For tools like compression programs, you’d generally prefer performance over everything (except data corruption, of course).

Probably you would prefer no backdoors also? Performance without correctness or trustworthiness is useless.

Yes, but my point was that at the level of performance tools like this are expected to operate at, it’s highly probable that you’ll need to get into incredibly esoteric code. Look at ffmpeg – tons of hand-written Assembly, because they need it.

To be clear, I have no idea how to solve this problem; I just don’t think saying that all code must be non-hacky is the right approach.

Performance can be bought with better hardware. It gets cheaper and cheaper every year. Trustworthiness cannot be purchased in the same way. I do not understand why performance would ever trumph clean code, especially for for code that processes user provided input.

This attitude is how we get streaming music players that consume in excess of 1 GiB of RAM.

Performant code needn’t be unclean; it’s just often using deeper parts of the language.

I have a small project that became absolute spaghetti. I rewrote it to be modular, using lots of classes, inheritance, etc. It was then slower, but eminently more maintainable and extensible. I’m addressing that by using more advanced features of the language (Python), like MemoryView for IPC between the C libraries it calls. I don’t consider this unclean, but it’s certainly not something you’re likely to find on a Medium article or Twitter take.

I value performant code above nearly everything else. I’m doing this for me, there are no other maintainers, and it’s what I enjoy. You’re welcome to prioritize something else in your projects, but it doesn’t make other viewpoints objectively worse.

Performant code does not need to be unclean, exactly! My original point was just to not put performance on a pedestal. Sure, prioritize it, but correct and clean should come first - at least for foundational libraries that others are supposed to build upon.

I maintain ML libraries that run on microcontrollers with kilobytes of memory, performance is a friend of mine ;)

I suggest you run your browsers Javascript engine in interpreter mode to understand how crippling the simple and sraight forward solution is to performance.

I guess because at server farm level, performance/efficiency translate to real million USD savings. In general, at scale ends (the cloud and the embedded) this matters a lot. In resource limited environments like raspberry pi, this design philosophy wins over many users between DIY and the private industry.

> Performance can be bought with better hardware.

I hate this argument. If current hardware promises you a theoretical throughput of 100 MB/s for an operation, someone will try to hit that limit. Your program that has no hard to understand code but gives me 5 MB/s will loose in the face of a faster one, even if that means writing harder to understand code.

There is no reason that understandable and safe code will hit just 5% of a theoretical max. It may be closer to 95%.

No, but often it is far worse than 95%. A good example is random.randint() vs math.ceil(random.random() * N) in Python. The former is approximately 5x slower than the latter, but they produce effectively the same result with large enough values of N. This isn’t immediately apparent from using them or reading docs, and it’s only really an issue in hot loops.

Another favorite of mine is bitshifting / bitwise operators. Clear and obvious? Depends on your background. Fast as hell? Yes, always. It isn’t always needed, but when it is, it will blow anything else out of the water.

The compiler will almost always do the bit shifting, and better for you. Just switch python with rust and you’ll get both performance and safe code

Those Python snippets are both clean options :)

Bitwise is highly context dependent. There are simple usages like shifts to divide/multiply by 2. Idiomatic patterns that are clean when wrapped in good reusable and restricted macros, like for common registers manipulation in microcontrollers. And other uses that are anything from involuntary obfuscation to competition grade obfuscation.

> There are simple usages like shifts to divide/multiply by 2.

Clean code should not do that as the compiler will do that.

Clean code should just say what it wants to do, not replace that with low-level performance optimizations. (Also wasn't performance to be obtained from newer hardware?)

Fair point about shifting being superfluous and not clean!

I never said performance should come only from newer hardware. Only that it is possible to trade vs hardware/costs - unlike correctness and trust.

Faster and more complex hardware can also have bugs or back doors, as can cheaper hardware. That said, I'm not happy with buggy and untrustworthy code either.