feat: add Fe4 struct-based field elements to eliminate array bounds checks
Prototype Fe4/Wide8 classes that replace LongArray(4)/LongArray(8) with
named @JvmField Long fields. Every LongArray access compiles to
laload/lastore (3 bytecode insns + implicit bounds check); Fe4 field
access compiles to getfield/putfield (2 insns, zero checks).
Bytecode impact (core arithmetic files):
LongArray: 464 bounds-checked array ops (U256: 150, FieldP: 119, Fused: 195)
Fe4: 0 bounds-checked ops, 203 direct field accesses
JVM benchmark results (HotSpot C2, best of 3 rounds, alternating order):
FieldP.mul: 40 ns → 41 ns (~equal, C2 fully optimizes hot mul)
FieldP.sqr: 48 ns → 30 ns (+60% faster)
FieldP.add: 9 ns → 6 ns (+34% faster)
FieldP.sub: 10 ns → 6 ns (+55% faster)
U256.sqrWide: 37 ns → 17 ns (+114% faster)
The gains are largest for sqr/add/sub which have high ratios of
array read-modify-write patterns. Expected to be even larger on
Android ART (limited inlining) and Kotlin/Native LLVM (AOT, no
profile-guided bounds check elimination).
https://claude.ai/code/session_01Sxi6Gpxbstuj3Y8TBY7XrU
Claude
ea8b693785