[OE-core] boost 1.56 compile fail

[Peter A. Bigot]

On 08/29/2014 04:18 PM, Dan McGregor wrote:
> On 29 August 2014 14:58, Peter A. Bigot <pab at pabigot.com> wrote:
>> On 08/29/2014 03:36 PM, Dan McGregor wrote:
>>> On 29 August 2014 05:48, Peter A. Bigot <pab at pabigot.com> wrote:
>>>> On 08/29/2014 06:28 AM, Yi Qingliang wrote:
>>>>
>>>> hardware: samsung s3c6410
>>>>
>>>> after updated to latest poky, the boost compile fail!
>>>>
>>>> error info:
>>>> libs/atomic/src/lockpool.cpp:127:5: error: 'thread_fence' is not a member
>>>> of
>>>> 'boost::atomics::detail'
>>>> libs/atomic/src/lockpool.cpp:138:5: error: 'signal_fence' is not a member
>>>> of
>>>> 'boost::atomics::detail'
>>>>
>>>>
>>>> after dig into it, I found that:
>>>> the marco 'BOOST_ATOMIC_FLAG_LOCK_FREE' is 0, so it don't include
>>>> 'operations_lockfree.hpp' which has 'thread_fence' and 'signal_fence',
>>>> but
>>>> pthread.h at line 21.
>>>>
>>>> in file 'caps_gcc_atomic.hpp', 'BOOST_ATOMIC_FLAG_LOCK_FREE' is set to
>>>> '0',
>>>> the author think if '__GCC_ATOMIC_BOOL_LOCK_FREE' is 1, the atomic serial
>>>> function gcc provided is not lock free.
>>>>
>>>>
>>>> This is the sort of GCC internal header indicator that would have changed
>>>> value as a result of:
>>>>
>>>>
>>>> http://cgit.openembedded.org/openembedded-core/commit/meta/recipes-devtools/gcc/gcc-configure-common.inc?id=0ba6ab39f187ecd4261f08e768f365f461384a3a
>>>>
>>>>
>>>>
>>>> at the end of 'caps_gcc_atomic.hpp', it defined
>>>> 'BOOST_ATOMIC_THREAD_FENCE'
>>>> as 2.
>>>>
>>>> so the conflict is: BOOST_ATOMIC_THREAD_FENCE and
>>>> BOOST_ATOMIC_FLAG_LOCK_FREE
>>>>
>>>> I don't know it is the new poky problem, or the boost problem, any idea?
>>>>
>>>>
>>>> My guess is that Boost is making assumptions about what the internal GCC
>>>> predefined symbols mean that aren't entirely accurate.  There are several
>>>> flags that are used in the libstdc++ headers to indicate whether the
>>>> compiler is using lock-free instructions.
>>>>
>>>> Boost-1.56 builds without error for my beaglebone target with poky at:
>>>>
>>>> * 669c07d (HEAD, origin/master, origin/HEAD, master/upstream, master/dev)
>>>> [Wed Aug 27 14:24:52 2014 +0100] bitbake: build/data: Write out more
>>>> complete python run files
>>>>
>>>> so it may have something to do with your target machine.
>>> It absolutely does. I found that armv6 breaks, but armv6zk and newer work.
>>
>> Interesting.  There are no armv6zk tune features I can see in poky, though
>> google suggests it applies to the Raspberry Pi.
>>
>> The problem then must be with the first override in this:
>>
>> # ARMv6+ adds atomic instructions that affect the ABI in libraries built
>> # with TUNE_CCARGS in gcc-runtime.  Make the compiler default to a
>> # compatible architecture.  armv6 and armv7a cover the minimum tune
>> # features used in OE.
>> EXTRA_OECONF_append_armv6 = " --with-arch=armv6"
>> EXTRA_OECONF_append_armv7a = " --with-arch=armv7-a"
>>
>> ARMv6 has LDREX/STREX, but ARMv6K adds {LD,ST}REX{B,H,D}.  The same problem
>> addressed above is likely to happen if the libraries are built with armv6k
>> but the compiler doesn't default to it.
>>
>> There are no armv6k tune parameters I can locate in poky.  What layers have
>> the tune configurations that are causing problems?
>>
> For me meta-raspberrypi failed to build. Its tuning is -march=armv6
> -mtune=arm1176zjf-s by default. I forced it to -march=armv6zk
> -mtune=arm1176jzf-s, and that worked.

tl;dr: for now, this can be claimed to be a boost problem, but it may 
rapidly become an OE problem.

OK, so there's several issues here.

Extracting the predefined symbols from gcc 4.9.1 with:

    arm-poky-linux-gnueabi-g++ -march=armv6 -dM -E -xc++ /dev/null

and similarly with -march=armv6k shows that the values of these 
atomic-related predefines are different (- = arvm6, + = armv6k):

-#define __GCC_ATOMIC_BOOL_LOCK_FREE 1
-#define __GCC_ATOMIC_CHAR16_T_LOCK_FREE 1
+#define __GCC_ATOMIC_BOOL_LOCK_FREE 2
+#define __GCC_ATOMIC_CHAR16_T_LOCK_FREE 2
-#define __GCC_ATOMIC_CHAR_LOCK_FREE 1
+#define __GCC_ATOMIC_CHAR_LOCK_FREE 2
-#define __GCC_ATOMIC_LLONG_LOCK_FREE 1
+#define __GCC_ATOMIC_LLONG_LOCK_FREE 2
-#define __GCC_ATOMIC_SHORT_LOCK_FREE 1
+#define __GCC_ATOMIC_SHORT_LOCK_FREE 2
+#define __GCC_HAVE_SYNC_COMPARE_AND_SWAP_1 1
+#define __GCC_HAVE_SYNC_COMPARE_AND_SWAP_2 1
+#define __GCC_HAVE_SYNC_COMPARE_AND_SWAP_8 1

(armv6zk is the same as armv6k for atomic-related capabilities.)

boost/atomic/detail/caps_gcc_atomic.hpp apparently does not provide an 
implementation of thread_fence or signal_fence for the armv6 
configuration, only for the armv6k and later ones.

That's a boost problem.

The fact that -mtune=arm1176jzf-s apparently doesn't enable the armv6k 
features even though gcc's source code implies it should is an anomaly.  
(Check this by substituting -mtune=arm1176jzf-s for -march=armv6 and 
verifying that the predefined symbols are the same for both configurations.)

If that anomaly is ever resolved, or if meta-raspberrypi chooses to 
switch to -march=armv6zk, then gcc-configure-common.inc almost certainly 
need to recognize armv6k as an override distinct from armv6: 
mutex-related code built for armv6k via gcc-runtime will result in a 
different ABI from mutex-related code built for armv6 (what gcc will 
produce for builds that do not use OE's tuning parameters).

If the solution to the boost problem is to change meta-raspberrypi to 
use -march=armv6k then gcc-configure-common.inc will need to be updated 
as well.  Probably OE should recognize it as a distinct ARM architecture 
too.

Peter