This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] i386: Don't use frame pointer without stack access
On August 8, 2017 6:38:30 PM GMT+02:00, "H.J. Lu" <hjl.tools@gmail.com> wrote:
>On Mon, Aug 7, 2017 at 1:05 PM, Richard Sandiford
><richard.sandiford@linaro.org> wrote:
>> Arjan van de Ven <arjan@linux.intel.com> 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 -fno-omit-frame-pointer.
>>>>> 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 with
>>>> -fno-omit-frame-pointer? Are you going to add
>>>> -fforce-no-omit-frame-pointer or something similar so that people
>can
>>>> actually get what they are asking for? This doesn't really make
>sense.
>>>> It is perfectly fine to omit frame pointer by default, when it
>isn't
>>>> required for something, but if the user asks for it, we shouldn't
>ignore his
>>>> request.
>>>>
>>>
>>>
>>> wanting a framepointer is very nice and desired... ... but if the
>>> optimizer/ins scheduler moves instructions outside of the frame'd
>>> portion, (it does it for cases like below as well), the value is
>>> 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 functions.
>> I think GCC has been doing this kind of scheduling of prologue and
>> epilogue instructions for a while, so there hasn*t really been a
>> guarantee which parts of the function will have a new FP and which
>> will still have the old one.
>>
>> Also, with an arbitrarily-picked host compiler (GCC 6.3.1),
>shrink-wrapping
>> kicks in when the following is compiled with -O3
>-fno-omit-frame-pointer:
>>
>> 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 HJ*s
>> patch does is OK...
>>
>
>In light of this, I am resubmitting my patch. I added 3 more
>testcases
>and also handle:
>
>typedef int v8si __attribute__ ((vector_size (32)));
>
>void
>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 eliminate it? I'd argue it's OK when neither -f nor -fno- is explicitly specified irrespective of the default in case we document the change but an explicit -fno- is pretty clear.
And yes, unfortunate placement of frame pointer init/de-init should be fixed.
Richard.
>
>Thanks.