Performance optimizations

[marius-bughiu]

Priority: High
Phase: 5 - Performance & Quality (Q4 2026)

Description

Profile and optimize hot paths in existing collections.

Tasks

• Profile with BenchmarkDotNet and profilers
• Reduce allocations in critical paths
• Improve cache locality
• Branch prediction optimization
• Inline critical methods with AggressiveInlining
• Reduce bounds checking where safe
• Optimize resize operations

Related

See ROADMAP.md for context

[marius-bughiu]

Closing as substantially complete — every discrete code-level subtask is now implemented across all 8 collections (IntDictionary, LongDictionary, CelerityDictionary, CelerityMultiMap, IntSet, LongSet, CeleritySet, SmallDictionary):

• Profile with BenchmarkDotNet and profilers — core --ci suite plus the extended on-demand suite (Comprehensive benchmark suite expansion #26); distribution / adversarial benches (Expanded benchmark suite: uniform / clustered / adversarial key distributions #60).
• Reduce allocations in critical paths — hot paths are allocation-free; struct enumerators / Keys / Values views avoid boxing; tracked by [MemoryDiagnoser].
• Improve cache locality — struct-of-arrays flat _keys / _values, power-of-two masking for index computation.
• Inline critical methods with AggressiveInlining — applied to the probe / insert / delete helpers on every collection.
• Reduce bounds checking where safe — probe loops walk Unsafe.Add against MemoryMarshal.GetArrayDataReference, eliding per-iteration bounds checks (the bound is structural via index & mask).
• Optimize resize operations — direct-reinsert resize that bypasses the public setter, plus bulk-constructor sizing for load-factor headroom so a known-count bulk fill never resizes (perf(collections): size bulk constructors for load factor so they never resize (#27) #168).
• Branch prediction optimization — no discrete remaining work; the hot loops are already tight, single-comparison probes.

The only genuinely open-ended item is perpetual micro-tuning, which an umbrella issue never truly retires. If a concrete profiling finding surfaces later, it'll get its own narrowly-scoped issue rather than reopening this catch-all.

Note: the Robin Hood probing experiment (#63) shipped separately as a new opt-in collection type (RobinHoodDictionary) — it complements this work but does not optimize the existing collections, so it was tracked under its own issue.