This changes the hash algorithm so that it does much less object
allocation and copying, and so that it is correct.
The old version of `_cmp_key()` would call `sorted_deps`, which would
call `flat_dependencies` to get a list of dependencies so that it
could sort them in alphabetical order. This isn't necessary in the
`_cmp_key()`, and in fact we want more DAG structure than that to be
included in the `_cmp_key()`.
The new version constructs a tuple without copying the Spec DAG, and
the tuple contains hashes of sub-DAGs that are computed recursively
in-place. This is way faster than the previous algorithm and reduces
the numebr of copies significantly. It is also a correct DAG hash.
Example timing and copy counts for the different hashing algorithms
we've tried:
Original (wrong) Spec hash:
```
106,170 copies
real 0m5.024s
user 0m4.949s
sys 0m0.104s
```
Spec hash using YAML `dag_hash()`:
```
3,794 copies
real 0m5.024s
user 0m4.949s
sys 0m0.104s
New no-copy, no-YAML hash:
```
3,594 copies
real 0m2.543s
user 0m2.435s
sys 0m0.104s
```
So now we have a hash that is correct AND faster.
The remaining ~3k copies happen mostly during concretization, and as
all packages are initially loaded. I believe this is because Spack
currently has to load all packages to figure out virtual dependency
information; it could also be becasue there ar a lot of lookups of
partial specs in concretize. I can investigate this further.
- _cross_provider_maps() had suffered some bit rot (map returned was
ill-formed but still worked for cases with one vdep)
- ProviderIndex.satisfies() was only checking whether the result map
was non-empty. It should check whether all common vdeps are *in*
the result map, as that indicates there is *some* way to satisfy
*all* of them. We were checking whether there was some way to
satisfy *any one* of them, which is wrong.
- Above would cause a problem when there is more than one vdep provider.
- Added test that covers this case.
- Added `constrained()` method to Spec. Analogous to `normalized()`:
`constrain():constrained() :: normalize():normalized()`
Added test for, e.g.:
import spack.pkg.builtin.mock.mpich
import spack.pkg.builtin.mock.mpich as mpich
from spack.pkg.builtin.mock.mpich import Mpich
Among others. These ensure that direct package imports work so that
packages can be extended.
Package repositories now look like this:
top-level-dir/
repo.yaml
packages/
libelf/
package.py
mpich/
package.py
...
This leaves room at the top level for additional metadata, source,
per-repo configs, indexes, etc., and it makes it easy to see that
something is a spack repo (just look for repo.yaml and packages).
unit tests, so tracking tests with sets wouldn't work unless I extracted the
details relevant to the particular test. For now a simple count will work so
using a set was unnecessary anyways.
1. Adding a plugin to keep track of the total number of tests run as well as the
number of tests with failures/errors.
2. Some nose plugins (including xunit which will be added in a future commit)
assign stdout to a stream object that does not have a .fileno attribute.
spack.util.executable.Executable now avoids passing stdout to subprocess (and
always uses subprocess.PIPE)
TODO:
1. Still need to figure out how to activate the plugin (as of now it is
being ignored by nose). Newer versions of nose appear to make this simpler
(e.g. the "addplugins" argument to nose.run)
2. Need to include new version of nose in order to use xunit
python install and lib/spack/external have the same library installed. This
requires modifying the names of some modules in lib/spack/external in cases
where both the system python and backported features of future python versions
(i.e. after 2.6) are used (previously distinguished by "from external import X"
and "import X").
- `remove_prefix` was modified to remove from the DB, but the package
may not have been added to the DB yet when `remove_prefix` is called
from `cleanup`.
- Made `remove_prefix` a pure utility function (it just removes the prefix)
- Added `installed_db.remove()` call only after the `remove_prefix` in
`uninstall`.
- Code simplification ignored case where exception was raised.
- If LockError was raised, read and write counts were incremented erroneously.
- updated lock test.
output path is specified. Test output files are placed in this directory.
Furthermore the filenames now have the prefix "test" (but otherwise are the
string representation of the spec ID as before).
- llnl.util.lock now uses fcntl.lockf instead of flock
- purported to have more NFS compatibility.
- Added an extensive test case for locks.
- tests acquiring, releasing, upgrading, timeouts, shared, & exclusive cases.
convert individual bytes, so therefore incorrectly converts utf-8 encoded
characters. Decoding byte strings to unicode objects results in correct HTML
number encodings.
2. The parent of a failed child process in build_environment.fork no longer
calls sys.exit - instead it raises a CommandError (from [1])
3. test-install command now attempts to install all packages even if one fails
command by having it consume exactly 1 positional argument (i.e. by removing
"nargs=argparse.REMAINDER") but this does not work when configuring dependencies
of a top-level package (which show up as additional positional args). Instead
now there is an explicit check to ensure there is only 1 top-level package.
argument). The default path is [package id].xml in the CWD where test-install
is called from.
2. Fixed a bug with package.build_log_path (which was added in this branch).
3. keep_stage for package.do_install is now set. This allows uninstalling and
reinstalling packages without (re) downloading them.
Also even though I calculated which installs are new (e.g. vs. packages that
have already been installed by a previous command) I forgot to make use of that
in create_test_output (so I was always generating test output even if a package
had been installed before running the test-install command).
Note to avoid confusion: the 'handled' variable (removed in this commit) did not
serve the same purpose as 'newInstalls': it was originally required because the
recursive approach would visit the same dependency twice if more than one
package depended on it.
packages are specified and a prior one fails, it will prevent any of the others
from succeeding (and generating test output) even if they don't share
dependencies.
the original intent was to generate output as if each package was a unit test,
but I noticed that I was only generating test output for top-level packages.
Still need to add output formatting (in a commonly parse-able format like Junit
or TAP). May want to adjust how the build log is accessed in case of a build
failure.
