aboutsummaryrefslogtreecommitdiffstatshomepage
path: root/docs/boost.md
diff options
context:
space:
mode:
Diffstat (limited to 'docs/boost.md')
-rw-r--r--docs/boost.md149
1 files changed, 149 insertions, 0 deletions
diff --git a/docs/boost.md b/docs/boost.md
new file mode 100644
index 0000000..6e91d42
--- /dev/null
+++ b/docs/boost.md
@@ -0,0 +1,149 @@
+Library naming
+--------------
+
+The way Boost names library files by default is insane. It's absolutely not compatible between
+OSs, compilers, Boost versions, etc. On Linux, for example, it would create
+stage/lib/libboost_filesystem.a, while on Windows it would become something insane like
+stage\lib\libboost_filesystem-vc142-mt-s-x64-1_72.lib. More than that, older Boost versions
+wouldn't include architecture information (the "x64" part) in the file name, so you couldn't
+store libraries for both x86 and x64 in the same directory. On Linux, on the other hand, you
+can't even store debug/release binaries in the same directory. What's worse is that older CMake
+versions don't support the architecture suffix, choking on the Windows example above.
+
+With all of that in mind, I decided to bring some uniformity by sacrificing some flexibility.
+b2 is called with --layout=system, and libraries are put to stage/<platform>/<configuration>/lib,
+where <platform> is x86/x64 and <configuration> is CMake's CMAKE_BUILD_TYPE. That means that I
+can't have libraries with different runtime-link values in the same directory, but I don't really
+care.
+
+Hate speech
+-----------
+
+Is there a person who doesn't hate Boost.Build? I'm not sure, I'm definitely
+_not_ one of these people. Maybe it's the lack of adoption (meaning that
+learning it is useless outside of Boost), maybe it's the incomprehensible
+syntax. Maybe it's the absolutely insane compiler-specific configuration
+files (tools/build/src/tools/*.jam), which are impossible to figure out.
+Maybe it's the fact that the implementation switched from C to C++ while some
+half-baked Python implementation has been there since at least 2015 (see the
+marvelous memo "Status: mostly ported." at the top of tools/build/src/build_system.py).
+
+What I hate the most though is how its various subtle, implicit and invisible
+decision-making heuristics changed thoughout the release history of Boost.
+You have a config and a compiler that will happily build version 1.65.0?
+Great! Want to use the same config and the same compiler to build version
+1.72.0? Well, too fucking bad, it doesn't work anymore. This I really do
+hate the most.
+
+Three kinds of toolsets
+-----------------------
+
+b2 accepts the toolset= parameter. What about building b2 itself though?
+Well, this is what the bootstrap.{sh,bat} scripts do. They also accept a
+toolset argument, but it is _completely_ different to that of b2. That's
+sort of OK, since e.g. cross-compiling b2 is something we rarely want to do
+(and hence there must typically be a native toolset available).
+
+bootstrap.sh and bootstrap.bat are completely different (of course!), and
+accept different arguments for their toolset parameters.
+
+Config file insanity
+--------------------
+
+Say, we're building Boost on Windows using the GCC from a MinGW-w64
+distribution. We can pass toolset=gcc and all the required flags on the
+command line no problem. What if we want to make a user configuration file
+so that 1) the command line is less polluted, and 2) it can possibly be
+shared? Well, if we put
+
+ using gcc : : : <name>value... ;
+
+there, Boost 1.65.0 will happily build everything, while Boost 1.72.0 will
+complain about "duplicate initialization of gcc". This is because when we
+ran `bootstrap.bat gcc` earlier, it wrote `using gcc ;` in project-config.jam.
+And while Boost 1.65.0 detects that toolset=gcc means we're going to use the
+MinGW GCC, and magically turns toolset=gcc to toolset=gcc-mingw, Boost 1.72.0
+does no such thing, and chokes on the "duplicate" GCC declaration.
+
+We also cannot put
+
+ using gcc : custom : : <options> ;
+
+without the executable path, since Boost insists that `g++ -dumpversion` must
+equal to "custom" (which makes total sense, lol). So we have to force it,
+and do provide the path.
+
+Windows & Clang
+---------------
+
+Building Boost using Clang on Windows is a sad story. As of 2020, there're
+three main ways to install the native Clang toolchain on Windows:
+
+ * download the installer from llvm.org (`choco install llvm` does this)
+ a.k.a. the upstream,
+ * install it as part of a MSYS2 installation (`pacman -S mingw-w64-x86_64-clang`),
+ * install as part of a Visual Studio installation.
+
+Using the latter method, you can switch a project to use the LLVM toolset
+using Visual Studio, but that's stupid. The former two, on the other hand,
+give us the the required clang/clang++/clang-cl executables, so everything
+seems to be fine.
+
+Except it's not fine. Let's start with the fact that prior to 1.66.0,
+toolset=clang is completely broken on Windows. It's just an alias for
+clang-linux, and it's hardcoded to require the ar & ranlib executables to
+create static libraries. Which is fine on Linux, since, and I'm quoting the
+source, "ar is always available". But it's not fine on Windows, since
+ar/ranlib are not, in fact, available there by default. Sure, you can
+install some kind of MinGW toolchain, and it might even work, but what the
+hell, honestly?
+
+Luckily, both the upstream distribution and the MSYS2 mingw-w64-x86_64-llvm
+package come with the llvm-ar and llvm-ranlib utilities. So we can put
+something like this in the config:
+
+ using clang : custom : clang++.exe : <archiver>llvm-ar <ranlib>llvm-ranlib.exe ;
+
+and later call
+
+ b2 toolset=clang-custom --user-config=path/to/config.jam ...
+
+But, as I mentioned, prior to 1.66.0, toolset=clang is _hardcoded_ to use ar
+& ranlib, these exact utility names. So either get them as part of some
+MinGW distribution or build Boost using another toolset.
+
+Now, it's all fine, but building stuff on Windows adds another thing into the
+equation: debug runtimes. When you build Boost using MSVC, for example, it
+picks one of the appropriate /MT[d] or /MD[d] flags to build the Boost
+libraries with. Emulating these flags with toolset=clang is complicated and
+inconvenient. Luckily, there's the clang-cl.exe executable, which aims to
+provide command line interface compatible with that of cl.exe.
+
+Boost.Build even supports toolset=clang-win, which should use clang-cl.exe.
+But alas, it's completely broken prior to 1.69.0. It just doesn't work at
+all. So, if you want to build w/ clang-cl.exe, either use Boost 1.69.0 or
+later, or build using another toolset.
+
+Cygwin & Clang
+--------------
+
+Now, a few words about Clang on Cygwin. When building 1.65.0, I encountered
+the following error:
+
+ /usr/include/w32api/synchapi.h:127:26: error: conflicting types for 'Sleep'
+ WINBASEAPI VOID WINAPI Sleep (DWORD dwMilliseconds);
+ ^
+ ./boost/smart_ptr/detail/yield_k.hpp:64:29: note: previous declaration is here
+ extern "C" void __stdcall Sleep( unsigned long ms );
+ ^
+
+GCC doesn't emit an error here because /usr/include is in a pre-configured
+"system" include directories list, and the declaration there take precedence,
+I guess? The root of the problem BTW is that sizeof(unsigned long) is
+
+ * 4 for MSVC and MinGW-born GCCs,
+ * 8 for Clang (and, strangely, Cygwin GCC; why don't we get runtime
+ errors?).
+
+The fix is to add `define=BOOST_USE_WINDOWS_H`. I don't even know what's the
+point of not having it as a default.