Saxon

I'll be capturing my observations here as a sort of personal notebook, but sharing it just so you can see how I'm getting on.

Tried it on the development 9.6 branch and it crashed, so I'll put that on one side to be investigated later.

On 9.5, I changed the stylesheet so it only processes the first 5 records. Ran this with -TP output; the dominant cost is function i2a:process on line 149.

Ran with Java profiling, all the cost is going into tree.tidy.SiblingEnumeration, mostly called from UnionEnumeration, but IntersectEnumeration also features. There's no obvious use of union or intersect operations in i2a:process.

By breakpointing UniionEnumeration in the debugger, I see that it's the match pattern on line 16 of semanticMetadata.xsl that's causing the trouble: the pattern is

<xsl:template match="metadata/*[. intersect ../(&SEMANTIC_METADATA;)][1]" priority="1">

where the entity is defined as

funding | event | product | patent

                      | contract | grant | individual | databank

                      | chemical | geneSymbol | spaceFlightMission

A union expression of the form child::A | child::B | child::c ought to be optimized as child::*[self::A or self::B or self::C], but it doesn't look as if this optimization has happened.

It turns out that in 9.5, the expressions used within a predicate in a match pattern are not put through the optimizer. In this particular case the effect of optimization is dramatic.

A patch is being developed. First attempt at a patch solves this performance problem, but causes regression for patterns that use local variables.

Thank you for the recommended fixes, those did work we were able to greatly reduce the time it takes to complete those 9.5 transformations.

We are still however seeing a difference that is big enough to warranty further investigation.

Are there any other updates we could make to the code to further make these numbers be closer?

Latest i2a.xsl and latest results excel uploaded.

I've just been checking on the status of this. It seems that it was fully resolved in the 9.6 branch. However, the attempt to patch 9.5 to optimize pattern predicates led to regression elsewhere in the system, and was abandoned. So the problem remains unresolved on 9.5. However, 9.5 is being managed for stability and any patch that is only for performance improvements has to be srutinized very carefully, so it's probably best to leave things this way.

Closing this bug according to comment 6. Bug fixed in Saxon 9.6.0.1