libunwind: add patch to fix behavior for ARM < v6

Since libatomic_ops does not implement real atomic operations for
ARMv4 and ARMv5, libunwind must define AO_REQUIRE_CAS do indicate it
requires compare-and-swap operations, even if not available as real
atomic operations for the current architecture. In this case,
libatomic_ops will rely on emulated atomic operations, which also
require linking against libatomic_ops, which was until now not done by
libunwind.

This fixes the mysterious ltrace build issue:

  http://autobuild.buildroot.org/results/e1b/e1b330abfa2d80f3f30bc3359428ea429c690eb8/

Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Cc: Baruch Siach <baruch@tkos.co.il>
---
 ...ests.patch => libunwind-01-disable-tests.patch} |  0
 ...Add-AO_REQUIRE_CAS-to-fix-build-on-ARM-v6.patch | 56 ++++++++++++++++++++++
 package/libunwind/libunwind.mk                     |  1 +
 3 files changed, 57 insertions(+)
 rename package/libunwind/{libunwind-disable-tests.patch => libunwind-01-disable-tests.patch} (100%)
 create mode 100644 package/libunwind/libunwind-02-Add-AO_REQUIRE_CAS-to-fix-build-on-ARM-v6.patch

>>>>> "Thomas" == Thomas Petazzoni <thomas.petazzoni@free-electrons.com> writes:

 > Since libatomic_ops does not implement real atomic operations for
 > ARMv4 and ARMv5, libunwind must define AO_REQUIRE_CAS do indicate it
 > requires compare-and-swap operations, even if not available as real
 > atomic operations for the current architecture. In this case,
 > libatomic_ops will rely on emulated atomic operations, which also
 > require linking against libatomic_ops, which was until now not done by
 > libunwind.

 > This fixes the mysterious ltrace build issue:

 >   http://autobuild.buildroot.org/results/e1b/e1b330abfa2d80f3f30bc3359428ea429c690eb8/

 > Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
 > Cc: Baruch Siach <baruch@tkos.co.il>

Committed, thanks.

Hi Thomas,

On Tue, May 13, 2014 at 11:47:37PM +0200, Thomas Petazzoni wrote:
> Since libatomic_ops does not implement real atomic operations for
> ARMv4 and ARMv5, libunwind must define AO_REQUIRE_CAS do indicate it
> requires compare-and-swap operations, even if not available as real
> atomic operations for the current architecture. In this case,
> libatomic_ops will rely on emulated atomic operations, which also
> require linking against libatomic_ops, which was until now not done by
> libunwind.
> 
> This fixes the mysterious ltrace build issue:
> 
>   http://autobuild.buildroot.org/results/e1b/e1b330abfa2d80f3f30bc3359428ea429c690eb8/
> 
> Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
> Cc: Baruch Siach <baruch@tkos.co.il>
> ---
>  ...ests.patch => libunwind-01-disable-tests.patch} |  0
>  ...Add-AO_REQUIRE_CAS-to-fix-build-on-ARM-v6.patch | 56 ++++++++++++++++++++++
>  package/libunwind/libunwind.mk                     |  1 +
>  3 files changed, 57 insertions(+)
>  rename package/libunwind/{libunwind-disable-tests.patch => libunwind-01-disable-tests.patch} (100%)
>  create mode 100644 package/libunwind/libunwind-02-Add-AO_REQUIRE_CAS-to-fix-build-on-ARM-v6.patch
> 
> diff --git a/package/libunwind/libunwind-disable-tests.patch b/package/libunwind/libunwind-01-disable-tests.patch
> similarity index 100%
> rename from package/libunwind/libunwind-disable-tests.patch
> rename to package/libunwind/libunwind-01-disable-tests.patch
> diff --git a/package/libunwind/libunwind-02-Add-AO_REQUIRE_CAS-to-fix-build-on-ARM-v6.patch b/package/libunwind/libunwind-02-Add-AO_REQUIRE_CAS-to-fix-build-on-ARM-v6.patch
> new file mode 100644
> index 0000000..2a37ed7
> --- /dev/null
> +++ b/package/libunwind/libunwind-02-Add-AO_REQUIRE_CAS-to-fix-build-on-ARM-v6.patch
> @@ -0,0 +1,56 @@
> +From 24484e80b3e329c9edee1995e102f8612eedb79c Mon Sep 17 00:00:00 2001
> +From: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
> +Date: Tue, 13 May 2014 23:32:27 +0200
> +Subject: [PATCH] Add AO_REQUIRE_CAS to fix build on ARM < v6
> +
> +ARM earlier than ARMv6, such as ARMv4 and ARMv5 do not provide
> +optimize atomic operations in libatomic_ops. Since libunwind is using
> +such operations, it should define AO_REQUIRE_CAS before including
> +<atomic_ops.h> so that libatomic_ops knows it should use emulated
> +atomic operations instead (even though they are obviously a lot more
> +expensive).
> +
> +Also, while real atomic operations are all inline functions and
> +therefore linking against libatomic_ops was not required, the emulated
> +atomic operations actually require linking against libatomic_ops, so
> +the commented AC_CHECK_LIB test in acinclude.m4 is uncommented to make
> +sure we link against libatomic_ops.
> +
> +Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
> +---
> + acinclude.m4          | 8 +-------
> + include/libunwind_i.h | 1 +
> + 2 files changed, 2 insertions(+), 7 deletions(-)
> +
> +diff --git a/acinclude.m4 b/acinclude.m4
> +index 497f7c2..9c15af1 100644
> +--- a/acinclude.m4
> ++++ b/acinclude.m4
> +@@ -22,11 +22,5 @@ fi])
> + AC_DEFUN([CHECK_ATOMIC_OPS],
> + [dnl Check whether the system has the atomic_ops package installed.
> +   AC_CHECK_HEADERS(atomic_ops.h)
> +-#
> +-# Don't link against libatomic_ops for now.  We don't want libunwind
> +-# to depend on libatomic_ops.so.  Fortunately, none of the platforms
> +-# we care about so far need libatomic_ops.a (everything is done via
> +-# inline macros).
> +-#
> +-#  AC_CHECK_LIB(atomic_ops, main)
> ++  AC_CHECK_LIB(atomic_ops, main)

For the sake of reproducibility shouldn't we add libatomic_ops to libunwind 
dependencies when libatomic_ops is enabled?

baruch

> + ])
> +diff --git a/include/libunwind_i.h b/include/libunwind_i.h
> +index 23f615e..deabdfd 100644
> +--- a/include/libunwind_i.h
> ++++ b/include/libunwind_i.h
> +@@ -95,6 +95,7 @@ WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.  */
> + 	(pthread_mutex_unlock != NULL ? pthread_mutex_unlock (l) : 0)
> + 
> + #ifdef HAVE_ATOMIC_OPS_H
> ++# define AO_REQUIRE_CAS
> + # include <atomic_ops.h>
> + static inline int
> + cmpxchg_ptr (void *addr, void *old, void *new)
> +-- 
> +1.9.2
> +
> diff --git a/package/libunwind/libunwind.mk b/package/libunwind/libunwind.mk
> index f573722..48dd0ec 100644
> --- a/package/libunwind/libunwind.mk
> +++ b/package/libunwind/libunwind.mk
> @@ -9,5 +9,6 @@ LIBUNWIND_SITE = http://download.savannah.gnu.org/releases/libunwind
>  LIBUNWIND_INSTALL_STAGING = YES
>  LIBUNWIND_LICENSE_FILES = COPYING
>  LIBUNWIND_LICENSE = MIT
> +LIBUNWIND_AUTORECONF = YES
>  
>  $(eval $(autotools-package))
> --

Dear Baruch Siach,

On Wed, 14 May 2014 23:47:12 +0300, Baruch Siach wrote:

> > +-# Don't link against libatomic_ops for now.  We don't want libunwind
> > +-# to depend on libatomic_ops.so.  Fortunately, none of the platforms
> > +-# we care about so far need libatomic_ops.a (everything is done via
> > +-# inline macros).
> > +-#
> > +-#  AC_CHECK_LIB(atomic_ops, main)
> > ++  AC_CHECK_LIB(atomic_ops, main)
> 
> For the sake of reproducibility shouldn't we add libatomic_ops to libunwind 
> dependencies when libatomic_ops is enabled?

Yes, you're right! I'll send a patch for that.

Thomas