child::* in non-final position causes XPathSyntaxException

I can confirm this one. I've added a test case for it.

As you say this is to do with 3.7. Short answer is that we had this in the javacc version of the parser ages ago, the implementation is at the bottom of this comment. Saxpath should have been binned years ago, etc etc.

Long answer: that tokenizer was written to be stateful, using the previous token to determine what the current token should be emitted as. The same fix would probably work here. A direct translation doesn't correspond to a change in 'star()', but a second switch statement in 'nextToken()'.

An implementation would be: instead of storing 'previousToken' in XPathLexer, store 'boolean expectOperator = false;'. Remove the switch statement from identifierOrOperatorName(), and place it at the end of nextToken(), use it to set 'expectOperator'. In the first switch statement in nextToken, the entry for '*' should read:

case '*':
{
if (expectOperator) { token = star(); } else { // or something less ambiguous; I havent checked the knock-on effects of this token = identifier(); }
break;

and the entry for default:
default:
{
if ( isIdentifierStartChar( LA(1) ) )
{
if (expectOperator) { token = operatorName(); } else { token = identifier(); }
}
}

this boils down to much the same change as in the first comment, but there should be a bit less duplication of the previousToken logic.

For completeness, here's how it looked in javacc:

/*
If there is a preceding token and the preceding token is not one of @, ::,
(, [, , or an Operator, then a * must be recognized as a MultiplyOperator
and an NCName must be recognized as an OperatorName.
*/
<OP> TOKEN: {
<MultiplyOperator: "*"> : DEFAULT

<*> TOKEN: {
<LEFT_PAREN:"("> : DEFAULT

ugh. jira completely messed up the formatting there, sorry. Unfortunately the preview button shows a completely different (also wrong) formatting, so I've no idea how to fix that up. You get the gist.

Sorry meant to get back to you on this Elliotte. I'll attach 2 variants of the patch, the first one leaves the getter and setter for PreviousToken in place - in case you want to preserve the API. Both variants add a second token type (STAR_OPERATOR) for * as a multiply op. I can get away without doing that but the changes started to look pretty ugly; you end up using Token.IDENTIFIER for '' instead, and having to test everywhere that's used to see if it matches ''. Bleh. Tests ran mostly ok for me except for a jdom issue which I believe is my environment - its picking up an older/newer jdom somewhere.

Second variant with the unused previousToken methods removed. BTW another issue you run into if you try to do this with identifier() is that it won't collect the token text for '*', you get lots of parse errors because the token text collected becomes ''. Or something. I didn't delve too far into that because this solution was neater.

Brian's patch works. Committed