Double Trouble with COM Interop

I have been calling FoxPro code from C# via COM Interop (using DotNet2Fox) for several years, but I ran across an interesting issue last week. I have some FoxPro code that performs calculations and returns the results to .NET. I expected one of those values to be zero, but according to C#, it was -3.5527136788005e-15. Huh? After repeated tests, I got the same result.

I was assigning the FoxPro results to C# variables of type “decimal”. The C# compiler had already informed me that FoxPro was returning values with type “double” over COM. I have heard about the tension between binary and floating point math in computers before, but in practice, I have never run into an issue. I initially thought that the double-to-decimal conversion in C# was exposing this tension. As a test, I changed my C# variables to type double, but the problem persisted. I stepped through my code in FoxPro, and according to the debugger, the value was indeed zero. What the heck was going on?

I came up with a simpler calculation to reproduce the issue: 1.23 + 4.56. The result should be 5.79, right? Enter ? 1.23 + 4.56 in the FoxPro command window, and that’s what you’ll get. However, run the code below in C# and the result will be 5.790000000000001.

        dynamic vfp = Activator.CreateInstance(Type.GetTypeFromProgID("VisualFoxPro.Application", true));
        var result = vfp.Eval("1.23 + 4.56");
        return result;

Had I run across some glitch with COM interop? I needed to be certain that calculations done in FoxPro would come over to C# accurately. This was disconcerting.

Upon further thought… someone would have surely caught an issue like this with general COM interop, right? Maybe this was more on the FoxPro side. Then, I remembered a blog entry Christof wrote years ago on Significant digits in Visual FoxPro that mentions some oddities with 15-digit floating point calculations in VFP. Hmmmm… 15 digits. That’s the same precision of double types in C#. And look how many digits are after the decimal place in 5.790000000000001. So, I tried another test in the FoxPro command window:

? 1.25 + 4.56

There it is! Now, FoxPro shows 5.790000000000001! Clearly, this is either some kind of binary-to-floating-point issue or something to do with that 15th digit, but SET DECIMALS masks that most of the time. SET DECIMALS is a mysterious command. It specifies the number of decimal places that will be displayed from calculations. It apparently also affects comparisons: ? 1.25 + 4.56 = 5.79 returns .T. when SET DECIMALS TO 2. But SET DECIMALS TO 15 and the same expression returns .F. Weird! SET DECIMALS appears to have no effect on calculated values returned over COM interop and always returns the full 15-digit precision of a double.

So, that explains why I had the issue when using COM interop. How do we fix it? Simple… round the calculated result to the desired number of decimals in FoxPro before returning the result, or in C# after the result is returned from FoxPro. In fact, I was able to resolve by rounding down to 14 digits (eliminating the troublesome 15th digit), but I can’t guarantee that would always work. By the way…if you want C# rounding to have the same behavior as FoxPro rounding, specify MidpointRounding.AwayFromZero:

    public static double FixDouble(double number, int roundDecimals = 14)
        return Math.Round(number, roundDecimals, MidpointRounding.AwayFromZero);