This patch attempts to address the problem described in the following OpenJDK issues:
- JDK-8039915: Wrong NumberFormat.format() HALF_UP rounding when last digit exactly at rounding position greater than 5
- JDK-8041961 (duplicate): DecimalFormat RoundingMode.HALF_UP is broken (HALF_EVEN is OK)
The problem was introduced in Java 8 at the same time as the fix for:
- JDK-7131459: DecimalFormat produces wrong format() results when close to a tie
The fix for JDK-7131459 has caused some confusion since it purposefully changes the
rounding behavior as compared to Java 7 and prior versions. A number of bugs were opened
and subsequently closed (duplicate, wontfix) because of "rounding differences" that were
assumed to be (and in many cases were) the result of the intentional change. However the
first two bugs listed above provide testcases that show obviously incorrect rounding
behavior in the HALF_UP
mode.
99.9989 -> 100.00
99.9990 -> 99.99
To put it simply, this demonstrates a case where a higher number rounds down,
and a lower number rounds up; (x <= y) != (round(x) <= round(y))
. It appears to
only affect the HALF_UP
rounding mode, which is the kind of rounding taught in
grade school arithmetic classes: 0.5
rounds away from zero, always.
This issue is known to exist in the Oracle/OpenJDK 1.8.0 GA release, as well as in released updates: u5, u11, u20, and u25.
Oracle has committed an official fix for Java 9 (bug 8039915, changeset), and backported to Java 8u40 (bug 8061380, changeset). The 8u40-b12 early access image from 28 October 2014 is the first available build that is not affected by the problem described above.
If you are running Java 8 update 40 or later, you do not require this patch. However, if you wish to support earlier versions of Java 8, read on.
Prior to update 40, software running on Java 8 that depends on the HALF_UP
rounding mode
for formatting decimal values into a String
will produce incorrect results. The root cause
appears to have been correctly identified on StackOverflow.com by the user
Holger with a link to the OpenJDK source file for java.text.DigitList
. Specifically, the
shouldRoundUp
method was modified as a part of JDK-7131459 and seemingly incorrectly so for
the HALF_UP
switch case block.
java.text.DigitList
is a system class and loaded by the boot classloader. In order to patch
it, one can decompile rt.jar
, apply the fix, recompile, and recreate rt.jar
. But that solution
has a number of drawbacks:
- Potential issues with the Oracle JDK/JVM license and/or terms of use
- Requires a moderate degree of technical skill outside the normal workflow for many sysadmins
- It's non-selective; it unconditionally applies to all programs that use that JVM (share that copy of
rt.jar
) - Changes will be lost when upgrading to a new version of the JVM that may also be buggy
- Not easily shipped with a product or application that doesn't already bundle its own JRE
Instead, this patch applies changes to java.text.DigitList
in memory and only for those
programs or environments that are explicitly configured to use it. That configuration will stay
local to that application/environment and so has the potential to survive JVM upgrades.
This is done by implementing a Java Agent, a feature introduced in JDK 5, that can
optionally redefine a class's bytecode when the target class is first loaded by any classloader.
The agent is injected before the application's (or container's) main
method and so should be
completely transparent to the application. Furthermore, the application requires no modification
other than to enable its startup script(s) to inject the agent as a JVM command line option.
Even though the bug only appears in Java 8, this patch should be compiled with Java 6 in order
to maintain maximum compatibility for those applications that may need to support running on a range
of older Java releases and not be limited only to the latest-and-greatest. This is possible as
long as nothing in the target class requires new bytecode formats/instructions, such as
INVOKEDYNAMIC
; thankfully, that condition appears to hold true at this time.
The patch, as implemented here, looks for specific bytecode signatures that changed at the time the bug was introduced and when it was fixed. It attempts to patch the bytecode only when it finds the particular bytecode signatures that suggest the bug is present. Otherwise, it stays out of the way. It is likely that this patch implementation is safe for use in all official Oracle releases of Java for versions 6 through 8.
You SHOULD NOT ship this patch with your application without testing it, first. There is a self-test included in the patch (see: Diagnostics below) to help determine if your JVM is affected and whether the patch works on that particular JVM, but it is not a replacement for due diligence and thorough testing of your own application.
This patch appears to work, but the authors have done neither exhaustive testing across
the vast range of decimal values nor have they tested every possible rounding path inside
the JDK libraries. More eyes on the patch code, and more real-world tests would be much
appreciated. Please consider opening an issue in Github with the keyword WFM
in the
title if this patch worked for your product, library, or application. Of course, if the patch
failed to work in a situation, definitely open an issue here in Github with as much detail
as you can provide.
Pull requests for new testcases, and of course bug fixes, are highly desired and welcome.
Pre-requisites:
- This patch library's JAR
- Its dependencies (currently ASM version >= 3.1 and < 4.0) must be on the classpath.
If your application does not currently use the ASM library, there is a version of this patch that
includes the ASM 3.1 classes already in the JAR. Or, use your favorite build system and package
management tool to get asm:asm:3.1
into your project's classpath. (If you need a newer version
of ASM or a different tool, contributions to the project are welcome.)
Precompiled JARs are available publicly:
| [jCenter](https://bintray.com/bintray/jcenter)
---------|------------------------------------------------ Group | com.pros.opensource.java Artifact | jdk8patch-halfupround-asm31 Version | 1.0
Always use the latest available version. Ironically, the logic in versions 0.9.1 and earlier did not account for the intentional and correct changes from JDK-7131459 in some edge cases.
Use the classifier all
if you want the JAR that bundles ASM 3.1 inside (in its original package
structure), then you can exclude the transitive dependency on the ASM library. (If you take
this approach, we assume you know how to manage your dependencies and avoid classpath collisions.)
Applying the patch is as simple as adding one additional entry to the beginning of the command line that starts your application's JVM:
java -javaagent:path/to/patch.jar ...
The patch library includes a self test you can use to determine if your particular JVM is affected by the bug, or if the patch will apply itself correctly. Execute the self test by hand using one of the following methods:
# Standalone patch JAR with no bundled dependencies
$ java -javaagent:path/to/patch.jar -cp <classpath_with_patch_and_ASM_jars> com.pros.java.text.DigitListPatch
# Patch JAR that includes ASM
$ java -javaagent:path/to/patch-with-asm.jar -jar path/to/patch-with-asm.jar
You may notice that the test case for JDK-7131459 does not match on this earlier version of Java. The test suite ignores the results for that particular test on versions prior to JDK 8.
JDK-8041961 test case:
99.9989 HALF_UP --> 100 OK
99.999 HALF_UP --> 100 OK
JDK-8039915 test case:
0.95000055 HALF_UP --> 0.950001 OK
0.9500006 HALF_UP --> 0.950001 OK
JDK-7131459 test case:
NOTE: These tests SHOULD NOT MATCH on earlier versions of Java.
0.15 is actually: 0.1499999999999999944488848768742172978818416595458984375
HALF_UP --> 0.2 expected: 0.1
0.35 is actually: 0.34999999999999997779553950749686919152736663818359375
HALF_UP --> 0.4 expected: 0.3
0.85 is actually: 0.84999999999999997779553950749686919152736663818359375
HALF_UP --> 0.9 expected: 0.8
0.95 is actually: 0.9499999999999999555910790149937383830547332763671875
HALF_UP --> 1 expected: 0.9
Above tests used Java 1.7.0_51 (Oracle Corporation)
installed at C:\JAVA\v7\jre
Agent installed: yes
Patch applied : NO
Overall result : OK (no patch necessary)
These are the results you should expect when this is working correctly on Java 8. These results should match Java 8u40 and later, without this patch, too.
JDK-8041961 test case:
99.9989 HALF_UP --> 100 OK
99.999 HALF_UP --> 100 OK
JDK-8039915 test case:
0.95000055 HALF_UP --> 0.950001 OK
0.9500006 HALF_UP --> 0.950001 OK
JDK-7131459 test case:
0.15 is actually: 0.1499999999999999944488848768742172978818416595458984375
HALF_UP --> 0.1 OK
0.35 is actually: 0.34999999999999997779553950749686919152736663818359375
HALF_UP --> 0.3 OK
0.85 is actually: 0.84999999999999997779553950749686919152736663818359375
HALF_UP --> 0.8 OK
0.95 is actually: 0.9499999999999999555910790149937383830547332763671875
HALF_UP --> 0.9 OK
Above tests used Java 1.8.0_20 (Oracle Corporation)
installed at C:\JAVA\v8\jre
Agent installed: yes
Patch applied : yes
Overall result : FIXED
Without the patch (without specifying the -javaagent
option) the self-test
on releases of Java 8 prior to update 40 will look similar to the following.
JDK-8041961 test case:
99.9989 HALF_UP --> 100 OK
99.999 HALF_UP --> 99.99 expected: 100
JDK-8039915 test case:
0.95000055 HALF_UP --> 0.950001 OK
0.9500006 HALF_UP --> 0.95 expected: 0.950001
JDK-7131459 test case:
0.15 is actually: 0.1499999999999999944488848768742172978818416595458984375
HALF_UP --> 0.1 OK
0.35 is actually: 0.34999999999999997779553950749686919152736663818359375
HALF_UP --> 0.3 OK
0.85 is actually: 0.84999999999999997779553950749686919152736663818359375
HALF_UP --> 0.8 OK
0.95 is actually: 0.9499999999999999555910790149937383830547332763671875
HALF_UP --> 0.9 OK
Above tests used Java 1.8.0_20 (Oracle Corporation)
installed at C:\JAVA\v8\jre
Agent installed: NO (missing -javaagent?)
Patch applied : NO
Overall result : NOT FIXED (expected on Java 1.8 < u40)
Copyright (c) 2014 by PROS, Inc. All Rights Reserved.
The original OpenJDK java.text.DigitList
source code is released under the
GNU Public License Version 2 with Classpath Exception
as designated by Oracle. Development of this patch required knowledge of that
source code, which makes the patch a derivative work and subject to the same
license terms. The authors of this patch wish to extend and apply the same
Classpath Exception to users of binary executable versions of this patch.
This code is free software; you can redistribute it and/or modify it
under the terms of the GNU General Public License version 2 only, as
published by the Free Software Foundation.
Linking this library statically or dynamically with other modules is making
a combined work based on this library. Thus, the terms and conditions of
the GNU General Public License cover the whole combination.
As a special exception, the copyright holders of this library give you
permission to link this library with independent modules to produce an
executable, regardless of the license terms of these independent modules,
and to copy and distribute the resulting executable under terms of your
choice, provided that you also meet, for each linked independent module,
the terms and conditions of the license of that module. An independent
module is a module which is not derived from or based on this library. If
you modify this library, you may extend this exception to your version of
the library, but you are not obligated to do so. If you do not wish to do
so, delete this exception statement from your version.
This code is distributed in the hope that it will be useful, but WITHOUT
ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
version 2 for more details (a copy is included in the LICENSE file that
accompanied this code).