Using for-each-group for high performance XSLT

Question

I have an XSLT (1.0) style sheet. It works with no problem. I want to make it to 2.0. I want to use xsl:for-each-group (and make it have high performance). It is possible? How? Please explain.

I have many places like

    <xsl:if test="test condition">
     <xsl:for-each select="wo:tent">
     <width aidwidth='{/wo:document/styles [@wo:name=current()/@wo:style-name]/@wo:width}'
</xsl:for-each>
    </xsl:if>

ADDED

<xsl:template match="wo:country">
            <xsl:for-each select="@*">
                <xsl:copy/>
            </xsl:for-each>
            <xsl:variable name="states" select="wo:pages[@xil:style = &quot;topstates&quot; or @xil:style = &quot;toppage-title&quot;]"/>
            <xsl:variable name="provinces" select="wo:pages[@xil:style = &quot;topprovinces&quot;]"/>
            <xsl:choose>
                <xsl:when test="$states">
                    <xsl:apply-templates select="$states[2]/preceding-sibling::*"/>
                    <xsl:apply-templates select="$states[2]" mode="states">
                        <xsl:with-param name="states" select="$states[position() != 0]"/>
                    </xsl:apply-templates>
                </xsl:when>
                <xsl:when test="$provinces">
                    <xsl:apply-templates select="$provinces[2]/preceding-sibling::*"/>
                    <xsl:apply-templates select="$provinces[2]" mode="provinces">
                        <xsl:with-param name="provinces" select="$provinces[position() != 2]"/>
                    </xsl:apply-templates>
                </xsl:when>
                <xsl:otherwise>
                    <xsl:apply-templates/>
                </xsl:otherwise>
            </xsl:choose>
    </xsl:template>

THE SOURCE

<?xml version="1.0" encoding="UTF-8"?>
<wo:country>
   some stuff
</wo:country>

Answer 1

I have assumed that you want an in-depth description of xsl:for-each-group and how to use it. If this is not what you are asking for, then please let me know.

The instruction, new in XSLT 2.0, takes a set of items and groups them. The set of items is called "the population", and the groups are just called groups. The instruction processes each group in turn.

Possible attributes of the xsl:for-each-group instruction include:

1. select
2. group-by
3. group-adjacent
4. group-starting-with
5. group-ending-with
6. collation

@select is mandatory. The others are optional. It can take any number of xsl:sort children (but they must come first), followed by a sequence constructor. A "sequence constructor" is the term for all the sequence emitting type instructions that go inside templates and the like.

@select

The select attribute specifies an XPATH expression which evaluates to the population to be grouped.

@group-by

The group-by attribute specifies an XPATH expression, which you use when the type of grouping is by common value. Every item in the population that evaluates to the same group-by value as another is in the same group as that other.

XSLT 1.0 Muenchian grouping is not too difficult when the type of grouping is group by common value. There are two more common forms of grouping: group adjacent items by similar value; and group an adjacent group of items whose group is either demarcated at the end or the at the beginning by some test. While both these forms of grouping are still possible with Muenchian, it becomes relatively complex. Muenchian on these types will also be less efficient at scale, because of the use of sibling axises (however you spell that!).

Another advantage of XSLT 2.0 that comes to mind is that Muenchian only works on node sets, whereas xsl:for-each-group is broader in application because it works on a sequence of items, not just nodes.

The result of the @group-by expression will be a sequence of items. This sequence is atomized and de-duped. The population item being tested will be a member of one group per value. It's a strange consequence, that with @group-by, and item may be a member of more than one group, or perhaps even none. Although I suspect that any thing that you can do in XSLT 2.0, you can, by some tortuous path, do in XSLT 1.0, the ability to put an item into two groups is something that would be quiet fiddly to do in XSLT 1.0 Muenchian.

@group-adjacent

The attributes group-by, group-adjacent, group-starting-with and group-ending-with are mutually exclusive because they specify different kinds of grouping. Items with commons values and adjacent in the population are grouped together. Unlike @group-by, @group-adjacent must evaluate to, after atomization, a single atomic value.

group-starting-with

Unlike select, group-adjacent and group-by, this attribute does not specify an XPATH select expression, but rather a pattern, in the same way the xsl:template/@match specifies a pattern, not a selection. If an item in the population passes the pattern test or is the first item in the population then it starts a new group. Otherwise the item continues the group from the previous item.

Martin mentioned the spec examples (w3.org/TR/xslt20/#grouping-example). From that reference, I am going to copy the example entitled "Identifying a Group by its Initial Element", but alter it slightly to emphasis the point about the initial item of the population.

So this is our input document (copied from w3 spec. The inclusion of the orphaned line is mine) ...

<body>
  <p>This is an orphaned paragraph.</p>
  <h2>Introduction</h2>
  <p>XSLT is used to write stylesheets.</p>
  <p>XQuery is used to query XML databases.</p>
  <h2>What is a stylesheet?</h2>
  <p>A stylesheet is an XML document used to define a transformation.</p>
  <p>Stylesheets may be written in XSLT.</p>
  <p>XSLT 2.0 introduces new grouping constructs.</p>
</body>

... what we want to do is define groups as nodes starting with h2 and include all the following p up until the next h2. The example solution given by w3 is to use @group-starting-with ...

<xsl:template match="body">
  <chapter>
        <xsl:for-each-group select="*" group-starting-with="h2"      >
          <section title="{self::h2}">
            <xsl:for-each select="current-group()[self::p]">
              <para><xsl:value-of select="."/></para>
            </xsl:for-each> 
          </section>
        </xsl:for-each-group>
  </chapter>
</xsl:template>

In the spec example, when the input does not contain an orphan line, this produces the desired result ...

<chapter>
  <section title="Introduction">
    <para>XSLT is used to write stylesheets.</para>
    <para>XQuery is used to query XML databases.</para>
  </section> 
  <section title="What is a stylesheet?">
    <para>A stylesheet is an XML document used to define a transformation.</para>
    <para>Stylesheets may be written in XSLT.</para>
    <para>XSLT 2.0 introduces new grouping constructs.</para>
  </section>
</chapter>

Although in our particular case we get instead ...

<chapter>
   <section title="">
      <para>This is an orphaned paragraph.</para>
   </section>
   <section title="Introduction">
      <para>XSLT is used to write stylesheets.</para>
      <para>XQuery is used to query XML databases.</para>
   </section>
   <section title="What is a stylesheet?">
      <para>A stylesheet is an XML document used to define a transformation.</para>
      <para>Stylesheets may be written in XSLT.</para>
      <para>XSLT 2.0 introduces new grouping constructs.</para>
   </section>
</chapter>

If the initial section for the orphaned lines is undesired, there are easy solutions. I won't go into them now. My point is just to high-light the fact that the first group resulting from @group-starting-with can be an 'orphan' group. By 'orphan', I mean a group whose head node does not fit the specified pattern.

@collation

The collation attribute specifies a collation URI and identifies a collation used to compare strings for equality.

current-group()

Within the xsl:for-each-group the current-group() function returns the current group being processed as a sequence of items.

current-grouping-key()

Within the xsl:for-each-group the current-group() function returns the current group key. I am not sure, but I believe that this can only be an atomic type. Also not sure, but I believe that this function is only applicable to @group-by and @group-adjacent type of grouping.

@group-by versus @group-adjacent

In some scenarios you will have a choice between these two sort types with the same functional result. When this is the case @group-adjacent is to be preferred over @group-by, because it will likely be more efficient to process.

Pattern versus Select

Some XSLT 2.0 instruction attributes contain select expressions. Michael Kay calls these "XPath expressions". Personally, when juxtaposing against patterns, I feel a better description would be "select expression". Other attributes contain patterns or "match expressions". While these two both contain the same syntax, they are very different beasts. The similarity between the two often makes XSLT beginners think of xsl:template/@match not as a pattern, but as a select expression. The consequence has been a lot of confusion from beginners about the value of the position() function within template's sequence constructors. As stated earlier, in xsl:for-each-group, @select, @group-by and @group-adjacent are select expressions, but @group-starting-with and @group-ending-with are patterns. So here is the difference:

1. Select expressions are a like a function. The input is a context document, context sequence, context item, context position and of course the actual expression. The output is a sequence of items. Depending where this is actually used, this could become the next context sequence. The default axis is child:: .
2. Unlike select expression, the default axis for a pattern is self:: . The pattern is also like a function. Its inputs are as before, and its output is not a sequence, but a boolean. Some item is being tested to see if it matches the pattern or not. The item being tested is made the context item. The match expression is temporarily evaluated as it were a select expression. Then the returned sequence is tested to see if the context item is a member or not. The returned sequence is then discarded. The result is true or 'match' if it was a member, and false otherwise.

Answer 2

Sean has provided a wonderful overview of xsl:for-each-group, which was very generous, but it doesn't really seem to be an answer to your question.

You've shown a fragment of XSLT code, and you've said you want faster performance. But the fragment you showed is not doing grouping, it is doing a join. There are two ways you can speed up a join. Either use an XSLT processor such as Saxon-EE that does automatic join optimization, or optimize it by hand using keys. For example, given this expression:

/wo:document/styles [@wo:name=current()/@wo:style-name]/@wo:width

you could define a key

<xsl:key name="style-name-key" match="styles" use="@wo:name"/>

and then replace the expression by

key('style-name-key', @wo:style-name)/@wo:width