Serializing supplementary unicode characters into XML documents with Java

I am trying to serialize DOM documents with supplementary unicode characters such as U+1D49C (๐’œ, mathematical script capital A). Creating a node with such a character is not a problem (I just set the node value to the UTF-16 equivalent, “uD835uDC9C”). When serializing, however, Xalan and XSLTC (with a Transformer) and Xerces (with LSSerializer) all create invalid character entities like “��” instead of “𝒜”. I tried the “normalize-characters” parameter for LSSerializer, but it is not supported. Only Saxon gets it right, without using a character entity when the encoding is unicode.

I cannot use Saxon in practice (among other reasons, I use Java applets and do not want to load another jar), so I am looking for a solution with the default JDK libraries. Is it possible to get valid XML documents serialized from a DOM document with supplementary unicode characters ?

[edit] I found someone else who encountered this problem : http://www.dragishak.com/?p=131

[edit2] actually, it seems to work with LSSerializer when I don’t have xerces on the classpath (the class used is com.sun.org.apache.xml.internal.serialize.DOMSerializerImpl). It does not work with a transformer and com.sun.org.apache.xalan.internal.xsltc.trax.TransformerFactoryImpl.

Answer

Since I didn’t see any answer coming, and other people seem to have the same problem, I looked into it further…

To find the origin of the bug, I used the serializer source code from Xalan 2.7.1, which is also used in Xerces.

org.apache.xml.serializer.dom3.LSSerializerImpl uses org.apache.xml.serializer.ToXMLStream, which extends org.apache.xml.serializer.ToStream.

ToStream.characters(final char chars[], final int start, final int length) handles the characters, and does not support unicode characters properly (note: org.apache.xml.serializer.ToTextSream (which can be used with a Transformer) does a better job in the characters method, but it only handles plain text and ignores all markup; one would think that XML files are text, but for some reason ToXMLStream does not extend ToTextStream).

org.apache.xalan.transformer.TransformerIdentityImpl is also using org.apache.xml.serializer.ToXMLStream (which is returned by org.apache.xml.serializer.SerializerFactory.getSerializer(Properties format)), so it suffers from the same bug.

ToStream is using org.apache.xml.serializer.CharInfo to check if a character should be replaced by a String, so the bug could also be fixed there instead of directly in ToStream. CharInfo is using a propery file, org.apache.xml.serializer.XMLEntities.properties, with a list of character entities, so changing this file could also be a way to fix the bug, although so far it is designed just for the special XML characters (quot,amp,lt,gt). The only way to make ToXMLStream use a different property file than the one in the package would be to add a org.apache.xml.serializer.XMLEntities.properties file before in the classpath, which would not be very clean…

With the default JDK (1.6 and 1.7), TransformerFactory returns a com.sun.org.apache.xalan.internal.xsltc.trax.TransformerImpl, which uses com.sun.org.apache.xml.internal.serializer.ToXMLStream. In com.sun.org.apache.xml.internal.serializer.ToStream, characters() is sometimes calling processDirty(), which calls accumDefaultEscape(), which could handle unicode characters better, but in practice it does not seem to work (maybe processDirty is not called for unicode characters)…

com.sun.org.apache.xml.internal.serialize.DOMSerializerImpl is using com.sun.org.apache.xml.internal.serialize.XMLSerializer, which supports unicode. Strangely enough, XMLSerializer comes from Xerces, and yet it is not used by Xerces when xalan or xsltc are on the classpath. This is because org.apache.xerces.dom.CoreDOMImplementationImpl.createLSSerializer is using org.apache.xml.serializer.dom3.LSSerializerImpl when it is available instead of org.apache.xerces.dom.DOMSerializerImpl. With serializer.jar on the classpath, org.apache.xml.serializer.dom3.LSSerializerImpl is used. Warning: xalan.jar and xsltc.jar both reference serializer.jar in the manifest, so serializer.jar ends up on the classpath if it is in the same directory and either xalan.jar or xsltc.jar is on the classpath ! If only xercesImpl.jar and xml-apis.jar are on the classpath, org.apache.xerces.dom.DOMSerializerImpl is used as the LSSerializer, and unicode characters are properly handled.

CONCLUSION AND WORKAROUND: the bug lies in Apache’s org.apache.xml.serializer.ToStream class (renamed com.sun.org.apache.xml.internal.serializer.ToStream inside the JDK). A serializer that handles unicode characters properly is org.apache.xml.serialize.DOMSerializerImpl (renamed com.sun.org.apache.xml.internal.serialize.DOMSerializerImpl inside the JDK). However, Apache prefers ToStream instead of DOMSerializerImpl when it is available, so maybe it behaves better for other things (or maybe it’s just a reorganization). On top of that, they went as far as deprecating DOMSerializerImpl in Xerces 2.9.0. Hence the following workaround, which might have side effects :

  • when Xerces and Apache’s serializer are on the classpath, replace “(doc.getImplementation()).createLSSerializer()” by “new org.apache.xerces.dom.DOMSerializerImpl()

  • when Apache’s serializer is on the classpath (for instance because of xalan) but not Xerces, try to replace “(doc.getImplementation()).createLSSerializer()” by “new com.sun.org.apache.xml.internal.serialize.DOMSerializerImpl()” (a fallback is necessary because this class might disappear in the future)

These 2 workarounds produce a warning when compiling.

I don’t have a workaround for XSLT transforms, but this is beyond the scope of the question. I guess one could do a transform to another DOM document and use DOMSerializerImpl to serialize.

Some other workarounds, which might be a better solution for some people :

  • use Saxon with a Transformer

  • use XML documents with UTF-16 encoding