SymCC is an LLVM-based symbolic execution tool that focuses on performance by compiling the LLVM bitcode instead of interpreting it (as is done by tools like KLEE). The result is an average speedup (relative to KLEE) of 12 and up to 3 orders of magnitude and a consequent improvement in coverage and bug discovery.
Use of compilation is justified by the analysis in [yun:usenix:2018] that (concrete) execution is a major bottleneck in symbolic execution. (The effect is similar to how the early tools DART, CUTE and EXE worked except that they operated by instrumenting C source code.)
The tool uses concolic execution: using concrete inputs, it executes as normal but constructs a symbolic path condition as a side-effect of execution; when execution completes, the path condition can be modified and then solved by an SMT solver to generate additional concrete inputs.
Unlike non-concolic tools such as KLEE, angr and Mayhem, SYMCC does not store multiple symbolic states and use backtracking, copy-on-write, etc. to avoid re-exploration of earlier execution: the argument is that this is only necessary because re-execution is too slow.
The compilation process inserts symbolic calculations alongside the concrete calculations. To reduce overhead, compile time checks test for constants (not sure if this really means concrete values?) and generates runtime checks for constants. The instrumentation process is written as an LLVM pass that can be dynamically loaded into the LLVM compiler.
Evaluation used the DARPA CGC and compared against QSYM and KLEE and considers (concrete) execution time, concolic execution time (with KLEE configure/modified to use concolic execution and coverage of test cases. On concrete execution, SymCC is 2-3x slower than native and around 30x faster than QSYM and KLEE. On concolic execution, SymCC is about 10x faster than QSYM and KLEE. On coverage, SymCC found more paths on 46 programs and less paths on 10 programs within the available time (compared with KLEE) and, compared with QSYM, more on 47, the same on 29 and less on 40. Part, but not all, of the performance difference is due to shifting initialization time from runtime to compile time.
Also tested a combination of AFL with SymCC (and comparing with AFL+QSYM) using the fuzzer to generate seeds and the concolic engine to generate variants of the seeds. The SymCC combination was significantly better.