Most of the optimizations listed have nothing to do with unsafe, and have to do with algorithmic changes that you could apply in any language, and could apply to safe code just as well as to unsafe code. Even if we pretend all unsafe used there is necessary, and interpret the source in the worst possible light for rust - that the optimization steps would be impossible without unsafe - I think we'd only be talking 4x to 5x slower from this article, not 25x.

But - I strongly suspect not all that unsafe is necessary, and the post certainly doesn't claim it is.

The first code snippet uses unsafe to skip bounds checks. There are plenty of other ways to get the optimizer to skip them for you in safe code - or at the very least, hoist them out of the loop instead. This seems like a good read showing some concrete examples:

https://coaxion.net/blog/2018/01/speeding-up-rgb-to-grayscal...

Explicitly using SIMD code like the second code snippet does might require unsafe for the transmuted slicing under the hood (implemented as raw pointer derefs in this case), but it's possible to build safe abstractions on top of that. That blog post links this, for example, although it didn't end up using it (EDIT: Actually, it initially didn't, but now does): https://github.com/AdamNiederer/faster . It does use unsafe internally: https://github.com/AdamNiederer/faster/search?q=unsafe&unsco...

The fragment of code in the article with `get_unchecked_mut()` shouldn't be necessary. It's a simple case that LLVM should be able to optimize. And if it didn't, it could be helped by either iterating both slices with `zip()` or a trick `let slice = &slice[0..len];` which proves to LLVM that you have the required length and it doesn't need to be checked again.

But overall the article seems in line what you'd expect from Rust: you have low-level control over memory layout and safety checks, and you can make trade-offs to squeeze maximum performance out of an algorithm if you need to.

I don't see why they used unsafe in `add_assign` - an assert!(octets.len() == other.len()) would likely have elided the bounds checks.

I would imagine without these simd optimizations you would see a 2X slowdown at least.