diff mbox

Re: [patch libstdc++] Optimize synchronization in std::future if futexes are available.

Message ID 20150117202356.GT3360@redhat.com
State New
Headers show

Commit Message

Jonathan Wakely Jan. 17, 2015, 8:23 p.m. UTC 
On 17/01/15 12:59 -0700, Sandra Loosemore wrote:
>Re:
>
>>On 17/01/15 01:45 -0500, Hans-Peter Nilsson wrote:
>>>On Fri, 16 Jan 2015, pinskia@gmail.com wrote:
>>>>> On Jan 16, 2015, at 9:57 PM, David Edelsohn <dje.gcc@gmail.com> wrote:
>>>>>
>>>>> This patch has broken bootstrap on AIX
>>>>>
>>>>> May I mention that this really should have been tested on systems
>>>>> other than x86 Linux.
>>>>
>>>>It also broke all newlib targets too. So you could have tested one listed in the sim-test web page.
>>>
>>>For those interested, PR64638.
>>
>>Should be fixed in trunk now, by this patch.
>
>I'm now getting this error in an arm-none-linux-gnueabi cross build:
>
>
>In file included from /scratch/sandra/arm-fsf2/obj/gcc-mainline-0-arm-none-linux-gnueabi-i686-pc-linux-gnu/arm-none-linux-gnueabi/libstdc++-v3/include/future:44:0,
>                 from /scratch/sandra/arm-fsf2/src/gcc-mainline/libstdc++-v3/src/c++11/functexcept.cc:34:
>/scratch/sandra/arm-fsf2/obj/gcc-mainline-0-arm-none-linux-gnueabi-i686-pc-linux-gnu/arm-none-linux-gnueabi/libstdc++-v3/include/bits/atomic_futex.h:71:3: 
>error: #error We require lock-free atomic operations on int
> # error We require lock-free atomic operations on int
>   ^
>
>It used to work a few days ago....  nothing changed in my build 
>environment except that I did "svn up" in my gcc source directory....

Well that file didn't exist until yesterday :-)

Does the attached patch fix it?

The new __atomic_futex_unsigned type is only used in <future> when
ATOMIC_LOCK_FREE > 1, so there's no point trying to define it and
then failing if int is not lock-free, as the type isn't going to be
used anyway.

Comments

Sandra Loosemore Jan. 17, 2015, 10:22 p.m. UTC | #1 
On 01/17/2015 01:23 PM, Jonathan Wakely wrote:
> On 17/01/15 12:59 -0700, Sandra Loosemore wrote:
>> Re:
>>
>>> On 17/01/15 01:45 -0500, Hans-Peter Nilsson wrote:
>>>> On Fri, 16 Jan 2015, pinskia@gmail.com wrote:
>>>>>> On Jan 16, 2015, at 9:57 PM, David Edelsohn <dje.gcc@gmail.com>
>>>>>> wrote:
>>>>>>
>>>>>> This patch has broken bootstrap on AIX
>>>>>>
>>>>>> May I mention that this really should have been tested on systems
>>>>>> other than x86 Linux.
>>>>>
>>>>> It also broke all newlib targets too. So you could have tested one
>>>>> listed in the sim-test web page.
>>>>
>>>> For those interested, PR64638.
>>>
>>> Should be fixed in trunk now, by this patch.
>>
>> I'm now getting this error in an arm-none-linux-gnueabi cross build:
>>
>>
>> In file included from
>> /scratch/sandra/arm-fsf2/obj/gcc-mainline-0-arm-none-linux-gnueabi-i686-pc-linux-gnu/arm-none-linux-gnueabi/libstdc++-v3/include/future:44:0,
>>
>>                 from
>> /scratch/sandra/arm-fsf2/src/gcc-mainline/libstdc++-v3/src/c++11/functexcept.cc:34:
>>
>> /scratch/sandra/arm-fsf2/obj/gcc-mainline-0-arm-none-linux-gnueabi-i686-pc-linux-gnu/arm-none-linux-gnueabi/libstdc++-v3/include/bits/atomic_futex.h:71:3:
>> error: #error We require lock-free atomic operations on int
>> # error We require lock-free atomic operations on int
>>   ^
>>
>> It used to work a few days ago....  nothing changed in my build
>> environment except that I did "svn up" in my gcc source directory....
>
> Well that file didn't exist until yesterday :-)
>
> Does the attached patch fix it?
>
> The new __atomic_futex_unsigned type is only used in <future> when
> ATOMIC_LOCK_FREE > 1, so there's no point trying to define it and
> then failing if int is not lock-free, as the type isn't going to be
> used anyway.

I'm getting a different series of build errors with this patch -- 
starting with

In file included from 
/scratch/sandra/arm-fsf2/src/gcc-mainline/libstdc++-v3/src/c++11/futex.cc:27:0:
/scratch/sandra/arm-fsf2/obj/gcc-mainline-0-arm-none-linux-gnueabi-i686-pc-linux-gnu/arm-none-linux-gnueabi/libstdc++-v3/include/bits/atomic_futex.h:221:5: 
error: 'mutex' does not name a type
      mutex _M_mutex;
      ^
/scratch/sandra/arm-fsf2/obj/gcc-mainline-0-arm-none-linux-gnueabi-i686-pc-linux-gnu/arm-none-linux-gnueabi/libstdc++-v3/include/bits/atomic_futex.h:222:5: 
error: 'condition_variable' does not name a type
      condition_variable _M_condvar;
      ^
/scratch/sandra/arm-fsf2/obj/gcc-mainline-0-arm-none-linux-gnueabi-i686-pc-linux-gnu/arm-none-linux-gnueabi/libstdc++-v3/include/bits/atomic_futex.h: 
In member function 'unsigned int 
std::__atomic_futex_unsigned<_Waiter_bit>::_M_load(std::memory_
order)':
/scratch/sandra/arm-fsf2/obj/gcc-mainline-0-arm-none-linux-gnueabi-i686-pc-linux-gnu/arm-none-linux-gnueabi/libstdc++-v3/include/bits/atomic_futex.h:230:7: 
error: 'unique_lock' was not declared in this scope
        unique_lock<mutex> __lock(_M_mutex);
        ^
and going on for several screenfuls.

Maybe the patch(es) causing all these problems should be reverted until 
the breakage is tracked down and fixed and regression-tested on a 
variety of targets?  It's not really blocking me at the moment, but it 
seems unfortunate that trunk is so unstable as we're entering Stage 4.....

-Sandra
Jonathan Wakely Jan. 17, 2015, 10:36 p.m. UTC | #2 
On 17/01/15 15:22 -0700, Sandra Loosemore wrote:
>On 01/17/2015 01:23 PM, Jonathan Wakely wrote:
>>On 17/01/15 12:59 -0700, Sandra Loosemore wrote:
>>>Re:
>>>
>>>>On 17/01/15 01:45 -0500, Hans-Peter Nilsson wrote:
>>>>>On Fri, 16 Jan 2015, pinskia@gmail.com wrote:
>>>>>>>On Jan 16, 2015, at 9:57 PM, David Edelsohn <dje.gcc@gmail.com>
>>>>>>>wrote:
>>>>>>>
>>>>>>>This patch has broken bootstrap on AIX
>>>>>>>
>>>>>>>May I mention that this really should have been tested on systems
>>>>>>>other than x86 Linux.
>>>>>>
>>>>>>It also broke all newlib targets too. So you could have tested one
>>>>>>listed in the sim-test web page.
>>>>>
>>>>>For those interested, PR64638.
>>>>
>>>>Should be fixed in trunk now, by this patch.
>>>
>>>I'm now getting this error in an arm-none-linux-gnueabi cross build:
>>>
>>>
>>>In file included from
>>>/scratch/sandra/arm-fsf2/obj/gcc-mainline-0-arm-none-linux-gnueabi-i686-pc-linux-gnu/arm-none-linux-gnueabi/libstdc++-v3/include/future:44:0,
>>>
>>>                from
>>>/scratch/sandra/arm-fsf2/src/gcc-mainline/libstdc++-v3/src/c++11/functexcept.cc:34:
>>>
>>>/scratch/sandra/arm-fsf2/obj/gcc-mainline-0-arm-none-linux-gnueabi-i686-pc-linux-gnu/arm-none-linux-gnueabi/libstdc++-v3/include/bits/atomic_futex.h:71:3:
>>>error: #error We require lock-free atomic operations on int
>>># error We require lock-free atomic operations on int
>>>  ^
>>>
>>>It used to work a few days ago....  nothing changed in my build
>>>environment except that I did "svn up" in my gcc source directory....
>>
>>Well that file didn't exist until yesterday :-)
>>
>>Does the attached patch fix it?
>>
>>The new __atomic_futex_unsigned type is only used in <future> when
>>ATOMIC_LOCK_FREE > 1, so there's no point trying to define it and
>>then failing if int is not lock-free, as the type isn't going to be
>>used anyway.
>
>I'm getting a different series of build errors with this patch -- 
>starting with
>
>In file included from /scratch/sandra/arm-fsf2/src/gcc-mainline/libstdc++-v3/src/c++11/futex.cc:27:0:
>/scratch/sandra/arm-fsf2/obj/gcc-mainline-0-arm-none-linux-gnueabi-i686-pc-linux-gnu/arm-none-linux-gnueabi/libstdc++-v3/include/bits/atomic_futex.h:221:5: 
>error: 'mutex' does not name a type
>     mutex _M_mutex;
>     ^
>/scratch/sandra/arm-fsf2/obj/gcc-mainline-0-arm-none-linux-gnueabi-i686-pc-linux-gnu/arm-none-linux-gnueabi/libstdc++-v3/include/bits/atomic_futex.h:222:5: 
>error: 'condition_variable' does not name a type
>     condition_variable _M_condvar;
>     ^
>/scratch/sandra/arm-fsf2/obj/gcc-mainline-0-arm-none-linux-gnueabi-i686-pc-linux-gnu/arm-none-linux-gnueabi/libstdc++-v3/include/bits/atomic_futex.h: 
>In member function 'unsigned int 
>std::__atomic_futex_unsigned<_Waiter_bit>::_M_load(std::memory_
>order)':
>/scratch/sandra/arm-fsf2/obj/gcc-mainline-0-arm-none-linux-gnueabi-i686-pc-linux-gnu/arm-none-linux-gnueabi/libstdc++-v3/include/bits/atomic_futex.h:230:7: 
>error: 'unique_lock' was not declared in this scope
>       unique_lock<mutex> __lock(_M_mutex);
>       ^
>and going on for several screenfuls.

Odd, that should already be fixed at rev 219799.

Are you still at a revision older than that?

>Maybe the patch(es) causing all these problems should be reverted 
>until the breakage is tracked down and fixed and regression-tested on 
>a variety of targets?  It's not really blocking me at the moment, but 
>it seems unfortunate that trunk is so unstable as we're entering Stage 
>4.....
Sandra Loosemore Jan. 17, 2015, 10:54 p.m. UTC | #3 
On 01/17/2015 03:36 PM, Jonathan Wakely wrote:
> On 17/01/15 15:22 -0700, Sandra Loosemore wrote:
> [snip snip]
>> I'm getting a different series of build errors with this patch --
>> starting with
>>
>> In file included from
>> /scratch/sandra/arm-fsf2/src/gcc-mainline/libstdc++-v3/src/c++11/futex.cc:27:0:
>>
>> /scratch/sandra/arm-fsf2/obj/gcc-mainline-0-arm-none-linux-gnueabi-i686-pc-linux-gnu/arm-none-linux-gnueabi/libstdc++-v3/include/bits/atomic_futex.h:221:5:
>> error: 'mutex' does not name a type
>>     mutex _M_mutex;
>>     ^
>> /scratch/sandra/arm-fsf2/obj/gcc-mainline-0-arm-none-linux-gnueabi-i686-pc-linux-gnu/arm-none-linux-gnueabi/libstdc++-v3/include/bits/atomic_futex.h:222:5:
>> error: 'condition_variable' does not name a type
>>     condition_variable _M_condvar;
>>     ^
>> /scratch/sandra/arm-fsf2/obj/gcc-mainline-0-arm-none-linux-gnueabi-i686-pc-linux-gnu/arm-none-linux-gnueabi/libstdc++-v3/include/bits/atomic_futex.h:
>> In member function 'unsigned int
>> std::__atomic_futex_unsigned<_Waiter_bit>::_M_load(std::memory_
>> order)':
>> /scratch/sandra/arm-fsf2/obj/gcc-mainline-0-arm-none-linux-gnueabi-i686-pc-linux-gnu/arm-none-linux-gnueabi/libstdc++-v3/include/bits/atomic_futex.h:230:7:
>> error: 'unique_lock' was not declared in this scope
>>       unique_lock<mutex> __lock(_M_mutex);
>>       ^
>> and going on for several screenfuls.
>
> Odd, that should already be fixed at rev 219799.
>
> Are you still at a revision older than that?

My source tree is at r219802 now.
>
>> Maybe the patch(es) causing all these problems should be reverted
>> until the breakage is tracked down and fixed and regression-tested on
>> a variety of targets?  It's not really blocking me at the moment, but
>> it seems unfortunate that trunk is so unstable as we're entering Stage
>> 4.....

-Sandra
diff mbox

Patch

diff --git a/libstdc++-v3/include/bits/atomic_futex.h b/libstdc++-v3/include/bits/atomic_futex.h
index 2673604..b4138ba 100644
--- a/libstdc++-v3/include/bits/atomic_futex.h
+++ b/libstdc++-v3/include/bits/atomic_futex.h
@@ -49,7 +49,7 @@  namespace std _GLIBCXX_VISIBILITY(default)
 _GLIBCXX_BEGIN_NAMESPACE_VERSION
 
 #if defined(_GLIBCXX_HAS_GTHREADS) && defined(_GLIBCXX_USE_C99_STDINT_TR1)
-#if defined(_GLIBCXX_HAVE_LINUX_FUTEX)
+#if defined(_GLIBCXX_HAVE_LINUX_FUTEX) && ATOMIC_INT_LOCK_FREE > 1
   struct __atomic_futex_unsigned_base
   {
     // Returns false iff a timeout occurred.
@@ -66,10 +66,7 @@  _GLIBCXX_BEGIN_NAMESPACE_VERSION
   {
     typedef chrono::system_clock __clock_t;
 
-    // XXX We expect this to be lock-free, and having the payload at offset 0.
-#if ATOMIC_INT_LOCK_FREE < 2
-# error We require lock-free atomic operations on int
-#endif
+    // This must be lock-free and at offset 0.
     atomic<unsigned> _M_data;
 
     __atomic_futex_unsigned(unsigned __data) : _M_data(__data)
@@ -281,7 +278,7 @@  _GLIBCXX_BEGIN_NAMESPACE_VERSION
 
   };
 
-#endif // _GLIBCXX_HAVE_LINUX_FUTEX
+#endif // _GLIBCXX_HAVE_LINUX_FUTEX && ATOMIC_INT_LOCK_FREE > 1
 #endif // _GLIBCXX_HAS_GTHREADS && _GLIBCXX_USE_C99_STDINT_TR1
 
 _GLIBCXX_END_NAMESPACE_VERSION