How is this different from tools like Valgrind and Address Sanitizer?

Fil-C seems to be more similar to ASAN than Valgrind in that the compiler outputs code with the instrumentation / checking present, rather than running already compiled code in a virtual machine as Valgrind does.

But it would seem to be more robust than ASAN; from reading about how ASAN works, it seems that it puts "poisoned" bytes around an allocation, so that memory accesses shortly after the end of a buffer hit those poisoned bytes and are caught. However, ASAN wouldn't catch an invalid access to a non-poisoned address of memory via a particular a pointer, if that address was allocated in a separate allocation. [1]

I assume Fil-C's pointer capability model is able to catch "provenance" violations like that.

[1] https://blog.gistre.epita.fr/posts/benjamin.peter-2022-10...

From the readme:

> Fil-C is engineered to prevent memory safety bugs from being used for exploitation rather than just simply flagging them often enough to find bugs. This makes Fil-C different from AddressSanitizer, HWAsan, or MTE, which can all be bypassed by attackers. The key difference that makes this possible is that Fil-C is capability based (so each pointer knows what range of memory it may access, and how it may access it) rather than tag based (where pointer accesses are allowed if they hit valid memory).

Clang says "AddressSanitizer's runtime was not developed with security-sensitive constraints in mind and may compromise the security of the resulting executable", so it should not be used in production.

Valgrind has much worse performance (the manual claims 10-50x slowdown, plus it's effectively single-threaded), which is probably bad enough to make it unusable in production, and similarly will miss many memory safety bugs.