Saxon

I should note that this issue also appears in releases 12.0 and 12.1 but is not present in 11.5.

I've reproduced this running SaxonJ-HE 12.2 from the command line, but it works fine when running SaxonJ-EE in IntelliJ.

Which makes it a bit tricky to debug...

For the record, it also works in SaxonJ-EE 12.2 when run from the command line (with a license file; but fails without one).

Now reproduced in IntelliJ by renaming my license file so I'm effectively running without one.

Noted that if I change the XSLT to say (1 to 20) I get no failure; if I change it to (1 to 24) then it still fails on file 22. So it's not just because it's the last one. Weird.

What seems to be happening is that the call on error() is loop-lifted because it has no dependencies on the context item. The execution strategy changes after the first 20 evaluations (or so) because (since Saxon 12) we're choosing between lazy and eager evaluation based on experience. But an expression that's been loop lifted in this way should force lazy evaluation to avoid spurious errors.

No idea why this is happening in HE and not EE, probably EE does some other optimization which takes things down a different path by chance.

Here's the -explain optimizer trace:

OPT : At line 29 of file:/Users/mike/bugs/2023/6022-Merchant/test.xsl
OPT : Inlined constant variable regex
OPT : Expression after rewrite: ^/tmp/xml/doc-\d+\.xml$
OPT : At line 38 of file:/Users/mike/bugs/2023/6022-Merchant/test.xsl
OPT : Lifted (fn:string-join(...)) above (analyzeString) on line 33
OPT : Expression after rewrite: let $Q{http://saxon.sf.net/generated-variable}v0 := string-join(convertTo_xs:string(data(error(QName(error, err:path-syntax), concat(The path ", $path, " does not match regex ", ^/tmp/xml/doc-\d+\.xml$, ")))),  ) return (exactly-one(convertTo_xs:string(data(AnalyzeString($path, ^/tmp/xml/doc-\d+\.xml$, , ValueOf($path), ValueOf($Q{http://saxon.sf.net/generated-variable}v0))))))

What should happen now is that the generated variable v0 is flagged as requiring lazy evaluation. Somewhere along the line this flag is being lost or ignored.

I've changed the lazily() method of Elaborator to pass a parameter indicating that lazy evaluation is required, but this isn't enough to fix the problem.

What seems to be happening is that we have in effect a LetExpression with two variables:

let $path := concat('/tmp/xml/doc-', $iter, '.xml'),
     $v0 := string-join(error(), ...)
return analyse-string(....)

and because it's decided to evaluate the first variable eagerly (which is reasonable) it's also evaluating the second one eagerly (which isn't).

Changing LetExprElaborator.eagerly() to invoke LetExprElaborator.lazily() when the needsLazyEvaluation flag is set solves the problem.

Regression testing.

Seeing some regression:

as : 3 catalog : 2 choose : 2 xslt-compat : 1

as-1303 does:

java.lang.ClassCastException: class net.sf.saxon.value.MemoClosure cannot be cast to class net.sf.saxon.om.GroundedValue (net.sf.saxon.value.MemoClosure and net.sf.saxon.om.GroundedValue are in unnamed module of loader 'app')
	at net.sf.saxon.expr.parser.ExpressionTool.eagerEvaluate(ExpressionTool.java:195)
	at net.sf.saxon.expr.instruct.GlobalVariable.getSelectValue(GlobalVariable.java:645)
	at net.sf.saxon.expr.instruct.GlobalVariable.actuallyEvaluate(GlobalVariable.java:739)
	at net.sf.saxon.expr.instruct.GlobalVariable.evaluateVariable(GlobalVariable.java:708)

Fixed by changing ExpressionTool.eagerEvaluate() to call Sequence.materialize() rather than assuming the returned Sequence will be a GroundedValue.