Profiling showed that we were spending 25s inside os.path.realpath()
on Windows while generating compile lines for build.ninja, inside
NinjaBackend.generate()
The real path for these will not (should not) change during a single
meson invocation, so cache all these.
Brings build.ninja generation from 73s to 47s on my machine.
Which adds the `use-set-for-membership` check. It's generally faster in
python to use a set with the `in` keyword, because it's a hash check
instead of a linear walk, this is especially true with strings, where
it's actually O(n^2), one loop over the container, and an inner loop of
the strings (as string comparison works by checking that `a[n] == b[n]`,
in a loop).
Also, I'm tired of complaining about this in reviews, let the tools do
it for me :)
Adds a new maximum warning level that is roughly equivalent to "all warnings".
This adds a way to use `/Wall` with MSVC (without the previous broken warning),
`-Weverything` with clang, and almost all general warnings in GCC with
strictness roughly equivalent to clang's `-Weverything`.
The GCC case must be implemented by meson since GCC doesn't provide a similar
option. To avoid maintenance headaches for meson, this warning level is
defined objectively: all warnings are included except those that require
specific values or are specific to particular language revisions. This warning
level is mainly intended for new code, and it is expected (nearly guaranteed)
that projects will need to add some suppressions to build cleanly with it.
More commonly, it's just a handy way to occasionally take a look at what
warnings are present with some compiler, in case anything interesting shows up
you might want to enable in general.
Since the warnings enabled at this level are inherently unstable with respect
to compiler versions, it is intended for use by developers and not to be set as
the default.
-Wnon-virtual-dtor is not what people think of as a standard warning
flag. It was previously removed from -Wall in
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=16190 on the grounds that
people didn't like it and were refusing to use -Wall at all because it
forced this warning. Instead, it is enabled by -Weffc++ which is
typically not enabled and even comes with GCC documentation warnings
stating that the standard library doesn't obey it, and you might need to
`grep -v` and filter out warnings. (!!!)
It doesn't fit into the typical semantics of Meson's warning_level
option, which usually aligns with compiler standard warning levels
rather than a niche ideological warning level.
It was originally added in commit 22af56e05a,
but without any specific rationale included, and has gone unquestioned
since then -- except by the Meson users who see it, assume there is a
finely crafted design behind it, and quietly opt out by rolling their own
warning options with `add_project_arguments('-Wall', ...)`.
Furthermore a GCC component maintainer for the C++ standard library
opened a Meson bug report specially to tell us that this warning flag is
a "dumb option" and "broken by design" and "doesn't warn about the right
thing anyway", thus it should not be used. This is a reasonably
authoritative source that maybe, just maybe, this flag... is too
opinionated to force upon Meson users without recourse. It's gone beyond
opinionated and into the realm of compiler vendors seem to think that
the state of the language would be better if the flag did not exist at
all, whether default or not.
Fixes#11096
ccache was used in all command lines but disabled using CCACHE_DISABLE
in Compiler.compile() method. Wrapping invokations still has a cost,
especially on Windows.
With sccache things are even worse because CCACHE_DISABLE was not
respected at all, making configure *extremely* slow on Windows when
sccache is installed.
https://github.com/mesonbuild/meson/pull/9287 changed the `optimization=0`
to pass `-O0` to the compiler. This change is reasonable by itself
but unfortunately, it breaks `buildtype=plain`, which promises
that “no extra build flags are used”.
`buildtype=plain` is important for distros like NixOS,
which manage compiler flags for optimization and hardening
themselves.
Let’s introduce a new optimization level that does nothing
and set it as the default for `buildtype=plain`.
We need to support cases where the library might be called "foo.so" and
therefore we check for exact matches too. But this also allows
`cc.find_library('libfoo')` to find libfoo.so, which is strange and
won't work in many cases. Emit a warning when this happens.
Fixes#10838
We *mostly* just need to do the same thing. Plug in one utility method
to make sanity_check_impl find the right compile args, and plug in
DEVNULL to the test run. It's that simple.
This solves a few inconsistencies. The main one is that fortran never
logged the sanity checks to the Meson debug log, making it hard to
debug.
There's also some interesting quirks we built up in the dedicated
fortran handling. For example:
- in commit 5b109c9ad2 we added cwd to
building the fortran executable, with a wordy comment about how the
compiler has defects. But the clike base has always done that on
general principle anyway, so we would never have had that bug in the
first place.
- in commit d6be7822a0 we added special
deletion of an old "bad existing exe file" just for fortran. Looking
at the PR discussion for this odd requirement, it turns out that the
real problem is mixing WSL and native Windows without deleting the
build directory. This is apparently fortran specific simply because
"contemporary Windows 10 Fortran users" switch between the two?
The actual problem is that this never used .exe as the output name, so
Windows thinks you want to run something other than the thing you
asked to run, because it's not even a Window executable. But... the
common clike handling could have fixed that without needing special
cases.
This used to be fine, until imports were removed from this file. Now a
function annotated as T.NoReturn doesn't actually tell mypy that it
cannot return, though, so we manually do it.
We no longer need these upfront at all, since we now import the ones we
need for the language we are detecting, at the time of actual detection.
This avoids importing 28 files, consisting of just under 9,000 lines of
code, at interpreter startup. Now, it is only imported depending on
which languages are invoked by add_languages, which may not even be
anything. And even if we do end up importing a fair chunk of it for
C/C++ projects, spreading the import cost around the interpreter runtime
helps responsiveness.
Only import the ones we need for the language we are detecting, once we
actually detect that language.
This will allow finally dropping the main imports of these files in a
followup commit.
Instead of comparing against specific compiler classes, check the
logical compiler id or language etc.
In a couple cases, we seem to be missing a couple things by being a bit
too strict about the exact class type.
This is wasteful and generally unneeded, since code can just use the
compiler they detected instead of manually poking at the internals of
this subpackage.
It also avoids importing an absolute ton of code the instant one runs
`from . import compilers`
In commit d326c87fe2 we added a special
holder object for cached compilation results, with some broken
attributes:
- "command", that was never set, but used to print the log
- "args", that was set to some, but not all, of the information a fresh
compilation would log, but never used
Remove the useless args attribute, call it command, and use it properly.
We compared a Visual Studio (the IDE) version, but we wanted a MSVC (the
compiler) version. This caused the option to be passed for a few too
many versions of MSVC, and emit a "D9002 : ignoring unknown option" on
those systems.
Compare the correct version using the version mapping from
https://walbourn.github.io/vs-2017-15-7-update/Fixes#10787
Co-authored-by: CorrodedCoder <38778644+CorrodedCoder@users.noreply.github.com>
Meson internally knows about many languages and tools, and *FLAGS
variables, and which languages to use them for. Instead of duplicating
this logic, import it from mesonbuild.*
This logic was originally standalone, but now that it is merged into the
Meson tree we can have a single source of truth.
If this command fails, for example when CXX is something not generic
enough to be a valid universal compiler command (clang -std=c++11
perhaps), we end up with two problems:
- it's impossible to figure out what Meson ran to get that error
- the error report isn't clear on what is stdout and what is stderr, or
even that that is what the message is about.
```
meson.build:1:0: ERROR: Unable to get clang pre-processor defines:
error: invalid argument '-std=c++11' not allowed with 'C'
```
What's C doing there and why is Meson talking about it? Answer: that's
compiler stdout. Say so.
We do something similar when running get_compiler() method checks from
the DSL. This ensures that if errors happen, the log file we tell people
to check actually works.
Specifically, this is a combination of the following:
- Revert "visualstudio.py: Apply /utf-8 only on clang or VS2015+"
This reverts commit 6e7c3efa79.
- Revert "Visual Studio: Only use /utf-8 on VS2015 or later or clang-cl"
This reverts commit 8ed151bbd7.
The changes were broken and untested, although this is because of a lack
of general CI testing for all languages on Windows. At least, this broke
the use of ifort, and possibly more.
The changes are fundamentally a bit "exciting", as they step out of the
hierarchy of compiler definitions and apply arguments almost willy-nilly.
And apparently it's leaky all over the place. I don't understand all of
what is going on with it, but it plainly failed to achieve its desired
goal and needs to be rolled back ASAP.
Transpilers need to run on the build machine in order to generate their
output, which can then be taken by a cross-compiler to create the final
output.
MachineChoice is a mesonlib object, not a compilers object, so it makes
no sense to import it from the latter simply because the latter imports
it too. This results in brittle module dependencies and everything
breaking when a refactor removes it from the latter.
... also it is a typing-only import so while we are fixing it to import
from the right place, we can also put it in a type-checking block.
In PR 10263, we didn't account for that we may have initialize the Visual
Studio-like compiler two times, once for a C compiler and once for the
C++ compiler, so we end up with Meson breaking on Visual Studio 2013
or earlier, such as when building GLib.
Fix this by setting up the always_args member of
the VisualStudioLikeCompiler instance during __init__() as needed, so that
we avoid falling into modifying shared objects.
Leak sanitizer can be enabled without the whole AddressSanitizer, this
can be done by passing -fsanitize=leak as documented at [1].
Meson doesn't support this, so add support for it.
[1] https://clang.llvm.org/docs/LeakSanitizer.html
This option is only valid for C++ and ObjC++, but the kwarg is useful
for mixed language targets. Asking for inlines as well, when the
compiler driver is trying to build the C components of a target, results
in gcc emitting:
```
cc1: warning: command-line option ‘-fvisibility-inlines-hidden’ is valid for C++/ObjC++ but not for C
```
Squelch this warning by filtering it out on Meson's side of things.
Otherwise it always returns the value for c++98, starting with MSVC 2017
15.7 or later. Earlier versions are not affected by this mis-feature.
See: https://docs.microsoft.com/en-us/cpp/build/reference/zc-cplusplus?view=msvc-160
This was originally applied as 0b97d58548
but later reverted because it made the CI red. Try it again, now.
Original-patch-by: Dylan Baker <dylan@pnwbakers.com>
Co-authored-by: Dylan Baker <dylan@pnwbakers.com>
[why]
Support for the relatively new mold linker is missing. If someone wants
to use mold as linker `LDFLAGS="-B/path/to/mold"` has to be added instead
of the usual `CC_LD=mold meson ...` or `CXX_LD=mold meson ...`.
[how]
Allow `mold' as linker for clang and newer GCC versions (that versions
that have support).
The error message can be a bit off, because it is generic for all GNU
like compilers, but I guess that is ok. (i.e. 'mold' is not listed as
possible linker, even if it would be possible for the given compiler.)
[note]
GCC Version 12.0.1 is not sufficient to say `mold` is supported. The
expected release with support will be 12.1.0.
On the other hand people that use the un-released 12.0.1 will probably
have built it from trunk. Allowing 12.0.1 is helping bleeding edge
developers to use mold in Meson already now.
Fixes: #9072
Signed-off-by: Fini Jastrow <ulf.fini.jastrow@desy.de>
The compiler flag only exists on Visual Studio 2015 or later, or clang-cl,
and using this always can interfere with compiler feature detection when
this flag is not supported.
So, remove '/utf-8' from always_args if we are on Visual Studio 2013 or
earlier.
This has been broken ever since the original implementation. Due to a
typo, the optimization flag used a zero instead of an uppercase "o",
which the compiler then breaks on during argument parsing because it is
an invalid argument.
Fixes#10267
When something goes wrong with running the compiler in
_symbols_have_underscore_prefix_searchbin, print stderr instead,
as it actually contains helpful output while stdout is usually empty
in this case.
Fixes the following ResourceWarnings:
ResourceWarning: subprocess 25556 is still running
_warn("subprocess %s is still running" % self.pid,
ResourceWarning: Enable tracemalloc to get the object allocation traceback
mesonbuild/compilers/mixins/gnu.py:195: ResourceWarning: unclosed file <_io.BufferedReader name=4>
return gnulike_default_include_dirs(tuple(self.exelist), self.language).copy()
ResourceWarning: Enable tracemalloc to get the object allocation traceback
It is always used as an immutable view so there is no point in doing
copies. However, mypy insist it must implement the same APIs as
Dict[OptionKey, UserOption[Any]] so keep faking it.
The code in the C++ and Fortran compilers' language_stdlib_only_link_flags
method is broken and cannot possibly have ever worked. Instead of
splitting by line, it splits by whitespace and therefore, instead of
the last line of the compiler output:
programs: =/Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin
libraries: =/Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/12.0.0
it is only the last field that has its first 11 characters removed.
Instead of reinventing the wheel with a new and brittle pattern,
reuse get_compiler_dirs.
Fixes: 64c267c49 ("compilers: Add default search path stdlib_only_link_flags", 2021-09-25)
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Using future annotations, type annotations become strings at runtime and
don't impact performance. This is not possible to do with T.cast though,
because it is a function argument instead of an annotation.
Quote the type argument everywhere in order to have the same effect as
future annotations. This also allows linters to better detect in some
cases that a given import is typing-only.
GCC with optimization set to 0 does not actually result in no
optimizations, which can be annoying when trying to use a debugger like
gdb, and finding that your variable has been optimized out. We already
do this with clang, so gcc is a bit of an outlier here.
We didn't consider that it has arguments following it, so the resulting
compiler command line ended up with stuff like:
-L=-rpath-link -L=-L=/path/to/directory -L=more-args
and the directory for rpath-link got eaten up as a regular -L path to
the compiler rather than being passed as -Xlinker to the linker.
Then the -rpath-link would consume the next -Xlinker argument, end up
with the wrong rpath-link (may or may not cause link errors) and then
disappear arguments we need.
As an example failure mode, if the next argument is -soname this treats
the soname text as an input file, which probably does not exist if it
was generated in a subdirectory, and also because it can never be
successfully built in the first place -- though if it did, it would link
to itself which is very wrong.
Kleis Auke Wolthuizen