This commit intends to fill out some of the remaining pieces of the
C-unwind ABI. This has a number of other changes with it though to move
this design space forward a bit. Notably contained within here is:
* On `panic=unwind`, the `extern "C"` ABI is now considered as "may
unwind". This fixes a longstanding soundness issue where if you
`panic!()` in an `extern "C"` function defined in Rust that's actually
UB because the LLVM representation for the function has the `nounwind`
attribute, but then you unwind.
* Whether or not a function unwinds now mainly considers the ABI of the
function instead of first checking the panic strategy. This fixes a
miscompile of `extern "C-unwind"` with `panic=abort` because that ABI
can still unwind.
* The aborting stub for non-unwinding ABIs with `panic=unwind` has been
reimplemented. Previously this was done as a small tweak during MIR
generation, but this has been moved to a separate and dedicated MIR
pass. This new pass will, for appropriate functions and function
calls, insert a `cleanup` landing pad for any function call that may
unwind within a function that is itself not allowed to unwind. Note
that this subtly changes some behavior from before where previously on
an unwind which was caught-to-abort it would run active destructors in
the function, and now it simply immediately aborts the process.
* The `#[unwind]` attribute has been removed and all users in tests and
such are now using `C-unwind` and `#![feature(c_unwind)]`.
I think this is largely the last piece of the RFC to implement.
Unfortunately I believe this is still not stabilizable as-is because
activating the feature gate changes the behavior of the existing `extern
"C"` ABI in a way that has no replacement. My thinking for how to enable
this is that we add support for the `C-unwind` ABI on stable Rust first,
and then after it hits stable we change the behavior of the `C` ABI.
That way anyone straddling stable/beta/nightly can switch to `C-unwind`
safely.
Whether for Rust's own `target_os`, LLVM's triples, or GNU config's, the
OS-related have fields have been for code running *on* that OS, not code
that is *part* of the OS.
The difference is huge, as syscall interfaces are nothing like
freestanding interfaces. Kernels are (hypervisors and other more exotic
situations aside) freestanding programs that use the interfaces provided
by the hardware. It's *those* interfaces, the ones external to the
program being built and its software dependencies, that are the content
of the target.
For the Linux Kernel in particular, `target_env: "gnu"` is removed for
the same reason: that `-gnu` refers to glibc or GNU/linux, neither of
which applies to the kernel itself.
Relates to #74247
Thanks @ojeda for catching some things.
