When debugging some code we discovered that some float math appears to be
incorrect at times.  Because of the size of the values we were using we should
have really been using doubles.  The values that get assigned differ if the
assignement is inside a for loop or external to the loop.  This 'problem' has
been demonstrated in gcc 3.4.3 and 4.0 on the x86 platform.  It does not show up
when cross compiled for the ARM core.  The x86 gcc packages used were fc3
distribution rpm packages.

Note: Compiling this sample code under Visual studio under windows demonstrates
similar issues when compiled as debug code.  When compiled as release code it
does not have issues.


------------------------

The 'problem' has bee reduced to the following example code:



#include "stdio.h"

const int GOOD_A = -85911104;
const int GOOD_B =  42537964;

const int BAD_A  = -85908672;
const int BAD_B  =  42539240;

class QPoint
{
public:
    QPoint()
    {
        m_x = 0;
        m_y = 0;
    }

    QPoint( int x, int y )
    {
        m_x = x;
        m_y = y;
    }

    int x() const { return ( m_x ); }
    int y() const { return ( m_y ); }
protected:
    int m_x;
    int m_y;
};

class QPointArray
{
public:
    QPointArray( int size )
    {
        m_points = new QPoint[size];
    }

    const QPoint& Point( int idx )
    {
        return ( m_points[idx] );
    }

    void SetPoint( int idx, int x, int y )
    {
        m_points[idx] = QPoint( x, y );
    }

protected:
    QPoint *m_points;
};

int m_points = 28;

int main(int argc, char *argv[])
{
    QPointArray m_pt_array( m_points );
    m_pt_array.SetPoint( 0,  -85911106, 42537898 );
    m_pt_array.SetPoint( 1,  -85910960, 42537964 );
    m_pt_array.SetPoint( 2,  -85910328, 42538390 );
    m_pt_array.SetPoint( 3,  -85908832, 42538685 );
    m_pt_array.SetPoint( 4,  -85908717, 42538755 );
    m_pt_array.SetPoint( 5,  -85908717, 42539009 );
    m_pt_array.SetPoint( 6,  -85908676, 42539131 );
    m_pt_array.SetPoint( 7,  -85908585, 42539238 );
    m_pt_array.SetPoint( 8,  -85908469, 42539331 );
    m_pt_array.SetPoint( 9,  -85908386, 42539440 );
    m_pt_array.SetPoint( 10, -85908361, 42539565 );
    m_pt_array.SetPoint( 11, -85908531, 42540442 );
    m_pt_array.SetPoint( 12, -85908610, 42540555 );
    m_pt_array.SetPoint( 13, -85908713, 42540656 );
    m_pt_array.SetPoint( 14, -85908849, 42540733 );
    m_pt_array.SetPoint( 15, -85909315, 42540895 );
    m_pt_array.SetPoint( 16, -85909479, 42540928 );
    m_pt_array.SetPoint( 17, -85911521, 42541069 );
    m_pt_array.SetPoint( 18, -85911792, 42541054 );
    m_pt_array.SetPoint( 19, -85911877, 42540895 );
    m_pt_array.SetPoint( 20, -85911996, 42540731 );
    m_pt_array.SetPoint( 21, -85912216, 42540613 );
    m_pt_array.SetPoint( 22, -85912609, 42540335 );
    m_pt_array.SetPoint( 23, -85912778, 42540178 );
    m_pt_array.SetPoint( 24, -85912884, 42539993 );
    m_pt_array.SetPoint( 25, -85912924, 42539795 );
    m_pt_array.SetPoint( 26, -85912974, 42538217 );
    m_pt_array.SetPoint( 27, -85911106, 42537898 );

    float area = 0.0;

    for ( int i = 0; i < m_points; i++ )
    {
        QPoint cur  = m_pt_array.Point( i );
        QPoint next = m_pt_array.Point( ( i + 1 ) % m_points );
        float f  = (float)cur.x() * (float)next.y();
        float af = cur.x();
        float bf = next.y();
        float ff = af * bf;
        area -= f;
    
        printf( "%d * %d = %f or %f\n", cur.x(), next.y(), f, ff );
        area += (float)cur.y() * next.x();
    }

    return 0;
}

-------------------------------------------

The output produced is as follows:

-85911106 * 42537964 = -3654483519209472.000000 or -3654483519209472.000000
-85910960 * 42538390 = -3654513852416000.000000 or -3654514120851456.000000
-85910328 * 42538685 = -3654512510238720.000000 or -3654512241803264.000000
-85908832 * 42538755 = -3654454796615680.000000 or -3654454796615680.000000
-85908717 * 42539009 = -3654471708049408.000000 or -3654471708049408.000000
-85908717 * 42539131 = -3654482177032192.000000 or -3654482445467648.000000
-85908676 * 42539238 = -3654489424789504.000000 or -3654489693224960.000000
-85908585 * 42539331 = -3654493719756800.000000 or -3654493719756800.000000
-85908469 * 42539440 = -3654498283159552.000000 or -3654498283159552.000000
-85908386 * 42539565 = -3654505262481408.000000 or -3654505262481408.000000
-85908361 * 42540442 = -3654579619102720.000000 or -3654579350667264.000000
-85908531 * 42540555 = -3654596530536448.000000 or -3654596530536448.000000
-85908610 * 42540656 = -3654608610131968.000000 or -3654608610131968.000000
-85908713 * 42540733 = -3654619615985664.000000 or -3654619615985664.000000
-85908849 * 42540895 = -3654639211773952.000000 or -3654639480209408.000000
-85909315 * 42540928 = -3654661760352256.000000 or -3654661760352256.000000
-85909479 * 42541069 = -3654681087705088.000000 or -3654681087705088.000000
-85911521 * 42541054 = -3654766718615552.000000 or -3654766718615552.000000
-85911792 * 42540895 = -3654764571131904.000000 or -3654764571131904.000000
-85911877 * 42540731 = -3654754102149120.000000 or -3654754370584576.000000
-85911996 * 42540613 = -3654749270310912.000000 or -3654749001875456.000000
-85912216 * 42540335 = -3654734506360832.000000 or -3654734506360832.000000
-85912609 * 42540178 = -3654737727586304.000000 or -3654737459150848.000000
-85912778 * 42539993 = -3654728869216256.000000 or -3654728869216256.000000
-85912884 * 42539795 = -3654716252749824.000000 or -3654716521185280.000000
-85912924 * 42538217 = -3654582840328192.000000 or -3654582571892736.000000
-85912974 * 42537898 = -3654557338959872.000000 or -3654557338959872.000000
-85911106 * 42537898 = -3654477882064896.000000 or -3654477613629440.000000

------------------------------

The sample code was built with the following make file:

floatcast: floatcast.cpp
	g++ -o floatcast floatcast.cpp

-------------------------------