I feel like neither of these fixes the underlying issue that we likely shouldn't be using a static mut here... It just side-steps the lint by using raw pointers which for some reason we don't lint on.

You are right. But I didn't see other ways to implement this. And we can guarantee that using raw pointers here is correct, right?

by using raw pointers which for some reason we don't lint on.

We don't lint against raw pointers because that's often what we're trying to push people toward (among other options). We lint against references because it's easier to get that wrong without realizing. E.g., it's fine to have aliasing pointers to some data as long as you're careful to not race the accesses, but it's never fine to have aliasing live mutable references. But taking potentially long-lived references to mutable statics makes it really easy to do that accidentally.

To make this particular case safer, I'd probably move STACK_TRACE into that let stack = unsafe { .. } block. Then it can't be directly accessed later while the reference to it via that slice is live, which is the main danger.

Probably I'd also expect this function to be unsafe extern "C" fn .. since there's a safety requirement here that must be upheld by callers (the function is not reentrant). Not sure if there's some practical reason preventing that here.

Assuming that safety invariant is upheld, the use of static mut here, as revised, seems fine to me in the context of what we were trying to achieve as a lang matter. This is plausibly the kind of use case that argues against deprecating static mut entirely (see #53639).

Did you check whether the library builds on stage 2? This should definitely still trigger the lint here since we haven't (afaict?) changed anything.

I found no reference used in this function's body, so I removed it and just try to build on stage 2 on my linux device.
But I just found it needs #[cfg(all(target_env = "msvc", not(target_arch = "arm")))], so I will test it on another windows device later today.

I tried ./x b library/panic_unwind --stage 3 and it succeed:

Building bootstrap
    Finished `dev` profile [unoptimized] target(s) in 0.17s
Building stage0 library artifacts (x86_64-pc-windows-msvc)
    Finished `release` profile [optimized + debuginfo] target(s) in 0.23s
Building compiler artifacts (stage0 -> stage1, x86_64-pc-windows-msvc)
    Finished `release` profile [optimized + debuginfo] target(s) in 1.09s
Creating a sysroot for stage1 compiler (use `rustup toolchain link 'name' build/host/stage1`)
Building stage1 library artifacts (x86_64-pc-windows-msvc)
    Finished `release` profile [optimized + debuginfo] target(s) in 0.36s
Building compiler artifacts (stage1 -> stage2, x86_64-pc-windows-msvc)
    Finished `release` profile [optimized + debuginfo] target(s) in 0.83s
Creating a sysroot for stage2 compiler (use `rustup toolchain link 'name' build/host/stage2`)
Uplifting library (stage1 -> stage2)
Uplifting rustc (stage1 -> stage3)
Creating a sysroot for stage3 compiler (use `rustup toolchain link 'name' build/host/stage3`)
Uplifting library (stage1 -> stage3)
Build completed successfully in 0:00:10