variant lite: A single-file header-only version of a C++17-like variant, a type-safe union for C++98, C++11 and later

Contents

• Example usage
• In a nutshell
• License
• Dependencies
• Installation
• Synopsis
• Features
• Reported to work with
• Building the tests
• Other implementations of variant
• Notes and references
• Appendix

Example usage

#include "nonstd/variant.hpp"

#include <cassert>
#include <string>

using namespace nonstd;

int main()
{
    std::string hello = "hello, world"; 

    variant< char, int, long, std::string > var;
    
    var =  'v' ; assert( get<  0 >( var ) == 'v' );
                 assert( get<char>( var ) == 'v' );
    var =   7  ; assert( get<int >( var ) ==  7  );
    var =  42L ; assert( get<long>( var ) == 42L );    
    var = hello; assert( get<std::string>( var ) == hello );
}

Compile and run

prompt>g++ -std=c++98 -Wall -I../include -o 01-basic.exe 01-basic.cpp && 01-basic.exe

In a nutshell

variant lite is a single-file header-only library to represent a type-safe union. The library aims to provide a C++17-like variant for use with C++98 and later. If available, std::variant is used.

Features and properties of variant lite are ease of installation (single header), freedom of dependencies other than the standard library and control over object alignment (if needed). variant lite shares the approach to in-place tags with any-lite, expected-lite and with optional-lite and these libraries can be used together.

Limitations of variant lite concern the number of alternative types and the number of visitor arguments. The maximum number of types and visitor arguments are configurable via script generate_header.py (default: 16 types, 5 visitor arguments). With C++98, the alternative types are required to be of different type and there's no move construction, move assignment and emplacement. variant lite does not provide allocator-extended constructors.

License

variant lite is distributed under the Boost Software License.

Dependencies

variant lite has no other dependencies than the C++ standard library.

Installation

variant lite is a single-file header-only library. Put variant.hpp in the include folder directly into the project source tree or somewhere reachable from your project.

Or, if you use the conan package manager, you might follow these steps:

1. Create source file ./main.cpp, e.g. with the contents of the example code above.

2. Create ./conanfile.txt file with a reference to variant-lite in the requires section:

   [requires]
   variant-lite/1.2.2  # variant-lite/2.0.0 when available
   
   [generators]
   cmake
   

3. Create ./CMakeLists.txt:

   cmake_minimum_required(VERSION 3.1)
   project(variant-example CXX)
   
   include(${CMAKE_BINARY_DIR}/conanbuildinfo.cmake)
   conan_basic_setup()
   
   add_executable(${PROJECT_NAME} main.cpp)
   target_link_libraries(${PROJECT_NAME} ${CONAN_LIBS})

4. Run the following commands:

   mkdir build && cd build
   conan install .. --settings arch=x86 --settings compiler=gcc
   cmake .. -G "Unix Makefiles" -DCMAKE_BUILD_TYPE=Release
   cmake --build . --config Release
   

Synopsis

Contents

• Types in namespace nonstd
• Interface of variant lite
• Algorithms for variant lite
• Information macros
• Configuration macros
• Macros to control alignment

Types in namespace nonstd

Interface of variant lite

Algorithms for variant lite

Note 1: visitor is limited to always return a Variant.

Information macros

variant_CONFIG_MAX_TYPE_COUNT
The maximum number of types thevariant can hold as configured via script generate_header.py.

variant_CONFIG_MAX_VISITOR_ARG_COUNT
The maximum number of visitor arguments as configured via script generate_header.py.

Configuration macros

Tweak header

If the compiler supports __has_include(), variant lite supports the tweak header mechanism. Provide your tweak header as nonstd/variant.tweak.hpp in a folder in the include-search-path. In the tweak header, provide definitions as documented below, like #define boolean_CPLUSPLUS 201103L.

Standard selection macro

-Dvariant_CPLUSPLUS=199711L
Define this macro to override the auto-detection of the supported C++ standard, or if your compiler does not set the __cplusplus macro correctly.

Select std::variant or nonstd::variant

At default, variant lite uses std::variant if it is available and lets you use it via namespace nonstd. You can however override this default and explicitly request to use std::variant or variant lite's nonstd::variant as nonstd::variant via the following macros.

-Dvariant_CONFIG_SELECT_VARIANT=variant_VARIANT_DEFAULT
Define this to variant_VARIANT_STD to select std::variant as nonstd::variant. Define this to variant_VARIANT_NONSTD to select nonstd::variant as nonstd::variant. Default is undefined, which has the same effect as defining to variant_VARIANT_DEFAULT.

Disable exceptions

-Dvariant_CONFIG_NO_EXCEPTIONS=0
Define this to 1 if you want to compile without exceptions. If not defined, the header tries and detect if exceptions have been disabled (e.g. via -fno-exceptions). Default is undefined.

Presence of variant_size_V() simulation macro

-Dvariant_CONFIG_OMIT_VARIANT_SIZE_V_MACRO=0
Define this macro to 1 to omit the variant_size_V(T) macro. Default is 0.

Presence of variant_alternative_T() simulation macro

-Dvariant_CONFIG_OMIT_VARIANT_ALTERNATIVE_T_MACRO=0
Define this macro to 1 to omit the variant_alternative_T(I,T) macro. Default is 0.

Macros to control alignment

If variant lite is compiled as C++11 or later, C++11 alignment facilities are used for storage of the underlying object. When compiled as pre-C++11, variant lite tries to determine proper alignment itself. If this doesn't work out, you can control alignment via the following macros. See also section Implementation notes.

-Dvariant_CONFIG_MAX_ALIGN_HACK=0
Define this to 1 to use the max align hack for alignment. Default is 0.

-Dvariant_CONFIG_ALIGN_AS=pod_type
Define this to the pod-type you want to align to (no default).

-Dvariant_CONFIG_ALIGN_AS_FALLBACK=pod_type
Define this to the pod-type to use for alignment if the algorithm of variant lite cannot find a suitable POD type to use for alignment. Default is double.

Reported to work with

The table below mentions the compiler versions variant lite is reported to work with.

Building the tests

To build the tests you need:

• CMake, version 2.8.12 or later to be installed and in your PATH.
• A suitable compiler.

The lest test framework is included in the test folder.

The following steps assume that the variant lite source code has been cloned into a directory named c:\variant-lite.

1. Create a directory for the build outputs for a particular architecture. Here we use c:\variant-lite\build-win-x86-vc10.

    cd c:\variant-lite
    md build-win-x86-vc10
    cd build-win-x86-vc10
   

2. Configure CMake to use the compiler of your choice (run cmake --help for a list).

    cmake -G "Visual Studio 10 2010" ..
   

3. Build the test suite in the Debug configuration (alternatively use Release).

    cmake --build . --config Debug
   

4. Run the test suite.

    ctest -V -C Debug
   

All tests should pass, indicating your platform is supported and you are ready to use variant lite.

Implementation notes

Object allocation and alignment

variant lite reserves POD-type storage for an object of the underlying type inside a union to prevent unwanted construction and uses placement new to construct the object when required. Using non-placement new (malloc) to obtain storage, ensures that the memory is properly aligned for the object's type, whereas that's not the case with placement new.

If you access data that's not properly aligned, it 1) may take longer than when it is properly aligned (on x86 processors), or 2) it may terminate the program immediately (many other processors).

Although the C++ standard does not guarantee that all user-defined types have the alignment of some POD type, in practice it's likely they do [10, part 2].

If variant lite is compiled as C++11 or later, C++11 alignment facilities are used for storage of the underlying object. When compiling as pre-C++11, variant lite tries to determine proper alignment using meta programming. If this doesn't work out, you can control alignment via three macros.

variant lite uses the following rules for alignment:

1. If the program compiles as C++11 or later, C++11 alignment facilities are used.

2. If you define -Dvariant_CONFIG_MAX_ALIGN_HACK=1 the underlying type is aligned as the most restricted type in struct max_align_t. This potentially wastes many bytes per variant if the actually required alignment is much less, e.g. 24 bytes used instead of the 2 bytes required.

3. If you define -Dvariant_CONFIG_ALIGN_AS=pod-type the underlying type is aligned as pod-type. It's your obligation to specify a type with proper alignment.

4. If you define -Dvariant_CONFIG_ALIGN_AS_FALLBACK=pod-type the fallback type for alignment of rule 5 below becomes pod-type. It's your obligation to specify a type with proper alignment.

5. At default, variant lite tries to find a POD type with the same alignment as the underlying type.

   The algorithm for alignment of 5. is:

   • Determine the alignment A of the underlying type using alignment_of<>.
   • Find a POD type from the list alignment_types with exactly alignment A.
   • If no such POD type is found, use a type with a relatively strict alignment requirement such as double; this type is specified in variant_CONFIG_ALIGN_AS_FALLBACK (default double).

Note that the algorithm of 5. differs from the one Andrei Alexandrescu uses in [10, part 2].

The class template alignment_of<> is gleaned from Boost.TypeTraits, alignment_of [14].

For more information on constructed unions and alignment, see [10-14].

Other implementations of variant

• Isabella Muerte. MNMLSTC Core (C++11).
• Anthony Williams. C++ variant class, matching the C++17 draft.
• Agustín Bergé. Eggs.Variant, a C++11/14 generic, type-safe, discriminated union.
• Chris Beck. strict_variant, a simpler type-safe union for C++11/C++14.
• Mapbox. C++11/C++14 Variant.
• Michael Park. Variant: A type-safe union (C++14).
• Gregor Burger. Typesafe tagged union with "don't call us, we'll call you" feature (C++14).
• Anatoliy Tomilov. C++1z variant, never empty, constexpr-enabled visit/multivisit, ... relational.

Notes and References

Acknowledgments

Thanks to @flexferrum for making the number of variant types and visitor arguments configurable. Thanks to @siffiejoe for contributing to fixing lifetime, noexcept and hash issues.

References

[1] CppReference. Variant.

[2] ISO/IEC WG21. N4606, section 20.7 Variants. July 2016.

[3] Axel Naumann. Variant: a type-safe union for C++17 (v8). June 2016.

[4] Peter Dimov. Valueless Variants Considered Harmful. March 2016.

[5] Anthony Williams. Standardizing Variant: Difficult Decisions. June 2015

[6] Andrzej Krzemieński. Constexpr unions. December 2012.

[7] Agustín Bergé. Eggs.Variant - Part I. August 2014.

[8] Agustín Bergé. Eggs.Variant - Part II (the constexpr experience). March 2015.

[9] Andrei Alexandrescu. An Implementation of Discriminated Unions in C++. August 2002.

[10] Andrei Alexandrescu. Generic: Discriminated Unions part 1, part 2, part 3. April 2002.

[11] Herb Sutter. Style Case Study #3: Construction Unions. GotW #85. 2009

[12] Kevin T. Manley. Using Constructed Types in C++ Unions. C/C++ Users Journal, 20(8), August 2002.

[13] StackOverflow. Determining maximum possible alignment in C++.

[14] Boost.TypeTraits, alignment_of ( code ).

Presentations

[15] Ben Deane. Using Types Effectively. CppCon 2016.

Appendix

A.1 Compile-time information

The version of variant lite is available via tag [.version]. The following tags are available for information on the compiler and on the C++ standard library used: [.compiler], [.stdc++], [.stdlanguage] and [.stdlibrary].

A.2 Variant lite test specification

variant: Disallows non-default constructible as first type
variant: Allows non-default constructible as second and later type (first: int)
variant: Allows non-default constructible as second and later type (first: monostate)
variant: Allows multiple identical types (C++11)
variant: Allows to default-construct variant
variant: Allows to copy-construct from variant
variant: Allows to move-construct from variant (C++11)
variant: Allows to move-construct if-noexcept from variant (C++11)
variant: Allows to obtain the index of the current type
variant: Allows to inspect if variant is "valueless by exception"
variant: Allows to copy-assign from variant
variant: Allows to copy-assign mutually empty variant
variant: Allows to copy-assign from empty variant
variant: Allows to copy-assign to empty variant
variant: Allows to move-assign from variant (C++11)
variant: Allows to move-assign mutually empty variant (C++11)
variant: Allows to move-assign from empty variant (C++11)
variant: Allows to move-assign to empty variant (C++11)
variant: Allows to construct from element value
variant: Allows to copy-construct from element
variant: Allows to move-construct from element (C++11)
variant: Allows to convert-copy-construct from element
variant: Allows to convert-move-construct from element (C++11)
variant: Allows to copy-assign from element value
variant: Allows to move-assign from element value
variant: Allows to copy-assign from element
variant: Allows to move-assign from element (C++11)
variant: Allows to convert-copy-assign from element value
variant: Allows to copy-construct from elements in intializer-list based on type (C++11)
variant: Allows to move-construct from elements in intializer-list based on type (C++11)
variant: Allows to in-place copy-construct element based on type (C++11)
variant: Allows to in-place move-construct element based on type (C++11)
variant: Allows to in-place copy-construct element based on index (C++11)
variant: Allows to in-place move-construct element based on index (C++11)
variant: Allows to in-place copy-construct elements from intializer-list based on type (C++11)
variant: Allows to in-place move-construct elements from intializer-list based on type (C++11)
variant: Allows to in-place copy-construct elements from intializer-list based on index (C++11)
variant: Allows to in-place move-construct elements from intializer-list based on index (C++11)
variant: Allows to copy-emplace element based on type (C++11)
variant: Disallows to copy-emplace non-unique element type on type (C++11)
variant: Allows to move-emplace element based on type (C++11)
variant: Allows to copy-emplace element based on index (C++11)
variant: Allows to move-emplace element based on index (C++11)
variant: Allows to copy-emplace elements from intializer-list based on type (C++11)
variant: Allows to move-emplace elements from intializer-list based on type (C++11)
variant: Allows to copy-emplace elements from intializer-list based on index (C++11)
variant: Allows to move-emplace elements from intializer-list based on index (C++11)
variant: Allows to swap variants, same index (member)
variant: Allows to swap variants, different index (member)
variant: Allows to visit contents (args: 1; configured max args: 5)
variant: Allows to visit contents (args: 2; configured max args: 5)
variant: Allows to visit contents (args: 3; configured max args: 5)
variant: Allows to visit contents, rvalue reference (args: 1; configured max args: 5)
variant: Allows to check for content by type
variant: Allows to get element by type
variant: Allows to get element by index
variant: Allows to get pointer to element or NULL by type
variant: Allows to get pointer to element or NULL by index
variant: Allows to compare variants
variant: Allows to swap variants, same index (non-member)
variant: Allows to swap variants, different index (non-member)
monostate: Allows to make variant default-constructible
bad_variant_access: Indicates invalid variant access
variant_size<>: Allows to obtain number of element types (configured max types: 16)
variant_size_v<>: Allows to obtain number of element types (C++14; configured max types: 16)
variant_size_V(): Allows to obtain number of element types (non-standard: macro; configured max types: 16)
variant_alternative<>: Allows to select type by index
variant_alternative_t<>: Allows to select type by index (C++11)
variant_alternative_T(): Allows to select type by index (non-standard: macro)
std::hash<>: Allows to obtain hash (C++11)
tweak header: reads tweak header if supported [tweak]