[as-j]

Thanks is, as the GitHub title says, for iot and embedded devices.

There’s a whole class of these that don’t have the resources to run C++ or rust. Source: develop on a Linux system with 7 megabyte roofs and 16 megs of ram. Oh! We have a 1 meg application partition.

[gpderetta]

If it can run C, it can run C++ (or rust I guess). The majority of abstractions have little to no overhead. Also 16 megabytes were enough for a very decent desktop computer in the '90, enough to run very complex C++ applications with ease.

[scottlamb]

I think there's some truth in what as-j is saying. Rust binaries (and C++ ones) tend to be larger than C ones. I think the major reasons are (a) Rust dependencies being statically linked due to a lack of ABI stability, (b) inclusion of portions of the (statically linked, see a) Rust standard library used by the program where C code uses libc, (c) code bloat due to monomorphization, (d) the ease of just using a full-featured library where someone writing in C might cheat a little bit. As an example of what I mean by the last point, see sdp_attribute_get_answer in this codebase. [1] It's writing JSON, but it doesn't use a JSON library that actually escapes the included string. It just assumes the included string doesn't have a quote character in it. Is that assumption valid? Will it always be valid? I'm not sure on quick inspection.

There are ways around all of these:

* a. Static vs dynamic linkage: in an embedded system, it'd be reasonable to just produce a single userspace binary that does everything. It can change its behavior based on argv[0]. I think this is not too unusual for constrained systems even with C binaries. Eg busybox does it. If you only have one binary, you don't need dynamic linking. Also, I think it's not strictly true that Rust doesn't support dynamic linking. I think you can dynamically link everything if you ensure the whole system is built with the same compiler version.

* b. Standard library. You don't have to use it at all, or you can use it sparingly, paying only for what you use.

* c. Monomorphization. You could write (for example) a Go-like map [2] rather than relying so heavily on monomorphization. I'd love to see someone take this idea as far as possible; it might be a good idea for a lot of non-inner-loop code in general, not just on tight embedded systems.

* d. Using full-featured libraries. Obviously no one is making you do this; the same cheats available in C are available in Rust.

but in fairness, the further you go down this path, the further you are from just being able to just take advantage of the whole Rust ecosystem.

Personally, I'd still rather develop or use a #![no_std] Rust codebase than a C one. Memory safety is important to me. IOT devices are no exception to that. Their security history is horrible, and I'd like their security future to be better...

[1] https://github.com/sepfy/pear/blob/b984c8dccaafdcdd1b181786a...

[2] https://dave.cheney.net/2018/05/29/how-the-go-runtime-implem...

[stjohnswarts]

Space is so cheap these days, it doesn't matter much unless you're on a really low budget. That said I still get a kick out of programming specialty cases with 8 and 16 bit micros. It's nice to have just about everything you need on a chip (a2d, d2a, CAN, I2C, SPI, timers, tons of general I/O). All you kids should do at least one project like that, you'd be surprised at how fun it can be to pick and choose exactly what the hardware is doing down to the clock cycle :)

[xfer]

Really? You have megabytes of memory but can't run C++? Which architecture is this?

[stjohnswarts]

lol none, someone who says that doesn't really know what they're talking about. You can use zero overhead subset of c++ just fine and it will take barely any more room than C. Now if you pull in the std library it's over :)