* sbang pushed back to callers;
star moved to util.lang
* updated unit test
* sbang test moved; local tests pass
Co-authored-by: Nathan Hanford <hanford1@llnl.gov>
fixes#20736
Before this one line fix we were erroneously deducing
that dependency conditions hold even if a package
was external.
This may result in answer sets that contain imposed
conditions on a node without the node being present
in the DAG, hence #20736.
At some point in the past, the skip_patch argument was removed
from the call to package.do_install() this broke the --skip-patch
flag on the dev-build command.
fixes#20679
In this refactor we have a single cardinality rule on the
provider, which triggers a rule transforming a dependency
on a virtual package into a dependency on the provider of
the virtual.
This adds a -i option to "spack python" which allows use of the
IPython interpreter; it can be used with "spack python -i ipython".
This assumes it is available in the Python instance used to run
Spack (i.e. that you can "import IPython").
Every other predicate in the concretizer uses a `_set` suffix to
implement user- or package-supplied settings, but compiler settings use a
`_hard` suffix for this. There's no difference in how they're used, so
make the names the same.
- [x] change `node_compiler_hard` to `node_compiler_set`
- [x] change `node_compiler_version_hard` to `node_compiler_version_set`
Previously, the concretizer handled version constraints by comparing all
pairs of constraints and ensuring they satisfied each other. This led to
INCONSISTENT ressults from clingo, due to ambiguous semantics like:
version_constraint_satisfies("mpi", ":1", ":3")
version_constraint_satisfies("mpi", ":3", ":1")
To get around this, we introduce possible (fake) versions for virtuals,
based on their constraints. Essentially, we add any Versions,
VersionRange endpoints, and all such Versions and endpoints from
VersionLists to the constraint. Virtuals will have one of these synthetic
versions "picked" by the solver. This also allows us to remove a special
case from handling of `version_satisfies/3` -- virtuals now work just
like regular packages.
This converts the virtual handling in the new concretizer from
already-ground rules to facts. This is the last thing that needs to be
refactored, and it converts the entire concretizer to just use facts.
The previous way of handling virtuals hinged on rules involving
`single_provider_for` facts that were tied to the virtual and a version
range. The new method uses the condition pattern we've been using for
dependencies, externals, and conflicts.
To handle virtuals as conditions, we impose constraints on "fake" virtual
specs in the logic program. i.e., `version_satisfies("mpi", "2.0:",
"2.0")` is legal whereas before we wouldn't have seen something like
this. Currently, constriants are only handled on versions -- we don't
handle variants or anything else yet, but they key change here is that we
*could*. For a long time, virtual handling in Spack has only dealt with
versions, and we'd like to be able to handle variants as well. We could
easily add an integrity constraint to handle variants like the one we use
for versions.
One issue with the implementation here is that virtual packages don't
actually declare possible versions like regular packages do. To get
around that, we implement an integrity constraint like this:
:- virtual_node(Virtual),
version_satisfies(Virtual, V1), version_satisfies(Virtual, V2),
not version_constraint_satisfies(Virtual, V1, V2).
This requires us to compare every version constraint to every other, both
in program generation and within the concretizer -- so there's a
potentially quadratic evaluation time on virtual constraints because we
don't have a real version to "anchor" things to. We just say that all the
constraints need to agree for the virtual constraint to hold.
We can investigate adding synthetic versions for virtuals in the future,
to speed this up.
This code in `SpecBuilder.build_specs()` introduced in #20203, can loop
seemingly interminably for very large specs:
```python
set([spec.root for spec in self._specs.values()])
```
It's deceptive, because it seems like there must be an issue with
`spec.root`, but that works fine. It's building the set afterwards that
takes forever, at least on `r-rminer`. Currently if you try running
`spack solve r-rminer`, it loops infinitely and spins up your fan.
The issue (I think) is that the spec is not yet complete when this is
run, and something is going wrong when constructing and comparing so many
values produced by `_cmp_key()`. We can investigate the efficiency of
`_cmp_key()` separately, but for now, the fix is:
```python
roots = [spec.root for spec in self._specs.values()]
roots = dict((id(r), r) for r in roots)
```
We know the specs in `self._specs` are distinct (they just came out of
the solver), so we can just use their `id()` to unique them here. This
gets rid of the infinite loop.
- [x] add `concretize.lp`, `spack.yaml`, etc. to licensed files
- [x] update all licensed files to say 2013-2021 using
`spack license update-copyright-year`
- [x] appease mypy with some additions to package.py that needed
for oneapi.py
This adds a new subcommand to `spack license` that automatically updates
the copyright year in files that should have a license header.
- [x] add `spack license update-copyright-year` command
- [x] add test
GCC looks for included files based on several env vars.
Remove C_INCLUDE_PATH, CPLUS_INCLUDE_PATH, and OBJC_INCLUDE_PATH
from the build environment to ensure it's clean and prevent
accidental clobbering.
Environment yaml files should not have default values written to them.
To accomplish this, we change the validator to not add the default values to yaml. We rely on the code to set defaults for all values (and use defaulting getters like dict.get(key, default)).
Includes regression test.
This creates a set of packages which all use the same script to install
components of Intel oneAPI. This includes:
* An inheritable IntelOneApiPackage which knows how to invoke the
installation script based on which components are requested
* For components which include headers/libraries, an inheritable
IntelOneApiLibraryPackage is provided to locate them
* Individual packages for DAL, DNN, TBB, etc.
* A package for the Intel oneAPI compilers (icx/ifx). This also includes
icc/ifortran but these are not currently detected in this PR
I lost my mind a bit after getting the completion stuff working and
decided to get Mypy working for spack as well. This adds a
`.mypy.ini` that checks all of the spack and llnl modules, though
not yet packages, and fixes all of the identified missing types and
type issues for the spack library.
In addition to these changes, this includes:
* rename `spack flake8` to `spack style`
Aliases flake8 to style, and just runs flake8 as before, but with
a warning. The style command runs both `flake8` and `mypy`,
in sequence. Added --no-<tool> options to turn off one or the
other, they are on by default. Fixed two issues caught by the tools.
* stub typing module for python2.x
We don't support typing in Spack for python 2.x. To allow 2.x to
support `import typing` and `from typing import ...` without a
try/except dance to support old versions, this adds a stub module
*just* for python 2.x. Doing it this way means we can only reliably
use all type hints in python3.7+, and mypi.ini has been updated to
reflect that.
* add non-default black check to spack style
This is a first step to requiring black. It doesn't enforce it by
default, but it will check it if requested. Currently enforcing the
line length of 79 since that's what flake8 requires, but it's a bit odd
for a black formatted project to be quite that narrow. All settings are
in the style command since spack has no pyproject.toml and I don't
want to add one until more discussion happens. Also re-format
`style.py` since it no longer passed the black style check
with the new length.
* use style check in github action
Update the style and docs action to use `spack style`, adding in mypy
and black to the action even if it isn't running black right now.
We have to repeat all the spec attributes in a number of places in
`concretize.lp`, and Spack has a fair number of spec attributes. If we
instead add some rules up front that establish equivalencies like this:
```
node(Package) :- attr("node", Package).
attr("node", Package) :- node(Package).
version(Package, Version) :- attr("version", Package, Version).
attr("version", Package, Version) :- version(Package, Version).
```
We can rewrite most of the repetitive conditions with `attr` and repeat
only for each arity (there are only 3 arities for spec attributes so far)
as opposed to each spec attribute. This makes the logic easier to read
and the rules easier to follow.
Co-authored-by: Massimiliano Culpo <massimiliano.culpo@gmail.com>
This PR does three related things to try to improve developer tooling quality of life:
1. Adds new options to `.flake8` so it applies the rules of both `.flake8` and `.flake_package` based on paths in the repository.
2. Adds a re-factoring of the `spack flake8` logic into a flake8 plugin so using flake8 directly, or through editor or language server integration, only reports errors that `spack flake8` would.
3. Allows star import of `spack.pkgkit` in packages, since this is now the thing that needs to be imported for completion to work correctly in package files, it's nice to be able to do that.
I'm sorely tempted to sed over the whole repository and put `from spack.pkgkit import *` in every package, but at least being allowed to do it on a per-package basis helps.
As an example of what the result of this is:
```
~/Workspace/Projects/spack/spack develop* ⇣
❯ flake8 --format=pylint ./var/spack/repos/builtin/packages/kripke/package.py
./var/spack/repos/builtin/packages/kripke/package.py:6: [F403] 'from spack.pkgkit import *' used; unable to detect undefined names
./var/spack/repos/builtin/packages/kripke/package.py:25: [E501] line too long (88 > 79 characters)
~/Workspace/Projects/spack/spack refactor-flake8*
1 ❯ flake8 --format=spack ./var/spack/repos/builtin/packages/kripke/package.py
~/Workspace/Projects/spack/spack refactor-flake8*
❯ flake8 ./var/spack/repos/builtin/packages/kripke/package.py
```
* qa/flake8: update .flake8, spack formatter plugin
Adds:
* Modern flake8 settings for per-path/glob error ignores, allows
packages to use the same `.flake8` as the rest of spack
* A spack formatter plugin to flake8 that implements the behavior of
`spack flake8` for direct invocations. Makes integration with
developer tooling nicer, linting with flake8 reports only errors that
`spack flake8` would report. Using pyls and pyls-flake8, or any other
non-format-dependent flake8 integration, now works with spack's rules.
* qa/flake8: allow star import of spack.pkgkit
To get working completion of directives and spack components it's
necessary to import the contents of spack.pkgkit. At the moment doing
this makes flake8 displeased. For now, allow spack.pkgkit and spack
both, next step is to ban spack * and require spack.pkgkit *.
* first cut at refactoring spack flake8
This version still copies all of the files to be checked as befire, and
some other things that probably aren't necessary, but it relies on the
spack formatter plugin to implement the ignore logic.
* keep flake8 from rejecting itself
* remove separate packages flake8 config
* fix failures from too many files
I ran into this in the PR converting pkgkit to std. The solution in
that branch does not work in all cases as it turns out, and all the
workarounds I tried to use generated configs to get a single invocation
of flake8 with a filename optoion to work failed. It's an astonishingly
frustrating config option.
Regardless, this removes all temporary file creation from the command
and relies on the plugin instead. To work around the huge number of
files in spack and still allow the command to control what gets checked,
it scans files in batches of 100. This is a completely arbitrary number
but was chosen to be safely under common line-length limits. One
side-effect of this is that every 100 files the command will produce
output, rather than only at the end, which doesn't seem like a terrible
thing.
Continuing to convert everything in `asp.py` into facts, make the
generation of ground rules for conditional dependencies use facts, and
move the semantics into `concretize.lp`.
This is probably the most complex logic in Spack, as dependencies can be
conditional on anything, and we need conditional ASP rules to accumulate
and map all the dependency conditions to spec attributes.
The logic looks complicated, but essentially it accumulates any
constraints associated with particular conditions into a fact associated
with the condition by id. Then, if *any* condition id's fact is True, we
trigger the dependency.
This simplifies the way `declared_dependency()` works -- the dependency
is now declared regardless of whether it is conditional, and the
conditions are handled by `dependency_condition()` facts.
There are currently no places where we do not want to traverse
dependencies in `spec_clauses()`, so simplify the logic by consolidating
`spec_traverse_clauses()` with `spec_clauses()`.
`version_satisfies/2` and `node_compiler_version_satisfies/3` are
generated but need `#defined` directives to avoid " info: atom does not
occur in any rule head:" warnings.
This PR addresses a number of issues related to compiler bootstrapping.
Specifically:
1. Collect compilers to be bootstrapped while queueing in installer
Compiler tasks currently have an incomplete list in their task.dependents,
making those packages fail to install as they think they have not all their
dependencies installed. This PR collects the dependents and sets them on
compiler tasks.
2. allow boostrapped compilers to back off target
Bootstrapped compilers may be built with a compiler that doesn't support
the target used by the rest of the spec. Allow them to build with less
aggressive target optimization settings.
3. Support for target ranges
Backing off the target necessitates computing target ranges, so make Spack
handle those properly. Notably, this adds an intersection method for target
ranges and fixes the way ranges are satisfied and constrained on Spec objects.
This PR also:
- adds testing
- improves concretizer handling of target ranges
Co-authored-by: Harmen Stoppels <harmenstoppels@gmail.com>
Co-authored-by: Gregory Becker <becker33@llnl.gov>
Co-authored-by: Massimiliano Culpo <massimiliano.culpo@gmail.com>
Currently, version range constraints, compiler version range constraints,
and target range constraints are implemented by generating ground rules
from `asp.py`, via `one_of_iff()`. The rules look like this:
```
version_satisfies("python", "2.6:") :- 1 { version("python", "2.4"); ... } 1.
1 { version("python", "2.4"); ... } 1. :- version_satisfies("python", "2.6:").
```
So, `version_satisfies(Package, Constraint)` is true if and only if the
package is assigned a version that satisfies the constraint. We
precompute the set of known versions that satisfy the constraint, and
generate the rule in `SpackSolverSetup`.
We shouldn't need to generate already-ground rules for this. Rather, we
should leave it to the grounder to do the grounding, and generate facts
so that the constraint semantics can be defined in `concretize.lp`.
We can replace rules like the ones above with facts like this:
```
version_satisfies("python", "2.6:", "2.4")
```
And ground them in `concretize.lp` with rules like this:
```
1 { version(Package, Version) : version_satisfies(Package, Constraint, Version) } 1
:- version_satisfies(Package, Constraint).
version_satisfies(Package, Constraint)
:- version(Package, Version), version_satisfies(Package, Constraint, Version).
```
The top rule is the same as before. It makes conditional dependencies and
other places where version constraints are used work properly. Note that
we do not need the cardinality constraint for the second rule -- we
already have rules saying there can be only one version assigned to a
package, so we can just infer from `version/2` `version_satisfies/3`.
This form is also safe for grounding -- If we used the original form we'd
have unsafe variables like `Constraint` and `Package` -- the original
form only really worked when specified as ground to begin with.
- [x] use facts instead of generating rules for package version constraints
- [x] use facts instead of generating rules for compiler version constraints
- [x] use facts instead of generating rules for target range constraints
- [x] remove `one_of_iff()` and `iff()` as they're no longer needed
I was keeping the old `clingo` driver code around in case we had to run
using the command line tool instad of through the Python interface.
So far, the command line is faster than running through Python, but I'm
working on fixing that. I found that if I do this:
```python
control = clingo.Control()
control.load("concretize.lp")
control.load("hdf5.lp") # code from spack solve --show asp hdf5
control.load("display.lp")
control.ground([("base", [])])
control.solve(...)
```
It's just as fast as the command line tool. So we can always generate the
code and load it manually if we need to -- we don't need two drivers for
clingo. Given that the python interface is also the only way to get unsat
cores, I think we pretty much have to use it.
So, I'm removing the old command line driver and other unused code. We
can dig it up again from the history if it is needed.
This fixes a logging error observed on macOS 11.0.1 (Big Sur).
When performing a Spack install in debugging mode (e.g.
`spack -d install py-scipy`) Spack is supposed to write a log of
compiler wrapper command line invocations to the current working
directory.
Due to a regression error introduced by #18205, these files were
no-longer generated, and Spack was printing errors such as
"No such file or directory: None/." This is because the log file
directory gets set from `spack.main.spack_working_dir`, but that
variable is not set in the spawned process.
This PR ensures that the working directory (at the time of the
"spack install" invocation) is persisted to the subprocess.
Track all the variant values mentioned when emitting constraints, validate them
and emit a fact that allows them as possible values.
This modification ensures that open-ended variants (variants accepting any string
or any integer) are projected to the finite set of values that are relevant for this
concretization.
Other parts of the concretizer code build up lists of things we can't
know without traversing all specs and packages, and they output these
list at the very end.
The code for this for variant values from spec literals was intertwined
with the code for traversing the input specs. This only covers the input
specs and misses variant values that might come from directives in
packages.
- [x] move ad-hoc value handling code into spec_clauses so we do it in
one place for CLI and packages
- [x] move handling of `variant_possible_value`, etc. into
`concretize.lp`, where we can automatically infer variant existence
more concisely.
- [x] simplify/clarify some of the code for variants in `spec_clauses()`
* [cmd versions] add spack versions --new flag to only fetch new versions
format
[cmd versions] rename --latest to --newest and add --remote-only
[cmd versions] add tests for --remote-only and --new
format
[cmd versions] update shell tab completion
[cmd versions] remove test for --remote-only --new which gives empty output
[cmd versions] final rename
format
* add brillig mock package
* add test for spack versions --new
* [brillig] format
* [versions] increase test coverage
* Update lib/spack/spack/cmd/versions.py
Co-authored-by: Massimiliano Culpo <massimiliano.culpo@gmail.com>
* Update lib/spack/spack/cmd/versions.py
Co-authored-by: Massimiliano Culpo <massimiliano.culpo@gmail.com>
Co-authored-by: Massimiliano Culpo <massimiliano.culpo@gmail.com>
* allow install of build-deps from cache via --include-build-deps switch
* make clear that --include-build-deps is useful for CI pipeline troubleshooting
fixes#20055
Compiler with custom versions like gcc@foo are not currently
matched to the appropriate targets. This is because the
version of spec doesn't match the "real" version of the
compiler.
This PR replicates the strategy used in the original
concretizer to deal with that and tries to detect the real
version of compilers if the version in the spec returns no
results.
* AOCC-2.3.0 is now added to spack
Change-Id: I18fd9606e6fd9a288cc7dc6c6ead11ea17839a7c
* Added flag and version tests for AOCC-2.3.0
* Addressed review comments
Co-authored-by: vkallesh <Vijay-teekinavar.Kallesh@amd.com>
fixes#20040
Matching compilers among nodes has been prioritized
in #20020. Selection of default variants has been
tuned in #20182. With this setup there is no need
to have an ad-hoc rule for external packages. On
the contrary it should be removed to prefer having
default variant values over more external nodes in
the DAG.
refers #20040
Before this PR optimization rules would have selected default
providers at a higher priority than default variants. Here we
swap this priority and we consider variants that are forced by
any means (root spec or spec in depends_on clause) the same as
if they were with a default value.
This prevents the solver from avoiding expected configurations
just because they contain directives like:
depends_on('pkg+foo')
and `+foo` is not the default variant value for pkg.
fixes#19981
This commit adds support for target ranges in directives,
for instance:
conflicts('+foo', when='target=x86_64:,aarch64:')
If any target in a spec body is not a known target the
following clause will be emitted:
node_target_satisfies(Package, TargetConstraint)
when traversing the spec and a definition of
the clause will then be printed at the end similarly
to what is done for package and compiler versions.
fixes#20019
Before this modification having a newer version of a node came
at higher priority in the optimization than having matching
compilers. This could result in unexpected configurations for
packages with conflict directives on compilers of the type:
conflicts('%gcc@X.Y:', when='@:A.B')
where changing the compiler for just that node is preferred to
lower the node version to less than 'A.B'. Now the priority has
been switched so the solver will try to lower the version of the
nodes in question before changing their compiler.
refers #20079
Added docstrings to 'concretize' and 'concretized' to
document the format for tests.
Added tests for the activation of test dependencies.
refers #20040
This modification emits rules like:
provides_virtual("netlib-lapack","blas") :- variant_value("netlib-lapack","external-blas","False").
for packages that provide virtual dependencies conditionally instead
of a fact that doesn't account for the condition.
This PR fixes two problems with clang/llvm's version detection. clang's
version output looks like this:
```
clang version 11.0.0
Target: x86_64-unknown-linux-gnu
```
This caused clang's version to be misdetected as:
```
clang@11.0.0
Target:
```
This resulted in errors when trying to actually use it as a compiler.
When using `spack external find`, we couldn't determine the compiler
version, resulting in errors like this:
```
==> Warning: "llvm@11.0.0+clang+lld+lldb" has been detected on the system but will not be added to packages.yaml [reason=c compiler not found for llvm@11.0.0+clang+lld+lldb]
```
Changing the regex to only match until the end of the line fixes these
problems.
Fixes: #19473
This adds a new `mark` command that can be used to mark packages as either
explicitly or implicitly installed. Apart from fixing the package
database after installing a dependency manually, it can be used to
implement upgrade workflows as outlined in #13385.
The following commands demonstrate how the `mark` and `gc` commands can be
used to only keep the current version of a package installed:
```console
$ spack install pkgA
$ spack install pkgB
$ git pull # Imagine new versions for pkgA and/or pkgB are introduced
$ spack mark -i -a
$ spack install pkgA
$ spack install pkgB
$ spack gc
```
If there is no new version for a package, `install` will simply mark it as
explicitly installed and `gc` will not remove it.
Co-authored-by: Greg Becker <becker33@llnl.gov>
Users can add test() methods to their packages to run smoke tests on
installations with the new `spack test` command (the old `spack test` is
now `spack unit-test`). spack test is environment-aware, so you can
`spack install` an environment and then run `spack test run` to run smoke
tests on all of its packages. Historical test logs can be perused with
`spack test results`. Generic smoke tests for MPI implementations, C,
C++, and Fortran compilers as well as specific smoke tests for 18
packages.
Inside the test method, individual tests can be run separately (and
continue to run best-effort after a test failure) using the `run_test`
method. The `run_test` method encapsulates finding test executables,
running and checking return codes, checking output, and error handling.
This handles the following trickier aspects of testing with direct
support in Spack's package API:
- [x] Caching source or intermediate build files at build time for
use at test time.
- [x] Test dependencies,
- [x] packages that require a compiler for testing (such as library only
packages).
See the packaging guide for more details on using Spack testing support.
Included is support for package.py files for virtual packages. This does
not change the Spack interface, but is a major change in internals.
Co-authored-by: Tamara Dahlgren <dahlgren1@llnl.gov>
Co-authored-by: wspear <wjspear@gmail.com>
Co-authored-by: Adam J. Stewart <ajstewart426@gmail.com>
The deprecatedProperties custom validator now can accept a function
to compute a better error message.
Improve error/warning message for deprecated properties
As of #18260, `spack load` and `spack env activate` now use
`prefix_inspections` from the modules configuration to decide
how to modify environment variables.
This updates the modules configuration documentation to describe
how to update environment variables with the `prefix_inspections`
section. This also updates the `spack load` and environments
documentation to refer to the new `prefix_inspections` documentation.
`spack load` and `spack env activate` now use the prefix inspections
defined in `modules.yaml`. This allows users to customize/override
environment variable modifications if desired.
If no `prefix_inspections` configuration is present, Spack uses the
values in the default configuration.
This PR reworks a few attributes in the container subsection of
spack.yaml to permit the injection of custom base images when
generating containers with Spack. In more detail, users can still
specify the base operating system and Spack version they want to use:
spack:
container:
images:
os: ubuntu:18.04
spack: develop
in which case the generated recipe will use one of the Spack images
built on Docker Hub for the build stage and the base OS image in the
final stage. Alternatively, they can specify explicitly the two
base images:
spack:
container:
images:
build: spack/ubuntu-bionic:latest
final: ubuntu:18.04
and it will be up to them to ensure their consistency.
Additional changes:
* This commit adds documentation on the two approaches.
* Users can now specify OS packages to install (e.g. with apt or yum)
prior to the build (previously this was only available for the
finalized image).
* Handles to avoid an update of the available system packages have been
added to the configuration to facilitate the generation of recipes
permitting deterministic builds.
This commit address the case of concretizing a root spec with a
transitive conditional dependency on a virtual package, provided
by an external. Before these modifications default variant values
for the dependency bringing in the virtual package were not
respected, and the external package providing the virtual was added
to the DAG.
The issue stems from two facts:
- Selecting a provider has higher precedence than selecting default variants
- To ensure that an external is preferred, we used a negative weight
To solve it we shift all the providers weight so that:
- External providers have a weight of 0
- Non external provider have a weight of 10 or more
Using a weight of zero for external providers is such that having
an external provider, if present, or not having a provider at all
has the same effect on the higher priority minimization.
Also fixed a few minor bugs in concretize.lp, that were causing
spurious entries in the final answer set.
Cleaned concretize.lp from leftover rules.
If a the default of a multi-valued variant is set to
multiple values either in package.py or in packages.yaml
we need to ensure that all the values are present in the
concretized spec.
Since each default value has a weight of 0 and the
variant value is set implicitly by the concretizer
we need to add a rule to maximize on the number of
default values that are used.
This commit introduces a new rule:
real_node(Package) :- not external(Package), node(Package).
that permits to distinguish between an external node and a
real node that shouldn't trim dependency. It solves the
case of concretizing ninja with an external Python.
`node_compiler_hard()` means that something explicitly asked for a node's
compiler to be set -- i.e., it's not inherited, it's required. We're
generating this in spec_clauses even for specs in rule bodies, which
results in conditions like this for optional dependencies:
In py-torch/package.py:
depends_on('llvm-openmp', when='%apple-clang +openmp')
In the generated ASP:
declared_dependency("py-torch","llvm-openmp","build")
:- node("py-torch"),
variant_value("py-torch","openmp","True"),
node_compiler("py-torch","apple-clang"),
node_compiler_hard("py-torch","apple-clang"),
node_compiler_version_satisfies("py-torch","apple-clang",":").
The `node_compiler_hard` there means we would have to *explicitly* set
py-torch's compiler to trigger the llvm-openmp dependency, rather than
just letting it be set by preferences. This is wrong; the dependency
should be there regardless of how the compiler was set.
- [x] remove fn.node_compiler_hard() call from spec_clauses when
generating rule body clauses.
If the version list passed to one_of_iff is empty, it still generates a
rule like this:
node_compiler_version_satisfies("fujitsu-mpi", "arm", ":") :- 1 { } 1.
1 { } 1 :- node_compiler_version_satisfies("fujitsu-mpi", "arm", ":").
The cardinality rules on the right and left above are never
satisfiale, and these rules do nothing.
- [x] Skip generating any rules at all for empty version lists.
As reported, conflicts with compiler ranges were not treated
correctly. This commit adds tests to verify the expected behavior
for the new concretizer.
The new rules to enforce a correct behavior involve:
- Adding a rule to prefer the compiler selected for
the root package, if no other preference is set
- Give a strong negative weight to compiler preferences
expressed in packages.yaml
- Maximize on compiler AND compiler version match
Variant of this kind don't have a list of possible
values encoded in the ASP facts. Since all we have
is a validator the list of possible values just includes
just the default value and possibly the value passed
from packages.yaml or cli.
This is done after the builder has actually built
the specs, to respect the semantics use with the
old concretizer.
Later we could move this to the solver as
a multivalued variant.
This is done after the builder has actually built
the specs, to respect the semantics use with the
old concretizer.
A better approach is to substitute the spec
directly in concretization.
The "none" variant value cannot be combined with
other values.
The '*' wildcard matches anything, including "none".
It's thus relevant in queries, but disregarded in
concretization.
- The test on concretization of anonymous dependencies
has been fixed by raising the expected exception.
- The test on compiler bootstrap has been fixed by
updating the version of GCC used in the test.
Since gcc@2.0 does not support targets later than
x86_64, the new concretizer was looking for a
non-existing spec, i.e. it was correctly trying
to retrieve 'gcc target=x86_64' instead of
'gcc target=core2'.
- The test on gitlab CI needed an update of the target
This commit adds support for specifying rules in
packages.yaml that refer to virtual packages.
The approach is to normalize in memory each
configuration and turn it into an equivalent
configuration without rules on virtual. This
is possible if the set of packages to be handled
is considered fixed.
The weight of the target used in concretization is, in order:
1. A specific per package weight, if set in packages.yaml
2. Inherited from the parent, if possible
3. The default target weight (always set)
Generate facts on externals by inspecting
packages.yaml. Added rules in concretize.lp
Added extra logic so that external specs
disregard any conflict encoded in the
package.
In ASP this would be a simple addition to
an integrity constraint:
:- c1, c2, c3, not external(pkg)
Using the the Backend API from Python it
requires some scaffolding to obtain a default
negated statement.
Conflict rules from packages are added as integrity
constraints in the ASP formulation. Most of the code
to generate them has been reused from PyclingoDriver.rules
The new concretizer and the old concretizer solve constraints
in a different way. Here we ensure that a SpackError is raised,
instead of a specific error that made sense in the old concretizer
but probably not in the new.
Instead of python callbacks, use cardinality constraints for package
versions. This is slightly faster and has the advantage that it can be
written to an ASP program to be executed *outside* of Spack. We can use
this in the future to unify the pyclingo driver and the clingo text
driver.
This makes use of add_weight_rule() to implement cardinality constraints.
add_weight_rule() only has a lower bound parameter, but you can implement
a strict "exactly one of" constraint using it. In particular, wee want to
define:
1 {v1; v2; v3; ...} 1 :- version_satisfies(pkg, constraint).
version_satisfies(pkg, constraint) :- 1 {v1; v2; v3; ...} 1.
And we do that like this, for every version constraint:
atleast1(pkg, constr) :- 1 {version(pkg, v1); version(pkg, v2); ...}.
morethan1(pkg, constr) :- 2 {version(pkg, v1); version(pkg, v2); ...}.
version_satisfies(pkg, constr) :- atleast1, not morethan1(pkg, constr).
:- version_satisfies(pkg, constr), morethan1.
:- version_satisfies(pkg, constr), not atleast1.
v1, v2, v3, etc. are computed on the Python side by comparing every
possible package version with the constraint.
Computing things like this has the added advantage that if v1, v2, v3,
etc. comprise *all* possible versions of a package, we can just omit the
rules for the constraint under consideration. This happens pretty
frequently in the Spack mainline.
- [x] Solver now uses the Python interface to clingo
- [x] can extract unsatisfiable cores from problems when things go wrong
- [x] use Python callbacks for versions instead of choice rules (this may
ultimately hurt performance)
There are now three parts:
- `SpackSolverSetup`
- Spack-specific logic for generating constraints. Calls methods on
`AspTextGenerator` to set up the solver with a Spack problem. This
shouln't change much from solver backend to solver backend.
- ClingoDriver
- The solver driver provides methods for SolverSetup to generates an ASP
program, send it to `clingo` (run as an external tool), and parse the
output into function tuples suitable for `SpecBuilder`.
- The interface is generic and should not have to change much for a
driver for, say, the Clingo Python interface.
- SpecBuilder
- Builds Spack specs from function tuples parsed by the solver driver.
The original implementation was difficult to read, as it only had
single-letter variable names. This converts all of them to descriptive
names, e.g., P -> Package, V -> Virtual/Version/Variant, etc.
To handle unknown compilers propely in tests (and elsewhere), we need to
add unknown compilers from the spec to the list of possible compilers.
Rework how the compiler list is generated and includes compilers from
specs if the existence check is disabled.
Specs like hdf5 ^mpi were unsatisfiable because we added a requierment
for `node("mpi").`. This can't be resolved because "mpi" is not a
package.
- [x] Introduce `virtual_node()`, which says *some* provider must be in
the DAG.
This adds compiler flags to the ASP solve so that we can have conditions
based on them in the solve. But, it keeps order out of the solve to
avoid unneeded complexity and combinatorial explosions.
The solver determines which flags are on a spec, but the order is
determined by DAG precedence (childrens' flags take precedence over
parents' and are added on the right) and order (order flags were
specified on the command line is respected).
The solver is responsible for determining when to propagate flags, when
to inheit them from other nodes, when to take them from compiler
preferences, etc.
Weight microarchitectures and prefers more rercent ones. Also disallow
nodes where the compiler does not support the selected target.
We should revisit this at some point as it seems like if I play around
with the compiler support for different architectures, the solver runs
very slowly. See notes in comments -- the bad case was gcc supporting
broadwell and skylake with clang maxing out at haswell.
We didn't have a cardinality constraint for multi-valued variants, so the
solver wasn't filling them in.
- [x] add a requirement for at least one value for multi-valued variants
Variants like `cpu_target` on `openblas` don't have defineed values, but
they have a default. Ensure that the default is always a possible value
for the solver.
Spack was generating the same dependency connstraints twice in the output ASP:
```
declared_dependency("abinit", "hdf5", "link")
:- node("abinit"),
variant_value("abinit", "mpi", "True"),
variant_value("abinit", "mpi", "True").
```
This was because `AspFunction` was modifying itself when called.
- [x] fix `AspFunction` so that every call returns a new object
- [x] Add support for packages.yaml and command-line compiler preferences.
- [x] Rework compiler version propagation to use optimization rather than
hard logic constraints
Technically the ASP output order does not matter, but it's hard to diff
two different solve fomulations unless we order it.
- [x] make sure ASP output is emitted in a deterministic order (by
sorting all hash keys)
This needs more thought, as I am pretty sure the weights are not correct.
Or, at least, I'm not convinced that they do what we want in all cases.
See note in concretize.lp.
Solver now prefers newer versions like the old concretizer. Prefer
package preferences from packages.yaml, preferred=True, package
definition, and finally each version itself.
Competition output only prints out one model, so we do not have to
unnecessarily parse all the non-optimal models. We'll just look at the
best model and bring that in.
In practice, this saves a lot of JSON parsing and spec construction time.
Clingo actually has an option to output JSON -- use that instead of
parsing the raw otuput ourselves.
This also allows us to pick the best answer -- modify the parser to
*only* construct a spec for that one rather than building all of them
like we did before.
- Instead of using default logic, handle variant defaults by minimizing
the number of non-default variants in the solution.
- This actually seems to be pretty fast, and it fixes the long-standing
issue that writing this:
spack install hdf5 ^mpich
will fail if you don't specify hdf5+mpi. With optimization and
allowing enums to be enumerated, the solver seems to be able to quickly
discover that +mpi is the only way hdf5 can depend on mpich, and it
forces the switch to be thrown.
Use '1 { version(x); version(y); version(z) } 1.' instead of declaring
conflicts for non-matching versions. This keeps the sense of version
clauses positive, which will allow them to be used more easily in
conditionals later.
Also refactor `spec_clauses()` method to return clauses that can be used
in conditions, etc. instead of just printing out facts.
