a possible bug related to '++' and 'cast'

[hujiajie]

Steps to reproduce:

var pa = new ParallelArray(1, 2, 3, 4);
pa.map(function (x) {
  var i = 0;
  i++;
  for (var j = 0; j < 1; j++) { i++; }
  return 0;
});

Expected result:

[0, 0, 0, 0]

Actual result:

The OpenCL compiler failed. Log was `:2:508: error: assignment to cast is illegal, lvalue casts are not supported

Remarks: Here is the generated kernel string:

__kernel void RT_nameless(__global int *_FAILRET,  __global double* opThisVect , int opThisVect__offset, __global double* retVal, int retVal__offset)
{
  int _sel_idx_tmp;
  int _FAIL = 0;
  int _id_0 = (int)get_global_id(0);
  int _writeoffset = 0+1 * _id_0;
  int _readoffset = 0 + _writeoffset;
  _writeoffset += retVal__offset;
  __global double*  tempThis;
  tempThis = &( opThisVect[opThisVect__offset]);
  double  tempResult;
  const double RTl_x = tempThis[_readoffset];
  {
    double RTl_i;
    double RTl_j;
    RTl_i = ((double) 0);
    ((int)RTl_i)++;; /* ERROR! */
    for ( RTl_j = ((double) 0); (RTl_j<((double) 1));(RTl_j += ((double) 1))) {
      (RTl_i += ((double) 1));;
    };
    tempResult = ((double) 0);
    retVal[_writeoffset] =  tempResult;
    if (_FAIL) {*_FAILRET = _FAIL;}
    return; ;
  }
}

1. By debugging the AST, I find that RTl_i and RTl_j are integers at first. Then, during the range analysis, the range information is invalidated here. I failed to understand this part though and don't know whether the result is right.
2. The i++ statement in the for loop is translated correctly. Why do they differ?

Thanks!

[sherhut]

The range information is invalidated as our loop analysis cannot handle unconstrained increments like the i++ in your example. We do not currently compute trip counts.

You found a bug in the code generation of ++ though. Although it was computed on doubles, the integer ++ was used in OpenCL. The wrong cast was just a symptom of that wrong behavior.

Thanks for reporting this. Stephan