Get a better (faster) parser, such as XML::LibXML. Using XML::XSH2, I loaded and saved that file, and the result was:

<?xml version="1.0" encoding="utf-8"?>
<AttributeDatabase>
     <attribclass toolonly="false" hidden="false" dbvault="true">
    <editlist name="default">
    <entry>*</entry>
    </editlist>
    <preferrededitlist>"default"</preferrededitlist>
    <attribdef name="mtl_param" synthetic="true" array="true">
        <desc>IsSynthetic</desc>
        <type>Attrib.Types.Vector4</type>
        <bound>True</bound>
        <properties platform="common">
        <mincount>2</mincount>
        <maxcount>2</maxcount>
        </properties>
    </attribdef>
    <attribdef name="mtl_Kdiff" toolonly="true">
        <memberkeyvalue path="" key="push(bits)-&gt;::mtl_param:common
+:0:w"/>
        <desc>controls --&gt; mtl_param[0].w </desc>
        <type>EA.Reflection.Float</type>
        <bound>False</bound>
        <increment>0.001</increment>
        <properties platform="common">
            <max>2.0</max>
            <min>0.0</min>
            <defaultreflectedobject>1.0</defaultreflectedobject>
        </properties>
    </attribdef>    
     </attribclass>
</AttributeDatabase>
[download]

And using the XSH2 language, I could have modified the tree with a mixture of XPath loops and Perl expressions. It's quite elegant, and even compiles down to Pure Perl code.

I have an article on using xsh2 to scrape HTML as well (embargoed at the moment, should go live in a few weeks).