SAVIOR is a hybrid testing tool that prioritizes bug-finding over coverage by using a pre-testing analysis that identifies code of interest (i.e., potentially containing bugs) and prioritizes code paths leading to that code. It combines KLEE and AFL fuzzer.
Each seed is scored according using a combination of this static information and runtime information by counting how many distinct UBSan labels are reachable along variants of the path that take unexplored branches weighted by the inverse of how many attempts Savior has made to solve those branches.
This scoring method is to ensure that SAVIOR always prioritizes seeds leading to more unverified bugs, while in the long run it would not trap into those with hard-to-solve branch conditions.
To handle function pointers, they use Andersen’s points-to-analysis.
KLEE is somehow used as a concolic execution tool – I am not completely clear how this works or whether it uses as modified version of KLEE. They also add partial C++ support to KLEE.
On average, SAVIOR discovers vulnerabilities 43.4% faster than DRILLER and 44.3% faster than QSYM.
To gain further insight, they look at how quickly the different tools find UBSan labels in the code.