A current rather untyped storage of options is one of the things that
contributes to the options code being so complex. This takes a small
step in synching down by storing the compiler options in dicts per
language.
Future work might be replacing the langauge strings with an enum, and
defaultdict with a custom struct, just like `PerMachine` and
`MachineChoice`.
I've combined the two ppc checks into one, since they're not very
complicated and added power machintosh, which is used by PPC macs
running older versions of osx.
Fixes: #6746
The previous diagnosis (that the wrapper that injects the -target
option to clang caused the issues) doesn't seem right; the issue
was that LLD's MinGW frontend (which is invoked when ld.lld is given
a COFF argument to the -m option) didn't handle the --version argument
before LLD 9.0.0.
Mips architectures may have `el` on the end, to differentiate the little
endian from the big endian variants. We don't encode endianness in the
cpu names, so ensure we've stripped that.
Fixes#6655
The 0.53.2 test suite has a new failure:
Traceback (most recent call last):
File "run_project_tests.py", line 1024, in <module>
detect_system_compiler()
File "run_project_tests.py", line 964, in detect_system_compiler
comp = env.compiler_from_language(lang, MachineChoice.HOST)
File "/build/meson/src/meson-0.53.2/mesonbuild/environment.py", line 1565, in compiler_from_language
comp = self.detect_cuda_compiler(for_machine)
File "/build/meson/src/meson-0.53.2/mesonbuild/environment.py", line 1131, in detect_cuda_compiler
linker = CudaLinker(compiler, for_machine, 'nvlink', CudaCompiler.LINKER_PREFIX, [], version=CudaLinker.parse_version())
File "/build/meson/src/meson-0.53.2/mesonbuild/linkers.py", line 963, in __init__
super().__init__('nvlink', *args, **kwargs)
TypeError: __init__() takes 6 positional arguments but 7 positional arguments (and 1 keyword-only argument) were given
Fixes: c708c52ca2 ("linkers: Update the linker names to be more consistent")
Emscripten does have a stand alone linker, wasm-ld. This patch adds the
linker, adds detection for the linker, and removes the IsLinkerMixin for
emscripten. This is a little more correct, and makes the code a lot
cleaner and more robust.
This is more correct, and forces the target(s) to be rebuilt if the
PDB files are missing. Increases the minimum required Ninja to 1.7,
which is available in Ubuntu 16.04 under backports.
We can't do the same for import libraries, because it is impossible
for us to know at configure time whether or not an import library will
be generated for a given DLL.
This makes two basic changes, 1 it moves the name of the linker into the
linker class, this should reduce the number of errors and typos, and
ensure that a linker always has one name. This then renames the linkers
to have more consistent names.
Posix/gnu linkers are called ld.<name>: ld.gold, ld.lld, ld.solaris.
Apple linkers are renamed ld64.
This allows Meson native-file [properties] to be used.
This avoids the need to call meson from a script file or have a
long command line invocation of `meson setup`
The method meson.get_native_property('prop', 'fallback') is added.
The native file can contain properties like
```
[properties]
myprop1 = 'foo'
mydir2 = 'lib/custom'
```
Then from within `meson.build`
```meson
x1 = meson.get_native_property('myprop1')
thedir = meson.get_native_property('mydir2', 'libs')
```
fallback values are optional
The rust code is ugly, because rust is annoying. It doesn't invoke a
linker directly (unless that linker is link.exe or lld-link.exe),
instead it invokes the C compiler (gcc or clang usually) to do it's
linking. Meson doesn't have good abstractions for this, though we
probably should because some of the D compilers do the same thing.
Either that or we should just call the c compiler directly, like vala
does.
This changes the public interface for meson, which we don't do unless we
absolutely have to. In this case I think we need to do it. A fair number
of projects have already been using 'ld' in their cross/native files to
get the ld binary and call it directly in custom_targets or generators,
and we broke that. While we could hit this problem again names like
`c_ld` and `cpp_ld` are far less likely to cause collisions than `ld`.
Additionally this gives a way to set the linker on a per-compiler basis,
which is probably in itself very useful.
Fixes#6442
* environment: Fix passing always args to a number of less common linkers
These are mostly (oops xilink) proprietary linkers I can't use for
various reasons.
Fixes: #6332
* Add intelfix from scivision.
* Ifort fix from scivision.
* PGI fix from scivision.
* Cuda fix from scivision.
* Fix linker passing for armclang.
If a user passes -fuse-ld=gold to gcc or clang, they expect that they'll
get ld.gold, not whatever the default is. Meson currently doesn't do
that, because it doesn't pass these arguments to the linker detection
logic. This patch fixes that. Another case that this is needed is with
clang's --target option
This is a bad solution, honestly, and it would be better to use $LD or a
cross/native file but this is needed for backwards compatability.
Fixes#6057
This dumps xild on mac and linux. After a lot of reading and banging my
head I've discovered we (meson) don't care about xild, xild is only
useful if your invoke ld directly (not through icc/icpc) and you want to
do ipo/lto/wpo. Instead just make icc report what it's actually doing,
invoking ld or ld64 (for linux and mac respectively) directly. This
allows us to get -fuse-ld working on linux.
Since I spent three days banging my head against this it seems
reasonable that other people might also run into this problem. It can
happen if you're trying to use microsoft's link.exe, but also have the
dmd bin directory at the tail of your %PATH%, among other reasons.
Since we pass the whole compiler class (as a type()) we don't need to
also pass it's LINKER_PREFIX attribute, we can just get it from the
type we're passing.
Rather than trying to figure out if we're using MSVC based on
environment variables, then trying to get the C compiler and test some
attributes, get the C compiler and see if it's MSVC. This is much more
reliable and we were already doing it anyway.
Since version 9.1, GCC provides support for the D programming language. Thus it
is easy to build a cross-compiler for D, such as aarch64-unknown-linux-gnu-gdc.
However to cross-compile a Meson project using D, using a cross build definition
such as the following is not enough:
```
[binaries]
d = '/path/to/aarch64-unknown-linux-gnu-gdc'
exe_wrapper = '/path/to/qemu-aarch64-static'
[properties]
needs_exe_wrapper = true
[host_machine]
system = 'linux'
cpu_family = 'aarch64'
cpu = 'cortex-a53'
endian = 'little'
```
Indeed, "exe_wrapper" is not be taken into account. Build will fail with:
```
Executables created by D compiler /path/to/aarch64-uknown-linux-gnu-gdc are not runnable.
```
This patch fixes this by reworking:
- detect_d_compiler() to properly get exe_wrapper and D compilers and detect the
one available.
- Dcompiler to properly handle exe_wrapper.
fixes regression for systems with nvcc installed--perhaps why not previously caught on CI.
Just a simple typo--missing a positional argument to CudaCompiler()
0c22798b1a is the first bad commit File
"meson\mesonbuild\environment.py", line 1066, in detect_cuda_compiler
return CudaCompiler(ccache + compiler, version, for_machine, is_cross, exe_wrap, host_compiler=cpp_compiler,
linker=linker)# Please enter the commit message for your changes. Lines starting
TypeError: __init__() missing 1 required positional argument: 'info'# with '#' will be ignored, and an empty message aborts the commit.
Now that the linkers are split out of the compilers this enum is
only used to know what platform we're compiling for. Which is
what the MachineInfo class is for
This is similar to what we currently do for scan-build except there is
no environment variable to choose a specific clang-format to run. If an
environment variable is needed for better control, we can add it later.
Both scan-build and llvm-config need the same list of LLVM version
suffixes. It is better to keep the list at a common place instead of
having several copies in different files, which is likely to become
out-of-sync when the list is updated.
Versioning of executables is not related to the operating system kernel.
It is possible for a Linux distribution to support multiple versions of
LLVM in a way similar to FreeBSD. For example, on Debian, you can use
'apt install clang-tools-7' to install the versioned 'scan-build-7'
executable without bringing the unversioned 'scan-build' into the
environment. Therefore, we should not skip the version list on Linux.
It also makes it consistent with the behavior of llvm dependency, which
does not change the search list depending on the operating system.
This commit also fixes the version suffix for Debian. Debian stops using
the minor version number on the executable after version 7, so it should
be 'scan-build-7', not 'scan-build-7.0'.
This is a follow-up of https://github.com/mesonbuild/meson/pull/5918.
Detect scan-build the same way when trying to launch it and when
generating the target.
The detection method is:
1. look within SCANBUILD env variable
2. shutil.which('scan-build')
3. *on non-linux platforms only*: go through all the possible
name candidates and test them individually.
The third step is added following this comment
https://github.com/mesonbuild/meson/pull/5857#issuecomment-528305788
However, going through a list of all the possible candidates is neither
easily maintainable nor performant, and is therefore skipped on
platforms that should not require such a step (currently, only Linux
platforms).
This is a follow-up to the issue raised by @lantw44 during PR:
https://github.com/mesonbuild/meson/pull/5857
Solaris 11.3 & earlier sent the --version output to stderr, but
Solaris 11.4 moved it to stdout in an attempt to be more compatible
with the GNU tools, so look for it in both streams of output.
Signed-off-by: Alan Coopersmith <alan.coopersmith@oracle.com>
llvm-mingw uses a wrapper script to inject (among other things) a
-target argument into the clang command, which breaks -Wl,--version.
This confuses Meson into thinking the linker is some unknown version of
Apple ld, which breaks builds.
This patch makes it detect and recover from the issue.
Fixes#5910
Instead of the DynamicLinker returning a hardcoded value like
`-Wl,-foo`, it now is passed a value that could be '-Wl,', or could be
something '-Xlinker='
This makes a few things cleaner, and will make it possible to fix using
clang (not clang-cl) on windows, where it invokes either link.exe or
lld-link.exe instead of a gnu-ld compatible linker.
The regex was incorrect, so it was matching 'ARM64' with 'ARM'.
Make the regex more specific so that it matches:
Microsoft (R) C/C++ Optimizing Compiler Version 19.16.27031.1 for x64
Microsoft (R) C/C++ Optimizing Compiler Version 19.16.27031.1 for x86
Microsoft (R) C/C++ Optimizing Compiler Version 19.16.27031.1 for ARM64
Microsoft (R) 32-bit C/C++ Optimizing Compiler Version 16.00.30319.01 for 80x86
etc.
These two functions are different only in the classes that they use, a
couple of simple in-line ternaries takes care of that and reduces code
duplication.
* coredata: Correctly handle receiving a pipe for native/cross files
In some cases a cross/native file may be a pipe, such as when using bash
process replacement `meson --native-file
<([binaries]llvm-config='/opt/bin/llvm-config')`, for example. In this
case we copy the contents of the pipe into a file in the meson-private
directory so we can create a proper ninja dependency, and be able to
reload the file on --wipe/--reconfigure. This requires some extra
negotiation to preserve these native/cross files.
Fixes#5505
* run_unitests: Add a unit test for native files that are pipes
Using mkfifo.
i86pc may be either 32-bit or 64-bit, so use existing compiler checks
to determine if it should return 'x86' or 'x86_64'.
Signed-off-by: Alan Coopersmith <alan.coopersmith@oracle.com>
In most cases instead pass `for_machine`, the name of the relevant
machines (what compilers target, what targets run on, etc). This allows
us to use the cross code path in the native case, deduplicating the
code.
As one can see, environment got bigger as more information is kept
structured there, while ninjabackend got a smaller. Overall a few amount
of lines were added, but the hope is what's added is a lot simpler than
what's removed.
the problem here is, that get_custom_target_provided_libraries iterated
over all generated sources of a target. In each output we check if this
is a library or not. In projects like EFL we have added a lot of
generated target to many different targets, so the iterating of the
output is rather consistent, with this commit we drop from 19% of the
time spending in get_custom_target_provided_libraries down to 3.51%.
Handling the PKG_CONFIG_PATH variable in meson introduces a new problem
for caching dependencies. We want to encode the pkg_config_path (or
cross_pkg_config_path if we're cross compiling) to be part of the key,
but we don't want to put that into the key for non-pkg-config
dependencies to avoid spurious cache misses (since pkg_config_path isn't
relevant to cmake, for example). However, on a cache lookup we can't
know that a dependency is a pkg-config dependency until we've looked in
the cache.
My solution is a two layer cache, the first layer remains the same as
before, the second layer is a dict-like object that encapsulates the
dependency type information and uses pkg_config_path and
cross_pkg_config_path as a sub key (and could be extended easily for
other types). A new object type is introduced to encapsulate this so
that callers don't need to be aware of the implementation details.
Meson itself *almost* only cares about the build and host platforms. The
exception is it takes a `target_machine` in the cross file and exposes
it to the user; but it doesn't do anything else with it. It's therefore
overkill to put target in `PerMachine` and `MachineChoice`. Instead, we
make a `PerThreeMachine` only for the machine infos.
Additionally fix a few other things that were bugging me in the process:
- Get rid of `MachineInfos` class. Since `envconfig.py` was created, it
has no methods that couldn't just got on `PerMachine`
- Make `default_missing` and `miss_defaulting` work functionally. That
means we can just locally bind rather than bind as class vars the
"unfrozen" configuration. This helps prevent bugs where one forgets
to freeze a configuration.
Intel helpfully provides a cl.exe that is indistinguishable from
Microsoft's cl.exe in output, but has the same behavior as icl.exe.
Since icl and ifort will only be present in your path if you've started
an Intel command prompt search for that first.
Currently C++ inherits C, which can lead to diamond problems. By pulling
the code out into a standalone mixin class that the C, C++, ObjC, and
Objc++ compilers can inherit and override as necessary we remove one
source of diamonding. I've chosen to split this out into it's own file
as the CLikeCompiler class is over 1000 lines by itself. This also
breaks the VisualStudio derived classes inheriting from each other, to
avoid the same C -> CPP inheritance problems. This is all one giant
patch because there just isn't a clean way to separate this.
I've done the same for Fortran since it effectively inherits the
CCompiler (I say effectively because was it actually did was gross
beyond explanation), it's probably not correct, but it seems to work for
now. There really is a lot of layering violation going on in the
Compilers, and a really good scrubbing would do this code a lot of good.
Some things, like `method[...](...)` or `x: ... = ...` python 3.5
doesn't support, so I made a comment instead with the intention that it
can someday be made into a real annotation.
The cl.exe from Visual Studio 2010 and earlier report '80x86', not
'x86', for the architecture that the compiler supports. So, we ought
to check for that as well to see whether we are building for 32-bit x86.
The out-of-source build syntax for gcovr 4.2 is different compared to
previous versions and therefore an update was needed. In researching the
most appropriate solution it was found that any gcovr version older than
3.3 always resulted in 0% coverage. Because of this, rather than adding
an additional layer of logic, some already existing logic was modified
to ensure correct syntax for the new version, while versions older than
3.3 are flagged as not supported.
Closes mesonbuild#5089.
From (almost) all points of view, the Xtensa toolchain can be treated as
a regular GCC toolchain.
This patch adds very basic support so that, at least, meson does not
fail when trying to use "xt-xcc" (which makes it possible to use it
without problems).
* coredata: store cross/native files in the same form they will be used
Currently they're forced to absolute paths when they're stored in the
coredata datastructure, then when they're loaded we de-absolute path
them to check if they're in the system wide directories. This doesn't
work at all, since the ninja backend will generat a dependency on a
file that is in the source directory unless the path was already given
as absolute. This results in builds being retriggereed forever due to
a non-existant file.
The right way to do this is to figure out whether the file is in the
build directory, is absolute, or is in one of the system paths at
creation time, and store that path as absolute. Then the code that
reads the file and the code that generates the dependencies in the
ninja backend just takes the computed list and there is no mismatch
between them.
Fixes#5257
* run_unittests: Add a test for correct native file storage
This tests the bug in #5257
When using clang as an objc/objc++ compiler, identify if it's a Windows
targeted compiler, so that GnuLikeCompiler::get_pic_args() doesn't use
'-fPIC', which clang considers an error for the Window target.
Future work: Factor out parsing the clang target string from the
detectors for various languages.
This can be useful to test a local ninja version (for example while developing
changes to ninja or samurai) without modifying the PATH.
The ninja binary that is detected is then hardcoded in the build.ninja
rules for scan-build and clean, so that it is always used until reconfiguration.
This function is used just once. It also seems all policy and no
mechanism (it raises, it calls the same function to do all the work
twice in a simple way). This makes it seem to be as a good candidate for
inlining.
`environment` and `coredata` are woefully intertwined and while this
change doesn't fix that, but at least it makes it easier to follow.
Instead of hard-coding the fact that load_configs() searches for files under
meson/native, pass in the subdirectory allowing the cross-file code to use the
same logic.
We need to match the "clang --version" output on OpenBSD:
$ clang --version | head -1
OpenBSD clang version 7.0.1 (tags/RELEASE_701/final) (based on LLVM 7.0.1)
1. They (and the others) all use PerMachineDefaultable. It's not the
best class, but consistency come first. (It and all of them can be
improved accross the board later.)
2. They use `None` as the default argument so as not to mutate what's
effectively a global variables. (Thanks @dcbaker!)
3. They have a `fallback` field to centralize authority on when
environment variables should be consulted.
First of all, I'd like compilers and other modules that environment.py
currently imports to be able to take these without creating
hard-to-follow module cycles.
Second of all, environment.py's exact purpose seems a bit obscured.
Splitting the data types (and basic pure functions) from the more
complex logic that infers that data seems like a good way to separate
concerns.
This allows the person running configure (either a developer, user, or
distro maintainer) to keep a configuration of where various kinds of
files should end up.
Instead use coredata.compiler_options.<machine>. This brings the cross
and native code paths closer together, since both now use that.
Command line options are interpreted just as before, for backwards
compatibility. This does introduce some funny conditionals. In the
future, I'd like to change the interpretation of command line options so
- The logic is cross-agnostic, i.e. there are no conditions affected by
`is_cross_build()`.
- Compiler args for both the build and host machines can always be
controlled by the command line.
- Compiler args for both machines can always be controlled separately.
John Ericson