| name | theory |
| description | Theoretical background for veb-in-c's design choices — memeff root, bitwise leaves, the O(u)-bits derivation from issue |
Theory — memory model
Two optimizations work together to keep the struct footprint sane on small-to-mid universes. Neither is a textbook vEB — both come from #2 and #4.
Scope: vebtree_init is publicly capped at universe_bits ≤ 32 (IPv4 address space). The recursion and bit-math below still reason in terms of arbitrary k, but no realistic k > 32 use case is supported.
1. Memeff root
Where: is_memeff_root branch at _vebtree_init, vebtrees.h:368.
Textbook vEB splits universe u = 2^k into √u clusters of √u each — so at the root, lower_bits = k/2 and there are 2^(k/2) cluster slots. That's catastrophic at large k: u=32 gives 2^16 slots whose cluster structs are each themselves non-trivial vEB nodes, recurring a handful of levels. Total bits: O(u log log u).
Trick from #2: at the root only, pin lower_bits to log u (in our code: the constant VEBTREE_LEAF_BITS = 6). Then:
- Each local holds ≤
2^6 = 64 keys — representable as a single bitboard.
- Number of locals:
2^(k - 6) — one flat layer, no recursion inside clusters.
- Global summary is a vEB over universe
2^(k - 6), recursing normally (√u split) from there.
Storage, per #2:
locals: 2^(k-6) structs × 48 B each
global: O(u / log u × log(u / log u)) bits = O(u) bits
The O(u) bits bound holds when lower_bits ≈ log u. We freeze it at 6 for ALU reasons (see §2), so for u ≥ 32 the locals array still explodes — that's the cliff #3's lazy allocation and #19's sparse allocation exist to solve. The root trick shifts the cliff from "impossible at u > ~16" to "impossible at u > ~30", which is the difference between "toy" and "usable on moderate universes."
2. Bitwise leaves at u ≤ 6
Where: vebtrees.h:238–280.
Issue #2's original suggestion was to use binary search trees as the local structure (where each local manages O(log u) keys). #4 replaces that with a 64-bit bitboard plus bit-scan intrinsics (__builtin_ctzll / __builtin_clzll / MSVC equivalents). All leaf ops — contains, insert, delete, min, max, successor, predecessor — become O(1) ALU ops, beating a BST's O(log log u) on every dimension: speed, branch-predictability, allocation count (zero per leaf).
3. Why they interlock
Since the memeff root already pins local-size to 6 bits = 64 keys, every locals slot fits in one bitboard. The two tricks are mutually reinforcing: memeff root caps cluster size at the ALU word width exactly because bitwise leaves efficiently fill that width. Change either and you need to rethink the other. #17 proposes widening leaves to 256/512 bits via SIMD — would let memeff root expand to lower_bits = 8 or 9, raising the cliff by 2–3 bits of universe without touching sparsity.
Implications for #3 (lazy) and #19 (sparse)
tree->locals[i] is always a leaf at the memeff root. That means sparse cluster storage (#19's job) only has to map cluster_idx → one 48 B leaf, not a recursively-allocated subtree. No deep recursion inside a cluster ever happens at the root layer. This is a huge simplification for the sparse-allocation design.- At non-memeff-root layers (the recursion inside
tree->global), lower_bits = upper_bits/2, so clusters there are internal nodes. Any sparse design has to handle both — but the leaf-cluster case is the hot path (it's where every key ultimately lives).
- Lazy (#3) alone doesn't break the cliff because
_ensure_subtrees still allocates the full 2^(k-6) locals array on the first second-key insert. Lazy helps with "never inserted" and "one-key" cases only. Full cliff removal needs sparse locals (#19).
