Fix two QualifiedName join orders

[jketema]

Before on StanfordLegion__legion with cpp/throwing-pointer:

Pipeline standard for QualifiedName::Namespace.getQualifiedName/0#cbc0648a@7ff329j5 was evaluated in 2 iterations totaling 0ms (delta sizes total: 70).
        162061  ~0%    {2} r1 = JOIN `QualifiedName::Namespace.getQualifiedName/0#cbc0648a#prev_delta` WITH namespacembrs ON FIRST 1 OUTPUT Rhs.1, Lhs.1
            70  ~2%    {4}    | JOIN WITH namespaces ON FIRST 1 OUTPUT Lhs.0, _, Lhs.1, Rhs.1
            70  ~0%    {2}    | REWRITE WITH Tmp.1 := "::", Out.1 := (In.2 ++ Tmp.1 ++ In.3) KEEPING 2
            70  ~0%    {2}    | AND NOT `QualifiedName::Namespace.getQualifiedName/0#cbc0648a#prev`(FIRST 2)
                       return r1

Pipeline standard for QualifiedName::Namespace.getAQualifierForMembers/0#132b16e1@cfd47189 was evaluated in 2 iterations totaling 3ms (delta sizes total: 85).
            12   ~0%    {2} r1 = JOIN `QualifiedName::Namespace.getAQualifierForMembers/0#132b16e1#prev_delta` WITH _#namespace_inlineMerge_#namespacembrsMerge#join_rhs ON FIRST 1 OUTPUT Rhs.1, Lhs.1

        162417   ~0%    {2} r2 = JOIN `QualifiedName::Namespace.getAQualifierForMembers/0#132b16e1#prev_delta` WITH namespacembrs ON FIRST 1 OUTPUT Rhs.1, Lhs.1
            73   ~1%    {4}    | JOIN WITH namespaces ON FIRST 1 OUTPUT Lhs.0, _, Lhs.1, Rhs.1
            73   ~0%    {2}    | REWRITE WITH Tmp.1 := "::", Out.1 := (In.2 ++ Tmp.1 ++ In.3) KEEPING 2

            85   ~0%    {2} r3 = r1 UNION r2
            85   ~0%    {2}    | AND NOT `QualifiedName::Namespace.getAQualifierForMembers/0#132b16e1#prev`(FIRST 2)
                        return r3

After:

Pipeline standard for QualifiedName::Namespace.getQualifiedName/0#cbc0648a@91677d3f was evaluated in 2 iterations totaling 0ms (delta sizes total: 70).
        70  ~0%    {4} r1 = JOIN `QualifiedName::Namespace.getQualifiedName/0#cbc0648a#prev_delta` WITH _#namespacembrsMerge_1#antijoin_rhs_#namespacembrsMerge_10#join_rhs_#namespacesMerge#join_rhs ON FIRST 1 OUTPUT Rhs.1, _, Lhs.1, Rhs.2
        70  ~0%    {2}    | REWRITE WITH Tmp.1 := "::", Out.1 := (In.2 ++ Tmp.1 ++ In.3) KEEPING 2
        70  ~0%    {2}    | AND NOT `QualifiedName::Namespace.getQualifiedName/0#cbc0648a#prev`(FIRST 2)
                   return r1

Pipeline standard for QualifiedName::Namespace.getAQualifierForMembers/0#132b16e1@3bbc99mb was evaluated in 2 iterations totaling 0ms (delta sizes total: 85).
        12   ~0%    {2} r1 = JOIN `QualifiedName::Namespace.getAQualifierForMembers/0#132b16e1#prev_delta` WITH _#namespace_inlineMerge_#namespacembrsMerge_1#antijoin_rhs__#namespacembrsMerge_#namespacembrsMerge___#join_rhs ON FIRST 1 OUTPUT Rhs.1, Lhs.1

        73   ~0%    {4} r2 = JOIN `QualifiedName::Namespace.getAQualifierForMembers/0#132b16e1#prev_delta` WITH _#namespacembrsMerge_1#antijoin_rhs_#namespacesMerge__#namespacembrsMerge_#namespacembrsMerge_10#joi__#join_rhs ON FIRST 1 OUTPUT Rhs.1, _, Lhs.1, Rhs.2
        73   ~1%    {2}    | REWRITE WITH Tmp.1 := "::", Out.1 := (In.2 ++ Tmp.1 ++ In.3) KEEPING 2

        85   ~0%    {2} r3 = r1 UNION r2
        85   ~0%    {2}    | AND NOT `QualifiedName::Namespace.getAQualifierForMembers/0#132b16e1#prev`(FIRST 2)
                    return r3

[Copilot]

Pull request overview

This PR improves evaluation performance for C++ qualified name construction by adjusting join order hints in the internal QualifiedName library, targeting expensive namespacembrs joins observed on large databases.

Changes:

• Add pragma[only_bind_out](this) to the namespacembrs calls in Namespace.getQualifiedName().
• Add pragma[only_bind_out](this) to the namespacembrs calls in Namespace.getAQualifierForMembers().

Show a summary per file

File	Description
cpp/ql/lib/semmle/code/cpp/internal/QualifiedName.qll	Adds only_bind_out pragmas to influence join order for namespace member lookups in qualified-name predicates.

Copilot's findings

• Files reviewed: 1/1 changed files
• Comments generated: 0

[aschackmull]

Whether this is a fix or whether this was even an issue beforehand is a bit unclear to me, since there were only two iterations with no indication of what the base delta count was, and in the "after" tuple counts the new loop-invariant prefix is unaccounted for. But the pragma does look like it enforces the new join-order (i.e. pushing the join between namespacembrs and namespaces into a loop-invariant prefix), so the QL change does appear to match the intention, hence I'll tentatively approve.

[jketema]

Thanks for the review. I agree. It's mostly because this is polluting the stable join-order table in DCA quite significantly. Maybe that table is not really measuring the right thing, but that's a separate discussion, I think.