Missing Fortran named constants for predefined datatypes

[kawashima-fj]

@bosilca, I found a Fortran/datatype bug and am trying to fix it but I want to hear your opinion.
The bug is that a Fortran program cannot be compiled if it uses some predefined datatypes.
The cause is that those datatypes are not defined or incorrectly defined in include/mpif-handles.h, which is created by ompi/include/mpif-values.pl.

The missing datatypes are:

• MPI_C_BOOL
• MPI_LONG_LONG

The incorrectly defined datatypes are:

• MPI_LONG_DOUBLE_INT
• MPI_CXX_FLOAT_COMPLEX
• MPI_CXX_DOUBLE_COMPLEX
• MPI_CXX_LONG_DOUBLE_COMPLEX

The incorrectly defined datatypes can be fixed easily:

--- a/ompi/include/mpif-values.pl
+++ b/ompi/include/mpif-values.pl
@@ -166,14 +166,14 @@ $handles->{MPI_DOUBLE} = 46;
 $handles->{MPI_LONG_DOUBLE} = 47;
 $handles->{MPI_FLOAT_INT} = 48;
 $handles->{MPI_DOUBLE_INT} = 49;
-$handles->{MPI_LONGDBL_INT} = 50;
+$handles->{MPI_LONG_DOUBLE_INT} = 50;
 $handles->{MPI_LONG_INT} = 51;
 $handles->{MPI_2INT} = 52;
 $handles->{MPI_SHORT_INT} = 53;
 $handles->{MPI_CXX_BOOL} = 54;
-$handles->{MPI_CXX_CPLEX} = 55;
-$handles->{MPI_CXX_DBLCPLEX} = 56;
-$handles->{MPI_CXX_LDBLCPLEX} = 57;
+$handles->{MPI_CXX_FLOAT_COMPLEX} = 55;
+$handles->{MPI_CXX_DOUBLE_COMPLEX} = 56;
+$handles->{MPI_CXX_LONG_DOUBLE_COMPLEX} = 57;
 $handles->{MPI_INT8_T} = 58;
 $handles->{MPI_UINT8_T} = 59;
 $handles->{MPI_INT16_T} = 60;

The problem is the two missing datatypes.

MPI_C_BOOL should be inserted between MPI_OFFSET and MPI_C_COMPLEX for natural ordering in ompi/include/mpif-values.pl (they are added in MPI-2.2; see ompi/datatype/ompi_datatype_module.c). But doing so will generate a incompatibility in user's Fortran program because indices of MPI_C_COMPLEX and others will be shifted. We have three options. Which will we take?

1. Insert MPI_C_BOOL between MPI_OFFSET and MPI_C_COMPLEX, and merge it to v2.x branch before v2.0.0 is released.
2. Insert MPI_C_BOOL as the last line (index: 73).
3. Insert MPI_C_BOOL between MPI_OFFSET and MPI_C_COMPLEX for master. And
   insert MPI_C_BOOL as the last line (index: 73) for v2.x

MPI_LONG_LONG is a synonym of MPI_LONG_LONG_INT and is not defined separately in ompi/datatype/ompi_datatype_module.c. So simply adding a line below in ompi/include/mpif-values.pl will solve the problem. But MPI_C_FLOAT_COMPLEX is also a synonym of MPI_C_COMPLEX and it is defined separately in ompi/datatype/ompi_datatype_module.c and has own index 68. The treatment is inconsistent between MPI_LONG_LONG and MPI_C_FLOAT_COMPLEX! Should we change the treatment of either synonym or keep the inconsistency?

--- a/ompi/include/mpif-values.pl
+++ b/ompi/include/mpif-values.pl
@@ -160,6 +160,7 @@ $handles->{MPI_UNSIGNED} = 40;
 $handles->{MPI_LONG} = 41;
 $handles->{MPI_UNSIGNED_LONG} = 42;
 $handles->{MPI_LONG_LONG_INT} = 43;
+$handles->{MPI_LONG_LONG} = $handles->{MPI_LONG_LONG_INT};
 $handles->{MPI_UNSIGNED_LONG_LONG} = 44;
 $handles->{MPI_FLOAT} = 45;
 $handles->{MPI_DOUBLE} = 46;

The program I checked the Fortran binding with is in my gist. This program needs --enable-mpi-cxx on configure for MPI_CXX_* datatypes.

[kawashima-fj]

@bosilca Could you comment?

[kawashima-fj]

@bosilca Could you comment?

[bosilca]

@kawashima-fj the fortran parameters must have the same values as the index of the C datatype in the ompi_datatype_f_to_c_table array, which normally corresponds to the OMPI datatype index, aka. OMPI_DATATYPE_MPI_* define (this is needed so that the PMPI_Type_f2c macro can convert between C and Fortran objects).

So your proposed fix for the first and third issues are correct. For the second issue I will suggest 1 as the best approach, or we will have a binary incompatibility between the trunk and the 2.0.