Find JSRs
Submit this Search

Ad Banner

Summary  |  Proposal  |  Detail (Summary & Proposal)
JSRs: Java Specification Requests
JSR 84: Floating Point Extensions

Stage Access Start Finish
Withdrawn   01 Mar, 2002  
Expert Group Formation   10 Oct, 2000  
JSR Review Ballot View results 26 Sep, 2000 09 Oct, 2000
Status: Withdrawn
Reason: Due to the general absence of interest in the community, the Specification lead withdrew the JSR.
JCP version in use: 2.1
Java Specification Participation Agreement version in use: 1.0

This JSR proposes extensions to the JavaTM Programming Language and Java Virtual Machine that support more efficient execution of floating point code.

Please direct comments on this JSR to the Spec Lead(s)

Specification Leads
  Marc Snir IBM
Expert Group

This JSR has been Withdrawn
Reason: Due to the general absence of interest in the community, the Specification lead withdrew the JSR.

Original Java Specification Request (JSR)

Identification | Request | Contributions | Additional Information

Floating Point Extensions


This JSR proposes extensions to the Java Programming Language and Java Virtual Machine that support more efficient execution of floating point code.

Section 1. Identification

Submitting Member: International Business Machines Corporation

Name of Contact Person: Marc Snir

E-Mail Address:

Telephone Number: 1-914-945-3204

Fax Number: 1-914-945-4425

Specification Lead: Marc Snir, IBM Corporation

E-Mail Address:

Telephone Number: 1-914-945-3204

Fax Number: 1-914-945-4425

Initial Expert Group Membership:

  • International Business Machines Corporation
  • Sun Microsystems, Inc

Endorsers of this JSR include individuals from:

  • National Institute of Standards and Technology
  • The Mathworks, Inc.

Section 2: Request

2.1 Please describe the proposed Specification:

Version 1.1 of the Java platform specification gave strict rules for floating point semantics, using the IEEE 754 Standard for Binary Floating-Point Arithmetic. These rules enforced bit-by-bit reproducibility of floating point results across implementations. As a result, in some cases, the rules also significantly impaired floating point performance by effectively prohibiting certain code generation optimizations and the use of certain native operations on some processors. Version 1.2 of the Java platform specification permitted a relaxation of the rules for floating point semantics. These new rules allow a larger exponent (larger than that specified by earlier version of Java) to be used in certain situations. This improved the achievable performance of Java platform implementations for certain popular microprocessors at the cost of bit-by-bit reproducibility of floating point calculations on those processors. However, the relaxed rules still impair performance in many important cases. In particular
  1. The current Java platform specifications prevent the use of hardware features such as the Fused Multiply Add (FMA) operation on systems such as Intel IA64, PowerPC, PA-RISC 2.0, and MIPS IV. The FMA operation computes a*b+c, where a, b and c are values representable in the IEEE 754 formats double (or float); the result is within 0.5 ulp's of the exact answer. The FMA operation does not round the intermediate product before performing the addition; therefore, the result may be (slightly) different than that obtained by computing the product, rounding back to an IEEE 754 double (resp., float), followed by performing the addition. On processors with FMA instructions, an FMA typically offers twice the throughput of a multiplication followed by an addition.

  2. The current Java platform specifications prohibit common code generation optimizations. Such optimizations transform programs using field axioms that hold for real arithmetic but that hold only approximately for floating-point arithmetic.

While it is sometimes desirable to maintain bit-by-bit reproducibility of floating-point operations, such strictness is not always required. Floating-point arithmetic is an approximation to real arithmetic, and rounding errors are unavoidable. From a numeric viewpoint, in order to improve performance and/or accuracy, it is often acceptable to replace a computation with another. This relaxation of floating-point rules can be adopted as an option by the Java platform, provided that strict reproducibility can be enforced when needed and that suitable restrictions are set on implementations exploiting the relaxed rules.

The Java Grande Numerics Work Group has discussed several proposals to fix this problem, focusing on proposals that

  1. Improve performance of codes without requiring significant, platform dependent, code rewrite.

  2. Introduce only limited platform dependence and do so only for codes that allow the relaxation of floating-point rules.

  3. Have limited impact on accuracy.

  4. Observe certain consistency and reproducibility constraints.

  5. Require modest implementation effort, and negligible implementation effort on platforms that are not affected by the current Java floating point restrictions.

  6. Add at most one additional keyword.

This JSR is also guided by the above constraints.

We propose to add an FP-fast floating point mode. This mode will be associated with methods declared using the fastfp modifier, and the methods of classes and interfaces declared using the fastfp modifier. It is thus superficially analogous to the FP-strict floating point mode and strictfp modifier introduced in Java platform version 1.2.

The fastfp modifier can be used as a modifier in a method declaration, in which case it applies to the method itself. It can also be used as a modifier in a class or interface declaration. It is a compile time error for a given method, class, or interface declaration to contain both the strictfp and fastfp modifiers. The detailed specification will define if and when method modifiers can override class modifiers.

A compiler for the Java programming language will recognize the fastfp modifier and will set accordingly an ACC_FAST bit flag in the method_info structure for each method within a declaration bearing the fastfp modifier. (That is, declaring a class to be fastfp causes all of its methods to be FP-fast; declaring an interface to be fastfp causes its static initializer to be FP-fast.) The ACC_FAST bit indicates to the Java virtual machine that the associated method can be executed so as to take best advantage of the underlying floating point hardware and advanced code generation optimizations, while respecting the constraints of the FP-fast mode. The detailed specification will define those constraints on the behavior of FP-fast methods, so that different implementations will produce the same results up to acceptable rounding errors.

A method that does not have its ACC_FAST bit set will be interpreted as having the default floating point mode or, if its ACC_STRICT flag is set, having the floating point mode FP-strict. Thus, the behavior of methods not declared using the fastfp modifier (and the behavior of preexisting binary classes) is not changed by this proposal. A VerifyError will be thrown at the verification phase of class linking if a method has both ACC_FAST and ACC_STRICT set.

The FP-strict and the default floating point mode are each a valid implementation of the FP-fast mode; thus, any Java virtual machine implementation can trivially provide support for FP-fast methods.

The class java.lang.reflect.Modifier, which currently supports querying whether a method is FP-strict, will be amended to permit querying whether a method is FP-fast.

The precise meaning of fastfp -- i.e., the extent to which the behavior of FP-fast methods can deviate from the behavior of methods of the existing floating point modes -- will be elaborated by the expert group. The following two options were specifically proposed by the Java Grande Forum:

  1. An expression of the form a*b+c can be replaced by an FMA.

  2. Floating point operations may be reordered, assuming floating-point arithmetic to be associative.

Other proposals for possible optimizations in the fastfp mode will be examined by the expert group. The preference will be for nonprocedural specifications of the allowed optimizations, as procedural specifications tend to get obsolete faster and hamper future technological enhancements. (An example of this situation is the procedural definition of the behavior of trigonometric functions in Java.) The specification for the use of FMA is an exception to its rule, given its fundamental importance to linear algebra codes and its availability as a hardware primitive with an axiomatic definition. The associative rule is an axiomatic specification, and thus nonprocedural. Of particular interest to this group are optimizations that also improve the accuracy of floating-point computations. (We define a computation as being more accurate if the final result it produces is closer to the exact result.) These optimizations can be valuable to some classes of applications (in a sense, they make floating-point arithmetic behave more like real arithmetic) and can be used when, and only if, appropriate for the particular computation.

It is important to note that the fastfp specification will provide a list (at least conceptually) of which optimizations are valid in this mode. Implementations are free to implement a subset of this list. However, any implementation of an optimization must respect some fundamental principles. One of these principles is that optimizations should be temporally and spatially consistent. That is, during a particular execution of a Java program, the same instance of a construct must always produce the same result, and two different instances of the same construct must also produce the same result. (This reproducibility, of course, only holds for the same set of inputs on constructs that are deterministic under the strictfp mode of operation.) Another fundamental principle is that the list of possible (not necessarily performed) optimizations that can be applied to a piece of code must be defined by the syntax of the code only. That is, the application developer must be able to tell which optimizations a particular implementation can or cannot perform on his/her code.

Implementations will provide information on the transformations enabled in fastfp mode using (read-only) java.util.Properties entries for each distinct optimization, e.g. java.fastfp.fma. An implementation can potentially choose, on startup, which optimizations it will exploit, and set the corresponding properties. Implementations can provide means of controlling which optimizations are exploited for a particular run and which property entries are set, e.g. with command line flags.

2.2 What is the target Java platform? (i.e., desktop, server, personal, embedded, card, etc.)

Desktop, server, personal, embedded (any platform with floating point).

2.3 What need of the Java community will be addressed by the proposed specification?

The proposal will enable Java programs to achieve competitive floating point performance by taking advantage of the hardware of various widespread microprocessors and better code generator restructuring.

2.4 Why isn't this need met by existing specifications?

Current specifications preclude using FMAs and disable many common code generation optimizations.

2.5 Please give a short description of the underlying technology or technologies:

See 2.1.

2.6 Is there a proposed package name for the API Specification? (i.e., javapi.something, org.something, etc.)

Not relevant.

2.7 Does the proposed specification have any dependencies on specific operating systems, CPUs, or I/O devices that you know of?

The proposal can be implemented on any system that currently supports Java, with minimal changes in compilers and Java virtual machine implementations. (The minimal change required in Java virtual machine implementations is support for the fastfp modifier in the reflection API.) The effect of the ACC_FAST flag is system dependent. (Note: Changes to a Java virtual machine implementation are necessary in order to exploit the additional optimizations enabled by the fastfp mode. However, since the effect of ACC_FAST is system-dependent, Java virtual machine implementations are not forced to perform any of these optimizations.)

2.8 Are there any security issues that cannot be addressed by the current security model?


2.9 Are there any internationalization or localization issues?


2.10 Are there any existing specifications that might be rendered obsolete, deprecated, or in need of revision as a result of this work?

This JSR proposes extensions to the current Java programming language and virtual machine specifications. It also proposes a minor extension to the java.lang.reflect.Modifier API.

2.11 Please describe the anticipated schedule for the development of this specification.

  • 12/2000: First specification of fastfp extensions.
  • 06/2001: Reference implementation of a Java compiler and Java Virtual Machine that are fastfp aware.
  • 06/2001 - 08/2001: Open review of the fastfp extension.
  • 12/2001: Final specification of the fastfp extension ready.
  • 03/2002: Reference implementation and test suite for final specification ready.

Section 3: Contributions

3.1 Please list any existing documents, specifications, or implementations that describe the technology. Please include links to the documents if they are publicly available.

This proposal is based on discussions of the Java Grande Numeric Group that can be accessed at

3.2 Explanation of how these items might be used as a starting point for the work.

The documents produced by the Java Grande Numerics Group include detailed discussions of design alternatives -- pros and cons.