optimize lookup to compete with abseil flat_hash_map#2

Draft

joshuaisaact wants to merge 187 commits intomainfrom

autoresearch/beat-abseil

Commits on Mar 23, 2026

Commits on Mar 24, 2026

add program-v3: paper-faithful multi-tier lookup research
joshuaisaact
committed
noinline getSlowPath for tier 1+ search (keep get() I-cache footprint small)
joshuaisaact
committed
limit getSlowPath to tier 1 only (tiers 2+ are empty at 99% load)
joshuaisaact
committed
Revert "limit getSlowPath to tier 1 only (tiers 2+ are empty at 99% load)"
joshuaisaact
committed
Reapply "limit getSlowPath to tier 1 only (tiers 2+ are empty at 99% load)"
joshuaisaact
committed
restore tier-0-only get() as clean base for insert-side experiments
joshuaisaact
committed
MAX_PROBES 7 -> 8 (more elements fit in tier 0)
joshuaisaact
committed
batch threshold 0.05 (95% fill) with MAX_PROBES=8
joshuaisaact
committed
Revert "batch threshold 0.05 (95% fill) with MAX_PROBES=8"
joshuaisaact
committed
Reapply "batch threshold 0.05 (95% fill) with MAX_PROBES=8"
joshuaisaact
committed
restore clean v3 baseline (MAX_PROBES=7, threshold=0.12)
joshuaisaact
committed
paper-faithful get(): noinline cold-hinted getOtherTiers for tier 1+
joshuaisaact
committed
use opaque function pointer for tier 1+ overflow (prevent LLVM codegen cascade)
joshuaisaact
committed
limit overflow to tier 1 only (tiers 2+ empty at 99% load)
joshuaisaact
committed
log v3 experiments
joshuaisaact
committed
early termination in overflow function (safe with 100% find rate)
joshuaisaact
committed
early termination in tier-0 loop (jump to overflow on empty slot)
joshuaisaact
committed
Revert "early termination in tier-0 loop (jump to overflow on empty slot)"
joshuaisaact
committed
log v3 early termination experiments
joshuaisaact
committed
v3 insights: elastic hash 40-65% faster than abseil at normal loads
joshuaisaact
committed
add verification program: are these results real?
joshuaisaact
committed
verification checks 1-3: capacity, shuffled access, size independence
joshuaisaact
committed
complete verification: realistic workloads, hash cost, shuffled access
joshuaisaact
committed
update README and PR with verified benchmark results
joshuaisaact
committed
verify: gcc vs clang abseil - no unfair compiler advantage
joshuaisaact
committed