Spack currently prepends include paths, library paths, and rpaths to the compile line. This causes problems when a header or library in the package has the same name as one exported by one of its dependencies. The *dependency's* header will be preferred over the package's, which is not what most builds expect. This also breaks some of our production codes.
This restores the original cc behavior (from *very* early Spack) of parsing compiler arguments out by type (`-L`, `-I`, `-Wl,-rpath`) and reconstituting the full command at the end.
`<includes> <other_args> <library dirs> <rpaths>`
This differs from the original behavior in one significant way, though: it *appends* the library arguments so that dependency libraries do not shadow those in the build.
This is safe because semantics aren't affected by *interleaving* `-I`, `-L`, and `-Wl,-rpath` arguments with others, only with each other (so the order fo two `-L` args affects the search path, but we search for all libraries on the command line using the same search path).
We preserve the following:
1. Any system directory in the paths will be listed last.
2. The root package's include/library/RPATH flags come before flags of the same type for any dependency.
3. Order will be preserved within flags passed by the build (except system paths, which are moved to be last)
4. Flags for dependencies will appear between the root flags and the system flags, and the flags for any dependency will come before those for *its* dependencies (this is for completeness -- we already guarantee this in `build_environment.py`)
If the OpenMPI build finds the infiniband drivers in /usr/lib64, it adds
-Wl,-rpath -Wl,/usr/lib64 to the OpenMPI wrappers. If the wrappers are using
a compiler outside of /usr, and the OpenMPI wrappers are used to build software
outside of Spack, they will rpath /usr/lib64 into the executable which then has
GLIBC, GLIBCXX runtime errors due to it picking up GCC libraries in /usr/lib64.
This adds the directories specified in "extra_rpaths" to the OpenMPI wrappers,
which allows them to use the correct compiler when invoked outside of Spack
builds.
- previously, output could be confusing when deptypes were only shown for
one dependent when a node had *multiple* dependents
- also fix default coverage of `Spec.tree()`: it previously defaulted to
cover only build and link dependencies, but this is a holdover from
when those were the only types.
* Updated llvm to version 6.0.1. The previous 6.0.0 had an incorrectly declared symbol, discussed at https://reviews.llvm.org/D44140, which, amongst other things, broke py-numba. This version works fine with py-numba.
* Flag the conflict between py-numba and llvm@6.0.0
* Removed a single trailing space to satisfy checks
* First draft of a templight recipe
* Let's be explicit for now, we can be clever later on
* The clang6 version does not work, drop it
* Reorder FIXMEs more logically
* Add patch for current templight develop branch
* Create an artificial stable release of templight
* Take some additional inspiration from llvm package
* Added the templight supporting tools
* Remove old notes from the no-url branch
* Avoid unnecessary join_path
* intel-tbb: Add variant `tm` to disable transactional memory.
Some AMD or very old Intel systems don't support transactional memory. This commit adds a variant `tm` that defaults to `True`, but can be switched off to allow running on those systems.
- Previously, Spack didn't check the arguments you put in version()
directives.
- So, you could do something like this, where there are arguments for a
URL fetcher AND for a git fetcher:
version('1.0', md5='abc123', git='https://foo.bar', commit='feda2343')
- Now, we check the arguments before constructing a fetcher, to ensure
that each package has *only* arguments for a single type of fetcher.
- Also added `test_package_version_consistency()` to the `package_sanity`
test, so that all builtin packages are required to have valid
`version()` directives.
- packagers can specify two top-level fetch URLs if one is `url`
- e.g., `url` and `git` or `url` and `svn`
- allow only one VCS fetcher so we can differentiate between URL and VCS.
- also clean up fetcher logic and class structure
- Packages can remove the top-level `url` attribute and still work
- These are now legal:
- Packages with *only* version-specific URLs (even with gaps)
- Packages with a top-level git/hg/svn attribute and `version`
directives for that.
- If a package has both a top-level hg/git/svn attribute AND a top-level
url attribute, the url attribute takes precedence.
Some packages do not have a `url` and are instead downloaded via `git`,
`hg`, or `svn`. Some packages like `spectrum-mpi` cannot be downloaded at
all, and are placeholder packages for system installations. Previously,
`__init__()` in `PackageBase` crashed if a package did not have a `url`
attribute defined.
I hacked this section of code out, but I have no idea what the
repercussions of that are.
- This hard-codes the hash lengths rather than computing them on import.
- Also cleans up the code in `spack.util.crypto` to make it easier to
understand.
