From saxon

When does Saxon build/use an index?

1. Saxon always builds an ID index for every document to support the id() function, whether you use it or not.
2. Saxon builds an IDREF index to support the idref() function the first time you use idref() on a given document tree.
3. In XSLT, Saxon builds an index to support xsl:key and the key() function the first time you use the key() function for that key on a given document tree. This index is reused if you use the same stylesheet repeatedly on the same source document.
4. The first time you use //x on a given document tree, Saxon keeps the result, linked to the document node. So effectively there is an index of elements in the document by name, but it is built incrementally one name at a time. This is used only to support expressions starting //x, it isn't used for example if you write y/x.
5. In Saxon-EE, any filter expression such as /a/b/c[@x=$y] is in effect translated into a key definition with match="/a/b/c" use="@x", together with a call on key('k', $y) - that is, a value index. There's of course a complex set of rules determining exactly when this is possible. The important thing is that the expression being filtered must be a path expression that can be statically determined to be rooted at a document node. This index is associated with the document node, and lasts as long as the document.
6. In Saxon-EE, a filter expression that isn't rooted at a document node, for example $abc[@x=$y], will result in creation of a temporary index for the variable $abc provided that the expression occurs in a loop relative to the creation of $abc - that is, if it appears the index will be used more than once. Such an index lasts only as long as the variable. (In many cases the variable won't be a user-written variable, but a system-created variable caused by extracting an expression from a loop). This often arises as a result of evaluating a join.
7. There's also a temporary index or hash-table used to support general comparisons A=B if both A and B are multi-valued.

Apart from the //x case, Saxon doesn't use structure indexes. Instead it relies on the fact that searching for elements with a known name in the TinyTree structure is extremely fast. (One researcher doing benchmarks concluded from his measurements that Saxon must be using indexes, but he was wrong.)