Saxon

The problem seems to be that the variable reference $last.day is being inlined - that is, we rewrite xsl:on-completion select="$last.day" as xsl:on-completion select="day[last()]", and this rewrite is incorrect because the focus for evaluating the xsl:on-completion is not the same as the focus for the variable declaration. The NPE occurs because we have constructed a focus-dependent expression in a context where there is no focus.

The change has had a side-effect which causes a number of streaming tests to fail, for example sf-boolean-100.

This test executes

<xsl:template name="c-100" use-when="true() or $RUN">

  <xsl:variable name="b" select="current-date() gt xs:date('1900-01-01')"/>

  <xsl:stream href="../docs/big-transactions.xml">

    <out>

      <xsl:value-of select="boolean(($b, account/transaction/dummy))"/>

    </out>

  </xsl:stream>

</xsl:template>

Previously variable $b was being inlined, and then the streamability analysis was performed.

Now $b is not inlined. The streamability analysis is OK, but then further optimization is carried out, and this causes the expression (current-date() gt xs:date('1900-01-01')) to be bound to a global variable; a global variable reference appears within the xsl:stream, and at run-time a failure occurs because this new expression has not been analyzed for streamability.

1. I think we should modify the patch so that inlining does happen in this case.

2. But this still leaves a design problem: we should really not change the tree after doing the inversion, otherwise we are going to hit this kind of problem.

I have fixed (1) by ensuring that the new rule inhibiting variable inlining when the focus changes applies only if the expression is focus-dependent.

As for (2), I'm just going to put a comment in the code noting the problem for now.