Optimizing GeneralComparison to ValueComparison
Saxon attempts to rewrite a GeneralComparison (x = y) as a ValueComparison (x eq y) when it is safe to do so. The reason for this is that a ValueComparison is simpler (and therefore fractionally faster) but more particularly because a ValueComparison is more amenable to further optimizations. For example an xsl:choose of the form
<xsl:choose> <xsl:when test="x = 1">...</xsl:when> <xsl:when test="x = 2">...</xsl:when> <xsl:when test="x = 3">...</xsl:when> <xsl:when test="x = 4">...</xsl:when> </xsl:choose>
is optimized internally to a SwitchExpression provided that the conditions are all ValueComparisons.
This rewrite is not happening for the very common expression of the form (@x = 'value'). In fact, it is only done where both operands have static cardinality exactly one and in this case the LHS cardinality is zero-or-one. It's not clear why this is done, because the result of a ValueComparison when one operand is an empty sequence is effectively false, which works out fine.
Probably it's because (@x = 'value') returns false when @x doesn't exist, while (@x eq 'value') returns an empty sequence; so the rewrite is only safe in a context where we are only interested in the effective boolean value. But ValueComparison has a flag that can be set to force it to return false on empty, so we could set this flag.
Updated by Michael Kay over 4 years ago
I have tried making this change to GeneralComparison.typeCheck() and it causes regression in a few tests, for example Count005. This test uses the query
count(//*[@name='John Doe 4']) lt 1.5
I've fixed this regression by changing the generated CastExpressions to permit an empty sequence.
Updated by Michael Kay almost 4 years ago
- Category set to Performance
- Status changed from New to Resolved
- Priority changed from Low to Normal
- Applies to branch 9.9 added
- Fix Committed on Branch 9.9 added
The change now causes no test regression, so I am committing it.
Specifically, the change is to GeneralComparison.typeCheck(). It now rewrites as a ValueComparison if both operands have cardinality 0-or-1, whereas previously it required exactly-one.
Please register to edit this issue