## Summary
Compilers stop being a *node attribute*, and become a *build-only* dependency.
Packages may declare a dependency on the `c`, `cxx`, or `fortran` languages, which
are now treated as virtuals, and compilers would be *providers* for one or more of
those languages. Compilers can also inject runtime dependency, on the node being
compiled. An example graph for something as simple as `zlib-ng` is the following:
<p align="center">
<img src="https://github.com/user-attachments/assets/ee6471cb-09fd-4127-9f16-b9fe6d1338ac" alt="zlib-ng DAG" width="80%" height="auto">
</p>
Here `gcc` is used for both the `c`, and `cxx` languages. Edges are annotated with
the virtuals they satisfy (`c`, `cxx`, `libc`). `gcc` injects `gcc-runtime` on the nodes
being compiled. `glibc` is also injected for packages that require `c`. The
`compiler-wrapper` is explicitly represented as a node in the DAG, and is included in
the hash.
This change in the model has implications on the semantics of the `%` sigil, as
discussed in #44379, and requires a version bump for our `Specfile`, `Database`,
and `Lockfile` formats.
## Breaking changes
Breaking changes below may impact users of this branch.
### 1. Custom, non-numeric version of compilers are not supported
Currently, users can assign to compilers any custom version they want, and Spack
will try to recover the "real version" whenever the custom version fails some operation.
To deduce the "real version" Spack must run the compiler, which can add needless
overhead to common operations.
Since any information that a version like `gcc@foo` might give to the user, can also
be suffixed while retaining the correct numeric version, e.g. `gcc@10.5.0-foo`, Spack
will **not try** anymore to deduce real versions for compilers.
Said otherwise, users should have no expectation that `gcc@foo` behaves as
`gcc@X.Y.Z` internally.
### 2. The `%` sigil in the spec syntax means "direct build dependency"
The `%` sigil in the spec syntax means *"direct build dependency"*, and is not a node
attribute anymore. This means that:
```python
node.satisfies("%gcc")
```
is true only if `gcc` is a direct build dependency of the node. *Nodes without a compiler
dependency are allowed.*
### `parent["child"]`, and `node in spec`, will now only inspect the link/run sub-DAG
and direct build dependencies
The subscript notation for `Spec`:
```python
parent["child"]
```
will look for a `child` node only in the link/run transitive graph of `parent`, and in its
direct build dependencies. This means that to reach a transitive build dependency,
we must first pass through the node it is associated with.
Assuming `parent` does not depend on `cmake`, but depends on a `CMakePackage`,
e.g. `hdf5`, then we have the following situation:
```python
# This one raises an Exception, since "parent" does not depend on cmake
parent["cmake"]
# This one is ok
cmake = parent["hdf5"]["cmake"]
```
### 3. Externals differing by just the compiler attribute
Externals are nodes where dependencies are trimmed, and that _is not planned to
change_ in this branch. Currently, on `develop` it is ok to write:
```yaml
packages:
hdf5:
externals:
- spec: hdf5@1.12 %gcc
prefix: /prefix/gcc
- spec: hdf5@1.12 %clang
prefix: /prefix/clang
```
and Spack will account for the compiler node attribute when computing the optimal
spec. In this branch, using externals with a compiler specified is allowed only if any
compiler in the dag matches the constraints specified on the external. _The external
will be still represented as a single node without dependencies_.
### 4. Spec matrices enforcing a compiler
Currently we can have matrices of the form:
```yaml
matrix:
- [x, y, z]
- [%gcc, %clang]
```
to get the cross-product of specs and compilers. We can disregard the nature of the
packages in the first row, since the compiler is a node attribute required on each node.
In this branch, instead, we require a spec to depend on `c`, `cxx`, or `fortran` for the
`%` to have any meaning. If any of the specs in the first row doesn't depend on these
languages, there will be a concretization error.
## Deprecations
* The entire `compilers` section in the configuration (i.e., `compilers.yaml`) has been
deprecated, and current entries will be removed in v1.2.0. For the time being, if Spack
finds any `compilers` configuration, it will try to convert it automatically to a set of
external packages.
* The `packages:compiler` soft-preference has been deprecated. It will be removed
in v1.1.0.
## Other notable changes
* The tokens `{compiler}`, `{compiler.version}`, and `{compiler.name}` in `Spec.format`
expand to `"none"` if a Spec does not depend on C, C++, or Fortran.
* The default install tree layout is now
`"{architecture.platform}-{architecture.target}/{name}-{version}-{hash}"`
## Known limitations
The major known limitations of this branch that we intend to fix before v1.0 is that compilers
cannot be bootstrapped directly.
In this branch we can build a new compiler using an existing external compiler, for instance:
```
$ spack install gcc@14 %gcc@10.5.0
```
where `gcc@10.5.0` is external, and `gcc@14` is to be built.
What we can't do at the moment is use a yet to be built compiler, and expect it will be
bootstrapped, e.g. :
```
spack install hdf5 %gcc@14
```
We plan to tackle this issue in a following PR.
---------
Signed-off-by: Massimiliano Culpo <massimiliano.culpo@gmail.com>
Signed-off-by: Todd Gamblin <tgamblin@llnl.gov>
Signed-off-by: Harmen Stoppels <me@harmenstoppels.nl>
Co-authored-by: Harmen Stoppels <me@harmenstoppels.nl>
Co-authored-by: Todd Gamblin <tgamblin@llnl.gov>
Supersedes #46792.
Closes#40018.
Closes#31026.
Closes#2700.
There were a number of feature requests for os-specific config. This enables os-specific
config without adding a lot of special sub-scopes.
Support `include:` as an independent configuration schema, allowing users to include
configuration scopes from files or directories. Includes can be:
* conditional (similar to definitions in environments), and/or
* optional (i.e., the include will be skipped if it does not exist).
Includes can be paths or URLs (`ftp`, `https`, `http` or `file`). Paths can be absolute or
relative . Environments can include configuration files using the same schema. Remote includes
must be checked by `sha256`.
Includes can also be recursive, and this modifies the config system accordingly so that
we push included configuration scopes on the stack *before* their including scopes, and
we remove configuration scopes from the stack when their including scopes are removed.
For example, you could have an `include.yaml` file (e.g., under `$HOME/.spack`) to specify
global includes:
```
include:
- ./enable_debug.yaml
- path: https://github.com/spack/spack-configs/blob/main/NREL/configs/mac/config.yaml
sha256: 37f982915b03de18cc4e722c42c5267bf04e46b6a6d6e0ef3a67871fcb1d258b
```
Or an environment `spack.yaml`:
```
spack:
include:
- path: "/path/to/a/config-dir-or-file"
when: os == "ventura"
- ./path/relative/to/containing/file/that/is/required
- path: "/path/with/spack/variables/$os/$target"
optional: true
- path: https://raw.githubusercontent.com/spack/spack-configs/refs/heads/main/path/to/required/raw/config.yaml
sha256: 26e871804a92cd07bb3d611b31b4156ae93d35b6a6d6e0ef3a67871fcb1d258b
```
Updated TODO:
- [x] Get existing unit tests to pass with Todd's changes
- [x] Resolve new (or old) circular imports
- [x] Ensure remote includes (global) work
- [x] Ensure remote includes for environments work (note: caches remote
files under user cache root)
- [x] add sha256 field to include paths, validate, and require for remote includes
- [x] add sha256 remote file unit tests
- [x] revisit how diamond includes should work
- [x] support recursive includes
- [x] add recursive include unit tests
- [x] update docs and unit test to indicate ordering of recursive includes with
conflicting options is deferred to follow-on work
---------
Signed-off-by: Todd Gamblin <tgamblin@llnl.gov>
Co-authored-by: Peter Scheibel <scheibel1@llnl.gov>
Co-authored-by: Todd Gamblin <tgamblin@llnl.gov>
On Windows, libraries search their directory for dependencies, and
we help libraries in Spack-built packages locate their dependencies
by symlinking them into the dependent's directory (we refer to this
as simulated RPATHing).
We extend the convenience functionality here to support base library
directories outside of the package prefix: this is primarily for
running tests in the build directory (which is not located inside
of the final install prefix chosen by spack).
Currently, the custom config scopes are pushed at the top when constructing
configuration, and are demoted whenever a context manager activating an
environment is used - see #48414 for details. Workflows that rely on the order
in the [docs](https://spack.readthedocs.io/en/latest/configuration.html#custom-scopes)
are thus fragile, and may break
This PR allows to assign priorities to scopes, and ensures that scopes of lower priorities
are always "below" scopes of higher priorities. When scopes have the same priority,
what matters is the insertion order.
Modifications:
- [x] Add a mapping that iterates over keys according to priorities set when
adding the key/value pair
- [x] Use that mapping to allow assigning priorities to configuration scopes
- [x] Assign different priorities for different kind of scopes, to fix a bug, and
add a regression test
- [x] Simplify `Configuration` constructor
- [x] Remove `Configuration.pop_scope`
---------
Signed-off-by: Massimiliano Culpo <massimiliano.culpo@gmail.com>
All the build jobs in pipelines are apparently relying on the bug that was fixed.
The issue was not caught in the PR because generation jobs were fine, and
there was nothing to rebuild.
Reverting to fix pipelines in a new PR.
This reverts commit 3ad99d75f9.
Currently, environments can end up with higher priority than `-C` custom
config scopes and `-c` command line arguments sometimes. This shouldn't
happen -- those explicit CLI scopes should override active environments.
Up to now configuration behaved like a stack, where scopes could be only be
pushed at the top. This PR allows to assign priorities to scopes, and ensures
that scopes of lower priorities are always "below" scopes of higher priorities.
When scopes have the same priority, what matters is the insertion order.
Modifications:
- [x] Add a mapping that iterates over keys according to priorities set when
adding the key/value pair
- [x] Use that mapping to allow assigning priorities to configuration scopes
- [x] Assign different priorities for different kind of scopes, to fix a bug, and
add a regression test
- [x] Simplify `Configuration` constructor
- [x] Remove `Configuration.pop_scope`
- [x] Remove `unify:false` from custom `-C` scope in pipelines
On the last modification: on `develop`, pipelines are relying on the environment
being able to override `-C` scopes, which is a bug. After this fix, we need to be
explicit about the unification strategy in each stack, and remove the blanket
`unify:false` from the highest priority scope
Signed-off-by: Massimiliano Culpo <massimiliano.culpo@gmail.com>
A directory and a symlink to it under the same relative path in a
different prefix
```
/prefix1/dir/
/prefix1/dir/file
/prefix2/dir -> /prefix1/dir/
```
are not a blocker to create a view. The view structure simply looks like
this:
```
/view/dir/
/view/dir/file
```
This should be the case independently of the order in which we visit
prefixes, so we could in principle create views order independently.
On macOS, prefix_a/file and prefix_b/FILE map to the same file view/file or view/FILE.
This commit ensures that we test whether a view is created on a case insensitive filesystem and handle projection conflicts accordingly.
* OSError.errno apparently is int | None, but already part of __str__ anyway so drop
* fix disambiguate_spec_from_hashes
* List[Spec].sort() complains, ignore it
* fix typing of KnownCompiler, and use it in asp.py
A few changes to tarball creation (for build caches):
- do not run file to distinguish binary from text
- file is slow, even when running it in a batched fashion -- it usually reads all bytes and has slow logic to categorize specific types
- we don't need a highly detailed file categorization; a crude categorization of elf, mach-o, text suffices.
detecting elf and mach-o is straightforward and cheap
- detecting utf-8 (and with that ascii) is highly accurate: false positive rate decays exponentially as file size increases. Further it's not only the most common encoding, but the most common file type in package prefixes.
iso-8859-1 is cheaply (but heuristically) detected too, and sufficiently accurate after binaries and utf-8 files are classified earlier
- remove file as a dependency of Spack in general, which makes Spack itself easier to install
- detect file type and need to relocate as part of creating the tarball, which is more cache friendly and thus faster
Add missing encoding=utf-8 to various open calls. This makes
files like spec.json, spack.yaml, spack.lock, config.yaml etc locale
independent w.r.t. text encoding. In practice this is not often an
issue since Python 3.7, where the C locale is promoted to
C.UTF-8. But it's better to enforce UTF-8 explicitly, since there is
no guarantee text files are written in the right encoding.
Also avoid opening in text mode if it can be avoided.
* `f.tell` on a `TextIOWrapper` does not return the offset in bytes, but
an opaque integer that can only be used for `f.seek` on the same
object. Spack assumes it's a byte offset.
* Do not open in a locale dependent way, but assume utf-8 (and allow
users to override that)
* Use tempfile to generate a backup/temporary file in a safe way
* Comparison between None and str is valid and on purpose.
If a package `foo` doesn't implement `libs`, the default was to search recursively for `libfoo` whenever asking for `spec[foo].libs` (this also happens automatically if a package includes `foo` as a link dependency).
This can lead to some strange behavior:
1. A package that is normally used as a build dependency (e.g. `cmake` at one point) is referenced like
`depends_on(cmake)` which leads to a fully-recursive search for `libcmake` (this can take
"forever" when CMake is registered as an external with a prefix like `/usr`, particularly on NFS mounts).
2. A similar hang can occur if a package is registered as an external with an incorrect prefix
- [x] Update the default library search to stop after a maximum depth (by default, search
the root prefix and each directory in it, but no lower).
- [x]
The following is a list of known changes to `find` compared to `develop`:
1. Matching directories are no longer returned -- `find` consistently only finds non-dirs,
even at `max_depth`
2. Symlinked directories are followed (needed to support max_depth)
3. `find(..., "dir/*.txt")` is allowed, for finding files inside certain dirs. These "complex"
patterns are delegated to `glob`, like they are on `develop`.
4. `root` and `files` arguments both support generic sequences, and `root`
allows both `str` and `path` types. This allows us to specify multiple entry points to `find`.
---------
Co-authored-by: Peter Scheibel <scheibel1@llnl.gov>
You can now provide multiple roots to a single `find()` call and all of
them will be searched. The roots can overlap (e.g. can be parents of one
another).
This also adds a library function for taking a set of regular expression
patterns and creating a single OR expression (and that library function
is used in `find` to improve its performance).
* `find(..., max_depth=...)` can be used to control how many directories at most to descend into below the starting point
* `find` now enters every unique (symlinked) directory once at the lowest depth
* `find` is now repeatable: it traverses the directory tree in a deterministic order
* Add a descriptor to have a class level constant
This descriptor helps intercept places where we set a value on instances.
It does not really behave like "const" in C-like languages, but is the
simplest implementation that might still be useful.
* Add a descriptor to deprecate properties/attributes of an object
This descriptor is used as a base class. Derived classes may implement a
factory to return an adaptor to the attribute being deprecated. The
descriptor can either warn, or raise an error, when usage of the deprecated
attribute is intercepted.
---------
Co-authored-by: Harmen Stoppels <me@harmenstoppels.nl>
Connection objects are Python version, platform and multiprocessing
start method independent, so better to use those than a mix of plain
file descriptors and inadequate guesses in the child process whether it
was forked or not.
This also allows us to delete the now redundant MultiProcessFd class,
hopefully making things a bit easier to follow.
Commit aa0825d642 accidentally added a semicolon
to the ANSI escape sequence even if the color code was `None` or unknown, breaking the
bold, uncolored font-face. This PR restores the old behavior.
---------
Co-authored-by: Todd Gamblin <tgamblin@llnl.gov>
Both `multiprocessing.connection.Connection.__del__` and `io.IOBase.__del__` called `os.close` on the same file descriptor. As of Python 3.13, this is an explicit warning. Ensure we close once by usef `os.fdopen(..., closefd=False)`
Some Windows Python installations may store the Python exe in Scripts/
rather than the base directory. Update `.command` to search in both
locations on Windows. On all systems, the search is now done
recursively from the search root: on Windows, that is the base install
directory, and on other systems it is bin/.
* Bugfix/Installer: properly track task queueing
* Move ordinal() to llnl.string; change time to attempt
* Convert BuildTask to use kwargs (after pkg); convert STATUS_ to BuildStats enum
* BuildTask: instantiate with keyword only args after the request
* Installer: build request is required for initializing task
* Installer: only the initial BuildTask cannnot have status REMOVED
* Change queueing check
* ordinal(): simplify suffix determination [tgamblin]
* BuildStatus: ADDED -> QUEUED [becker33]
* BuildTask: clarify TypeError for 'installed' argument
Spack can now bootstrap two new dependencies on Windows: GnuPG, and file.
These dependencies are modeled as a separate package, and they install a cross-compiled binary.
Details on how they binaries are built are in https://github.com/spack/windows-bootstrap-resources
* Add missing MultiMethodMeta metaclass in builders
and remove the Python 2 fallback option in favor of hard errors to catch
similar issues going forward.
The fallback option can cause about 10K stat calls due to use of
`realpath` in the inspect module, depending on how deep Spack itself is
nested in the file system, which is ... undesirable.
* code shuffling to avoid circular import
* more reshuffling
* move reserved variant names into variants module
