This makes attribute_value and lexical_cast use the "new" to/from_chars api.
Its main advantages are:
- It's guaranteed to be locale independent, hopefully fixing all cases of #3945 and similar
- It fixes some cases config serialization, in particular the test
```
cfg["x"] = "9.87654321";
BOOST_CHECK_EQUAL(cfg["x"], 9.87654321);
```
- Previously the lexical_cast implementation used exception
handling for invalid formats (catching std::invalid_argument)
which made noise during debugging (and is also slower if it
is not optimized out).
- It's faster
So far afaik the only compiler which has a complete and proper to/from_chars implementation is msvc, gccs implementation of from_chars sometimes uses strtod under the hood and clang simply hasn't implemented from_chars for floating point numbers yet at all (actually the upcomig clang 20 will have it). Luckily for us, there is now also boost::charconv that can be used. So this raises to minimum build requirement to have at least one of:
- msvc 2019 update 5
- gcc 11
- clang 14 (i have added a fallback implementation of from_chars for this case, that doesn't support all of its features, and is probably certainly not as fast as the boost version, but supports the features that we use from it)
- boost 1.85
Since in particular the gcc implementation isn't that good (at least it on gcc11), boost charconv is the preferred implementation that is used if available.
This also removes a strange overload for
pointers to integers in lexical_cast while changing lexical_cast to use the new api.
Original code from mesilliac in #6826. Made a separate PR for this since the relevant code moved to an entirely separate file and trying to "properly" fix the merge conflicts resulting from that isn't worth the time.
The super tag dependency forms a directed graph. Boost Graph's depth first search implementation was used with a back edge detector to find the cycles.
This is a preparation for enumerating all keys that a tag can use, including the keys from the super tags. If cycles aren't handled, it is impossible to validate mandatory keys without entering an infinite loop.
scons/lua.py makes use of the vestigial luadir option from commit
e94dcecf17.
Like FindLua.cmake, scons/lua.py searches for the Lua headers and
library, instead of using pkg-config like the old scons/lua.py (removed
in commit 9929d3ca1c) did, because even though distributions typically
provide .pc files for Lua, upstream Lua doesn't. It's likely that all
distributions that compile Lua as a C++ library will also provide .pc
files, but this check doesn't rely on that (just as the CMake module
doesn't).
Unfortunately, SCons.Conftest.CheckLib() prints up to eight messages
like "Checking for C++ library lua54-c++... no" until a working library
name is found.
Also conditionally include system Lua headers in src/lua/*.h and update
documentation in src/modules/lua_README.md, src/wesnoth_lua_config.h,
and src/wesnoth_lua_config.md. The two lines about "The primary commit,
after replacing the sources," in src/wesnoth_lua_config.md don't make
sense since the instructions were updated for submodule Lua in commit
d32cfb88c4 and make even less sense now with preceding commits for
updating CMake modules.
Version 13 of GCC added the -Wdangling-reference option, however
due to false positives it had to be moved from -Wall to -Wextra.
Discussion about its status in GCC-14 is in
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110075
The warning is for C++ only and the flag will generate a warning
when compiling C files if the flag is set in Scon's CCFLAGS.
The docs don't explain exactly what triggers the warning, but here's the
explanation from https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106393#c1
> The warning works by checking if a reference is initialized with a function
> that returns a reference, and at least one parameter of the function is a
> reference that is bound to a temporary. It assumes that such a function
> actually returns one of its arguments! ... suppress the warning when we've
> seen the definition of the function and we can say that it can return a
> variable with static storage duration.
Consistent with that, we're getting warnings on functions of the form
find(container, const std::string&). All of the things triggering it
seem to be false positives:
* Calls to find_widget<...>.
* Calls to theme::get_theme_config.
* Calls to race::gender_value.
* In src/actions/attack.cpp, any `attacker->attacks()[i]`.
Although there's an iterator in there, the eventual reference
is to a member of the array, not the iterator.
* In src/gui/dialogs/unit_advance.cpp, `cfg.child_range("...").back()`
That returns an object indirectly owned by cfg.
We'd like the other warnings of -Wextra, so turn off -Wdangling-reference.
This is a complement to #7416 where it replaces opening the replay's download URL in a browser with directly downloading it into the player's save folder.
Turning GLIBCXX_ASSERTIONS adds some bounds checks, such as checking
in std::string::operator[]. It's a subset of the checks that GLIBCXX_DEBUG
adds, but has the advantage that it doesn't require a rebuild of the Boost
libraries.
None of the CI builds enable these assertions, as they slow down the code
which makes the unit tests time out.
msys2 pkg-config adds root system include path among -I flags and
passing it with -isystem isn't supported by gcc, and
the workaround of passing --dont-define-prefix to pkg-config doesn't
work on msys2's "subsystem" install of scons because posix style
absolute paths aren't handled properly when expanding CPPPATH.
Most likely related to https://github.com/msys2/MINGW-packages/issues/10761
In CMake this is the FORCE_COLOR_OUTPUT option. In SCons this is the
force_color option. They both pass -fdiagnostics-color=always to gcc,
or -fcolor-diagnostics to clang.
This is useful when building with Ninja, or running piped CI tools.
This is a more reliable way to detect distro with systemd support.
"systemd" binary may not be in PATH while systemctl would have to be.
This helps with part of issue #5692
This removes the build-time dependencies on SDL_ttf and FriBidi,
alongside the SDL_ttf wrappers, the SDL_ttf text surface class, the
SDL_ttf render cache, and the SDL_ttf (de)initialization code.
* Convert server_base class to use coroutine instead of handlers
* Rework wesnothd's client login to use coroutine
* Merge 3 player handling functions into a single coroutine
* update cmakelists too
* Implement send_doc_queued in terms of coroutine
* Use brace initialization for making asio buffers
* Implement campaignd's request handling in coroutine
* Brace-initialize entire vector
* Remove old handler based send/receive helpers
* Document coroutine send/receive helpers
* Made coro_send_doc() helper take wml doc by reference
In most cases there is no need to rely on shared pointers to ensure
object lifetime if using coroutines since even when coroutine is
suspended args are still kept alive by its context.
* Document coro_send_file()
* Silence deprecation warning to fix build on earlier versions of boost
* Explicitly check for boost.context to allow linking against static boost libs
* Add boost.coroutine to flatpak manifest
* Port winapi TransmitFile codepath to coroutines
* Exception safety fix
* Add boost.scope_exit to vcpkg
* Fix build with pre-1.66 boost
* Move coro_* helpers into server_base class
Those helpers were in .ipp solely because they were templated on handler
types, this is no longer true after coroutine based rework.
* Make server_base::coro_send_file non-inline
* CleanUp Xcode project
Co-authored-by: Martin Hrubý (hrubymar10) <hrubymar10@gmail.com>
