Re: [PATCH] pr 63354 - gcc -pg -mprofile-kernel creates unused stack frames on leaf functions on ppc64le

[Iain Sandoe <iain@codesourcery.com>]

Hi Martin,

I've applied your latest patch to top of trunk and looked at the code gen on powerpc-darwin9 (and a cross from x86-64-darwin12 => powerpc64-linux-gnu).

On 15 Mar 2015, at 23:39, Martin Sebor wrote:

> On 03/14/2015 08:34 AM, Segher Boessenkool wrote:
>> On Fri, Mar 13, 2015 at 03:54:57PM -0600, Martin Sebor wrote:
>>> Attached is a patch that eliminates the unused stack frame
>>> allocated by gcc 5 with -pg -mprofile-kernel on powepc64le
>>> and brings the code into parity with previous gcc versions.
>>> 
>>> The patch doesn't do anything to change the emitted code
>>> when -mprofile-kernel is used without -pg. Since the former
>>> option isn't fully documented (as noted in pr 65372) it's
>>> unclear what effect it should be expected to have without
>>> -pg.
>> 
>> -mprofile-kernel does nothing without profiling enabled.  Maybe it
>> should just have been called -pk or something horrid like that.
>> 
>> The effect it should have is to do what the only user of the option
>> (the 64-bit PowerPC Linux kernel) wants.  The effect it does have
>> is to make the 64-bit ABI more like the 32-bit ABI for mcount.
> 
> Thanks for the review and the clarification. FWIW, I mentioned
> -pg because the reporter had noted that in prior versions of
> GCC specifying -pg in addition to -mprofile-kernel wasn't
> necessary to get the expected effect.
> 
>> 
>> 
>>> 2015-03-13  Anton Blanchard  <anton@samba.org>
>>> 
>>> 	PR target/63354
>>> 	* gcc/config/rs6000/linux64.h (ARGET_KEEP_LEAF_WHEN_PROFILED): Define.
>>                                        ^ typo

This ^ will cause a bootstrap fail for every rs6000 target that doesn't include linux64.h.
(because rs6000_keep_leaf_when_profiled will be "defined but unused").

Since ISTM you intend this to apply to all rs6000 sub-targets, you might as well move it to rs6000.h?

>> 
>>> 	* cc/config/rs6000/rs6000.c (rs6000_keep_leaf_when_profiled). New
>>           ^ typo                                                    ^ typo
>> 
>> It shouldn't have "gcc/" in the path names at all, actually.
> 
> Sorry, I must have mangled the ChangeLog sopmehow while copying
> it from one terminal to another. I fixed it in the new patch
> (attached) along with the other issues you pointed out.
> 
> I tested the changes in powerpc64*-linux-* native builds and on
> an x86_64 host in a build for the powerpc-unknown-linux-gnu and
> powerpc64-apple-darwin targets. Of these, the -mprofile-kernel
> option is only accepted for powerpc64*-linux-* (which was also
> confirmed by inspecting the sources) so I adjusted the test
> target accordingly and kept the body of
> rs6000_keep_leaf_when_profiled you suggested.
> 
> Martin
> 
>> 
>>> +/* -mprofile-kernel code calls mcount before the function prolog,
>> 
>> "prologue".
>> 
>>> +   so a profiled leaf function should stay a leaf function.  */
>>> +
>>> +static bool
>>> +rs6000_keep_leaf_when_profiled (void)
>>> +{
>>> +  return TARGET_PROFILE_KERNEL;
>>> +}
>> 
>> Something like
>> 
>>   switch (DEFAULT_ABI)
>>     {
>>       case ABI_AIX:
>>       case ABI_ELFv2:
>> 	return TARGET_PROFILE_KERNEL;
>> 
>>       default:
>> 	return true;
>>     }
>> 
>> although I'm not sure about Darwin here.  More conservative is to
>> return false for anything untested, of course.

The change is 'no-op' on Darwin, since we pass a parameter to mcount a stack frame is always forced.

>>> --- /dev/null
>>> +++ b/gcc/testsuite/gcc.target/powerpc/pr63354.c
>>> @@ -0,0 +1,10 @@
>>> +/* { dg-do compile { target { powerpc*-*-* } } } */
>>> +/* { dg-options "-O2 -pg -mprofile-kernel" } */
>>> +
>>> +int foo (void)
>>> +{
>>> +  return 1;
>>> +}
>>> +
>>> +/* { dg-final { scan-assembler "bl _mcount" } } */
>>> +/* { dg-final { scan-assembler-not "\(addi|stdu\) 1," } } */
>> 
>> Either you should run this only on AIX/ELFv2 ABIs, or you want to
>> test for "stwu" as well.  Bare "1" does not work for all assemblers
>> (only Darwin again?)

a bare register # will, indeed, fail for Darwin's native assembler (which expects r#).
cheers
Iain

>> 
>> 
>> Segher
>> 
> 
> <gcc-63354.patch>