Accord.Net: Looking for a Bug that Could Help Machines Conquer Humankind

Sergey Vasiliev
Articles: 24



Articles discussing the results of analysis of open-source projects are a good thing as they benefit everyone: some, including project authors themselves, can find out what bugs lurk in a project; others discover for themselves the static analysis technology and start using it to improve their code's quality. For us, it is a wonderful means to promote PVS-Studio analyzer, as well as to put it through some additional testing. This time I have analyzed Accord.Net framework and found lots of interesting issues in its code.

Picture 5

About the project and the analyzer

Accord.Net is a .NET machine learning framework written in C#. It is comprised of several libraries covering a wide range of tasks such as static data processing, machine learning, pattern recognition, and so forth. The source code can be downloaded from the GitHub repository.

Рисунок 2

The project was scanned with PVS-Studio static code analyzer, which can be downloaded here. I also encourage you to take a look at other articles on analysis of open-source projects, and the "bug database", where we collect bugs found by our tool.

A few words about warnings

The analyzer issued 91 first-level and 141 second-level warnings. In this article, I discuss or mention 109 warnings out of the total number of warnings issued. Looking through the other warnings, I found 23 more issues that looked like errors, but I don't mention them here because they are of little interest or look very similar to those already discussed. As for the remaining warnings, they are a bit harder to classify and would require more thorough investigation. So, out of 232 warnings, at least 132 report real errors. This figure tells us that the ratio of false positives for this project is about 46%. Oh, wait, sorry... It actually tells us that half the warnings deal with real bugs! It looks like a rather weighty argument for why we need to use static analysis tools. At the end of the article, I will talk about how to and how not to use static analysis, but for now let's see what interesting issues were be found in Accord.Net.

Errors found

Identical subexpressions

It's quite easy to let in errors detected by diagnostic V3001, especially when you use the copy-paste technique or when variables used in an expression have similar names. This type of errors is one of the most common and can be found in this project, too. Try to spot the error in the following fragment without reading the warning description.

public Blob[] GetObjects(UnmanagedImage image, 
                         bool extractInOriginalSize)
{
  ....
  if ((image.PixelFormat != PixelFormat.Format24bppRgb)    &&
      (image.PixelFormat != PixelFormat.Format8bppIndexed) &&
      (image.PixelFormat != PixelFormat.Format32bppRgb)    &&
      (image.PixelFormat != PixelFormat.Format32bppArgb)   &&
      (image.PixelFormat != PixelFormat.Format32bppRgb)    &&
      (image.PixelFormat != PixelFormat.Format32bppPArgb)
      )
  ....
}

PVS-Studio warning: V3001 There are identical sub-expressions 'image.PixelFormat != PixelFormat.Format32bppRgb' to the left and to the right of the '&&' operator. Accord.Imaging BlobCounterBase.cs 670

The image.PixelFormat != PixelFormat.Format32bppRgb subexpression is repeated twice. When you have names like that for enumeration elements, making a mistake becomes very easy - and that's exactly what happened in this example. An extra subexpression like that is very difficult to notice when just skimming the code. A trickier question is whether one of the comparisons is really redundant or it was meant to work with some other enumeration value instead. The first case simply deals with redundant code, while the second implies a logical error.

The same code snippet was found one more time in the same file, line 833. Copy-paste... Copy-paste never changes.

Wrong loop termination condition

We all are used to naming loop counters with names like i, j, k, etc. It's a convenient technique, and quite a common one, but sometimes it may backfire on you, as shown in the following example.

public static void Convert(float[][] from, short[][] to)
{
  for (int i = 0; i < from.Length; i++)
    for (int j = 0; i < from[0].Length; j++)
      to[i][j] = (short)(from[i][j] * (32767f));
}

PVS-Studio warning: V3015 It is likely that a wrong variable is being compared inside the 'for' operator. Consider reviewing 'i' Accord.Audio SampleConverter.cs 611

Variable i is used in the termination condition of the second loop, while variable j is used as its counter. As for i, it does not change within the nested loop. Therefore, the j variable will be incremented until it goes beyond the array bounds, causing an exception to be thrown.

Different logic blocks for identical conditions

The following code fragment contains two identical if statements with different logic blocks.

public void Fit(double[][] observations, 
                double[] weights, 
                MultivariateEmpiricalOptions options)
{
  if (weights != null)
    throw new ArgumentException("This distribution does not support  
                                 weighted  samples.", "weights");
  ....
  if (weights != null)
      weights = inPlace ? weights : (double[])weights.Clone();
  ....
}

PVS-Studio warning: V3021 There are two 'if' statements with identical conditional expressions. The first 'if' statement contains method return. This means that the second 'if' statement is senseless Accord.Statistics MultivariateEmpiricalDistribution.cs 653

Strange code, isn't it? Especially considering that the weights variable is a method parameter and is not used at all between the conditions. Therefore, the second if statement will never execute because if the weights != null expression is true, ArgumentException will be thrown.

The same code was found one more time in the same file, line 687.

Conditions that are always false

Diagnostic V3022 has grown much nicer since the analyzer's first release and keeps surprising me. Let's see if it can surprise you as well. First try to find the error in the code below without reading the diagnostic message. Mind that this is an abridged version of the code, with some lines left out.

private static void dscal(int n, double da, double[] dx, 
                          int _dx_offset, int incx)
{
  ....
  if (((n <= 0) || (incx <= 0)))
  {
    return;
  }
  ....
  int _i_inc = incx;
  for (i = 1; (_i_inc < 0) ? i >= nincx : i <= nincx; i += _i_inc)
  ....
}

PVS-Studio warning: V3022 Expression '(_i_inc < 0)' is always false. Accord.Math BoundedBroydenFletcherGoldfarbShanno.FORTRAN.cs 5222

Finding the error now that irrelevant lines have been removed is very easy of course. However, you still can't say right away where exactly the error is hiding. The point is (as you might have guessed after reading the warning) that the (_i_inc < 0) expression is always false. Note also that the _i_inc variable is initialized to the value of variable incx, which is a positive number at the moment of initializing _i_inc because the method executed a bit earlier would terminate if it were otherwise. Therefore, the _i_inc variable can have only a positive value, so the _i_inc < 0 comparison will always evaluate to false, and the loop termination condition will always be i <= nincx.

Such profound analysis has become possible thanks to the mechanism of virtual values, which has significantly improved some diagnostics of the analyzer.

private void hqr2()
{
  ....
  int low = 0;
  ....
  for (int i = 0; i < nn; i++)
  {
      if (i < low | i > high)
        ....
  }
  ....
}

PVS-Studio warning: V3063 A part of conditional expression is always false: i < low. Accord.Math JaggedEigenvalueDecompositionF.cs 571

The i < low subexpression will always be false, as the least value the i variable can take is 0, while low, too, will always be referring to 0 when this comparison is evaluated. That is, the i < low subexpression will be "running idle" all the time.

There were a lot of defects like that. Here's just a few:

  • V3063 A part of conditional expression is always false: i < low. Accord.Math JaggedEigenvalueDecompositionF.cs 972
  • V3063 A part of conditional expression is always false: i < low. Accord.Math JaggedEigenvalueDecomposition.cs 571
  • V3063 A part of conditional expression is always false: i < low. Accord.Math JaggedEigenvalueDecomposition.cs 972
  • V3063 A part of conditional expression is always false: i < low. Accord.Math EigenvalueDecomposition.cs 567

Integer division with casting to real type

The analyzer detected suspicious calculations. Programmers often forget that division of integer values is performed as integer division by default. If it was meant to be real division instead, you might get a nasty and elusive error. It's hard sometimes for a programmer who is not involved in a project to tell when such expressions are incorrect, but they must be checked anyway. Let's examine a few more similar cases.

public static double GetSpectralResolution(int samplingRate, 
                                           int samples)
{
  return samplingRate / samples;
}

PVS-Studio warning: V3041 The expression was implicitly cast from 'int' type to 'double' type. Consider utilizing an explicit type cast to avoid the loss of a fractional part. An example: double A = (double)(X) / Y;. Accord.Audio Tools.cs 158

The method above performs division of two integers, but the result of this operation is implicitly cast to type double, which looks strange.

Рисунок 0

The next example is even stranger:

public static int GreatestCommonDivisor(int a, int b)
{
  int x = a - b * (int)Math.Floor((double)(a / b));
  while (x != 0)
  {
    a = b;
    b = x;
    x = a - b * (int)Math.Floor((double)(a / b));
  }
  return b;    
}

PVS-Studio warnings:

  • V3041 The expression was implicitly cast from 'int' type to 'double' type. Consider utilizing an explicit type cast to avoid the loss of a fractional part. An example: double A = (double)(X) / Y;. Accord.Math Tools.cs 137
  • V3041 The expression was implicitly cast from 'int' type to 'double' type. Consider utilizing an explicit type cast to avoid the loss of a fractional part. An example: double A = (double)(X) / Y;. Accord.Math Tools.cs 142

The analyzer didn't like the (double)(a / b) expression. The Floor method returns the largest integer less than or equal to the specified double-precision floating-point number (MSDN. Math.Floor). The variables a and b, however, are of type int, so integer division will be performed, producing an integer. It turns out that explicitly casting this value to type double and calling the Floor method make no sense.

To execute that operation correctly, the programmer should have cast one of the operands to type double. In that case, it would execute as real division and calling the Floor method would make sense:

Math.Floor((double)a / b)

The value of a method parameter is constantly overwritten

Let's go on. Errors of this type are rather rare, but they still show up every now and then.

private static double WeightedMode(double[] observations, 
                                   double[] weights, 
                                   double mode, 
                                   int imax, 
                                   int imin)
{
  ....
  var bestValue = currentValue;
  ....
  mode = bestValue;
  return mode;
}

PVS-Studio warning: V3061 Parameter 'mode' is always rewritten in method body before being used. Accord.Statistics TriangularDistribution.cs 646

One of the method's parameters, mode, is overwritten and returned, although it's not used at all within the method (except when it is overwritten). I can't say for sure if this is an error (some of the similar issues found in other projects obviously were bugs), but this code does look strange.

There is, by the way, one interesting thing about this project: almost each triggered diagnostic is triggered more than once. The same code as in the example above was found in several other parts of the project. Indeed, copy-paste never changes...

  • V3061 Parameter 'mode' is always rewritten in method body before being used. Accord.Statistics TriangularDistribution.cs 678
  • V3061 Parameter 'mode' is always rewritten in method body before being used. Accord.Statistics TriangularDistribution.cs 706
  • V3061 Parameter 'mode' is always rewritten in method body before being used. Accord.Statistics TriangularDistribution.cs 735

Null dereference

public override string ToString(string format, 
                                IFormatProvider formatProvider)
{
  ....
  var fmt = components[i] as IFormattable;
  if (fmt != null)
    sb.AppendFormat(fmt.ToString(format, formatProvider));
  else
    sb.AppendFormat(fmt.ToString());
  ....
}

PVS-Studio warning: V3080 Possible null dereference. Consider inspecting 'fmt'. Accord.Statistics MultivariateMixture'1.cs 697

Bugs that result in throwing exceptions are usually caught during the development process if the code is tested thoroughly enough. But sometimes they slip by, as proved by the example above. If the fmt != null condition is false, the instance method ToString of the fmt object is called. What's the result? Raising NullReferenceException.

As you have probably guessed already, this diagnostic was triggered one more time: MultivariateMixture'1.cs 697

Mutual assignment of references

public class MetropolisHasting<T> : IRandomNumberGenerator<T[]>
{
  ....        
  T[] current;
  T[] next;
  ....
  public bool TryGenerate()
  {
    ....
    var aux = current;
    current = next;
    next = current;
   ....
  }
  ....
}

PVS-Studio warning: V3037 An odd sequence of assignments of this kind: A = B; B = A;. Check lines: 290, 289. Accord.Statistics MetropolisHasting.cs 290

In the fragment of method TryGenerate above, the programmer obviously wanted to swap the references to arrays next and current (the aux variable is not used anywhere else) but made a mistake assigning a reference to the same array to both variables current and next - the array that was previously referred to by the reference stored in next.

This is what the fixed code should look like:

var aux = current;
current = next;
next = aux;

Potential division by zero

There were a few potential division-by-zero errors. Let's check them briefly:

public BlackmanWindow(double alpha, int length) 
    : base(length)
{
    double a0 = (1.0 - alpha) / 2.0;
    double a1 = 0.5;
    double a2 = alpha / 2.0;

    for (int i = 0; i < length; i++)
        this[i] = (float)(a0 - 
          a1 * Math.Cos((2.0 * System.Math.PI * i) / (length - 1)) +
          a2 * Math.Cos((4.0 * System.Math.PI * i) / (length - 1)));
}

PVS-Studio warnings:

  • V3064 Potential division by zero. Consider inspecting denominator '(length - 1)'. Accord.Audio BlackmanWindow.cs 64
  • V3064 Potential division by zero. Consider inspecting denominator '(length - 1)'. Accord.Audio BlackmanWindow.cs 65

The warning was triggered by the following code:

(2.0 * System.Math.PI * i) / (length - 1)

In reality, this potential error might never show up (length is the length of some window and must be 1 for the error to occur), but who knows? We'd better play safe; otherwise we risk getting a nasty error, which could also be hard to track down.

There was another interesting code fragment with potential division by zero.

public static double[,] Centering(int size)
{
  if (size < 0)
  {
      throw new ArgumentOutOfRangeException("size", size,
          "The size of the centering matrix must 
           be a positive integer.");
  }

  double[,] C = Matrix.Square(size, -1.0 / size);

  ....
}

PVS-Studio warning: V3064 Potential division by zero. Consider inspecting denominator 'size'. Accord.Math Matrix.Construction.cs 794

Personally I find this error very interesting and funny. The analyzer warned us about potential division by zero in the -1.0 / size expression. Now take a look at the check a bit earlier. If size < 0, an exception will be thrown, but if size == 0, there will be no exception, but we will get division by 0. At the same time, it is mentioned in the literal passed to the exception constructor that the matrix size must be a positive integer, while the check is done against non-negative values; and positive and non-negative are different things, after all. It seems like we could fix the error by simply adjusting the check:

if (size <= 0)

Using a bitwise operator instead of a logical one

At times you have to deal with the problem of some programmers not knowing the difference between bitwise and logical operators ('|' and '||', '&' and '&&'). Possible implications range from extra computations to crashes. In this project, the analyzer found a few strange fragments with bitwise operations:

public JaggedSingularValueDecompositionF(
         Single[][] value,
         bool computeLeftSingularVectors, 
         bool computeRightSingularVectors, 
         bool autoTranspose, 
         bool inPlace)
{
  ....
  if ((k < nct) & (s[k] != 0.0))
  ....
}

PVS-Studio warning: V3093 The '&' operator evaluates both operands. Perhaps a short-circuit '&&' operator should be used instead. Accord.Math JaggedSingularValueDecompositionF.cs 461

The body of the if statement will execute if both subexpressions (k < nct and s[k] != 0.0) are true. However, even if the first subexpression (k < nct) is false, the second will be evaluated anyway, which wouldn't happen if the programmer used the && operator. So, if they wanted the value of k to be checked to avoid going beyond the array bounds, they failed.

Other similar issues:

  • V3093 The '&' operator evaluates both operands. Perhaps a short-circuit '&&' operator should be used instead. Accord.Math JaggedSingularValueDecompositionF.cs 510
  • V3093 The '&' operator evaluates both operands. Perhaps a short-circuit '&&' operator should be used instead. Accord.Math JaggedSingularValueDecompositionF.cs 595
  • V3093 The '&' operator evaluates both operands. Perhaps a short-circuit '&&' operator should be used instead. Accord.Math JaggedSingularValueDecomposition.cs 461
  • V3093 The '&' operator evaluates both operands. Perhaps a short-circuit '&&' operator should be used instead. Accord.Math JaggedSingularValueDecomposition.cs 510
  • V3093 The '&' operator evaluates both operands. Perhaps a short-circuit '&&' operator should be used instead. Accord.Math JaggedSingularValueDecomposition.cs 595
  • V3093 The '&' operator evaluates both operands. Perhaps a short-circuit '&&' operator should be used instead. Accord.Math Gamma.cs 296

Checking one and the same element in a loop

One bug was found through a diagnostic added in the recent analyzer version.

public override int[] Compute(double[][] data, double[] weights)
{
  ....
  int cols = data[0].Length;
  for (int i = 0; i < data.Length; i++)
    if (data[0].Length != cols)
      throw new DimensionMismatchException("data", 
                  "The points matrix should be rectangular. 
                   The vector at position {} has a different
                   length than previous ones.");
  ....
}

PVS-Studio warning: V3102 Suspicious access to element of 'data' object by a constant index inside a loop. Accord.MachineLearning BinarySplit.cs 121

It's quite an interesting bug. The programmer wanted to make sure that the jagged array data is two-dimensional (i.e. is a matrix) but made a mistake in the data[0].Length != cols expression and indexed into it with an integer literal, 0, instead of the loop counter i. As a result, the data[0].Length != cols expression is always false, as it is equivalent to expression data[0].Length != data[0].Length. Had the data parameter been a two-dimensional array (Double[,]), this error, as well as the entire check, could have been avoided. However, the use of the jagged array might be determined by some specifics of the application's architecture.

Passing a calling object as an argument to a method

The following code fragment looks strange, too.

public static double WeightedMean(this double[] values, 
                                       double[] weights)
{
  ....
}

public override void Fit(double[] observations, 
                         double[] weights, 
                         IFittingOptions options)
{
  ....
  mean = observations.WeightedMean(observations);
  ....
}

PVS-Studio warning: V3062 An object 'observations' is used as an argument to its own method. Consider checking the first actual argument of the 'WeightedMean' method. Accord.Statistics InverseGaussianDistribution.cs 325

The analyzer didn't like that the WeightedMean method receives as an argument the same object it is called from. It's even stranger considering that WeightedMean is an extension method. I did some additional investigation to see how this method was used in other parts of the application. Everywhere it is used, the second argument is represented by array weights (note that this array is also present in the Fit method, which we are discussing), so it does look like an error, and then the fixed code should look like this:

mean = observations.WeightedMean(weights);

Potential serialization error

The analyzer detected a potential issue related to serialization of one of the classes.

public class DenavitHartenbergNodeCollection :  
  Collection<DenavitHartenbergNode>
{ .... }

[Serializable]
public class DenavitHartenbergNode
{
  ....
  public DenavitHartenbergNodeCollection Children 
  { 
    get; 
    private set; 
  }
  ....
}

PVS-Studio warning: V3097 Possible exception: the 'DenavitHartenbergNode' type marked by [Serializable] contains non-serializable members not marked by [NonSerialized]. Accord.Math DenavitHartenbergNode.cs 77

When serializing an instance of class DenavitHartenbergNode, SerializationException exception may be thrown - it depends on what serializer type is selected. If it is, for example, an instance of type BinaryFormatter, the exception will be thrown because all serializable members (and this property is such a member) are required to be annotated with the attribute [Serializable].

Here are some ways to fix this error:

  • implement this property through a field annotated with the [NonSerialized] attribute. In that case, the field (and therefore the associated property) won't be serialized;
  • implement the ISerializable interface and set the GetObjecData method to ignore serialization of this property;
  • annotate the DenavitHartenbergNodeCollection type with the attribute [Serializable].

Judging by the surrounding code (all the other properties are serializable), it is the third scenario that must be implemented.

However, if instances of this type are serialized by serializers that do not require all serializable members to be annotated with the [Serializable] attribute, there is nothing to worry about.

The analyzer found lots of unsafe event calls. How many? 75 V3083 warnings! Let's examine only one such example because they all look almost the same.

private void timeUp_Elapsed(object sender, ElapsedEventArgs e)
{
  ....
  if (TempoDetected != null)
    TempoDetected(this, EventArgs.Empty);
}

PVS-Studio warning: V3083 Unsafe invocation of event 'TempoDetected', NullReferenceException is possible. Consider assigning event to a local variable before invoking it. Accord.Audio Metronome.cs 223

This code checks if there are any subscribers to the TempoDetected event and calls it if the check proves true. The programmer assumed that the check would help avoid throwing an exception if no subscribers to TempoDetected were found. However, there is some chance that at the moment between testing TempoDetected for null and the call to the event, no subscribers will be left (for example, they could unsubscribe from it in other threads). In that case, NullReferenceException will be thrown. To avoid problems like that, you may use, for example, the null-conditional operator '?.', which was added in C# 6.0. To learn more about this issue and other ways to resolve it, see the documentation on the diagnostic rule.

How to and how not to use static analyzers

Before finishing the article, I'd like to say a few words about how one should use static analysis tools. The following approach is very common: "We've tested our project before a release and haven't found anything of much interest". No, no, no! It's the worst way of using static analysis. To make it clearer, here's an analogy: stop using the IDE when developing applications and write all your code in Notepad instead; then, before the very release, switch back to the IDE. Sounds crazy, doesn't it? Of course it does! The IDE wouldn't be of much use if you left it sitting idle on your SSD/HDD for most of the development time when it could have really helped. It's just the same thing with static analyzers - they must be applied regularly, not occasionally.

Picture 6

When you run an analyzer on your code just before the release, it's obvious that most of the bugs have been fixed already. But at what cost? At the cost of the developers' nerves and time, and numerous tests designed to catch those very bugs. Taking all that into account, the cost of fixing these errors is, to put it mildly, quite a big one.

All these troubles, however, could be avoided if you integrated a static analyzer into the development process in the right way. Having it installed on every developer's machine, you can set it all up in such a way that most of the bugs that can be detected by the analyzer are found and fixed before they get a chance to make it to the repository. Moreover, finding and fixing an error that hasn't yet become overgrown with various dependencies is much cheaper. The incremental analysis mode, which allows you to catch errors right after they have appeared, makes analysis even more efficient.

Another good technique is to integrate static analysis into night builds. It can help catch errors quicker and also find out who let them slip in the repository, which is also a good way to motivate developers to be more careful when writing code.

To sum up, it is the regular use of static analysis tools that allows developers to benefit from them in the best way possible.

Conclusion

It was one more opportunity for me to scan an interesting project and find as interesting errors to share with you so that you could take note of something we have discussed, or learn something new, or just try to be more careful when writing code. Nevertheless, we are all humans, and to err is human. PVS-Studio can help fix errors in your code, and remember that regular use of static analyzers helps reduce the number of troubles you are faced with when hunting for bugs and fixing them.



Use PVS-Studio to search for bugs in C, C++, and C# code

We offer you to check your project code with PVS-Studio. Just one bug found in the project will show you the benefits of the static code analysis methodology better than a dozen of the articles.

goto PVS-Studio;

Sergey Vasiliev
Articles: 24


Do you make errors in the code?

Check your code
with PVS-Studio

Static code analysis
for C, C++, and C#

goto PVS-Studio;
On our website we use a cookie to collect information of a technical nature.
If you do not agree, please leave the site. Learn More →
Do not show again