g++ 4.2.x x86: code generation for __sync_lock_test_and_set() - builtin

[Daniel Lohmann]

On 25.02.2008, at 16:03, Andrew Haley wrote:

> Daniel Lohmann wrote:
>> g++ 4.2.3
>> Hi,
>> I have the following code, which uses the new  
>> __sync_lock_test_and_set() builtin:
>> class Mutex {
>>  int locked;
>> public:
>>  Mutex() {
>>    locked = 0;
>>  }
>>  void lock();
>>  void unlock();
>> };
>> void Mutex::lock() {
>>  while( __sync_lock_test_and_set( &locked, 1) == 0 )
>>    ;
>> }
>> void Mutex::unlock() {
>>  __sync_lock_release( &locked );
>> }
>> After compiling with -03 -fomit-frame-pointer, the resulting code  
>> for the Mutex::lock() method looks as follows:
>> 00000010 <Mutex::lock()>:
>>  10:    8b 54 24 04              mov    0x4(%esp),%edx
>>  14:    b8 01 00 00 00           mov    $0x1,%eax
>>  19:    87 02                    xchg   %eax,(%edx)
>>  1b:    85 c0                    test   %eax,%eax
>>  1d:    74 f5                    je     14 <Mutex::lock()+0x4>
>>  1f:    f3 c3                    repz ret
>> I am wondering about the repz prefix before the ret. A "do RET  
>> until Z-Flag is set" obviously does not make sense from the  
>> functional point of view. So I assume that it actually is a side  
>> effects of the repz prefix that is exploited here to guarantee  
>> "something" with respect to instruction reordering, fetching,  
>> caching, or ...?
>> So what exactly is this "something"?
>> And what exactly could happen under which circumstances if we don't  
>> use it?
>> Google does not reveal much. If one googles for "repz ret" one gets  
>> a *load* of hits --  but just because of the fact that "ret"  
>> appears immediately after "repz" in the alphabetically sorted list  
>> of x86 instructions :-)
>
> If you grep the gcc source you'll find
>
> ;; Used by x86_machine_dependent_reorg to avoid penalty on single  
> byte RET
> ;; instruction Athlon and K8 have.
>
> (define_insn "return_internal_long"
> [(return)
>  (unspec [(const_int 0)] UNSPEC_REP)]
> "reload_completed"
> "rep\;ret"
> [(set_attr "length" "1")
>  (set_attr "length_immediate" "0")
>  (set_attr "prefix_rep" "1")
>  (set_attr "modrm" "0")])


Oh, it is something *that* machine specific...

Thanks Andrew!