binding: generate embiggened functions

We will call dump_mpi_c twice for functions with POLY types. First we
call with map_type="SMALL-poly", then call with map_type="BIG".
Functions without POLY types will just call dump_mpi_c once with
map_type="SMALL".

First of all, if there is a custom "large_count" block, we'll always
use that in large version.

If a POLY function only has scalar POLY type, we assume the
implementation use MPI_Aint -- need fix any exceptions.

If a POLY function has scalar output or inout POLY type, we assume
internally use MPI_Aint, and python generrates temporary MPI_Aint
variable to copy in and out value in the *small* version.

If the function uses POLY arrays, we currently always assume there is a
large mpir impl version with `_c_impl` suffix, and will call that. The
compilation will fail if that (_c_impl) routine is missing.

We will also rely on the -Wconversion warnings to find all cases when we
assume internal version use MPI_Aint but it doesn't.

maint/gen_binding_c.py

@@ -48,10 +48,13 @@ def main():
         else:
             G.out = []
             G.err_codes = {}
-            mapping = G.MAPS['SMALL_C_KIND_MAP']
 
             # dumps the code to G.out array
-            dump_mpi_c(func, mapping)
+            if function_has_POLY_parameters(func):
+                dump_mpi_c(func, "SMALL-poly")
+                dump_mpi_c(func, "BIG")
+            else:
+                dump_mpi_c(func, "SMALL")
 
             file_path = get_func_file_path(func, binding_dir)
             dump_c_file(file_path, G.out)

maint/local_python/binding_c.py

@@ -12,11 +12,28 @@
 
 # ---- top-level routines ----
 
-def dump_mpi_c(func, mapping):
+def dump_mpi_c(func, map_type="SMALL"):
     """Dumps the function's C source code to G.out array"""
+    if map_type == "SMALL":
+        # single version, business as usual
+        mapping = G.MAPS['SMALL_C_KIND_MAP']
+    elif map_type == "SMALL-poly":
+        # small version of poly type, difference in man page and body of routines
+        mapping = G.MAPS['SMALL_C_KIND_MAP']
+    elif map_type == "BIG":
+        # big version of poly type, difference in man page and body of routines
+        mapping = G.MAPS['BIG_C_KIND_MAP']
+    else:
+        raise Exception("dump_mpi_c is called with map_type = " + map_type)
+
+    # save as attribute so we don't have pass it around
+    func['_map_type'] = map_type
+
     check_func_directives(func)
     filter_c_parameters(func)
 
+    check_large_parameters(func)
+
     G.mpi_declares.append(get_declare_function(func, mapping))
 
     check_params_with_large_only(func, mapping)
@@ -48,6 +65,11 @@ def dump_mpi_c(func, mapping):
     else:
         dump_function(func, mapping, kind="normal")
 
+    if "decl" in func:
+        push_impl_decl(func, func['decl'])
+    elif 'body' not in func and 'impl' not in func:
+        push_impl_decl(func)
+
 def get_func_file_path(func, root_dir):
     file_path = None
     dir_path = root_dir + '/' + func['dir']
@@ -247,6 +269,9 @@ def process_func_parameters(func, mapping):
     # Note: we'll attach the lists to func at the end
     validation_list, handle_ptr_list, impl_arg_list, impl_param_list = [], [], [], []
 
+    # init to empty list or we will have double entries due to being called twice (small and large)
+    func['_has_handle_out'] = []
+
     func_name = func['name']
     n = len(func['c_parameters'])
     i = 0
@@ -296,7 +321,7 @@ def process_func_parameters(func, mapping):
                     t += ","
                 t += temp_p['name']
                 impl_arg_list.append(temp_p['name'])
-                impl_param_list.append(get_C_param(temp_p, mapping))
+                impl_param_list.append(get_impl_param(temp_p, mapping))
             validation_list.append({'kind': group_kind, 'name': t})
             i += group_count
             continue
@@ -437,11 +462,9 @@ def process_func_parameters(func, mapping):
             impl_arg_list.append(name + "_ptr")
             impl_param_list.append("%s *%s_ptr" % (G.handle_mpir_types[kind], name))
         elif do_handle_ptr == 2:
-            if '_has_handle_out' in func:
-                # Just MPI_Comm_idup[_with_info] have 2 handle output, but anyway...
-                func['_has_handle_out'].append(p)
-            else:
-                func['_has_handle_out'] = [p]
+            # Just MPI_Comm_idup[_with_info] have 2 handle output, but use a list anyway
+            func['_has_handle_out'].append(p)
+
             impl_arg_list.append('&' + name + "_ptr")
             impl_param_list.append("%s **%s_ptr" % (G.handle_mpir_types[kind], name))
         elif do_handle_ptr == 3:
@@ -460,7 +483,7 @@ def process_func_parameters(func, mapping):
             impl_param_list.append("%s *%s_ptr" % (G.handle_mpir_types[kind], name))
         else:
             impl_arg_list.append(name)
-            impl_param_list.append(get_C_param(p, mapping))
+            impl_param_list.append(get_impl_param(p, mapping))
         i += 1
 
     if RE.match(r'MPI_(Wait|Test)', func_name):
@@ -641,10 +664,35 @@ def dump_function(func, mapping, kind):
     G.out.append("DEDENT")
     G.out.append("}")
 
-    if "decl" in func:
-        push_impl_decl(func, func['decl'])
-    elif 'body' not in func and 'impl' not in func:
-        push_impl_decl(func)
+def check_large_parameters(func):
+    # whether we need *separate* large version
+    # whether we always use large version internally
+    func['_need_large_impl'] = False
+    func['_impl_uses_large'] = False
+    # for large parameters that needs copy in & out, e.g. output, inout, array
+    func['_poly_in_list'] = []
+    func['_poly_out_list'] = []
+    func['_poly_inout_list'] = []
+    for p in func['c_parameters']:
+        if RE.match(r'POLY', p['kind']):
+            if p['param_direction'] == 'out':
+                func['_poly_out_list'].append(p)
+            elif p['param_direction'] == 'inout':
+                # MPI_{Pack,Unpack}[_external]
+                func['_poly_inout_list'].append(p)
+            elif p['length']:
+                # For perforance reason, default internal code will use small
+                # counts array to avoid extra array copy for the common case
+                func['_need_large_impl'] = True
+            else:
+                func['_poly_in_list'].append(p)
+
+    if 'code-large_count' in func:
+        func['_need_large_impl'] = True
+
+    if not func['_need_large_impl']:
+        if func['_poly_in_list'] or func['_poly_out_list'] or func['_poly_inout_list']:
+            func['_impl_uses_large'] = True
 
 def dump_function_normal(func, state_name, mapping):
     func['exit_routines'] = []
@@ -724,10 +772,20 @@ def dump_function_normal(func, state_name, mapping):
 
     check_early_returns(func)
     G.out.append("")
+
+    # out or inout POLY scalar may need temporary variable to copy in and out values
+    dump_poly_pre_filter(func)
+
     G.out.append("/* ... body of routine ... */")
     if 'body' in func:
-        for l in func['body']:
-            G.out.append(l)
+        if func['_map_type'] == "BIG" and func['_need_large_impl']:
+            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 and RE.match(r'mpid', func['impl'], re.IGNORECASE):
         dump_body_impl(func, "mpid")
     elif func['dir'] == 'coll':
@@ -736,6 +794,9 @@ def dump_function_normal(func, state_name, mapping):
         dump_body_impl(func, "mpir")
 
     G.out.append("/* ... end of body of routine ... */")
+
+    dump_poly_post_filter(func)
+
     G.out.append("")
     G.out.append("fn_exit:")
     for l in func['exit_routines']:
@@ -757,15 +818,61 @@ def dump_function_normal(func, state_name, mapping):
         dump_mpi_fn_fail(func, mapping)
     G.out.append("goto fn_exit;")
 
+def dump_poly_pre_filter(func):
+    if not func['_need_large_impl']:
+        if func['_map_type'] != "BIG":
+            # internal code will use MPI_Aint
+            for p in func['_poly_out_list']:
+                G.out.append("MPI_Aint %s_c;" % p['name'])
+            for p in func['_poly_inout_list']:
+                G.out.append("MPI_Aint %s_c = *%s;" % (p['name'], p['name']))
+
+            for p in func['_poly_out_list'] + func['_poly_inout_list']:
+                for i in range(len(func['_impl_arg_list'])):
+                    if func['_impl_arg_list'][i] == p['name']:
+                        func['_impl_arg_list'][i] = "&%s_c" % p['name']
+
+def dump_poly_post_filter(func):
+    if not func['_need_large_impl']:
+        if func['_map_type'] != "BIG":
+            for p in func['_poly_out_list'] + func['_poly_inout_list']:
+                # check if value fits
+                dump_if_open("%s_c > INT_MAX" % p['name'])
+                G.out.append("*%s = MPI_UNDEFINED;" % p['name'])
+                errname = "\"**too_big_for_output\", \"**too_big_for_output %s\""
+                G.out.append("mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE,")
+                G.out.append("                                 __func__, __LINE__, MPI_ERR_ARG,")
+                G.out.append("                                 %s, \"%s\");" % (errname, p['name']))
+                G.out.append("goto fn_fail;")
+                dump_if_close()
+                # copy the value
+                G.out.append("*%s = %s_c;" % (p['name'], p['name']))
+
 def push_impl_decl(func, impl_name=None):
     if not impl_name:
         impl_name = re.sub(r'^MPIX?_', 'MPIR_', func['name']) + "_impl"
+
+    if func['_map_type'] != "BIG":
+        # SMALL
+        if func['_impl_uses_large']:
+            # skip since large interface will be used
+            return
+    else:
+        # BIG
+        if func['_need_large_impl']:
+            # add suffix to differentiate
+            impl_name = re.sub(r'_impl$', '_c_impl', impl_name)
+        elif not func['_impl_uses_large']:
+            # skip since small interface will be used
+            return
+
     if func['_impl_param_list']:
         params = ', '.join(func['_impl_param_list'])
         if func['dir'] == 'coll' and not RE.match(r'MPI_I', func['name']):
             params = params + ", MPIR_Errflag_t *errflag"
     else:
         params="void"
+
     G.impl_declares.append("int %s(%s);" % (impl_name, params))
 
 def dump_CHECKENUM(var, errname, t, type="ENUM"):
@@ -798,11 +905,17 @@ def dump_body_coll(func):
     if not name.startswith('I'):
         G.out.append("MPIR_Errflag_t errflag = MPIR_ERR_NONE;")
         args = args + ", " + "&errflag"
-    G.out.append("if (%s || (%s && %s)) {" % (cond_a, cond_b1, cond_b2))
-    dump_line_with_break("    mpi_errno = MPID_%s(%s);" % (name, args))
-    G.out.append("} else {")
-    dump_line_with_break("    mpi_errno = MPIR_%s_impl(%s);" % (name, args))
-    G.out.append("}")
+
+    if func['_map_type'] == "BIG" and func['_need_large_impl']:
+        # BIG v-count only do MPIR_Xxx_c_impl for now
+        dump_line_with_break("    mpi_errno = MPIR_%s_c_impl(%s);" % (name, args))
+    else:
+        G.out.append("if (%s || (%s && %s)) {" % (cond_a, cond_b1, cond_b2))
+        dump_line_with_break("    mpi_errno = MPID_%s(%s);" % (name, args))
+        G.out.append("} else {")
+        dump_line_with_break("    mpi_errno = MPIR_%s_impl(%s);" % (name, args))
+        G.out.append("}")
+
     dump_error_check("")
     if name.startswith('I'):
         G.out.append("if (!request_ptr) {")
@@ -812,7 +925,7 @@ def dump_body_coll(func):
 
 def dump_body_impl(func, prefix='mpir'):
     # mpi_errno = MPIR_Xxx_impl(...);
-    if '_has_handle_out' in func:
+    if func['_has_handle_out']:
         for p in func['_has_handle_out']:
             (name, kind) = (p['name'], p['kind'])
             mpir_type = G.handle_mpir_types[kind]
@@ -824,17 +937,19 @@ def dump_body_impl(func, prefix='mpir'):
         if p['kind'] == "DATATYPE" and p['param_direction'] == 'out':
             G.out.append("*%s = MPI_DATATYPE_NULL;" % p['name'])
 
+    impl = func['name']
+    if func['_map_type'] == "BIG" and func['_need_large_impl']:
+        impl += '_c'
     if prefix == 'mpid':
-        impl = func['name']
         impl = re.sub(r'^MPIX?_', 'MPID_', impl)
     else:
-        impl = func['name'] + "_impl"
-        impl = re.sub(r'^MPIX?_', 'MPIR_', impl)
+        impl = re.sub(r'^MPIX?_', 'MPIR_', impl) + "_impl"
+
     args = ", ".join(func['_impl_arg_list'])
     dump_line_with_break("mpi_errno = %s(%s);" % (impl, args))
     dump_error_check("")
 
-    if '_has_handle_out' in func:
+    if func['_has_handle_out']:
         # assuming we are creating new handle(s)
         for p in func['_has_handle_out']:
             (name, kind) = (p['name'], p['kind'])
@@ -1633,6 +1748,13 @@ def get_C_param(param, mapping):
 
     return s
 
+def get_impl_param(param, mapping):
+    s = get_C_param(param, mapping)
+    if RE.match(r'POLY', param['kind']):
+        # internally we always use MPI_Aint for large poly type
+        s = re.sub(r'MPI_Count', 'MPI_Aint', s)
+    return s
+
 def get_polymorph_param_and_arg(s):
     # s is a polymorph spec, e.g. "MPIR_Attr_type attr_type=MPIR_ATTR_PTR"
     # Note: we assum limit of single extra param for now (which is sufficient)

maint/local_python/mpi_api.py

@@ -175,3 +175,9 @@ def parse_param_attributes(p):
         p['constant'] = True
     else:
         p['constant'] = False
+
+def function_has_POLY_parameters(func):
+    for p in func['parameters']:
+        if p['kind'].startswith('POLY'):
+            return True
+    return False