atomic128

128-bit atomic operations for Golang, using CMPXCHG16B when available.

Partially based on github.com/tmthrgd/atomic128, without the problematic parts.

Performance

$ benchstat <(go test -bench=. -cpu=8 -benchtime=0.1s -count=5)
name              time/op
Load/native-8     28.1ns ± 2%
Load/fallback-8    106ns ± 3%
Store/native-8     106ns ± 2%
Store/fallback-8   109ns ± 5%
Swap/native-8      110ns ± 1%
Swap/fallback-8    127ns ± 2%
Add/native-8       129ns ± 2%
Add/fallback-8     142ns ± 2%
CAS/native-8      31.4ns ± 1%
CAS/fallback-8     135ns ± 2%

$ benchstat <(go test -bench=. -cpu=1 -benchtime=0.1s -count=5)
name            time/op
Load/native     16.2ns ± 3%
Load/fallback   17.0ns ± 2%
Store/native    16.1ns ± 2%
Store/fallback  19.9ns ± 2%
Swap/native     19.3ns ± 1%
Swap/fallback   23.5ns ± 3%
Add/native      20.3ns ± 2%
Add/fallback    31.6ns ± 2%
CAS/native      17.3ns ± 6%
CAS/fallback    25.3ns ± 2%

TODO

• Add ARM/aarch64 assembly version
• Add shift/rotate operations