Windows, FreeBSD: the pointer passed to `_swift_willThrow` cannot be weakly referenced

[grynspan]

Description
On Windows, the pointer passed to _swift_willThrow does not appear to be a valid heap object, and attempting to weakly reference it causes a crash.

Steps to reproduce
Implement a program that sets the _swift_willThrow hook to a function like so:

struct S {
  weak var error: AnyObject
}

func swiftWillThrowHandler(_ errorAddress: UnsafeMutableRawPointer) {
  let errorObject = unsafeBitCast(errorAddress, to: AnyObject.self)
  var s = S()
  s.error = errorObject // crash
}

Expected behavior
As on Darwin/Linux, this pointer should be safe to bitcast to an object and then weakly reference.

Environment

• Swift compiler version info compnerd.org Swift version 5.8-dev (LLVM ab856b0fa3177ee, Swift 71c62c0)
• Deployment target: x86_64-unknown-windows-msvc

[grynspan]

More detailed STR:

1. Create a Swift package with a C++ target that declares _swift_setWillThrowHandler() appropriately. _swift_willThrow isn't getting exported correctly on Windows right now (runtime: export _swift_willThrow #62979), so you can't use it directly.
2. Create a Swift test target in the package that imports the C++ target and calls _swift_setWillThrowHandler().
3. In the closure you pass, bitcast (or Unmanaged-cast) the first argument from UnsafeMutableRawPointer to AnyObject.
4. Assign the pointer to a weak reference somewhere. Crash should then happen.

[al45tair]

Debugging this is proving somewhat tricky; the symbols in libswiftCore seem to be incomplete, or LLDB is buggy on Windows, or maybe both. It looks like it's actually blowing up in the second swift_release() call that gets generated at the end of the catch{} block. (Though the symbols are pointing to something else entirely, but they're clearly wrong.)

[grynspan]

Perhaps this is failing on Darwin/Linux, but since their allocators may allow for more liberal use-after-free, it's less explosive? If so, adding a retain in the handling code manually might solve the problem (i.e. not a Swift bug.)

[grynspan]

Trying at desk, just manually inserting a strong retain doesn't appear to prevent the crash.

[grynspan]

This issue affects @michael-yuji's FreeBSD experiments too.