- Setting specs from lockfiles was not correctly stringifying concretized
user specs.
- Fix `_set_user_specs_from_lockfile`
- Add some validation code to `SpecList` constructor
Spack has evolved to have three types of hash functions, and it's
becoming hard to tell when each one is called. Whlie we aren't yet ready
to get rid of them, we can refactor them so that the code is clearer and
easier to track.
- Add a `hash_types` module with concise descriptors for hashes.
- Consolidate hashing logic in a private `Spec._spec_hash()` function.
- `dag_hash()`, `build_hash()`, and `full_hash()` all call `_spec_hash()`
- `to_node_dict()`, `to_dict()`, `to_yaml()` and `to_json()` now take a
`hash` parameter consistent with the one that `_spec_hash()` requires.
Co-authored-by: Todd Gamblin <tgamblin@llnl.gov>
- ensure that Spec._build_hash attr is defined
- add logic to compute _build_hash when it is not already computed (e.g. for specs prior to this PR)
- add test to ensure that different instance of a build dep are preserved
- test conversion of old env lockfile format to new format
- tests: avoid view creation, DAG display in tests using MockPackage
- add regression test for more-general bug also fixed by this PR
- update lockfile version since the way we are maintaining hashes has changed
- write out backup for version-1 lockfiles and test that
* openjdk
* Removing preferred from latest and adding a few lines for readability.
* Remove bad whitesapce
* removing formatting changes
* stable url for openjdk 1.8
* remove extra blank line
* actual binary
Add version for 'gpu' and variant for 'cuda'. These are currently available from the 'master-gpu' branch.
* hpctoolkit: add version gpu and variant cuda
* The gcc >= 4.8.x requirement only applies for dyninst >= 10.x.
* fix docker builds/remove extra builds/add ci builds
* preinstall vim in CI builder images
* simplify & streamline docker resources
* restore os-container-mapping.yaml file
The database and mutable_database fixtures were installing and uninstalling the same specs multiple times to ensure the database for tests has the correct state.
This commit optimizes the procedure by caching the state in an external directory, and copying it in instead of going through the installation or uninstallation again.
The database fixture is meant not to be modified by tests. This commit enforces this invariant by making the database read-only before starting the test.
* Added missing db markers to tests
* Added test for uninstall_by_spec
* `database` fixture now returns a read-only database
* Tests that modify the DB now use `mutable_database` fixture
ibm-java: fix the URL to only use a machine type that ibm-java
supports so that 'spack info' doesn't report a non-existent URL on
platforms that it doesn't support.
Summary:
- Allow multiple definitions of compiler in compilers.yaml (use first instance)
- Still print debug messages when there are duplicates, to assist users in finding this issue.
Merging configs from different scopes can result in multiple compiler being present in the same configuration list. Instead of raising when there are duplicates, take the one with highest precedence.
Print a debug message instead of raising, so that we can still diagnose this. We don't have a good way of warning the user about inconsistent configuration *in the same file* -- we'd need to dig into YAML file/line info for that.
- [x] Add shell tests to ensure that `spack env activate`, `spack env
deactivate`, and `despacktivate` continue to work.
- [x] Also ensure that activate and deactivate both work with `set -u`
* extends mkdirs with permissions for intermediate folders
Does not use os.makedirs mode parameter because its behavior is changed
with Python 3.7 (it ignores it for intermediate dirs), and moreover it
was not possible to set different modes for newly-created folders
and leaf folder.
reference:
- https://bugs.python.org/issue19930
- https://docs.python.org/3.7/library/os.html#os.makedirs
* comment mkdirp step easing code understanding
* revert mkdir to default for package metapath
since metapath is nested in package folder, there is no need
to specify permissions for intermediate folders because the prefix
already exists.
* comment create_install_directory package modes
Bug relates to the interplay between:
1. random dict orders in python 3.5
2. bugfix in initial implementation of stacks for `_concretize_dependencies`
when `self._dependencies` is empty
3. bug in coconcretization algorithm computation of split specs
Result was transient hang in coconcretization.
Fixed#3 (bug in coconcretization) to resolve.
- remove redundant code in Environment.__init__
- use socket.gethostname() instead of which('hostname')
- refactor updating SpecList references
- refactor 'specs' literals to a single variable for default list name
- use six.string_types for python 2/3 compatibility
* Fix scikit-learn build with Apple clang
* Update compiler unit tests
* Fix unit tests
* Fix OpenMP dep
* recursive=True no longer necessary
* Add myself as a maintainer
* Specify which versions require OpenMP
* Typo fixes
* LLVM-OpenMP is the correct dependency, no OMPT
* Flake8 fix
* Undo Apple Clang OpenMP flag change, split into a separate PR
Later versions of libjpeg-turbo build with CMake; to build with
user-specified cflags, the user must supply these to CMake as
-DCMAKE_C_FLAGS (Spack's typical approach of injecting these flags
into the compiler wrapper invocation is insufficient in this case).
Currently libjpeg-turbo cannot be implemented as a CMakePackage
(and thereby take advantage of the flag_handler implementation it
provides) because not all versions of libjpeg-turbo use CMake, so
this adds a custom implementation of flag_handler and
flags_to_build_system_args to libjpeg-turbo.
