[PATCH 2/4] gas: sframe: i386: have the backend specify the RA too

[Jan Beulich]

On 13.05.2025 21:01, Indu Bhagat wrote:
> On 5/13/25 7:54 AM, Jens Remus wrote:
>> Hello Jan,
>>
>> I agree with your feedback.  This is just to add some context.
>>
>> On 13.05.2025 14:57, Jan Beulich wrote:
>>> On 11.05.2025 09:35, Indu Bhagat wrote:
>>>> To process some CFI directives like .cfi_undefined and .cfi_same_value,
>>>> it is necessary for correctness to detect all cases when the register
>>>> used is one of SP, FP or RA.
>>>>
>>>> Currently, the backends needed to specify the RA only in the case of
>>>> those ABIs where RA tracking was necessary, e.g. AArch64.  For AMD64,
>>>> since the return address is always at a fixed offset from the CFA, RA
>>>> tracking was disabled.  That now needs to change as we should detect the
>>>> cases when RA is used like so:
>>>>     .cfi_undefined <RA>
>>>> or,
>>>>     .cfi_same_value <RA>
>>>
>>> I can't really make sense of this: As you say, the return address on x86
>>> is on the stack, not in a register. Yet both directives want a register
>>> specified. See also below. (In fact it's unclear to me what
>>>
>>> 	.cfi_same_value rip
>>>
>>> would actually mean: %rip is neither preserved nor can it remain
>>> unmodified.)
>>>
>>>> @@ -3656,6 +3657,7 @@ md_begin (void)
>>>>   #ifdef OBJ_ELF
>>>>         x86_sframe_cfa_sp_reg = REG_SP;
>>>>         x86_sframe_cfa_fp_reg = REG_FP;
>>>> +      x86_sframe_cfa_ra_reg = 16;
>>>
>>> Along the lines of the earlier comment asking to have a #define for this
>>> number - what would the identifier be? Not REG_RA, I would say, as that's
>>> not appropriate on x86, where there's conceptually no such register. Yet
>>> REG_IP also doesn't look right, as you're after the return address here.
>>
>> The background is to represent .cfi_undefined <RA> (for x86-64, due to
>> the lack of a RA register, potentially .cfi_undefined <IP> instead) in
>> SFrame, which is used in e.g. Glibc to identify the outermost frame,
>> which cannot be further unwound.  See instances of .cfi_undefined in
>> Glibc start.S:
>>
>> sysdeps/aarch64/start.S: cfi_undefined (x30): LR (=RA) is undefined
>> sysdeps/s390/s390-64/start.S: cfi_undefined (r14): RA is undefined
>> sysdeps/x86_64/start.S: cfi_undefined (rip): IP is undefined
>>
> 
> (Adding more context to what Jens mentions)
> 
> Its a provision in the DWARF standard:
> "If a Return Address register is defined in the virtual unwind table, 
> and its rule is
> undefined (for example, by DW_CFA_undefined), then there is no return 
> address
> and no call address, and the virtual unwind of stack activations is 
> complete."
> 
> The SysV ABI AMD64 doc does mention 16 to be the return address register 
> in the "DWARF Register Number Mapping" table:
> 
> Return Address RA 16
> 
> So, REG_RA can be mentioned as a pseudo register in the comments. 
> Something like:
> 
> /* DWARF register number of the (pseudo) return-address register in 64-bit
>     mode.  */
> #define REG_RA 16

Well, okay. The comment would better associate this with i386-reg.tbl's

rip, Qword, RegRex64, RegIP, Dw2Inval, 16

though, imo. (Eventually we may want to actually use the REG_* constants
there, but that's for a separate change.)

Jan