Re: Divide-by-zero in post_init_entity_util_avg
From: Peter Zijlstra
Date: Fri Jun 17 2016 - 04:24:10 EST
On Fri, Jun 17, 2016 at 11:16:24AM +0300, Andrey Ryabinin wrote:
> > I suppose instead we'd want something like:
> >
> > tmp = READ_ONCE(*ptr);
> > if (tmp > val)
> > tmp -= val;
> > else
> > tmp = 0;
> > WRITE_ONCE(*ptr, tmp);
> >
> > In order to generate:
> >
> > xchg %rax,0xa0(%r13)
> > mov 0x78(%r13),%rcx
> > sub %rax,%rcx
> > cmovae %r15,%rcx
> > mov %rcx,0x78(%r13)
> >
> > however, GCC isn't smart enough and generates:
> >
> > xchg %rax,0x98(%r13)
> > mov 0x70(%r13),%rsi
> > mov %rsi,%rcx
> > sub %rax,%rcx
> > cmp %rsi,%rax
> > cmovae %r15,%rcx
> > mov %rcx,0x70(%r13)
> >
> > Doing a CMP with the _same_ values it does the SUB with, resulting in
> > exactly the same CC values.
> >
>
> FYI - https://gcc.gnu.org/bugzilla/show_bug.cgi?id=3507 (Reported: 2001-07-01)
>
I found this one when I was googling yesterday:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=30315
But yes, it seems this is a 'known' issue.