Meson accidentally strips '-isystem' from C build args like ['-isystem',
'/path/to/headers'] if the compiler includes the current working directory
in the header search paths. The root cause is that '-isystem'[8:] evaluates
to an empty string and os.path.realpath('') returns the absolute path to
the current working directory, causing meson to think that '-isystem'
specifies a default include path.
Different compiler versions varies whether the current working directory is
in its search paths. For example, on Ubuntu 21.04:
# gcc -xc -v -E -
gcc version 10.3.0 (Ubuntu 10.3.0-1ubuntu1)
#include "..." search starts here:
#include <...> search starts here:
.
/usr/lib/gcc/x86_64-linux-gnu/10/include
/usr/local/include
/usr/include/x86_64-linux-gnu
/usr/include
End of search list.
While on Ubuntu 24.04:
# gcc -xc -v -E -
gcc version 13.2.0 (Ubuntu 13.2.0-23ubuntu4)
...
#include "..." search starts here:
#include <...> search starts here:
/usr/lib/gcc/x86_64-linux-gnu/13/include
/usr/local/include
/usr/include/x86_64-linux-gnu
/usr/include
End of search list.
Do not check the '-isystem=' and '-isystem/path/to/header' cases when the
option is '-isystem' but the path that follows is not a default search
path.
Signed-off-by: Junjie Mao <junjie.mao@hotmail.com>
In particular, it is only ever constructed once, and all but two of its
arguments are always passed. So there's no reason or valid justification
to initialize them as None or pretend that you have to check them for
None values. Nor in fact does the command list need to be a
default_factory.
There are also two instance attributes which it is never instantiated
with, but which are tracked after the fact. Don't add this to the
signature for `__init__` as that is misleading.
Fallout from the OptionStore refactor, and specifically commit
9a6fcd4d9a. The `std` object was migrated
from having an option itself, to having the value fetched and saved
directly. In most cases, this also meant avoiding `.value`, but in a
couple cases this refactor went overlooked, and crashed at runtime.
Only affects Elbrus and Intel C++ compilers, seemingly.
Fixes#13401
For other reasons, Meson transforms "-Wno-x" into "-Wx -Wno-x" for GCC,
but this breaks with "-Wno-attributes=x" with:
```
cc1plus: error: arguments ignored for '-Wattributes='; use '-Wno-attributes=' instead
```
Suppress that workaround for -Wno-attributes=.
Closes: https://github.com/mesonbuild/meson/issues/13022
c1076241af changed the logic in multiple
places, in particular it looks like it was assumed that is_cross is always
the same as need_exe_wrapper(), but that's not true.
Also the commit only talks about mypy, so this was definitely not intended.
This reverts all the cases where need_exe_wrapper() was introduced back to
is_cross.
The change in backends.py could be a correct simplification, but I don't know
the code base enough, so reverting that too.
See #13403 and #13410
_get_gnu_compiler_defines and _get_clang_compiler_defines were broken
by not defining the language they used.
Neither GCC nor Clang infer the language based on the driver name which means
`self.defines` isn't populated correctly in compilers/cpp.py.
e.g.
```
$ echo "" | g++ -E -dM - | grep -i cplus
$ echo "" | g++ -x c++ -E -dM - | grep -i cplus
#define __cplusplus 201703L
```
Fix that by passing '-cpp -x LANGUAGE' as a first pass. If it fails, try
again without '-cpp -x LANGUAGE' as before, as its portability isn't
certain. We do '-cpp' because during testing, I found Fortran needs this,
although per below, I had to drop Fortran in the end and leave it to the
fallback (existing) path.
Without this change, a63739d394 is only
partially effective. It works if the system has injected Clang options
via /etc/clang configuration files, but not by e.g. patching the driver
(or for GCC there too).
Unfortunately, we have to wimp out for Fortran and fallback to the
old method because you need the language standard (e.g. -x f95).
Make the debug & error message strings consistent between the GCC and Clang probes.
Copy-paste error. Here, we're scraping pre-processor tokens, not checking
for the compiler type.
There are two issues:
1. has_header() wants just the header name without surrounding <>
or similar; it fails otherwise.
2. has_header() returns a tuple of two bools, where the first element
determines whether or not the header has been found. So use
that element specifically, otherwise the tuple will always evaluate
to true because it is not empty.
Fixes: 675b47b069 ("compilers: cpp: improve libc++ vs libstdc++ detection (again)")
Clang is such a great compiler! Not.
Compilers have enhanced diagnostics for some kinds of "well known"
undeclared identifiers, telling you exactly which header you might have
forgotten to include. The reason why clang needs an option GCC doesn't
need is because clang's fixit suggestions, unlike GCC's actually
*changes the type of the error*, as a result of a fixit of all things.
After the fixit suggestion grants this error the right to be ignored,
we start having to add clang-specific options.
Follow-up to https://github.com/mesonbuild/meson/issues/9140
Upstream clang bug, which appears to be going nowhere:
https://github.com/llvm/llvm-project/issues/33905
* Previously, cuda would just plainly prepend `-l` to the libname.
* By relying on the host compiler to find libraries, we now get
more subtle failures, such as CUDA modules not being found
anymore.
* We need to simplify these CUDA modules when nvcc is used for
linking, since this may have side-effects from the cuda toolchain.
Closes: #13240
Emscripten version numbers are unrelated to Clang version numbers,
so it is necessary to change the version checks for `c_std=c17` & co.
Without that, no project that defaults to C17 or newer will build with
Emscripten.
See https://github.com/pyodide/pyodide/discussions/4762 for more
context. Also note that this bug caused defaulting to C17 in
scikit-learn to be reverted (scikit-learn#29015), and it may be a
problem for SciPy 1.14.0 too since that release will upgrade from C99
to C17.
Co-authored-by: Loic Esteve <loic.esteve@ymail.com>
Using the keyword argument dependencies with compiler.links() for vala doesn't work as the library being linked to needs to be prefixed with --pkg= before being passed to valac.
GCC only has very limited support for Objective-C and doesn't support
any of the modern features, so whenever Clang is available, it should be
used instead. Essentially, the only reason to ever use GCC for
Objective-C is that Clang simply does not support the target system.
C#, Java, and Swift targets were manually collecting compiler arguments
rather than using the helper function for this purpose. This caused each
target type to add arguments in a different order, and to forget to add
some arguments respectively:
C#: /nologo, -warnaserror
Java: warning level (-nowarn, -Xlint:all, -Xdoclint:all), debug arguments
(-g, -g:none), -Werror
Swift: -warnings-as-errors
Fix this. Also fix up some no-longer-unused argument processing in the
Compiler implementations.
If nasm is not defined in cross file binaries we can fallback to build
machine nasm.
When cross compiling C code we need a different gcc binary for native
and cross targets, e.g. `gcc` and `x86_64-w64-mingw32-gcc`. But when
cross compiling NASM code the compiler is the same, it is the source
code that has to be made for the target platform. We can thus use nasm
from build machine's PATH to cross compile for Windows on Linux for
example. The difference is the arguments Meson will pass when invoking
nasm e.g. `-fwin64`. That is already handled by NasmCompiler class.
Junjie Mao
Jonathan Schleifer
Kacper Michajłow
Eli Schwartz
H. Vetinari
Sam James
Jussi Pakkanen
Christoph Reiter
L. E. Segovia
Dylan Baker
Barnabás Pőcze
TheHillBright
Rohit Goswami
David Seifert
Axel Ricard
Ralf Gommers
Tristan Partin
Ben Corby
Benjamin Gilbert
Xavier Claessens
William D. Jones