Surprising code generated for vdso_read_begin()

[Christophe Leroy]

Le 09/01/2020 à 21:07, Segher Boessenkool a écrit :
> On Thu, Jan 09, 2020 at 05:52:34PM +0000, Christophe Leroy wrote:
>> Wondering why we get something so complicated/redundant for
>> vdso_read_begin() <include/vdso/helpers.h>
>>
>> static __always_inline u32 vdso_read_begin(const struct vdso_data *vd)
>> {
>> 	u32 seq;
>>
>> 	while ((seq = READ_ONCE(vd->seq)) & 1)
>> 		cpu_relax();
>>
>> 	smp_rmb();
>> 	return seq;
>> }
>>
>>
>>   6e0:   81 05 00 f0     lwz     r8,240(r5)
>>   6e4:   71 09 00 01     andi.   r9,r8,1
>>   6e8:   41 82 00 10     beq     6f8 <__c_kernel_clock_gettime+0x158>
>>   6ec:   81 05 00 f0     lwz     r8,240(r5)
>>   6f0:   71 0a 00 01     andi.   r10,r8,1
>>   6f4:   40 82 ff f8     bne     6ec <__c_kernel_clock_gettime+0x14c>
>>   6f8:
>>
>> r5 being vd pointer
>>
>> Why the first triplet, not only the second triplet ? Something wrong
>> with using READ_ONCE() for that ?
> 
> It looks like the compiler did loop peeling.  What GCC version is this?
> Please try current trunk (to become GCC 10), or at least GCC 9?
> 

It is with GCC 5.5

https://mirrors.edge.kernel.org/pub/tools/crosstool/ doesn't have more 
recent than 8.1

With 8.1, the problem doesn't show up.

Thanks
Christophe