There was a bit of mystery surrounding the arguments for `_setup_pkg_and_run`. It passes
two file descriptors for handling the `gmake`'s job server in child processes, but they are
unsed in the method.
It turns out that there are good reasons to do this -- depending on the multiprocessing
backend, these file descriptors may be closed in the child if they're not passed
directly to it.
- [x] Document all args to `_setup_pkg_and_run`.
- [x] Document all arguments to `_setup_pkg_and_run`.
- [x] Add type hints for `_setup_pkg_and_run`.
- [x] Refactor exception handling in `_setup_pkg_and_run` so it's easier to add type
hints. `exc_info()` was problematic because it *can* return `None` (just not
in the context where it's used). `mypy` was too dumb to notice this.
Signed-off-by: Todd Gamblin <tgamblin@llnl.gov>
* Revert "`cc`: ensure that RPATHs passed to linker are unique"
This reverts commit 2613a14c43.
* Revert "`cc`: simplify ordered list handling"
This reverts commit a76a48c42e.
Updated the terminology for the two types of environments to be
consistent with that used in the tutorial for the last three years.
Additionally:
* changed 'anonymous' to 'independent in environment command+test for consistency.
#45205 already removed previous use of single letter packages
from unit tests, in view of reserving `c` as a language (see #45191).
Some use of them has been re-introduced accidentally in #46382, and
is making unit-tests fail in the feature branch #45189 since there
`c` is a virtual package.
Signed-off-by: Massimiliano Culpo <massimiliano.culpo@gmail.com>
On sysroot systems like gentoo prefix, as well as nix/guix, our "is
system path" logic is broken cause it's static.
Talking about "the system paths" is not helpful, we have to talk
about default search paths of the dynamic linker instead.
If glibc is recent enough, we can query the dynamic loader's default
search paths, which is a much more robust way to avoid registering
rpaths to system dirs, which can shadow Spack dirs.
This PR adds an **additional** filter on rpaths the compiler wrapper
adds, dropping rpaths that are default search paths. The PR **does
not** remove any of the original `is_system_path` code yet.
This fixes issues where build systems run just-built executables
linked against their *not-yet-installed libraries*, typically:
```
LD_LIBRARY_PATH=. ./exe
```
which happens in `perl`, `python`, and other non-cmake packages.
If a default path is rpath'ed, it takes precedence over
`LD_LIBRARY_PATH`, and a system library gets loaded instead
of the just-built library in the stage dir, breaking the build. If
default paths are not rpath'ed, then LD_LIBRARY_PATH takes
precedence, as is desired.
This PR additionally fixes an inconsistency in rpaths between
cmake and non-cmake packages. The cmake build system
computed rpaths by itself, but used a different order than
computed for the compiler wrapper. In fact it's not necessary
to compute rpaths at all, since we let cmake do that thanks to
`CMAKE_INSTALL_RPATH_USE_LINK_PATH`. This covers rpaths
for all dependencies. The only install rpaths we need to set are
`<install prefix>/{lib,lib64}`, which cmake unfortunately omits,
although it could also know these. Also, cmake does *not*
delete rpaths added by the toolchain (i.e. Spack's compiler
wrapper), so I don't think it should be controversial to simplify
things.
The current `Spec.splice` model is very limited by the inability to splice specs that
contain multiple nodes with the same name. This is an artifact of the original
algorithm design predating the separate concretization of build dependencies,
which was the first feature to allow multiple specs in a DAG to share a name.
This PR provides a complete reimplementation of `Spec.splice` to avoid that
limitation. At the same time, the new algorithm ensures that build dependencies
for spliced specs are not changed, since the splice by definition cannot change
the build-time information of the spec. This is handled by splitting the dependency
edges and link/run edges into separate dependencies as needed.
Signed-off-by: Gregory Becker <becker33@llnl.gov>
* CI: Add documentation for adding new stacks and runners
* More docs for runner registration
---------
Co-authored-by: Zack Galbreath <zack.galbreath@kitware.com>
Co-authored-by: Bernhard Kaindl <contact@bernhard.kaindl.dev>
This PR shorten the string representation for concrete specs,
in order to make it more legible.
Signed-off-by: Massimiliano Culpo <massimiliano.culpo@gmail.com>
macOS Sequoia's linker will complain if RPATHs on the CLI are specified more than once.
To avoid errors due to this, make `cc` only append unique RPATHs to the final args list.
This required a few improvements to the logic in `cc`:
1. List functions in `cc` didn't have any way to append unique elements to a list. Add a
`contains()` shell function that works like our other list functions. Use it to implement
an optional `"unique"` argument to `append()` and an `extend_unique()`. Use that to add
RPATHs to the `args_list`.
2. In the pure `ld` case, we weren't actually parsing `RPATH` arguments separately as we
do for `ccld`. Fix this by adding *another* nested case statement for raw `RPATH`
parsing. There are now 3 places where we deal with `-rpath` and friends, but I don't
see a great way to unify them, as `-Wl,`, `-Xlinker`, and raw `-rpath` arguments are
all ever so slightly different.
3. Fix ordering of assertions to make `pytest` diffs more intelligible. The meaning of
`+` and `-` in diffs changed in `pytest` 6.0 and the "preferred" order for assertions
became `assert actual == expected` instead of the other way around.
Signed-off-by: Todd Gamblin <tgamblin@llnl.gov>
`cc` divides most paths up into system paths, spack managed paths, and other paths.
This gets really repetitive and makes the code hard to read. Simplify the script
by adding some functions to do most of the redundant work for us.
Signed-off-by: Todd Gamblin <tgamblin@llnl.gov>
SimpleFilesystemView was producing an error due to looking for a
<prefix>/lib/.spack folder. Also, view_destination had no effect and
wasn't called. Changed this by instead patching in the correct
installation prefix for dictionaries.
Since aspell is using the resolved path of the executable prefix, the
runtime environment variable ASPELL_CONF is set to correct the prefix
when in a view. With this change aspell can now find installed
dictionaries. Verified with:
aspell dump config
aspell dump dicts
If `add_padding()` is allowed to return a path with a trailing path
separator, it will get collapsed elsewhere in Spack. This can lead to
buildcache entries that have RPATHS that are too short to be replaced by
other users whose install root happens to be padded to the correct
length. Detect this and replace the trailing path separator with a
concrete path character.
Signed-off-by: Samuel E. Browne <sebrown@sandia.gov>
We've seen `getfqdn()` cause slowdowns on macOS in CI when added elsewhere. It's also
called by database.py every time we write the DB file.
- [x] replace the call with a memoized version so that it is only called once per process.
Signed-off-by: Todd Gamblin <tgamblin@llnl.gov>
This PR introduces a new heuristic for the solver, which behaves better when
compilers are treated as nodes. Apparently, it performs better also on `develop`,
where compilers are still node attributes.
The new heuristic:
- Sets an initial priority for guessing a few attributes. The order is "nodes" (300),
"dependencies" (150), "virtual dependencies" (60), "version" and "variants" (30), and
"targets" and "compilers" (1). This initial priority decays over time during the solve, and
falls back to the defaults.
- By default, it considers most guessed facts as "false". For instance, by default a node
doesn't exist in the optimal answer set, or a version is not picked as a node version etc.
- There are certain conditions that override the default heuristic using the _priority_ of
a rule, which previously we didn't use. For instance, by default we guess that a
`attr("variant", Node, Variant, Value)` is false, but if we know that the node is already
in the answer set, and the value is the default one, then we guess it is true.
Signed-off-by: Massimiliano Culpo <massimiliano.culpo@gmail.com>
The `spack.target.Target` class is a weird entity, that is just needed to:
1. Sort microarchitectures in lists deterministically
2. Being able to use microarchitectures in hashed containers
This PR removes it, and uses `archspec.cpu.Microarchitecture` directly. To sort lists, we use a proper `key=` when needed. Being able to use `Microarchitecture` objects in sets is achieved by updating the external `archspec`.
Signed-off-by: Massimiliano Culpo <massimiliano.culpo@gmail.com>
While the existing getting started guide does in fact reference the
powershell support, it's a footnote and easily missed. This PR adds
explicit, upfront mentions of the powershell support. Additionally
this PR adds notes about some of the issues with certain components
of the spec syntax when using CMD.
Removes `spack.package_base.PackageBase.do_{install,deprecate}` in favor of
`spack.installer.PackageInstaller.install` and `spack.installer.deprecate` resp.
That drops a dependency of `spack.package_base` on `spack.installer`, which is
necessary to get rid of circular dependencies in Spack.
Also change the signature of `PackageInstaller.__init__` from taking a dict as
positional argument, to an explicit list of keyword arguments.
Continuing the work started in #40326, his changes the structure
of Variant metadata on Packages from a single variant definition
per name with a list of `when` specs:
```
name: (Variant, [when_spec, ...])
```
to a Variant definition per `when_spec` per name:
```
when_spec: { name: Variant }
```
With this change, everything on a package *except* versions is
keyed by `when` spec. This:
1. makes things consistent, in that conditional things are (nearly)
all modeled in the same way; and
2. fixes an issue where we would lose information about multiple
variant definitions in a package (see #38302). We can now have,
e.g., different defaults for the same variant in different
versions of a package.
Some notes:
1. This required some pretty deep changes to the solver. Previously,
the solver's job was to select value(s) for a single variant definition
per name per package. Now, the solver needs to:
a. Determine which variant definition should be used for a given node,
which can depend on the node's version, compiler, target, other variants, etc.
b. Select valid value(s) for variants for each node based on the selected
variant definition.
When multiple variant definitions are enabled via their `when=` clause, we will
always prefer the *last* matching definition, by declaration order in packages. This
is implemented by adding a `precedence` to each variant at definition time, and we
ensure they are added to the solver in order of precedence.
This has the effect that variant definitions from derived classes are preferred over
definitions from superclasses, and the last definition within the same class sticks.
This matches python semantics. Some examples:
```python
class ROCmPackage(PackageBase):
variant("amdgpu_target", ..., when="+rocm")
class Hipblas(ROCmPackage):
variant("amdgpu_target", ...)
```
The global variant in `hipblas` will always supersede the `when="+rocm"` variant in
`ROCmPackage`. If `hipblas`'s variant was also conditional on `+rocm` (as it probably
should be), we would again filter out the definition from `ROCmPackage` because it
could never be activated. If you instead have:
```python
class ROCmPackage(PackageBase):
variant("amdgpu_target", ..., when="+rocm")
class Hipblas(ROCmPackage):
variant("amdgpu_target", ..., when="+rocm+foo")
```
The variant on `hipblas` will win for `+rocm+foo` but the one on `ROCmPackage` will
win with `rocm~foo`.
So, *if* we can statically determine if a variant is overridden, we filter it out.
This isn't strictly necessary, as the solver can handle many definitions fine, but
this reduces the complexity of the problem instance presented to `clingo`, and
simplifies output in `spack info` for derived packages. e.g., `spack info hipblas`
now shows only one definition of `amdgpu_target` where before it showed two, one of
which would never be used.
2. Nearly all access to the `variants` dictionary on packages has been refactored to
use the following class methods on `PackageBase`:
* `variant_names(cls) -> List[str]`: get all variant names for a package
* `has_variant(cls, name) -> bool`: whether a package has a variant with a given name
* `variant_definitions(cls, name: str) -> List[Tuple[Spec, Variant]]`: all definitions
of variant `name` that are possible, along with their `when` specs.
* `variant_items() -> `: iterate over `pkg.variants.items()`, with impossible variants
filtered out.
Consolidating to these methods seems to simplify the code a lot.
3. The solver does a lot more validation on variant values at setup time now. In
particular, it checks whether a variant value on a spec is valid given the other
constraints on that spec. This allowed us to remove the crufty logic in
`update_variant_validate`, which was needed because we previously didn't *know* after
a solve which variant definition had been used. Now, variant values from solves are
constructed strictly based on which variant definition was selected -- no more
heuristics.
4. The same prevalidation can now be done in package audits, and you can run:
```
spack audit packages --strict-variants
```
This turns up around 18 different places where a variant specification isn't valid
given the conditions on variant definitions in packages. I haven't fixed those here
but will open a separate PR to iterate on them. I plan to make strict checking the
defaults once all existing package issues are resolved. It's not clear to me that
strict checking should be the default for the prevalidation done at solve time.
There are a few other changes here that might be of interest:
1. The `generator` variant in `CMakePackage` is now only defined when `build_system=cmake`.
2. `spack info` has been updated to support the new metadata layout.
3. split out variant propagation into its own `.lp` file in the `solver` code.
4. Add better typing and clean up code for variant types in `variant.py`.
5. Add tests for new variant behavior.
