I am not familiar with NSan (and impatient to read the paper) but a common problem with tools based on higher precision shadow values is that they can be easily

I am not familiar with NSan (and impatient to read the paper) but a common problem with tools based on higher precision shadow values is that they can be easily defeated by a large cancelation (`(1e200 + 1) - 1e200` will return 0 in both 64 and 200 bits arithmteic). My solution to this problem (which also has performance benefits) is to use an error-free transform based arithmetic (they give you roughly doubled precision but are also very resilient to cancelation). See Shaman [0] for a C++ reference implementation, you can change your types and it will display only the significant digits of your results. If you want something that does not require modifying your source code, I would recommend Verrou [1], a Valgrind-based tool using stochastic arithmetic. For those interested in the subject, my PhD [1] (which is written in English despite the French title) gives further information on all those methods, their pros and cons. [0]: https://gitlab.com/numerical_shaman/shaman https://gitlab.com/numerical_shaman/shaman [1]: https://github.com/edf-hpc/verrou https://github.com/edf-hpc/verrou [2]: https://tel.archives-ouvertes.fr/tel-03116750 https://tel.archives-ouvertes.fr/tel-03116750