ASE '19: extended abstract

December 20, 2019

Get rid of inline assembly through verification-oriented lifting

General topic

Automatic analysis of C & inline assembly; Precise decompilation of inline assembly


Have you ever seen this kind of code?

# 76 "ffmpeg/libavcodec/x86/mathops.h"
static inline int mid_pred(int a, int b, int c)
    int i=b;
    __asm__ (
        "cmp    %2, %1 \n\t"
        "cmovg  %1, %0 \n\t"
        "cmovg  %2, %1 \n\t"
        "cmp    %3, %1 \n\t"
        "cmovl  %3, %1 \n\t"
        "cmp    %1, %0 \n\t"
        "cmovg  %1, %0 \n\t"
        :"+&r"(i), "+&r"(a)
        :"r"(b), "r"(c)
    return i;

The fragment above is C code using the GNU inline assembly extension, which allows to embed assembly instructions into your C programs. A 2018 survey reported that more than 1/4 of trending GitHub projects actually contain it. We actually did find it in more than 10% of package sources written in C/C++ of the Debian distribution. This is usually used either to implement more efficient routines or access low-level hardware instructions.

State-of-the-art C program analyzers do not behave very well in presence of such code.

For instance, the KLEE symbolic execution engine simply stops when it hits an assembly chunk. Frama-C plugins, on the other hand, will handle it, severely over-approximating its behavior in the process. This leads to either incomplete or imprecise results – when they are not simply wrong.

That is why we proposed TInA – for Taming Inline Assembly – an automated, generic, verification-friendly and trustworthy lifting technique such that we can reuse, as is, our favorite C analyzers.


In summary, this paper makes the following contributions:

  • A new principled method lifting inline assembly to high-level C amenable to further formal analysis;
  • The automated validation of said method to make the lifter trustworthy;
  • Thorough experiments of a prototype implementation on real-world examples.

Thus, we have shown TInA to be widely applicable and trustworthy: in less than 1/2 hour, it was able to lift 2k assembly chunks coming from a typical Debian distribution (including FFMPEG, ALSA, GMP, …) and validate all of the lifted code. Moreover, these results stand whatever the compiler was (GCC or Clang) while the technique also handles different instruction sets (x86 or ARM). Last but not least, experiments have shown the produced C code actually enhances for automatic analyses (for both KLEE and Frama-C).

Further information