Re: [Swig-user] argout for Java/C#?

[William S F. <ws...@fu...>]

riemann tolow wrote:
> Hi,
> 
> I am trying to write Java/C# typemaps for std::vector. They should map the 
> vector into the language arrays. I know that there is an overhead of copying 
> between the types but that's ok for me. I intend to use the std::vector 
> wrapper generated by SWIG to do the marshal from/to array.
> 
> The problem I am facing is that there is no way to insert code before the 
> $jnicall/$imcall or after (unless uisng javaout/csout). The typemaps 
> javain/csin are only one liner code that is embedded in the calls 
> $jnicall/$imcall.
> 
> For example in C#:
> %include std_vector.i
> %typemap(cstype) std::vector<double>"double[]";
> %typemap(csin) std::vector<double> "DoubleVector.getCPtr(new 
> DoubleVector($csinput))";
> %template(DoubleVector) std::vector<double>
> 
> If the vector is used as an output argument, there is no way to copy back 
> the result into the array (assume there is a "&" and "ref")
> 
> Am I doing it wrong? Is there some other way to do it? I suppose using the 
> "in" typemap give better possibilities but this requires using the native 
> API of the language.
> 

I'm afraid that the typemaps containing C#/Java code do not allow for
temporary variables for parameters. For Java, use helper methods in the
javain typemap, eg see the CheckForNaN example in Java.html. For C# this
is a known problem for C# ref and out types. We need to modify SWIG to
work better with ref and out values.

What I propose is making changes to the csin/javain typemaps addin in
two new optional attributes, 'pre' and 'post'. The idea is that the
'pre' attribute contains code that gets added before the $imcall and
'post' contains code executed after the $imcall call.

Let's presume you are wrapping:

void f(std::vector<double> & v);
%template(DoubleVector) std::vector<double>;

You probably want the C# code to look like this:

  public static bool f(double[] v) {
    DoubleVector dv = new DoubleVector();
    foreach (double d in v) {
      dv.Add(d);
    }
    bool ret = examplePINVOKE.f(DoubleVector.getCPtr(dv));
    if (examplePINVOKE.SWIGPendingException.Pending)
      throw examplePINVOKE.SWIGPendingException.Retrieve();
    for (int i=0; i<v.Length; ++i) {
      v[i] = dv[i];
    }
    return ret;
  }

I can't see this being possible because it splits the current csout bool
typemap (it inserts code between the "bool ret=..." and "return ret".
Instead we could generate this:

  public static bool f(double[] v) {
    DoubleVector dv = new DoubleVector();
    foreach (double d in v) {
      dv.Add(d);
    }
    try {
      bool ret = examplePINVOKE.f(DoubleVector.getCPtr(dv));
      if (examplePINVOKE.SWIGPendingException.Pending)
        throw examplePINVOKE.SWIGPendingException.Retrieve();
      return ret;
    } finally {
      for (int i=0; i<v.Length; ++i) {
        v[i] = dv[i];
      }
    }
  }

Current csin typemap:
%typemap(csin) SWIGTYPE &
  "$csclassname.getCPtr($csinput)"

Proposed new attributes in the csin typemap:
%typemap(csin,
         pre="    DoubleVector dv = new DoubleVector();\n    foreach
(double d in $input) {    \n dv.Add(d);\n    }"
         post="    for (int i=0; i<$input.Length; ++i) {\n    $input[i]
= dv[i];\n    }"
) std::vector<double> &
  "$csclassname.getCPtr(dv)"

It looks a bit horrible formatted in this email, but is simply:

%typemap(csin, pre="...code...", post="...code...") std::vector<double>&
  "$csclassname.getCPtr(dv)"

and uses \n for line delimeters. The variable names will probably need
$argnum or $input adding to them in the typemap, so that there are no
duplicate variable names in when using a 2nd parameter of the same type.

Fixing this issue is a bit overdue and I'll probably code it up in the
next few days unless you see any problems.

William