3rdParty/boost/1.78.0/libs/outcome/doc/html/changelog.html
#255Restore Experimental Outcome constexpr compatibility in C++ 20 which was an undocumented property of the Outcome v2.1 series, and which had been dropped in the v2.2 series.GCC Coroutines supportCoroutine support in GCCs after 10 is now correctly detected.
Fix failure to compile Boost.Outcome when the compiler declares support for C++ modules.
Don’t use [[likely]] in GCCs before 9.
Make support for C++ modules opt-in.
BREAKING CHANGE As announced for a year and three Boost releases, Outcome v2.2 became the default, replacing v2.1.All v2.1 Outcome code will need to be upgraded as described in the v2.1 => v2.2 upgrade guide. This branch has a number of major breaking changes to Outcome v2.1, see the list of v2.2 major changes.
VS2019.8 compatibilityVS2019.8 changed how to enable Coroutines, which caused Outcome to not compile on that compiler.#237If on C++ 20, we now use C++ 20 [[likely]] instead of compiler-specific markup to indicate when TRY has likely success or failure.BREAKING CHANGE #247
Previously the value of spare_storage(const basic_result|basic_outcome *) noexcept was not propagated over BOOST_OUTCOME_TRY, which causes things like stack backtraces captured at the point of construction of an errored result to get dropped at every TRY point. This has been fixed by adding an optional spare_storage to success_type<T> and failure_type<T> , as well as to auto success(T &&, ...) and auto failure(T &&, ...) .
You should not notice this in your code, except that where before spare storage values did not propagate through TRY, now they do, which is a breaking change.
BREAKING CHANGE #244
It came as a shock to learn that BOOST_OUTCOME_TRY had been broken since the inception of this library for certain corner case code:
outcome::result<Foo> get_foo();
outcome::result<Foo> filter1(outcome::result<Foo> &&);
outcome::result<Foo> && filter2(outcome::result<Foo> &&);
// This works fine, and always has
BOOST_OUTCOME_TRY(auto v, filter1(get_foo()))
// This causes UB due to result<Foo> being destructed before move of value into v
BOOST_OUTCOME_TRY(auto v, filter2(get_foo()))
Whilst reference passthrough filter functions are not common, they can turn up in highly generic code, where destruction before copy/move is not helpful.
The cause is that TRY used to work by binding the result of the expression to an auto &&unique, testing if that unique if successful or not, and if successful then moving from unique.value() into the user’s output variable. If the expression returned is a prvalue, the Result’s lifetime is extended by the bound reference to outside of the statement, and all is good. If the expression returned is an xvalue or lvalue, then the lifetime extension does not exceed that of the statement, and the Result is destructed after the semicolon succeeding the assignment to auto &&unique.
This bug has been fixed by TRY deducing the value category of its input expression as follows:
auto unique = (expr) (breaking change)auto unique = (expr) (breaking change)auto unique = (expr) (breaking change)This ensures that xvalue and lvalue inputs do not cause unhelpfully early lifetime end, though it does silently change the behaviour of existing code which relied on rvalues and lvalues being passed through, as a new construct-move-destruct or construct-copy-destruct cycle is introduced to where there was none before. Also, before C++ 17, there is now an added copy/move for prvalue inputs, which does not affect runtime codegen due to Return Value Optimisation (RVO), but does cause Results containing non-copying non-moving types to fail to compile, which is a breaking change from beforehand.
If one wishes rvalues or lvalues to be passed through, one can avail of a new TRY syntax based on preprocessor overloading:
BOOST_OUTCOME_TRY((refspec, varname), expr)BOOST_OUTCOME_TRYV2(refspec, expr)Here refspec is the storage to be used for both the internal temporary unique, AND varname. So if you write:
Foo &&foo;
BOOST_OUTCOME_TRY((auto &&, v), filter2(foo))
… then the internal unique is declared as auto &&unique = (filter2(foo)), and the output variable is declared as auto &&v = std::move(unique).assume_value(). This passes through the rvalue referencing, and completely avoids copies and moves of Foo. If you wish to not extract the value but also specify unique storage, there is a new BOOST_OUTCOME_TRYV2(refspec, expr).
My thanks to KamilCuk from https://stackoverflow.com/questions/66069152/token-detection-within-a-c-preprocessor-macro-argument for all their help in designing the new overloaded TRY syntax. My thanks also to vasama for reporting this issue and working through how best to fix it with me.
The preprocessor logic for choosing when to use bool with concept on GCC was yet again refactored. This should fix those choices of GCC configuration which caused failure due to the wrong combination being chosen.
The ADL discovered event hooks have been replaced with policy-specified event hooks insteadThis is due to brittleness (where hooks would quietly self-disable if somebody changed something), compiler bugs (a difference in compiler settings causes the wrong hooks, or some but not all hooks, to get discovered), and end user difficulty in using them at all. The policy-specified event hooks can be told to default to ADL discovered hooks for backwards compatibility: set BOOST_OUTCOME_ENABLE_LEGACY_SUPPORT_FOR to less than 220 to enable emulation.Improve configuring BOOST_OUTCOME_GCC6_CONCEPT_BOOLOlder GCCs had boolean based concepts syntax, whereas newer GCCs are standards conforming. However the precise logic of when to use legacy and conforming syntax was not well understood, which caused Outcome to fail to compile depending on what options you pass to GCC. The new logic always uses the legacy syntax if on GCC 8 or older, otherwise we use conforming syntax if and only if GCC is in C++ 20 mode or later. This hopefully will resolve the corner case build failures on GCC.
Boost.Outcome should now compile with BOOST_NO_EXCEPTIONS definedThanks to Emil, maintainer of Boost.Exception, making a change for me, Boost.Outcome should now compile with C++ exceptions globally disabled. You won’t be able to use boost::exception_ptr as it can’t be included if C++ exceptions are globally disabled.#236In the Coroutine support the final_suspend() was not noexcept, despite being required to be so in the C++ 20 standard. This has been fixed, but only if your compiler implements noop_coroutine. Additionally, if noop_coroutine is available, we use the much more efficient coroutine handle returning variant of await_suspend() which should significantly improve codegen and context switching performance.
BREAKING CHANGE void results and outcomes no longer default construct types during explicit constructionPreviously if you explicitly constructed a result<T> from a non-errored result<void>, it default constructed T. This was found to cause unhelpful surprise, so it has been disabled.New macro BOOST_OUTCOME_ENABLE_LEGACY_SUPPORT_FORThe macro BOOST_OUTCOME_ENABLE_LEGACY_SUPPORT_FOR can be used to enable aliasing of older naming and features to newer naming and features when using a newer version of Outcome.Concepts now have snake case style naming instead of camel case styleWhen Outcome was first implemented, it was thought that C++ 20 concepts were going to have camel case style. This was changed before the C++ 20 release, and Outcome’s concepts have been renamed similarly. This won’t break any code in Outcome v2.1, as compatibility aliases are provided. However code compiled against Outcome v2.2 will need to be upgraded, unless BOOST_OUTCOME_ENABLE_LEGACY_SUPPORT_FOR is set to less than 220.Concepts now live in BOOST_OUTCOME_V2_NAMESPACE::concepts namespacePreviously concepts lived in the convert namespace, now they live in their own namespace.New concepts basic_result<T> and basic_outcome<T> addedEnd users were finding an unhelpful gap in between is_basic_result<T> and value_or_error<T> where they wanted a concept that matched types which were basic_result, but not exactly one of those. Concepts filling that gap were added.Operation TRY works differently from Outcome v2.2 onwardsThis is a severely code breaking change which change the syntax of how one uses BOOST_OUTCOME_TRY(). A regular expression suitable for upgrading code can be found in the list of changes between Outcome v2.1 and v2.2.
#224The clang Apple ships in Xcode 11.4 (currently the latest) has not been patched with the fixes to LLVM clang that fix noexcept(std::is_constructible<T, void>) failing to compile which I originally submitted years ago. So give up waiting on Apple to fix their clang, add a workaround to Outcome.Use of void in T or E caused noexcept(false)Direct traits examination of void was causing nothrow detection to return false, fixed.Spare storage could not be used from within no-value policy classesDue to an obvious brain fart when writing the code at the time, the spare storage APIs had the wrong prototype which prevented them working from within policy classes. Sorry.
Performance of Outcome-based code compiled by clang has been greatly improved
The previous implementation of Outcome’s status bitfield confused clang’s optimiser, which caused low quality codegen. Unlike most codegen issues, this was noticeably in empirical benchmarks of real world code, as was shown by P1886 Error speed benchmarking.
The safe part of the better_optimisation Outcome v2.2.0 future branch was merged to Outcome v2.1.3 which includes a new status bitfield implementation. This appears to not confuse clang’s optimiser, and clang 9 produces code which routinely beats GCC 9’s code for various canned use cases.
Precompiled headers are automatically enabled on new enough cmake’s for standalone Outcome
If on cmake 3.16 or later, its new precompiled headers build support is used to tell consumers of the outcome::hl cmake target to precompile Outcome, if and only if PROJECT_IS_DEPENDENCY is false. PROJECT_IS_DEPENDENCY is set by Outcome’s CMakeLists.txt if it detects that it was included using add_subdirectory(), so for the vast majority of Outcome end users, the use of precompiled headers will NOT be enabled.
Exported targets do NOT request precompilation of headers, as it is assumed that importers of the Outcome cmake targets will configure their own precompiled headers which incorporate Outcome.
Installability is now CI tested per commit
Due to installability of standalone Outcome (e.g. make install) breaking itself rather more frequently than is ideal, installability is now tested on CI per commit.
Coroutines support has been documented
The coroutines support added in v2.1.2 has now been properly documented.
#214Newer Concepts implementing compilers were unhappy with the early check for destructibility of T and E, so removed template constraints, falling back to static assert which runs later in the type instantiation sequence.#215For standalone Outcome, CMAKE_TOOLCHAIN_FILE is now passed through during dependency superbuild. This should solve build issues for some embedded toolchain users.#220A false positive undefined behaviour sanitiser failure in some use cases of Experimental Outcome was worked around to avoid the failure message.#221Restored compatibility with x86 on Windows, which was failing with link errors. It was quite surprising that this bug was not reported sooner, but obviously almost nobody is using Outcome with x86 on Windows.#223Fix a segfault in Debug builds only when cloning a status_code_ptr in Experimental.Outcome only.
Improved compatibility with cmake toolingStandalone outcome is now make install-able, and cmake find_package() can find it. Note that you must separately install and find_package() Outcome’s dependency, quickcpplib, else find_package() of Outcome will fail.Non-permissive parsing is now default in Visual StudioThe default targets in standalone Outcome’s cmake now enable non-permissive parsing. This was required partially because VS2019 16.3’s quite buggy Concepts implementation is unusuable in permissive parsing mode. Even then, lazy ADL two phase lookup is broken in VS2019 16.3 with /std:latest, you may wish to use an earlier language standard. Breaking change! The git submodule mechanism used by standalone Outcome of specifying dependent libraries has been replaced with a cmake superbuild of dependencies mechanism instead. Upon cmake configure, an internal copy of quickcpplib will be git cloned, built and installed into the build directory from where an internal find_package() uses it. This breaks the use of the unconfigured Outcome repo as an implementation of Outcome, one must now do one of:
ExternalProject_Add()) to build and install Outcome into a local installation.Breaking change! For standalone Outcome, the current compiler is now checked for whether it will compile code containing C++ Concepts, and if it does, all cmake consumers of Outcome will enable C++ Concepts. Set the cmake variable BOOST_OUTCOME_C_CONCEPTS_FLAGS to an empty string to prevent auto detection and enabling of C++ Concepts support occurring.BOOST_OUTCOME_TRY operation now hints to the compiler that operation will be successfulP1886 Error speed benchmarking showed that there is considerable gain in very small functions by hinting to the compiler whether the expression is expected to be successful or not. BOOST_OUTCOME_TRY previously did not hint to the compiler at all, but now it does. A new suite of macros BOOST_OUTCOME_TRY_FAILURE_LIKELY hint to the compiler that failure is expected. If you wish to return to the previously unhinted behaviour, define BOOST_OUTCOME_TRY_LIKELY(expr) to (!!expr).#199Support for C++ Coroutines has been added. This comes in two parts, firstly there is now an BOOST_OUTCOME_CO_TRY() operation suitable for performing the TRY operation from within a C++ Coroutine. Secondly, in the header outcome/coroutine_support.hpp there are implementations of eager<OutcomeType> and lazy<OutcomeType> which let you more naturally and efficiently use basic_result or basic_outcome from within C++ Coroutines – specifically, if the result or outcome will construct from an exception pointer, exceptions thrown in the coroutine return an errored or excepted result with the thrown exception instead of throwing the exception through the coroutine machinery (which in current compilers, has a high likelihood of blowing up the program). Both eager<T> and lazy<T> can accept any T as well. Both have been tested and found working on VS2019 and clang 9.#210make_error_code() and make_exception_ptr() are now additionally considered for compatible copy and move conversions for basic_result<>. This lets you construct a basic_result<T, E> into a basic_result<T, error_code>, where E is a custom type which has implemented the ADL discovered free function error_code make_error_code(E), but is otherwise unrelated to error_code. The same availability applies for exception_ptr with make_exception_ptr() being the ADL discovered free function. basic_outcome<> has less support for this than basic_result<> in order to keep constructor count down, but it will accept via this mechanism conversions from basic_result<> and failure_type<>.
#184The detection of [[nodiscard]] support in the compiler was very mildly broken.
#184As per request from Boost release managers, relocated version.hpp and revision.hpp into detail, and added the Boost licence boilerplate to the top of every source file which was missing one (I think). Also took the opportunity to run the licence restamping script over all Outcome, so copyright dates are now up to date.#185Add FAQ item explaining issue #185, and why we will do nothing to fix it right now.#189Refactored the BOOST_OUTCOME_TRY implementation to use more clarified customisation points capable of accepting very foreign inputs. Removed the std::experimental::expected<T, E> specialisations, as those are no longer necessary. Fixed the documentation for the customisation points which previously claimed that they are ADL discovered, which they are not. Added a recipe describing how to add in support for foreign input types.#183Added a separate motivation/plug_error_code specifically for Boost.
-BOOST_OUTCOME_VERSION_MINOR hadn’t been updated to 1.#181Fix issue #181 where Outcome didn’t actually implement the strong swap guarantee, despite being documented as doing so.#190Fix issue #190 in Boost edition where unit test suite was not runnable from the Boost release distro.#182Fix issue #182 where trait::is_exception_ptr_available<T> was always true, thus causing much weirdness, like not printing diagnostics and trying to feed everything to make_exception_ptr().#194Fix issue #192 where the std::basic_outcome_failure_exception_from_error() was being defined twice for translation units which combine standalone and Boost Outcome’s.
include/outcome/outcome.natvis is now namespace permuted like the rest of Outcome, so debugging Outcome based code in Visual Studio should look much prettier than before.
Added experimental status_result and status_outcome based on experimental status_code.
Boost edition is now 100% Boost, so defaults for result and outcome are boost::system::error_code::errc_t and boost::exception_ptr. Moreover, the test suite in the Boost edition now exclusively tests the Boost edition. One can, of course, freely use the standalone edition with Boost, and the Boost edition with std types.
Renamed ADL discovered customisation point throw_as_system_error_with_payload() to outcome_throw_as_system_error_with_payload().
Inexplicably outcome’s error + exception constructor had been removed. Nobody noticed during the Boost peer review, which is worrying seeing as that constructor is needed for one of the main advertised features to Boost!
Added reference dump of v2.1 ABI so we can check if ABI breakage detection works in the next set of changes, plus Travis job to check ABI and API compatibility per commit.