llvm dev - Feb 2020 - The AnghaBench collection of compilable programs

Dear LLVMers,

    we, at UFMG, have been building a large collection of compilable
benchmarks. Today, we have one million C files, mined from open-source
repositories, that compile into LLVM bytecodes (and from there to
object files). To ensure compilation, we perform type inference on the
C programs. Type inference lets us replace missing dependencies.

The benchmarks are available at: http://cuda.dcc.ufmg.br/angha/

We have a technical report describing the construction of this
collection: http://lac.dcc.ufmg.br/pubs/TechReports/LaC_TechReport012020.pdf

Many things can be done with so many LLVM bytecodes. A few examples
follow below:

* We can autotune compilers. We have trained YaCoS, a tool used to
find good optimization sequences. The objective function is code size.
We find the best optimization sequence for each program in the
database. To compile an unknown program, we get the program in the
database that is the closest, and apply the same optimization
sequence. Results are good: we can improve on clang -Oz by almost 10%
in MiBench, for instance.

* We can perform many types of explorations on real-world code. For
instance, we have found that 95.4% of all the interference graphs of
these programs, even in machine code (no phi-functions and lots of
pre-colored registers), are chordal.

* We can check how well different tools are doing on real-world code.
For instance, we can use these benchmarks to check how many programs
can be analyzed by Ultimate Buchi Automizer
(https://ultimate.informatik.uni-freiburg.de/downloads/BuchiAutomizer/).
This is a tool that tries to prove termination or infinite execution
for some programs.

* We can check how many programs can be compiled by different
high-level synthesis tools into FPGAs. We have tried LegUp and Vivado,
for instance.

* Our webpage contains a search box, so that you can get the closest
programs to a given input program. Currently, we measure program
distance as the Euclidian distance on Namolaru feature vectors.

We do not currently provide inputs for those programs. It's possible
to execute the so called "leaf-functions", e.g., functions that do not
call other routines. We have thousands of them. However, we do not
guarantee the absence of undefined behavior during the execution.

Regards,

Fernando

Hi Fernando,

That sounds like a very useful resource to improve testing and also get easier
access to good stress tests (e.gQuite a few very large functions have proven to
surface compile time problems in some backend passes).

From a quick look on the website I couldn’t find under which license the code is
published. That may be a problem for some users.

Have you thought about integrating the benchmarks as external tests into LLVM’s
test-suite? That would make it very easy to play around with.

Cheers,
Florian
> On 22 Feb 2020, at 14:56, Fernando Magno Quintao Pereira via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
> 
> Dear LLVMers,
> 
>    we, at UFMG, have been building a large collection of compilable
> benchmarks. Today, we have one million C files, mined from open-source
> repositories, that compile into LLVM bytecodes (and from there to
> object files). To ensure compilation, we perform type inference on the
> C programs. Type inference lets us replace missing dependencies.
> 
> The benchmarks are available at: http://cuda.dcc.ufmg.br/angha/
> 
> We have a technical report describing the construction of this
> collection:
http://lac.dcc.ufmg.br/pubs/TechReports/LaC_TechReport012020.pdf
> 
> Many things can be done with so many LLVM bytecodes. A few examples
> follow below:
> 
> * We can autotune compilers. We have trained YaCoS, a tool used to
> find good optimization sequences. The objective function is code size.
> We find the best optimization sequence for each program in the
> database. To compile an unknown program, we get the program in the
> database that is the closest, and apply the same optimization
> sequence. Results are good: we can improve on clang -Oz by almost 10%
> in MiBench, for instance.
> 
> * We can perform many types of explorations on real-world code. For
> instance, we have found that 95.4% of all the interference graphs of
> these programs, even in machine code (no phi-functions and lots of
> pre-colored registers), are chordal.
> 
> * We can check how well different tools are doing on real-world code.
> For instance, we can use these benchmarks to check how many programs
> can be analyzed by Ultimate Buchi Automizer
> (https://ultimate.informatik.uni-freiburg.de/downloads/BuchiAutomizer/).
> This is a tool that tries to prove termination or infinite execution
> for some programs.
> 
> * We can check how many programs can be compiled by different
> high-level synthesis tools into FPGAs. We have tried LegUp and Vivado,
> for instance.
> 
> * Our webpage contains a search box, so that you can get the closest
> programs to a given input program. Currently, we measure program
> distance as the Euclidian distance on Namolaru feature vectors.
> 
> We do not currently provide inputs for those programs. It's possible
> to execute the so called "leaf-functions", e.g., functions that
do not
> call other routines. We have thousands of them. However, we do not
> guarantee the absence of undefined behavior during the execution.
> 
> Regards,
> 
> Fernando
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

Hi Florian,

    we though about using UIUC, like in LLVM. Do you guys know if that
could be a problem, given that we are mining the functions from
github?
>  Have you thought about integrating the benchmarks as external tests into
LLVM’s test-suite? That would make it very easy to play around with.
We did not think about it actually. But we would be happy to do it, if
the community accepts it.

Regards,

Fernando

On Sat, Feb 22, 2020 at 5:16 PM Florian Hahn <florian_hahn at apple.com>
wrote:
>
> Hi Fernando,
>
> That sounds like a very useful resource to improve testing and also get
easier access to good stress tests (e.gQuite a few very large functions have
proven to surface compile time problems in some backend passes).
>
> From a quick look on the website I couldn’t find under which license the
code is published. That may be a problem for some users.
>
> Have you thought about integrating the benchmarks as external tests into
LLVM’s test-suite? That would make it very easy to play around with.
>
> Cheers,
> Florian
>
> > On 22 Feb 2020, at 14:56, Fernando Magno Quintao Pereira via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
> >
> > Dear LLVMers,
> >
> >    we, at UFMG, have been building a large collection of compilable
> > benchmarks. Today, we have one million C files, mined from open-source
> > repositories, that compile into LLVM bytecodes (and from there to
> > object files). To ensure compilation, we perform type inference on the
> > C programs. Type inference lets us replace missing dependencies.
> >
> > The benchmarks are available at: http://cuda.dcc.ufmg.br/angha/
> >
> > We have a technical report describing the construction of this
> > collection:
http://lac.dcc.ufmg.br/pubs/TechReports/LaC_TechReport012020.pdf
> >
> > Many things can be done with so many LLVM bytecodes. A few examples
> > follow below:
> >
> > * We can autotune compilers. We have trained YaCoS, a tool used to
> > find good optimization sequences. The objective function is code size.
> > We find the best optimization sequence for each program in the
> > database. To compile an unknown program, we get the program in the
> > database that is the closest, and apply the same optimization
> > sequence. Results are good: we can improve on clang -Oz by almost 10%
> > in MiBench, for instance.
> >
> > * We can perform many types of explorations on real-world code. For
> > instance, we have found that 95.4% of all the interference graphs of
> > these programs, even in machine code (no phi-functions and lots of
> > pre-colored registers), are chordal.
> >
> > * We can check how well different tools are doing on real-world code.
> > For instance, we can use these benchmarks to check how many programs
> > can be analyzed by Ultimate Buchi Automizer
> >
(https://ultimate.informatik.uni-freiburg.de/downloads/BuchiAutomizer/).
> > This is a tool that tries to prove termination or infinite execution
> > for some programs.
> >
> > * We can check how many programs can be compiled by different
> > high-level synthesis tools into FPGAs. We have tried LegUp and Vivado,
> > for instance.
> >
> > * Our webpage contains a search box, so that you can get the closest
> > programs to a given input program. Currently, we measure program
> > distance as the Euclidian distance on Namolaru feature vectors.
> >
> > We do not currently provide inputs for those programs. It's
possible
> > to execute the so called "leaf-functions", e.g., functions
that do not
> > call other routines. We have thousands of them. However, we do not
> > guarantee the absence of undefined behavior during the execution.
> >
> > Regards,
> >
> > Fernando
> > _______________________________________________
> > LLVM Developers mailing list
> > llvm-dev at lists.llvm.org
> > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev