[PATCH] gas: fix silent failure of .cfi_sections .sframe on unsupported targets

[Indu Bhagat]

On 3/13/26 6:23 AM, Jens Remus wrote:
> On 3/13/2026 9:16 AM, Jens Remus wrote:
>> On 3/12/2026 8:54 PM, Indu Bhagat wrote:
>>> On 3/10/26 5:23 AM, Jens Remus wrote:
>>>> On 3/10/2026 7:07 AM, Indu Bhagat via Binutils wrote:
>>>>> [Resending as I messed up one of the email in the to list in the
>>>>> previous attempt].
>>>>>
>>>>> Starting in Binutils 2.46, the .cfi_sections .sframe directive failed
>>>>> silently when targeting unsupported architectures (e.g., i386 or even
>>>>> x32 on x86_64).  This happened because the sframe_as_bad macro only
>>>>> checked for the command-line flag --gsframe (GEN_SFRAME_ENABLED) and
>>>>> ignored the explicit request via the assembler directive
>>>>> (CFI_EMIT_sframe).
>>>>>
>>>>> This patch ensures that an error is emitted (for unsupported targets) if
>>>>> SFrame is requested via either the command line or the .cfi_sections
>>>>> directive, while maintaining silence for 'opportunistic' sframe
>>>>> generation (GEN_SFRAME_CONFIG_ENABLED).
>>>>>
>>>>> Add two testcases:
>>>>>     - cfi-sframe-x86_64-err-1.d: usage of --32 with .cfi_sections .sframe,
>>>>>     - cfi-sframe-x86_64-err-2.d: usage of --gsframe --32 without
>>>>>       .cfi_sections .sframe
>>>>> GAS is expected to error out in both the cases.
>>>>>
>>>>> This fixes PR gas/33962 - .cfi_sections directive with .sframe on
>>>>> unsupported targets does not error.
>>>>>
>>>>> gas/
>>>>>      PR gas/33962
>>>>>      * dw2gencfi.c (dot_cfi_sections): Ensure as_bad is called if
>>>>>      SFrame was explicitly requested via directive.
>>>>>      * gen-sframe.h: Remove sframe_as_bad macro.
>>>>> gas/testsuite/
>>>>>      PR gas/33962
>>>>>      * gas/cfi-sframe/cfi-sframe-x86_64-err-1.s: New test.
>>>>>      * gas/cfi-sframe/cfi-sframe-x86_64-err-1.d: New test.
>>>>>      * gas/cfi-sframe/cfi-sframe-x86_64-err-2.s: New test.
>>>>>      * gas/cfi-sframe/cfi-sframe-x86_64-err-1.d: New test.
>>>>>      * gas/cfi-sframe/cfi-sframe.exp: Run the new tests.
>>>>
>>>> Thank you for fixing this issue!
>>>>
>>>>
>>>>> diff --git a/gas/dw2gencfi.c b/gas/dw2gencfi.c
>>>>
>>>>> @@ -2620,9 +2620,18 @@ cfi_finish (void)
>>>>>        }
>>>>>          else
>>>>>    #endif
>>>>> -    /* Avoid erroring with DEFAULT_SFRAME for non-default options, like
>>>>> -       -32 on x86_64.  */
>>>>> -    sframe_as_bad ("%s", _(".sframe not supported for target"));
>>>>> +    /* Issue an error for unsupported targets, like --32 on x86_64.  Avoid
>>>>> +       erroring when default-enabled at configure-time though, because we
>>>>> +       interpret default-enabled as "opportunistic SFrames".  Users don't
>>>>> +       want to be bothered by something preventing emission of SFrames in
>>>>> +       such a case.  */
>>>>> +    {
>>>>> +      if (flag_gen_sframe == GEN_SFRAME_ENABLED
>>>>
>>>> This issues an error for --gsframe[-N].
>>>>
>>>
>>> Yes. Because command line takes precedence.
>>>
>>>>> +          || ((all_cfi_sections & CFI_EMIT_sframe) != 0
>>>>> +          && flag_gen_sframe != GEN_SFRAME_DISABLED
>>>>> +          && flag_gen_sframe != GEN_SFRAME_CONFIG_ENABLED))
>>>>
>>>> IIUC this issues an error for .cfi_sections .sframe, but only if not
>>>> --gsframe=no, which takes precedence, and not 'opportunistic' SFrame
>>>> generation (GEN_SFRAME_CONFIG_ENABLED).
>>>>
>>>
>>> Correct.
>>>
>>>> Why should .cfi_sections .sframe be ignored when used with
>>>> 'opportunistic' SFrame generation (GEN_SFRAME_CONFIG_ENABLED)?
>>>>
>>>
>>> Umm, .cfi_sections .sframe is not being ignored with
>>> GEN_SFRAME_CONFIG_ENABLED, its only that the error is silenced when
>>> this occurs for unsupported target.
>>
>> Sorry, my wording was odd.  I meant: Why is an error for .cfi_sections
>> .sframe being silenced with GEN_SFRAME_CONFIG_ENABLED?
>>
>>> I think you meant to ask why silence that error:
>>> 'Opportunistic' SFrame is for the use case of distro wide builds.  If
>>> there is user code that builds with --32 or --x32 (in case x86_64),
>>> and gas is built with --enable-default-sframe, there will be errors of
>>> unsupported target.  These are considered noise in this setup, and
>>> hence the desire to silence them.
>>
>> I understand and completely agree that if gas is built with
>> --enable-default-sframe it should silence any SFrame related warnings/
>> errors, that result from the 'opportunistic' SFrame generation
>> (GEN_SFRAME_CONFIG_ENABLED).  That is whenever SFrame is generated
>> solely because of that configure option.
>>
>> My point is that if gas option --gsframe[-3] is specified or
>> .cfi_sections .sframe is coded, then SFrame is generated because the
>> user explicitly requested so and then it should not matter whether the
>> assembler is built with 'opportunistic' SFrame generation
>> (GEN_SFRAME_CONFIG_ENABLED) whether issue an error.
>>
>>> Does that make sense ?
>>
>> No.  Whenever explicitly requesting SFrame generation and this is not
>> supported an error/warning should be the result.  The only exception
>> should be 'opportunistic' SFrame generation (GEN_SFRAME_CONFIG_ENABLED)
>> from my POV (but not if it is explicitly requested).
> 
> Suppressing the ".sframe not supported for target" error in case of
> .cfi_sections .sframe if gas is configured with 'opportunistic' SFrame
> generation (--enable-default-sframe / GEN_SFRAME_CONFIG_ENABLED) can't
> be right, as that would suppress an error that would otherwise be
> emitted.  With your patch applied and gas built for x86-64:
> 
> gas with --disable-default-sframe (default) errors out, which is
> expected:
> 
> $ printf ".cfi_sections .sframe\n.cfi_startproc\n.cfi_endproc\n" | as --32 ; echo $?
> Assembler messages:
> {standard input}: Error: .sframe not supported for target
> 1
> 
> gas with --enable-default-sframe suppresses this error, which is
> unexpected as 'opportunistic' SFrame generation should not suppress
> errors that would otherwise be issued:
> 
> $ printf ".cfi_sections .sframe\n.cfi_startproc\n.cfi_endproc\n" | as --32 ; echo $?
> 0
> 

Jens, I understand your point of view; for now I mostly tried to keep 
the behavior as per the existing code and comments which hinted at no 
error is desired when opportunistic sframe IIUC.

Sam, I have a vague recollection of us talking about whats desirable 
here and having chosen the following behavior:
   - when .cfi-sections .sframe is present
   - User uses --32 or --x32 (unsupported target)
   - Opportunistic sframe enabled
   => no error

Switching to error does make sense: If user had --32 for some reason 
together with .cfi_sections .sframe, it was likely because they are 
relying on the error to be present.

Thanks
Indu

>>>>> +        as_bad (_(".sframe not supported for target"));
>>>>> +    }
>>>>>        }
>>>>>        if ((all_cfi_sections & CFI_EMIT_debug_frame) != 0)
>>>>
>>>>> diff --git a/gas/gen-sframe.h b/gas/gen-sframe.h
>>>>
>>>>> @@ -21,16 +21,6 @@
>>>>>    #ifndef GENSFRAME_H
>>>>>    #define GENSFRAME_H
>>>>>    -/* Errors shouldn't be emitted either if SFrames are default-enabled, as
>>>>> -   we interpret default-enabled as "opportunistic SFrames".  Users don't
>>>>> -   want to be bothered by something preventing emission of SFrames in
>>>>> -   such a case.  */
>>>>> -#define sframe_as_bad(format, ...) \
>>>>> -  do {                           \
>>>>> -    if (flag_gen_sframe == GEN_SFRAME_ENABLED) \
>>>>> -      as_bad (format, __VA_ARGS__);            \
>>>>> -  } while (0)
>>>>> -
>>>>>    /* The entity is not tracked.  */
>>>>>    #define SFRAME_FRE_ELEM_LOC_NONE    0
>>>>>    /* The location of the tracked entity is based on a register.  May or may not
>> Thanks and regards,
>> Jens
> 
> Regards,
> Jens