This avoids us having no compilers at all for targets that are composed
entirely of objects with no sources.
Now we will always have a compiler for a target even if it is composed
entirely of objects generated with custom targets unless it has
completely unknown sources.
When doing a compile-only self.compile() check, cl.exe expects an output
file of type *.obj, and passing it `output.exe` makes it write the
compiled object to `output.exe.obj`. Detect when we're doing
a compily-only check and change the suffix ourselves.
This is fine most of the time, but fails when you need to inspect the
compiled file manually.
Also store stdout/stderr on the returned object.
Check if the compiler prefixes an underscore to global symbols. This is
useful when linking to compiled assembly code, or other code that does
not have its C symbol mangling handled transparently by the compiler.
C symbol mangling is platform and architecture dependent, and a helper
function is needed to detect it. Eg: Windows 32-bit prefixes underscore,
but 64-bit does not. Linux does not prefix an underscore but OS X does.
Can't just #include them and use them directly in unity builds. Inline
assembly is a thing, but it's not trivial and is deprecated with some
compilers. Just build them separately and link them in. Ideally the user
would then use LTO to ensure the same result.
Also C++ compilers can build .S assembly files. This wasn't noticed
earlier because most people were also using C compilers in their C++
projects and we would fall back to using the C compiler for building the
assembly files. Now we have a test for this.
This was trivial to add; except that we needed a new LLVM IR rule
because the compiler emits warnings if you pass any special arguments to
it such as include arguments or dependency arguments.
Closes#1089
Instead of adding it everywhere manually, create a wrapper called
mesonlib.Popen_safe and use that everywhere that we call an executable
and extract its output.
This will also allow us to tweak it to do more/different things if
needed for some locales and/or systems.
Closes#1079
We were checking for builtins explicitly like this because the ordinary
checks don't work for builtins at all. We do exactly the same check as
Autoconf and it doesn't work with Autoconf either (Autoconf is broken!)
So now we check for it in two ways: if there's no #include in prefix, we
check if `__builtin_symbol` exists (has_function allows checking for
functions without providing includes). If there's a #include, we check
if `symbol` exists.
The old method was causing problems with some buggy toolchains such as
MSYS2 which define some builtins in the C library but don't expose them
via headers which meant that `__builtin_symbol` would be found even
though `symbol` is not available.
Doing this allows people to always get the correct answer as long as
they specify the includes that are required to find a function while
also not forcing people to always specify includes to find a function
which is cumbersome.
Closes#1083
Unlike Linux and OS X, when a library is loaded, all the symbols aren't
loaded into a single namespace. You must fetch the symbol by iterating over
all loaded modules.
So, we shouldn't use /FORCE:UNRESOLVED since that is not what modules do
on Windows. Instead, we now do exactly what GModule does on Windows.
Also use `void` for functions that take no arguments.
We want compiler check arguments (-O0, -fpermissive, etc) to override
all other arguments, and we want extra_args passed in by the build file
to always override everything.
To do this properly, we must split include arguments out, append them
first, append all other arguments as usual, and then append the rest.
As part of this, we also add the compiler check flags to the
cc.compiles() and cc.links() helper functions since they also most
likely need them.
Also includes a unit test for all this.
We currently pass cross-compiler options to the native compiler too and
when cross-compiling from Windows to Linux, `options` will contain
Linux-specific options which doesn't include `c_winlibs`.
The proper fix is to allow cross-info files to specify compiler options
and to maintain both cross and native compiler options in coredata, but
that will have to be done after the 0.36.0 release.
Also fixes a typo in MinGW cpp_winlibs option setting.
Closes#1029
Move CCompiler.compile to Compiler.compile so that ValaCompiler can use
it. Also rewrite ValaCompiler.sanity_check to use it since it does
a simple compile check.
At the same time, it enhances ExternalLibrary to support arguments for
languages other than C-like.
Includes a test for this that links against zlib through Vala.
Closes#983
With MSVC 2013 and newer, using pre-compiled headers with .pdb debugging
fails with the following error message:
fatal error C1041: cannot open program database '[...]\prog.pdb'; if multiple CL.EXE write to the same .PDB file, please use /FS
So we use /FS when PCH is enabled. When PCH is enabled and debugging is
disabled, this will have no effect since .pdb files will not be written.
Without this, we error out with an exception if `javac` is found but
`java` isn't:
[...]
File "mesonbuild/interpreter.py", line 1759, in detect_compilers
comp.sanity_check(self.environment.get_scratch_dir(), self.environment)
File "mesonbuild/compilers.py", line 1279, in sanity_check
pe = subprocess.Popen(cmdlist, cwd=work_dir)
File "/usr/lib64/python3.5/subprocess.py", line 947, in __init__
restore_signals, start_new_session)
File "/usr/lib64/python3.5/subprocess.py", line 1551, in _execute_child
raise child_exception_type(errno_num, err_msg)
FileNotFoundError: [Errno 2] No such file or directory: 'java'
Clang++ doesn't allow that, but void* will always be allowed because
lots of projects depend on that.
error: cast from pointer to smaller type 'int' loses information
Need to pass -fpermissive to force C++ compilers to only warn about our
non-conformant code that tests for a symbol being defined.
Also do a simple #ifdef check first in has_header_symbol to allow
arbitrary macros to be detected which would not have been detected
earlier. This follows what AC_CHECK_DECL does.
Closes#958
Every other build system does this, and at least OS X, iOS, and Android
depend on this to select the OS versions that your application is
targetting.
At the same time, just use a wrapper for self.run and self.links to
share the common (identical) code.
Simply placing a reference to it isn't enough for the linker to try and
think it's being used and do a symbol availability check with
-Wl,-no_weak_imports on OS X ld.
This is needed to ensure that symbol availability checks actually fail
at link time (instead of at runtime) which is necessary for has_function
to work correctly.
The Autoconf-style check we were doing gives false positives when the
linker uses the prototype defined in the SDK header to decide whether
a function is available or not.
For example, with macOS 10.12, clock_gettime is now implemented
(alongwith other functions). These functions are always defined in the
XCode 8 SDK as weak imports and you're supposed to do a runtime check to
see if the symbols are available and use fallback code if they aren't.
The linker will always successfully link if you use one of those symbols
(without a runtime fallback) even if you target an older OS X version
with -mmacosx-version-min. This is the intended behaviour by Apple.
But this makes has_function useless because to test if the symbol is
available, we must know at link-time whether it is available.
To force the linker to do the check at link-time you must use
'-Wl,-no_weak_imports` *and* use the prototype in time.h which has an
availability macro which tells the linker whether the symbol is
available or not based on the -mmacosx-version-min flag.
An autoconf-style check would override this prototype and use its own
which would result in the linker thinking that the function is always
available (a false positive). Worse, this would manifest at runtime and
might not be picked up immediately.
We now use the function prototype in the user-provided includes if the
'prefix' kwarg contains a `#include` and use the old Autoconf-style
check if not. I've tested that the configure checks done by GStreamer
and GLib are completely unaffected by this; at least on Linux.
The next commit will also add `-Wl,-no_weak_imports` to extra_args by
default so that Meson avoids this mess completely. We always want this
because the user would not do a has_function check if they have
a runtime fallback for the function in their code.
This is done by adding a new compiler method called 'no_warn_args' which
returns the argument(s) required for the compiler to emit no warnings
for that compilation.
This take care of not disabling warnings for C code compiled into the
BuildTarget. Also includes tests for ensuring all this.
https://github.com/mesonbuild/meson/issues/864
Jussi Pakkanen