https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87206
Bug ID: 87206
Summary: Suboptimal code generation for
__atomic_compare_exchange_n followed by a comparison
Product: gcc
Version: 8.2.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: iii at linux dot ibm.com
CC: krebbel at gcc dot gnu.org
Target Milestone: ---
I tried to build the example #5 from
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80080 on x86_64 and observed a
similar issue:
$ cat 1.c
extern void bar (int *);
void foo5(int *mem)
{
int oldval = 0;
__atomic_compare_exchange_n (mem, (void *) &oldval, 1,
1, __ATOMIC_ACQUIRE, __ATOMIC_RELAXED);
if (oldval != 0)
bar (mem);
}
$ gcc-8 -c 1.c -O3 -g
$ objdump -d 1.o
# skip
0000000000000000 <_foo5>:
0: 31 c0 xor %eax,%eax
2: ba 01 00 00 00 mov $0x1,%edx
7: f0 0f b1 17 lock cmpxchg %edx,(%rdi)
b: 85 c0 test %eax,%eax
d: 75 01 jne 10 <_foo5+0x10>
f: c3 retq
10: e9 00 00 00 00 jmpq 15 <_foo5+0x15>
We don't have to do "test %eax,%eax", because this information is already
available through ZF, which is set by CMPXCHG.
I wonder if it would be possible to come up with a common solution for all
architectures, including x86_64 and s390?
