[PATCH] Fix x86 atomic_fetch_xor_release.

[PATCH] Fix x86 atomic_fetch_xor_release.

[Carlos O'Donell]

Torvald,

No code uses atomic_fetch_xor_release except for the upcoming conditional
variable rewrite. Therefore there is no user visible bug here.

The use of atomic_compare_and_exchange_bool_rel is removed (since it doesn't
exist anymore), and is replaced by atomic_compare_exchange_weak_release.

We use weak_release because it provides better performance in the loop
(the weak semantic) and because the xor is release MO (the release semantic).

We don't reload expected in the loop because atomic_compare_and_exchange_weak_release
does this for us as part of the CAS failure.

It is otherwise a fairly plain conversion that fixes building the new condvar
for 32-bit x86.

I have pushed the new condvar into Fedora Rawhide for testing.

OK to checkin?

Cheers,
Carlos.

2016-10-17  Carlos O'Donell  <carlos@redhat.com>

	* include/atomic.h
	[USE_COMPILER_ATOMIC_BUILTINS && !atomic_fetch_xor_release]
	(atomic_fetch_xor_release): Use atomic_compare_exchange_weak_release.

Index: glibc-2.24-256-g5140d03/include/atomic.h
===================================================================
--- glibc-2.24-256-g5140d03.orig/include/atomic.h
+++ glibc-2.24-256-g5140d03/include/atomic.h
@@ -777,18 +777,22 @@ void __atomic_link_error (void);
 # endif

 # ifndef atomic_fetch_xor_release
+/* Failing the atomic_compare_exchange_weak_release reloads the value in
+   __atg104_expected, so we need only do the XOR again and retry.  */
 # define atomic_fetch_xor_release(mem, operand) \
-  ({ __typeof (*(mem)) __atg104_old;                                         \
-     __typeof (mem) __atg104_memp = (mem);                                   \
+  ({ __typeof (mem) __atg104_memp = (mem);                                   \
+     __typeof (*(mem)) __atg104_expected = (*__atg104_memp);                 \
+     __typeof (*(mem)) __atg104_desired;                                     \
      __typeof (*(mem)) __atg104_op = (operand);                                      \
                                                                              \
      do                                                                              \
-       __atg104_old = (*__atg104_memp);                                              \
-     while (__builtin_expect                                                 \
-           (atomic_compare_and_exchange_bool_rel (                           \
-               __atg104_memp, __atg104_old ^ __atg104_op, __atg104_old), 0));\
+       __atg104_desired = __atg104_expected ^ __atg104_op;                   \
+     while (__glibc_unlikely                                                 \
+           (atomic_compare_exchange_weak_release (                           \
+               __atg104_memp, &__atg104_expected, __atg104_desired)          \
+            == 0));                                                          \
                                                                              \
-     __atg104_old; })
+     __atg104_expected; })
 #endif

 #endif /* !USE_ATOMIC_COMPILER_BUILTINS  */
---

-- 
Cheers,
Carlos.

Re: [PATCH] Fix x86 atomic_fetch_xor_release.

[Torvald Riegel]

On Mon, 2016-10-17 at 23:18 -0400, Carlos O'Donell wrote:
> Torvald,
> 
> No code uses atomic_fetch_xor_release except for the upcoming conditional
> variable rewrite. Therefore there is no user visible bug here.
> 
> The use of atomic_compare_and_exchange_bool_rel is removed (since it doesn't
> exist anymore), and is replaced by atomic_compare_exchange_weak_release.
> 
> We use weak_release because it provides better performance in the loop
> (the weak semantic) and because the xor is release MO (the release semantic).
> 
> We don't reload expected in the loop because atomic_compare_and_exchange_weak_release
> does this for us as part of the CAS failure.
> 
> It is otherwise a fairly plain conversion that fixes building the new condvar
> for 32-bit x86.
> 
> I have pushed the new condvar into Fedora Rawhide for testing.
> 
> OK to checkin?

OK.

Re: [PATCH] Fix x86 atomic_fetch_xor_release.

[Carlos O'Donell]

On 10/18/2016 07:02 AM, Torvald Riegel wrote:
> On Mon, 2016-10-17 at 23:18 -0400, Carlos O'Donell wrote:
>> Torvald,
>>
>> No code uses atomic_fetch_xor_release except for the upcoming conditional
>> variable rewrite. Therefore there is no user visible bug here.
>>
>> The use of atomic_compare_and_exchange_bool_rel is removed (since it doesn't
>> exist anymore), and is replaced by atomic_compare_exchange_weak_release.
>>
>> We use weak_release because it provides better performance in the loop
>> (the weak semantic) and because the xor is release MO (the release semantic).
>>
>> We don't reload expected in the loop because atomic_compare_and_exchange_weak_release
>> does this for us as part of the CAS failure.
>>
>> It is otherwise a fairly plain conversion that fixes building the new condvar
>> for 32-bit x86.
>>
>> I have pushed the new condvar into Fedora Rawhide for testing.
>>
>> OK to checkin?
> 
> OK.
> 

Checked in. Verified it fixes everything with the other dependent
patches applied.

-- 
Cheers,
Carlos.