Somewhat notable is that `char8_t` is banned with very reasonable motivation that applies to most codebases:
> Use char and unprefixed character literals. Non-UTF-8 encodings are rare enough in Chromium that the value of distinguishing them at the type level is low, and char8_t* is not interconvertible with char* (what ~all Chromium, STL, and platform-specific APIs use), so using u8 prefixes would obligate us to insert casts everywhere. If you want to declare at a type level that a block of data is string-like and not an arbitrary binary blob, prefer std::string[_view] over char*.
> there is no guarantee `char` is 8 bits, nor that it represents text, or even a particular encoding.
True, but sizeof(char) is defined to be 1. In section 7.6.2.5:
"The result of sizeof applied to any of the narrow character types is 1"
In fact, char and associated types are the only types in the standard where the size is not implementation-defined.
So the only way that a C++ implementation can conform to the standard and have a char type that is not 8 bits is if the size of a byte is not 8 bits. There are historical systems that meet that constraint but no modern systems that I am aware of.
char8_t also isn't guaranteed to be 8-bits, because sizeof(char) == 1 and sizeof(char8_t) >= 1. On a platform where char is 16 bits, char8_t will be 16 bits as well
The cpp standard explicitly says that it has the same size, typed, signedness and alignment as unsigned char, but its a distinct type. So its pretty useless, and badly named
Wouldn't it be rather the case that char8_t just wouldn't exist on that platform? At least that's the case with the uintN_t types, they are just not available everywhere. If you want something that is always available you need to use uintN_least_t or uintN_fast_t.
It is pretty consistent. It is part of the C Standard and a feature meant to make string handling better, it would be crazy if it wasn't a complete clusterfuck.
> There's no guarantee char8_t is 8 bits either, it's only guaranteed to be at least 8 bits.
Have you read the standard? It says: "The result of sizeof applied to any of the narrow character types is 1." Here, "narrow character types" means char and char8_t. So technically they aren't guaranteed to be 8 bits, but they are guaranteed to be one byte.
Well platforms with CHAR_BIT != 8. In c and c++ char and there for byte is atleast 8 bytes not 8 bytes. POSIX does force CHAR_BIT == 8. I think only place is in embeded and that to some DSPs or ASICs like device. So in practice most code will break on those platforms and they are very rare. But they are still technically supported by c and c++ std. Similarly how c still suported non 2's complement arch till 2023.
That's where the standard should come in and say something like "starting with C++26 char is always 1 byte and signed. std::string is always UTF-8" Done, fixed unicode in C++.
But instead we get this mess. I guess it's because there's too much Microsoft in the standard and they are the only ones not having UTF-8 everywhere in Windows yet.
Of course it can be made UTF-8. Just add a codepoints_size() method and other helpers.
But it isn't really needed anyway: I'm using it for UTF-8 (with helper functions for the 1% cases where I need codepoints) and it works fine. But starting with C++20 it's starting to get annoying because I have to reinterpret_cast to the useless u8 versions.
First, because of existing constraints like mutability though direct buffer access, a hypothetical codepoints_size() would require recomputation each time which would be prohibitively expensive, in particular because std::string is virtually unbounded.
Second, there is also no way to be able to guarantee that a string encodes valid UTF-8, it could just be whatever.
You can still just use std::string to store valid encoded UTF-8, you just have to be a little bit careful. And functions like codepoints_size() are pretty fringe -- unless you're not doing specialized Unicode transformations, it's more typical to just treat strings as opaque byte slices in a typical C++ application.
I never said always. Just add some new methods for which it has to be UTF-8. All current functions that need an encoding (e.g. text IO) also switch to UTF-8.
Of course you could still save arbitrary binary data in it.
> That's where the standard should come in and say something like "starting with C++26 char is always 1 byte and signed. std::string is always UTF-8" Done, fixed unicode in C++.
> I never said always
Yes you totally did. And regarding "add some new methods for which it has to be UTF-8", there is no need at all to add UTF-8 methods to std::string. It would be a bad idea. UTF-8 is not bound to a particular type (or C++ type). It works on _any_ byte sequence.
Is there a single esoteric DSP in active use that supports C++20? This is the umpteenth time I've seen DSP's brought up in casual conversations about C/C++ standards, so I did a little digging:
Aside from that, from what I can tell, those esoteric architectures are being phased out in lieu of running DSP workloads on Cortex-M, which is just ARM.
I'd love it if someone who was more familiar with DSP workloads would chime in, but it really does seem that trying to be the language for all possible and potential architectures might not be the right play for C++ in 202x.
Besides, it's not like those old standards or compilers are going anywhere.
Cadence DSPs have C++17 compatible compiler and will be c++20 soon, new CEVA cores also (both are are clang based).
TI C7x is still C++14 (C6000 is ancient core, yet still got c++14 support as you mentioned).
AFIR Cadence ASIP generator will give you C++17 toolchain and c++20 is on roadmap, but not 100% sure.
But for those devices you use limited subset of language features and you would be better of not linking c++ stdlib and even c stdlib at all (so junior developers don't have space for doing stupid things ;))
How common is it to use Green Hills compilers for those DSP targets? I was under the impression that their bread was buttered by more-familiar-looking embedded targets, and more recently ARM Cortex.
Dunno! My last project there was to add support for one of the TI DSPs, but as I said, that's decades past now.
Anyway, I think there are two takeaways:
1. There probably do exist non-8-bit-byte architectures targeted by compilers that provide support for at-least-somewhat-recent C++ versions
2. Such cases are certainly rare
Where that leaves things, in terms of what the C++ standard should specify, I don't know. IIRC JF Bastien or one of the other Apple folks that's driven things like "twos complement is the only integer representation C++ supports" tried to push for "bytes are 8 bits" and got shot down?
Judging by the lack of modern C++ in these crufty embedded compilers, maybe modern C++ is throwing too much good effort after bad. C++03 isn't going away, and it's not like these compilers always stuck to the standard anyway in terms of runtime type information, exceptions, and full template support.
Besides, I would argue that the selling point of C++ wasn't portability per se, but the fact that it was largely compatible with existing C codebases. It was embrace, extend, extinguish in language form.
> Judging by the lack of modern C++ in these crufty embedded compilers,
Being conservative with features and deliberately not implementing them are two different thing. Some embedded compilers go through certification, to be allowed to be used producing mission critical software. Chasing features is prohibitively expensive, for no obvious benefit. I'd bet in 2030s most embedded compiler would support C++ 14 or even 17. Good enough for me.
> Being conservative with features and deliberately not implementing them are two different thing.
There is no version of the C++ standard that lacks features like exceptions, RTTI, and fully functional templates.
If the compiler isn't implementing all of a particular standard then it's not standard C++. If an implementation has no interest in standard C++, why give those implementations a seat at the table in the first place? Those implementations can continue on with their C++ fork without mandating requirements to anyone else.
> Then they will diverge too much, like it happened with countless number of other languages, like Lisp.
Forgive me if I am unconvinced that the existence of DSP-friendly dialects of C++ will cause the kinds of language fracturing that befell Lisp.
DSP workloads are relatively rare compared to the other kinds of workloads C++ is tasked with, and even in those instances a lot of DSP work is starting to be done on more traditional architectures like ARM Cortex-M.
A cursory Chromium code search does not find anything outside third_party/ forcing either signed or unsigned char.
I suspect if I dug into the archives, I'd find a discussion on cxx@ with some comments about how doing this would result in some esoteric risk. If I was still on the Chrome team I'd go looking and see if it made sense to reraise the issue now; I know we had at least one stable branch security bug this caused.
Related: in C at least (C++ standards are tl;dr), type names like `int32_t` are not required to exist. Most uses, in portable code, should be `int_least32_t`, which is required.
Isn't the real reason to use char8_t over char that it that char8_t* are subject to the same strict aliasing rules as all other non-char primitive types? (i.e., the compiler doesn't have to worry that a char8_t* could point to any random piece of memory like it would for char*?).
At least in Chromium that wouldn't help us, because we disable strict aliasing (and have to, as there are at least a few core places where we violate it and porting to an alternative looks challenging; some of our core string-handling APIs that presume that wchar_t* and char16_t* are actually interconvertible on Windows, for example, would have to begin memcpying, which rules out certain API shapes and adds a perf cost to the rest).
The shortest double-to-string algorithm is basically Schubfach or, rather, it's variation Tejú Jaguá with digit output from Dragonbox. Schubfach is a beautiful algorithm: I implemented and wrote about it in https://vitaut.net/posts/2025/smallest-dtoa/. However, in terms of performance you can do much better nowadays. For example, https://github.com/vitaut/zmij does 1 instead of 2-3 costly 128x64-bit multiplications in the common case and has much more efficient digit output.
Schubfach, Ryū, Dragonbox etc support round-tripping and shortest-width, which (it sounds like) is not important for you. The idea of round-tripping is that if you convert a double to a string and then parse that, you get the exact same value. Shortest-width is to correctly round and generate the shortest possible text. I tried to implement a version that does _just_ round-tripping but is not shortest-width, it is around 290 lines for both parsing and toString [1]
I implemented Teju Jaguá in Rust, based of the original C impl https://github.com/andrepd/teju-jagua-rs. Comparing to Zmij, I do wonder how much speedup is there on the core part of the algorithm (f2^e -> f10^e) vs on the printing part of the problem (f*10^e -> decimal string)! Benchmarks on my crate show a comparable amount of time spent on each of those parts.
I don't have exact numbers but from measuring perf changes per commit it seemed that most improvements came from "printing" (e.g. switching to BCD and SIMD, branchless exponent output) and microoptimizations rather than algorithmic improvements.
What about reasonably fast but smallest code, for running on a microcontroller? Anything signifactly better in terms of compiled size (including lookups)?
If you compress the table (see my earlier comment) and use plain Schubfach then you can get really small binary size and decent perf. IIRC Dragonbox with the compressed table was ~30% slower which is a reasonable price to pay and still faster than most algorithms including Ryu.
Note that ~3-6ns is on modern desktop CPUs where extra few kB matter less. On microcontrollers it will be larger in absolute terms but I would expect the relative difference to also be moderate.
It's easier to write faster code in a language with compile-time facilities such as C++ or Rust than in C. For example, doing this sort of platform-specific optimization in C is a nightmare https://github.com/vitaut/zmij/blob/91f07497a3f6e2fb3a9f999a... (likely impossible without an external pass to generate multiple lookup tables).
It depends on the input distribution, specifically exponents. It is also possible to compress the table at the cost of additional computation using the method from Dragonbox.
Thank you! The simplicity is mostly thanks to Schubfach although I did simplify it a bit more. Unfortunately the paper makes it appear somewhat complex because of all the talk about generic bases and Java workarounds.
I've just started a Julia port and I think it will be even cleaner than the C version (mostly because Julia gives you a first class (U)Int128 and count leading zeros (and also better compile time programming that lets you skip on writing the first table out explicitly).
I considered computing the table at compile time (you can do it in C++ using constexpr) but decided against it not to add compile-time overhead, however small. The table never changes so I'd rather not make users pay for recomputing it every time.
It converts 1.0 to "1.e-01" which reminds me to remove the trailing decimal point =). dtoa-benchmark tests that the algorithm produces valid results on its dataset.
My bad, you are right. The small integer optimization should be switched to a different output method (or disabled since it doesn't provide much value). Thanks for catching this!
reply