[lopster568]

Rebased on master since the August approval. The cost code is ported to OverloadModel (see Implementation notes), and a repl test that was failing CI is fixed (see Testing). The API and semantics are unchanged.

Description

CEL has no way to combine two maps. The + operator concatenates strings, bytes, and lists, but has no map overload, and there is no maps extension.

This adds one, with a single member function:

<map(K, V)>.putAll(<map(K, V)>) -> map(K, V)

It returns a new map holding the entries of both maps, and the argument's values win on conflicting keys. Neither input is modified. The copy is shallow, so a value that is itself a map is replaced rather than merged recursively.

{}.putAll({})                              // {}
{'a': 1}.putAll({'b': 2})                  // {'a': 1, 'b': 2}
{'a': 1}.putAll({'a': 2})                  // {'a': 2}
{'a': {'x': 1}}.putAll({'a': {'y': 2}})    // {'a': {'y': 2}}

Scope

This is deliberately only the replace step from #1240. @TristonianJones suggested starting there:

I think replace semantics is where I'd start -- last write wins, followed by set-if-absent as the deep merge is the trickiest case

So set-if-absent and recursive merge are left for follow-ups.

Implementation notes

• Cost is described by an OverloadModel registered with cel.CostModel, following the ext/lists migration in Migrate ext/lists to OverloadModel and add strategy soundness tests #1566. The call charges for visiting every entry of both maps plus a map allocation. The result size ranges from the larger input's entry count (every key collides) to the sum of both (no key collides), with key and value sizes taken from either input.
• At cost model version 0 the estimates and tracked costs are unchanged from the earlier revision of this PR.
• Entries are read with Find rather than Fold, for the reason given in the review thread: Fold hands back native Go values for reflect and string backed maps, which would then have to be adapted again.
• Registered in ext/extension_option_factory.go (factory and alias) and in repl/evaluator.go. ext/README.md gains a Maps section.

Testing

ext/maps_test.go covers empty inputs, disjoint keys, conflicting keys, nested maps being replaced rather than merged, non-string keys, a variable receiver, compile-time rejection of non-map arguments, non-map arguments at runtime, and cost estimates and tracked costs at the latest model version and at version 0. The putAll cases are also added to TestCostSoundnessBySizingStrategy.

The earlier cel-go-builder run failed in its bazel-test step. Running cd repl && go test ./... reproduced a failure in TestProcessOptionError, whose expected list of extension options did not include maps. repl is a separate module, so my go test ./... from the root never covered it. That is fixed.

go build ./... and go test ./... pass in every module except conformance, where the policy tests need Bazel runfiles and fail the same way on master. go vet ./... is clean for this change; its only findings are unreachable code warnings in the generated ANTLR parsers, the same count as on master. I did not run the Bazel build locally.

[TristonianJones]

/gcbrun

[TristonianJones]

/gcbrun

[TristonianJones]

/gcbrun

[lopster568]

Rebased onto master and moved the cost estimate to OverloadModel, following #1566. I couldn't read the cel-go-builder log, but TestProcessOptionError in repl failed locally because its expected list didn't include maps; that's fixed too. @TristonianJones could you /gcbrun again when you have a moment?