Floating-point numbers can take on three exceptional values:
NaN (not-a-number). These values are produced as a result of exceptional or otherwise unresolvable floating-point operations, such as division by zero. These exceptional values can also be obtained directly from user input through methods such as
Double.valueOf(String s). Failure to detect and handle such exceptional values can result in inconsistent behavior.
Double.valueOf(String s) can return
NaN or an infinite
double, as specified by its contract. Programs must ensure that all floating-point inputs (especially those obtained from the user) are free of unexpected exceptional values. The methods
Double.isNaN(double d) and
Double.isInfinite(double d) can be used for this purpose.
NaN values are particularly problematic because they are unordered. That is, the expression
NaN == NaN always returns
false (see NUM07-J. Do not attempt comparisons with NaN for more information).
Noncompliant Code Example
This noncompliant code example accepts user data without validating it:
This code produces unexpected results when an exceptional value is entered for
val and subsequently used in calculations or as control values. The user could, for example, input the strings
NaN on the command line, which would be parsed by
Double.valueOf(String s) into the floating-point representations of either
NaN. All subsequent calculations using these values would be invalid, possibly causing runtime exceptions or enabling denial-of-service (DoS) attacks.
In this noncompliant example, entering
val would cause
currentBalance to be set to
NaN, corrupting its value. If this value were used in other expressions, every resulting value would also become
NaN, possibly corrupting important data.
This compliant solution validates the floating-point input before using it. The value is tested to ensure that it is neither
-infinity may be acceptable as expected inputs to a program. In such cases, explicit checks might not be necessary. However, such programs must be prepared to handle these exceptional values gracefully and should prevent propagation of the exceptional values to other code that fails to handle exceptional values. The choice to permit input of exceptional values during ordinary operation should be explicitly documented.
Incorrect or missing validation of floating-point input can result in miscalculations and unexpected results, possibly leading to inconsistent program behavior and denial of service.
Automated detection is infeasible in the general case. It could be possible to develop a taint-like analysis that detects many interesting cases.