3.0.0rc2: Borked OPAL_CHECK_GCC_ATOMIC_BUILTINS definitions

[hppritcha]

@PHHargrove reports seeing problems using xlc12.1 to build 3.0.0rc2

With xlc-12.1 I am seeing spurious failures of configure tests due to multiple definitions of OPAL_ASM_SYNC_HAVE_64BIT:

"conftest.c", line 211.9: 1506-236 (W) Macro name OPAL_ASM_SYNC_HAVE_64BIT has been redefined.
"conftest.c", line 211.9: 1506-358 (I) "OPAL_ASM_SYNC_HAVE_64BIT" is defined on line 209 of conftest.c.

Indeed the generated boilerplate in each conftest.c past a certain point contains two *different* definitions of this macro on lines 209 and 211:

 [...]
 #define OPAL_ASM_SYNC_HAVE_64BIT
 #define OPAL_HAVE_GCC_BUILTIN_CSWAP_INT128 0
 #define OPAL_ASM_SYNC_HAVE_64BIT 1
 [...]

reported on devel:
https://www.mail-archive.com/devel@lists.open-mpi.org//msg20323.html

[PHHargrove]

It appears that both OPAL_CHECK_GCC_ATOMIC_BUILTINS and OPAL_CHECK_SYNC_BUILTINS both set OPAL_ASM_SYNC_HAVE_64BIT. That alone may be erroneous.

However, I have tracked the empty (first) definition to the OPAL_CHECK_GCC_ATOMIC_BUILTINS macro which is consuming opal_asm_sync_have_64bit which has yet to be set:

AC_DEFUN([OPAL_CHECK_GCC_ATOMIC_BUILTINS], [
  AC_MSG_CHECKING([for __atomic builtin atomics])

  AC_TRY_LINK([
#include <stdint.h>
uint32_t tmp, old = 0;
uint64_t tmp64, old64 = 0;], [
__atomic_thread_fence(__ATOMIC_SEQ_CST);
__atomic_compare_exchange_n(&tmp, &old, 1, 0, __ATOMIC_RELAXED, __ATOMIC_RELAXED);
__atomic_add_fetch(&tmp, 1, __ATOMIC_RELAXED);
__atomic_compare_exchange_n(&tmp64, &old64, 1, 0, __ATOMIC_RELAXED, __ATOMIC_RELAXED);
__atomic_add_fetch(&tmp64, 1, __ATOMIC_RELAXED);],
    [AC_MSG_RESULT([yes])
     $1],
    [AC_MSG_RESULT([no])
     $2])

  AC_DEFINE_UNQUOTED([OPAL_ASM_SYNC_HAVE_64BIT],[$opal_asm_sync_have_64bit],
                     [Whether 64-bit is supported by the __sync builtin atomics])

  # Check for 128-bit support
  OPAL_CHECK_GCC_BUILTIN_CSWAP_INT128
])

[hjelmn]

Hmm, thats an error. Let me take a look now.

[hjelmn]

copy-and-paste error. fix incoming.

[PHHargrove]

👍 I am not well equipped to autogen, and so have not tested #4000 directly. However, manually removing the expansion of the offending AC_DEFINE_UNQUOTED from the configure in the 3.0.0rc2 tarball lets me progress to the next bug ( #3811 ).