[evanchooly]

Plan item 3.2 (review A2/B4).

The problem

bestConstructor picks a no-arg constructor whenever one exists, so most entities (every POJO-style one) get a ConstructorCreator with an empty parameter plan. That creator still buffered every decoded property in two ArrayLists and replayed them into the instance through setValue once the document was done. It also cloned a zero-length parameter array per instance. The InstanceCreatorFactoryImpl branch that would have used NoArgCreator was unreachable, because bestConstructor had already returned any no-arg constructor.

Change

• No-arg fast path. When the constructor takes no parameters, ConstructorCreator.set constructs the instance on the first value and sets each value through model.setValue as it arrives. Conversions still apply, unlike in the deleted NoArgCreator.
• Lazy pending lists. They're allocated only when a partial constructor actually has values to hold until it runs.
• Shared empty parameter array for no-arg creators.
• arg[0-9]+ is a precompiled Pattern, instead of String.matches compiling it once per constructor parameter.
• NoArgCreator deleted (@MorphiaInternal), along with the unreachable factory branch and its LOG.info("using old creator approach").

Tests

• ConstructorCreatorTest.noArgConstructorSetsValuesAsTheyArrive checks that a no-arg creator is eager-safe, that values land on the one instance, and that a value needing conversion (42.0 into an int) still goes through Conversions. It pins behavior. It would also pass on master, because the old path replayed the same values, so the speedup itself is shown by the benchmark below.
• The harness is unchanged and passes. The core and Kotlin suites pass with -Dmorphia.mapper=critter and with reflection: 1486 + 5 tests, 0 failures.

JMH: CodecBenchmark, master (67dc8c46533, after 3.0 + 3.1) vs this branch

Same machine as the baseline (i9-11900H, Temurin 17), -prof gc, the benchmark's own settings (3 forks × 5 × 1 s). Before and after ran back to back with nothing else running. Critter is the AOT variant.

reflection

benchmark	model	before ns/op	after ns/op	Δ time	before B/op	after B/op	Δ alloc
decode	ARRAYS	68,627 ± 1,327	68,718 ± 1,396	+0.1%	77,770	77,706	-0.1%
decode	COLLECTIONS	9,441 ± 211	9,362 ± 170	-0.8%	19,376	18,840	-2.8%
decode	CONCRETE_COLLECTIONS	3,092 ± 34	3,048 ± 59	-1.4%	6,664	6,608	-0.8%
decode	LAZY_REFERENCE	3,355 ± 113	3,465 ± 42	+3.3%	8,176	8,131	-0.6%
decode	LIFECYCLE	1,188 ± 29	1,134 ± 23	-4.5%	3,035	2,960	-2.5%
decode	NESTED	3,261 ± 85	3,128 ± 35	-4.1%	7,024	6,776	-3.5%
decode	POLYMORPHIC	6,076 ± 96	5,879 ± 117	-3.2%	12,672	12,024	-5.1%
decode	SIMPLE	1,545 ± 381	1,274 ± 30	-17.5%	2,648	2,584	-2.4%
decodeInCursor	ARRAYS	69,156 ± 1,411	69,003 ± 1,204	-0.2%	77,458	77,394	-0.1%
decodeInCursor	COLLECTIONS	9,703 ± 222	9,372 ± 139	-3.4%	19,056	18,520	-2.8%
decodeInCursor	CONCRETE_COLLECTIONS	3,102 ± 91	3,115 ± 65	+0.4%	6,368	6,283	-1.3%
decodeInCursor	LAZY_REFERENCE	3,639 ± 140	3,564 ± 72	-2.0%	7,907	7,843	-0.8%
decodeInCursor	LIFECYCLE	1,218 ± 26	1,197 ± 16	-1.7%	2,712	2,648	-2.4%
decodeInCursor	NESTED	3,322 ± 28	3,277 ± 56	-1.3%	6,720	6,464	-3.8%
decodeInCursor	POLYMORPHIC	6,145 ± 98	6,060 ± 130	-1.4%	12,344	11,704	-5.2%
decodeInCursor	SIMPLE	1,321 ± 51	1,305 ± 26	-1.3%	2,336	2,272	-2.7%
encode	ARRAYS	96,921 ± 1,860	99,863 ± 2,915	+3.0%	137,187	137,187	-0.0%
encode	COLLECTIONS	7,757 ± 128	7,747 ± 149	-0.1%	7,336	7,336	+0.0%
encode	CONCRETE_COLLECTIONS	2,053 ± 25	2,177 ± 129	+6.0%	3,104	3,147	+1.4%
encode	LAZY_REFERENCE	1,809 ± 28	1,909 ± 81	+5.5%	4,072	4,072	+0.0%
encode	LIFECYCLE	837 ± 24	839 ± 16	+0.3%	3,043	3,021	-0.7%
encode	NESTED	2,374 ± 113	2,311 ± 42	-2.7%	1,856	1,856	-0.0%
encode	POLYMORPHIC	4,443 ± 233	4,197 ± 80	-5.5%	6,024	6,024	-0.0%
encode	SIMPLE	792 ± 18	784 ± 16	-1.1%	1,216	1,216	-0.0%

critter

benchmark	model	before ns/op	after ns/op	Δ time	before B/op	after B/op	Δ alloc
decode	ARRAYS	72,482 ± 1,907	73,698 ± 4,112	+1.7%	77,770	77,706	-0.1%
decode	COLLECTIONS	10,260 ± 349	10,747 ± 142	+4.7%	19,376	18,840	-2.8%
decode	CONCRETE_COLLECTIONS	3,313 ± 113	3,242 ± 41	-2.1%	6,664	6,616	-0.7%
decode	LAZY_REFERENCE	3,715 ± 98	3,396 ± 70	-8.6%	8,131	8,120	-0.1%
decode	LIFECYCLE	1,250 ± 28	1,209 ± 25	-3.3%	3,024	2,939	-2.8%
decode	NESTED	3,762 ± 134	3,829 ± 80	+1.8%	7,024	6,784	-3.4%
decode	POLYMORPHIC	6,817 ± 218	6,855 ± 88	+0.6%	12,648	12,016	-5.0%
decode	SIMPLE	1,342 ± 41	1,441 ± 60	+7.4%	2,648	2,584	-2.4%
decodeInCursor	ARRAYS	74,118 ± 2,296	72,694 ± 1,617	-1.9%	77,458	77,394	-0.1%
decodeInCursor	COLLECTIONS	10,521 ± 290	10,200 ± 156	-3.1%	19,056	18,520	-2.8%
decodeInCursor	CONCRETE_COLLECTIONS	3,438 ± 91	3,405 ± 95	-1.0%	6,347	6,312	-0.5%
decodeInCursor	LAZY_REFERENCE	3,750 ± 106	3,670 ± 109	-2.1%	7,848	7,816	-0.4%
decodeInCursor	LIFECYCLE	1,299 ± 22	1,297 ± 14	-0.2%	2,712	2,659	-2.0%
decodeInCursor	NESTED	3,947 ± 150	3,750 ± 80	-5.0%	6,720	6,464	-3.8%
decodeInCursor	POLYMORPHIC	7,123 ± 167	6,784 ± 101	-4.8%	12,344	11,704	-5.2%
decodeInCursor	SIMPLE	1,448 ± 43	1,454 ± 35	+0.4%	2,336	2,272	-2.7%
encode	ARRAYS	107,407 ± 1,909	106,766 ± 2,456	-0.6%	145,763	145,785	+0.0%
encode	COLLECTIONS	8,437 ± 120	8,245 ± 96	-2.3%	7,336	7,336	-0.0%
encode	CONCRETE_COLLECTIONS	2,252 ± 44	2,172 ± 35	-3.6%	3,125	3,104	-0.7%
encode	LAZY_REFERENCE	1,940 ± 55	1,811 ± 48	-6.7%	4,072	4,072	-0.0%
encode	LIFECYCLE	886 ± 34	851 ± 13	-3.9%	3,085	3,043	-1.4%
encode	NESTED	2,613 ± 88	2,475 ± 40	-5.3%	1,856	1,856	-0.0%
encode	POLYMORPHIC	4,882 ± 241	4,613 ± 66	-5.5%	5,640	5,640	-0.0%
encode	SIMPLE	856 ± 154	796 ± 5	-7.0%	1,216	1,216	-0.0%

Reading it:

• Decode allocation drops wherever a no-arg entity is built: −64 B per entity on SIMPLE (−2.4%), −248/256 B on NESTED (−3.5%), −536 B on COLLECTIONS (−2.8%) and −632/640 B on POLYMORPHIC (−5%). That's the same on both mappers and in decodeInCursor. Models whose entities use full constructors barely move, as expected.
• This is less than the review estimated. It predicted ~216 B per no-arg entity, from a microbenchmark of the creator alone. In a full decode, the JIT probably already eliminated part of the old buffering, so the measured saving is about 64 B per entity. I haven't confirmed that cause.
• Decode time is flat to a few percent faster. Two critter decode cells looked slower (SIMPLE +7.4%, COLLECTIONS +4.7%). I reran both twice, alternating master and this branch: SIMPLE 1,348 vs 1,373 and 1,374 vs 1,392 ns/op, within the error bars; COLLECTIONS 10,247 vs 10,205 and 10,467 vs 10,330 ns/op, slightly faster. So both were noise.
• Encode doesn't touch the creator. Its swings in either direction, of up to ±7%, are run-to-run noise.

Decisions to review

1. The no-arg instance is now constructed when the first property arrives, not after the whole document has been read. Field values end up the same, because the constructor runs before any setValue in both versions. If decoding fails partway through a document, an instance has already been constructed and is then discarded. EntityDecoder's eager path for reference cycles already did this.
2. The fast path lives in ConstructorCreator, keyed on the plan having no parameters. I didn't add a separate creator class. That keeps one creator type and one factory path, which matters because 3.17 has to hold values that arrive before _t for exactly these creators.
3. The gain is smaller than the review's estimate. The numbers above are what was measured.

[Copilot]

🟢 Approval recommended

The focused optimization has no identified blocking issues in constructor selection, value conversion, or instance isolation.

No open issues 🎉

What changed in this PR

Optimizes Morphia’s no-argument entity construction to reduce decoding allocations while preserving value conversion.

Changes:

• Assigns no-arg entity values immediately and allocates pending buffers lazily.
• Reuses an empty parameter array, precompiles parameter-name matching, and removes the unused creator.
• Adds assignment/conversion coverage and marks plan item 3.2 complete.

File	Description
design/​mapping-game-plan.md	Marks the optimization complete.
core/​src/​test/​java/​dev/​morphia/​test/​mapping/​internal/​ConstructorCreatorTest.java	Tests instance reuse and value conversion.
core/​src/​main/​java/​dev/​morphia/​mapping/​NoArgCreator.java	Removes the unused creator.
core/​src/​main/​java/​dev/​morphia/​mapping/​internal/​ConstructorCreator.java	Adds immediate assignment and reduces allocations.
core/​src/​main/​java/​dev/​morphia/​mapping/​InstanceCreatorFactoryImpl.java	Removes the unreachable fallback and logging.

_{Review effort: Balanced}