Instad of having special casing of threads in the backends and
everywehre else, do what we did for openmp, create a real
dependency. Then make use of the fact that dependencies can now have
sub dependencies to add threads.
In Fedora 30 the llvm-config binary has the number of bits in the isa
added as a suffic to the name llvm-config-64 or llvm-config-32, and no
naked llvm-config binary is provided. This commit moves the list of
llvm-config binaries to the constructor so that we can modify the list
on a per-instance basis, and adds the new Feodra names.
Fixes#5059
Instead of trying to hardcode which versions of which OSes are
misconfiguring llvm-config, lets try to use a generic mechanism that
catches most of the broken cases.
If a dynamic lib is built but the tools are linked staticly
(LLVM_BUILD_LLVM_DYLIB=true and LLVM_LINK_LLVM_DYLIB=false) and you
don't enabled shared libs (LLVM_BUILD_SHARED_LIBS=false, which should be
false), then llvm-config will mistakenly return a each static library
with .(so|dll|dylib) appended, instead of -lLLVM-x.y.z. if
LLVM_LINK_LLVM_DYlIB is true this works fine. There's been a bug opened
in upstream since 2014 about this with no responses[1].
So, this patches looks at --shared-mode (which tells us whether
llvm-config was linked staticly or dynamicly), then if the shared-mode
is static it will check and see what it gets. If it gets valid results
then it will go with those, if it doesn't then it will try to build a
set of valid link arguments to link with libLLVM.(so|dll|dylib).
[1] https://bugs.llvm.org/show_bug.cgi?id=19937
The versions of LLVM provided by brew have the same configuration problem that
dragonflyBSD and FreeBSD do, which means that they produce wrong results for
shared linking.
It appears that debian has changed their naming scheme. This might
require revisiting in the future to add llvm-config-7.0 if debian
decides to do that.
If successful, we should identify the method which was successful
If successful, we should report the version found (if known)
If failing, we should identify the methods we tried
Some dependency detectors which had no reporting now gain it
There's all kinds of complexities, inconsistencies and special cases hidden
in the existing behaviour, e.g.:
- boost reports modules requested, and BOOST_ROOT (if set)
- gtest/gmock report if they are a prebuilt library or header only
- mpi reports the language
- qt reports modules requested, and the config tool used or tried
- configtool reports the config tool used
- llvm reports if missing modules are optional (one per line)
We add some simple hooks to allow the dependency object to expose the
currently reported information into the consolidated reporting
Note that PkgConfigDependency() takes a silent: keyword which is used
internallly to suppress reporting. This behaviour isn't needed in
find_external_dependency().
Those tools use our arguments to build a file and execute it to
introspect it at runtime. However, they do not know that you can pass
the full path to the library to use, and ignore the arguments.
The long-term fix for this is to have them output a .c file that Meson
will build for them, which they can then run, but that will require
upstream changes:
https://gitlab.gnome.org/GNOME/gtk-doc/merge_requests/1
Closes https://github.com/mesonbuild/meson/issues/3774
All dependencies were using find_library, has_header, get_define, etc on
self.compiler assuming that it's a compiler that outputs and consumes
C-like libraries. This is not true for D (and in the future, for Rust)
since although they can consume C libraries, they do not use the
C ecosystem.
For such purposes, we now have self.clib_compiler. Nothing uses
self.compiler anymore as a result, and it has been removed.
This replaces calls to .rstrip('git'), .rstrip('svn') with a regex that
takes the leading numbers and dots, and throws away the rest. This also
moves the code up to the ConfigToolDepdency level, since these config
tools are of various quality and some of them are good, and some are
not. This shouldn't affect well behaved tools.
This should future proof LLVM against future suffixes (like someone
doing something strange like using Mercurial as a VCS).
Because FreeBSD's llvm-config adds -l/usr/lib/libexecinfo.so when asked
for system-libs, which is bogus. We'll remove the leading -l from any
argument that also ends with .so.
If LLVM is built from a git mirror instead of from SVN it will have
"git-<very short sha>" at the end of the version. We need to remove that
so that version comparison will work correctly.
Fixes: #2786
f.ex when you don't have the llvm-static package installed, the error
message when generating libs is cryptic and uninformative since we
discard stderr.
LLVM >= 3.9 provides an llvm-config that has a sane mechanism for
selecting static vs dynamic linking. LLVM < 3.9 (but >= 3.5) not so
much. For those older LLVM versions, llvm-config will always provide
arguments for statically linking LLVM, even if there is a library for
dynamic linking.
Fixes#2442
Daniel Mensinger
Dylan Baker
Jon Turney
Jakub Wilk
David Seifert
Nirbheek Chauhan
Jussi Pakkanen
Josh Soref