This is the official repository for Microsoft's implementation of the C++ Standard Library (also known as the STL), which ships as part of the MSVC toolset and the Visual Studio IDE.
- Our Changelog tracks which updates to this repository appear in each VS release.
- Our Status Chart displays our overall progress over time.
- Join our Discord server.
If you're a programmer who just wants to use the STL, you don't need this repo. Simply install the Visual Studio IDE and select the "Desktop development with C++" workload.
If you want to participate in the STL's development, welcome! You can report issues, comment on pull requests, and learn about what we're working on. You can also submit pull requests to fix bugs or add features: see CONTRIBUTING.md for more information.
Finally, you can take our code and use it in other apps and libraries (according to the terms of our license, like everything else).
We're in the process of moving all of our work on the STL to GitHub. Current status:
-
Code: Done. Our source code is available under the Apache License v2.0 with LLVM Exception. (See LICENSE.txt and NOTICE.txt for more information.)
-
Build System: In progress. We're working on a CMake build system, which is currently capable of building one flavor of the STL (native desktop). We need to extend this to build all of the flavors required for the MSVC toolset (e.g.
/clr
,/clr:pure
, OneCore, Spectre). Until that's done, we're keeping our legacy build system around in thestl/msbuild
subdirectory. (We're keeping those files in this repo, even though they're unusable outside of Microsoft, because they need to be updated whenever source files are added/renamed/deleted. We'll delete the legacy machinery as soon as possible.) -
Tests: In progress. We rely on three test suites: std, tr1, and libcxx. We've partially ported std and tr1, and fully ported libcxx to run under lit using the various configurations/compilers we test internally.
-
Continuous Integration: In progress. We've set up Azure Pipelines to validate changes to the repository. Currently, it builds the STL (native desktop for x86, x64, ARM, and ARM64). Also, it strictly verifies that all of our files have been formatted with clang-format and follow our other whitespace conventions.
-
Contribution Guidelines: Coming soon. Working on the STL's code involves following many rules. We have codebase conventions, Standard requirements, Microsoft-specific requirements, binary compatibility (ABI) requirements, and more. We're eager to begin accepting features and fixes from the community, but in addition to setting up a CI system, we need to write down all of the rules that are currently stored in our brains. (The ABI rules may be useful to other C++ libraries.)
-
Issues: In progress. We're going to use GitHub issues to track all of the things that we need to work on. This includes C++20 features, LWG issues, conformance bugs, performance improvements, and other todos. There are approximately 200 active bugs in the STL's Microsoft-internal database; we need to manually replicate all of them to GitHub issues. Currently, the cxx20 tag and LWG tag are done; every remaining work item is tracked by a GitHub issue. The bug tag and enhancement tag are being populated.
-
Plans: In progress. We're writing up our Roadmap.
We're implementing the latest C++ Working Draft, currently N4950, which will eventually become the next C++
International Standard. The terms Working Draft (WD) and Working Paper (WP) are interchangeable; we often
informally refer to these drafts as "the Standard" while being aware of the difference. (There are other relevant
Standards; for example, supporting /std:c++14
and /std:c++17
involves understanding how the C++14 and C++17
Standards differ from the Working Paper, and we often need to refer to the C Standard Library and ECMAScript regular
expression specifications.)
Our primary goals are conformance, performance, usability, and compatibility.
-
Conformance: The Working Paper is a moving target; as features and LWG issue resolutions are added, we need to implement them. That can involve a lot of work because the STL is required to behave in very specific ways and to handle users doing very unusual things.
-
Performance: The STL needs to be extremely fast at runtime; speed is one of C++'s core strengths and most C++ programs use the STL extensively. As a result, we spend more time on optimization than most general-purpose libraries. (However, we're wary of changes that improve some scenarios at the expense of others, or changes that make code significantly more complicated and fragile. That is, there's a "complexity budget" that must be spent carefully.)
-
Usability: This includes parts of the programming experience like compiler throughput, diagnostic messages, and debugging checks. For example, we've extensively marked the STL with
[[nodiscard]]
attributes because this helps programmers avoid bugs. -
Compatibility: This includes binary compatibility and source compatibility. We're keeping VS 2022 binary-compatible with VS 2015/2017/2019, which restricts what we can change in VS 2022 updates. (We've found that significant changes are possible even though other changes are impossible, which we'll be documenting in our Contribution Guidelines soon.) While there are a few exceptions to this rule (e.g. if a feature is added to the Working Paper, we implement it, and then the feature is significantly changed before the International Standard is finalized, we reserve the right to break binary compatibility because
/std:c++latest
offers an experimental preview of such features), binary compatibility generally overrides all other considerations, even conformance. Source compatibility refers to being able to successfully recompile user code without changes. We consider source compatibility to be important, but not all-important; breaking source compatibility can be an acceptable cost if done for the right reasons in the right way (e.g. in a controlled manner with escape hatches).
There are things that we aren't interested in doing with this project, for various reasons (most importantly, we need to focus development effort on our goals). Some examples:
-
Non-goal: Porting to other platforms.
-
Non-goal: Adding non-Standard extensions.
-
Non-goal: Implementing Technical Specifications. (We're prioritizing features in the Working Paper. Occasionally, we might implement some or all of a TS, often when we're working on the specification itself.)
If you're proposing a feature to WG21 (the C++ Standardization Committee), you're welcome (and encouraged!) to use our code as a base for a proof-of-concept implementation. These non-goals simply mean that we're unable to consider pull requests for a proposed feature until it has been voted into a Working Paper. After that happens, we'll be delighted to review a production-ready pull request.
You can report STL bugs here, where they'll be directly reviewed by maintainers. You can also report STL bugs through Developer Community, or the VS IDE (Help > Send Feedback > Report a Problem...).
Please help us efficiently process bug reports by following these rules:
-
Only STL bugs should be reported here. If it's a bug in the compiler, CRT, or IDE, please report it through Developer Community or Report A Problem. If it's a bug in the Windows SDK, please report it through the Feedback Hub app. If you aren't sure, try to reduce your test case and see if you can eliminate the STL's involvement while still reproducing the bug.
-
You should be reasonably confident that you're looking at an actual implementation bug, instead of undefined behavior or surprising-yet-Standard behavior. Comparing against other implementations can help (but remember that implementations can differ while conforming to the Standard); try Compiler Explorer. If you still aren't sure, ask the nearest C++ expert.
-
You should prepare a self-contained command-line test case, ideally as small as possible. We need a source file, a command line, what happened (e.g. a compiler error, runtime misbehavior), and what you expected to happen. By "self-contained", we mean that your source file has to avoid including code that we don't have. Ideally, only CRT and STL headers should be included. If you have to include other MSVC libraries, or the Windows SDK, to trigger an STL bug, that's okay. But if you need parts of your own source code to trigger the STL bug, you need to extract that for us. (On Developer Community, we'll accept zipped IDE projects if you have no other way to reproduce a bug, but this is very time-consuming for us to reduce.)
-
A good title is helpful. We prefer "
<header_name>
: Short description of your issue". You don't usually need to mentionstd::
or C++. For example, "<type_traits>
:is_cute
should be true forenum class FluffyKittens
".
It's okay if you report an apparent STL bug that turns out to be a compiler bug or surprising-yet-Standard behavior. Just try to follow these rules, so we can spend more time fixing bugs and implementing features.
- Install Visual Studio 2022 17.8 Preview 1 or later.
- Select "Windows 11 SDK (10.0.22000.0)" in the VS Installer.
- We recommend selecting "C++ CMake tools for Windows" in the VS Installer. This will ensure that you're using supported versions of CMake and Ninja.
- Otherwise, install CMake 3.26.0 or later, and Ninja 1.11.0 or later.
- We recommend selecting "Python 3 64-bit" in the VS Installer.
- Otherwise, make sure Python 3.9 or later is available to CMake.
- Open Visual Studio, and choose the "Clone or check out code" option. Enter the URL of this repository,
https://github.com/microsoft/STL
. - Open a terminal in the IDE with
Ctrl + `
(by default) or press on "View" in the top bar, and then "Terminal". - In the terminal, invoke
git submodule update --init --progress
- Choose the architecture you wish to build in the IDE, and build as you would any other project. All necessary CMake
settings are set by
CMakePresets.json
.
- Install Visual Studio 2022 17.8 Preview 1 or later.
- Select "Windows 11 SDK (10.0.22000.0)" in the VS Installer.
- We recommend selecting "C++ CMake tools for Windows" in the VS Installer. This will ensure that you're using supported versions of CMake and Ninja.
- Otherwise, install CMake 3.26.0 or later, and Ninja 1.11.0 or later.
- We recommend selecting "Python 3 64-bit" in the VS Installer.
- Otherwise, make sure Python 3.9 or later is available to CMake.
- Open a command prompt.
- Change directories to a location where you'd like a clone of this STL repository.
git clone https://github.com/microsoft/STL --recurse-submodules
To build the x86 target:
- Open an "x86 Native Tools Command Prompt for VS 2022 Preview".
- Change directories to the previously cloned
STL
directory. cmake --preset x86
cmake --build --preset x86
To build the x64 target (recommended):
- Open an "x64 Native Tools Command Prompt for VS 2022 Preview".
- Change directories to the previously cloned
STL
directory. cmake --preset x64
cmake --build --preset x64
Consumption of the built library is largely based on the build system you're using. There are at least 2 directories
you need to hook up. Assuming you built the x64 target with the Visual Studio IDE, with the STL repository cloned to
C:\Dev\STL
, build outputs will end up at C:\Dev\STL\out\x64\out
. Ensure that the inc
directory is searched
for headers, and that lib\{architecture}
is searched for link libraries, before any defaults supplied by MSVC. The
names of the import and static libraries are the same as those that ship with MSVC. As a result, the compiler /MD
,
/MDd
, /MT
, or /MTd
switches will work without modification of your build scripts or command-line muscle memory.
Should you choose to use the DLL flavors, the DLLs to deploy are built to bin\{architecture}
. Note that the DLLs
generated by the CMake build system here have a suffix, defaulting to _oss
, which distinguishes them from the binaries
that ship with MSVC. That avoids any conflict with the DLLs installed by the redistributables into System32 and
ensures that other components wanting to be a "guest in your process", like print drivers and shell extensions, see the
export surface of the STL they were built with. Otherwise, the "msvcp140.dll
" you deployed in the same directory as
your .exe would "win" over the versions in System32.
The compiler looks for include directories according to the INCLUDE
environment variable, and the linker looks for
import library directories according to the LIB
environment variable, and the Windows loader will (eventually) look
for DLL dependencies according to directories in the PATH
environment variable.
The build generates a batch script named set_environment.bat
in the output directory. If you run this script in a VS
Developer Command Prompt, it will insert the proper directories into the INCLUDE
, LIB
, and PATH
environment
variables to ensure that the built headers and libraries are used.
From an "x64 Native Tools Command Prompt for VS 2022 Preview":
C:\Users\username\Desktop>C:\Dev\STL\out\x64\set_environment.bat
C:\Users\username\Desktop>type example.cpp
#include <iostream>
int main() {
std::cout << "Hello STL OSS world!\n";
}
C:\Users\username\Desktop>cl /nologo /EHsc /W4 /WX /MDd /std:c++latest .\example.cpp
example.cpp
C:\Users\username\Desktop>.\example.exe
Hello STL OSS world!
C:\Users\username\Desktop>dumpbin /DEPENDENTS .\example.exe | findstr msvcp
msvcp140d_oss.dll
- Follow either How To Build With A Native Tools Command Prompt or How To Build With The Visual Studio IDE.
- Acquire Python 3.9 or newer and have it on the
PATH
(or run it directly using its absolute or relative path). - Have LLVM's
bin
directory on thePATH
(soclang-cl.exe
is available).- We recommend selecting "C++ Clang tools for Windows" in the VS Installer. This will automatically add LLVM to the
PATH
of the x86 and x64 Native Tools Command Prompts, and will ensure that you're using a supported version. - Otherwise, use LLVM's installer and choose to add LLVM to your
PATH
during installation.
- We recommend selecting "C++ Clang tools for Windows" in the VS Installer. This will automatically add LLVM to the
- Follow the instructions below.
After configuring and building the project, running ctest
from the build output directory will run all the tests.
CTest will only display the standard error output of tests that failed. In order to get more details from CTest's
lit
invocations, run the tests with ctest -V
.
${PROJECT_BINARY_DIR}\tests\utils\stl-lit\stl-lit.py
can be invoked on a subdirectory of a test suite and will execute
all the tests under that subdirectory. This can mean executing the entirety of a single test suite, running all tests
under a category in libcxx, or running a single test in std
and tr1
.
These examples assume that your current directory is C:\Dev\STL\out\x64
.
- This command will run all of the test suites with verbose output.
ctest -V
- This command will also run all of the test suites.
python tests\utils\stl-lit\stl-lit.py ..\..\llvm-project\libcxx\test ..\..\tests\std ..\..\tests\tr1
- This command will run all of the std test suite.
python tests\utils\stl-lit\stl-lit.py ..\..\tests\std
- If you want to run a subset of a test suite, you need to point it to the right place in the sources. The following
will run the single test found under VSO_0000000_any_calling_conventions.
python tests\utils\stl-lit\stl-lit.py ..\..\tests\std\tests\VSO_0000000_any_calling_conventions
- You can invoke
stl-lit
with any arbitrary subdirectory of a test suite. In libcxx this allows you to have finer control over what category of tests you would like to run. The following will run all the libcxx map tests.python tests\utils\stl-lit\stl-lit.py ..\..\llvm-project\libcxx\test\std\containers\associative\map
When running the tests via CTest, all of the test suites are considered to be a single test. If any single test in a
test suite fails, CTest will simply report that the stl
test failed.
Example:
0% tests passed, 1 tests failed out of 1
Total Test time (real) = 2441.55 sec
The following tests FAILED:
1 - stl (Failed)
The primary utility of CTest in this case is to conveniently invoke stl-lit.py
with the correct set of arguments.
CTest will output everything that was sent to stderr for each of the failed test suites, which can be used to identify
which individual test within the test suite failed. It can sometimes be helpful to run CTest with the -V
option in
order to see the stdout of the tests.
When running the tests directly via the generated stl-lit.py
script the result of each test will be printed. The
format of each result is {Result Code}: {Test Suite Name} :: {Test Name}:{Configuration Number}
.
Example:
-- Testing: 28 tests, 12 workers --
PASS: tr1 :: tests/cwchar1:01 (1 of 28)
PASS: tr1 :: tests/cwchar1:11 (2 of 28)
PASS: tr1 :: tests/cwchar1:02 (3 of 28)
PASS: tr1 :: tests/cwchar1:03 (4 of 28)
PASS: tr1 :: tests/cwchar1:00 (5 of 28)
PASS: tr1 :: tests/cwchar1:04 (6 of 28)
PASS: tr1 :: tests/cwchar1:05 (7 of 28)
PASS: tr1 :: tests/cwchar1:09 (8 of 28)
PASS: tr1 :: tests/cwchar1:06 (9 of 28)
UNSUPPORTED: tr1 :: tests/cwchar1:20 (10 of 28)
UNSUPPORTED: tr1 :: tests/cwchar1:21 (11 of 28)
UNSUPPORTED: tr1 :: tests/cwchar1:22 (12 of 28)
UNSUPPORTED: tr1 :: tests/cwchar1:23 (13 of 28)
UNSUPPORTED: tr1 :: tests/cwchar1:24 (14 of 28)
PASS: tr1 :: tests/cwchar1:07 (15 of 28)
PASS: tr1 :: tests/cwchar1:08 (16 of 28)
PASS: tr1 :: tests/cwchar1:10 (17 of 28)
PASS: tr1 :: tests/cwchar1:16 (18 of 28)
PASS: tr1 :: tests/cwchar1:17 (19 of 28)
PASS: tr1 :: tests/cwchar1:14 (20 of 28)
PASS: tr1 :: tests/cwchar1:12 (21 of 28)
PASS: tr1 :: tests/cwchar1:13 (22 of 28)
PASS: tr1 :: tests/cwchar1:19 (23 of 28)
PASS: tr1 :: tests/cwchar1:18 (24 of 28)
PASS: tr1 :: tests/cwchar1:15 (25 of 28)
PASS: tr1 :: tests/cwchar1:25 (26 of 28)
PASS: tr1 :: tests/cwchar1:26 (27 of 28)
PASS: tr1 :: tests/cwchar1:27 (28 of 28)
Testing Time: 3.96s
Expected Passes : 23
Unsupported Tests : 5
In the above example, we see that 23 tests succeeded and 5 were unsupported.
Our tests use the standard lit result codes, and an undocumented result code: SKIPPED
. For our tests, only the
PASS
, XFAIL
, XPASS
, FAIL
, UNSUPPORTED
, and SKIPPED
result codes are relevant.
The PASS
and FAIL
result codes are self-explanatory. We want our tests to PASS
and not FAIL
.
The XPASS
and XFAIL
result codes are less obvious. XPASS
is actually a failure result and indicates that we
expected a test to fail but it passed. XFAIL
is a successful result and indicates that we expected the test to fail
and it did. Typically an XPASS
result means that the expected_results.txt
file for the test suite needs to be
modified. If the XPASS
result is a test legitimately passing, the usual course of action would be to remove a FAIL
entry from the expected_results.txt
. However, some tests from libcxx
mark themselves as XFAIL
(meaning they
expect to fail) for features they have added tests for but have yet to implement in libcxx
. If the STL implements
those features first the tests will begin passing unexpectedly for us and return XPASS
results. In order to resolve
this it is necessary to add a PASS
entry to the expected_results.txt
of the test suite in question.
The UNSUPPORTED
result code means that the requirements for a test are not met and so it will not be run. Currently,
all tests which use the /clr
or /clr:pure
options are unsupported. Also, the /BE
option is unsupported for x86.
The SKIPPED
result code indicates that a given test was explicitly skipped by adding a SKIPPED
entry to the
expected_results.txt
. A test may be skipped for a number of reasons, which include, but are not limited to:
- being an incorrect test
- taking a very long time to run
- failing or passing for the incorrect reason
While stl-lit
is super awesome in finding out that something is wrong or not even compiling, it is not really
helpful in debugging what is going wrong. However, debugging individual tests is rather simple given some additional
steps. Let's assume we want to debug a new feature with tests located in tests\std\tests\GH_XXXX_meow
.
As always, build the STL from your branch and run the tests:
C:\STL\out\x64> ninja
C:\STL\out\x64> python tests\utils\stl-lit\stl-lit.py -v C:\STL\tests\std\tests\GH_XXXX_meow
Let's assume one of the tests fails an assert and we want to debug that configuration. stl-lit
will conveniently print
the build command, which is far too long to provide here in full. The important part is to add the following options to
provide debug symbols: /Zi /Fdbark.pdb
.
You can replace bark
with any descriptive name you like. Add these before the "-link"
option in the command line
and recompile. Example:
C:\STL\out\x64>cl "C:\STL\tests\std\tests\GH_XXXX_meow\test.cpp" [... more arguments ...]
"-FeC:\STL\out\x64\tests\std\tests\GH_XXXX_meow\Output\02\GH_XXXX_meow.exe" /Zi /Fdbark.pdb "-link"
[... more arguments ...]
You can now start debugging the test via:
devenv "C:\STL\out\x64\tests\std\tests\GH_XXXX_meow\Output\02\GH_XXXX_meow.exe"
"C:\STL\tests\std\tests\GH_XXXX_meow\test.cpp"
However, this might not work right away, as Visual Studio may complain about a missing msvcp140_oss.dll
. The reason
is that the STL builds those and other DLLs itself and we should under no circumstances overwrite the installed ones.
If you are testing one of the configurations with dynamic linkage (/MD
or /MDd
) the easiest solution is to add the
build folder to your path:
set PATH=C:\STL\out\x64\out\bin\amd64;%PATH%
For performance-sensitive code – containers, algorithms, and the like – you may wish to write and/or run benchmarks, and the STL team will likely run any benchmarks we do have in our PR process. Additionally, if you are writing a "performance improvement" PR, please add and run benchmarks to show that the PR does, in fact, improve performance.
The benchmarking code is located in benchmarks
. Adding a new benchmark is as easy as adding a new file
to benchmarks/src
, and then adding add_benchmark(<name> <source_file>)
to benchmarks/CMakeLists.txt
.
You may also modify an existing benchmark file. We use Google's Benchmark library,
so you may find their documentation helpful, and you can also read the existing code
for how we use it.
To run benchmarks, you'll need to first build the STL, then build the benchmarks:
cmake --preset x64
cmake --build --preset x64
cmake -B out\bench -S benchmarks -G Ninja -DSTL_BINARY_DIR=out\x64
cmake --build out\bench
You can then run your benchmark with:
out\bench\benchmark-<benchmark-name> --benchmark_out=<file> --benchmark_out_format=csv
And then you can copy this CSV file into Excel, or another spreadsheet program. For example:
out\bench\benchmark-std_copy --benchmark_out=benchmark-std_copy-results.csv --benchmark_out_format=csv
If you want to see all the other flags you can pass, run:
out\bench\benchmark-<benchmark-name> --help
To modify how components are visualized in the debugger, edit the file stl\debugger\STL.natvis
. For more information
on how to modify this file, check the natvis documentation.
You can add the natvis file to any Visual Studio C++ project if you right-click your project > Add > Existing Item and select the STL.natvis file. After doing this you should be able to see your changes in a Visual Studio debugging session.
The STL is built atop other compiler support libraries that ship with Windows and Visual Studio, like the UCRT, VCRuntime, and VCStartup. The following diagram describes the dependencies between those components and their ship vehicles.
flowchart TB
classDef default text-align:left
subgraph VisualStudioSubgraph[Visual Studio]
direction TB
STLNode("<b>STL</b>
This repo; provides C++ Standard Library headers, separately
compiled implementations of most of the iostreams functionality,
and a few runtime support components like std::exception_ptr.")
subgraph VCRuntimeSubgraph[VCRuntime]
direction TB
VCStartupNode("<b>VCStartup</b>
Provides compiler support mechanisms that
live in each binary; such as machinery to
call constructors and destructors for global
variables, the entry point, and the /GS cookie.
Merged into static and import libraries of VCRuntime.")
VCRuntimeNode("<b>VCRuntime</b>
Provides compiler support mechanisms that can be
shared between binaries; code that the compiler calls
on your behalf, such as the C++ exception handling
runtime, string.h intrinsics, math intrinsics, and
declarations for CPU-vendor-specific intrinsics.")
end
end
subgraph WindowsSDKSubgraph[Windows SDK]
UniversalCRTNode("<b>Universal CRT</b>
Windows component that provides C library support, such as printf,
C locales, and some POSIX-like shims for the Windows API, like _stat.")
end
STLNode ==> VCRuntimeSubgraph & UniversalCRTNode
VCStartupNode ==> VCRuntimeNode ==> UniversalCRTNode
This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.
When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.
See CONTRIBUTING.md for more information.
This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.
Copyright (c) Microsoft Corporation.
SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception