February 21, 2018

A cake for your cherry: what should go in the C++ standard library?

This is a reply to Guy Davidson’s article “Batteries not included: what should go in the C++ standard library?”.

Over the past few years there has been a push to include a graphics library into the C++ standard. It would be something a bit like cairo. Or SDL. The proposal, in its current form, is here

In its current state, the library proposal can draw some shapes on a pre-allocated surface, has some support for images, and there are of courses projects to add text, maybe some input in the form of mouse / keyboard handling.

The primary goal of the library seems to be teaching. The argument put forward is that it’s cool and ludic for kids to have pretty blinky pixies on screen. Of course there already exists libraries to do that and more, but you see, C++ has no decent, idiomatic package manager, so, of course the conclusion was reached by some prominent committee members that the C++ standard should offer a 2D graphics library, out of the box.

I do think this is a path that should not be pursued and that doing so would be, at best, a waste of time. Let me tell you why.

But, first, some needed clarifications

Guy Davidson and others have, on all account, put a great amount of work, time and energy in that proposal. The people pushing for that proposal to be rushed through standardization are way more experts than I will ever be.

I did not contribute anything to C++ so what will follow are just the opinions of one guy.

I also want to make clear that I do not have a negative opinion of that particular library. My issue is with the inclusion of a 2D painting library, any painting library in the C++ standard, at this point in time.

I hope I won’t be misconstrued

Anyway, let’s get to it.

The C++ Standard Library is not a library

The C++ Standard is exactly that: A well specified document that describes in the most detailed and unambiguous possible way what C++ is, and how it works. The goal being that anyone can implement a C++ compiler for themselves by implementing that specification. However, it happens that the specification is not specific enough, or implemented not quite properly, or implemented opinionatedly and so various C++ compilers end up having some slight differences of behavior from one implementation to the next. Sometimes it cannot be implemented at all because the people doing the implementation and the people doing the specification forgot to talk to each others.

Now, a big part of that specification describes the Standard Template Library, a library that ships with every conforming compiler.

There exist at least 5 implementations of that specification, maintained by as many entities. Some are open source, some are not. They each work in a select subset of platforms and system. And even if they sit at the very bottom of about any C++ program, they are, like any other library , subject to bugs.

In this context, what should, or should not be included in the C++ standard library is a very important question. What should come as standard bundled with the compiler? What do most people need to be productive with C++ ?

Guy’s article describes the positions one can have. Maybe we need about nothing ? Maybe we need some vocabulary types ? Maybe containers ? Maybe not ? Do we need filesystem support ? sockets ? json ? xml ? rpg making tools ? sql ? html ? javascript vm ? 2d graphics ? 3d graphics ? soap ? IPC ? windowing ? Should $\pi$ be defined ? What about websockets ? ftp ? ssh ? VR ? AR ? crypto ? ssl ? Do we need ssl but no other crypto ? Deep learning ? Sound ? 3d sound ? Video Decoding ? gif ?

Clearly we need to draw a line.

Somewhere ?

Where ?

Let’s look at .Net. Or Java. When the STL is mentioned, comparing C++ and Java is customary. Java is cool, right? It has sockets and HTTP and crypto and everything, basically.

But Java is mostly maintained by a single entity. So someone at Oracle decides Java should have sockets and they implement it, there are internal reviews and now Java has Sockets. Sometimes later, Google want to have sockets using the same API, and before they can say “Ahead of time”, they are sued for 9 billions USD.

Meanwhile, the C++ specification undergoes a long, painful process until there is a vote and there is a majority consensus on every single feature, every single method. Should that be called data? get? “At Bloomberg we have experience using data on our 2 millions line code base” will say the guy working at Bloomberg. “We noticed it’s faster to use type get on EBCDIC keyboards” Will object the guy at IBM. “And we have a 3 millions lines code base”.

I don’t have an opinion on which model is best. Benevolent dictatorship obviously only works if the dictator is benevolent.

However, I will argue that democracy is unfit to the birth of a good graphic library.

The Committee Resources are limited

Even if sleep deprived proposal authors sweat blood, a big part of the work and the voting takes places in week-long quarterly meetings where people go through an ever growing pile of proposals. As the committee learn to be more transparent, more people contribute, leading to more work for people attending. There is little to no money in that work. At best you can hope for someone to pay you the plane tickets to the beaches of Florida, the green hills of Switzerland or the pools of Hawaii at which the meeting happen. You will reportedly never see neither the beaches, the hills nor the pools.

And because resources are limited and time is limited there is a need to sort, prioritize and even discard proposal. Directions for ISO C++ attempts to describes how that sorting and prioritizing should happen.

The question then becomes : can the committee spares the time to work on a 2D graphics library and is that a priority ?

In its current form, which is limited to drawing shapes, the proposal is about 150 pages long. It’s one of the biggest proposal submitted for the next meeting.

It can only grow bigger. There is no end to the complexity of a “small and simple graphics library”. Every second spend on that proposal will not be on some other work. Of course people discuss proposals they have an interested in and discussions happen in parallel. Still. There is maybe one person in those meetings for every 200’000 c++ developers.

Let’s draw a triangle

A 2D graphics is the complete opposite of what the Standardization process is good at. Standardization is all about formalism, so it works best at describing formal things, Math, Algorithms. The more reality gets messy, the more it’s hard to describe it put it on paper and have that paper serves as source of Truth for decades.

The first thing one need to do play with pretty pixes is to get a “surface”. A canvas where pixels get drawn.

So hopefully you have a surface class to which you give dimensions and that gets you a canvas on which to paint.

But wait. On most desktop systems, if you want a surface you need to put it in a window. It’s customary for windows to have titles so a graphics API should probably handle that, right ?

You probably also want the window to have an icon. An icon is a file on most system, the format of which is system specific. But sometime it’s not a path, it’s a name corresponding to a path.

The size of a window can change during the execution of the program on some desktop operating system.

Sometimes the window can be moved to another screen that has another resolution. And there is this weird new screens where there are virtual pixels that are bigger than true pixels ? Unless you are rendering an images or something then you should make sure you uses all the power of the small crispy pixes since the customer paid a premium for boasting about how crispy his screen is.

That woman over there was jealous so she bought a TV with 40 bits per pixels. You can’t really see the difference but are you going to tell her she wasted 5000 bucks ?

And then there is a screen in your pocket and IT ROTATES in all the directions and now the surface is all wonky. But it has no window so it has no title or icon.

What time is it ? OH GOSH THAT THING HAS A SCREEN TOO BUT IT’S SO SMALL… Better go and read a book WTF ELECTRONIC INK that you should refresh as little as possible and that only is black ?

World is crazy, right ? Let’s stick to Linux, shall we? So on Linux there is this thing called X11 which you request a surface to… oh sorry, while you are writing the paper X11 is being deprecated and now you should use Wayland… unless you rather have a frame buffer? It can be accelerated using opengl. or embedded opengl. totally different thing. But really, Vulkan is faster than both these things. Oh and on this system we prefer that you draw the windows yourself, there is a war raging on about CSD vs SSD it has been going on for years and you can’t take side.

And if you have CSD please make sure that I can drag the windows properly and I have set up sticky corners so that the windows can be aligned nicely. Make sure to handle them. Properly. And when you drag the window it should be a bit transparent, you know about windows compositing right ?

Okay so, you start to tell yourself that maybe drawing stuff is complicated. Let the implementors compiler writers and library vendors deal with all that crap. So you provide an API that works everywhere, so it handles absolutely nothing, which is to say it works nowhere.

Now the compiler writers are a bit pissed. All they wanted in life was to write compilers and there they are, trying to understand how GDI work. Plus Microsoft is maybe not really interested in providing a drawing framework, they rather have their users use the WinRT xml-based tools. Meanwhile the GCC guys are still trying to have std::thread work on windows.

Clang people get bug reports that “it doesn’t work”. People have expectations that the STL will work perfectly, consistently, anywhere

No problem. We will make the graphics library optional. So now there are bits of the Standard Library that are not standard. If and when they are implemented, they don’t behave quite the same on every platform. So now the code written using standard tools is not portable. So we need to have a copy of the STL in the repository along with messy build scripts. Back to square one.

Maybe we messed up somewhere? Let’s look at what exists on Internet. People have displays so surely they do write libraries for them, right ?

Turns out Qt is pretty popular. It does a lot more than displaying triangle though. It was released in 1995. It has strings, threads, tons of stuffs. People really did not came up with anything better since ?

wxWidgets is even older. It too has strings and threads and a lot of things that don’t have business being in a graphic library. GTK is the exact same thing.

But C++ goals are more aligned with things like SDL. Released in 1995 with threads and strings and weird things. Allegro, released in 1990. Same thing

You look at other languages. Surely the Rust community has an awesome painting framework, right ? Or the Go people ? Turns out they write wrappers around Qt, or SDL or something, like they deemed to complicated to start from scratch.

So 20 years later you manage to draw a triangle on all platforms. For some definition of all.

It’s quite the achievement, so you want to share your joy with the world. People communicate mostly using languages [citation needed] so you are going to display some words on the screen, how hard can it be to go from a triangle to that ?

void draw_text(std::point2d, std::string);

You learn that there is a standard called “Unicode” that describe all the letters people around the world use. So many letters. The Unicode standard is about 10 times the size of the proposal you have been working on for 5 years. Fortunately most programming languages have backed in support for at least parts of Unicode. Except C++. Well, okay let’s put that aside for now.

So text is rendered using fonts. The fonts are often installed on the system. There is that thing called a font database that tells what the font are. Unless the systems has no font database. Or no fonts. Or no system. People also like to use their own fonts.

A font is a file, whose format is standard. There are 5 or so competing standards.

A font file can contain glyph tables, PNGs, SVGs, scripts executed in a virtual machine, a mix of all that. Some fonts have color, but not all people like colors. Your children like colors. They sent you a 🐈. You will add support for cats, right ?

You learn about subpixel rendering. You spend a few months in jail for patent infringement. You figure you can use that time to learn about ligatures in the encyclopædia. You start regretting being a developer and consider a new career as a monastic scribe.

There is a lot of math involved in font rendering so you pick up a mathematics book written a dead guy called AL-Khwarizmi. You realize everything is written from right to left. How does that even work ?

So maybe the optional 2D graphics library should have optional text support ?

At the next committee meeting in Toronto ( Hawaii sank in the ocean long ago), someone is trying to write a complex graphic application with network and lots of input and to avoid spaghetti code they like to have some kind of event loop with maybe some threading. It’s a theoretical concern obviously as there is no input support. Consensus was never reached on how to name the keyboard keys.

You think back on all the existing framework like Qt, now on version 8.0, that provide an event loop, a message passing system and a Unicode string type. Maybe they were up to something.

During all this time, people continued using Qt. People were hired for knowing Qt. They used it in their school projects. Of course Qt still sucks because the C++ reflection features that got added in the standard were never quite enough to replace their code generator. But people don’t care that its sucks. People who do use QML. Or Electron.

Having not displayed a 🐅, let’s go back to 2018.

Has the committee anything better to do anyway ?

To be considered, a proposal has to be written and put forward, and the library proposal exists because someone put a lot of work in it.

However, currently, C++ has

  • Poor threading support ( no executors or facilities to use coroutines )
  • No support for launching processes
  • No support for Unicode
  • Poor I/O facilities
  • Poor locale facilities
  • No support for dynamic loaded libraries
  • No HTTP support
  • Nothing crypto related

The list, of course goes on. I don’t know what is a good candidate for a C++ library, but, according to the committee itself, a library proposal should

  • Be useful to most people
  • Have a stable API that isn’t too subject to frequent change
  • Have real world experience and feedback. That’s why most C++ library started their lives as boost library.

Proposal are often dismissed from the get go for not being useful enough or not being battle tested enough. Which is reasonable given the expectation people have regarding the stability of the STL but then those criteria should apply consistently.

And of course there are a lot of languages features that are still in the pipelines after years and years of work, and they should take priority over library features since pure library addition can be polyfilled by boost or other.

The teaching argument

One of the argument put forward for the inclusion of that library is that it would make C++ more teachable and that people are more interested in graphics-based projects.
I sympathize and fully agree with the goal of making C++ more teachable. However, there is a difference between making sure a given feature is teachable and adding a major feature to the language with the primary goal of being used in classrooms.

Teachability implies easy to use, hard to misuse, and a sane mapping between a concept and its implementation, and generally behaving accordingly to the expectation of the majority of users. Quality that should be looked for in any new feature.

It is also to be expected that some features are targeted to advanced users, library writers and experts.

However, the “teaching friendly part” of C++ should be a subset of the features used in a professional settings rather than a different set.

I would prefer that people learn to use Qt (for example) as that is a skill they can use in their professional careers, rather that something dedicated to teaching purposes.

I also think that a library whose scope is too limited may give a bad image to the language. If people are told that they can’t draw emojis or gifs or use a gamepad, they may end up thinking that C++ is not powerful enough and switch to another language like C#, java, javascript, swift… But if they get to use an existing, battle tested framework, that is powerful enough to let them implement their design (Qt, SDL) even if the code isn’t “modern”, they will get a better grasp of what c++ can and does do.

In other words, I’m afraid that if people are introduced to a toy library, they will think that C++ is a toy language.

Beside, “Teaching” needs to be better define.

Are we talking about middle schoolers? And if so, is teaching them C++ a good idea ? In some cases Python, Javascript, Lua are more suitable, easier to grasp choices. I think that’s okay.

Are we talking about college CS 101 ? In this case, it’s probably desirable to introduce students to build systems, libraries and package management. tools are important. And in my experience, a lot of juniors developers don’t know how to use their tool and that’s as important as knowing languages. It’s also important that people know and are taught the ecosystem. Qt, boost, wxwidgets, SDL…

The “We need a standard library because using third party libraries is hard” argument

I think most people agree on that. Including a library in a C++ project is a bad, often painful experience. Investing a lot of resources on a 2d graphic library don’t solve that problem. Unless every single library that exist or will exist gets folded into the standard, so, where do we stop ?

And I’m sorry to say, things won’t improve on their own, it’s just not possible. The number one requirement for a package manager of any sort is to be authoritative. It doesn’t even necessarily need to be good. But until individual entities are left to deal with the issue, we will continue to have myriad of incompatible, half backed tools. I understand that the committee prerogatives don’t necessarily extend outside the definition of the language and therefore the issue of package management may not be solvable. But tools, not UI is the big challenge that C++ must tackle.

Note that there are ways the committee can help improving tooling without extending its prerogatives, notably:

  • Finding ways to supersede all reasonable uses of the preprocessor ( the work on reflection / code injection is very important for that )
  • Defining a portable C++ ABI (N4028)
  • Defining a portable module representation

Sure, these works may not be as glamorous as a 2D API, but they are more fundamental, and more importantly, can’t happen independently of the committee.

Things should move forward, somehow.

After looking at P0939 and P0267, I wanted to share my wishes for work to be done in related areas. Of course, I’m not in a position of doing more than wishing and I can only hope to inspire someone ! But I’m interested in what you think is important in those areas !

Take the Unicode bull by the horns

I would not have suggested that, as I understand why C++ lacks Unicode but if we are seriously considering 2D graphics, then we absolutely need to have proper Unicode support.

  • A first step is the char8_t paper . Of course that is not sufficient, but it is necessary.
  • We need a set of algorithm to normalize, compare, sanitize and transform Unicode strings, count characters. Something range-based could work nicely
  • class of characters, Regexps… We might not need as many features as ICU but we do need some. That could be a <unicode> header. I’m not certain that proper Unicode support is a goal aligned with the constraints outlined in P0939, however it would be beneficial to any application dealing with user input/output, including GUI, databases, (web) server, console application…

I don’t know if we can qualify Unicode strings of vocabulary type but handling the world languages is certainly something everybody needs and it would be easier if there was a universal, idiomatic tool to do that.

Add geometry primitives to the standard

It could be interesting to extract the vocabulary types introduced in p0267 and standardize them independently of graphics. Types like point_2d , matrix_2d ( and eventually point_3d, matrix_3d ) are useful for graphics but may have other use, for example scientific computation, plot manipulation. They could be accompanied by a set of methods to perform widely used analytic geometry computation. All of that could live in a <geometry> header.

There are multiple reasons why this would be beneficial

  • It’s something every library dealing with painting or surfaces need SDL_Point, QPoint, wxPoint. Converting from one type to the other is cumbersome, error prone. All these frameworks could benefit from speaking the same language in the same coordinate system. It’s the definition of a vocabulary type.
  • It’s guarantee to stand the test of time. Math are not affected by new technology trends and as such the API would remain stable for decades.
  • For the same reason, it would be probably be easy to reach consensus, it’s hard to bike-shed basic maths.

Help improving existing graphics library

What can be done by the committee to improve Qt, wxWwidgets SDL and other graphics frameworks ? It turns out a lot of frameworks rely on boiler plate code that is either generated by extensive and invasive use of macros or through a code generator. Reflection and code injection are fundamental to the modernization and betterment of these frameworks, and this is fundamentally a language feature that should take priority over purely library work.

Let grow the Graphics proposal on its own

Maybe we do need another graphics framework. Who am I to say otherwise ? But existing framework have been battle tested for 20 years. I think the 2D graphics could thrive and grow as an independent or boost library over the next few years. Most importantly, it could provide a single implementation working on a wide variety of platforms rather than having 5 implementations or more of the same thing.

It would be free to experiment with text rendering, inputs, events, back-end, threading models…

I feel that this proposal as well as the package management issue calls for something that is authoritative without being ISO and I have no idea what that could or should be.

In the mean time, Visual Studio and Xcode could ship with more third party libraries and that would solve at least half of the issues that this proposal is trying to solve.

Share on