Fix backtraces with -C panic=abort on linux; emit unwind tables by default

[Enselic]

The linux backtrace unwinder relies on unwind tables to work properly, and generating and printing a backtrace is done by for example the default panic hook.

Begin emitting unwind tables by default again with -C panic=abort (see history below) so that backtraces work.

Closes #81902 which is regression-from-stable-to-stable
Closes #94815

History

Backtraces with -C panic=abort used to work in Rust 1.22 but broke in Rust 1.23, because in 1.23 we stopped emitting unwind tables with -C panic=abort (see #45031 and #81902 (comment)).

In 1.45 a workaround in the form of -C force-unwind-tables=yes was added (see #69984).

-C panic=abort was added in Rust 1.10 and the motivation was binary size and compile time. But given how confusing that behavior has turned out to be, it is better to make binary size optimization opt-in with -C force-unwind-tables=no rather than default since the current default breaks backtraces.

Besides, if binary size is a primary concern, there are many other tricks that can be used that has a higher impact.

Release Note Entry Draft:

Compatibility Notes

• Fix backtraces with -C panic=abort on Linux by generating unwind tables by default. Build with -C force-unwind-tables=no to keep omitting unwind tables.

[rustbot]

r? @petrochenkov

rustbot has assigned @petrochenkov.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

[rustbot]

These commits modify compiler targets.
(See the Target Tier Policy.)

This PR modifies run-make tests.

cc @jieyouxu

[hanna-kruppe]

Note that it’s currently pretty tricky to change this setting from whatever the default is on a given target, since there’s no Cargo profile setting for it (unlike the panic strategy), cc rust-lang/cargo#15333

This observation can probably be used as argument for or against this change.

[bjorn3]

If you don't want unwind tables to save space you probably want to strip it out after linking anyway. -Cpanic=abort still gives unwind tables for the standard library on targets supporting panic=unwind unless -Zbuild-std is used. Would be a good idea to document this stripping in the min-sized-rust guide. You have to strip .eh_frame, .eh_frame_hdr and .gcc_except_table. One complication however is that we do have to keep unwind tables for functions that call extern "C-unwind" functions for soundness though.

[hanna-kruppe]

If there's a way to drop all unwind tables that can safely be dropped but no more, then that would be a nice new feature (-Cstrip variant?) or addition to min-sized-rust. But if you're going all out on minimizing binary size, then you'll already want -Zbuild-std for a dozen other reasons anyway. I mostly brought it up from the POV of: if the default is changed, how easily can an ordinary user of stable Rust get back to the current behavior?

[Enselic]

Changes since the initial version d113be9:

• Add a regression test specifically for backtrace with -C panic=abort.
• Adjust tests/run-make/panic-abort-eh_frame/rmake.rs. I will also ping the people that added it in Prevent .eh_frame from being emitted for -C panic=abort #112403 for review. Edit: Done.
• I added a draft release note entry to the PR description.

if the default is changed, how easily can an ordinary user of stable Rust get back to the current behavior?

I don't think ordinary users want to go back to the regression introduced in Rust 1.23, because I expect ordinary users to want backtraces to work.

For advanced users I don't think it's too much to ask to require building with -C force-unwind-tables=no to opt back in to the current (arguably broken) behavior.

[nbdd0121]

If you don't want unwind tables to save space you probably want to strip it out after linking anyway.

Stripping unwind tables is going to change the behaviour if you link in some foreign static library (e.g. a C++ library) that depends on unwinding.

One complication however is that we do have to keep unwind tables for functions that call extern "C-unwind" functions for soundness though.

I think there is no soundness concern, it's just correctness? Unwinding will stop with error if it cannot find unwind tables.

[bjorn3]

I think there is no soundness concern, it's just correctness? Unwinding will stop with error if it cannot find unwind tables.

Right, only keeping .eh_frame without the LSDA would cause soundness issues I think.

[petrochenkov]

The implementation LGTM, but is there some consensus about doing this?
I haven't heard from the "we should keep minimizing size" camp yet.

[petrochenkov]

Let's start a compiler FCP, see the PR description for what is changed.
@rfcbot merge

[rfcbot]

Team member @petrochenkov has proposed to merge this. The next step is review by the rest of the tagged team members:

• @Mark-Simulacrum
• @Noratrieb
• @SparrowLii
• @cjgillot
• @compiler-errors
• @davidtwco
• @estebank
• @jieyouxu
• @matthewjasper
• @nagisa
• @oli-obk
• @petrochenkov
• @spastorino
• @wesleywiser

No concerns currently listed.

Once a majority of reviewers approve (and at most 2 approvals are outstanding), this will enter its final comment period. If you spot a major issue that hasn't been raised at any point in this process, please speak up!

See this document for info about what commands tagged team members can give me.

[nagisa]

I'm +1 on this change if for no other reason that my personal use-cases generally tend to prefer -Cpanic=abort for perf/compile time reasons, but I also almost always care a bunch about keeping unwind tables for stack traces for profiling and just generally.

AT least my experience suggests it to be a more common use-case to want these working out of the box which is a good argument to making it work by default. Another is that people pursuing minimum binary size tend to have to apply modifications to how they build anyway, so another flag/step doesn't sound like too onerous a problem (so long as we figure out how to strip only those uwtables that don't affect soundness.)