lno-branch vs ptr-to-member

[Dale Johannesen <dalej at apple dot com>]

I'm having a problem demonstrated by the example below.  The key 
difficulty is

       h = h + rowData[r].*size;

which is gimplified into

    size.6<D1654>_15 = (int<D2>)size<D1640>_14;
    T.7<D1655>_16 = T.5<D1653>_13 + size.6<D1654>_15;

The code that looks for induction variables tries to analyze
(int<D2>)size<D1640>_14
and winds up trying to do the conversion at compile time through
this chain of calls:

interpret_rhs_modify_expr
chrec_convert
convert
ocp_convert
convert_to_integer

which crashes with
3683113.sm.cc:25: error: aggregate value used where an integer was 
expected

I'm not familiar enough with ptr-to-member to figure out what it ought 
to be doing,
but this isn't it.  Please help.  (It seems wrong to call those 
particular conversion
routines here at all, since they call 'error' for problems, which looks 
like user errors
not compiler internal errors.)

I'll file a PR on this shortly.

class QGArray
{
public:
  char *at (unsigned int index) const { return &shd[index]; }
  char *shd;
};
template<class type>
class QMemArray : public QGArray
{
public:
    type& operator[]( int i ) const
 { return (type &)(*(type *)QGArray::at(i*sizeof(type))); }
};
struct QLayoutStruct
{
  bool empty;
};
class QGridLayoutData
{
    int rr;
    int findSize (int QLayoutStruct::*) const;
    QMemArray<QLayoutStruct> rowData;
};
int QGridLayoutData::findSize( int QLayoutStruct::*size ) const
{
    int h = 0;
    int n = 0;
    for ( int r = 0; r < rr; r++ ) {
       h = h + rowData[r].*size;
       if ( !rowData[r].empty )
         n++;
    }
  return n * h;
}