30151295ad
The inline uLt() using XOR trick in commonMain regressed JVM by ~30% because HotSpot's Long.compareUnsigned is a JIT intrinsic (single unsigned CMP + SETB), while the XOR trick generates 2 extra XOR insns that HotSpot doesn't optimize away. Convert uLt to expect/actual: - JVM: Long.compareUnsigned (HotSpot intrinsic) - Android: XOR trick (avoids ULong.constructor-impl NOOP calls) - Native: XOR trick (no JVM intrinsics available) JVM verify: 2.1x → 1.5x (restored, slightly better than 1.6x baseline) Android: unchanged (still uses XOR trick) https://claude.ai/code/session_01EMY5RnXb9rnsyU2KbXrSaY