Silent Compiler Bug De-duplication via Three-Dimensional Analysis

Welcome to the homepage of $D^3$! This the implementation of our research "Silent Compiler Bug De-duplication via Three-Dimensional Analysis".

Introduction

Compiler testing is an important task for assuring the quality of compilers, but investigating test failures is very time-consuming. This is because many test failures are caused by the same compiler bug (known as bug duplication problem). In this work, we propose a novel technique (called $D^3$) to solve the duplication problem on silent compiler bugs. Its key insight is to characterize the silent bugs from the testing process and identify three-dimensional information (i.e., test program, optimizations, and test execution) for bug de-duplication.

Figure 1 presents the workflow of $D^3$. Specifically, for a silent bug (denoted as b), $D^3$ considers three-dimensional information to depict the test failure, i.e., test program (denoted as p), optimizations (denoted as o), and test execution (denoted as c). Then, $D^3$ measures the distance among test failures based on the three-dimensional information for silent bug de-duplication.

Figure 1: Overview of D3

To evaluate the effectiveness of $D^3$, we conducted an empirical study based on four released datasets, which contain 2,024 test failures caused by 62 uniques bugs from four versions of GCC and LLVM. Our results show that $D^3$ significantly outperforms the two state-of-the-art compiler bug de-duplication techniques (called Tamer and Transformer in our work). Table 1 shows the overall experimental results.

Table 1: Overall experimental results

Getting Started

The failing test programs and their coverage can be downloaded from the link, the coverage data of test suites (used in the execution feature extraction) is in the passing testsuite coverage folder, the generated passing programs and the corresponding differences extracted from pairs of failing program and passing programs (used in the program feature extraction) can be downloaded from this link

Use pre-calculated distances

If you don't want to process the row data by youself, we provide the pre-calculated distances in folder distances.

Unzip the data.zip under your project directory first, then just run the following command to get the result, which will be saved in the results folder.

python test.py --dataset llvm280 --loop_time 100

Note that --dataset indicates which data set to test on. The options are 'gcc430', 'gcc440', 'gcc450', and 'llvm280'. --loop_time indicates the number of times the test is repeated, and the results are averaged. We recommend at least 100 times to avoid random factors.

Process the row data on your own

Here we take the data processing of llvm-2.8.0 as an example.

First download the generated passing programs and the differences and unzip it, then place it in the mutation-data folder.

To get the three features, namely program features, optimization features, and exexcution features, just run the following three commands.

# Program features
python get-program-features.py
# Optimization features
python get-optimization-features.py
# Execution features
python get-execution-features.py

The required configuration for each feature is in the header of the file, such as get-execution-features.py. The first few lines of the file look like this:

names = './data/llvm280/names'
wrongat = './data/llvm280/wrongat.txt'
cov_prefix = './data/llvm280/'
ep_file = './passing-testsuite-coverage/testsuite_280_ep.txt'
np_file = './passing-testsuite-coverage/testsuite_280_np.txt'
dis_url = './coverage.npy'

If you want to extend to other data sets, you only need to modify the path here (the required files are provided in our data).

Directory description

├── distances                     :  the calculated distances
├── figures                       :  figures in the README.md
├── passing-testsuite-coverage    :  preprocessed coverage information of testsuites
├── results                       :  testing results
├── Config.py                     :  Configurations and dataset descriptions
├── get-execution-features.py     :  get execution features and save the distance 
├── get-program-features.py       :  get program features and save the distance 
├── get-optimization-features.py  :  get optimization features and save the distance 
└── test.py                       :  use the calculated distances to reproduce the results