simulation of dapys2 is wrong

I agree that your fix is correct. 

The host intrinsics package, which supplies this code, is no longer maintained.  Nonetheless, I filed a bug report against it.  The system which contains these bug reports is not available publicly.  For the record, the ID number of the bug is 3069. 

The following fix was applied to the source file c66_ag_intrins.c ...

static dword dapys2(dword op1, dword op2)
{
    
    word res0, res1, res2, res3 ;
    dword result ;
    
    if ( op1.low & 0x00008000 ) {
      if ( (op2.low & 0xFFFF) == 0x8000 ) 
        res0 = 0x7FFF;
      else
        res0 = (-((word)op2.low)) & 0xFFFF ;
    } else {
      res0 = op2.low & 0xFFFF ;
    }
    
    if ( op1.low & 0x80000000 ) {
      if ( (op2.low & 0xFFFF0000) == 0x80000000 ) 
        res1 = 0x7FFF0000;
      else
        res1 = ((-((word)op2.low>>16)) & 0xFFFF) << 16 ;
    } else {
      res1 = op2.low & 0xFFFF0000 ;
    }
    
    if ( op1.high & 0x00008000 ) {
      if ( (op2.high & 0xFFFF) == 0x8000 ) 
        res2 = 0x7FFF;
      else
        res2 = (-((word)op2.high)) & 0xFFFF ;
    } else {
      res2 = op2.high & 0xFFFF ;
    }
    
    if ( op1.high & 0x80000000 ) {
      if ( (op2.high & 0xFFFF0000) == 0x80000000 ) 
        res3 = 0x7FFF0000;
      else
        res3 = ((-((word)op2.high>>16)) & 0xFFFF) << 16 ;
    } else {
      res3 = op2.high & 0xFFFF0000 ;
    }
    
    result.high = res3 | res2 ;
    result.low  = res1 | res0 ;
    
    return result ;
}

The change is the addition of the 4 casts to type (word).  I verified that all test vectors pass with this change.  Since this change is unlikely to appear in a release, I am publishing it here for all to use.

Thanks and regards,

-George