[LLVMdev] Question about using steensgaard's pointer analysis in poolalloc

William Chan ustccmchen at gmail.com
Sat Apr 10 02:21:27 PDT 2010


Hi, LLVM dev team:
Thanks for your suggestion, I have done the experiment to compare the two
pointer analysis(Andersen and Steensgaard) methods in LLVM, but the result
was unexpected. In each test, I compare these two methods using same
optimization; There are several tests, each with a different optimization.
The benchmark is all the 11 C programs in CINT2000 of SPEC. In all the
tests, I found very little performance difference between Andersen and
Steensgaard. Here is an example of the options in one of the tests:
  llvmc -opt -Wo,=-O3 -Wo,=-load=path_of_libLLVMDataStructure.so
-Wo,=-steens-aa
  llvmc -opt -Wo,=-O3 -Wo,=-anders-aa
the other optimizations are: dead code elimination, dead store elimination,
const propogation, O1, etc.

And now what confusing me are:
1 Did the optimizations really use the result of Andersen or Steensgaard?
2 It seems both Andersen and Steensgaard have a little trouble to compile
some programs, such as: 175.vpr, 176.gcc and 197.parser for Andersen,
 253.perlbmk for Steensgaard.
3 Are the difference between O1 and O3 in LLVM similar with that in gcc?

thank you very much.

-congming

在 2010年3月23日 下午9:59,John Criswell <criswell at uiuc.edu>写道:

> 聪明陈 wrote:
> > Hi LLVM dev team:
> > I am now doing an experiment to comparing Steensgaard-style and
> > Andersen-style pointer analysis on LLVM. Since steensgaard pointer
> > analysis is in module "poolalloc", so I installed poolalloc release
> > 2.6 on my machine(intel X86_64 RedHatEnterpriseLinux 5.1, gcc-4.2.4),
> > two directories "include" and "lib" were created after installation
> > but no binary files generated.
>
> Generally, we don't use the files created by a "make install." Instead,
> we generally just compile the code and use the files directly out of the
> Release/bin (or Debug/bin) directories of the LLVM object tree.
>
> > I loaded poolalloc module into opt program according to the poolalloc
> > "README" file:
> > /opt -load <path to pool allocator> -poolalloc <other opt options>/
> > Here's my command:
> > *opt -load=/home/cmchen/INSTALL/llvm-common/lib/libpoolalloc_rt.so
> > -poolalloc -analyze -print-alias-sets test.bc*
> > and here's the error message:
> > *opt: Unknown command line argument '-poolalloc' Try: 'opt --help' *
> > So, my questions are:
> > 1 Did I install the poolalloc module in a wrong way? or I just did not
> > completely installed the module?
> > 2 How should I load the module correctly? I just use the path of
> > poolalloc shared object file to be the <path to pool allocator>, cause
> > no binary file has been generated.
> > Could you give me some suggestion? Thank you.
>
> There are two problems:
>
> 1) You need to load the library containing DSA first. To do that, you
> need to use the -load <path>/libLLVMDataStructure.so option.
>
> 2) I believe you are loading the wrong library. You want to load
> libpoolalloc.so and not libpoolalloc_rt.so. The former is the LLVM
> poolalloc transform pass; the latter is the run-time library
> implementing the poolallocation functions.
>
> If you're only interested in DSA (for points-to and alias analysis),
> then you don't need poolalloc. Just use:
>
> opt -load <path>/libLLVMDataStructure.so <dsa passes you want to run>
>
> If you want to run poolalloc, then you do the following:
>
> opt -load <path>/libLLVMDataStructure.so -load <path>/libpoolalloc.so
> -poolalloc input.bc -f -o output.bc
>
> A couple of warnings about the alias analysis passes in DSA:
>
> 1) I believe all the alias analysis implementations are based on DSA.
> DSA is a unification-based algorithm, so I think you might see
> unification even in non-unification algorithms like Andersons. Andrew,
> does this sound correct?
>
> 2) I have not used the alias analysis passes in DSA, so I don't know how
> well they work.
>
> -- John T.
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20100410/dc928d19/attachment.html>


More information about the llvm-dev mailing list