absl::flat_hash_map

Google's reference swiss-table implementation, introduced in 2017 and the design that started everything. Still the most widely deployed high-performance hash map in production, powering Google's infrastructure. As of 2025 it is narrowly beaten on most operations by boost-unordered-flat-map, but remains the right choice within Google's ecosystem and for large maps with integer keys — its original sweet spot.

The original Swiss Table

Each 64-bit hash splits into H1 (slot/group selector) and H2 (a 7-bit fingerprint). H2 is broadcast across 16 metadata bytes via _mm_set1_epi8 and compared simultaneously with _mm_cmpeq_epi8, making one SIMD instruction probe 16 slots. Only on H2 match does the implementation touch the actual key. Elasticsearch reported 6× fewer LLC misses after adopting it; insertions ran ~3× faster than std::unordered_map while using 60% less memory.

The header-only clone phmap::flat_hash_map (Greg Popovitch) reproduces the design without the Abseil dependency.

Where it now ranks

In Jackson Allan's May 2024 benchmark, boost-unordered-flat-map beat it on insert, erase, and lookup-hit. Abseil claws back ground on lookup-miss (1.00× vs Boost's 1.02×) and remains highly competitive overall — but the gap on erase (Abseil 2.17× vs Boost 1.36×) reflects Abseil's structural weakness: tombstones.

The tombstone problem

When a slot is erased, Abseil marks it with a tombstone sentinel rather than reorganizing. Tombstones accumulate under heavy insert/erase cycling and degrade probe-chain performance over time. boost-unordered-flat-map sidestepped this with an overflow byte; ankerl-unordered-dense uses Robin Hood backward-shift; emhash tolerates load factor 0.999 with no tombstones. Abseil is also more sensitive to hash quality than Boost (which auto-bit-mixes weak hashes).

When it is still the right choice

• Large maps with integer keys (its original benchmark sweet spot).
• Anywhere in Google's C++ ecosystem (it is the reference).
• When you want the most battle-tested production deployment.
• As a known fast default — within ~1.1× of Boost on most operations is still excellent.

In the broader convergence

The Swiss Table architecture pioneered here now underpins hashbrown (Rust std::HashMap since 1.36), Go 1.24's built-in map, and experimentally Java. See swiss-table for the architecture and fastest-hash-map-2025 for the full landscape.