This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] i386: Don't use frame pointer without stack access

On Wed, Aug 9, 2017 at 4:22 AM, Richard Biener
<> wrote:
> On August 9, 2017 9:53:05 AM GMT+02:00, Richard Sandiford <> wrote:
>>Richard Biener <> writes:
>>> On August 8, 2017 7:36:35 PM GMT+02:00, Richard Sandiford
>>> <> wrote:
>>>>Richard Sandiford <> writes:
>>>>> Richard Biener <> writes:
>>>>>> On August 8, 2017 6:38:30 PM GMT+02:00, "H.J. Lu"
>>>><> wrote:
>>>>>>>On Mon, Aug 7, 2017 at 1:05 PM, Richard Sandiford
>>>>>>><> wrote:
>>>>>>>> Arjan van de Ven <> writes:
>>>>>>>>> On 8/7/2017 8:43 AM, Jakub Jelinek wrote:
>>>>>>>>>> On Mon, Aug 07, 2017 at 08:39:24AM -0700, H.J. Lu wrote:
>>>>>>>>>>> When Linux/x86-64 kernel is compiled with
>>>>>>>>>>> this optimization removes more than 730
>>>>>>>>>>> pushq %rbp
>>>>>>>>>>> movq %rsp, %rbp
>>>>>>>>>>> popq %rbp
>>>>>>>>>> If you don't want the frame pointer, why are you compiling
>>>>>>>>>> -fno-omit-frame-pointer?  Are you going to add
>>>>>>>>>> -fforce-no-omit-frame-pointer or something similar so that
>>>>>>>>>> actually get what they are asking for?  This doesn't really
>>>>>>>>>> It is perfectly fine to omit frame pointer by default, when it
>>>>>>>>>> required for something, but if the user asks for it, we
>>>>>>>ignore his
>>>>>>>>>> request.
>>>>>>>>> wanting a framepointer is very nice and desired...  ... but if
>>>>>>>>> optimizer/ins scheduler moves instructions outside of the
>>>>>>>>> portion, (it does it for cases like below as well), the value
>>>>>>>>> already negative for these functions that don't have stack use.
>>>>>>>>> <MPIDU_Sched_are_pending@@Base>:
>>>>>>>>> mov    all_schedules@@Base-0x38460,%rax
>>>>>>>>> push   %rbp
>>>>>>>>> mov    %rsp,%rbp
>>>>>>>>> pop    %rbp
>>>>>>>>> cmpq   $0x0,(%rax)
>>>>>>>>> setne  %al
>>>>>>>>> movzbl %al,%eax
>>>>>>>>> retq
>>>>>>>> Yeah, and it could be even weirder for big single-block
>>>>>>>> I think GCC has been doing this kind of scheduling of prologue
>>>>>>>> epilogue instructions for a while, so there hasn*t really been a
>>>>>>>> guarantee which parts of the function will have a new FP and
>>>>>>>> will still have the old one.
>>>>>>>> Also, with an arbitrarily-picked host compiler (GCC 6.3.1),
>>>>>>>> kicks in when the following is compiled with -O3
>>>>>>>>     void f (int *);
>>>>>>>>     void
>>>>>>>>     g (int *x)
>>>>>>>>     {
>>>>>>>>       for (int i = 0; i < 1000; ++i)
>>>>>>>>         x[i] += 1;
>>>>>>>>       if (x[0])
>>>>>>>>         {
>>>>>>>>           int temp;
>>>>>>>>           f (&temp);
>>>>>>>>         }
>>>>>>>>     }
>>>>>>>> so only the block with the call to f sets up FP.  The relatively
>>>>>>>> long-running loop runs with the caller's FP.
>>>>>>>> I hope we can go for a target-independent position that what
>>>>>>>> patch does is OK...
>>>>>>>In light of this,  I am resubmitting my patch.  I added 3 more
>>>>>>>and also handle:
>>>>>>>typedef int v8si __attribute__ ((vector_size (32)));
>>>>>>>foo (v8si *out_start, v8si *out_end, v8si *regions)
>>>>>>>    v8si base = regions[3];
>>>>>>>    *out_start = base;
>>>>>>>    *out_end = base;
>>>>>>>OK for trunk?
>>>>>> The invoker specified -fno-omit-frame-pointer, why did you
>>>>>> I'd argue it's OK when neither -f nor -fno- is explicitly
>>>>>> irrespective of the default in case we document the change but an
>>>>>> explicit -fno- is pretty clear.
>>>>> I don't buy that we're ignoring the user.  -fomit-frame-pointer
>>>>> that, when you're creating a frame, it's OK not to set up the frame
>>>>> pointer.  Forcing it off means that if you create a frame, you need
>>>>> to set up the frame pointer too.  But it doesn't say anything about
>>>>> whether the frame itself is needed.  I.e. it's
>>>>> rather than -fno-omit-frame.
>>> Isn't that a bit splitting hairs if you look at (past) history?
>>I guess it would have been splitting hairs in the days when they
>>amounted to the same thing, i.e. when there was no behaviour that
>>would match "-fomit-frame" and when the prologue and epilogue were
>>glued to the start and end of the function.  But that was quite a
>>long time ago.  Shrink-wrapping at least means that omitting the frame
>>and omitting the frame pointer are different things, and it seems
>>fair that -fomit-frame-pointer has followed the natural meaning.
>>> You could also interpret -fno-omit-frame-pointer as obviously forcing
>>> frame as otherwise there's nothing to omit...
>>But applying that kind of interpretation to something like
>>-maccumulate-outgoing-args would make inlining all calls within a
>>function invalid, since there'd no longer be arguments to accumulate.
>>I think this kind of disagreement just emphasises that if we really
>>need a "always emit a prologue at the very start, an epilogue at the
>>very end, and always use a frame pointer" option, we should add it
>>and document exactly what the guarantees are.  I don't think
>>-fno-omit-frame-pointer should be it, since as the replies earlier in
>>the thread said, the natural meaning of that option has its uses too.
> OK, but then both -f[no-]omit-frame-pointer do not have clearly defined semantics and thus shouldn't be exposed to the user?

-f[no-]omit-frame-pointer apply to cases where a new stack frame
is needed.  -fno-omit-frame-pointer allows you to unwind each
stack frame, not necessarily each function, via frame pointer.
 -fno-omit-frame-pointer may not create a new stack frame for
each function, similar to LTO or function inlining.  But you can still
unwind via frame pointer.  We never guarantee that we will create
a new stack frame for each function.  Some functions are inlined
completely.  Some just use the caller's stack frame.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]