Re: [PATCH, RFC] x86: also CFI-annotate certain inline asm()s

["Jan Beulich" <JBeulich@suse.com>]

>>> On 05.11.14 at 18:23, <luto@amacapital.net> wrote:
> On Wed, Nov 5, 2014 at 9:13 AM, Jan Beulich <jbeulich@suse.com> wrote:
>>>>> Andy Lutomirski <luto@amacapital.net> 11/04/14 8:40 PM >>>
>>>On 11/04/2014 01:24 AM, Jan Beulich wrote:
>>>> The main obstacle to having done this long ago was the need to
>>>> determine whether annotations are needed in the first place: They need
>>>> to be avoided when a frame pointer got set up. Since I can't see a way
>>>> to determine this before the compilation phase, this is being achieved
>>>> by inspecting the memory address generated by the compiler in an
>>>> interposed assembler macro. Of course this isn't really nice code, and
>>>> this the main reason I'm posting this as RFC only at this point (with
>>>> the hope that maybe someone has an idea of how to achieve the same
>>>> thing in a more elegant way).
>>>
>>>Ask binutils for help?
>>
>> Binutils know as little about the code the compiler generated as we do.
> 
> Could binutils add a
> .cfi_adjust_cfa_offset_if_the_cfa_depends_on_sp_right_now directive?
> IIUC, the issue is that, when you push, you don't want the canonical
> frame address to change as a result, but you just changed the stack
> pointer, so if the CFA is computed as an offset from the stack pointer
> in the current context, that offset needs to change.

While that's theoretically doable, I don't think this would be a
reasonable approach.

> Alternatively, is there any sane way to get the inline asm to act as
> though it creates an entirely new frame?  It would have CFA == rsp
> initially (or rsp + 8 or whatever -- I can never keep track of what
> the CFA is actually supposed to point to) and unwind instructions that
> tell the unwinder that the caller pc is at a known address instead of
> being stuck in the stack frame?

No, that can't work: You'd have to
- end the previous function (from the CFI engine's pov)
- start a new function
- do what you suggest above
- end the "nested" function
- start a continuation function for the subsequent compiler
  generated code
- magically know the state of things at the point the original
  function got (artificially) ended

>>>Is the issue that the CFI annotation you need is different depending on
>>>whether there's a frame pointer or not?
>>
>> No - as said above, they need to be avoided altogether when there's a
>> frame pointer.
>>
>>> If so, can you add some
>>>comments so that mere asm mortals have some prayer of understanding how
>>>your magic works and what the desired output annotations are in the
>>>various cases?
>>
>> Honestly I have a hard time seeing where comments would help here. Plus
>> the difficult part isn't how the annotations look like, but (see above) 
> simply
>> whether to emit them at all.
> 
> It could be something simple like an example of what the inputs to the
> asm macros are in the two cases.  Currently even figuring out where
> those inputs come from involves following a big tangle of C
> preprocessor stuff, and I don't know what it's supposed to output, and
> what that's supposed to do the expansions in the inline asm, and how
> that ends up influencing the gas macros.
> 
> I.e. I sort of expect I'll need to want to one of these things some
> day, and I'd like a couple pointers :)

I'll see what I can do (but I'll invest time into doing so only if there
are at least signs of this having a remote chance of going in at some
point).

Jan