The Lost Atomic Update on Loongson CPU

In 2026, Wang Miao encountered an unusual issue while compiling normaliz for Debian on a LoongArch server. The math software's built-in test would time out due to an infinite loop that could not be resolved. The loop's exit condition depended on the accumulated value matching a specific number, but the accumulated result was always lower, resulting in an endless loop. Due to the large and complex codebase, a minimal example that could be understood by a human was never achieved, so the matter was put aside.

Six months later, Wang Miao returned to discuss the issue once more. This time, an AI was employed to help find a minimal reproducible example. After two days, a stable reproducer was developed, leading to the discovery of a new CPU erratum: the CPU's atomic add instruction occasionally fails to execute as an atomic operation. Loongson learned of this issue and provided a fix with minimal performance impact, expected to be released before National Day (October 1).

The story of the lost atomic update on Loongson CPU unfolds from this point onward, detailing the discovery, investigation, and resolution of the issue.