diff --git a/.gitignore b/.gitignore
index 07252f6710e..3424dffb881 100644
--- a/.gitignore
+++ b/.gitignore
@@ -384,6 +384,7 @@ Makefile.am-stamp
 
 # /src/binding/
 /src/binding/c/c_binding.c
+/src/binding/c/io.c
 
 # /src/binding/cxx/
 /src/binding/cxx/Makefile.sm
diff --git a/Makefile.am b/Makefile.am
index 1d4afd58ed4..ab103004b40 100644
--- a/Makefile.am
+++ b/Makefile.am
@@ -51,9 +51,22 @@ errnames_txt_files =
 external_subdirs = @mplsrcdir@ @hwlocsrcdir@ @jsonsrcdir@ @yaksasrcdir@ @pmisrcdir@
 external_ldflags = @mpllibdir@ @yaksalibdir@
 external_libs = @WRAPPER_LIBS@
-mpi_convenience_libs =
+
+# convenience libs that do not depend on MPI
 pmpi_convenience_libs = @mpllib@ @hwloclib@ @jsonlib@ @yaksalib@ @pmilib@
 
+# convenience libs that depend on MPI (either MPI functions or MPI constants)
+mpi_convenience_libs =
+abi_convenience_libs =
+
+if BUILD_ROMIO
+    external_subdirs += src/mpi/romio
+    mpi_convenience_libs += src/mpi/romio/libromio.la
+if BUILD_ABI_LIB
+    abi_convenience_libs += src/mpi/romio/libromio_abi.la
+endif
+endif
+
 # NOTE on our semi-unconventional usage of DIST_SUBDIRS:
 # The automake manual recommends thinking of DIST_SUBDIRS as the list of all
 # *configured* subdirectories.  The normal autotools model involves
@@ -124,8 +137,8 @@ if BUILD_PROFILING_LIB
 lib_LTLIBRARIES += lib/lib@PMPILIBNAME@.la
 lib_lib@PMPILIBNAME@_la_SOURCES = $(mpi_sources) $(mpi_core_sources)
 lib_lib@PMPILIBNAME@_la_LDFLAGS = $(external_ldflags) $(ABIVERSIONFLAGS)
-lib_lib@PMPILIBNAME@_la_LIBADD = $(external_libs) $(pmpi_convenience_libs)
-EXTRA_lib_lib@PMPILIBNAME@_la_DEPENDENCIES = $(pmpi_convenience_libs)
+lib_lib@PMPILIBNAME@_la_LIBADD = $(external_libs) $(pmpi_convenience_libs) $(mpi_convenience_libs)
+EXTRA_lib_lib@PMPILIBNAME@_la_DEPENDENCIES = $(pmpi_convenience_libs) $(mpi_convenience_libs)
 
 # lib@MPILIBNAME@.la might depend on lib@PMPILIBNAME@.la.  We add them
 # in that order to lib_LTLIBRARIES so libtool doesn't get
@@ -134,14 +147,8 @@ lib_LTLIBRARIES += lib/lib@MPILIBNAME@.la
 lib_lib@MPILIBNAME@_la_SOURCES = $(mpi_sources)
 lib_lib@MPILIBNAME@_la_LDFLAGS = $(ABIVERSIONFLAGS)
 lib_lib@MPILIBNAME@_la_CPPFLAGS = $(AM_CPPFLAGS) -DMPICH_MPI_FROM_PMPI
-lib_lib@MPILIBNAME@_la_LIBADD = lib/lib@PMPILIBNAME@.la $(mpi_convenience_libs)
+lib_lib@MPILIBNAME@_la_LIBADD = lib/lib@PMPILIBNAME@.la
 
-if BUILD_ROMIO
-    # libpromio contains the PMPI symbols (unlike libpmpi, which contains MPI
-    # symbols) and should be added to libpmpi
-    lib_lib@PMPILIBNAME@_la_LIBADD += src/mpi/romio/libpromio.la
-    lib_lib@MPILIBNAME@_la_LIBADD += src/mpi/romio/libromio.la
-endif
 else !BUILD_PROFILING_LIB
 
 lib_LTLIBRARIES += lib/lib@MPILIBNAME@.la
@@ -151,9 +158,6 @@ lib_lib@MPILIBNAME@_la_CPPFLAGS = $(AM_CPPFLAGS)
 lib_lib@MPILIBNAME@_la_LIBADD = $(external_libs) $(pmpi_convenience_libs) $(mpi_convenience_libs)
 EXTRA_lib_lib@MPILIBNAME@_la_DEPENDENCIES = $(pmpi_convenience_libs) $(mpi_convenience_libs)
 
-if BUILD_ROMIO
-    lib_lib@MPILIBNAME@_la_LIBADD += src/mpi/romio/libromio.la
-endif
 endif !BUILD_PROFILING_LIB
 
 if BUILD_ABI_LIB
@@ -165,8 +169,8 @@ lib_LTLIBRARIES += lib/lib@PMPIABILIBNAME@.la
 lib_lib@PMPIABILIBNAME@_la_SOURCES = $(mpi_abi_sources) $(mpi_f77_sources) $(mpi_core_sources)
 lib_lib@PMPIABILIBNAME@_la_LDFLAGS = $(external_ldflags) $(ABIVERSIONFLAGS)
 lib_lib@PMPIABILIBNAME@_la_CPPFLAGS = $(AM_CPPFLAGS) -DF77_USE_PMPI $(abi_cppflags)
-lib_lib@PMPIABILIBNAME@_la_LIBADD = $(external_libs) $(pmpi_convenience_libs)
-EXTRA_lib_lib@PMPIABILIBNAME@_la_DEPENDENCIES = $(pmpi_convenience_libs)
+lib_lib@PMPIABILIBNAME@_la_LIBADD = $(external_libs) $(pmpi_convenience_libs) $(abi_convenience_libs)
+EXTRA_lib_lib@PMPIABILIBNAME@_la_DEPENDENCIES = $(pmpi_convenience_libs) $(abi_convenience_libs)
 
 # lib@MPIABILIBNAME@.la might depend on lib@PMPIABILIBNAME@.la.  We add them
 # in that order to lib_LTLIBRARIES so libtool doesn't get
@@ -175,24 +179,17 @@ lib_LTLIBRARIES += lib/lib@MPIABILIBNAME@.la
 lib_lib@MPIABILIBNAME@_la_SOURCES = $(mpi_abi_sources)
 lib_lib@MPIABILIBNAME@_la_LDFLAGS = $(ABIVERSIONFLAGS)
 lib_lib@MPIABILIBNAME@_la_CPPFLAGS = $(AM_CPPFLAGS) -DMPICH_MPI_FROM_PMPI $(abi_cppflags)
-lib_lib@MPIABILIBNAME@_la_LIBADD = lib/lib@PMPIABILIBNAME@.la $(mpi_convenience_libs)
+lib_lib@MPIABILIBNAME@_la_LIBADD = lib/lib@PMPIABILIBNAME@.la
 
-if BUILD_ROMIO
-    lib_lib@PMPIABILIBNAME@_la_LIBADD += src/mpi/romio/libpromio_abi.la
-    lib_lib@MPIABILIBNAME@_la_LIBADD += src/mpi/romio/libromio_abi.la
-endif
 else !BUILD_PROFILING_LIB
 
 lib_LTLIBRARIES += lib/lib@MPIABILIBNAME@.la
 lib_lib@MPIABILIBNAME@_la_SOURCES = $(mpi_abi_sources) $(mpi_core_sources)
 lib_lib@MPIABILIBNAME@_la_LDFLAGS = $(external_ldflags) $(ABIVERSIONFLAGS)
 lib_lib@MPIABILIBNAME@_la_CPPFLAGS = $(AM_CPPFLAGS) $(abi_cppflags)
-lib_lib@MPIABILIBNAME@_la_LIBADD = $(external_libs) $(pmpi_convenience_libs) $(mpi_convenience_libs)
-EXTRA_lib_lib@MPIABILIBNAME@_la_DEPENDENCIES = $(pmpi_convenience_libs) $(mpi_convenience_libs)
+lib_lib@MPIABILIBNAME@_la_LIBADD = $(external_libs) $(pmpi_convenience_libs) $(abi_convenience_libs)
+EXTRA_lib_lib@MPIABILIBNAME@_la_DEPENDENCIES = $(pmpi_convenience_libs) $(abi_convenience_libs)
 
-if BUILD_ROMIO
-    lib_lib@MPIABILIBNAME@_la_LIBADD += src/mpi/romio/libromio_abi.la
-endif
 endif !BUILD_PROFILING_LIB
 endif BUILD_ABI_LIB
 
diff --git a/maint/gen_abi.py b/maint/gen_abi.py
index d449339aabd..826bc747b94 100644
--- a/maint/gen_abi.py
+++ b/maint/gen_abi.py
@@ -25,6 +25,7 @@ def search(pat, str, flags=0):
 def main():
     load_mpi_abi_h("src/binding/abi/mpi_abi.h")
     dump_mpi_abi_internal_h("src/binding/abi/mpi_abi_internal.h")
+    dump_io_abi_internal_h("src/binding/abi/io_abi_internal.h")
     dump_romio_abi_internal_h("src/mpi/romio/include/romio_abi_internal.h")
     dump_mpi_abi_util_c("src/binding/abi/mpi_abi_util.c")
 
@@ -101,7 +102,7 @@ def gen_mpi_abi_internal_h(out):
         for line in output_lines:
             print(line, file=Out)
         print("", file=Out)
-        
+
         print("#endif /* MPI_ABI_INTERNAL_H_INCLUDED */", file=Out)
 
 def dump_romio_abi_internal_h(romio_abi_internal_h):
@@ -181,6 +182,52 @@ def add_other(out):
         print("#define ROMIO_ABI_INTERNAL_H_INCLUDED", file=Out)
         print("", file=Out)
 
+        for line in output_lines:
+            print(line, file=Out)
+        print("", file=Out)
+
+        print("#endif /* ROMIO_ABI_INTERNAL_H_INCLUDED */", file=Out)
+
+# similar to romio_abi_internal.h but for use in the mpich io binding
+def dump_io_abi_internal_h(io_abi_internal_h):
+    def gen_io_abi_internal_h(out):
+        for line in G.abi_h_lines:
+            if RE.search(r'MPI_ABI_H_INCLUDED', line):
+                # skip the include guard, harmless
+                pass
+            elif RE.match(r'typedef struct.*\bMPI_File;\s*$', line):
+                out.append("typedef struct ADIOI_FileD *MPI_File;")
+            elif RE.match(r'(int|double|MPI_\w+) (P?MPI\w+)\((.*)\);', line):
+                # prototypes, rename param prefix, add MPICH_API_PUBLIC
+                (T, name, param) = RE.m.group(1,2,3)
+                if RE.match(r'P?MPI_(File_\w+|Register_datarep\w*)', name):
+                    out.append("%s %s(%s) MPICH_API_PUBLIC;" % (T, name, param))
+                else:
+                    out.append("%s %s(%s);" % (T, name, param))
+            else:
+                # replace param prefix
+                out.append(line.rstrip())
+
+    def add_mpich_visibility(out):
+        out.append("#if defined(HAVE_VISIBILITY)")
+        out.append("#define MPICH_API_PUBLIC __attribute__((visibility (\"default\")))")
+        out.append("#else")
+        out.append("#define MPICH_API_PUBLIC")
+        out.append("#endif")
+        out.append("")
+
+    # ----
+    output_lines = []
+    add_mpich_visibility(output_lines)
+    gen_io_abi_internal_h(output_lines)
+
+    print(" --> [%s]" % io_abi_internal_h)
+    with open(io_abi_internal_h, "w") as Out:
+        dump_copyright(Out)
+        print("#ifndef ROMIO_ABI_INTERNAL_H_INCLUDED", file=Out)
+        print("#define ROMIO_ABI_INTERNAL_H_INCLUDED", file=Out)
+        print("", file=Out)
+
         for line in output_lines:
             print(line, file=Out)
         print("", file=Out)
diff --git a/maint/gen_binding_c.py b/maint/gen_binding_c.py
index d2f50ef74b8..ccb37681a2b 100644
--- a/maint/gen_binding_c.py
+++ b/maint/gen_binding_c.py
@@ -50,10 +50,11 @@ def main():
             repl_func['_replaces'].append(func)
 
 
+    # We generate io functions separately for now
+    io_func_list = [f for f in func_list if f['dir'] == 'io']
+    func_list = [f for f in func_list if f['dir'] != 'io']
+
     # -- Generating code --
-    G.out = []
-    G.out.append("#include \"mpiimpl.h\"")
-    G.out.append("")
     G.doc3_src_txt = []
     G.poly_aliases = [] # large-count mansrc aliases
     G.need_dump_romio_reference = True
@@ -64,9 +65,6 @@ def dump_out(file_path):
         dump_c_file(file_path, G.out)
         # add to mpi_sources for dump_Makefile_mk()
         G.mpi_sources.append(file_path)
-        G.out = []
-        G.out.append("#include \"mpiimpl.h\"")
-        G.out.append("")
         G.need_dump_romio_reference = True
 
     def dump_func(func, manpage_out):
@@ -95,52 +93,91 @@ def dump_func_abi(func):
             dump_mpi_c(func, True)
         del func['_is_abi']
 
-    # ----
-    for func in func_list:
-        G.err_codes = {}
-        manpage_out = []
-
-        if 'replace' in func and 'body' not in func:
-            continue
+    def dump_c_binding():
+        G.out = []
+        G.out.append("#include \"mpiimpl.h\"")
+        G.out.append("")
+        for func in func_list:
+            G.err_codes = {}
+            manpage_out = []
 
-        dump_func(func, manpage_out)
-        if '_replaces' in func:
-            for t_func in func['_replaces']:
-                dump_func(t_func, manpage_out)
+            if 'replace' in func and 'body' not in func:
+                continue
 
-        if 'single-source' not in G.opts:
-            # dump individual functions in separate source files
-            dump_out(get_func_file_path(func, c_dir))
+            dump_func(func, manpage_out)
+            if '_replaces' in func:
+                for t_func in func['_replaces']:
+                    dump_func(t_func, manpage_out)
 
-    if 'single-source' in G.opts:
-        # otherwise, dump all functions in binding.c
-        dump_out(c_dir + "/c_binding.c")
+            if 'single-source' not in G.opts:
+                # dump individual functions in separate source files
+                dump_out(get_func_file_path(func, c_dir))
+                G.out = []
+                G.out.append("#include \"mpiimpl.h\"")
+                G.out.append("")
 
-    # ---- dump c_binding_abi.c
-    G.out = []
-    G.out.append("#include \"mpiimpl.h\"")
-    G.out.append("#include \"mpi_abi_util.h\"")
-    G.out.append("")
+        if 'single-source' in G.opts:
+            # otherwise, dump all functions in binding.c
+            dump_out(c_dir + "/c_binding.c")
 
-    for func in func_list:
-        if 'replace' in func and 'body' not in func:
-            continue
+    def dump_c_binding_abi():
+        G.out = []
+        G.out.append("#include \"mpiimpl.h\"")
+        G.out.append("#include \"mpi_abi_util.h\"")
+        G.out.append("")
 
-        if re.match(r'MPIX_', func['name']):
-            if re.match(r'MPIX_(Grequest_|Type_iov)', func['name']):
-                # needed by ROMIO
-                pass
-            else:
+        for func in func_list:
+            if 'replace' in func and 'body' not in func:
                 continue
-        dump_func_abi(func)
-        if '_replaces' in func:
-            for t_func in func['_replaces']:
-                dump_func_abi(t_func)
 
-    abi_file_path = abi_dir + "/c_binding_abi.c"
-    G.check_write_path(abi_file_path)
-    dump_c_file(abi_file_path, G.out)
-    G.out = []
+            if re.match(r'MPIX_', func['name']):
+                if re.match(r'MPIX_(Grequest_|Type_iov)', func['name']):
+                    # needed by ROMIO
+                    pass
+                else:
+                    continue
+            dump_func_abi(func)
+            if '_replaces' in func:
+                for t_func in func['_replaces']:
+                    dump_func_abi(t_func)
+
+        abi_file_path = abi_dir + "/c_binding_abi.c"
+        G.check_write_path(abi_file_path)
+        dump_c_file(abi_file_path, G.out)
+
+    def dump_io_funcs():
+        G.out = []
+        G.out.append("#include \"mpiimpl.h\"")
+        G.out.append("#include \"mpir_io_impl.h\"")
+        G.out.append("")
+
+        for func in io_func_list:
+            G.err_codes = {}
+            manpage_out = []
+            dump_func(func, manpage_out)
+
+        dump_out(c_dir + "/io.c")
+
+    def dump_io_funcs_abi():
+        G.out = []
+        G.out.append("#include \"mpichconf.h\"")
+        G.out.append("#include \"io_abi_internal.h\"")
+        G.out.append("#include \"mpir_io_impl.h\"")
+        G.out.append("#include <limits.h>")
+        G.out.append("")
+
+        for func in io_func_list:
+            dump_func_abi(func)
+
+        abi_file_path = abi_dir + "/io_abi.c"
+        G.check_write_path(abi_file_path)
+        dump_c_file(abi_file_path, G.out)
+
+    # ----
+    dump_c_binding()
+    dump_c_binding_abi()
+    dump_io_funcs()
+    dump_io_funcs_abi()
 
     if 'output-mansrc' in G.opts:
         f = c_dir + '/mansrc/' + 'poly_aliases.lst'
@@ -154,6 +191,7 @@ def dump_func_abi(func):
     G.check_write_path("src/include/mpi_proto.h")
     dump_Makefile_mk("%s/Makefile.mk" % c_dir)
     dump_mpir_impl_h("src/include/mpir_impl.h")
+    dump_mpir_io_impl_h("src/include/mpir_io_impl.h")
     dump_errnames_txt("%s/errnames.txt" % c_dir)
     dump_qmpi_register_h("src/mpi_t/qmpi_register.h")
     dump_mpi_proto_h("src/include/mpi_proto.h")
diff --git a/maint/gen_binding_f08.py b/maint/gen_binding_f08.py
index 77df0964c94..0af0904b764 100644
--- a/maint/gen_binding_f08.py
+++ b/maint/gen_binding_f08.py
@@ -19,13 +19,6 @@ def main():
     G.check_write_path("%s/wrappers_f/" % f08_dir)
     G.check_write_path("%s/wrappers_c/" % f08_dir)
     func_list = load_C_func_list(binding_dir, True) # suppress noise
-    if "no-mpiio" in G.opts:
-        # a few MPI_File_xxx functions are already in (MPI_File_xxx_errhandler)
-        func_list = [f for f in func_list if not f['name'].startswith('MPI_File_')]
-    else:
-        # FIXME: until romio interface is generated
-        func_list.extend(get_mpiio_func_list())
-    func_list.append(G.FUNCS['mpi_f_sync_reg'])
 
     # preprocess
     get_real_POLY_kinds()
diff --git a/maint/gen_binding_f77.py b/maint/gen_binding_f77.py
index c4bbd43c367..dc56f779c31 100644
--- a/maint/gen_binding_f77.py
+++ b/maint/gen_binding_f77.py
@@ -19,14 +19,7 @@ def main():
 
     func_list = load_C_func_list(binding_dir, True) # suppress noise
 
-    if "no-mpiio" in G.opts:
-        # a few MPI_File_xxx functions are already in (MPI_File_xxx_errhandler)
-        func_list = [f for f in func_list if not f['name'].startswith('MPI_File_')]
-    else:
-        # FIXME: until romio interface is generated
-        func_list.extend(get_mpiio_func_list())
     func_list.extend(get_f77_dummy_func_list())
-    func_list.append(G.FUNCS['mpi_f_sync_reg'])
 
     # preprocess
     for func in func_list:
diff --git a/maint/gen_binding_f90.py b/maint/gen_binding_f90.py
index f71e6a1bdf4..2aa95c9defa 100644
--- a/maint/gen_binding_f90.py
+++ b/maint/gen_binding_f90.py
@@ -17,10 +17,6 @@ def main():
     f90_dir = "src/binding/fortran/use_mpi"
     G.check_write_path("%s/" % f90_dir)
     func_list = load_C_func_list(binding_dir, True) # suppress noise
-    if "no-mpiio" in G.opts:
-        func_list = [f for f in func_list if not f['name'].startswith('MPI_File_')]
-    else:
-        func_list.extend(get_mpiio_func_list())
     func_list.extend(get_f77_dummy_func_list())
 
     def has_cptr(func):
diff --git a/maint/local_python/__init__.py b/maint/local_python/__init__.py
index 08c41ff53c2..6a025b387da 100644
--- a/maint/local_python/__init__.py
+++ b/maint/local_python/__init__.py
@@ -62,6 +62,7 @@ def check_write_path(path):
     mpi_sources = []
     mpi_declares = []
     impl_declares = []
+    io_impl_declares = []
     mpi_errnames = []
     mpix_symbols = {}
 
diff --git a/maint/local_python/binding_c.py b/maint/local_python/binding_c.py
index f219ee4e194..e1ea7fe68d5 100644
--- a/maint/local_python/binding_c.py
+++ b/maint/local_python/binding_c.py
@@ -81,14 +81,18 @@ def dump_romio_reference(name):
 
     dump_profiling(func)
 
-    if 'polymorph' in func:
+    skip_wrappers = False
+    if func['dir'] == 'io' and '_is_abi' in func:
+        # The mpi-abi version of io bindings does not have access to MPICH internals (i.e. mpiimpl.h)
+        dump_function_io(func)
+        skip_wrappers = True
+    elif 'polymorph' in func:
         # MPII_ function to support C/Fortran Polymorphism, eg MPI_Comm_get_attr
         G.out.append("#ifndef MPICH_MPI_FROM_PMPI")
         dump_function_internal(func, kind="polymorph")
         G.out.append("#endif /* MPICH_MPI_FROM_PMPI */")
         G.out.append("")
 
-    if 'polymorph' in func:
         dump_function_internal(func, kind="call-polymorph")
     elif 'replace' in func and 'body' not in func:
         pass
@@ -96,7 +100,9 @@ def dump_romio_reference(name):
         dump_function_internal(func, kind="normal")
     G.out.append("")
 
-    if '_is_abi' in func:
+    if skip_wrappers:
+        pass
+    elif '_is_abi' in func:
         dump_abi_wrappers(func, func['_is_large'])
     else:
         # Create the MPI and QMPI wrapper functions that will call the above, "real" version of the
@@ -208,6 +214,40 @@ def dump_mpix_symbols():
         print("", file=Out)
         print("#endif /* MPIR_IMPL_H_INCLUDED */", file=Out)
 
+def dump_mpir_io_impl_h(f):
+    print("  --> [%s]" %f)
+    with open(f, "w") as Out:
+        for l in G.copyright_c:
+            print(l, file=Out)
+        print("#ifndef MPIR_IO_IMPL_H_INCLUDED", file=Out)
+        print("#define MPIR_IO_IMPL_H_INCLUDED", file=Out)
+        print("", file=Out)
+
+        # io_abi.c doesn't have access to MPICH internal
+        print("#ifdef BUILD_MPI_ABI", file=Out)
+        print("#define MPIR_ERR_RECOVERABLE 0", file=Out)
+        print("#define MPIR_ERR_FATAL 1", file=Out)
+        print("int MPIR_Err_create_code(int, int, const char[], int, int, const char[], const char[], ...);", file=Out)
+        print("int MPIR_Err_return_comm(void *, const char[], int);", file=Out)
+        print("#endif", file=Out)
+        print("", file=Out)
+
+        print("void MPIR_Ext_cs_enter(void);", file=Out)
+        print("void MPIR_Ext_cs_exit(void);", file=Out)
+        print("#ifndef HAVE_ROMIO", file=Out)
+        print("#define MPIO_Err_return_file(fh, errorcode) MPIR_Err_return_comm((void *)0, __func__, errorcode)", file=Out)
+        print("#else", file=Out)
+        print("MPI_Fint MPIR_File_c2f_impl(MPI_File fh);", file=Out)
+        print("MPI_File MPIR_File_f2c_impl(MPI_Fint fh);", file=Out)
+        print("int MPIO_Err_return_file(MPI_File fh, int errorcode);", file=Out)
+        print("#endif", file=Out)
+
+        print("", file=Out)
+        for l in G.io_impl_declares:
+            print(l, file=Out)
+        print("", file=Out)
+        print("#endif /* MPIR_IO_IMPL_H_INCLUDED */", file=Out)
+
 def filter_out_abi():
     funcname = None
     for l in G.out:
@@ -552,8 +592,10 @@ def process_func_parameters(func):
 
         do_handle_ptr = 0
         (kind, name) = (p['kind'], p['name'])
-        if '_has_comm' not in func and kind == "COMMUNICATOR" and p['param_direction'] == 'in':
+        if '_has_comm' not in func and kind == "COMMUNICATOR" and p['param_direction'] == 'in' and func['name'] != "MPI_File_open":
             func['_has_comm'] = name
+        elif kind == "FILE" and p['param_direction'] == 'in' and func['dir'] == 'io':
+            func['_has_file'] = name
         elif name == "win":
             func['_has_win'] = name
         elif name == "session":
@@ -733,7 +775,7 @@ def process_func_parameters(func):
         elif RE.match(r'F90_(COMM|ERRHANDLER|FILE|GROUP|INFO|MESSAGE|OP|REQUEST|SESSION|DATATYPE|WIN)', kind):
             # no validation for these kinds
             pass
-        elif RE.match(r'(POLY)?(DTYPE_STRIDE_BYTES|DISPLACEMENT_AINT_COUNT)$', kind):
+        elif RE.match(r'(POLY)?(DTYPE_STRIDE_BYTES|DISPLACEMENT_AINT_COUNT|OFFSET)$', kind):
             # e.g. stride in MPI_Type_vector, MPI_Type_create_resized
             pass
         elif is_pointer_type(p):
@@ -741,7 +783,11 @@ def process_func_parameters(func):
         else:
             print("Missing error checking: func=%s, name=%s, kind=%s" % (func_name, name, kind), file=sys.stderr)
 
-        if do_handle_ptr == 1:
+        if func['dir'] == 'io':
+            # pass io function parameters as is
+            impl_arg_list.append(name)
+            impl_param_list.append(get_impl_param(func, p))
+        elif do_handle_ptr == 1:
             if p['param_direction'] == 'inout':
                 # assume only one such parameter
                 func['_has_handle_inout'] = p
@@ -810,6 +856,48 @@ def dump_copy_right():
     G.out.append(" */")
     G.out.append("")
 
+def dump_function_io(func):
+    is_large = func['_is_large']
+    parameters = ""
+    for p in func['c_parameters']:
+        parameters = parameters + ", " + p['name']
+
+    func_decl = get_declare_function(func, is_large)
+
+    dump_line_with_break(func_decl)
+    G.out.append("{")
+    G.out.append("INDENT")
+    if "impl" in func and func['impl'] == "direct":
+        dump_function_direct(func)
+    else:
+        G.out.append("int mpi_errno = MPI_SUCCESS;")
+        if not '_skip_global_cs' in func:
+            G.out.append("MPIR_Ext_cs_enter();")
+        G.out.append("")
+        G.out.append("#ifndef HAVE_ROMIO")
+        if not '_is_abi' in func:
+            G.out.append("mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE, __func__, __LINE__, MPI_ERR_OTHER, \"**notimpl\", 0);")
+        else:
+            G.out.append("mpi_errno = MPI_ERR_INTERN;")
+        G.out.append("goto fn_fail;");
+        G.out.append("#else")
+        dump_body_of_routine(func)
+        G.out.append("#endif")
+        G.out.append("")
+        G.out.append("fn_exit:")
+        if not '_skip_global_cs' in func:
+            G.out.append("MPIR_Ext_cs_exit();")
+        G.out.append("return mpi_errno;")
+        G.out.append("fn_fail:")
+        if '_has_file' in func:
+            G.out.append("mpi_errno = MPIO_Err_return_file(%s, mpi_errno);" % func['_has_file'])
+        else:
+            G.out.append("mpi_errno = MPIO_Err_return_file(MPI_FILE_NULL, mpi_errno);")
+        G.out.append("goto fn_exit;")
+    G.out.append("DEDENT")
+    G.out.append("}")
+    G.out.append("")
+
 def dump_qmpi_wrappers(func, is_large):
     parameters = ""
     for p in func['c_parameters']:
@@ -1106,7 +1194,7 @@ def out_can_be_undefined(p):
         G.out.append("int ret = " + static_call + ";")
         for l in post_filters:
             G.out.append(l)
-        if re.match(r'MPI_(Init|Init_thread|Session_init)$', func_name, re.IGNORECASE):
+        if re.match(r'MPI_(Init|Init_thread|Session_init|T_init_thread)$', func_name, re.IGNORECASE):
             G.out.append("ABI_init_builtins();")
         G.out.append("return ret;")
     G.out.append("DEDENT")
@@ -1387,6 +1475,25 @@ def check_large_parameters(func):
                 func['_poly_in_list'].append(p)
 
 def dump_function_normal(func):
+    def dump_global_cs_enter():
+        if not '_skip_global_cs' in func:
+            G.out.append("")
+            if func['dir'] == 'mpit':
+                G.out.append("MPIR_T_THREAD_CS_ENTER();")
+            elif func['dir'] == 'io':
+                G.out.append("MPIR_Ext_cs_enter();")
+            else:
+                G.out.append("MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);")
+    def dump_global_cs_exit():
+        if not '_skip_global_cs' in func:
+            G.out.append("")
+            if func['dir'] == 'mpit':
+                G.out.append("MPIR_T_THREAD_CS_EXIT();")
+            elif func['dir'] == 'io':
+                G.out.append("MPIR_Ext_cs_exit();")
+            else:
+                G.out.append("MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);")
+    # ----
     G.out.append("int mpi_errno = MPI_SUCCESS;")
     if '_handle_ptr_list' in func:
         for p in func['_handle_ptr_list']:
@@ -1405,12 +1512,7 @@ def dump_function_normal(func):
         else:
             G.out.append("MPIR_ERRTEST_INITIALIZED_ORDIE();")
 
-    if not '_skip_global_cs' in func:
-        G.out.append("")
-        if func['dir'] == 'mpit':
-            G.out.append("MPIR_T_THREAD_CS_ENTER();")
-        else:
-            G.out.append("MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);")
+    dump_global_cs_enter()
     G.out.append("MPIR_FUNC_TERSE_ENTER;")
 
     if '_handle_ptr_list' in func:
@@ -1466,41 +1568,6 @@ def dump_function_normal(func):
     check_early_returns(func)
     G.out.append("")
 
-    # ----
-    def dump_body_of_routine():
-        do_threadcomm = False
-        if RE.search(r'threadcomm', func['extra'], re.IGNORECASE):
-            do_threadcomm = True
-            G.out.append("#ifdef ENABLE_THREADCOMM")
-            dump_if_open("comm_ptr->threadcomm")
-            dump_body_threadcomm(func)
-            dump_else_open()
-            G.out.append("#endif")
-            dump_else_close()
-
-        if 'body' in func:
-            if func['_is_large'] and func['_poly_impl'] == "separate":
-                if 'code-large_count' not in func:
-                    raise Exception("%s missing large count code block." % func['name'])
-                for l in func['code-large_count']:
-                    G.out.append(l)
-            else:
-                for l in func['body']:
-                    G.out.append(l)
-        elif 'impl' in func:
-            if RE.match(r'mpid', func['impl'], re.IGNORECASE):
-                dump_body_impl(func, "mpid")
-            elif RE.match(r'topo_fns->(\w+)', func['impl'], re.IGNORECASE):
-                dump_body_topo_fns(func, RE.m.group(1))
-            else:
-                print("Error: unhandled special impl: [%s]" % func['impl'])
-        elif func['dir'] == 'coll':
-            dump_body_coll(func)
-        else:
-            dump_body_impl(func, "mpir")
-
-        if do_threadcomm:
-            dump_if_close()
     # ----
     G.out.append("/* ... body of routine ... */")
 
@@ -1511,34 +1578,18 @@ def dump_body_of_routine():
         G.out.append("goto fn_exit;")
         dump_if_close()
 
-    if func['_is_large'] and 'code-large_count' not in func:
-        # BIG but internally is using MPI_Aint
-        impl_args_save = copy.copy(func['_impl_arg_list'])
-
-        dump_if_open("sizeof(MPI_Count) == sizeof(MPI_Aint)")
-        # Same size, just casting the _impl_arg_list
-        add_poly_impl_cast(func)
-        dump_body_of_routine()
+    need_endif = False
+    if func['dir'] == 'io' and not 'return' in func:
+        G.out.append("#ifndef HAVE_ROMIO")
+        G.out.append("mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE, __func__, __LINE__, MPI_ERR_OTHER, \"**notimpl\", 0);")
+        G.out.append("goto fn_fail;")
+        G.out.append("#else")
+        need_endif = True
 
-        dump_else()
-        # Need filter to check limits and potentially swap args
-        func['_impl_arg_list'] = copy.copy(impl_args_save)
-        G.out.append("/* MPI_Count is bigger than MPI_Aint */")
-        dump_poly_pre_filter(func)
-        dump_body_of_routine()
-        dump_poly_post_filter(func)
-
-        dump_if_close()
+    dump_body_of_routine(func)
 
-    elif not func['_is_large'] and func['_has_poly'] and func['_poly_impl'] != "separate":
-        # SMALL but internally is using MPI_Aint
-        dump_poly_pre_filter(func)
-        dump_body_of_routine()
-        dump_poly_post_filter(func)
-
-    else:
-        # normal
-        dump_body_of_routine()
+    if need_endif:
+        G.out.append("#endif")
 
     G.out.append("/* ... end of body of routine ... */")
 
@@ -1548,12 +1599,8 @@ def dump_body_of_routine():
     for l in func['_clean_up']:
         G.out.append(l)
     G.out.append("MPIR_FUNC_TERSE_EXIT;")
+    dump_global_cs_exit()
 
-    if not '_skip_global_cs' in func:
-        if func['dir'] == 'mpit':
-            G.out.append("MPIR_T_THREAD_CS_EXIT();")
-        else:
-            G.out.append("MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);")
     G.out.append("return mpi_errno;")
     G.out.append("")
     G.out.append("fn_fail:")
@@ -1697,7 +1744,10 @@ def push_impl_decl(func, impl_name=None):
         mpir_name = re.sub(r'^MPIX?_', 'MPIR_', func['name'])
         G.impl_declares.append("int %s(%s);" % (mpir_name, params))
     # dump MPIR_Xxx_impl(...)
-    G.impl_declares.append("int %s(%s);" % (impl_name, params))
+    if func['dir'] == 'io':
+        G.io_impl_declares.append("int %s(%s);" % (impl_name, params))
+    else:
+        G.impl_declares.append("int %s(%s);" % (impl_name, params))
 
 def push_threadcomm_impl_decl(func):
     impl_name = re.sub(r'^mpix?_(comm_)?', 'MPIR_Threadcomm_', func['name'].lower())
@@ -1959,6 +2009,75 @@ def dump_body_reduce_equal(func):
     dump_line_with_break("mpi_errno = %s(%s);" % (impl, args))
     dump_error_check("")
 
+def dump_body_of_routine(func):
+    def dump_body_normal():
+        if 'body' in func:
+            if func['_is_large'] and func['_poly_impl'] == "separate":
+                if 'code-large_count' not in func:
+                    raise Exception("%s missing large count code block." % func['name'])
+                for l in func['code-large_count']:
+                    G.out.append(l)
+            else:
+                for l in func['body']:
+                    G.out.append(l)
+        elif 'impl' in func:
+            if RE.match(r'mpid', func['impl'], re.IGNORECASE):
+                dump_body_impl(func, "mpid")
+            elif RE.match(r'topo_fns->(\w+)', func['impl'], re.IGNORECASE):
+                dump_body_topo_fns(func, RE.m.group(1))
+            else:
+                print("Error: unhandled special impl: [%s]" % func['impl'])
+        elif func['dir'] == 'coll':
+            dump_body_coll(func)
+        else:
+            dump_body_impl(func, "mpir")
+
+    def dump_body_internal():
+        do_threadcomm = False
+        if RE.search(r'threadcomm', func['extra'], re.IGNORECASE):
+            do_threadcomm = True
+            G.out.append("#ifdef ENABLE_THREADCOMM")
+            dump_if_open("comm_ptr->threadcomm")
+            dump_body_threadcomm(func)
+            dump_else_open()
+            G.out.append("#endif")
+            dump_else_close()
+
+        dump_body_normal()
+
+        if do_threadcomm:
+            dump_if_close()
+
+    # ----
+    if func['_is_large'] and 'code-large_count' not in func:
+        # BIG but internally is using MPI_Aint
+        impl_args_save = copy.copy(func['_impl_arg_list'])
+
+        dump_if_open("sizeof(MPI_Count) == sizeof(MPI_Aint)")
+        # Same size, just casting the _impl_arg_list
+        add_poly_impl_cast(func)
+        dump_body_internal()
+
+        dump_else()
+        # Need filter to check limits and potentially swap args
+        func['_impl_arg_list'] = copy.copy(impl_args_save)
+        G.out.append("/* MPI_Count is bigger than MPI_Aint */")
+        dump_poly_pre_filter(func)
+        dump_body_internal()
+        dump_poly_post_filter(func)
+
+        dump_if_close()
+
+    elif not func['_is_large'] and func['_has_poly'] and func['_poly_impl'] != "separate":
+        # SMALL but internally is using MPI_Aint
+        dump_poly_pre_filter(func)
+        dump_body_internal()
+        dump_poly_post_filter(func)
+
+    else:
+        # normal
+        dump_body_internal()
+
 def dump_function_replace(func, repl_call):
     G.out.append("int mpi_errno = MPI_SUCCESS;")
 
@@ -1995,6 +2114,8 @@ def dump_mpi_fn_fail(func):
             G.out.append("mpi_errno = MPIR_Err_return_comm(%s_ptr, __func__, mpi_errno);" % func['_has_comm'])
         elif '_has_win' in func:
             G.out.append("mpi_errno = MPIR_Err_return_win(win_ptr, __func__, mpi_errno);")
+        elif '_has_file' in func:
+            G.out.append("mpi_errno = MPIO_Err_return_file(%s, mpi_errno);" % func['_has_file'])
         elif RE.match(r'mpi_session_init', func['name'], re.IGNORECASE):
             G.out.append("mpi_errno = MPIR_Err_return_session_init(errhandler_ptr, __func__, mpi_errno);")
         elif '_has_session' in func:
@@ -2003,6 +2124,8 @@ def dump_mpi_fn_fail(func):
             G.out.append("mpi_errno = MPIR_Err_return_comm_create_from_group(errhandler_ptr, __func__, mpi_errno);")
         elif '_has_group' in func:
             G.out.append("mpi_errno = MPIR_Err_return_group(%s_ptr, __func__, mpi_errno);" % func['_has_group'])
+        elif re.match(r'MPI_File_(delete|open|close)', func['name']):
+            G.out.append("mpi_errno = MPIO_Err_return_file(MPI_FILE_NULL, mpi_errno);")
         else:
             G.out.append("mpi_errno = MPIR_Err_return_comm(0, __func__, mpi_errno);")
 
@@ -2027,9 +2150,9 @@ def get_fn_fail_create_code(func):
             fmt = 'p'
         elif kind in fmt_codes:
             fmt = fmt_codes[kind]
-        elif mapping[kind] == "int":
+        elif mapping[kind] == "int" or mapping[kind] == "MPI_Fint":
             fmt = 'd'
-        elif mapping[kind] == "MPI_Aint":
+        elif mapping[kind] == "MPI_Aint" or mapping[kind] == "MPI_Offset":
             fmt = 'L'
         elif mapping[kind] == "MPI_Count":
             fmt = 'c'
diff --git a/maint/local_python/binding_common.py b/maint/local_python/binding_common.py
index f53603e5c05..edaff6ca917 100644
--- a/maint/local_python/binding_common.py
+++ b/maint/local_python/binding_common.py
@@ -119,7 +119,9 @@ def get_C_param(param, func, mapping):
     if param['func_type']:
         param_type = param['func_type']
         if mapping['_name'].startswith("BIG_"):
-            param_type += "_c"
+            # hard list the limited number of large count function types
+            if re.match(r'MPI_User_function|MPI_Datarep_conversion_function', param_type):
+                param_type += "_c"
 
     if param_type in G.mpix_symbols:
         param_type = re.sub(r'MPI_', 'MPIX_', param_type)
diff --git a/src/binding/abi/Makefile.mk b/src/binding/abi/Makefile.mk
index 197a48e3c19..199d860b5e9 100644
--- a/src/binding/abi/Makefile.mk
+++ b/src/binding/abi/Makefile.mk
@@ -9,6 +9,7 @@ include_HEADERS += src/binding/abi/mpi_abi.h
 
 mpi_abi_sources += \
     src/binding/abi/mpi_abi_util.c \
-    src/binding/abi/c_binding_abi.c
+    src/binding/abi/c_binding_abi.c \
+    src/binding/abi/io_abi.c
 
 endif BUILD_ABI_LIB
diff --git a/src/binding/c/async_api.txt b/src/binding/c/async_api.txt
index b3d0e0f9d01..9a948c2faef 100644
--- a/src/binding/c/async_api.txt
+++ b/src/binding/c/async_api.txt
@@ -13,6 +13,7 @@ MPIX_Async_get_state:
     .return: EXTRA_STATE
     async_thing: ASYNC_THING, [opaque pointer for async thing]
     .impl: direct
+    .skip: validate-ASYNC_THING
 {
     return MPIR_Async_thing_get_state(async_thing);
 }
@@ -22,6 +23,7 @@ MPIX_Async_spawn:
     poll_fn: FUNCTION, func_type=MPIX_Async_poll_function, [user defined poll function for progressing async things]
     extra_state: EXTRA_STATE, [extra state for callback function]
     stream: STREAM, [stream object]
+    .skip: validate-ASYNC_THING
 {
     mpi_errno = MPIR_Async_thing_spawn(async_thing, poll_fn, extra_state, stream_ptr);
     if (mpi_errno) goto fn_fail;
diff --git a/src/binding/c/datatype_api.txt b/src/binding/c/datatype_api.txt
index 63234904e81..495023c5dbb 100644
--- a/src/binding/c/datatype_api.txt
+++ b/src/binding/c/datatype_api.txt
@@ -201,15 +201,6 @@ MPI_Unpack_external:
     }
 }
 
-MPI_Register_datarep: not_implemented
-    .desc: Register a set of user-provided data conversion
-{
-    /* FIXME UNIMPLEMENTED */
-    mpi_errno =
-        MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE, __func__, __LINE__, MPI_ERR_OTHER,
-                             "**notimpl", 0);
-}
-
 MPI_Status_set_elements:
     .desc: Set the number of elements in a status
 {
diff --git a/src/binding/c/io_api.txt b/src/binding/c/io_api.txt
new file mode 100644
index 00000000000..02d3257b74f
--- /dev/null
+++ b/src/binding/c/io_api.txt
@@ -0,0 +1,208 @@
+# vim: set ft=c:
+
+MPI_File_open:
+    .desc: Opens a file
+    .skip: validate-amode
+
+MPI_File_close:
+    .desc: Closes a file
+
+MPI_File_delete:
+    .desc: Deletes a file
+
+MPI_File_c2f:
+    .desc: Translates a C file handle to a Fortran file handle
+    .impl: direct
+{
+#ifndef HAVE_ROMIO
+    return 0;
+#else
+    return MPIR_File_c2f_impl(file);
+#endif
+}
+
+MPI_File_f2c:
+    .desc: Translates a Fortran file handle to a C file handle
+    .impl: direct
+{
+#ifndef HAVE_ROMIO
+    return 0;
+#else
+    return MPIR_File_f2c_impl(file);
+#endif
+}
+
+MPI_File_sync:
+    .desc: Causes all previous writes to be transferred
+
+MPI_File_get_amode:
+    .desc: Returns the file access mode
+    .skip: global_cs
+
+MPI_File_get_atomicity:
+    .desc: Returns the atomicity mode
+    .skip: global_cs
+
+MPI_File_get_byte_offset:
+    .desc: Returns the absolute byte position in
+    .skip: global_cs
+
+MPI_File_get_type_extent:
+    .desc: Returns the extent of datatype in the file
+    .poly_impl: use_aint
+
+MPI_File_get_group:
+    .desc: Returns the group of processes that
+
+MPI_File_get_info:
+    .desc: Returns the hints for a file that are actually being used by MPI
+
+MPI_File_get_position:
+    .desc: Returns the current position of the
+
+MPI_File_get_position_shared:
+    .desc: Returns the current position of the
+
+MPI_File_get_size:
+    .desc: Returns the file size
+    .skip: global_cs
+
+MPI_File_get_view:
+    .desc: Returns the file view
+
+MPI_File_iread_all:
+    .desc: Nonblocking collective read using individual file pointer
+
+MPI_File_iread_at_all:
+    .desc: Nonblocking collective read using explicit offset
+
+MPI_File_iread_at:
+    .desc: Nonblocking read using explicit offset
+
+MPI_File_iread:
+    .desc: Nonblocking read using individual file pointer
+
+MPI_File_iread_shared:
+    .desc: Nonblocking read using shared file pointer
+
+MPI_File_iwrite_all:
+    .desc: Nonblocking collective write using individual file pointer
+
+MPI_File_iwrite_at_all:
+    .desc: Nonblocking collective write using explicit offset
+
+MPI_File_iwrite_at:
+    .desc: Nonblocking write using explicit offset
+
+MPI_File_iwrite:
+    .desc: Nonblocking write using individual file pointer
+
+MPI_File_iwrite_shared:
+    .desc: Nonblocking write using shared file pointer
+
+MPI_File_open:
+    .desc: Opens a file
+
+MPI_File_preallocate:
+    .desc: Preallocates storage space for a file
+
+MPI_File_read_at_all_begin:
+    .desc: Begin a split collective read using explicit offset
+
+MPI_File_read_at_all_end:
+    .desc: Complete a split collective read using
+
+MPI_File_read_all_begin:
+    .desc: Begin a split collective read using individual file pointer
+
+MPI_File_read_all:
+    .desc: Collective read using individual file pointer
+
+MPI_File_read_all_end:
+    .desc: Complete a split collective read using
+
+MPI_File_read_at_all:
+    .desc: Collective read using explicit offset
+
+MPI_File_read_at:
+    .desc: Read using explicit offset
+
+MPI_File_read:
+    .desc: Read using individual file pointer
+
+MPI_File_read_ordered_begin:
+    .desc: Begin a split collective read using shared file pointer
+
+MPI_File_read_ordered:
+    .desc: Collective read using shared file pointer
+
+MPI_File_read_ordered_end:
+    .desc: Complete a split collective read using shared file pointer
+
+MPI_File_read_shared:
+    .desc: Read using shared file pointer
+
+MPI_File_seek:
+    .desc: Updates the individual file pointer
+    .skip: validate-whence
+
+MPI_File_seek_shared:
+    .desc: Updates the shared file pointer
+    .skip: validate-whence
+
+MPI_File_set_atomicity:
+    .desc: Sets the atomicity mode
+
+MPI_File_set_info:
+    .desc: Sets new values for the hints associated with a file
+
+MPI_File_set_size:
+    .desc: Sets the file size
+
+MPI_File_set_view:
+    .desc: Sets the file view
+{ -- error_check -- disp
+    if (disp != MPI_DISPLACEMENT_CURRENT) {
+        MPIR_ERRTEST_ARGNEG(disp, "disp", mpi_errno);
+    }
+}
+
+MPI_File_write_at_all_begin:
+    .desc: Begin a split collective write using
+
+MPI_File_write_at_all_end:
+    .desc: Complete a split collective write using explicit offset
+
+MPI_File_write_all_begin:
+    .desc: Begin a split collective write using
+
+MPI_File_write_all:
+    .desc: Collective write using individual file pointer
+
+MPI_File_write_all_end:
+    .desc: Complete a split collective write using individual file pointer
+
+MPI_File_write_at_all:
+    .desc: Collective write using explicit offset
+
+MPI_File_write_at:
+    .desc: Write using explicit offset
+
+MPI_File_write:
+    .desc: Write using individual file pointer
+
+MPI_File_write_ordered_begin:
+    .desc: Begin a split collective write using shared file pointer
+
+MPI_File_write_ordered:
+    .desc: Collective write using shared file pointer
+
+MPI_File_write_ordered_end:
+    .desc: Complete a split collective write using shared file pointer
+
+MPI_File_write_shared:
+    .desc: Write using shared file pointer
+
+MPI_Register_datarep:
+    .desc: Register a set of user-provided data conversion
+    .poly_impl: separate
diff --git a/src/binding/cxx/buildiface b/src/binding/cxx/buildiface
index 1c6ae9e4784..83f4106235c 100755
--- a/src/binding/cxx/buildiface
+++ b/src/binding/cxx/buildiface
@@ -1736,14 +1736,7 @@ sub print_call_args {
 		if (defined($mytopclass{$lctype})) {
 		    $lctype = $mytopclass{$lctype};
 		}
-		# Handle the MPIO_Request problem (temp until ROMIO uses
-		# MPI_Requests)
-		$cast = "";
-		if ($parm =~ /MPI_Request/ &&
-		    $class eq "file") {
-		    $cast = "(MPIO_Request *)";
-		}
-		print $OUTFD "$cast&(v$count.the_real_$lctype)";
+		print $OUTFD "&(v$count.the_real_$lctype)";
 	    }
 	    else {
 		print $OUTFD "&v$count";
diff --git a/src/glue/romio/all_romio_symbols b/src/glue/romio/all_romio_symbols
index 84ee707b680..fc6696376fb 100755
--- a/src/glue/romio/all_romio_symbols
+++ b/src/glue/romio/all_romio_symbols
@@ -138,7 +138,9 @@ while (<FD>) {
 
     print OUTFD "${tab}{\n";
     for ($x = 0; $x <= $#arglist; $x++) {
-	print OUTFD "${tab}${tab}$arglist[$x]";
+        my $a = $arglist[$x];
+        $a =~ s/MPIO_Request/MPI_Request/;
+	print OUTFD "${tab}${tab}$a";
 	if ($arglist[$x] =~ /\*/) {
 	    print OUTFD " = NULL;\n";
 	}
diff --git a/src/include/mpi.h.in b/src/include/mpi.h.in
index 3e28e5307db..80b71af2ff2 100644
--- a/src/include/mpi.h.in
+++ b/src/include/mpi.h.in
@@ -504,6 +504,9 @@ static const MPI_Datatype mpich_mpi_offset MPICH_ATTR_TYPE_TAG(MPI_Offset) = MPI
 #define MPI_MAX_PSET_NAME_LEN  256
 #define MPI_MAX_INFO_KEY       255
 #define MPI_MAX_INFO_VAL      1024
+#define MPI_MAX_DATAREP_STRING 128
+
+#define MPI_DISPLACEMENT_CURRENT ((MPI_Offset) -54278278)
 
 /* MPI-3 window flavors */
 typedef enum MPIR_Win_flavor {
@@ -622,6 +625,21 @@ enum MPIR_Combiner_enum {
 #define MPI_THREAD_SERIALIZED 2
 #define MPI_THREAD_MULTIPLE 3
 
+/* MPI-IO constants */
+#define MPI_MODE_RDONLY              2  /* ADIO_RDONLY */
+#define MPI_MODE_RDWR                8  /* ADIO_RDWR  */
+#define MPI_MODE_WRONLY              4  /* ADIO_WRONLY  */
+#define MPI_MODE_CREATE              1  /* ADIO_CREATE */
+#define MPI_MODE_EXCL               64  /* ADIO_EXCL */
+#define MPI_MODE_DELETE_ON_CLOSE    16  /* ADIO_DELETE_ON_CLOSE */
+#define MPI_MODE_UNIQUE_OPEN        32  /* ADIO_UNIQUE_OPEN */
+#define MPI_MODE_APPEND            128  /* ADIO_APPEND */
+#define MPI_MODE_SEQUENTIAL        256  /* ADIO_SEQUENTIAL */
+
+#define MPI_SEEK_SET            600
+#define MPI_SEEK_CUR            602
+#define MPI_SEEK_END            604
+
 /* MPI's error classes */
 #define MPI_SUCCESS          0      /* Successful return code */
 /* Communication argument parameters */
@@ -1027,8 +1045,6 @@ typedef struct MPIX_Iov {
 /* We require that the C compiler support prototypes */
 #include <mpi_proto.h>
 
-@HAVE_ROMIO@
-
 /* The f2c and c2f APIs exist as real functions, but these macros allows
  * for backward MPICH ABI compatibility.
  */
diff --git a/src/mpi/Makefile.mk b/src/mpi/Makefile.mk
index 8d7c4236ee6..b0fa296ff04 100644
--- a/src/mpi/Makefile.mk
+++ b/src/mpi/Makefile.mk
@@ -22,30 +22,5 @@ include $(top_srcdir)/src/mpi/topo/Makefile.mk
 include $(top_srcdir)/src/mpi/stream/Makefile.mk
 include $(top_srcdir)/src/mpi/threadcomm/Makefile.mk
 
-if BUILD_ROMIO
-SUBDIRS += src/mpi/romio
-DIST_SUBDIRS += src/mpi/romio
-MANDOC_SUBDIRS += src/mpi/romio
-HTMLDOC_SUBDIRS += src/mpi/romio
-
-# This was previously a hard copy (not a symlink) performed by config.status
-# (specified via AC_CONFIG_COMMANDS).  Ideally we would eliminate this "copy"
-# altogether and just set -Iromio_include_dir, but MPE2's build system uses
-# $(top_builddir)/bin/mpicc that can't handle more than one include dir.
-#
-# Using a symlink allows us to avoid trying to capture the full dependency chain
-# of MPICH/mpio.h --> ROMIO/mpio.h --> ROMIO/mpio.h.in --> ROMIO/config.status --> ...MORE_AUTOTOOLS...
-BUILT_SOURCES += $(top_builddir)/src/include/mpio.h
-$(top_builddir)/src/include/mpio.h: $(top_builddir)/src/mpi/romio/include/mpio.h
-	if test ! -h $(top_builddir)/src/include/mpio.h ; then \
-	    rm -f $(top_builddir)/src/include/mpio.h ; \
-	    ( cd $(top_builddir)/src/include &&       \
-	        $(LN_S) ../mpi/romio/include/mpio.h ) ; \
-	fi
-
-DISTCLEANFILES += $(top_builddir)/src/include/mpio.h
-
-endif BUILD_ROMIO
-
 # dir is present but currently intentionally unbuilt
 #include $(top_srcdir)/src/mpi/io/Makefile.mk
diff --git a/src/mpi/errhan/errnames.txt b/src/mpi/errhan/errnames.txt
index 9198104035d..35c8c01fac6 100644
--- a/src/mpi/errhan/errnames.txt
+++ b/src/mpi/errhan/errnames.txt
@@ -923,8 +923,6 @@ is too big (> MPIU_SHMW_GHND_SZ)
 **mpi_attr_get %C %d %p %p:MPI_Attr_get(%C, keyval=%d, attr_value=%p, flag=%p) failed
 **mpi_attr_delete:MPI_Attr_delete failed
 **mpi_attr_delete %C %d:MPI_Attr_delete(%C, keyval=%d) failed
-**mpi_register_datarep:MPI_Register_datarep failed
-**mpi_register_datarep %s %p %p %p %p:MPI_Register_datarep(datarep=%s, read_conversion_fn=%p, write_conversion_fn=%p, dtype_file_extent_fn=%p, extra_state=%p) failed
 
 # MPIR functions
 
diff --git a/src/mpi/romio/configure.ac b/src/mpi/romio/configure.ac
index da28bbd5761..6a42e3254d1 100644
--- a/src/mpi/romio/configure.ac
+++ b/src/mpi/romio/configure.ac
@@ -21,7 +21,7 @@ dnl AC_CONFIG_AUX_DIR(../../../confdb)
 dnl Set the directory that contains the required install-sh, config.sub,
 dnl and config.guess .  Make sure that these are updated (in MPICH, use
 dnl the top-level confdb files).  This separate directory is used for
-dnl the moment to allow ROMIO to be separatedly distributed.
+dnl the moment to allow ROMIO to be separately distributed.
 dnl scripts.
 AC_CONFIG_AUX_DIR([confdb])
 AC_CONFIG_MACRO_DIR([confdb])
@@ -105,6 +105,8 @@ AC_ARG_VAR([FROM_OMPI],[set to "yes" if building ROMIO inside of Open MPI])
 FROM_OMPI=${FROM_OMPI:-no}
 if test "$FROM_OMPI" = 1 ; then FROM_OMPI=yes ; fi
 
+AM_CONDITIONAL([BUILD_BINDING], [test "$FROM_MPICH" != "yes"])
+
 # MPL
 AC_ARG_VAR([MPLLIBNAME],[can be used to override the name of the MPL library (default: "mpl")])
 MPLLIBNAME=${MPLLIBNAME:-"mpl"}
@@ -516,8 +518,6 @@ fi
 ])
 if test "$pac_cv_int_hold_pointer" != yes ; then
     AC_DEFINE(INT_LT_POINTER,1,[Define if int smaller than pointer])
-    dnl Switch to a conforming name (start with HAVE or USE)
-    AC_DEFINE(HAVE_INT_LT_POINTER,1,[Define if int smaller than pointer])
 fi
 
 # LL is the printf-style format name for output of a MPI_Offset.
@@ -693,7 +693,9 @@ AM_CONDITIONAL([BUILD_MPIO_ERRHAN],[false])
 
 # if we don't have weak symbol support, we must build a separate convenience
 # library in order to provide the "PMPI_" symbols
-AM_CONDITIONAL([BUILD_PROFILING_LIB],[test "x$HAVE_WEAK_SYMBOLS" = "x0"])
+# MPICH will generate separate binding layer, thus we can skip PROFILING LIB
+# if build inside MPICH
+AM_CONDITIONAL([BUILD_PROFILING_LIB],[test "x$HAVE_WEAK_SYMBOLS" = "x0" -a "$FROM_MPICH" != "yes"])
 # disable visibility if building profiling library
 if test "x$HAVE_WEAK_SYMBOLS" = "x0" ; then
     VISIBILITY_CFLAGS=""
diff --git a/src/mpi/romio/include/mpio.h.in b/src/mpi/romio/include/mpio.h.in
index c0322fdb1ad..6aa84e38b6e 100644
--- a/src/mpi/romio/include/mpio.h.in
+++ b/src/mpi/romio/include/mpio.h.in
@@ -82,6 +82,8 @@ typedef int MPI_Fint;
 # define MPI_MAX_INFO_VAL      1024
 #endif
 
+#ifndef MPI_MODE_RDONLY
+/* only define these if they are not defined in mpi.h */
 #define MPI_MODE_RDONLY              2  /* ADIO_RDONLY */
 #define MPI_MODE_RDWR                8  /* ADIO_RDWR  */
 #define MPI_MODE_WRONLY              4  /* ADIO_WRONLY  */
@@ -92,26 +94,28 @@ typedef int MPI_Fint;
 #define MPI_MODE_APPEND            128  /* ADIO_APPEND */
 #define MPI_MODE_SEQUENTIAL        256  /* ADIO_SEQUENTIAL */
 
-#define MPI_DISPLACEMENT_CURRENT   -54278278
-
-#ifndef MPICH
-/* FIXME: Make sure that we get a consistent definition of MPI_FILE_NULL
-	in MPICH */
-/* MPICH defines null object handles differently */
-#define MPI_FILE_NULL           ((MPI_File) 0)
-#endif
-#define MPIO_REQUEST_NULL       ((MPIO_Request) 0)
-
 #define MPI_SEEK_SET            600
 #define MPI_SEEK_CUR            602
 #define MPI_SEEK_END            604
 
+#define MPI_DISPLACEMENT_CURRENT   -54278278
+
 /* Open MPI: don't define MPI_MAX_DATAREP_STRING here; it's defined in
    OMPI's mpi.h. */
 #ifndef OPEN_MPI
 #define MPI_MAX_DATAREP_STRING  128
 #endif
 
+#endif
+
+#ifndef MPICH
+/* FIXME: Make sure that we get a consistent definition of MPI_FILE_NULL
+	in MPICH */
+/* MPICH defines null object handles differently */
+#define MPI_FILE_NULL           ((MPI_File) 0)
+#endif
+#define MPIO_REQUEST_NULL       ((MPIO_Request) 0)
+
 #ifndef HAVE_MPI_DARRAY_SUBARRAY
 /* *INDENT-OFF* */
 @HAVE_MPI_DARRAY_SUBARRAY@
diff --git a/src/mpi/romio/mpi-io/Makefile.mk b/src/mpi/romio/mpi-io/Makefile.mk
index 3bde1b94256..621fa85016f 100644
--- a/src/mpi/romio/mpi-io/Makefile.mk
+++ b/src/mpi/romio/mpi-io/Makefile.mk
@@ -9,6 +9,7 @@ include $(top_srcdir)/mpi-io/fortran/Makefile.mk
 AM_CPPFLAGS += -I$(top_builddir)/mpi-io -I$(top_srcdir)/mpi-io
 noinst_HEADERS += mpi-io/mpioimpl.h mpi-io/mpioprof.h
 
+if BUILD_BINDING
 romio_mpi_sources +=          \
     mpi-io/close.c            \
     mpi-io/delete.c           \
@@ -68,6 +69,7 @@ romio_mpi_sources +=          \
     mpi-io/write_ordb.c       \
     mpi-io/write_orde.c       \
     mpi-io/write_sh.c
+endif BUILD_BINDING
 
 
 # non-MPI/PMPI sources that will be included in libromio
diff --git a/src/mpi/romio/mpi-io/glue/mpich/mpio_file.c b/src/mpi/romio/mpi-io/glue/mpich/mpio_file.c
index 0ffb5efcf1d..54906393b8a 100644
--- a/src/mpi/romio/mpi-io/glue/mpich/mpio_file.c
+++ b/src/mpi/romio/mpi-io/glue/mpich/mpio_file.c
@@ -39,7 +39,7 @@ void MPIO_File_free(MPI_File * mpi_fh)
 MPI_File MPIO_File_f2c(MPI_Fint fh)
 {
 #ifndef INT_LT_POINTER
-    return (MPI_File) ((void *) fh);
+    return (MPI_File) ((void *) (intptr_t) fh);
     /* the extra cast is to get rid of a compiler warning on Exemplar.
      * The warning is because MPI_File points to a structure containing
      * longlongs, which may be 8-byte aligned. But MPI_Fint itself
@@ -58,7 +58,7 @@ MPI_File MPIO_File_f2c(MPI_Fint fh)
 MPI_Fint MPIO_File_c2f(MPI_File fh)
 {
 #ifndef INT_LT_POINTER
-    return (MPI_Fint) fh;
+    return (MPI_Fint) (intptr_t) fh;
 #else
     int i;