[libcxx-commits] [compiler-rt] [libcxx] [libcxxabi] [libunwind] [runtimes][PAC] Harden unwinding when possible (PR #143230)

[Daniil Kovalev via libcxx-commits]

================

----------------
kovdan01 wrote:

As I mentioned in our prior discussion, just putting old signed RA handling logic under `!defined(LIBUNWIND_PTRAUTH_CALLS_AND_RETURNS)` is fine - see my [explanation](https://github.com/llvm/llvm-project/commit/e2f8b9d3859eff96442ce04662aefb40debbef3f#r165673683) (I kindly ask you to continue discussion here since it's easy to lose comments in individual commits).

Also, I've ensured that pac-ret remains working properly after applying this patch and my proposed fix ced8b99373c9b0756f1718999996f44a74bdf46d.

Particularly:

1. Mainline llvm, **mainline** runtime libraries built **with** pac-ret, user code built with pac-ret - unwinding works fine (`RA_SIGN_STATE` detected in all stack frames)

2. Mainline llvm, **mainline** runtime libraries built **without** pac-ret, user code built with pac-ret - unwinding works fine (`RA_SIGN_STATE` detected in user stack frames and not detected in stack frames of runtime libraries)

3. Mainline llvm, **mainline + this PR + my fix** runtime libraries built **with** pac-ret, user code built with pac-ret - unwinding works fine (`RA_SIGN_STATE` detected in all stack frames)

4. Mainline llvm, **mainline + this PR + my fix** runtime libraries built **without** pac-ret, user code built with pac-ret - unwinding works fine (`RA_SIGN_STATE` detected in user stack frames and not detected in stack frames of runtime libraries)

So, the proposed change ced8b99373c9b0756f1718999996f44a74bdf46d does not break anything. Those using pac-ret instead of full arm64e/pauthtest would not even notice something changed.

P.S. In future, we want to unify libunwind logic for handling signed RA. I've created a task for this: #160110.

https://github.com/llvm/llvm-project/pull/143230