Re: best practices for using XSLT modes

On 05/12/2019 14:51, Eliot Kimber ekimber@xxxxxxxxxxxx wrote:
> Not really about modes, but I would replace the choice that acts on different @val values with templates applied to the @val attribute, i.e.:
>
>          <xsl:template match="a">
>            <val><xsl:apply-templates select="@val"/></val>
>         </xsl:template>
>
>       <xsl:template match="@val[. ge 0]">
>          <xsl:value-of select="@val || ': positive'"/>
>      </xsl:template>
>
>       <xsl:template match="@val[. lt  0]">
>          <xsl:value-of select="@val || ': negative"/>
>      </xsl:template>
>
> Note that I handle the bug in the original in that it would produce no result when @val is "0" (zero).
>
> The use of templates rather than xsl:choose makes the code cleaner, I think, puts the focus at the template level on the @val attribute, which is the focus of the business logic, and enables extension and override. For example, if you want a value of exactly zero to have a different result, you could do that by adding an override template. The original use of xsl:choose would require overriding the entire template for the <a> element.

I'm of a different opinion, somewhat in line with some remarks Mike made 
earlier. (Note that this topic can sometimes get a little 'heated' in a 
gentle XSLT manner.)

The design issues in part depend on the 'scope' and 'size' of the 
operation you're invoking and the likelihood that you'll need to 'add 
additional semantics' at a later date or by overriding with 'library' 
importation. Lets look at a slight rewriting of the 'choose' method in 
XSLT3:

    <xsl:template match="a" expand-text="true">
     B  B  B  <val>
     B B B B B B B  <xsl:choose>
     B  B  B  B  <xsl:when test="@val ge 0">{@val}: positive</xsl:when>
     B  B  B  B  <xsl:when test="@val lt 0">{@val}: negative</xsl:when>
     B  B  B  </xsl:choose>
     B B B  </val>
    </xsl:template>

(Yes it could be terser or it could be conditional XPath, but bear with 
me.) The point of the choose over the template is that the /entirety /of 
the operation choice semantics (identifying the signum of a/@val) is 
contained in a single place, in fact inside a single XML element, and 
won't be effected by any other additional code /outside/ this segment. 
You're pretty safe in the knowledge that if something's going wrong 
(e.g. the zero case Eliot pointed out and another possible fault 
described later), then the fault should /only be within this code 
section, /nowhere else. And choose also has a totally defined order of 
checking - each when is tested in turn, so you can control accurately 
the sequence of checks.

If that is what you wanted (contained scope), then if you used 
templates, even moded ones, an additional matching template, almost 
anywhere else in your stylesheets, might confound your nice solution. 
For example:

    <xsl:template match="a/@val[. mod 2 eq 0]" priority="2">{@val}:
    even</xsl:template>

would alter the result. And similarly the 'default priority order' 
between template rules is sometimes not as straightforward as it seems. 
Believe me, with large stylesheets spread over many files it's extremely 
easy to find that an errant template somewhere in the back of beyond is 
trumping the code sections you're tearing your hair out debugging. And 
with stylesheet importation this becomes even more prevalent as 
importation precedence wins out over priority every time. This means 
that for example if your stylesheet was imported into another, which 
contained:

    <xsl:template match="a/@val[. eq 0]"><zero/></xsl:template>

then your design will have been 'broken' (perhaps unintentionally) by 
someone else - the designer of the importing stylesheet. (Currently 
there are no mechanisms, outside use of packages, to make templates 
local, even as children of a mode, so all are effectively global in scope.)

But if you do expect to have a large and variable set of conditions that 
in different cases will supercede each other, such as for example in the 
design of DITA-OT, then templates are certainly the way. And if you're 
using next-match to pre/post-process special cases, then you will need 
to use templates.

    Note that in this particular example lurks a perhaps more insidious
    design issue for which a choose may have advantage. What if
    a/@val="abc", assuming no schema-awareness checking for integer
    values. With choose an error would be thrown during evaluation of
    the first when/@test ('abc' cannot be cast to an integer); in the
    template case nothing would happen - errors in patterns mean 'no
    match' and the usual default behaviour for @* is never matching:
    hence the result would be just <val/> and you might be left
    scratching your head.

For me the choices really come down to:

  * How local is this operation? How many separable sections? How large
    are they?
  * Are some parts likely to be used from another section of the program?
  * Do I expect to modify it, or update it, or override it?

etc... For some (most?) templates are the answer, for others choose wins 
out. The most interesting are of course a toss-up!



-- 
*John Lumley* MA PhD CEng FIEE
john@xxxxxxxxxxxx <mailto:john@xxxxxxxxxxxx>
on behalf of Saxonica Ltd