diff --git a/config/oshmem_config_files.m4 b/config/oshmem_config_files.m4
index a2c6bb35c9..74d80cd163 100644
--- a/config/oshmem_config_files.m4
+++ b/config/oshmem_config_files.m4
@@ -15,10 +15,11 @@ AC_DEFUN([OSHMEM_CONFIG_FILES],[
     oshmem/Makefile
     oshmem/include/Makefile
     oshmem/shmem/c/Makefile
-    oshmem/shmem/fortran/Makefile
-
     oshmem/shmem/c/profile/Makefile
 
+    oshmem/shmem/fortran/Makefile
+    oshmem/shmem/fortran/profile/Makefile
+
     oshmem/tools/oshmem_info/Makefile
     oshmem/tools/wrappers/Makefile
     oshmem/tools/wrappers/shmemcc-wrapper-data.txt
diff --git a/oshmem/Makefile.am b/oshmem/Makefile.am
index f48b50d583..96a3767e0f 100644
--- a/oshmem/Makefile.am
+++ b/oshmem/Makefile.am
@@ -20,8 +20,14 @@ endif
 # Do we have the Fortran bindings?
 if OSHMEM_BUILD_FORTRAN_BINDINGS
 fortran_oshmem_lib = shmem/fortran/liboshmem_fortran.la
+
+if OSHMEM_PROFILING
+fortran_pshmem_lib = shmem/fortran/profile/liboshmem_fortran_pshmem.la
+endif
+
 else
 fortran_oshmem_lib =
+fortran_pshmem_lib =
 endif
 
 SUBDIRS = \
@@ -56,6 +62,7 @@ liboshmem_la_LIBADD = \
 	shmem/c/liboshmem_c.la \
 	$(c_pshmem_lib) \
 	$(fortran_oshmem_lib) \
+	$(fortran_pshmem_lib) \
 	$(MCA_oshmem_FRAMEWORK_LIBS) \
 	$(top_ompi_builddir)/ompi/libmpi.la
 liboshmem_la_DEPENDENCIES = $(liboshmem_la_LIBADD)
diff --git a/oshmem/op/op.c b/oshmem/op/op.c
index c84f5a082e..f5350e4cd2 100644
--- a/oshmem/op/op.c
+++ b/oshmem/op/op.c
@@ -47,6 +47,7 @@ oshmem_op_t* oshmem_op_and_short = NULL;
 oshmem_op_t* oshmem_op_and_int = NULL;
 oshmem_op_t* oshmem_op_and_long = NULL;
 oshmem_op_t* oshmem_op_and_longlong = NULL;
+oshmem_op_t* oshmem_op_and_fint2 = NULL;
 oshmem_op_t* oshmem_op_and_fint4 = NULL;
 oshmem_op_t* oshmem_op_and_fint8 = NULL;
 
@@ -55,6 +56,7 @@ oshmem_op_t* oshmem_op_or_short = NULL;
 oshmem_op_t* oshmem_op_or_int = NULL;
 oshmem_op_t* oshmem_op_or_long = NULL;
 oshmem_op_t* oshmem_op_or_longlong = NULL;
+oshmem_op_t* oshmem_op_or_fint2 = NULL;
 oshmem_op_t* oshmem_op_or_fint4 = NULL;
 oshmem_op_t* oshmem_op_or_fint8 = NULL;
 
@@ -63,6 +65,7 @@ oshmem_op_t* oshmem_op_xor_short = NULL;
 oshmem_op_t* oshmem_op_xor_int = NULL;
 oshmem_op_t* oshmem_op_xor_long = NULL;
 oshmem_op_t* oshmem_op_xor_longlong = NULL;
+oshmem_op_t* oshmem_op_xor_fint2 = NULL;
 oshmem_op_t* oshmem_op_xor_fint4 = NULL;
 oshmem_op_t* oshmem_op_xor_fint8 = NULL;
 
@@ -74,6 +77,7 @@ oshmem_op_t* oshmem_op_max_longlong = NULL;
 oshmem_op_t* oshmem_op_max_float = NULL;
 oshmem_op_t* oshmem_op_max_double = NULL;
 oshmem_op_t* oshmem_op_max_longdouble = NULL;
+oshmem_op_t* oshmem_op_max_fint2 = NULL;
 oshmem_op_t* oshmem_op_max_fint4 = NULL;
 oshmem_op_t* oshmem_op_max_fint8 = NULL;
 oshmem_op_t* oshmem_op_max_freal4 = NULL;
@@ -88,6 +92,7 @@ oshmem_op_t* oshmem_op_min_longlong = NULL;
 oshmem_op_t* oshmem_op_min_float = NULL;
 oshmem_op_t* oshmem_op_min_double = NULL;
 oshmem_op_t* oshmem_op_min_longdouble = NULL;
+oshmem_op_t* oshmem_op_min_fint2 = NULL;
 oshmem_op_t* oshmem_op_min_fint4 = NULL;
 oshmem_op_t* oshmem_op_min_fint8 = NULL;
 oshmem_op_t* oshmem_op_min_freal4 = NULL;
@@ -104,6 +109,7 @@ oshmem_op_t* oshmem_op_sum_double = NULL;
 oshmem_op_t* oshmem_op_sum_longdouble = NULL;
 oshmem_op_t* oshmem_op_sum_complexf = NULL;
 oshmem_op_t* oshmem_op_sum_complexd = NULL;
+oshmem_op_t* oshmem_op_sum_fint2 = NULL;
 oshmem_op_t* oshmem_op_sum_fint4 = NULL;
 oshmem_op_t* oshmem_op_sum_fint8 = NULL;
 oshmem_op_t* oshmem_op_sum_freal4 = NULL;
@@ -120,6 +126,7 @@ oshmem_op_t* oshmem_op_prod_double = NULL;
 oshmem_op_t* oshmem_op_prod_longdouble = NULL;
 oshmem_op_t* oshmem_op_prod_complexf = NULL;
 oshmem_op_t* oshmem_op_prod_complexd = NULL;
+oshmem_op_t* oshmem_op_prod_fint2 = NULL;
 oshmem_op_t* oshmem_op_prod_fint4 = NULL;
 oshmem_op_t* oshmem_op_prod_fint8 = NULL;
 oshmem_op_t* oshmem_op_prod_freal4 = NULL;
@@ -156,6 +163,7 @@ FUNC_OP_CREATE(and, short, short, __and_op)
 FUNC_OP_CREATE(and, int, int, __and_op)
 FUNC_OP_CREATE(and, long, long, __and_op)
 FUNC_OP_CREATE(and, longlong, long long, __and_op)
+FUNC_OP_CREATE(and, fint2, ompi_fortran_integer4_t, __and_op)
 FUNC_OP_CREATE(and, fint4, ompi_fortran_integer4_t, __and_op)
 FUNC_OP_CREATE(and, fint8, ompi_fortran_integer8_t, __and_op)
 
@@ -165,6 +173,7 @@ FUNC_OP_CREATE(or, short, short, __or_op)
 FUNC_OP_CREATE(or, int, int, __or_op)
 FUNC_OP_CREATE(or, long, long, __or_op)
 FUNC_OP_CREATE(or, longlong, long long, __or_op)
+FUNC_OP_CREATE(or, fint2, ompi_fortran_integer2_t, __or_op)
 FUNC_OP_CREATE(or, fint4, ompi_fortran_integer4_t, __or_op)
 FUNC_OP_CREATE(or, fint8, ompi_fortran_integer8_t, __or_op)
 
@@ -174,6 +183,7 @@ FUNC_OP_CREATE(xor, short, short, __xor_op)
 FUNC_OP_CREATE(xor, int, int, __xor_op)
 FUNC_OP_CREATE(xor, long, long, __xor_op)
 FUNC_OP_CREATE(xor, longlong, long long, __xor_op)
+FUNC_OP_CREATE(xor, fint2, ompi_fortran_integer4_t, __xor_op)
 FUNC_OP_CREATE(xor, fint4, ompi_fortran_integer4_t, __xor_op)
 FUNC_OP_CREATE(xor, fint8, ompi_fortran_integer8_t, __xor_op)
 
@@ -186,6 +196,7 @@ FUNC_OP_CREATE(max, longlong, long long, __max_op)
 FUNC_OP_CREATE(max, float, float, __max_op)
 FUNC_OP_CREATE(max, double, double, __max_op)
 FUNC_OP_CREATE(max, longdouble, long double, __max_op)
+FUNC_OP_CREATE(max, fint2, ompi_fortran_integer4_t, __max_op)
 FUNC_OP_CREATE(max, fint4, ompi_fortran_integer4_t, __max_op)
 FUNC_OP_CREATE(max, fint8, ompi_fortran_integer8_t, __max_op)
 FUNC_OP_CREATE(max, freal4, ompi_fortran_real4_t, __max_op)
@@ -203,6 +214,7 @@ FUNC_OP_CREATE(min, longlong, long long, __min_op)
 FUNC_OP_CREATE(min, float, float, __min_op)
 FUNC_OP_CREATE(min, double, double, __min_op)
 FUNC_OP_CREATE(min, longdouble, long double, __min_op)
+FUNC_OP_CREATE(min, fint2, ompi_fortran_integer4_t, __min_op)
 FUNC_OP_CREATE(min, fint4, ompi_fortran_integer4_t, __min_op)
 FUNC_OP_CREATE(min, fint8, ompi_fortran_integer8_t, __min_op)
 FUNC_OP_CREATE(min, freal4, ompi_fortran_real4_t, __min_op)
@@ -222,6 +234,7 @@ FUNC_OP_CREATE(sum, double, double, __sum_op)
 FUNC_OP_CREATE(sum, longdouble, long double, __sum_op)
 FUNC_OP_CREATE(sum, complexf, float complex, __sum_op)
 FUNC_OP_CREATE(sum, complexd, double complex, __sum_op)
+FUNC_OP_CREATE(sum, fint2, ompi_fortran_integer4_t, __sum_op)
 FUNC_OP_CREATE(sum, fint4, ompi_fortran_integer4_t, __sum_op)
 FUNC_OP_CREATE(sum, fint8, ompi_fortran_integer8_t, __sum_op)
 FUNC_OP_CREATE(sum, freal4, ompi_fortran_real4_t, __sum_op)
@@ -241,6 +254,7 @@ FUNC_OP_CREATE(prod, double, double, __prod_op)
 FUNC_OP_CREATE(prod, longdouble, long double, __prod_op)
 FUNC_OP_CREATE(prod, complexf, float complex, __prod_op)
 FUNC_OP_CREATE(prod, complexd, double complex, __prod_op)
+FUNC_OP_CREATE(prod, fint2, ompi_fortran_integer2_t, __prod_op)
 FUNC_OP_CREATE(prod, fint4, ompi_fortran_integer4_t, __prod_op)
 FUNC_OP_CREATE(prod, fint8, ompi_fortran_integer8_t, __prod_op)
 FUNC_OP_CREATE(prod, freal4, ompi_fortran_real4_t, __prod_op)
@@ -267,6 +281,7 @@ int oshmem_op_init(void)
     OBJ_OP_CREATE(and, int, int, OSHMEM_OP_AND, OSHMEM_OP_TYPE_INT);
     OBJ_OP_CREATE(and, long, long, OSHMEM_OP_AND, OSHMEM_OP_TYPE_LONG);
     OBJ_OP_CREATE(and, longlong, long long, OSHMEM_OP_AND, OSHMEM_OP_TYPE_LLONG);
+    OBJ_OP_CREATE(and, fint2, ompi_fortran_integer2_t, OSHMEM_OP_AND, OSHMEM_OP_TYPE_FINT2);
     OBJ_OP_CREATE(and, fint4, ompi_fortran_integer4_t, OSHMEM_OP_AND, OSHMEM_OP_TYPE_FINT4);
     OBJ_OP_CREATE(and, fint8, ompi_fortran_integer8_t, OSHMEM_OP_AND, OSHMEM_OP_TYPE_FINT8);
 
@@ -275,6 +290,7 @@ int oshmem_op_init(void)
     OBJ_OP_CREATE(or, int, int, OSHMEM_OP_OR, OSHMEM_OP_TYPE_INT);
     OBJ_OP_CREATE(or, long, long, OSHMEM_OP_OR, OSHMEM_OP_TYPE_LONG);
     OBJ_OP_CREATE(or, longlong, long long, OSHMEM_OP_OR, OSHMEM_OP_TYPE_LLONG);
+    OBJ_OP_CREATE(or, fint2, ompi_fortran_integer2_t, OSHMEM_OP_OR, OSHMEM_OP_TYPE_FINT2);
     OBJ_OP_CREATE(or, fint4, ompi_fortran_integer4_t, OSHMEM_OP_OR, OSHMEM_OP_TYPE_FINT4);
     OBJ_OP_CREATE(or, fint8, ompi_fortran_integer8_t, OSHMEM_OP_OR, OSHMEM_OP_TYPE_FINT8);
 
@@ -283,6 +299,7 @@ int oshmem_op_init(void)
     OBJ_OP_CREATE(xor, int, int, OSHMEM_OP_XOR, OSHMEM_OP_TYPE_INT);
     OBJ_OP_CREATE(xor, long, long, OSHMEM_OP_XOR, OSHMEM_OP_TYPE_LONG);
     OBJ_OP_CREATE(xor, longlong, long long, OSHMEM_OP_XOR, OSHMEM_OP_TYPE_LLONG);
+    OBJ_OP_CREATE(xor, fint2, ompi_fortran_integer2_t, OSHMEM_OP_XOR, OSHMEM_OP_TYPE_FINT2);
     OBJ_OP_CREATE(xor, fint4, ompi_fortran_integer4_t, OSHMEM_OP_XOR, OSHMEM_OP_TYPE_FINT4);
     OBJ_OP_CREATE(xor, fint8, ompi_fortran_integer8_t, OSHMEM_OP_XOR, OSHMEM_OP_TYPE_FINT8);
 
@@ -294,6 +311,7 @@ int oshmem_op_init(void)
     OBJ_OP_CREATE(max, float, float, OSHMEM_OP_MAX, OSHMEM_OP_TYPE_FLOAT);
     OBJ_OP_CREATE(max, double, double, OSHMEM_OP_MAX, OSHMEM_OP_TYPE_DOUBLE);
     OBJ_OP_CREATE(max, longdouble, long double, OSHMEM_OP_MAX, OSHMEM_OP_TYPE_LDOUBLE);
+    OBJ_OP_CREATE(max, fint2, ompi_fortran_integer2_t, OSHMEM_OP_MAX, OSHMEM_OP_TYPE_FINT2);
     OBJ_OP_CREATE(max, fint4, ompi_fortran_integer4_t, OSHMEM_OP_MAX, OSHMEM_OP_TYPE_FINT4);
     OBJ_OP_CREATE(max, fint8, ompi_fortran_integer8_t, OSHMEM_OP_MAX, OSHMEM_OP_TYPE_FINT8);
     OBJ_OP_CREATE(max, freal4, ompi_fortran_real4_t, OSHMEM_OP_MAX, OSHMEM_OP_TYPE_FREAL4);
@@ -310,6 +328,7 @@ int oshmem_op_init(void)
     OBJ_OP_CREATE(min, float, float, OSHMEM_OP_MIN, OSHMEM_OP_TYPE_FLOAT);
     OBJ_OP_CREATE(min, double, double, OSHMEM_OP_MIN, OSHMEM_OP_TYPE_DOUBLE);
     OBJ_OP_CREATE(min, longdouble, long double, OSHMEM_OP_MIN, OSHMEM_OP_TYPE_LDOUBLE);
+    OBJ_OP_CREATE(min, fint2, ompi_fortran_integer2_t, OSHMEM_OP_MIN, OSHMEM_OP_TYPE_FINT2);
     OBJ_OP_CREATE(min, fint4, ompi_fortran_integer4_t, OSHMEM_OP_MIN, OSHMEM_OP_TYPE_FINT4);
     OBJ_OP_CREATE(min, fint8, ompi_fortran_integer8_t, OSHMEM_OP_MIN, OSHMEM_OP_TYPE_FINT8);
     OBJ_OP_CREATE(min, freal4, ompi_fortran_real4_t, OSHMEM_OP_MIN, OSHMEM_OP_TYPE_FREAL4);
@@ -328,6 +347,7 @@ int oshmem_op_init(void)
     OBJ_OP_CREATE(sum, longdouble, long double, OSHMEM_OP_SUM, OSHMEM_OP_TYPE_LDOUBLE);
     OBJ_OP_CREATE(sum, complexf, float complex, OSHMEM_OP_SUM, OSHMEM_OP_TYPE_FCOMPLEX);
     OBJ_OP_CREATE(sum, complexd, double complex, OSHMEM_OP_SUM, OSHMEM_OP_TYPE_DCOMPLEX);
+    OBJ_OP_CREATE(sum, fint2, ompi_fortran_integer2_t, OSHMEM_OP_SUM, OSHMEM_OP_TYPE_FINT2);
     OBJ_OP_CREATE(sum, fint4, ompi_fortran_integer4_t, OSHMEM_OP_SUM, OSHMEM_OP_TYPE_FINT4);
     OBJ_OP_CREATE(sum, fint8, ompi_fortran_integer8_t, OSHMEM_OP_SUM, OSHMEM_OP_TYPE_FINT8);
     OBJ_OP_CREATE(sum, freal4, ompi_fortran_real4_t, OSHMEM_OP_SUM, OSHMEM_OP_TYPE_FREAL4);
@@ -346,6 +366,7 @@ int oshmem_op_init(void)
     OBJ_OP_CREATE(prod, longdouble, long double, OSHMEM_OP_PROD, OSHMEM_OP_TYPE_LDOUBLE);
     OBJ_OP_CREATE(prod, complexf, float complex, OSHMEM_OP_PROD, OSHMEM_OP_TYPE_FCOMPLEX);
     OBJ_OP_CREATE(prod, complexd, double complex, OSHMEM_OP_PROD, OSHMEM_OP_TYPE_DCOMPLEX);
+    OBJ_OP_CREATE(prod, fint2, ompi_fortran_integer2_t, OSHMEM_OP_PROD, OSHMEM_OP_TYPE_FINT2);
     OBJ_OP_CREATE(prod, fint4, ompi_fortran_integer4_t, OSHMEM_OP_PROD, OSHMEM_OP_TYPE_FINT4);
     OBJ_OP_CREATE(prod, fint8, ompi_fortran_integer8_t, OSHMEM_OP_PROD, OSHMEM_OP_TYPE_FINT8);
     OBJ_OP_CREATE(prod, freal4, ompi_fortran_real4_t, OSHMEM_OP_PROD, OSHMEM_OP_TYPE_FREAL4);
diff --git a/oshmem/op/op.h b/oshmem/op/op.h
index d102061f52..6eeba2d7b2 100644
--- a/oshmem/op/op.h
+++ b/oshmem/op/op.h
@@ -41,6 +41,7 @@ enum {
     OSHMEM_OP_TYPE_FCOMPLEX,    /** Complex: float */
     OSHMEM_OP_TYPE_DCOMPLEX,    /** Complex: double */
 
+    OSHMEM_OP_TYPE_FINT2,       /** Fortran integer: int2 */
     OSHMEM_OP_TYPE_FINT4,       /** Fortran integer: int4 */
     OSHMEM_OP_TYPE_FINT8,       /** Fortran integer: int8 */
     OSHMEM_OP_TYPE_FREAL4,      /** Fortran integer: real4 */
@@ -91,6 +92,7 @@ OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_and_short;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_and_int;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_and_long;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_and_longlong;
+OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_and_fint2;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_and_fint4;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_and_fint8;
 
@@ -99,6 +101,7 @@ OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_or_short;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_or_int;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_or_long;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_or_longlong;
+OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_or_fint2;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_or_fint4;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_or_fint8;
 
@@ -107,6 +110,7 @@ OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_xor_short;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_xor_int;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_xor_long;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_xor_longlong;
+OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_xor_fint2;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_xor_fint4;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_xor_fint8;
 
@@ -118,6 +122,7 @@ OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_max_longlong;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_max_float;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_max_double;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_max_longdouble;
+OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_max_fint2;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_max_fint4;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_max_fint8;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_max_freal4;
@@ -132,6 +137,7 @@ OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_min_longlong;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_min_float;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_min_double;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_min_longdouble;
+OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_min_fint2;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_min_fint4;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_min_fint8;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_min_freal4;
@@ -148,6 +154,7 @@ OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_sum_double;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_sum_longdouble;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_sum_complexf;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_sum_complexd;
+OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_sum_fint2;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_sum_fint4;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_sum_fint8;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_sum_freal4;
@@ -164,6 +171,7 @@ OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_prod_double;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_prod_longdouble;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_prod_complexf;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_prod_complexd;
+OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_prod_fint2;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_prod_fint4;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_prod_fint8;
 OSHMEM_DECLSPEC extern oshmem_op_t* oshmem_op_prod_freal4;
diff --git a/oshmem/shmem/c/profile/Makefile.am b/oshmem/shmem/c/profile/Makefile.am
index cbfab0dff4..06a7f9847f 100644
--- a/oshmem/shmem/c/profile/Makefile.am
+++ b/oshmem/shmem/c/profile/Makefile.am
@@ -72,7 +72,7 @@ nodist_liboshmem_c_pshmem_la_SOURCES = \
 	$(OSHMEM_API_SOURCES)
 
 #
-# Sym link in the sources from the real MPI directory
+# Sym link in the sources from the real OSHMEM directory
 #
 $(nodist_liboshmem_c_pshmem_la_SOURCES):
 	$(OMPI_V_LN_S) if test ! -r $@ ; then \
diff --git a/oshmem/shmem/fortran/Makefile.am b/oshmem/shmem/fortran/Makefile.am
index bd9f5c916c..8d974275ed 100644
--- a/oshmem/shmem/fortran/Makefile.am
+++ b/oshmem/shmem/fortran/Makefile.am
@@ -1,5 +1,5 @@
 #
-# Copyright (c) 2013      Mellanox Technologies, Inc.
+# Copyright (c) 2014      Mellanox Technologies, Inc.
 #                         All rights reserved
 # Copyright (c) 2013-2014 Cisco Systems, Inc.  All rights reserved.
 # $COPYRIGHT$
@@ -9,9 +9,14 @@
 # $HEADER$
 #
 
-AM_CPPFLAGS     = -DOSHMEM_PROFILING_DEFINES=0 -DOSHMEM_HAVE_WEAK_SYMBOLS=0
-
 # This is guaranteed to be false if we're not building OSHMEM at all
+#
+if OSHMEM_BUILD_FORTRAN_BINDINGS
+if OSHMEM_PROFILING
+    SUBDIRS = profile
+endif
+endif
+
 if OSHMEM_BUILD_FORTRAN_BINDINGS
 oshmem_fortran_lib       = liboshmem_fortran.la
 else
@@ -24,11 +29,13 @@ headers = prototypes_shmem.h \
           bindings.h \
           shmem_fortran_pointer.h
 
-liboshmem_fortran_la_SOURCES = \
+liboshmem_fortran_la_SOURCES = shmem_finalize_f.c
+
+if ! OSHMEM_PROFILING
+liboshmem_fortran_la_SOURCES += \
                     start_pes_f.c \
                     num_pes_f.c \
                     my_pe_f.c \
-                    shmem_finalize_f.c \
                     shmem_barrier_all_f.c \
                     shpalloc_f.c \
                     shpdeallc_f.c \
@@ -47,6 +54,7 @@ liboshmem_fortran_la_SOURCES = \
                     shmem_put8_f.c \
                     shmem_put32_f.c \
                     shmem_put64_f.c \
+                    shmem_put128_f.c \
                     shmem_putmem_f.c \
                     shmem_complex_iput_f.c \
                     shmem_double_iput_f.c \
@@ -115,6 +123,7 @@ liboshmem_fortran_la_SOURCES = \
                     shmem_int8_inc_f.c \
                     shmem_quiet_f.c \
                     shmem_fence_f.c
+endif
 
 if PROJECT_OSHMEM
 if WANT_INSTALL_HEADERS
diff --git a/oshmem/shmem/fortran/my_pe_f.c b/oshmem/shmem/fortran/my_pe_f.c
index 263e4df9e2..abc07ff4e9 100644
--- a/oshmem/shmem/fortran/my_pe_f.c
+++ b/oshmem/shmem/fortran/my_pe_f.c
@@ -13,6 +13,13 @@
 #include "oshmem/shmem/fortran/bindings.h"
 #include "oshmem/include/shmem.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(MY_PE, my_pe)
+#pragma weak _my_pe_ = p_my_pe_
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_FUNCTION (MPI_Fint,
         MY_PE,
         my_pe_,
@@ -21,6 +28,11 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_FUNCTION (MPI_Fint,
         (void), 
         () )
 
+MPI_Fint _my_pe_(void)
+{
+    return my_pe_f();
+}
+
 MPI_Fint my_pe_f(void)
 {
     MPI_Fint rc;
diff --git a/oshmem/shmem/fortran/num_pes_f.c b/oshmem/shmem/fortran/num_pes_f.c
index 8e981be4b6..711c643870 100644
--- a/oshmem/shmem/fortran/num_pes_f.c
+++ b/oshmem/shmem/fortran/num_pes_f.c
@@ -13,6 +13,12 @@
 #include "oshmem/shmem/fortran/bindings.h"
 #include "oshmem/include/shmem.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(NUM_PES, num_pes)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_FUNCTION (MPI_Fint,
         NUM_PES,
         num_pes_,
diff --git a/oshmem/shmem/fortran/profile/Makefile.am b/oshmem/shmem/fortran/profile/Makefile.am
new file mode 100644
index 0000000000..af4e2a1945
--- /dev/null
+++ b/oshmem/shmem/fortran/profile/Makefile.am
@@ -0,0 +1,145 @@
+#
+# Copyright (c) 2013      Mellanox Technologies, Inc.
+#                         All rights reserved
+# Copyright (c) 2013-2014 Cisco Systems, Inc.  All rights reserved.
+# $COPYRIGHT$
+# 
+# Additional copyrights may follow
+# 
+# $HEADER$
+#
+
+AM_CPPFLAGS     = -DOSHMEM_PROFILING=1
+
+# This is guaranteed to be false if we're not building OSHMEM at all
+if OSHMEM_BUILD_FORTRAN_BINDINGS
+pshmem_fortran_lib       = liboshmem_fortran_pshmem.la
+else
+pshmem_fortran_lib       =
+endif
+
+noinst_LTLIBRARIES = $(pshmem_fortran_lib)
+
+headers = prototypes_pshmem.h \
+	pbindings.h \
+	defines.h
+
+nodist_liboshmem_fortran_pshmem_la_SOURCES = \
+	pstart_pes_f.c \
+	pnum_pes_f.c \
+	pmy_pe_f.c \
+	pshmem_barrier_all_f.c \
+	pshpalloc_f.c \
+	pshpdeallc_f.c \
+	pshpclmove_f.c \
+	pshmem_ptr_f.c \
+	pshmem_pe_accessible_f.c \
+	pshmem_addr_accessible_f.c \
+	pshmem_put_f.c \
+	pshmem_character_put_f.c \
+	pshmem_double_put_f.c \
+	pshmem_complex_put_f.c \
+	pshmem_logical_put_f.c \
+	pshmem_integer_put_f.c \
+	pshmem_real_put_f.c \
+	pshmem_put4_f.c \
+	pshmem_put8_f.c \
+	pshmem_put32_f.c \
+	pshmem_put64_f.c \
+	pshmem_put128_f.c \
+	pshmem_putmem_f.c \
+	pshmem_complex_iput_f.c \
+	pshmem_double_iput_f.c \
+	pshmem_integer_iput_f.c \
+	pshmem_iput128_f.c \
+	pshmem_iput32_f.c \
+	pshmem_iput4_f.c \
+	pshmem_iput64_f.c \
+	pshmem_iput8_f.c \
+	pshmem_logical_iput_f.c \
+	pshmem_real_iput_f.c \
+	pshmem_character_get_f.c \
+	pshmem_complex_get_f.c \
+	pshmem_double_get_f.c \
+	pshmem_get128_f.c \
+	pshmem_get32_f.c \
+	pshmem_get4_f.c \
+	pshmem_get64_f.c \
+	pshmem_get8_f.c \
+	pshmem_getmem_f.c \
+	pshmem_integer_get_f.c \
+	pshmem_logical_get_f.c \
+	pshmem_real_get_f.c \
+	pshmem_complex_iget_f.c \
+	pshmem_double_iget_f.c \
+	pshmem_iget128_f.c \
+	pshmem_iget32_f.c \
+	pshmem_iget4_f.c \
+	pshmem_iget64_f.c \
+	pshmem_iget8_f.c \
+	pshmem_integer_iget_f.c \
+	pshmem_logical_iget_f.c \
+	pshmem_real_iget_f.c \
+	pshmem_swap_f.c \
+	pshmem_int4_swap_f.c \
+	pshmem_int8_swap_f.c \
+	pshmem_real4_swap_f.c \
+	pshmem_real8_swap_f.c \
+	pshmem_int4_cswap_f.c \
+	pshmem_int8_cswap_f.c \
+	pshmem_int4_fadd_f.c \
+	pshmem_int8_fadd_f.c \
+	pshmem_int4_finc_f.c \
+	pshmem_int8_finc_f.c \
+	pshmem_int4_add_f.c \
+	pshmem_int8_add_f.c \
+	pshmem_int4_wait_f.c \
+	pshmem_int8_wait_f.c \
+	pshmem_wait_f.c \
+	pshmem_int4_wait_until_f.c \
+	pshmem_int8_wait_until_f.c \
+	pshmem_wait_until_f.c \
+	pshmem_barrier_f.c \
+	pshmem_and_to_all_f.c \
+	pshmem_or_to_all_f.c \
+	pshmem_xor_to_all_f.c \
+	pshmem_max_to_all_f.c \
+	pshmem_min_to_all_f.c \
+	pshmem_sum_to_all_f.c \
+	pshmem_prod_to_all_f.c \
+	pshmem_collect_f.c \
+	pshmem_broadcast_f.c \
+	pshmem_lock_f.c \
+	pshmem_cache_f.c \
+	pshmem_int4_inc_f.c \
+	pshmem_int8_inc_f.c \
+	pshmem_quiet_f.c \
+	pshmem_fence_f.c
+
+#
+# Sym link in the sources from the real OSHMEM directory
+#
+$(nodist_liboshmem_fortran_pshmem_la_SOURCES):
+	if test ! -r $@ ; then \
+		pname=`echo $@ | cut -b '2-'` ; \
+		$(LN_S) $(top_srcdir)/oshmem/shmem/fortran/$$pname $@ ; \
+	fi
+
+if PROJECT_OSHMEM
+if WANT_INSTALL_HEADERS
+oshmemdir = $(ompiincludedir)/$(subdir)
+oshmem_HEADERS = $(headers)
+endif
+endif
+
+# These files were created by targets above
+
+MAINTAINERCLEANFILES = $(nodist_liboshmem_fortran_pshmem_la_SOURCES)
+
+# Don't want these targets in here
+
+tags-recursive:
+tags:
+TAGS:
+GTAGS:
+ID:
diff --git a/oshmem/shmem/fortran/profile/defines.h b/oshmem/shmem/fortran/profile/defines.h
new file mode 100644
index 0000000000..bb598611b5
--- /dev/null
+++ b/oshmem/shmem/fortran/profile/defines.h
@@ -0,0 +1,577 @@
+/*
+ * Copyright (c) 2014      Mellanox Technologies, Inc.
+ *                         All rights reserved.
+ * $COPYRIGHT$
+ *
+ * Additional copyrights may follow
+ *
+ * $HEADER$
+ */
+
+#ifndef OSHMEM_FORTRAN_PROFILE_DEFINES_H
+#define OSHMEM_FORTRAN_PROFILE_DEFINES_H
+/*
+ * This file is included in the top directory only if
+ * profiling is required. Once profiling is required,
+ * this file will replace all shmem_* symbols with
+ * pshmem_* symbols
+ */
+
+#define MY_PE PMY_PE
+#define my_pe_ pmy_pe_
+#define my_pe__ pmy_pe__
+#define _my_pe_ p_my_pe_
+
+#define NUM_PES PNUM_PES
+#define num_pes_ pnum_pes_
+#define num_pes__ pnum_pes__
+
+#define SHMEM_ADDR_ACCESSIBLE PSHMEM_ADDR_ACCESSIBLE
+#define shmem_addr_accessible_ pshmem_addr_accessible_
+#define shmem_addr_accessible__ pshmem_addr_accessible__
+
+#define SHMEM_INT2_AND_TO_ALL PSHMEM_INT2_AND_TO_ALL
+#define shmem_int2_and_to_all_ pshmem_int2_and_to_all_
+#define shmem_int2_and_to_all__ pshmem_int2_and_to_all__
+
+#define SHMEM_INT4_AND_TO_ALL PSHMEM_INT4_AND_TO_ALL
+#define shmem_int4_and_to_all_ pshmem_int4_and_to_all_
+#define shmem_int4_and_to_all__ pshmem_int4_and_to_all__
+
+#define SHMEM_INT8_AND_TO_ALL PSHMEM_INT8_AND_TO_ALL
+#define shmem_int8_and_to_all_ pshmem_int8_and_to_all_
+#define shmem_int8_and_to_all__ pshmem_int8_and_to_all__
+
+#define SHMEM_BARRIER_ALL PSHMEM_BARRIER_ALL
+#define shmem_barrier_all_ pshmem_barrier_all_
+#define shmem_barrier_all__ pshmem_barrier_all__
+
+#define SHMEM_BARRIER PSHMEM_BARRIER
+#define shmem_barrier_ pshmem_barrier_
+#define shmem_barrier__ pshmem_barrier__
+
+#define SHMEM_BROADCAST4 PSHMEM_BROADCAST4
+#define shmem_broadcast4_ pshmem_broadcast4_
+#define shmem_broadcast4__ pshmem_broadcast4__
+
+#define SHMEM_BROADCAST8 PSHMEM_BROADCAST8
+#define shmem_broadcast8_ pshmem_broadcast8_
+#define shmem_broadcast8__ pshmem_broadcast8__
+
+#define SHMEM_BROADCAST32 PSHMEM_BROADCAST32
+#define shmem_broadcast32_ pshmem_broadcast32_
+#define shmem_broadcast32__ pshmem_broadcast32__
+
+#define SHMEM_BROADCAST64 PSHMEM_BROADCAST64
+#define shmem_broadcast64_ pshmem_broadcast64_
+#define shmem_broadcast64__ pshmem_broadcast64__
+
+#define SHMEM_SET_CACHE_INV PSHMEM_SET_CACHE_INV
+#define shmem_set_cache_inv_ pshmem_set_cache_inv_
+#define shmem_set_cache_inv__ pshmem_set_cache_inv__
+
+#define SHMEM_SET_CACHE_LINE_INV PSHMEM_SET_CACHE_LINE_INV
+#define shmem_set_cache_line_inv_ pshmem_set_cache_line_inv_
+#define shmem_set_cache_line_inv__ pshmem_set_cache_line_inv__
+
+#define SHMEM_CLEAR_CACHE_INV PSHMEM_CLEAR_CACHE_INV
+#define shmem_clear_cache_inv_ pshmem_clear_cache_inv_
+#define shmem_clear_cache_inv__ pshmem_clear_cache_inv__
+
+#define SHMEM_CLEAR_CACHE_LINE_INV PSHMEM_CLEAR_CACHE_LINE_INV
+#define shmem_clear_cache_line_inv_ pshmem_clear_cache_line_inv_
+#define shmem_clear_cache_line_inv__ pshmem_clear_cache_line_inv__
+
+#define SHMEM_UDCFLUSH PSHMEM_UDCFLUSH
+#define shmem_udcflush_ pshmem_udcflush_
+#define shmem_udcflush__ pshmem_udcflush__
+
+#define SHMEM_UDCFLUSH_LINE PSHMEM_UDCFLUSH_LINE
+#define shmem_udcflush_line_ pshmem_udcflush_line_
+#define shmem_udcflush_line__ pshmem_udcflush_line__
+
+#define SHMEM_CHARACTER_GET PSHMEM_CHARACTER_GET
+#define shmem_character_get_ pshmem_character_get_
+#define shmem_character_get__ pshmem_character_get__
+
+#define SHMEM_CHARACTER_PUT PSHMEM_CHARACTER_PUT
+#define shmem_character_put_ pshmem_character_put_
+#define shmem_character_put__ pshmem_character_put__
+
+#define SHMEM_COLLECT4 PSHMEM_COLLECT4
+#define shmem_collect4_ pshmem_collect4_
+#define shmem_collect4__ pshmem_collect4__
+
+#define SHMEM_COLLECT8 PSHMEM_COLLECT8
+#define shmem_collect8_ pshmem_collect8_
+#define shmem_collect8__ pshmem_collect8__
+
+#define SHMEM_COLLECT32 PSHMEM_COLLECT32
+#define shmem_collect32_ pshmem_collect32_
+#define shmem_collect32__ pshmem_collect32__
+
+#define SHMEM_COLLECT64 PSHMEM_COLLECT64
+#define shmem_collect64_ pshmem_collect64_
+#define shmem_collect64__ pshmem_collect64__
+
+#define SHMEM_FCOLLECT4 PSHMEM_FCOLLECT4
+#define shmem_fcollect4_ pshmem_fcollect4_
+#define shmem_fcollect4__ pshmem_fcollect4__
+
+#define SHMEM_FCOLLECT8 PSHMEM_FCOLLECT8
+#define shmem_fcollect8_ pshmem_fcollect8_
+#define shmem_fcollect8__ pshmem_fcollect8__
+
+#define SHMEM_FCOLLECT32 PSHMEM_FCOLLECT32
+#define shmem_fcollect32_ pshmem_fcollect32_
+#define shmem_fcollect32__ pshmem_fcollect32__
+
+#define SHMEM_FCOLLECT64 PSHMEM_FCOLLECT64
+#define shmem_fcollect64_ pshmem_fcollect64_
+#define shmem_fcollect64__ pshmem_fcollect64__
+
+#define SHMEM_COMPLEX_GET PSHMEM_COMPLEX_GET
+#define shmem_complex_get_ pshmem_complex_get_
+#define shmem_complex_get__ pshmem_complex_get__
+
+#define SHMEM_COMPLEX_IGET PSHMEM_COMPLEX_IGET
+#define shmem_complex_iget_ pshmem_complex_iget_
+#define shmem_complex_iget__ pshmem_complex_iget__
+
+#define SHMEM_COMPLEX_IPUT PSHMEM_COMPLEX_IPUT
+#define shmem_complex_iput_ pshmem_complex_iput_
+#define shmem_complex_iput__ pshmem_complex_iput__
+
+#define SHMEM_COMPLEX_PUT PSHMEM_COMPLEX_PUT
+#define shmem_complex_put_ pshmem_complex_put_
+#define shmem_complex_put__ pshmem_complex_put__
+
+#define SHMEM_DOUBLE_GET PSHMEM_DOUBLE_GET
+#define shmem_double_get_ pshmem_double_get_
+#define shmem_double_get__ pshmem_double_get__
+
+#define SHMEM_DOUBLE_IGET PSHMEM_DOUBLE_IGET
+#define shmem_double_iget_ pshmem_double_iget_
+#define shmem_double_iget__ pshmem_double_iget__
+
+#define SHMEM_DOUBLE_IPUT PSHMEM_DOUBLE_IPUT
+#define shmem_double_iput_ pshmem_double_iput_
+#define shmem_double_iput__ pshmem_double_iput__
+
+#define SHMEM_DOUBLE_PUT PSHMEM_DOUBLE_PUT
+#define shmem_double_put_ pshmem_double_put_
+#define shmem_double_put__ pshmem_double_put__
+
+#define SHMEM_FENCE PSHMEM_FENCE
+#define shmem_fence_ pshmem_fence_
+#define shmem_fence__ pshmem_fence__
+
+#define SHMEM_GET128 PSHMEM_GET128
+#define shmem_get128_ pshmem_get128_
+#define shmem_get128__ pshmem_get128__
+
+#define SHMEM_GET32 PSHMEM_GET32
+#define shmem_get32_ pshmem_get32_
+#define shmem_get32__ pshmem_get32__
+
+#define SHMEM_GET4 PSHMEM_GET4
+#define shmem_get4_ pshmem_get4_
+#define shmem_get4__ pshmem_get4__
+
+#define SHMEM_GET64 PSHMEM_GET64
+#define shmem_get64_ pshmem_get64_
+#define shmem_get64__ pshmem_get64__
+
+#define SHMEM_GET8 PSHMEM_GET8
+#define shmem_get8_ pshmem_get8_
+#define shmem_get8__ pshmem_get8__
+
+#define SHMEM_GETMEM PSHMEM_GETMEM
+#define shmem_getmem_ pshmem_getmem_
+#define shmem_getmem__ pshmem_getmem__
+
+#define SHMEM_IGET128 PSHMEM_IGET128
+#define shmem_iget128_ pshmem_iget128_
+#define shmem_iget128__ pshmem_iget128__
+
+#define SHMEM_IGET32 PSHMEM_IGET32
+#define shmem_iget32_ pshmem_iget32_
+#define shmem_iget32__ pshmem_iget32__
+
+#define SHMEM_IGET4 PSHMEM_IGET4
+#define shmem_iget4_ pshmem_iget4_
+#define shmem_iget4__ pshmem_iget4__
+
+#define SHMEM_IGET64 PSHMEM_IGET64
+#define shmem_iget64_ pshmem_iget64_
+#define shmem_iget64__ pshmem_iget64__
+
+#define SHMEM_IGET8 PSHMEM_IGET8
+#define shmem_iget8_ pshmem_iget8_
+#define shmem_iget8__ pshmem_iget8__
+
+#define SHMEM_INT4_ADD PSHMEM_INT4_ADD
+#define shmem_int4_add_ pshmem_int4_add_
+#define shmem_int4_add__ pshmem_int4_add__
+
+#define SHMEM_INT4_CSWAP PSHMEM_INT4_CSWAP
+#define shmem_int4_cswap_ pshmem_int4_cswap_
+#define shmem_int4_cswap__ pshmem_int4_cswap__
+
+#define SHMEM_INT4_FADD PSHMEM_INT4_FADD
+#define shmem_int4_fadd_ pshmem_int4_fadd_
+#define shmem_int4_fadd__ pshmem_int4_fadd__
+
+#define SHMEM_INT4_FINC PSHMEM_INT4_FINC
+#define shmem_int4_finc_ pshmem_int4_finc_
+#define shmem_int4_finc__ pshmem_int4_finc__
+
+#define SHMEM_INT4_INC PSHMEM_INT4_INC
+#define shmem_int4_inc_ pshmem_int4_inc_
+#define shmem_int4_inc__ pshmem_int4_inc__
+
+#define SHMEM_INT4_SWAP PSHMEM_INT4_SWAP
+#define shmem_int4_swap_ pshmem_int4_swap_
+#define shmem_int4_swap__ pshmem_int4_swap__
+
+#define SHMEM_INT4_WAIT PSHMEM_INT4_WAIT
+#define shmem_int4_wait_ pshmem_int4_wait_
+#define shmem_int4_wait__ pshmem_int4_wait__
+
+#define SHMEM_INT4_WAIT_UNTIL PSHMEM_INT4_WAIT_UNTIL
+#define shmem_int4_wait_until_ pshmem_int4_wait_until_
+#define shmem_int4_wait_until__ pshmem_int4_wait_until__
+
+#define SHMEM_INT8_ADD PSHMEM_INT8_ADD
+#define shmem_int8_add_ pshmem_int8_add_
+#define shmem_int8_add__ pshmem_int8_add__
+
+#define SHMEM_INT8_CSWAP PSHMEM_INT8_CSWAP
+#define shmem_int8_cswap_ pshmem_int8_cswap_
+#define shmem_int8_cswap__ pshmem_int8_cswap__
+
+#define SHMEM_INT8_FADD PSHMEM_INT8_FADD
+#define shmem_int8_fadd_ pshmem_int8_fadd_
+#define shmem_int8_fadd__ pshmem_int8_fadd__
+
+#define SHMEM_INT8_FINC PSHMEM_INT8_FINC
+#define shmem_int8_finc_ pshmem_int8_finc_
+#define shmem_int8_finc__ pshmem_int8_finc__
+
+#define SHMEM_INT8_INC PSHMEM_INT8_INC
+#define shmem_int8_inc_ pshmem_int8_inc_
+#define shmem_int8_inc__ pshmem_int8_inc__
+
+#define SHMEM_INT8_SWAP PSHMEM_INT8_SWAP
+#define shmem_int8_swap_ pshmem_int8_swap_
+#define shmem_int8_swap__ pshmem_int8_swap__
+
+#define SHMEM_INT8_WAIT PSHMEM_INT8_WAIT
+#define shmem_int8_wait_ pshmem_int8_wait_
+#define shmem_int8_wait__ pshmem_int8_wait__
+
+#define SHMEM_INT8_WAIT_UNTIL PSHMEM_INT8_WAIT_UNTIL
+#define shmem_int8_wait_until_ pshmem_int8_wait_until_
+#define shmem_int8_wait_until__ pshmem_int8_wait_until__
+
+#define SHMEM_INTEGER_GET PSHMEM_INTEGER_GET
+#define shmem_integer_get_ pshmem_integer_get_
+#define shmem_integer_get__ pshmem_integer_get__
+
+#define SHMEM_INTEGER_IGET PSHMEM_INTEGER_IGET
+#define shmem_integer_iget_ pshmem_integer_iget_
+#define shmem_integer_iget__ pshmem_integer_iget__
+
+#define SHMEM_INTEGER_IPUT PSHMEM_INTEGER_IPUT
+#define shmem_integer_iput_ pshmem_integer_iput_
+#define shmem_integer_iput__ pshmem_integer_iput__
+
+#define SHMEM_INTEGER_PUT PSHMEM_INTEGER_PUT
+#define shmem_integer_put_ pshmem_integer_put_
+#define shmem_integer_put__ pshmem_integer_put__
+
+#define SHMEM_IPUT128 PSHMEM_IPUT128
+#define shmem_iput128_ pshmem_iput128_
+#define shmem_iput128__ pshmem_iput128__
+
+#define SHMEM_IPUT32 PSHMEM_IPUT32
+#define shmem_iput32_ pshmem_iput32_
+#define shmem_iput32__ pshmem_iput32__
+
+#define SHMEM_IPUT4 PSHMEM_IPUT4
+#define shmem_iput4_ pshmem_iput4_
+#define shmem_iput4__ pshmem_iput4__
+
+#define SHMEM_IPUT64 PSHMEM_IPUT64
+#define shmem_iput64_ pshmem_iput64_
+#define shmem_iput64__ pshmem_iput64__
+
+#define SHMEM_IPUT8 PSHMEM_IPUT8
+#define shmem_iput8_ pshmem_iput8_
+#define shmem_iput8__ pshmem_iput8__
+
+#define SHMEM_SET_LOCK PSHMEM_SET_LOCK
+#define shmem_set_lock_ pshmem_set_lock_
+#define shmem_set_lock__ pshmem_set_lock__
+
+#define SHMEM_TEST_LOCK PSHMEM_TEST_LOCK
+#define shmem_test_lock_ pshmem_test_lock_
+#define shmem_test_lock__ pshmem_test_lock__
+
+#define SHMEM_CLEAR_LOCK PSHMEM_CLEAR_LOCK
+#define shmem_clear_lock_ pshmem_clear_lock_
+#define shmem_clear_lock__ pshmem_clear_lock__
+
+#define SHMEM_LOGICAL_GET PSHMEM_LOGICAL_GET
+#define shmem_logical_get_ pshmem_logical_get_
+#define shmem_logical_get__ pshmem_logical_get__
+
+#define SHMEM_LOGICAL_IGET PSHMEM_LOGICAL_IGET
+#define shmem_logical_iget_ pshmem_logical_iget_
+#define shmem_logical_iget__ pshmem_logical_iget__
+
+#define SHMEM_LOGICAL_IPUT PSHMEM_LOGICAL_IPUT
+#define shmem_logical_iput_ pshmem_logical_iput_
+#define shmem_logical_iput__ pshmem_logical_iput__
+
+#define SHMEM_LOGICAL_PUT PSHMEM_LOGICAL_PUT
+#define shmem_logical_put_ pshmem_logical_put_
+#define shmem_logical_put__ pshmem_logical_put__
+
+#define SHMEM_INT2_MAX_TO_ALL PSHMEM_INT2_MAX_TO_ALL
+#define shmem_int2_max_to_all_ pshmem_int2_max_to_all_
+#define shmem_int2_max_to_all__ pshmem_int2_max_to_all__
+
+#define SHMEM_INT4_MAX_TO_ALL PSHMEM_INT4_MAX_TO_ALL
+#define shmem_int4_max_to_all_ pshmem_int4_max_to_all_
+#define shmem_int4_max_to_all__ pshmem_int4_max_to_all__
+
+#define SHMEM_INT8_MAX_TO_ALL PSHMEM_INT8_MAX_TO_ALL
+#define shmem_int8_max_to_all_ pshmem_int8_max_to_all_
+#define shmem_int8_max_to_all__ pshmem_int8_max_to_all__
+
+#define SHMEM_REAL4_MAX_TO_ALL PSHMEM_REAL4_MAX_TO_ALL
+#define shmem_real4_max_to_all_ pshmem_real4_max_to_all_
+#define shmem_real4_max_to_all__ pshmem_real4_max_to_all__
+
+#define SHMEM_REAL8_MAX_TO_ALL PSHMEM_REAL8_MAX_TO_ALL
+#define shmem_real8_max_to_all_ pshmem_real8_max_to_all_
+#define shmem_real8_max_to_all__ pshmem_real8_max_to_all__
+
+#define SHMEM_REAL16_MAX_TO_ALL PSHMEM_REAL16_MAX_TO_ALL
+#define shmem_real16_max_to_all_ pshmem_real16_max_to_all_
+#define shmem_real16_max_to_all__ pshmem_real16_max_to_all__
+
+#define SHMEM_INT2_MIN_TO_ALL PSHMEM_INT2_MIN_TO_ALL
+#define shmem_int2_min_to_all_ pshmem_int2_min_to_all_
+#define shmem_int2_min_to_all__ pshmem_int2_min_to_all__
+
+#define SHMEM_INT4_MIN_TO_ALL PSHMEM_INT4_MIN_TO_ALL
+#define shmem_int4_min_to_all_ pshmem_int4_min_to_all_
+#define shmem_int4_min_to_all__ pshmem_int4_min_to_all__
+
+#define SHMEM_INT8_MIN_TO_ALL PSHMEM_INT8_MIN_TO_ALL
+#define shmem_int8_min_to_all_ pshmem_int8_min_to_all_
+#define shmem_int8_min_to_all__ pshmem_int8_min_to_all__
+
+#define SHMEM_REAL4_MIN_TO_ALL PSHMEM_REAL4_MIN_TO_ALL
+#define shmem_real4_min_to_all_ pshmem_real4_min_to_all_
+#define shmem_real4_min_to_all__ pshmem_real4_min_to_all__
+
+#define SHMEM_REAL8_MIN_TO_ALL PSHMEM_REAL8_MIN_TO_ALL
+#define shmem_real8_min_to_all_ pshmem_real8_min_to_all_
+#define shmem_real8_min_to_all__ pshmem_real8_min_to_all__
+
+#define SHMEM_REAL16_MIN_TO_ALL PSHMEM_REAL16_MIN_TO_ALL
+#define shmem_real16_min_to_all_ pshmem_real16_min_to_all_
+#define shmem_real16_min_to_all__ pshmem_real16_min_to_all__
+
+#define SHMEM_INT2_OR_TO_ALL PSHMEM_INT2_OR_TO_ALL
+#define shmem_int2_or_to_all_ pshmem_int2_or_to_all_
+#define shmem_int2_or_to_all__ pshmem_int2_or_to_all__
+
+#define SHMEM_INT4_OR_TO_ALL PSHMEM_INT4_OR_TO_ALL
+#define shmem_int4_or_to_all_ pshmem_int4_or_to_all_
+#define shmem_int4_or_to_all__ pshmem_int4_or_to_all__
+
+#define SHMEM_INT8_OR_TO_ALL PSHMEM_INT8_OR_TO_ALL
+#define shmem_int8_or_to_all_ pshmem_int8_or_to_all_
+#define shmem_int8_or_to_all__ pshmem_int8_or_to_all__
+
+#define SHMEM_PE_ACCESSIBLE PSHMEM_PE_ACCESSIBLE
+#define shmem_pe_accessible_ pshmem_pe_accessible_
+#define shmem_pe_accessible__ pshmem_pe_accessible__
+
+#define SHMEM_INT2_PROD_TO_ALL PSHMEM_INT2_PROD_TO_ALL
+#define shmem_int2_prod_to_all_ pshmem_int2_prod_to_all_
+#define shmem_int2_prod_to_all__ pshmem_int2_prod_to_all__
+
+#define SHMEM_INT4_PROD_TO_ALL PSHMEM_INT4_PROD_TO_ALL
+#define shmem_int4_prod_to_all_ pshmem_int4_prod_to_all_
+#define shmem_int4_prod_to_all__ pshmem_int4_prod_to_all__
+
+#define SHMEM_INT8_PROD_TO_ALL PSHMEM_INT8_PROD_TO_ALL
+#define shmem_int8_prod_to_all_ pshmem_int8_prod_to_all_
+#define shmem_int8_prod_to_all__ pshmem_int8_prod_to_all__
+
+#define SHMEM_COMP4_PROD_TO_ALL PSHMEM_COMP4_PROD_TO_ALL
+#define shmem_comp4_prod_to_all_ pshmem_comp4_prod_to_all_
+#define shmem_comp4_prod_to_all__ pshmem_comp4_prod_to_all__
+
+#define SHMEM_COMP8_PROD_TO_ALL PSHMEM_COMP8_PROD_TO_ALL
+#define shmem_comp8_prod_to_all_ pshmem_comp8_prod_to_all_
+#define shmem_comp8_prod_to_all__ pshmem_comp8_prod_to_all__
+
+#define SHMEM_REAL4_PROD_TO_ALL PSHMEM_REAL4_PROD_TO_ALL
+#define shmem_real4_prod_to_all_ pshmem_real4_prod_to_all_
+#define shmem_real4_prod_to_all__ pshmem_real4_prod_to_all__
+
+#define SHMEM_REAL8_PROD_TO_ALL PSHMEM_REAL8_PROD_TO_ALL
+#define shmem_real8_prod_to_all_ pshmem_real8_prod_to_all_
+#define shmem_real8_prod_to_all__ pshmem_real8_prod_to_all__
+
+#define SHMEM_REAL16_PROD_TO_ALL PSHMEM_REAL16_PROD_TO_ALL
+#define shmem_real16_prod_to_all_ pshmem_real16_prod_to_all_
+#define shmem_real16_prod_to_all__ pshmem_real16_prod_to_all__
+
+#define SHMEM_PTR PSHMEM_PTR
+#define shmem_ptr_ pshmem_ptr_
+#define shmem_ptr__ pshmem_ptr__
+
+#define SHMEM_PUT32 PSHMEM_PUT32
+#define shmem_put32_ pshmem_put32_
+#define shmem_put32__ pshmem_put32__
+
+#define SHMEM_PUT4 PSHMEM_PUT4
+#define shmem_put4_ pshmem_put4_
+#define shmem_put4__ pshmem_put4__
+
+#define SHMEM_PUT64 PSHMEM_PUT64
+#define shmem_put64_ pshmem_put64_
+#define shmem_put64__ pshmem_put64__
+
+#define SHMEM_PUT128 PSHMEM_PUT128
+#define shmem_put128_ pshmem_put128_
+#define shmem_put128__ pshmem_put128__
+
+#define SHMEM_PUT8 PSHMEM_PUT8
+#define shmem_put8_ pshmem_put8_
+#define shmem_put8__ pshmem_put8__
+
+#define SHMEM_PUT PSHMEM_PUT
+#define shmem_put_ pshmem_put_
+#define shmem_put__ pshmem_put__
+
+#define SHMEM_PUTMEM PSHMEM_PUTMEM
+#define shmem_putmem_ pshmem_putmem_
+#define shmem_putmem__ pshmem_putmem__
+
+#define SHMEM_QUIET PSHMEM_QUIET
+#define shmem_quiet_ pshmem_quiet_
+#define shmem_quiet__ pshmem_quiet__
+
+#define SHMEM_REAL4_SWAP PSHMEM_REAL4_SWAP
+#define shmem_real4_swap_ pshmem_real4_swap_
+#define shmem_real4_swap__ pshmem_real4_swap__
+
+#define SHMEM_REAL8_SWAP PSHMEM_REAL8_SWAP
+#define shmem_real8_swap_ pshmem_real8_swap_
+#define shmem_real8_swap__ pshmem_real8_swap__
+
+#define SHMEM_REAL_GET PSHMEM_REAL_GET
+#define shmem_real_get_ pshmem_real_get_
+#define shmem_real_get__ pshmem_real_get__
+
+#define SHMEM_REAL_IGET PSHMEM_REAL_IGET
+#define shmem_real_iget_ pshmem_real_iget_
+#define shmem_real_iget__ pshmem_real_iget__
+
+#define SHMEM_REAL_IPUT PSHMEM_REAL_IPUT
+#define shmem_real_iput_ pshmem_real_iput_
+#define shmem_real_iput__ pshmem_real_iput__
+
+#define SHMEM_REAL_PUT PSHMEM_REAL_PUT
+#define shmem_real_put_ pshmem_real_put_
+#define shmem_real_put__ pshmem_real_put__
+
+#define SHMEM_INT2_SUM_TO_ALL PSHMEM_INT2_SUM_TO_ALL
+#define shmem_int2_sum_to_all_ pshmem_int2_sum_to_all_
+#define shmem_int2_sum_to_all__ pshmem_int2_sum_to_all__
+
+#define SHMEM_INT4_SUM_TO_ALL PSHMEM_INT4_SUM_TO_ALL
+#define shmem_int4_sum_to_all_ pshmem_int4_sum_to_all_
+#define shmem_int4_sum_to_all__ pshmem_int4_sum_to_all__
+
+#define SHMEM_INT8_SUM_TO_ALL PSHMEM_INT8_SUM_TO_ALL
+#define shmem_int8_sum_to_all_ pshmem_int8_sum_to_all_
+#define shmem_int8_sum_to_all__ pshmem_int8_sum_to_all__
+
+#define SHMEM_COMP4_SUM_TO_ALL PSHMEM_COMP4_SUM_TO_ALL
+#define shmem_comp4_sum_to_all_ pshmem_comp4_sum_to_all_
+#define shmem_comp4_sum_to_all__ pshmem_comp4_sum_to_all__
+
+#define SHMEM_COMP8_SUM_TO_ALL PSHMEM_COMP8_SUM_TO_ALL
+#define shmem_comp8_sum_to_all_ pshmem_comp8_sum_to_all_
+#define shmem_comp8_sum_to_all__ pshmem_comp8_sum_to_all__
+
+#define SHMEM_REAL4_SUM_TO_ALL PSHMEM_REAL4_SUM_TO_ALL
+#define shmem_real4_sum_to_all_ pshmem_real4_sum_to_all_
+#define shmem_real4_sum_to_all__ pshmem_real4_sum_to_all__
+
+#define SHMEM_REAL8_SUM_TO_ALL PSHMEM_REAL8_SUM_TO_ALL
+#define shmem_real8_sum_to_all_ pshmem_real8_sum_to_all_
+#define shmem_real8_sum_to_all__ pshmem_real8_sum_to_all__
+
+#define SHMEM_REAL16_SUM_TO_ALL PSHMEM_REAL16_SUM_TO_ALL
+#define shmem_real16_sum_to_all_ pshmem_real16_sum_to_all_
+#define shmem_real16_sum_to_all__ pshmem_real16_sum_to_all__
+
+#define SHMEM_SWAP PSHMEM_SWAP
+#define shmem_swap_ pshmem_swap_
+#define shmem_swap__ pshmem_swap__
+
+#define SHMEM_WAIT PSHMEM_WAIT
+#define shmem_wait_ pshmem_wait_
+#define shmem_wait__ pshmem_wait__
+
+#define SHMEM_WAIT_UNTIL PSHMEM_WAIT_UNTIL
+#define shmem_wait_until_ pshmem_wait_until_
+#define shmem_wait_until__ pshmem_wait_until__
+
+#define SHMEM_INT2_XOR_TO_ALL PSHMEM_INT2_XOR_TO_ALL
+#define shmem_int2_xor_to_all_ pshmem_int2_xor_to_all_
+#define shmem_int2_xor_to_all__ pshmem_int2_xor_to_all__
+
+#define SHMEM_INT4_XOR_TO_ALL PSHMEM_INT4_XOR_TO_ALL
+#define shmem_int4_xor_to_all_ pshmem_int4_xor_to_all_
+#define shmem_int4_xor_to_all__ pshmem_int4_xor_to_all__
+
+#define SHMEM_INT8_XOR_TO_ALL PSHMEM_INT8_XOR_TO_ALL
+#define shmem_int8_xor_to_all_ pshmem_int8_xor_to_all_
+#define shmem_int8_xor_to_all__ pshmem_int8_xor_to_all__
+
+#define SHMEM_COMP4_XOR_TO_ALL PSHMEM_COMP4_XOR_TO_ALL
+#define shmem_comp4_xor_to_all_ pshmem_comp4_xor_to_all_
+#define shmem_comp4_xor_to_all__ pshmem_comp4_xor_to_all__
+
+#define SHMEM_COMP8_XOR_TO_ALL PSHMEM_COMP8_XOR_TO_ALL
+#define shmem_comp8_xor_to_all_ pshmem_comp8_xor_to_all_
+#define shmem_comp8_xor_to_all__ pshmem_comp8_xor_to_all__
+
+#define SHPALLOC PSHPALLOC
+#define shpalloc_ pshpalloc_
+#define shpalloc__ pshpalloc__
+
+#define SHPCLMOVE PSHPCLMOVE
+#define shpclmove_ pshpclmove_
+#define shpclmove__ pshpclmove__
+
+#define SHPDEALLC PSHPDEALLC
+#define shpdeallc_ pshpdeallc_
+#define shpdeallc__ pshpdeallc__
+
+#define START_PES PSTART_PES
+#define start_pes_ pstart_pes_
+#define start_pes__ pstart_pes__
+
+#endif /* OSHMEM_FORTRAN_PROFILE_DEFINES_H */
diff --git a/oshmem/shmem/fortran/profile/pbindings.h b/oshmem/shmem/fortran/profile/pbindings.h
new file mode 100644
index 0000000000..ad680a92d4
--- /dev/null
+++ b/oshmem/shmem/fortran/profile/pbindings.h
@@ -0,0 +1,24 @@
+/*
+ * Copyright (c) 2014      Mellanox Technologies, Inc.
+ *                         All rights reserved.
+ * Copyright (c) 2013 Cisco Systems, Inc.  All rights reserved.
+ * $COPYRIGHT$
+ * 
+ * Additional copyrights may follow
+ * 
+ * $HEADER$
+ */
+
+#ifndef SHMEM_FORTRAN_PBINDINGS_H
+#define SHMEM_FORTRAN_PBINDINGS_H
+
+#include "prototypes_pshmem.h"
+
+#define SHMEM_GENERATE_WEAK_PRAGMA(x) _Pragma(#x)
+
+#define SHMEM_GENERATE_WEAK_BINDINGS(UPPER_NAME, lower_name)                 \
+    SHMEM_GENERATE_WEAK_PRAGMA(weak UPPER_NAME = P ## UPPER_NAME)            \
+    SHMEM_GENERATE_WEAK_PRAGMA(weak lower_name ## _ = p ## lower_name ## _)  \
+    SHMEM_GENERATE_WEAK_PRAGMA(weak lower_name ## __ = p ## lower_name ## __)
+
+#endif /*SHMEM_FORTRAN_PBINDINGS_H*/
diff --git a/oshmem/shmem/fortran/profile/prototypes_pshmem.h b/oshmem/shmem/fortran/profile/prototypes_pshmem.h
new file mode 100644
index 0000000000..34882028bf
--- /dev/null
+++ b/oshmem/shmem/fortran/profile/prototypes_pshmem.h
@@ -0,0 +1,167 @@
+/*
+ * Copyright (c) 2014      Mellanox Technologies, Inc.
+ *                         All rights reserved.
+ * $COPYRIGHT$
+ *
+ * Additional copyrights may follow
+ *
+ * $HEADER$
+ */
+
+#ifndef OSHMEM_F77_PROTOTYPES_PSHMEM_H
+#define OSHMEM_F77_PROTOTYPES_PSHMEM_H
+#include "oshmem_config.h"
+#include "oshmem/shmem/fortran/shmem_fortran_pointer.h"
+
+BEGIN_C_DECLS
+
+#define PN(ret, lower_name, upper_name, args) \
+  OSHMEM_DECLSPEC ret lower_name##_f args;    \
+  OSHMEM_DECLSPEC ret lower_name##_ args;     \
+  OSHMEM_DECLSPEC ret lower_name##__ args;    \
+  OSHMEM_DECLSPEC ret upper_name args
+
+PN (void, pstart_pes, PSTART_PES, (MPI_Fint npes));
+PN (MPI_Fint, pnum_pes, PNUM_PES, (void));
+PN (MPI_Fint, pmy_pe, PMY_PE, (void));
+OSHMEM_DECLSPEC MPI_Fint p_my_pe(void);
+PN (void, shmem_finalize, SHMEM_FINALIZE, (void));
+PN (void, pshmem_barrier_all, PSHMEM_BARRIER_ALL, (void));
+PN (void, pshpalloc, PSHPALLOC, (FORTRAN_POINTER_T *addr, MPI_Fint *length, MPI_Fint *errcode, MPI_Fint *abort));
+PN (void, pshpdeallc, PSHPDEALLC, (FORTRAN_POINTER_T *addr, MPI_Fint *errcode, MPI_Fint *abort));
+PN (void, pshpclmove, PSHPCLMOVE, (FORTRAN_POINTER_T *addr, MPI_Fint *length, MPI_Fint *status, MPI_Fint *abort));
+PN (FORTRAN_POINTER_T*, pshmem_ptr, PSHMEM_PTR, (FORTRAN_POINTER_T target, MPI_Fint *pe));
+PN (ompi_fortran_logical_t, pshmem_pe_accessible, PSHMEM_PE_ACCESSIBLE, (MPI_Fint *pe));
+PN (MPI_Fint, pshmem_addr_accessible, PSHMEM_ADDR_ACCESSIBLE, (FORTRAN_POINTER_T addr, MPI_Fint *pe));
+PN (void, pshmem_put, PSHMEM_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_character_put, PSHMEM_CHARACTER_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_complex_put, PSHMEM_COMPLEX_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_double_put, PSHMEM_DOUBLE_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_logical_put, PSHMEM_LOGICAL_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_integer_put, PSHMEM_INTEGER_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_real_put, PSHMEM_REAL_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_put4, PSHMEM_PUT4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_put8, PSHMEM_PUT8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_put32, PSHMEM_PUT32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_put64, PSHMEM_PUT64, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_put128, PSHMEM_PUT128, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_putmem, PSHMEM_PUTMEM, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, pshmem_iput4, PSHMEM_IPUT4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_iput8, PSHMEM_IPUT8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_iput32, PSHMEM_IPUT32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_iput64, PSHMEM_IPUT64, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_iput128, PSHMEM_IPUT128, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_complex_iput, PSHMEM_COMPLEX_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_double_iput, PSHMEM_DOUBLE_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_integer_iput, PSHMEM_INTEGER_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_logical_iput, PSHMEM_LOGICAL_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_real_iput, PSHMEM_REAL_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_character_get, PSHMEM_CHARACTER_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_complex_get, PSHMEM_COMPLEX_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_double_get, PSHMEM_DOUBLE_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_integer_get, PSHMEM_INTEGER_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_get4, PSHMEM_GET4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_get8, PSHMEM_GET8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_get32, PSHMEM_GET32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_get64, PSHMEM_GET64, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_get128, PSHMEM_GET128, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_getmem, PSHMEM_GETMEM, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_logical_get, PSHMEM_LOGICAL_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_real_get, PSHMEM_REAL_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_iget4, PSHMEM_IGET4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_iget8, PSHMEM_IGET8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_iget32, PSHMEM_IGET32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_iget64, PSHMEM_IGET64, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_iget128, PSHMEM_IGET128, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_complex_iget, PSHMEM_COMPLEX_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_double_iget, PSHMEM_DOUBLE_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_integer_iget, PSHMEM_INTEGER_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_logical_iget, PSHMEM_LOGICAL_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, pshmem_real_iget, PSHMEM_REAL_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (MPI_Fint, pshmem_swap, PSHMEM_SWAP, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));
+PN (ompi_fortran_integer4_t, pshmem_int4_swap, PSHMEM_INT4_SWAP, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));
+PN (ompi_fortran_integer8_t, pshmem_int8_swap, PSHMEM_INT8_SWAP, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));
+PN (ompi_fortran_real4_t, pshmem_real4_swap, PSHMEM_REAL4_SWAP, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));
+PN (ompi_fortran_real8_t, pshmem_real8_swap, PSHMEM_REAL8_SWAP, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));
+PN (ompi_fortran_integer4_t, pshmem_int4_cswap, PSHMEM_INT4_CSWAP, (FORTRAN_POINTER_T target, MPI_Fint *cond, FORTRAN_POINTER_T value, MPI_Fint *pe));
+PN (ompi_fortran_integer8_t, pshmem_int8_cswap, PSHMEM_INT8_CSWAP, (FORTRAN_POINTER_T target, MPI_Fint *cond, FORTRAN_POINTER_T value, MPI_Fint *pe));
+PN (ompi_fortran_integer4_t, pshmem_int4_fadd, PSHMEM_INT4_FADD, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));
+PN (ompi_fortran_integer8_t, pshmem_int8_fadd, PSHMEM_INT8_FADD, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));
+PN (void, pshmem_int4_inc, PSHMEM_INT4_INC, (FORTRAN_POINTER_T target, MPI_Fint *pe));
+PN (void, pshmem_int8_inc, PSHMEM_INT8_INC, (FORTRAN_POINTER_T target, MPI_Fint *pe));
+PN (ompi_fortran_integer4_t, pshmem_int4_finc, PSHMEM_INT4_FINC, (FORTRAN_POINTER_T target, MPI_Fint *pe));
+PN (ompi_fortran_integer8_t, pshmem_int8_finc, PSHMEM_INT8_FINC, (FORTRAN_POINTER_T target, MPI_Fint *pe));
+PN (void, pshmem_int4_add, PSHMEM_INT4_ADD, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));
+PN (void, pshmem_int8_add, PSHMEM_INT8_ADD, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));
+PN (void, pshmem_int4_wait, PSHMEM_INT4_WAIT, (ompi_fortran_integer4_t *var, ompi_fortran_integer4_t *value));
+PN (void, pshmem_int8_wait, PSHMEM_INT8_WAIT, (ompi_fortran_integer8_t *var, ompi_fortran_integer8_t *value));
+PN (void, pshmem_wait, PSHMEM_WAIT, (MPI_Fint *var, MPI_Fint *value));
+PN (void, pshmem_int4_wait_until, PSHMEM_INT4_WAIT_UNTIL, (ompi_fortran_integer4_t *var, MPI_Fint *cmp, ompi_fortran_integer4_t *value));
+PN (void, pshmem_int8_wait_until, PSHMEM_INT8_WAIT_UNTIL, (ompi_fortran_integer8_t *var, MPI_Fint *cmp, ompi_fortran_integer8_t *value));
+PN (void, pshmem_wait_until, PSHMEM_WAIT_UNTIL, (MPI_Fint *var, MPI_Fint *cmp, MPI_Fint *value));
+PN (void, pshmem_barrier, PSHMEM_BARRIER, (MPI_Fint *PE_start, MPI_Fint *logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int2_and_to_all, PSHMEM_INT2_AND_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int4_and_to_all, PSHMEM_INT4_AND_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int8_and_to_all, PSHMEM_INT8_AND_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int2_or_to_all, PSHMEM_INT2_OR_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int4_or_to_all, PSHMEM_INT4_OR_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int8_or_to_all, PSHMEM_INT8_OR_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int2_xor_to_all, PSHMEM_INT2_XOR_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int4_xor_to_all, PSHMEM_INT4_XOR_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int8_xor_to_all, PSHMEM_INT8_XOR_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_comp4_xor_to_all, PSHMEM_COMP4_XOR_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_comp8_xor_to_all, PSHMEM_COMP8_XOR_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int2_max_to_all, PSHMEM_INT2_MAX_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int4_max_to_all, PSHMEM_INT4_MAX_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int8_max_to_all, PSHMEM_INT8_MAX_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_real4_max_to_all, PSHMEM_REAL4_MAX_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_real8_max_to_all, PSHMEM_REAL8_MAX_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_real16_max_to_all, PSHMEM_REAL16_MAX_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int2_min_to_all, PSHMEM_INT2_MIN_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int4_min_to_all, PSHMEM_INT4_MIN_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int8_min_to_all, PSHMEM_INT8_MIN_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_real4_min_to_all, PSHMEM_REAL4_MIN_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_real8_min_to_all, PSHMEM_REAL8_MIN_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_real16_min_to_all, PSHMEM_REAL16_MIN_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int2_sum_to_all, PSHMEM_INT2_SUM_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int4_sum_to_all, PSHMEM_INT4_SUM_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int8_sum_to_all, PSHMEM_INT8_SUM_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_comp4_sum_to_all, PSHMEM_COMP4_SUM_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_comp8_sum_to_all, PSHMEM_COMP8_SUM_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_real4_sum_to_all, PSHMEM_REAL4_SUM_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_real8_sum_to_all, PSHMEM_REAL8_SUM_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_real16_sum_to_all, PSHMEM_REAL16_SUM_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int2_prod_to_all, PSHMEM_INT2_PROD_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int4_prod_to_all, PSHMEM_INT4_PROD_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_int8_prod_to_all, PSHMEM_INT8_PROD_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, shmem_comp4_prod_to_all, SHMEM_COMP4_PROD_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, shmem_comp8_prod_to_all, SHMEM_COMP8_PROD_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_real4_prod_to_all, PSHMEM_REAL4_PROD_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_real8_prod_to_all, PSHMEM_REAL8_PROD_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_real16_prod_to_all, PSHMEM_REAL16_PROD_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_collect4, PSHMEM_COLLECT4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nlong, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_collect8, PSHMEM_COLLECT8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nlong, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_collect32, PSHMEM_COLLECT32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nlong, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_collect64, PSHMEM_COLLECT64, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nlong, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_fcollect4, PSHMEM_FCOLLECT4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nlong, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_fcollect8, PSHMEM_FCOLLECT8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nlong, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_fcollect32, PSHMEM_FCOLLECT32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nlong, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_fcollect64, PSHMEM_FCOLLECT64, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nlong, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_broadcast4, PSHMEM_BROADCAST4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nlong, MPI_Fint *PE_root, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_broadcast8, PSHMEM_BROADCAST8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nlong, MPI_Fint *PE_root, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_broadcast32, PSHMEM_BROADCAST32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nlong, MPI_Fint *PE_root, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_broadcast64, PSHMEM_BROADCAST64, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nlong, MPI_Fint *PE_root, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync));
+PN (void, pshmem_set_lock, PSHMEM_SET_LOCK, (FORTRAN_POINTER_T lock));
+PN (void, pshmem_clear_lock, PSHMEM_CLEAR_LOCK, (FORTRAN_POINTER_T lock));
+PN (MPI_Fint, pshmem_test_lock, PSHMEM_TEST_LOCK, (FORTRAN_POINTER_T lock));
+PN (void, pshmem_set_cache_inv, PSHMEM_SET_CACHE_INV, (void));
+PN (void, pshmem_set_cache_line_inv, PSHMEM_SET_CACHE_LINE_INV, (FORTRAN_POINTER_T target));
+PN (void, pshmem_clear_cache_inv, PSHMEM_CLEAR_CACHE_INV, (void));
+PN (void, pshmem_clear_cache_line_inv, PSHMEM_CLEAR_CACHE_LINE_INV, (FORTRAN_POINTER_T target));
+PN (void, pshmem_udcflush, PSHMEM_UDCFLUSH, (void));
+PN (void, pshmem_udcflush_line, PSHMEM_UDCFLUSH_LINE, (FORTRAN_POINTER_T target));
+PN (void, pshmem_fence, PSHMEM_FENCE, (void));
+PN (void, pshmem_quiet, PSHMEM_QUIET, (void));
+
+END_C_DECLS
+#endif /*OSHMEM_F77_PROTOTYPES_PSHMEM_H*/
diff --git a/oshmem/shmem/fortran/prototypes_shmem.h b/oshmem/shmem/fortran/prototypes_shmem.h
index 567a3482ff..4670d1ca07 100644
--- a/oshmem/shmem/fortran/prototypes_shmem.h
+++ b/oshmem/shmem/fortran/prototypes_shmem.h
@@ -1,10 +1,10 @@
 /*
- * Copyright (c) 2013      Mellanox Technologies, Inc.
+ * Copyright (c) 2014      Mellanox Technologies, Inc.
  *                         All rights reserved.
  * $COPYRIGHT$
- * 
+ *
  * Additional copyrights may follow
- * 
+ *
  * $HEADER$
  */
 
@@ -19,64 +19,66 @@ BEGIN_C_DECLS
   OSHMEM_DECLSPEC ret lower_name##_f args;    \
   OSHMEM_DECLSPEC ret lower_name##_ args;     \
   OSHMEM_DECLSPEC ret lower_name##__ args;    \
-  OSHMEM_DECLSPEC ret upper_name args        
+  OSHMEM_DECLSPEC ret upper_name args
 
 PN (void, start_pes, START_PES, (MPI_Fint npes));
 PN (MPI_Fint, num_pes, NUM_PES, (void));
 PN (MPI_Fint, my_pe, MY_PE, (void));
+OSHMEM_DECLSPEC MPI_Fint _my_pe_(void);
 PN (void, shmem_finalize, SHMEM_FINALIZE, (void));
 PN (void, shmem_barrier_all, SHMEM_BARRIER_ALL, (void));
-PN (void, shpalloc, SHPALLOC, (FORTRAN_POINTER_T *addr, MPI_Fint *length, MPI_Fint *errcode, MPI_Fint *abort)); 
-PN (void, shpdeallc, SHPDEALLC, (FORTRAN_POINTER_T *addr, MPI_Fint *errcode, MPI_Fint *abort)); 
-PN (void, shpclmove, SHPCLMOVE, (FORTRAN_POINTER_T *addr, MPI_Fint *length, MPI_Fint *status, MPI_Fint *abort)); 
-PN (FORTRAN_POINTER_T*, shmem_ptr, SHMEM_PTR, (FORTRAN_POINTER_T target, MPI_Fint *pe)); 
-PN (ompi_fortran_logical_t, shmem_pe_accessible, SHMEM_PE_ACCESSIBLE, (MPI_Fint *pe)); 
-PN (MPI_Fint, shmem_addr_accessible, SHMEM_ADDR_ACCESSIBLE, (FORTRAN_POINTER_T addr, MPI_Fint *pe)); 
-PN (void, shmem_put, SHMEM_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)); 
-PN (void, shmem_character_put, SHMEM_CHARACTER_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)); 
-PN (void, shmem_complex_put, SHMEM_COMPLEX_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)); 
-PN (void, shmem_double_put, SHMEM_DOUBLE_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)); 
-PN (void, shmem_logical_put, SHMEM_LOGICAL_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)); 
-PN (void, shmem_integer_put, SHMEM_INTEGER_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)); 
-PN (void, shmem_real_put, SHMEM_REAL_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)); 
-PN (void, shmem_put4, SHMEM_PUT4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)); 
-PN (void, shmem_put8, SHMEM_PUT8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)); 
-PN (void, shmem_put32, SHMEM_PUT32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)); 
-PN (void, shmem_put64, SHMEM_PUT64, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)); 
-PN (void, shmem_putmem, SHMEM_PUTMEM, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)); 
-PN (void, shmem_iput4, SHMEM_IPUT4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_iput8, SHMEM_IPUT8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_iput32, SHMEM_IPUT32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_iput64, SHMEM_IPUT64, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_iput128, SHMEM_IPUT128, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_complex_iput, SHMEM_COMPLEX_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_double_iput, SHMEM_DOUBLE_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_integer_iput, SHMEM_INTEGER_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_logical_iput, SHMEM_LOGICAL_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_real_iput, SHMEM_REAL_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_character_get, SHMEM_CHARACTER_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_complex_get, SHMEM_COMPLEX_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_double_get, SHMEM_DOUBLE_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_integer_get, SHMEM_INTEGER_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_get4, SHMEM_GET4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_get8, SHMEM_GET8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_get32, SHMEM_GET32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_get64, SHMEM_GET64, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_get128, SHMEM_GET128, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_getmem, SHMEM_GETMEM, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_logical_get, SHMEM_LOGICAL_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_real_get, SHMEM_REAL_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_iget4, SHMEM_IGET4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_iget8, SHMEM_IGET8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_iget32, SHMEM_IGET32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_iget64, SHMEM_IGET64, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_iget128, SHMEM_IGET128, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_complex_iget, SHMEM_COMPLEX_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_double_iget, SHMEM_DOUBLE_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_integer_iget, SHMEM_INTEGER_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_logical_iget, SHMEM_LOGICAL_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (void, shmem_real_iget, SHMEM_REAL_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe)); 
-PN (MPI_Fint, shmem_swap, SHMEM_SWAP, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe)); 
+PN (void, shpalloc, SHPALLOC, (FORTRAN_POINTER_T *addr, MPI_Fint *length, MPI_Fint *errcode, MPI_Fint *abort));
+PN (void, shpdeallc, SHPDEALLC, (FORTRAN_POINTER_T *addr, MPI_Fint *errcode, MPI_Fint *abort));
+PN (void, shpclmove, SHPCLMOVE, (FORTRAN_POINTER_T *addr, MPI_Fint *length, MPI_Fint *status, MPI_Fint *abort));
+PN (FORTRAN_POINTER_T*, shmem_ptr, SHMEM_PTR, (FORTRAN_POINTER_T target, MPI_Fint *pe));
+PN (ompi_fortran_logical_t, shmem_pe_accessible, SHMEM_PE_ACCESSIBLE, (MPI_Fint *pe));
+PN (MPI_Fint, shmem_addr_accessible, SHMEM_ADDR_ACCESSIBLE, (FORTRAN_POINTER_T addr, MPI_Fint *pe));
+PN (void, shmem_put, SHMEM_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_character_put, SHMEM_CHARACTER_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_complex_put, SHMEM_COMPLEX_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_double_put, SHMEM_DOUBLE_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_logical_put, SHMEM_LOGICAL_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_integer_put, SHMEM_INTEGER_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_real_put, SHMEM_REAL_PUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_put4, SHMEM_PUT4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_put8, SHMEM_PUT8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_put32, SHMEM_PUT32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_put64, SHMEM_PUT64, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_put128, SHMEM_PUT128, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_putmem, SHMEM_PUTMEM, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe));
+PN (void, shmem_iput4, SHMEM_IPUT4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_iput8, SHMEM_IPUT8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_iput32, SHMEM_IPUT32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_iput64, SHMEM_IPUT64, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_iput128, SHMEM_IPUT128, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_complex_iput, SHMEM_COMPLEX_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_double_iput, SHMEM_DOUBLE_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_integer_iput, SHMEM_INTEGER_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_logical_iput, SHMEM_LOGICAL_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_real_iput, SHMEM_REAL_IPUT, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_character_get, SHMEM_CHARACTER_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_complex_get, SHMEM_COMPLEX_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_double_get, SHMEM_DOUBLE_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_integer_get, SHMEM_INTEGER_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_get4, SHMEM_GET4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_get8, SHMEM_GET8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_get32, SHMEM_GET32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_get64, SHMEM_GET64, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_get128, SHMEM_GET128, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_getmem, SHMEM_GETMEM, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_logical_get, SHMEM_LOGICAL_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_real_get, SHMEM_REAL_GET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_iget4, SHMEM_IGET4, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_iget8, SHMEM_IGET8, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_iget32, SHMEM_IGET32, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_iget64, SHMEM_IGET64, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_iget128, SHMEM_IGET128, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_complex_iget, SHMEM_COMPLEX_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_double_iget, SHMEM_DOUBLE_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_integer_iget, SHMEM_INTEGER_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_logical_iget, SHMEM_LOGICAL_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (void, shmem_real_iget, SHMEM_REAL_IGET, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *tst, MPI_Fint *sst, MPI_Fint *len, MPI_Fint *pe));
+PN (MPI_Fint, shmem_swap, SHMEM_SWAP, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));
 PN (ompi_fortran_integer4_t, shmem_int4_swap, SHMEM_INT4_SWAP, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));
 PN (ompi_fortran_integer8_t, shmem_int8_swap, SHMEM_INT8_SWAP, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));
 PN (ompi_fortran_real4_t, shmem_real4_swap, SHMEM_REAL4_SWAP, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T value, MPI_Fint *pe));
@@ -98,31 +100,42 @@ PN (void, shmem_int4_wait_until, SHMEM_INT4_WAIT_UNTIL, (ompi_fortran_integer4_t
 PN (void, shmem_int8_wait_until, SHMEM_INT8_WAIT_UNTIL, (ompi_fortran_integer8_t *var, MPI_Fint *cmp, ompi_fortran_integer8_t *value));
 PN (void, shmem_wait_until, SHMEM_WAIT_UNTIL, (MPI_Fint *var, MPI_Fint *cmp, MPI_Fint *value));
 PN (void, shmem_barrier, SHMEM_BARRIER, (MPI_Fint *PE_start, MPI_Fint *logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T pSync));
+PN (void, shmem_int2_and_to_all, SHMEM_INT2_AND_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_int4_and_to_all, SHMEM_INT4_AND_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_int8_and_to_all, SHMEM_INT8_AND_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, shmem_int2_or_to_all, SHMEM_INT2_OR_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_int4_or_to_all, SHMEM_INT4_OR_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_int8_or_to_all, SHMEM_INT8_OR_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, shmem_int2_xor_to_all, SHMEM_INT2_XOR_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_int4_xor_to_all, SHMEM_INT4_XOR_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_int8_xor_to_all, SHMEM_INT8_XOR_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_comp4_xor_to_all, SHMEM_COMP4_XOR_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_comp8_xor_to_all, SHMEM_COMP8_XOR_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, shmem_int2_max_to_all, SHMEM_INT2_MAX_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_int4_max_to_all, SHMEM_INT4_MAX_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_int8_max_to_all, SHMEM_INT8_MAX_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_real4_max_to_all, SHMEM_REAL4_MAX_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_real8_max_to_all, SHMEM_REAL8_MAX_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_real16_max_to_all, SHMEM_REAL16_MAX_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, shmem_int2_min_to_all, SHMEM_INT2_MIN_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_int4_min_to_all, SHMEM_INT4_MIN_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_int8_min_to_all, SHMEM_INT8_MIN_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_real4_min_to_all, SHMEM_REAL4_MIN_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_real8_min_to_all, SHMEM_REAL8_MIN_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_real16_min_to_all, SHMEM_REAL16_MIN_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, shmem_int2_sum_to_all, SHMEM_INT2_SUM_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_int4_sum_to_all, SHMEM_INT4_SUM_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_int8_sum_to_all, SHMEM_INT8_SUM_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, shmem_comp4_sum_to_all, SHMEM_COMP4_SUM_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, shmem_comp8_sum_to_all, SHMEM_COMP8_SUM_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_real4_sum_to_all, SHMEM_REAL4_SUM_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_real8_sum_to_all, SHMEM_REAL8_SUM_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_real16_sum_to_all, SHMEM_REAL16_SUM_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, shmem_int2_prod_to_all, SHMEM_INT2_PROD_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_int4_prod_to_all, SHMEM_INT4_PROD_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_int8_prod_to_all, SHMEM_INT8_PROD_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, shmem_comp4_prod_to_all, SHMEM_COMP4_PROD_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
+PN (void, shmem_comp8_prod_to_all, SHMEM_COMP8_PROD_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_real4_prod_to_all, SHMEM_REAL4_PROD_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_real8_prod_to_all, SHMEM_REAL8_PROD_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
 PN (void, shmem_real16_prod_to_all, SHMEM_REAL16_PROD_TO_ALL, (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync));
@@ -151,4 +164,4 @@ PN (void, shmem_fence, SHMEM_FENCE, (void));
 PN (void, shmem_quiet, SHMEM_QUIET, (void));
 
 END_C_DECLS
-#endif /*OSHMEM_F77_PROTOTYPES_SHMEM_H*/ 
+#endif /*OSHMEM_F77_PROTOTYPES_SHMEM_H*/
diff --git a/oshmem/shmem/fortran/shmem_addr_accessible_f.c b/oshmem/shmem/fortran/shmem_addr_accessible_f.c
index 0e947841d2..0e4eb8c437 100644
--- a/oshmem/shmem/fortran/shmem_addr_accessible_f.c
+++ b/oshmem/shmem/fortran/shmem_addr_accessible_f.c
@@ -15,6 +15,12 @@
 #include "oshmem/shmem/shmem_api_logger.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_ADDR_ACCESSIBLE, shmem_addr_accessible)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_FUNCTION (MPI_Fint,
         SHMEM_ADDR_ACCESSIBLE,
         shmem_addr_accessible_,
diff --git a/oshmem/shmem/fortran/shmem_and_to_all_f.c b/oshmem/shmem/fortran/shmem_and_to_all_f.c
index 8e86b2830c..3803b25e0e 100644
--- a/oshmem/shmem/fortran/shmem_and_to_all_f.c
+++ b/oshmem/shmem/fortran/shmem_and_to_all_f.c
@@ -18,6 +18,22 @@
 #include "oshmem/proc/proc_group_cache.h"
 #include "oshmem/op/op.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT2_AND_TO_ALL, shmem_int2_and_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT4_AND_TO_ALL, shmem_int4_and_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT8_AND_TO_ALL, shmem_int8_and_to_all)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_INT2_AND_TO_ALL,
+        shmem_int2_and_to_all_,
+        shmem_int2_and_to_all__,
+        shmem_int2_and_to_all_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
+        (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INT4_AND_TO_ALL,
         shmem_int4_and_to_all_,
@@ -98,5 +114,6 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
     }\
 }
 
+SHMEM_AND_TO_ALL(shmem_int2_and_to_all_f, oshmem_op_and_fint2, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_AND_TO_ALL(shmem_int4_and_to_all_f, oshmem_op_and_fint4, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_AND_TO_ALL(shmem_int8_and_to_all_f, oshmem_op_and_fint8, OSHMEM_GROUP_CACHE_ENABLED)
diff --git a/oshmem/shmem/fortran/shmem_barrier_all_f.c b/oshmem/shmem/fortran/shmem_barrier_all_f.c
index 4f43b0af30..1798eb530e 100644
--- a/oshmem/shmem/fortran/shmem_barrier_all_f.c
+++ b/oshmem/shmem/fortran/shmem_barrier_all_f.c
@@ -13,6 +13,12 @@
 #include "oshmem/shmem/fortran/bindings.h"
 #include "oshmem/include/shmem.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_BARRIER_ALL, shmem_barrier_all)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_BARRIER_ALL,
         shmem_barrier_all_,
diff --git a/oshmem/shmem/fortran/shmem_barrier_f.c b/oshmem/shmem/fortran/shmem_barrier_f.c
index e6cbee4440..ceea64b154 100644
--- a/oshmem/shmem/fortran/shmem_barrier_f.c
+++ b/oshmem/shmem/fortran/shmem_barrier_f.c
@@ -13,6 +13,12 @@
 #include "oshmem/shmem/fortran/bindings.h"
 #include "oshmem/include/shmem.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_BARRIER, shmem_barrier)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_BARRIER,
         shmem_barrier_,
diff --git a/oshmem/shmem/fortran/shmem_broadcast_f.c b/oshmem/shmem/fortran/shmem_broadcast_f.c
index 24bcc80061..9ef8104478 100644
--- a/oshmem/shmem/fortran/shmem_broadcast_f.c
+++ b/oshmem/shmem/fortran/shmem_broadcast_f.c
@@ -18,6 +18,15 @@
 #include "oshmem/proc/proc_group_cache.h"
 #include "oshmem/op/op.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_BROADCAST4, shmem_broadcast4)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_BROADCAST8, shmem_broadcast8)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_BROADCAST32, shmem_broadcast32)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_BROADCAST64, shmem_broadcast64)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_BROADCAST4,
         shmem_broadcast4_,
diff --git a/oshmem/shmem/fortran/shmem_cache_f.c b/oshmem/shmem/fortran/shmem_cache_f.c
index 66d0fe4dc0..e2eab6ab88 100644
--- a/oshmem/shmem/fortran/shmem_cache_f.c
+++ b/oshmem/shmem/fortran/shmem_cache_f.c
@@ -15,6 +15,17 @@
 #include "oshmem/shmem/shmem_api_logger.h"
 #include "ompi/datatype/ompi_datatype.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_SET_CACHE_INV, shmem_set_cache_inv)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_SET_CACHE_LINE_INV, shmem_set_cache_line_inv)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_CLEAR_CACHE_INV, shmem_clear_cache_inv)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_CLEAR_CACHE_LINE_INV, shmem_clear_cache_line_inv)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_UDCFLUSH, shmem_udcflush)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_UDCFLUSH_LINE, shmem_udcflush_line)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_SET_CACHE_INV,
         shmem_set_cache_inv_,
diff --git a/oshmem/shmem/fortran/shmem_character_get_f.c b/oshmem/shmem/fortran/shmem_character_get_f.c
index 28aca18f6c..c43865d94e 100644
--- a/oshmem/shmem/fortran/shmem_character_get_f.c
+++ b/oshmem/shmem/fortran/shmem_character_get_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_CHARACTER_GET, shmem_character_get)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_CHARACTER_GET,
         shmem_character_get_,
diff --git a/oshmem/shmem/fortran/shmem_character_put_f.c b/oshmem/shmem/fortran/shmem_character_put_f.c
index 85536f4753..76bd89215c 100644
--- a/oshmem/shmem/fortran/shmem_character_put_f.c
+++ b/oshmem/shmem/fortran/shmem_character_put_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_CHARACTER_PUT, shmem_character_put)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_CHARACTER_PUT,
         shmem_character_put_,
diff --git a/oshmem/shmem/fortran/shmem_collect_f.c b/oshmem/shmem/fortran/shmem_collect_f.c
index 62549eb403..7927267a5c 100644
--- a/oshmem/shmem/fortran/shmem_collect_f.c
+++ b/oshmem/shmem/fortran/shmem_collect_f.c
@@ -18,6 +18,19 @@
 #include "oshmem/proc/proc_group_cache.h"
 #include "oshmem/op/op.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_COLLECT4, shmem_collect4)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_COLLECT8, shmem_collect8)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_COLLECT32, shmem_collect32)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_COLLECT64, shmem_collect64)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_FCOLLECT4, shmem_fcollect4)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_FCOLLECT8, shmem_fcollect8)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_FCOLLECT32, shmem_fcollect32)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_FCOLLECT64, shmem_fcollect64)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_COLLECT4,
         shmem_collect4_,
diff --git a/oshmem/shmem/fortran/shmem_complex_get_f.c b/oshmem/shmem/fortran/shmem_complex_get_f.c
index be9a744efc..ba589cd2c2 100644
--- a/oshmem/shmem/fortran/shmem_complex_get_f.c
+++ b/oshmem/shmem/fortran/shmem_complex_get_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_COMPLEX_GET, shmem_complex_get)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_COMPLEX_GET,
         shmem_complex_get_,
diff --git a/oshmem/shmem/fortran/shmem_complex_iget_f.c b/oshmem/shmem/fortran/shmem_complex_iget_f.c
index 05da4a09f5..b8a1e75da6 100644
--- a/oshmem/shmem/fortran/shmem_complex_iget_f.c
+++ b/oshmem/shmem/fortran/shmem_complex_iget_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_COMPLEX_IGET, shmem_complex_iget)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_COMPLEX_IGET,
         shmem_complex_iget_,
diff --git a/oshmem/shmem/fortran/shmem_complex_iput_f.c b/oshmem/shmem/fortran/shmem_complex_iput_f.c
index 5cfa4fbfe1..095abb64f8 100644
--- a/oshmem/shmem/fortran/shmem_complex_iput_f.c
+++ b/oshmem/shmem/fortran/shmem_complex_iput_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_COMPLEX_IPUT, shmem_complex_iput)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_COMPLEX_IPUT,
         shmem_complex_iput_,
diff --git a/oshmem/shmem/fortran/shmem_complex_put_f.c b/oshmem/shmem/fortran/shmem_complex_put_f.c
index 1f09c6e839..228838264e 100644
--- a/oshmem/shmem/fortran/shmem_complex_put_f.c
+++ b/oshmem/shmem/fortran/shmem_complex_put_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_COMPLEX_PUT, shmem_complex_put)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_COMPLEX_PUT,
         shmem_complex_put_,
diff --git a/oshmem/shmem/fortran/shmem_double_get_f.c b/oshmem/shmem/fortran/shmem_double_get_f.c
index 022796e3a8..cd7b998ced 100644
--- a/oshmem/shmem/fortran/shmem_double_get_f.c
+++ b/oshmem/shmem/fortran/shmem_double_get_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_DOUBLE_GET, shmem_double_get)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_DOUBLE_GET,
         shmem_double_get_,
diff --git a/oshmem/shmem/fortran/shmem_double_iget_f.c b/oshmem/shmem/fortran/shmem_double_iget_f.c
index 21a6eb0c6a..2655ac1a1e 100644
--- a/oshmem/shmem/fortran/shmem_double_iget_f.c
+++ b/oshmem/shmem/fortran/shmem_double_iget_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_DOUBLE_IGET, shmem_double_iget)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_DOUBLE_IGET,
         shmem_double_iget_,
diff --git a/oshmem/shmem/fortran/shmem_double_iput_f.c b/oshmem/shmem/fortran/shmem_double_iput_f.c
index 48cda8b53d..7c20b5be21 100644
--- a/oshmem/shmem/fortran/shmem_double_iput_f.c
+++ b/oshmem/shmem/fortran/shmem_double_iput_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_DOUBLE_IPUT, shmem_double_iput)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_DOUBLE_IPUT,
         shmem_double_iput_,
diff --git a/oshmem/shmem/fortran/shmem_double_put_f.c b/oshmem/shmem/fortran/shmem_double_put_f.c
index b904ec5b79..ee74987bca 100644
--- a/oshmem/shmem/fortran/shmem_double_put_f.c
+++ b/oshmem/shmem/fortran/shmem_double_put_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_DOUBLE_PUT, shmem_double_put)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_DOUBLE_PUT,
         shmem_double_put_,
diff --git a/oshmem/shmem/fortran/shmem_fence_f.c b/oshmem/shmem/fortran/shmem_fence_f.c
index d2abe4e281..f73856f401 100644
--- a/oshmem/shmem/fortran/shmem_fence_f.c
+++ b/oshmem/shmem/fortran/shmem_fence_f.c
@@ -14,6 +14,12 @@
 #include "oshmem/include/shmem.h"
 #include "oshmem/mca/spml/spml.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_FENCE, shmem_fence)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_FENCE,
         shmem_fence_,
diff --git a/oshmem/shmem/fortran/shmem_get128_f.c b/oshmem/shmem/fortran/shmem_get128_f.c
index 2d05edf535..51c94c1a44 100644
--- a/oshmem/shmem/fortran/shmem_get128_f.c
+++ b/oshmem/shmem/fortran/shmem_get128_f.c
@@ -17,6 +17,12 @@
 #include "oshmem/mca/spml/spml.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_GET128, shmem_get128)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_GET128,
         shmem_get128_,
diff --git a/oshmem/shmem/fortran/shmem_get32_f.c b/oshmem/shmem/fortran/shmem_get32_f.c
index c973455106..a54f5dc85b 100644
--- a/oshmem/shmem/fortran/shmem_get32_f.c
+++ b/oshmem/shmem/fortran/shmem_get32_f.c
@@ -17,6 +17,12 @@
 #include "oshmem/mca/spml/spml.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_GET32, shmem_get32)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_GET32,
         shmem_get32_,
diff --git a/oshmem/shmem/fortran/shmem_get4_f.c b/oshmem/shmem/fortran/shmem_get4_f.c
index e6367524d1..8ffc98bf0a 100644
--- a/oshmem/shmem/fortran/shmem_get4_f.c
+++ b/oshmem/shmem/fortran/shmem_get4_f.c
@@ -17,6 +17,12 @@
 #include "oshmem/mca/spml/spml.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_GET4, shmem_get4)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_GET4,
         shmem_get4_,
diff --git a/oshmem/shmem/fortran/shmem_get64_f.c b/oshmem/shmem/fortran/shmem_get64_f.c
index 3a9d576404..9627052225 100644
--- a/oshmem/shmem/fortran/shmem_get64_f.c
+++ b/oshmem/shmem/fortran/shmem_get64_f.c
@@ -17,6 +17,12 @@
 #include "oshmem/mca/spml/spml.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_GET64, shmem_get64)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_GET64,
         shmem_get64_,
diff --git a/oshmem/shmem/fortran/shmem_get8_f.c b/oshmem/shmem/fortran/shmem_get8_f.c
index 9eafbcdbe9..d24a74045b 100644
--- a/oshmem/shmem/fortran/shmem_get8_f.c
+++ b/oshmem/shmem/fortran/shmem_get8_f.c
@@ -17,6 +17,12 @@
 #include "oshmem/mca/spml/spml.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_GET8, shmem_get8)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_GET8,
         shmem_get8_,
diff --git a/oshmem/shmem/fortran/shmem_getmem_f.c b/oshmem/shmem/fortran/shmem_getmem_f.c
index b1f1bbf7df..fae84ee1b3 100644
--- a/oshmem/shmem/fortran/shmem_getmem_f.c
+++ b/oshmem/shmem/fortran/shmem_getmem_f.c
@@ -17,6 +17,12 @@
 #include "oshmem/mca/spml/spml.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_GETMEM, shmem_getmem)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_GETMEM,
         shmem_getmem_,
diff --git a/oshmem/shmem/fortran/shmem_iget128_f.c b/oshmem/shmem/fortran/shmem_iget128_f.c
index 47393c95a1..a80161bf42 100644
--- a/oshmem/shmem/fortran/shmem_iget128_f.c
+++ b/oshmem/shmem/fortran/shmem_iget128_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_IGET128, shmem_iget128)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_IGET128,
         shmem_iget128_,
diff --git a/oshmem/shmem/fortran/shmem_iget32_f.c b/oshmem/shmem/fortran/shmem_iget32_f.c
index 9c7e5be9bf..ddc15fbba1 100644
--- a/oshmem/shmem/fortran/shmem_iget32_f.c
+++ b/oshmem/shmem/fortran/shmem_iget32_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_IGET32, shmem_iget32)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_IGET32,
         shmem_iget32_,
diff --git a/oshmem/shmem/fortran/shmem_iget4_f.c b/oshmem/shmem/fortran/shmem_iget4_f.c
index 383f8476ed..4f64952efd 100644
--- a/oshmem/shmem/fortran/shmem_iget4_f.c
+++ b/oshmem/shmem/fortran/shmem_iget4_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_IGET4, shmem_iget4)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_IGET4,
         shmem_iget4_,
diff --git a/oshmem/shmem/fortran/shmem_iget64_f.c b/oshmem/shmem/fortran/shmem_iget64_f.c
index 03c60ea38b..eff5d3cb59 100644
--- a/oshmem/shmem/fortran/shmem_iget64_f.c
+++ b/oshmem/shmem/fortran/shmem_iget64_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_IGET64, shmem_iget64)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_IGET64,
         shmem_iget64_,
diff --git a/oshmem/shmem/fortran/shmem_iget8_f.c b/oshmem/shmem/fortran/shmem_iget8_f.c
index 0c9cc34d12..4f15841f26 100644
--- a/oshmem/shmem/fortran/shmem_iget8_f.c
+++ b/oshmem/shmem/fortran/shmem_iget8_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_IGET8, shmem_iget8)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_IGET8,
         shmem_iget8_,
diff --git a/oshmem/shmem/fortran/shmem_int4_add_f.c b/oshmem/shmem/fortran/shmem_int4_add_f.c
index 5cd1aa32c3..82fb5fe611 100644
--- a/oshmem/shmem/fortran/shmem_int4_add_f.c
+++ b/oshmem/shmem/fortran/shmem_int4_add_f.c
@@ -19,6 +19,12 @@
 #include "oshmem/op/op.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT4_ADD, shmem_int4_add)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INT4_ADD,
         shmem_int4_add_,
diff --git a/oshmem/shmem/fortran/shmem_int4_cswap_f.c b/oshmem/shmem/fortran/shmem_int4_cswap_f.c
index 6a0fda8bb0..2a8ac75227 100644
--- a/oshmem/shmem/fortran/shmem_int4_cswap_f.c
+++ b/oshmem/shmem/fortran/shmem_int4_cswap_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT4_CSWAP, shmem_int4_cswap)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_FUNCTION (ompi_fortran_integer4_t,
         SHMEM_INT4_CSWAP,
         shmem_int4_cswap_,
diff --git a/oshmem/shmem/fortran/shmem_int4_fadd_f.c b/oshmem/shmem/fortran/shmem_int4_fadd_f.c
index 17e44c7d8a..f962ebb8f3 100644
--- a/oshmem/shmem/fortran/shmem_int4_fadd_f.c
+++ b/oshmem/shmem/fortran/shmem_int4_fadd_f.c
@@ -19,6 +19,12 @@
 #include "oshmem/op/op.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT4_FADD, shmem_int4_fadd)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_FUNCTION (ompi_fortran_integer4_t,
         SHMEM_INT4_FADD,
         shmem_int4_fadd_,
diff --git a/oshmem/shmem/fortran/shmem_int4_finc_f.c b/oshmem/shmem/fortran/shmem_int4_finc_f.c
index 9bf5165422..2d1ff2419d 100644
--- a/oshmem/shmem/fortran/shmem_int4_finc_f.c
+++ b/oshmem/shmem/fortran/shmem_int4_finc_f.c
@@ -19,6 +19,12 @@
 #include "oshmem/op/op.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT4_FINC, shmem_int4_finc)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_FUNCTION (ompi_fortran_integer4_t,
         SHMEM_INT4_FINC,
         shmem_int4_finc_,
diff --git a/oshmem/shmem/fortran/shmem_int4_inc_f.c b/oshmem/shmem/fortran/shmem_int4_inc_f.c
index b48e979531..24f58e3bb8 100644
--- a/oshmem/shmem/fortran/shmem_int4_inc_f.c
+++ b/oshmem/shmem/fortran/shmem_int4_inc_f.c
@@ -19,6 +19,12 @@
 #include "oshmem/op/op.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT4_INC, shmem_int4_inc)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INT4_INC,
         shmem_int4_inc_,
diff --git a/oshmem/shmem/fortran/shmem_int4_swap_f.c b/oshmem/shmem/fortran/shmem_int4_swap_f.c
index c27de698fa..806bb4629e 100644
--- a/oshmem/shmem/fortran/shmem_int4_swap_f.c
+++ b/oshmem/shmem/fortran/shmem_int4_swap_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT4_SWAP, shmem_int4_swap)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_FUNCTION (ompi_fortran_integer4_t,
         SHMEM_INT4_SWAP,
         shmem_int4_swap_,
diff --git a/oshmem/shmem/fortran/shmem_int4_wait_f.c b/oshmem/shmem/fortran/shmem_int4_wait_f.c
index 8611699593..c128251de6 100644
--- a/oshmem/shmem/fortran/shmem_int4_wait_f.c
+++ b/oshmem/shmem/fortran/shmem_int4_wait_f.c
@@ -16,6 +16,12 @@
 #include "oshmem/mca/spml/spml.h"
 #include "ompi/datatype/ompi_datatype.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT4_WAIT, shmem_int4_wait)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INT4_WAIT,
         shmem_int4_wait_,
diff --git a/oshmem/shmem/fortran/shmem_int4_wait_until_f.c b/oshmem/shmem/fortran/shmem_int4_wait_until_f.c
index b1a07024e2..cac9b52ad4 100644
--- a/oshmem/shmem/fortran/shmem_int4_wait_until_f.c
+++ b/oshmem/shmem/fortran/shmem_int4_wait_until_f.c
@@ -16,6 +16,12 @@
 #include "oshmem/mca/spml/spml.h"
 #include "ompi/datatype/ompi_datatype.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT4_WAIT_UNTIL, shmem_int4_wait_until)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INT4_WAIT_UNTIL,
         shmem_int4_wait_until_,
diff --git a/oshmem/shmem/fortran/shmem_int8_add_f.c b/oshmem/shmem/fortran/shmem_int8_add_f.c
index 36bd945d7f..e2d701f4e6 100644
--- a/oshmem/shmem/fortran/shmem_int8_add_f.c
+++ b/oshmem/shmem/fortran/shmem_int8_add_f.c
@@ -19,6 +19,12 @@
 #include "oshmem/op/op.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT8_ADD, shmem_int8_add)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INT8_ADD,
         shmem_int8_add_,
diff --git a/oshmem/shmem/fortran/shmem_int8_cswap_f.c b/oshmem/shmem/fortran/shmem_int8_cswap_f.c
index 07a47a9ebf..9ef5da984d 100644
--- a/oshmem/shmem/fortran/shmem_int8_cswap_f.c
+++ b/oshmem/shmem/fortran/shmem_int8_cswap_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT8_CSWAP, shmem_int8_cswap)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_FUNCTION (ompi_fortran_integer8_t,
         SHMEM_INT8_CSWAP,
         shmem_int8_cswap_,
diff --git a/oshmem/shmem/fortran/shmem_int8_fadd_f.c b/oshmem/shmem/fortran/shmem_int8_fadd_f.c
index c643d3c5f7..61c75303d0 100644
--- a/oshmem/shmem/fortran/shmem_int8_fadd_f.c
+++ b/oshmem/shmem/fortran/shmem_int8_fadd_f.c
@@ -19,6 +19,12 @@
 #include "oshmem/op/op.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT8_FADD, shmem_int8_fadd)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_FUNCTION (ompi_fortran_integer8_t,
         SHMEM_INT8_FADD,
         shmem_int8_fadd_,
diff --git a/oshmem/shmem/fortran/shmem_int8_finc_f.c b/oshmem/shmem/fortran/shmem_int8_finc_f.c
index 8afeed32db..62b2596a6d 100644
--- a/oshmem/shmem/fortran/shmem_int8_finc_f.c
+++ b/oshmem/shmem/fortran/shmem_int8_finc_f.c
@@ -19,6 +19,12 @@
 #include "oshmem/op/op.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT8_FINC, shmem_int8_finc)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_FUNCTION (ompi_fortran_integer8_t,
         SHMEM_INT8_FINC,
         shmem_int8_finc_,
diff --git a/oshmem/shmem/fortran/shmem_int8_inc_f.c b/oshmem/shmem/fortran/shmem_int8_inc_f.c
index a2fd42b408..5198a7fc69 100644
--- a/oshmem/shmem/fortran/shmem_int8_inc_f.c
+++ b/oshmem/shmem/fortran/shmem_int8_inc_f.c
@@ -19,6 +19,12 @@
 #include "oshmem/op/op.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT8_INC, shmem_int8_inc)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INT8_INC,
         shmem_int8_inc_,
diff --git a/oshmem/shmem/fortran/shmem_int8_swap_f.c b/oshmem/shmem/fortran/shmem_int8_swap_f.c
index 727f92eb95..55d8d43020 100644
--- a/oshmem/shmem/fortran/shmem_int8_swap_f.c
+++ b/oshmem/shmem/fortran/shmem_int8_swap_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT8_SWAP, shmem_int8_swap)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_FUNCTION (ompi_fortran_integer8_t,
         SHMEM_INT8_SWAP,
         shmem_int8_swap_,
diff --git a/oshmem/shmem/fortran/shmem_int8_wait_f.c b/oshmem/shmem/fortran/shmem_int8_wait_f.c
index 9b488468d0..4cbd853f17 100644
--- a/oshmem/shmem/fortran/shmem_int8_wait_f.c
+++ b/oshmem/shmem/fortran/shmem_int8_wait_f.c
@@ -16,6 +16,12 @@
 #include "oshmem/mca/spml/spml.h"
 #include "ompi/datatype/ompi_datatype.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT8_WAIT, shmem_int8_wait)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INT8_WAIT,
         shmem_int8_wait_,
diff --git a/oshmem/shmem/fortran/shmem_int8_wait_until_f.c b/oshmem/shmem/fortran/shmem_int8_wait_until_f.c
index f28e8f1e35..cf23f85750 100644
--- a/oshmem/shmem/fortran/shmem_int8_wait_until_f.c
+++ b/oshmem/shmem/fortran/shmem_int8_wait_until_f.c
@@ -16,6 +16,12 @@
 #include "oshmem/mca/spml/spml.h"
 #include "ompi/datatype/ompi_datatype.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT8_WAIT_UNTIL, shmem_int8_wait_until)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INT8_WAIT_UNTIL,
         shmem_int8_wait_until_,
diff --git a/oshmem/shmem/fortran/shmem_integer_get_f.c b/oshmem/shmem/fortran/shmem_integer_get_f.c
index 5c09de8986..d140d37889 100644
--- a/oshmem/shmem/fortran/shmem_integer_get_f.c
+++ b/oshmem/shmem/fortran/shmem_integer_get_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INTEGER_GET, shmem_integer_get)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INTEGER_GET,
         shmem_integer_get_,
diff --git a/oshmem/shmem/fortran/shmem_integer_iget_f.c b/oshmem/shmem/fortran/shmem_integer_iget_f.c
index 9ff1ea36f9..ac4e67a042 100644
--- a/oshmem/shmem/fortran/shmem_integer_iget_f.c
+++ b/oshmem/shmem/fortran/shmem_integer_iget_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INTEGER_IGET, shmem_integer_iget)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INTEGER_IGET,
         shmem_integer_iget_,
diff --git a/oshmem/shmem/fortran/shmem_integer_iput_f.c b/oshmem/shmem/fortran/shmem_integer_iput_f.c
index 30b11730e9..823459136e 100644
--- a/oshmem/shmem/fortran/shmem_integer_iput_f.c
+++ b/oshmem/shmem/fortran/shmem_integer_iput_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INTEGER_IPUT, shmem_integer_iput)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INTEGER_IPUT,
         shmem_integer_iput_,
diff --git a/oshmem/shmem/fortran/shmem_integer_put_f.c b/oshmem/shmem/fortran/shmem_integer_put_f.c
index 8fad6cfc8e..38de14bba8 100644
--- a/oshmem/shmem/fortran/shmem_integer_put_f.c
+++ b/oshmem/shmem/fortran/shmem_integer_put_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INTEGER_PUT, shmem_integer_put)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INTEGER_PUT,
         shmem_integer_put_,
diff --git a/oshmem/shmem/fortran/shmem_iput128_f.c b/oshmem/shmem/fortran/shmem_iput128_f.c
index 822f1d9fb7..c9a98fa4ea 100644
--- a/oshmem/shmem/fortran/shmem_iput128_f.c
+++ b/oshmem/shmem/fortran/shmem_iput128_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_IPUT128, shmem_iput128)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_IPUT128,
         shmem_iput128_,
diff --git a/oshmem/shmem/fortran/shmem_iput32_f.c b/oshmem/shmem/fortran/shmem_iput32_f.c
index 6b5d033644..574f205d5d 100644
--- a/oshmem/shmem/fortran/shmem_iput32_f.c
+++ b/oshmem/shmem/fortran/shmem_iput32_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_IPUT32, shmem_iput32)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_IPUT32,
         shmem_iput32_,
diff --git a/oshmem/shmem/fortran/shmem_iput4_f.c b/oshmem/shmem/fortran/shmem_iput4_f.c
index c53557a5be..db3a603761 100644
--- a/oshmem/shmem/fortran/shmem_iput4_f.c
+++ b/oshmem/shmem/fortran/shmem_iput4_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_IPUT4, shmem_iput4)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_IPUT4,
         shmem_iput4_,
diff --git a/oshmem/shmem/fortran/shmem_iput64_f.c b/oshmem/shmem/fortran/shmem_iput64_f.c
index 35b1606b6a..d6e534599b 100644
--- a/oshmem/shmem/fortran/shmem_iput64_f.c
+++ b/oshmem/shmem/fortran/shmem_iput64_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_IPUT64, shmem_iput64)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_IPUT64,
         shmem_iput64_,
diff --git a/oshmem/shmem/fortran/shmem_iput8_f.c b/oshmem/shmem/fortran/shmem_iput8_f.c
index b7439d47a0..649c76a08d 100644
--- a/oshmem/shmem/fortran/shmem_iput8_f.c
+++ b/oshmem/shmem/fortran/shmem_iput8_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_IPUT8, shmem_iput8)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_IPUT8,
         shmem_iput8_,
diff --git a/oshmem/shmem/fortran/shmem_lock_f.c b/oshmem/shmem/fortran/shmem_lock_f.c
index 23b1827f6f..e7b7620fa8 100644
--- a/oshmem/shmem/fortran/shmem_lock_f.c
+++ b/oshmem/shmem/fortran/shmem_lock_f.c
@@ -16,6 +16,14 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "oshmem/shmem/shmem_lock.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_SET_LOCK, shmem_set_lock)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_CLEAR_LOCK, shmem_clear_lock)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_TEST_LOCK, shmem_test_lock)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_SET_LOCK,
         shmem_set_lock_,
diff --git a/oshmem/shmem/fortran/shmem_logical_get_f.c b/oshmem/shmem/fortran/shmem_logical_get_f.c
index 6564d67559..fd5a005e67 100644
--- a/oshmem/shmem/fortran/shmem_logical_get_f.c
+++ b/oshmem/shmem/fortran/shmem_logical_get_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_LOGICAL_GET, shmem_logical_get)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_LOGICAL_GET,
         shmem_logical_get_,
diff --git a/oshmem/shmem/fortran/shmem_logical_iget_f.c b/oshmem/shmem/fortran/shmem_logical_iget_f.c
index efb817637c..76674c0ec9 100644
--- a/oshmem/shmem/fortran/shmem_logical_iget_f.c
+++ b/oshmem/shmem/fortran/shmem_logical_iget_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_LOGICAL_IGET, shmem_logical_iget)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_LOGICAL_IGET,
         shmem_logical_iget_,
diff --git a/oshmem/shmem/fortran/shmem_logical_iput_f.c b/oshmem/shmem/fortran/shmem_logical_iput_f.c
index 76a74c709f..c8e4c061ee 100644
--- a/oshmem/shmem/fortran/shmem_logical_iput_f.c
+++ b/oshmem/shmem/fortran/shmem_logical_iput_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_LOGICAL_IPUT, shmem_logical_iput)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_LOGICAL_IPUT,
         shmem_logical_iput_,
diff --git a/oshmem/shmem/fortran/shmem_logical_put_f.c b/oshmem/shmem/fortran/shmem_logical_put_f.c
index 81a19611fc..4409e1197c 100644
--- a/oshmem/shmem/fortran/shmem_logical_put_f.c
+++ b/oshmem/shmem/fortran/shmem_logical_put_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_LOGICAL_PUT, shmem_logical_put)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_LOGICAL_PUT,
         shmem_logical_put_,
diff --git a/oshmem/shmem/fortran/shmem_max_to_all_f.c b/oshmem/shmem/fortran/shmem_max_to_all_f.c
index be4ab41f99..60762d6234 100644
--- a/oshmem/shmem/fortran/shmem_max_to_all_f.c
+++ b/oshmem/shmem/fortran/shmem_max_to_all_f.c
@@ -18,6 +18,25 @@
 #include "oshmem/proc/proc_group_cache.h"
 #include "oshmem/op/op.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT2_MAX_TO_ALL, shmem_int2_max_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT4_MAX_TO_ALL, shmem_int4_max_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT8_MAX_TO_ALL, shmem_int8_max_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL4_MAX_TO_ALL, shmem_real4_max_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL8_MAX_TO_ALL, shmem_real8_max_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL16_MAX_TO_ALL, shmem_real16_max_to_all)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_INT2_MAX_TO_ALL,
+        shmem_int2_max_to_all_,
+        shmem_int2_max_to_all__,
+        shmem_int2_max_to_all_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
+        (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INT4_MAX_TO_ALL,
         shmem_int4_max_to_all_,
@@ -119,6 +138,7 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
     }\
 }
 
+SHMEM_MAX_TO_ALL(shmem_int2_max_to_all_f, oshmem_op_max_fint2, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_MAX_TO_ALL(shmem_int4_max_to_all_f, oshmem_op_max_fint4, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_MAX_TO_ALL(shmem_int8_max_to_all_f, oshmem_op_max_fint8, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_MAX_TO_ALL(shmem_real4_max_to_all_f, oshmem_op_max_freal4, OSHMEM_GROUP_CACHE_ENABLED)
diff --git a/oshmem/shmem/fortran/shmem_min_to_all_f.c b/oshmem/shmem/fortran/shmem_min_to_all_f.c
index abf973b37e..73084eed49 100644
--- a/oshmem/shmem/fortran/shmem_min_to_all_f.c
+++ b/oshmem/shmem/fortran/shmem_min_to_all_f.c
@@ -18,6 +18,25 @@
 #include "oshmem/proc/proc_group_cache.h"
 #include "oshmem/op/op.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT2_MIN_TO_ALL, shmem_int2_min_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT4_MIN_TO_ALL, shmem_int4_min_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT8_MIN_TO_ALL, shmem_int8_min_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL4_MIN_TO_ALL, shmem_real4_min_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL8_MIN_TO_ALL, shmem_real8_min_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL16_MIN_TO_ALL, shmem_real16_min_to_all)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_INT2_MIN_TO_ALL,
+        shmem_int2_min_to_all_,
+        shmem_int2_min_to_all__,
+        shmem_int2_min_to_all_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
+        (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INT4_MIN_TO_ALL,
         shmem_int4_min_to_all_,
@@ -50,6 +69,15 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
         (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
 
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_REAL16_MIN_TO_ALL,
+        shmem_real16_min_to_all_,
+        shmem_real16_min_to_all__,
+        shmem_real16_min_to_all_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
+        (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
+
+
 #define SHMEM_MIN_TO_ALL(F_NAME, T_NAME, OSHMEM_GROUP_CACHE_ENABLED) void F_NAME(FORTRAN_POINTER_T target, \
     FORTRAN_POINTER_T source, \
     MPI_Fint *nreduce,\
@@ -111,6 +139,7 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
     }\
 }
 
+SHMEM_MIN_TO_ALL(shmem_int2_min_to_all_f, oshmem_op_min_fint2, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_MIN_TO_ALL(shmem_int4_min_to_all_f, oshmem_op_min_fint4, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_MIN_TO_ALL(shmem_int8_min_to_all_f, oshmem_op_min_fint8, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_MIN_TO_ALL(shmem_real4_min_to_all_f, oshmem_op_min_freal4, OSHMEM_GROUP_CACHE_ENABLED)
diff --git a/oshmem/shmem/fortran/shmem_or_to_all_f.c b/oshmem/shmem/fortran/shmem_or_to_all_f.c
index 0404a4e6ac..16747429a2 100644
--- a/oshmem/shmem/fortran/shmem_or_to_all_f.c
+++ b/oshmem/shmem/fortran/shmem_or_to_all_f.c
@@ -18,6 +18,22 @@
 #include "oshmem/proc/proc_group_cache.h"
 #include "oshmem/op/op.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT2_OR_TO_ALL, shmem_int2_or_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT4_OR_TO_ALL, shmem_int4_or_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT8_OR_TO_ALL, shmem_int8_or_to_all)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_INT2_OR_TO_ALL,
+        shmem_int2_or_to_all_,
+        shmem_int2_or_to_all__,
+        shmem_int2_or_to_all_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
+        (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INT4_OR_TO_ALL,
         shmem_int4_or_to_all_,
@@ -95,5 +111,6 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
     }\
 }
 
+SHMEM_OR_TO_ALL(shmem_int2_or_to_all_f, oshmem_op_or_fint2, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_OR_TO_ALL(shmem_int4_or_to_all_f, oshmem_op_or_fint4, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_OR_TO_ALL(shmem_int8_or_to_all_f, oshmem_op_or_fint8, OSHMEM_GROUP_CACHE_ENABLED)
diff --git a/oshmem/shmem/fortran/shmem_pe_accessible_f.c b/oshmem/shmem/fortran/shmem_pe_accessible_f.c
index 82e38338c7..e24028b3d4 100644
--- a/oshmem/shmem/fortran/shmem_pe_accessible_f.c
+++ b/oshmem/shmem/fortran/shmem_pe_accessible_f.c
@@ -15,6 +15,12 @@
 #include "oshmem/shmem/shmem_api_logger.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_PE_ACCESSIBLE, shmem_pe_accessible)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_FUNCTION (ompi_fortran_logical_t,
         SHMEM_PE_ACCESSIBLE,
         shmem_pe_accessible_,
diff --git a/oshmem/shmem/fortran/shmem_prod_to_all_f.c b/oshmem/shmem/fortran/shmem_prod_to_all_f.c
index 4332e3f9b0..80a0f28ef5 100644
--- a/oshmem/shmem/fortran/shmem_prod_to_all_f.c
+++ b/oshmem/shmem/fortran/shmem_prod_to_all_f.c
@@ -18,6 +18,27 @@
 #include "oshmem/proc/proc_group_cache.h"
 #include "oshmem/op/op.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT2_PROD_TO_ALL, shmem_int2_prod_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT4_PROD_TO_ALL, shmem_int4_prod_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT8_PROD_TO_ALL, shmem_int8_prod_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_COMP4_PROD_TO_ALL, shmem_comp4_prod_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_COMP8_PROD_TO_ALL, shmem_comp8_prod_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL4_PROD_TO_ALL, shmem_real4_prod_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL8_PROD_TO_ALL, shmem_real8_prod_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL16_PROD_TO_ALL, shmem_real16_prod_to_all)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_INT2_PROD_TO_ALL,
+        shmem_int2_prod_to_all_,
+        shmem_int2_prod_to_all__,
+        shmem_int2_prod_to_all_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
+        (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INT4_PROD_TO_ALL,
         shmem_int4_prod_to_all_,
@@ -34,6 +55,23 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
         (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
 
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_COMP4_PROD_TO_ALL,
+        shmem_comp4_prod_to_all_,
+        shmem_comp4_prod_to_all__,
+        shmem_comp4_prod_to_all_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
+        (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_COMP8_PROD_TO_ALL,
+        shmem_comp8_prod_to_all_,
+        shmem_comp8_prod_to_all__,
+        shmem_comp8_prod_to_all_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
+        (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
+
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_REAL4_PROD_TO_ALL,
         shmem_real4_prod_to_all_,
@@ -50,6 +88,14 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
         (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
 
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_REAL16_PROD_TO_ALL,
+        shmem_real16_prod_to_all_,
+        shmem_real16_prod_to_all__,
+        shmem_real16_prod_to_all_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
+        (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
+
 #define SHMEM_PROD_TO_ALL(F_NAME, T_NAME, OSHMEM_GROUP_CACHE_ENABLED) void F_NAME(FORTRAN_POINTER_T target, \
     FORTRAN_POINTER_T source, \
     MPI_Fint *nreduce,\
@@ -111,8 +157,11 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
     }\
 }
 
+SHMEM_PROD_TO_ALL(shmem_int2_prod_to_all_f, oshmem_op_prod_fint2, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_PROD_TO_ALL(shmem_int4_prod_to_all_f, oshmem_op_prod_fint4, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_PROD_TO_ALL(shmem_int8_prod_to_all_f, oshmem_op_prod_fint8, OSHMEM_GROUP_CACHE_ENABLED)
+SHMEM_PROD_TO_ALL(shmem_comp4_prod_to_all_f, oshmem_op_prod_complexf, OSHMEM_GROUP_CACHE_ENABLED)
+SHMEM_PROD_TO_ALL(shmem_comp8_prod_to_all_f, oshmem_op_prod_complexd, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_PROD_TO_ALL(shmem_real4_prod_to_all_f, oshmem_op_prod_freal4, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_PROD_TO_ALL(shmem_real8_prod_to_all_f, oshmem_op_prod_freal8, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_PROD_TO_ALL(shmem_real16_prod_to_all_f, oshmem_op_prod_freal16, OSHMEM_GROUP_CACHE_ENABLED)
diff --git a/oshmem/shmem/fortran/shmem_ptr_f.c b/oshmem/shmem/fortran/shmem_ptr_f.c
index 14a00a2aa3..6b5fbe36dd 100644
--- a/oshmem/shmem/fortran/shmem_ptr_f.c
+++ b/oshmem/shmem/fortran/shmem_ptr_f.c
@@ -15,6 +15,12 @@
 #include "oshmem/shmem/shmem_api_logger.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_PTR, shmem_ptr)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_FUNCTION (FORTRAN_POINTER_T *,
         SHMEM_PTR,
         shmem_ptr_,
diff --git a/oshmem/shmem/fortran/shmem_put128_f.c b/oshmem/shmem/fortran/shmem_put128_f.c
new file mode 100644
index 0000000000..f753c91168
--- /dev/null
+++ b/oshmem/shmem/fortran/shmem_put128_f.c
@@ -0,0 +1,42 @@
+/*
+ * Copyright (c) 2014      Mellanox Technologies, Inc.
+ *                         All rights reserved.
+ * Copyright (c) 2013 Cisco Systems, Inc.  All rights reserved.
+ * $COPYRIGHT$
+ * 
+ * Additional copyrights may follow
+ * 
+ * $HEADER$	
+ */
+
+#include "oshmem_config.h"
+#include "oshmem/shmem/fortran/bindings.h"
+#include "oshmem/include/shmem.h"
+#include "oshmem/shmem/shmem_api_logger.h"
+#include "oshmem/runtime/runtime.h"
+#include "oshmem/mca/spml/spml.h"
+#include "ompi/datatype/ompi_datatype.h"
+#include "stdio.h"
+
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_PUT128, shmem_put128)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_PUT128,
+        shmem_put128_,
+        shmem_put128__,
+        shmem_put128_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe), 
+        (target,source,length,pe) )
+
+void shmem_put128_f(FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *length, MPI_Fint *pe)
+{
+    MCA_SPML_CALL(put(FPTR_2_VOID_PTR(target), 
+        OMPI_FINT_2_INT(*length) * 16, 
+        FPTR_2_VOID_PTR(source), 
+        OMPI_FINT_2_INT(*pe)));
+}
+   
diff --git a/oshmem/shmem/fortran/shmem_put32_f.c b/oshmem/shmem/fortran/shmem_put32_f.c
index 4159d4951a..57756332be 100644
--- a/oshmem/shmem/fortran/shmem_put32_f.c
+++ b/oshmem/shmem/fortran/shmem_put32_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_PUT32, shmem_put32)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_PUT32,
         shmem_put32_,
diff --git a/oshmem/shmem/fortran/shmem_put4_f.c b/oshmem/shmem/fortran/shmem_put4_f.c
index c8a0e49c38..edbd662aa3 100644
--- a/oshmem/shmem/fortran/shmem_put4_f.c
+++ b/oshmem/shmem/fortran/shmem_put4_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_PUT4, shmem_put4)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_PUT4,
         shmem_put4_,
diff --git a/oshmem/shmem/fortran/shmem_put64_f.c b/oshmem/shmem/fortran/shmem_put64_f.c
index c4191a1aee..fb476bcad1 100644
--- a/oshmem/shmem/fortran/shmem_put64_f.c
+++ b/oshmem/shmem/fortran/shmem_put64_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_PUT64, shmem_put64)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_PUT64,
         shmem_put64_,
diff --git a/oshmem/shmem/fortran/shmem_put8_f.c b/oshmem/shmem/fortran/shmem_put8_f.c
index f9b0ab6ff7..9fc6bcf78e 100644
--- a/oshmem/shmem/fortran/shmem_put8_f.c
+++ b/oshmem/shmem/fortran/shmem_put8_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_PUT8, shmem_put8)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_PUT8,
         shmem_put8_,
diff --git a/oshmem/shmem/fortran/shmem_put_f.c b/oshmem/shmem/fortran/shmem_put_f.c
index 09c96979f5..babb14a6e6 100644
--- a/oshmem/shmem/fortran/shmem_put_f.c
+++ b/oshmem/shmem/fortran/shmem_put_f.c
@@ -16,6 +16,12 @@
 #include "oshmem/runtime/runtime.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_PUT, shmem_put)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_PUT,
         shmem_put_,
diff --git a/oshmem/shmem/fortran/shmem_putmem_f.c b/oshmem/shmem/fortran/shmem_putmem_f.c
index 77639e973b..482ba407f9 100644
--- a/oshmem/shmem/fortran/shmem_putmem_f.c
+++ b/oshmem/shmem/fortran/shmem_putmem_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_PUTMEM, shmem_putmem)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_PUTMEM,
         shmem_putmem_,
diff --git a/oshmem/shmem/fortran/shmem_quiet_f.c b/oshmem/shmem/fortran/shmem_quiet_f.c
index 996dcc3c97..70db6cbe07 100644
--- a/oshmem/shmem/fortran/shmem_quiet_f.c
+++ b/oshmem/shmem/fortran/shmem_quiet_f.c
@@ -14,6 +14,12 @@
 #include "oshmem/include/shmem.h"
 #include "oshmem/mca/spml/spml.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_QUIET, shmem_quiet)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_QUIET,
         shmem_quiet_,
diff --git a/oshmem/shmem/fortran/shmem_real4_swap_f.c b/oshmem/shmem/fortran/shmem_real4_swap_f.c
index 1e69ac5f08..d0a6c2c059 100644
--- a/oshmem/shmem/fortran/shmem_real4_swap_f.c
+++ b/oshmem/shmem/fortran/shmem_real4_swap_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL4_SWAP, shmem_real4_swap)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_FUNCTION (ompi_fortran_real4_t,
         SHMEM_REAL4_SWAP,
         shmem_real4_swap_,
diff --git a/oshmem/shmem/fortran/shmem_real8_swap_f.c b/oshmem/shmem/fortran/shmem_real8_swap_f.c
index 9a0b5bac59..f18870e321 100644
--- a/oshmem/shmem/fortran/shmem_real8_swap_f.c
+++ b/oshmem/shmem/fortran/shmem_real8_swap_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL8_SWAP, shmem_real8_swap)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_FUNCTION (ompi_fortran_real8_t,
         SHMEM_REAL8_SWAP,
         shmem_real8_swap_,
diff --git a/oshmem/shmem/fortran/shmem_real_get_f.c b/oshmem/shmem/fortran/shmem_real_get_f.c
index 3698505398..eb179400a8 100644
--- a/oshmem/shmem/fortran/shmem_real_get_f.c
+++ b/oshmem/shmem/fortran/shmem_real_get_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL_GET, shmem_real_get)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_REAL_GET,
         shmem_real_get_,
diff --git a/oshmem/shmem/fortran/shmem_real_iget_f.c b/oshmem/shmem/fortran/shmem_real_iget_f.c
index 082afef9aa..db3411d9c9 100644
--- a/oshmem/shmem/fortran/shmem_real_iget_f.c
+++ b/oshmem/shmem/fortran/shmem_real_iget_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL_IGET, shmem_real_iget)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_REAL_IGET,
         shmem_real_iget_,
diff --git a/oshmem/shmem/fortran/shmem_real_iput_f.c b/oshmem/shmem/fortran/shmem_real_iput_f.c
index 34ac5f0b3d..827c52b79c 100644
--- a/oshmem/shmem/fortran/shmem_real_iput_f.c
+++ b/oshmem/shmem/fortran/shmem_real_iput_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL_IPUT, shmem_real_iput)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_REAL_IPUT,
         shmem_real_iput_,
diff --git a/oshmem/shmem/fortran/shmem_real_put_f.c b/oshmem/shmem/fortran/shmem_real_put_f.c
index e786a4fbe8..fd6c6a63ad 100644
--- a/oshmem/shmem/fortran/shmem_real_put_f.c
+++ b/oshmem/shmem/fortran/shmem_real_put_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL_PUT, shmem_real_put)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_REAL_PUT,
         shmem_real_put_,
diff --git a/oshmem/shmem/fortran/shmem_sum_to_all_f.c b/oshmem/shmem/fortran/shmem_sum_to_all_f.c
index 8e914ef8bd..b3815d8b74 100644
--- a/oshmem/shmem/fortran/shmem_sum_to_all_f.c
+++ b/oshmem/shmem/fortran/shmem_sum_to_all_f.c
@@ -18,6 +18,27 @@
 #include "oshmem/proc/proc_group_cache.h"
 #include "oshmem/op/op.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT2_SUM_TO_ALL, shmem_int2_sum_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT4_SUM_TO_ALL, shmem_int4_sum_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT8_SUM_TO_ALL, shmem_int8_sum_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_COMP4_SUM_TO_ALL, shmem_comp4_sum_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_COMP8_SUM_TO_ALL, shmem_comp8_sum_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL4_SUM_TO_ALL, shmem_real4_sum_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL8_SUM_TO_ALL, shmem_real8_sum_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_REAL16_SUM_TO_ALL, shmem_real16_sum_to_all)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_INT2_SUM_TO_ALL,
+        shmem_int2_sum_to_all_,
+        shmem_int2_sum_to_all__,
+        shmem_int2_sum_to_all_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
+        (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INT4_SUM_TO_ALL,
         shmem_int4_sum_to_all_,
@@ -34,6 +55,22 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
         (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
 
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_COMP4_SUM_TO_ALL,
+        shmem_comp4_sum_to_all_,
+        shmem_comp4_sum_to_all__,
+        shmem_comp4_sum_to_all_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
+        (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_COMP8_SUM_TO_ALL,
+        shmem_comp8_sum_to_all_,
+        shmem_comp8_sum_to_all__,
+        shmem_comp8_sum_to_all_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
+        (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_REAL4_SUM_TO_ALL,
         shmem_real4_sum_to_all_,
@@ -50,6 +87,14 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
         (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
 
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_REAL16_SUM_TO_ALL,
+        shmem_real16_sum_to_all_,
+        shmem_real16_sum_to_all__,
+        shmem_real16_sum_to_all_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
+        (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
+
 #define SHMEM_SUM_TO_ALL(F_NAME, T_NAME, OSHMEM_GROUP_CACHE_ENABLED) void F_NAME(FORTRAN_POINTER_T target, \
     FORTRAN_POINTER_T source, \
     MPI_Fint *nreduce,\
@@ -111,8 +156,11 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
     }\
 }
 
+SHMEM_SUM_TO_ALL(shmem_int2_sum_to_all_f, oshmem_op_sum_fint2, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_SUM_TO_ALL(shmem_int4_sum_to_all_f, oshmem_op_sum_fint4, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_SUM_TO_ALL(shmem_int8_sum_to_all_f, oshmem_op_sum_fint8, OSHMEM_GROUP_CACHE_ENABLED)
+SHMEM_SUM_TO_ALL(shmem_comp4_sum_to_all_f, oshmem_op_sum_complexf, OSHMEM_GROUP_CACHE_ENABLED)
+SHMEM_SUM_TO_ALL(shmem_comp8_sum_to_all_f, oshmem_op_sum_complexd, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_SUM_TO_ALL(shmem_real4_sum_to_all_f, oshmem_op_sum_freal4, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_SUM_TO_ALL(shmem_real8_sum_to_all_f, oshmem_op_sum_freal8, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_SUM_TO_ALL(shmem_real16_sum_to_all_f, oshmem_op_sum_freal16, OSHMEM_GROUP_CACHE_ENABLED)
diff --git a/oshmem/shmem/fortran/shmem_swap_f.c b/oshmem/shmem/fortran/shmem_swap_f.c
index 4c171dcffc..8fd34ff3f8 100644
--- a/oshmem/shmem/fortran/shmem_swap_f.c
+++ b/oshmem/shmem/fortran/shmem_swap_f.c
@@ -18,6 +18,12 @@
 #include "ompi/datatype/ompi_datatype.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_SWAP, shmem_swap)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_FUNCTION (MPI_Fint,
         SHMEM_SWAP,
         shmem_swap_,
diff --git a/oshmem/shmem/fortran/shmem_wait_f.c b/oshmem/shmem/fortran/shmem_wait_f.c
index 1b429b76af..de5deb00dd 100644
--- a/oshmem/shmem/fortran/shmem_wait_f.c
+++ b/oshmem/shmem/fortran/shmem_wait_f.c
@@ -16,6 +16,12 @@
 #include "oshmem/mca/spml/spml.h"
 #include "ompi/datatype/ompi_datatype.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_WAIT, shmem_wait)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_WAIT,
         shmem_wait_,
diff --git a/oshmem/shmem/fortran/shmem_wait_until_f.c b/oshmem/shmem/fortran/shmem_wait_until_f.c
index 9b9ad7bba3..b12bf2ec45 100644
--- a/oshmem/shmem/fortran/shmem_wait_until_f.c
+++ b/oshmem/shmem/fortran/shmem_wait_until_f.c
@@ -16,6 +16,12 @@
 #include "oshmem/mca/spml/spml.h"
 #include "ompi/datatype/ompi_datatype.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_WAIT_UNTIL, shmem_wait_until)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_WAIT_UNTIL,
         shmem_wait_until_,
diff --git a/oshmem/shmem/fortran/shmem_xor_to_all_f.c b/oshmem/shmem/fortran/shmem_xor_to_all_f.c
index 077e18a7ec..d1b42eb26d 100644
--- a/oshmem/shmem/fortran/shmem_xor_to_all_f.c
+++ b/oshmem/shmem/fortran/shmem_xor_to_all_f.c
@@ -18,6 +18,24 @@
 #include "oshmem/proc/proc_group_cache.h"
 #include "oshmem/op/op.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT2_XOR_TO_ALL, shmem_int2_xor_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT4_XOR_TO_ALL, shmem_int4_xor_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_INT8_XOR_TO_ALL, shmem_int8_xor_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_COMP4_XOR_TO_ALL, shmem_comp4_xor_to_all)
+SHMEM_GENERATE_WEAK_BINDINGS(SHMEM_COMP8_XOR_TO_ALL, shmem_comp8_xor_to_all)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_INT2_XOR_TO_ALL,
+        shmem_int2_xor_to_all_,
+        shmem_int2_xor_to_all__,
+        shmem_int2_xor_to_all_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
+        (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHMEM_INT4_XOR_TO_ALL,
         shmem_int4_xor_to_all_,
@@ -34,6 +52,22 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
         (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
 
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_COMP4_XOR_TO_ALL,
+        shmem_comp4_xor_to_all_,
+        shmem_comp4_xor_to_all__,
+        shmem_comp4_xor_to_all_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
+        (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
+
+SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
+        SHMEM_COMP8_XOR_TO_ALL,
+        shmem_comp8_xor_to_all_,
+        shmem_comp8_xor_to_all__,
+        shmem_comp8_xor_to_all_f,
+        (FORTRAN_POINTER_T target, FORTRAN_POINTER_T source, MPI_Fint *nreduce, MPI_Fint *PE_start, MPI_Fint * logPE_stride, MPI_Fint *PE_size, FORTRAN_POINTER_T *pWrk, FORTRAN_POINTER_T pSync), 
+        (target,source,nreduce,PE_start,logPE_stride,PE_size,pWrk,pSync) )
+
 #define SHMEM_XOR_TO_ALL(F_NAME, T_NAME, OSHMEM_GROUP_CACHE_ENABLED) void F_NAME(FORTRAN_POINTER_T target, \
     FORTRAN_POINTER_T source, \
     MPI_Fint *nreduce,\
@@ -97,6 +131,7 @@ SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
     }\
 }
 
+SHMEM_XOR_TO_ALL(shmem_int2_xor_to_all_f, oshmem_op_xor_fint2, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_XOR_TO_ALL(shmem_int4_xor_to_all_f, oshmem_op_xor_fint4, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_XOR_TO_ALL(shmem_int8_xor_to_all_f, oshmem_op_xor_fint8, OSHMEM_GROUP_CACHE_ENABLED)
 SHMEM_XOR_TO_ALL(shmem_comp4_xor_to_all_f, oshmem_op_xor_fint4, OSHMEM_GROUP_CACHE_ENABLED)
diff --git a/oshmem/shmem/fortran/shpalloc_f.c b/oshmem/shmem/fortran/shpalloc_f.c
index d627ca46e8..e3e8dedf62 100644
--- a/oshmem/shmem/fortran/shpalloc_f.c
+++ b/oshmem/shmem/fortran/shpalloc_f.c
@@ -16,6 +16,12 @@
 #include "oshmem/runtime/runtime.h"
 #include "stdio.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHPALLOC, shpalloc)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHPALLOC,
         shpalloc_,
diff --git a/oshmem/shmem/fortran/shpclmove_f.c b/oshmem/shmem/fortran/shpclmove_f.c
index ea6b8f720a..d1e79d376b 100644
--- a/oshmem/shmem/fortran/shpclmove_f.c
+++ b/oshmem/shmem/fortran/shpclmove_f.c
@@ -16,6 +16,12 @@
 #include "stdio.h"
 #include "oshmem/runtime/runtime.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHPCLMOVE, shpclmove)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHPCLMOVE,
         shpclmove_,
diff --git a/oshmem/shmem/fortran/shpdeallc_f.c b/oshmem/shmem/fortran/shpdeallc_f.c
index 67c052441c..49eac47307 100644
--- a/oshmem/shmem/fortran/shpdeallc_f.c
+++ b/oshmem/shmem/fortran/shpdeallc_f.c
@@ -14,6 +14,13 @@
 #include "oshmem/include/shmem.h"
 #include "oshmem/shmem/shmem_api_logger.h"
 #include "stdio.h"
+
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(SHPDEALLC, shpdeallc)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         SHPDEALLC,
         shpdeallc_,
diff --git a/oshmem/shmem/fortran/start_pes_f.c b/oshmem/shmem/fortran/start_pes_f.c
index 43de77ae89..3720dfedec 100644
--- a/oshmem/shmem/fortran/start_pes_f.c
+++ b/oshmem/shmem/fortran/start_pes_f.c
@@ -13,6 +13,12 @@
 #include "oshmem/shmem/fortran/bindings.h"
 #include "oshmem/include/shmem.h"
 
+#if OSHMEM_PROFILING
+#include "oshmem/shmem/fortran/profile/pbindings.h"
+SHMEM_GENERATE_WEAK_BINDINGS(START_PES, start_pes)
+#include "oshmem/shmem/fortran/profile/defines.h"
+#endif
+
 SHMEM_GENERATE_FORTRAN_BINDINGS_SUB (void,
         START_PES,
         start_pes_,