[Bug target/84521] [8 Regression] aarch64: Frame-pointer corruption with __builtin_setjmp/__builtin_longjmp and -fomit-frame-pointer

[wilco at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84521

--- Comment #22 from Wilco <wilco at gcc dot gnu.org> ---
(In reply to Eric Botcazou from comment #21)
> > You can use __builtin_setjmp and __builtin_longjmp in the same function,
> > only if they use a different buffer.  Otherwise it is invalid.
> 
> Yes, that's invalid.

Where exactly is this documented? I can't find no references to
__builtin_setjmp or __builtin_longjmp in
https://gcc.gnu.org/onlinedocs/gcc-7.3.0/gcc.pdf or
https://gcc.gnu.org/onlinedocs/gccint.pdf.

> > CCing Eric for the other __builtin_setjmp related issues.
> 
> I don't see anything special with Aarch64 that would make it impossible to
> support __builtin_setjmp/__builtin_longjmp properly.  You can parameterize a
> lot of things in the back-end with TARGET_BUILTIN_SETJMP_FRAME_VALUE and
> various patterns like:
>   save_stack_nonlocal
>   restore_stack_nonlocal,
>   nonlocal_goto_receiver,
>   builtin_setjmp_setup,
>   builtin_setjmp_receiver,
>   builtin_longjmp
> See the documentation and existing ports.

Which documentation? Like you say, AArch64 doesn't need anything special, it's
the mid-end implementation that is broken here. I bet most of my examples fail
on other targets too.