This repo contains all the code and test cases for the LLift project

For verification, we provide a docker image to run the program with several cases (a super set of Table 4 in the paper)

0. set OPENAI_API_KEY in the docker-compose file with your own key.

1. docker-compose build app to build image

2. docker-compose up -d start the containers

3. docker exec -it llift_ae_container /bin/bash to go to the container "app"

4. cd /app/app to go to the project directory

5. /bin/bash run_min_expr.sh to run the program

The program will run for a while (30min), and then put back all the results to the database

connect to the database with the following configuration:

"host": "127.0.0.1",
"database": "ubidb1",
"user": "ubiuser1",
"password": "ubitect",
"port": 5433

(the port is weird, but it is the port we use in the docker-compose file)

And in the table "sampling_res", you can find the result.

The expected result should be:

focusing on the "must_init", it should be either empty or "something" (i.e., a varaible)

Note: the result may be slightly different due to the randomness of GPT-4

use python run.py --id <n> --group <m> to run a single case, where n is the case number and m is the group number. The group number has the following rule:

• 20: TP cases of UBITect, use for test (bug-50)
• 3: selective cases for comparison (Cmp-40)
• 11: random-1000
• 2: a small subset of "20", for quick test

For Cmp-40. The id from -412 to -400 are 13 real bugs, and -626 to -600 are 27 false positives from UBITect.