Thriving in ACrowded and Changing World

Thriving in a Crowded and Changing World: C++ 2006–2020
BJARNE STROUSTRUP, Morgan Stanley and Columbia University, USA

Shepherd: Yannis Smaragdakis, University of Athens, Greece
By 2006, C++ had been in widespread industrial use for 20 years. It contained parts that had survived unchanged
since introduced into C in the early 1970s as well as features that were novel in the early 2000s. From 2006 to
2020, the C++ developer community grew from about 3 million to about 4.5 million. It was a period where
new programming models emerged, hardware architectures evolved, new application domains gained massive
importance, and quite a few well-financed and professionally marketed languages fought for dominance. How
did C++ ś an older language without serious commercial backing ś manage to thrive in the face of all that?
This paper focuses on the major changes to the ISO C++ standard for the 2011, 2014, 2017, and 2020 revisions.
The standard library is about 3/4 of the C++20 standard, but this paper’s primary focus is on language features
and the programming techniques they support.
The paper contains long lists of features documenting the growth of C++. Significant technical points are
discussed and illustrated with short code fragments. In addition, it presents some failed proposals and the
discussions that led to their failure. It offers a perspective on the bewildering flow of facts and features across
the years. The emphasis is on the ideas, people, and processes that shaped the language.
Themes include efforts to preserve the essence of C++ through evolutionary changes, to simplify its use, to
improve support for generic programming, to better support compile-time programming, to extend support
for concurrency and parallel programming, and to maintain stable support for decades’ old code.
The ISO C++ standard evolves through a consensus process. Inevitably, there is competition among proposals
and clashes (usually polite ones) over direction, design philosophies, and principles. The committee is now
larger and more active than ever, with as many as 250 people turning up to week-long meetings three times
a year and many more taking part electronically. We try (not always successfully) to mitigate the effects of
design by committee, bureaucratic paralysis, and excessive enthusiasm for a variety of language fashions.
Specific language-technical topics include the memory model, concurrency and parallelism, compile-time
computation, move-semantics, exceptions, lambda expressions, and modules. Designing a mechanism for
specifying a template’s requirements on its arguments that is sufficiently flexible and precise yet doesn’t
impose run-time costs turned out to be hard. The repeated attempts to design łconceptsž to do that have their
roots back in the 1980s and touch upon many key design issues for C++ and for generic programming.
The description is based on personal participation in the key events and design decisions, backed by the
thousands of papers and hundreds of meeting minutes in the ISO C++ standards committee’s archives.
CCS Concepts: · Software and its engineering → Software notations and tools; General programming
languages;
Additional Key Words and Phrases: C++, programming language design and evolution, standardization, generic 70
programming, resource management, concurrency and parallelism, simplification of language use
ACM Reference Format:
Bjarne Stroustrup. 2020. Thriving in a Crowded and Changing World: C++ 2006ś2020. Proc. ACM Program.
Lang. 4, HOPL, Article 70 (June 2020), 167 pages. https://doi.org/10.1145/3386320
Author’s address: Bjarne Stroustrup, Morgan Stanley and Columbia University, USA, [email protected].
Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee
provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and
the full citation on the first page. Copyrights for third-party components of this work must be honored. For all other uses,
contact
This worktheisowner/author(s).
licensed under a Creative Commons Attribution-NoDerivatives 4.0 International License.
© 2020 Copyright held by the owner/author(s).
2475-1421/2020/6-ART70
https://doi.org/10.1145/3386320
Proc. ACM Program. Lang., Vol. 4, No. HOPL, Article 70. Publication date: June 2020.
70:2 Bjarne Stroustrup
Contents
Abstract 1
Contents 2
1 Introduction 5
1.1 Chronology 6
1.2 Overview 7
2 Background: C++ 1979ś2006 7
2.1 The First Decade 7
2.2 The Second Decade 10
2.2.1 Language Features 10
2.2.2 Standard-Library Components 12
2.3 C++ in 2006 13
2.4 Other Languages 16
3 The C++ Standards Committee 17
3.1 The Standard 18
3.2 Organization 18
3.3 Impact on Design 21
3.4 Proposal Checklists 28
4 C++11: It Feels Like a New Language 29
4.1 C++11: Support for Concurrency 31
4.1.1 Memory Model 31
4.1.2 Threads and Locks 33
4.1.3 Futures 35
4.2 C++11: Simplifying Use 37
4.2.1 auto and decltype 38
4.2.2 Range-for 41
4.2.3 Move Semantics 42
4.2.4 Resource-Management Pointers 44
4.2.5 Uniform Initialization 45
4.2.6 nullptr 48
4.2.7 constexpr Functions 48
4.2.8 User-Defined Literals 50
4.2.9 Raw literals 51
4.2.10 Attributes 51
4.2.11 Garbage Collection 52
4.3 C++11: Improving Support for Generic Programming 52
4.3.1 Lambda 53
4.3.2 Variadic Templates 55
4.3.3 Aliases 56
4.3.4 tuples 58
4.4 C++11: Increase Static Type Safety 59
4.5 C++11: Support for Library Building 60
4.5.1 Implementation Techniques 60
4.5.2 Metaprogramming Support 61
4.5.3 noexcept Specifications 61
4.6 C++11: Standard-Library Components 62
5 C++14: Completing C++11 63
Thriving in a Crowded and Changing World: C++ 2006ś2020 70:3
5.1 Digit Separators 64

5.2 Variable Templates 65
5.3 Function Return Type Deduction 65
5.4 Generic Lambdas 65
5.5 Local Variables in constexpr Functions 66
6 Concepts 67
6.1 The Prehistory of Concepts 68
6.2 C++0x Concepts 69
6.2.1 Concept Definitions 70
6.2.2 Concept Use 72
6.2.3 Concept Maps 72
6.2.4 Definition Checking 73
6.2.5 Lessons Learned 73
6.2.6 What Went Wrong? 75
6.3 The Concepts TS 77
6.3.1 Definition Checking 78
6.3.2 Concept Use 79
6.3.3 Concept Definition 81
6.3.4 Concept Name Introducers 82
6.3.5 Concepts and Types 82
6.3.6 Improvements 84
6.3.7 Syntax Equivalences 84
6.3.8 Why No Concepts in C++17? 86
6.4 C++20 Concepts 86
6.5 Naming of Concepts 88
7 Error Handling 89
7.1 Background 89
7.2 Real-World Problems 90
7.3 noexcept Specifications 92
7.4 Type System Support 92
7.5 Back to Basics 93
8 C++17: Lost at Sea 95
8.1 Constructor Template Argument Deduction 96
8.2 Structured Bindings 96
8.3 variant, optional, and any 99
8.4 Concurrency 100
8.5 Parallel STL 101
8.6 File System 102
8.7 Explicit Tests in Conditions 103
8.8 Proposals That Didn’t Make C++17 104
8.8.1 Networking 104
8.8.2 Operator Dot 104
8.8.3 Uniform Call Syntax 105
8.8.4 Default Comparisons 106
9 C++20: A Struggle for Direction 109
9.1 Design Principles 109
9.2 My C++17 List 110
9.3 C++20 Features 111
9.3.1 Modules 111

9.3.2 Coroutines 114
9.3.3 Compile-Time Computation Support 117
9.3.4 <=> 118
9.3.5 Ranges 119
9.3.6 Dates and Time Zones 120
9.3.7 Format 120
9.3.8 Span 121
9.4 Concurrency 123
9.5 Minor Features 123
9.6 Work in Progress 124
9.6.1 Contracts 124
9.6.2 Static Reflection 127
10 C++ in 2020 128
10.1 What Is C++ Used For? 129
10.2 The C++ Community 130
10.3 Education and Research 131
10.4 Tools 132
10.5 Programming Styles 133
10.5.1 Generic Programming 134
10.5.2 Metaprogramming 134
10.6 Coding Guidelines 136
10.6.1 General Approach 136
10.6.2 Static Analysis 137
11 Retrospective 139
11.1 The C++ Model 139
11.2 Technical Successes 139
11.3 Areas That Need Work 140
11.4 Lessons Learned 141
11.5 The Future 143
Acknowledgments 144
References 145
1 INTRODUCTION
Originally, I designed C++ to answer to the question łHow do you directly manipulate hardware
and also support efficient high-level abstraction?ž Over the years, C++ has grown from a relatively
simple solution based on a combination of facilities from the C and Simula languages aimed
at systems programming on 1980s computers to a far more complex and effective tool for an
extraordinary range of applications. It retains its dual focus on:
• Direct mapping of language constructs to hardware facilities
• Zero-overhead abstraction
This combination is the defining characteristic that sets C++ apart from most languages.
łZero overheadž was explained like this [Stroustrup 1994]:
• What you don’t use, you don’t pay for (aka łno distributed fatž).
• What you do use, you couldn’t hand-code any better.
Abstractions are represented in code as functions, classes, templates, concepts, and aliases.
C++ is a living language, so it changes to meet new challenges and the styles of use evolve. These
challenges and changes in the 2006-to-2020 time-frame are the focus of this paper. Of course, a
language doesn’t change by itself; it is changed by people. So this is also the story of the people
involved in the evolution of C++, the way they perceived the challenges, interpreted the constraints
on solutions, organized their work, and resolved their inevitable differences. When presenting a
language or standard-library feature, I do so in the context of the general evolution of C++ and
the concerns of the individuals involved at the time. For many features accepted early in the time
period, we now have the benefit of hindsight from massive industrial use.
C++ is primarily an industrial language, a tool for building systems. For a user, łC++ž is not just
a language as defined by a specification; it is part of a tool set with many parts:
• The language
• The standard library
• Many other libraries
• Massive ś often old ś code bases
• Tools (including other languages)
• Teaching and training
• Community support
Where possible and relevant, I will consider the interactions among those łparts.ž
There is a myth, a very popular myth, that programmers want their languages to be simple.
That’s obviously the case when you have to learn a new language, have to design a programming
course or curriculum, or describe a language in an academic paper. For such uses, having a language
cleanly embody a few clear principles is an obvious advantage and the ideal. When the focus shifts
from learning to delivering and maintaining significant applications, the demands from developers
shift from simplicity to comprehensive support, stability (compatibility), and familiarity. People
invariably confuse familiarity with simplicity and prefer familiarity over simplicity if given a choice.
One way of looking at C++ is as the result of decades of three contradictory demands:
• Make the language simpler!
• Add these two essential features now!!
• Don’t break (any of) my code!!!
I added the exclamation marks because these points are often delivered with a fair bit of emotion.
I wanted to make simple things simple and to ensure that complex things are not impossible or
unnecessarily hard. The former is essential for developers who are not language lawyers; the later
for implementers of foundational code. Stability is an essential property for all systems meant to
last for decades, yet a living language must adapt to a changing world.
There are overarching ideals for C++. I articulate some (e.g., The Design and Evolution of C++
[Stroustrup 1994] (ğ2), design principles (ğ9.1), and the C++ model (ğ11.1)) and try to make the
evolution of the language follow them. However, long lists of new features and very detailed
practical concerns are the main focus of C++’s development as controlled by its ISO standards
committee. That’s what the most vocal and influential people in the community insist on, and it
would be foolhardy to deem their concerns and opinions wrong solely based on philosophical or
theoretical views.
1.1 Chronology
To give a quick overview, here is a rough chronology. If you are not familiar with C++, many of the
terms, construct, and libraries will be obscure; most are explained in length in the previous HOPL
papers [Stroustrup 1993, 2007] or in this paper.
• 1979: Start of work on łC with Classesž that became C++; first non-research user;
ś Language: classes, constructors/destructors, public/private, simple inheritance, function
argument type checking
ś Library: tasks (coroutines and simulation support), vector parameterized with macros
• 1985: First commercial release of C++; TC++PL1 [Stroustrup 1985b]
ś Language: virtual functions, operator overloading, references, const
ś Library: complex arithmetic, stream I/O
• 1989-91: ANSI and ISO standardization start; TC++PL2 [Stroustrup 1991]
ś Language: abstract classes, multiple inheritance, exceptions, templates
ś Library: iostreams (but no tasks)
• 1998: C++98, the first ISO C++ standard [Koenig 1998], TC++PL3 [Stroustrup 1997]
ś Language: namespaces, named casts, bool, dynamic_cast
ś Library: the STL (containers and algorithms), string, bitset
• 2011: C++11 [Becker 2011], TC++PL4 [Stroustrup 2013]
ś Language: memory model, auto, range-for, constexpr, lambdas, user-defined literals, ...
ś Library: threads and locks, future, unique_ptr, shared_ptr, array, time and clocks,
random numbers, unordered containers (hash tables), ...
• 2014: C++14 [du Toit 2014]
ś Language: generic lambdas, local variables in constexpr functions, digit separators, ...
ś Library: user-defined literals, ...
• 2017: C++17 [Smith 2017]
ś Language: structured bindings, variable templates, template argument deduction from
constructors, ...
ś Library: file system, scoped_lock, shared_mutex (reader-writer locks), any, variant,
optional, string_view, parallel algorithms, ...
• 2020: C++20 [Smith 2020]
ś Language: concepts, modules, coroutines, three-way comparisons, improved support for
compile-time computation, ...
ś Library: concepts, ranges, dates and time zones, span, formats, improved concurrency and
parallelism support, ...
Note the poverty of libraries in the early years. There were, in fact, many libraries (including
GUI libraries), but few were widely used and many were proprietary. This was before open-
source development became widespread. This left the C++ community without a significant shared
foundation library. In the retrospective of my HOPL2 paper [Stroustrup 1993], I deemed that the
worst mistake of early C++.
The task library [Stroustrup 1985a,c] was a coroutine-based library with some support for event
driven simulations (e.g., random number generation) that was very efficient when compared to
alternatives, even on tiny computers. For example, I ran simulations with 700 tasks in a 256KB
memory. The task library was immensely important in C++’s early years as the base of many
important applications in Bell Labs and elsewhere. However, it was a bit ugly and couldn’t easily
be ported to Sun’s SPARC architecture so it wasn’t supported by most post-1989 implementations.
In 2020, coroutines are only just coming back (ğ9.3.2).
Basically, the feature set grows contiguously. The ISO committee deprecated a few features
in attempts to clean up the language but given the massive use of C++ (łmany billions of lines
of codež), nothing significant ever goes away. Stability is a key feature. One way of addressing
problems related to the growing size and complexity is through coding guidelines (ğ10.6).
1.2 Overview
The paper is organized in rough chronological order around the sequence of ISO standard releases.
• ğ1: Introduction
• ğ2: Background: C++ 1979-2006
• ğ3: The C++ standards committee
• ğ4: C++11: It feels like a new language
• ğ5: C++14: Completing C++11
• ğ6: Concepts
• ğ7: Error handling
• ğ8: C++17: Lost at sea
• ğ9: C++20: A struggle for direction
• ğ10: C++ in 2020
• ğ11: Retrospective
Where a topic, such as łconceptsž and the standards process, spans a longer period of time, I
cover it in one place, giving the contents priority over the chronological format.
This paper is extraordinarily long, a monograph really. However, from 2006 to 2020, C++ went
through two major revisions, C++11 and C++20, and early readers all requested more information;
that led to almost doubling the page count. Even at the current size, readers will find important
topics, such as concurrency and the standard library, underrepresented.
2 BACKGROUND: C++ 1979–2006

The history of C++ from 1979 to 2006 is documented in my HOPL papers [Stroustrup 1993, 2007].
In that period, C++ grew from a one-person research project to a community of about 3 million
programmers.
2.1 The First Decade

The work on what became C++ started in April 1979 under the name of C with Classes. I wanted a
tool that combined the ability to deal directly and efficiently with hardware (e.g., to write memory
managers, process schedulers, and device drivers) with Simula-like facilities for organizing code
(e.g., łstrongž static extensible type checking, classes, class hierarchies, and coroutines). I wanted
that tool to write a version of the Unix kernel that could work on multiple processors connected
through a local area network or a shared memory.
I chose C as a base for my work because it was good enough and very well supported locally:
my office was just across the corridor from Dennis Ritchie’s and Brian Kernighan’s. However, C
wasn’t the only language I considered. I was very attracted by Algol68 and was at the time quite
expert at BCPL and a few other machine-level languages. The later success of C was at that time
nowhere near certain, but Brian Kernighan and Dennis Ritchie’s superb introduction and manual
[Kernighan and Ritchie 1978] had just appeared and Unix was starting its victory run.
The initial implementation was a preprocessor that translated łC with Classesž into C more-
or-less line-for-line. In 1982, that approach proved unmanageable as the łC with Classesž user
population grew to several dozen people. So I wrote a conventional compiler, called Cfront, that
was first used by others in October 1983. Cfront was a traditional compiler in that it had a lexical
analyzer, a syntax analyzer that built an abstract syntax tree, a type checker that decorated that tree
with types, and a high-level optimizer that rearranged the AST to improve the run-time efficiency
of the generated code. There has been a lot of confusion about the nature of Cfront because it then
finally output C (optimized and not particularly human-readable C). I generated C so that I did not
have to directly deal with the myriad of (non-standardized) linkers and optimizers then in current
use. Cfront was nothing like a traditional preprocessor, though. You can find a Cfront source with
documentation in the Computer History Museum’s source code collection [McJones 2007ś2020].
Cfront was bootstrapped to C++ from C with Classes, so the first C++ compiler was written in
(simple) C++ for tiny computers (less than 1MB of memory and less than 1MHz of processor speed).
The first feature added to C for łC with Classesž was classes. I knew their power from earlier
use in Simula, where they were key to the strict static, but extensible type system. I immediately
added constructors and destructors. They were novel, but from my machine architecture and
operating systems background, I considered it obvious that I needed a mechanism to set up a
working environment (a constructor) and an inverse operation to release resources acquired while
running (a destructor). From my 1979 lab book:
• A łnew functionž creates the run-time environment for member functions

• A łdelete functionž reverses that
The terms łnew functionž and łdelete functionž were the original terms for łconstructorž and
łdestructor.ž To this day, I consider constructor/destructor pairs the real heart of C++. See also
(ğ2.2.1) and (ğ10.6).
At the time, essentially every language except C had proper function-argument type checking.
I didn’t think I could do anything significant without that. So, with the encouragement of my
department head, Alexander Fraser, I immediately added (optional) function argument declarations
and argument checking. That’s what in C is now called function prototypes. In 1982, after seeing
the effects of leaving the function argument checking optional, I made it compulsory. That caused
a decade or two’s loud howls of complaints about incompatibility with C. People wanted their type
errors, or at least many loudly said they didn’t want checking and used that as an excuse for not
using C++. This factoid may give people an idea of the problems involved in evolving a language
in significant use.
Given the occasional nasty words exchanged between overly parochial C and C++ aficionados, it
may be worth pointing out that I was always friends with Dennis Ritchie and Brian Kernighan,
eating lunch with them most days for 16 years. I learned a lot from them and still see Brian regularly.
I credit both with contributions to C++ [Stroustrup 1993] and I am a major contributor to C myself
(e.g., the function definition syntax, function prototypes, const, and //-comments).
To be able to think rationally about the growth of C++, I devised a set of design rules. These are
featured in [Stroustrup 1993, 1994], so here I will just mention a small sample:
• Don’t get involved in a sterile quest for perfection.
• Always provide a transition path.
• Say what you mean (i.e., enable direct expression of higher-level ideas) .
• No implicit violations of the static type system.
• Provide as good support for user-defined types as for built-in types.
• Preprocessor usage should be eliminated.
• Leave no room for a lower-level language below C++ (except assembler).
These were not unambitious goals. Some, I am still working on in 2020. In the early-to-mid-1980s,
I added more language facilities to C++:
• 1981: const ś to support immutability in interfaces and symbolic constants.
• 1982: virtual functions ś to offer run-time polymorphism.
• 1984: References ś to support operator overloading and simplify argument passing.
• 1984: Operator and function overloading ś including allowing the user to define = (assign-
ment), () (application; enabling łfunction objectsž (ğ4.3.1)), [] (subscripting), and -> (smart
pointers) in addition to the arithmetic and logical operators.
• 1987: Type-safe linkage ś to eliminate many errors coming from inconsistent declarations in
separate translation units.
• 1987: abstract classes - to offer pure interfaces.
In the late 1980s, as the power of computers increased dramatically, I got more interested in
larger-scale software and added
• Templates ś to better support generic programming after years of suffering with writing
generic programming using macros.
• Exceptions ś to try to bring some order to the chaos of error-handling; RAII (ğ2.2.1) was
articulated for that design.
These later facilities were not universally well received (e.g., see (ğ7)). Part of the reason was
that the community had grown large and unmanageable. The ANSI standardization had started so
I was no longer able to implement and experiment in private. People insisted on large elaborate
designs and on debating them extensively before serious implementation. I could no longer start
with a minimal proposal and grow it into a more complete facility while knowing that it isn’t
possible to please everybody. For example, people insisted on the łheavyž template syntax with the
template<class T> prefix everywhere.
In the late 1980s, the łobject-orientedž hype became deafening and stole the message of C++
from me. My opinion of what C++ was and was meant to become was widely ignored ś many
never heard it. All new languages were to be łpure object-oriented,ž for some definition of łobject
oriented.ž Not being łtruly OOž was deemed bad without the need for argument.
The fact that I never used the phrase łC++ is an object-oriented programming languagež was
not known or ignored as a bit embarrassing. At the time, my standard description was
C++ is a general-purpose programming language with a bias towards systems program-
ming that
• is a better C
• supports data abstraction
• supports object-oriented programming
• supports generic programming
This was (and is) accurate, but not as exciting as slogans such as łEverything is an object!ž
2.2 The Second Decade

The ANSI C++ committee was founded at a meeting in Washington D.C. in December of 1989,
just over 10 years after the first beginnings of łC with Classes.ž About 25 C++ programmers were
present. I was there, as were a small handful ISO C++ standard committee members who were still
active in the 2010s.
The committee delivered its first standard, C++98, after the conventional about a decade’s work.
Naturally, I ś and many others ś would have preferred a standard sooner, but committee rules,
over-ambition, and various delays brought us into line with the schedules of Fortran, C, and other
formally standardized languages.
The work that led to C++98 is the core of the HOPL3 paper [Stroustrup 2007], so here I only
briefly summarize.
2.2.1 Language Features. The major language features of C++98 were:

• Templates ś unconstrained, Turing-complete, compile-time support for generic programming
following up on my early work (ğ2.1) with many elaborations and refinements; that work
continues (ğ6).
• Exceptions ś a mechanism for returning error-values on a separate (łinvisiblež) path to be
handled by code łelsewherež up the stack of callers; see (ğ7).
• dynamic_cast and typeid ś a very simple form of run-time reflection (łRun-time Type
Identificationž aka RTTI).
• namespaces ś allowing programmers to avoid name clashes when composing larger pro-
grams out of separate parts.
• Declarations in conditions ś tightening up notation and limiting scope of variables.
• Named casts (static_cast, reinterpret_cast, and const_cast) ś eliminating ambiguities from
C-style casts and making explicit type conversion far more visible.
• bool ś a Boolean type that proved surprisingly useful and popular; C and C++ had used
integers as Boolean variables and constants.
Consider a simple C++98 example. The dynamic_cast is C++’s version of what is often called
something like łisKindOfž in object-oriented languages:
void do_something ( Shape * p)

{
if ( Circle * pc = dynamic_cast < Circle * >( p )) { // is p a kind of Circle ?
// ... use the Circle pointed to by pc ...
}
else {
// ... it wasn 't a Circle , do something else ...
}
}
The dynamic_cast is a run-time operation relying on data stored in the Shape’s virtual function
table. It is general, easy to use, and about as efficient as equivalent facilities in other languages.
However, dynamic_cast became quite unpopular because its implementations tended to be compli-
cated and special cases can be hand-code more efficiently (so that dynamic_cast arguably violated
the zero-overhead principle). The use of declarations in conditions was novel, though at the time, I
thought that I had just copied the idea from Algol68.
A simpler variant uses references instead of pointers:

void do_something2 ( Shape & r)
{
Circle & rc = dynamic_cast < Circle & >( r ); // r is a kind of Circle !
// ... use the Circle referred to by rc ...
}
This simply asserts that r refers to a Circle and throws an exception if it does not. The idea was
to use pointers and tests if the łerrorž could reasonably be handled locally and to rely on references
and exceptions if not.
One of the most important techniques in C++98 was RAII (Resource Acquisition Is initialization).
That was my clumsy name for the idea that every resource should have an owner represented by
a scoped object: A constructor acquires the resource and a destructor implicitly releases it. This
idea was present in the earliest C with Classes (ğ2), but not named until a decade later. Here is an
example I often used to illustrate the idea that not every resource is memory:
void my_fct ( const char * name ) // C - style resource management
{
FILE * p = fopen ( name ," r "); // open File ' name ' for reading
// ... use p ...
fclose (p );
}
The problem here is that if (between the calls of fopen() and fclose()) we return from the
function, throw an exception, or use C’s longjmp, the file handle pointed to by p is leaked.
Leaking file handles exhausts an operating system even faster than memory leaks. That file handle
is an example of a non-memory resource.
The solution is to represent the file handle as a class with a constructor and a destructor:
class File_handle {
FILE * p;
public :
File_handle ( const char * name , const char * permissions ); // open file
~ File_handle (); // close file
// ...
};
We can now simplify our use:

void my_fct2 ( const char * name ) // RAII - style resource management
{
File_handle p( name ," r "); // open File 'name ' for reading
// ... use p ...
} // p is implicitly closed
With the introduction of exceptions, such resource handles became pervasive. In particular, the
standard-library file stream is such a resource handle, so using the C++98 standard library, this
example becomes:
void my_fct3 ( const string & name )
{
ifstream p( name ); // open File 'name ' for reading
// ... use p ...
} // p is implicitly closed
Note that the RAII code differs from the traditional use of functions by allowing the łcleanupž
to be defined once-and-for-all in the library, rather than having to be remembered and explicitly
written by the programmer for each use of a resource. It is critical that the correct and robust code
is simpler, shorter, and at least as efficient as the conventional style. Over the next 20 years, RAII
permeated C++ libraries.
The implication of having non-memory resources is that garbage collection isn’t by itself sufficient
for resource management. In addition, RAII plus smart pointers (ğ4.2.4) eliminate much of the need
for Garbage collection. See also (ğ10.6).
2.2.2 Standard-Library Components. The C++98 standard-library provided:
• The STL ś the innovative, general, elegant, and efficient framework of containers, iterators,
and algorithms by Alexander Stepanov.
• Traits ś sets of compile time properties useful for programming with templates (ğ4.5.1).
• string ś a type for holding and manipulating a sequence of characters. The character type is
a template parameter defaulted to char.
• iostreams ś an elaboration by Jerry Schwartz and the standards committee of my simple 1984
streams library to handle a wide variety of character types, locales, and buffering strategies.
• bitset ś a type for holding and manipulating sets of bits.
• locales ś an elaborate framework of cultural conventions, mostly related to I/O.
• valarray ś a numeric array with optimizable vector operations that unfortunately didn’t see
much use.
• auto_ptr ś an early pointer representing exclusive ownership; in C++11, it was replaced by
shared_ptr (for shared ownership) and unique_ptr (for exclusive ownership) (ğ4.2.4).
The STL framework was by far the most important standard-library component. I think it fair
to say that it ś and the generic programming techniques it pioneered ś saved C++ as a living
modern language. Like all the C++98 facilities, the STL has been extensively described elsewhere
(e.g., [Stroustrup 1997, 2007]) so here I will present just a single, short example:
void test ( vector < string >& v , list < int >& lst )
{
vector < string >:: iterator p
= find_if (v. begin () , v. end () , Less_than < string >(" falcon "));
if (p != v. end ()) { // p points to ' falcon '
// ... use *p ...
}
else { // ' falcon ' not found
// ...
}
vector < int >:: iterator q

= find_if ( lst . begin () , lst . end () , Greater_than < int >(42));
// ...
}
The standard-library algorithm find_if traverses a sequence (delimited by a begin/end pair)

looking for an element for which a predicate is true. The algorithm is generic in three dimensions:
• The way elements of the sequence are stored (here, vector and list)
• The type of elements (here, string and int)
• The predicate used to determine when an element is found (here, Less_than and Greater_than)
Note the absence of object-oriented techniques. This is generic programming relying on templates,
sometimes referred to as compile-time polymorphism.
The notation was still primitive, but from about 2017, I could use auto (ğ4.2.1), ranges (ğ9.3.5),
and lambdas (ğ4.3.1) to simplify that code:
void test2 ( vector < string >& v , list < int >& lst )
{
if ( auto p = find_if (v ,[]( const string & s) { return s <" falcon "; })) {
// ...
}
// ...
if ( auto q = find_if ( lst ,[]( int x) { return x >42; })) {
// ...
}
// ...
}
2.3 C++ in 2006

In 2006, I and most other members of the ISO C++ committee had high hopes for a feature-rich C++0x
standard. A feature freeze was planned for 2007, so we had a reasonable expectation that C++0x
would be C++08 or C++09. In fact, C++0x became C++11, causing jokes about the hexadecimal
C++0xB.
In my 2006 HOPL paper [Stroustrup 2007], I listed 39 proposals and predicted that the first
21 would make it into C++0x. Interestingly, 24 of the first 25 proposals on my list made it into
C++11. Proposals 22-25, I listed as łbeing developed aiming for votes in 2007.ž To my surprise ś and
immense pleasure ś they all made it. None of proposals 26-39 even made it into C++17. That leaves
proposal 10, łconceptsž, which has its own long sad story with a happy ending in C++20 (ğ6).
I and many others were frustrated by the delays of C++0x and feared that an unimproved C++
might not survive as a living language in the face of competition from more modern and better
financed alternatives. In 2006, Java use was still increasing and Microsoft’s C# was heavily supported
and marketed. My estimate in 2006 was that C++ use had ś for the first time ś declined slightly
over the previous 4 years. It is hard to obtain real numbers and my best estimate (a 7% decline) is
well within the error margins, but there was certainly reason to worry. Languages, such as Java and
C#, were based on the assumption ś often loudly proclaimed ś that C++ didn’t have an ecological
niche:
• łLow-level stuffž could be handled by a small amount of C or assembler.
• łHigh-level stuffž was better, cheaper, and more efficiently done in a safer, smaller, garbage-
collected language with a huge run-time support system.
• Managed languages, such as Java and C#, using garbage collection and consistent runtime
range checking, made less-expert programmers more productive and highly-skilled developers
less needed.
• Deep integration of a programming language into a platform and supported by an integrated
tool set were essential for productivity and the construction of large systems.
Obviously, I and many others didn’t agree, but these were (and are) serious arguments, that
(if correct) should lead to the abandonment of C++. C++ is based on the traditional model of
a programming language separate from the underlying operating system and supported by a
multitude of independent tool suppliers. The managed languages tended to be proprietary; only a
large and rich organization could develop the massive infrastructure and libraries required. I and
many others in the C++ community prefer languages to be free of corporate control; this was one
reason I took part in the ISO standards effort.
In retrospect, 2006 might have been the nadir for C++, but important technological events
had just happened: For the first time in history, in about 2005, single-processor (single-core)
performance stopped improving and energy efficiency (łPerformance per watt’) became a key
measure (especially for server farms and hand-held devices). The economics of computing shifted
to favor better software. No longer could inefficiencies in languages or programming techniques
be completely hidden by hardware advances. Now, a highly skilled developer using a łsharp toolž
could (again, after a decade or so) gain an economic order-of-magnitude advantage over weaker
programmers or programmers hobbled by overheads in their tool chains. Even today, these facts
have not yet worked their way through all the educational and management systems, but there are
now many significant tasks for which spending time on carefully crafting performant code pays off
massively.
Another turning point came from vendors trying to impose their favorite language on all users by
defining standard interfaces to, say GUI, that could be met only by using their favored ś and often
proprietary ś languages. Examples were Google’s use of Java for Android, Apple’s Objective-C for
iOS, and Microsoft’s C# for Windows. Application vendors could try to dodge the lock-in by using
dialects, such as Objective C++ [Objective C++ Wikipedia 2020] or C++/CLI [ECMA International
2005], but the resulting code was still not portable. Many organizations, such as Adobe, Google,
and Microsoft, responded by writing the major parts of their demanding applications in C++ and
then using thin interface layers for the various platforms (e.g., Android, iOS, and Windows). In
2006, this trend was barely noticeable.
On portable devices (in particular, smartphones), the need for energy efficiency and platform
independence combined. One effect is that by my best estimates in 2018, the number of C++
programmers was up about 50% since 2006 to about 4.5 million developers [Kazakova 2015]. That’s
a 150,000 developers/year increase; about 4%/year for a decade.
In 2006, few people had spotted the significant hardware trends feeding into C++’s inherent
strengths. Instead, the community and the standards committee were focusing on novel language
features and libraries to increase C++’s usefulness and to raise some enthusiasm. Some committee
members, including me, felt an urgency and a dire need for significant improvements. Others
were more focused on stabilizing the language and improving its implementations. A standards
committee needs both groups, but the constant tug of war between innovation and retrenchment is
a source of tension. As in any large organization, there is an organizational advantage to people
defending status quo and serving current users. In The C++ Programming Language (3rd Edition)
[Stroustrup 1997], I cited Niccolò Machiavelli on that topic:
łthere is nothing more difficult to carry out, nor more doubtful of success, nor more
dangerous to handle, than to initiate a new order of things. For the reformer makes
enemies of all those who profit by the old order, and only lukewarm defenders in all those
who would profit by the new order.ž
My opinion was that C++ needed significant improvement to serve its community well. C++
applications were massively deployed, but new projects often chose more fashionable languages
and some successful C++ projects were being rewritten into such languages. For example, much of
Google’s large-scale applications, such as search, was (and is) based on their map-reduce frame-
work [Dean and Ghemawat 2004, 2008]. That’s a C++ program. However, as it is proprietary for
commercial reasons, people replicated it, but ś sadly for the C++ community ś the open-source
map-reduce framework (Hadoop) was for a variety of reasons done in Java.
Another significant reason for development moving to other languages is that the flexibility
of interfaces offered by templates makes it extremely difficult to provide a stable ABI using all
C++ features: you can be flexible or you can offer stable binary interfaces, but it is beyond most
organizations to do both. I consider this a contributing reason that people require C, Java, C#, etc.,
interfaces to programs written in C++. ABI stability for C++ is a genuinely hard technical problem,
especially as the C++ standard has to be platform independent.
To add to the problems of the C++ community, by 2006, most professional software magazines
covering C++ had died as publishing on paper was declining and journalists followed fashion and
advertising revenues. Dr. Dobbs Journal lasted another few years (stopped in print in February 2009).
The C++ conferences were being absorbed into łObject-orientedž or general software development
conferences, depriving the C++ community of venues for exposure of new developments. Books
were still being written, but programmers were reading fewer books (or at least buying far fewer as
pirating was becoming easier and consequently statistics were becoming less reliable) and on-line
sources were becoming more popular.
An even more serious problem was that C++’s role in education was sharply decreasing. C++ was
no longer łnew and interestingž and Java was being marketed directly to universities as an easier
and more powerful language. The US high-school Computer Science test suddenly changed from
C++ to Java. The use of Java as the introductory language in universities increased dramatically. The
quality of C++ teaching was also decreasing with most courses choosing a C-first approach or taking
the view that Object-Oriented programming relying heavily on class hierarchies was the one true
way. Both approaches minimized C++ strengths and required heavy use of macros. The standard
library (relying on generic programming; (ğ2.2)) and RAII (relying on constructor/destructor pairs
(ğ2.2.1)) were often completely left out of foundational courses or delegated to an ładvanced
featuresž section that most students either never reached or considered scary. Textbooks often
bogged down in obscure details. There were exceptions, of course, but on average the C++ presented
to students was far inferior to the best industrial practices. In 2005, I accepted the challenge to
teach programming to first-year university students. I surveyed about twenty of the most popular
C++ programming textbooks and ended up loudly complaining:
łIf that’s C++, I don’t like it either!ž
After teaching a year using a well-reputed textbook, I changed to using just my own notes, and in
2008 published Programming: Principles and Practice using C++ [Stroustrup 2008a] but to this day,
much C++ teaching has a 1980s flavor.
Despite this, C++ usage was starting to increase again. I think the reason was the fundamental
technology trends again favored C++ and towards the end of the decade the beginnings of C++11
were helping.
The Boost libraries and the Boost organization were important [Boost 1998ś2020]. In 1998, Beman
Dawes, an experienced developer and influential member of WG21, had started a łC++ library
repository web sitež [Dawes 1998] with the explicit aim of developing C++ libraries to establish
existing practice that future standardization could build upon. Before that, C++ had never even
had a common repository for libraries. Boost slowly grew into an active organization with peer
review of new libraries and a yearly conference. The Boost libraries became very widely used and
the most popular were absorbed into the standard (e.g., regex (ğ4.6), thread (ğ4.1.2), shared_ptr
(ğ4.6), variant (ğ8.3), and file system (ğ8.6)). It was important for the C++ community that the
Boost libraries were available more than a decade earlier than their ISO standard versions but were
trusted as a kind of łjunior standard.ž Many committee members, notably, Dave Abrahams, Doug
Gregor, Jaakko Järvi, Andrew Sutton, and of course Beman Dawes, were involved with Boost.
By 2006, C++ was no longer new and exciting in industry, but spread over many industries.
Usage was ś and still is ś strong in the telecommunications industry where C++ was born. From
there, it had spread into the games (e.g., Unreal, PlayStation, Xbox, and Douglas Adams’ łSpaceship
Titanicž), finance (e.g., Morgan Stanley and Renaissance Technologies), microelectronics (e.g., Intel
and Mentor Graphics), movies (e.g., Pixar and Maya), aerospace (e.g., Lockheed-Martin and NASA),
and many other industries.
Personally, I was particularly fond of C++’s widespread use in science and engineering, such as
High Energy Physics (e.g., CERN, SLAC, FermiLab), biology (e.g., the human genome project), space
exploration (e.g., Mars Rovers and the deep space communication network), medicine and biology
(e.g., tomography, general imaging, the human genome project, and monitoring equipment), and
much more.
2.4 Other Languages

People often look for direct technical influences from other programming languages on C++. There
are few. Typically, influences bubble up through history from common ancestors and shared ideas.
Decisive arguments for extending C++ tend to relate to observed problems in the C++ community.
Direct borrowing from a fashionable language is rare and far more difficult than people imagine.
Most members of the standards committee master many languages and keep an eye out for useful
facilities, libraries, and techniques.
Consider some real and conjectured influences on C++ in the 2000s:
• auto ś the ability to deduce a type from an initializer. This is popular in modern languages,
but old as the mountains. I don’t really know where it came from, but I implemented it in
1983 and didn’t consider it novel then (ğ4.2.1).
• tuple ś many languages, especially from the functional programming tradition have tuples,
usually as a built-in type. The C++ standard-library tuple and many of its uses are inspired
from those. The std::tuple was derived from the boost::tuple [Boost 1998ś2020] (ğ4.3.4).
• regex ś The standard library regex added in C++11 was copied (via Boost with proper
acknowledgements) from facilities in Unix and JavaScript (ğ4.6).
• Functional programming ś there are many obvious similarities between FP features and C++
constructs. Most are not simple language features, but programming techniques. The STL
was inspired by functional programming and first tried (unsuccessfully) in Scheme [Stepanov
1986] and Ada [Musser and Stepanov 1987].
• future and promise ś from Multilisp via other Lisp dialects (ğ4.1.3).
• Range-for ś it has equivalents in many languages, but one direct inspiration was STL
sequences (ğ4.2.2).
• variant, any, and optional ś clearly inspired by a variety of languages (ğ8.3).
• Lambdas ś clearly the use of lambda expressions in functional languages was part of the
inspiration. However, in C++, the roots of lambdas also include blocks of code used as
expressions going back to BCPL, local functions (which had been repeatedly rejected for C and
C++ because they were seen as error-prone and adding complexity), and (most importantly)
function objects (ğ4.3.1).
• final and override ś for more explicit management of class hierarchies and present in many
object-oriented languages. They had been considered ś and considered unnecessary ś from
the earliest days of C++.
• The three-way comparison operator, <=>, was inspired by C’s strcmp and operators in
various languages, including PERL, PHP, Python, and Ruby (ğ9.3.4).
• await ś the original C++ coroutines (ğ1.1) were inspired by uses of Simula, but offered as
a library rather than as a language feature to leave room for several alternative ways of
expressing concurrency. The idea for the C++20 stackless coroutines came primarily from F#
(ğ9.3.2).
Even when a feature is borrowed from another language in a pretty direct manner, it mutates.
Typically, the syntax changes quite a bit to fit into C++. When borrowing from a garbage-collected
language, lifetime issues must be dealt with, and often the C++ distinction between objects and
references to objects needs to be addressed in ways that differ from the original. Often novel uses
are discovered during the łtranslationž into C++. The story of the introduction of łlambdasž into
C++ offers examples of most of these phenomena (ğ4.3.1).
Many people imagine that I (and others involved with C++) sit around all day strategizing in a
complicated war for dominance among popular languages. In fact, I spend no time on such. Most
days, I don’t think about other languages except when I happen to study one out of general technical
interest or use one to get some work done. What I do is to talk with software developers, consider
problems that people encounter using C++, and consider the flood of suggested improvements in
the standards committee. Of course, I also write code to experience problems and test out ideas
for improvements. The problem is to find time to calmly consider what’s fundamental, what’s just
fashion, and what would do harm.
Similarly, C++’s contributions to other languages are hard to pinpoint. Often, similar features
are parallel evolution or have common roots. Consider:
• Generics in Java and C# ś they modeled their generics from other languages but took the C++
syntax and added generics only after C++ demonstrated the utility of generic programming
on a large scale.
• The dispose idiom in Java, Python, etc. ś that is about the best you can do to approximate
destructors in a garbage-collected language.
• Compile time evaluation in the D programming language ś I explained the early constexpr
design to Walter Bright.
• C++’s model of object lifetime based on constructors and destructors was part of the inspira-
tion for Rust. Amusingly, these days, C++ is often accused of having borrowed such ideas
from Rust.
• C adopted the C++11 memory model, the function declaration and definition syntax, declara-
tions as statements, const, //-comments, inline, and initializers in for-loops.
Many differences between C++ and other languages stem from C++’s use of destructors. That
makes it hard for garbage-collected languages to borrow directly from C++.
3 THE C++ STANDARDS COMMITTEE

The international C++ standards committee, officially named ISO/IEC JTC1/SC22/WG21, is central
to the development of C++. This has been the case since its founding in 1991 and before that the
focus of C++ development was the ANSI C++ standards committee from 1989 [Stroustrup 1993].
C++ has no rich owner or other significant funding sources, so the community relies on corporate
development and open-source projects. WG21 and the National Standards committees are the only
venues where people from otherwise competing organizations can meet to jointly solve problems.
The members are all volunteers ś there is no paid secretariat ś even if many members do come as
representatives of the organizations they work for. At every meeting, there are people who proudly
claim to represent łself.ž That is, they are not sponsored and represent only themselves. It is not
uncommon for someone turning up to represent a new organization after a job change. There are
many examples of people making łattendance of the C++ standards committeež as a condition for
accepting a new job. People have joined the committee to learn about C++ and łmember of the C++
committeež has been quoted as a qualification (not always truthfully).
Some participate only for a couple of meetings or infrequently. On the other hand, there are
people who have been at most meetings over decades. In the beginning and currently, there are three
meetings a year. During a few years after the 1998 standard, we met only twice a year. Currently,
the face-to-face meetings are supplemented with several teleconferences and many, many emails
every day.
Here, I describe
• The role of the standard (ğ3.1)
• The organization of the committee (ğ3.2)
• The impact of the committee structure on the design of C++ (ğ3.3)
3.1 The Standard

The purpose of the standards committee is to write a standard. One official rationale for standards
is łfacilitating trade, particularly in reducing technical barriers and artificial obstacles to international
tradež and łproviding a framework for achieving economies, efficiencies and interoperability.ž A stan-
dard is a specification, not an implementation. Its purpose is to keep multiple implementations in
agreement, and to decide exactly what łagreementž means in a world where diverse underlying
hardware must be exploited effectively. Many programmers have a problem understanding that.
They consider their current compiler the definition of the language or have trouble understanding
why it is hard to get 100% agreement among many separate ś and typically competing ś orga-
nizations. In the 1990s, the committee considered formal specification but after consulting with
world-class experts concluded that neither specification technology nor the committee members
were up to a formal specification of C++. Obviously, the idea of a reference implementation was
considered, but the complexity of the language and especially issues related to hardware use and
optimization have defeated such ideas. It would be too complicated and too expensive. Alterna-
tively, it would be simplified to the point where it couldn’t help with the hardest problems where it
would be most needed. Also, a complicated reference implementation would be as likely to hide
surprises as N competing implementation teams documenting their decisions, running extensive
conformance tests, and discussing where they differ. For C++, N is at least four when it comes to
front-ends (Clang, EDG, GCC, and Microsoft) and at least a dozen for backends.
So, the standards committee is grappling with the problems of having many implementations.
The alternative would be to take the risks of a monoculture. If a single organization was the source
of C++ technology, everybody would get the same, for good and bad. An organization controlling
the łone true implementationž would have a dominant voice in the community and problems
there would affect all. In particular, funding problems, commercial concerns, political opinions, and
technical single-mindedness could seriously damage the language and its community.
For good and bad, the C++ community chose the semi-organized chaos of a large committee
plus multiple compiler, tools, and library suppliers over a unified ownership/dictatorship model.
3.2 Organization
For the work on C++17 and C++20, as many as 250 people turned up at each face-to-face WG21
meeting, out of a membership of about twice that. In addition, there are national standards commit-
tees and C++ standards interest groups supporting members in a dozen or more countries, including
Canada, Finland, France, Germany, Russia, Spain, the UK, and the USA. The members represent
more than a hundred organizations. To give an idea, here is a selection: Apple, Bloomberg, CERN,
Codeplay, EDG (Edison Design Group), Facebook, Google, IBM, Intel, Microsoft, Morgan Stanley,
Nvidia, Qt, Qualcom, Red Hat, Ripple, Sandia National Labs, University of Applied Sciences HSR
Rapperswil, and University of Carlos III, Madrid. There is a solid representation from compiler
suppliers, hardware suppliers, finance, games, library providers, platform suppliers, national labs
(physics), and more. The telecom presence that was prominent in early C++ has decreased, whereas
university presence, which used to be minimal, seems to be on the increase.
Obviously, such a large group of organizations and individuals, representing widely varying
interest and technical backgrounds, need an organizational structure to function. The meetings are
organized around working groups (WGs) and study groups (SGs). In the summer of 2019, we had:
• Core WG (CWG) ś writes the final standards text for the language ś chair, Michael Miller
(EDG).
• Library WG (LWG) ś writes the final standards text for the standard library ś chair, Marshall
Clow (The C++ Alliance, formerly Qualcom).
• Evolution WG (EWG) ś processes language proposals ś chair, Ville Voutilainen (Qt, formerly
Symbio).
• Library Evolution WG (LEWG) ś processes standard-library proposals ś chair, Titus Winters
(Google).
Study groups explore new areas and designs for possible standardization:
• SG1, Concurrency ś concurrency and parallelism topics ś chair, Olivier Giroux (Nvidia).
• SG5, Transactional Memory ś Exploring transactional memory constructs ś chair, Michael
Wong (Codeplay, formerly IBM).
• SG6, Numerics ś including but not limited to fixed point, decimal floating point, and fractions
ś chair, Lawrence Crowl ("self," formerly Google and Sun).
• SG7, Compile-time programming ś Initially focused on compile-time reflection, then expanded
to compile-time programming in general ś chair, Chandler Carruth (Google).
• SG12, Undefined behavior and Vulnerabilities ś a systematic review of vulnerabilities and
undefined/unspecified behavior ś chair, Gabriel Dos Reis (Microsoft, formerly Texas A&M
University).
• SG13, Human/Machine Interface and I/O ś selected low-level output (e.g., graphics, audio) and
input (e.g., keyboard, pointing) I/O primitives ś chair, Roger Orr (British Standards (BSI)).
• SG14, Game Development and Low Latency ś topics of interest to game developers and others
with low-latency requirements ś chair, Michael Wong (Codeplay, formerly IBM).
• SG15, Tooling ś topics related to creation of developer tools for standard C++, including but
not limited to modules and package management ś chair, Titus Winters (Google).
• SG16, Unicode ś topics related to Unicode text processing in C++ ś chair, Tom Honermann
(Synopsis).
• SG19, Machine Learning ś chair, Michael Wong (CodePlay, formerly IBM)
• SG20, Education ś looking for ways to support learners and teachers approach C++ as it is
today ś chair, Jan Christiaan van Winkel (Google)
• SG21, Contracts ś trying to design a contact system after the failure to do so for C++20 (ğ9.6.1)
ś chair John Spicer (EDG).
In 2017, a small group was established to address problems to do with the lack of direction in
the design of the language and standard library [Dawes et al. 2018]. The members of this Direction
Group (DG) are appointed by the convener in consultation with the WG chairs. Its members are
long-term contributors to the committee, language, and standard library. The initial members were
Beman Dawes, Howard Hinnant, Bjarne Stroustrup, David Vandevoorde, and Michael Wong; later,
Beeman retired and Roger Orr joined). The DG chairmanship is rotating, starting with me. The DG
is advisory and has the policy of only giving opinions when its members agree unanimously. It
maintains a document presenting its recommendations [Dawes et al. 2018; Hinnant et al. 2019]
The WGs persist over decades with slowly changing membership. The SGs come and go as interest
dictates and/or they complete their work and hand over proposals to WGs for final processing. For
example, four of the most significant SGs have declared victory and disbanded:
• SG2, Modules ś chair, Gabriel Dos Reis (Microsoft, formerly Texas A&M University).
• SG3, File system ś chair, Beman Dawes ("self").
• SG8, Concepts ś chair, Andrew Sutton (University of Akron, Ohio, formerly Texas A&M
University).
• SG9, Ranges ś updating the STL to use concepts, simplify notation, and provide infinite
sequences and pipelining ś chair, Eric Niebler (FaceBook).
SG4, Networking is dormant as its results are waiting to be merged into the standard (ğ8.8.1).
Another SG, SG11, databases, disbanded for lack of consensus and lack of critical mass of volunteers
to get the work done.
Some SGs produce Technical Specifications (TSs) that can be significant documents in the style
of the standard itself. They have some official (ISO) standing but don’t offer the long-term stability
of an international standard (IS). The Concurrency SG (SG1) has been active since 2006, headed
by Hans-J Boehm (Google, formerly HP Labs, formerly SGI) for most of its existence, and has a
standing very similar to the WGs.
In addition to these groups, there is a semi-official C/C++ liaison group consisting of people who
are members of both the C++ committee and the C committee (ISO/SC22/WG14). This group tries
to minimize C/C++ incompatibilities and the C++ standard documents every incompatibility. C
and C++ are far more compatible than they would have been without the constant effort of the
liaison group, but even then, most of the many features imported into C from C++ were modified
so that they introduced some incompatibility.
There are just three official officers required by and recognized by the ISO:
• Convener ś chairs the WG, sets the WG meeting schedule (łconvenesž meetings), appoints
Study Groups, and is responsible to higher levels of ISO (SC22, JTC1, and ITTF) for the WG’s
work ś Herb Sutter (Microsoft) who has held that position since 2002 with a break from
2008-2009 when PJ. Plauger (Dinkumware) held it.
• Project Editor ś ultimately responsible for applying committee-approved changes to the
standard’s working draft ś Richard Smith (Google); for C++11, Pete Becker (Dinkumware);
for C++14, Stephanus Du Toit (Intel).
• Secretary ś responsible for taking and distributing minutes of WG21 meetings ś Nina Ranns
(Edison Design Group, formerly Symantec).
The National standards committees have their own officers and procedures.
Obviously, the positions are held by various people over the years, but it is rare for someone to
hold a position for less than 5 years, despite the typically heavy workload. I was chair of the EWG
for 24 years before handing over to Ville Voutilainen in 2014.
In general, smaller proposals go directly to EWG and/or LEWG and larger proposals start out in
a SG. Proposals need to come in writing and be presented by someone. Typically, processing of
a significant proposal takes several meetings (often years) and requires several papers, revisions
of papers, and repeated presentations. Finally, a proposal that has gained strong support will be
presented to the committee as a whole for a final vote. The convener looks at the vote and makes the
call whether there is consensus. Consensus is not just a majority. The committee prefers unanimity
after the WG processing and votes, and if that’s not the case, a 9:1 or 8:2 majority is usually required.
The convener may very well deem a 8:2 majority łnot consensus.ž This happens if heads of national
standards bodies or several prominent members voice strong objections, so that there is a danger
that issues will linger or lead to patchy adoption.
A standards meeting is exhausting. Usually members talk shop from breakfast until midnight,
with formal sessions 8:30-12:30 and 14:00-17:30 plus evening sessions (19:00-22:00) most days.
Members who are preparing proposals work even longer hours. WG and SG chairs typically have
meetings over most mealtimes. Monday to Friday are full days, but if nothing surprising happens,
most members are done by about 15:00 on Saturday. Still, when meetings are in nice places, such as
Kona Hawaii, nobody outside the committee seems to be willing to believe that a meeting isn’t
some sort of vacation.
Voting in WGs and SGs is by one vote to everyone present. Formal voting in full committee is
one vote to every organization present (so that large organizations don’t get multiple votes) plus a
count of the national body positions. The łtechnical votež and the national body vote must agree
for a consensus to be established.
The history of the committee before 2006 is documented in [Stroustrup 1993, 1994, 2007]. The C++
Foundation (ğ10.2) keeps a reasonably up-to-date description of the organization, key individuals,
and processes of the committee on its website (isocpp.org/std).
An almost complete collection of committee papers from 1989 onwards is available [WG21
1989ś2020]. Currently that collection increases by more than 500 papers a year. In addition, many
of the committee’s discussions are on archived mailing lists (called łreflectorsž). There can be more
than a hundred messages a day. It is very hard to keep up with all that goes on in the committee,
especially as much require specialized technical knowledge to follow. I keep the collection of my
WG21 papers on my home pages [Stroustrup 1990ś2020].
Traditionally, ISO standards were revised every ten years or so. For example, we have C89, C99,
and C11. The problem with such a long revision cycle is that if a new feature misses a feature freeze,
we must wait another 12 years or so for it to become standard. Naturally, people then argue for
slipping the next standard by a year or two: łThis feature is so important that it can’t wait, so we
must delay the standard!ž That was how C++0x became C++11, 13 years after C++98.
After C++11, several members wanted a shorter cycle and Herb Sutter, the convener, suggested
we adopt a łtrain model.ž That is, łthe train leaves at its scheduled time and anyone not on board
will have to wait for the next scheduled departure.ž People liked that and there was a long discussion
about what would be the right interval between standards revisions. I argued for a short interval,
3 years, because anything longer (e.g., 5 years) would be vulnerable to the łthis feature is too
important to waitž argument for delay. We agreed on a three-year łrelease cyclež and Herb Sutter
added that we should adopt the łIntel tick-tockž model of alternating major and minor releases.
That too was agreed, so three years after C++11 (ğ4) we delivered C++14 (ğ5) adopting delayed
features and remedying minor problems discovered in early use. C++17 was also delivered on time,
but sadly not as a major upgrade (ğ8). C++20 was voted feature complete in February 2019 and the
final technical vote was done in Prague in February 2020.
3.3 Impact on Design

How do this organization of work, the elaborate decision processes, and the large number of
participants affect the development of C++? Looking at the size of the committee, its composition,
and its processes, I consider it amazing that anything constructive ever emerges. This is not just
łdesign by committee;ž it is łdesign by a confederation of committees.ž
In addition, the committee has only the weakest of management structures, lacking even the
most basic management tools:
• There are no qualifications (say, of education or of practical experience) required for member-
ship, speaking, or voting. Pay the ISO membership fee ($1280 for US members in 2018) and
attend two meetings and you are a full voting member. In SGs and WGs anyone can speak
and vote, even at their first meeting.
• There are no rewards beyond getting a proposal accepted and the satisfaction of seeing an
improved standard. That is a major motivator, though.
• There are no real ways to discourage disruptive behavior. All the unofficial committee
management can do is to politely encourage people not to do what others consider disruptive.
Members have different opinions about what is disruptive.
Before considering the problems of evolving a language in a large committee, please remember
that most of the time and work in the committee is done to resolve łminor problems;ž that is,
problems that don’t rise to the level of language design philosophy, academic publication, or
conference presentation. They are essential for keeping the language and its standard library from
fragmenting into dialects and for portability across compilers and platforms. These issues include
naming, name lookup, overload resolution, grammar details, exact meaning of constructs, lifetime
of temporaries, linkage, and much, much more. Many are tricky to resolve and poor resolutions can
have surprising and damaging consequences. Solutions tend to be crafted to minimize breakage
of existing code. The committee resolves hundreds of issues a year. I estimate that the fraction of
the committee members’ time and effort spent at this is at least a third and maybe as high as two
thirds. This work tends to be overlooked and underappreciated. If you have used a computer or a
computerized gadget (e.g., a phone or a car), you can thank the people in the CWG and LWG for it
working.
When focusing on problems caused by a huge committee, please also remember that these
problems are caused by an embarrassment of riches: The C++ standards process is driven by
hundreds of enthusiastic people from a wide variety of backgrounds, with a wealth of diverse
experience, and a massive dose of idealism.
The committee is a filter supposedly preventing bad proposals from getting into the standard
while improving the quality of the proposals that make it through. The existence of the committee
encourages people to make proposals and come forward to help. However, there is no formal
request-for-proposal process.
There are no full-time C++ designers, though there are many full-time C++ compiler, library, and
tool implementers. Until recently, there were relatively few application builders on the committee.
This is a problem because it biases the committee towards language lawyering, advanced features,
and implementation issues, rather than directly addressing the needs of the mass of C++ developers,
which many committee members know only indirectly. This problem may be partly alleviated by
the recent sharp increase of new members.
There are relatively few educators on the committee. This can be a problem because the committee
(rightly) prioritizes łease of learningž highly but members have very different ideas (and often
strong opinions) of what that means. This often confounds discussions about łsimplicityž and łease
of use.ž
When considering the impact of organizational problems on the development of C++, please
remember that the ISO process was not designed for 200-people meetings ś the typical ISO pro-
gramming language committee is one or two dozen people. On average we manage, partly by
recognizing the problems and addressing them. Consider some observed problems:
• Delays: The multi-stage process offers many opportunities for delays, blocking of proposals,
and for proposals to mutate. Dozens of members will insist that their needs be met, often
by elaboration, extension, and special cases. One person’s excessive delay is another’s due
diligence.
Examples: Concepts (6 years for the current approach (ğ6)), contracts (6 years from start to
failure (ğ9.6.1)), networking (15 years and still in progress (ğ8.8.1)) and constexpr (5 years
(ğ4.2.7)). Even getting nullptr accepted took three years (ğ4.2.6).
• Isolated features: Most committee members like to see improvement; that is, to see features
added. On the other hand, they are ś quite reasonably ś terrified of breaking existing code.
This gives a systematic advantage to isolated features, small proposals that supposedly don’t
affect the rest of the language or the standard library. Such proposals rarely have major
impact on how the language is used but add to the complexity of learning and implementation.
Also, they often turn out to have surprising feature interactions after all.
Examples: mostly features that were not worth mentioning in this summary of language
evolution. Structured bindings (ğ8.2) and operator <=> (ğ8.8.4) both required many meetings
to complete.
• Late alternatives: When ś sometimes after years of work ś a proposal comes near a vote,
members who have so far not taken an interest enter the discussions and alternative proposals
are made. Such proposals can be dramatically different from the original proposal or just a
stream of requests for minor changes. This usually leads to delays, confusion, and sometimes
acrimony as issues considered settled are resurrected and close-to-equal weight is given
to untried (and usually unimplemented) new ideas and to proposals that are the result of
years of work. For an older proposal, imperfections will have been uncovered and technical
tradeoffs worked out. It is far too easy to imagine the benefits of something new and forgetting
the law of unintended consequences: there will be unintended consequences. The new and
relatively unexamined always looks better than the old. This makes the proponents of the
older proposals defensive and diverts from efforts to refine the łold proposal.ž Here łoldž could
mean just a couple of years, or ś as in the case of concepts (ğ6) a dozen years. Sometimes
untried late changes (łimprovementsž) are accepted to appease opposition; this often has
unintended consequences. People entering into a discussion late often don’t see łthe need
to rushž and naturally want to see their ideas seriously considered (often without seriously
considering the details and rationale of an older proposal). This can cause friction with people
who have already invested years of work on the older proposal.
Examples: Structured bindings (syntax change, added support for bitfields, clumsy get() (ğ8.2)),
concepts (ğ6). Digit separators (ğ5.1), operator dot (ğ8.8.2), modules (ğ9.3.1), coroutines (ğ9.3.2),
contracts (ğ9.6.1).
• Enthusiasm favors the new: It is easier to summon enthusiasm for something new than for
opposing it. Every proposal will solve something for someone and its proponents are willing
to spend massive amounts of time demonstrating its value. To oppose, someone has to say
things like
ś No, this problem isn’t all that important.
ś No, this solution has flaws.
ś No, you have not documented your solution sufficiently.
ś No, you have not examined the alternatives carefully.
However politely phrased, this can easily make an opponent appear to be a łthe bad guyž
blocking progress and denying the validity of the proponents’ needs. Worse, the proponents
invariably spend much more time preparing papers and presentations than opponents. Most
people prefer to work constructively on something they believe in, rather than carefully de-
molishing other peoples’ work. So, the proponents are usually enthusiastic and well-prepared,
whereas the opponents can easily end up sounding vague and demonstrating ignorance of
details. Yet every new feature has a cost: e.g., design, specification, implementation, revision,
deployment, and teaching (ğ9.5). I fear Thursday afternoons in the evolution working group.
That’s when EWG members are tired after working hard on major proposals for days, many
of the regulars (such as me) have been dragged into other groups, and members are impatient
to see something done. This is where minor proposals slip by with relatively minor scrutiny.
Examples: explicit tests in conditions (ğ8.7), inline variables (ğ8), late changes to structured
bindings (ğ8.2).
• Overconfidence: Relative to the complexity of the complete language, the complete standard
library, and especially the complexity of the problems facing the users of C++ in diverse
application areas, an individual’s experience from everyday work is inadequate. Not all
committee members see that or adequately compensate by doubting the wider relevance
of their own experience. This can lead to proposals of limited generality being pushed too
hard. Worse, it can lead to proposals being vigorously opposed by members who can’t
see the need to solve the problems addressed. Language design requires some intellectual
humility [Stroustrup 2019b]. First solutions are rarely the best and off-the-cuff objections
and suggestions rarely lead to improvements without further serious thought.
Examples: No examples, to protect the guilty.
• Poor implementation timing: Implementing a proposal late in the standards process risks
a seriously flawed feature, potentially with unimplementable parts and lack of feedback
from use. Implementing a proposal early risks that the feature gets frozen in an incomplete,
suboptimal, and hard-to-use form. This is a hard, practical dilemma: many in the committee
will not vote for a proposal that has not been implemented or is implemented in a way they
do not trust. On the other hand, many implementers are unwilling to devote implementation
resources on a proposal that has not been approved by the committee. The committee often
hear the question łhas it been implemented?ž Frequently, łHas it been designed?ž and łHow
will this be used?ž are more important questions. People easily get lost in details. My suggested
way out of this dilemma is to agree on a direction, a general scope of a proposal, then start
with the detailed design and implementation of a relatively small subset guided by critical
use cases. That way, we can gain user experience relatively early and see how the feature
interacts with other features. This requires a long-term view of what that language should
be [Stroustrup 1993, 1994, 2007] (ğ1), (ğ11.2) or it deteriorates to mere opportunistic hacking.
When it works, the language benefits from feedback and organic growth.
Examples: modules (ğ9.3.1), C++0x concepts (ğ6), and <=> (ğ8.8.4).
• Feature interaction: One of the hardest issues to deal with is the use of features in combination.
This is partly a technical issue of specification and implementation. As such, it soaks up
much committee time. From a design perspective, the harder problem is to anticipate the
use of a new feature in the context of the whole language, including other new language
and library features under consideration. Every feature should be designed to be used in
combination with other features. I fear that point is underappreciated. Few proposal papers
offer detailed discussions and committee discussions about feature interactions tend to be
short or confused. One consequence of this is that individual features tend to bloat to be
usable in isolation from the rest of the language.
Examples: tuple (ğ4.3.4) and <=> (ğ9.3.4). The (failed) proposals for having specialized syntax
for actions in lambdas (ğ4.3.1).
• Volume and distractions: So much is going on, often concurrently, that nobody can keep up
with all. Those of us who try, can easily get distracted from important issues by what turns out
to be unimportant. There are now more than 500 committee papers a year, some are dozens
or even hundreds of pages long. This represents about a doubling of the volume of documents
compared to the early 2010s. I noted that the fall-2018 pre-meeting mailing (collection of new
papers) had three times as many words as the complete works of Shakespeare.
The flood of email messages can be most distracting as many members like to conduct
technical discussions through bursts of short messages. If you fall behind in such a discussion,
you lose track of issues and an apparent consensus can emerge from a handful of people.
That kind of discussion does not lend itself to a calm and systematic weighing of alternatives.
Sometimes, it leads to unfortunate features slipping through. Sometimes, it leads to differ-
ent parts of the language and standard library reflecting different design philosophies and
compromising interoperability.
Examples: The differing interfaces to any, optional, and variant (ğ8.3). Concepts (ğ6).
• Precise specification: The standard is a specification, not an implementation. However, the
standard is written in English, so mathematical precision escapes us. Many members of the
committee are mathematically inclined, but more are not, so mathematical notation cannot be
used in the specification. The attempts to make the English text precise and comprehensive
make it stilted and damage comprehensibility. I often have a hard time understanding the
standard’s description of my own proposals.
Most members are programmers, more than designers, so the specification sometimes ends
up looking like a program ś a program written in a low-level language without a type system
or a compiler. There are elaborate if-then-else explanations and few statements of invariants.
Worse, much of the vocabulary is inherited from C and based on tokens in the program
source text so that higher-level concepts are stated only indirectly.
Curiously, the standard-library specification is noticeably more formal in structure than the
language specification.
• Scholasticism: A heavy emphasis on having the text of the standard correct and precise is of
course necessary. However, people sometimes forget that the standard might be wrong and
discuss correctness exclusively based on arguments from the text. Thus, arguments based on
the models and uses that the standard’s text supposedly reflects can be ignored.
• Direction: Which problems are real? Important? For whom? Which are urgent? Which
solutions will still be relevant in a decade’s time? That something can be a problem doesn’t
imply that it must have a direct solution in the language. In particular, it is hard for a
committee to remember that a language cannot be all things to all people. It is even harder
to accept that it cannot solve even the most urgent problems of every member [Stroustrup
2018d].
Examples: C++17 (ğ8) and C++20 (ğ9).
• Exclusive focus: Some members focus exclusively on one or two issues, such as language
technicalities, ease of use, łteachability,ž efficiency, use in a single style of programming, use
in a single industry, use in a single firm, a single language feature, etc. This can be a very
effective technique for a member with an exclusive focus, but can make broad, balanced
progress difficult. Excessive trust in theory or in personal experience are other examples of
this. A good proposal offers progress in many areas but is typically not perfect along all of
these axes.
• Inappropriate application of principle: Applying general principles to a concrete example is
often difficult. Sometimes, a principle is applied rigidly without the necessary tradeoff with
other principles. The need for tradeoff is one reason D&E [Stroustrup 1994] refers to design
principles as łrules of thumb.ž Sometimes, a principle seems to appear out of nowhere without
empirical basis. Sometimes, a principle is rigidly adhered to for one proposal, yet disregarded
for another. Principled design is difficult; it requires taste and experience as well as principles.
Practical language design is not just an exercise in deduction from first principles. Often,
principles have to be balanced against each other.
• Pro-expert bias: It is hard to imagine the problems of someone different from yourself. The
committee members are almost all experts in something or other. In their day jobs, they are
typically the persons who deal with the most subtle and intricate problems. Such problems
are often rare in the billions of lines of C++ code łout therež and not the problems that the
majority of C++ programmers struggle with. Yet, the expert-level problems are typically the
ones that are urgent to the committee and have the easiest passage through the process.
Examples: The ease with which support for the use of enable_if and type traits (ğ4.5.1)
permeated the standard library compared to the trouble getting concepts (ğ6) accepted.
• Cleverness: The average committee member is clever and many show a weakness for clever
solutions. Further, they have a hard time to decide that not every problem is worth solving
and that having a solution doesn’t imply that we have to put it into the standard. This leads
to overelaborate features and to features that most programmers could happily live without.
It is only fair to point out that many programmers are also clever and sometimes delight in
overly clever language and standard-library features.
Example: proposals that require serious template metaprogramming even for simple uses.
• Unwillingness to compromise: Most members have strong opinions but gaining consensus
in a large group requires compromise. It can be hard to distinguish between compromise
over something inessential and over fundamental principles. The latter could damage the
language and should be avoided. Unfortunately, members who are firmly convinced that
their concerns are fundamental and essential have a critical tactical advantage compared
to people with a more open mind. People who care more about the language as a whole
than about any individual issue tend to give in to people who don’t. Similarly, people who
never seriously question their own principles or needs can mount a ferocious attack against
technical compromises that others consider necessary. Making progress requires concern for
the community as a whole, self-awareness, and a dose of humility [Stroustrup 2019b].
• Lack of priority: From a technical point of view, all problems are equal: an imprecise specifi-
cation is an imprecise specification independently of what it fails to properly specify and any
error that might slip through a hole in the type system could in principle cause death and
destruction. However, the real-world implications can be dramatically different. In fact, most
obscure details have essentially no effect. That’s hard for some people to remember when
working on details of a design.
Example: More time was spent on digit separators (ğ5.1) than on range-for (ğ4.2.2).
• Perfectionism: A standard is expected to be used by millions and be stable for decades.
Naturally, people would like it to be perfect. This leads to feature bloat (too many features)
and in particular to the bloat of individual features. Programmers are excellent at imagining
problems and as a feature progresses through the committee, members insist that it solve
them all. This can lead to serious mission creep and to features only an expert could love. It
can also lead to a feature not making it into the standard.
Examples: Operator dot (ğ8.8.2), networking library (ğ8.8.1), and exception specifications
(ğ4.5.3).
• Minority blocking: The consensus process protects against quite a few kinds of mistakes and
especially against the tyranny of the majority. However, it is vulnerable to individuals and
small groups blocking progress. That can be good (avoiding mistakes), but when it happens
repeatedly at the various stages of the proposal pipeline or just in the last minute, it can be
disruptive.
Examples: constexpr (ğ4.2.7), Operator dot (ğ8.8.2), modules (ğ9.3.1), and coroutines (ğ9.3.2).
• Cohesive groups: Many WGs and SGs have a stable core set of people who over the years
develop a cohesive outlook, a shared vocabulary, and specific ways of operating. This can
make it difficult for łoutsidersž to communicate and contribute. It can also make it difficult to
design features that cross WG boundaries, such as facilities with both library and language
parts. Each group is likely to design something that fits into its own organizational domain,
thus adding confirmation to the old dictum that the structure of a system resembles the
structure of the organization that created it.
Examples: range-for (ğ4.2.2) and concurrency mechanisms that might require language
changes (ğ4.1.3). The differing interfaces to any, optional, and variant (ğ8.3).
On the positive side, actions based on personal animosity or tit-for-tat are rare. In that sense, the
committee is quite professional.
Fortunately, not every proposal suffers from the effects of all of these phenomena and most of
these problems are shared with many other large projects. However, C++, as represented by its
ISO standard, reflects these phenomena. They are not brand new, but have been increasing since
C++11. I suspect they are caused by a combination of
• the increased size of the committee
• the influx of new people
• the specialization (fragmentation) of the membership
• the decrease of knowledge of C++’s history among the members
One reason the standards process has repeatedly succeeded despite these serious problems is that
many people mount a constant effort to minimize the negative effects. The creation of the Direction
Group is part of the effort (ğ3.2) [Dawes et al. 2018; Stroustrup 2018d]. See also (ğ11.4). The constant
efforts of working group chairs, note takers, meeting organizers, and editing groups are invisible,
but essential. For example, Jens Maurer has for decades taken notes in the CWG, helped proposers
write standards text, arranged for web access, organized phone access to members unable to attend,
organized meeting rooms, informed members about local travel possibilities, and more.
What would be an alternative? In an ideal world, I would recommend restricting decision making
to a small group (on the order of 5 people) of full-time trusted experts but have the discussions,
the ability to propose, and most of the processing done by a large group (like the 350+ member
committee). No, I don’t see something like that happening for C++:
• Nobody likes to give up power (in this case voting power).
• Maintaining stable funding for a permanent staff of full-time experts requires non-trivial
skills (and such skills haven’t surfaced in the C++ community).
• Radical change doesn’t happen during times of success; only a marked drop in C++ use could
motivate the committee to dramatic organizational innovation (and then it would probably
be too late).
I don’t see corporate control as a viable alternative:
• Corporations expect return on investment.
• Corporate support often evaporates after a few years.
• Corporations tend to prefer differential advantages over progress that benefits all.
Nor do I consider fully open processing (by thousands of voters) viable:
• A cast of thousands does not have taste.
• The membership and opinions of a large group are not stable over decades.
The hierarchical approval process that works for many large open-source projects might have
been at least part of an answer, but there was little experience with that when the standardization
of C and C++ started. When such a system works well, the higher you get in the approval hierarchy
the broader a base of knowledge of the approvers and the wider their area of concern. At the top,
we find one or more people with some knowledge of just about everything and concern for all
users. To contrast, the ISO process dilutes expertise and areas of concern as a proposal approaches
final approval: In plenary, many members vote on proposals that they have little interest in, have
limited experience with the problem area, and haven’t followed closely. People try hard to do
so responsibly, but that is hard and seeing each proposal as part of the bigger picture is almost
impossible.
Looking at it this way, WG21 hasn’t done that badly. I do worry whether this model can keep C++
coherent and relevant for much longer. The 200+ people who turn up for a C++ standards meeting
is an order of magnitude larger than for other standards groups, for which the ISO process was
designed. Also, the membership is far more diverse than the groups of grizzled experts, corporate
representatives, and national body representatives of the past. Chaos could erupt.
I take some comfort from Winston Churchill’s dictum that łdemocracy is the worst form of
Government except for all those other forms that have been tried from time to time.ž
In particular, I don’t see the often suggested łBenevolent Dictator For Lifež model scaling, and
that model has never been relevant for C++ anyway.
An individual or a small coherent group of friends is my ideal model for starting a language
design project, but I don’t see that approach scaling. A mature language needs dozens or even
hundreds of people working on the huge variety of problems that must be faced. Even coordination
with relevant standards groups and industry groups would swamp the capacity of a tiny coherent
group.
3.4 Proposal Checklists

There was a łHow to write a proposalž guide for C++98 [Stroustrup et al. 1992], but curiously the
evolution group did not have a checklist for proposals for C++14, C++17, or C++20. There was
one for standard-library proposals [Meredith 2012]. For C++20, a note from the heads of National
Standards Bodies [van Winkel et al. 2017] and a paper from the Direction Group [Hinnant et al.
2019] gave some guidance. Here is a short and incomplete list of questions that were almost always
raised for a proposal:
• What is the problem to be solved? What kind of users will be served? Novices? Experts?
• What is the solution? Articulate the principles that it is based on. Give simple use cases and
examples of expert-level use.
• What are alternative solutions? Could a library solution be sufficient? Why are current
facilities not good enough?
• Why does the solution need to be in the standard?
• What barriers to adoption are there? How long is a transition from existing techniques likely
to take?
• Has it been implemented? What implementation problems were encountered or can be
expected? Is there any user experience?
• Will there be significant compile-time overheads?
• Does the feature fit into the frameworks of existing tools and compilers?
• Will there be run-time overheads compared to workarounds? In time? In space?
• Will there be compatibility problems? Breakage of existing code? ABI breakage?
• How will the new feature interact with existing and other new features?
• Is the solution teachable? To whom? By whom?

• How will the standard library be affected?
• Will the proposal lead to demands for further extension in future standards?
• How does the feature fit into the wording of the standard?
• What mistakes are users likely to make with the new feature?
• Is the proposal among the top-20 in terms of benefits to the C++ community at large? Top-10?
• Is the proposal among the top-3 in terms of a specific sub-community? Which sub-community?
• Is the proposal for a general mechanism to solve a class of problems or a specific solution to
a specific problem? If to a class, which class of problems?
• Is the proposal coherent with the rest of the language in terms of semantics, syntax, and
naming?
Ideally, a proposal would have answers to all these questions, and more, but that rarely happened.
In particular, the rationale was often very weak in an initial proposal as proposers thought that the
importance of the problems addressed and their suggested solution were pretty obvious. However,
follow-up papers, revisions, email discussions, and face-to-face discussions in the Evolution Group
typically cover those questions, but rarely systematically or consistently across proposals. Members
have a tendency to focus on technical details (e.g., grammar, ambiguities, optimization opportunities,
and naming), rather than revisiting fundamental questions. Sometimes, what I consider a bad
proposal slips through. The reason is usually great enthusiasm from proposers combined with
distraction, politeness, and exhaustion among opponents [Stroustrup 2019b].
4 C++11: IT FEELS LIKE A NEW LANGUAGE

The release of C++11 [Becker 2011] and the relatively quick deployment of implementations
led to much enthusiasm, increased use, an influx of new people into the C++ world, and much
experimentation. Three complete or almost complete implementations of C++11 were available
in 2013. My comment at the time, C++11 feels like a new language [Stroustrup 2014d], was widely
perceived as accurate. Why and how did C++11 do such a good job helping programmers?
C++11 introduced a bewildering number of language features, including:
• memory model ś an efficient low level-model of modern hardware as a foundation for
concurrency (ğ4.1.1)
• auto and decltype ś avoiding redundant repetition of type names (ğ4.2.1)
• range-for ś simple linear traversal of ranges (ğ4.2.2)
• move semantics and rvalue references ś minimizing copying of data (ğ4.2.3)
• uniform initialization ś an (almost) completely general syntax and semantics for initializing
objects of all kinds and types (ğ4.2.5)
• nullptr ś a name for the null pointer (ğ4.2.6)
• constexpr functions ś compile-time evaluated functions (ğ4.2.7)
• user-defined literals ś literals for user-defined types (ğ4.2.8)
• raw string literals ś literals where escape characters are not needed, mostly for regular
expressions (ğ4.2.9)
• attributes ś associating essentially arbitrary information with a name (ğ4.2.10)
• lambdas ś unnamed function objects (ğ4.3.1)
• variadic templates ś templates that can handle an arbitrary number of arguments of
arbitrary types (ğ4.3.2)
• template aliases ś the ability to rename a template and to bind some template arguments for
the new name (ğ4.3.3)
• noexcept ś a way of ensuring that an exception isn’t thrown from a function (ğ4.5.3)
• override and final ś explicit syntax for managing large class hierarchies
• static_assert ś compile-time assertions
• long long ś a longer integer type
• default member initializers ś give a data member a default value that can be superseded by
initialization in a constructor
• enum classes ś strongly typed enumerations with scoped enumerators
And here is a list of the major standard-library components (ğ4.6):
• unique_ptr and shared_ptr ś resource-management pointers (ğ4.2.4) relying on RAII
(ğ2.2.1)
• memory model and atomic variables (ğ4.1.1)
• thread, mutex, condition_variable, etc. ś type-safe and portable support for basic system-
level concurrency (ğ4.1.2)
• future, promise, and packaged_task, etc. ś slightly higher-level concurrency (ğ4.1.3)
• tuple ś unnamed simple composite types (ğ4.3.4)
• type traits ś testable properties of types for use in metaprogramming (ğ4.5.1)
• regular expression matching (ğ4.6)
• random numbers ś with many generators (engines) and distributions (ğ4.6)
• Time ś time_point and duration (ğ4.6)
• unordered_map, etc. ś hash tables
• forward_list ś a singly-linked list
• array ś a fixed-constant-sized array that knows its size
• emplace operations ś construct objects right within a container to avoid copying
• exception_ptr ś enables transfer of exceptions between threads
There are more, but these are the most significant changes. All are described in [Stroustrup 2013]
and much information is available online (e.g., [Cppreference 2011ś2020]).
How could these apparently disconnected extensions make a coherent whole? How could this
actually change the way we write code for the better? C++11 did achieve that. In a relatively short
period of time (say, 5 years), huge amounts of C++ were upgraded to C++11 (and further to C++14
and C++17) and the presentation of C++ at conferences and blogs completely changed.
This dramatic change in the łfeelž of the language and the styles of its use is not the result of
a traditional careful design process guided by a master craftsman, but the result of a huge set of
suggestions filtered through layers of decisions by a large and changing set of individuals.
In my HOPL3 paper [Stroustrup 2007], I correctly described many of the C++11 language features.
The notable exception was łconceptsž which is addressed in (ğ6). Instead of going into details again,
I will describe a classification of facilities into łthemesž based on what programmer needs they
address. I think this way of looking at proposals is the root of C++11’s success:
• ğ4.1: Support concurrency
• ğ4.2: Simplify use
• ğ4.3: Improve support for generic programming
• ğ4.4: Increase static type safety
• ğ4.5: Support for library building
• ğ4.6: Standard library components
These łthemesž are not disjoint. In fact, my conjecture is that C++11 was a success because it
adds to up a fine mesh of interrelated facilities addressing genuine needs. My favorite features
belong in every theme. I suspect that my articulated aims for C++ in writing (e.g., [Stroustrup
1993, 1994, 2007]) and presentations helped the design remain reasonably focused. For me, a crucial
measure of every new feature is whether it brings C++ closer to its ideals, e.g., by making the
support for built-in types and user-defined types more similar (ğ2.1).
Looking at C++11, we see suggested improvements from about 2002 and quite a few libraries
emerging early, often as part of Boost [Boost 1998ś2020]. However, complete C++11 implemen-
tations were not available until 2013. In 2020, some organizations still struggle with an upgrade
to C++11 because of huge code bases, programmers stuck in the past, outdated teaching, and
(especially in the embedded systems world) seriously outdated compilers. Adoption of C++17 is
noticeably faster than the adoption of C++98 and C++11, though, and as early as 2018 some major
C++20 features were already in production use.
As late as 2018, I have seen pre-C++98 compilers used in teaching. I consider that abuse of
students, depriving them of 20+ years of progress.
What is the distant past according to the standards committee, to the major compiler vendors,
and to most vocal C++ proponents is still the present ś or even the future ś for many. The result is
a continuing confusion about what C++ really is. This confusion will last as long as C++ continues
to evolve.
4.1 C++11: Support for Concurrency

C++11 had to support concurrency. It was both obvious and a common requirement from all major
users and platform suppliers. C++ was (and is) heavily used as the foundation for most of the
software industry, and in the first decade of 2000 concurrency was becoming pervasive. Exploiting
hardware concurrency well was essential. Like C, C++ had of course always supported various
forms of concurrency, but that support had not been standardized and was generally low level.
Machine architectures were using increasingly subtle memory architectures and compiler writers
were applying increasingly aggressive optimization techniques, making life extremely difficult for
the writers of the lower levels of software. A treaty between machine architects and optimizer
writers was badly needed. Only with a well-specified memory model could writers of foundational
libraries have a stable base and some degree of portability.
The work on concurrency was spun off from the EWG into the concurrency group with expert
membership led by Hans-J Boehm (HP, later Google). It had three main trusts:
• ğ4.1.1: Memory model
• ğ4.1.2: Threads and locks
• ğ4.1.3: Futures
In addition, parallel algorithms (ğ8.5), networking (ğ8.8.1), and coroutines (ğ9.3.2) were dealt
with in separate groups and (as expected) not ready for C++11.
4.1.1 Memory Model. One of the most urgent issues was to precisely specify the rules for accessing
memory in a world of multi-cores, caches, speculative execution, instruction reordering, etc. Paul
Mckenney from IBM was very active on topics of memory guarantees. The research of Mark
Batty from Cambridge University [Batty et al. 2013, 2012, 2010, 2011] helped our formalizations as
described in [McKenney et al. 2010] by P. McKenney, M. Batty, C. Nelson, H. Boehm, A. Williams,
S. Owens, S. Sarkar, P. Sewell, T. Weber, M. Wong, L. Crowl, and B. Kosnik. It was a massive and
essential part of C++11.
In C11, C adopted the C++ memory model. However, in the very last minute before the C standard
was put out to vote and after the last opportunity to change the C++11 standard, the C committee
introduced notational incompatibilities that became a pain for C and C++ implementers and users.
Much of the memory model was motivated by the needs of the Linux and Windows kernels. It
is now used there and also far more widely. The memory model is widely underrated because it
isn’t seen by most programmers. To a first order of approximation, it simply makes code work as
anyone would expect.
Initially, I think that most of the committee members underestimated the problem. We knew
that Java had a good memory model [Pugh 2004] and hoped to adopt that. I was highly amused to
find that representatives from Intel and IBM effectively vetoed that idea by pointing out that by
adopting the Java memory model for C++ we would slow down all JVMs by a factor of at least two.
Consequently, to preserve the performance of Java, we had to adopt a far more complex model for
C++. Ironically and predictably, C++ was then criticized for having a more complicated memory
model than Java.
Basically, the C++11 model is based on happens-before relations [Lamport 1978] and supports
relaxed memory models as well as sequentially consistent [Lamport 1979] ones. On top of that
and integrated with it, C++11 provides atomic types and support for lock-free programming. The
details are far beyond the scope of this paper (e.g., see [Williams 2018]).
Unsurprisingly, the memory model discussions in the concurrency group got a bit heated at times.
Significant interests of hardware manufacturers and compiler vendors were at stake. One of the
hardest decisions was to accept both Intel’s x86 primitives (a Total Store Order (TSO) model [TSO
Wikipedia 2020] plus a few atomic operations) and IBM’s PowerPC primitives (weak consistency
plus fences) for the lowest level synchronization. Logically, only one set of primitives were needed,
but Paul Mckenney convinced me that there was far too much code written using fences deep
in complex algorithms for IBM to adopt something like Intel’s model. One day, I literally did
shuttle diplomacy between two corners of a large room. Eventually, I suggested that both had to
be supported and that was what C++11 adopted. I ś and others ś were very pleased when later
people found that fences and atomics could be used together to create better solutions than either
in isolation.
A bit later, we added support for data-dependency-based consistency represented in source code
through attributes (ğ4.2.10), such as [[carries_dependency]].
C++11 introduced atomic types on which simple operations are atomic:
atomic < int > x;
void increment ()
{
x ++; // not x = x + 1
}
Obviously, these are widely useful. For example, using an atomic makes the notoriously tricky
double-checked locking optimization trivial:
mutex mutex_x ;
atomic < bool > init_x ; // initially false .
int x;
if (! init_x ) {
lock_guard < mutex > lck ( mutex_x );
if (! init_x ) x = 42;
init_x = true ;
} // implicitly release mutex_x here ( RAII )
// ... use x ...
The point of double-checked locking is to use the relatively cheap atomic to guard the use of
the much more expensive mutex.
The lock_guard is a RAII type (ğ2.2.1) that guarantees the unlocking of the mutex it controls.
Hans-J Boehm has described the atomic types as łsurprisingly popular,ž but I can’t say I’m
surprised. Being less expert than Hans, I appreciate the simplification more. C++11 also introduced
the key operations for lock-free programming, such as compare and swap:
template < typename T >
class stack {
std :: atomic < node <T >* > head ;
public :
void push ( const T& data )
{
node <T >* new_node = new node <T >( data );
new_node -> next = head . load ( std :: memory_order_relaxed );
while (! head . compare_exchange_weak ( new_node -> next , new_node ,
std :: memory_order_release , std :: memory_order_relaxed )) ;
}
// ...
};
Even with C++11 support, I consider lock-free programming expert-level work.

4.1.2 Threads and Locks. On top of the memory model, a threads-and-locks model of concurrency
was provided. I consider the threads-and-locks level of concurrency the worst model for application
use of concurrency, but it is essential for a language like C++. Whatever else it is, C++ is (and always
was) a systems programming language capable of interacting directly with the operating system,
usable for kernel code and device drivers. Therefore, it has to support what the systems support at
their lowest levels. On top of that, we can build a variety of concurrency models more suitable for
specific applications. Personally, I’m particularly fond of message-based systems because they can
eliminate the data races that are the root of the most subtle concurrency bugs.
C++’s support for the threads-and-locks level of programming is a type-safe variant of what
POSIX and Windows offer. It is described in [Stroustrup 2013] and in greater depth in Anthony
Williams’ book [Williams 2012, 2018]:
• thread ś a system’s thread of execution, with join() and detach()
• mutex ś a system’s mutex with lock(), unlock(), and RAII ways of getting guaranteed
unlock()
• condition_variable ś a system’s condition variable for communicating events between
threads
• thread_local ś thread-local storage
Compared to the C versions, the type safety makes the code much simpler and cleaner, e.g., no
more void**s and macros. Consider a simple example of having a function execute on a different
thread and returning a result:
class F { // traditional function object

public :
F( const vector < double >& vv , double * p) :v{ vv }, res {p} { }
void operator ()(); // place result in * res
private :
const vector < double >& v; // source of input
double * res ; // target for output
};
double f( const vector < double >& v ); // traditional function
void g( const vector < double >& v , double * res ); // put result into * res
int comp ( vector < double >& vec1 , vector < double >& vec2 , vector < double >& vec3 )
{
double res1 ;
double res2 ;
double res3 ;
// ...
thread t1 {F{ vec1 , res1 }}; // function object
thread t2 {[&](){ res2 =f( vec2 );} }; // lambda
thread t3 {g , vec3 ,& res3 }; // ordinary function
t1 . join ();
t2 . join ();
t3 . join ();
cout << res1 << ' ' << res2 << ' ' << res3 << '\n ';
}
The design of the type-safe library support relies critically on variadic templates (ğ4.3.2). For ex-
ample, the constructor for std::thread is a variadic template. It can distinguish different executable
first arguments and check that they are followed by a correct number of arguments of correct types.
Similarly, lambdas (ğ4.3.1) made many uses of the <thread> library much simpler. For example,
the argument for t2 is a piece of code (a lambda expression) that accesses the surrounding local
scope.
It was difficult to get novel features accepted and used in the standard library in the same release
of the standard. There were voices raised that doing this was too aggressive and that it could lead
to long-term problems. Introducing new language features and using them at the same time was
undoubtedly risky, but it added significantly to the quality of the standard by
• giving users a better standard library
• giving users examples of good use of the language features
• saving users from having to implement low-level facilities
• forcing designers of language features to cope with difficult real-world uses
The threads-and-locks model requires the use of some form of synchronization to avoid race
conditions. C++11 provides standard mutexes for that:
mutex m; // controlling mutex

int sh ; // shared data
void access ()
{
unique_lock < mutex > lck {m }; // acquire mutex
sh += 7; // manipulate shared data
} // release mutex implicitly
The unique_lock is a RAII object ensuring that the user cannot forget to unlock() the mutex.
These lock objects also provided a way of protecting against the most common form of deadlock:
void f ()
{
// ...
unique_lock < mutex > lck1 {m1 , defer_lock }; // don 't yet acquire m1
unique_lock < mutex > lck2 {m2 , defer_lock };
unique_lock < mutex > lck3 {m3 , defer_lock };
// ...
lock ( lck1 , lck2 , lck3 ); // acquire all three mutexes
// ... manipulate shared data ...
} // implicitly release all mutexes
Here, the lock() function acquires all mutexes łsimultaneouslyž and releases all implicitly (RAII
(ğ2.2.1). C++17 has an even more elegant solution (ğ8.4).
The threads library was first proposed for C++0x by Pete Becker (Dinkumware) [Becker 2004]
in 2004 based on a Dinkumware implementation of the interface offered by boost::thread [Boost
1998ś2020]. It is probably not a coincidence that this was presented at the same meeting (Redmond
WA, September 2004) as the first proposal for a memory model [Alexandrescu et al. 2004].
The biggest controversy was over cancellation, which is the ability to stop a thread from running
to completion. Essentially every C++ programmer on the committee wanted that in some form or
other. However, the C committee objected to thread cancellation in a formal note to WG21 [WG14
2007], the only formal note ever sent from WG14 (the ISO C standards committee) to WG21. I noted,
łbut C doesn’t have destructors and RAII for systematic resource management and clean-up.ž The
Austin Group who manages POSIX sent representatives who were 100% against any form of the
idea, insisting that cancellation was neither necessary nor possible to do safely. Observing that
Windows and other operating systems offer variants of the idea and that C++ isn’t C made no
difference to the POSIX people. I fear that they were defending their business and their C-language
world view, rather than trying to come up with the best solution for C++. The lack of standard
cancellation has repeatedly been a problem. For example, in a parallel search (ğ8.5), the thread that
first finds the answer would like to trigger cancellation (by any name) of the other such threads.
C++20 provides the stop-token mechanism to support this use case (ğ9.4).
4.1.3 Futures. A type-safe and standard POSIX/Windows-like thread library was a major im-
provement over the incompatible C-style libraries in use, but that was still 1980s style low-level
programming. Some members, notably me, argued that C++ badly needed something more modern
and higher level. For example, Matt Austern (Google, formerly SGI) and I argued for message queues
(łchannelsž) and a thread pool. That made little progress against objections that there wasn’t time
to do that right. I pleaded and pointed out that if the experts in the committee didn’t offer such
facilities, they would end up having to use facilities łcooked up in a hurry by my students!ž The
committee could certainly do a much better job than that. łIf you won’t do that, please give me one
way, just one way, to pass information between threads without explicit synchronization!ž
The committee were split between members who basically wanted POSIX with improved typing
(notably, PJ Plauger) and members pointing out that POSIX was basically a 1970s design and
łeverybodyž was using higher-level facilities already. At the 2007 Kona meeting we agreed on a
compromise: C++0x (still expected to become C++09) would offer promises and futures and a
launcher for asynchronous tasks, async(), that allows but does not require a thread pool. Like
most compromises, łthe Kona compromisež pleased nobody and led to some technical problems.
However, many users considered it a success ś most didn’t know it was a compromise ś and over
the years, improvements have emerged.
In the end, C++11 offered:

• future ś a handle from which you can get() a value from a shared one-object buffer, possibly
after a wait for the value to be put there by a promise.
• promise ś a handle through which you can put() a value to a shared one-object buffer,
possibly waking up a thread waiting on a future.
• packaged_task ś a class that makes it easy to set up a function to be executed asynchronously
on a thread with a future for its result and a promise to return the result.
• async() ś a function that can launch a task to be executed on another thread.
The easiest way of using all this is to use async(). Given an ordinary function as an argument,
async() runs it on a thread handling all the details of thread launching and communication:
double comp4 ( vector < double >& v)
// spawn many tasks if v is large enough
{
if (v. size () <10000) // is it worth using concurrency ?
return accum (v. begin () , v. end () ,0.0);
auto v0 = &v [0];
auto sz = v. size ();
auto f0 = async ( accum ,v0 , v0 + sz /4 ,0.0); // first quarter

auto f1 = async ( accum , v0 + sz /4 , v0 + sz /2 ,0.0); // second quarter
auto f2 = async ( accum , v0 + sz /2 , v0 + sz *3/4 ,0.0); // third quarter
auto f3 = async ( accum , v0 + sz *3/4 , v0 +sz ,0.0); // fourth quarter
return f0 . get ()+ f1 . get ()+ f2 . get ()+ f3 . get (); // collect the results
}
The async wraps the code in a packaged_task and manages the setup of a future and its
promise to transmit the result.
Either a value or an exception can be passed from one thread to another through a future/promise
pair. For example:
X f(Y ); // ordinary function
void ff (Y y , promise <X >& p) // to execute f(y) asynchronously

{
try {
X res = f(y ); // ... compute a value for res ...
p. set_value ( res );
}
catch (...) { // oops : couldn 't compute res
p. set_exception ( current_exception ());
}
}
For simplicity, I have not used perfect forwarding of the argument (ğ4.2.3).
A get() on the corresponding future will now either get a value or throw an exception ś exactly
as for an equivalent synchronous call of f().
void user (Y arg )
{
auto pro = promise <X >{};
auto fut = pro . get_future ();

thread t {ff , arg , pro }; // run ff on a different thread
// ... do something else for a while ...
X x = fut . get ();
// ...
}
The standard library packaged_task automates this way of wrapping an ordinary function in a
function object that handles the promise/future setup and deals with returns and exceptions.
I hoped that this would lead to a work-stealing implementation supported by a thread pool, but I
was disappointed.
See also (ğ8.4).
4.2 C++11: Simplifying Use

C++ is łexpert friendly.ž I think I was the first to use that phrase as a gentle criticism and to
popularize the slogan Make simple things simple! in the context of C++. A language primarily aimed
at industrial use should be friendly to experts, of course, but a language cannot be just expert
friendly. Most people who use a programming language are not ś and do not want to become ś
expert in all aspects of that language. They want to get their job done sufficiently well with the
least distraction from the language. A programming language is there to allow application ideas to
be expressed, not to turn programmers into language lawyers. It follows that a language design
should strive to make simple things simple. A language that is also for experts must additionally
ensure that nothing essential is impossible or unreasonably expensive.
Another criterion commonly applied in discussions about potential C++ language extensions
and standard-library components is łIs it easy to teach?ž This question is now very common; it
was pioneered by Francis Glassborow and me. The idea of łbeing easy to teachž was there from the
earliest days of C++; you find it in The Design and Evolution of C++ [Stroustrup 1994].
Naturally, proponents of something new inevitably deem their design simple, easy to use, reason-
ably safe, efficient, easy to teach, and useful for most programmers. Opponents tend to doubt some
or all of those claims. However, it is important that this discussion takes place for every new feature
proposed for C++: in person at meetings, in papers [WG21 1989ś2020], and in email discussions. In
these discussions, I often point out that I’m a novice most of the time. That is, whenever I learn a
new feature, technique, or application domain, I am a novice and can use all the help I can get from
the language and standard library. One result was that C++11 offered some facilities specifically
aimed to simplify the use of C++ by learners, novices, and non-language-experts.
Every new feature makes something simpler for somebody. The łSimplifying usež theme focuses
on features for which making it simpler to express a known idiom was a major part of its motivation.
Here is a selection:
• ğ4.2.1: auto ś avoiding redundant repetition of type names
• ğ4.2.2: Range-for ś simple linear traversal of ranges
• ğ4.2.3: Move semantics and rvalue references ś minimizing copying of data
• ğ4.2.4: Resource-management pointers ś łsmartž pointers managing the lifetime of the objects
they point to (unique_ptr and shared_ptr)
• ğ4.2.5: Uniform initialization ś an (almost) completely general syntax and semantics for
initializing objects of all kinds and types
• ğ4.2.6: nullptr ś a name for the null pointer
• ğ4.2.7: constexpr functions ś compile-time evaluated functions
• ğ4.2.8: User-defined literals ś literals for user-defined types
• ğ4.2.9: Raw string literals ś literals where the escape character (’\’) is not used to escape,
mostly for regular expressions
• ğ4.2.10: Attributes ś associating essentially arbitrary information with a name
• ğ4.2.11: An interface to an optional garbage collector
• ğ4.3.1: Lambdas ś unnamed function objects
After C++11 was beginning to see serious use, I started taking small unscientific surveys as I
traveled widely to talk with C++ users: Which C++11 features do you most like? The top three
were invariably:
• ğ4.2.1: auto
• ğ4.2.2: Range-for
• ğ4.3.1: Lambdas
These three are among the simplest additions to C++11 and do not offer any new fundamental
functionality. What they do could be done in C++98, but not as elegantly.
I interpret this to imply that programmers of all abilities really appreciate concise notation for
simple common uses. They will happily abandon a more general notation for a simpler and more
specialized one for the cases where it applies. A common rallying cry is łThere should be only one
way of saying something!ž This łdesign principlež simply doesn’t reflect real-world user preferences.
I tend to rely on the onion principle instead [Stroustrup 1994]. You design so that doing simple tasks
is simple and when you need to do something less simple, you need to be more detailed, using a
more complicated technique or notation. That is, you peel one layer off the onion. The more layers
you peel off, the more you cry.
Please note that here simple does not mean low level. It is superficially simple to learn low-level
facilities such as void*, macros, C-style strings, and casts but it is hard to use them to produce
quality, maintainable software.
4.2.1 auto and decltype. The oldest new feature in C++11 was the ability to specify that an
object should have the type of its initializer. For example:
auto i = 7; // i is an int
auto d = 7.2; // d is a double
auto p = v. begin (); // p gets the type of v 's iterator
// ( begin () returns an iterator )
auto is a static facility allowing the deduction of the (static) type of an object from its initializer.
If you need a dynamically typed variable, use variant or any (ğ8.3).
I implemented auto in the winter of 1982/83 but was forced to remove it to preserve C compati-
bility.
In the context of C++11, people proposed a typeof operator to replace typeof macros and
compiler extensions that had become popular. Unfortunately, the different typeof macros were
incompatible in their treatment of references, so none of those could be adopted without major
code breakage. Introducing a new keyword is always hard because if it is short with a reasonably
obvious meaning, it will already have been used many thousands of times. If the suggested keyword
is ugly and long, people will dislike it.
Jaakko Järvi, one of the most prolific contributors of interesting libraries to Boost and at the
time my colleague at Texas A&M, led the discussion of typeof. We realized that the problem with
semantics could be summarized as łis typeof of a reference the reference itself or the referenced
type?ž Also, we felt that typeof was a bit verbose and error prone:
typeof (x+y) z = y+x;
Here, I thought I was repeating x+y, but I wasn’t (with potentially bad effects) and anyway, why
should I have to repeat anything? At this point, I realized that I had solved this problem back in
1982; we could eliminate the repetition and łhighjackž the keyword auto:
auto z = y+x; // z gets the type of y+x
In C, and then in C++, auto meant łallocate in automatic storage (i.e., on the stack)ž and was
never used. Looking through many millions of lines of C and C++, we verified that auto was only
used in test suites and in error, so we could recycle it with my 1982 meaning łget the type of the
initializer expression.ž
That left the problem of what to do with the use cases where we would like to deduce the type
of a reference as a reference. This is not uncommon in template-based foundation libraries. We
proposed an operator decltype with the reference-preserving semantics:
template < typename T > void f(T& r)
{
auto v = r; // v is a T
decltype (r) r2 = r; // r2 is a T&
// ...
}
Why decltype? Unfortunately, I don’t remember who suggested that name, but I know why:
• We couldn’t use typeof because that would break a lot of code.
• We couldn’t find a nice, short, and unused name.
• decltype is sufficiently mnemonic to remember (łdeclared typež) and sufficiently odd not to
have been used in existing code.
• decltype is reasonably short.
The paper proposing this was from 2003 [Järvi et al. 2003b]. The paper accepted by vote was from
2006 [Järvi et al. 2007]. Jaakko Järvi did most of the detailed work getting decltype through the
committee. Doug Gregor, Gabriel Dos Reis, Jeremy Siek, and I also helped and appear as co-authors
on various papers. It turned out that specifying the exact semantics of decltype is far harder
than I have made it sound here. Spending years on the details of an apparently simple feature is
unfortunately not uncommon. Part of the reason is inherent complexity of features; part is the
number of people who have to agree that every detail is designed and specified to their satisfaction.
I consider auto to be a purely simplifying feature whereas the primary purpose of decltype is
to enable sophisticated metaprogramming in foundation libraries. They are however very closely
related when looked at from a language-technical point of use.
There were two obvious generalizations of auto that I had explored elsewhere [Stroustrup and
Dos Reis 2003b]: as return types and argument types. This is obvious because argument passing
and value return are defined as initialization in C++. In 2003, when I first presented those ideas
to the committee, the members of the evolution working group reacted with undisguised horror.
Consider:
auto f( auto arg )
{
return arg ;
}
auto x = f (1); // x is an int

auto s = f( string (" Hello ")); // s is a string
Such examples were received more negatively than any other idea I have ever presented to the
committee. łIt was like the duchess who saw the mousež was my description: łEeeeek!ž However,
that was not the end of the story. C++17 offered auto for both arguments and return types for
lambdas (ğ4.3.1), but for functions C++17 offered auto only for return types. C++20 added auto
for function arguments as part of concepts (ğ6.4), thus finally completing support for my 2003
suggestion.
A weaker use of auto was added in C++11 to move the specification of the return type after the
arguments. For example, in C++98, we’d write
vector <T >:: iterator vector <T >:: begin () { /* ... */ }
The repetition of vector<T>:: is annoying and there was no way a return type could depend on
an argument type (as is useful in some generic programming). C++11 remedied that and improves
readability:
auto vector <T >:: begin () -> iterator { /* ... */ }
So, after years of work, we had auto. It immediately became very popular because it saved
programmers from spelling out long type names and from thinking about details of types in generic
code. For example:
for ( auto p = v. begin (); p != v. end (); ++ p) ... // traditional STL loop
It allowed people to line up names:

class X {
public :
auto f () -> int ;
auto gpr ( int ) -> void ;
// ...
};
void use ( int x , char * p)

{
auto x2 = x *2; // x2 is an int
auto ch = p[x ]; // ch is a char
auto p2 = p +2; // p2 is a char *
// ...
}
There were even calls to use auto everywhere or almost everywhere [Sutter 2013b]. This is
classic: Every new useful feature is initially overused and misused. After a while, parts of the
community find a balance. Articulating such balanced use as a best practice is one of the reasons I
(and many others) work on programming guidelines (ğ10.6). For auto, I received many comments
about lack of readability when people used it with initializers that did not have obvious types.
Consequently, the C++ Core Guidelines [Stroustrup and Sutter 2014ś2020] (ğ10.6) has this rule:
ES.11: Use auto to avoid redundant repetition of type names
My books have advice along the same line [Stroustrup 2013, 2014d]. Consider:
auto n = 1; // OK n is an int
auto x = make_unique < Gadget >( arg ); // OK : x is a std :: unique_ptr < Gadget >
auto y = flopscomps (x ,3); // Bad : what might flopscomps () return ?
It’s not 100% clear how to apply such a rule in every case, but it is far better than no rule and leads
to more readable code than the absolute rules łnever use auto!ž and łalways use auto!ž Real-world
programming tends to take more skill than simple examples illustrating language features.
If flopscomps() isn’t part of a generic computation, it would be better to explicitly state the
desired type. We had to wait for C++20 to use a concept to constrain the return type (ğ6.3.5):
Channel auto y = flopscomps (x ,3); // y can be used as a Channel
So, was the work on auto worthwhile? It is a small facility that for the simple cases can be
implemented in a day, yet it took 4 years to get through the committee. It is not even novel: many
languages have had facilities like this for the last 40 years. Even C with Classes had it 35 years ago!
I often despair about the time it takes to get even the smallest feature through the C++ standards
committee, and the often-painful discussions it involves. On the other hand, once it is done well,
millions of programmers benefit. When something is done really well, the most common comment
is łObvious! What took you so long?ž
4.2.2 Range-for. A range-for is a statement that loops over all elements of a sequence from first
to last. For example:
void use ( vector < int >& v , list < string >& lst )
{
for ( int x : v) cout << x << '\n ';
int sum = 0;
for ( auto i : {1 ,2 ,3 ,5 ,8}) sum += i; // an initializer list is a sequence
for ( string & s : lst ) s += ". cpp "; // use a reference allow modification
}
It was first proposed by Thorsten Ottosen (University of Aalborg, Denmark) because łJust about
any modern language has some form of ‘for each’ built into it.ž [Ottosen 2005]. I don’t usually
consider łeverybody else has onež as a good argument, but in this case the real point is that a simple
loop over a range simplifies one of the most common operations and offers some optimization
opportunities. Thus, range-for fits my overall aims for C++ perfectly. It directly expresses what
should be done, rather than spelling out in detail how it is done. It was also a nice clean syntax and
the semantics is obvious.
Being simpler and more specific, the range-for notation eliminates the possibility for making a
few łtrivialž but common errors:
void use ( vector < int >& v , list < string >& lst )
{
for ( int i =0; i < imax ; ++ i)
for ( int j =0; i < imax ; ++ j) ... // bad nested loop
for ( int i =0; i <= max ; ++ i) ... // off -by - one error ?
}
Still, over the years there were changes. Doug Gregor suggested a change to use C++0x concepts
which was nice and approved [Ottosen et al. 2007]. I remember him writing the proposal in my office
in Texas, but unfortunately, we had to back out that change when we removed C++0x concepts
(ğ6). In 2018, a small change was made to accommodate the infinite sequences supported by the
Ranges TS (ğ9.3.5).
4.2.3 Move Semantics. Traditionally, in C and C++, to get a large amount of data out of a function as
a result, you allocate it on the free store (heap, dynamic memory) and pass a pointer to it. Examples
are factory functions and functions returning containers (e.g., vectors and maps. This seemed
natural to the community and is suitably efficient. Unfortunately, it is one of the major sources of
explicit use of pointers, leading to notational inconvenience, explicit memory management, and
hard-to-find errors.
For years many experts had used łtrickeryž to eliminate this problem by using classes that were
handles and were passed around as simple values (often called value types). For example:
Matrix operator +( const Matrix &, const Matrix &);
void use ( const Matrix & m1 , const Matrix & m2 )

{
Matrix m3 = m1 + m2 ;
// ...
}
Here, operator+ offers the conventional mathematical notation and is an example of a factory
function returning a large object.
Passing Matrixes into a function by const-reference was (and is) conventional and efficient. The
problem is to return a Matrix by value without copying all the elements. As early as 1982, I had
partly addressed this problem with an optimization that simply allocated the return value on the
caller’s stack frame. This worked very well, but it was only an optimization technique and didn’t
handle more complex return statements. What users needed to return "large objects" by value was
a guarantee that large amounts of data were never copied.
To do provide that, it was observed that a "large object" typically was a handle to data on the
free store. To avoid copying of large amounts of data we then simply had to ensure that that the
constructor used to implement the return must copy the handle, rather than all the elements. The
C++11 solution to this problem looks like this:
class Matrix {
double * elements ; // pointer to the elements
// ...
public :
Matrix ( Matrix && a) // a move constructor
{
elements = a. elements ; // copy the handle
a. elements = nullptr ; // now a 's destructor has nothing to do
}
// ...
};
A move is preferred over a copy when the source of an initialization or an assignment is about to
be destroyed anyway: a move operation simply łstealsž the representation. The && indicates that
the constructor is a move constructor and Matrix&& is called an rvalue reference. The && notation
for an rvalue reference, called a forwarding reference when used for a template parameter, was
suggested by John Spicer in a 2002 meeting with Dave Abrahams and Howard Hinnant.
The Matrix example is particularly interesting because if we return a pointer from a Matrix
addition, the conventional mathematical notation (a+b) cannot be used.
The use of move semantics can have significant implications for efficiency: it eliminates expensive
temporary variables. For example:
Matrix mx = m1 + m2 + m3 ; // requires no temporary
string sx = s1 + s2 + s3 ; // requires no temporary
I added the string example because move semantics was immediately added to all standard-
library containers. This speeded up some C++98 programs without a source-code change.
Allowing the designer of a class to define move operations completes the control of object
lifetimes and resource management that started in 1979 with the introduction of constructors and
destructors. Move semantics is a cornerstone of C++’s model of resource management [Stroustrup
et al. 2015]. It’s the mechanism that enables simple and cheap movement of objects between scopes.
This important general point may have been obscured by an early emphasis on argument
passing, perfect forwarding, and smart pointers. Howard Hinnant, Dave Abrahams, and Peter
Dimov proposed the general version of move semantics in 2002 [Hinnant et al. 2004, 2002]:
łThe rvalue reference can be used to easily add move semantics to an existing class. By
this we mean that the copy constructor and assignment operator can be overloaded based
on whether the argument is an lvalue or an rvalue. When the argument is an rvalue, the
author of the class knows that he has a unique reference to the argument.ž
A prominent example was factory functions yielding łsmart pointersž:
template < class T , class A1 >
std :: shared_ptr <T > factory ( A1 && a1 )
{
return std :: shared_ptr <T >( new T ( std :: forward <A1 >( a1 )));
}
The (now) standard-library function forward tells the compiler to treat its argument as an rvalue
reference, so that a T’s move constructor is used to steal that argument (rather than T’s copy
constructor). It is basically a cast (explicit type conversion) to an rvalue reference.
In C++98, without rvalue references, such łsmart pointersž were very difficult to implement. In
C++11, the solution is simple [Hinnant et al. 2006]:
template < class T >
class clone_ptr
{
private :
T* ptr ;
public :
// ...
clone_ptr ( clone_ptr && p) // move constructor
: ptr (p. ptr ) // copy representation
{
p. ptr = 0; // " zero out " the source 's representation
}
clone_ptr & operator =( clone_ptr && p) // move assignment

{
std :: swap ( ptr , p. ptr );
return * this ; // destroy old value of the target
}
};
Soon the move semantics technique was applied to all the standard-library containers, such as
vector, string, and map. The shared_ptr and unique_ptr are smart, but still pointers. I prefer to
emphasize move constructors and move assignments that allow efficient movement of large objects
(represented as handles) from scope to scope.
The rvalue reference had a rough passage through the committee. Some people thought it likely
that the rvalue references and move semantics would not make it into C++11 because the notions
were novel and we didn’t have suitable terminology for them. Partly as a consequence of the
terminology problems [Miller 2010], the use of the term rvalue reference in the core language and in
the standard library diverged, thus making the draft standard inconsistent. At the Pittsburg meeting
in March 2010, I got involved in discussions about that in the Core Working Group (CWG). As the
CWG broke for lunch, it seemed to me that łwe were headed for an impasse or a mess or both.ž
Instead of going to lunch, I did an analysis of the problems and concluded that there were only two
fundamental concepts involved: has identity and can be moved from. From these two primitives, I
derived the conventional categories of lvalue and rvalue [Barron et al. 1963] as well as three new
ones needed to resolve our definitional problems. When the CWG returned, I presented my solution.
It was promptly accepted so that we could keep move semantics for C++11 [Stroustrup 2010a].
4.2.4 Resource-Management Pointers. C++11 offered łsmart pointersž (ğ4.2.4):
• shared_ptr ś representing shared ownership
• unique_ptr ś representing unique ownership (replacing the C++98 auto_ptr)
The addition of these resource-management łsmartž pointersž representing ownership had great
impact on programming style. For many, their use meant the end of resource leaks and a meaningful
decrease in dangling pointer problems. They were the most visible part of the effort to automate
resource management and minimize the use of raw pointers for that (ğ4.2.3).
shared_ptr is a conventional counted pointer: all shared pointers to an object share a counter.
When the last shared pointer to an object is destroyed, the object pointed to is destroyed. That’s a
simple, general, and effective form of garbage collection. It handles non-memory resources correctly
(ğ2.2.1). To deal with circular data structures, there is also a weak_ptr. It is often suboptimal, though.
People often used (and use) a shared_pointer simply to safely return data from a factory function:
shared_ptr < Blob > make_Blob ( Args a)
{
auto p = shared_ptr < Blob >( new Blob (a ));
// ... fill *p with lots of good stuff ...
return p;
}
When moving an object out of a function, the use count just goes from 1 to 2 and back to 1. In a
multi-threaded program, that typically is a slow operation involving synchronization. Also, when
naively used and/or implemented the use-count adds allocation and deallocation overhead.
As expected, shared_ptr quickly became very popular and seriously overused in places. Conse-
quently, unique_ptr that does not imply any overheads was provided. A unique_ptr has exclusive
ownership of the object it refers to and simply deletes that object when it itself is destroyed.
unique_ptr < Blob > make_Blob ( Args a)
{
auto p = unique_ptr < Blob >( new Blob (a ));
// ... fill *p with lots of good stuff ...
return p;
}
The shared_ptr and weak_ptr were the work of Peter Dimov [Dimov et al. 2003]. Howard
Hinnant contributed unique_ptr as an improvement on the C++98 auto_ptr [Hinnant et al. 2002].
Given that unique_ptr is a drop-in replacement for auto_ptr, it offered a rare opportunity for
(eventually) removing a flawed facility from the standard. The resource management pointers are
closely related to the work on move semantics, perfect forwarding, and rvalue references (ğ4.2.3).
The resource management pointers were (and are) extensively used to hold onto objects so that
exceptions (and the like) don’t cause a leak (ğ2.2). For example:
void old_use ( Args a)
{
auto q = new Blob (a );
// ...
if ( foo ) throw Bad (); // can leak
if ( bar ) return ; // can leak
// ...
delete q; // easy to forget
}
That old style using explicit new and delete is error-prone and not recommended in modern
C++ (e.g., the C++ Core Guidelines (ğ10.6)). Instead, we can write:
void newer_use ( Args a)
{
auto p = unique_ptr < Blob >( new Blob (a ));
// ...
if ( foo ) throw Bad (); // doesn 't leak
if ( bar ) return ; // doesn 't leak
// ...
}
This is shorter, safer, and quickly became very popular. However, łsmart pointersž are still
overused: łthey are smart but they are still pointers.ž Unless we actually need a pointer, simply
using a local variable is better still:
void simplest_use ( Args a)
{
Blob b(a );
// ...
if ( foo ) throw Bad (); // doesn 't leak
if ( bar ) return ; // doesn 't leak
// ...
}
The primary use for smart pointers for representing ownership of resources is object-oriented
programming where a pointer (or reference) is used to access objects for which the exact type is
unknown at compile time.
4.2.5 Uniform Initialization. For historical reasons, C++ has a variety of notations for initialization
and their semantics vary in surprising ways.
From C, C++ inherited three forms of initialization and added a fourth:
int x; // default initialization ( for static variables only )
int x = 7; // value initialization
int a [] = {7 ,8}; // aggregate initialization
string s; // initialization by default constructor

vector < int > v (10); // initialization by constructor
The notion used for initialization can depend on both the type of the object being initialized and
the context of the initialization. This is a mess and was recognized as such. For example, why could
we initialize a built-in array with a list, but not a vector?
int a [] = {7 ,8}; // OK
vector < int > v = {7 ,8}; // should work ( obviously , but it did not )
The last example bothered me a lot because it violates the fundamental C++ design aim of
providing equally good support for built-in and user-defined types. In particular, by providing
better support for array initialization than for vector, it encourages the use of the error-prone
built-in arrays.
From 2002, when the work on C++0x started, there were many discussions and proposals to
address parts of the problem from Daniel Gutson, Francis Glassborow, Alisdair Meredith, Bjarne
Stroustrup, and Gabriel Dos Reis. In 2005, Gabriel Dos Reis and I proposed a uniform initialization
notation that could be used for every type and have the same meaning everywhere in a program
[Stroustrup and Dos Reis 2005b]. This notation held the promise of great simplification of user
code and the elimination of many subtle errors. The notation was (and still is) based on the list
notation using braces. For example:
int a = {5}; // built - in type
int a [] {7 ,8}; // array
vector < int > v = {7 ,8}; // user - defined type with constructor
Braces ({}) are optional for single values and = is optional before a brace initializer list. For
uniformity the brace-style initialization is accepted in many places where C++98 didn’t allow it or
= initialization:
int f( vector < int >);
int i = f ({1 ,2 ,3}); // function argument
struct X {
vector < int > v;
int a [];
X () : v {1 ,2} , a {3 ,4} {} // member initializers
X( int );
// ...
}
vector < int >* p = new vector < int >{1 ,2 ,3 ,4}; // new expression
X x {}; // default initialization
template < typename T > int foo (T );

int z = foo (X {1}); // explicit construction
Many of these cases, such as an initializer list for an object created by new, simply couldn’t be
done using earlier notations.
Unfortunately, this ideal was only incompletely approximated, so we have a scheme that is only
almost uniform. Some people found the use of {...} łweirdž unless the ... was a homogeneous list,
others clung to the conventional C distinction between aggregates and non-aggregates, and many
were worried that lists without an explicit type marker would lead to ambiguities and errors. For
example, this was considered dangerous, but eventually accepted:
struct S { string s; int i; };
S foo (S s)
{
// ...
return { string {" foo "} ,13};
}
S x = foo ({ string {" alpha "} ,12.3});
In one case, the quest for uniform notation was defeated by a common practice. Consider:
vector < int > v1 (10); // 10 elements
vector < int > v2 {10}; // 10 elements or 1 element with the value 10?
vector < int > v3 {1 ,2 ,3 ,4 ,5}; // vector with 5 elements
There are many millions of lines of code using size initializers like vector<int> v1(10) and from
first principles vector<int> v2 {10} really is ambiguous. In a new language, I would not use a plain
integer to indicate a size, I’d introduce a specific type for that (e.g., Size or Extent); for example:
vector < int > v1 { Extent {10}}; // 10 elements with the default value 0
vector < int > v2 {10}; // 1 element with the value 10
However, C++ is not a new language, so we decided to give the initializer list interpretation
priority when choosing among constructors. That makes vector<int> v2 {10} a vector with one
element and the interpretation of {...} initializers consistent. However, that forces us to use the (...)
notation when we want to avoid the initializer-list constructor.
One problem with initialization is exactly that it is everywhere so that essentially all problems
with programs and the language rules manifest themselves in the context of initialization. Consider:
int x = 7.2; // traditional initialization
int y {7.2}; // brace initialization
From the introduction of floating-point numbers into C (about 1974), the value of x has been 7;
that is, 7.2 is implicitly truncated resulting in a loss of information. This is a source of errors. The
brace initialization does not allow narrowing conversions (here, the truncation). That’s good, but
makes it harder to upgrade old code:
double d = 7.2;
int x = d; // OK : truncates
int y {d }; // error
This is an example of a common problem. People want a simple upgrade path, but unless some
effort and change is needed, the result of a perfectly simple upgrade is that the old problems and
errors are preserved. Improving a language in wide use is harder than we tend to think.
After many heated debates and many modifications (not all of which I consider improvements),
uniform initialization was approved for C++0x in 2008 [Stroustrup 2008b].
As ever, notation was a contentious issue, but eventually we agreed to have a standard-library
type initializer_list to be used as the argument type of initializer-list constructors. For example:
template < typename T > class vector {

public :
vector ( initializer_list <T >); // initializer - list constructor
// ...
};
vector < int > v3 {1 ,2 ,3 ,4 ,5}; // vector with 5 elements
Sadly, uniform initialization ({}-initialization) isn’t as widely used as I had hoped for. People seem
to prefer the familiar notations and familiar bugs. I seem to have fallen prey to the N+1 problem: You
have N incompatible and incomplete solutions to a problem, so you add a new and better solution.
Unfortunately, the original N solutions don’t go away so now you have N+1 solutions. To be fair,
there were subtle problems beyond the scope of this paper that were only slowly being remedied
(in C++14, C++17, and C++20). My impression is that generic programming and a general push for
more concise notation is slowly increasing the appeal of uniform initialization. All standard-library
containers (e.g., vector) have initializer-list constructors.
4.2.6 nullptr. In C and C++, the literal 0 denotes the null pointer if assigned to a pointer or
compared to a pointer. More confusingly, any integer constant expression that evaluates to zero
denotes the null pointer if assigned to a pointer or compared to one. For example:
int * p = 99 -55 -44; // the null pointer

int * q = 2; // error : 2 is an int , not a pointer
This has annoyed and confused many, so there is a standard-library macro NULL (adopted from
C), which in standard C++ is defined as 0. Some compilers warn against int* p = 0; but you still
can’t overload a function for a pointer and an integer without getting ambiguity for 0.
This could be easily fixed by giving a name to the null pointer, but somehow nobody had gotten
around to making a proposal that people could agree to. Sometime in 2003, I was attending a
meeting over the phone where people discussed how to name the null pointer. Suggestions, such
as NULL, null, nil, nullptr, and 0p, were among the alternatives. As usual, all short and łnicež
names had been used thousands of times and thus could not be used without breaking millions
of lines of code. Being a bit bored having heard variants of this discussion dozens of times, I was
listening with only half an ear. People were saying variations of null pointer, null ptr, nullputter. I
woke up and said: łYou are all saying nullptr; I don’t think I have seen that in code.ž
Herb Sutter and I wrote up that suggestion [Sutter and Stroustrup 2003] and it passed relatively
easily in 2007 (after only four minor revisions), so now we can say:
int p0 = nullptr ;
int * p1 = 99 -55 -44; // OK for compatibility
int * p2 = NULL ; // OK for compatibility
int f( char *);

int f( int );
int x1 = f( nullptr ); // f( char *)

int x2 = f (0); // f( int )
I pronounce nullptr as łnull pointer.ž

I still think that defining the macro NULL to be nullptr would have eliminated a significant
class of problems, but the committee deemed that change too radical.
4.2.7 constexpr Functions. In 2003, Gabriel Dos Reis and I proposed a radically different and
significant better mechanism for doing constant expression evaluation in C++ [Dos Reis 2003].
People were using (typeless) macros and the impoverished C definition of constant expressions.
Others were beginning to use template metaprogramming to compute values (ğ10.5.2). łThis is
tedious and error-pronež [Dos Reis and Stroustrup 2010]. Our aim was to
• Make compile-time computation type safe
• Generally, improve efficiency by moving computation to compile-time
• Support embedded systems programming (especially ROMs)
• Directly support metaprogramming (as opposed to template metaprogramming (ğ10.5.2))
• Make compile-time programming very similar to łordinary programmingž
The idea was simple: allow functions prefixed with constexpr to be used in constant expressions
and also allow simple user-defined types, called literal types, to be used in constant expressions. A
literal type is basically a type for which all operations are constexpr.
Consider an application where we for efficiency, ROMability, or reliability wanted to use a unit
system [Dos Reis and Stroustrup 2010]:
struct LengthInKM {
constexpr explicit LengthInKM ( double d) : val (d) { }
constexpr double getValue () { return val ; }
private :
double val ;
};
struct LengthInMile {
constexpr explicit LengthInMile ( double d) : val (d) { }
constexpr double getValue () { return val ; }
constexpr operator LengthInKM () { return LengthInKM (1.609344 * val ); }
private :
double val ;
};
Given that, we can make a table of constants without fear of unit errors or conversion errors:
LengthInKM marks [] = { LengthInMile (2.3) , LengthInMile (0.76) };
The traditional solution would either require more run time or have the programmer work out
the values on a doodle pad. My interest in unit systems was stimulated by the loss of the 1999 Mars
Climate Orbiter due to an undetected unit mismatch [Stephenson et al. 1999].
A constexpr function can be evaluated at compile time, so it cannot access non-local objects
(they don’t exist at compile-time) so C++ acquired a kind of pure functions.
Why did we require that programmers should use constexpr to mark functions that can be
executed at compile time? In principle, the compiler can figure out what can be computed at
compile time, but without an annotation, users would be at the mercy of variations in cleverness of
compilers and a compiler would need to keep bodies of all functions around łforeverž in case they
were needed for constant expression evaluation. We picked the word constexpr because it was
sufficiently mnemonic, yet łsufficiently oddž not to break existing code.
In a few places, C++ requires constant expressions (e.g., array bounds and case labels). In addition,
we can require a variable to be initialized at compile time by declaring it constexpr:
constexpr LengthInKM marks [] = { LengthInMile (2.3) , LengthInMile (0.76) };
void f( int x)
{
int y1 = x;
constexpr int y2 = x; // error : x is not a constant
constexpr int y3 = 77; // OK
}
The early discussions focused on simple examples of performance and embedded systems. Only
much later (from about 2015) did constexpr functions become a mainstay of metaprogramming
(ğ10.5.2). C++14 allowed local variables, and therefore loops, to be used in constexpr functions;
before that they had to be purely functional. C++20 (finally, about 10 years after it was first proposed)
allows literal types as value template parameter types [Maurer 2012]. Thus, C++20 will be very
close to the original (1979) goal of being able to use user-defined types wherever built-in types can
be used (ğ2.1).
The constexpr functions quickly became wildly popular. They are all over the C++14, C++17,
and C++20 standard libraries and there is a constant stream of suggestions for allowing more
language constructs in constexpr functions, for applying constexpr to more functions in the
standard library, and for more support for compile-time computation (ğ9.3.3).
However, constexpr functions were not easy to get into the standard. They were repeatedly
deemed useless and unimplementable. Implementing constexpr functions obviously required work
on older compilers, but soon the writers of all major compiler proved the łunimplementablež claims
wrong. The discussions about constexpr were about the most heated and unpleasant ever. It took
four years to get the initial version through the standards process [Dos Reis and Stroustrup 2007]
and twelve more to complete the process.
4.2.8 User-Defined Literals. łUser-defined literalsž is very a small feature. However, it fits into the
general thrust to make the support of user-defined types similar to what built-in types receive.
Built-in types have literals, for example, 10 is an integer and 10.9 is a floating-point number. I
tried to convince people that the equivalent for user-defined types was explicit use of constructors;
for example, complex<double>(1.2,3.4) is the complex equivalent to a literal. However, many
didn’t consider that good enough: the notation is not conventional and there was no guarantee
that the constructor would be evaluated at compile time (though it was from the earliest days). For
complex, people wanted 1.2+3.4i.
Compared to other problems, this didn’t seem important, so nothing happened for decades. One
day in 2006, David Vandevoorde (EDG), Mike Wong (IBM), and I were out for a good dinner in a
Chinese restaurant in Berlin. We were talking shop, and a design emerged on a paper napkin. The
discussion was prompted by the need for suffixes in an IBM proposal for decimal-floating-point that
eventually became its own international standard [(editor) 2007]. After significant mutation, that
design became user-defined literals (often referred to as UDLs) in 2008 [McIntosh et al. 2008]. The
significant development that made UDLs interesting just then was the progress of the constexpr
function proposal (ğ4.2.7). Given that, we could guarantee compile-time evaluation.
As is often the case, finding an acceptable notation was a problem. We decided that the cryptic
operator"" was acceptable as a notation for literal operator, after all "" is a literal. Then, ""x is
the notation for a literal followed by the suffix x. Given that and an Imaginary type for use with
complex numbers, we can define:
constexpr Imaginary operator "" i( long double x) { return Imaginary (x ); }
Now, 3.4i is an Imaginary and 1.2+3.4i is complex<double>(1.2,3.4). Mission accomplished!
The language technical details are quite quirky, but I consider that reasonable for a relatively
rarely-used feature. Even when UDLs are used heavily, there are few definitions of literal operators.
What matters most is the elegance and ease of use of the suffixes. For many types, it is important
that the conversion from the built-in types to user-defined types can be done at compile time.
Naturally, people used UDLs to define literals for many useful types, including several from the
standard library (e.g., s for seconds and s for std::string). There was discussion about support for
binary literals, and Peter Sommerlad (HSR) suggested what I consider a candidate for a łbest abuse
of the rulesž award: define operator""_01(long int) appropriately and 101010_01 is a binary
literal! When the astonishment and laughter had died down, the committee decided to define binary
literals in the language itself and use 0b as the prefix meaning łbinaryž (e.g., 0b101010) in analogy
to the use of 0x to mean łhexadecimalž (e.g., 0xDEADBEEF).
4.2.9 Raw literals. This is a rare simple feature with the single purpose of providing an alternative
to an error-prone notation. Like C, C++ uses the backslash as an escape character. This means that
to represent a backslash in a string literal, you need to use a double backslash (\\) and when you
want a double quote in a string, you need to use \". However, the usual regular expression patterns
use backslashes extensively and double quotes are common, so patterns quickly get messy and
error prone. Consider a simple example (a US postal code):
regex pattern1 {"\\ w {2}\\ s *\\ d {5}( -\\ d {4})?"}; // ordinary literal
regex pattern2 {R "(\ w {2}\ s *\ d {5}( -\ d {4})?)"}; // raw literal
The two patterns are identical. The raw literal R"( ... )" bracketing can be elaborated to hold
more complicated patterns, but when you use regular expressions (ğ4.6) the simplest version is
sufficient and a major convenience. Providing raw literals is a minor detail, of course, but (similar
to digit separators (ğ5.1) much loved by people who need many literals.
The raw literals were proposed by Beman Dawes in 2006 [Dawes 2006] based on experience
with boost::regex [Maddock 2002].
4.2.10 Attributes. Attributes provide a way of associating essentially arbitrary information with
an entity in a program. For example:
[[ noreturn ]] void forever ()

{
for (;;) {
do_work ();
wait (10 s );
}
}
The [[noreturn]] informs a compiler or other tool that forever() is never supposed to return
so that it can suppress warnings against the lack of a return. An attribute is bracketed by [[ ... ]].
Attributes were first proposed in 2007 by Alisdair Meredith [Meredith 2007], the head of the
Library Working Group, to eliminate incompatibilities among vendor attribute notations (e.g.,
__declspecs and __attribute__s) that complicated library implementation. In response to that,
Jens Maurer and Michael Wong did an analysis of the problems and proposed the [[ ... ]] syntax
based on an implementation that Michael had done for IBM’s XL compiler [Maurer and Wong
2007]. In addition to standardizing a multitude of non-portable practices, this would also allow
language extension to be done with fewer keywords and new keywords are always controversial.
The proposal mentioned possible uses: explicit syntax to override virtual functions, dynamic
libraries, user-controlled garbage collection, thread local storage, controlling alignment, identifying
classes that are łplain old data,ž defaulted and deleted functions, strong enums, strong typedefs,
pure side-effect free functions, final overrides, sealed classes, fine grained control over concurrency,
support for runtime reflection, and lightweight support for contract programming. Many more
were mentioned in early discussions.
łAttributesž is certainly a feature that makes certain things simpler, but I am not sure that it
encourages good design or that the łthingsž it makes simpler are always what yield the most
benefits. I had visions of attributes opening the floodgates for a mass of unrelated, half understood,
minor features. Instead of proposing a feature to WG21, anyone would be able to add a attribute
to a compiler and lobby for its adoption everywhere. Many programmers just love such minor
features. The elimination of the need to introduce keywords and to modify the grammar would
lower the barrier to entry. So would the inevitable insufficient attention to feature interactions and
overlapping, but incompatible, similar features in different compilers. This had already happened
with proprietary extensions, but I considered those inevitable, localized, and often transient.
To attempt to limit potential damage, we decided that an attribute should imply no change of a
program’s semantics. That is, a compiler that ignored an attribute would do no harm. Over the
years, this łrulež almost worked. Most standard attributes ś though not all ś have no semantic
effects even though they help with optimizations and error detection.
In the end, most of the initially suggested uses of attributes were addressed by ordinary syntax
and language rules.
C++11 added the standard attributes [[noreturn]] and [[carries_dependency]].
C++17 added [[fallthrough]], [[nodiscard]], and [[maybe_unused]].
C++20 added [[likely]], [[unlikely]], [[deprecated(message)]], [[no_unique_address]],
and [[using: ...]].
I still see attribute proliferation as a potential danger, but so far the floodgates haven’t opened.
The C++ standard-library uses attributes liberally; [[nodiscard]] is particularly popular, especially
to protect against potential leaks from unused return values that are resource handles.
The attribute syntax was adopted for the (failed) C++20 contract design (ğ9.6.1).
4.2.11 Garbage Collection. From the earliest days of C++, optional garbage collection was consid-
ered (for a variety of definitions of łoptionalž) [Stroustrup 1993, 2007]. After much debate, C++11
offered an interface to conservative garbage collectors designed by Mike Spertus and Hans-J Boehm
[Boehm and Spertus 2005; Boehm et al. 2008]. However, few people noticed that and fewer used
garbage collection (despite the availability of good collectors). The approach was to
support both garbage collected implementations and reachability-based leak detectors.
This is done by giving undefined behavior to programs that łhide a pointerž by, for example,
xor-ing it with another value, and then later turn it back into an ordinary pointer and
dereference it. [Boehm et al. 2008]
This work became a boon to the precise specification of the semantics of C++ and a few uses of
garbage collection in C++ exist (e.g., by Macaulay2 [Eisenbud et al. 2001; Macaulay2 2005ś2020]).
However, garbage collectors don’t address non-memory resources and the C++ community in
general chose to use combinations of resource-management pointers (ğ4.2.4) and RAII (ğ2.2.1).
4.3 C++11: Improving Support for Generic Programming

Generic programming (and its offspring template metaprogramming (ğ10.5.2)) became a run-away
success with C++98. Its use seriously strained the language and inadequate language support led
to baroque programming techniques and horrendous error messages. It is a testimony of the utility
of generic programming and metaprogramming that so many sane programmers were willing to
suffer to gain the benefits. Those benefits were
• Flexibility beyond what could be obtained in C-style or object-oriented style
• Cleaner code
• Finer granularity of static type checking
• Efficiency (mostly from inlining, having the compiler see code from several sources at once,
and better type checking)
The main new features supporting generic programming in C++11 are:
• ğ4.3.1: Lambdas
• ğ4.3.2: Variadic templates
• ğ4.3.3: template aliases
• ğ4.3.4: tuples
• ğ4.2.5: uniform initialization
Concepts should have been the centerpiece of improved support for generic programming in
C++11, but that didn’t happen (ğ6.2.6). We had to wait until C++20 (ğ6.4).
4.3.1 Lambda. BCPL allowed blocks of code as expressions, but to save space in the compiler,
Dennis Ritchie didn’t adopt this feature into C. I followed C in this, but added inline functions
to (re)gain the ability to get code executed without the cost of a function call. However, this still
didn’t offer the ability to
• Write code exactly where it was needed (usually as a function argument).
• Access the context of the code from within the code.
During the work on C++98, there had been proposals for local functions to address that second
point, but they were voted down as a likely bug source.
Instead of allowing function definitions inside functions, C++ relied on functions defined inside
classes. That allowed the context of a function to be represented as class members and function
objects became very popular. A function object is simply a class with an application operator
(operator()()). It was a very efficient and effective technique and I (and others) expressed the
opinion that named objects resulted in clearer code than unnamed operations. However, that is
only the case when something can be given a reasonable name outside the context of its use and
especially if it is used more than once.
In 2002, Jaakko Järvi and Gary Powell wrote the Boost lambda library [Järvi and Powell 2002]
that allowed us to write something like this
find_if (v. begin () , v. end () , _1 <i ); // find element with a value less than i
Here, _1 is the name of a first argument to the code fragment _1<i and i is a variable in the
enclosing scope. The _1<i expands into a function object, where i is bound to a reference and _1
becomes the argument to an operator()():
struct Less_than {
int & i;
Less_than ( int & ii ) :i( ii ) {} // bind to i
bool operator ()( int x) { return x <i; } // compare to argument
}
The lambda library was a masterpiece of early template metaprogramming (ğ10.5.2) and very
convenient and popular. Unfortunately, it wasn’t particularly efficient. For years, I tracked its
performance relative to hand-coded equivalents and found a fairly consistent 2.5 times overhead. I
could not recommend something that was convenient, but slow. Doing so would damage C++’s
reputation as a language for efficient code. Obviously, the slowdown was partly a result of poor
optimization, but for this and other reasons a group of us, led by Jaakko Järvi, decided to propose
lambda expressions as a language feature [Willcock et al. 2006]. For example:
template < typename Oper >
void g( Oper op )
{
int xx = op (7);
// ...
}
void f ()
{
int y = 3;
g ( < >( int x) -> int { return x + y ;}); // call g () with a lambda argument
}
Here, xx will become 3+7.

The <> was the lambda introducer. We did not dare to propose a new keyword.
That proposal generated quite a lot of excitement and many lively discussions:
• Should the syntax be expressive or terse?
• What names from which scope can a lambda refer to? [Crowl 2009].
• Should the function object generated from a lambda be mutable? Not by default.
• Can a lambda be polymorphic? Not until C++14 (ğ5.4).
• What is the type of a lambda? A unique type unless it basically is a local function.
• Can a lambda have a name? No. If you need a name just assign it to a variable.
• Are names bound by value or by reference? You have a choice.
• Can variables be moved into a lambda (as opposed to copied)? Not until C++14 (ğ5).
• Would the syntax clash with various non-standard extensions? (Not seriously).
By the time lambdas were approved in 2009, the syntax had mutated and become more conven-
tional [Vandevoorde 2009]:
void abssort ( float *x , unsigned N)
{
std :: sort (x , x+N ,
[]( float a , float b) { return std :: abs (a) < std :: abs (b ); }
);
}
The switch from <> to [] was suggested by Herb Sutter and implemented by Jonathan Caves.
The change was in parts motivated by the need for a simple way of designating which names from
the surrounding scopes could be used by the lambda. Herb Sutter recalls:
I was motivated by needing lambdas for my parallel algorithms project ... and seeing
that the lambdas EWG had adopted were just butt-ugly to use and poorly designed from
a syntax consistency/cleanness point of view (e.g., captures appeared in two separate
places, syntax elements were used inconsistently, the ordering was wrong because all
the łconstructorž elements should come first followed by the łoperatorž elements that get
invoked later, and various smaller issues).
By default, a lambda cannot refer to names in the local environment, so they are just ordinary
functions. However, we can specify that a lambda should łcapturež some or all of the variables from
its environment. Callbacks are a common use case for lambdas because often the action is unique
and need to use some information from the context of the code installing the callback. Consider:
void test ()
{
string s;
// ... compute a suitable value for s ...
w. foo_callback ([& s ]( int i ){ do_foo (i ,s ); });
w. bar_callback ([= s ]( double d ){ return do_bar (d ,s ); });
}
The [&s] says that do_foo(i,s) can use s and that s is passed (łcapturedž) by reference. The [=s]
says that do_bar(d,s) can use s and that s is passed by value. If the callback is invoked on the same
thread as test, [&s] capture is potentially efficient because s is not copied. If the callback is invoked
on a different thread, [&s] capture could be a disaster because s could go out of scope before it
was used; in that case, we want a copy. A [=] capture list means łcopy all local variables into the
lambda.ž A [&] capture list means łthe lambda can refer to all local variables by referencež and
implies that the lambda can be implemented as simply a local function. The flexibility of the capture
mechanism has proven very valuable. The capture mechanism allows control of what names can
be referred to from a lambda, and how. This is an answer to the 1990s worries about local functions
being error prone.
The implementation of a lambda is basically that the compiler builds a suitable function object
and passes it. The captured local variables become members initialized by a constructor and the
lambda’s code becomes the function object’s application operator. For example, the bar_callback
becomes:
struct __XYZ {
string s;
__XYZ ( const string & ss ) : s{ ss } {}
int operator ()( double d) { return do_bar (d ,s ); }
};
The return type of a lambda can be deduced from its return statement. If there is no return
statement, the lambda doesn’t return anything.
I classify lambdas as support for generic programming because one of the most common uses ś
and a major motivation ś was for the use as arguments to STL algorithms:
// sort in descending order :
sort (v. begin () , v. end () ,[]( int x , int y) { return x >y; });
As such, lambdas added significantly to the attraction of generic programming.

After C++11, C++14 added generic lambdas (ğ5.4) and move capture (ğ5).
4.3.2 Variadic Templates. In 2004, Douglas Gregor, Jaakko Järvi, and Gary Powell (all then at
Indiana University) proposed a feature called variadic templates [Gregor et al. 2004] to
łdirectly addresses two problems:
• The inability to instantiate class and function templates with an arbitrarily-long list of
template parameters.
• The inability to pass an arbitrary number of arguments to a function in a type-safe
manner.ž
These are important goals, but I initially found the solution overly complex, the notation too cryptic,
and the programming style too recursive for my taste. However, after a brilliant presentation by
Doug Gregor in 2004, I changed my mind and backed the proposal to the hilt so that variadic
templates had a relatively smooth passage through the committee. Part of what convinced me was
measurements of compile times comparing variadic templates with (then) current workarounds.
Variadic templates was a major (sometimes 20 times) improvement over what was an increasingly
serious problem as template metaprogramming was beginning to see major use (ğ10.5.2). Unfortu-
nately, variadic templates became so popular and an essential part of the C++ standard library, so
the compile-time problem reappeared. However, that penalty for success was (then) still in the far
future.
The basic idea is to build up a parameter pack recursively and then use it by another recursive
pass. The recursive technique was necessary because each element of a parameter pack has its own
type (and size).
Consider an implementation of printf that can handle every type that can be output using the
standard-library iostream output operator << [Gregor 2006]:
To build our type-safe printf(), we use the following strategy: write out the string
up until the first format specifier is reached, print its corresponding value, then call
printf() recursively to print the rest of the string and remaining values.
template < typename T , typename ... Args >
void printf ( const char * s , const T& value , const Args &... args )
{
while (* s) {
if (* s == '%' && *++ s != '% ') { // ignore the char that follows
// the '% ': we already know the type !
std :: cout << value ;
return printf (++ s , args ...);
}
std :: cout << *s ++;
}
throw std :: runtime error (" extra arguments provided to printf ");
}
The <typename T, typename... Args> specifies a traditional list with a head (T) and a tail
(Args). Each call handles the head and then calls itself with the tail. Ordinary characters are simply
printed and the format character % indicates that an argument is to be printed. The test case offered
by Doug (then a resident of Indiana) was:
const char * msg = " The value of %s is about %g ( unless you live in %s ).\ n ";
printf ( msg , std :: string (" pi ") , 3.14159 , " Indiana ");
That prints
The value of pi is about 3.14159 ( unless you live in Indiana ).
One nice thing about this implementation is that, in contrast to the standard printf, user-defined
types are handled as well as built-in ones. By using << it also avoids errors from mismatch between
the type indicator and the argument type, e.g., printf("%g %c","Hello",7.2).
The technique illustrated by this printf is a basis for the C++20 format (ğ9.3.7).
The weakness of variadic templates is that they can easily lead to code bloat as N parameters
imply N instantiations of the template.
4.3.3 Aliases. The C mechanism for defining an alias for a type is a typedef. For example:
typedef double (* pf )( int ); // pf is a pointer to a function
// taking an int and returning a double
This is a bit convoluted, but type aliases are very useful and ubiquitous in C and C++ code. From
the first days of C++ templates, people were considering whether we could have typedef templates
and if so, what they would be. In 2002, Herb Sutter proposed a solution [Sutter 2002]:
template < typename A , typename B > class X { /* ... */ };
template < typename T > typedef X <T , int > Xi ; // the alias
Xi < double > Ddi ; // equivalent to X < double , int >
Building on that and after lengthy email reflector (archived email group) discussions, Gabriel
Dos Reis (then at INRIA in France) and Matt Marcus (Adobe) resolved some nasty problems related
to specialization and introduced a simplified syntax for what David Vandevoorde named alias
templates [Dos Reis and Marcus 2003]. For example:
template < typename T , typename A > class MyVector { /* ... */};
template < typename T > using Vec = MyVector <T , MyAlloc <T > >;
The using syntax, where the name being introduce always comes first, was my suggestion.
Finally, together with Gabriel Dos Reis, I generalized this to an (almost) complete aliasing
mechanism, that was accepted [Stroustrup and Dos Reis 2003c]. This gave people a choice of
notation even where templates are not involved:
typedef double (* analysis_fp )( const vector < Student_info >&);
using analysis_fp = double (*)( const vector < Student_info >&);
Type and template aliases are key to some of the most effective techniques for zero-overhead
abstraction and modularization. An alias allows a user to refer to a set of standard names while
various implementations use their own (differing) implementation techniques and names. That
way we can have true zero-overhead abstractions while maintaining convenient user interfaces.
Consider a real-world example from a communications library (using the concepts TS [Sutton 2017]
and C++20 notational simplifications):
template < InputTransport Transport , MessageDecoder MessageAdapter >
class InputChannel {
public :
using InputMessage = MessageAdapter :: InputMessage < Transport :: InputBuffer >;
using MessageCallback = function < void ( InputMessage &&) >;
using ErrorCallback = function < void ( const error_code &) >;
// ...
};
We have found concepts and aliases invaluable for managing such composition at scale.
The user-interface for InputChannel primarily consists of the three aliases InputMessage,
MessageCallback, and ErrorCallback, that are initialized from its template arguments.
The InputChannel needs to initialize its transport layer, represented by a Transport object.
However, the InputChannel should not know about the details of the transport implementation,
so it cannot directly initialize its Transport member. Variadic templates (ğ4.3.2) come to the rescue:
template < InputTransport Transport , MessageDecoder MessageAdapter >

class InputChannel {
public :
template < typename ... TransportArgs >
InputChannel ( TransportArgs &&... transportArgs )
: _transport { forward < TransportArgs >( transportArgs )... }
{}
// ...
Transport _transport ;
};
Without variadic templates, we would need to either define a common interface for initialization
of transports or expose the transport to users.
I consider this an excellent example of how the C++11 features (plus concepts) combine to
provide an elegant zero-overhead solution to a hard problem.
4.3.4 tuples. C++98 had a pair<T,U> template; it was mainly used for returning pairs of values,
such as two iterators or a pointer and a success indicator. In 2002, with reference to Haskell, ML,
Python, and Eiffel, Jaakko Järvi proposed to generalize this idea to a tuple [Järvi 2002]:
łTuples are fixed-size heterogeneous containers. They are a general-purpose utility, adding
to the expressiveness of the language. Some examples of common uses for tuple types are:
• Return types for functions that need to have more than one return type.
• Grouping related types or objects (such as entries in parameter lists) into single entities.
• Simultaneous assignment of multiple values.ž
For a specific purpose, a class with appropriate names for members and semantic relations
among members is usually best. Alisdair Meredith forcefully presented that view in the committee,
discouraging us from overusing unnamed types in interfaces. However, when writing generic
code, bundling values together in a tuple and manipulating that tuple as an entity often simplify
implementation. Tuples are most useful for intermediate groupings that are not worthy of a name
and not worth designing a class for.
For example, consider a matrix decomposition that simply returns three values:
auto SVD ( const Matrix & A) -> tuple < Matrix , Vector , Matrix >
{
Matrix U , V;
Vector S;
// ...
return make_tuple (U ,S ,V );
};
void use ()
{
Matrix A , U , V;
Vector S;
// ...
tie (U ,S ,V) = SVD (A ); // using the tuple form
}
Here, make_tuple() is a standard-library function that makes a tuple with element types
deduced from its function arguments and tie() is a standard-library function that assigns a tuple’s
members to named variables.
In C++17 using structured bindings (ğ8.2), this example reduces to:

auto SVD ( const Matrix & A) -> tuple < Matrix , Vector , Matrix >
{
Matrix U , V;
Vector S;
// ...
return {U ,S ,V };
};
void use ()
{
Matrix A;
// ...
auto [U ,S ,V] = SVD (A ); // using the tuple form and structured bindings
}
A further notational simplification was proposed for C++20 [Spertus 2018], but didn’t make it
on time:
tuple SVD ( const Matrix & A) // deduce the tuple template arguments
// from the return statement
{
Matrix U , V;
Vector S;
// ...
return {U ,S ,V };
};
Why is tuple not a language feature? I don’t remember that question being asked at the time,
though someone must have thought of it. It has long (since 1979) been the policy not to add a
feature to C++ if we could reasonably add it as a library; if not, improve the abstraction mechanisms
to make it possible. There are obvious advantages to this policy:
• It is usually far easier to experiment with a library than with a language feature, so we get
better feedback sooner.
• A library can see serious use long before all compilers can be upgraded to support the new
feature.
• Improvements to the abstraction mechanisms (classes, templates, etc.) help beyond the
immediate problem.
For tuple we had boost::tuple to build on and people were proud of the clever implementations.
There appears not to be run-time efficiency reasons to prefer a language implementation over a
library implementation in this case. That’s somewhat impressive.
Parameter packs is an example of tuples provided with a compiler-supported interface (ğ4.3.2).
Tuples are extensively used in libraries interfacing C++ with other languages (e.g., Python).
4.4 C++11: Increase Static Type Safety

There are two reasons to rely on static type safety
• Make it clear what is intended
ś to help the programmer directly express ideas
ś to help the compiler catch more errors
• Help the compiler generate better code.
The second point is a consequence of the first. Inspired by Simula, my aim for C++ was to
provide a flexible and extensible static type system for C++. The aim is not just type safety, but
the ability to directly express fine-grained distinctions, such as physical unit checking (ğ4.2.7). A
program written exclusively using built-in types, such as integers and floating-point types, can be
type-safe without delivering significant benefits. Such code doesn’t directly express the concepts of
the application. In particular, an int or a string can represent just about anything so that passing
one doesn’t give a clue about the semantics of the value passed.
The C++11 improvements relating directly to type safety are:
• Type-safe interface to threading and locking ś avoid POSIX and Windows reliance on void**
and macros in concurrent code (ğ4.1.2)
• Range-for ś avoid erroneous specification of ranges (ğ4.2.2)
• Move semantics ś addresses the problem of overuse of pointers (ğ4.2.3)
• Resource management pointers (unique_ptr and shared_ptr (ğ4.2.4))
• uniform initialization ś make initialization more general, more uniform, and safer (ğ4.2.5)
• constexpr ś eliminate many uses of (untyped and unscoped) macros (ğ4.2.7)
• User-defined literals ś make user-defined types more like built-in types (ğ4.2.8)
• enum classes ś eliminate some weakly-typed practices involving integer constants
• std::array ś avoid unsafe łdecayž of built-in arrays to pointers
It has repeatedly been suggested that the committee should improve type safety by banning
unsafe features (e.g., abandoning łC-isms,ž such as built-in arrays and casts). However, attempts to
remove features (to łdeprecatež them) have repeatedly failed as users ignored the warnings about
removal and insisted that their implementation providers continue to supply them. A more viable
approach seems to be to provide guidelines for use and the means to enforce them while leaving
the standard itself compatible with earlier standards (ğ10.6).
4.5 C++11: Support for Library Building

Foundational libraries for C++ are often designed to compete in performance and ease-of-use with
built-in facilities in C++ and other languages. This is where subtleties of lookup-rules, overload
resolution, access control, template instantiation rules, and more combine to yield great expressive
power but also to expose fearsome complexity.
4.5.1 Implementation Techniques. Some implementation techniques are essentially łdark artsž to
which non-experts should not be exposed. Most programmers can happily write good C++ for years
without knowing the sophisticated trickery and esoteric techniques. Unfortunately, beginners flock
to examine the most horrendously specialized code and take great pride in explaining it (often
erroneously) to others. Bloggers and speakers enhance their reputations by showing off hair-raising
examples. This is a major source of C++’s reputation for complexity. In other languages, such
optimization opportunities are either not offered or the trickery is hidden inside optimizers.
I cannot go into details here, so I will mention just one technique that emerged as key during the
development of C++11 and has become almost universal in template-based libraries (including the
C++ standard library) known by its odd acronym: SFINAE (Substitution Failure Is Not An Error).
How do you provide an operation if and only if some predicate is true? Concepts provide that
for C++20 (and have been available in GCC since 2015), but in the early 2000s, people had to rely
on obscure language rules. For example:
template < typename T , typename U >
struct pair {
T first ;
U second ;
// ...
enable_if < is_copy_assignable <T >:: value
&& is_copy_assignable <U >:: value , pair & >:: type
operator =( const pair &);
// ...
};
Now, pair has a copy assignment if and only if both of its members have one. This is extraordinar-
ily ugly, but ś in the absence of concepts ś also extraordinarily useful for defining and implementing
foundational libraries.
The idea is that enable_if<...,pair&>::type will become plain pair& if the members have copy
assignments and fail to instantiate otherwise (because enable_if didn’t provide a return type for
the assignment). This is where SFINAE comes in: failure to instantiate is not an error; the result of
failure is as if the whole declaration wasn’t there.
The is_copy_assignable is a type trait. C++11 provided dozens of such traits to allow pro-
grammer to inquire about the properties of types at compile time.
The enable_if metafunction was pioneered by Boost and is part of C++11. A plausible imple-
mentation is
template < bool B , typename T = void >
struct enable_if {}; // the false case : no mention of 'type '

struct enable_if { typedef T type ; }; // type T
The precise rules for SFINAE are very subtle and hard to craft, but under steady pressure from
users, they became simpler and more general during the development of C++11. As a side effect,
this significantly improved the internals of compilers as they had to be able to back out of failed
template instantiations without side effects. This seriously discouraged their use of non-local state.
4.5.2 Metaprogramming Support. The first decade of the 2000s were a bit like the Wild West for
metaprogramming in C++ with new techniques and applications being tried with essentially no
support beyond the basic template mechanisms. Those mechanisms were sorely stressed. Error
messages were atrocious, compile times often extraordinarily long, and it was easy to exhaust
the compiler’s resources (e.g., memory, recursion depth, and identifier length). Also, individuals
repeatedly re-discovered problems and re-invented basic techniques. Clearly, better support was
needed. Attempts to help took two complementary (at least in theory) tracks:
• Language: concepts (ğ6), compile-time functions (ğ4.2.7), lambdas (ğ4.3.1), template aliases
(ğ4.3.3), and more precise specification of template instantiation (ğ4.5.1).
• Standard library: tuples (ğ4.3.4), type traits (ğ4.5.1), and enable_if (ğ4.5.1).
Unfortunately, concepts failed for C++11 (ğ6.2) leaving the field open for (often horribly complex
and error-prone) workarounds, typically involving traits and enable_if (ğ4.5.1).
4.5.3 noexcept Specifications. The original exception design did not have any way to indicate that
an exception might be thrown from a function. I still consider that the correct design. To get excep-
tions accepted for C++98, we had to add exception specifications, a way of listing which exceptions
a function could throw [Stroustrup 1993]. Their use was optional and they were checked at run time.
As I feared, that led to maintenance problems, run-time overhead as an exception was repeatedly
checked along the unwinding path, and bloated source code. For C++11, exception specifications
were deprecated [Gregor 2010] and for C++17, we finally removed them (unanimously).
There was always a group of people who wanted compile-time checks of which exceptions a
function could throw. That, of course, works well in type theory, with small programs, with fast
compilers, and with full control of the source code. The committee repeatedly rejected that idea on
the grounds that it doesn’t scale to million-line programs developed and maintained by dozens (or
more) organizations [Stroustrup 1994]. See also (ğ7.4).
Without exception specifications, library implementers faced a performance problem: In many
important cases, a library implementer needs to know if a copy operation can throw. If it can, taking
a copy is necessary to avoid leaving an invalid object behind (violating the exception guarantees
[Stroustrup 1993]). If not, we can write straight into a target. The performance difference can
be very significant and the simplest exception specification throw(), throw nothing, was helpful
here. So, as exception specifications were pushed into disuse and eventually removed from the
standard, we introduce the notion of noexcept based on proposals from David Abrahams and
Doug Gregor[Abrahams et al. 2010; Gregor 2010; Gregor and Abrahams 2009].
A noexcept function is still dynamically checked. For example:
void do_something ( int n) noexcept
{
vector < int > v(n );
// ...
}
If do_something() throws, the program is terminated. Doing so happens to have very close to
zero overhead because it simply short-circuits the usual exception propagation mechanism. See
also (ğ7.3).
There is also a conditional version of noexcept so that people can write templates where the
implementation depends on whether a parameter may throw. That’s the original use case that
motivated noexcept. For example, here is an implementation of a move constructor for pair that
is defined if and only if both elements of the pair have move constructors:
template < typename First , typename Second >
class pair {
// ...
template < typename First2 , typename Second2 >
pair ( pair < First2 , Second2 >&& rhs )
noexcept ( is_nothrow_constructible < First , First2 && >:: value
&& is_nothrow_constructible < Second , Second2 && >:: value )
: first ( move ( rhs . first )) ,
second ( move ( rhs . second ))
{}
// ...
};
The is_nothrow_constructible<> is one of the C++11 standard-library type traits (ğ4.5.1).

Writing optimal code at this relatively low and very general level is non-trivial. At the founda-
tional level, knowing whether to bitwise copy, move, or memberwise copy can make large factors
of difference.
4.6 C++11: Standard-Library Components

C++ has always had a tiny standard library compared to other modern languages. Furthermore,
most standard-library components are foundational rather than addressing application-level tasks.
However, C++11 added a few key library components supporting specific tasks:
• threads ś thread-and-lock-based concurrency

• regex ś regular expressions
• chrono ś time
• random ś random number generators and distributions
Compared to the massive corporate support libraries this is obviously pitiful, but the components
are of high quality and massive compared to what had previously been offered as standard for C++.
These components all provide significant help to programmers for the tasks they were designed
for. Unfortunately, the backgrounds of these library components show up as differences in interface
styles; there was no coherent overall design philosophy beyond aiming for flexibility and high
performance. There were no articulated criteria for incorporating a component into the C++11
standard library (C++98 had some [Stroustrup 1994]). Rather, components were adopted from what
was available and had proven successful in the community. Many came from Boost (ğ2.3).
If you need to use regular expressions, the addition of regex to the standard library was a massive
improvement. Similar, adding unordered containers (hash tables), such as unordered_map, saved
many programmers a lot of tedious work and yielded better programs. However, these library
components didn’t seriously affect the way people organized their code, so I will not discuss such
library components in detail.
The regex library was primarily the work of John Maddock [Maddock 2002].
Hash tables just barely missed the cutoff for C++98 and were one of the very first proposals for
C++0x [Austern 2001]. They are called unordered (e.g., unordered_map) to distinguish them from
the older, ordered, standard containers (e.g., map) and because the obvious names (e.g., hash_map)
had been used extensively in other libraries before C++11. Also, unordered_map is arguably a
better name because it refers to what the type offers, rather than to how it is implemented.
The random library provides distributions and random number generators of a sophistication
that has earned in the description łwhat every random-number library wants to be when it grows
up,ž but it is not easy to use for beginners or casual users (who often need random numbers). It was
first proposed by Jens Maurer in 2002 [Maurer 2002] and finally accepted in 2006 after a revision
by a group of people from the Fermi National Lab [Brown et al. 2006].
In contrast, Howard Hinnant’s chrono library [Hinnant et al. 2008] for dealing with time points
and durations is sophisticated, but also very easy to use. For example:
using namespace std :: chrono ; // in sub - namespace std :: chrono

auto t0 = system_clock :: now ();
do_work ();
auto t1 = system_clock :: now ();
cout << duration_cast < milliseconds >( t1 - t0 ). count () << " msec \n ";
The duration_cast converts the clock-dependent łticksž into the time unit of the programmer’s
choice.
Using such simple code, you can get even first-year students to appreciate the different costs of
different algorithms and data structures. chrono provides the notion of time in the thread library
(ğ4.1.2).
For C++20, chrono was enhanced with facilities for dealing with dates and time zones (ğ9.3.6).
C++20 also allows us to simplify that example:
cout << t1 - t0 << '\n ';
This will output the time elapsed between t0 and t1 with an appropriate unit.
5 C++14: COMPLETING C++11

According to the plan of alternating major and minor releases, C++14 [du Toit 2014] was aimed at
łcompleting C++11ž (ğ3.2); that is, to include ideas accepted as good after the 2009 feature freeze
and to remedy problems discovered during initial large-scale use of the C++11 standard. In this
limited aim, it succeeded.
Importantly, it demonstrated that WG21 could deliver a standard on time. This, in turn, allowed
implementers to deliver on time. Before the end of 2014, three major C++ implementers (Clang, GCC,
and Microsoft) delivered what most people could consider feature complete. The conformance
wasn’t perfect, but people could experiment with essentially all features and combinations of
features. The ability to compile libraries using łall advanced featuresž lagged a bit (until 2018
for Microsoft), but for most users the conformance was good enough for real use. The standards
effort and the implementation efforts had become closely tied. That made a big difference to the
community.
The C++14 feature set can be summarized as:
• Binary literals ś e.g., 0b1001000011110011
• ğ5.1: Digit separators ś for readability, e.g., 0b1001’0000’1111’0011
• ğ5.2: Variable templates ś parameterized constants and variables
• ğ5.3: Function return type deduction
• ğ5.4: Generic lambdas
• ğ5.5: Local variables in constexpr functions
• Move capture ś e.g., [p = move(ptr)] { /* ... */ }; move a value into a lambda
• Accessing a tuple by type, e.g., x = get<int>(t);
• User-defined literals in the standard library ś e.g., 10i, "Hello!"s, 10s, 3ms, 55us, 17ns
Most of these features were met with a combination of łGood, what took you so long?ž and łHuh?
who needs that?ž My impression is that every addition was well motivated by some significant
need ś even if that need was not universal ś and that the addition of local variables in constexpr
functions and generic lambdas caused significant improvements to many people’s code.
It was important that upgrading from C++11 to C++14 was relatively painless, with no ABI
breakage. People who had done the major ś and often difficult ś upgrade from C++98 to C++11
were in for a pleasant surprise: they could upgrade sooner than expected and with less effort.
5.1 Digit Separators

Curiously enough, the digit separators led to the most heated debates. Lawrence Crowl repeatedly
presented analyses of the alternatives [Crowl 2013]. Many, including me, argued for using the
underscore as a separator (as in several other languages). For example:
auto a = 1 _234_567 ; // 1234567
Unfortunately, people were using the underscore as part of user-defined literal suffixes:
auto a = 1 _234_567_s ; // 1234567 seconds
This could cause ambiguities. For example, is that last underscore a redundant separator or the
start of a suffix? To my surprise this potential ambiguity made the underscore unacceptable to
many. One reason was that to protect programmers from surprises, the library group had reserved
suffixes not starting with an underscore for the standard library. After many long discussions,
including debates in full committee (about 100 people), we agreed to use the single quote:
auto a = 1 '234 '567; // 1234567 ( an int )
auto b = 1 '234 '567 s; // 1234567 seconds
Despite dire warnings against this use of single quotes would break uncountable numbers of
tools, this seems to work well. The single quote was suggested by David Vandevoorde [Crowl et al.
2013]. He pointed out that a single quote used as a separator in some countries, notably in Swiss
financial notation.
My other suggestion, whitespace, never gained traction:
int a = 1 234 567; // 1234567

int b = 1 234 567 s; // 1234567 seconds
Many people thought it was a joke relating to an old April fool’s article [Stroustrup 1998]. In
fact, it mirrors the old rule that adjacent string literals are concatenated, so that "abc" "def" means
"abcdef".
5.2 Variable Templates

In 2012, Gabriel Dos Reis proposed to extend the template mechanism to offer template variables
in addition to template classes, functions, and aliases [Dos Reis 2012]. For example:

constexpr T pi = T (3.1415926535897932385);

T circular_area (T r)
{
return pi <T > * r * r;
}
At first, this struck me as an obvious language-technical generalization of no particular impor-

tance. However, the history of problems with workarounds for specifying constants of various
precisions is long and littered with uncomfortable workarounds and compromises. After this sim-
ple language generalization, significant amounts of code were simplified. In particular, variable
templates emerged as the main way of defining concepts (ğ6.3.6). For example:
// expression :
concept SignedIntegral = Signed <T > && Integral <T >;
The C++20 standard-library offers a set of mathematical constants defined as variable templates
with the most common cases defined as a plain constexpr [Minkovsky and McFarlane 2019]. For
example:
template < typename T > constexpr T pi_v = unspecified ;

constexpr double pi = pi_v < double >;
5.3 Function Return Type Deduction

C++11 had introduced the ability to deduce the return type of a lambda from its return statement.
C++14 extended that to functions

auto size ( const T& a) { return a. size (); }
This notational convenience can be significant for small functions in generic code. Users have to
be careful, though, because such a function does not provide a stable interface because its type
now depends on its implementation and that implementation must now be visible when a use of
the function is compiled.
5.4 Generic Lambdas

Lambdas are function objects (ğ4.3.1) and as such they could obviously be templates. The problems
related to generic (polymorphic) lambdas had been extensively discussed for C++11, but were
deemed to be not quite ready then (ğ4.3.1).
In 2012, Faisal Vali, Herb Sutter, and Dave Abrahams proposed generic lambdas [Vali et al. 2012].
The proposed notation simply omitted the type from the syntax:
auto get_size = [](& m ){ return m. size (); };
This was strongly opposed by many in the committee (including me) who pointed out that
this syntax was unique and didn’t generalize to constrained generic lambdas, so the notation was
changed to use auto as a token indicating that a type was to be deduced:
auto get_size = []( auto & m ){ return m. size (); };
This brought generic lambdas into line with concepts proposals and suggestions for generic
functions stretching back as far as 2002 [Stroustrup 2003; Stroustrup and Dos Reis 2003a,b].
This direction of integrating the lambda syntax with the syntax used for the rest of the language
was counter to the efforts of some who wanted a unique (ultra-terse) syntax for generic lambdas
similar to what was found in other languages [Vali et al. 2012]:
C# 3.0 (2007): x => x * x;
Java 1.8 (~2013): x -> x * x;
D 2.0 (~2009): (x) { return x * x; };
I think that the decision to use auto and in general not to introduce special notation for lambdas
that is not shared with functions was correct. Further, I think that it was a mistake to introduce
generic lambdas in C++14 without also introducing concepts so that the rules and notations for
constrained and unconstrained lambda arguments and function arguments were not considered
together. The language-technical irregularities stemming from this are (finally) remedied in C++20
(ğ6.4). However, we now have a generation of programmers accustomed to use unconstrained
generic lambdas and proud of that. It will take significant time to overcome that.
From the brief discussion here, it might look as if notation/syntax are given an outsized impor-
tance in the committee process. That may be so, but syntax is not unimportant. The syntax is the
programmer’s user interface and debates about syntax often reflect differences over semantics or
over the desired use of a feature. A notation should reflect the underlying semantics and a syntax
typically favors one kind of use over another. For example, a fully general and verbose notation
favors experts wanting to express subtle distinctions whereas a notation optimized for expressing
simple cases favors novices and casual users. I am typically on the side of the latter and often in
favor of supplying both (ğ4.2).
5.5 Local Variables in constexpr Functions

By 2012, people had stopped being scared of constexpr functions and started to demand relaxations
of the restrictions on their implementations. There were people who essentially wanted to be able
to do anything in constexpr functions. However, neither the users nor the compiler implementers
were ready for that.
After some discussions, Richard Smith (Google) proposed a relatively modest set of relaxations
[Smith 2013]. In particular, local variables and for-loops were allowed. For example:
constexpr int min ( std :: initializer_list < int > xs )

{
int low = std :: numeric_limits < int >:: max ();
for ( int x : xs )
if (x < low )
low = x;
return low ;
}
constexpr int m = min ({1 ,3 ,2 ,4});
Given a constant expression as argument, this min() can be evaluated at compile time. The local
variables (here, low and x) simply łlivež in the compiler. The evaluation cannot have side effects
on the environment of a caller. The original (academic) constexpr paper by Gabriel Dos Reis and
Bjarne Stroustrup pointed to this possibility [Dos Reis and Stroustrup 2010].
This relaxation simplified many constexpr functions and pleased the many C++ programmers
who had not been happy to find that only purely functional expressions of algorithms could
be evaluated at compile time. In particular, they wanted loops to avoid recursion. In the longer
term, this unleashed demands for further relaxations in C++17 and C++20 (ğ9.3.3). To illustrate
the potential power of compile-time evaluation, I have pointed out that constexpr threads are
possible, though I am not in a hurry to propose that.
6 CONCEPTS
Generic programming and metaprogramming with templates have been runaway successes for
C++. However, the proper specification of interfaces to generic components was slow reaching a
satisfactory state. For example, in C++98, the standard-library algorithm was specified roughly like
this:
template < typename Forward_iterator , typename Value >

ForwardIterator find ( Forward_iterator first , Forward_iterator last ,
const Value & val )
{
while ( first != last && * first == val )
++ first ;
return first
}
The standard says that

• the first template argument must be a forward iterator.
• the second template argument type must be comparable to the value type of that iterator
using ==.
• the first two function arguments must denote a sequence.
These requirements are implicit in the code: all the compiler has to go by is the use of the
template parameters in the function body. The result is great flexibility, splendid generated code
for correct calls, and spectacularly bad error messages for incorrect calls. The obvious solution is to
specify the first two requirements as part of the template’s interface:
template < forward_iterator Iter , typename Value >

requires equality_comparable < Value , Iter :: value_type >
forward_iterator find ( Iter first , Iter last , const Value & val );
This is roughly what C++20 offers. Note the equality_comparable concept. It captures the
required relationship between the two template arguments. Such multi-argument concepts are
very common.
To express the third requirement (that [first:last) is a sequence) requires a library extension.
C++20 offers that in the Ranges standard-library component (ğ9.3.5):
template < range R , typename Value >
requires equality_comparable < Value , Range :: value_type >
forward_iterator find (R r , const Value & val )
{
auto first = begin (r );
auto last = end (r );
while ( first != last && * first == val )
++ first ;
return first
}
This section describes the attempts to provide good support for the specification of a template’s
requirements on its template arguments:
• ğ6.1: The prehistory of concepts
• ğ6.2: C++0x concepts
• ğ6.3: The concepts TS
• ğ6.4: C++20 concepts
6.1 The Prehistory of Concepts

In 1980, I conjectured that generic programming could be effectively supported through C-style
macros [Stroustrup 1982]. I was flat wrong; a few useful simple generic abstractions could be
expressed that way and the 1980s pre-standard C++ supported generic programming with a set
of macros in <generic.h>, but macros weren’t manageable in larger projects or in wide use. I did
identify a problem that needed solving to meet my aims for łC with Classes,ž though, even though
generic programming didn’t have a place in łobject-oriented thinkingž as was then becoming
fashionable.
In about 1987, I tried to design templates with proper interfaces [Stroustrup 1994]. I failed. I
wanted three fundamental properties to support generic programming:
• Full generality/expressiveness ś I explicitly didn’t want facilities that could express only what
I could imagine.
• Zero overhead compared to hand coding ś for example, I wanted to build a vector that could
compete with C arrays for time and space performance.
• Well-specified interfaces ś I wanted facilities for type checking and overloading comparable
to what we had for non-generic code.
Then, nobody could figure out how to get all three, so C++ got:
• Turing completeness [Veldhuizen 2003]
• Better than hand-coding performance
• Lousy interfaces (basically compile-time duck typing), but still statically type safe
The first two properties made templates a run-away success.
The lack of well-specified interfaces led to the spectacularly bad error messages we saw over the
years and still in C++17. The lack of well-specified interfaces bothered me and many others over
the years. It bothered me a lot because templates failed to meet the fundamental design criteria of
C++ [Stroustrup 1994]. We (obviously) needed a simple way of specifying the requirements of a
template on its template arguments that didn’t imply run-time overheads.
For years, several people (including me) believed that requirements for template arguments
could adequately be specified in the C++ itself. In 1994, I documented the basic idea in [Stroustrup
1994] and published examples on my website [Stroustrup 2004ś2020]. Since 2006, Boost provided a
variant of that idea, the Boost concept check library [Siek and Lumsdaine 2000ś2007], based on
the work of Jeremy Siek. Somehow, this never caught on as widely as I had hoped. I suspect the
reason is that it wasn’t sufficiently general, sufficiently elegant (Boost felt obliged to hide details in
macros), and not supported in the standard. Many saw it as a hack.
Concepts, as defined for C++, go back to Alex Stepanov’s work on generic programming starting
in the late 1970s and first documented under the name łAlgebraic structuresž [Kapur et al. 1981].
Note that’s almost a decade before the design of Haskell’s type classes [Wadler and Blott 1989] and
about 5 years before I tried to address the problem for C++. Alex Stepanov used the name łconceptž
for such requirements in lectures in the late 1990s and documented that in [Dehnert and Stepanov
2000]. I mention this because many have conjectured that concepts were derived from Haskell type
classes and misnamed. Alex used the name łconceptž because concepts were meant to represent
fundamental concepts from an application domain, such as algebra.
The current use of concepts as type predicates relying of use patterns to describe operations
has its origins in the work of Bjarne Stroustrup and Gabriel Dos Reis in the early 2000s and
documented in [Dos Reis and Stroustrup 2005b, 2006; Stroustrup and Dos Reis 2003b, 2005a]. The
approach is even mentioned in D&E from 1994 [Stroustrup 1994], but I don’t remember when I
first experimented with it. The reason for basing concepts on use patterns was primarily to handle
implicit conversions and overloading in a simple and general manner. We knew of Haskell type
classes, but they were not significant influences on the current C++ design because we considered
them too inflexible.
Precise specification of a template’s requirements on its arguments and the checking of those
were to have been the centerpiece of C++0x and the crucial support for generic programming. It
didn’t even make C++17.
The 2003 papers by Bjarne Stroustrup and Gabriel Dos Reis [Stroustrup 2003; Stroustrup and Dos
Reis 2003a,b] make it clear that concepts were part of an ambitious program to simplify generic
programming. For example, a concept can be defined as a set of constraints specified as use patterns;
that is, as language constructs that should be valid for a type [Stroustrup and Dos Reis 2003b]:
concept Value_type {
constraints ( Value_type a)
{
Value_type b = a; // copy initialization
a = b; // copy assignment
Value_type v [] = {a }; // not a reference
}
};
template < Value_type V >

void swap (V& a , V& b ); // the arguments to swap () must be value types
The syntax and semantics were still very immature, though. We were primarily trying to establish
design criteria [Stroustrup and Dos Reis 2003a]. From a modern (2018) perspective, [Stroustrup
2003; Stroustrup and Dos Reis 2003a,b] have many flaws. However, they offered design constraints
for concepts together with suggestions for
• Concepts ś compile-time predicates for specifying requirements on template arguments.
• Use patterns for specifying primitive constraints ś to handle overloading and implicit type
conversions.
• Multi-argument concepts ś e.g., Mergeable<In1,In2,Out>.
• Both type and value concepts ś that is, concepts can take values as well as types as arguments,
e.g., Buffer<unsigned char,128>.
• A shorthand notation for łtype of typež uses of templates ś e.g., template<Iterator Iter> ....
• A łsimplified notation for template definitionsž ś e.g., void f(Comparable&); to bring
generic programming closer to łordinary programming.ž
• auto as the least constrained type in function arguments and return values.
• Uniform function call (ğ8.8.3) ś to alleviate problems with style differences between generic
programming and object-oriented programming (e.g., x.f(y), f(x,y), and x+y).
Curiously enough, we didn’t suggest general requires-clauses (ğ6.2.2). Those have been part of
all later variations of concepts.
6.2 C++0x Concepts

In 2006, essentially everyone expected that the version of concepts described in [Gregor et al.
2006; Stroustrup 2007] and voted into the draft standard (working paper) would be part of C++09.
However, C++0x became C++11 and in 2009 the committee agreed by a massive voting majority
to abandon the concept design [Becker 2009] as mired in complexity and usability problems
[Stroustrup 2009a,b]. The reasons for this failure are varied and may have lessons beyond the C++
standards effort.
In 2004 there were two independent efforts to add concepts to C++. There were often referred to
as łIndianaž and łTexas,ž respectively because the major proponents came from Indiana University
and Texas A&M University:
• Indiana: An approach related to Haskell type classes and primarily relying on tables of
operations to define concepts. It was deemed important for a programmer to explicitly state
that a type łmodeledž a concept; that is, that the type offered a set of operations specified
by the concept [Gregor et al. 2006]. Key people were Andrew Lumsdaine (professor) and
Douglas Gregor (postdoc and compiler writer).
• Texas: An approach based on compile-time type predicates and predicate logic. It was deemed
important for usability that a programmer should not have to explicitly specify which types
matched which concepts (those matches could be computed by the compiler). It was consid-
ered essential for C++ to elegantly and efficiently handle implicit conversions, overloading,
and mixed-type expressions [Dos Reis and Stroustrup 2006; Stroustrup and Dos Reis 2003b].
Key people were Bjarne Stroustrup (professor) and Gabriel Dos Reis (postdoc, later professor).
Given these descriptions, it may seem obvious that the approaches were irreconcilable, but
that was not obvious to the people involved at the time. In fact, I argued that the approaches
were theoretically equivalent [Stroustrup and Dos Reis 2003b]. That argument may indeed be
true, but the practical implications for detailed language design and use in the context of C++
were not equivalent. Separately, the WG21 consensus process as interpreted by the committee
members strongly encourages collaboration and joint proposals as opposed to working for years
on competing proposals and ending up with a grand shoot-out between them (ğ3.2). I consider
the latter approach a recipe for dialect creation because a losing side is unlikely to just go away
abandoning their implementation and users. Note that all the people mentioned above together
with Jeremy Siek (grad student in Indiana and my summer intern in AT&T Labs) and Jaakko Järvi
(postdoc in Indiana and later professor in Texas A&M) were co-authors of the OOPSLA paper
presenting the first version of the compromise design. The Indiana and Texas groups were never
completely disjoint and we tried hard for genuine consensus. In addition, I had known Andrew
Lumsdaine for years before this work. We really wanted the compromise design to work.
The Indiana design was far ahead of the Texas design in terms of implementation and had more
people involved, so we proceeded primarily based on that. The Indiana design was also more
conventional, based on function signatures and had obvious similarities to Haskell type classes.
Given the number of academics involved, it was important that the Indiana design was seen as
more conventional and academically respectable. It seemed that we łjustž had to
• make the compiler acceptably fast
• generate efficient code
• handle overloading and implicit conversions.
That decision cost us three years of hard work and much controversy.
The C++0x concept design is described in [Gregor et al. 2006; Stroustrup 2007]. The former paper
contains a standard academic łrelated workž section comparing the design to facilities offered by
Java, C#, Scala, Cecil, ML, Haskell, and G. Here I summarize, using examples from [Gregor et al.
2006].
6.2.1 Concept Definitions. A concept was defined as a set of operations and associated types:
concept EqualityComparable < typename T > {
bool operator ==( const T& x , const T& y );
bool operator !=( const T& x , const T& y) { return !( x == y ); }
}
concept InputIterator < typename Iter > {

// Iter must have a member value_type :
typename value_type = Iter :: value_type ;
// ...
}
The similarity between a concept and a class was by some (Indiana) considered an advantage.
Functions specified in concepts were not exactly like functions defined in classes, though. For
example, an operator declared within a concept doesn’t have the implicit argument (łthisž) that
an operator defined within a class has.
There was a serious problem lurking in the approach of defining a concept as a set of operations.
Consider the ways an argument can be passed in C++:
void f(X );
void f(X &);
void f( const X &);
void f(X &&);
Ignore for a moment volatile because that’s rarely seen for arguments in generic code, but we
still have four alternatives. In a concept, do we
• represent f as one function and have the users pick the right kind of argument for their calls?
• overload f with all alternatives?
• represent f as one function and let the users define a concept_map (ğ6.2.3) to map to f’s
desired argument type?
• have the language implicitly map user’s argument types to the template’s parameter type?
For two arguments, we would have 16 alternatives. Three argument generic functions are rare,
but we would have 4*4*4 alternatives. What about variadic templates? (ğ4.3.2); for those, we could
have 4N alternatives.
The semantics of the different ways of passing an argument are not equivalent, so we naturally
drifted towards accepting the argument type as specified, pushing the burden of matching onto
designers of types and writers of concept_maps (ğ6.2.3).
Similarly, we have the problem whether to specify a concept for x.f(y) (object-oriented style) or
f(x,y) (functional style), or both. This problem occurs immediately when we try to specify a binary
operator, such as +.
In retrospect, we were far too optimistic about solving these problems within the framework of
concepts defined in terms of operations with specific types or specific łpseudo signatures,ž where a
łpseudo signaturež somehow represented a solution to the problems outlined here.
Relationships among concepts were defined by explicit refinement:
concept BidirectionalIterator < typename Iter > // A BidirectionalIterator is
: ForwardIterator < Iter > { // a kind of ForwardIterator
// ...
}
Refinement was almost, but not quite, like class derivation. The idea was for programmers to
explicitly build up hierarchies of concepts. Unfortunately, that introduces a serious inflexibility
into the system. Concepts (in the conventional English meaning) are often not strictly hierarchical.
6.2.2 Concept Use. A concept could be used either as a predicate in a where-clause or in a
shorthand notation:
where LessThanComparable <T > // explicit predicate
const T& min ( const T& x , const T& y)
{
return x <y ? x : y;
}
template < GreaterThanComparable T > // shorthand notation

const T& max ( const T& x , const T& y)
{
return x >y ? x : y;
}
For simple łtype of typež concepts, the shorthand notation (first suggested in [Stroustrup 2003])
quickly became very popular. However, we soon found that where was far too popular as an
identifier in existing code and renamed it to requires.
6.2.3 Concept Maps. Relationships between concepts and types were defined by specializations of
concept_maps:
concept_map EqualityComparable < int > {}; // int is EqualityComparable
// student_record is EqualityComparable :
concept_map EqualityComparable < student_record > {

bool operator ==( const student_record & a , const student_record & b)
{
return a. id_equal (b );
}
};
For int, we can simply say that the type int has the properties required by EqualityComparable
(that is, it has == and !=). However, student_record doesn’t have a ==, but we can add one in the
concept_map. Thus, a concept_map is a very powerful mechanism for non-intrusively adding
properties to a type in specific circumstances.
Why do we have to tell the compiler that ints can be compared? The compiler already knows.
This was a constant point of contention. The łIndiana groupž generally felt that being explicit was
good (always) and the łTexas groupž tended to consider a concept map worse than useless unless it
added functionality. Would explicit statements save use from significant errors from łaccidentalž
syntactic matches that were semantically meaningless? Alternatively, would such errors be rare and
the explicit modeling statements be mostly an annoyance to write and opportunities for making
mistakes? The compromise solution was to allow the definer of a concept to declare the use of a
concept_map optional by adding auto:
auto concept EqualityComparable < typename T > {
bool operator ==( const T& x , const T& y );
bool operator !=( const T& x , const T& y) { return !( x == y ); }
}
Now, the concept_map for EqualityComparable is automatically used when EqualityCompa-

rable required for a type even when a user didn’t supply a specialization for that type.
6.2.4 Definition Checking. The code in a template definition was checked against the concepts of
its template parameters:
template < InputIterator Iter , typename Val >
requires EqualityComparable < Iter :: value_type , Val >
Iter find ( Iter first , Iter last , Val v)
{
while ( first < last && !(* first == v )) // error : no < in EqualityComparable
++ first ;
return first ;
}
Here, we use < to compare iterators, but EqualityComparable guarantees only == so this defi-
nition won’t compile. Catching such used of non-guaranteed operations was seen as an important
benefit, but is turned out to have serious negative implications: (ğ6.2.5) and (ğ6.3.1).
6.2.5 Lessons Learned. The years after the initial proposal and relatively fast approval was filled
with plugging holes in this initial design and addressing comments about generality, implementabil-
ity, quality of specification, and usability.
Doug Gregor, as the primary implementer, performed heroics to generate quality code, but at
the end, the concepts compiler was still more than 10 times slower than a compiler implementing
just unconstrained templates. I suspect that the implementation problems had their ultimate roots
in the adoption of a class-like structure for representing concepts in the compiler. This enabled
quick early results, but left the concepts represented in a way carefully crafted for classes, and
concepts are not classes. Having a concept represented as a set of functions (like virtual member
functions), led to problems handling implicit conversions and mixed-type operations. The very
flexible combinations of code from different contexts that is the łsecretž of some of the powerful
generic programming and metaprogramming code-generation techniques became impracticable to
specify with C++0x concepts. It is essential for matching (unconstrained) template performance
that the functions used to specify a concept do not appear in the generated code as function calls
(or worse still, as indirect function calls).
I was unpleasantly reminded of the problems that many early C++ compiler writers had had from
adopting the structure and code base of a C compiler. The handling of C++ scopes and overloading
didn’t fit well into a C compiler framework. Based on the idea that design concepts should be
represented directly in code, Cfront (ğ2.1) used specific scope classes to avoid such problems.
However, most compiler writers with a C background thought they could take a shortcut using
familiar C techniques, but ended up having to rewrite their C++ front-end from scratch anyway.
Language design and implementation techniques can strongly influence each other.
It soon became obvious that we needed language support for the transition from unconstrained
templates to templates using concepts. In the C++0x design, they were very different:
• A constrained template cannot call an unconstrained one because it is not known what
operations an unconstrained template uses, so the definition check of the constrained template
cannot be done.
• An unconstrained template can call a constrained one, but checking must be postponed until
instantiation time because not until then do we know what types the unconstrained template
uses in its calls.
The solution to the first problem was to allow the programmer to say łdon’t check these calls
from a constrained templatež using a late_check block [Gregor et al. 2008]:
template < Semigroup T >
T add (T x , T y) {
T r = x + y; // uses Semigroup <T >:: operator +
late_check {
r = x + y; // uses operator + found at instantiation time
// ( not considering Semigroup <T >:: operator +)
}
return r;
}
This łsolutionž was at best a patch and had the extraordinary problem that a concept_map for
Semigroup wouldn’t be known in the unconstrained called template. This had the łinteresting
effectž that an object could be used in exactly the same way in two places in a program with
different semantics. Thus, the type system had been violated in a really hard-to-trace way.
As we used concepts more, the role of semantics in the design of concepts (and indeed of types and
libraries) became increasingly clear and many in the committee started pushing for a mechanism
for expressing semantics rules. This was not a surprise. Alex Stepanov was fond of saying łconcepts
are all about semantics.ž However, most people were approaching concepts like any other language
facility and were more concerned with syntax and name lookup rules.
In 2009, a notation, called axioms was proposed by Gabriel Dos Reis (strongly supported by me)
and approved [Dos Reis et al. 2009]:
concept TotalOrdering < typename Op , typename T > {
bool operator ()( Op , T , T );
axiom Antisymmetry ( Op op , T x , T y) {
if ( op (x , y) && op (y , x ))
x <=> y;
}
axiom Transitivity ( Op op , T x , T y , T z) {
if ( op (x , y) && op (y , z ))
op (x , z );
}
axiom Totality ( Op op , T x , T y) {
op (x , y) || op (y , x );
}
}
Curiously enough, it was hard to get the notion of an axiom accepted. The main objection
seemed to be that the proposers explicitly rejected the idea of compilers testing axioms against
the types for which they were used łto catch errors.ž Apparently, the notion that axioms were
axioms in the mathematical sense (that is, assumptions that you are allowed to make because you
in general can’t check them) was alien to some committee members. Others weren’t convinced
that specifying axioms could help tools other than the compilers. However, axioms were adopted
as part of concept specifications.
There were obvious problems with our definition and implementation of concepts, but we had a
pretty complete facility and plowed along trying to solve the problems and to gain experience by
using concepts in the definition of the standard library [Gregor and Lumsdaine 2008] and other
libraries.
6.2.6 What Went Wrong? In 2009, I had reluctantly come to the conclusion that the concepts effort
was in deep trouble. The problems that I had expected us to solve were still festering and new ones
had been discovered:
• We still didn’t have an agreement whether implicit or explicit modeling (implicit or explicit
use of concept_maps) was the right approach in most cases.
• We still didn’t have an agreement whether to rely on implicit or explicit statement of relations
among concepts (should we explicitly build hierarchies of łrefinementž relations in a way
very similar to object-oriented inheritance?).
• We were still seeing examples where code generated from concept-constrained code was
inferior to code generated from unconstrained templates. The late composition opportunities
from templates kept showing surprising strengths.
• It was still difficult to write concepts to capture all the conversions and overload cases that
we were used to from generic and non-generic C++.
• We were seeing increasing numbers of examples where the combination of non-trivial
concept_maps and late_checks led to inconsistent views of types (aka surprising and
almost invisible violations of the type system).
• The complexity of specification in the draft standard had ballooned beyond all expectation (91
pages, excluding any uses of concepts in the library) and some of us considered it essentially
unreadable.
• The set of concepts used to specify the standard library was getting large (about 125 concepts,
103 for the STL alone).
• The compiler was getting better at code generation (because of heroic efforts from Doug
Gregor) but not faster. Some major compiler vendors were telling me in private and confidence
that if a concept-enabled compiler was more than 20% slower than older compilers they’d
have to oppose concepts however nice they were. At the time, the concept-enabled compiler
was more than 10 times slower.
In the spring of 2009, a wide-ranging discussion occurred on the standard reflectors (archived
email groups). It started when Howard Hinnant asked an eminently practical question about the
use of concepts: A utility that he was designing could be done in two ways: One would require
quite a lot of users ś not necessarily expert users ś to write concept maps. The other ś far less
elegant ś design would avoid concept maps (and concepts) so as not to require users to understand
anything significant about concepts. Should łaverage usersž understand concepts? Just enough to
use them? Enough to define them?
This became known as the łAre concepts required of Joe Coder?ž thread. Who is łJoe Coder?ž
asked Peter Gottschling. Great question, I answered:
łI think most C++ programmers are łJoe Coderž (I again register my opposition to that
term). I’m Joe Coder most of the time and with most libraries. I expect to remain so as
long as I keep learning new techniques and libraries. Yet, I want to use concepts (and,
when I must, concept maps). I want the łdoctrine of usež radically simpler than the subtle
expert-only use of facilities we have now.ž
In other words, should we design concepts as a delicate instrument for fine control by a small
group of language experts or as a robust tool for most programmers? This is a question that comes
up again and again in the design of language features and standard-library components. I heard it
for years about classes; for some, defining a class was obviously something that most programmers
should be discouraged from doing. The łaverage programmerž (sometimes derisively referred to
as łJoe Coderž) was to some obviously not smart enough or educated enough to use sophisticated
features and techniques. I was (and am) strongly of the opinion that most programmers can learn
to use features like classes and concepts well. Once they do, their programming become easier and
their code better. It may take years for the community at large to absorb a lesson, but if that can’t
be done, we ś as language and library designers ś have failed.
In response to that thread and reflecting my growing concern about the direction of the work on
C++0x concepts, I wrote a paper Simplifying the use of concepts [Stroustrup 2009c] outlining what I
saw as the minimum improvements necessary for concepts to be acceptable in C++0x:
• Make concept_maps rare.
• Make all concept_maps implicit/automatic.
• Make a concept that requires begin(x) accept x.begin() and vice versa (uniform function
call; (ğ6.1), (ğ8.8.3)
• Make all standard-library concepts implicit/automatic.
The paper is quite detailed with many examples and suggestions that had emerged over the
years.
One reason that I insisted on making all concepts implicit/automatic was the observation that
given a choice, the least flexible and least trusting programmers could force everyone to live with
their choice of explicit concepts. Library writers were showing a strong tendency to push resolution
of even the most obvious choices onto their users by using explicit (non-automatic) concepts.
I observed that the then recent Elements of Programming Style [Stepanov and McJones 2009] by
Alex Stepanov, the father of C++ generic programming, didn’t use a single concept map to describe
a superset of the facilities from the STL and a superset of the then common generic programming
techniques.
The committee responded with a discussion almost exclusively focused on whether a consensus
was likely for a timely standard and came to the obvious conclusions that it was not likely. We
could not agree to łfixž concepts to make them usable by most programmers and also ship the
standard (more or less) on time. Thus, łconceptsž ś the result of years of work by many competent
people ś was removed from the draft standard. My summary of the łremove conceptsž decision
[Stroustrup 2009a,b] is more readable than the technical papers and discussions.
As the committee voted to remove concepts by a large majority (I too voted for removal), everyone
who spoke up reconfirmed that they wanted concepts. The vote just reflected that the concept
design wasn’t ready for standardization. I think that the problem was far worse: The committee
wanted concepts, but the members didn’t agree about what kind of concepts they wanted. The
committee did not have a shared set of design aims. This is still a problem, and not just for concepts.
There are deep łphilosophicalž differences among members. In particular:
• Explicit vs. implicit: Should programmers ś in the interest of safety and the avoidance of
surprises ś be explicit about how choices among potential alternatives are resolved? This
discussion turns up in discussion about overload resolution, scope resolution, matching of
types to concepts, relationships among concepts, and more.
• Experts vs. average: Should key language and standard-library facilities be designed for the use
of experts? If so, should łaverage programmersž be encouraged to exclusively use a limited
subset of the language and should separate libraries be designed for łaverage programmersž?
This discussion turns up in the contexts of the design and use of classes, class hierarchies,
exceptions, templates, and more.
In both cases, a łyesž will bias the design of facilities towards complicated features that require
great expertise and frequent use of notation to get right. I tend systematically to be on the other
side of such arguments, trusting łaverage programmersž more and relying on regular language
rules and checking by compilers and other tools to avoid nasty surprises. Programmers are at least
as likely to get explicit resolution of tricky issues wrong as are (implicit) language rules.
Different people have drawn different conclusions from the C++0x concepts failure. I drew three
main ones:
• We put too much weight on early implementation. We should have spent more effort on
working out the requirements, the constraints, the desired use patterns, and a relatively
simple implementation model. After that, we could have grown an implementation relying
on feedback from use.
• Some disagreements are fundamental (philosophical) and cannot be resolved through com-
promises. We must identify and articulate such issues early on.
• No set of facilities can serve all the diverse wishes from a large committee of experts without
becoming so large that the bloat becomes a problem for implementers and an impediment
to users. We must identify core needs and serve them well with a simple notation; more
complex uses and rarer use cases can be served with facilities and notation that puts greater
requirements on the expertise of their users (ğ4.2).
These conclusions do not have anything in particular to do with concepts. They are general
observations about design aims and decision making within a large group.
6.3 The Concepts TS

In 2009, almost immediately after the removal of concepts from C++0x, Gabriel Dos Reis, Andrew
Sutton, and I started to re-design concepts based on our original ideas, the experience gained from
the C++0x language design, the experience from use of C++0x concepts, and the feedback from the
standards committee. We concluded
• concepts must have semantic meaning
• there should be relatively few concepts
• concepts should be fundamental, rather than minimal
We deemed most of the concepts included in the C++ standard library meaningless in isolation
[Sutton and Stroustrup 2011]. łThe STL does not have 103 concepts for any reasonable definition
of ‘concepts’!ž I cried during a discussion with Andrew Sutton, łeven basic algebra doesn’t have
more than about a dozen!ž Language design discussions can become quite animated.
In 2011, at the urging of Andrew Lumsdaine, Alex Stepanov called a week-long meeting in Palo
Alto. A largish group, including most of the people closely associated with the C++0x concepts
effort, Sean Parent, and Alex Stepanov, met to attack the problem from the user’s perspective: What
would a properly constrained set of STL algorithms ideally look like? Then, we went home to
document our user-oriented design and invent language mechanisms to approximate that ideal
design [Stroustrup and Sutton 2012]. This effort re-booted the standards effort based on a new,
fundamentally different and better, approach than the C++0x effort. The 2016 ISO TS (Technical
Specification) for concepts [Sutton 2017] and the C++20 concepts (ğ6.4) are direct results of that
meeting. Andrew Sutton’s implementation has been used for experimentation since 2012 and
shipping as part of GCC 6.0 and higher.
In the concepts TS [Sutton 2017]
• Concepts are based on compile-time predicates (including multi-argument predicates and
value arguments).
• Primitive requirements specified in terms of use patterns [Dos Reis and Stroustrup 2006]
(requires-expressions).
• A concept can be used in general requires-clauses, as an alternative to a typename in
template parameter definitions, and as an alternative to a type name in function parameter
definitions.
• Matching of types to concepts is implicit (no concept_maps).
• Relations among concepts for overloading are implicit (computed; there are no explicit
refinement of concepts).
• There is no definition checking (for now at least, so no late_check).
• There are no axioms, but that is just because we didn’t want to complicate and delay the
design with a potentially controversial feature. The C++0x axioms would make a good start.
Compared to the C++0x concepts there is a strong emphasis on simplifying the use of concepts.
A major part of that has been to eliminate requirements for programmers to be explicit and rely on
the compiler resolving issues based on well-specified and simple algorithms.
People who favor explicit resolution by users tend to deem that an emphasis of semantics over
syntax and warn against łaccidental matchesž and łsurprises.ž The most common example is that
a Forward_iterator differs from an Input_iterator only in its semantics: a Forward_iterator
allows multiple passes over its sequence. Nobody disagrees that such examples exist, but endless
debates have raged (and still rage) over how important such examples are and how to resolve them.
I consider it a bad mistake to let a few rare complicated examples dominate a design.
The Concept TS design is based on the opinion (backed by much experience) that such examples
are very rare (especially in well-designed concepts [Stroustrup 2017a]), are usually well understood
by concept writers, and can often be resolved by adding an operation to the most constrained
concept to reflect the sematic difference. For example, a simple solution to the Forward_iterator/
Input_iterator problem is to require a Forward_iterator to offer a can_multipass() operation.
It doesn’t even have to do anything; it would just be there so that the concept resolution mechanism
can check for its existence. Thus, no new language features need to be added specifically to resolve
possible accidental ambiguities.
Because the point is often missed, I must emphasize that concepts are predicates. They are not
classes or class hierarchies. Fundamentally, we are just asking a type simple questions, such as "are
you an Iterator?", and asking sets of types questions about their interoperation, such as "can you
compare against each other using ==?" (ğ6.3.2). Using concepts, we only ask questions that can be
answered at compile time; no run-time evaluation is involved. Potential ambiguities are detected
by comparing the predicates involved for a type (or set of types); not by having the programmer
write resolution rules (ğ6.3.2).
Sensitive to the problems with C++0x concepts (ğ6.2.6), we were careful to design concepts
so that their use would not imply significant compile-time overheads. Even early versions of
Andrew Sutton’s compiler was able to compile templates using concepts faster than programs
using workarounds (e.g., enable_if (ğ4.5.1)).
6.3.1 Definition Checking. Sometime during the months after that Palo Alto meeting, Andrew
Sutton, Gabriel Dos Reis and I decided to attack the design and implementation of the concept
language features in stages. That way, we could learn from the implementation experience and
gain early feedback from use before łfreezing the designž. In particular, we decided to postpone
the implementation of definition checking (ğ6.2.4); that is, the checking that a template doesn’t
use facilities not specified for its arguments. Consider a simplified version of std::advance() that
moves an iterator n positions forward in a sequence:
template < Forward_iterator Iter >

void advance ( Iter p , int n)
{
p += n; // advance p n positions
}
A Forward_iterator doesn’t offer +=, only ++, so definition checking will catch this as an
error. Without checking the body of advance() in isolation (before use) we will only get poor
instantiation-time error message from the (mis)use of +=. Note that the code generated from
template instantiation is always type checked, so not doing definition checking will not lead to
run-time errors.
We decided that on the order of 90% of the benefits from concepts would accrue from point-
of-use checking and that the relatively expert implementers of constrained templates could do
without definition checking for a while. This 90% was obviously an ad hoc estimate based on limited
information, but with the benefit of a decade’s work with concepts I consider it a good guess. It was
more important to us as designers of language features and libraries to gain experience with use,
starting with the STL algorithm examples from the Palo Alto Technical Memorandum [Stroustrup
and Sutton 2012]. We valued feedback over theoretical completeness. This was radical. Looking
back over the documents about concepts (in C++ and for other languages), definition checking
was always emphasized as a major reason for offering concepts as a language feature [Gregor et al.
2006; Stroustrup and Dos Reis 2003b].
For a while, this new design was referred to as Concepts Lite and many deemed it incomplete,
or even useless. However, we quickly discovered that not doing definition checking offered real
benefits [Sutton and Stroustrup 2011].
• With definition checking we can’t use partial concept checking during development. It is
very common not to know the full set of requirements during the initial construction of
a larger program. Partial checking allows many errors to be caught early and encourages
gradual improvement of a design based on feedback from early use.
• Definition checking makes it hard to have stable interfaces. In particular, it is not possible
to add debug statements, statistics gathering, tracing, or łtelemetryž to a class or a function
without changing its interface to include the facilities for that. Such facilities are rarely
fundamental to a class or a function and tend to change over time.
• When we don’t use definition checking, existing templates can be gradually converted
to use concepts. However, if we have definition checking, a constrained template cannot
use an unconstrained one because we cannot in general know which facilities the uncon-
strained template uses. Also, an unconstrained template using a constrained one implies late
(instantiation-time) checking in either case.
Ville Voutilainen, the chair of the EWG from 2014 onwards, said it stronger:
I cannot support any proposal for concepts that includes definition checking.
We might eventually get a form of definition checking, but only if we can design a mechanism for
escaping it to address transition and data gathering needs. That will have to be carefully thought
out and experimentation will be needed. The C++0x late_check is not sufficient.
The problems with definition checking are problems of use, not implementation. Gabriel Dos
Reis designed and implemented an experimental language, called Liz, testing out the facilities of
the Concepts TS design [Dos Reis 2012], including definition checking. If we find an acceptable
form of definition checking, we can implement it.
6.3.2 Concept Use. Simple examples look much like they did for C++0x and before:
template < Sequence Seq , Number Num >
Num sum ( Seq s , Num v)
{
for ( const auto & x : s)
v += x;
return v;
}
Here Sequence and Number are concepts. Using a concept instead of typename to introduce
the name of a type means that the type used must satisfy the requirements of the concept. Note
that because the concepts TS doesn’t offer definition checking, the use of += will not be checked
by the concepts, but only late, at instantiation time. This is likely during initial development. Later
we would most likely be more explicit:
using Value_type = typename T :: value_type ; // simplified alias
template < Sequence Seq , typename Num >

requires Arithmetic < Value_type < Seq >, Num >
Num sum ( Seq s , Num n)
{
for ( const auto & x : s)
v += x;
return v;
}
That is, we must have arithmetic operations, including +=, for combinations of the Sequence’s
value type and the type we use as the accumulator. We no longer have to specify Num to be
a Number; Arithmetic checks that Num has the needed properties. Here, Arithmetic is used
explicitly as a predicate in a (C++0x-style) requires-clause.
Overloading is handled by picking the function that has the strictest requirements. Consider a
simple version of the classical advance example from the standard library:
template < Forward_iterator Iter >

void advance ( Iter p , int n) // move p n elements forward
{
while (--n)
++ p; // a forward iterator has ++ , but not + or +=
}
template < Random_access_iterator Iter >

void advance ( Iter p , int n) // move p n elements forward
{
p += n; // a random - access iterator has +=
}
That is, we should use the second version for sequences that offer random access and the first
for sequences that offer only forward iteration:
void user ( vector < int >:: iterator vip , list < string >:: iterator lsp )
{
advance ( vip ,10); // use the fast advance ()
advance ( lsp ,10); // use the slow advance ()
}
The compiler breaks the concepts for the two functions into primitive (łatomicž) requirements
and since the requirements for forward iteration are a strict subset of those for random access
iteration, this example can be resolved.
Overlapping requirements that are not strict subsets of each other give ambiguities (compile-time
errors) when an argument type match both. For example:
requires Copyable <T > && Integral <T >
T fct (T x );

requires Copyable <T > && Swappable <T >
T fct (T x );
int x = fct (2); // ambiguous : int is Copyable , Integral , and Swappable

auto y = fct ( complex < double >{1 ,2}); // OK : complex is not integral
The only control mechanism offered to programmers is the ability to add operations when
defining a concept. That seems to be sufficient for real-world examples, though. Of course, you can
define concepts that differ only semantically so that there would be no way of distinguishing them
based on our syntax-only concepts. However, it’s not difficult to avoid doing that.
6.3.3 Concept Definition. Primitive requirements are specified by use patterns in requires-expressions:
template < typename T , typename U =T >
concept Equality_comparable =
requires (T a , U b) {
{ a == b } -> bool ; // compare Ts and Us with == yielding a bool
{ a != b } -> bool ; // compare Ts and Us with != yielding a bool
};
The requires-expressions were invented by Andrew Sutton as part of his implementation of

the Concepts TS. They turned out to be so useful that the users insisted they become part of the
standard.
The =T gives a default for the second type argument so that Equality_comparable can be used
for a single type.
The notation for the use patterns were invented by Bjarne Stroustrup in real time at the Palo
Alto meeting based on ideas from 2003 [Stroustrup and Dos Reis 2003b]. This notation and the
ideas that it is based upon do not involve function signatures or function-table implementation.
There is no specific mechanism for saying that a type matches a concept, but if someone wants
to, the general C++11 static_assert can be used:
static_assert ( Equality_comparable < int >); // succeeds
static_assert < Equality_comparable < int , long >); // succeeds
struct S { int a; };
static_assert ( Equality_comparable <S >); // fails because structs don 't
// automatically get == and !=
The associated type notion from C++0x (and earlier [Stroustrup 2003]) is also supported:
template < typename S >
concept Sequence = requires (S a) {
typename Value_type <S >; // S must have a value type .
typename Iterator_type <S >; // S must have an iterator type .
{ begin (a) } -> Iterator_type <S >; // begin (a) must return an iterator
{ end (a) } -> Iterator_type <S >; // end (a) must return an iterator
{ a. begin () } -> Iterator_type <S >; // a. begin () must return an iterator
{ a. end () } -> Iterator_type <S >; // a. end () must return an iterator
requires Same_type < Value_type <S >, Value_type < Iterator_type <S > > >;
requires Input_iterator < Iterator_type <S > >;
};
Note the repetition to be able to accept both a.begin() and begin(a). The lack of uniform function
call hurts (ğ6.1), (ğ8.8.3).
6.3.4 Concept Name Introducers. One thing we learned from use was that primitive uses of concepts
were repetitive. We used too many requires-expressions directly in requires-clauses and we used
too many łsmallž concepts. Our requirements looked a lot like code written by novice programmers:
too few functions, too little abstraction, and too few symbolic names.
Consider the standard merge family of functions. These functions all take three sequences and
need to specify the relationships among those sequences. That’s three requirements on the type of
sequences plus three requirements on the relationships among the elements of those sequences.
First try:
template < Input_iterator In1 , Input_iterator In2 , Output_iterator Out >
requires Comparable < Value_type < In1 >, Value_type < In2 >>
&& Assignable < Value_type < In1 >, Value_type < Out >
&& Assignable < Value_type < In2 >, Value_type < Out >
Out merge ( In1 , In1 , In2 , In2 , Out );
That’s tedious and such patterns of type name introduction are very common; for example, there
are at least four merge functions in the STL. Tedious and repetitive code is error-prone and hard
to maintain. We soon learned to use more multi-argument concepts to define common patterns of
requirements among types:
template < Input_iterator In1 , Input_iterator In2 , Output_iterator Out >
requires Mergeable < In1 , In2 , Out >
This then became too messy for Andrew Sutton, who in 2012 probably had written more code
using concepts than anyone else. He suggested a mechanism for saying łintroduce a set of type
names for types that must satisfy a concept.ž That reduced the merge example to its logical
minimum:
Mergeable { In1 , In2 , Out } // concept name introducer
It is amazing what you can learn just by trying! It is also amazing how much resistance novel
notations and solutions can encounter from people who have yet to experience the problem.
6.3.5 Concepts and Types. Many still saw (and see) concepts as a variant of the type-of-type idea.
Yes, a concept taking a single type argument can be seen as a type of a type, but only the simplest
uses fit that pattern.
Most generic functions (algorithms) take more than one template argument and for such a
function to make any sense, those argument types must somehow relate to each other. Then, we
must use multi-argument concepts. For example:
template < Forward_iterator Iter , typename Val )
requires Equality_comparable < Value_type < Iter >, Val >
Forward_iterator find ( Iter first , Iter last , Val v)
{
while ( first != last && * first != v)
++ first ;
return first ;
}
Crucially, the multi-argument concepts directly address the need to handle implicit conversions
and mixed-type operations. Together with Gabriel Dos Reis, I had considered the possibility of
specifying all constraints on each argument in isolation from other arguments as early as 2003
[Stroustrup 2003; Stroustrup and Dos Reis 2003b]. This would involve
• parameterization (e.g., an Iterator parameterized with its value type)
• some form of inheritance (e.g., a Random_access_iterator is a Forward_iterator)
• the ability to apply more than one concept to a template argument type (e.g., the element of
a Container must be Value_type and a Comparable)
• combinations of those three techniques.
The result was very complex template argument type constraints. We deemed that complexity
unnecessarily and unmanageable. Consider x+y and y+x where x and y are of different template
argument types, X and Y. Dealing with each template argument on its own, we would have to
parameterize X with Y and Y with X. In a pure object-oriented language, that might seem natural;
after all, there would be two methods to cope with +, one in X’s hierarchy and one in Y’s hierarchy.
However, I rejected that solution for C++ back in 1982. To complete the picture, we must add
implicit conversions (e.g., to handle x+2 and 2+x) . The multi-argument concepts exactly match
the way C++ resolves such examples and avoid most of the complexity.
This decision was revisited repeatedly over the years and confirmed. In the context of the C++0x
concepts effort, people tried applying the standard academic systems as found in Haskell typeclasses
and Java constraints. However, that failed to deliver the simplicity of implementation and use
needed for large scale use.
Where a generic use fits the type-of-type pattern, concepts supports it quite elegantly.
• A type specifies the set of operations that can be applied to an object (implicitly and explicitly),
relies on function declarations and language rules, and specifies how an object is laid out in
memory.
• A concept specifies the set of operations that can be applied to an object (implicitly and
explicitly), relies on use patterns reflecting function declarations and language rules, and
says nothing about the layout of the object. Thus, a concept is a kind of interface.
My ideal is to be able use concepts wherever we use a type and in the same way. Except for
defining layout, they are very similar. Concepts can even be used to constrain the type of variables
with their type determined by their initializer (constrained auto variables (ğ4.2.1)). For example:
concept Integer = Same <T , short > || Same <T , int > || Same <T , long >;
Integer x1 = 7;
int x2 = 9;
Integer y1 = x1 + x2 ;
int y2 = x2 + x1 ;
void f( int &); // a function

void f( Integer &); // a function template
void ff ()
{
f( x1 );
f( x2 );
}
C++20 comes close to fulfilling this ideal. To get that example to work in C++20, we have to add a
logically redundant auto after each use of the concept Integer (ğ6.4). On the other hand, in C++20,
we can use the standard-library concept integral instead of the obviously incomplete Integer.
6.3.6 Improvements. At the start of the Concepts TS effort, a concept was a constexpr function
(ğ4.2.7) returning a bool. That made sense since we saw concepts as compile-time predicates. Then
Gabriel Dos Reis got variable templates accepted into C++14 (ğ5.2). Now, we had a choice:
// function style :
concept bool Sequence () { return Has_begin <T >() && Has_end <T >(); }
// expression style :
concept bool Sequence = Has_begin <T > && Has_end <T >;
We could live happily with either style, but having both, users of a concept would have to know
which style was used in its definition to use parentheses correctly. That quickly became impractical.
The functional style allows for overloading of concepts, but we had only few examples of concept
overloading and decided we could do without those, so we simplified to use variable templates
exclusively for concepts. Andrew Sutton pioneered the consistent use of the expression form of
concepts.
We (Andrew Sutton, Gabriel Dos Reis, and I) always knew that explicitly mentioning that a
concept returned a bool was redundant. After all, a concept is by definition a predicate. However,
we decided not to mess with the grammar and concentrate on semantically significant topics. Later,
people were latching on to the redundant mention of bool as an argument against the concept
design. So we fixed it and no longer have to mention bool.
The removal of bool was part of a set of suggested improvements from Richard Smith that
included a more precise specification of what constituted an atomic predicate and simplifications
of the matching rules [Smith and Sutton 2017]. We now use the expression style:
// expression style :
concept Sequence = Has_begin <T > && Has_end <T >;
6.3.7 Syntax Equivalences. The concepts TS supports three notations for using concepts in function
declarations:
• explicit requires-clauses for full generality
• the shorthand notation for type-of-type uses
• the natural notation (also known as the terse notation, the conventional notation, and more)
The fundamental idea was to let the programmer use a notation that closely matched the needs
of a particular declaration without drowning that definition with notation needed by more complex
declarations. To allow the programmer a free choice and in particular to allow the notation to be
adjusted as functions change during initial development or maintenance, these notation styles were
defined as equivalent:
void sort ( Sortable &); // natural notation
is equivalent to
template < Sortable S > void sort (S &); // shorthand notation
is equivalent to
template < typename S > requires Sortable <S > void sort ( Sortable &);
Users were rather happy with this and tended to prefer the natural and shorthand notations for
most declarations. However, some committee members reacted in horror to the natural notation (łI
can’t see that it’s a template!ž) and rather liked the most explicit notation using requires because
it can express even the most complex examples (łwhy would you want anything more than that?ž).
My interpretation is that we have a clash of two views of what’s simple:
• I can write my code in the simplest and shortest way
• I need only to learn one notation
I am in favor of the former view, considering it a good example of the Onion Principle (ğ4.2).
The natural notation became a focus of strong opposition to concepts. I ś and others ś insisted
on the elegance of
void sort ( Sortable &); // natural notation
We saw (and see) it being a useful and elegant step towards making generic programming
łjust ordinary programmingž rather than a dark art with different syntax, different source code
organization (łheader onlyž), and different coding styles (e.g., template metaprogramming (ğ10.5.2).
Modules addresses the source code organization issues (ğ9.3.1). In addition, that łnaturalž syntax
addresses frequent loud complaints about the template syntax being verbose and clumsy. I agree
with those complaints. The prefix template<...> syntax was not my first choice when I designed
templates. It was forced upon me by people worried that templates would be misused by less
competent programmers, leading to confusion and errors. The heavy syntax for exception handling,
try { ... } catch( ... ) { ... }, was a similar story [Stroustrup 2007]. It seems that for every new feature
many people demand a LOUD syntax to protect against real and imagined potential problems. After
a while, they then complain about verbosity.
However, a large minority of committee members insisted that the natural syntax would lead to
confusion and misuse because people, especially less experienced programmers, would not realize
that functions defined in this way were templates and therefore different from other functions.
Having used and taught concepts for years without observing these problems, I wasn’t particularly
concerned about such hypothetical problems, but the opposition turned out to be solid. People
simply knew that such code was dangerous. The prime example was
void f(C &&); // Danger : Is C a concept or a type ?
The meaning of C&& is somewhat different dependent on whether f is a function template or

an łordinaryž function. To my mind, this difference in the semantics of C&& is a most unfortunate
design mistake in C++11 that we should try to correct, rather than let it affect the definition of
concepts. The possibility of a misunderstanding is undeniably real and would certainly happen to
someone once the facility became used by millions. However, I haven’t seen it in real life and doubt
that the relatively experienced programmers who would write code where the difference matters
would have real trouble with it. In other words, I considered it an example of łthe tail wagging the
dog;ž that is, an obscure example blocking a feature that could benefit large numbers of users.
I’m also pretty sure that my aim of making generic programming as similar to łordinaryž pro-
gramming as possible wasn’t universally shared. There are still people who think that generic
programming is beyond that vast mass of programmers. I still see no evidence of that.
6.3.8 Why No Concepts in C++17? I had hoped and expected to see concepts in C++17. Of the
extensions that I considered feasible in the 2017 time frame (ğ9.2), I saw concepts as the one that
would yield the most significant improvement to C++ programmers’ basic vocabulary. It would
eliminate the need for much ugly and error-prone template metaprogramming (ğ10.5.2), would
simplify the precise specification of libraries, and would significantly improve the design of libraries.
I fear that was part of the problem: concepts would directly affect essentially all of the voting
members. Some were more comfortable with their old ways, many had no experience with concepts,
and some considered them an untried (łacademicž/łtheoreticalž) idea.
It was such worries ś fueled by the fiasco of C++0x concepts (ğ6.2) ś that led to us having a
TS [Sutton 2017] in the first place. We didn’t have experience with Technical Specifications for
language features, but it seemed worth trying and Andrew Sutton’s concepts implementation in
GCC was still new so caution seemed warranted. At the Bristol standards meeting (2013), Herb
Sutter strongly argued for the TS route whereas J-Daniel Garcia and I warned against likely delays.
I also pointed to dangers of considering concepts separately from generic lambdas (ğ4.3.1), but
łcautionž and łwe need more experiencež are strong arguments in a standards committee. In the
end, I voted in favor of concepts TS. I now consider that a mistake.
In 2013, we had an implementation and a pretty good specification (primarily thanks to Andrew
Sutton), but completing the concepts TS still took three years. I am unable to discern a difference
in rigor in the treatment of the TS and inclusion into the ISO standard proper. However, at the 2016
Jacksonville meeting, when concepts as described in the TS came up for a vote for inclusion into
the standard, all the earlier arguments against were repeated. It seemed that the opponents had
just ignored concepts for three years. I even heard arguments against concepts that were valid for
C++0x concepts but had never been relevant to the concept TS design. People again argued for
łcautionž and łwe need more experience.ž As far as I could tell, there were more people present in
Jacksonville who had not tried concepts than there had been in Bristol ś partly an effect of the
increased size of the committee. In addition to all the objections I had heard over the previous
decade, novel objections were raised and untried designs were suggested in full committee, yet
taken seriously.
At the meeting in February 2016 Jacksonville, Ville Voutilainen (the EWG chair) proposed to
move the concepts as specified in the Concepts TS [Voutilainen 2016c]:
... łprogrammers are aching to get the language feature into their hands, and it’s high
time we ship it to them. Conceptifying the standard library will take its time, and we will
not find major design issues for Concepts while doing it. We shouldn’t keep programmers
waiting for the language feature out of concern for hypothetical design issues that we
have no proof of, have some amounts of counter-proof of, and very likely don’t exist.
For the benefit of C++ users everywhere, let’s ship Concepts the language feature in C++17.ž
He was strongly supported by many, notably Gabriel Dos Reis, Alasdair Meredith (formerly, the
LWG chair) and me, but (despite a positive EWG vote earlier in the week), the vote went against us:
25 in favor, 31 against, 8 abstained. My interpretation is that the users voted for and the language
technicians voted against, but that may be considered sour grapes.
At the same meeting, uniform call syntax (ğ8.8.3) was voted down and coroutines (ğ9.3.2) sent to
a TS, basically ensuring that C++17 would be a minor release of the standard (ğ8).
6.4 C++20 Concepts

In 2017, as one of the very first features for C++20, WG21 voted what they considered the foun-
dational and uncontroversial parts of the Concepts TS [Sutton 2017] into the Working paper
(ğ6.3.2):
• explicit requires-clauses for full generality; e.g., requires Sortable<S>
• the shorthand notation for type-of-type uses; e.g., template<Sortable S>
The natural notation (e.g., void sort(Sortable&); (ğ6.3.7) was left out as controversial. The
reasons given for wanting it left out were:
• It is not obvious that void sort(Sortable&); is a template.
• The meaning of void f(C&&); depends on whether C is a concept or a type.
• In Iterator foo(Iterator,Iterator); must Iterator be the same type in all three cases or can
the types be separately constrained?
• The natural syntax is confusing and hard to teach.
• How do we constrain the parameter in template<auto N> void f();?
These objections were not new, but they were accompanied with a lot of proposals for novel
syntax [Honermann 2017; Keane et al. 2017; Köppe 2017a; Riedle 2017; Sutter 2018a]. These proposals
were all different and all incompatible with the Concepts TS. They were presented with significant
amount of emotion in meetings. None were supported by actual experience. To contrast, my
position was based on about four years of teaching, much experimental use, some industrial use,
and several uses in proposed standard-library components (e.g., iterator facade [Dawes et al. 2016],
tuple implementation [Voutilainen 2016b], ranges [Niebler et al. 2014]).
At the Jacksonville meeting (2018), Tom Honerman proposed to remove the natural syntax
and offered an alternative [Honermann 2017]. I defended my position and the concept TS design
[Stroustrup 2017a,b]. My basic defense was
• The natural syntax has not caused problems in real-world teaching and use for more than
five years.
• Users like it.
• There are no technical ambiguities.
• It simplifies common cases.
• It is part of the drive to make generic programming more like ordinary programming.
That failed to convince anyone who was opposed in the first place, so the natural syntax was
not moved to the working paper for C++20.
The last objection came from a new C++17 minor feature, auto for value arguments [Touton
and Spertus 2015], and became a focal point for objections:
template < auto N > void f ();
People wanted to syntactically distinguish value and type template arguments. Typically, that
would imply that the shorthand syntax that had been the stable of proposal and uses since 2002
would no longer be valid.
template < typename T > void f(T &); // proposed banned
In mid-2018, I suggested a minimal compromise [Stroustrup 2018b] that:

• Preserved the meaning of template<Concept T> void f(T&);
• Used a template prefix to identify template using the natural notation, e.g., template void
f(Concept&);
That proposal succeeded, but so did a very different proposal from Herb Sutter [Sutter 2018a].
We were now in the extraordinary position of having two very different and incompatible proposals
each supported by a large majority of the EWG. This deadlock opened the door for Ville Voutilainen
(the EWG chair) to suggest a variant that was accepted with massive support in November 2018
[Voutilainen et al. 2018]:
• Preserved the meaning of template<Concept T> void f(T&);
• Used an auto to identify template arguments using the natural notation, e.g., void f(Concept
auto&);
For example
// almost natural notation :
void sort ( Sortable auto & x ); // x must be Sortable
Integral auto ch = f( val ); // the result of f( val ) must be Integral
Integral auto add ( Integral auto x , Integral auto x ); // can use a wider type
// to prevent overflow
The łnatural notationž was renamed the łabbreviated syntaxž even though it is not simply an
abbreviation.
I supported the compromise even though I consider that use of auto redundant, distracting,
and compromising my aim of making generic programming łjust ordinary programming.ž Maybe
sometime in the future, people will (as Herb Sutter suggested at the time) agree and make auto after
a concept name redundant. I am not holding my breath, though; there is a large group of people
who consider syntactic markers for implementation techniques important. Maybe auto-completion
features of IDEs will save users from writing the redundant autos.
Sadly, there was no consensus for re-introducing concept name introducers (ğ6.3.4). The lack of
a sufficiently conventional syntax was a major stumbling block. Also, quite a few people still seem
not to believe in their utility.
The years of delay introducing concepts caused long-lasting harm. Ad-hoc designs based on traits
and enable_if proliferated. A generation of programmers was brought up on low-level, typeless
metaprogramming.
6.5 Naming of Concepts

One of the final discussions about concepts before shipping C++20 was about naming conventions.
Naming is always a tricky topic. In my early work with concepts, I named concepts the way I
usually name types that are not in the standard: capitalize the first letter as for proper nouns and
separate words with an underscore for readability (e.g., Sortable and Forward_iterator). Others
(notably the Indiana team) used CamelCase (e.g., Sortable and ForwardIterator). Unfortunately,
this naming convention crept into the standard text [Carter 2018] and caused some confusion by
being different from all other names in the standard library. There, underscores and no capital
letters are used (except for MACROS and three obscure, hard to find, examples). Some people
then thought that the different naming convention aimed to distinguish the łnovel and difficultž
concepts from łordinary constructsž, such as functions and types.
When I noticed that rationalization, I strongly disliked it. In C++, we generally don’t encode
types in the names of entities, but I thought it too late to change the naming style. In 2019, Herb
Sutter responded to my grumbles by proposing to rename all standard-library concepts to follow
the usual standard-library naming convention [Sutter et al. 2019]. All the major concept designers
and the range-library (ğ9.3.5) designers signed on as co-authors. Another reason for the change
was that we were beginning to see clashes between the CamelCase names of standard-library
concepts and CamelCase names in other libraries. One reason for using CamelCase (or using my
convention of capitalizing types) is exactly to avoid clashes with the standard library. So, we now
have sortable, forward_iterator, etc.
The C++20 standard-library has some 70 concepts covering the needs of invocation of oper-
ations, use of basic types, ranges, and the standard algorithms, including constructible_from,
convertible_to, derived_from, equality_comparable, invocable, mergeable, range, regular,
same_as, signed_integral, semiregular, sortable, swappable, and totally_ordered. They will
guide much C++ library design. Note that many of those 70 concepts are not fundamental but
simply there for notational convenience or as building blocks.
7 ERROR HANDLING
Error handling is ś and I suspect will remain ś a hotly debated topic. Many people have strong
opinions ś some based on solid experience in a variety of application areas ś and many different
techniques have been developed over the last 50 years or so. This is an area where the demands of
performance, generality, and reliability easily get into conflict.
As is often the case with C++, the problem is not that we don’t have a solution, but that we
have many. It is fundamentally hard to address the diverse needs of the C++ community with a
single mechanism, but looking at a subset of the problem people often think they have the solution
[Stroustrup 2019a].
7.1 Background
From C, C++ inherited a variety of techniques based on error return codes, errors represented as
special values, global state, local state, and callbacks. For example:
double sqrt ( double d ); // set errno ==33 if d is negative

int getchar (); // return -1 if end of file
char * malloc ( int ); // return 0 if allocation failed
Early users of C++ (1980s) found those techniques confusing and insufficient. Returning a
(value,error-code) pair became popular, adding to the variety and confusion. For example:
Result r = make_window ( arguments ); // Result is a ( value , error ) pair

If (r. error ) {
// ... handle error ...
}
Shape * p = r. value ;
That led to code becoming cluttered with tedious repeated error checks. When using error codes,
it can be hard to distinguish the main logic of a program from error handling. The main line of a
program (łthe business logicž) gets deeply intertwined with the handling of a multitude of rare and
obscure errors. This can be a serious problem in the not unusual systems where the error-handling
is the majority of code and the most complex part of the code.
The use of a class containing a (value,error-code) pair came with a significant cost. In addition to
the cost of the test of the error-code, many ABI’s did not pass even small structures in registers, so
not only was more information passed (often twice the amount), but it was passed in a way that
could be an order of magnitude slower. Sadly, this problem persists to this day (2020) in many ABIs,
especially ABIs for embedded systems (designed for C code).
Furthermore, there was no really good way of using error codes to handle failure in a constructor
(there is no return value in which to pass an error-code) and the then-fashionable large complicated
class hierarchies led to complex and error-prone handling of the variety of potential errors from
sub-object creation.
In addition, all traditional error handling techniques suffer from people forgetting to check for
errors. This was ś and still is in 2020 ś a major source of errors. Minimizing errors from incomplete
or complex error handling was a major goal for C++ exceptions.
Exceptions for C++ were designed in 1988-89 to address the mess of complicated and error-prone
error-handling techniques then prevalent. They were documented in the ARM [Ellis and Stroustrup
1990] and adopted into ANSI C++ as part of the base document for standardization [Stroustrup
1993].
Compared to some exception designs, the one for C++ was complicated by the need to use C++
code in combination with code in other languages, especially C. Consider a C++ function, f() calling
a C function g() that in turn calls a C++ function h(). Now h() throws an exception for f() to catch.
In general, a C++ function does not know the implementation language of a called function. This
scenario precludes adoption of implementation schemes that modify all function signatures to add
an łexception propagation argumentž or implicitly add a return code to the return type.
Exceptions together with RAII (ğ2.2) did solve many of the most vexing error-handling problems
(such as how to handle an error in a constructor and how to handle errors detected far from code
that could handle them) with a very minor run-time cost compared to use of other techniques
(usually less than 3% by the mid-1990s and sometimes even cheaper than alternatives). Exceptions
were never uncontroversial, though, and I underestimated their potential for controversy.
7.2 Real-World Problems

There have always been applications for which the use of exceptions was unsuitable. Examples
include
• Systems where the memory is so limited that the run-time support needed for exception
handling crowds out needed application functionality.
• Hard-real time systems where the tool chains cannot guarantee prompt response after a
throw (e.g., [Lockheed Martin Corporation 2005]).
• Systems relying on multiple unreliable computers so that immediate crash-and-restart is a
reasonable (and almost necessary) way of dealing with errors that cannot be handled locally.
Consequently, most C++ implementations always had a no-exception switch. On the other hand,
there are problems for which error codes provide no good solution:
• Constructor failures ś there is no return value (except the constructed object itself). Pure RAII
must be replaced by explicit checks on object states.
• Operators ś there is no place to return an error indicator from ++, *, or ->. You will have to use
non-local error indicators or live with an impoverished notation, e.g., multiply(add(a,b),c)
rather than (a+b)*c.
• Callbacks ś where the function using the callback should be able to invoke functions with a
wide variety of possible errors (often, callbacks are lambdas (ğ4.3.1)).
• Non-C++ code ś there is no way to propagate an error through a function that is not C++ and
hasn’t been written specifically to deal with error codes.
• Addition of a new kind of error deep in a call chain ś every function on the call chain must be
prepared to handle or propagate a new kind of error (e.g., a network error in a program not
specifically designed to access data across a network).
• People forgetting to test a return code ś there are clever schemes to try to ensure that error-
codes are checked consistently, but they are either incomplete or rely on exceptions or
termination once a failure to check is detected (e.g., [Botet and Bastien 2018]).
In addition, there were practical problems related to the use of exceptions:
• Some people could not introduce the use of exceptions because their code already was a large
irreparable mess of unprincipled pointer use. Quite often, such people directed their critique
toward exceptions rather than their old code.
• Some people (many) simply didn’t understand or even didn’t know of RAII (ğ2.2) and used
exceptions as merely as an alternative return mechanism mirroring the use of error-return
codes. Typically, code using try-catch as a form of if-then is uglier, larger, and slower than
proper use of error codes or RAII.
• Many implementations of exceptions were slow because implementers generalized to handle
other kinds of exceptions (e.g., Microsoft’s łstructured exceptionsž), prioritized debugging (e.g.,
GCC walks the stack twice after a throw to preserve backtraces), used a single mechanism
to serve a variety of languages (equally badly), or simply didn’t expend much development
effort on optimization.
• Exception handling have become relatively slower over the years because significant efforts
have been spent optimizing non-exception cases. I suspect there are significant optimization
opportunities left. For example, Gor Nishanov reported up to 1000 times speed improvements
for some simple optimizations related to coroutine implementations on Windows and Linux
[Nishanov 2019a]. Significant space improvements will likely be harder to achieve, though.
Some recent experiments are promising [Renwick et al. 2019].
• To get exceptions accepted, we had to add exception specifications [Stroustrup 2007]. They
never provided the improved maintainability that their proponents claimed, but they did
deliver the verbosity and overhead that their opponents (including me) insisted they would.
Once exception specifications were in the language, many people felt encouraged to use
them and blamed exceptions in general for the resulting problems. Ironically, the people
who insisted most loudly on getting exception specifications went on to help design Java.
Exception specifications were deprecated in 2010 and finally removed in 2017 (ğ4.5.3). Partly
as an alternative, C++11 introduced noexcept as a simpler and more efficient mechanism
for controlling exceptions (ğ4.5.3).
• Catching an exception is done by specifying the type of exception to be caught. As a result,
the implementations of throw and catch got entangled with the mechanism for runtime type
information (RTTI [Stroustrup 2007]). This caused inefficiencies and complexity. In particular,
it caused memory to be consumed (by the data needed for RTTI) even if an application never
relied on RTTI for distinguishing exceptions and precluded optimization for simple cases.
Furthermore, relying on RTTI made it hard to optimize the type matching where dynamic
linking was used. Basically, exception handling implementations were tuned for the rare most
complicated case. This was made worse when an exception class hierarchy was added to the
standard library and people were encouraged to use that for even the simplest cases. For class
hierarchies that can be statically analyzed (as in many embedded systems) fast constant-time
type matching is possible [Gibbs and Stroustrup 2006]. The fact that exceptions are part of
the platform ABIs makes it very hard to change early overdesigned implementations.
• Some people insist that only a single method of error handling be used and usually concludes
that since exceptions are not suitable for every case, that method must be error-codes. The
problems with error codes are then deemed łjust inconveniences.ž
• Some people believed the persistent rumors of inefficiencies based on worst-case scenarios
and/or unrealistic comparisons, such as leaving error-code handling in place after adding
exceptions, comparing incomplete error-handling to exception-based handling, or using
exceptions to handle ordinary choices rather than for just handling errors that cannot be
handled locally. There has been far too little serious investigation into the cost of exception
and its alternatives. I suspect that the myths about exceptions have had more influence than
any fact.
The net effect was a bifurcation of the C++ community into exception and non-exception camps.
De facto, łuse no exceptionsž is a dialect and dialects is one of the things standards are meant to
avoid (ğ3.1). A dialect can be an advantage for an individual organization or community, but harms
the C++ community as a whole by complicating sharing of code and competences.
It has been claimed that the problem with exceptions comes from it violating the zero-overhead
principle (e.g., [Sutter 2018b]). For an error-handling scheme that responds to every error by
terminating, any error-handling mechanism is obviously overhead, so the zero-overhead principle
is violated (unless you take into account the cost of handling termination, e.g., in another processor).
At the time of the design of exceptions, we considered that and thought it acceptable based on
the argument that such cases were rare, that there was no run-time cost unless an exception was
thrown, and that the tables used to implement exceptions could be kept in virtual memory [Koenig
and Stroustrup 1989]. Where virtual memory isn’t available and memory is scarce, the use of tables
to implement exceptions can be a serious problem. The main concern at the time was systems where
some form of error propagation and error handling were needed. In such cases, the zero-overhead
was interpreted łas no overhead for exceptions compared with the use of error codes for the same
degree of rigor of error handling.ž
Today, the confusion about error handling is worse than ever and the number of alternative
techniques for handling errors is higher than ever, causing much confusion and harm. Given N
ways of handling errors, someone comes up with a solution just to find that the old solutions
don’t go away so now we have to cope with N+1 ways (łthe N+1 problemž). If an organization
has M separately developed programs using N libraries, we may even have an N*M problem. The
introduction of exceptions could be seen as having increased the number of popular ways of
handling errors from 7 to 8. In 2015, Lawrence Crowl wrote an analysis of the problems [Crowl
2015a].
The problem of multiple error-reporting schemes is felt most acutely by writers of foundational
libraries. They cannot know what their users prefer and their users may very well have many
different preferences. The authors of the C++17 filesystem library (ğ8.6) chose to duplicate their
interfaces: for each operation, they offer two functions, one that throws in case of error and another
that sets a standard-library error_code passed to it as an argument:
bool create_directory ( const filesystem :: path & p ); // throws in case of error
bool create_directory ( const filesystem :: path & p , error_code & ec ) noexcept ;
This, of course, is a bit verbose and pleases only people who like either exceptions or error_codes.
Note how the bool return value is supplied so that people don’t have to use try or directly test
the error_code all the time. The fact that the file systems (quite properly, IMO) uses exception for
rare errors doesn’t please people who consider exceptions fundamentally flawed. In particular, it
requires that exception support be present.
7.3 noexcept Specifications

Using noexcept (ğ4.5.3), people can suppress all exception throws from a function and allow callers
to ignore the possibility of exceptions being thrown.
The use of noexcept can reassure people who worry about performance problems (real and
imaginary). It can also improve optimization by decreasing the number of control paths, but only if
the programmers don’t add those paths back by testing return codes. There are lots of low-level
functions that are naturally noexcept, such as most C functions.
The use of noexcept can simplify error handling (if a function can’t throw, we don’t need to
catch any exceptions) or complicate it (if a function cannot throw, yet may fail, we have to use
another error-reporting mechanism). In particular, a noexcept on a path between a throw and its
handler can turn a throw into termination. Thus, making a function noexcept during maintenance
can cause failure of a previously correct program.
Note that one of the significant reasons that exceptions were added to C++ was to support
applications where an error must never lead to an unconditional abort. An exception simply signals
that a failure has happened and any code on the path from main() to the throw-point can take
control. In particular, this supports the important use case of doing some local cleanup (e.g., flushing
some output buffers or adding an error-report to a log file) before terminating.
7.4 Type System Support

The traditional approach to logical and performance problems in C++ is to move computation
from run time to compile time. Obviously, the possibility of integrating exceptions with the static
type system was seriously considered in the 1980s and repeatedly reconsidered later. If exceptions
were part of every function type, programs would be better type checked, functions more self-
documenting, and exception handling easier to optimize.
A major reason not to make exceptions part of the type system was the observation that if
exceptions were part of a function’s type, a change to the set of exceptions it could possibly throw
would require all functions calling the function to be recompiled. In a world where most major
programs are composed from many separately developed libraries, this would lead to disastrous
brittleness and unmanageable dependencies [Stroustrup 1994].
There are also obvious problems related to pointers to functions. There was ś and still is ś a lot
of C-style code in most major C++ programs. The main parameterization mechanism for C-style
generic code (e.g., qsort parameterized by its comparison criteria) and callbacks (e.g., in GUIs) is
pointers to functions.
If I need a pointer to function and exceptions are part of the type system, I either have to decide
to always require exceptions from the function pointed to, never accept exceptions, or somehow
handle both alternatives. Handling both is hard unless support for type inquiry or overloading
based on exceptions are added to the language. Having decided what kind of pointer-to-function
argument(s) to accept, I now have to adjust the error checking in the calling function to match.
Even if this could be handled in C++, interaction with C would be impeded: how would a pointer
to a C++ function be passed to C? For example, what would callbacks from C to a C++ program
relying on exceptions look like? Obviously, the original C++ exceptions would not go away, so we
would have four alternatives: error codes, compile-time checked exceptions (e.g., [Sutter 2018b]),
current exceptions, and noexcept. Only the current exceptions and non-local error-codes do not
affect the type system or the calling conventions (ABIs). Fortunately, few functions require two
pointers to functions or we would risk having to deal with 16 alternatives. If different exception
types were accepted (as for the current exceptions), the chaos would be complete.
In modern C++, this kind of problem would persist in different guises for other callback mecha-
nism, such as objects with member functions meant to be called, function objects, and lambdas.
My conclusion (as endorsed by WG21) was and is that adding exceptions to C++’s static type sys-
tem would lead to brittleness, significant increases in complexity of code, serious incompatibilities,
and problems with interaction with C code. This was appreciated in 1989.
7.5 Back to Basics

Fundamentally, I think C++ needs two error-handling mechanisms:
• Exceptions ś for errors that are rare or cannot be handled by an immediate caller.
• Error codes for errors that can be handled by an immediate caller (often łdisguisedž in easy-
to-use test operations or returned from a function as a (value,error-code) pair).
Consider:
void user ()
{
vector < string > v {" hello !"};
for ( string s; cin >>s; )
v. push_back (s );
auto ps = make_unique < Shape >( read_shape ( cin ));
Smiley_face face { Point {0 ,0} ,20};
// ...
}
This example is artificial, but stylistically not atypical. The user() function offers many oppor-
tunities for unlikely errors: memory exhaustion, read errors, construction failures (e.g., deep in
the hierarchy of Smiley_face). In addition, the use of a unique_ptr<Shape> protects against a
memory leak. If we used explicit error-codes instead of exceptions, we would need at least five
error-checks in this function, doubling the amount of source code, plus a few more checks inside
the various constructors. Without RAII (and its integration with exceptions) the code would bloat
further still. On average, more code implies more errors. This is especially so when the added code
complicates control flows. This point is often underappreciated by people who argue from small
examples. For small examples, łjust one testž doesn’t matter much and is relatively hard to forget.
On the other hand, some errors should be expected and for those checks of some form of error
code is preferred:
ifstream f {" Myfile "};
if (! f) {
// ... deal with error ...
}
// ... use f ...
Here, the error-code is hidden inside the stream state for ease of use.
So, ideally, there would be just two ways of handling an error: However, I do not know how to
get to such an ideal state. There are on the order of a dozen variants of the (value,error-code) pair
idea in wide use (e.g., std::map::insert()) and several new ones were being discussed in WG21 in
2018 (e.g., [Botet and Bastien 2018; Sutter 2018b]). Even if the committee could agree on one, there
would still be at least a dozen widely used error-handling schemes łout therež each supported by
large groups of devoted followers, many with millions of lines of hard-to-change code.
There has been little serious research on the topics of performance of exceptions and reliability
of the resulting code in C++ ([Renwick et al. 2019] is an exception). There are, however, many
small unscientific studies and lots of loudly expressed opinions ś often claiming exceptions to be
inherently slower than various forms of checking of error-codes. That is not my experience. To
the best of my knowledge no half-way serious study has failed to observe that there are realistic
examples where error codes win big and realistic examples where exceptions win big. In this
context, łbigž means integer factors, rather than a few percent.
Run a simple performance test: go N levels deep into a call sequence and then report an error.
If the error is rare, say 1:1000 or 1:10000 and the call nesting is deep, say 100 or 1000, exception
handling is much faster than explicit tests. If the call depth is 1 and the error happens 50% of the
time, explicit tests win big. The call depth and error probability determine the difference between
the such examples. My naive, but potentially useful question is łhow rare must an error be to be
considered exceptional?ž Unfortunately, the answer is łthat depends.ž It depends on the complexity
of the code, the hardware, the optimizer, the implementation of exception handling, and more. C++
exceptions were designed assuming an answer at least in the 1:100 region. In other words, that
propagation of error indicators is far more common than explicit handling.
The space problem is likely to be harder to solve than the run-time problem. For systems
relying on immediate termination for all errors that cannot be handled locally, I could imagine an
implementation simply terminating on a throw but if errors are to be propagated and handled, the
tradeoff difficulties won’t go away.
Any solution of the error-handling mess is liable to hit the N+1 problem (ğ4.2.5) [Stroustrup
2018a].
Curiously enough, one of the worries at the time where exceptions were included in C++ was
that they were not sufficiently general. A significant number of people considered resumption
semantics essential [Stroustrup 1993]. My guess at the time was that allowing resumption would
have slowed down exception handling by at least a factor of two.
8 C++17: LOST AT SEA

After the minor standards revision, C++14, C++17 [Smith 2017] was to have been a major revision
of the standard. C++17 has quite a few new features, but none that I would deem major. The key
question about C++17 is łWhy did all of that hard work not lead to more significant improvements?ž
The processes that led to the C++11 and C++14 successes were in place, the standards community
was more experienced, larger, and more enthusiastic.
C++17 had about 21 new language features (depending how you count), including:
• Constructor template argument deduction ś simplify object definitions (ğ8.1)
• Deduction guides ś an explicit notation for resolving constructor template argument deduc-
tion ambiguities (ğ8.1)
• Structured bindings ś simplify notation and eliminate a source of uninitialized variables
(ğ8.2)
• inline variables ś simplify the use of statically allocated variables in header-only libraries
[Finkel and Smith 2016]
• Fold expressions ś simplify some uses of variadic templates [Sutton and Smith 2014]
• Explicit test in conditions ś a bit like conditions in for-statements (ğ8.7)
• Guaranteed copy elision ś eliminate many redundant copy operations [Smith 2015]
• Stricter expression evaluation order ś prevents some subtle order-of-evaluation mistakes
[Dos Reis et al. 2016b]
• auto as a template argument type ś type deduction for value template arguments [Touton
and Spertus 2016]
• Standard attributes to catch common mistakes ś [[maybe_unused]], [[nodiscard]], and
[[fallthrough]] [Tomazos 2015]
• Hexadecimal floating-point literals [Köppe 2016a]
• Constant expression if ś simplify compile-time evaluated code [Voutilainen and Vandevoorde
2016]
Unfortunately, this is not quite the full list of extensions. Quite a few are so small that they are
not easy to briefly describe.
The C++17 standard-library added about 13 new features plus many minor modifications:
• optional, variant, and any ś standard-library types for expressing alternatives (ğ8.3)
• shared_mutex and shared_lock (reader-writer locks) and scoped_lock (ğ8.4)
• parallel STL ś multi-threaded and/or vectorized versions of standard-library algorithms (ğ8.5)
• file system ś the ability to portably manipulate file-system paths and directories (ğ8.6)
• string_view ś a non-owning reference to an immutable sequence of characters [Yasskin
2014]
• Mathematical special functions ś including Laguerre and Legendre polynomials, beta func-
tions, Riemann zeta function [Reverdy 2012]
I have not been able to identify any unifying themes for the C++17 features. These features
look to be merely a set of łbright ideasž thrown into the language and standard library as voting
majorities could be found. That bothers me a lot. It bothers me a lot even when I like an individual
feature. There was no overall plan. That boded ill for the future ś something had to be done
[Stroustrup 2018d]. The creation of the Direction Group was part of WG21’s response (ğ3.2) (ğ9.1).
C++17 undeniably offers something that can help most programmers in minor ways, but nothing
that I can deem major. In this context, I define łmajorž as łmakes a difference to the way we think
about programming and organize our code.ž Here, I describe the facilities that I suspect will have
the largest positive impact.
I also examine a few examples that, despite serious consideration, did not make it into C++17:
• ğ6.3.8: Concepts (C++20)
• ğ8.8.1: Networking library
• ğ8.8.2: Operator dot (operator.())
• ğ8.8.3: Uniform Function Call

• ğ8.8.4: Default comparison operators ==, !=,<, <=, >, and >= for sufficiently simple types
• ğ9.3.2: Coroutines (C++20)
I suspect that had they been adopted, each of these would have been among the most significant
features of C++17. They fit a coherent view of what C++ should become (ğ9.2) and even a couple of
those would have drastically changed the ways C++17 could have been used.
Looking at C++11, I see a dense web of mutually supportive features leading to support of better
ways of writing code. I don’t see that for C++17. However, C++20 completes such a web to bring
C++ another major step forward (ğ9). It could be argued that C++17 was just a stepping stone on
the way to C++20, but the committee discussions gave little hint on that; the focus was firmly
on individual features. I have even heard claims that the train model (ğ3.2) precludes long-term
planning; it does not.
8.1 Constructor Template Argument Deduction

For decades, people have wondered why template arguments could be deduced from function
arguments, but not from constructor arguments. For example, in C++98, C++11, and C++14:
pair < string , int > p0 ( string (" Hi !") ,129); // no deduction
auto p1 = make_pair (" Hi !" s ,129); // p1 is a pair < string , int >
pair p2 (" Hi !"s ,129); // error : template arguments missing for pair
Naturally, I had considered the possibility of deducing template arguments from constructor
arguments when I first designed templates, but I was deterred by fear of ambiguities. There were also
technical obstacles to a solution, but Michael Spertus and Richard Smith overcame them [Spertus
and Smith 2015] so in C++17 we can write that last example (p2) without an error, removing the
need for the make_pair() workaround.
This simplifies the use of types, such as pair and tuple, and also writing of concurrent programs
using locks and mutexes (ğ8.4).
shared_lock lck {m }; // no explicit lock type
This is a rare example of mutually supportive features in C++17 leading to significantly simpler
code. Unfortunately, these simplifications were accepted (or not) on a case by case basis, rather
than in general, so the efforts to łfill holesž in the type deduction rules continues [Spertus et al.
2018].
In addition to what is described here, the facility provides a notation for resolving ambiguities.
For an example, see (ğ8.3).
8.2 Structured Bindings

Structured bindings started out as a simple proposal by Herb Sutter, Bjarne Stroustrup, and Gabriel
Dos Reis [Sutter et al. 2015] to simplify notation and eliminate one of the few remaining sources of
uninitialized variables. For example:
template < typename T , typename U >
void print ( vector < pair <T ,U > >& v)
{
for ( auto [x ,y] : v)
cout << '{' << x << ' ' << y << "}\ n ";
}
The names x and y are bound to the first and second elements of the pair. That can be a
significant notational convenience.
C++14 had given us convenient ways of returning multiple values. For example:
tuple <T1 ,T2 ,T3 > f (/*...*/) // nice declaration syntax

{
// ...
return {a ,b ,c }; // nice return syntax
}
I consider tuples somewhat overused in current C++ and prefer specifically defined classes
when the multiple values are not independent, but notationally that makes no difference. However,
C++14 did not offer a convenient way of unpacking such multiple return values to match the
convenient ways of creating them. This led to verbose workarounds, uninitialized variables, or
run-time overhead. For example:
Tuple <T1 ,T2 ,T3 > res = f ();

T1 & alpha = get <0 >( res ); // indirect access through alpha
T2 & val = get <1 >( res );
T3 err_code = get <2 >( res ); // copy
Many experts preferred to use the standard-library function tie() for unpacking tuples:
T1 x;
T2 y;
T3 z;
// ...
tie (x ,y ,z) = f (); // nice call syntax , with existing variables
Assigning to a tie() assigns to the variables used as arguments to the tie(). However, to use
tie, you have to separately declare the variables and name their types to match those of the
members of the object returned by f() (here, T1, T2, and T3). Unfortunately, that leaves the local
variables open to use-before-set errors and initialization-followed-by-assignment overhead. Also,
most programmers don’t know that tie() exists or consider it too odd to use in real code.
Herb Sutter suggested a solution that basically mirrors the return syntax:
auto {x ,y ,z} = f (); // nice call syntax , introducing aliases
This would work for any struct holding three members, not just for tuples. The elimination of
the second-to-last source of uninitialized variables in the Core Guidelines (ğ10.6) was the major
motivation for me. Yes, I liked the notation, but the important issue was the improved approximation
to the ideals of C++.
Not everybody liked the idea and we almost didn’t get to discuss it in time for C++17. The paper
proposing structured bindings [Sutter et al. 2015] was late and Ville Voutilainen was about to close
the EWG at the end of the November 2015 meeting in Kona when I noticed that we had about
45 minutes left before lunch and I thought the group would like to see this proposal. Kona 2015
was when we froze the feature set to be worked on for C++17, so those 45 minutes were critical.
We didn’t even have time to fetch Herb from another group, so I presented. The EWG liked the
proposal; the minutes say Encouragement by acclamation; EWG wants something like this.
Now, the real work started.
In this and later meetings, several people ś notably Chandler Carruth ś pointed out that to meet
the ideals for C++, we needed to extend the ability of break an object into N values to cope with
types that are not tuples or plain structs. For example:
complex < double > z = 2+3 i;

auto {re , im } = sqrt (z ); // sqrt () returns a complex value
The standard-library complex does not expose its representation.

For C++17 we solved this problem by allowing the user to define a set of get functions, e.g.,
get<0> and get<1> effectively pretending the result being a tuple. This works but requires the
user to supply some inelegant boilerplate code. The discussions about potential improvements
continues but no significant simplifications made it into C++20.
There were demands for having this work with functions returning arrays and functions returning
structs with bitfields. Support for that was added, so the resulting design was at least twice as
complicated as the proposed one.
There was a longish debate (spread over several meetings) over whether it should be possible (or
required) to explicitly specify the type of the local variables introduced. For example:
auto { int x , const double * y , string & z} = f (); // not C ++
The arguments for that, most eloquently expressed by Ville Voutilainen, was that without explicit
types, the notation decreased readability, could damage maintainability, and could lead to errors.
These are very similar to the arguments against auto in general and having explicit types brings
their own problems. What if the types don’t match what is being returned? Some said it should be an
error. Some said that conversions to the specified type would be very useful (e.g., char[20] returned
into a string). I pointed out that structured binding was supposed to introduce zero-overhead
aliases and any conversion that implied a change in representation would cause significant overhead.
Also, one purpose of structured binding was to improve notation and requiring explicit types would
lead to code that was more verbose than existing alternatives.
The original proposal used braces ({}) to group the names being introduced:
auto {x ,y ,z} = f (); // nice call syntax , introducing aliases
However, some members, notably Chandler Carruth and David Vandevoorde, feared syntactic
ambiguities and insisted that would be confusing łbecause {} means scopež. So we got the [] syntax:
auto [x ,y ,z] = f (); // call syntax , introducing aliases
This is a minor change, but I think a mistake. It was a last-minute change and led to minor
grammar complications with the attribute notation (e.g., [[fallthrough]] (ğ4.2.10). I don’t buy the
aesthetic or scope arguments, and in 2014 I had presented ideas of adding functional-programming-
style pattern matching to C++ using { ... } for patterns to break out values (ğ8.3). The structured
bindings had been designed to fit into that general scheme.
These were not the only proposals for late changes. Every proposal added or would have added
complexity.
It is dangerous to consider one addition to a language at a time. Last-minute changes are
dangerous unless they fit into a larger scheme. They also easily lead to łbloatž through demands
for łcompleteness.ž In this case of structured bindings, I am not convinced that agreeing to allow
structured bindings to refer to bitfields offer sufficient utility to warrant the increased complexity.
8.3 variant, optional, and any

Alternative types can be represented without run-time overhead by a union. For example:
union U {
int i;
char * p;
};
U u;
// ...
int x = u.i; // OK : iff u holds an integer
char * p = u.p; // OK : iff u holds a pointer
This has been used and misused since the earliest days of C as the fundamental way of different
types łtimesharingž a memory location. There are no compile-time or run-time checks to ensure
that a location is used only as the type it actually holds. Ensuring consistent use of union members
is the programmers’ job, and programmers get it wrong with sickening regularity.
Experienced programmers avoid the problem by encapsulating a union in a class that guarantees
proper use. In particular Boost offered three such types
• optional<T> ś holds a T or nothing
• variant<T,U> ś holds a T or a U
• any ś holds any type
The great utility of these types has been demonstrated in C++ and many other languages.
The committee decided to standardize those three types. Unfortunately, the design of those three
types was discussed separately as if their use cases were disjoint. The possibility of direct language,
as opposed to standard-library, support appears not to have been seriously considered. The result
was three standard-library types that were (like their Boost ancestors) dramatically different. So,
despite the undoubted utility of these types, they are a classic example of design by committee.
Consider:
optional < int > var1 = 7;
variant < int , string > var2 = 7;
any var3 = 7;
auto x1 = * var1 ; // dereference the optional

auto x2 = get < int >( var2 ); // access the variant as a tuple
auto x3 = any_cast < int >( var3 ); // convert the any
To extract the value stored, I have to use one of three incompatible notations. That’s a burden
on programmers. Yes, experienced programmers will get used to it, but there shouldn’t be an
irregularity for people to get used to.
To simplify the use of variants, a visitor mechanism is provided. First we need a helper template
to define overloat an set:
// boilerplate for simple visitation :
template < class ... Ts > struct overloaded : Ts ... { using Ts :: operator ()...; };
template < class ... Ts > overloaded ( Ts ...) -> overloaded < Ts ... >;
That overloaded template really ought to be standard. It is simple only to people comfortable
with variadic templates (ğ4.3.2) and template argument deduction guides (ğ8.1). However, given
overloaded, I can construct a switch on the type of a variant:
using var_t = std :: variant < int , long , double , std :: string >; // a variant type
// boilerplate for simple visitation :

template < class ... Ts > struct overloaded : Ts ... { using Ts :: operator ()...; };
template < class ... Ts > overloaded ( Ts ...) -> overloaded < Ts ... >;
void use ()
{
std :: vector < var_t > vec = {10 , 20 L , 30.40 , " hello "};
for ( auto & var : vec ) {

std :: visit ( overloaded {
[]( auto arg ) { cout << arg << '\n '; }, // handle the integers
[]( double arg ) { cout << " double : " << arg << '\n '; },
[]( const std :: string & arg ) { cout << "\"" << arg << '\"\n '; },
}, var );
}
}
Indisputably, variant and friends address an important problem, but inelegantly. Maybe future
work can reduce the confusing variations in the interfaces to what genuinely needs to differ. In the
meantime, the problem is to get the wider C++ community to use these new types well so as to
eliminate most of the decades’ old problems with unions.
I see these three variants of the idea of a discriminated union as a stop-gap measure. Functional-
programming-style pattern matching is a far more elegant, general, and potentially more efficient
solution to the problems with unions. At the November 2014 meeting in the University of Illinois
at Urbana-Champain, I gave a presentation on design issues related to pattern matching [Solodkyy
et al. 2014] partly based on research done with Yuriy Solodkyy and Gabriel Dos Reis in Texas A&M
University [Solodkyy et al. 2013]. We had a library implementation that performed comparably to
functional programming languages despite lack of integration into the compiler. That library could
cope with both closed sets of alternatives (algebraic types) and open sets (class hierarchies). One
aim was to eliminate use of the visitor pattern [Gamma et al. 1994]. However, we did not have an
acceptable syntax. My purpose for the presentation was to raise interest and set longer-term goals.
There was significant interest and after the completion of C++17 work started [Murzin et al. 2019,
2020]. Maybe pattern matching can make it into C++23 (ğ11.5).
8.4 Concurrency
In C++17, the use of locking became significantly easier by the addition of
• scoped_lock ś to acquire an arbitrary number of locks without the possibility of deadlock
• shared_mutex and shared_lock ś to implement reader-writer locks
For example, we can acquire multiple locks without fear of deadlock:
void f ()
{
scoped_lock lck { mutex1 , mutex2 , mutex3 }; // acquire all three locks
// ... manipulate shared data ...
} // implicitly release all mutexes
C++11 and C++14 failed to give us reader-writer locks. That was obviously a serious omission,
caused by the pressure of proposals and the length of time needed to process proposals. C++17
remedied that by adding shared_mutex:
shared_mutex mx ; // a mutex that can be shared
void reader ()
{
shared_lock lck { mx }; // willing to share access with other readers
// ... read ...
}
void writer ()
{
unique_lock lck { mx }; // a writer needs exclusive ( unique ) access
// ... write ...
}
Many readers can łsharež the lock (i.e., enter the critical section at the same time) whereas the
writer requires exclusive access.
I consider these good examples of the łmake simple things simplež philosophy. Sometimes, I
wonder ś with many C++ programmers ś łwhat took them so long?ž
Note how the notation was simplified using template argument deduction from constructor
arguments (ğ8.1).
8.5 Parallel STL

In the longer term, the use of parallel algorithms will be of prime importance, because from a
user’s point of view, there is nothing simpler than just saying łplease do this algorithm.ž From an
implementer’s perspective having a specific interface and no serial constraint on the algorithm
is an opportunity. C++17 offers only a small start, but that’s far better than no start because it
sets a direction. Unsurprisingly, there was some opposition in the committee, mostly from people
wanting more complex interfaces for expert users. Some people expressed serious doubt that such
a simple solution would be viable and argued for delays.
The basic idea is to provide an extra argument to each standard-library algorithm, allowing the
user to request vectorization and/or multithreading. For example:
sort ( par_unseq , begin (v), end (v )); // consider parallelizing and vectorizing
This was done only for the STL algorithms, so the important find_any and find_all algorithms
are missing. In the future we will see algorithms specifically designed for parallel use. This is
already happening for C++20.
Another weakness is that there was still no standard way of cancelling a thread. For example,
having found an item in a search, a thread cannot stop other parallel searches. This is a result of the
POSIX intervention against all forms of cancellation (ğ4.1.2). C++20 offers cooperative cancellation
(ğ9.4).
The parallel algorithms the C++17 support vectorization. This is important because improved
support for SIMD is one of the few areas where hardware still (post-2017) delivers dramatic growth
in single-thread performance.
In C++20, we can also use the ranges library (ğ6.3) to (finally) avoid the explicit mention of a
container’s sequence of elements and just write:
sort (v );
Unfortunately, the range versions of the parallel algorithms didn’t get finished in time for C++20,
so we’ll have to wait for C++23 before writing:
sort ( par_unseq ,v ); // sort v using parallelism and vectorization
Alternatively, we can write our own adapter:

concept execution policy = std :: is_execution_policy <T >:: value ;
void sort ( execution_policy auto && ex , std :: random_access_range auto & r)

{
sort (ex , begin (r), end (r )); // sort with the execution policy ex
}
After all, the standard library is extensible.
8.6 File System

In 2002, Beman Dawes wrote the Boost file system library that became one of the most popular
Boost libraries [Boost 1998ś2020]. In 2014, the Boost file system library [Dawes 2002ś2014] was
(with modifications) made into a TS [Dawes 2014, 2015] and after further modifications moved
into the standard for C++17. Dealing with file names and file systems is tricky because it involves
concurrency, many natural languages, and differences among operating systems. It is nice finally
to be able to manipulate directories (folders) in a standard manner (as has been done using Boost
for 15 years). The key type offered is that of a path that abstracts away notions of character sets
and filesystem notations. For example:
void do_something ( const string & name )

{
path p { name }; // name could be in Russian or Arabic
// name could use Linux or Windows file notation
try {
if ( exists (p )) {
if ( is_regular_file (p ))
cout << p << " regular file , size : " << file_size (p) << '\n ';
else if ( is_directory (p )) {
cout << p << " directory , containing :\ n ";
for ( auto & x : directory_iterator (p ))

cout << " " << x. path () << '\n ';
}
else
cout << p << " exists \n ";
}
else
cout << p << " does not exist \n ";
}
catch ( const filesystem_error & ex ) {
cerr << ex . what () << '\n ';
throw ;
}
// ... use p ...
}
Catching the exception protects against rare errors, such as someone removing a file after the
exists(p) check and before the more detailed inquiries. The file system’s interfaces offer support
for handling both rare (exceptional) and common (expected) errors (ğ7.2).
8.7 Explicit Tests in Conditions

I consider łmany small proposalsž a danger, even if each can help someone. Consider the ability to
add an explicit test to a condition [Köppe 2016b]:
if ( auto p = f(y ); p ->m >0) {
// ...
}
The p->m>0 is an explicit test, so the meaning is

{
auto p = f(y );
if (p ->m >0) {
// ...
}
}
It is a generalization of the C++98 combined declaration and test (ğ2.2.1):

if ( auto pd = dynamic_cast < Derived * >( pb )) { // true if pb points to a Derived
// ...
}
The question is whether that generalization is sufficiently obvious and useful to be worthwhile.
My answer was łno.ž However, this is a (not all that uncommon) example of me being voted down.
My opinion is that the explicit test is best expressed within the if-statement. There, it is harder
to overlook and being conventional has its benefits, especially for people who don’t spend all of
their time writing C++. On the other hand, the explicit test seems to be popular with people who
design their code so that the result of every function needs to be tested for errors - a style I strongly
dislike (ğ7.5).
There are people who aggressively rewrite code to use novel features. I heard of several cases of
people who saw this:
if ( auto p = f(y )) {
if (p ->m >2) {
// ...
}
// ...
}
and immediately re-wrote it to this:

if ( auto p = f(y ); p ->m >2) {
// ...
}
Claiming elegance and shorter code. Naturally, it crashed when p==nullptr, which the original
code didn’t. Thus, whatever benefits we might get, we introduced a new opportunity for errors and
confusion.
For generality, the explicit test was also allowed in switch and while conditions. For C++20,
this facility was extended to include initializations in range-for statements [Köppe 2017c].
8.8 Proposals That Didn’t Make C++17

In addition to concepts (ğ6.3.8), a few proposals that I consider important didn’t make C++17. The
story of C++ would not be complete without a mention of them.
• ğ6.3.8: concepts (C++20)
• ğ8.8.1: networking
• ğ8.8.2: operator dot
• ğ8.8.3: uniform call syntax
• ğ8.8.4: default comparison
• ğ9.3.2: coroutines (C++20)
Static reflection was processed in a study group (ğ3) and wasn’t scheduled for C++17, but it was
important work started in this period.
8.8.1 Networking. In 2003, Christopher M. Kohlhoff started developing a library, called asio, to
support networking [Kohlhoff 2018]:
łAsio is a cross-platform C++ library for network and low-level I/O programming that
provides developers with a consistent asynchronous model using a modern C++ approachž
In 2005, it became part of Boost [Kohlhoff 2005] and in 2006 it was proposed for the standard
[Kohlhoff 2006]. In 2018 it became a TS [Wakely 2018]. Despite 13 years of heavy production use, it
didn’t make it into C++17. Worse, work to get the networking library into C++20 also stalled. This
implies that, after 15 years of production use of asio we still have to wait for at least until 2023
for it to become standard. The reason for the delay is that there are still serious discussions about
how best to generalize the way concurrency is handled in asio and elsewhere. A proposal for that,
called łexecutorsž, has broad support and some hoped it would make it into C++20 [Hoberock et al.
2019, 2018]. I consider the lack of executors and networking in C++20 an example of łthe best is
the enemy of the good.ž
8.8.2 Operator Dot. The very first proposal for an extension to C++ at the start of the standards
process was by Jim Adcock in 1990 to allow the overloading of operator dot (.) [Adcock 1990].
We have been able to overload operator arrow (->) since 1984 and it is heavily used to implement
łsmart pointersž (e.g., shared_ptr). People wanted (and still want) operator dot to implement
smart references (proxies). Basically, people asked for a way of making x.f() mean x.operator.().f()
so that operator.() could control access to members. However, discussions tended to deadlock
over the issue of whether an overloaded operator should be applied to an implicit use of dot. For
example: ++x for a user-defined type is interpreted as x.operator++(). Now, if the user-define
type has operator.() defined, should ++x mean x.operator.().operator++()? Andrew Koenig and
Bjarne Stroustrup tried to resolve that in 1991 [Koenig and Stroustrup 1991a] just to be vehemently
opposed by the original proposer, Jim Adcock. Gary Powell, Doug Gregor, and Jaakko Järvi tried
again in 2004 for C++0x [Powell et al. 2004], but again the committee deadlocked. Finally, in 2014
Bjarne Stroustrup and Gabriel Dos Reis tried again for C++17 with what I consider a more complete
and better reasoned approach [Stroustrup and Dos Reis 2014]. For example:
template < class X >
class Ref { // smart reference ( with ownership )
public :
explicit Ref ( int a) :p{ new X{a }} {}
X& operator .() { /* maybe some code here */ return *p; }
~ Ref () { delete p; }
void rebind (X* pp ) { delete p; p= pp ; }
// ...
private :
X* p;
};
Ref <X > x {99};

x.f (); // means (x. operator .()). f () means (* x.p ). f ()
x = X {9}; // means x. operator .() = X {9} means (* x.p )= X {9}
x. rebind ( new X {77}); // means that x holds and own that new X
The basic idea is that operations defined on łthe handlež (here Ref) are applied to the handle (e.g.,
the constructor, destructor, operator.(), and rebind()) whereas operations that are not defined on
łthe handlež are applied to łthe value,ž that is, to the result of operator.().
After much work [Stroustrup and Dos Reis 2016], that also failed. The reasons for the failure of
the 2014 proposal are interesting. There was of course the usual wording problems and obscure
łdark cornersž of the design, but I think the proposal would have succeeded had the committee not
become so excited by the idea of smart references that mission creep set in and alternative proposals
were made by Mathias Gaunard and Dietmar Kühl [Gaunard and Kühl 2015] and Hubert Tong
and Faisal Vali [Tong and Vali 2016], respectively. The former required a heavy dose of template
metaprogramming from all who wanted to define an operator.() while the latter was basically
object-oriented, introducing a new form of inheritance and implicit conversions.
Should the action of an operator.() depend on the member to be accessed or should operator.()
be a unary operator depending only on the object it was applied to (just like operator->())? The
former was central to Gaunard and Kühl’s proposal. Bjarne Stroustrup and Gabriel Dos Reis had
also considered making operator.() binary but concluded that it was far too complex and that
matching operator arrow (->) was important.
In the end, the initial proposal wasn’t really rejected (it was approved by EWG, but never brought
to a full committee vote), but further progress stalled for lack of new input to gain consensus among
the competing proposals. Also, the original proposers (Bjarne Stroustrup and Gabriel Dos Reis)
got distracted by even more important proposals, such as concepts (ğ6) and modules (ğ9.3.1), and
their łday jobs.ž I consider operator dot a prime example of the members of the committee lacking
a shared view of what C++ is and what it should become (ğ9.1). Thirty years, six proposals, many
discussions, much design and implementation work, and we have nothing.
8.8.3 Uniform Call Syntax. The very first discussion of concepts in 2003, mentioned the need for a
uniform syntax for function calls [Stroustrup and Dos Reis 2003b]. That is, x.f(y) and f(x,y) should
ideally mean the same. The point is that when writing a generic library, you have to decide whether
to call operations on arguments using the object-oriented or the functional notation (x.f(y) or
f(x,y)). As a user, you have to adjust to the choice made by the library designer. Different libraries
and different organizations differ on this. For operators, such as + and *, uniform resolution has
always been the rule; that is, a use (e.g., x+y) would find both member functions and free-standing
functions. In the standard library, we have a plague of pairs of functions to deal with this dilemma
(e.g., to make both begin(x) and x.begin() work).
I should have settled that issue in 1985 or so before the committee could tie itself into knots over
details and potential problems, but I failed to generalize from the operator case.
In 2014, Herb Sutter and I independently proposed a łuniform function call syntaxž [Stroustrup
2014a; Sutter 2014]. The proposals were incompatible, of course, but we immediately solved that by
merging them into a joint proposal [Stroustrup and Sutter 2015].
Herb was partly motivated by a wish to support autocompletion in IDEs and was leaning
toward the łobject-orientedž notation (e.g., x.f(y)) whereas I was primarily motivated by generic
programming concerns and leaned toward the traditional mathematical notation (e.g., f(x,y)).
The first serious objection was, as ever, compatibility; that is, that we might break existing code.
The original proposal could indeed break some code by preferring a better match or making a call
ambiguous, but we argued that it would be worthwhile and often beneficial. We lost that argument
and a revised version worked by the principle that x.f(y) would look into the class of x first and
only if no f was found there would it consider f(x,y). Similarly, f(x,y) would look into the class of
x only if no free-standing function was called. That would not make f(x,y) and x.f(y) completely
equivalent but obviously it would break no existing code.
This looked most promising but was met with a howl of outrage: it would mean the end of
stable interfaces! The argument, primarily presented by people from Google, was that no interface
that depended on overload resolution could be stable because adding a function could change the
meaning of existing code. That is of course true. Consider:
void print ( int );

void print ( double );
print ('a '); // prints the integer value of 'a '
void print ( char ); // add a print () to change the overload set
print ('a '); // print the character 'a '
My answer to that argument was that just about any program can have its meaning changed
by quite a few different kinds of added declarations. Also, a common use of overloading is to add
functions that provide semantically better resolutions (often, to fix bugs). We have always strongly
recommended against adding overloads that (like print(char) above) changes the semantics of
calls of an overload set in the middle of a program. In other words, this definition of łstablež is
unrealistic. I (and others) pointed out that the problem already existed for class members. The
answer was basically that the set of class members is closed so that version of the problem is
manageable. I observed that by using namespaces, the set of free-standing functions associated
with a class can be identified much as members are [Stroustrup 2015b].
By then, much controversy and confusion had erupted and novel proposals started to appear to
compete with the ones under discussion. The UK delegation suggested C# style extension methods
[Coe and Orr 2015] and several people, notably John Spicer insisted that if we needed a unified
function call notation, it should be a new notation separate from the two we already have. I still
fail to see how adding a third notation (e.g., .f(x,y) as suggested) would unify anything. It would
become yet another example of the N+1 problem (ğ4.2.5).
After the defeat of the proposal, I was asked to reexamine the issue once we had modules (ğ9.3.1).
Then, the reach of the name lookup for a free-standing function could be limited to the module of
the class of its first argument. That might resurrect the unified function call proposal, but I don’t
see how that would address the (IMO vastly overstated) concerns about stability of interfaces.
Again, the lack of a shared view of C++’s role and future blocked progress (ğ9.1).
In retrospect, I don’t think that the object-oriented notation (e.g., x.f(y)) should ever have
been introduced. The traditional mathematical notation f(x,y) is sufficient. As a side benefit, the
mathematical notation would naturally have given us multi-methods, thereby saving us from the
visitor pattern workaround [Solodkyy et al. 2012].
8.8.4 Default Comparisons. Like C, C++ doesn’t offer default comparisons for data structures. For
example:
struct S {
char a;
int b;
};
S s1 = { 'a ' ,1};

S s2 = { 'a ' ,1};
void text ()
{
S s3 = s1 ; // OK , initialization
s2 = s1 ; // OK , assignment
if ( s1 == s2 ) { /* ... */} // error : == undefined for Ss

}
The reason is that, given the usual memory layout of S, there will be łunused bits,ž in the memory
holding an S so that a naive implementation of s1==s2 comparing the bits of the words holding s1
and s2 might yield false. Had it not been for those łunused bits,ž C would have had at least default
equality. I discussed this with Dennis Ritchie back in the early 1980s, but we were both too busy to
do anything about it at the time. This problem doesn’t occur for copying (e.g., s1=s2), where the
naive and traditional solution simply copies all the bits.
Allowing assignment but not comparison because of the efficiency of simple implementations
was appropriate for the 1970s, but odd for the 2010s. Our optimizers could easily handle that, and I ś
like many others ś were tired of explaining why such comparisons weren’t provided. In particular,
many STL algorithms require == or < and therefore don’t work for simple data structures without
the user explicitly defining operator==() and/or operator<() for them.
In 2014, Oleg Smolsky [Smolsky 2014] proposed a simpler way of defining comparison operators
struct Thing {
int a , b , c;
std :: string d;
bool operator ==( const Thing &) const = default ;

bool operator <( const Thing &) const = default ;
bool operator !=( const Thing &) const = default ;
bool operator >=( const Thing &) const = default ;

bool operator >( const Thing &) const = default ;
bool operator <=( const Thing &) const = default ;
};
This addresses the right problem, but it is verbose (six long lines to say łI want the default
operatorsž) and definitely second best to getting comparison operators by default. There were
other technical problems (e.g., łbut that solution is intrusive: if I can’t modify a class, I can’t add
comparisonsž) but the race was now on to better support operators for C++17.
I wrote a paper discussing the problem [Stroustrup 2014c] and proposed to supply the comparisons
by default for simple classes [Stroustrup 2014b]. It turned out that defining what it meant for a
class to be simple in this context was hard and Jens Maurer discovered some nasty scope problems
with user-defined comparison operators in the presence of default operators (e.g., łwhat does it
mean if we use a default == and then later define an operator==() in a different scope?).
Many more papers were written by Oleg, me, and others but the proposals stalled. People
started to heap more requirements on the proposal. For example, the performance of default
comparisons should equal three-way comparisons for naive uses. Lawrence Crowl wrote an analysis
of comparisons in general [Crowl 2015b] touching upon issues such as handling total, weak, and
partial orders. The general opinion of the EWG was that Lawrence’s analysis was great, but he’d
need a time machine to get the mechanisms in place in C++.
Finally, in 2017, Herb Sutter made a proposal (partially based on Lawrence Crowl’s work) based
on a three-way comparison operator <=> (as found in a variety of languages) from which the usual
operators could be generated [Sutter 2017a]. It didn’t give us the default operators, but at least it
offered a one-line formula for defining it:
struct S {
char a;
int b;
friend std :: strong_order operator <= >(S ,S) = default ;
};
S s1 = { 'a ' ,1};

S s2 = { 'a ' ,1};
bool b0 = s1 == s2 ; // true
int b1 = s1 <= > s2 ; // 0
bool b2 = s1 < s2 ; // false
The solution above is recommended by Herb Sutter as causing the fewest problems (e.g., with
overloading and scope), but it is intrusive. I can’t use it for a class that I cannot modify. In that case,
a non-member <=> can be defined:
struct S {
char a;
int b;
};
std :: strong_order operator <= >(S ,S) = default ;
The <=> proposal contained an option for implicitly defining <=> for simple classes, but unsur-
prisingly people who consider it safer to be explicit about everything voted that down.
So, instead of a facility that made trivial examples work as expected for novices, we got an
elaborate facility that allows experts to carefully craft subtle comparisons.
The <=> had an easier pass through the committee than any other recent proposal I can think of.
This was the case even though there were no usable implementations and the proposal had strong
implications for the standard library. Predictably, this led to many surprises (ğ9.3.4) including the
kind of lookup problems that had help scuttle the earlier proposals for ==. My guess is that the
comparison operator discussions had convinced many that something was worth doing and the
<=> proposal addressed a variety of issues and fitted with what was familiar from other languages.
Sometime in the future, I will most likely again propose that == and <=> be defined by default
for simple classes. The novice and casual users of C++ deserve that simplicity.
Being proposed in 2017, <=> missed C++17, but after much further work, it is in C++20 (ğ9.3.4).
9 C++20: A STRUGGLE FOR DIRECTION

Design by a 350+ member committee is unlikely to produce a coherent result. People from vastly
different backgrounds (including different education) and differing pressures from their łday jobsž
naturally differ on direction, priorities, and committee procedures. At a first approximation, for
every proposal there are a dozen or so members who strongly object to something. Given that
WG21 wants 80% or 90% in favor to declare consensus, it’s a amazing that C++ has succeeded so far.
9.1 Design Principles

What does C++ want to be when it grows up? In other words, does WG21 have a clear view of
what it is trying to do? I don’t think so. Individual members have ideas, but none that are generally
accepted and sufficiently concrete to guide individual discussions and decisions.
The ISO C++ standards committee doesn’t have a generally accepted set of design criteria or set
of criteria for accepting a feature. This is not for lack of trying. I have repeatedly and consistently
articulated design criteria:
• The łrules of thumbž from The Design and Evolution of C++ [Stroustrup 1994] (ğ2.1) including
RAII (ğ2.2.1), object-oriented programming, generic programming, and static type safety.
• Make simple things simple! (ğ4.2) leading to the Onion principle (ğ4.2).
• Direct map to hardware and Zero-overhead abstraction (ğ1) (ğ11.2).
• Grow C++ based on feedback to solve real-world problems (ğ11.2).
• Stay stable and compatible [Koenig and Stroustrup 1991b; Stroustrup 1994].
• Deal directly with hardware, powerful compositional abstraction, and a minimal run-time
system (the retrospective in my HOPL3 paper [Stroustrup 2007]).
The problem is that people find it too hard to agree on interpretation and too easy to ignore
what they don’t like. This allows for fundamental disagreements about what is important to fester.
People make design decisions based on what they understand from their education and their day
jobs. One problem is the variety of such backgrounds combined with the uneven coverage of the
wide field of C++ use in the standards committee (ğ3.3). Many are simply too certain of their
opinions [Stroustrup 2019b]. It is really hard to know exactly what’s a fad and what will help the
C++ community in the long term. Often, the first solution found is not the best.
It is very easy to get lost in details and lose sight of the bigger picture. It is very easy to focus
on current problems and forget about the long term (decades). Conversely, members can get so
focused on general principles and the distant future that they neglect urgent practical problems.
In 2017, at the urging of a group of heads of national standards body delegations [van Winkel
et al. 2017] to become serious about direction, the WG21 established the Direction Group (often
called the DG) to try to address issues of design aims and direction (ğ3.2). The DG issued its first
extensive statement in 2018 [Dawes et al. 2018] emphasizing adherence to articulated principles,
consistency, and encouraging processes to ensure those. For example:
łWe fundamentally need:
• Stability: Useful code łlivesž for decades.
• Evolution: The world changes and C++ must change to face new challenges.
There is an inherent tension here.ž
The DG emphasizes the need for coherence across the whole standard:
ł Today, some of the most powerful design techniques combine aspects of traditional
object-oriented programming, aspects of generic programming, aspects of functional
programming, and some traditional imperative techniques. Such combinations, rather
than theoretical purity, are the ideal.
• Provide features that are coherent in style (syntax and semantics) and style of use.
This applies to libraries, to language features, and combinations of the two.ž
And of course, static types:
ł C++ relies critically on static type safety for expressiveness, performance, and safety. The
ideal is
• Complete type-safety and resource-safety (no memory corruption and no leaks)
This is achievable without added overhead, in particular without adding a garbage collector,
and without restricting expressibility. ž
Both the NB head request [van Winkel et al. 2017] and the DG document [Dawes et al. 2018]
emphasized the need for committee members to know the history of C++ to ensure a degree of
continuity. An ahistorical group cannot maintain a coherent view of what they are designing. Thus,
the HOPL papers [Stroustrup 1993, 2007] and The Design and Evolution of C++ [Stroustrup 1994]
play an essential role.
Traditionally, in line with the ISO charter for WG21, the work on the evolution of C++ has focused
exclusively on language and library issues. However, a developer is not concerned exclusively with
the language: a program is the product of a tool chain (ğ1). Astoundingly, C++ does not have a
standard for dynamically linked libraries or a standard build system. The tools study group SG15
was founded in 2018 to try to grapple with the diverse issues related to tooling (ğ3.2).
9.2 My C++17 List

As part of my continuing effort to encourage the committee to focus on significant improvements
ś rather than on just what can be easily done and easily agreed upon ś I made a list of what I
considered important and feasible for C++17 together with its rationale:
• Concepts ś they allow us to precisely specify our generic programs and address the most vocal
complaints about the quality of error messages.
• Modules ś provided they can demonstrate significant isolation from macros and a significant
improvement in compile times.
• Ranges and other key STL components using concepts ś to improve error messages for mainstream
users and improve the precision of the library specification (łSTL2ž).
• Uniform call syntax ś to simplify the specification and use of template libraries.
• Coroutines ś should be very fast and simple.
• Networking support ś based on the asio library as described in the TS.
• Contracts ś not necessarily used in the C++17 library specification.
• SIMD vector and parallel algorithms.
• Library vocabulary types, such as optional, variant, string_view, and array_view.
• A łmagic typež providing arrays on the stack (stack_array) with support for reasonable safe
and convenient use.
In April 2015, I presented that list at an evening session at the WG21 meeting in Lenexa, Kansas,
to a sympathetic audience. However, few seemed sufficiently motivated to change the focus of their
work to align with that list. The list łleakedž and led to confused discussions on the web, so I had
to write it up [Stroustrup 2015a].
In a united committee, everything on that list could be ready for C++17. In reality, I thought
it would be feasible to complete about half of the proposals on that list if we focused on them.
However, I was far too optimistic. We were only able to get consensus on the standard-library
vocabulary types. The array_view was renamed span and is part of C++20 (ğ9.3.8).
Fortunately, most items on that list are in C++20. The exceptions are
• the networking library (ğ8.8.1) ś now a TS [Wakely 2018]
• contracts (ğ9.6.1) ś almost made C++20
• uniform function call (ğ8.8.3)
• the SIMD vector ś now in a TS [Hoberock 2019]
• the stack_array
This list led to scheduling debates. As the 2016 defeat of the concepts proposal (ğ6.3.8) seemed
inevitable, I was asked ś in full committee ś whether I would propose a one-or-two-year delay to
get concepts in, making the standard C++18 or C++19. I declined because I consider a predictable
release cycle more important to the community than any individual improvement. Also, there was
no guarantee that a consensus would emerge and one schedule slip would most likely cause further
slips. If one proposal is deemed worth delaying for, it will be argued that others are equally worth
waiting for. Such logic caused C++0x to become C++11, even as some had hoped for C++06.
9.3 C++20 Features

WG21 set the deadline for new proposals for C++20 to November 2018 and declared a łfeature
freezež after the February 2019 meeting. In February 2020, at a meeting in Prague in the Czech
Republic, the technical vote was 79-0 with one abstention [Smith 2020]. All 15 national body heads
of delegation voted for. The official standard will be issued by the ISO in late 2020. The C++20
features are:
• ğ6.4: Concepts ś specification of requirements for generic code
• ğ9.3.1: Modules ś support for modularity for code hygiene and improved compile times
• ğ9.3.2: Coroutines ś stackless coroutines
• ğ9.3.3: Compile-time computation support
• ğ9.3.4: <=> ś a three-way comparison operator
• ğ9.3.5: Ranges ś a library flexible range abstractions
• ğ9.3.6: Date ś a library providing date types, calendar, and time zones
• ğ9.3.8: Span ś a library providing efficient and safe access to arrays
• ğ9.3.7: Format ś a library providing type-safe printf-like output
• ğ9.4: Concurrency improvements ś such as scoped threads and stop tokens
• ğ9.5: Many minor features ś such as C99-style designated initializers and string literals as
template arguments
The following are not ready for C++20 but may become major features of C++23:
• ğ8.8.1: Networking ś a networking library (sockets, etc.)
• ğ9.6.2: Static reflection ś facilities for generating code based on the surrounding program
• Pattern Matching ś selecting code to be executed based on types and object values [Murzin
et al. 2019]
C++20 offers a set of features reflecting long-standing aims for C++ and addresses some funda-
mental problems. For example, modules and concepts are mentioned in The Design and Evolution of
C++ [Stroustrup 1994] from 1994 and coroutines were part of łC with Classesž and C++ throughout
the 1980s. C++20 will have as major an impact on C++ as C++11 had.
Unfortunately, C++20 doesn’t have standard library support for modules and coroutines. That
could become a significant problem, but there wasn’t time to get that ready in time to ship on time.
C++23 should provide that (ğ4.1.3).
9.3.1 Modules. There is an obvious need for improved modularity in C++ programs. From C, C++
inherited the #include mechanism that relies on textually including C++ source from łheader filesž
containing textual definitions of interfaces. A popular header can be #included hundreds of times
in various separately compiled sections of a large program. The fundamental problems are:
• Lack of code hygiene: the code in one header can affect the meaning of the code in another
#included in the same translation unit so that #includes are not order independent. Macros
are a major problem here, though not the only one.
• Separate-compilation inconsistencies: declarations of the same entity in two translation units
can be inconsistent, but not all such errors are caught by the compiler or linker.
• Excessive compile times: Compiling an interface from source text is relatively slow. Compiling
an interface from source text repeatedly is very slow.
This has been known from łthe dawn of timež (e.g., see The Design and Evolution of C++ [Strous-
trup 1994] Chapter 18), but the problems have been steadily growing over the years as more and
more information is put into header files (inline functions, constexpr functions, and especially
templates). In the early days of C++, typically 10% of the text came from headers, but currently it is
more like 90% or even 99%. Consider:
# include < iostream >
int main ()
{
std :: cout << " Hello , World \n ";
}
This canonical program is 70 characters, but after the #include it yields 419,909 characters for
the compiler to digest. Despite the impressive processing speeds of modern C++ compilers, the
modularity problem has become urgent.
Encouraged by the committee (and supported by me) David Vandevoorde produced a series of
module designs in the 2000s [Vandevoorde 2007, 2012], but progress was very slow. Finishing C++0x,
rather than making progress on modules, was the priority of the committee. David mostly struggled
on his own getting little more than moral support. In 2012, Doug Gregor presented a completely
different module system design from Apple [Gregor 2012]. That design had been done for C and
Objective C in the Clang compiler infrastructure [Team 2014] and relied on extra-linguistic file
mapping directives, rather than C++ language constructs. There was also an emphasis on keeping
header files unmodified.
In 2014 a group of people from Microsoft led by Gabriel Dos Reis presented a proposal based on
their work [Dos Reis et al. 2014]. It was closer in spirit to David Vandevoorde’s designs than the
Clang/Apple proposal and to a large degree based on work on an optimal graph representation
of C++ source code done by Gabriel Dos Reis and Bjarne Stroustrup at Texas A&M University
(published and made open-source in 2007 [Dos Reis 2009; Dos Reis and Stroustrup 2009, 2011]).
This set the scene for major progress with modules, but also for a series of clashes between the
Apple/Google/Clang approach (and implementation) and the Microsoft approach (and implementa-
tion).
A study group was created for modules and after 3 years, it produced a TS primarily based on
Gabriel Dos Reis’ design [Dos Reis 2018].
In 2017 and again in 2018, suggestions to move the Modules TS into the standard for C++20 was
blocked by proposals for different designs from Google [Smith 2018a,b]. A major bone of contention
was that in Gabriel Dos Reis’ design macros could not be exported. The Google people argued
that this was a fatal flaw, whereas Gabriel Dos Reis (and I) considered it essential for modularity
[Stroustrup 2018c]:
łWhat do I mean by modularity? Order independence: import X; import Y; should be

the same as import Y; import X; In other words, nothing can implicitly łleakž from one
module into another. That’s one key problem with #include files. Just about anything in
an #include can affect any subsequent #include.ž
I consider order independence key to both łcode hygienež and performance. By adhering to this
Gabriel Dos Reis’s implementation of modules achieved on the order of 10-times compile-time
improvements over header files ś even when pre-compiled headers were used for the old-style
compilations. Approaches that catered to traditional headers and conventional uses of macros
struggle to match that because of the need to keep module units in a form that allows macro
substitution (łtoken soupž) as opposed to graphs of C++ logical entities.
A series of compromises was crafted to converge on a solution with wide acceptance. The key
people in that multi-year effort were Richard Smith (Google) and Gabriel Dos Reis (Microsoft) with
GCC’s module implementer, Nathan Sidwell (Facebook), and others contributing [Dos Reis and
Smith 2018a,b; Smith and Dos Reis 2018]. From mid-2018, most discussions focused on technical
details that needed precise specification to ensure portability among implementations [Sidwell
2018; Sidwell and Herring 2019].
Consider a simple example of C++20 modules:
export module map_printer ; // we are defining a module
import iostream ; // use iostreams

import containers ; // use my containers
using namespace std ;
export // make print_map () available to users of map_printer

template < Sequence S >
requires Printable < Key_type <S >> && Printable < Value_type <S >>
void print_map ( const S& m) {
for ( const auto & [ key , val ] : m) // break out key and value
cout << key << " -> " << val << '\n ';
}
This code fragment defines a module map_printer that offers the function print_map as its
user interface and implements it using facilities imported from modules iostream and containers.
To emphasize the difference from older C++ styles, I use concepts (ğ6) and structured bindings
(ğ8.2).
Key ideas:
• An export directive makes an entity available for import into another module.
• An import directive makes exported entities from another module accessible for use.
• An imported entity is not implicitly exported.
• An import doesn’t add entities to a context; it only makes entities accessible (thus, an unused
import is essentially cost free).
The two last points differ from #includes and are essential for modularity and compile-time
performance.
That simple example is purely module based; this is the ideal. However, there may be half a trillion
lines of C++ deployed and header files and #includes will not go away any day soon, possibly not
for decades. Several individuals and organizations pointed to the need for mechanisms to allow for
transition, for coexistence of header files and modules in programs, and for libraries to offer both
header file and module interfaces to users with code bases of different maturity. Remember that at
any given time, there are users relying on 10-year old compilers.
Consider implementing the map_printer under the constraint that you cannot modify the
iostream and container headers:
export module map_printer ; // we are defining a module
import < iostream > // use iostream header

import " containers " // use my containers header
using namespace std ;
export // make print_map () available to users of map_printer

template < Sequence S >
requires Printable < Key_type <S >> && Printable < Value_type <S >>
void print_map ( const S& m) {
for ( const auto & [ key , val ] : m) // break out key and value
cout << key << " -> " << val << '\n ';
}
An import directive that names a header file works almost exactly as an #include would ś
macros, implementation details, and recursively #included headers. However, the compiler ensures
that imported łlegacy headersž do not have mutual dependencies. That is, imports of headers is
order independent, thus providing some, but not nearly all of the benefits of modularity. For example,
importing individual headers, such as import<iostream>, leaves the programmer with the task of
deciding which to import, slows down compilation by having unnecessarily many interactions with
the file system, and limits pre-compilation of standard-library components from different headers.
Personally, I’d like to see modules with a coarser granularity, e.g., a standard import std directive
for making the complete standard library available. However, a more ambitious refactoring of the
standard library [Clow et al. 2018] had to be postponed until C++23 (ğ11.5).
Facilities like import of headers were an essential part of the Google/Clang proposals. One
reason for that is that there are libraries for which the primary interface is a set of macros.
Late in the design/implementation/standardization effort objections focused on the modules
likely impact on build systems. Current build systems for C and C++ are heavily optimized for
handling header files. Decades of work has been expended on optimizing that and several people
associated with traditional build systems expressed doubts that modules could be fitted in without
(unaffordable) major redesigns and/or that builds using modules would not allow for parallel
compilation (because an importing module depends on the imported module being previously
compiled) [Bindels et al. 2018; Lopes et al. 2019; Rivera 2019a]. Fortunately, early impressions had
been overly pessimistic [Rivera 2019b], the build2 system had been modified to handle modules,
Microsoft and Google reported that their build systems showed good results with modules, and
finally Nathan Sidwell reported that he had modified GNU build to handle modules in only two
weeks of his spare time [Sidwell 2019]. A final presentation of these experiences and a joint paper
of the key module implementers (Gabriel Dos Reis, Nathan Sidwell, Richard Smith, and David
Vandevoorde) swayed almost all nay-sayers [Dos Reis et al. 2019].
In February 2019, modules were voted into C++20 by a 46-6 majority including all implementers
[Smith 2019]. By then, the major C++ implementations were already close approximations to the
C++20 standard. Modules promise to be the most important single improvement offered by C++20.
9.3.2 Coroutines. Coroutines offer a model of cooperative multi-tasking that can be far more
efficient than use of threads or processes. Coroutines were an essential part of early C++. Without
the task library offering coroutines, C++ would have been stillborn, but for a variety of reasons
coroutines didn’t make it into C++98 (ğ1.1).
The history of what became C++20 coroutines starts with a proposal from Niklas Gustafsson
(Microsoft) for łresumable functionsž [Gustafsson 2012]. The primary aim was to support asynchro-
nous I/O; łserver applications that handle many thousands or millions of clientsž [Kohlhoff 2013].
It was equivalent to the async/await facility being introduced into C# at the time (in release 6.0 in
2015). Similar facilities have been introduced into Python, JavaScript, and other languages. Niklas’
proposal triggered a competing proposal based on Boost::coroutine from Oliver Kowalke and Nat
Goodspeed [Kowalke and Goodspeed 2013] and lots of interest. The await design was stackless,
asymmetric, and language-based whereas the Boost-derived one used stacks, had symmetric control
primitives, and was library-based. A stackless coroutine is one that can only be suspended in its
own body and not in a function called from it. That way, a suspension involves only saving a
single stackframe (łthe coroutine statež) and not a whole stack. For performance, this is a massive
advantage.
The design space for coroutines is huge, so consensus was hard to achieve. Many in the committee
(including me) hoped for a synthesis that gave the best of both approaches, so a group of interested
members did an analysis of the alternatives [Goodspeed 2014]. The conclusion was that it might be
possible to get the best of both approaches, but that would require serious study. That study took
years, yielded no clear results, and in the meantime more proposals emerged.
As for the closely related topic of concurrency (ğ8.4), a complete explanation of the proposals
written, presented, and discussed is beyond the scope of this paper. Here, I present only an outline.
There are simply too many complex details for anything else; the papers alone run into many
hundreds of pages and many discussions hinge on the performance of (sometimes hypothetical)
highly-optimized implementations for advanced use cases. Discussions occurred in SG1 (concur-
rency), EWG (evolution), LEWG (library evolution), CWG (core language), LWG (library), and even
in evening sessions and plenary.
Three ideas emerged repeatedly in these discussions and proposals:
• Represent the state of a coroutine and its operations as a lambda, thus fitting coroutines
elegantly into the C++ type system and not requiring some of the łcompiler magicž employed
by the await-style coroutines [Kohlhoff 2013].
• Provide a common interface to both stackless and stackful coroutines ś and possibly also to
other kinds of concurrency mechanisms, such as threads and fibers. [Kowalke 2015; Riegel
2015].
• To gain optimal performance (runtime and space) for the simplest and most critical uses (gen-
erators and pipelines), stackless coroutines need compiler support and must have interfaces
that are not compromised by the need to support more advanced use cases [Nishanov 2018,
2019b].
You can’t have all three. I’m a great fan of the idea of common interfaces because that minimizes
learning efforts and greatly eases experimentation. Similarly, using perfectly normal objects to
represent coroutines would open up the whole language to support coroutines. However, in the
end the performance argument won out.
In 2017, a proposal from Gor Nishanov based on the await stackless approach was accepted as a
TS [Nishanov 2017]. The reason this proposal (inevitably nicknamed łGor-routinesž) was approved
was that its implementation had demonstrated superior performance for its key use cases (pipelines
and generators) [Jonathan et al. 2018; Psaropoulos et al. 2017]. The reason that it was made a
TS, rather than put into the standard was that many liked the more general (but slower) stackful
coroutines and some still hoped for a zero-overhead synthesis of the two approaches. My opinion
at the time (and still today) was that a synthesis was impossible on a reasonable timescale. Having
waited almost 30 year to get coroutines back into C++, I didn’t want to wait for a breakthrough
that might never come: łThe best is the enemy of the good.ž
As usual, naming was a contentious issue. In particular, the draft TS used the keyword yield,
which was quickly determined to be a popular identifier (e.g., in finance and agriculture). Also,
the result produced by a coroutine needs to be wrapped into a structure that a caller can wait on
(e.g., a future (ğ4.1.3)) so the semantics of a coroutine return-statement is not identical to that
of an ordinary return-statement. Consequently, some people objected to the łre-usež of return.
In response, the Evolution Working Group introduced the keywords co_return, co_yield, and
co_await for the three key operations in a coroutine. The underscores were introduced to stop
native English speakers from misreading coreturn, coyield, and coawait as core-turn, coy-ield,
and coa-wait. Making yield, and await context-sensitive keywords was explored, but that didn’t
gain consensus. These new keywords aren’t pretty and they quickly became a rallying point for
people who disliked the TS coroutines for any reason.
In 2018, the TS coroutines were proposed to be put into the standard for C++20, but in the
very last minute, Geoff Romer, James Dennett, and Chandler Carruth from Google proposed a
rather novice-unfriendly proposal [Romer et al. 2018]. Like Gor’s proposal, the Google proposal,
named łCore Coroutinesž, required library support to make the basic mechanisms friendly to
non-expert users. This library had yet to be designed. Core Coroutines were claimed to be more
efficient than the TS coroutines and addressed a use case that Google had for non-exception-based
error propagation. It was based on the idea of representing the state of a coroutine as a lambda.
To avoid the widely despised keywords co_return, co_yeld, and co_await, the Core Coroutines
offered the supposedly friendlier operators [->] and [<-]. Surprisingly for an operator, [->] was four
characters long and took four operands; the ł[ž and ł]ž were part of the token. Unfortunately, the
Core coroutines were not implemented so claims of usability and efficiency could not be verified.
This delayed further decisions about coroutines.
One significant and potentially fatal problem with the TS coroutines was that they relied on free
store (dynamic memory, heap) allocation. That’s a significant overhead in some applications. Worse,
for many critical real-time and embedded applications, free store use is simply not allowed because
of its potential for unpredictable response time and the possibility of memory fragmentation. The
Core coroutines didn’t have this problem. However, Gor Nishanov and Richard Smith demonstrated
that the TS coroutines could guarantee the absence of free store use for almost all uses (and detect
and prevent the rest) in one of several ways [Smith and Nishanov 2018]. In particular, it is possible
to optimize free store use into stack allocation (łHalo optimizationž) for almost all critical use cases.
The Core Coroutines evolved and improved over time [Romer et al. 2019a], but a complete
implementation never emerged. In 2018, the Bulgarian National Body objected to the TS coroutine
design [Mihaylov and Vassilev 2018] and proposed yet another design [Mihaylov and Vassilev
2019]. Again, elegance, generality, and performance were claimed, but again, no implementation
existed.
At this point, the head of the Evolution group, Ville Voutilainen, requested that the authors of
the three still-active proposals write two evaluation and comparison papers:
• Coroutines: Use-cases and Trade-offs [Romer et al. 2019b].

• Coroutines: Language and Implementation Impact [Smith et al. 2019].
All three proposals (Gor, Google, and Bulgaria) were stackless and use cases requiring stacks
were left for future proposals. All had a bewildering number of customization points [Nishanov
2018], which were deemed essential by their implementers and expert users. It emerged that the
expression of key use cases was not dramatically different in the different proposals. Consequently,
the differences could be dismissed as largely cosmetic. For example, is co_await uglier than [<-]?
This left performance. Gor’s proposal had the clear advantage of four years of production use and
being implemented in both the Microsoft and Clang compilers. Over the last few meetings before
the key votes for C++20, the committee heard experience reports from people from Sandia [Hollman
2019], Microsoft [Jonathan et al. 2018], and Facebook [Howes et al. 2018] and considered some
suggestions about improvements and simplifications based on such uses [Baker 2019]. However,
the point that (as far as I can judge) swayed the committee to a solid 48-4 vote in favor was that
a fundamental flaw was discovered in the strategy of using łordinary lambdasž to represent the
state of a coroutine. For a lambda representing a coroutine state to be just like other lambdas, its
size must be known in the first stage of compilation. Only then can we allocate coroutine states on
the stack, copy them, move them around, and use them in the various ways the language allows.
However, the size of a stack frame (and that’s what the state of a stackless coroutine fundamentally
is) isn’t known until the optimizer has run. There is no information path from the optimizer back
to the early stages of the compiler. An optimizer can decrease the size of a frame by eliminating
variables and increase it by adding useful temporaries. Thus, a lambda represented a coroutine
state from cannot be łordinary.ž
Finally, consider a trivial example of a C++20 coroutine:
generator < int > fibonacci () // generate 0 ,1 ,1 ,2 ,3 ,5 ,8 ,13 ...
{
int a = 0; // initial values
int b = 1;
while ( true ) {
int next = a+b;
co_yield a; // return next Fibonacci number
a = b; // update values
b = next ;
}
}
int main ()
{
for ( auto v: fibonacci ())
cout << v << '\n ';
}
The use of co_yield makes fibonnacci() a coroutine. The generator<int> return value will
hold the next int generated and the minimal state needed for fibonacci() to wait for the next call.
For asynchronous use, we’d use a future<int> instead of generator<int>. The standard library
support for coroutine return types is still incomplete, but mature libraries are in production use.
Could the committee have handled coroutines better? Probably; the C++20 coroutines are
remarkably similar to Niklas Gustafsson’s 2012 proposal. It was good that we explored alternatives,
but did we really need 7 years? Could the massive efforts by many competent people have been
more collaborative and less competitive? I feel that better scholarship would have helped in the
early stages. After all, coroutines have about 60 years of history, e.g., [Conway 1963]. People
knew modern approaches in C++ and related languages, but our understanding wasn’t shared or
systematic. Had we spent a few months or a year on a thorough review of fundamental design
choices, implementation techniques, key use cases, and the literature, I suspect we could have
reached the conclusions we arrived at in February 2019 as early as 2014. The years afterwards
could then have been spent on incremental improvements and further functionality for our chosen
fundamental approaches.
Much of the credit for the progress made and the final success go to Gor Nishanov. Without his
tenacity and his solid implementation work (he did both the Microsoft and the Clang implemen-
tations), we would not have the C++20 coroutines. Perseverance is a key factor in success in the
committee.
9.3.3 Compile-Time Computation Support. The importance of compile-time evaluation has been
steadily increasing in C++ over the years. The STL relied critically on compile-time dispatch
[Stroustrup 2007] and template metaprogramming was mostly aimed at moving computation from
run time to compile time (ğ10.5.2). Even the reliance on overloading and the use of virtual function
tables in early C++ can be seen as gaining performance by moving computation from run time to
compile time. Thus, compile-time computation has always been a key part of C++.
From C, C++ inherited constant expressions that were restricted to integers and could not call
functions. Macros were needed for anything nontrivial. This did not scale well. Once templates
were introduced and template metaprogramming was discovered, template metaprogramming
became widely used to compute values and types at compile time (ğ10.5.2). In 2010, Gabriel Dos
Reis and Bjarne Stroustrup published a paper that pointed out that compile-time computation of
values could (and should) be expressed just like other computations, relying on the usual rules
for expressions and functions, including the use of user-defined types [Dos Reis and Stroustrup
2010]. This became constexpr functions in C++11 (ğ4.2.7) and the basis for modern compile-time
programming. C++14 generalized constexpr functions (ğ5.5) and C++20 added several related
features:
• consteval ś a constexpr function guaranteed to be compile-time evaluated [Smith et al.
2018a]
• constinit ś a declaration modifier to guarantee compile-time initialization [Fiselier 2019]
• the use of matching pairs of new and delete in constexpr functions [Dimov et al. 2019]
• constexpr string and constexpr vector [Dionne 2018]
• the use of virtual functions [Dimov and Vassilev 2018]
• the use of unions, exceptions, dynamic_cast, and typeid [Dionne and Vandevoorde 2018]
• user-defined types as value template arguments ś finally allowing user-defined types wher-
ever built-in types are [Maurer 2012]
• is_constant_evaluated() predicate ś to enable library implementers to optimize code with
far fewer platform dependent intrinsics [Smith et al. 2018b]
Along with this effort, the standard library is being made more friendly to compile-time evalua-
tion.
The ultimate aim for much of this effort is to support static reflection for C++23 or later (ğ9.6.2).
The use of user-defined types as template argument types and the use of strings as template
arguments were anticipated when I first designed templates, but then they were beyond my ability
to design and implement properly.
There are people who would like every C++ construct to be available at compile time. In particular,
they would like to be able to use the complete standard-library in constexpr functions. That may
be too much of a good thing. For example, do you really need threads at compile time? Yes, it is
feasible. Not making all functions usable at compile time, leaves us with the problem of deciding
which should be usable and which should not. So far, the answer is a bit ad hoc and not coherent.
Further refinement is needed.
To make a language construct or library component constexpr, it has to be very precisely

specified and stripped of the possibility of undefined behavior. Thus, the push for compile-time eval-
uation has been a major driver of more precise specification, for analysis of platform dependencies,
and examination of the roots of undefined behavior.
Obviously, this push for compile-time computation creates more work for compilers. The increase
in the amount of information in interfaces to allow the compiler to do all that work is being addressed
through modules (ğ9.3.1). Compilers also compensate by caching results and systems relying on
parallel builds are common. However, C++ programmers must learn to limit their use of compile-
time computation and metaprogramming to code where the compactness and run-time performance
makes it worthwhile.
9.3.4 <=>. See (ğ8.8.4). Immediately after the łspaceship operatorž (<=>) was voted in for C++20,
it became clear that serious further work was needed on both the language rules and its integration
into the standard library. Out of over-enthusiasm and a desire to solve the thorny problems related
to comparisons, the committee had fallen victim to the law of unintended consequences. Some
committee members (including me) worried that the introduction of <=> had been too hurried.
However, by the time our worries became concrete, work had been done assuming the availability
of <=>. Also, many committee members and others in the wider C++ community had become
excited by the possible performance advantages of a three-way comparison. It therefore came
as an unpleasant surprise to find that <=> caused significant inefficiencies in important cases.
Given <=> for a type, == was generated from <=>. For strings, == is usually optimized by first
comparing sizes: if the number of characters differ, the strings are not equal. A == generated from
<=> must read enough of the strings to determine their lexicographical order and that can be far
more expensive. After long discussions, it was decided not to generate == from <=> . That and
several other corrections [Crowl 2018; Revzin 2018, 2019; Smith 2018c] solved the problems in
hand, but compromised the fundamental promise of <=>: that all comparison operators could be
generated from a single, simple line of code. Also, thanks to <=>, == and < now have many rules
that differ from the rules for other operators (e.g., == is assumed to be symmetric). For better and
worse, most rules related to operator overloading have <=> as a special case.
9.3.5 Ranges. The Ranges library started as the work of Eric Niebler to generalize and modernize
the STL notion of a sequence [Niebler et al. 2014]. It provides standard-library algorithms that
are simpler to use, more general, and better performing . For example, the C++20 standard library
offers the long-awaited simpler notation for operations on whole containers:
void test ( vector < string >& vs )

{
sort ( vs ); // rather than sort ( vs . begin () , vs . end ())
}
The original STL as adopted for C++98 [Stroustrup 1993] defined a sequence as a pair of iterators.
That left out two important ways of specifying a sequence. The range library provides the three
major alternatives (now called ranges):
• (first,one_beyond_last) for when we know where the beginning and the end of the sequence
are (e.g., łsort from the beginning to the end of a vectorž).
• (first,number_of_elements) for when we don’t actually need to have the end of a sequence
computed (e.g., łlook at the first 10 elements of a listž).
• (first,termination_criteria) for when we use a predicate (e.g., a sentinel) to define the end
of the sequence (e.g., łread until end of inputž).
A range is a concept (ğ6). All C++20 standard-library algorithms are now precisely specified
using concepts. This is in itself a major improvement and it is what enables the generalization to
use ranges, rather than just iterators. That generalization allows for pipelining of algorithms:
vector < int > vec = {1 , 2, 3, 4, 5, 6, 7 ,8, 9, 10 };
auto even = []( int i ){ return i %2 == 0; }
for ( int i : vec | view :: filter ( even )

| view :: transform ( []( int i) { return i*i; } )
| view :: take (5))
cout << i << '\n '; // print the squares of the first 5 even integers
Like in Unix, the pipeline operator, |, delivers the output of its left-hand operand as the input
to its right-hand operand (e.g., A|B means B(A). This will get much more interesting once people
start using coroutines (ğ9.3.2) to write pipeline filters.
In 2017, the ranges library became a TS [Niebler and Carter 2017] and February 2019 it was voted
into C++20 [Niebler et al. 2018].
9.3.6 Dates and Time Zones. The date library is the work of Howard Hinnant (Ripple, formerly
Apple) providing standard calendar and time zone support for C++ [Hinnant and Kamiński 2018].
It is based on the chrono standard-library time support. Howard was also the main person behind
chrono (ğ4.6). The date library is the result of years of work and real-world use. In 2018, it was
voted into C++20 and placed in <chrono> together with the older time utilities.
Consider how to express a point in time (a time_point):
constexpr auto tp = 2016 y/ May /29 d + 7h + 30 min + 6s + 153 ms ;
cout << tp << '\n '; // 2016 -05 -29 07:30:06.153
The notation is conventional (using user-define literals (ğ4.2.8)) and a date is represented as a
year,month,day structure. However, when needed, a date maps to a point on the standard time
line (system_time) at compile time (using constexpr functions (ğ4.2.7)), so it is blindingly fast
and can be used in constant expressions. For example:
static_assert (2016 y/ May /29== Thursday ); // check at compile time
By default, the time zone is UTC (aka Unix time), but conversion to different time zones is easy:
zoned_time zt = {" Asia / Tokyo ", tp };
cout << zt << '\n '; // 2016 -05 -29 16:30:06.153 JST
The date library can also handle days of the week (e.g., Monday and Friday), multiple calendars
(e.g., Gregorian and Julian), and more esoteric (but necessary) notions, such as leap seconds.
In addition to being useful and fast, the date library is interesting because it offers very fine-
grained static type checking. Common mistakes are caught at compile time. For example:
auto d1 = 2019 y /5/4; // error : May 4 or April 12?
auto d2 = 2019 y/ May /4; // OK
auto d2 = May /4/2019; // OK ( the day follows the month )
auto d3 = d2 +10; // error : add 10 days , 10 months , or 10 years ?
The date library is a rare example of a standard-library component that directly addresses an
application domain, rather than łjustž offering a supporting łcomputer sciencež abstraction. I hope
to see more such in future standards.
9.3.7 Format. The iostream library offers type-safe extensible I/O, but it’s formatting facilities are
weak. In addition, some people dislike the use of << to separate output values. The format library
provides a printf-like way of composing strings and formatting values that is typesafe, fast, and
works with iostreams. It is primarily the work of Victor Zverovich [Zverovich 2019].
Types that have a << operator can be output within a format string:
string s = " foo ";
cout << format (" The string '{} ' has {} characters ",s ,s. size ());
This outputs The string ’foo’ has 3 characters.

This is a variant of the łtypesafe printfž variadic template idea (ğ4.3.2). The {} simply indicates
that a default representation of the value of an argument is to be inserted.
The argument values can be used in any order:
// s before s. size ():
cout << format (" The string '{0} ' has {1} characters ",s ,s. size ());
// s. size () before s:
cout << format (" The string '{1} ' has {0} characters ",s. size () , s );
Like printf(), format() offers a whole little programming language for expressing formatting
details, such as field width, floating-point precision, integer number base, and alignment in a field.
Unlike printf(), format() is extensible and can handle user-defined types. Here is an example
printing a date from the <chrono> library (ğ9.3.6) [Zverovich et al. 2019]:
string s1 = format ("{}" , birthday );
string s2 = format ("{0: >15% Y -%m -% d }" , birthday );
The year-month-day format is the default. The >15 means use 15 characters and left-align the
text. The date library contains yet-another small formatting language for use with format(). It can
even handle time zones and locales:
std :: format ( std :: locale {" fi_FI "} , "{}" , zt );
This will give us the local time in Finland. By default, formatting is not locale dependent, but you
can opt in. This significantly improves performance compared to traditional iostream performance
when you don’t need locale information.
There is no equivalent to format for input (istream).
9.3.8 Span. Out-of-range access, sometimes referred to as buffer overflow, has been a serious
problem from the early days of C. Consider:
void f( int * p , int n) // what is n?
{
for ( int i =0; i <n; ++ i)
p[i] = 7; // OK ?
}
How would a tool, e.g., a compiler, know that n was meant to be the number of elements in the
array pointed to? How can a programmer get it consistently right in a large program?
int x = 100;
int a [100];
f(a ,x ); // OK
f(a ,x /2); // OK : first half of a
f(a ,x +1); // disaster !
For decades, that łdisasterž comment has been accurate and range errors have been the root
of major security problems. Compilers do not catch range-errors and run-time checking of all
subscripting has generally been deemed too expensive for production code.
The obvious solution is to supply an abstraction that holds a pointer plus a size. For example, in
1990, Dennis Ritchie proposed that to the C standard committee: ł‘fat pointers’ whose representation
will include the space to store the adjustable bounds at run-timež [Ritchie 1990]. For various reasons,
the C committee didn’t approve. At the time, I heard the priceless comment: łDennis isn’t a C
expert; he never comes to the meetings.ž It is probably good that I don’t remember who said that.
In 2015, Neil MacIntosh (then at Microsoft), revived that idea for the C++ Core Guidelines (ğ10.6)
where we needed a mechanism to encourage and optionally enforce effective programming styles.
This span<T> class template was put into the Core Guidelines support library and promptly ported
to the Microsoft, Clang, and GCC C++ compilers. In 2018, it was voted into C++20.
Using span, that example can now be written:
void f( span < int > a) // span holds a pointer and a size
{
for ( int & x : a)
x = 7; // OK
}
The range-for extracts the range from the span and traverses exactly the right number of
elements (without costly range checking). This is an example of how an appropriate abstraction
can simultaneously simplify notation and improve performance. It is easier and cheaper for an
algorithm to explicitly use a range (here, span) than to check each individual element access.
If necessary, you can explicitly specify a size (e.g., to operate on a subrange), but then the risk is
yours and the notation stands out as a warning:
int x = 100;
int a [100];
f(a ); // template argument deduction : f( span < int >{a ,100})
f ({a ,x /2}); // OK : first half of a
f ({a ,x +1}); // disaster
Naturally, simple element access can also be done, e.g., a[7]=9, and can be checked at run time.
Range checking of span is the default in the C++ Core Guideline Support Library (GSL).
The most contentious part of getting span into C++20 turned out to be the type of subscripts
and sizes. The Core Guidelines span::size() was defined to return a signed int as opposed to the
unsigned used by the standard-library containers. Similarly, subscripts were ints like for arrays,
rather than unsigned as they are for standard-library containers. This led to a revival of an old
and contentious issue:
• Some consider it obvious that subscripts, being non-negative, should be represented as
unsigned.
• Some consider consistency with standard-library containers more important than any argu-
ments about past mistakes related to unsigned.
• Some consider the use of unsigned to represent non-negative numbers misguided (giving a
false sense of safety) and a significant source of errors.
Over the strenuous objections of the original designers (including me) and implementers of span,
the second group won the votes, enthusiastically supported by the first group, so std::span has
unsigned sizes and subscripts. I consider that a sad failure to take advantage of a rare opportunity
to remedy a nasty old mistake [Stroustrup 2018e]. Somewhat predictably and not irrationally, the
committee chose bug compatibility over the work of removing a significant bug source.
But what could be wrong about representing subscripts as unsigned? This appears to be a rather
emotional topic. I received several hate mails about this. There are two fundamental problems:
• unsigned does not model natural numbers: it has modular arithmetic and subtraction. For
example, if ch is an unsigned char, ch+100 will never overflow.
• int and unsigned convert to each other at the slightest provocation, turning negative values
into huge signed values and vice versa. For example, -2<2u is false; 2u is unsigned, so -2 is
converted into a huge positive integer before the comparison.
Here is an infinite loop occasionally seen łin the wildž:
for ( size_t i = n -1; i >= 0; --i) { /* ... */ } // " reverse loop "
Unfortunately, the standard-library type size_t is unsigned and then obviously always >=0.
Basically, the rules for conversions among signed and unsigned types, as inherited by C++ from C,
have been a major source of hard-to-find errors for decades, but it is hard to convince a committee
to address old problems.
9.4 Concurrency
Despite valiant efforts and an emerging broad consensus, the hoped-for general concurrency
model (łexecutorsž) wasn’t ready for C++20 (ğ8.8.1). This was not for lack of effort, including a
special two-day September 2018 meeting in Bellevue WA attended by about 25 people including
representative from NVIDIA, Facebook, and the US National Labs. However, several less dramatic
useful improvements were completed in time, including
• jthread and stop tokens [Josuttis et al. 2019a]
• atomic<shared_ptr<T>> [Sutter 2017b]
• classic semaphores [Lelbach et al. 2019]
• barriers and latches [Lelbach et al. 2019]
• minor memory model repairs and improvements to [Meredith and Sutter 2017]
The jthread (short for łjoining threadž) is a thread that obeys RAII; that is, it’s destructor joins,
rather than terminates, if the jthread goes out of scope:
void some_fct ()
{
thread t1 ;
jthread t2 ;
// ...
}
At the end of scope, t1’s destructor terminates the program unless t1’s task has completed,
joined, or detached, whereas t2’s destructor will wait for its task to complete.
From the start (pre-C++11), many (incl. me) had wanted thread to have what is now jthread’s
behavior, but people grounded in traditional operating systems threads insisted that terminating
a program was far preferable to a deadlock. In 2012 and 2013, Herb Sutter proposed a joining
thread [Sutter 2012, 2013a]. This led to much discussion, but no decisions. In 2016, Ville Voutilainen
summarized the issues and conducted votes for including joining threads into C++17 [Voutilainen
2016a]. The votes were so massively in favor that I (only partly joking) suggested that we could
submit joining threads to C++14 as a bugfix, but somehow, the progress again stalled. In 2017, Nico
Josuttis, re-raised the issue and eventually ś after eight revisions and the addition of stop tokens ś
the proposal made it into C++20 [Josuttis et al. 2019a].
łStop tokensž is a solution to the old problem of how to stop a thread when we are no longer
interested in its result. The basic idea is to make thread cancellation (ğ4.1.2) cooperative. If I want a
jthread to terminate, I set its stop token. It is the thread’s obligation to occasionally test whether
the stop token has been set and if so, clean up and exit. This technique is as old as the mountains
and works nicely and efficiently for just about every thread that has a main loop where the test of
the stop token can be placed.
As ever, naming became an issue: safe_thread, ithread (’i’ for interruptible), raii_thread,
joining_thread, and finally jthread. The Guideline Support Library calls it gsl::thread. Really,
the proper name is thread, but unfortunately, that name was already taken for a less useful kind
of thread.
9.5 Minor Features

C++20 offers many minor new features, such as:
• C99-style designated initializers [Shen et al. 2016]
• Refinements to lambda capture [Köppe 2017b]
• Template parameter lists for generic lambdas [Dionne 2017]
• Initialization of an additional variable within a range-for (ğ8.7)
• Lambdas in unevaluated contexts [Dionne 2016]
• Pack expansions in lambda capture [Revzin 2017]
• Removing the need for typename in some cases [Vandevoorde 2017]
• More attributes: [[likely]] and [[unlikely]] [Trychta 2016]
• source_location to give the source code location of a piece of code without the use of macros
[Douglas and Jabot 2019]
• Feature test macros [Voutilainen and Wakely 2018]
• Conditional explicit [Revzin and Lavavej 2018]
• Signed integers are guaranteed to be two’s complement [Bastien 2018]
• Mathematical constants, such as pi and sqrt2 [Minkovsky and McFarlane 2019]
• Operations on bits, such as rotations and counting ones [Maurer 2019]
Some are improvements, but I worry that the sheer volume of obscure novelties does harm
[Stroustrup 2018d]. They make the language harder to learn and code harder to read for non-
experts. I opposed several features as likely to do as much harm as good (e.g., designated initializers
will be used where constructors would lead to more maintainable code). Many are special-purpose
and some łexpert-only.ž However, there are people who have a hard time understanding that a
feature that can do some good for someone, can be a net liability for C++. Small features that
increase generality and uniformity of notation and semantics are of course always welcome.
From a standardization point of view, even the smallest feature takes time to process, document,
and implement. That time cannot be used for something else.
9.6 Work in Progress

Naturally, much work is going on aimed at releases beyond C++20 and some work aimed at C++20
didn’t complete in time. Notably:
• ğ8.8.1: Networking and executors ś delayed yet again.
• ğ9.6.1: Contracts ś assertions, preconditions, and postconditions; aimed for C++20, but
delayed.
• ğ9.6.2: Reflection ś injecting code into a program based on the code being compiled; aimed
for C++23.
In addition, the working group and study group pipelines are not empty (ğ3.2) [Stroustrup 2018d].
9.6.1 Contracts. Contracts are special in that not only did most people expect them to make it
into C++20, but contracts were voted into the working paper for C++20 just to be yanked out at
the very last moment. A new Study Group, SG-21 chaired by John Spicer, has been created to try to
get some form of contracts back for C++23 or C++26. The story of contracts for C++20 is sad but
possibly enlightening.
Contracts in various guises have a long history in C++ and other languages. I remember being
very attracted to Peter Naur’s invariants [Naur 1966] when I first encountered them in the early
1970s. A system of assertions, called A++, was considered for C++ in the early 1990s, but was
considered too extensive to be practical. In the late 1980s, Bertrand Meyer popularized the notion of
łcontractsž with Eiffel [Meyer 1994]. As part of the C++0x effort, a couple of proposals [Crowl and
Ottosen 2006] received serious attention in the committee but eventually failed, mostly because of
perceived excess complexity and inelegant notation.
For years, Bloomberg (the New York City financial information company) had used a system
of run-time assertions, called contracts, to catch problems in their code. In 2013, John Lakos from
Bloomberg proposed that system for standardization [Lakos and Zakharov 2013]. It was well
received, but it ran into two problems:
• It was based on macros
• It was strictly assertions in implementation code, rather than something that enhanced
interfaces.
Revisions followed, but no consensus emerged. To try to break the deadlock a group of people
from Microsoft, Facebook, Google, and University Carlos III in Madrid proposed a system of łsimple
contractsž that did not use macros and added support for pre-conditions and post-conditions (as
had the C++0x attempts) [García et al. 2015]. Like the proposal from Bloomberg, this proposal was
backed by many years of large-scale industrial use, but had an emphasis on the use of contracts in
static analysis. J-Daniel Garcia (University Carlos III) worked hard to produce a design that would
satisfy all, but that proposal also ran into opposition.
After many meetings, papers, and (occasionally heated) discussions, it became clear that a
compromise was elusive. The two groups then asked me to coordinate and to prove my conjecture
that the discussions were focused on łdetailsž and that there had to be a minimal proposal that
included the essentials of both groups stated needs and no controversial łdetails.ž After a fair bit of
work where I alternately met with representatives of the two groups, we finally produced a joint
proposal coauthored by łall partiesž [Dos Reis et al. 2016a]. I think that design was technically
sound rather than a political compromise. It aimed to serve three needs (in order of importance):
• Systematic and controlled run-time tests
• Information for static analyzers
• Information for optimizers
After further work led by J-Daniel Garcia, the proposal was adopted for C++20 in June 2018 [Dos
Reis et al. 2018].
To avoid introducing new keywords, we used the attribute syntax. For example, [[assert: x+y>0]].
A contract has no effect on a valid program, so this fits the original conception of attributes (ğ4.2.10).
There are three kinds of contracts:
• assert ś assertions in executable code
• expects ś pre-conditions on function declaration
• ensures ś post-conditions on function declarations
There are three levels of contract checking:

• audit ś for łexpensivež predicates checked only in some łdebug modež
• default ś for łcheapž predicates that you could plausibly check even in production code
• axiom ś for predicates intended primarily for static analyzers and never checked at run time.
Upon a contract violation a (possibly user-installed) violation handler is executed. The default
action is immediate termination.
I found one aspect interesting: There is a build-mode that allows a program to continue after a
contract failure. My first reaction to that was łbut that’s insane! Contracts are intended to prevent
programs with contract violations from running.ž That was by far the most common reaction.
However, John Lakos insisted, based on experience with Bloomberg code, that when you add
contracts to a large old code base, you invariably get violations:
• Some code will violate a contract without actually doing anything it protects against.
• Some new contracts will contain errors.
• Some new contracts will have unintended effects.
With continuation, you can use a violation handler to log violations and continue. That way, you
can detect many violations in a single run and also leave contracts active in supposedly correct old
code. This was accepted to be critical for gradual adoption of contracts.
We saw no sufficient reasons to add class invariants, to allow weakening pre-conditions for
overriding functions, or to allow strengthening post-conditions for overriding functions. The
emphasis was on simplicity. The ideal was to provide a minimal initial design for C++20 and then
add to it later if needed.
This design was implemented by J-Daniel Garcia and in June 2018 it was voted into the commit-
tee’s working paper for C++20. As usual, a few bugs remained in the specification but we were
confident that those could be dealt with in the two years left before the release of the final standard.
For example, it was discovered that the working paper text allowed a compiler to optimize based
on all contracts (checked or not). That was unintended. It made sense from the point of view that
all contracts are valid in a correct program, but disastrous for a program with contracts specifically
written to catch łimpossible errors.ž Consider:
[[ assert : p != nullptr ]]
p ->m = 7;
If p==nullptr, p->m is undefined behavior. A compiler is allowed to assume that undefined

behavior doesn’t happen; it can therefore optimize away code that leads to the undefined behavior.
The results of that can be very surprising. In this case, if the execution might continue after a
contract violation, the compiler would be allowed to assume that p->m is valid so that p!=nullptr;
then, it could eliminate the contract’s check of p==nullptr. This (łtime-travel optimizationž) is of
course contrary of the aims of contracts and several proposals to remedy that were submitted in a
timely manner [Garcia 2018; Stroustrup 2019c; Voutilainen 2019a].
In October 2018, after the deadline for new proposals for C++20, a team from Bloomberg led
by John Lakos, including Hyman Rosen and Joshua Berne presented a stream of proposals for
redesigns [Berne et al. 2018; Berne and Lakos 2018a,b; Lakos 2018]. The date of the feature freeze
(the last day to consider new proposals) had been fixed by a full plenary committee vote. The
proposals were based on a scheme to specify the behavior of a contract in the contract itself. For
example, [[assert check_maybe_continue: x>0]] and [[assert assume: p!=nullptr]].
Instead of using build modes to control the behavior of all contracts (e.g., enable all default
contracts or turn off all contract-based run-time checking), you would modify the source code of
individual contracts. In this, these new schemes differed radically from the agreed-upon design in
the working paper. Consider:
[[ assert assume : p != nullptr ]]
This would bring back the macro-based scheme that had been rejected in 2014 because the
obvious way of managing code changes would involve macros in the code. For example:
[[ assert MODE1 : p != nullptr ]]
Here, MODE1 could be #defined as one of the supported alternatives such as assume and
default. Alternatively and roughly equivalently, the meaning of qualifiers such as assume could
be defined by assignments on the command line (acting like command-line macros).
In essence, the combination of the possibility of continuation after a contract violation and
programmer control of the meaning of a contract, would turn the contract mechanism from a
system of assertions to a novel control-flow mechanism.
Some proposals even suggested to abolish support for static analysis. There were dozens of
variations of these proposals, all late and none capable of increasing consensus.
The flood of novel proposals (from the Bloomberg team and others, e.g., [Berne 2019; Berne
and Lakos 2019; Khlebnikov and Lakos 2019; Lakos 2019; Rosen et al. 2019]) and the hundreds of
email messages discussing them blocked the needed discussion of bug fixes to the status quo design
in the working paper. As I had repeatedly warned (e.g., in [Stroustrup 2019c]), the result of the
proposals for a redesign was that contracts were removed from C++20 after a proposal to do so by
Nico Josuttis [Josuttis et al. 2019b]. I consider the last year of discussions about contracts a classic
case of nobody getting anything because someone wanted it exactly their way. Time will tell if the
new study group, SG21, can deliver something more widely acceptable for C++23 or C++26.
9.6.2 Static Reflection. In 2013, a study group for łreflectionž (SG-7) was formed and a call for
ideas was issued [Snyder and Carruth 2013]. There was broad agreement that C++ needed a static
reflection mechanism. That is, we needed a way to write code that examined the program of which
it was part and to inject code based on that analysis into that program. That way, we could replace
tedious and tricky boilerplate, macros, and extra-linguistic generators with clean code. For example,
we could generate functions for stream I/O, logging, comparison, marshalling for storage and
networking, building and using object maps, łstringificationž of enumerators, test support, and
more [Chochlík et al. 2017; Stroustrup 2018g]. The aim of the reflection study group was to get
something ready for C++20 or C++23; it was understood that C++17 was not a realistic target.
It was agreed that reflection/introspection relying on run-time traversal of an ever-present data
structure was unsuitable for C++ because of the size of such data, the complexity of a complete
representation of the language constructs, and the run-time cost of traversal.
Several proposals promptly emerged [Chochlík 2014; Silva and Auresco 2014; Tomazos and
Spertus 2014] and over the following years the study group, chaired by Chandler Carruth, held
several meetings to try to decide on scope and direction. The approach chosen was based on types
organized in classical object-oriented class hierarchies supported by concepts (ğ6) where they
needed to be generic [Chochlík 2015; Chochlík and Naumann 2016; Chochlík et al. 2017]. This
approach was developed and implemented primarily by Matóš Chochlík, Axel Naumann, and David
Sankel. It resulted in a technical specification approved in 2019 [Sankel 2018].
During the (expected) long gestation period for static reflection, compile-time computation
based on constexpr functions (ğ9.3.3) developed steadily and eventually proposals to base static
reflection on functions rather than class hierarchies emerged. The main proponents were Andrew
Sutton, Daveed Vandevoorde, Herb Sutter, and Faisal Vali [Sutton and Sutter 2018; Sutton et al.
2018]. The main arguments for the shift in focus was partly that analyzing and generating code are
inherently functional and that the compile-time computation based on constexpr functions had
developed to the point where metaprogramming and reflection merged. Another strong point (first
presented by Daveed Vandevoorde) of that approach was that a compiler’s internal data structures
for functions were inherently smaller and more transient than for type hierarchies so they used
significantly less memory and compilation was significantly faster.
At the February 2019 standards meeting in Cologne, David Sankel and Michael Park presented a
design that combined the strengths of the two approaches [Sankel and Vandevoorde 2019]. At the
most fundamental level only a single type exists. This maximizes flexibility and minimizes compiler
overheads.
On top of that, a statically typed interface can be imposed by a form of safe-type conversion (from
the low-level monotype meta::info to more specific types such as meta::type_ and meta::class_).
Here is an example based on [Sankel and Vandevoorde 2019]. It achieves the conversion from
meta::info to more specific types through concept overloading (ğ6.3.2). Consider:
namespace meta {
consteval std :: span < type_ > get_member_types ( class_ c) const ;
}
struct baz {
enum E { /*...*/ };
class Buz { /*...*/ };
using Biz = int ;
};
void print ( meta :: enum_ ); // print an enumeration

void print ( meta :: class_ ); // print a class
void print ( meta :: type_ ); // print any type
void f ()
{
constexpr meta :: class_ metaBaz = reflexpr ( baz );
template for ( constexpr meta :: type member_ : get_member_types ( metaBaz ))
print ( meta :: most_derived ( member_ ));
}
The key new language features here is the reflexpr operator that returns a (meta) object
describing its argument and template for [Sutton et al. 2019] that iterates over the elements of a
heterogeneous structure by expanding each element according to its type.
In addition, there is a mechanism for injecting code into the program being compiled.
The likelihood is that something like this will become standard in C++23 or C++26.
As a side effect, the work on the ambitious reflection schemes spurred improvements to compile-
time evaluation facilities:
• The set of type traits in the standard (ğ4.5.1)
• The macros for source location (e.g., __FILE__ and __LINE__) were replaced by an intrinsic
mechanism [Douglas and Jabot 2019]
• The facilities for compile-time computation (e.g., consteval for guaranteed compile-time
evaluation)
• Expansion statements (template for ś for iterating over tuple elements [Sutton et al. 2019])
for C++23.
10 C++ IN 2020
This section looks at how C++ was used in the second decade of 2000 and for what:
• ğ10.1: What is C++ used for?
• ğ10.2: The C++ community
• ğ10.3: Education and Research
• ğ10.4: Tools
• ğ10.5: Programming styles
• ğ10.6: Core Guidelines
Areas of use are much as in 2006 (ğ2.3). Some new areas been added, but mostly what we see is
a wider and deeper use in the same and similar areas. C++ hasn’t suddenly become a łweb-app
language,ž though there is some use even there [Obiltschnig et al. 2005]. For most programmers, C++
is still something in the background that offers stability, reliability, portability, and performance.
For end users, C++ is invisible.
The change in programming styles has been more dramatic. C++11 is a far better language than
C++98. It is easier to use well, more expressive, and deliver better performance. C++20, as deployed
in 2020, is a similar improvement over C++11.
10.1 What Is C++ Used For?

To a first approximation, C++ is used everywhere and for everything. However, most uses are
invisible, deep in the infrastructure of important systems.
Nobody knows everything about where C++ is used, or how. In 2015, the Czech company
JetBrains commissioned a study [Kazakova 2015] that showed massive use in North America,
Europe and the Middle East, and the Asia Pacific areas plus some use in South America. łSome use
in South Americaž was 400,000 developers. The total was 4.4 million developers. The industries
listed were (in order) finance, banking, games, front office, telecom, electronics, investment banking,
marketing, manufacturing, and retail. All indications are that the size of the user population and
the areas of use have steadily increased since 2015.
Here I will give a ś necessarily somewhat personal, impressionistic, and very incomplete ś view
of the range of C++ usage in the 2006-2020 time span:
• Industry: Telecom (e.g., AT&T, Ericsson, Huawei, and Siemens), mobile devices (essentially
all; signal processing, screen rendering, apps with significant performance or portability
requirements), microelectronics (e.g., AMD, Intel, Mentor Graphics, and Nvidia), finance (e.g.,
Morgan Stanley and Renaissance), games (almost all), graphics and animation (e.g., Maya,
Disney, and SideFx), block chain implementation (e.g., Ripple), databases (e.g., SAP, Mongo,
MySQL, and Oracle), cloud (e.g., Google, Microsoft, IBM, and Amazon), AI and ML (e.g., the
TensorFlow library), operations support (e.g., Maersk and AT&T).
• Science: Aerospace (e.g., Space X, Mars Rovers, the Orion crew vehicle, the James Webb Space
Telescope), high Energy Physics (e.g., CERN, SLAC, FermiLab), biology (genetics, genome
sequencing), exascale computing.
• Teaching: Most engineering schools worldwide.
• Software development: TensorFlow, tools, libraries, compilers, Emscripten (generating asm.js
and WebAssembly from C++), run-time code generation, LLVM (the backend backbone of
many new languages and of much tool building), XML and JSON parsers, heterogenous
computing (e.g., SYCL [Khronos Group 2014ś2020] and HPX [Stellar Group 2014ś2020]).
• Web infrastructure: Browsers (Chrome, Edge, FireFox, and Safari), JavaScript engines (V8 and
SpiderMonkey), JVMs (HotSpot and J9), Google and similar organizations (search, map-reduce,
and file system).
• Major Web applications: Alibaba, Amadeus (airline ticketing), Amazon, Apple, Facebook,
PayPal, Tencent (WeChat), Yandex.
• Engineering: Dassault (CAD/CAM), Lockheed Martin (airplanes).
• Automotive: Assisted driving [ADAS Wikipedia 2020; Mobileye 2020; NVIDIA 2020], software
architecture [Autosar 2020; Autosar Wikipedia 2020], machine vision [OpenCV 2020; OpenCV
Wikipedia 2020], BMW, GM, Mercedes, Tesla, Toyota, Volvo, Volkswagen, Waymo (Google
self-driving cars).
• Embedded systems: smart watches and health monitors (e.g., Garmin), cameras and video
equipment (e.g., Olympus and Canon), navigation aids (e.g., TomTom), coffee machines
(e.g., Nespresso), farm-animal monitoring (e.g., Big Dutchman), production-line temperature
control (e.g., Carlsberg).
• Security: Kaspersky, NSA, Symantec.
• Medicine and biology: Medical monitoring and imaging (e.g., Siemens, GE, Toshiba, and
Phillips), tomography (e.g., CAT scanners), genome analysis, bioinformatics, radiation oncol-
ogy (e.g., Elekta and Varian).
This only scratches the surface, but it demonstrates the breadth and depth of C++ use. Most
C++ uses are invisible to its (indirect) users. Some of these uses started pre-2006, but many were
initiated later. No major modern system is written exclusively in a single language, but C++ plays a
major role in what is mentioned here.
It is easy to forget that many uses are quite mundane, yet important in our lives. Yes, C++ helps
run NASA’s deep space network, but it also runs in familiar gadgets, such as coffee machines, stereo
speakers, and dishwashers. I was surprised to find it used in the advanced systems used to run
modern pig farms.
10.2 The C++ Community

Compared to 2006, the 2020 C++ community is larger, growing, optimistic, vibrant, productive, and
impatient for further improvements.
Compared to most programming language communities, the C++ community has always been
amazingly disorganized and fragmented. The problem started early because I had (and have) no
talent for building organizations. My employer at the time, AT&T Bell Labs, had no desire to develop
a C++ community, but it seemed that everybody else was keenly interested and willing to spend
money to build up their user base. The net effect was that many companies, such as Apple, Borland,
GNU, IBM, Microsoft, and Zortech formed C++ communities focused on their customers, but there
were no general C++ community ś no center to the C++ community. There were magazines, but few
(relative to the size of the C++ community) read them. There were conferences, but they tended to
get absorbed into or mutated into general łobject orientedž or łsoftware developmentž conferences.
There was no general C++ users’ group.
Today there are many dozen local, national, and international C++ users’ groups with some
cooperation, and dozens of C++ conferences each with attendance into the hundreds:
• The C++ Foundation ś founded in 2014 as non-profit organization to promote ISO C++ (and
not any particular vendor’s C++). It hosts the CppCon annual conference.
• Boost ś a collection of peer-reviewed libraries and the community that builds and uses them.
Boost holds an annual conference. Founded in 1999.
• Meeting C++ ś a very active network of user groups holding conferences (initially active
in Germany). There are dozens of Meeting C++ conferences and meetings (łmeetupsž) in
various places. Founded in 2012.
• ACCU ś the grandfather of all surviving C++ organizations; it publishes two magazines and
holds annual conferences (primarily active in England). Founded as a C users’ group in 1984.
• isocpp.org ś the website of the C++ Foundation with C++-related news, information of the
standards process, and useful links.
• cppreference.com ś an excellent online reference; it even has a history section!
• Conferences ś CppCon, ACCU, Meeting++, C++ Now (formerly BoostCon), Qt, NDC, std::cpp,
plus several in Poland, Russia, China, Israel, and elsewhere. C++ tracks at general software
conferences are also on the rise.
• Blogs ś many, and podcasts.
• Videos ś Videos have become a major source of information about new developments in
C++. The major C++ conferences typically video the talks and post them for free access
(e.g., CppCon, C++ Now, and Meeting++). Video interviews have become popular. The most
popular hosting site is YouTube, which unfortunately is blocked in some countries with large
C++ developer communities (e.g., China).
• GitHub ś makes it easier to share code and to organize joint projects.
This is nowhere near a match for the centralized organizations of some languages and vendors,
but it is a lively and varied set of interrelated communities and far, far more active than what the
C++ community had in 2006. Also, several corporate user’s groups and conferences are still active.
10.3 Education and Research

Has C++ education improved since the sorry state in 2006 (ğ2.3)? Maybe, but it is certainly still
not a strength for C++ and most new efforts focus on information and training for people in
industry. In most countries, most students emerge from universities with only weak and inaccurate
understanding of C++ and the key techniques for using it. This is a serious problem for the C++
community. No language can succeed at an industrial scale without a steady stream of enthusiastic
programmers mastering its key design and implementation techniques. So much could be done
to improve software if more developers using C++ knew how to use it better! So much would be
easier if graduates entered the workforce with a more accurate view of C++!
A problem for C++ is that educational establishments often treat programming as a low-level
skill, rather than a foundational topic. Good software is essential to our civilization. To manage,
we need to treat software development for critical systems as seriously as we treat mathematics
and physics. A łone size fits allž approach to education and software development doesn’t work.
One semester’s teaching isn’t sufficient. We would never expect to teach English in just a few
months and then expect the students to appreciate Shakespeare. Similarly, there is a difference
between knowing the basics mechanics of a language and mastering the idioms and techniques
used by professionals. C++ ś like any major modern programming language ś needs a variety of
teaching approaches adjusted to the learners backgrounds and needs. Even when an educational
institution is aware of these problems and would like to compensate, curricula are overcrowded
and teachers have a hard time keeping up with industrial practice. SG20 (Education) is trying to
help by outlining approaches to teaching and using modern C++. SG15 (Tooling) could become
important by making tools supporting teaching more available.
Since C++11, there has been an increasing awareness of this. For example, Kate Gregory has
produced some great videos on how to teach C++ [Gregory 2015, 2017, 2018]. Some recent books
address the education problems head on by recognizing that there are several constituencies with
different needs when it comes to support for education:
• Programming: Principles and Practice using C++ [Stroustrup 2008a] ś a textbook aimed at
beginning university students and people doing self-study
• A Tour of C++ [Stroustrup 2014d, 2018f] ś a short (200 pages) overview aimed at more
experienced programmers
• Discovering Modern C++ [Gottschling 2015] ś a book specifically for students with a strong
mathematical background
I wrote a couple of semi-academic papers (Software Development for Infrastructure [Stroustrup
2012] and What should we teach software developers? Why? [Stroustrup 2010b]) and devoted the
opening keynote of CppCon 2017 to education (Learning and Teaching Modern C++ [Stroustrup
2017c]).
The use of videos and online courses has dramatically increased since about 2014. This serves
C++ well because that doesn’t require a central organization or significant funding.
Here is a sampling of academic research related to the C++ language in the 2006-2020 timeframe:
• Concepts: generic programming [Dehnert and Stepanov 2000], C++0x Concepts [Gregor et al.
2006], use patterns [Dos Reis and Stroustrup 2006], library design [Sutton and Stroustrup
2011].
• Theory and formalism: inheritance model [Wasserrab et al. 2006], templates and overloading
[Dos Reis and Stroustrup 2005a], template semantics [Siek and Taha 2006], object layout
[Ramananandro et al. 2011], construction and destruction [Ramananandro et al. 2012], a
representation for code manipulation [Dos Reis and Stroustrup 2009, 2011], resource model
[Stroustrup et al. 2015].
• Dynamic lookup: fast dynamic cast [Gibbs and Stroustrup 2006], pattern matching [Solodkyy
et al. 2012], multimethods [Pirkelbauer et al. 2010].
• Static Analysis: sound representation [Yang et al. 2012], practical experience [Bessey et al.
2010].
• Performance: code bloat [Bourdev and Järvi 2006, 2011], exception implementation [Renwick
et al. 2019].
• Language comparisons: generic programming [Garcia et al. 2007].
• Concurrent and parallel programming: memory model [Batty et al. 2013, 2012, 2011], HPX (a
general-purpose C++ runtime system for parallel and distributed applications of any scale
[Kaiser et al. 2009Sept.]), STAPL (An adaptive, generic parallel C++ library [Zandifar et al.
2014]), TBB (Intel’s task parallelism library [Reinders 2007]).
• Coroutines: Database optimization [Jonathan et al. 2018; Psaropoulos et al. 2017].
• Software engineering: Code organization and optimization [Garcia and Stroustrup 2015],
Constant expression evaluation [Dos Reis and Stroustrup 2010].
There seem to be opportunities for more academic research related to C++’s features and
techniques (e.g., exception handling, compile-time programming, and resource management) as
well as to the effectiveness of its use (e.g., static analysis or real-world code and empirical studies).
Few of the most active members of the C++ community would ever consider writing an academic
paper; writing books seems more popular (e.g., [Čukić 2018; Gottschling 2015; Meyers 2014; Stepanov
and McJones 2009; Vandevoorde et al. 2018; Williams 2018]).
10.4 Tools
In the early-to-mid 1990s, compared to other languages, C++ was doing reasonably well in the areas
of tools and programming environments for industrial use. For example, graphical user interfaces
and integrated software development environments were pioneered for C++. Later, the focus of
development and investment shifted to proprietary languages, such as Java (Sun), C# (Microsoft),
and Objective-C (Apple), and to simpler languages, such as C (GNU).
I see two major reasons:
• Funding: Organizations tend to prefer languages and tools that they can control and that give
then a differential advantage over competitors. From that perspective, the fact that C++ is
controlled by a formal standards committee emphasizing benefits for all is a disadvantage ś
a variant of the tragedy of the commons.
• Macros and textual definition: C++ did not have a simple, widely available internal representa-
tion to simplify tool building based on source code and the heavy use of macros ensured that
what a programmer saw wasn’t what the compiler analyzed. Like C, C++ is defined in terms
of sequences of characters, rather than constructs that directly represent abstractions and are
easier to manipulate. Together with Gabriel Dos Reis, I defined such a representation [Dos
Reis and Stroustrup 2009, 2011], but the character-oriented traditions in the C++ community
proved hard to overcome. It is hard to retrofit a regular structure on something not built with
that in mind.
Consequently, in the 2006-2020 time frame, C++ suffered badly in the area of support tools compared
to other languages. However, the situation also improved slightly by the emergence of
• Industrial-strength integrated software development environments: e.g., Microsoft’s Visual

Studio [Microsoft 2020; VStudio Wikipedia 2020] and JetBrain’s Clion [Clion Wikipedia 2020;
JetBrains 2020]. These environments support not just editing and debugging, but also various
forms of analysis and simple code transformation.
• Online compilers: e.g., Compiler Explorer [Godbolt 2016] and Wandbox [Wandbox 2016ś2020].
These systems allow the compilation of and sometimes execution of C++ programs from any
browser. They are used for experimentation, examination of code quality, and comparison of
different compilers and versions of compilers and libraries.
• GUI libraries and tools: e.g., Qt [Qt 1991ś2020], GTKmm [GTKmm 2005ś2020], and wxWidgets
[wxWidgets 1992ś2020]. Unfortunately, Qt relies on a meta-object protocol (MOP), so that
Qt programs are not just ISO C++. Static reflection (ğ9.6.2) should eventually allow us to
solve that problem. The problem for the C++ community is not that there are no good GUI
libraries, but that there are so many that choosing one is hard.
• Analyzers: e.g., coverity [Coverity 2002ś2020], Visual Studio’s analyser for the C++ Core
Guidelines (ğ10.6), and Clang tidy [Clang Tidy 2007ś2020].
• Compiler tool support: e.g., the LLVM compiler back-end infrastructure simplifying code gen-
eration and code analysis [LLVM 2003ś2020]. This provided a boon to many new languages,
in addition to C++ itself.
• Build systems: e.g., build2 [Build2 2014ś2020] and Cmake [Cmake 2000ś2020], and GNUmake
[GNUmake 2006ś2020]. Again, in the absence of a standard, choosing one is hard.
• Package managers: e.g., Conan [Conan 2016ś2020] and vcpkg [vcpkg 2016ś2020].
• Run-time environments: e.g., WebAssembly: a system for compiling ISO C+ into bytecodes
for deployment in browsers [WebAssembly 2017ś2020].
• Run-time compilation, JIT-ing, and linking: e.g., Cling [Cling 2014ś2020; Naumann 2012;
Naumann et al. 2010] and RC++ [RC++ 2010ś2020].
What is listed above are just examples. As usual, the problem for the C++ users is the number of
alternatives: [RC++ 2010ś2020] lists 26 systems for generating code at compile time and there are
dozens of package managers. What is needed is some form of standardization.
As of 2020, tools are still not a strength of C++, but progress is being made over a broad front.
10.5 Programming Styles

A key driver of C++’s evolution is that optimal solutions to most real-world problems involve a
combination of techniques. Naturally, this offends everybody who claims to have a single simple
best solution (a łprogramming paradigmž), but supporting a variety of styles has always been
a fundamental strength of C++. Consider the łdraw all shapesž example that has been used to
illustrate object-oriented programming since the early days of Simula (where the drawing device
was a wet-ink plotter). In C++20, we might write:
void draw_all ( range auto & seq )
{
for ( Shape & s : seq )
s. draw ();
}
What programming paradigm is that code?

• It is clearly object-oriented programming: it uses a virtual function and a class hierarchy.
• It is clearly generic programming: it uses a template (by parameterizing with the range
concept, we get a template).
• It is clearly ordinary imperative programming: it uses a for-loop and defines a function to be
called with the conventional f(x) syntax.
I could elaborate this example: A Shape usually has mutable state. I could use a lambda. I could
call a C function. I could further constrain the argument with a Drawable concept. For a variety
of definitions of łbetterž, a suitable mix of techniques is a better solution than any I could come up
with using a single paradigm.
The idea behind C++’s support of several programming styles (łparadigmsž if you must) is not
that you can choose to program in one favorite style, but that you can use them in combination to
express better solutions than you could in a single style.
10.5.1 Generic Programming. In 2006, much C++ code was still a mixture of object-oriented styles
and C-style programming. Naturally, there is still a lot of that in 2020. However, with C++98, the
STL style of generic programming (often referred to as GP) became widely known and slowly user
code started using GP beyond simple applications of the standard library. The improved support for
GP in C++11 opened the floodgates for its use in production code. However, the lack of concepts
(ğ6) in C++17 is still a drag on the use of generic programming in C++.
Essentially all experts read Alex Stepanov’s The Elements of Programming (often referred to as
EoP) [Stepanov and McJones 2009] and were influenced by it.
Generic programming using templates is the backbone of the C++ standard library: containers,
ranges (ğ9.3.5), algorithms, iostreams, file system (ğ8.6), random numbers (ğ4.6), threads (ğ4.1.2)
(ğ9.4), locking (ğ4.1.2) (ğ8.4), time (ğ4.6) (ğ9.3.6), strings, regular expressions (ğ4.6), and formats
(ğ9.3.7).
10.5.2 Metaprogramming. Metaprogramming in C++ grew out of generic programming in the
sense that both rely on templates. Its roots go back to the early days of templates in C++ when
people discovered that templates were Turing complete [Vandevoorde and Josuttis 2002; Veldhuizen
2003] and offered compile-time pure functional programming in a somewhat useful form.
Template metaprogramming (often referred to as TMP) is often very ugly. Sometimes, that
ugliness is hidden through the use of macros, creating further problems. It is a testimony to TMP’s
utility that it became almost universal. For example, you can’t implement the C++14 standard library
without metaprogramming. Many techniques and experiments were pre-2006, but it was C++11
with better compilers, variadic templates (ğ4.3.2), and lambdas (ğ4.3.1) that gave TMP the boost
needed for mainstream use. The C++ standard library also added support, such as the compile-time
selection template, conditional, type traits allowing code to depend on properties of types, such
as łcan type X safely be bitwise copied?ž (ğ4.5.1), and enable_if (ğ4.5.1). For example:
conditional <( sizeof ( int ) <4) , double , int > x; // if ints are small use double
Computing types to precisely mirror requirements is arguably the essential part of TMP. We can
also calculate values:
template < unsigned n >
struct fac {
enum { val = n * fac <n -1 >:: val };
};
template <>
struct fac <0 > { // specialization for 0: fac <0 > is 1
enum { val = 1 };
};
constexpr int fac7 = fac <7 >:: val ; // 5040
Note the key role of template specialization; it is essential in most TMP. It has been used to
compute non-trivial values and to express control flows (e.g., to compute decisions tables at compile
time and unroll loops). Specialization is a largely underappreciated feature of C++98 [Stroustrup
2007].
In clever libraries and also in real-world code, primitives like enable_if have become the basis of
programs of many hundreds or even thousands of lines. Early examples of TMP included a complete
compile-time Lisp interpreter [Czarnecki and Eisenecker 2000]. Debugging is extremely hard,
maintenance of such code is horrendous, and I have seen a couple of hundred lines of (admittedly
clever) TMP lead to compile times in minutes and failure to compile by running out of memory
on a 30GB computer. The compiler error messages from even simple errors can run into many
thousands of lines. And yet TMP is widely used. Sane programmers have found TMP, warts and
all, preferable to alternatives. I have seen code generated from TMP that was better than I would
expect a competent human to write in assembler.
Thus, the problem becomes to better serve such needs. I worry when people start to consider
code like fac<> normal. That is not a good way to express an ordinary numerical algorithm.
Concepts (ğ6) and compile-time evaluated functions (constexpr (ğ4.2.7)) can dramatically simplify
metaprogramming. For example:
constexpr int fac ( int n)
{
int r = 1;
while (n >1) r *=n - -;
return r;
};
constexpr int fac7 = fac (7); // 5040
This example illustrates that when you want a value, a function is the best way to compute it,
even ś and especially ś at compile time. Traditional template metaprogramming is best reserved
for computing new types and control structures.
Jaakko Järvi’s boost::lambda [Järvi and Powell 2002; Järvi et al. 2003a] was an early use of TMP
that helped convince people that lambdas were useful and also that they needed direct language
support.
The Boost metaprogramming library, boost::MPL [Gurtovoy and Abrahams 2002ś2020], offered
the best and worst of traditional TMP. A more modern library, boost::Hana [Boost Hana 2015ś
2020], uses constexpr functions. WG21’s SG7 (ğ3.2) is trying to develop a better and standard
metaprogramming system that also includes compile-time reflection (ğ9.6.2).
10.6 Coding Guidelines

My ultimate aim for C++ is a language that is
• Significantly simpler to use and to learn than C or current C++
• Completely type safe ś no implicit type violations and no dangling pointers
• Completely resource safe ś no leaks and no need for a garbage collector
• Relatively simple to build tools for ś no macros
• As fast or faster than current C++ ś the zero-overhead principle
• Predictable performance ś suitable for embedded systems
• Not less expressive than current C++ ś handle hardware well
This is not all that different from the design aims articulated in The Design and Evolution of C++
[Stroustrup 1994] and earlier. Obviously, it’s a tall order and incompatible with most uses of older
C and C++.
From the earliest days of łC with Classesž, people have suggested creating safe subsets of the
language and compiler switches to enforce such safety. However, those suggestions failed for one
of many reasons:
• Not enough people could agree on a definition of łsafe.ž
• The unsafe features (for every definition of łunsafež) are the ones needed to build the basic
safe abstractions.
• The safe subset is insufficiently expressive.
• The safe subset is inefficient.
The second reason implies that you cannot define a safe C++ simply by banning unsafe features.
łPerfection through restrictionž approaches to programming language design at best works in very
limited situations. You need to consider the context and nature of the use of features that in general
are unsafe but have safe uses. Furthermore, the standard cannot abandon backwards compatibility
(ğ1.1), so we need a different approach.
From the start, C++ adopted a different philosophy [Stroustrup 1994]:
It is more important to enable good programming than to prevent errors.
That implies that we need guidelines for łgood usež, rather than language rules. However, to be
useful at an industrial scale, guidelines must be enforceable by tools. For example, from the earliest
days of C and C++, we have known of the problems with dangling pointers. For example:
int * p = new int {7 ,9 ,11 ,13};

// ...
delete p; // delete the array pointed to by p
// p now doesn 't point to a valid object ; it " dangles "
// ...
*p = 7; // likely disaster
Many programmers have developed techniques to prevent pointers from dangling. However,
dangling pointers are still a major problem in most large code bases and security issues are much
more critical than in the past. Some dangling pointers can be exploited as a security hole.
10.6.1 General Approach. In 2004, I helped develop a set of coding guidelines for flight control
software [Lockheed Martin Corporation 2005] that approximated my ideals of safety, flexibility, and
performance. In 2014, I started writing a set of coding guidelines to address this for use in a broad
range of applications. This was in part a response to loud demands for practical guidelines for using
C++11 well and part in horror over seeing what some people considered good C++11. Talking to
people, I soon discovered the obvious: I wasn’t the only one thinking and working along these
lines. So, some experienced C++ programmers, tool builders, and library builders joined forces and
started the C++ Core Guidelines project [Stroustrup and Sutter 2014ś2020] with contributors from a
broad spectrum of the C++ community. The project is open-source (MIT license) and the list of
contributors can be found on GitHub. Early on, contributors from Morgan Stanley (notably me),
Microsoft (notably Herb Sutter, Gabriel Dos Reis, and Neil Macintosh), Red Hat (notably Jonathan
Wakely), CERN, Facebook, and Google were prominent.
The Core Guidelines are by no means the only C++ guidelines project, but they are the most
prominent and most ambitious. They have the explicit and articulated aim of dramatically improving
the quality of C++ code. For example, the ideals and basic model for complete type and resource
safety was articulated early on in a paper by Bjarne Stroustrup, Herb Sutter, and Gabriel Dos Reis
[Stroustrup et al. 2015].
To reach the ambitions goals, we apply a łcocktailž of approaches applied in concert:
• Rules: A large set of rules, aimed at type-safe and resource-safe use of C++, recommending
known effective practices, and banning known sources of errors and inefficiencies.
• Foundation libraries: A set of library components to allow programmers to write efficient
low-level programs without using known error-prone features and in general to provide a
higher-level basis for programming. Most components are from the standard library and
some from a small Guidelines Support Library (GSL) written in ISO standard C++.
• Static analysis: Tools to detect violations and enforce key parts of the guidelines.
Each of these techniques has a long history and each cannot by itself handle the problems at
an industrial scale. For example, I am a great fan of static analysis, but the kinds of analysis I am
most interested in (e.g., the elimination of dangling pointers) cannot be solved if a programmer
can write arbitrarily complex code in a separately compiled program potentially using dynamic
linking. Here łcannotž means łtheoretically impossible in generalž as well as łtoo computationally
expensive for industrial-scale programs.ž
The basic approach is not simple restriction, but what I called the subset-of-superset approach
or SELL [Stroustrup 2005]:
• First, extend the language with library facilities to makes a solid base for use.
• Then, subset by removing unsafe, error-prone, and overly-costly facilities.
For libraries, we primarily rely on parts of the standard library, such as variant (ğ8.3) and vector.
The small Guidelines Support Library (GSL) offer support for type-safe access, such as span for
range checked access to a contiguous sequence of elements of a given type (ğ9.3.8). The idea is to
eventually eliminate the GSL by getting it absorbed into the ISO standard library. For example,
span was added to the C++20 standard library and the feeble contract support in the GSL should
be replaced by proper contracts if and when they become available (ğ9.6.1).
10.6.2 Static Analysis. To scale, the static analysis is completely local (one function or one class
at a time). The hardest problems relate to lifetime of objects. RAII is essential: manual resource
management has repeatedly showed itself seriously error-prone in the many languages in which it
is used. Furthermore, there are plenty of programs łout therež that use pointers and iterators in
principled ways. Such uses must be accepted. To make a program safe is easy, you simply prohibit
everything. However, maintaining both the expressiveness of C++ and its performance are among
the aims of the Core Guidelines, so we can’t gain safety simply through restriction. The aim is a
better C++, not a slow or neutered subset.
These guidelines can help focus education on the more effective aspects of C++ by articulating
principles, making good practices explicit, and mechanizing checking for known problems. They
also help relieve pressures on the language to directly accommodate fashions.
For object lifetimes, there are two major needs:
• Never point to an object that has gone out of scope.
• Never access an invalidated object.
Consider an example from łthe basic modelž paper [Stroustrup et al. 2015]):
int glob = 666;
int * f( int * p)
{
int x = 4; // local variable
// ...
return &x; // No ! would point to destroyed stack frame
// ...
return & glob ; // OK : points to something that " lives forever "
// ...
return new int {7}; // OK ( sort of : doesn 't dangle ,
// but returns an owner as an int *)
// ...
return p; // OK : came from caller
}
We can return a pointer that points to something known to outlive the function (e.g., came to
the function as an argument) but not a pointer to a local. In a program following the guidelines, it
is guaranteed that a pointer passes as an argument points to something or be the nullptr).
To avoid leaks, łnaked newsž as in the example above are eliminated through the use of resource
handles (RAII) or ownership annotations.
A pointer can be invalidated if the object it points to is reallocated. For example:
vector < int > v = { 1 ,2 ,3 };
int * p = &v [2];
v. push_back (4); // v ' s elements may be relocated
*p = 5; // bad : p might have been invalidated
int * q = &v [2];

v. clear (); // all v 's elements are deleted
*q = 7; // bad : q is invalidated
Checking for invalidation is even harder than checking for simple dangling pointers because it
is hard to be sure which functions move objects and whether that is to be considered invalidation
(the pointer p still points to something, but conceptually to a different element). It is not yet clear
if invalidation checking can be completely handled without annotation or non-local state in the
static analyzer. In initial implementations, every function manipulating an object as a non-const
is assumed to invalidate, but that is too conservative, leading to too many false positives. The
initial detailed specification of lifetime checking was written by Herb Sutter [Sutter 2019] and
implemented by his colleagues at Microsoft.
Range checking and checking for nullptr is done with library support (GSL). Static analysis is
then used to ensure that the libraries are used consistently.
The initial implementation of the static analysis ideas was by Neil Macintosh and is currently
deployed as part of Microsoft Visual Studio. Some rules are checked as part of Clang and HSR’s
Cevelop (Eclipse plugins) [Cevelop 2014ś2020]. Some courses and books are incorporating rules
(e.g., [Stroustrup 2018f]).
The Core Guidelines are designed for gradual and selective adoption. As a result, we see wide-
spread adoption of some guidelines in industry and education, but little complete adoption. For
complete adoption good tool support is necessary.
11 RETROSPECTIVE
The ultimate aim of a programming language design is to improve the way programmers think
and work when delivering useful applications. There are languages deemed łjust experimentalž
but as soon as a language is used for work that is not related to the language itself, the language
designers acquire a responsibility to their users. Correctness, appropriateness, stability, and adequate
performance become important issues. For C++, that happened in 1979 after only 6 months. C++
has thrived for 40 years. Why and how?
My previous HOPL papers [Stroustrup 1993, 2007] offer answers from the perspective of 1991
and 2006. Apart from the language features and library components, what has changed since then
is mostly the role and impact of the standards committee (ğ3).
Here, I consider
• ğ11.1: The C++ model
• ğ11.2: Technical successes
• ğ11.3: Areas that need work
• ğ11.4: Lessons learned
• ğ11.5: The future
11.1 The C++ Model

C++ emerged as a major ś and in some areas dominant ś language for demanding applications.
It did so without serious commercial backing and with no marketing. Many modern languages
copied features and ideas from it (ğ2.4). The key language-technical areas of contribution are:
• A static type system with equal support for built-in types and user-defined types (ğ2.1)
• Value and reference semantics (ğ4.2.3)
• Systematic and general resource management (RAII) (ğ2.2)
• Support for efficient object-oriented programming (ğ2.1)
• Support for flexible and efficient generic programming (ğ10.5.1)
• Support for compile-time programming (ğ4.2.7)
• Direct use of machine and operating system resources (ğ1)
• Concurrency support through libraries (often implemented using intrinsics) (ğ4.1) (ğ9.4)
C++ offers a different ś and for many application areas better ś model for software than the
currently dominant łmanagedž model relying on garbage collectors and extensive run-time support
typified by languages such as Java, C#, Python, and JavaScript (ğ2.3). By łbetterž I mean łsimpler
to write, more likely to be correct, more maintainable, using less memory, using less energy, and
faster.ž
The areas of contribution are mutually supportive. For example

• Reference semantics (e.g., pointers and smart pointers) enables the efficient implementation
of advanced types with value semantics (e.g., jthread and vector).
• Uniform rules for built-in and user-defined types simplify generic programming (built-in
types are not special cases).
• Compile-time programming makes a range of abstraction techniques affordable for effective
use of hardware.
• RAII allows use of user-defined types without taking specific actions to support their imple-
mentations’ use of resources (including non-memory resources).
11.2 Technical Successes

The fundamental reason for C++’s success is simple: It fills an important łnichež in the programming
landscape:
Applications that need to use hardware efficiently and to manage significant complexity.
If you can afford to łwastež 25% or 99% of your hardware capacity, there is no shortage of
programming languages and environments to choose from and if your low-level needs represent
only a few thousand lines of low-level code, C or assembler will serve. For 40 years, C++’s łnichež
has been sufficient to keep its community growing.
Here is a modern (2014) summary of C++:
• Direct map to hardware
ś of instructions and fundamental data types
ś Initially from C
• Zero-overhead abstraction
ś Classes with constructors and destructors, inheritance, generic programming, function
objects
ś Initially from Simula (where it wasn’t zero-overhead)
Simula pioneered abstraction mechanisms and a flexible type system, but they came with heavy
costs in run-time and space. Compared to the 1995 description of C++ (ğ2.1), the focus has shifted
from programming techniques to problem areas. That’s more a difference in explanation style and
people’s interest than in language design. Both summaries are accurate now and were then.
Building on the previous decades, the key technical advances of the 2000s include:
• The memory model (ğ4.1.1)
• Type-safe concurrency support: threads and locks (ğ4.1.2), parallel algorithms (ğ8.5), joining
threads (ğ9.4)
• Type deduction: auto (ğ4.2.1), concepts (ğ6), template argument deduction (ğ8.1), variadic
templates (ğ4.3.2)
• Simplification of use: auto (ğ4.2.1), range-for, (ğ4.2.2), parallel algorithms(ğ8.5), ranges(ğ9.3.5),
lambdas (ğ4.3.1)
• Move semantics (ğ4.2.3)
• Compile-time programming: constexpr (ğ4.2.7), compile-time loops (ğ5.5), guaranteed compile-
time evaluation and containers (ğ9.3.3), metaprogramming (ğ10.5.2)
• Generic programming: the STL (ğ10.5.1), concepts (ğ6), user-defined types as template pa-
rameters (ğ9.3.3), lambdas (ğ4.3.1)
• Metaprogramming (ğ10.5.2)
They all relate to the zero-overhead principle, but the last two are a bit surprising because of
C++’s incomplete support for them in the 2006-2020 timeframe.
None of this would have mattered had C++ broken into incompatible dialects or become some-
thing you couldn’t rely on for the long term:
• Stability and compatibility are essential (ğ1.1) (ğ11.4)
The new features (from C++11 onwards) led to improvements in the standard library (e.g.,
unique_ptr, chrono, format, and scoped_lock) and in many other libraries.
The purpose of C++ is to be a tool for building applications, so the many great applications of
C++, such as those mentioned in (ğ2.3) and (ğ10.1), are the real successes of C++.
11.3 Areas That Need Work

No language is perfect for everybody and everything. Nobody knows this better than people who
know several languages, seriously use a language, and try to support it. It is rarely simple ignorance
that stands in the way of progress. Rather, the major obstacles to significant improvement are lack
of direction, lack of development resources, and fear of breaking existing code.
C++ suffers from having been born before modern IDEs, build systems, GUI systems, and Unicode.
I expect C++ to slowly catch up. For example:
• Tooling: Starting from C with a semantics specified in terms of characters and lexical tokens
and source code organized with #includes and macros have been major barriers to effec-
tive tool building. Modules should help (ğ9.3.1) and it is possible to devise a sane internal
representation for C++ [Dos Reis and Stroustrup 2009, 2011].
• Education: The C++ taught today is still mostly outdated and backwards looking (ğ2.3). The
Core Guidelines (ğ10.6) is one approach to modernizing practice. WG21’s education study
group (ğ3.2) and the many education-oriented conference presentations (ğ10.3) show that
the problems are appreciated and being addressed.
• Packaging and distribution: C++ was born before composing software out of independently
developed and maintained parts was common. Today, there are build systems and package
managers for C++. However, none are standard, some are hard to use for simple tasks, and
others don’t generalize to cope with the massive systems built using C++. My 2017 CppCon
keynote challenged the C++ community to address this [Stroustrup 2017c] and I think we
are seeing progress. In addition, the C++ community suffers from a lack of a standard place
to find useful libraries. Boost [Boost 1998ś2020] was an effort to address this and GitHub is
emerging as a common repository, but there is a long way to go before a relative novice can
find, download, install, and run a couple of major libraries.
• Character sets and graphics: The C++ language and standard library depend on ASCII, yet
most applications use some form of Unicode. WG21 now has a study group to find a way
to standardize Unicode support (ğ3.2). The lack of standard Graphics and GUI is a harder
problem.
• Clean-up of old messes: This is unpleasantly hard. For example, we know that the implicit
narrowing conversions among the built-in types cause endless problems (ğ9.3.8), but there are
hundreds of billions of lines of C++ code that depend on those conversions in unpredictable
ways. Attempts to improve by adding łmore modernł features to replace the old ones easily
fall prey to the N+1 problem (ğ4.2.5). Improved tooling (e.g., static program analysis and
program transformation) offers hope.
The challenges for a large language community are diverse and not amenable to a single simple
solution. It is not just an issue of syntax, type theory, or basic language design. Some problems are
commercial. The range of skills needed for success at an industrial scale is daunting. Time will tell
whether the C++ community will get to grips with all of this, and more. I’m modestly optimistic
because there are initiatives in all areas (ğ3.2).
11.4 Lessons Learned

C++ is controlled by a large committee with a diverse and changing membership (ğ3.2). So in
addition to technical issues, we must consider what works in the process of evolving the language:
• Problem driven: C++’s development should be driven by needs of specific real-world problems.
• Simple: C++ should be grown by generalization from simple, efficient, easy-to-use solutions.
• Efficient: The C++ language and standard library should obey the zero-overhead principle.
• Stable: Don’t break my code!
The development of most (all?) of the most successful parts of C++ obeyed those łrules of thumb.ž
They naturally limit the scope of the language, but that’s good. C++ is not meant to be all things to
all people. Furthermore, these principles forces C++ to grow relatively slowly based on real-world
challenges and enable it to benefit from feedback. See also the other łrules of thumbž in The Design
and Evolution of C++ [Stroustrup 1994] and my HOPL2 paper [Stroustrup 1993]. There is continuity.
To contrast, facilities designed without a clear focus on the problems of the larger community
usually fail:
• Experts only: A facility has to serve all the experts’ needs from the start.
• Imitative: We need this facility because it is popular in łlanguage X.ž
• Theoretical: Language theory says that a language must have this feature.
• Revolutionary: This feature is so important that we must break compatibility or deprecate
łthe bad old waysž of doing things.
I conclude that setting direction and expectations early is essential. Later, there will be too many
people with too diverse opinions to get agreement on a coherent set of ideas.
Given a direction and a set of principles, a language can grow based on feedback, user experience,
experiments, and theory as tools. This is good engineering as opposed to unprincipled pragmatism
or dogmatic idealism.
The C++ standards committee’s charter focuses almost exclusively on language and library
design. That’s limiting. Important topics such as dynamic linking, build systems, and static analysis
have been mostly ignored. That’s a mistake. Tools are a huge part of a software developer’s world
and it would be nice if they were not peripheral to language design.
Enthusiasm for a diverse set of idea is dangerous. In a 2018 paper [Stroustrup 2018d], I listed 51
recent proposals:
łI list papers that I think have the potential to significantly change the way we write code,
so that each has significant implications on teaching, maintenance, and coding guidelines.
Many also have implications for implementations.
Individually, many (most?) proposals make sense. Together they are insanity to the point
of endangering the future of C++.ž
That paper was entitled Remember the Vasa!. The Vasa was a magnificent 17th century Swedish
battleship that because of late additions to its design and insufficient testing sank in Stockholm
harbor on its maiden voyage. In the 1990s, the committee often reminded itself of the Vasa, but in
the 2010s that lesson seemed to have been forgotten.
In an attempt to put organizational constraints on the committee process, the DG proposed łThe
C++ Programmers’ Bill of Rightsž [Dawes et al. 2018]:
(1) Compile-time stability: Every significant change in behavior in a new version of the
standard is detectable by a compiler for the previous version.
(2) Link-time stability: ABI breakage is avoided except in rare cases, which will be well
documented and supported by a written rationale.
(3) Compiler performance stability: Changes will not imply significant added compile-
time costs for existing code.
(4) Run-time Performance stability: Changes will not imply significant added run-time
costs to existing code.
(5) Progress: Every revision of the standard will offer improved support for some signifi-
cant programming activity or community.
(6) Simplicity: Every revision of the standard will offer simplification for some significant
programming activity.
(7) Timeliness: Every revision of the standard will be shipped on time according to a
published schedule.
The next decades will show how this works out.
11.5 The Future

For the near term, C++20 will be as great a boon to the C++ community as C++11 was. At the
February 2020 meeting in Prague where the committee finalized C++20, it also voted in favor of
Ville Voutilainen’s łbold plan for C++23ž [Voutilainen 2019b]:
łwork towards having the following things in C++23: ž
• Library support for coroutines (ğ9.3.2)
• A modular standard library (ğ9.3.1)
• The general asynchronous computing model (executors) (ğ8.8.1)
• Networking (ğ8.8.1)
Note the focus is on libraries. łAlso make progress on:ž
• Static reflection facilities (ğ9.6.2)
• Functional-programming style pattern matching (ğ8.2)
• Contracts (ğ9.6.1)
Given that work on these topics is quite advanced, the odds are that the committee will accomplish
most. What else that large group of enthusiasts can develop and agree upon is less predictable. For
the next few years, the direction group (of which I am a member) mentions some promising areas
for further work [Hinnant et al. 2020]:
• Improved Unicode support
• Support for simple graphics and simple user interaction
• Support for new kinds of hardware
• Exploration of improved styles and implementations of error-handling
Outside the committee, I expect to see significant progress on build systems, package management,
and static analysis (ğ10.4).
Beyond that ś five, ten, or more years into the future ś my crystal ball gets a bit cloudy. For
that time scale, we need to look at fundamentals rather than specific language features. I hope the
standards committee will heed the lessons learned (ğ11.4) and focus on the fundamentals (ğ11.1):
• Pursue the goal of a completely resource-safe and type-safe C++
• Support a wide variety of hardware well
• Maintain C++’s record of stability (compatibility)
Maintaining stability requires a focus on compatibility as well as resisting the urge to try
to radically improve C++ by adding a multitude of łperfectž features to replace imperfect or
unfashionable older ways of doing things. New features invariably cause surprises (some pleasant,
some not so pleasant) and the older features will not simply disappear. Remember the Vasa!
[Stroustrup 2018d] (ğ11.4). Often libraries, guidelines, and tools are superior alternatives to language
changes.
Hardware isn’t getting any faster for single threaded computation, so the premium on efficiency
will remain and the pressure to effectively support various forms of concurrency and parallelism
will increase (ğ2.3). Specialized hardware will proliferate (e.g., a variety of memory architectures
and special-purpose processors); that will benefit languages, such as C++, that can take advantage
of it. The only thing that increases faster than hardware performance is human expectation.
Systems are getting increasingly complex, so affordable abstraction mechanisms will also increase
in importance. For systems relying on real-time interaction, predictable performance is essential
(e.g., many real-time systems ban the use of free store (dynamic memory)).
Security concerns will only increase in importance as our dependence on computerized systems
increases and the number of sophisticated hackers grow. To defend, I’d bet on hardware protection
and on more structured systems supporting better static analysis, rather than endless ad-hoc
run-time checks and low-level code.
Interoperability among languages and systems will remain essential; few major systems will be
written in a single language.
As systems become more complex and the requirements for dependability increase, the need
for quality of design and coding grow dramatically. I think C++ is well prepared for this and the
plans for C++23 are to strengthen it further. However, language features alone are not sufficient to
satisfy future demands. We need guidelines supported by tools to ensure effective use (ğ10.6). In
particular, we need to ensure complete type safety and resource safety. This has to be reflected
in education. To thrive, C++ needs better educational materials for novices and some to help
experienced programmers master modern C++. Presentations of clever tricks and advanced uses
are not sufficient and can harm the language by reinforcing its reputation of complexity.
For many reasons, we need to simplify the majority of C++ use. C++ has evolved to make
that possible and I expect this trend will continue (ğ4.2). Improved optimizers ś capable of taking
advantage of the type system and abstractions used in the code ś make a difference. Over the last few
years, this has dramatically changed the way I approach optimizing code: I start by chucking away
the clever and complicated stuff. That’s where the bugs hide and if I have trouble understanding
what’s going on, so will the compiler and optimizer. I find that this approach usually gives me
modest to spectacular performance improvements, as well as simplifying future maintenance. Only
if this doesn’t give me the performance I want, do I resort to advanced (aka complicated) data
structures and algorithms. This is a triumph of the design of the C++ abstraction mechanisms.
I look forward to seeing many more exciting applications built with C++ and to seeing new
programming idioms and design techniques developed.
I hope other languages learn from C++’s successes. It would be sad if the lessons learned from
C++’s evolution were limited to the C++ community. I hope and expect to see key aspects of C++’s
model in other languages and systems; that would be a true measure of success. To a limited extent,
this has already happened (ğ2.4).
ACKNOWLEDGMENTS
I am painfully aware that
• This paper is far too long.
• The description of most technical topics leaves out much that could be considered essential.
Often, the work of many people over many months spread over many years is reduced to a
single page or even a single sentence. In particular, I shortchanged the immensely important
topic of concurrency; it deserves a long, detailed paper of its own.
Thanks to the millions of programmers who made C++ a success. Their applications are critical
parts of our world.
Thanks to reviewers of drafts of this paper, including Al Aho, A. Bratterud, Shigeru Chiba, J.
Daniel Garcia, Brent Hailpern, Howard Hinnant, Roger Orr, Aaron Satlow, Yannis Smaragdakis,
David Vandevoorde, J.C. Van Winkel, and Michael Wong. This paper relies strongly on reviewers
for completeness and accuracy. Errors are of course my own.
Thanks to Guy Steele who helped me navigate the mysteries of LaTex and BibTex to bring the
references into the form required by the ACM.
Thanks to all who worked so hard on the standards. The names of many more than I could
mention can be found as authors of WG21 papers and in the acknowledgement sections of those
papers. The many ‘P’ and ‘N’ numbered papers that I reference and quote can all be found at
open-std.org/jtc1/sc22/wg21/docs/papers/. Without those papers, this paper would have had to
rely far too much on my memory.
REFERENCES
David Abrahams, Rani Sharoni, and Doug Gregor. 2010. Allowing Move Constructors to Throw (Rev. 1). ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper N3050. 12 March 2010. http://www.open-std.org/jtc1/sc22/wg21/
docs/papers/2010/n3050.html (also at Internet Archive 4 Sept. 2019 02:05:46).
ADAS Wikipedia 2020. ADAS: Advanced driver-assistance systems. https://en.wikipedia.org/wiki/Advanced_driver-
assistance_systems (also at Internet Archive 9 May 2019 04:16:15).
James Adcock. 1990. Request for Consideration ś Overloadable Unary operator.(). 8 Oct. 1990. http://www.open-
std.org/jtc1/sc22/wg21/docs/papers/1990/WG21%201990/X3J16_90%20WG21%20Request%20for%20Consideration%20-
%20Overloadable%20Unary%20operator.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
Andrei Alexandrescu, Hans Boehm, Kevlin Henney, Doug Lea, and Bill Pugh. 2004. Memory model for multithreaded C++.
ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N1680. 10 Sept. 2004. http://www.open-std.org/jtc1/sc22/
wg21/docs/papers/2004/n1680.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
Matthew Austern. 2001. A Proposal to Add Hashtables to the Standard Library. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper N1326. 17 Oct. 2001. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2001/n1326.html (also at
Internet Archive 4 Sept. 2019 02:05:46).
Autosar 2020. AUTOSAR: The standardized software framework for intelligent mobility (website). https://www.autosar.org/
(also at Internet Archive 23 Aug. 2019 18:35:20).
AUTOSAR (AUTomotive Open System ARchitecture) is a worldwide development partnership of vehicle manufacturers,
suppliers, service providers and companies from the automotive electronics, semiconductor and software industry.
Autosar Wikipedia 2020. AUTOSAR (AUTomotive Open System ARchitecture). https://en.wikipedia.org/wiki/AUTOSAR
(also at Internet Archive 7 Aug. 2019 19:39:43).
Lewis Baker. 2019. Coroutines TS Simplifications. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1477R1.
12 Feb. 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1477r1.pdf (also at Internet Archive 4 Sept.
2019 02:05:46).
D. W. Barron, J. N. Buxton, D. F. Hartley, E. Nixon, and C. Strachey. 1963. The Main Features of CPL. The Computer Journal
6, 2 (Aug.), 134ś143. 0010-4620 https://doi.org/10.1093/comjnl/6.2.134 Also at https://academic.oup.com/comjnl/article-
pdf/6/2/134/1041447/6-2-134.pdf
JF Bastien. 2018. Signed Integers are Two’s Complement. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
P0907R0. 9 Feb. 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0907r0.html (also at Internet Archive
4 Sept. 2019 02:05:46).
Mark Batty, Mike Dodds, and Alexey Gotsman. 2013. Library Abstraction for C/C++ Concurrency. In Proceedings of the
40th Annual ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages (Rome, Italy, Jan.) (POPL ’13).
Association for Computing Machinery, New York, NY, USA, 235ś248. 978-1450318327 https://doi.org/10.1145/2429069.
2429099
Mark Batty, Kayvan Memarian, Scott Owens, Susmit Sarkar, and Peter Sewell. 2012. Clarifying and Compiling C/C++
Concurrency: From C++11 to POWER. In Proceedings of the 39th Annual ACM SIGPLAN-SIGACT Symposium on Principles
of Programming Languages (Philadelphia, PA, USA, Jan.) (POPL ’12). Association for Computing Machinery, New York,
NY, USA, 509ś520. 978-1450310833 https://doi.org/10.1145/2103656.2103717
M. Batty, S. Owens, S. Sarkar, P. Sewell, and T. Weber. 2010. Mathematizing C++ Concurrency: The Post-Rapperswil Model.
ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N3132. 23 Aug. 2010. http://www.open-std.org/jtc1/sc22/

Mark Batty, Scott Owens, Susmit Sarkar, Peter Sewell, and Tjark Weber. 2011. Mathematizing C++ Concurrency. In
Proceedings of the 38th Annual ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages (Austin,
Texas, USA, Jan.) (POPL ’11). Association for Computing Machinery, New York, NY, USA, 55ś66. 978-1450304900
https://doi.org/10.1145/1926385.1926394
Pete Becker. 2004. A Multi-threading Library for Standard C++. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
N1682. 10 Sept. 2004. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2004/n1682.html (also at Internet Archive
4 Sept. 2019 02:05:46).
Pete Becker (Ed.). 2009. Working Draft, Standard for Programming Language C++. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper N2914. 2 June 2009. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2009/n2914.pdf (also at
Pete Becker (Ed.). 2011. ISO/IEC 14882:2011: Information Technology Ð Programming languages Ð C++. ISO (International
Organization for Standardization), Geneva, Switzerland (Sept.). 1338 book pages. https://www.iso.org/standard/50372.
html Status: withdrawn.
Joshua Berne. 2019. łAxiomž is a False Friend. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1672R0. 16 June
2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1672r0.pdf (also at Internet Archive 4 Sept. 2019
02:05:46).
Joshua Berne, Nathan Burgers, Hyman Rosen, and John Lakos. 2018. Contract Checking in C++: A (long-term) Road Map.
ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1332R0. 26 Nov. 2018. http://www.open-std.org/jtc1/sc22/
wg21/docs/papers/2018/p1332r0.txt (also at Internet Archive 4 Sept. 2019 02:05:46).
Joshua Berne and John Lakos. 2018a. Assigning Concrete Semantics to Contract-Checking Levels at Compile Time. ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper P1333R0. 26 Nov. 2018. http://www.open-std.org/jtc1/sc22/wg21/
docs/papers/2018/p1333r0.txt (also at Internet Archive 4 Sept. 2019 02:05:46).
Joshua Berne and John Lakos. 2018b. Specifying Concrete Semantics Directly in Contract-Checking Statements. ISO/IEC
docs/papers/2018/p1334r0.txt (also at Internet Archive 4 Sept. 2019 02:05:46).
Joshua Berne and John Lakos. 2019. Contracts That Work. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
P1429R2. 16 June 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1429r2.pdf (also at Internet Archive
4 Sept. 2019 02:05:46).
Al Bessey, Ken Block, Ben Chelf, Andy Chou, Bryan Fulton, Seth Hallem, Charles Henri-Gros, Asya Kamsky, Scott McPeak,
and Dawson Engler. 2010. A Few Billion Lines of Code Later: Using Static Analysis to Find Bugs in the Real World.
Commun. ACM 53, 2 (Feb.), 66ś75. 0001-0782 https://doi.org/10.1145/1646353.1646374
Peter Bindels, Ben Craig, Steve Downey, Rene Rivera, Tom Honermann, Corentin Jabot, and Stephen Kelly. 2018. Concerns
about module toolability. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1427R0. 20 Nov. 2018. http:
//www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1427r0.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
Hans Boehm and Michael Spertus. 2005. Transparent Garbage Collection for C++. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper N1833. 24 June 2005. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2005/n1833.pdf (also at
Hans-J. Boehm, Mike Spertus, and Clark Nelson. 2008. Minimal Support for Garbage Collection and Reachability-Based
Leak Detection (revised). ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N2586. 16 March 2008. http:
//www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2586.html (also at Internet Archive 4 Sept. 2019 02:05:46).
Boost 1998ś2020. Boost C++ Libraries (website). http://www.boost.org/ (also at Internet Archive 7 April 2020 20:27:31).
Boost provides free peer-reviewed portable C++ source libraries.
We emphasize libraries that work well with the C++ Standard Library. Boost libraries are intended to be widely useful,
and usable across a broad spectrum of applications. The Boost license encourages the use of Boost libraries for all
users with minimal restrictions.
We aim to establish "existing practice" and provide reference implementations so that Boost libraries are suitable for
eventual standardization. Beginning with the ten Boost Libraries included in the Library Technical Report (TR1) and
continuing with every release of the ISO standard for C++ since 2011, the C++ Standards Committee has continued
to rely on Boost as a valuable source for additions to the Standard C++ Library.
Boost Hana 2015ś2020. Boost Hana (website). https://www.boost.org/doc/libs/1_61_0/libs/hana/doc/html/index.html
Archived at https://web.archive.org/web/2019*/https://www.boost.org/doc/libs/1_61_0/libs/hana/doc/html/index.html
Hana is a header-only library for C++ metaprogramming suited for computations on both types and values. The
functionality it provides is a superset of what is provided by the well established Boost.MPL and Boost.Fusion
libraries. By leveraging C++11/14 implementation techniques and idioms, Hana boasts faster compilation times and
runtime performance on par or better than previous metaprogramming libraries, while noticeably increasing the
level of expressiveness in the process. Hana is easy to extend in a ad-hoc manner and it provides out-of-the-box
inter-operation with Boost.Fusion, Boost.MPL and the standard library.

Vicente Botet and JF Bastien. 2018. std::expected. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0323R6.
2 April 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0323r6.html (also at Internet Archive 4 Sept.
2019 02:05:46).
Lubomir Bourdev and Jaakko Järvi. 2006. Efficient Run-Time Dispatching in Generic Programming with Minimal Code
Bloat. In Proceedings of the Second International Workshop on Library-Centric Software Design (LCSD ’06), co-located with
OOPSLA 2006 (Portland, Oregon, USA, 22 Oct.), Andreas Priesnitz and Sibylle Schupp (Eds.). Computer Science and
Engineering Department, Chalmers University of Technology, Göteborg, Sweden, 15ś24. http://sms.cs.chalmers.se/
bibliography/proceedings/2006-LCSD.pdf (also at Internet Archive 6 Feb. 2007 22:16:12). Technical Report 06-18.
Lubomir Bourdev and Jaakko Järvi. 2011. Efficient Run-Time Dispatching in Generic Programming with Minimal Code
Bloat. Science of Computer Programming 76, 4 (April), 243ś257. https://doi.org/10.1016/j.scico.2008.06.003
Walter E. Brown, Mark Fischler, Jim Kowalkowski, and Marc Paterno. 2006. Random Number Generation in C++0X:
A Comprehensive Proposal. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N1932. 23 Feb. 2006. http:
//www.open-std.org/jtc1/sc22/wg21/docs/papers/2006/n1932.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
Build2 2014ś2020. Build2, an open source, cross-platform build toolchain for developing and packaging C and C++ code.
(website). http://build2.org/ (also at Internet Archive 17 Dec. 2019 16:03:15).
build2 is an open source (MIT), cross-platform build toolchain for developing and packaging C and C++ code. It is a
hierarchy of tools that includes the build system, package dependency manager (for package consumption), and
project dependency manager (for project development).
Casey Carter. 2018. Standard Library Concepts. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0898R0. 12 Feb.
02:05:46).
Cevelop 2014ś2020. Cevelop: An Eclipse-based IDE (website; initial release: 21 July 2014). https://www.cevelop.com/ (also
at Internet Archive 10 Oct. 2019 02:56:01).
IDE++: Cevelop extends Eclipse CDT with many additional features: CUTE unit testing with Test Driven Development
support, new refactorings and quick fixes, and much more. Cevelop comes with many tools to upgrade your code to
C++11/14.
Matúš Chochlík. 2014. Static reflection. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N3996. 26 May 2014.
http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n3996.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
Matúš Chochlík. 2015. A case for strong static reflection. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N4452.
11 April 2015. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/n4452.pdf (also at Internet Archive 4 Sept.
2019 02:05:46).
Matúš Chochlík and Axel Naumann. 2016. Static reflection (revision 4). ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper P0194R0. 8 Feb. 2016. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/p0194r0.pdf (also at Internet
Archive 4 Sept. 2019 02:05:46).
Matúš Chochlík, Axel Naumann, and David Sankel. 2017. Static Reflection in a Nutshell. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper P0578R1. 18 June 2017. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0578r1.
html (also at Internet Archive 4 Sept. 2019 02:05:46).
Clang Tidy 2007ś2020. Clang Tidy: A collection of rule checkers for the Clang compiler (website). https://clang.llvm.org/
extra/clang-tidy/ (also at Internet Archive 4 April 2020 14:39:53).
clang-tidy is a clang-based C++ łlinterž tool. Its purpose is to provide an extensible framework for diagnosing and
fixing typical programming errors, like style violations, interface misuse, or bugs that can be deduced via static
analysis. clang-tidy is modular and provides a convenient interface for writing new checks.
Cling 2014ś2020. Cling, an interactive C++ interpreter from CERN (website). https://root.cern.ch/cling (also at Internet
Archive 28 Jan. 2020 05:14:14).
Cling is an interactive C++ interpreter, built on the top of LLVM and Clang libraries. Its advantages over the standard
interpreters are that it has command line prompt and uses just-in-time (JIT) compiler for compilation. Many of the
developers (e.g. Mono in their project called CSharpRepl) of such kind of software applications name them interactive
compilers.
One of Cling’s main goals is to provide contemporary, high-performance alternative of the current C++ interpreter in
the ROOT project - CINT. The backward-compatibility with CINT is major priority during the development.
Clion Wikipedia 2020. JetBrains (formerly IntelliJ Software). https://en.wikipedia.org/wiki/JetBrains (also at Internet
Archive 10 April 2020 22:12:43).
Marshall Clow, Beman Dawes, Gabriel Dos Reis, Stephan T. Lavavej, Billy O’Neal, Bjarne Stroustrup, and Jonathan Wakely.
2018. Standard Library Modules. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0581R1. 11 Feb. 2018.
http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0581r1.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
Cmake 2000ś2020. Cmake, an open-source, cross-platform family of tools designed to build, test and package software
(website). https://cmake.org/ (also at Internet Archive 8 April 2020 14:02:11).
CMake is an open-source, cross-platform family of tools designed to build, test and package software. CMake is used
to control the software compilation process using simple platform and compiler independent configuration files, and
generate native makefiles and workspaces that can be used in the compiler environment of your choice. The suite
of CMake tools were created by Kitware in response to the need for a powerful, cross-platform build environment
for open-source projects such as ITK and VTK. CMake is part of Kitware’s collection of commercially supported
open-source platforms for software development.
Jonathan Coe and Roger Orr. 2015. Extension methods for C++ (Uniform-function-calling syntax lite). ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper P0079R0. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/
2015/p0079r0.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
Conan 2016ś2020. Conan, the C / C++ Package Manager for Developers (website). https://conan.io/ (also at Internet
Archive 28 Jan. 2020 04:44:48).
The open source, decentralized and multi-platform package manager to create and share all your native binaries.
Melvin E. Conway. 1963. Design of a Separable Transition-Diagram Compiler. Commun. ACM 6, 7 (July), 396ś408. 0001-0782
Coverity 2002ś2020. Coverity: static analysis tools (website). https://scan.coverity.com/ (also at Internet Archive 5 March
2020 17:57:39).
Find and fix defects in your Java, C/C++, C#, JavaScript, Ruby, or Python open source project for free. Test every line
of code and potential execution path. The root cause of each defect is clearly explained, making it easy to fix bugs.
Integrated with GitHub. More than 6600 open source projects and 33000 developers use Coverity Scan.
Cppreference 2011ś2020. Cppreference: An online reference for the C++ language and standard library (website). https:
//en.cppreference.com/w/Main_Page (also at Internet Archive 31 March 2020 00:33:10).
Our goal is to provide programmers with a complete online reference for the C and C++ languages and standard
libraries, i.e., a more convenient version of the C and C++ standards.
Lawrence Crowl. 2009. Reaching Scope of Lambda Expressions. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
N2957. 25 Sept. 2009. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2009/n2957.html (also at Internet Archive
4 Sept. 2019 02:05:46).
Lawrence Crowl. 2013. Digit Separators. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N3661. 19 April 2013.
http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2013/n3661.html (also at Internet Archive 4 Sept. 2019 02:05:46).
Lawrence Crowl. 2015a. Comparison in C++. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N4367. 8 Feb. 2015.
Lawrence Crowl. 2015b. Handling Disappointment in C++. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
P0157R0. 7 Nov. 2015. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/p0157r0.html (also at Internet Archive
4 Sept. 2019 02:05:46).
Lawrence Crowl. 2018. Ambiguity and Insecurities with Three-Way Comparison. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper P1380R0. 26 Nov. 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p1380r0.html (also
at Internet Archive 4 Sept. 2019 02:05:46).
Lawrence Crowl and Thorsten Ottosen. 2006. Proposal to add Contract Programming to C++ (revision 4). ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper N1962. 25 Feb. 2006. http://www.open-std.org/jtc1/sc22/wg21/docs/
papers/2006/n1962.html (also at Internet Archive 4 Sept. 2019 02:05:46).
Lawrence Crowl, Richard Smith, Jeff Snyder, and Daveed Vandevoorde. 2013. Single-Quotation-Mark as a Digit Separator.
Ivan Čukić. 2018. Functional Programming in C++: How to improve your C++ programs using functional techniques. Manning
Publications, Shelter Island, NY, USA. 320 book pages. 978-1617293818
Krysztof Czarnecki and Ulrich Eisenecker. 2000. Generative Programming: Methods, Tools, and Applications. Addison-Wesley
Professional, Reading, Massachussets, USA. 864 book pages. 978-0201309775
Beman Dawes. 2002ś2014. The Boost Filesystem Library Design (website). https://www.boost.org/doc/libs/1_67_0/libs/
filesystem/doc/design.htm (also at Internet Archive 10 April 2020 20:04:42).
The key feature C++ lacked for script-like applications was the ability to perform portable filesystem operations on
directories and their contents. The Filesystem Library was developed to fill that void.
The intent is not to compete with traditional scripting languages, but to provide a solution for situations where C++
is already the language of choice.
Beman Dawes. 2006. Raw String Literals. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N2053. 6 Sept. 2006.
Beman Dawes (Ed.). 2014. Programming languages Ð C++ Ð File System Technical Specification. ISO/IEC JTC1/SC2/WG21:
C++ Standards Committee paper N4100. 4 July 2014. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n4100.
pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
Beman Dawes (Ed.). 2015. ISO/IEC TS 18822:2015: Programming languages Ð C++ Ð File System Technical Specification.
ISO (International Organization for Standardization), Geneva, Switzerland (July). 63 book pages. https://www.iso.org/
standard/63483.html Status: withdrawn.
Beman Dawes, Howard Hinnant, Bjarne Stroustrup, David Vandevoorde, and Michael Wong. 2018. Direction for ISO C++.
ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0939R0. 10 Feb. 2018. http://www.open-std.org/jtc1/sc22/
wg21/docs/papers/2018/p0939r0.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
Beman Dawes, Eric Niebler, and Casey Carter. 2016. Iterator Facade Library Proposal for Ranges. ISO/IEC JTC1/SC2/WG21:
C++ Standards Committee paper P0186R0. 11 Feb. 2016. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/
p0186r0.html (also at Internet Archive 4 Sept. 2019 02:05:46).
Beman G. Dawes. 1998. Proposal for a C++ Library Repository Web Site. 6 May 1998. https://www.boost.org/users/proposal.
pdf (also at Internet Archive 16 Dec. 2019 15:10:40).
Jeffrey Dean and Sanjay Ghemawat. 2004. MapReduce: Simplified Data Processing on Large Clusters. In Proceedings of the 6th
Symposium on Operating Systems Design and Implementation (San Francisco, California, USA, Dec.) (OSDI ’04). USENIX
Association, Berkeley, California, USA, 137ś149. https://www.usenix.org/legacy/events/osdi04/tech/full_papers/dean/
dean.pdf (also at Internet Archive 8 Feb. 2020 05:00:02).
Jeffrey Dean and Sanjay Ghemawat. 2008. MapReduce: Simplified Data Processing on Large Clusters. Commun. ACM 51, 1
(Jan.), 107ś113. 0001-0782 https://doi.org/10.1145/1327452.1327492
James C. Dehnert and Alexander Stepanov. 2000. Fundamentals of Generic Programming. LNCS, Vol. 1766. Springer-Verlag,
Berlin, 1ś11. 978-3-540-41090-4 https://doi.org/10.1007/3-540-39953-4_1 Also at http://stepanovpapers.com/DeSt98.pdf
(also at Internet Archive 6 June 2019 02:08:06).
Peter Dimov, Beman Dawes, and Greg Colvin. 2003. A Proposal to Add General Purpose Smart Pointers to the Library
Technical Report. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N1431. 28 Feb. 2003. http://www.open-
std.org/jtc1/sc22/wg21/docs/papers/2003/n1431.htm (also at Internet Archive 24 Dec. 2016 19:20:32).
Peter Dimov, Louis Dionne, Nina Ranns, Richard Smith, and Daveed Vandevoorde. 2019. More constexpr containers. ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper P0784R5. 21 Jan. 2019. http://www.open-std.org/jtc1/sc22/wg21/
docs/papers/2019/p0784r5.html (also at Internet Archive 4 Sept. 2019 02:05:46).
Peter Dimov and Vassil Vassilev. 2018. Allowing Virtual Function Calls in Constant Expressions. ISO/IEC JTC1/SC2/WG21:
C++ Standards Committee paper P1064R0. 4 May 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/
Louis Dionne. 2016. Lambdas in unevaluated contexts. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0315R0.
http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/p0315r0.pdf (also at Internet Archive 21 April 2020 02:00:28).
Louis Dionne. 2017. Familiar template syntax for generic lambdas. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper P0428R2. 13 July 2017. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0428r2.pdf (also at Internet
Archive 4 Sept. 2019 02:05:46).
Louis Dionne. 2018. Making std::vector constexpr. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1004R1.
7 Oct. 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p1004r1.pdf (also at Internet Archive 4 Sept.
2019 02:05:46).
Louis Dionne and David Vandevoorde. 2018. Changing the active member of a union inside constexpr. ISO/IEC
docs/papers/2018/p1330r0.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
Gabriel Dos Reis. 2003. Generalized Constant Expressions. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
N1521. 21 Sept. 2003. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2003/n1521.pdf (also at Internet Archive
4 Sept. 2019 02:05:46).
Gabriel Dos Reis. 2009. IPR: Compiler-neutral Internal Program Representation for C++ (code repository). http://github.
com/GabrielDosReis/ipr (also at Internet Archive 10 Jan. 2019 00:58:20).
Gabriel Dos Reis. 2012. A System for Axiomatic Programming. In Intelligent Computer Mathematics: 11th International
Conference (CICM 2012) (Bremen, Germany, July). Springer-Verlag, Berlin, Germany, 295ś309. 978-3-642-31373-8 https:
//doi.org/10.1007/978-3-642-31374-5_20 LNCS 7362.
Gabriel Dos Reis (Ed.). 2018. Working Draft, Extensions to C++ for Modules. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper N4720. 29 Jan. 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/n4720.pdf (also at
Gabriel Dos Reis, J-Daniel Garcia, John Lakos, Alisdair Meredith, Nathan Myers, and Bjarne Stroustrup. 2016a. A Contract
Design. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0380R0. 28 May 2016. http://www.open-std.org/
jtc1/sc22/wg21/docs/papers/2016/p0380r0.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
Gabriel Dos Reis, J-Daniel Garcia, John Lakos, Alisdair Meredith, Nathan Myers, and Bjarne Stroustrup. 2018. Support for
contract-based programming in C++. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0542R4. 2 April 2018.
http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0542r4.html (also at Internet Archive 4 Sept. 2019 02:05:46).
Gabriel Dos Reis, Mark Hall, and Gor Nishanov. 2014. A Module System for C++. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper N4047. 27 May 2014. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n4047.pdf (also at
Gabriel Dos Reis and Mat Marcus. 2003. Proposal to add template aliases to C++. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper N1449. 7 April 2003. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2003/n1449.pdf (also at
Gabriel Dos Reis, Nathan Sidwell, Richard Smith, and David Vandevoorde. 2019. Modules are ready. 14 Feb. 2019. Note on
the WG21 mailing list.
Gabriel Dos Reis and Richard Smith. 2018a. Modules for Standard C++. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper P1087R0. 7 May 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p1087r0.pdf (also at Internet
Archive 21 April 2020 03:04:09).
Gabriel Dos Reis and Richard Smith. 2018b. Plan of Record for Making C++ Modules Available in C++ Standards. ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper P0983R0. 1 April 2018. http://www.open-std.org/jtc1/sc22/wg21/
docs/papers/2018/p0983r0.pdf (also at Internet Archive 21 April 2020 16:08:45).
Gabriel Dos Reis and Bjarne Stroustrup. 2005a. A Formalism for C++. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper N1885. 20 Oct. 2005. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2005/n1885.pdf (also at Internet
Archive 4 Sept. 2019 02:05:46).
Gabriel Dos Reis and Bjarne Stroustrup. 2005b. Specifying C++ concepts. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper N1886. 20 Oct. 2005. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2005/n1886.pdf (also at
Gabriel Dos Reis and Bjarne Stroustrup. 2006. Specifying C++ Concepts. In Conference Record of the 33rd ACM SIGPLAN-
SIGACT Symposium on Principles of Programming Languages (Charleston, South Carolina, USA, Jan.) (POPL ’06). Associa-
tion for Computing Machinery, New York, NY, USA, 295ś308. 1595930272 https://doi.org/10.1145/1111037.1111064
Gabriel Dos Reis and Bjarne Stroustrup. 2007. Constant Expressions in the Standard Library. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper N2219. 11 March 2007. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2007/n2219.pdf
(also at Internet Archive 4 Sept. 2019 02:05:46).
Gabriel Dos Reis and Bjarne Stroustrup. 2009. A Principled, Complete, and Efficient Representation of C++. In Proceedings of
the Joint Conference of the 9th Asian Symposium on Computer Mathematics (ASCM 2009) and 3rd International Conference
on Mathematical Aspects of Computer and Information Sciences (MACIS 2009) (Fukuoka, Japan, 14 Dec.), Masakazu Suzuki,
Hoon Hong, Hirokazu Anai, Chee Yap, Yousuke Sato, and Hiroshi Yoshida (Eds.). Kyushu University, Fukuoka, Japan,
407ś421. https://catalog.lib.kyushu-u.ac.jp/opac_download_md/16844/miln_22.pdf (also at Internet Archive 13 April
2020 14:57:31). Also at http://stroustrup.com/macis09.pdf http://hdl.handle.net/2324/16844 MI Lecture Note Series
(formerly COE Lecture Note Series) Vol. 22.
Gabriel Dos Reis and Bjarne Stroustrup. 2010. General Constant Expressions for System Programming Languages. In
Proceedings of the 2010 ACM Symposium on Applied Computing (Sierre, Switzerland, March) (SAC ’10). Association for
Computing Machinery, New York, NY, USA, 2131ś2136. 978-1605586397 https://doi.org/10.1145/1774088.1774537
Gabriel Dos Reis and Bjarne Stroustrup. 2011. A Principled, Complete, and Efficient Representation of C++. Mathematics in
Computer Science 5, 3 (Sept.), 335ś356. 1661-8270 https://doi.org/10.1007/s11786-011-0094-1
Gabriel Dos Reis, Bjarne Stroustrup, and Alisdair Meredith. 2009. Axioms: Semantics Aspects of C++ Concepts. ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper N2887. 21 June 2009. http://www.open-std.org/jtc1/sc22/wg21/docs/
papers/2009/n2887.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
Gabriel Dos Reis, Herb Sutter, and Jonathan Caves. 2016b. Refining Expression Evaluation Order for Idiomatic C++. ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper P0145R2. 3 March 2016. http://www.open-std.org/jtc1/sc22/wg21/
Robert Douglas and Corentin Jabot. 2019. Adopt source location from Library Fundamentals V3 for C++20. ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper P1208R5. 17 June 2019. http://www.open-std.org/jtc1/sc22/wg21/
Stefanus du Toit (Ed.). 2014. ISO/IEC 14882:2014: Information Technology Ð Programming languages Ð C++. ISO (International
Organization for Standardization), Geneva, Switzerland (Dec.). 1358 book pages. https://www.iso.org/standard/64029.html
Status: withdrawn.
ECMA International. 2005. C++/CLI Language Specification. ECMA International, Geneva, Switzerland (Dec.). https:
//www.ecma-international.org/publications/files/ECMA-ST/ECMA-372.pdf (also at Internet Archive 13 April 2019
01:09:02). ECMA Standard ECMA-372.
Robert Klarer (editor). 2007. Extension for the programming language C++ to support decimal floating-point arithmetic.
ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N2198. 12 March 2007. http://www.open-std.org/jtc1/sc22/
wg21/docs/papers/2007/n2198.pdf (also at Internet Archive 4 Sept. 2019 02:05:46). ISO/IEC WDTR 24733.
David Eisenbud, Daniel R. Grayson, Michael Stillman, and Bernd Sturmfel (Eds.). 2001. Computations in Algebraic Geometry
with Macaulay 2. Algorithms and Computation in Mathematics, Vol. 8. Springer, Berlin, Germany (Oct.). 345 book pages.
978-3540422303
Margaret A. Ellis and Bjarne Stroustrup. 1990. The C++ Annotated Reference Manual. Addison-Wesley, Reading, Massachusetts,
USA. 0-201-51459-1
Hal Finkel and Richard Smith. 2016. Inline Variables. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0386R0.
30 May 2016. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/p0386r0.pdf (also at Internet Archive 4 Sept.
2019 02:05:46).
Eric Fiselier. 2019. Adding the ’constinit’ keyword. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1143R1.
21 Jan. 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1143r1.md (also at Internet Archive 4 Sept.
2019 02:05:46).
Erich Gamma, Richard Helm, Ralph Johnson, and John Vlissides. 1994. Design Patterns: Elements of Reusable Object-Oriented
Software. Addison-Wesley Professional, Reading, Massachussets, USA. 416 book pages. 978-0201633610
J. Daniel Garcia. 2018. Avoiding undefined behavior in contracts. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper P1290R0. 26 Nov. 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p1290r0.pdf (also at Internet
Archive 4 Sept. 2019 02:05:46).
J. Daniel García, Francesco Logozzo, Gabriel Dos Reis, Manuel Fähndrich, and Shuvendu Lahiri. 2015. Simple Contracts for
C++. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N4415. 12 April 2015. http://www.open-std.org/jtc1/
sc22/wg21/docs/papers/2015/n4415.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
J. Daniel Garcia and Bjarne Stroustrup. 2015. Improving performance and maintainability through refactoring in C++11.
27 Aug. 2015. https://pdfs.semanticscholar.org/1a84/826ad4968082761952b7ef0f5f6bc65f39e7.pdf (also at Internet
Archive 19 Feb. 2019 20:02:29). Also at Research Gate: https://www.researchgate.net/publication/285576155_Improving_
performance_and_maintainability_through_refactoring_in_C11_Improving_performance_and_maintainability_
through_refactoring_in_C11
Ronald Garcia, Jaakko Järvi, Andrew Lumsdaine, Jeremy Siek, and Jeremiah Willcock. 2007. An Extended Comparative
Study of Language Support for Generic Programming. Journal of Functional Programming 17, 2 (March), 145ś205.
https://doi.org/10.1017/S0956796806006198
Mathias Gaunard and Dietmar Kühl. 2015. Function Object-Based Overloading of Operator Dot. ISO/IEC JTC1/SC2/WG21:
C++ Standards Committee paper P0060R0. 18 Sept. 2015. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/
Michael Gibbs and Bjarne Stroustrup. 2006. Fast Dynamic Casting. Software: Practice and Experience 36, 2, 139ś156.
https://doi.org/10.1002/spe.686
GNUmake 2006ś2020. GNUmake, a tool for generating executable code from multiple source files (website). http:
//www.gnu.org/software/make/ (also at Internet Archive 8 April 2020 02:17:10).
GNU Make is a tool which controls the generation of executables and other non-source files of a program from the
program’s source files.
Matt Godbolt. 2016. Compiler Explorer, an online interactive environment for compiling C++ using a variety of compilers
and seeing the assembly code they generate (blog post). 4 Sept. 2016. https://xania.org/201609/how-compiler-explorer-
runs-on-amazon (also at Internet Archive 1 April 2020 22:07:49). https://godbolt.org
Nat Goodspeed. 2014. Stackful Coroutines and Stackless Resumable Functions. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper N4232. 13 Oct. 2014. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n4232.pdf (also at
Peter Gottschling. 2015. Discovering Modern C++. Addison-Wesley, Boston, Massachussets, USA. 978-0-13-438358-3
Douglas Gregor. 2006. A Brief Introduction to Variadic Templates. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper N2087. 10 Sept. 2006. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2006/n2087.pdf (also at Internet
Archive 4 Sept. 2019 02:05:46).
Doug Gregor. 2010. Deprecating Exception Specifications. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
N3051. 12 March 2010. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2010/n3051.html (also at Internet Archive
4 Sept. 2019 02:05:46).
Douglas Gregor. 2012. Modules (slides). Nov. 2012. http://llvm.org/devmtg/2012-11/Gregor-Modules.pdf (also at Internet
Archive 20 Dec. 2019 10:43:30). Video at https://youtu.be/586c_QMXir4 Meeting proceedings at http://llvm.org/devmtg/
2012-11/ Presentation at 2012 LLVM Developers’ Meeting, San Jose, California, USA.
Douglas Gregor and David Abrahams. 2009. Rvalue References and Exception Safety. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper N2855. 23 March 2009. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2009/n2855.

Douglas Gregor, Jaakko Järvi, and Gary Powell. 2004. Variadic Templates. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper N1603. 17 Feb. 2004. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2004/n1603.pdf (also at
Douglas Gregor, Jaakko Järvi, Jeremy Siek, Bjarne Stroustrup, Gabriel Dos Reis, and Andrew Lumsdaine. 2006. Concepts:
Linguistic Support for Generic Programming in C++. In Proceedings of the 21st Annual ACM SIGPLAN Conference
on Object-Oriented Programming Systems, Languages, and Applications (Portland, Oregon, USA, Oct.) (OOPSLA ’06).
Association for Computing Machinery, New York, NY, USA, 291ś310. SIGPLAN Notices 41, 10. 1595933484 0362-1340
Douglas Gregor and Andrew Lumsdaine. 2008. Core Concepts for the C++0x Standard Library. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper N2502. 3 Feb. 2008. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2502.pdf
Douglas Gregor, Bjarne Stroustrup, James Widman, and Jeremy Siek. 2008. Proposed Wording for Concepts (Revision 6).
ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N2676. 30 June 2008. http://www.open-std.org/jtc1/sc22/
Kate Gregory. 2015. Stop Teaching C (video). 25 Sept. 2015. non-archival https://youtu.be/YnWhqhNdYyk (63 minutes).
non-archival https://cppcon2015.sched.com/event/3vm1/stop-teaching-c
To this day most people who set out to help others learn C++ start with łintroduction to Cž material. I think
this actively contributes to bad C++ code in the world. For the past few years I’ve been teaching C++ (and making
suggestions to folks who intend to teach themselves) in an entirely different way. No char* strings, no strlen, strcmp,
strcpy, no printf, and no [] arrays. Pointers introduced very late. References before pointers, and polymorphism
with references rather than with pointers. Smart pointers as the default pointer with raw pointers (whether from
new or &) reserved for times they’re needed. Drawing on the Standard Library sooner rather than later, and writing
modern C++ from lesson 1.
Talk at CppCon: The C++ Conference.
Kate Gregory. 2017. 10 Core Guidelines You Need to Start Using Now (video). 27 Sept. 2017. non-archival https:
//youtu.be/XkDEzfpdcSg (63 minutes). non-archival https://cppcon2017.sched.com/event/Bgt1/10-core-guidelines-
you-need-to-start-using-now
The C++ Core Guidelines were announced at CppCon 2015, yet some developers have still never heard of them. It’s
time to see what they have to offer for you, no matter how much C++ experience you have. You don’t need to read
and learn the whole thing: in this talk I am pulling out some highlights of the Guidelines to show you why you should
be using these selected guidelines. For each one I’ll show some examples, and discuss the benefit of adopting them
for new code or going back into old code to make a change.
Talk at CppCon: The C++ Conference.
Kate Gregory. 2018. Simplicity: Not Just For Beginners (video). 26 Sept. 2018. non-archival https://youtu.be/n0Ak6xtVXno
(89 minutes). non-archival https://cppcon2018.sched.com/event/FnKB/simplicity-not-just-for-beginners
Many people say that simple code is better code, but fewer put it into practice. In this talk I’ll spend a little time on
why simpler is better, and why we resist simplicity. Then I’ll provide some specific approaches that are likely to make
your code simpler, and discuss what you need to know and do in order to consistently write simpler code and reap
the benefits of that simplicity.
Keynote talk at CppCon: The C++ Conference.
GTKmm 2005ś2020. GTKmm: an open-source, cross-platform GUI library (used for GTK+ and Gnome) (website). http:
//www.gtkmm.org/en/ (also at Internet Archive 25 Jan. 2020 02:17:50).
gtkmm is the official C++ interface for the popular GUI library GTK+. Highlights include typesafe callbacks, and
a comprehensive set of widgets that are easily extensible via inheritance. You can create user interfaces either in
code or with the Glade User Interface designer, using Gtk::Builder. There’s extensive documentation, including API
reference and a tutorial.
Aleksey Gurtovoy and David Abrahams. 2002ś2020. The Boost MPL Library (website). https://www.boost.org/doc/libs/1_
73_0/libs/mpl/doc/index.html Archived at https://web.archive.org/web/*/https://www.boost.org/doc/libs/1_73_0/libs/
mpl/doc/index.html
The Boost.MPL library is a general-purpose, high-level C++ template metaprogramming framework of compile-time
algorithms, sequences and metafunctions. It provides a conceptual foundation and an extensive set of powerful and
coherent tools that make doing explict metaprogramming in C++ as easy and enjoyable as possible within the current
language.
Niklas Gustafsson. 2012. Resumable Functions. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N3328. 12 Jan.
2012. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2012/n3328.pdf (also at Internet Archive 4 Sept. 2019
02:05:46).
Howard Hinnant, Roger Orr, Bjarne Stroustrup, David Vandevoorde, and Michael Wong. 2019. Direction for ISO C++.
ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0939R2. 21 Jan. 2019. http://www.open-std.org/jtc1/sc22/
Howard Hinnant, Roger Orr, Bjarne Stroustrup, David Vandevoorde, and Michael Wong. 2020. Direction for ISO C++.
ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P2000R0. 13 Jan. 2020. http://www.open-std.org/jtc1/sc22/
Howard E. Hinnant, Dave Abrahams, and Peter Dimov. 2004. A Proposal to Add an Rvalue Reference to the C++ Language.
wg21/docs/papers/2004/n1690.html (also at Internet Archive 4 Sept. 2019 02:05:46).
Howard E. Hinnant, Walter E. Brown, Jeff Garland, and Marc Paterno. 2008. A Foundation to Sleep On: Clocks, Points
in Time, and Time Durations. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N2615. 18 May 2008. http:
//www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2615.html (also at Internet Archive 4 Sept. 2019 02:05:46).
Howard E. Hinnant, Peter Dimov, and Dave Abrahams. 2002. A Proposal to Add Move Semantics Support to the C++
Language. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N1377. 10 Sept. 2002. http://www.open-
std.org/jtc1/sc22/wg21/docs/papers/2002/n1377.htm (also at Internet Archive 4 Sept. 2019 02:05:46).
Howard E. Hinnant and Tomasz Kamiński. 2018. Extending <chrono> to Calendars and Time Zones. ISO/IEC JTC1/SC2/WG21:
C++ Standards Committee paper P0355R7. 16 March 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/
Howard E. Hinnant, Bjarne Stroustrup, and Bronek Kozicki. 2006. A Brief Introduction to Rvalue References. ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper N2027. 12 June 2006. http://www.open-std.org/jtc1/sc22/wg21/docs/
Jared Hoberock (Ed.). 2019. Working Draft, C++ Extensions for Parallelism Version 2. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper N4796. 21 Jan. 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/n4796.pdf
Jared Hoberock, Michael Garland, Chris Kohlhoff, Chris Mysen, Carter Edwards, Gordon Brown, David Hollman, and other
contributors: Hans Boehm, Thomas Heller, Lee Howes, Bryce Lelbach, Hartmut Kaiser, Bryce Lelbach, Gor Nishanov,
Thomas Rodgers, and Michael Wong. 2019. A Unified Executors Proposal for C++. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper P0443R10. 21 Jan. 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p0443r10.
Jared Hoberock, Michael Garland, Chris Kohlhoff, Chris Mysen, Carter Edwards, Gordon Brown, and Michael Wong.
2018. Executors Design Document. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0761R2. 12 Feb. 2018.
http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0761r2.pdf (also at Internet Archive 21 April 2020 17:01:39).
David S. Hollman. 2019. Experience Report: Implementing a Coroutines TS Frontend to an Existing Tasking Library
Draft Proposal. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1403R0. 16 Jan. 2019. http://www.open-
std.org/jtc1/sc22/wg21/docs/papers/2019/p1403r0.html (also at Internet Archive 4 Sept. 2019 02:05:46).
Tom Honermann. 2017. Remove abbreviated functions and template-introduction syntax from the Concepts TS. ISO/IEC
Lee Howes, Eric Niebler, and Lewis Baker. 2018. In support of merging coroutines into C++20. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper P1241. 8 Oct. 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p1241r0.html
Jaakko Järvi. 2002. Proposal for adding tuple types into the standard library. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper N1382. 10 Sept. 2002. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2002/n1382.pdf (also at
Jaakko Järvi and Gary Powell. 2002. Boost.Lambda: The Boost Lambda Library (website). https://www.boost.org/doc/libs/
1_72_0/doc/html/lambda.html (also at Internet Archive 29 March 2020 02:26:54). http://boost.org/libs/lambda
The Boost Lambda Library (BLL in the sequel) is a C++ template library, which implements a form of lambda
abstractions for C++. The term originates from functional programming and lambda calculus, where a lambda
abstraction defines an unnamed function. The primary motivation for the BLL is to provide flexible and convenient
means to define unnamed function objects for STL algorithms.
Chapter 20 of the Boost Library Documentation.
Jaakko Järvi, Gary Powell, and Andrew Lumsdaine. 2003a. The Lambda Library: Unnamed functions in C++. Software:
Practice and Experience 33, 3, 259ś291. https://doi.org/10.1002/spe.504
Jaakko Järvi, Bjarne Stroustrup, and Gabriel Dos Reis. 2007. Decltype (revision 7): proposed wording. ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper N2343. 18 July 2007. http://www.open-std.org/jtc1/sc22/wg21/
docs/papers/2007/n2343.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
Jaakko Järvi, Bjarne Stroustrup, Doug Gregor, and Jeremy Siek. 2003b. Decltype and auto. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper N1478. 28 April 2003. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2003/n1478.pdf
JetBrains. 2020. CLion: A cross-platform IDE for C and C++ (website). https://www.jetbrains.com/clion/ (also at Internet
Archive 10 April 2020 00:41:01).
Smart C and C++ editor: Thanks to native C and C++ support, including modern C++ standards, libc++ and Boost,
CLion knows your code through and through and takes care of the routine while you focus on the important things.
Christopher Jonathan, Umar Farooq Minhas, James Hunter, Justin Levandoski, and Gor Nishanov. 2018. Exploiting
Coroutines to Attack the łKiller Nanosecondsž. Proceedings of the VLDB Endowment 11, 11 (July), 1702ś1714. 2150-8097
Nicolai Josuttis, Lewis Baker, Billy O’Neal, Herb Sutter, and Anthony Williams. 2019a. Stop Token and Joining Thread,
Rev 9. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0660R9. 10 March 2019. http://www.open-
std.org/jtc1/sc22/wg21/docs/papers/2019/p0660r9.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
Nicolai Josuttis, Ville Voutilainen, Roger Orr, Daveed Vandevoorde, John Spicer, and Christopher Di Bella. 2019b. Remove
Contracts from C++20. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1823R0. 18 July 2019. http:
//www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1823r0.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
Hartmut Kaiser, Maciek Brodowicz, and Thomas Sterling. 2009Sept.. ParalleX: An Advanced Parallel Execution Model for
Scaling-Impaired Applications. In 38th International Conference on Parallel Processing Workshops (ICPPW 2009) (Vienna,
Austria). IEEE Computer Society, Los Alamitos, California, USA, 394ś401. 978-0-7695-3803-7 https://doi.org/10.1109/
ICPPW.2009.14
D. Kapur, D. R. Musser, and A. A. Stepanov. 1981. Operators and Algebraic Structures. In Proceedings of the 1981 Conference
on Functional Programming Languages and Computer Architecture (Portsmouth, New Hampshire, USA) (FPCA ’81).
Association for Computing Machinery, New York, NY, USA, 59ś64. 0897910605 https://doi.org/10.1145/800223.806763
Anastasia Kazakova. 2015. Infographic: C/C++ Facts We Learned Before Going Ahead with CLion (blog post). 27 July
2015. http://blog.jetbrains.com/clion/2015/07/infographics-cpp-facts-before-clion/ (also at Internet Archive 28 Jan. 2020
05:02:34).
Erich Keane, Adam David Alan Martin, and Allan Deutsch. 2017. A Case for Simplifying/Improving Natural Syntax
Concepts. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0782R0. 25 Sept. 2017. http://www.open-
Brian Kernighan and Dennis M. Ritchie. 1978. The C Programming Language. Prentice Hall, Englewood Cliffs, NJ, USA.
978-0131101630
Rostislav Khlebnikov and John Lakos. 2019. Contracts, Undefined Behavior, and Defensive Programming. ISO/IEC
The Khronos Group. 2014ś2020. C++ Single-source Heterogeneous Programming for OpenCL (website). https://www.
khronos.org/sycl/ Archived at https://web.archive.org/web/*/https://www.khronos.org/sycl/
SYCL (pronounced ’sickle’) is a royalty-free, cross-platform abstraction layer that builds on the underlying concepts,
portability and efficiency of OpenCL that enables code for heterogeneous processors to be written in a "single-source"
style using completely standard C++. SYCL single-source programming enables the host and kernel code for an
application to be contained in the same source file, in a type-safe way and with the simplicity of a cross-platform
asynchronous task graph. SYCL includes templates and generic lambda functions to enable higher-level application
software to be cleanly coded with optimized acceleration of kernel code across the extensive range of shipping
OpenCL 1.2 implementations.
Andrew Koenig and Bjarne Stroustrup. 1989. Exception Handling for C++. In Proceedings of the USENIX łC++ at Workž
Conference (Nov.). USENIX Association, Berkeley, California, USA, 31 pages. http://www.rajatorrent.com.stroustrup.
com/except89.pdf (also at Internet Archive 11 April 2020 01:00:36).
Andrew Koenig and Bjarne Stroustrup. 1991a. Analysis of Overloaded operator.(). ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper N0054. 21 Sept. 1991. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/1991/WG21%201991/
X3J16_91-0121%20WG21_N0054.pdf (also at Internet Archive 12 April 2016 17:17:26).
Andrew Koenig and Bjarne Stroustrup. 1991b. C++: As close to C as possible but no closer. ISO/IEC JTC1/SC2/WG21:
C++ Standards Committee paper N0089. May 1991. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/1991/WG21%
201991/X3J16_91-0089%20WG21_N0007.pdf (also at Internet Archive 4 Sept. 2019 02:05:46). Also bears the document
number N0007 and the date 11 May 1989.
Andrew R. Koenig (Ed.). 1998. ISO/IEC 14882:1998: Programming languages Ð C++. ISO (International Organization
for Standardization), Geneva, Switzerland (Sept.). 732 book pages. https://www.iso.org/standard/25845.html Status:
withdrawn.
Christopher Kohlhoff. 2005. Boost ASIO (website). https://www.boost.org/doc/libs/1_67_0/doc/html/boost_asio.html (also

at Internet Archive 11 April 2020 01:04:37).
Boost.Asio is a cross-platform C++ library for network and low-level I/O programming that provides developers with
a consistent asynchronous model using a modern C++ approach.
Christopher Kohlhoff. 2006. Networking Library Proposal for TR2. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper N2054. 8 Sept. 2006. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2006/n2054.pdf (also at Internet Archive
4 Sept. 2019 02:05:46).
Christopher Kohlhoff. 2013. A Universal Model for Asynchronous Operations. ISO/IEC JTC1/SC2/WG21: C++ Standards
Christopher M. Kohlhoff. 2018. ASIO C++ Library (website). http://think-async.com/Asio/ (also at Internet Archive
23 March 2020 15:37:13).
Asio is a cross-platform C++ library for network and low-level I/O programming that provides developers with a
consistent asynchronous model using a modern C++ approach.
Thomas Köppe. 2016a. Hexadecimal floating literals for C++. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
4 Sept. 2019 02:05:46).
Thomas Köppe. 2016b. Selection statements with initializer. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
P0305R1. 24 June 2016. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/p0305r1.html (also at Internet Archive
4 Sept. 2019 02:05:46).
Thomas Köppe. 2017a. An Adjective Syntax for Concepts. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
P0807R0. 12 Oct. 2017. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0807r0.html (also at Internet Archive
4 Sept. 2019 02:05:46).
Thomas Köppe. 2017b. Allow lambda capture [=, this]. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0409R2.
4 March 2017. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0409r2.html (also at Internet Archive 4 Sept.
2019 02:05:46).
Thomas Köppe. 2017c. Range-based for statements with initializer. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper P0614R1. 6 Nov. 2017. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0614r1.html (also at Internet
Archive 4 Sept. 2019 02:05:46).
Oliver Kowalke. 2015. A low-level API for stackful coroutines. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
N4397. 9 April 2015. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/n4397.pdf (also at Internet Archive
4 Sept. 2019 02:05:46).
Oliver Kowalke and Nat Goodspeed. 2013. A proposal to add coroutines to the C++ standard library. ISO/IEC JTC1/SC2/WG21:
C++ Standards Committee paper N3708. 4 March 2013. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2013/
n3708.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
John Lakos. 2018. łAvoiding undefined behavior in contractsž [P1290R0] Explained. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper P1335R0. 26 Nov. 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p1335r0.txt (also at
John Lakos. 2019. United Amendment to Contracts Facility for C++20. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper P1486R1. 21 Feb. 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1486r1.txt (also at Internet
Archive 4 Sept. 2019 02:05:46).
John Lakos and Alexei Zakharov. 2013. Centralized Defensive-Programming Support for Narrow Contracts. ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper N3604. 18 March 2013. http://www.open-std.org/jtc1/sc22/wg21/
docs/papers/2013/n3604.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
Leslie Lamport. 1978. Time, Clocks and the Ordering of Events in a Distributed System. Commun. ACM 21, 7 (July), 558ś565.
0001-0782 https://doi.org/10.1145/359545.359563
Leslie Lamport. 1979. How to Make a Multiprocessor Computer That Correctly Executes Multiprocess Programs. IEEE
Trans. Comput. C-28, 9 (Sept.), 690ś691. 2326-3814 https://doi.org/10.1109/TC.1979.1675439
Bryce Adelstein Lelbach, Olivier Giroux, JF Bastien, Detlef Vollmann, and David Olsen. 2019. The C++20 Synchronization
Library. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1135R4. 4 March 2019. http://www.open-
LLVM 2003ś2020. LLVM, a collection of tools and libraries for building compilers and tools (website). https://llvm.org/
(also at Internet Archive 7 April 2020 15:19:33).
The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. Despite its name,
LLVM has little to do with traditional virtual machines. The name "LLVM" itself is not an acronym; it is the full name
of the project.
LLVM began as a research project at the University of Illinois, with the goal of providing a modern, SSA-based
compilation strategy capable of supporting both static and dynamic compilation of arbitrary programming languages.
Since then, LLVM has grown to be an umbrella project consisting of a number of subprojects, many of which are
being used in production by a wide variety of commercial and open source projects as well as being widely used in
academic research.
Lockheed Martin Corporation. 2005. Joint Strike Fighter Air Vehicle C++ Coding Standards for the System Development and
Demonstration Program. Lockheed Martin Corporation (Dec.). http://stroustrup.com/JSF-AV-rules.pdf (also at Internet
Archive 8 Jan. 2020 08:09:03). Document Number 2RDU00001 Rev C.
Bruno Cardoso Lopes, Michael Spencer, and JF Bastien. 2019. Modules Feedback. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper P1482R0. 8 Feb. 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1482r0.html (also
Macaulay2 2005ś2020. Macaulay2 (website). http://www2.macaulay2.com/Macaulay2/ (also at Internet Archive 21 Sept.
2019 08:34:37).
Macaulay2 is a software system devoted to supporting research in algebraic geometry and commutative algebra,
whose creation has been funded by the National Science Foundation since 1992.
Macaulay2 includes core algorithms for computing Gröbner bases and graded or multi-graded free resolutions
of modules over quotient rings of graded or multi-graded polynomial rings with a monomial ordering. The core
algorithms are accessible through a versatile high level interpreted user language with a powerful debugger supporting
the creation of new classes of mathematical objects and the installation of methods for computing specifically with
them Macaulay2 can compute Betti numbers, Ext, cohomology of coherent sheaves on projective varieties, primary
decomposition of ideals, integral closure of rings, and more.
John Maddock. 2002. A Proposal to add Regular Expressions to the Standard Library. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper N1386. 6 Sept. 2002. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2002/n1386.htm
Jens Maurer. 2002. A Proposal to Add an Extensible Random Number Facility to the Standard Library. ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper N1398. 10 Nov. 2002. http://www.open-std.org/jtc1/sc22/wg21/docs/
Jens Maurer. 2012. Allowing arbitrary literal types for non-type template parameters. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper N3413. 19 Sept. 2012. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2012/n3413.html
Jens Maurer. 2019. Bit operations. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0553R4. 1 March 2019.
Jens Maurer and Michael Wong. 2007. Towards support for attributes in C++. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper N2236. 4 May 2007. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2007/n2236.pdf (also at
Ian McIntosh, Michael Wong, Raymond Mak, Robert Klarer, Jens Maurer, Alisdair Meredith, Bjarne Stroustrup, and David
Vandevoorde. 2008. User-defined Literals (aka. Extensible Literals (revision 5)). ISO/IEC JTC1/SC2/WG21: C++ Standards
Paul McJones (Ed.). 2007ś2020. C++ Historical Sources Archive (website). Computer History Museum Software Preservation
Group. http://www.softwarepreservation.org/projects/c_plus_plus/ (also at Internet Archive 29 March 2020 16:45:18).
This is a collection of design documents, source code, and other materials concerning the birth, development,
standardization, and use of the C++ programming language.
C++ added January 2007.
Paul E. McKenney, Mark Batty, Clark Nelson, Hans Boehm, Anthony Williams, Scott Owens, Susmit Sarkar, Peter Sewell,
Tjark Weber, Michael Wong, Lawrence Crowl, and Benjamin Kosnik. 2010. Omnibus Memory Model and Atomics Paper.
ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N3196. 11 Nov. 2010. http://www.open-std.org/jtc1/sc22/
wg21/docs/papers/2010/n3196.htm (also at Internet Archive 4 Sept. 2019 02:05:46).
Alisdair Meredith. 2007. Seeking a Syntax for Attributes in C++09. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper N2224. 12 March 2007. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2007/n2224.html (also at Internet
Archive 4 Sept. 2019 02:05:46).
Alisdair Meredith. 2012. Call for Library Proposals. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N3370.
26 Feb. 2012. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2012/n3370.html (also at Internet Archive 4 Sept.
2019 02:05:46).
Alisdair Meredith and Herb Sutter. 2017. Revising atomic_shared_ptr for C++20. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper P0718R2. 10 Nov. 2017. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0718r2.html (also
Bertrand Meyer. 1994. Object-Oriented Software Construction. Prentice Hall, Englewood Cliffs, NJ, USA. 534 book pages.
978-0136290490
Scott Meyers. 2014. Effective Modern C++: 42 Specific Ways to Improve Your Use of C++11 and C++14. O’Reilly Media,
Sebastopol, California. 978-1491903995
Microsoft. 2020. Visual Studio: Best-in-class tools for any developer (website). https://visualstudio.microsoft.com/ (also at
Internet Archive 13 April 2020 16:16:48).
Visual Studio: Full-featured IDE to code, debug, test, and deploy to any platform
Visual Studio Code: Editing and debugging on any OS
Visual Studio for Mac: Develop apps and games for iOS, Android, and web using .NET
Mihail Mihaylov and Vassil Vassilev. 2018. On the Coroutines TS. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper P1329R0. 2 Nov. 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1329r0.pdf (also at Internet
Archive 4 Sept. 2019 02:05:46).
Mihail Mihaylov and Vassil Vassilev. 2019. First-class symmetric coroutines in C++. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper P1430R0. 21 Jan. 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1430r0.pdf (also at
William M. Miller. 2010. Rvalue References as łFunnyž Lvalues. Edison Design Group. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper N3030. 14 Feb. 2010. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2010/n3030.pdf
Lev Minkovsky and John McFarlane. 2019. Math Constants. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
P0631R7. 24 May 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p0631r7.pdf (also at Internet Archive
4 Sept. 2019 02:05:46).
Mobileye. 2020. ADAS (Advanced Driver Assistance Systems) (website). https://www.mobileye.com/our-technology/adas/
(also at Internet Archive 13 Jan. 2020 03:29:15). Mobileye is an Intel company.
Sergei Murzin, Michael Park, David Sankel, and Dan Sarginson. 2019. Pattern Matching. ISO/IEC JTC1/SC2/WG21: C++
Sergei Murzin, Michael Park, David Sankel, and Dan Sarginson. 2020. Pattern Matching. ISO/IEC JTC1/SC2/WG21: C++
David R. Musser and Alexander A. Stepanov. 1987. A Library of Generic Algorithms in Ada. In Proceedings of the 1987
Annual ACM SIGAda International Conference on Ada (Boston, Massachusetts, USA, Dec.) (SIGAda ’87). Association
for Computing Machinery, New York, NY, USA, 216ś225. 0897912438 https://doi.org/10.1145/317500.317529 Also at
non-archival http://stepanovpapers.com/p216-musser.pdf (retrieved 28 Feb. 2020)
Axel Naumann. 2012. Introducing cling, a C+++ Interpreter Based on clang/LLVM (video). 15 March 2012. https:
//youtu.be/f9Xfh8pv3Fs Google Tech Talk (48 minutes).
Axel Naumann, Philippe Canal, Paul Russo, and Vassil Vassilev. 2010. Creating cling, an interactive interpreter interface for
clang (video). 4 Nov. 2010. http://llvm.org/devmtg/2010-11/ (also at Internet Archive 2 April 2020 04:03:23). Video at
https://youtu.be/BjmGOMJWeAo Slides at http://llvm.org/devmtg/2010-11/Naumann-Cling.pdf Talk at 2010 LLVM
Developers’ Meeting (26 minutes).
Peter Naur. 1966. Proof of Algorithms by General Snapshots. BIT (Nordisk Tidskrift för Informationsbehandling) Numerical
Mathematics 6, 4, 310ś316. https://doi.org/10.1007/BF01966091
Eric Niebler and Casey Carter. 2017. C++ Extensions for Ranges. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper N4651. 15 March 2017. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/n4651.pdf (also at Internet
Archive 4 Sept. 2019 02:05:46).
Eric Niebler, Casey Carter, and Christopher Di Bella. 2018. The One Ranges Proposal (was Merging the Ranges TS). ISO/IEC
docs/papers/2018/p0896r2.pdf (also at Internet Archive 21 April 2020 20:59:57).
Eric Niebler, Sean Parent, and Andrew Sutton. 2014. Ranges for the Standard Library, Revision 1. ISO/IEC JTC1/SC2/WG21:
C++ Standards Committee paper N4128. 10 Oct. 2014. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n4128.
Gor Nishanov. 2017. ISO/IEC TS 22277:2017: Technical Specification Ð C++ Extensions for Coroutines. ISO (International
Organization for Standardization), Geneva, Switzerland (Nov.). 18 book pages. https://www.iso.org/standard/73008.html
Gor Nishanov. 2018. Incremental Approach: Coroutine TS + Core Coroutines. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper P1362R0. 15 Nov. 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p1362r0.pdf (also
at Internet Archive 21 April 2020 21:01:42).
Gor Nishanov. 2019a. C++ Exception Optimizations. An experiment. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper P1676R0. 4 June 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1676r0.pdf (also at Internet
Archive 4 Sept. 2019 02:05:46).

Gor Nishanov. 2019b. Response to response to łFibers under the magnifying glassž. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper P1520R0. 8 March 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1520r0.pdf (also
NVIDIA. 2020. Advanced Driver Assistance Systems (ADAS) (website). https://www.nvidia.com/en-us/self-driving-
cars/adas/ (also at Internet Archive 10 April 2020 17:56:28).
Günter Obiltschnig et al. 2005. POCO C++ Libraries (website). http://pocoproject.org/ (also at Internet Archive 3 March
2020 05:39:34).
The POCO C++ Libraries are powerful cross-platform C++ libraries for building network- and internet-based applica-
tions that run on desktop, server, mobile, IoT, and embedded systems.
Initial release: February 21, 2005.
Objective C++ Wikipedia 2020. Objective C++ (a language variant accepted by GNU GCC and Clang). https://en.wikipedia.
org/wiki/Objective-C#Objective-C++ (also at Internet Archive 9 April 2020 21:16:54).
OpenCV 2020. OpenCV (website). https://opencv.org/about/ (also at Internet Archive 25 March 2020 14:29:00).
OpenCV (Open Source Computer Vision Library) is an open source computer vision and machine learning software
library. OpenCV was built to provide a common infrastructure for computer vision applications and to accelerate the
use of machine perception in the commercial products.
The library has more than 2500 optimized algorithms, which includes a comprehensive set of both classic and
state-of-the-art computer vision and machine learning algorithms.
OpenCV Wikipedia 2020. OpenCV (Open source Computer Vision). https://en.wikipedia.org/wiki/OpenCV (also at Internet
Archive 3 April 2020 14:24:23).
Thorsten Ottosen. 2005. Proposal for new for-loop. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N1796.
27 April 2005. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2005/n1796.html (also at Internet Archive 4 Sept.
2019 02:05:46).
Thorsten Ottosen, Lawrence Crowl, and Douglas Gregor. 2007. Wording for range-based for-loop (revision 1). ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper N2196. 7 March 2007. http://www.open-std.org/jtc1/sc22/wg21/docs/
Peter Pirkelbauer, Yuriy Solodkyy, and Bjarne Stroustrup. 2010. Design and Evaluation of C++ Open Multi-Methods. Science
of Computer Programming 75, 7 (July), 638ś667. 0167-6423 https://doi.org/10.1016/j.scico.2009.06.002 Accepted June 2009.
Expanded version of a paper presented at the conference on Generative Programming and Component Engineering
(GPCE 2007).
Gary Powell, Doug Gregor, and Jaakko Järvi. 2004. Overloading Operator.() & Operator.*(). ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper N1671. 10 Sept. 2004. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2004/n1671.pdf
Georgios Psaropoulos, Thomas Legler, Norman May, and Anastasia Ailamaki. 2017. Interleaving with Coroutines: A Practical
Approach for Robust Index Joins. In Proceedings of the 44th International Conference on Very Large Data Bases (Rio de
Janeiro, Brazil, 31 Aug.). VLDB Endowment Inc., Los Angeles, California, 230ś242. Proceedings of the VLDB Endowment
11, 2. https://doi.org/10.14778/3149193.3149202
William Pugh. 2004. JSR 133: Java Memory Model and Thread Specification Revision (website). Sept. 2004. http:
//jcp.org/en/jsr/detail?id=133 (also at Internet Archive 29 Sept. 2019 09:53:52).
The proposed specification describes the semantics of threads, locks, volatile variables and data races. This includes
what has been referred to as the Java memory model.
Qt 1991ś2020. Qt, a platform for GUI and related software (website). http://www.qt.io/ (also at Internet Archive 3 April
2020 19:32:36).
One framework. One codebase. Any platform. Everything you need for your entire software development life cycle.
Qt is the fastest and smartest way to produce industry-leading software that users love.
Tahina Ramananandro, Gabriel Dos Reis, and Xavier Leroy. 2011. Formal Verification of Object Layout for C++ Multiple
Inheritance. In Proceedings of the 38th Annual ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages
(Austin, Texas, USA) (POPL ’11). Association for Computing Machinery, New York, NY, USA, 67ś80. 978-1450304900
Tahina Ramananandro, Gabriel Dos Reis, and Xavier Leroy. 2012. A Mechanized Semantics for C++ Object Construction
and Destruction, with Applications to Resource Management. In Proceedings of the 39th Annual ACM SIGPLAN-SIGACT
Symposium on Principles of Programming Languages (Philadelphia, PA, USA) (POPL ’12). Association for Computing
Machinery, New York, NY, USA, 521ś532. 978-1450310833 https://doi.org/10.1145/2103656.2103718
RC++ 2010ś2020. RC++, a system for runtime-compiled C++: Alternatives (website). https://github.com/
RuntimeCompiledCPlusPlus/RuntimeCompiledCPlusPlus/wiki/Alternatives (also at Internet Archive 27 April 2015
23:18:58).
Runtime-Compiled C++ (RCC++) enables you to reliably make major changes to your C++ code at runtime and see
the results immediately. The technique is aimed at games development but can also be useful in any industry where
turnaround times are a bottleneck.
We believe RCC++ is a good overall standard C++ approach to runtime editing of code which is easy to use, available
across a number of platforms and relatively easy to port. There are however a few alternative implementations for
compiling C++ at runtime.
James Reinders. 2007. Intel Threading Building Blocks: Outfitting C++ For Multi-Core Processor Parallelism. O’Reilly Media,
Sebastopol, California (July). 344 book pages. 978-0596514808 http://shop.oreilly.com/product/9780596514808.do Code
repository at http://threadingbuildingblocks.org/
James Renwick, Tom Spink, and Björn Franke. 2019. Low-Cost Deterministic C++ Exceptions for Embedded Systems.
In Proceedings of the 28th International Conference on Compiler Construction (Washington, DC, USA, Feb.) (CC 2019).
Association for Computing Machinery, New York, NY, USA, 76ś86. 978-1450362771 https://doi.org/10.1145/3302516.
3307346
Vincent Reverdy. 2012. A proposal to add special mathematical functions according to the ISO/IEC 80000-2:2009 standard.
ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N3494. 19 Dec. 2012. http://www.open-std.org/jtc1/sc22/
Barry Revzin. 2017. Allow pack expansion in lambda init-capture. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper P0780R0. 8 Oct. 2017. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0780r0.html (also at Internet
Archive 4 Sept. 2019 02:05:46).
Barry Revzin. 2018. <=> != ==. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1185R0. 7 Oct. 2018.
Barry Revzin. 2019. Spaceship needs a tune-up: Addressing some discovered issues with P0515 and P1185. ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper P1630R1. 17 July 2019. http://www.open-std.org/jtc1/sc22/wg21/
Barry Revzin and Stephan T. Lavavej. 2018. explicit(bool). ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
P0892R2. 8 June 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0892r2.html (also at Internet Archive
4 Sept. 2019 02:05:46).
Jakob Riedle. 2017. Concepts are Adjectives, not Nouns. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0791R0.
10 Oct. 2017. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0791r0.pdf (also at Internet Archive 4 Sept.
2019 02:05:46).
Torvald Riegel. 2015. On unifying the coroutines and resumable functions proposals. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper P0073R0. 25 Sept. 2015. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/p0073r0.
Dennis M. Ritchie. 1990. Variable-Size Arrays in C. Journal of C Language Translation 2, 2 (Sept.), 81ś86. http://jclt.iecc.com/
Jct22.pdf (also at Internet Archive 2 April 2016 11:29:25). See also https://www.bell-labs.com/usr/dmr/www/vararray.pdf
Rene Rivera. 2019a. Are modules fast? (revision 0). ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1441R0.
21 Jan. 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1441r0.pdf (also at Internet Archive 4 Sept.
2019 02:05:46).
Rene Rivera. 2019b. Are modules fast? (revision 1). ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1441R1.
6 March 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1441r1.pdf (also at Internet Archive 4 Sept.
2019 02:05:46).
Geoff Romer, James Dennett, and Chandler Carruth. 2018. Core Coroutines: Making coroutines simpler, faster, and
more general. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1063R0. 6 May 2018. http://www.open-
Geoff Romer, James Dennett, and Chandler Carruth. 2019a. Core Coroutines: Making coroutines simpler, faster, and
more general. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1063R2. 16 Jan. 2019. http://www.open-
Geoffrey Romer, Gor Nishanov, Lewis Baker, and Mihail Mihailov. 2019b. Coroutines: Use-cases and Trade-offs. ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper P1493R0. 19 Feb. 2019. http://www.open-std.org/jtc1/sc22/wg21/
Hyman Rosen, John Lakos, and Alisdair Meredith. 2019. Adding a global contract assumption mode. ISO/IEC JTC1/SC2/WG21:
C++ Standards Committee paper P1730R0. 14 June 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/
p1730r0.md (also at Internet Archive 4 Sept. 2019 02:05:46).
David Sankel (Ed.). 2018. Working Draft, C++ Extensions for Reflection. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper N4766. 11 Aug. 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/n4766.pdf (also at Internet
Archive 4 Sept. 2019 02:05:46).
David Sankel and Daveed Vandevoorde. 2019. User-friendly and Evolution-friendly Reflection: A Compromise Document.
ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1733R0. 15 June 2019. http://www.open-std.org/jtc1/sc22/
Tim Shen, Richard Smith, Zhihao Yuan, and Chandler Carruth. 2016. Designated Initialization. ISO/IEC JTC1/SC2/WG21:
C++ Standards Committee paper P0329R0. 9 May 2016. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/
p0329r0.pdf (also at Internet Archive 21 April 2020 21:37:11).
Nathan Sidwell. 2018. Module Preamble is Unnecessarily Fragile (re N4720 Merging Modules). ISO/IEC JTC1/SC2/WG21:
C++ Standards Committee paper P1299R3. 13 Nov. 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/
p1299r3.pdf (also at Internet Archive 21 April 2020 21:39:18).
Nathan Sidwell. 2019. Make Me A Module (re P1184 A Module Mapper). ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper P1602R0. 1 March 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1602r0.pdf (also at Internet
Archive 4 Sept. 2019 02:05:46).
Nathan Sidwell and Davis Herring. 2019. Modules: ADL & Internal Linkage (re P1103R1 Merging Modules). ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper P1347R1. 17 Jan. 2019. http://www.open-std.org/jtc1/sc22/wg21/
Jeremy Siek and Andrew Lumsdaine. 2000ś2007. The Boost Concept Check Library (BCCL) (website). https://www.boost.
org/doc/libs/1_69_0/libs/concept_check/concept_check.htm (also at Internet Archive 16 Feb. 2019 12:45:11).
The Concept Check library allows one to add explicit statement and checking of concepts in the style of the proposed
C++ language extension.
The mechanisms use standard C++ and introduce no run-time overhead. The main cost of using the mechanism is in
compile-time.
Jeremy Siek and Walid Taha. 2006. A Semantic Analysis of C++ Templates. In Proceedings of the 20th European Conference
on Object-Oriented Programming (ECOOP) (Nantes, France). Springer-Verlag, Berlin and Heidelberg, Germany, 304ś327.
3540357262 https://doi.org/10.1007/11785477_19 LNCS 4067.
Cleiton Santoia Silva and Daniel Auresco. 2014. C++ type reflection via variadic template expansion. ISO/IEC JTC1/SC2/WG21:
C++ Standards Committee paper N3951. 7 Feb. 2014. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n3951.
Richard Smith. 2013. Relaxing constraints on constexpr functions. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper N3597. 15 March 2013. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2013/n3597.html (also at Internet
Archive 4 Sept. 2019 02:05:46).
Richard Smith. 2015. Guaranteed copy elision through simplified value categories. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper P0135R0. 27 Sept. 2015. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/p0135r0.html (also
Richard Smith (Ed.). 2017. ISO/IEC 14882:2017: Programming languages Ð C++. ISO (International Organization for
Standardization), Geneva, Switzerland (Dec.). 1605 book pages. https://www.iso.org/standard/68564.html
Richard Smith. 2018a. Another take on Modules. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0947R0.
12 Feb. 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0947r0.html (also at Internet Archive 4 Sept.
2019 02:05:46).
Richard Smith. 2018b. Another take on Modules (Revision 1). ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
P0947R1. 6 March 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0947r1.html (also at Internet
Archive 4 Sept. 2019 02:05:46).
Richard Smith. 2018c. Towards consistency between <=> and other comparison operators. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper P0946R0. 10 Feb. 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0946r0.
Richard Smith. 2019. Merging Modules. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1103R3. 22 Feb. 2019.
Richard Smith (Ed.). 2020. Working Draft, Standard for Programming Language C++. Number N4861. (April). 1834 book
pages. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/n4861.pdf (also at Internet Archive 27 April 2020
14:58:44).
Richard Smith and Gabriel Dos Reis. 2018. Merging Modules. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
21 April 2020 22:21:27).
Richard Smith and Gor Nishanov. 2018. Halo: coroutine Heap Allocation eLision Optimization: The joint response. ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper P0981R0. 18 March 2018. http://www.open-std.org/jtc1/sc22/wg21/
Richard Smith and Andrew Sutton. 2017. Semantic constraint matching for concepts. ISO/IEC JTC1/SC2/WG21: C++

Richard Smith, Andrew Sutton, and Daveed Vandevoorde. 2018a. Immediate functions. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper P1073R3. 6 Nov. 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p1073r3.
Richard Smith, Andrew Sutton, and Daveed Vandevoorde. 2018b. std::is_constant_evaluated(). ISO/IEC
Richard Smith, Daveed Vandevoorde, Geoffrey Romer, Gor Nishanov, Nathan Sidwell, Iain Sandoe, and Lewis Baker. 2019.
Coroutines: Language and Implementation Impact. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1492R0.
19 Feb. 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1492r0.pdf (also at Internet Archive 4 Sept.
2019 02:05:46).
Oleg Smolsky. 2014. Defaulted comparison operators. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N3950.
19 Feb. 2014. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n3950.html (also at Internet Archive 4 Sept.
2019 02:05:46).
Jeff Snyder and Chandler Carruth. 2013. Call for Compile-Time Reflection Proposals. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper N3814. 6 Oct. 2013. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2013/n3814.html
Yuriy Solodkyy, Gabriel Dos Reis, and Bjarne Stroustrup. 2013. Open Pattern Matching for C++. In Proceedings of the 2013
Companion Publication for Conference on Systems, Programming, & Applications: Software for Humanity (Indianapolis,
Indiana, USA, Oct.) (SPLASH ’13). Association for Computing Machinery, New York, NY, USA, 97ś98. 978-1450319959
Yuriy Solodkyy, Gabriel Dos Reisl, and Bjarne Stroustrup. 2012. Open and Efficient Type Switch for C++. In Proceedings of
the ACM International Conference on Object Oriented Programming Systems Languages and Applications (Tucson, Arizona,
USA, Oct.) (OOPSLA ’12). Association for Computing Machinery, New York, NY, USA, 963ś982. SIGPLAN Notices 47, 10
(Nov. 2012). 978-1450315616 https://doi.org/10.1145/2384616.2384686
Yuriy Solodkyy, Gabriel Dos Reisl, and Bjarne Stroustrup. 2014. Pattern Matching for C++. Nov. 2014. http://wiki.
edg.com/pub/Wg21urbana-champaign/EveningSessions/pattern_matching.pdf Evening session at WG21 meeting in
Urbana-Champaign, Illinois.
Mike Spertus. 2018. Extensions to Class Template Argument Deduction. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper P1021R0. 7 May 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p1021r0.html (also at Internet
Archive 4 Sept. 2019 02:05:46).
Mike Spertus, Timur Doumler, and Richard Smith. 2018. Filling holes in Class Template Argument Deduction. ISO/IEC
Mike Spertus and Richard Smith. 2015. Template parameter deduction for constructors (Rev. 3). ISO/IEC JTC1/SC2/WG21:
C++ Standards Committee paper P0091R0. 24 Sept. 2015. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/
The Stellar Group. 2014ś2020. HPX (High Performance ParalleX) (website). http://stellar.cct.lsu.edu/projects/hpx/ (also at
Internet Archive 27 Dec. 2019 05:13:30).
HPX (High Performance ParalleX) is a general purpose C++ runtime system for parallel and distributed applications
of any scale. It strives to provide a unified programming model which transparently utilizes the available resources to
achieve unprecedented levels of scalability. This library strictly adheres to the C++11 Standard and leverages the Boost
C++ Libraries which makes HPX easy to use, highly optimized, and very portable. HPX is developed for conventional
architectures including Linux-based systems, Windows, Mac, and the BlueGene/Q, as well as accelerators such as the
Xeon Phi. High Performance ParalleX is the first open-source implementation of the ParalleX execution model.
Alexander Stepanov. 1986. Scheme higher order programming library (code repository plus documentation). Aug. 1986.
http://stepanovpapers.com/schemenotes/index.html (also at Internet Archive 16 Aug. 2018 07:27:46).
Alex Stepanov and Paul McJones. 2009. Elements of Programming. Addison-Wesley, Reading, Massachusetts, USA. 288 pages.
978-0321635372
Arthur G. Stephenson et al. 1999. Mars Climate Orbiter Mishap Investigation Board Phase I Report. Technical Report. NASA
(10 Nov.). https://llis.nasa.gov/llis_lib/pdf/1009464main1_0641-mr.pdf (also at Internet Archive 5 March 2020 02:00:59).
Bjarne Stroustrup. 1982. Classes: An Abstract Data Type Facility for the C Language. SIGPLAN Notices 17, 1 (Jan.), 42ś51.
0362-1340 https://doi.org/10.1145/947886.947893 Published internally in Bell Labs in 1981.
Bjarne Stroustrup. 1985a. AT&T C++ Translator Release Notes. Nov. 1985. Archived at the Computer History Museum:
http://www.softwarepreservation.org/projects/c_plus_plus/ (also at Internet Archive 29 March 2020 16:45:18). Contains
the manual and tutorials (e.g., for stream I/O and tasks).
Bjarne Stroustrup. 1985b. The C++ Programming Language. Addison-Wesley, Reading, Massachusetts, USA. 0-201-12078-X
Bjarne Stroustrup. 1985c. A Set of C++ Classes for Co-routine Style Programming. In C++ Translator Release Notes.
AT&T, New York, NY, USA (Nov.). http://www.softwarepreservation.org/projects/c_plus_plus/cfront/release_e/doc/
ClassesForCoroutines.pdf (also at Internet Archive 18 July 2019 02:33:30). First published internally in 1980.
Bjarne Stroustrup. 1990ś2020. My C++ Standards committee papers (website with links to documents). http://www.
stroustrup.com/WG21.html (also at Internet Archive 20 Feb. 2020 04:39:09).
This is an incomplete list of papers that I (Bjarne Stroustrup) have written or co-authored for the C++ standards
committee. Please remember that these are incomplete proposals, discussions of alternatives, and exploration of the
design space. Note also that these papers were all written to a specific audience (the members of the C++ standards
committee), at a specific time (relative to the then current draft of the standard), and as part of a specific discussion.
They are typically not ideal introcutions for a general audience. Not all has led to or will lead to changes to C++, and
most of those will łmutatež significantly as they progress from idea/proposal to final language change.
Bjarne Stroustrup. 1991. The C++ Programming Language (2th Edition). Addison-Wesley, Reading, Massachusetts, USA.
0-201-53992-6
Bjarne Stroustrup. 1993. A History of C++. In The Second ACM SIGPLAN Conference on History of Programming Languages
(Cambridge, Massachusetts, USA, March) (HOPL-II ). Association for Computing Machinery, New York, NY, USA, 271ś297.
0897915704 https://doi.org/10.1145/154766.155375 Also published in History of Programming Languages (edited by T. J.
Bergin and R. G. Gibson), Addison-Wesley, ISBN 1-201-89502-1, 1996.
Bjarne Stroustrup. 1994. The Design and Evolution of C++. Addison-Wesley, Reading, Massachusetts, USA. 0-201-54330-3
Bjarne Stroustrup. 1997. The C++ Programming Language (3rd Edition). Addison-Wesley, Reading, Massachusetts, USA.
0-201-88954-4
Bjarne Stroustrup. 1998. Generalizing Overloading for C++2000. Overload Journal 25 (1 April), 20ś24. 1354-3172 https:
//accu.org/var/uploads/journals/overload25.pdf (also at Internet Archive 13 May 2011 09:37:09).
Bjarne Stroustrup. 2003. Concept checking Ð A more abstract complement to type checking. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper N1510. 22 Oct. 2003. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2003/n1510.pdf
Bjarne Stroustrup. 2004ś2020. Why can’t I define constraints for my template parameters? (website). http://www.stroustrup.
com/bs_faq2.html#constraints (also at Internet Archive 10 April 2020 05:20:05).
Bjarne Stroustrup. 2005. A rationale for semantically enhanced library languages. In Proceedings of the Workshop on Library-
Centric Software Design (LCSD’05), co-located with OOPSLA 2005 (San Diego, California, USA, 16 Oct.). Department
of Computer Science & Engineering, Texas A&M University, College Station, Texas, USA, 44. Archived at https:
//web.archive.org/web/20051118062506/http://lcsd05.cs.tamu.edu/papers/stroustrup.pdf
Bjarne Stroustrup. 2007. Evolving a Language in and for the Real World: C++ 1991ś2006. In Proceedings of the Third ACM
SIGPLAN Conference on History of Programming Languages (San Diego, California, June) (HOPL III ). Association for
Computing Machinery, New York, NY, USA, 4ś1ś4ś59. 978-1595937667 https://doi.org/10.1145/1238844.1238848
Bjarne Stroustrup. 2008a. Programming ś Principles and Practice Using C++. Addison-Wesley, Boston, Massachusetts, USA.
978-0321543721 There is also a second edition, ISBN 978-0321992789, 2014.
Bjarne Stroustrup. 2008b. Uniform initialization design choices (Revision 2). ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper N2532. 2 Feb. 2008. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2532.pdf (also at
Bjarne Stroustrup. 2009a. The C++0x łRemove Conceptsž Decision. Dr. Dobb’s Journal (July). https://www.drdobbs.com/
cpp/the-c0x-remove-concepts-decision/218600111 (also at Internet Archive 9 Oct. 2012 01:02:52).
Bjarne Stroustrup. 2009b. The C++0x łRemove Conceptsž Decision. Overload Journal 92 (Aug.). https://accu.org/index.php/
journals/1576 (also at Internet Archive 4 March 2010 22:57:32). Reprinted with permission from Dr. Dobb’s Journal.
Bjarne Stroustrup. 2009c. Simplifying the use of concepts. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
N2906. 21 June 2009. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2009/n2906.pdf (also at Internet Archive
4 Sept. 2019 02:05:46).
Bjarne Stroustrup. 2010a. łNewž Value Terminology. April 2010. http://stroustrup.com/terminology.pdf (also at Internet
Archive 2 Oct. 2012 20:01:36).
Bjarne Stroustrup. 2010b. What Should We Teach New Software Developers? Why? Communications of the ACM 53, 1
(Jan.), 40ś42. 0001-0782 https://doi.org/10.1145/1629175.1629192
Bjarne Stroustrup. 2012. Software Development for Infrastructure. Computer 45, 1 (Jan.), 47ś58. 0018-9162 https:
//doi.org/10.1109/MC.2011.353
Bjarne Stroustrup. 2013. The C++ Programming Language (4th Edition). Addison-Wesley, Boston, Massachusetts, USA.
978-0321563842
Bjarne Stroustrup. 2014a. Call syntax: x.f(y) vs. f(x,y). ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N4174.
11 Oct. 2014. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n4174.pdf (also at Internet Archive 4 Sept.
2019 02:05:46).
Bjarne Stroustrup. 2014b. Default comparisons. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N4175. 11 Oct.
2014. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n4175.pdf (also at Internet Archive 4 Sept. 2019
02:05:46).
Bjarne Stroustrup. 2014c. Thoughts about Comparisons. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N4176.
11 Oct. 2014. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n4176.pdf (also at Internet Archive 4 Sept.
2019 02:05:46).
Bjarne Stroustrup. 2014d. A Tour of C++. Addison-Wesley, Boston, Massachusetts, USA. 978-0321958310
Bjarne Stroustrup. 2015a. Thoughts about C++17. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N4492.
15 May 2015. http://open-std.org/jtc1/sc22/wg21/docs/papers/2015/n4492.pdf (also at Internet Archive 4 Sept. 2019
02:05:46).
Bjarne Stroustrup. 2015b. Unified call syntax concerns. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0131R0.
27 Sept. 2015. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/p0131r0.pdf (also at Internet Archive 4 Sept.
2019 02:05:46).
Bjarne Stroustrup. 2017a. Concepts: The Future of Generic Programming; or, How to Design Good Concepts and Use
Them Well. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0557R1. 31 Jan. 2017. http://www.open-
Bjarne Stroustrup. 2017b. Function declarations using concepts. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
4 Sept. 2019 02:05:46).
Bjarne Stroustrup. 2017c. Learning and Teaching Modern C++ (video). 25 Sept. 2017. https://youtu.be/fX2W3nNjJIo
Opening keynote for CppCon (the C++ Conference), Bellevue, Washington, USA (99 minutes).
Bjarne Stroustrup. 2018a. The Evils of Paradigms; or, Beware of one-solution-fits-all thinking. ISO/IEC JTC1/SC2/WG21:
C++ Standards Committee paper P0976R0. 6 March 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/
p0976r0.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
Bjarne Stroustrup. 2018b. A minimal solution to the concepts syntax problems. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper P1079. 6 May 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p1079r0.pdf (also at
Bjarne Stroustrup. 2018c. Modules and macros. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0955R0. 11 Feb.
02:05:46).
Bjarne Stroustrup. 2018d. Remember the Vasa! ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0977R0. 6 March
02:05:46).
Bjarne Stroustrup. 2018e. Subscripts and sizes should be signed. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
P1428R0. 18 Jan. 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1428r0.pdf (also at Internet Archive
4 Sept. 2019 02:05:46).
Bjarne Stroustrup. 2018f. A Tour of C++ (Second Edition). Addison-Wesley, Boston, Massachusetts, USA. 978-0134997834
Bjarne Stroustrup. 2018g. What do we want to do with reflection? ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper P0954R0. 11 Feb. 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0954r0.pdf (also at Internet
Archive 4 Sept. 2019 02:05:46).
Bjarne Stroustrup. 2019a. C++ exceptions and alternatives. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
P1947R0. 18 Nov. 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1947r0.pdf (also at Internet Archive
4 Sept. 2019 02:05:46).
Bjarne Stroustrup. 2019b. How can you be so certain? ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1962R0.
18 Nov. 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1962r0.pdf (also at Internet Archive 4 Sept.
2019 02:05:46).
Bjarne Stroustrup. 2019c. What to do about contracts? ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1711R0.
13 June 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1711r0.pdf (also at Internet Archive 4 Sept.
2019 02:05:46).
Bjarne Stroustrup et al. 1992. How to Write a C++ Language Extension Proposal for ANSI-X3J16/ISO-WG21. SIGPLAN
Notices 27, 6 (June), 64ś71. 0362-1340 https://doi.org/10.1145/130981.130989
Bjarne Stroustrup and Gabriel Dos Reis. 2003a. Concepts Ð Design choices for template argument checking. ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper N1522. 22 Oct. 2003. http://www.open-std.org/jtc1/sc22/wg21/docs/
Bjarne Stroustrup and Gabriel Dos Reis. 2003b. Concepts ś Syntax and composition. ISO/IEC JTC1/SC2/WG21: C++
Bjarne Stroustrup and Gabriel Dos Reis. 2003c. Templates aliases for C++. ISO/IEC JTC1/SC2/WG21: C++ Standards
Bjarne Stroustrup and Gabriel Dos Reis. 2005a. A concept design (Rev. 1). ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper N1782. April 2005. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2005/n1782.pdf (also at
Bjarne Stroustrup and Gabriel Dos Reis. 2005b. Initialization and initializers. ISO/IEC JTC1/SC2/WG21: C++ Standards
Bjarne Stroustrup and Gabriel Dos Reis. 2014. Operator Dot. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
N4173. 11 Oct. 2014. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n4173.pdf (also at Internet Archive
4 Sept. 2019 02:05:46).
Bjarne Stroustrup and Gabriel Dos Reis. 2016. Operator Dot (R3). ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper P0416R1. 16 Oct. 2016. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/p0416r1.pdf (also at Internet
Archive 4 Sept. 2019 02:05:46).
Bjarne Stroustrup and Herb Sutter (Eds.). 2014ś2020. C++ Core Guidelines (online document repository). http://github.
com/isocpp/CppCoreGuidelines/blob/master/CppCoreGuidelines.md (also at Internet Archive 24 Feb. 2020 12:59:15).
Continually updated.
Bjarne Stroustrup and Herb Sutter. 2015. Unified Call Syntax: x.f(y) and f(x,y). ISO/IEC JTC1/SC2/WG21: C++ Standards
Bjarne Stroustrup, Herb Sutter, and Gabriel Dos Reis. 2015. A brief introduction to C++’s model for type- and resource-safety.
Isocpp.org. Oct. 2015. http://www.stroustrup.com/resource-model.pdf (also at Internet Archive 10 April 2020 05:20:23).
Revised Dec. 2015.
Bjarne Stroustrup and Andrew Sutton (Eds.). 2012. A Concept Design for the STL. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper N3351. 13 Jan. 2012. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2012/n3351.pdf (also at
Herb Sutter. 2002. Proposed Addition to C++: Typedef Templates. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper N1406. 21 Oct. 2002. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2002/n1406.pdf (also at Internet
Archive 4 Sept. 2019 02:05:46).
Herb Sutter. 2012. async and ~future. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N3451. 23 Sept. 2012.
Herb Sutter. 2013a. ~thread Should Join. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N3636. 17 April 2013.
Herb Sutter. 2013b. AAA Style (Almost Always Auto) (blog post). 12 Aug. 2013. http://herbsutter.com/2013/08/12/gotw-94-
solution-aaa-style-almost-always-auto/ (also at Internet Archive 3 Dec. 2019 14:58:42).
Herb Sutter. 2014. Unified Call Syntax. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N4165. 4 Oct. 2014.
Herb Sutter. 2017a. Consistent comparison. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0515R0. 5 Feb.
02:05:46).
Herb Sutter. 2017b. Merge Concurrency TS atomic pointers into C++20 working draft. ISO/IEC JTC1/SC2/WG21: C++
pdf (also at Internet Archive 22 April 2020 02:23:22).
Herb Sutter. 2018a. Concepts in-place syntax. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0745R0. 11 Feb.
02:05:46).
Herb Sutter. 2018b. Zero-overhead deterministic exceptions: Throwing values. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper P0709R0. 2 May 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0709r0.pdf (also at
Herb Sutter. 2019. Lifetime safety: Preventing common dangling Ð version 1.1. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper P1179R1. 22 Nov. 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1179r1.pdf (also
Herb Sutter, Casey Carter, Gabriel Dos Reis, Eric Niebler, Bjarne Stroustrup, Andrew Sutton, and Ville Voutilainen. 2019.
Rename concepts to standard_case for C++20, while we still can. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper P1754R0. 16 June 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1754r0.pdf (also at Internet
Archive 4 Sept. 2019 02:05:46).
Herb Sutter and Bjarne Stroustrup. 2003. A name for the null pointer: nullptr. ISO/IEC JTC1/SC2/WG21: C++ Standards
Herb Sutter, Bjarne Stroustrup, and Gabriel Dos Reis. 2015. Structured bindings. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper P0144R0. 14 Oct. 2015. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/p0144r0.pdf (also
Andrew Sutton. 2017. Working Draft, C++ extensions for Concepts. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper N4674. 19 June 2017. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/n4674.pdf (also at Internet
Archive 4 Sept. 2019 02:05:46).
Andrew Sutton, Sam Goodrick, and Daveed Vandevoorde. 2019. Expansion statements. ISO/IEC JTC1/SC2/WG21: C++
Andrew Sutton and Richard Smith. 2014. Folding expressions. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
N4295. 7 Nov. 2014. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n4295.html (also at Internet Archive
4 Sept. 2019 02:05:46).
Andrew Sutton and Bjarne Stroustrup. 2011. Design of Concept Libraries for C++. In Software Language Engineering: 4th
International Conference (SLE 2011) (3 July 2011ś). Springer, Berlin and Heidelberg, Germany, 97ś118. 978-3-642-28829-6
https://doi.org/10.1007/978-3-642-28830-2_6 LNCS 6940.
Andrew Sutton and Herb Sutter. 2018. Value-based Reflection. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
P0993R0. 2 April 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0993r0.pdf (also at Internet Archive
22 April 2020 02:28:24).
Andrew Sutton, Faisal Vali, and Daveed Vandevoorde. 2018. Scalable Reflection in C++. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper P1240R0. 8 Oct. 2018. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p1240r0.
Clang Team. 2014. Clang 3.5 Documentation: Modules. http://clang.llvm.org/docs/Modules.html (also at Internet Archive
5 Dec. 2014 14:06:59).
Andrew Tomazos. 2015. Proposal of [[unused]], [[nodiscard]] and [[fallthrough]] attributes. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper P0068R0. 3 Sept. 2015. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/p0068r0.
Andrew Tomazos and Michael Spertus. 2014. Defaulted Comparison Using Reflection. ISO/IEC JTC1/SC2/WG21: C++
Hubert Tong and Faisal Vali. 2016. Smart References through Delegation: An Alternative to N4477’s Operator Dot. ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper P0352R0. 30 May 2016. http://www.open-std.org/jtc1/sc22/wg21/
James Touton and Mike Spertus. 2015. Template Argument Type Deduction. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper N4469. 10 April 2015. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/n4469.html (also at
James Touton and Mike Spertus. 2016. Declaring non-type template arguments with auto. ISO/IEC JTC1/SC2/WG21: C++
Standards Committee paper P0127R1. 4 March 2016. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/p0127r1.
Clay Trychta. 2016. Attributes for Likely and Unlikely Branches. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper P0479R0. 16 Oct. 2016. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/p0479r0.html (also at Internet
Archive 4 Sept. 2019 02:05:46).
TSO Wikipedia 2020. Memory ordering. https://wikipedia.org/wiki/Memory_ordering Archived at https://web.archive.
org/web/20200301003203/https://en.wikipedia.org/wiki/Memory_ordering
Faisal Vali, Herb Sutter, and Dave Abrahams. 2012. Proposal for Generic (Polymorphic) Lambda Expressions. ISO/IEC
JTC1/SC2/WG21: C++ Standards Committee paper N3418. 21 Sept. 2012. http://www.open-std.org/jtc1/sc22/wg21/docs/
Jan Christiaan van Winkel, Jose Daniel Garcia, Ville Voutilainen, Roger Orr, Michael Wong, and Sylvain Bonnal. 2017.
Operating principles for evolving C++. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0559R0. 31 Jan. 2017.
Daveed Vandevoorde. 2007. Modules in C++ (Revision 5). ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
N2316. 19 June 2007. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2007/n2316.pdf (also at Internet Archive
4 Sept. 2019 02:05:46).
Daveed Vandevoorde. 2009. New wording for C++0x Lambdas (rev. 2). ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper N2927. 15 July 2009. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2009/n2927.pdf (also at Internet
Archive 4 Sept. 2019 02:05:46).

Daveed Vandevoorde. 2012. Modules in C++ (Revision 6). ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
N3347. 11 Jan. 2012. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2012/n3347.pdf (also at Internet Archive
4 Sept. 2019 02:05:46).
Daveed Vandevoorde. 2017. Down with typename! ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0634.
5 March 2017. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0634r0.pdf (also at Internet Archive 4 Sept.
2019 02:05:46).
David Vandevoorde and Nicolai M. Josuttis. 2002. C++ Templates: The Complete Guide. Addison-Wesley, Reading, Mas-
sachussets, USA. 978-0201734843
David Vandevoorde, Nicolai M. Josuttis, and Douglas Gregor. 2018. C++ Templates Ð The Complete Guide (Second Edition).
Addison-Wesley, Boston, Massachussets, USA. 978-0-321-71412-1
vcpkg 2016ś2020. vcpkg, a C++ package manager for Windows, Linux and MacOS (part of Microsoft’s Visual Studio)
(website). https://docs.microsoft.com/en-us/cpp/build/vcpkg?view=vs-2019 (also at Internet Archive 25 Dec. 2019
14:16:14).
vcpkg is a command-line package manager for C++. It greatly simplifies the acquisition and installation of third-party
libraries on Windows, Linux, and MacOS. If your project uses third-party libraries, we recommend that you use vcpkg
to install them. vcpkg supports both open-source and proprietary libraries. All libraries in the vcpkg Windows catalog
have been tested for compatibility with Visual Studio 2015, Visual Studio 2017, and Visual Studio 2019. Between the
Windows and Linux/MacOS catalogs, vcpkg now supports over 1900 libraries. The C++ community is adding more
libraries to both catalogs on an ongoing basis.
Todd L. Veldhuizen. 2003. C++ Templates are Turing Complete. Indiana University Computer Science. Archived at
http://web.archive.org/web/20040919170502/http://osl.iu.edu/~tveldhui/papers/2003/turing.pdf Also at http://rtraba.files.
wordpress.com/2015/05/cppturing.pdf
Ville Voutilainen. 2016a. Discussion about std::thread and RAII. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
P0206R0. 27 Jan. 2016. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/p0206r0.html (also at Internet Archive
4 Sept. 2019 02:05:46).
Ville Voutilainen. 2016b. Presentation at WG21 meeting in Jacksonville, Florida. Feb. 2016.
Ville Voutilainen. 2016c. Why I want Concepts, and why I want them sooner rather than later. ISO/IEC JTC1/SC2/WG21:
C++ Standards Committee paper P0225R0. 5 Feb. 2016. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/
Ville Voutilainen. 2019a. Adding a global contract assumption mode. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper P1710R0. 17 June 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p1710r0.html (also at Internet
Archive 4 Sept. 2019 02:05:46).
Ville Voutilainen. 2019b. To boldly suggest an overall plan for C++23. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee
paper P0592R4. 25 Nov. 2019. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p0592r4.html (also at Internet
Archive 4 Sept. 2019 02:05:46).
Ville Voutilainen, Thomas Köppe, Andrew Sutton, Herb Sutter, Gabriel Dos Reis, Bjarne Stroustrup, Jason Merrill, Hubert
Tong, Eric Niebler, Casey Carter, Tom Honermann, and Erich Keane. 2018. Yet another approach for constrained
declarations. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P1141R0. 23 June 2018. http://www.open-
Ville Voutilainen and Daveed Vandevoorde. 2016. constexpr if. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper
4 Sept. 2019 02:05:46).
Ville Voutilainen and Jonathan Wakely. 2018. Integrating feature-test macros into the C++ WD (rev. 2). ISO/IEC
VStudio Wikipedia 2020. Microsoft Visual Studio. https://en.wikipedia.org/wiki/Microsoft_Visual_Studio (also at Internet
Archive 10 April 2020 22:12:27).
Philip Wadler and Stephen Blott. 1989. How to Make Ad-Hoc Polymorphism Less Ad Hoc. In Proceedings of the 16th ACM
SIGPLAN-SIGACT Symposium on Principles of Programming Languages (Austin, Texas, USA, Jan.) (POPL ’89). Association
for Computing Machinery, New York, NY, USA, 60ś76. 0897912942 https://doi.org/10.1145/75277.75283
Jonathan Wakely. 2018. Working Draft, C++ Extensions for Networking. ISO/IEC JTC1/SC2/WG21: C++ Standards
Wandbox 2016ś2020. Wandbox, a Japanese online environment for compiling and running C++ using a variety of compilers
(website). https://wandbox.org/ (also at Internet Archive 2 Jan. 2020 21:23:50).
Daniel Wasserrab, Tobias Nipkow, Gregor Snelting, and Frank Tip. 2006. An Operational Semantics and Type Safety Proof for
Multiple Inheritance in C++. In Proceedings of the 21st Annual ACM SIGPLAN Conference on Object-Oriented Programming
Systems, Languages, and Applications (Portland, Oregon, USA, Oct.) (OOPSLA ’06). Association for Computing Machinery,
New York, NY, USA, 345ś362. 1595933484 https://doi.org/10.1145/1167473.1167503
WebAssembly 2017ś2020. WebAssembly (Wasm), a portable binary code format primarily for executing in browsers
(website). https://webassembly.org/ (also at Internet Archive 1 April 2020 05:02:38).
WebAssembly (abbreviated Wasm) is a binary instruction format for a stack-based virtual machine. Wasm is designed
as a portable target for compilation of high-level languages like C/C++/Rust, enabling deployment on the web for
client and server applications.
WG14. 2007. Thread Cancellation. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N2455. 11 Oct. 2007.
WG21. 1989ś2020. C++ Standards Committee Papers (website). http://open-std.org/jtc1/sc22/wg21/docs/papers/ (also at
Internet Archive 4 Sept. 2019 02:05:46). Links to all official papers for ISO/IEC JTC1/SC2/WG21.
Jeremiah Willcock, Jaakko Järvi, Doug Gregor, Bjarne Stroustrup, and Andrew Lumsdaine. 2006. Lambda expressions and
closures for C++. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper N1968. 26 Feb. 2006. http://www.open-
std.org/jtc1/sc22/wg21/docs/papers/2006/n1968.pdf (also at Internet Archive 4 Sept. 2019 02:05:46).
Anthony Williams. 2012. C++ Concurrency in Action ś Practical Multithreading. Manning Publishing, Shelter Island, NY,
USA. 978-1933988771
Anthony Williams. 2018. C++ Concurrency in Action ś Practical Multithreading (2nd edition). Manning Publishing, Shelter
Island, NY, USA. 978-1617294693
wxWidgets 1992ś2020. wxWidgets, an open-source, cross-platform GUI library (website). http://www.wxwidgets.org/
(also at Internet Archive 9 April 2020 07:40:49).
wxWidgets is a C++ library that lets developers create applications for Windows, macOS, Linux and other platforms
with a single code base. It has popular language bindings for Python, Perl, Ruby and many other languages, and
unlike other cross-platform toolkits, wxWidgets gives applications a truly native look and feel because it uses the
platform’s native API rather than emulating the GUI. It’s also extensive, free, open-source and mature.
Jing Yang, Gogul Balakrishnan, Naoto Maeda, Franjo Ivančić, Aarti Gupta, Nishant Sinha, Sriram Sankaranarayanan, and
Naveen Sharma. 2012. Object Model Construction for Inheritance in C++ and Its Applications to Program Analysis.
In Compiler Construction: 21st International Conference (CC 2012) (Tallinn, Estonia), Michael O’Boyle (Ed.). Springer,
Berlin and Heidelberg, Germany, 144ś164. 978-3-642-28651-3 https://doi.org/10.1007/978-3-642-28652-0_8 Also at
https://link.springer.com/content/pdf/10.1007/978-3-642-28652-0_8.pdf (also at Internet Archive 29 July 2018 19:36:26).
LNCS 7210. Held as part of the European Joint Conferences on Theory and Practice of Software (ETAPS 2012).
Jeffrey Yasskin. 2014. String_view: a non-owning reference to a string, revision 7. ISO/IEC JTC1/SC2/WG21: C++ Standards
Committee paper N3921. 14 Feb. 2014. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n3921.html (also at
Mani Zandifar, Nathan Thomas, Nancy M. Amato, and Lawrence Rauchwerger. 2014. The STAPL Skeleton Framework. In
The 27th International Workshop on Languages and Compilers for Parallel Computing (LCPC 2014) (Hillsboro, Oregon,
USA, 17 Sept.), James Brodman and Peng Tu (Eds.). Springer International Publishing, Cham, Switzerland, 176ś190.
978-3-319-17472-3 https://doi.org/10.1007/978-3-319-17473-0_12 Also at https://www.researchgate.net/profile/Mani_
Zandifar/publication/279286000_The_stapl_Skeleton_Framework/links/56327c9208ae242468d9f8df.pdf LNCS 8967.
Victor Zverovich. 2019. Text Formatting. ISO/IEC JTC1/SC2/WG21: C++ Standards Committee paper P0645R9. 16 June 2019.
Victor Zverovich, Daniela Engert, and Howard E. Hinnant. 2019. Integration of chrono with text formatting. ISO/IEC

Thriving in ACrowded and Changing World

Uploaded by

Copyright:

Available Formats

Thriving in ACrowded and Changing World

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Thriving in ACrowded and Changing World

Uploaded by

Copyright:

Available Formats

Thriving in a Crowded and Changing World: C++ 2006–2020

BJARNE STROUSTRUP, Morgan Stanley and Columbia University, USA

5.1 Digit Separators 64

9.3.1 Modules 111

2 BACKGROUND: C++ 1979–2006

2.1 The First Decade

• A łnew functionž creates the run-time environment for member functions

2.2 The Second Decade

2.2.1 Language Features. The major language features of C++98 were:

void do_something ( Shape * p)

A simpler variant uses references instead of pointers:

We can now simplify our use:

vector < int >:: iterator q

The standard-library algorithm find_if traverses a sequence (delimited by a begin/end pair)

2.3 C++ in 2006

łIf that’s C++, I don’t like it either!ž

2.4 Other Languages

3 THE C++ STANDARDS COMMITTEE

3.1 The Standard

3.3 Impact on Design

3.4 Proposal Checklists

• Is the solution teachable? To whom? By whom?

4 C++11: IT FEELS LIKE A NEW LANGUAGE

4.1 C++11: Support for Concurrency

// ... use x ...

Even with C++11 support, I consider lock-free programming expert-level work.

class F { // traditional function object

double f( const vector < double >& v ); // traditional function

mutex m; // controlling mutex

In the end, C++11 offered:

auto f0 = async ( accum ,v0 , v0 + sz /4 ,0.0); // first quarter

void ff (Y y , promise <X >& p) // to execute f(y) asynchronously

auto fut = pro . get_future ();

4.2 C++11: Simplifying Use

auto x = f (1); // x is an int

It allowed people to line up names:

void use ( int x , char * p)

Channel auto y = flopscomps (x ,3); // y can be used as a Channel

Matrix operator +( const Matrix &, const Matrix &);

void use ( const Matrix & m1 , const Matrix & m2 )

clone_ptr & operator =( clone_ptr && p) // move assignment

string s; // initialization by default constructor

template < typename T > int foo (T );

S x = foo ({ string {" alpha "} ,12.3});

template < typename T > class vector {

vector < int > v3 {1 ,2 ,3 ,4 ,5}; // vector with 5 elements

int * p = 99 -55 -44; // the null pointer

int f( char *);

int x1 = f( nullptr ); // f( char *)

I pronounce nullptr as łnull pointer.ž

constexpr Imaginary operator "" i( long double x) { return Imaginary (x ); }

Now, 3.4i is an Imaginary and 1.2+3.4i is complex<double>(1.2,3.4). Mission accomplished!

regex pattern2 {R "(\ w {2}\ s *\ d {5}( -\ d {4})?)"}; // raw literal

[[ noreturn ]] void forever ()

4.3 C++11: Improving Support for Generic Programming

Here, xx will become 3+7.

As such, lambdas added significantly to the attraction of generic programming.

The value of pi is about 3.14159 ( unless you live in Indiana ).

using analysis_fp = double (*)( const vector < Student_info >&);

tuple <T1 ,T2 ,T3 > f (/.../) // nice declaration syntax