[evanchooly]

Plan item 3.7 (review B8).

The problem

Every decoded property goes through Conversions.convert, even when the value already has the right type. The early return, fromType.equals(target), misses the common cases: an Integer for an int field, an ArrayList for a List, a HashMap for a Map. Each miss then ran computeIfAbsent(fromType, …), which also inserted an empty ConcurrentHashMap for every new value class. Then .get(target) came back empty and the value was returned unchanged.

When a stored value doesn't match the codec (schema drift), the decoder's fallback path converted the value, and the creator converted it again even though it already fit.

Change

• Fast path. PropertyModel.setValue and ConstructorCreator.set now set the value directly when it's already an instance of the target type, with primitives treated as their wrappers. Only values that need converting reach Conversions.convert.
• get instead of computeIfAbsent in Conversions.convert. An unregistered source type is no longer added to the map.
• One conversion on the fallback path. EntityDecoder still converts the fallback value to the property's type, and the creator's fast path now skips the second conversion because the value fits. (The first version of this PR dropped the decoder's conversion instead; see the review thread for why that was wrong.)
• Conversions.boxedType(Class), a small @MorphiaInternal helper (MethodType.methodType(t).wrap()), shared by both call sites.

Tests

• TestConversions: boxing; registered conversions still apply ("42" → int, 42.0 → long, "c" → Character); values without a conversion pass through; null → 0 for primitives.
• The harness fixtures already cover drifted fields (the fallback path), constructor-built entities and records under both mappers, and are unchanged.
• TestDriftedValues: a drifted value reaching a constructor parameter whose type differs from the property's (Integer → String property → Character parameter) still converts through the property type.
• The core and Kotlin suites pass with -Dmorphia.mapper=critter and with reflection: 1498 + 5 tests, 0 failures.

JMH

Expect a small decode-only gain. The review estimated ~10–20 ns per primitive- or interface-typed property, with SIMPLE having 4 such properties.

decode, master (f41533b8735) and this branch alternating, three rounds each

variant	benchmark	model	master ns/op (3 runs)	3.7 ns/op (3 runs)	Δ median
reflection	CodecBenchmark.decode	COLLECTIONS	8,087, 8,113, 8,326	8,057, 8,167, 8,287	+0.7%
reflection	CodecBenchmark.decode	NESTED	2,710, 2,742, 2,785	2,618, 2,645, 2,664	-3.5%
reflection	CodecBenchmark.decode	POLYMORPHIC	4,980, 5,017, 5,091	4,842, 4,893, 4,928	-2.5%
reflection	CodecBenchmark.decode	SIMPLE	1,092, 1,126, 1,171	1,010, 1,030, 1,042	-8.5%
reflection	MixedCodecBenchmark.decode	-	26,760, 26,957, 27,410	26,552, 26,594, 26,689	-1.3%
critter	CodecBenchmark.decode	COLLECTIONS	8,586, 8,606, 8,654	8,486, 8,649, 8,699	+0.5%
critter	CodecBenchmark.decode	NESTED	3,063, 3,101, 3,135	3,002, 3,011, 3,028	-2.9%
critter	CodecBenchmark.decode	POLYMORPHIC	5,396, 5,430, 5,586	5,260, 5,416, 5,439	-0.3%
critter	CodecBenchmark.decode	SIMPLE	1,126, 1,167, 1,197	1,045, 1,058, 1,072	-9.3%
critter	MixedCodecBenchmark.decode	-	29,106, 29,181, 29,247	28,039, 28,302, 28,929	-3.0%

Reading it:

• SIMPLE decodes 8.5–9.3% faster under both mappers, and every branch run beats every master run. ~95 ns over 4 primitive properties is ~24 ns each, at the top of the review's estimate.
• NESTED is ~3% faster. POLYMORPHIC is faster under reflection; under critter the runs overlap.
• COLLECTIONS is flat, within noise. Its ~8 µs goes almost entirely to decoding the collections' elements, so saving ~20 ns on each of its few properties doesn't show.
• Allocation is unchanged. The empty maps computeIfAbsent inserted were per value class, not per value, so after warm-up they never showed up in B/op.

Full CodecBenchmark + MixedCodecBenchmark, one run each (master reflection run was disturbed)

The first master reflection run was noisy: decode error bars reached ±2,264 ns, which inflates the reflection "gains" to 20–35%. That's why the decode rerun above exists. The critter run and all encode rows are clean; encode code is untouched.

reflection

class	benchmark	model	before ns/op	after ns/op	Δ time	before B/op	after B/op	Δ alloc
CodecBenchmark	decode	ARRAYS	71,752 ± 1,984	68,881 ± 1,549	-4.0%	74,858	74,858	-0.0%
CodecBenchmark	decode	COLLECTIONS	8,904 ± 259	8,940 ± 448	+0.4%	16,680	16,704	+0.1%
CodecBenchmark	decode	CONCRETE_COLLECTIONS	4,474 ± 621	2,913 ± 52	-34.9%	6,144	6,107	-0.6%
CodecBenchmark	decode	LAZY_REFERENCE	4,888 ± 1,195	3,280 ± 67	-32.9%	7,288	7,309	+0.3%
CodecBenchmark	decode	LIFECYCLE	1,318 ± 179	1,066 ± 29	-19.1%	2,840	2,840	-0.0%
CodecBenchmark	decode	NESTED	4,036 ± 460	2,742 ± 85	-32.1%	5,640	5,640	-0.0%
CodecBenchmark	decode	POLYMORPHIC	7,869 ± 2,264	5,273 ± 190	-33.0%	9,520	9,536	+0.2%
CodecBenchmark	decode	SIMPLE	1,492 ± 278	1,124 ± 54	-24.7%	2,136	2,136	-0.0%
CodecBenchmark	decodeInCursor	ARRAYS	71,158 ± 1,833	71,091 ± 1,925	-0.1%	74,546	74,546	+0.0%
CodecBenchmark	decodeInCursor	COLLECTIONS	10,100 ± 701	8,912 ± 332	-11.8%	16,376	16,392	+0.1%
CodecBenchmark	decodeInCursor	CONCRETE_COLLECTIONS	3,320 ± 306	3,115 ± 85	-6.2%	5,819	5,840	+0.4%
CodecBenchmark	decodeInCursor	LAZY_REFERENCE	3,598 ± 81	3,445 ± 148	-4.3%	7,016	6,981	-0.5%
CodecBenchmark	decodeInCursor	LIFECYCLE	1,204 ± 31	1,151 ± 67	-4.4%	2,557	2,557	+0.0%
CodecBenchmark	decodeInCursor	NESTED	3,738 ± 943	2,886 ± 120	-22.8%	5,344	5,352	+0.1%
CodecBenchmark	decodeInCursor	POLYMORPHIC	5,833 ± 118	5,225 ± 221	-10.4%	9,216	9,224	+0.1%
CodecBenchmark	decodeInCursor	SIMPLE	1,367 ± 77	1,109 ± 44	-18.9%	1,840	1,824	-0.9%
CodecBenchmark	encode	ARRAYS	103,436 ± 5,576	98,424 ± 1,948	-4.8%	135,587	135,587	-0.0%
CodecBenchmark	encode	COLLECTIONS	7,966 ± 120	7,539 ± 248	-5.4%	7,336	7,336	-0.0%
CodecBenchmark	encode	CONCRETE_COLLECTIONS	2,121 ± 54	2,057 ± 33	-3.0%	3,104	3,104	-0.0%
CodecBenchmark	encode	LAZY_REFERENCE	1,853 ± 67	1,736 ± 25	-6.3%	3,843	3,843	-0.0%
CodecBenchmark	encode	LIFECYCLE	847 ± 25	813 ± 13	-4.0%	3,021	3,021	-0.0%
CodecBenchmark	encode	NESTED	2,385 ± 62	2,276 ± 58	-4.5%	1,856	1,856	-0.0%
CodecBenchmark	encode	POLYMORPHIC	4,227 ± 78	4,065 ± 73	-3.8%	4,200	4,200	-0.0%
CodecBenchmark	encode	SIMPLE	805 ± 17	794 ± 20	-1.4%	1,216	1,216	-0.0%
MixedCodecBenchmark	decode	-	32,748 ± 10,096	27,063 ± 768	-17.4%	50,261	50,429	+0.3%
MixedCodecBenchmark	encode	-	23,095 ± 380	21,914 ± 305	-5.1%	24,657	24,657	+0.0%

critter

class	benchmark	model	before ns/op	after ns/op	Δ time	before B/op	after B/op	Δ alloc
CodecBenchmark	decode	ARRAYS	67,317 ± 2,236	68,414 ± 1,112	+1.6%	74,858	74,858	-0.0%
CodecBenchmark	decode	COLLECTIONS	8,579 ± 219	8,688 ± 123	+1.3%	16,704	16,712	+0.0%
CodecBenchmark	decode	CONCRETE_COLLECTIONS	2,922 ± 43	2,979 ± 59	+2.0%	6,128	6,128	+0.0%
CodecBenchmark	decode	LAZY_REFERENCE	3,172 ± 95	3,188 ± 119	+0.5%	7,245	7,245	-0.0%
CodecBenchmark	decode	LIFECYCLE	1,080 ± 29	1,061 ± 61	-1.8%	2,861	2,813	-1.7%
CodecBenchmark	decode	NESTED	3,048 ± 152	3,010 ± 147	-1.2%	5,648	5,648	-0.0%
CodecBenchmark	decode	POLYMORPHIC	5,524 ± 106	5,297 ± 121	-4.1%	9,528	9,552	+0.3%
CodecBenchmark	decode	SIMPLE	1,150 ± 59	1,030 ± 23	-10.4%	2,136	2,136	-0.0%
CodecBenchmark	decodeInCursor	ARRAYS	67,781 ± 1,003	66,964 ± 996	-1.2%	74,546	74,546	-0.0%
CodecBenchmark	decodeInCursor	COLLECTIONS	8,823 ± 165	8,740 ± 212	-0.9%	16,376	16,392	+0.1%
CodecBenchmark	decodeInCursor	CONCRETE_COLLECTIONS	2,967 ± 94	3,002 ± 121	+1.2%	5,840	5,840	-0.0%
CodecBenchmark	decodeInCursor	LAZY_REFERENCE	3,216 ± 145	3,307 ± 54	+2.8%	6,947	6,965	+0.3%
CodecBenchmark	decodeInCursor	LIFECYCLE	1,156 ± 14	1,102 ± 24	-4.7%	2,528	2,523	-0.2%
CodecBenchmark	decodeInCursor	NESTED	3,034 ± 85	3,150 ± 72	+3.8%	5,344	5,352	+0.1%
CodecBenchmark	decodeInCursor	POLYMORPHIC	5,525 ± 121	5,426 ± 117	-1.8%	9,216	9,240	+0.3%
CodecBenchmark	decodeInCursor	SIMPLE	1,082 ± 21	1,104 ± 38	+2.0%	1,824	1,824	+0.0%
CodecBenchmark	encode	ARRAYS	98,288 ± 1,998	98,250 ± 2,806	-0.0%	144,163	144,163	-0.0%
CodecBenchmark	encode	COLLECTIONS	7,594 ± 140	7,520 ± 94	-1.0%	7,336	7,336	+0.0%
CodecBenchmark	encode	CONCRETE_COLLECTIONS	2,013 ± 49	2,020 ± 31	+0.3%	3,104	3,104	+0.0%
CodecBenchmark	encode	LAZY_REFERENCE	1,750 ± 109	1,671 ± 36	-4.5%	3,885	3,832	-1.4%
CodecBenchmark	encode	LIFECYCLE	799 ± 21	829 ± 16	+3.7%	3,043	3,064	+0.7%
CodecBenchmark	encode	NESTED	2,307 ± 39	2,268 ± 34	-1.7%	1,856	1,856	-0.0%
CodecBenchmark	encode	POLYMORPHIC	4,090 ± 68	4,116 ± 199	+0.6%	4,200	4,200	-0.0%
CodecBenchmark	encode	SIMPLE	731 ± 14	726 ± 3	-0.7%	1,216	1,216	-0.0%
MixedCodecBenchmark	decode	-	29,606 ± 744	28,342 ± 722	-4.3%	50,185	50,429	+0.5%
MixedCodecBenchmark	encode	-	22,521 ± 944	22,649 ± 609	+0.6%	24,657	24,657	+0.0%

Decisions to review

1. The boxed type lives on PropertyModel and ConstructorCreator.Plan, not on CodecPlan slots as the plan item said. The conversion happens in setValue and the creator, which never see the codec plan. The boxed type depends only on the property's declared type, not on the datastore, so it fits the shared model (it's computed lazily on first use and isn't copied).
2. A registered conversion from a type to one of its own supertypes is now skipped. For example, an ArrayList → List conversion never runs because an ArrayList already fits a List property. No default conversion is like that; the review flagged the same risk. A user who needs to rewrite a value of the right type should use a codec.
3. Found along the way, not fixed here: refresh() of an entity with a drifted field fails on master too. refresh decodes through DocumentReader, whose readString throws ClassCastException on a stored integer instead of BsonInvalidOperationException, so the fallback is never reached. It belongs in its own PR.

[Copilot]

🟡 Changes recommended

Removing fallback property-type conversion breaks schema-drift refreshes and constructor mappings that require an intermediate conversion.

1 open issue

• Restore fallback value conversion through the property type · New

What changed in this PR

This PR implements plan item 3.7 to reduce conversion overhead when Morphia decodes values that already fit their target types.

Changes:

• Adds boxed-type fast paths for property assignment and constructor arguments.
• Avoids caching empty conversion maps and removes decoder-side fallback conversion.
• Adds conversion tests and marks the plan item complete.

File	Description
design/​mapping-game-plan.md	Marks item 3.7 complete.
core/​src/​test/​java/​dev/​morphia/​mapping/​codec/​TestConversions.java	Tests boxing, registered conversions, pass-through values, and null handling.
core/​src/​main/​java/​dev/​morphia/​mapping/​internal/​ConstructorCreator.java	Caches boxed parameter types and skips unnecessary conversions.
core/​src/​main/​java/​dev/​morphia/​mapping/​codec/​pojo/​PropertyModel.java	Adds a boxed-type assignment fast path.
core/​src/​main/​java/​dev/​morphia/​mapping/​codec/​pojo/​EntityDecoder.java	Delegates fallback conversion to instance creators.
core/​src/​main/​java/​dev/​morphia/​mapping/​codec/​Conversions.java	Adds a boxing helper and avoids empty-map insertion.

_{Review effort: Balanced}

[Copilot]

🟢 Approval recommended

The localized optimizations preserve required fallback conversions and null handling, with regression coverage and no unresolved blocking issues.

No open issues 🎉

1 resolved since last review

• Restore fallback value conversion through the property type

_{Review effort: Balanced}