`noexcept` affects libstdc++'s `unordered_set`

The article discusses how the `noexcept` specifier impacts the performance of `std::unordered_set` in GNU libstdc++. It explains that the layout of nodes in hash-based associative containers can change based on whether the hash function is marked as `noexcept`. The author illustrates how the `erase` function can be optimized by precomputing and storing the hash value in each node, which can improve performance during element removal and rehashing. The article highlights that libstdc++ does not cache hash values by default, assuming that recomputing them is inexpensive, which can lead to performance issues when using non-trivial hash functions. The author provides benchmarks showing the performance differences based on the type of hash function used, emphasizing that marking a non-trivial hash function as `noexcept` can lead to increased calls to the hash function and slower performance. The conclusion suggests that while these details may not be critical for most programming tasks, they highlight the importance of understanding how `noexcept` can affect performance in specific scenarios. The author advocates for improvements in libstdc++ to better handle these cases.

- The `noexcept` specifier can significantly affect the performance of `std::unordered_set`.

- Storing hash values in nodes can optimize performance during element removal and rehashing.

- libstdc++ assumes recomputing hash values is cheap, which can lead to inefficiencies with non-trivial hash functions.

- Marking non-trivial hash functions as `noexcept` can result in performance degradation.

- The author calls for improvements in libstdc++ to address these performance issues.