[JasperGeurtz]

Summary

This PR reduces per-frame CPU overhead, shared-memory ByteBuffer reads, and heap/GC allocation pressure

1. Lazy Per-Frame UnitListCache (Game, Player):

   • Replaces eager stream filtering on Game.getMinerals(), Game.getGeysers(), Game.getNeutralUnits(), Player.getUnits(), and internal selfPylons (Game.hasPower) with lazy per-frame UnitListCache instances that compute at most once per frame on demand and return the cached unmodifiable list on subsequent calls.

2. Lazy Write-Only Latency Compensation (UnitSelf, PlayerSelf):

   • Unit.self (37 Cache objects per unit) and Player.self (520 Cache objects per player) are now allocated lazily only when CommandTemp mutates latency-compensated state, rather than on the first getter read or unconditionally on construction.
   • Uncommanded, neutral, and enemy units/players no longer pay the latency-compensation memory and initialization cost.

3. Allocation-Free Unit.getDistance & Stationary Unit.updatePosition (Point, Unit):

   • Unit.getDistance(Position) and Unit.getDistance(Unit) now read getType() and position once per unit (instead of 4–8 times via getLeft()/getTop()/getRight()/getBottom()) and call package-private Point.getApproxDistance(0, 0, xDist, yDist) directly instead of allocating a temporary Position.
   • Unit.updatePosition(int) only allocates a new Position when ((position.x ^ newX) | (position.y ^ newY)) != 0, avoiding heap allocations and Old-Gen → Young-Gen card-table dirtying for stationary units (minerals, geysers, buildings, sieged/holding units).

4. Single-Pass getUnitsInRectangle & canBuildHere (Game, Unit):

   • Replaces stream/lambda allocation and 4× getType() calls per unit in Game.getUnitsInRectangle with a direct loop using Unit.overlapsRectangle.
   • Replaces intermediate Position/TilePosition allocations and full getUnitsInRectangle list materialization in Game.canBuildHere with primitive coordinate math and an early-exit ground unit overlap check.

5. Cached Immutable Return Wrappers (UnitType, TechType, UpgradeType):

   • Pre-builds and caches tileSize(), whatBuilds(), requiredUnits(), abilities(), upgrades(), buildsWhat(), researchesWhat(), upgradesWhat(), maxUnitWidth(), and maxUnitHeight() on UnitType / UnitTypeContainer, as well as whatUses() on TechType and UpgradeType.
   • Replaces internal getTrainingQueue().size() and getTrainingQueue().get(i) calls with allocation-free getTrainingQueueCount() and getTrainingQueueAt(i).

6. BitSet for Game.visibleUnits (Game):

   • Replaces HashSet<Integer> and visibleUnits.stream() in Game.onFrame with a BitSet and nextSetBit iteration into a pre-sized ArrayList.

---

JMH Benchmark Results

Measured with bwapi.PerformanceBenchmark on JDK 1.8.0_452 ((4)Andromeda.scx, 356 visible units: 236 stationary resources/buildings + 120 mobile combat units):

Benchmark (bwapi.PerformanceBenchmark)	Workload	Baseline (master)	Optimized (perf-optimizations)	Speedup	GC Impact
unitGetDistanceToPosition	2,400 calls	27.58 ± 3.61 µs	15.52 ± 2.78 µs	-43.7% (1.78×)	0 allocs
unitGetDistanceToUnit	3,600 calls	103.29 ± 26.84 µs	64.01 ± 4.98 µs	-38.0% (1.61×)	0 allocs
canBuildHere	96 checks (12 bases × 8 types)	336.88 ± 44.05 µs	232.74 ± 43.41 µs	-30.9% (1.45×)	44 → 5 GCs (-89%)
getUnitsInRectangle	32 spatial queries	184.84 ± 45.07 µs	131.03 ± 3.93 µs	-29.1% (1.41×)	38 → 28 GCs (-26%)
unitUpdatePositionPerFrame	356 units (236 static, 120 moving)	3.21 ± 0.45 µs	2.65 ± 0.22 µs	-17.4% (1.21×)	140 → 54 GCs (-61%)
visibleUnitsWorstCaseOnFrame	700 visible units + 40 fog toggles + onFrame	16.09 ± 1.75 µs	13.30 ± 0.48 µs	-17.4% (1.21×)	41 → 12 GCs (-71%)