4 ms·

STL should “just” be a library written against the language spec. The fact that there are so many different, unrelated implementations is crazy. The STL spec i

STL should “just” be a library written against the language spec. The fact that there are so many different, unrelated implementations is crazy. The STL spec is so over designed that they strongly imply, if not mandate, a particular implementation. But they don’t actually provide an implementation so every platform has to implement one. However those implementations select different trade-offs which means different behaviors so actually developers wind up writing their own version that is consistent. It’s basically the worst outcome.

jltsiren 4y ago
STL is simultaneously overspecified and underspecified. In some situations, like std::unordered_map, the specification mandates an implementation that is suboptimal for many applications. In other situations, the specification is so vague that those parts of the library are often useless. For example, random number distributions are underspecified. The standard only defines the distributions but not the algorithms. If you need deterministic behavior after choosing the rng and the seed – for example for testing, reproducibility, or procedural content generation – you can't use the standard distributions. C++ is typically used in applications where performance is important. If you want a "good enough" standard library, it must be good enough for most people who care about performance. A cross-platform implementation is rarely good enough, because performance often depends on platform-specific parameters and instructions. Data structure APIs are particularly difficult if you care about performance. If you use a generic API, you will often end up using the data structure suboptimally. Your mental model is likely wrong, and the API may not pass through information that could help you avoid redundant work. You want an API that matches the actual implementation closely – but not so closely that every implementation change becomes a breaking API change.
- forrestthewoods 4y ago
  > A cross-platform implementation is rarely good enough, because performance often depends on platform-specific parameters and instructions. I respectfully disagree with this. We’re talking about std::unordered_map here. I don’t know of any platform that has done implementation work to make std::unordered_map not mediocre. Let’s be specific. What platforms are we talking about? How many STL implementations are there? Which ones offer a significant improvement for which platforms? C++ has been my daily driver for 15 years. I’ve shipped quite a lot of code that uses std::vector and std::unordered_map, amongst others! If I really cared about performance I’d pick a hash map implementation that wasn’t mediocre. I would definitely not consider any std::unordered_map to be highly tuned for any platform. Quite the opposite. It’s over-specified but still stuck with different implementations. IMHO it would be much better if the C++ committee released C++XX libraries that were considered “pretty good” with support for common platforms. New or proprietary platforms can then add the relatively small bits of code necessary to support their platform (think PS5) or release a completely separate library with different constraints. Part of the problem here is C++ also really, really sucks to compile. Maybe someday modules will make it easy to drop in new libraries and containers. Jury’s still out on that one though.
  - jltsiren 4y ago
    I also have something like 15 years of serious C++ experience, but I've used it more for data processing than software engineering. While std::vector has almost always been good enough for me, I've often had to replace std::unordered_map with something smaller or faster, or with something where I have more control over memory layout or serialization. std::unordered_map is not a particularly good example of things that need platform-specific optimizations. The implementations I have seen are pretty close to standard C++. std::vector implementations, on the other hand, are often really weird and rely extensively on compiler-specific annotations.
- kwillets 4y ago
  random was what came to my mind after your first sentence. It could really use a deterministic test harness for the different layers.

jltsiren 4y ago

STL is simultaneously overspecified and underspecified. In some situations, like std::unordered_map, the specification mandates an implementation that is suboptimal for many applications. In other situations, the specification is so vague that those parts of the library are often useless. For example, random number distributions are underspecified. The standard only defines the distributions but not the algorithms. If you need deterministic behavior after choosing the rng and the seed – for example for testing, reproducibility, or procedural content generation – you can't use the standard distributions. C++ is typically used in applications where performance is important. If you want a "good enough" standard library, it must be good enough for most people who care about performance. A cross-platform implementation is rarely good enough, because performance often depends on platform-specific parameters and instructions. Data structure APIs are particularly difficult if you care about performance. If you use a generic API, you will often end up using the data structure suboptimally. Your mental model is likely wrong, and the API may not pass through information that could help you avoid redundant work. You want an API that matches the actual implementation closely – but not so closely that every implementation change becomes a breaking API change.

forrestthewoods 4y ago
> A cross-platform implementation is rarely good enough, because performance often depends on platform-specific parameters and instructions. I respectfully disagree with this. We’re talking about std::unordered_map here. I don’t know of any platform that has done implementation work to make std::unordered_map not mediocre. Let’s be specific. What platforms are we talking about? How many STL implementations are there? Which ones offer a significant improvement for which platforms? C++ has been my daily driver for 15 years. I’ve shipped quite a lot of code that uses std::vector and std::unordered_map, amongst others! If I really cared about performance I’d pick a hash map implementation that wasn’t mediocre. I would definitely not consider any std::unordered_map to be highly tuned for any platform. Quite the opposite. It’s over-specified but still stuck with different implementations. IMHO it would be much better if the C++ committee released C++XX libraries that were considered “pretty good” with support for common platforms. New or proprietary platforms can then add the relatively small bits of code necessary to support their platform (think PS5) or release a completely separate library with different constraints. Part of the problem here is C++ also really, really sucks to compile. Maybe someday modules will make it easy to drop in new libraries and containers. Jury’s still out on that one though.
- jltsiren 4y ago
  I also have something like 15 years of serious C++ experience, but I've used it more for data processing than software engineering. While std::vector has almost always been good enough for me, I've often had to replace std::unordered_map with something smaller or faster, or with something where I have more control over memory layout or serialization. std::unordered_map is not a particularly good example of things that need platform-specific optimizations. The implementations I have seen are pretty close to standard C++. std::vector implementations, on the other hand, are often really weird and rely extensively on compiler-specific annotations.
kwillets 4y ago
random was what came to my mind after your first sentence. It could really use a deterministic test harness for the different layers.