http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59785
Bug ID: 59785
Summary: atomic_store should load source operand atomically?
Product: gcc
Version: 4.8.2
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: ma.jiang at zte dot com.cn
gcc provide a atomic_store interface as below.
void __atomic_store (type *ptr, type *val, int memmodel)
In manual, There is only a simple description ---"This is the generic version
of an atomic store. It stores the value of *val into *ptr."
But it seemed that this interface could not do what the description said when
*val is a global memory which might be messed up by other threads.
unsigned long long int ta=123;
unsigned long long int tb=321;
int test()
{
__atomic_store(&tb, &ta, __ATOMIC_RELAXED);
return 0;
}
On ARM,the codes above give out a assembly instruction sequence as below.
movw r3, #:lower16:ta
movt r3, #:upper16:ta
ldrd r0, [r3]
movw r3, #:lower16:tb
movt r3, #:upper16:tb
.L3:
ldrexd r4, r5, [r3]
strexd r2, r0, r1, [r3]
cmp r2, #0
bne .L3
the source operand ta was not fetched atomically, so we might get a totally
wrong num from tb when ta were modified by other threads.
I guess the atomic_store interface only provide a function as--"fetch val
non-atomically ,and then store it into ptr atomically."
This is exactlly what the gimple tree shows.
test ()
{
long long unsigned int ta.0;
int D.4127;
ta.0 = ta;
__atomic_store_8 (&tb, ta.0, 0);
D.4127 = 0;
return D.4127;
}
I think it would be better to give a more detailed desctription in the manual.
Or make a atomical load in the atomic_store.
