- [x] Our wrapper error messages are sometimes hard to differentiate from other build
output, so prefix all errors from `die()` with '[spack cc] ERROR:'
- [x] The error we raise when running, say, `fc` without a Fortran compiler was not
clear enough. Clarify the message and the comment.
This converts everything in cc to POSIX sh, except for the parts currently
handled with bash arrays. Tests are still passing.
This version tries to be as straightforward as possible. Specifically, most conversions
are kept simple -- convert ifs to ifs, handle indirect expansion the way we do in
`setup-env.sh`, only mess with the logic in `cc`, and don't mess with the python code at
all.
The big refactor is for arrays. We can't rely on bash's nice arrays and be ignorant of
separators anymore. So:
1. To avoid complicated separator logic, there are three types of lists. They are:
* `$lsep`-separated lists, which end with `_list`. `lsep` is customizable, but we
picked `^G` (alarm bell) for `$lsep` because it's ASCII and it's unlikely that it
would actually appear in any arguments. If we need to get fancier (and I will lose
faith in the world if we do) then we could consider XON or XOFF.
* `:`-separated directory lists, which end with `_dirs`, `_DIRS`, `PATH`, or `PATHS`
* Whitespace-separated lists (like flags), which can have any other name.
Whitespace and colon-separated lists come with the territory with PATHs from env
vars and lists of flags. `^G` separated lists are what we use for most internal
variables, b/c it's more likely to work.
2. To avoid subshells, use a bunch of functions that do dirty `eval` stuff instead. This
adds 3 functions to deal with lists:
* `append LISTNAME ELEMENT [SEP]` will put `ELEMENT` at the end of the list called
`LISTNAME`. You can optionally say what separator you expect to use. Note that we
are taking advantage of everything being global and passing lists by name.
* `prepend LISTNAME ELEMENT [SEP]` like append, but puts `ELEMENT` at the start of
`LISTNAME`
* `extend LISTNAME1 LISTNAME2 [PREFIX]` appends everything in LISTNAME2 to
LISTNAME1, and optionally prepends `PREFIX` to every element (this is useful for
things like `-I`, `-isystem `, etc.
* `preextend LISTNAME1 LISTNAME2 [PREFIX]` prepends everything in LISTNAME2 to
LISTNAME1 in order, and optionally prepends `PREFIX` to every element.
The routines determine the separator for each argument by its name, so we don't have to
pass around separators everywhere. Amazingly, as long as you do not expand variables'
values within an `eval` environment, you can do all this and still preserve quoting.
When iterating over lists, the user of this API still has to set and unset `IFS`
properly.
We ended up having to ignore shellcheck SC2034 (unused variable), because using evals
all over the place means that shellcheck doesn't notice that our list variables are
actually used.
So far this is looking pretty good. I took the most complex unit test I could find
(which runs a sample link line) and ran the same command line 200 times in a shell
script. Times are roughly as follows:
For this invocation:
```console
$ bash -c 'time (for i in `seq 1 200`; do ~/test_cc.sh > /dev/null; done)'
```
I get the following performance numbers (the listed shells are what I put in `cc`'s
shebang):
**Original**
* Old version of `cc` with arrays and `bash v3.2.57` (macOS builtin): `4.462s` (`.022s` / call)
* Old version of `cc` with arrays and `bash v5.1.8` (Homebrew): `3.267s` (`.016s` / call)
**Using many subshells (#26408)**
* with `bash v3.2.57`: `25.302s` (`.127s` / call)
* with `bash v5.1.8`: `27.801s` (`.139s` / call)
* with `dash`: `15.302s` (`.077s` / call)
This version didn't seem to work with zsh.
**This PR (no subshells)**
* with `bash v3.2.57`: `4.973s` (`.025s` / call)
* with `bash v5.1.8`: `4.984s` (`.025s` / call)
* with `zsh`: `2.995s` (`.015s` / call)
* with `dash`: `1.890s` (`.0095s` / call)
Dash, with the new posix design, is easily the winner.
So there are several interesting things to note here:
1. Running the posix version in `bash` is slower than using `bash` arrays. That is to be
expected because it's doing a bunch of string processing where it likely did not have
to before, at least in `bash`.
2. `zsh`, at least on macOS, is significantly faster than the ancient `bash` they ship
with the system. Using `zsh` with the new version also makes the posix wrappers
faster than `develop`. So it's worth preferring `zsh` if we have it. I suppose we
should also try this with newer `bash` on Linux.
3. `bash v5.1.8` seems to be significantly faster than the old system `bash v3.2.57` for
arrays. For straight POSIX stuff, it's a little slower. It did not seem to matter
whether `--posix` was used.
4. `dash` is way faster than `bash` or `zsh`, so the real payoff just comes from being
able to use it. I am not sure if that is mostly startup time, but it's significant.
`dash` is ~2.4x faster than the original `bash` with arrays.
So, doing a lot of string stuff is slower than arrays, but converting to posix seems
worth it to be able to exploit `dash`.
- [x] Convert all but array-related portions to sh
- [x] Fix basic shellcheck issues.
- [x] Convert arrays to use a few convenience functions: `append` and `extend`
- [x] Get `cc` tests passing.
- [x] Add `cc` tests where needed passing.
- [x] Benchmarking.
Co-authored-by: Tom Scogland <scogland1@llnl.gov>
Co-authored-by: Danny McClanahan <1305167+cosmicexplorer@users.noreply.github.com>
When using modules for compiler (and/or external package), if a
package's `setup_[dependent_]build_environment` sets `PYTHONHOME`, it
can influence the python subprocess executed to gather module
information.
The error seen was:
```
json.decoder.JSONDecodeError: Expecting value: line 1 column 1 (char 0)
```
But the actual hidden error happened in the `python -c 'import
json...'` subprocess, which made it return an empty string as json:
```
ModuleNotFoundError: No module named 'encodings'
```
This fix uses `python -E` to ignore `PYTHONHOME` and
`PYTHONPATH`. Should be safe here because the python subprocess code
only use packages built-in python.
The python subprocess in `environment.py` was also patched to be safe
and consistent.
* Remove redundant preserve environment code in build environment
* Remove fix for a bug in a module
See https://github.com/spack/spack/issues/3153#issuecomment-280460041,
this shouldn't be part of core spack.
* Don't module unload cray-libsci on all platforms
Spack has logic to preserve an installation prefix when it is being
overwritten: if the new install fails, the old files are restored.
This PR adds error handling for when this backup restoration fails
(i.e. the new install fails, and then some unexpected error prevents
restoration from the backup).
* Remove vestigial code to be compatible with Spack v0.9.X
* ArchSpec: reworked __repr__ to be more adherent to common Python idioms
* ArchSpec: simplified __init__.py and copy()
The logic to perform detection of already installed
packages has been extracted from cmd/external.py
and put into the spack.detection package.
In this way it can be reused programmatically for
other purposes, like bootstrapping.
The new implementation accounts for cases where the
executables are placed in a subdirectory within <prefix>/bin
* Use gnuconfig package for config file replacement
Currently the autotools build system tries to pick up config.sub and
config.guess files from the system (in /usr/share) on arm and power.
This is introduces an implicit system dependency which we can avoid by
distributing config.guess and config.sub files in a separate package,
such as the new `gnuconfig` package which is very lightweight/text only
(unlike automake where we previously pulled these files from as a
backup). This PR adds `gnuconfig` as an unconditional build dependency
for arm and power archs.
In case the user needs a system version of config.sub and config.guess,
they are free to mark `gnuconfig` as an external package with the prefix
pointing to the directory containing the config files:
```yaml
gnuconfig:
externals:
- spec: gnuconfig@master
prefix: /tmp/tmp.ooBlkyAKdw/lol
buildable: false
```
Apart from that, this PR gives some better instructions for users when
replacing config files goes wrong.
* Mock needs this package too now, because autotools adds a depends_on
* Add documentation
* Make patch_config_files a prop, fix the docs, add integrations tests
* Make macOS happy
- Match failed autotest tests show the word "FAILED" near the end
- Match "FAIL: ", "FATAL: ", "failed ", "Failed test" of other suites
- autotest " ok"$ means the test passed, independend of text before.
- autoconf messages showing missing tools are fatal later, show them.
* autotoolspackage.rst: No depends_on('m4') with depends_on('autoconf')
- Remove `m4` from the example depends_on() lines for the autoreconf phase.
- Change the branch used as example from develop to master as it is
far more common in the packages of spack's builtin repo.
- Fix the wrong info that libtoolize and aclocal are run explicitly
in the autoreconf phase by default. autoreconf calls these internally
as needed, thus autotools.py also does not call them directly.
- Add that autoreconf() also adds -I<aclocal-prefix>/share/aclocal.
- Add an example how to set autoreconf_extra_args.
- Add an example of a custom autoreconf phase for running autogen.sh.
Co-authored-by: Harmen Stoppels <harmenstoppels@gmail.com>
This commit shows a template for cut-and-paste into the package to fix it:
```py
==> fast-global-file-status: Executing phase: 'autoreconf'
==> Error: RuntimeError: Cannot generate configure: missing dependencies autoconf, automake, libtool.
Please add the following lines to the package:
depends_on('autoconf', type='build', when='@master')
depends_on('automake', type='build', when='@master')
depends_on('libtool', type='build', when='@master')
Update the version (when='@master') as needed.
```
Co-authored-by: Harmen Stoppels <harmenstoppels@gmail.com>
clean_environment(): Unset three more environment variables:
MAKEFLAGS: Affects make, can eg indirectly inhibit enabling parallel build
DISPLAY: Tests of GUI widget libraries might try to connect to an X server
TERM: Could make testsuites attempt to color their output
fixes#25992
Currently the bootstrapping process may need a compiler.
When bootstrapping from sources the need is obvious, while
when bootstrapping from binaries it's currently needed in
case patchelf is not on the system (since it will be then
bootstrapped from sources).
Before this PR we were searching for compilers as the
first operation, in case they were not declared in
the configuration. This fails in case we start
bootstrapping from within an environment.
The fix is to defer the search until we have swapped
configuration.
While debugging #24508, I noticed that we call `basename` in `cc`. The
same can be achieved by using Bash's parameter expansion, saving one
external process per call.
Parameter expansion cannot replace basename for directories in some
cases, but is guaranteed to work for executables.
Git 2.24 introduced a feature flag for repositories with many files, see:
https://github.blog/2019-11-03-highlights-from-git-2-24/#feature-macros
Since Spack's Git repository contains roughly 8,500 files, it can be
worthwhile to enable this, especially on slow file systems such as NFS:
```
$ hyperfine --warmup 3 'cd spack-default; git status' 'cd spack-manyfiles; git status'
Benchmark #1: cd spack-default; git status
Time (mean ± σ): 3.388 s ± 0.095 s [User: 256.2 ms, System: 625.8 ms]
Range (min … max): 3.168 s … 3.535 s 10 runs
Benchmark #2: cd spack-manyfiles; git status
Time (mean ± σ): 168.7 ms ± 10.9 ms [User: 98.6 ms, System: 126.1 ms]
Range (min … max): 144.8 ms … 188.0 ms 19 runs
Summary
'cd spack-manyfiles; git status' ran
20.09 ± 1.42 times faster than 'cd spack-default; git status'
```
Modifications:
- [x] Change `defaults/config.yaml`
- [x] Add a fix for bootstrapping patchelf from sources if `compilers.yaml` is empty
- [x] Make `SPACK_TEST_SOLVER=clingo` the default for unit-tests
- [x] Fix package failures in the e4s pipeline
Caveats:
1. CentOS 6 still uses the original concretizer as it can't connect to the buildcache due to issues with `ssl` (bootstrapping from sources requires a C++14 capable compiler)
1. I had to update the image tag for GitlabCI in e699f14.
1. libtool v2.4.2 has been deprecated and other packages received some update
This will allow a user to (from anywhere a Spec is parsed including both name and version) refer to a git commit in lieu of
a package version, and be able to make comparisons with releases in the history based on commits (or with other commits). We do this by way of:
- Adding a property, is_commit, to a version, meaning I can always check if a version is a commit and then change some action.
- Adding an attribute to the Version object which can lookup commits from a git repo and find the last known version before that commit, and the distance
- Construct new Version comparators, which are tuples. For normal versions, they are unchanged. For commits with a previous version x.y.z, d commits away, the comparator is (x, y, z, '', d). For commits with no previous version, the comparator is ('', d) where d is the distance from the first commit in the repo.
- Metadata on git commits is cached in the misc_cache, for quick lookup later.
- Git repos are cached as bare repos in `~/.spack/git_repos`
- In both caches, git repo urls are turned into file paths within the cache
If a commit cannot be found in the cached git repo, we fetch from the repo. If a commit is found in the cached metadata, we do not recompare to newly downloaded tags (assuming repo structure does not change). The cached metadata may be thrown out by using the `spack clean -m` option if you know the repo structure has changed in a way that invalidates existing entries. Future work will include automatic updates.
# Finding previous versions
Spack will search the repo for any tags that match the string of a version given by the `version` directive. Spack will also search for any tags that match `v + string` for any version string. Beyond that, Spack will search for tags that match a SEMVER regex (i.e., tags of the form x.y.z) and interpret those tags as valid versions as well. Future work will increase the breadth of tags understood by Spack
For each tag, Spack queries git to determine whether the tag is an ancestor of the commit in question or not. Spack then sorts the tags that are ancestors of the commit by commit-distance in the repo, and takes the nearest ancestor. The version represented by that tag is listed as the previous version for the commit.
Not all commits will find a previous version, depending on the package workflow. Future work may enable more tangential relationships between commits and versions to be discovered, but many commits in real world git repos require human knowledge to associate with a most recent previous version. Future work will also allow packages to specify commit/tag/version relationships manually for such situations.
# Version comparisons.
The empty string is a valid component of a Spack version tuple, and is in fact the lowest-valued component. It cannot be generated as part of any valid version. These two characteristics make it perfect for delineating previous versions from distances. For any version x.y.z, (x, y, z, '', _) will be less than any "real" version beginning x.y.z. This ensures that no distance from a release will cause the commit to be interpreted as "greater than" a version which is not an ancestor of it.
Signed-off-by: vsoch <vsoch@users.noreply.github.com>
Co-authored-by: vsoch <vsoch@users.noreply.github.com>
Co-authored-by: Gregory Becker <becker33@llnl.gov>
Co-authored-by: Todd Gamblin <tgamblin@llnl.gov>
This PR coincides with tiny changes to spack to support spack monitor using the new spec
the corresponding spack monitor PR is at https://github.com/spack/spack-monitor/pull/31.
Since there are no changes to the database we can actually update the current server
fairly easily, so either someone can test locally or we can just update and then
test from that (and update as needed).
Signed-off-by: vsoch <vsoch@users.noreply.github.com>
Co-authored-by: vsoch <vsoch@users.noreply.github.com>
#22845 revealed a long-standing bug that had never been triggered before, because the
hashing algorithm had been stable for multiple years while the bug was in production. The
bug was that when reading a concretized environment, Spack did not properly read in the
build hashes associated with the specs in the environment. Those hashes were recomputed
(and as long as we didn't change the algorithm, were recomputed identically). Spack's
policy, though, is never to recompute a hash. Once something is installed, we respect its
metadata hash forever -- even if internally Spack changes the hashing method. Put
differently, once something is concretized, it has a concrete hash, and that's it -- forever.
When we changed the hashing algorithm for performance in #22845 we exposed the bug.
This PR fixes the bug at its source, but properly reading in the cached build hash attributes
associated with the specs. I've also renamed some variables in the Environment class
methods to make a mistake of this sort more difficult to make in the future.
* ensure environment build hashes are never recomputed
* add comment clarifying reattachment of env build hashes
* bump lockfile version and include specfile version in env meta
* Fix unit-test for v1 to v2 conversion
Co-authored-by: Massimiliano Culpo <massimiliano.culpo@gmail.com>
* Refactor platform etc. to avoid circular dependencies
All the base classes in spack.architecture have been
moved to the corresponding specialized subpackages,
e.g. Platform is now defined within spack.platforms.
This resolves a circular dependency where spack.architecture
was both:
- Defining the base classes for spack.platforms, etc.
- Collecting derived classes from spack.platforms, etc.
Now it dopes only the latter.
* Move a few platform related functions to "spack.platforms"
* Removed spack.architecture.sys_type()
* Fixup for docs
* Rename Python modules according to review