Document the tradeoffs of the `Dict` memory optimization opportunity in a code comment.
PiperOrigin-RevId: 437299999
diff --git a/src/main/java/net/starlark/java/eval/Dict.java b/src/main/java/net/starlark/java/eval/Dict.java
index b1a0c16..c7a72ec 100644
--- a/src/main/java/net/starlark/java/eval/Dict.java
+++ b/src/main/java/net/starlark/java/eval/Dict.java
@@ -122,7 +122,13 @@
Preconditions.checkState(mutability != Mutability.IMMUTABLE);
this.mutability = mutability;
// TODO(bazel-team): Memory optimization opportunity: Make it so that a call to
- // `mutability.freeze()` causes `contents` here to become an ImmutableMap.
+ // `mutability.freeze()` causes `contents` here to become an ImmutableMap. Benchmarks show that
+ // for many targets, this can save a small amount of retained heap (up to 1%). But for some
+ // targets the bookkeeping required for this causes unacceptably increased temporary heap, and
+ // the CPU overhead of the bookkeeping and the CPU cost of the ImmutableMap#copyOf call cause
+ // unacceptably increased CPU. In other words, the overall tradeoff is not obviously worth it
+ // in all cases. So be careful making this optimization! See comment #12 of b/225469491 for
+ // details.
this.contents = contents;
}
