Record each discovered defect with an identifier, a minimal reproduction, expected and actual behavior, and links to the affected specification and code.
Do not use this log as a substitute for feature planning; planned work belongs in the backlog.
FF-P1 — CLMul at GF(2^64) loses its speed in the element wrapper (open, 2026-11-19)
- Where:
julia/FiniteFieldArithmetic.jl/src/nimber_fast.jl,nim_mul_bits(::CLMul, …, Val(6))andmul(CLMul(), a::GF2_64, b::GF2_64); measurements insrc/nimber_fast.md. - Reproduction: broadcast over 4096 random pairs, in one process. On
UInt64bit patterns,nim_mul_bits(CLMul(), x, y, Val(6))takes 10.4–11.7 ns per product against 18–20 ns forMulTable. Wrapped asGF2_64(mul(CLMul(), a, b)orunsafe_frombits(GF2_64, nim_mul_bits(…))) it takes 17–19 ns. The native code of both functions is nearly identical (168 vs 179 instructions, the same single out-of-line call). - Expected: the wrapper (an
isbitsstruct around theUInt64) costs nothing. - Actual: about 7 ns per product in some runs. The VM’s run-to-run noise (up to about 40%) makes it unclear how much is real: one run of the bare variant also took 18 ns.
- Consequence: GF(2^64) keeps
MulTableas default. Investigate on a quiet machine (code_llvmof the whole broadcast loop,perf), then reconsider the default.