NUM06-J. Use the strictfp modifier for floating-point calculation consistency across platforms

The Java language allows platforms to use available floating-point hardware that can provide floating-point support with exponents that contain more bits than the standard Java primitive type double (in the absence of the strictfp modifier). Consequently, these platforms can represent a superset of the values that can be represented by the standard floating-point types. Floating-point computations on such platforms can produce different results than would be obtained if the floating-point computations were restricted to the standard representations of float and double. According to the Java Language Specification (JLS), §15.4, "FP-strict Expressions",

the net effect [of non-fp-strict evaluation], roughly speaking, is that a calculation might produce "the correct answer" in situations where exclusive use of the float value set or double value set might result in overflow or underflow.

Programs that require consistent results from floating-point operations across different JVMs and platforms must use the strictfp modifier. This modifier requires the JVM and the platform to behave as though all floating-point computations were performed using values limited to those that can be represented by a standard Java float or double, guaranteeing that the result of the computations will match exactly across all JVMs and platforms.

Using the strictfp modifier leaves execution unchanged on platforms that lack platform-specific floating-point behavior. It can have substantial impact, however, on both the efficiency and the result values of floating-point computations when executing on platforms that implement platform-specific floating-point behavior. On these platforms, using the strictfp modifier increases the likelihood that intermediate operations will overflow or underflow because it restricts the range that can be represented and the precision of intermediate values; it can also reduce computational efficiency. These issues are unavoidable when portability is the main concern.

The strictfp modifier can be used with a class, method, or interface:

An expression is strict when any of the containing classes, methods, or interfaces is declared to be strictfp. Constant expressions containing floating-point operations are also evaluated strictly. All compile-time constant expressions are by default, strictfp.

Strict behavior cannot be inherited by a subclass that extends a strictfp superclass. An overriding method can independently choose to be strictfp when the overridden method is not or vice versa.

Noncompliant Code Example

This noncompliant code example does not mandate strictfp computation. Double.MAX_VALUE is multiplied by 1.1 and reduced back by dividing by 1.1, according to the evaluation order. If Double.MAX_VALUE is the maximum value permissible by the platform, the calculation will yield the result infinity.

However, if the platform provides extended floating-point support, this program might print a numeric result roughly equivalent to Double.MAX_VALUE.

JVM implementations are not required to report an overflow resulting from the initial multiplication, although they can choose to treat this case as strictfp. The ability to use extended exponent ranges to represent intermediate values is implementation defined.

class Example {
  public static void main(String[] args) {
    double d = Double.MAX_VALUE;
    System.out.println("This value \"" + ((d * 1.1) / 1.1) + "\" cannot be represented as double.");
  }
}

Compliant Solution

For maximum portability, use the strictfp modifier within an expression (class, method, or interface) to guarantee that intermediate results do not vary because of implementation-defined compiler optimizations or by design. The calculation in this compliant solution is guaranteed to produce infinity because of the intermediate overflow condition, regardless of what floating-point support is provided by the platform.

strictfp class Example {
  public static void main(String[] args) {
    double d = Double.MAX_VALUE;
    System.out.println("This value \"" + ((d * 1.1d) / 1.1d) + "\" cannot be represented as double.");
  }
}

This compliant solution also specifies that the floating-point literals (1.1) are of type double to clarify their expected type and minimize their lack of precision.

Noncompliant Code Example

On platforms whose native floating-point hardware provides greater precision than double, the JIT is permitted to use floating-point registers to hold values of type float or type double (in the absence of the strictfp modifier), even though the registers support values with greater exponent range than that of the primitive types. Consequently, conversion from float to double can cause an effective loss of magnitude.

class Example {
  double d = 0.0;

  public void example() {
    float f = Float.MAX_VALUE;
    float g = Float.MAX_VALUE;
    this.d = f * g;
    System.out.println("d (" + this.d + ") might not be equal to " + (f * g));
  }
  public static void main(String[] args) {
    Example ex = new Example();
    ex.example();
  }
}

The lost magnitude would also have been lost if the value were stored to memory, for example to a field of type float.

Compliant Solution

This compliant solution uses the strictfp keyword to require exact conformance with standard Java floating-point. Consequently, the intermediate value of both computations of f * g will be identical to the value stored in this.d, even on platforms that support extended range exponents.

strictfp class Example {
  double d = 0.0;

  public void example() {
    float f = Float.MAX_VALUE;
    float g = Float.MAX_VALUE;
    this.d = f * g;
    System.out.println("d (" + this.d + ") might not be equal to " + (f * g));
  }

  public static void main(String[] args) {
    Example ex = new Example();
    ex.example();
  }
}

Exceptions

NUM06-EX0: This rule applies only to calculations that require consistent floating-point results on all platforms. Applications that lack this requirement need not comply.

NUM06-EX1: The strictfp modifier may be omitted when competent numerical analysis demonstrates that the computed values will meet all accuracy and behavioral requirements that are appropriate to the application. Note that "competent numerical analysis" generally requires a specialized professional numerical analyst; lesser levels of rigor fail to qualify for this exception.

Risk Assessment

Failure to use the strictfp modifier can result in implementation-defined behavior with respect to the behavior of floating-point operations.

Automated Detection

Sound automated detection of violations of this rule is not feasible in the general case.

Related Guidelines

Bibliography

NUM05-J. Do not use denormalized numbers      03. Numeric Types and Operations (NUM)