuan > > > > > > _______________________________________________ > LLVM Developers mailing listLLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.eduhttp://lists.cs.uiuc.edu/mailman/listinfo/llvmdev > > > > -- > John Criswell > Assistant Professor > Department of Computer Science, University of Rochesterhttp://www.cs.rochester.edu/u/criswell > >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150505/813b03f8/attachment.html>