Add a config option to strip `-Werror*` or `-Werror=*` from compile lines everywhere.
```yaml
config:
keep_werror: false
```
By default, we strip all `-Werror` arguments out of compile lines, to avoid unwanted
failures when upgrading compilers. You can re-enable `-Werror` in your builds if
you really want to, with either:
```yaml
config:
keep_werror: all
```
or to keep *just* specific `-Werror=XXX` args:
```yaml
config:
keep_werror: specific
```
This should make swapping in newer versions of compilers much smoother when
maintainers have decided to enable `-Werror` by default.
Parse error information is kept for specs, but it doesn't seem like we propagate it
to the user when we encounter an error. This fixes that.
e.g., for this error in a package:
```python
depends_on("python@:3.8", when="0.900:")
```
Before, with no context and no clue that it's even from a particular spec:
```
==> Error: Unexpected token: ':'
```
With this PR:
```
==> Error: Unexpected token: ':'
Encountered when parsing spec:
0.900:
^
```
* Introduce concretizer:unify option to replace spack:concretization
* Deprecate concretization
* Make spack:concretization overrule concretize:unify for now
* Add environment update logic to move from spack:concretization to spack:concretizer:reuse
* Migrate spack:concretization to spack:concretize:unify in all locations
* For new environments make concretizer:unify explicit, so that defaults can be changed in 0.19
The oneapi and dpcpp compilers are essentially the same except for which
binary is used foc CXX. Spack will detect them as "mixed toolchain" and
not inject compiler optimization flags. This will be needed once
archspec has entries for the oneapi and dpcpp compilers. This PR detects
when dpcpp and oneapi are in the toolchains list and explicitly sets
`is_mixed_toolchain` to `False`.
Error messages for the clingo concretizer have proven challenging. The current messages are incredibly vague and often don't help users at all. Unsat cores in clingo are not guaranteed to be minimal, and lead to cores that are either not useful or need to be post-processed for hours to reach a minimal core.
Following up on an idea from a slack conversation with kwryankrattiger on slack, this PR takes a new approach. We eliminate most integrity constraints and minima/maxima on choice rules in clingo, and instead force invalid states to imply an error predicate. The error predicate can include context on the cause of the error (Package, Version, etc). These error predicates are then heavily optimized against, to ensure that we do not include error facts in the solution when a solution with no error facts could be generated. When post-processing the clingo solution to construct specs, any error facts cause the program to raise an exception. This leads to much more legible error messages. Each error predicate includes a priority and an error message. The error message is formatted by the remaining arguments to produce the error message. The priority is used to ensure that when clingo has a choice of which rules to violate, it chooses the one which will be most informative to the user.
Performance:
"fresh" concretizations appear to suffer a ~20% performance penalty under this branch, while "reuse" concretizations see a speedup of around 33%.
Possible optimizations if users still see unhelpful messages:
There are currently 3 levels of priority of the error messages. Additional priorities are possible, and can allow us finer granularity to ensure more informative error messages are provided in lieu of less informative ones.
Future work:
Improve tests to ensure that every possible rule implying an error message is exercised
A non-existent upstream should not be fatal: it could only mean it is
not deployed yet. In the meantime, it should not block the user to
rebuild anything it needs.
A warning is still emitted, to let the user decide if this is ok or not.
Fixes missing chgrp on symlinks in package installations, and errors on
symlinks referencing non-existent or non-writable locations.
Note: `os.chown(.., follow_symlinks=False)` is python3 only, but
`os.lchown` exists in both versions.
* Change license dir from hard-coded to a configurable item
* Change config item to be a string not an array
Co-authored-by: Todd Gamblin <tgamblin@llnl.gov>
Trying to compute `dag_hash()` or `package_hash()` on a concrete spec that doesn't have
a `_package_hash` attribute would attempt to recompute the package hash.
This most commonly manifests as a failed lookup of a namespace if you attempt to uninstall
or compute the hashes of packages in exsternal repositories that aren't registered, e.g.:
```console
> spack spec --json c/htno
==> Error: Unknown namespace: myrepo
```
While it wouldn't change the already-assigned `dag_hash` value, this behavior is
incorrect, since the package file for a previously concrete spec:
1. might have changed since concretization,
2. might not exist anymore, or
3. might just not be findable by Spack.
This PR ensures that the package hash can't be computed on older concrete specs. Instead
of calling `package_hash()` from within `to_node_dict()`, we now check for the `_package_hash`
attribute and only add the package_hash to the spec record if it's there.
This PR also handles the tricky semantics of computing `package_hash()` at concretization
time. We have to compute it *before* marking the spec concrete so that `to_node_dict` can
use it. But this means that the logic for `package_hash()` can't rely on `spec.concrete`,
as it is called *during* concretization. Instead of checking for concreteness, `package_hash()`
now checks `_patches_assigned()` to determine whether it should add them to the package
hash.
- [x] Add an assert to `package_hash()` so it can't be called on specs for which it
would be wrong.
- [x] Add an `_assign_hash()` method to handle tricky semantics of `package_hash`
and `dag_hash`.
- [x] Rework concretization to call `_assign_hash()` before and after marking specs
concrete.
- [x] Rework content hash part of package hash to check for `_patches_assigned()`
instead of `spec.concrete`.
- [x] regression test
Previously we sorted by hash values for `spack graph`, but changing hashes can make the
test brittle and the node order seem nondeterministic to users.
- [x] Sort nodes in `spack graph` by the default edge order, which takes into account
parent and child names as well as dependency types.
- [x] Update ASCII test output for new order.
The dependency check currently checks whether there are only build
dependencies left for a particular package. However, the database also
contains uninstalled packages, which can cause the check to fail.
For instance, with `bison` and `flex` having already been uninstalled,
`m4` will have the following dependents:
```
bison ('build', 'run')--> m4
flex ('build',)--> m4
libnl ('build',)--> m4
```
`bison` and `flex` should be ignored in this case because they are not
installed anymore.
Fixes#30673
#24556 merged in support for Python's .zip file support via ZipFile.
However as per #30200 ZipFile does not preserve file permissions of
the extracted contents. This PR returns to using the `unzip`
executable on non-Windows systems (as was the case before #24556)
and now uses `tar` on Windows to extract .zip files.
We previously had checks in `directory_layout` to check for build-dependency
conflicts when we weren't storing build dependencies. We don't need
those anymore; we can just rely on the DAG hash now that it includes everything
we know about each spec.
- [x] Remove vestigial code for checking installed spec against concrete spec
in `ensure_installed()`
- [x] Remove `SpecHashCollisionError` -- if specs have the same hash now, they're
the same as far as `DirectoryLayout` should be concerned.
- [x] Convert spec comparison to `dag_hash()` comparison when adding extensions.
The database now stores full hashes, so we need to adjust the criteria we use to
determine if something can be uninstalled. Specifically, it's ok to uninstall thing that
have remaining build-only dependents.
With the original DAG hash, we did not store build dependencies in the database, but
with the full DAG hash, we do. Previously, we'd never tell the concretizer about build
dependencies of things used by hash, because we never had them. Now, we have to avoid
telling the concretizer about them, or they'll unnecessarily constrain build
dependencies for new concretizations.
- [x] Make database track all dependencies included in the `dag_hash`
- [x] Modify spec_clauses so that build dependency information is optional
and off by default.
- [x] `spack diff` asks `spec_clauses` for build dependencies for completeness
- [x] Modify `concretize.lp` so that reuse optimization doesn't affect fresh
installations.
- [x] Modify concretizer setup so that it does *not* prioritize installed versions
over package versions. We don't need this with reuse, so they're low priority.
- [x] Fix `test_installed_deps` for full hash and new concretizer (does not work
for old concretizer with full hash -- leave this for later if we need it)
- [x] Move `test_installed_deps` mock packages to `builtin.mock` for easier debugging
with `spack -m`.
- [x] Fix `test_reuse_installed_packages_when_package_def_changes` for full hash
- [x] update test to use `build_hash` instead of `dag_hash`, as we're testing for
graph structure, and specifically NOT testing for package changes.
- [x] make hash descriptors callable on specs to simplify syntax for invoking them
- [x] make `Spec.spec_hash()` public
This removes all but one usage of runtime hash. The runtime hash was being used to write
historical lockfiles for tests, but we don't need it for that; we can just save those
lockfiles.
- [x] add legacy lockfiles for v1, v2, v3
- [x] fix bugs with v1 lockfile tests (the dummy lockfile we were writing was not actually
a v1 lockfile because it used the new spec file format).
- [x] remove all but one runtime_hash usage -- that one needs a small rework of the
concretizer to really fix, as it's about separate concretization of build
dependencies.
- [x] Document the history of the lockfile format in `environment/__init__.py`
Some test cases had to be modified in a kludgy way so that abstract specs made
concrete would have versions on them. We shouldn't *need* to do this, as the
only reason we care is because the content hash has to be able to get an archive
for a version.
This modifies the content hash so that it can be called on abstract specs,
including only relevant content.
This does NOT add a partial content hash to the DAG hash, as we do not really
want that -- we don't need in-memory spec hashes to need to load package files.
It just makes `Package.content_hash()` less prickly and tests easier to
understand.
`spack monitor` expects a field called `spec_full_hash`, so we shouldn't change that.
Instead, we can pass a `dag_hash` (which is now the full hash) but not change the field
name.
`hashes_final` was used to indicate when a spec was concrete but possibly lacked
`full_hash` or `build_hash` fields. This was only necessary because older Spacks
didn't generate them, and we want to avoid recomputing them, as we likely do not
have the same package files as existed at concretization time.
Now, we don't need to do that -- there is only the DAG hash and specs are either
concrete and have a `dag_hash`, or not concrete and have no `dag_hash`. There's
no middle ground.
Without some enforcement of spec ordering, python 2 produced
different results in the affected test than did python 3. This
change makes the arbitrary but reproducible decision to sort
the specs by their lockfile key alphabetically.
The full hash appears twice in the spec dict now, replacing just
the value replaces it under "hash" and "full_hash". Only replace
the one that appears after "full_hash".
I'm actually not sure what purpose this test served, so maybe it
could be removed, as it may be testing some distinction between
full and dag hash which no longer exists.
For a long time, Spack has used a coarser hash to identify packages
than it likely should. Packages are identified by `dag_hash()`, which
includes only link and run dependencies. Build dependencies are
stripped before hashing, and we have notincluded hashes of build
artifacts or the `package.py` files used to build. This means the
DAG hash actually doesn't represent all the things Spack can build,
and it reduces reproducibility.
We did this because, in the early days, users were (rightly) annoyed
when a new version of CMake, autotools, or some other build dependency
would necessitate a rebuild of their entire stack. Coarsening the hash
avoided this issue and enabled a modicum of stability when only reusing
packages by hash match.
Now that we have `--reuse`, we don't need to be so careful. Users can
avoid unnecessary rebuilds much more easily, and we can add more
provenance to the spec without worrying that frequent hash changes
will cause too many rebuilds.
This commit starts the refactor with the following major change:
- [x] Make `Spec.dag_hash()` include build, run, and link
dependencides and the package hash (it is now equivalent to
`full_hash()`).
It also adds a couple of bugfixes for problems discovered during
the switch:
- [x] Don't add a `package_hash()` in `to_node_dict()` unless
the spec is concrete (fixes breaks on abstract specs)
- [x] Don't add source ids to the package hash for packages without
a known fetch strategy (may mock packages are like this)
- [x] Change how `Spec.patches` is memoized. Using
`llnl.util.lang.memoized` on `Spec` objects causes specs to
be stored in a `dict`, which means they need a hash. But,
`dag_hash()` now includes patch `sha256`'s via the package
hash, which can lead to infinite recursion
`spack pkg list` tests were broken by #29593 for cases when your `builtin.mock` repo
still has stale backup files (or, really, stale directories) sitting around. This
happens if you switch branches a lot. In this case, things like this were causing
erroneous packages in the mock listing:
```
var/spack/repos/builtin.mock/packages/
foo/
package.py~
```
- [x] make `list_packages` consider only directories with one-deep `package.py` files.
Reworking lua to allow easier substitution of the base lua implementation.
Also adding in a maintained version of luajit and re-factoring the entire stack
to use a custom build-system to centralize functionality like environment
variable management and luarocks installation.
The `lua-lang` virtual is now versioned so that a package that requires
Lua 5.1 semantics can get any lua, but one that requires 5.2 will only
get upstream lua.
The luaposix package requires lua-bit32, but only when built with a
lua conforming to version 5.1. This adds the package, and the
dependencies, but exposed a problem with luarocks dependency
detection. Since we're installing each package in its own "tree" and
there's no environment variable to list extra trees, spack now
generates a luarocks config file that lists all the trees of all the
dependencies, and references it by setting `LUAROCKS_CONFIG`
in the build environment of every LuaPackage. This allows luarocks
to find the spack installed dependencies correctly rather than
trying (and failing) to download them.
Co-authored-by: Adam J. Stewart <ajstewart426@gmail.com>
Co-authored-by: Tom Scogland <tscogland@llnl.gov>
Co-authored-by: Massimiliano Culpo <massimiliano.culpo@gmail.com>
