OSHMEM/AMO: added int/uint/32/64 atomics calls

- added int/uint/32/64 atomics calls
- added SHMEM_SYNC_SIZE macro

Signed-off-by: Sergey Oblomov <[email protected]>

Author: Sergey Oblomov

oshmem/include/pshmem.h

@@ -218,8 +218,6 @@ OSHMEM_DECLSPEC void pshmemx_int64_prod_to_all(int64_t *target, const int64_t *s
 
 #define pshmem_int32_wait            pshmemx_int32_wait
 #define pshmem_int64_wait            pshmemx_int64_wait
-#define pshmem_int32_wait_until      pshmemx_int32_wait_until
-#define pshmem_int64_wait_until      pshmemx_int64_wait_until
 
 #define pshmem_int16_and_to_all      pshmemx_int16_and_to_all
 #define pshmem_int32_and_to_all      pshmemx_int32_and_to_all

oshmem/include/pshmemx.h

@@ -204,8 +204,6 @@ OSHMEM_DECLSPEC void shmemx_int64_prod_to_all(int64_t *target, const int64_t *so
 
 #define shmem_int32_wait            shmemx_int32_wait
 #define shmem_int64_wait            shmemx_int64_wait
-#define shmem_int32_wait_until      shmemx_int32_wait_until
-#define shmem_int64_wait_until      shmemx_int64_wait_until
 
 #define shmem_int16_and_to_all      shmemx_int16_and_to_all
 #define shmem_int32_and_to_all      shmemx_int32_and_to_all

oshmem/include/shmem.h.in

@@ -532,13 +532,21 @@
 #define shmem_uint_atomic_fetch_and       pshmem_uint_atomic_fetch_and
 #define shmem_ulong_atomic_fetch_and      pshmem_ulong_atomic_fetch_and
 #define shmem_ulonglong_atomic_fetch_and  pshmem_ulonglong_atomic_fetch_and
+#define shmem_int32_atomic_fetch_and      pshmem_int32_atomic_fetch_and
+#define shmem_int64_atomic_fetch_and      pshmem_int64_atomic_fetch_and
+#define shmem_uint32_atomic_fetch_and     pshmem_uint32_atomic_fetch_and
+#define shmem_uint64_atomic_fetch_and     pshmem_uint64_atomic_fetch_and
 
 #define shmem_ctx_int_atomic_fetch_and    pshmem_ctx_int_atomic_fetch_and
 #define shmem_ctx_long_atomic_fetch_and   pshmem_ctx_long_atomic_fetch_and
 #define shmem_ctx_longlong_atomic_fetch_and pshmem_ctx_longlong_atomic_fetch_and
 #define shmem_ctx_uint_atomic_fetch_and   pshmem_ctx_uint_atomic_fetch_and
 #define shmem_ctx_ulong_atomic_fetch_and  pshmem_ctx_ulong_atomic_fetch_and
 #define shmem_ctx_ulonglong_atomic_fetch_and pshmem_ctx_ulonglong_atomic_fetch_and
+#define shmem_ctx_int32_atomic_fetch_and  pshmem_ctx_int32_atomic_fetch_and
+#define shmem_ctx_int64_atomic_fetch_and  pshmem_ctx_int64_atomic_fetch_and
+#define shmem_ctx_uint32_atomic_fetch_and pshmem_ctx_uint32_atomic_fetch_and
+#define shmem_ctx_uint64_atomic_fetch_and pshmem_ctx_uint64_atomic_fetch_and
 
 #define shmemx_int32_atomic_fetch_and     pshmemx_int32_atomic_fetch_and
 #define shmemx_int64_atomic_fetch_and     pshmemx_int64_atomic_fetch_and
@@ -552,13 +560,21 @@
 #define shmem_uint_atomic_fetch_or        pshmem_uint_atomic_fetch_or
 #define shmem_ulong_atomic_fetch_or       pshmem_ulong_atomic_fetch_or
 #define shmem_ulonglong_atomic_fetch_or   pshmem_ulonglong_atomic_fetch_or
+#define shmem_int32_atomic_fetch_or       pshmem_int32_atomic_fetch_or
+#define shmem_int64_atomic_fetch_or       pshmem_int64_atomic_fetch_or
+#define shmem_uint32_atomic_fetch_or      pshmem_uint32_atomic_fetch_or
+#define shmem_uint64_atomic_fetch_or      pshmem_uint64_atomic_fetch_or
 
 #define shmem_ctx_int_atomic_fetch_or     pshmem_ctx_int_atomic_fetch_or
 #define shmem_ctx_long_atomic_fetch_or    pshmem_ctx_long_atomic_fetch_or
 #define shmem_ctx_longlong_atomic_fetch_or pshmem_ctx_longlong_atomic_fetch_or
 #define shmem_ctx_uint_atomic_fetch_or    pshmem_ctx_uint_atomic_fetch_or
 #define shmem_ctx_ulong_atomic_fetch_or   pshmem_ctx_ulong_atomic_fetch_or
 #define shmem_ctx_ulonglong_atomic_fetch_or pshmem_ctx_ulonglong_atomic_fetch_or
+#define shmem_ctx_int32_atomic_fetch_or   pshmem_ctx_int32_atomic_fetch_or
+#define shmem_ctx_int64_atomic_fetch_or   pshmem_ctx_int64_atomic_fetch_or
+#define shmem_ctx_uint32_atomic_fetch_or  pshmem_ctx_uint32_atomic_fetch_or
+#define shmem_ctx_uint64_atomic_fetch_or  pshmem_ctx_uint64_atomic_fetch_or
 
 #define shmemx_int32_atomic_fetch_or      pshmemx_int32_atomic_fetch_or
 #define shmemx_int64_atomic_fetch_or      pshmemx_int64_atomic_fetch_or
@@ -572,13 +588,21 @@
 #define shmem_uint_atomic_fetch_xor       pshmem_uint_atomic_fetch_xor
 #define shmem_ulong_atomic_fetch_xor      pshmem_ulong_atomic_fetch_xor
 #define shmem_ulonglong_atomic_fetch_xor  pshmem_ulonglong_atomic_fetch_xor
+#define shmem_int32_atomic_fetch_xor      pshmem_int32_atomic_fetch_xor
+#define shmem_int64_atomic_fetch_xor      pshmem_int64_atomic_fetch_xor
+#define shmem_uint32_atomic_fetch_xor     pshmem_uint32_atomic_fetch_xor
+#define shmem_uint64_atomic_fetch_xor     pshmem_uint64_atomic_fetch_xor
 
 #define shmem_ctx_int_atomic_fetch_xor    pshmem_ctx_int_atomic_fetch_xor
 #define shmem_ctx_long_atomic_fetch_xor   pshmem_ctx_long_atomic_fetch_xor
 #define shmem_ctx_longlong_atomic_fetch_xor pshmem_ctx_longlong_atomic_fetch_xor
 #define shmem_ctx_uint_atomic_fetch_xor   pshmem_ctx_uint_atomic_fetch_xor
 #define shmem_ctx_ulong_atomic_fetch_xor  pshmem_ctx_ulong_atomic_fetch_xor
 #define shmem_ctx_ulonglong_atomic_fetch_xor pshmem_ctx_ulonglong_atomic_fetch_xor
+#define shmem_ctx_int32_atomic_fetch_xor  pshmem_ctx_int32_atomic_fetch_xor
+#define shmem_ctx_int64_atomic_fetch_xor  pshmem_ctx_int64_atomic_fetch_xor
+#define shmem_ctx_uint32_atomic_fetch_xor pshmem_ctx_uint32_atomic_fetch_xor
+#define shmem_ctx_uint64_atomic_fetch_xor pshmem_ctx_uint64_atomic_fetch_xor
 
 #define shmemx_int32_atomic_fetch_xor     pshmemx_int32_atomic_fetch_xor
 #define shmemx_int64_atomic_fetch_xor     pshmemx_int64_atomic_fetch_xor
@@ -664,13 +688,21 @@
 #define shmem_uint_atomic_and        pshmem_uint_atomic_and
 #define shmem_ulong_atomic_and       pshmem_ulong_atomic_and
 #define shmem_ulonglong_atomic_and   pshmem_ulonglong_atomic_and
+#define shmem_int32_atomic_and       pshmem_int32_atomic_and
+#define shmem_int64_atomic_and       pshmem_int64_atomic_and
+#define shmem_uint32_atomic_and      pshmem_uint32_atomic_and
+#define shmem_uint64_atomic_and      pshmem_uint64_atomic_and
 
 #define shmem_ctx_int_atomic_and     pshmem_ctx_int_atomic_and
 #define shmem_ctx_long_atomic_and    pshmem_ctx_long_atomic_and
 #define shmem_ctx_longlong_atomic_and pshmem_ctx_longlong_atomic_and
 #define shmem_ctx_uint_atomic_and    pshmem_ctx_uint_atomic_and
 #define shmem_ctx_ulong_atomic_and   pshmem_ctx_ulong_atomic_and
 #define shmem_ctx_ulonglong_atomic_and pshmem_ctx_ulonglong_atomic_and
+#define shmem_ctx_int32_atomic_and   pshmem_ctx_int32_atomic_and
+#define shmem_ctx_int64_atomic_and   pshmem_ctx_int64_atomic_and
+#define shmem_ctx_uint32_atomic_and  pshmem_ctx_uint32_atomic_and
+#define shmem_ctx_uint64_atomic_and  pshmem_ctx_uint64_atomic_and
 
 #define shmemx_int32_atomic_and      pshmemx_int32_atomic_and
 #define shmemx_int64_atomic_and      pshmemx_int64_atomic_and
@@ -685,13 +717,21 @@
 #define shmem_uint_atomic_or         pshmem_uint_atomic_or
 #define shmem_ulong_atomic_or        pshmem_ulong_atomic_or
 #define shmem_ulonglong_atomic_or    pshmem_ulonglong_atomic_or
+#define shmem_int32_atomic_or        pshmem_int32_atomic_or
+#define shmem_int64_atomic_or        pshmem_int64_atomic_or
+#define shmem_uint32_atomic_or       pshmem_uint32_atomic_or
+#define shmem_uint64_atomic_or       pshmem_uint64_atomic_or
 
 #define shmem_ctx_int_atomic_or      pshmem_ctx_int_atomic_or
 #define shmem_ctx_long_atomic_or     pshmem_ctx_long_atomic_or
 #define shmem_ctx_longlong_atomic_or pshmem_ctx_longlong_atomic_or
 #define shmem_ctx_uint_atomic_or     pshmem_ctx_uint_atomic_or
 #define shmem_ctx_ulong_atomic_or    pshmem_ctx_ulong_atomic_or
 #define shmem_ctx_ulonglong_atomic_or pshmem_ctx_ulonglong_atomic_or
+#define shmem_ctx_int32_atomic_or    pshmem_ctx_int32_atomic_or
+#define shmem_ctx_int64_atomic_or    pshmem_ctx_int64_atomic_or
+#define shmem_ctx_uint32_atomic_or   pshmem_ctx_uint32_atomic_or
+#define shmem_ctx_uint64_atomic_or   pshmem_ctx_uint64_atomic_or
 
 #define shmemx_int32_atomic_or       pshmemx_int32_atomic_or
 #define shmemx_int64_atomic_or       pshmemx_int64_atomic_or
@@ -706,13 +746,21 @@
 #define shmem_uint_atomic_xor        pshmem_uint_atomic_xor
 #define shmem_ulong_atomic_xor       pshmem_ulong_atomic_xor
 #define shmem_ulonglong_atomic_xor   pshmem_ulonglong_atomic_xor
+#define shmem_int32_atomic_xor       pshmem_int32_atomic_xor
+#define shmem_int64_atomic_xor       pshmem_int64_atomic_xor
+#define shmem_uint32_atomic_xor      pshmem_uint32_atomic_xor
+#define shmem_uint64_atomic_xor      pshmem_uint64_atomic_xor
 
 #define shmem_ctx_int_atomic_xor     pshmem_ctx_int_atomic_xor
 #define shmem_ctx_long_atomic_xor    pshmem_ctx_long_atomic_xor
 #define shmem_ctx_longlong_atomic_xor pshmem_ctx_longlong_atomic_xor
 #define shmem_ctx_uint_atomic_xor    pshmem_ctx_uint_atomic_xor
 #define shmem_ctx_ulong_atomic_xor   pshmem_ctx_ulong_atomic_xor
 #define shmem_ctx_ulonglong_atomic_xor pshmem_ctx_ulonglong_atomic_xor
+#define shmem_ctx_int32_atomic_xor   pshmem_ctx_int32_atomic_xor
+#define shmem_ctx_int64_atomic_xor   pshmem_ctx_int64_atomic_xor
+#define shmem_ctx_uint32_atomic_xor  pshmem_ctx_uint32_atomic_xor
+#define shmem_ctx_uint64_atomic_xor  pshmem_ctx_uint64_atomic_xor
 
 #define shmemx_int32_atomic_xor      pshmemx_int32_atomic_xor
 #define shmemx_int64_atomic_xor      pshmemx_int64_atomic_xor
@@ -764,13 +812,34 @@
 #define shmem_int_wait_until         pshmem_int_wait_until
 #define shmem_long_wait_until        pshmem_long_wait_until
 #define shmem_longlong_wait_until    pshmem_longlong_wait_until
+#define shmem_ushort_wait_until      pshmem_ushort_wait_until
+#define shmem_uint_wait_until        pshmem_uint_wait_until
+#define shmem_ulong_wait_until       pshmem_ulong_wait_until
+#define shmem_ulonglong_wait_until   pshmem_ulonglong_wait_until
+#define shmem_int32_wait_until       pshmem_int32_wait_until
+#define shmem_int64_wait_until       pshmem_int64_wait_until
+#define shmem_uint32_wait_until      pshmem_uint32_wait_until
+#define shmem_uint64_wait_until      pshmem_uint64_wait_until
+#define shmem_size_wait_until        pshmem_size_wait_until
+#define shmem_ptrdiff_wait_until     pshmem_ptrdiff_wait_until
+
 #define shmemx_int32_wait_until      pshmemx_int32_wait_until
 #define shmemx_int64_wait_until      pshmemx_int64_wait_until
 
 #define shmem_short_test             pshmem_short_test
 #define shmem_int_test               pshmem_int_test
 #define shmem_long_test              pshmem_long_test
 #define shmem_longlong_test          pshmem_longlong_test
+#define shmem_ushort_test            pshmem_ushort_test
+#define shmem_uint_test              pshmem_uint_test
+#define shmem_ulong_test             pshmem_ulong_test
+#define shmem_ulonglong_test         pshmem_ulonglong_test
+#define shmem_int32_test             pshmem_int32_test
+#define shmem_int64_test             pshmem_int64_test
+#define shmem_uint32_test            pshmem_uint32_test
+#define shmem_uint64_test            pshmem_uint64_test
+#define shmem_size_test              pshmem_size_test
+#define shmem_ptrdiff_test           pshmem_ptrdiff_test
 
 /*
  * Barrier sync routines

oshmem/include/shmemx.h

@@ -31,13 +31,21 @@
 #pragma weak shmem_uint_atomic_and      = pshmem_uint_atomic_and
 #pragma weak shmem_ulong_atomic_and     = pshmem_ulong_atomic_and
 #pragma weak shmem_ulonglong_atomic_and = pshmem_ulonglong_atomic_and
+#pragma weak shmem_int32_atomic_and     = pshmem_int32_atomic_and
+#pragma weak shmem_int64_atomic_and     = pshmem_int64_atomic_and
+#pragma weak shmem_uint32_atomic_and    = pshmem_uint32_atomic_and
+#pragma weak shmem_uint64_atomic_and    = pshmem_uint64_atomic_and
 
 #pragma weak shmem_ctx_int_atomic_and   = pshmem_ctx_int_atomic_and
 #pragma weak shmem_ctx_long_atomic_and  = pshmem_ctx_long_atomic_and
 #pragma weak shmem_ctx_longlong_atomic_and = pshmem_ctx_longlong_atomic_and
 #pragma weak shmem_ctx_uint_atomic_and  = pshmem_ctx_uint_atomic_and
 #pragma weak shmem_ctx_ulong_atomic_and = pshmem_ctx_ulong_atomic_and
 #pragma weak shmem_ctx_ulonglong_atomic_and = pshmem_ctx_ulonglong_atomic_and
+#pragma weak shmem_ctx_int32_atomic_and  = pshmem_ctx_int32_atomic_and
+#pragma weak shmem_ctx_int64_atomic_and  = pshmem_ctx_int64_atomic_and
+#pragma weak shmem_ctx_uint32_atomic_and = pshmem_ctx_uint32_atomic_and
+#pragma weak shmem_ctx_uint64_atomic_and = pshmem_ctx_uint64_atomic_and
 
 #pragma weak shmemx_int32_atomic_and    = pshmemx_int32_atomic_and
 #pragma weak shmemx_int64_atomic_and    = pshmemx_int64_atomic_and
@@ -52,12 +60,22 @@ OSHMEM_TYPE_OP(longlong, long long, shmem, and)
 OSHMEM_TYPE_OP(uint, unsigned int, shmem, and)
 OSHMEM_TYPE_OP(ulong, unsigned long, shmem, and)
 OSHMEM_TYPE_OP(ulonglong, unsigned long long, shmem, and)
+OSHMEM_TYPE_OP(int32, int32_t, shmem, and)
+OSHMEM_TYPE_OP(int64, int64_t, shmem, and)
+OSHMEM_TYPE_OP(uint32, uint32_t, shmem, and)
+OSHMEM_TYPE_OP(uint64, uint64_t, shmem, and)
+
 OSHMEM_CTX_TYPE_OP(int, int, shmem, and)
 OSHMEM_CTX_TYPE_OP(long, long, shmem, and)
 OSHMEM_CTX_TYPE_OP(longlong, long long, shmem, and)
 OSHMEM_CTX_TYPE_OP(uint, unsigned int, shmem, and)
 OSHMEM_CTX_TYPE_OP(ulong, unsigned long, shmem, and)
 OSHMEM_CTX_TYPE_OP(ulonglong, unsigned long long, shmem, and)
+OSHMEM_CTX_TYPE_OP(int32, int32_t, shmem, and)
+OSHMEM_CTX_TYPE_OP(int64, int64_t, shmem, and)
+OSHMEM_CTX_TYPE_OP(uint32, uint32_t, shmem, and)
+OSHMEM_CTX_TYPE_OP(uint64, uint64_t, shmem, and)
+
 OSHMEM_TYPE_OP(int32, int32_t, shmemx, and)
 OSHMEM_TYPE_OP(int64, int64_t, shmemx, and)
 OSHMEM_TYPE_OP(uint32, uint32_t, shmemx, and)

oshmem/shmem/c/profile/defines.h

@@ -33,13 +33,21 @@
 #pragma weak shmem_uint_atomic_fetch_and      = pshmem_uint_atomic_fetch_and
 #pragma weak shmem_ulong_atomic_fetch_and     = pshmem_ulong_atomic_fetch_and
 #pragma weak shmem_ulonglong_atomic_fetch_and = pshmem_ulonglong_atomic_fetch_and
+#pragma weak shmem_int32_atomic_fetch_and     = pshmem_int32_atomic_fetch_and
+#pragma weak shmem_int64_atomic_fetch_and     = pshmem_int64_atomic_fetch_and
+#pragma weak shmem_uint32_atomic_fetch_and    = pshmem_uint32_atomic_fetch_and
+#pragma weak shmem_uint64_atomic_fetch_and    = pshmem_uint64_atomic_fetch_and
 
 #pragma weak shmem_ctx_int_atomic_fetch_and   = pshmem_ctx_int_atomic_fetch_and
 #pragma weak shmem_ctx_long_atomic_fetch_and  = pshmem_ctx_long_atomic_fetch_and
 #pragma weak shmem_ctx_longlong_atomic_fetch_and = pshmem_ctx_longlong_atomic_fetch_and
 #pragma weak shmem_ctx_uint_atomic_fetch_and  = pshmem_ctx_uint_atomic_fetch_and
 #pragma weak shmem_ctx_ulong_atomic_fetch_and = pshmem_ctx_ulong_atomic_fetch_and
 #pragma weak shmem_ctx_ulonglong_atomic_fetch_and = pshmem_ctx_ulonglong_atomic_fetch_and
+#pragma weak shmem_ctx_int32_atomic_fetch_and  = pshmem_ctx_int32_atomic_fetch_and
+#pragma weak shmem_ctx_int64_atomic_fetch_and  = pshmem_ctx_int64_atomic_fetch_and
+#pragma weak shmem_ctx_uint32_atomic_fetch_and = pshmem_ctx_uint32_atomic_fetch_and
+#pragma weak shmem_ctx_uint64_atomic_fetch_and = pshmem_ctx_uint64_atomic_fetch_and
 
 #pragma weak shmemx_int32_atomic_fetch_and    = pshmemx_int32_atomic_fetch_and
 #pragma weak shmemx_int64_atomic_fetch_and    = pshmemx_int64_atomic_fetch_and
@@ -48,20 +56,28 @@
 #include "oshmem/shmem/c/profile/defines.h"
 #endif
 
- unsigned int shmem_uint_atomic_fand(unsigned int *target, unsigned int value, int pe);
-
 OSHMEM_TYPE_FOP(int, int, shmem, and)
 OSHMEM_TYPE_FOP(long, long, shmem, and)
 OSHMEM_TYPE_FOP(longlong, long long, shmem, and)
 OSHMEM_TYPE_FOP(uint, unsigned int, shmem, and)
 OSHMEM_TYPE_FOP(ulong, unsigned long, shmem, and)
 OSHMEM_TYPE_FOP(ulonglong, unsigned long long, shmem, and)
+OSHMEM_TYPE_FOP(int32, int32_t, shmem, and)
+OSHMEM_TYPE_FOP(int64, int64_t, shmem, and)
+OSHMEM_TYPE_FOP(uint32, uint32_t, shmem, and)
+OSHMEM_TYPE_FOP(uint64, uint64_t, shmem, and)
+
 OSHMEM_CTX_TYPE_FOP(int, int, shmem, and)
 OSHMEM_CTX_TYPE_FOP(long, long, shmem, and)
 OSHMEM_CTX_TYPE_FOP(longlong, long long, shmem, and)
 OSHMEM_CTX_TYPE_FOP(uint, unsigned int, shmem, and)
 OSHMEM_CTX_TYPE_FOP(ulong, unsigned long, shmem, and)
 OSHMEM_CTX_TYPE_FOP(ulonglong, unsigned long long, shmem, and)
+OSHMEM_CTX_TYPE_FOP(int32, int32_t, shmem, and)
+OSHMEM_CTX_TYPE_FOP(int64, int64_t, shmem, and)
+OSHMEM_CTX_TYPE_FOP(uint32, uint32_t, shmem, and)
+OSHMEM_CTX_TYPE_FOP(uint64, uint64_t, shmem, and)
+
 OSHMEM_TYPE_FOP(int32, int32_t, shmemx, and)
 OSHMEM_TYPE_FOP(int64, int64_t, shmemx, and)
 OSHMEM_TYPE_FOP(uint32, uint32_t, shmemx, and)

oshmem/shmem/c/shmem_and.c

@@ -33,13 +33,21 @@
 #pragma weak shmem_uint_atomic_fetch_or      = pshmem_uint_atomic_fetch_or
 #pragma weak shmem_ulong_atomic_fetch_or     = pshmem_ulong_atomic_fetch_or
 #pragma weak shmem_ulonglong_atomic_fetch_or = pshmem_ulonglong_atomic_fetch_or
+#pragma weak shmem_int32_atomic_fetch_or     = pshmem_int32_atomic_fetch_or
+#pragma weak shmem_int64_atomic_fetch_or     = pshmem_int64_atomic_fetch_or
+#pragma weak shmem_uint32_atomic_fetch_or    = pshmem_uint32_atomic_fetch_or
+#pragma weak shmem_uint64_atomic_fetch_or    = pshmem_uint64_atomic_fetch_or
 
 #pragma weak shmem_ctx_int_atomic_fetch_or   = pshmem_ctx_int_atomic_fetch_or
 #pragma weak shmem_ctx_long_atomic_fetch_or  = pshmem_ctx_long_atomic_fetch_or
 #pragma weak shmem_ctx_longlong_atomic_fetch_or = pshmem_ctx_longlong_atomic_fetch_or
 #pragma weak shmem_ctx_uint_atomic_fetch_or  = pshmem_ctx_uint_atomic_fetch_or
 #pragma weak shmem_ctx_ulong_atomic_fetch_or = pshmem_ctx_ulong_atomic_fetch_or
 #pragma weak shmem_ctx_ulonglong_atomic_fetch_or = pshmem_ctx_ulonglong_atomic_fetch_or
+#pragma weak shmem_ctx_int32_atomic_fetch_or  = pshmem_ctx_int32_atomic_fetch_or
+#pragma weak shmem_ctx_int64_atomic_fetch_or  = pshmem_ctx_int64_atomic_fetch_or
+#pragma weak shmem_ctx_uint32_atomic_fetch_or = pshmem_ctx_uint32_atomic_fetch_or
+#pragma weak shmem_ctx_uint64_atomic_fetch_or = pshmem_ctx_uint64_atomic_fetch_or
 
 #pragma weak shmemx_int32_atomic_fetch_or    = pshmemx_int32_atomic_fetch_or
 #pragma weak shmemx_int64_atomic_fetch_or    = pshmemx_int64_atomic_fetch_or
@@ -54,12 +62,22 @@ OSHMEM_TYPE_FOP(longlong, long long, shmem, or)
 OSHMEM_TYPE_FOP(uint, unsigned int, shmem, or)
 OSHMEM_TYPE_FOP(ulong, unsigned long, shmem, or)
 OSHMEM_TYPE_FOP(ulonglong, unsigned long long, shmem, or)
+OSHMEM_TYPE_FOP(int32, int32_t, shmem, or)
+OSHMEM_TYPE_FOP(int64, int64_t, shmem, or)
+OSHMEM_TYPE_FOP(uint32, uint32_t, shmem, or)
+OSHMEM_TYPE_FOP(uint64, uint64_t, shmem, or)
+
 OSHMEM_CTX_TYPE_FOP(int, int, shmem, or)
 OSHMEM_CTX_TYPE_FOP(long, long, shmem, or)
 OSHMEM_CTX_TYPE_FOP(longlong, long long, shmem, or)
 OSHMEM_CTX_TYPE_FOP(uint, unsigned int, shmem, or)
 OSHMEM_CTX_TYPE_FOP(ulong, unsigned long, shmem, or)
 OSHMEM_CTX_TYPE_FOP(ulonglong, unsigned long long, shmem, or)
+OSHMEM_CTX_TYPE_FOP(int32, int32_t, shmem, or)
+OSHMEM_CTX_TYPE_FOP(int64, int64_t, shmem, or)
+OSHMEM_CTX_TYPE_FOP(uint32, uint32_t, shmem, or)
+OSHMEM_CTX_TYPE_FOP(uint64, uint64_t, shmem, or)
+
 OSHMEM_TYPE_FOP(int32, int32_t, shmemx, or)
 OSHMEM_TYPE_FOP(int64, int64_t, shmemx, or)
 OSHMEM_TYPE_FOP(uint32, uint32_t, shmemx, or)