xl: Use C11 builtin atomics if available.

If unavailable, fallback to gcc builtins.
Some configury work was needed to force include
stdatomic.h. Currently xl doesn't have it in its search
path.

Tested with xl V16.1.1.

Signed-off-by: Austen Lauria <[email protected]>

Author: awlauria

config/opal_setup_cc.m4

@@ -47,10 +47,46 @@ AC_DEFUN([OPAL_PROG_CC_C11_HELPER],[
 
     opal_prog_cc_c11_helper_CFLAGS_save=$CFLAGS
     CFLAGS="$CFLAGS $1"
+    OPAL_C_COMPILER_VENDOR([opal_c_vendor])
 
     OPAL_CC_HELPER([if $CC $1 supports C11 _Thread_local], [opal_prog_cc_c11_helper__Thread_local_available],
                    [],[[static _Thread_local int  foo = 1;++foo;]])
 
+
+    OPAL_CC_HELPER([if $CC $1 has stdatomic.h], [opal_prog_cc_c11_helper_atomic_has_stdatomic_h],
+                   [[#include <stdatomic.h>]], [])
+    if test $opal_prog_cc_c11_helper_atomic_has_stdatomic_h -eq 0; then
+        if test "$opal_cv_c_compiler_vendor" = "ibm"; then
+            AC_REQUIRE([AC_PROG_GREP])
+            tmp=`which gcc`
+            if test $? -eq 0; then
+                stdatomic_include_path=""
+                AC_MSG_CHECKING([if gcc has stdatomic.h])
+                AC_LANG_CONFTEST([AC_LANG_SOURCE([
+                                                  #include<stdatomic.h>
+                                                  int main() { }
+                                                 ]
+                                                )])
+                stdatomic_include_path=`gcc -M conftest.c | $GREP stdatomic.h`
+                if test -z "$stdatomic_include_path"; then
+                    AC_MSG_WARN([stdatomic.h cannot be found. Fallback to C99 atomics.])
+                else
+                    stdatomic_include_path="${stdatomic_include_path#"${stdatomic_include_path%%[![:space:]]*}"}"
+                    stdatomic_include_path=${stdatomic_include_path%stdatomic.h}
+                    CFLAGS="$CFLAGS -I$stdatomic_include_path"
+                    opal_prog_cc_c11_helper_CFLAGS_save="$opal_prog_cc_c11_helper_CFLAGS -I$stdatomic_include_path"
+                    opal_prog_cc_c11_helper_atomic_has_stdatomic_h=1
+                    AC_MSG_RESULT([stdatomic.h is available])
+                    AC_MSG_RESULT([Adding $stdatomic_include_path to include path. Using C11 atomics.])
+                fi
+            else
+                AC_MSG_WARN([No gcc found. Not checking for stdatomic.h include. Fallback to C99 atomics.])
+            fi
+        else
+            AC_MSG_WARN([Not checking for gcc stdatomic.h include. Fallback to C99 atomics.])
+        fi
+    fi
+
     OPAL_CC_HELPER([if $CC $1 supports C11 atomic variables], [opal_prog_cc_c11_helper_atomic_var_available],
                    [[#include <stdatomic.h>]], [[static atomic_long foo = 1;++foo;]])
 
@@ -65,9 +101,9 @@ AC_DEFUN([OPAL_PROG_CC_C11_HELPER],[
 
     OPAL_CC_HELPER([if $CC $1 supports C11 atomic_fetch_xor_explicit], [opal_prog_cc_c11_helper_atomic_fetch_xor_explicit_available],
                    [[#include <stdatomic.h>
-#include <stdint.h>]],[[_Atomic uint32_t a; uint32_t b; atomic_fetch_xor_explicit(&a, b, memory_order_relaxed);]])
+                     #include <stdint.h>]],[[_Atomic uint32_t a; uint32_t b; atomic_fetch_xor_explicit(&a, b, memory_order_relaxed);]])
 
-    AS_IF([test $opal_prog_cc_c11_helper__Thread_local_available -eq 1 && test $opal_prog_cc_c11_helper_atomic_var_available -eq 1 && test $opal_prog_cc_c11_helper_atomic_fetch_xor_explicit_available -eq 1],
+    AS_IF([test $opal_prog_cc_c11_helper__Thread_local_available -eq 1 && test $opal_prog_cc_c11_helper_atomic_var_available -eq 1],
           [$2],
           [$3])
 
@@ -166,7 +202,7 @@ AC_DEFUN([OPAL_SETUP_CC],[
     if test $opal_cv_c11_supported = no ; then
         # It is not currently an error if C11 support is not available. Uncomment the
         # following lines and update the warning when we require a C11 compiler.
-        # AC_MSG_WARNING([Open MPI requires a C11 (or newer) compiler])
+        # AC_MSG_WARN([Open MPI requires a C11 (or newer) compiler])
         # AC_MSG_ERROR([Aborting.])
         # From Open MPI 1.7 on we require a C99 compiant compiler
         AC_PROG_CC_C99
@@ -287,16 +323,20 @@ AC_DEFUN([OPAL_SETUP_CC],[
     OPAL_CFLAGS_BEFORE_PICKY="$CFLAGS"
 
     if test $WANT_PICKY_COMPILER -eq 1; then
-        _OPAL_CHECK_SPECIFIC_CFLAGS(-Wundef, Wundef)
+        if test "$opal_cv_c_compiler_vendor" != "portland group"; then
+            _OPAL_CHECK_SPECIFIC_CFLAGS(-Wundef, Wundef)
+            _OPAL_CHECK_SPECIFIC_CFLAGS(-Wmissing-prototypes, Wmissing_prototypes)
+            _OPAL_CHECK_SPECIFIC_CFLAGS(-Wstrict-prototypes, Wstrict_prototypes)
+        fi
         _OPAL_CHECK_SPECIFIC_CFLAGS(-Wno-long-long, Wno_long_long, int main() { long long x; })
         _OPAL_CHECK_SPECIFIC_CFLAGS(-Wsign-compare, Wsign_compare)
-        _OPAL_CHECK_SPECIFIC_CFLAGS(-Wmissing-prototypes, Wmissing_prototypes)
-        _OPAL_CHECK_SPECIFIC_CFLAGS(-Wstrict-prototypes, Wstrict_prototypes)
         _OPAL_CHECK_SPECIFIC_CFLAGS(-Wcomment, Wcomment)
         _OPAL_CHECK_SPECIFIC_CFLAGS(-Werror-implicit-function-declaration, Werror_implicit_function_declaration)
         _OPAL_CHECK_SPECIFIC_CFLAGS(-Wno-long-double, Wno_long_double, int main() { long double x; })
         _OPAL_CHECK_SPECIFIC_CFLAGS(-fno-strict-aliasing, fno_strict_aliasing, int main() { long double x; })
-        _OPAL_CHECK_SPECIFIC_CFLAGS(-pedantic, pedantic)
+        if test "$opal_cv_c_compiler_vendor" != "ibm" && test "$opal_cv_c_compiler_vendor" != "portland group"; then
+            _OPAL_CHECK_SPECIFIC_CFLAGS(-pedantic, pedantic)
+        fi
         _OPAL_CHECK_SPECIFIC_CFLAGS(-Wall, Wall)
     fi
 

ompi/mca/osc/rdma/osc_rdma_active_target.c

@@ -240,6 +240,20 @@ static void ompi_osc_rdma_check_posts (ompi_osc_rdma_module_t *module)
     }
 }
 
+#if defined(__ibmxl__)
+// Work around an xl optimization bug that would cause the compiler to segv.
+//
+// xl doesn't like something about caching and checking the return value of
+// ompi_osc_rdma_lock_compare_exchange() below when the opt-level is high.
+// For now work around this bug by lowering the optimization on this function
+// with xl compilers.
+//
+// Found on:
+// $. xlc --version
+//   IBM XL C/C++ for Linux, V16.1.1 (5725-C73, 5765-J13)
+//   Version: 16.01.0001.0008
+#pragma option_override(ompi_osc_rdma_post_peer, "opt(level,0)")
+#endif
 static int ompi_osc_rdma_post_peer (ompi_osc_rdma_module_t *module, ompi_osc_rdma_peer_t *peer)
 {
     uint64_t target = (uint64_t) (intptr_t) peer->state + offsetof (ompi_osc_rdma_state_t, post_index);

opal/include/opal/opal_portable_platform.h

@@ -180,7 +180,7 @@
 #    else
 #        define PLATFORM_COMPILER_XLC_C  1
 #    endif
-#    define PLATFORM_COMPILER_VERSION __xlC__
+#    define PLATFORM_COMPILER_VERSION __ibmxl_version__
 #    define PLATFORM_COMPILER_VERSION_INT(maj,min,pat) \
          ( ((maj) << 8) | ((min) << 4) | (pat) )
 

opal/include/opal/sys/atomic.h

@@ -97,6 +97,7 @@ BEGIN_C_DECLS
  * to use an int or unsigned char as the lock value - the user is not
  * informed either way.
  */
+#ifndef OPAL_USE_ATOMIC_FLAG_OPAL_LOCK
 struct opal_atomic_lock_t {
     union {
         opal_atomic_int32_t lock;     /**< The lock address (an integer) */
@@ -105,7 +106,7 @@ struct opal_atomic_lock_t {
     } u;
 };
 typedef struct opal_atomic_lock_t opal_atomic_lock_t;
-
+#endif
 /**********************************************************************
  *
  * Set or unset these macros in the architecture-specific atomic.h
@@ -164,7 +165,7 @@ enum {
  *********************************************************************/
 #if defined(DOXYGEN)
 /* don't include system-level gorp when generating doxygen files */
-#elif OPAL_ASSEMBLY_BUILTIN == OPAL_BUILTIN_GCC
+#elif OPAL_ASSEMBLY_BUILTIN == OPAL_BUILTIN_GCC || OPAL_ASSEMBLY_ARCH == OPAL_POWERPC32 || OPAL_ASSEMBLY_ARCH == OPAL_POWERPC64
 #include "opal/sys/gcc_builtin/atomic.h"
 #elif OPAL_ASSEMBLY_ARCH == OPAL_X86_64
 #include "opal/sys/x86_64/atomic.h"
@@ -174,10 +175,6 @@ enum {
 #include "opal/sys/arm64/atomic.h"
 #elif OPAL_ASSEMBLY_ARCH == OPAL_IA32
 #include "opal/sys/ia32/atomic.h"
-#elif OPAL_ASSEMBLY_ARCH == OPAL_POWERPC32
-#include "opal/sys/powerpc/atomic.h"
-#elif OPAL_ASSEMBLY_ARCH == OPAL_POWERPC64
-#include "opal/sys/powerpc/atomic.h"
 #endif
 
 #ifndef DOXYGEN

opal/include/opal/sys/atomic_stdc.h

@@ -97,10 +97,28 @@ static inline void opal_atomic_rmb (void)
 #define opal_atomic_compare_exchange_strong_acq(addr, oldval, newval)  atomic_compare_exchange_strong_explicit (addr, oldval, newval, memory_order_acquire, memory_order_relaxed)
 #define opal_atomic_compare_exchange_strong_rel(addr, oldval, newval)  atomic_compare_exchange_strong_explicit (addr, oldval, newval, memory_order_release, memory_order_relaxed)
 
+#if defined(__PGI) || defined(__ibmxl__)
+#define opal_atomic_swap_32(addr, value) atomic_exchange_explicit (addr, value, memory_order_relaxed)
+#define opal_atomic_swap_64(addr, value) atomic_exchange_explicit (addr, value, memory_order_relaxed)
+#define opal_atomic_swap_ptr(addr, value) atomic_exchange_explicit (addr, value, memory_order_relaxed)
+#else
 #define opal_atomic_swap_32(addr, value) atomic_exchange_explicit ((_Atomic unsigned int *)addr, value, memory_order_relaxed)
 #define opal_atomic_swap_64(addr, value) atomic_exchange_explicit ((_Atomic unsigned long *)addr, value, memory_order_relaxed)
 #define opal_atomic_swap_ptr(addr, value) atomic_exchange_explicit ((_Atomic unsigned long *)addr, value, memory_order_relaxed)
+#endif
 
+#if defined(__PGI) || defined(__ibmxl__)
+#define OPAL_ATOMIC_STDC_DEFINE_FETCH_OP(op, bits, type, operator)      \
+    static inline type opal_atomic_fetch_ ## op ##_## bits (opal_atomic_ ## type *addr, type value) \
+    {                                                                   \
+        return atomic_fetch_ ## op ## _explicit ((type *)addr, value, memory_order_relaxed); \
+    }                                                                   \
+                                                                        \
+    static inline type opal_atomic_## op ## _fetch_ ## bits (opal_atomic_ ## type *addr, type value) \
+    {                                                                   \
+        return atomic_fetch_ ## op ## _explicit ((type *)addr, value, memory_order_relaxed) operator value; \
+    }
+#else
 #define OPAL_ATOMIC_STDC_DEFINE_FETCH_OP(op, bits, type, operator)      \
     static inline type opal_atomic_fetch_ ## op ##_## bits (opal_atomic_ ## type *addr, type value) \
     {                                                                   \
@@ -111,6 +129,7 @@ static inline void opal_atomic_rmb (void)
     {                                                                   \
         return atomic_fetch_ ## op ## _explicit (addr, value, memory_order_relaxed) operator value; \
     }
+#endif
 
 OPAL_ATOMIC_STDC_DEFINE_FETCH_OP(add, 32, int32_t, +)
 OPAL_ATOMIC_STDC_DEFINE_FETCH_OP(add, 64, int64_t, +)
@@ -206,22 +225,39 @@ static inline int64_t opal_atomic_max_fetch_64 (opal_atomic_int64_t *addr, int64
 #define OPAL_ATOMIC_LOCK_UNLOCKED false
 #define OPAL_ATOMIC_LOCK_LOCKED true
 
-#define OPAL_ATOMIC_LOCK_INIT ATOMIC_FLAG_INIT
+#define OPAL_USE_C11_ATOMIC_LOCK 1
 
+#define OPAL_USE_ATOMIC_FLAG_OPAL_LOCK 1
+#if defined(__ibmxl__) || defined(__PGI)
+#define OPAL_ATOMIC_LOCK_INIT 0
+typedef _Atomic bool opal_atomic_lock_t;
+#else
+#define OPAL_ATOMIC_LOCK_INIT ATOMIC_FLAG_INIT
 typedef atomic_flag opal_atomic_lock_t;
+#endif
 
 /*
  * Lock initialization function. It set the lock to UNLOCKED.
  */
 static inline void opal_atomic_lock_init (opal_atomic_lock_t *lock, bool value)
 {
+#if defined(__PGI)
+    atomic_flag_clear ((volatile void *) lock);
+#else
     atomic_flag_clear (lock);
+#endif
 }
 
 
 static inline int opal_atomic_trylock (opal_atomic_lock_t *lock)
 {
+
+#if defined(__PGI)
+    return (int) atomic_flag_test_and_set ((volatile void *) lock);
+#else
     return (int) atomic_flag_test_and_set (lock);
+#endif
+
 }
 
 
@@ -234,7 +270,11 @@ static inline void opal_atomic_lock(opal_atomic_lock_t *lock)
 
 static inline void opal_atomic_unlock (opal_atomic_lock_t *lock)
 {
+#if defined(__PGI)
+    atomic_flag_clear ((volatile void *) lock);
+#else
     atomic_flag_clear (lock);
+#endif
 }
 
 

opal/include/opal/sys/powerpc/Makefile.am

@@ -19,5 +19,4 @@
 # This makefile.am does not stand on its own - it is included from opal/include/Makefile.am
 
 headers += \
-	opal/sys/powerpc/atomic.h \
 	opal/sys/powerpc/timer.h