Adapt tcl modulefile template to call "module load" on autoload
dependency without testing if this dependency is already loaded or not.
The is-loaded test is not necessary, as module commands know how to cope
with an already loaded module. With environment-modules 4.2+ (released
in 2018) it is also important to have this "module load" command even if
dependency is already loaded in order to record that the modulefile
declares such dependency. This is important if you want to keep a
consistent environment when a dependent module is unloaded.
The "Autoloading" verbose message is also removed as recent module
commands will report such information to the user (depending on the
verbosity configured for the module command).
Such change has been test successfully with Modules 3.2 (EL7), 4.5 (EL8)
and 5.2 (latest) and also with Lmod 7 and 8 (as it is mentionned in
Spack docs that Lmod can be used along with tcl modules). Dependencies
are correctly loaded or unloaded, whether they are loaded/unloaded or
not.
This change fixes Tcl quoting issue introduced in #32853.
Fixes#19155.
In the Windows filesystem logic for creating a symlink, we intend to
fall back to a copy when the symlink cannot be created (for some
configuration settings on Windows it is not possible for the user
to create a symlink). It turns out we were overly-broad in which
exceptions lead to this fallback, and the subsequent copy would
also fail: at least one case where this occurred is when we
attempted to create a symlink that already existed.
The updated logic expressly avoids falling back to a copy when the
file/symlink already exists.
* ASP-based solver: use satisfies instead of intersects
They are semantically equivalent for concrete versions,
but the GitVersion.intersects implementation is buggy
* Mitigation for git version bug
fixes#36134
This commit works around the issue in #36134, by using
GitVersion.satisfies instead of GitVersion.intersects
There are still underlying issues when trying to infer the
"reference version" when no explicit one is given, but:
1. They are not reproducible with our synthetic repo
2. They occur only when the `git.<xxx>` form of Git version
is used
Here we just work around the user facing issue and ensure
the tests are correct with our synthetic repository.
This PR does 2 unrelated things:
1. It changes the encoding of the compilers
2. It tweaks the heuristic for the solves in a0d88179074f81d13a3fad629a43d86334e7f982
Both were initially motivated by trying to get a performance gain but, while 2 showed significant speed-ups[^1], 1 instead didn't. I kept it anyhow, since I think the code related to compilers is more consolidated with the new encoding and we might get some performance improvement out of it if we can base our errors on the `node_compiler(Package, CompilerID)` atoms instead of `attrs`.
[^1]: In general the changes in the heuristic brought a ~10% speed-up on the tests I did. I'll post detailed results below.
Add a warning about compilers.yaml that is triggered if there are multiple compilers with the same spec, os and
target (since they can't be selected by users with the spec syntax only).
Since GPG clear-sign cannot deal with lines longer than 19995 characters
and doesn't even error but simply truncates those linese (don't ask me
why...), we have to be careful not to hit that line limit when reducing
the filesize.
So, instead this PR sets the indent level to 0 and drops the whitespace
after `: `, which still reduces file size by 50% or so.
Adds builders appropriate for building these packages on Windows.
It is intended that builds on other platforms are unaffected (e.g.
they build with Autotools as before on Linux).
... and use colors in disambiguate message for clarity.
This commit avoids the loop:
```
for root in roots:
for dep in deps(root):
...
```
instead it ensures each node is visited once and only once.
Also adds a small optimization when searching for concrete specs, since
we can assume uniqueness of dag hash, so it's fine to early exit.
This adds a new mode for `concretizer:reuse` called `dependencies`,
which only reuses dependencies. Currently, `spack install foo` will
reuse older versions of `foo`, which might be surprising to users.
* CI configuration boilerplate reduction and refactor
Configuration:
- New notation for list concatenation (prepend/append)
- New notation for string concatenation (prepend/append)
- Break out configuration files for: ci.yaml, cdash.yaml, view.yaml
- Spack CI section refactored to improve self-consistency and
composability
- Scripts are now lists of lists and/or lists of strings
- Job attributes are now listed under precedence ordered list that are
composed/merged using Spack config merge rules.
- "service-jobs" are identified explicitly rather than as a batch
CI:
- Consolidate common, platform, and architecture configurations for all CI stacks into composable configuration files
- Make padding consistent across all stacks (256)
- Merge all package -> runner mappings to be consistent across all
stacks
Unit Test:
- Refactor CI module unit-tests for refactor configuration
Docs:
- Add docs for new notations in configuration.rst
- Rewrite docs on CI pipelines to be consistent with refactored CI
workflow
* Script verbose environ, dev bootstrap
* Port #35409
When untouched spec pruning is enabled, specs possibly affected
by a change cannot be pruned from a pipeline.
Previously spack looked at all specs matching changed package
names, and traversed dependents of each, all the way to the
environment root, to compute the set of environment specs
possibly affected by a change (and thus, not candidates for
pruning).
With this PR, when untouched spec pruning is enabled, a new
environment variable can control how far towards the root spack
traverses to compute the set of specs possibly affected by a
change. SPACK_UNTOUCHED_PRUNING_DEPENDENT_DEPTH can be set
to any numeric value before the "spack ci generate" command
is called to control this traversal depth parameter. Setting
it to "0" traverses only touched specs, setting it to "1"
traverses only touched specs and their direct dependents, and
so on. Omitting the variable results in the previous behavior
of traversing all the way to the root. Setting it to a negative
value means no traversal is done, and always yields an empty
set of possibly affected specs (which would result in the max
pruning possible).
Currently `spack buildcache create` creates compressed tarballs that
differ between each invocation, thanks to:
1. The gzip header containing mtime set to time.time()
2. The generated buildinfo file which has a different mtime every time.
To avoid this, you have to explicitly construct GZipFile yourself, since
the Python API doesn't expose the mtime arg, and we have to manually
create the tarinfo object for the buildinfo metadata file.
Normalize mode: regular files & hardlinks executable by user, dirs, symlinks: set 0o755 permissions in tarfile; other files use 0o644
This commit formalizes `satisfies(lhs, rhs, strict=True/False)`
and splits it into two functions: `satisfies(lhs, rhs)` and
`intersects(lhs, rhs)`.
- `satisfies(lhs, rhs)` means: all concrete specs matching the
left hand side also match the right hand side
- `intersects(lhs, rhs)` means: there exist concrete specs
matching both lhs and rhs.
`intersects` now has the property that it's commutative,
which previously was not guaranteed.
For abstract specs, `intersects(lhs, rhs)` implies that
`constrain(lhs, rhs)` works.
What's *not* done in this PR is ensuring that
`intersects(concrete, abstract)` returns false when the
abstract spec has additional properties not present in the
concrete spec, but `constrain(concrete, abstract)` will
raise an error.
To accomplish this, some semantics have changed, as well
as bugfixes to ArchSpec:
- GitVersion is now interpreted as a more constrained
version
- Compiler flags are interpreted as strings since their
order is important
- Abstract specs respect variant type (bool / multivalued)
Two fixes:
1. `-Wl,a,b,c,d` is a comma separated list of linker arguments, we
incorrectly assume key/value pairs, which runs into issues with for
example `-Wl,--enable-new-dtags,-rpath,/x`
2. `-Xlinker,xxx` is not a think, so it shouldn't be parsed.
Currently, if two compilers with the same spec differ on the flags, the concretizer will:
1. mix both sets of flags for the spec in the ASP program
2. error noting that the set of flags from the compiler (both of them) doesn't match the set from the lower priority compiler
This PR fixes both -- only flags from the highest priority compiler with a given spec are considered.
`mypy` only understands `sys.platform == "win32"`, not indirect assignments of that
value to things like `is_windows`. If we don't use the accepted platform checks, `mypy`
registers many Windows-only symbols as not present on Linux, when it should skip the
checks for platform-specific code.
Update `spack.util.environment` to remove legacy idioms.
* Remove kwargs from method signature and use a class for traces
* Uppercase a few global variables
* spack.util.environment: add type-hints
* Improve docstrings
* Fixed most style issues reported by pylint
a shared library /lib64/libcxi.so, which seems to also appear on other
non-slingshot systems. This patch also checks to make sure that there
is a Cray programming enviornment in /opt/cray/pe in addition to the
shared library.
if dump file existed it was not truncating the file, resulting in
a file with unaltered filesize, with the new content at the beginning,
"padded" with the tail of the old content, since the new content was
not enough to overwrite it.
`colify` is an old module in Spack that still uses `**kwargs` liberally.
We should be more explicit. Doing this eliminates the need for many
checks (can't pass the wrong arg if it isn't allowed) and makes the
function documentation more clear.
Fixes a bug introduced in 44ed0de8c0
where the push method of binary_distribution now takes named args
include_root and include_depedencies, to avoid the **kwarg hole.
But the call site wasn't update and we passed a dict of keys/values instead
of arguments, which resulted in a call like this:
```
push(include_root={"include_root": True, "include_dependencies": False})
```
This commit fixes that, and adds a test to see if we push the correct packages.
This error shows up a lot, typically it's harmless because an error
happened before the source build even started, in which case we don't
have build logs to copy. So, warn instead of error, cause it distracts
from the actual CI error.
Currently we attempt to setup the build environment even when
dependencies are not installed, which typically results in error while
searching for libraries or executables in a dependency's prefix.
With this change, we get a more user friendly error:
```
$ spack build-env perl
==> Error: Not all dependencies of perl are installed, cannot setup build environment:
- qpj6dw5 perl@5.36.0%apple-clang@14.0.0+cpanm+open+shared+threads build_system=generic arch=darwin-ventura-m1
- jq2plbe ^berkeley-db@18.1.40%apple-clang@14.0.0+cxx~docs+stl build_system=autotools patches=26090f4,b231fcc arch=darwin-ventura-m1
...
$ echo $?
1
```
* Allow users to specify root env dir
Environments managed by spack have some advantages over anonymous Environments
but they are tucked away inside spack's directory tree. This PR gives
users the ability to specify where the environments should live.
See #32823
This is also taken as an opportunity to ensure that all references are to "managed environments",
rather than "named environments". Prior to this PR some references to the latter persisted.
Co-authored-by: Tom Scogland <scogland1@llnl.gov>
Co-authored-by: Tamara Dahlgren <35777542+tldahlgren@users.noreply.github.com>
Co-authored-by: Gregory Becker <becker33@llnl.gov>
The call:
```
x.satisfies(y[, strict=False])
```
is commutative, and tests non-empty intersection, whereas:
```
x.satsifies(y, strict=True)
```
is not commutative, and tests set-inclusion.
There are 2 fast paths. When strict=False both self and other need to
be concrete, when strict=True we can optimize when other is concrete.
Spack generally ignores file-file projection clashes in environment
views, but would eventually error when linking the `.spack` directory
for two specs of the same package.
This leads to obscure errors where users have no clue what the issue is
and how to fix it. On top of that, the error comes very late, since it
happens when the .spack dir contents are linked (which happens after
everything else)
This PR improves that by doing a quick check ahead of time if clashes
are going to be anticipated (by simply checking for clashes in the
projection of each spec's .spack metadir). If there are clashes, a
human-readable error is thrown which shows two of the conflicting specs,
and tells users to user unify:true, view:false, or set up custom
projections.
The checksum exception was not detailed enough and not reraised when using cache only, resulting in useless error messages.
Now it dumps the file path, expected
hash, computed hash, and the downloaded file summary.
* Style: black 23, skip magic trailing commas
* isort should use same line length as black
* Fix unused import
* Update version of black used in CI
* Update new packages
* Update new packages
Specs that did not contribute any files to an env view caused a problem
where zip(specs, files grouped by prefix) got "out of sync", causing the
wrong merge map to be passed to a package's `add_files_to_view`, which
specifically caused an issue where *sometimes* bin/python ended up as a
symlink instead of a copy.
One such example is kokkos + kokkos-nvcc-wrapper, as the latter package
only provides the file bin/nvcc_wrapper, which is also added to view by
kokkos, causing kokkos-nvcc-wrapper to contribute 0 files.
The test feels a bit contrived, but it captures the problem... pkg a is
added first and has 0 files to contribute, pkg b adds a single file, and
we check if pkg b receives a merge map (and a does not).
`spack gc` removes build deps of explicitly installed specs, but somehow
if you take one of the specs that `spack gc` would remove, and feed it
to `spack uninstall /<hash>` by hash, it complains about all the
dependents that still rely on it.
This resolves the inconsistency by only following run/link type deps in
spack uninstall.
That way you can finally do `spack uninstall cmake` without having to
remove all packages built with cmake.
Default package requirements might contain
variants that are not defined in each package,
so we shouldn't verify them when emitting facts
for the ASP solver.
Account for group when enforcing requirements
packages:all : don't emit facts for requirement conditions
that can't apply to current spec
#35098 added the correct extraction of toolset version for the MSVC
compiler. This updates the associated method in MSBuilder to retrieve
the (now correct) property.
This PR enables the successful execution of the spack binary cache
tutorial on Windows. It assumes gnupg and file are available (they
can be installed with choco).
* Fix handling of args with quotes in spack.bat
* `file` utility can be installed on Windows (e.g. with choco): update
error message accordingly
At least with ZSH, prefix inspections containing `./bin` result in a
`$PREFIX/./bin` and result in strange `$PATH` handling.
I.e., `module load git` will prepend `/path/to/git/./bin`, `which git`
will find the right executable, but `git --version` will print the
system one. Normalize the relative path to avoid this behavior.
See also spack/spack#31867.
fixes#34879
This commit adds a new maintainer directive,
which by default extend the list of maintainers
for a given package.
The directive is backward compatible with the current
practice of having a "maintainers" list declared at
the class level.
Move the relocation of binary text in its own class
Drop threaded text replacement, since the current bottleneck
is decompression. It would be better to parallellize over packages,
instead of over files per package.
A small improvement with separate classes for text replacement is that we
now compile the regex in the constructor; previously it was compiled per
binary to be relocated.
This commit makes explicit the format version of the spec file
we are reading from.
Before there were different functions capable of reading some
part of the spec file at multiple format versions. The decision
was implicit, since checks were based on the structure of the
JSON without ever checking a format version number.
The refactor makes also explicit which spec file format is used
by which database and lockfile format, since the information is
stored in global mappings.
To ensure we don't change the hash of old specs, JSON representations
of specs have been added as data. A unit tests checks that we read
the correct hash in, and that the hash stays the same when we
re-serialize the spec using the most recent format version.
Co-authored-by: Harmen Stoppels <me@harmenstoppels.nl>
Currently we print "sha256 checksum failed for [file]. Expected X but
got Y".
This PR extends that message with file size and contents info:
"... but got Y. File size = 123456 bytes. Contents = b'abc...def'"
That way we can immediately see if the file was downloaded only
partially, or if we downloaded a text page instead of a binary, etc.
Co-authored-by: Massimiliano Culpo <massimiliano.culpo@gmail.com>
When running unit-test the test/ci.py module is leaving
garbage (help.sh, test.sh files) in the current working
directory.
This commit changes the current working directory to a
temporary path before those files are created.
* environments: don't rewrite relative view path, expand path on cli ahead of time
Currently if you have a spack.yaml that specifies a view by relative
path, Spack expands it to an absolute path on `spack -e . install` and
persists that to disk.
This is rather annoying when you have a `spack.yaml` file inside a git
repo, cause you want to use relative paths to make it relocatable, but
you constantly have to undo the changes made to spack.yaml by Spack.
So, as an alternative:
1. Always stick to paths as they are provided in spack.yaml, never
replace them with a canonicalized version
2. Turn relative paths on the command line into absolute paths before
storing to spack.yaml. This way you can do `spack env create --dir
./env --with-view ./view` and both `./env` and `./view` are resolved
to the current working dir, as expected (not `./env/view`). This
corresponds to the old behavior of `spack env create`.
* create --with-view always takes a value
All packages with explicit Windows support can be found with
`spack list --tags=windows`.
This also removes the documentation which explicitly lists
supported packages on Windows (which is currently out of date and
is now unnecessary with the added tags).
Note that if a package does not appear in this list, it *may*
still build on Windows, but it likely means that no explicit
attempt has been made to support it.
Since SPACK_PACKAGE_IDS is now also "namespaced" with <prefix>, it makes
more sense to call the flag `--make-prefix` and alias the old flag
`--make-target-prefix` to it.
Normally when using external packages in concretization, Spack ignores
all dependencies of the external. #33777 updated this logic to attach
a Python Spec to external Python extensions (most py-* packages), but
as implemented there were a couple issues:
* this did not account for concretization groups and could generate
multiple different python specs for a single DAG
* in some cases this created a fake Python spec with insufficient
details to be usable (concretization/installation of the
extension would fail)
This PR addresses both of these issues:
* For environment specs that are concretized together, external python
extensions in those specs will all be assigned the same Python spec
* If Spack needs to "invent" a Python spec, then it will have all the
needed details (e.g. compiler/architecture)
With the new variable [prefix/]SPACK_PACKAGE_IDS you can conveniently execute
things after each successful install.
For example push just-built packages to a buildcache
```
SPACK ?= spack
export SPACK_COLOR = always
MAKEFLAGS += -Orecurse
MY_BUILDCACHE := $(CURDIR)/cache
.PHONY: all clean
all: push
ifeq (,$(filter clean,$(MAKECMDGOALS)))
include env.mk
endif
# the relevant part: push has *all* example/push/<pkg identifier> as prereqs
push: $(addprefix example/push/,$(example/SPACK_PACKAGE_IDS))
$(SPACK) -e . buildcache update-index --directory $(MY_BUILDCACHE)
$(info Pushed everything, yay!)
# and each example/push/<pkg identifier> has the install target as prereq,
# and the body can use target local $(HASH) and $(SPEC) variables to do
# things, such as pushing to a build cache
example/push/%: example/install/%
@mkdir -p $(dir $@)
$(SPACK) -e . buildcache create --allow-root --only=package --unsigned --directory $(MY_BUILDCACHE) /$(HASH) # push $(SPEC)
@touch $@
spack.lock: spack.yaml
$(SPACK) -e . concretize -f
env.mk: spack.lock
$(SPACK) -e . env depfile -o $@ --make-target-prefix example
clean:
rm -rf spack.lock env.mk example/
``
In the past we checked remote binary mirrors for existence of a spec
before attempting to download it. That changed to only checking local
copies of index.jsons (if available) to prioritize certain mirrors where
we expect to find a tarball. That was faster for CI since fetching
index.json and loading it just to order mirrors takes more time than
just attempting to fetch tarballs -- and also if we have a direct hit
there's no point to look at other mirrors.
Long story short: the info message only makes sense in the old version
of Spack, so it's better to remove it.
