Introduction to C++11 - FSU Computer Science - Florida State ...

19 downloads 91 Views 171KB Size Report
Feb 1, 2012 ... C++11. Introduction to the New Standard. Alejandro Cabrera. February 1 ... C++ 98. ○ Boost C++ Libraries. ○ C++03 and C++TR1. ○ C++11 ...
C++11 Introduction to the New Standard

Alejandro Cabrera February 1, 2012 Florida State University Department of Computer Science

Overview ●

A Brief History of C++



Features Improving: ●

Overall Use



Meta-programming



Resource-critical/High-performance Programming



Class and Object-Oriented Design



New Libraries



How to Use it Today

A Brief History ●

C With Classes – C++ v2.0



C++98



Boost C++ Libraries



C++03 and C++TR1



C++11

C With Classes (1979) ●



Started as a result of Ph.D. thesis work of Bjarne Stroustrup Added to C: ●

Classes



Derived classes



Strong type checking



Inlining



Default function arguments

C++ v 1.0 (1983) ●



Added to C With Classes: ●

Virtual functions



Function name and operator overloading



References



Constants



User-controlled, free-store memory control



Improved type checking



Single-line comments

The C++ Programming Language 1e published (1985)

C++ v 2.0 (1989) ●



The C++ Programming Language 2e published (1991) Added to C++ v 1.0:



Late additions:



Multiple inheritance



Templates



Abstract classes





Static member functions

Exceptions



Namespaces



New-style casts



Boolean type

● ●

const member functions Protected members

C++ 98 (1998) ● ●



C++ is standardized (ISO/IEC 14882:1998) Standard includes language core and standard library Standard library includes C library, containers, I/O streams, and much more

Boost C++ Libraries (199x) ●



The Boost C++ libraries provide portable, efficient implementations of many C++ library components Often times, libraries that appear in later standards make their debut in Boost



Boost releases go back as far as 1999



For more information, visit: http://www.boost.org/

C++ 03 (2003) ●



Correction to C++ standard is published (ISO/IEC 14882:2003) TR1, an extension to the standard library, was also published around this time (2005) ● ●

TR1 is only a draft, not a standard Most features detailed in TR1 became part of 2011 standard

C++ 11 (2011) ●





Unanimously approved by C++ committee in August 2011 C++11 brings many new features, many that make it more apparent that C++ is a hybrid language This presentation will cover most of the new features in detail

History Summary ●

Early C++ - 1979 – 1998



C++, first standard – 1998 ●

Birth of Boost C++ Libraries - 1999



C++, corrected standard – 2003



C++ TR1 – 2005



C++11 – 2011



C++ TR2 – W.I.P.

Features Overview ●

Easier programming



Easier meta-programming



Facilitated, performance-critical programming



Better class design

Easier Programming Introduction ●



C++11 introduces: ●

new keywords



new constructs



fixes parsing issues



adds support for garbage collection

I'll demonstrate how this leads to easier programming

Easier Programming auto keyword ● ●



auto is not a new keyword in C++ auto used to declare a given variable as a local variable Now, auto is used to have the compiler perform type-inference for you

Easier Programming auto keyword ●

Uses of auto: ● ●

● ●

Simplify iterator use in iteration constructs Simplify type declarations in meta-programming contexts

For more examples, refer to: auto.cpp auto also meshes well with the new range-for loop!

Easier Programming enum class ●

Old-style enumerations have three problems: ●

they implicitly covert to integers



they are globally visible, introducing name-clashing



the underlying type cannot be declared: – –

introducing compatibility problems cannot be forward declared

Easier Programming enum class ●

While old-style enums are still supported, newstyle enums address all their weaknesses: ●

Using an enum class variable where an integer is expected raises a compiler error: –







Explicit cast required

Scope of enum class declaration is limited to scope of declaration Underlying type can be specified

For examples, refer to: enum_class.cpp

Easier Programming constexpr keyword ●





For initializing complicated constants, sometimes functions are desired For example, initializing the size of a static buffer that depends on: ●

Number of CPUs available on system



Amount of memory available on system



Speed of CPUs on system



Maximum stack-size on system

However, functions cannot be used to accomplish this in C or C++ - macros must be used ●

Macros are not type-safe

Easier Programming constexpr keyword ●







Introducing constexpr: ●

Used to provide compile-time expressions



Can be used with user-defined types



Guarantees initialization at compile-time

Used as a function qualifier before the function's return type Also allows for more extensive compileroptimization opportunities! For more details, refer to: constexpr.cpp

Easier Programming nullptr keyword ● ●

Minor syntactic sugar improvement Previous efforts to define a clear use of null revolved around macros ●





#define NULL ((void *) 0)

Now, nullptr can be used to express null-pointer initialization where that is the intent. For more details, refer to: nullptr.cpp

Easier Programming range-for Loop Construct ● ●







Major syntactic sugar upgrade Provides a for loop similar to for-each loop provided in other high-level languages Can be used on any sequence for which a begin() and end() function are defined! Especially useful for sequential one-pass algorithms and function applications For more details, refer to: rangefor.cpp

Easier Programming Initializer Lists ●









Initializer lists allow the construction of any type using a curly-brace enclosed, comma-separated list of elements Previously, this style of construction was limited to primitive arrays with unspecified size Especially useful for providing easier to use interfaces for container types Critical for the support of a new feature, unified initialization For more details, refer to: ilist.cpp

Easier Programming Unified Initialization ●

Currently, there are many ways to initialize



C:





X a = {v}; // structs, arrays



X a = v; // primitives

C++: ●

new X(v); // heap allocation



X a(v); // for classes with constructors



X(v); // temporaries



X(v); // functional-style casts

Easier Programming Unified Initialization ●





The problem is, looking at a huge code base, it's impossible to tell just by looking if X(v) is a casting operation or a construction Solution: unified initialization ●

X v{1, 2, 3};



vector vec{1,2,3,4};



float *f{new float[10]};



int three{3};



ShadingMode shade{ShadingMode::SMOOTH};

For more details, refer to: unified_initialization.cpp

Easier Programming Right-Angle Bracket Parse Fix ●

C++98/03 syntax specification did not account for: ●





vector x;

Frequently, this was interpreted as either a right shift operation or input operator C++98/03 solution: ●

vector x; // space them out



C++11 fixes this.



An example is provided: right_bracket_parse.cpp

Easier Programming Improved POD Rules ●

POD: Plain Old Data type, a.k.a., standard layout types ●





Allows an object to be initialized using memcpy, and serialized/loaded using read/write directly An important concept for serialization and optimization

C++98/03 only considered objects avoiding use of certain language features PODs ●

Cannot use virtual functions



Cannot have constructor/destructor



Cannot allocate memory on heap

Easier Programming Improved POD Rules ●





C++11 expands set of objects considered PODs: ●

Recursively, if all members are PODs, so is the whole



No virtual functions



No virtual bases



No references



No multiple access specifiers (protected, public, private)

The biggest addition is that POD types may now have constructors/destructors For more details, refer to: improved_pod.cpp ●

Similar improvements have been made for unions!

Easier Programming Long Longs ●



Guarantees availability of [unsigned] long long type See long_long.cpp

Easier Programming User-Defined Literals ●



Various literal types have built-in support ●

12 // int



'a' // char



“as” // null-terminated string



1.2f // float



1.2lf // double



0xD0 // hexadecimal

However, there is no support for literals of user-defined types ●

123s // seconds?



“hello!”s // std::string

Easier Programming User-Defined Literals ●



C++11 introduces literal operators to create user-defined literals By overloading a single function, one can create literals with a given suffix for: ●

integers



floats



strings



characters

Easier Programming User-Defined Literals ●



This is extremely useful for creating DomainSpecific Languages (DSLs), or say, implementing a unit system ●

25.1s // seconds



12.5mps // meters per seconds



Distance m = 12.5mps * 25.1s

For more details, refer to: user_defined_literals.cpp ●

Requires GCC 4.7 (svn trunk, unreleased)

Easier Programming Raw Strings ●

C++98 provides no means to escape strings for use with regular expression engines ●



Example: match all lines containing two words separated by a backslash ●



Language escaping rules get in the way of correctly writing regular expressions

Solution: “\\w\\\\\\w”

C++11 makes this much easier by providing raw strings.

Easier Programming Raw Strings ●

Again, with raw strings: ●





Solution: R”(\w\\\w)”

More examples: ●

Quoted string: R”(“quoted string”)”



Regex: R”([\d\d\d]{3})”

Even the delimiter pair “()” is only a default: ●

Parenthetical string: R”*(“(Tough cookie.)”)*” –



Delimiter string is this color to ease reading

For more examples, see: raw_strings.cpp

Easier Programming Lambda Functions ●



In C++, to work with the standard algorithms library, one often had to create function objects For example, to count all elements less than 5: 1. Create a function object LessThan returns true if an input argument is less than 5 2.Pass LessThan to count_if as the final parameter.

Easier Programming Lambda Functions ●

This has a few undesirable consequences ●

LessThan potentially pollutes the global namespace –





Furthermore, function objects created at non-global scope cannot be used as template arguments

To implement simple logic, an entire class had to be implemented (~5 LOC for easy cases)

Lambdas were created to address these problems

Easier Programming Lambda Functions ●



Lambdas are anonymous, shorthand function objects Note: C++11 did not invent lambdas. They appear in many modern languages, including: ●

Haskell, Python, C#, …

Easier Programming Structure of Lambda Functions ●

C++11 lambdas have a structure that takes some getting used to ● ●

● ●

The capture list (optionally) The argument list [same as function argument list] The function body (optionally) The return type using suffix return type syntax –

Only if the return type cannot be deduced, which is true most of the time for function bodies longer than a single statement

Easier Programming Lambda Functions: Capture List ●

It can take on a few forms: ●







[] take none of the variables from the enclosing scope [&] take all the variables from the enclosing scope by reference [=] take all the variables from the enclosing scope by value

There is also support for capture-by-name, to capture only part of the enclosing scope

Easier Programming Lambda Functions: Return Type ●



C++11 introduces a new way to specify return types that apply for the decltype keyword and for lambdas: suffix return style In a simple lambda, it looks like: ●



[] (int a, int b) -> bool { return a < b; }

For many examples, refer to: lambda.cpp

Easier Programming Lambda Functions: Advice ●



Benefits: ●

Concise, terse



Does not pollute namespace



Great for one- or two-liners

Cons: ●

Less readable – intent may not be clear



Can easily create obfuscated code!





Capturing entire enclosing scope by value can be very expensive – be careful!

Keep in mind that C++11 now allows local-types as template arguments ●

A local, function object class may be the best solution!

Easier Programming Generalized Attributes ●





Crafted as an attempt to unify various compilerspecific extensions Generalized attributes will not be discussed further in this presentation, as support is nearly non-existent For further information: ●

Generalized Attributes ISO/IEC Paper

Easier Programming Garbage Collection ●







Garbage collection serves as a means to automatically clean up heap memory that cannot be reached Support for this has been experimental for a long time in C/C++ The standard now makes provisions for adding support if a compiler wishes to implement garbage collection For more details, see: ●

GC ISO/IEC Paper

Easier Programming Summary ●

New keywords ●



New constructs ●



Right-angle bracket parse

Improved usability ●



range for-loop, initializer lists, unified initialization

Language fixes ●



auto, enum class, constexpr, nullptr

Less restrictive unions and PODs, long longs, user-defined literals, raw strings, lambdas, generalized attributes, enum class, local types as template args

Support for garbage collection

Easier Meta-programming Introduction ●





Meta-programming and related techniques allow for entire libraries of highly-optimized, highly-customizable software to be generated from a limited source base Meta-programming also adds new ways to compose classes and functions C++11 adds a few features to facilitate metaprogramming

Easier Meta-programming Overview ●

decltype



constexpr



static_assert



variadic templates



library

Easier Meta-programming decltype keyword ●

decltype used in conjunction with the new return type syntax can facilitate template programs

Consider: template ●

??? mul(T x, U y) { return x * y; } ●

What is the return type? (assuming operator*(x,y) is defined)

Easier Meta-programming decltype keyword Here is the solution using C++11: template ●

auto (T x, U y) -> decltype(x*y) { return x * y; } ●

decltype expressions are evaluated at compile-time



No examples are provided for decltype

Easier Meta-programming static_assert keyword ●





static_assert is extremely valuable for placing compile-time constraints on templateparameters For example, in combination with the new type_traits library, one could...: ●

Allow only integral parameters using is_integer



Allow only values of N less than 16 for Factorial



Allow only POD types for MakePacket

For an example, refer to: static_assert.cpp

Easier Meta-programming Variadic Templates ●



One of the features behind some of the most amazing wizardry that goes on in metatemplate programming Previously, recursive template instantiation was required to emulate type-lists ●



Could get very expensive for compiler

Variadic templates essentially allow recursion to be replaced by iteration

Easier Meta-programming Variadic Templates ●

Two components: ●

Declaration

template void printf(const char *format, T value, Args... args); ●

Iteration – –



Uses either iteration over run-time arguments or iteration over compile-time size of Args parameter pack In latter case, uses sizeof...() operator on Args

For more information, search for implementation inside your compiler source.

Easier Meta-programming Library ●



Provides various utilities to assist with metaprogramming Used primarily to ask questions about a type: ●

Is type T integral?



Is type T default constructible?



Does type T have a particular operator defined?

Easier Meta-programming Summary ● ●

● ●

decltype – return type inference constexpr – generalized compile-time expressions static_assert – compile-time constraints variadic templates – more efficient and usable template instantiation

Efficient Programming Introduction ● ●





C++11 was designed with efficiency in mind A few features were added to the language to enable more efficient implementations of various familiar operations Perhaps the most significant feature added in this regard is the rvalue reference, that enables perfect forwarding This feature and more will be explained over the next few slides

Efficient Programming Overview ●

std::move and rvalue references



noexcept expression



alignment support ●

alignas



alignof

Efficient Programming rvalue references ●

An rvalue is any temporary that occurs on the right side of an assignment ●





As compared to lvalues, which are storage locations on the left side of an assignment

C++11 brings rvalue references, denoted by a type signature of T&& These are meant to be used in conjunction with the std::move function

Efficient Programming std::move Function ●

std::move takes as an argument any one type and returns an rvalue reference of that type ●



Does not trigger copy constructor

This is very useful for: ●

Implementing move constructors



Implementing swap operations

Efficient Programming Move: What Does it Mean? ●

A move operation is intended to be a destructive copy



Instead of copying data over, move implemented correctly...







Assigns the address of source pointers to destination pointers



Copies over primitive values



Sets source pointers to nullptr



Either ignores source primitive values or sets them to 0

Since all pointers in the source are set to nullptr, they are not destructed This allows the execution of a technique known as perfect function forwarding

Efficient Programming Perfect Function Forwarding ●

A move operation is intended to be a destructive copy



Instead of copying data over, move implemented correctly...







Assigns the address of source pointers to destination pointers



Copies over primitive values



Sets source pointers to nullptr



Either ignores source primitive values or sets them to 0

Since all pointers in the source are set to nullptr, they are not destructed This allows the execution of a technique known as perfect function forwarding ●



Very useful for factory functions, in particular

For more details, refer to: ●

C++ Rvalue References Explained

Efficient Programming noexcept Expression ●

Declares to the compiler that a given function will NEVER propagate an exception ●



A function declared as noexcept that encounters an exception will immediately terminate the program

Allows the compiler to optimize those functions better

Actual form is: function_type ●

name(...) noexcept[(expression)] { }

Efficient Programming noexcept Expression ●



By default, it occurs as noexcept(true), and is used as noexcept If the expression given in the parentheses can throw, then noexcept is disabled ●



Allows for flexible conditional enabling of the noexcept feature Very important for generic programming!

Efficient Programming noexcept Suggestions ●

The following are great functions to decorate with noexcept: ●

Destructors – these should never throw



Move constructors



Functions that were not designed to handle exceptions –



noexcept serves as both an indicator to the compiler and documentation for humans!

For more details, refer to: noexcept.cpp

Efficient Programming Alignment Support ●







Discussion of this feature will be limited as support for it is currently very limited Two new operators are added to C++: ●

alignas



alignof

alignas allows for a memory region to be aligned on a specified boundary ●

alignas(double) unsigned char c[1024];



alignas(16) float[100];

alignof returns the alignment of a give type ●

const size_t n = alignof(float);

Efficient Programming Summary ●

rvalue references ●



std::move ●



Function that performs a move operation

noexcept expression ●



Enable perfect forwarding, expressed as T&&

Guarantees a function will not throw an exception

alignment support ●

alignas, alignof – Particularly relevant for serialization and SIMD programming

Class Design Introduction ●

Good class design is fundamental for making the most of C++



C++11 adds several features that target the realm of classes







Some of these facilitate the implementation, some assist/enforce the interface Some particularly salient features include: ●

Delegating constructors



Constructor control



Non-static member initialization

These features and more are covered in the following slides

Class Design Overview ●

Controlling defaults: ●



default and delete

New constructors: move and initializer list constructor, move assignment



Delegating constructors



Inheriting constructors



Non-static member initialization



Inheritance control: override, final



Explicit conversion operators



Inline name spaces for version support

Class Design Controlling Silent Defaults ●



The compiler has always silently generated various class members if any one of them is defined ●

Constructor: Default and copy



Destructor



Copy assignment

Sometimes, you don't want to allow copying of a resource

Class Design Controlling Silent Defaults ●

C++98/03 solution: ●



Declare class NonCopyable with private copy assignment and copy constructor Have new class publicly inherit from NonCopyable

C++11 solution: class X { ●

X(const X&) = delete; const X& operator=(const X&) = delete; };

Class Design Controlling Silent Defaults C++11 also allows you to communicate whether a default member is used: class X { ●

X() = default; ~X() = default; }; ●

For more details, refer to: class_control.cpp

Class Design New Constructors ●







C++11 gives you access to two new types of constructors: ●

Initializer list constructor



Move constructor



Move assignment operator

The first can be used to provide convenient use of your class The latter two are important to avoid unnecessary memory copying Implementation examples are given in: move.cpp

Class Design Delegating Constructors ●

Many modern OO languages allow you to implement other constructors in terms of one constructor ●

C++98/03 does not



Beginning with C++11, this is now possible!



For details, refer to: delegating_constructor.cpp ●

Support requires GCC >= 4.7

Class Design Inheriting Constructors ●

Allows members available in base class to be used in derived class



Not currently supported by any compiler



For details, refer to: inheriting_constructors.cpp

Class Design Non-static Member Initialization ●



Previously, it was a compilation error to give default values to non-static class members C++11 makes this possible using the new initialization syntax ●



Allows for simplified constructors

For details, refer to: member_defaults.cpp ●

Support requires GCC >= 4.7

Class Design Inheritance Control ●

C++11 introduces two new keywords for constraining designs through inheritance ●

final –



override –





Prevent virtual function from being overriden. Used to clearly express the intent that a given derived class function is meant to provide a new implementation for a base class function Helps the compiler flag errors where one accidentally overloads rather than overrides

For more details, refer to: inheritance_control.cpp

Class Design Explicit Conversion ●

Allows the use of the explicit keyword in conversion operators now ●



This feature has a limited range of use ●



Disables conversion in implicit context, if that is the desired result It primarily targets the case where you want to only allow conversion from one class to another during construction, but not in function call or copy contexts

For details, refer to: explicit_conversion.cpp

Class Design Inline Namespaces ●



Used primarily to provide version control within the language Rules: ● ●

Can only be used within an enclosing namespace To reference elements of inline namespace, need only use name of enclosing namespace –



Inline namespace is invisible to external code

For examples, refer to: inline_namespace.cpp

Class Design Summary ●

Improvements to constructors ●



Ability to control available constructors, new move and initializer list constructors, delegating constructors, inheriting constructors

Additional improvements: ●

Non-static member initialization, ability to control inheritance, explicit conversion operators, inline namespaces

New Libraries ●



New containers: ●

- hash-implemented [multi]sets and [multi]maps



– singly-linked list



– compile-time-sized array class



– type container

Concurrency support: ●

threads, mutexes, locks, condition variables, promises, futures, atomics



Random number generators



Regular expressions



Compile-time rational arithmetic



Smart pointers



Time management



For examples, refer to: libraries/*.cpp

Using C++11 Now ●

C++11 depends on the availability of a fairly recent compiler ●





GCC >= 4.6, LLVM Clang >= 3.0, MSVC >= 11.0

In many cases, the feature that you hope to use is available on only GCC and/or LLVM Clang ●



Standard draft was published in early 2011, final draft in August 2011.

For a few features, no compiler implements them

If your goal is portability across compilers, DO NOT use C++11 ●

C++`11 is not available on linprog

Using C++11 Now ●

Compiler support



Features lacking support

Using C++11 Now Compiler Support ●

GCC C++11 Page



Apache Compiler Support Matrix



Clang LLVM C++11 Support



Intel Compiler C++11 Support



Visual Studio C++11 Compiler Support

Using C++11 Now Features Lacking Support ●



The following is a list of features that no or almost no compiler supports: ●

alignas (Clang 3.0)



alignof (GCC 4.5, Clang 2.9)



constexpr (GCC 4.6, Clang 3.1)



Initializer Lists (GCC 4.4)



Raw-string Literals (GCC 4.5, Clang All)



Template alias (GCC 4.7, Clang 3.0, MSVC 12.1)



Unrestricted unions (GCC 4.6, Clang 3.0)



Range-for loop (GCC 4.6, Clang 3.0)



Generalized attributes (MSVC 12.1 )



Non-static member initialization (GCC 4.7, Clang 3.0)

In short, if you want the most salient C++11 features now, use GCC ●

If portability is important, use Boost C++11 emulation layers

#include #include using namespace std; int main() { string msg = R”(“Thanks!”)”; for (auto x : msg) cout