[PATCH] Mark swapcontext returns_twice without indirect_return [BZ #23130]

[Adhemerval Zanella Netto]

On 02/09/26 09:43, Stian Halseth wrote:
> swapcontext does not restore every callee-saved register on all targets.
> On sparc the %l and %i registers are not in the saved context at all:
> they live in the register window save area of the calling frame, which
> another pass through that frame can overwrite.  stdlib/tst-setcontext2
> fails on sparc64 for this reason at -O2 with PIC, where gcc keeps
> &global in %i5 across swapcontext and reuses %i5 as the GOT base on
> another path through the same frame.
> 
> Make the generic bits/indirect-return.h fall back to returns_twice where
> indirect_return is unavailable, covering every affected target.  x86 and
> aarch64 keep their own headers, which gate on CET and on GCS, so those
> targets are unchanged.
> 
> string/tst-xbzero-opt.c calls swapcontext from an always_inline
> function, which a returns_twice callee forbids; key its guard on the
> availability of indirect_return rather than on __CET__.  It is the only
> such site in the tree.
> 
> Tested on sparc64: full make check is clean, with no test that passed
> before failing after.
> 
> Signed-off-by: Stian Halseth <stian@itx.no>
> ---
>  bits/indirect-return.h  | 15 +++++++++++++--
>  string/tst-xbzero-opt.c |  7 +++----
>  2 files changed, 16 insertions(+), 6 deletions(-)
> 
> diff --git a/bits/indirect-return.h b/bits/indirect-return.h
> index 819bd25b..4bfe3d35 100644
> --- a/bits/indirect-return.h
> +++ b/bits/indirect-return.h
> @@ -21,5 +21,16 @@
>  #endif
>  
>  /* __INDIRECT_RETURN is used on swapcontext to indicate if it requires
> -   special compiler treatment.  */
> -#define __INDIRECT_RETURN
> +   special compiler treatment.  swapcontext does not restore every
> +   callee-saved register on all targets -- on sparc the %l and %i
> +   registers are not in the saved context at all -- so where
> +   indirect_return is unavailable, returns_twice stops the compiler
> +   assuming such a value survives the call.  Targets that need this only
> +   under an optional feature override this file.  */

I don't think it make much sense to add sparc-specific rationale on an
installed generic header. I think something like would suffice.

  /* swapcontext does not restore every callee-saved register on all targets.  */

> +#if __glibc_has_attribute (__indirect_return__)
> +# define __INDIRECT_RETURN __attribute__ ((__indirect_return__))

I think we should not use __indirect_return__ as the preferable attribute,
It was added as a way to mark a function may return via an indirect branch,
and as a way to optimize the swapcontext on x86_64/aarch64.  

The optimization rationale is moot because the call will issue a syscall
anyway (sigprocmask).

The rest looks ok.

> +#elif __glibc_has_attribute (__returns_twice__)
> +# define __INDIRECT_RETURN __attribute__ ((__returns_twice__))
> +#else
> +# define __INDIRECT_RETURN
> +#endif
> diff --git a/string/tst-xbzero-opt.c b/string/tst-xbzero-opt.c
> index 3aa13d03..26e2ec1c 100644
> --- a/string/tst-xbzero-opt.c
> +++ b/string/tst-xbzero-opt.c
> @@ -111,10 +111,9 @@ use_test_buffer (unsigned char *buf)
>  /* Always check the test buffer immediately after filling it; this
>     makes externally visible side effects depend on the buffer existing
>     and having been filled in.  */
> -#if defined __CET__ && !__glibc_has_attribute (__indirect_return__)
> -/* Note: swapcontext returns via indirect branch when SHSTK is enabled.
> -   Without indirect_return attribute, swapcontext is marked with
> -   returns_twice attribute, which prevents always_inline to work.  */
> +#if !__glibc_has_attribute (__indirect_return__)
> +/* Without indirect_return, swapcontext is marked returns_twice, which
> +   prevents always_inline from working.  */
>  # define ALWAYS_INLINE
>  #else
>  # define ALWAYS_INLINE	__attribute__ ((always_inline))