The SAILR evaluation pipeline, sailreval
, is a tool for measuring various aspects of decompilation quality.
This evaluation pipeline was originally developed for the USENIX 2024 paper "Ahoy SAILR! There is No Need to DREAM of C:
A Compiler-Aware Structuring Algorithm for Binary Decompilation". It supports 26 different C packages from Debian,
for compiling, decompiling, and measuring. Currently, angr, Hex-Rays (IDA Pro), and Ghidra are supported as decompilers.
If you are only looking to use the SAILR version of angr, then jump to the using SAILR on angr section.
This repo contains the sailreval
Python package and information about the SAILR paper artifacts.
sailreval
is the Python package that contains all the code for running the evaluation pipeline.
sailreval
evaluates the quality of decompilation by comparing it to the original source code.
This evaluation is done in four phases:
- Compilation: a project described in the targets directory is downloaded, preprocessed, and compiled into object files.
- Decompilation: decompilers supported in sailreval are used to decompile the object files into C source files.
- Measurement: the preprocessed source and decompiled source are compared using metrics in
sailreval
. - Aggregation: the results from the measurement are normalized for functions that had a metric on all decompilers.
Each phase requires the phase directly before it runs; however, you can skip stages if you manually provide the required files. For example, you can skip the decompilation phase if you already have the object files and preprocessed source.
The sailreval
package can be used in two ways: locally or in a docker container.
If you plan on reproducing the results of the SAILR paper or using some pre-packaged decompiler like Ghidra, then you
will need both. Below are two methods for installing: one is heavy (docker and local), and one is light (only local).
Make sure you have Docker installed on your system.
On Linux and MacOS:
./setup.sh
This will build the Docker container, install system dependencies, and install the Python package locally.
If you want to use only local decompilers and you have the build dependencies installed for your compiled project, you can install the Python package without the Docker container. For an example of this use case, see our CI runner.
pip3 insatll -e .
Note: you will need to install the system dependencies for the Python project yourself, listed [here](CI runner. The package is also available on PyPi, so remote installation works as well.
Verify the installation by running:
./scripts/verify_pipeline.sh
This will use both the Docker container and your local install to run the Pipeline. If you installed it correctly, you should see some final output like:
# Evaluation Data
## Stats
Layout: ('sum', 'mean', 'median')
### O2
Metric | source | angr_sailr | angr_dream
---------- | ----------- | ----------- | -----------
gotos | 1/0.12/0.0 | 1/0.12/0.0 | 0/0/0.0
...
After installation, if you used the script normally (i.e. the docker install), than you can use the docker-eval.sh
script
which is a proxy to the eval.py
script, but inside the container.
As an example you can use:
./docker-eval.sh --help
./eval.py --help
They should both produce the same result.
Using the steps below, you can run the entire pipeline stage-by-stage. In each evaluated target in targets
you will
be able to find a sailr_compiled
, sailr_decompiled
, and sailr_measured
folder in the package folder.
Each folder will contain the results of the respective stage. All targets are places in the results
directory under
their respective optimization.
For coreutils compiled with O2, you'll see results/O2/coreutils
.
To compile a package it must be described in the targets
folder by a target.toml
. Here is coreutils:
package_name = "coreutils"
source_remote = "git://git.sv.gnu.org/coreutils.git"
remote_type = "git"
download = true
post_download_cmds = ["./bootstrap"]
version = "v9.1"
package_dir = "coreutils"
pre_make_cmds = ["./configure --quiet"]
make_cmd = "make"
post_make_cmds = []
source_dir = "src"
There are many flags that you can set which are defined in the sailr_target class.
We compile just coreutils using the docker wrapper:
docker-eval.sh --compile coreutils --cores 8 --opt-levels O2
After compiling is done, you can find the results in the results/O2/coreutils
directory.
In the sailr_compiled
folder located in coreutils
you will find all the object files, preprocessed source, and normal source.
The next phase will destroy the normal source and replace it with the preprocessed source.
It's critical that you do not edit the preprocessed source in any way.
The target must contain the sailr_compiled
folder with .o
files in it. In the case of coreutils that would be:
./results/O2/coreutils/sailr_compiled/
. The source must also be present in that folder.
For the very first time you decompile a target, you must "decompile" the source, which creates normalized preprocessed source. Do it like so:
./eval.py --decompile coreutils --use-dec source --cores 20 --opt-levels O2
Highly recommend to run locally for speed. After this is done, you don't need to do it again even if you re-decompile for other decompilers.
Next, you decompile all the decompilers you want:
./docker-eval.sh --decompile coreutils --use-dec ghidra angr_sailr angr_phoenix --cores 20 --opt-levels O2
All the decompilation files, including the preprocessed source, will be found inside the sailr_decompiled
folder.
For coreutils that would be: ./results/O2/coreutils/sailr_decompiled/
.
You will find the preprocessed source as source_*.c
and the decompilation as <decompiler>_*.c
.
You will also notice files like angr_sailr_mv.linemaps
, angr_sailr_mv.toml
, and mv.dwarf.linemaps
.
These files contain the line mappings for decompiled source to original source and pre-computed metrics like goto
counts.
If you plan on using IDA Pro, you must mount it into the container.
Please mount the idat64
binary directly into the container at /tools/
.
To do that, add -v /path/to/idat64_folder/:/tools/
to the docker run
command in the docker-eval.sh script.
Like the decompilation phase, this phase requires the sailr_decompiled
to exist with .o
and .c
files in it.
If you plan on using cfged
, then the sailr_decompild
folder must contain the linemaps, toml, and dwarf files for each targeted object file.
If you ran the decompilation step above, you should automatically have that. Measure with:
./eval.py --measure coreutils --use-metric gotos cfged --use-dec source angr_sailr --cores 15
NOTE: you must put source
as one of the targeted decompilers if you are using cfged
.
After runing, you will find files in the sailr_measured
folder.
For coreutils that would be: ./results/O2/coreutils/sailr_measured/
.
In the folder you will find various toml
files that look like the following:
binary = "mv"
total_time = 231.44658088684082
timeout = false
[source.cfged]
main = "0.0"
[source.gotos]
main = 0
[angr_sailr.cfged]
main = "309.0"
[angr_sailr.gotos]
main = 11
# ...
You can use the toml
library in Python to load these files into a dictionary.
The dictionary is keyed by [decompiler][metric][function]
and the value is the metric value.
After measuring, you can aggregate the results like so:
./eval.py --summarize-targets coreutils --use-dec source angr_sailr --use-metric gotos cfged
The results will look something like, which is all sums:
# Evaluation Data
## Stats
### O2
Decompiler | gotos_sum | cfged_sum
---------- | --------- | ---------
source | 46 | 0
angr_sailr | 668 | 39701
## Metadata
total_unique_functions_in_src | total_unique_functions_in_all_metrics
----------------------------- | -------------------------------------
1152 | 918
Only the last printed table matters. Tables printed before that are intermediate results.
You can also show Sum/Average/Median
by using the --show-stats
arg.
The above summarization is the normalized results where each count is based on functions that successfully decompiled and measured on all decompilers specified in the command. You can also do multiple targets at once:
./eval.py --summarize-targets coreutils diffutils ...
There is a special case summarization for projects that are the same but may have different names.
This happens in the case of coreutils
and coreutils_gcc5
. Both are Coreutils compiled with different decompilers.
You can normalize across both projects for binaries and functions that only exist across both projects with:
./eval.py --merge-results ./results/O2/coreutils*/sailr_measured --use-dec source angr_sailr --use-metric gotos cfged
The SAILR paper introduced four artifacts:
- The angr decompiler, found in the angr repo.
- The SAILR algorithm, built on the angr decompiler as optimization passes.
- The SAILR evaluation pipeline, found in the
sailreval
Python package - The results of
sailreval
for the paper (tomls and decompilation outputs)
Below you will find instructions for using each of these artifacts.
Currently, SAILR is being slowly integrated into the angr master branch. Until then, you can use the angr-sailr fork of angr inside our provided stripped down Dockerfile found in misc/angr_sailr_dec/Dockerfile. You can also use the pre-built docker image found on Dockerhub (~3.5gb). Note, this fork will not receive updates and is the exact version used in the paper. The commit is be3855762a84983137696aa14efe2431a86a7e97.
To build the decompiler docker image, run, from the root of the repo:
docker build -t angr-sailr-dec -f misc/angr_sailr_dec/Dockerfile .
You could now run the image with docker run --rm -it angr-sailr-dec
, but we recommend using the wrapper script.
You can use the wrapper script that will run this image for you:
./scripts/docker-angr-sailr-dec.sh --help
NOTE: this mounts the current directory into the container so the decompiler can access the binary.
Verify your version works by running it on the motivating_example
binary in the root of the repo:
./scripts/docker-angr-sailr-dec.sh ./tests/binaries/motivating_example schedule_job --structuring-algo sailr
If working correctly, you should see the following output, which matches the papers example:
long long schedule_job(unsigned long a0, unsigned long long a1, unsigned long a2) {
if (a0 && a1) {
complete_job();
if (EARLY_EXIT == a2)
goto LABEL_40126b;
next_job();
}
refresh_jobs();
if (a1 || a1)
fast_unlock();
LABEL_40126b:
complete_job();
log_workers();
return job_status(a1);
}
In the SAILR paper, we run an evaluation on all 26 packages in the targets directory.
We generated data for optimization levels O0
, O1
, O2
, and O3
.
We also recorded the decompilation on all these targets with the following decompilers:
SAILR
, IDA Pro 8.0
, Ghidra 10.1
, Phoenix
, DREAM
, and rev.ng
. In each sailr_decompiled
folder,
you will find files likes so <decompiler>_<binary_name>.c
. For example, sailr_decompiled/ida_mv.c
is the decompilation
of mv
from coreutils with IDA Pro 8.0. angr based decompilation starts with angr_
and then the structuring algorithm.
All the files, which are about 11gbs in total, can be downloaded from this Dropbox link.
After downloading, you can extract the files with:
tar xf results.tar.gz --use-compress-program=pigz
The will looks like the following (but with more files):
results/
├── O0
├── O1
├── O2
│ └── coreutils
│ ├── sailr_compiled
│ ├── sailr_decompiled
│ └── sailr_measured
└── O3
To further understand what is contained in the sailr_*
folders, see the usage section above.
You can now use the sailreval
package to aggregate the results like so to get the results from the paper:
./eval.py --summarize-targets ./eval.py --measure bash libselinux shadow libedit base-passwd openssh-portable \
dpkg dash grep diffutils findutils gnutls iproute2 gzip sysvinit bzip2 libacl libexpat libbsd tar rsyslog \
cronie zlib e2fsprogs coreutils \
--use-dec source ida ghidra angr_sailr angr_phoenix angr_dream angr_comb \
--use-metric gotos cfged bools func_calls \
--opt-levels O0 O1 O2 O3 \
--show-stats
We ran the entire pipeline of SAILR on an Ubuntu 22.04 machine that had 40 logical cores and 64 GB of RAM. With these specs, it took about 8 hours to run the entire pipline for all 26 packages on the O2 optimization level. If you intend to reproduce the results as they were in the paper, checkout this repo to commit 8442959e99c9d386c2cdfaf11346bf0f56e959eb, which was the last version with minor fixes to the pipeline, but not edits to CFGED. If you plan on evaluating modernly, use the latest commit, since it will have stability, speed, and other fixes to components of SAILR.
Due to slowness in processing of source with Joern, we recommend running the Joern stage LOCALLY and not in the container. Here is an example run of only coreutils:
./docker-eval.sh --compile coreutils --cores 20 && \
./eval.py --decompile coreutils --use-dec source --cores 20 && \
./docker-eval.sh --decompile coreutils --use-dec ghidra angr_sailr angr_phoenix angr_dream angr_comb --cores 20 && \
./eval.py --measure coreutils --use-metric gotos cfged bools func_calls --use-dec source ghidra angr_sailr angr_phoenix angr_dream angr_comb --cores 20 && \
./eval.py --summarize-targets coreutils --use-dec source ghidra angr_sailr angr_phoenix angr_dream angr_comb --use-metric gotos cfged bools func_calls --show-stats
Take note of when eval.py
is run instead of the docker version.
To reproduce the results from the paper, you run the following evaluation scripts that will run the entire pipeline for you:
./scripts/evaluation/all_packages_o2_table3.sh
./scripts/evaluation/coreutils_o2_table4.sh
Run them one at a time to observe their output.
Note, you will likely not get the exact numbers shown in the paper, but the final conclusions on the numbers (i.e. the relative distance of each score) should be the same. This is due to a fundamental limitation in CFGED, which relies on GED to compute the edit distance between two CFGs. Since we never know if GED will conclude, we must use a timeout, which can be affected by the machine you run on. However, for most cases the timeout should not be triggered.
Windows targets, like libz_windows
, will not be compiled by this pipeline, so you must compile them yourself.
Follow the following the steps to compile a windows target:
- Download the source code for the target specified in the targets toml file
- Make a new configuration in MSVC
Project->Properties->Configuration Manager->Active Solution Configuration->New
- Name is
SAILR
- Go to
Project->Properties->C/C++->Preprocessor
and enable Preprocessor Definitions to File - Hit compile with SAILR config, the copy all
*.i
,*.c
, and*.obj
files into thesrc
folder you need to make - Rename the
*.obj
to*.o
- If step
5
failed, then just remove the preprocessor option after running once
To run the full pipeline for Widnows targets, you must have llvm-pdbutil installed on the system.