Bug #4765: NPE in LoopLifter (from StackOverflow) - Saxon - Saxonica Developer Community

Actions

Send by e-mail Copy link

Bug #4765

closed

NPE in LoopLifter (from StackOverflow)

Added by Norm Tovey-Walsh over 3 years ago. Updated over 3 years ago.

Status:

Closed

Priority:

Normal

Assignee:

Michael Kay

Category:

Internals

Sprint/Milestone:

Start date:

2020-09-29

Due date:

% Done:

100%

Estimated time:

Legacy ID:

Applies to branch:

10, 9.9, trunk

Fix Committed on Branch:

10, 9.9, trunk

Fixed in Maintenance Release:

9.9.1.8, 10.3

Platforms:

Description

From StackOverflow, https://stackoverflow.com/questions/64114488/

declare function local:test($input as element(input)) as element(output)
{
  <output>
      <details>
          <service>
          {
              for $i in 1 to count($input/foo)
              return
                  for $j in 1 to count($input/bar)
                  return if ($input/foo[$i]/data = 'A' and $input/baz[$i]/type = 'B')
                         then <result/>
                         else ()
          }
          </service>
      </details>
  </output>
};

declare variable $input as element(input) external := <input>
</input>;

local:test($input)

raises NPE:

java.lang.NullPointerException
        at net.sf.saxon.expr.parser.LoopLifter.markDependencies(LoopLifter.java:168)
        at net.sf.saxon.expr.parser.LoopLifter.gatherInfo(LoopLifter.java:112)
        at net.sf.saxon.expr.parser.LoopLifter.gatherInfo(LoopLifter.java:122)
        at net.sf.saxon.expr.parser.LoopLifter.gatherInfo(LoopLifter.java:122)
        at net.sf.saxon.expr.parser.LoopLifter.gatherInfo(LoopLifter.java:122)
...

History
Notes
Property changes

Actions

Copy link

Updated by Michael Kay over 3 years ago

Priority changed from Low to Normal

Running with -explain triggers additional verification of the expression tree at various stages during optimisation, leading to earlier failure:

java.lang.IllegalStateException: Invalid parent pointer in ($input/(child::element(Q{}foo)[$i]))/data = "A" and ($input/(child::element(Q{}baz)[$i]))/type = "B" subexpression ($input/(child::element(Q{}foo)[$i]))/data = "A"
	at net.sf.saxon.expr.Expression.verifyParentPointers(Expression.java:229)
	at net.sf.saxon.expr.Expression.verifyParentPointers(Expression.java:239)
	at net.sf.saxon.expr.parser.Optimizer.trace(Optimizer.java:555)
	at net.sf.saxon.expr.ForExpression.optimize(ForExpression.java:132)
	at net.sf.saxon.expr.flwor.FLWORExpression.rewriteForOrLet(FLWORExpression.java:782)
	at net.sf.saxon.expr.flwor.FLWORExpression.optimize(FLWORExpression.java:495)
	at net.sf.saxon.expr.flwor.FLWORExpression.optimize(FLWORExpression.java:488)
	at net.sf.saxon.expr.Operand.optimize(Operand.java:228)
	at net.sf.saxon.expr.Expression.optimizeChildren(Expression.java:618)
	at net.sf.saxon.expr.instruct.ParentNodeConstructor.optimize(ParentNodeConstructor.java:191)
	at net.sf.saxon.expr.instruct.FixedElement.optimize(FixedElement.java:107)
	at net.sf.saxon.expr.Operand.optimize(Operand.java:228)
	at net.sf.saxon.expr.Expression.optimizeChildren(Expression.java:618)
	at net.sf.saxon.expr.instruct.ParentNodeConstructor.optimize(ParentNodeConstructor.java:191)
	at net.sf.saxon.expr.instruct.FixedElement.optimize(FixedElement.java:107)
	at net.sf.saxon.expr.Operand.optimize(Operand.java:228)
	at net.sf.saxon.expr.Expression.optimizeChildren(Expression.java:618)
	at net.sf.saxon.expr.instruct.ParentNodeConstructor.optimize(ParentNodeConstructor.java:191)
	at net.sf.saxon.expr.instruct.FixedElement.optimize(FixedElement.java:107)
	at net.sf.saxon.query.XQueryFunction.optimize(XQueryFunction.java:448)
	at net.sf.saxon.query.XQueryFunctionLibrary.optimizeGlobalFunctions(XQueryFunctionLibrary.java:327)
	at net.sf.saxon.query.QueryModule.optimizeGlobalFunctions(QueryModule.java:1207)
	at net.sf.saxon.expr.instruct.Executable.fixupQueryModules(Executable.java:458)
	at net.sf.saxon.query.XQueryParser.makeXQueryExpression(XQueryParser.java:177)
	at net.sf.saxon.query.StaticQueryContext.compileQuery(StaticQueryContext.java:568)
	at net.sf.saxon.query.StaticQueryContext.compileQuery(StaticQueryContext.java:630)
	at net.sf.saxon.s9api.XQueryCompiler.compile(XQueryCompiler.java:609)
	at net.sf.saxon.Query.compileQuery(Query.java:804)
	at net.sf.saxon.Query.doQuery(Query.java:317)
	at net.sf.saxon.Query.main(Query.java:97)

Actions

Copy link

Updated by Michael Kay over 3 years ago

The rewrite that corrupts the tree is in ForExpression.rewriteWhereClause() (there isn't actually a where clause here, but the "return if..." compiles to exactly the same tree as if a where clause had been used). The rewrite extracts all the terms (=operands of AND) in the condition, and moves any of them that don't depend on the control variable of the "for" (that is, $j) up a level, recombining the extracted terms (in this case, all of them) into a new AndExpression. It then notices that all of the terms of the (imaginary) where clause have been promoted, which means the where clause itself can be dropped, and it replaces the entire construct for $i in X return if C then Y else () with if C then (for $i in X return Y) else (). However the C that it uses here is the original AndExpression, rather than the reconstituted AndExpression, and this is no longer valid, because its subexpressions have been transferred to the reconstituted AndExpression, which means their parent pointers are now corrupt.

Actions

Copy link

Updated by Michael Kay over 3 years ago

Category set to Internals
Status changed from New to In Progress
Assignee set to Michael Kay
Applies to branch 9.9, trunk added

The simplest fix is to use the reconstituted AndExpression rather than the original, that is, change ForExpression#281 from

return Choose.makeConditional(condition, this);

return Choose.makeConditional(promotedCondition, this);

Actions

Copy link

Updated by Michael Kay over 3 years ago

In trying to write a regression test for this, I've found there is another precondition for the bug to trigger: the "where" clause must not be rewritten as a predicate by an earlier phase of optimisation, specifically in FLWORExpression.rewriteWhereClause(). This means it must not actually be a where clause (where C return X), it has to be written as return if C then X else () which looks like a where clause but isn't caught be the earlier optimisation phase.

QT3 Test case WhereExpr035 created: this fails without the patch, and succeeds with it.

Actions

Copy link

Updated by Michael Kay over 3 years ago

On the 9.9 branch, my first attempt to run the new test didn't fail: this turned out to be because I was running the test with option -unfolded, which complicates the query sufficiently to prevent the optimization taking place. Running without that option, I confirmed the patch is needed and works.

However, regression testing the patch under 9.9, I get a new failure in test -s:prod-WhereClause -t:cbcl-left-outer-join-004. The failure occurs whether or not the new patch is present. The failure occurs because we're trying to rewrite the "order-by" expression of a FLWORExpression with something that isn't a TupleExpression.

The problem is in OrderByClause#31, which reads

this.sortKeysOp = new Operand(flwor, new SortKeyDefinitionList(sortKeys), OperandRole.REPEAT_NAVIGATE_CONSTRAINED);

whereas the 10 branch has, correctly,

this.sortKeysOp = new Operand(flwor, new SortKeyDefinitionList(sortKeys), SORT_KEYS_ROLE);

I'm having difficulty searching the history here because of the change from Subversion to Git.

Actions

Copy link

Updated by Michael Kay over 3 years ago

Status changed from In Progress to Resolved
Fix Committed on Branch 10, 9.9, trunk added

The bug in OrderByClause was fixed in the 10 branch somewhere between 10.0 and 10.1, but I haven't been able to track it down to a particular bug entry. I have retrofitted the change to the 9.9 branch, however, so 9.9 now passes these tests.

Actions

Copy link