Rust: Recent nightly versions break thumb* targets

[jamesmunns]

Hey all,

I've noticed in nightly-2019-07-18, rustc has begun producing symbols in the .ARM.exidx region, which it did not previously do. As our standard linker script does not have a linker rule for this region, building fails with with the following error:

  = note: rust-lld: error: no memory region specified for section '.ARM.exidx'

To reproduce this, this project can be used: https://github.com/ferrous-systems/embedded-trainings/tree/master/beginner/templates/segment-1. You will need to add the thumbv7em-none-eabihf target first. This builds with the current and previous stable releases (1.32 <= version <= 1.36), but does not build with nightly-2019-07-18. This also seems fine with rustc 1.37.0-beta.3 (2ba6de7e2 2019-07-12). I have not yet attempted to bisect this.

We've had a previous similar issue here: https://github.com/rust-embedded/cortex-m-rt/issues/157 - and the fix was to discard that section in our linker script template. However I first wanted to check if this new output was expected, because it would cause a stable to stable regression if folks don't update the cortex-m-rt crate, which includes the linker script template.

[jamesmunns]

Bisecting, this behavior showed up between:

• nightly-2019-07-17 - rustc 1.38.0-nightly (07e0c3651 2019-07-16) - OK
• nightly-2019-07-18 - rustc 1.38.0-nightly (bc2e84ca0 2019-07-17) - NOT OK

EDIT:

link to diff of commits

[jonas-schievink]

Possibly caused by the LLVM update in https://github.com/rust-lang/rust/pull/62592 – cc @nikic

It seems unexpected to me that this section is now included, since embedded projects are built with panic = abort

[nikic]

Based on https://lists.llvm.org/pipermail/llvm-dev/2019-March/130669.html it seems like the emission of .ARM.exidx is by design, even if no unwinding is used and the suggestion there is to discard in the linker.

[nikomatsakis]

Update from compiler team triage:

I can't tell quite what's going on here. Is this a compiler bug or something to be fixed downstream? @nikic's comment suggests the latter -- @jamesmunns if that is true, can you close the issue?

[nikomatsakis]

Marking as P-high so we check up on this

[jamesmunns]

So in general:

• This does appear to be an expected change in LLVM9
  
  
  
  • It's not a bug, just not something that was emitted by rustc/llvm before, so we have no handling for this section in our linker scripts
  
  
• Current versions of widely (nearly universally) used crates that are fundamental for embedded systems (e.g. cortex-m-rt, which provides a linker script generation template, among other things) do not handle this change, and are broken by versions of rustc that include LLVM9.
• We have submitted and landed a PR to cortex-m-rt to handle this (updating the linker script template we use), but have not yet released an updated crate for this
  
  
  
  • EDIT: We have tagged a release candidate as of 2019-07-25, but have not yet published on crates.io.
  
  
• Users will experience broken builds with confusing errors if they update from stable to stable, unless they also update their version of cortex-m-rt, once an update has been released.

This is an odd case, since I suppose the list of linker sections produced by rustc isn't really a "stabilized decision", but this will be mildly annoying for embedded users. I'd say our best handling responses to this would be:

• Find a way to disable this section when abort = panic, either through Rust or configuration of LLVM9
  
  
  
  • It is unclear if this is possible at all
  
  
• Patch LLVM to not emit .ARM.exidx sections when abort = panic
  
  
  
  • This has an unknown, and likely unfortunate amount of effort associated with it
  
  
• Do some kind of post-processing (invocation flags to LLD maybe?) to prevent this from happening
  
  
  
  • No idea if this is possible
  
  
• Deal with the breakage, do our best to put out information that upgrades to cortex-m-rt will be needed, hopefully making the fix searchable, and the knowledge well-known before LLVM9 lands in stable.
  
  
  
  • Annoying stable-to-stable regression for embedded users
  
  

[jonas-schievink]

Deal with the breakage, do our best to put out information that upgrades to cortex-m-rt will be needed

To help out with this, it might help to backport the fix and publish 0.5.8, since 0.5.x is still used by quite a few crates.

[skade]

Phew, I see multiple lines of arguing where this would break the language semver.

While not strictly compiler breakage, I think linking behaviour is also a big part of our compiler story.

[nagisa]

I personally do not thing this should be considered a breaking change (much like inference changes are not breaking), but passing on the ball to T-lang to have their say on the topic.

[Centril]

Me, @cramertj, and @qmx (=P) discussed this on the language team meeting and we agree with @nagisa. Attendance was spotty however and @joshtriplett may have thoughts.

[nikic]

After seeing the fix I realize that I misunderstood the original report -- this isn't about .ARM.exidx sections being emitted, but specifically about .ARM.exidx without a suffix, which is a special-case that was not handled in the cortex-m-rt linker script. (The relevant change for this on LLVM side is likely https://reviews.llvm.org/D59216.)

The ARM exception handling ABI specification says "The index table section name must be .ARM.exidx optionally followed by further characters", which permits using the name without a suffix, so LLVM isn't doing anything wrong here -- the linker script was relying on an implementation detail that was not guaranteed by the relevant specification.

[jamesmunns]

Thanks all for taking a look at this! I'll circle up with the embedded-wg folks to make sure there isn't any other detail that I've neglected to mention, but this seems like a reasonable response.

I'm guessing the LLVM9 update is riding the trains currently, so will be in Beta in 2 weeks, and stable in 8 weeks? That gives us a timeline to make sure we have this patched and released in cortex-m-rt 0.6.x (and likely backport changes to 0.5.x), and publish some "heads up" wherever possible.

[skade]

Reading @nikic response: ignore my concerns above.

[jamesmunns]

Closing this issue, I haven't heard anything more from embedded folks. Thanks again all.