On Thu, Aug 10, 2017 at 9:09 AM, Richard Sandiford
<richard.sandif...@linaro.org> wrote:
> "H.J. Lu" <hjl.to...@gmail.com> writes:
>> On Wed, Aug 9, 2017 at 10:28 AM, H.J. Lu <hjl.to...@gmail.com> wrote:
>>> On Wed, Aug 9, 2017 at 8:26 AM, Andi Kleen <a...@firstfloor.org> wrote:
>>>>> This must be much more specific.  How does it impact:
>>>>> 1. LTO
>>>>> 2. Function inlining.
>>>>> 3. Partial function inlining.
>>>>> 4. Shrink-wrapping.
>>>>> Any of them can impact function stack frame.
>>>> It doesn't. It's just to get back to the previous state.
>>>> Also these others already have explicit options to disable them.
>>> How about
>>> item -fkeep-frame-pointer
>>> @opindex fkeep-frame-pointer
>>> Keep the frame pointer in a register for functions.  This option always
>>> forces a new stack frame for all functions regardless of whether
>>> @option{-fomit-frame-pointer} is enabled or not.  Disabled by default.
>> Here is the updated patch with -fkeep-frame-pointer.
> It sounds like there's still some disagreement about what this option
> should mean; Andi's and Arjan's replies didn't seem to be asking for
> the same thing.
> I think as implemented the patch just retains the GCC 7 x86 behaviour of
> -fno-omit-frame-pointer, i.e. forces a frame to be created *somewhere*
> between the start and end addresses of the function, but makes no
> guarantee where.  It could be a bundle of three instructions somewhere
> in the middle of a basic block, and the code might not be executed for
> all paths through the function (e.g. it might only be executed on
> error paths).
> I think even if we think that's useful, it should be documented clearly.
> Otherwise people might assume that a function f is guaranteed to set up
> a frame every time it's called.

As said earlier, I think we should proceed with the previous patch
(that documents -fno-omit-frame-pointer limitations), It was
demonstrated that the patch does not make current situation worse.

-fkeep-frame-pointer is an orthogonal issue, and this option should
guarantee frame formation in *all* situations.This means that the
option should disable (partial) inlining, shrink-wrapping, etc.
Actually, it would disable so much optimizations, that its usefullnes
is questioned. OTOH, nobody ever complained about limited FP debugging
"experience" when mentioned optimizations were activated.

BTW: The option should be called -fforce-frame-pointer, but I really
doubt about its usefullnes.


Reply via email to