Trying to XInclude parts from OpenDocument files fails due to a bug in
libxml2 (snapshot 2014-05-31) in xmlXPathNodeCollectAndTest().  The
cause is that xmlXPtrEvalXPtrPart() normalizes namespace URIs using
xmlSaveUri() before registering them whereas the normal XML parser
does no normalization.  Thus namespace URI comparison with
xmlStrEqual() fails for some URIs, e.g. ones containing ':' in their
path part.  The following shows how to trigger and locate the bug in
gdb.


$ > cat data.xml
<rootB xmlns="abc://d/e:f"/>

$ > cat xinc.xml
<rootA xmlns="wxy://z" xmlns:xi="http://www.w3.org/2001/XInclude">
<xi:include href="data.xml"
xi:xpointer="xmlns(b=abc://d/e:f) xpointer(/b:rootB)"/>
</rootA>

$ > xmllint --xinclude xinc.xml
xinc.xml:3: element include: XInclude error : XPointer evaluation failed: #xmlns(b=abc://d/e:f) xpointer(/b:rootB)
xinc.xml:3: element include: XInclude error : could not load data.xml, and no fallback was found
<?xml version="1.0"?>
<rootA xmlns="wxy://z" xmlns:xi="http://www.w3.org/2001/XInclude">
<xi:include href="data.xml" xi:xpointer="xmlns(b=abc://d/e:f) xpointer(/b:rootB)"/>
</rootA>


$ > gdb xmllint
...
(gdb) start --xinclude xinc.xml
...
(gdb) break xmlStrEqual if (str1 && strcmp(str1,"abc://d/e%3Af")==0)
Breakpoint 2 at 0xb7f10fed: file xmlstring.c, line 158.
(gdb) cont
...
(gdb) finish
...
(gdb) print URI
$2 = (const xmlChar *) 0x8001ea28 "abc://d/e%3Af"
(gdb) print cur->ns->href
$3 = (const xmlChar *) 0x8001e4b0 "abc://d/e:f"


Here in xmlXPathNodeCollectAndTest() the value of URI comes from
xmlXPathNsLookup() whose values are normalized in
xmlXPtrEvalXPtrPart() using xmlSaveUri() before being fed to
xmlXPathRegisterNs().  Because the value of cur->ns->href is never
normalized in the same manner, xmlStrEqual() fails if the normalized
value differs from the non-normalized value.

xpointer.c: 979 xmlXPtrEvalXPtrPart(...) {
xpointer.c: ...
xpointer.c:1085         value = xmlParseURI((const char *)ctxt->cur);
xpointer.c: ...
xpointer.c:1092         URI = xmlSaveUri(value);
xpointer.c: ...
xpointer.c:1101         xmlXPathRegisterNs(ctxt->context, prefix, URI);

xpath.c:12075 xmlXPathNodeCollectAndTest(...
xpath.c:  ...
xpath.c:12149         URI = xmlXPathNsLookup(xpctxt, prefix);
xpath.c:  ...
xpath.c:12433              (xmlStrEqual(URI, cur->ns->href)))
xpath.c:  ...
xpath.c:12450              (xmlStrEqual(URI, cur->ns->href)))
xpath.c:  ...
xpath.c:12482                      (xmlStrEqual(URI, cur->ns->href)))
xpath.c:  ...
xpath.c:12501                          (xmlStrEqual(URI,


I see several approaches to this:

 1. Change xmlXPtrEvalXPtrPart() so it doesn't call xmlSaveUri() when
    registering a new namespace.

 2. Add a new function for comparing URIs which ensures normalization
    and use it everywhere URIs are compared.  It could use a hash
    table of already seen URIs for speed.

 3. Ensure everywhere that only normalized URIs make it into all trees
    during parsing.  Or store normalized URIs alongside the
    non-normalized ones.

 4. Maybe something else entirely.

Personally, I'd prefer a solution with URI normalization.  Otherwise
documents like the following won't be recognized as invalid even
though the xi:include element has the same attribute twice.

<rootA xmlns="abc://a"
       xmlns:xi="http://www.w3.org/2001/XInclude"
       xmlns:xo="http://www.w3.org/2001/XInclud%65">
<xi:include href="data.xml" xi:xpointer=" " xo:xpointer=" "/>
</rootA>