But that’s sort of tangential: the goal is to write software that’s correct to begin with, not poke holes in it with the testsuite after the fact. With Rust, I can pick third-party dependencies that I can be (vanishingly) confident don’t have memory safety issues; when I start a greenfield C++ project, any dependencies I bring in are now potential sources of memory unsafety in my code.
In other words: I very rarely write unsafe code myself. When I bring in others’ unsafe code, I use cargo-geiger and siderophile (which I help maintain) to quantify it.
Or such bugreports would make no sense: https://github.com/denoland/deno/issues/15020
This is exactly the same as in C++, except stronger: in C++, any piece of well-formed code can violate safe invariants. In Rust, only unsafe code can violate safe invariants, and can only do so if it isn't itself well-formed.
> so long as the only unsafe code in a well-typed λRust program is confined to libraries that satisfy their verification conditions, the program is safe to execute.
I think the main caveat is that IIRC, the RustBelt unsafe rules do not cover ALL uses of unsafe in the wild, they analyze only a subset of Rust as whole, and of course, unsafe usage actually has to obey the rules. But I'm hardly an expert here.
That bug report basically says they broke the unsafe rules and thus the guarantee no longer holds (and there is UB).