Arm32 EABI exception handling

[workingjubilee]

I am opening this issue to track work regarding improving Arm 32-bit EABI exception-handling. There are a few fixmes that were never addressed and some things that I found a bit unclear. These are currently mostly questions, and it is possible this issue may be resolved fully by merely answering the questions (and hopefully committing an answer somewhere).

• How bad is reusing r12, anyways?
• What's up with NetBSD?

Arm, GCC, exceptions, and r12

We find this comment in our source:

rust/library/std/src/sys/personality/gcc.rs

Lines 126 to 136 in 6106b05

	// The DWARF unwinder assumes that _Unwind_Context holds things like the function
	// and LSDA pointers, however ARM EHABI places them into the exception object.
	// To preserve signatures of functions like _Unwind_GetLanguageSpecificData(), which
	// take only the context pointer, GCC personality routines stash a pointer to
	// exception_object in the context, using location reserved for ARM's
	// "scratch register" (r12).
	uw::_Unwind_SetGR(context, uw::UNWIND_POINTER_REG, exception_object as uw::_Unwind_Ptr);
	// ...A more principled approach would be to provide the full definition of ARM's
	// _Unwind_Context in our libunwind bindings and fetch the required data from there
	// directly, bypassing DWARF compatibility functions.

It's unclear what benefits this has, to me, and if our codegen backend could even exploit them, theoretically, if we did this. Mere correctness? Perhaps mere correctness is worth it, it might just be a fairly simple def.

NetBSD

According to this lovingly detailed commit message, NetBSD on 32-bit Arm using the EABI does not use the Arm EABI exception handling routines.

However, perhaps because it was not seen as relevant, it was not documented what it does use as its approach for unwinding. This is unfortunate, as we have recently seen Arm bugs on NetBSD that appear to be caused by assuming NetBSD on Arm must operate exactly like it does on x86, and discovering we are wrong, because of bugs or other reasons.

@he32 Do all NetBSD architectures use the same unwinding? Do we even unwind correctly on the not-so-x86 NetBSD architectures? Apparently not if this is to be believed? #96914

[he32]

## NetBSD According to this [lovingly detailed commit message](32f9b53), NetBSD on 32-bit Arm using the EABI does not use the Arm EABI exception handling routines. However, perhaps because it was not seen as relevant, it was not documented what it does use as its approach for unwinding. This is unfortunate, as we have recently seen Arm bugs on NetBSD that appear to be caused by assuming NetBSD on Arm must operate exactly like it does on x86, and discovering we are wrong, because of bugs or other reasons. @he32 Do all NetBSD architectures use the same unwinding? Do we even unwind correctly on the not-so-x86 NetBSD architectures? Apparently not if this is to be believed? #96914

You ask good questions for putting me on the spot. :) I cannot claim to comprehend the first commit above and its implications. My attempt along these lines are perhaps somewhat cargo-cult driven, and https://mail-index.netbsd.org/pkgsrc-changes/2022/05/18/msg254732.html was my best attempt at providing rust with backtrace() and associated functions, and the intention from my side was that all the NetBSD ports should link with -lexecinfo and get the support for doing this from that library. Some if not most of these changes have also been upstreamed. However, what I have never found documentation for is "this is how you interface the rust compiler to your OS / platform", it's apparent that that's supposed to be self-evident from the code, which makes following along quite the struggle. At least the original PR appears to be fixed for amd64: ``` $ uname -rp 10.0_RC3 x86_64 $ rustc --version rustc 1.73.0 (cc66ad4 2023-10-03) $ cat t.rs fn main() { panic!("explode!") } $ rustc t.rs $ ./t thread 'main' panicked at t.rs:1:13: explode! note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace $ env RUST_BACKTRACE=1 ./t thread 'main' panicked at t.rs:1:13: explode! stack backtrace: 0: std::panicking::begin_panic 1: t::main 2: core::ops::function::FnOnce::call_once note: Some details are omitted, run with `RUST_BACKTRACE=full` for a verbose backtrace. $ ``` However, my arm systems are out of reach at the moment, will regain access over the weekend. And ... I've also lately had problems getting rust to work at all on the 32-bit arm systems since about rust 1.77.1. I used to be able to self-host rust on 32-bit armv6 and armv7, but not so anymore. I tend to put some distrust in "cross compiling is a thing" as sufficient for adequate stress-testing...

[workingjubilee]

Yes! I do happen to ask incisive questions, and thank you for answering them! Even when you give an "I don't know", that is quite a lot better than having no idea whatsoever. At least then we know it is an unknown, and might cost effort to find an answer.

Fortunately I also have almost no such belief that things should be "self-evident from the code", especially when it is a matter of a complicated system interface. You will notice that I try to add documentation for later reference as I go: workingjubilee@fcf7539

[workingjubilee]

Ooh, things get even more exciting soon! #130077