The Lost Atomic Update on Loongson CPU

While packaging the math library normaliz on a LoongArch server, maintainers hit an infinite loop: an OpenMP-parallel loop never terminated because a shared counter never reached its expected total. Debugging showed the code used #pragma omp atomic increments compiled to the LoongArch64 amadd.d instruction, but per-run counter totals diverged unpredictably. Human-driven reduction failed to produce a minimal reproducer, so the bug was shelved until AI-assisted investigation resumed months later. The AI-aided approach quickly produced a stable minimal test and revealed the root cause: the CPU’s atomic add sometimes loses updates, so atomics that should be indivisible occasionally fail under specific runtime conditions.

Further tests showed CAS and a wide range of atomic operations can suffer the same lost-update behavior. The trigger was linked to vectorized memcpy using LASX instructions (xvld); LSX vector loads and ordinary scalar reads usually did not trigger it, though later work found certain memory-address relationships can make scalar reads trigger the bug at lower probability. Loongson received the report, produced a firmware patch within two weeks that fixes the erratum with almost no performance loss, and provided test firmware that reproduces the fix; a public firmware release was expected before October 1.