- [x] remove master from github actions
- [x] remove master from .travis.yml
- [x] make `develop` the default branch for `spack ci`

* Move flake8 tests on Github Actions

* Move shell test to Github Actions

* Moved documentation build to Github Action

* Don't run coverage on Python 2.6

Since we get connection errors consistently on Travis
when trying to upload coverage results for Python 2.6,
avoid computing coverage entirely to speed-up tests.

This is needed because libcuda is used by the driver,
whereas libcudart is used by the runtime. CMake searches
for cudart instead of cuda.

On LLNL LC systems, libcuda is only found in compat and
stubs directories, meaning that the lookup of libraries
fails.

`spack -V` stopped working when we added the `releases/latest` tag to
track the most recent release. It started just reporting the version,
even on a `develop` checkout. We need to tell it to *only* search for
tags that start with `v`, so that it will ignore `releases/latest`.

`spack -V` also would print out unwanted git eror output on a shallow
clone.

- [x] add `--match 'v*'` to `git describe` arguments
- [x] route error output to `os.devnull`

* Change buildcache create `NoOverwriteException` back to a warning.

* Hotfix: move CUDAHOSTCXX

Set only in dependent packages.

* dependent compiler

`spack buildcache list` was trying to construct an `Arch` object and
compare it to `arch_for_spec(<spec>)`. for each spec in the buildcache.
`Arch` objects are only intended to be constructed for the machine they
describe. The `ArchSpec` object (part of the `Spec`) is the descriptor
that lets us talk about architectures anywhere.

- [x] Modify `spack buildcache list` and `spack buildcache install` to
      filter with `Spec` matching instead of using `Arch`.

- [x] Make it easier to get a `Spec` with a proper `ArchSpec` from an
      `Arch` object via new `Arch.to_spec()` method.

- [x] Pull `spack.architecture.default_arch()` out of
      `spack.architecture.sys_type()` so we can get an `Arch` instead of
      a string.

Relative paths in views have been broken since #17608 or earlier.

- [x] Fix by passing base path of the environment into the `ViewDescriptor`.
      Relative paths are calculated from this path.

Relocation of sbang needs to be done when the spack prefix changes even if the install tree has not changed. (#17455)

Relocate rpaths for all binaries, then do text bin replacement if the rpaths still exist after running patchelf/otool (#17418)

* add tutorial setup script to share/spack

* Add check for Ubuntu 18, fix xvda check, fix apt-get errors
  - now works on t2.micro, t2.small, and m instances
  - apt-get needs retries around it to work

Co-authored-by: Todd Gamblin <tgamblin@llnl.gov>

A bug was introduced in #13100 where ChildErrors would be redundantly
printed when raised during a build. We should eventually revisit error
handling in builds and figure out what the right separation of
responsibilities is for distributed builds, but for now just skip
printing.

- [x] SpackErrors were designed to be printed by the forked process, not
      by the parent, so check if they've already been printed.
- [x] update tests

Fixes #17299

Cray Shasta systems appear to use an unmodified Sles or other Linux operating system on the backend (like Cray "Cluster" systems and unlike Cray "XC40" systems that use CNL).

This updates the CNL version detection to properly note that this is the underlying OS instead of CNL and delegate to LinuxDistro.

* environment-views: fix bug where missing recipe/repo breaks env commands

When a recipe or a repo has been removed from Spack and an environment
is active, it causes the view activation to crash Spack before any
commands can be executed. Further, the error message it not at all clear
in explaining the issue.

This forces view regeneration to always start from scratch to avoid the
missing package recipes, and defaults add_view=False in main for views activated
by the `spack -e` option.

* add messages to env status and deactivate

Warn users that a view may be corrupt when deactivating an environment
or checking its status while active. Updated message for activate.

* tests for view checking

Co-authored-by: Gregory Becker <becker33@llnl.gov>

* switch from bool to int debug levels

* Added debug options and changed lock logging to use more detailed values

* Limit installer and timestamp PIDs to standard debug output

* Reduced verbosity of fetch/stage/install output, changing most to debug level 1

* Combine lock log methods; change build process install to debug

* Changed binary cache install messages to extraction messages

* bugfix: make compiler preferences slightly saner

This fixes two issues with the way we currently select compilers.

If multiple compilers have the same "id" (os/arch/compiler/version), we
currently prefer them by picking this one with the most supported
languages.  This can have some surprising effects:

* If you have no `gfortran` but you have `gfortran-8`, you can detect
  `clang` that has no configured C compiler -- just `f77` and `f90`. This
  happens frequently on macOS with homebrew. The bug is due to some
  kludginess about the way we detect mixed `clang`/`gfortran`.

* We can prefer suffixed versions of compilers to non-suffixed versions,
  which means we may select `clang-gpu` over `clang` at LLNL. But,
  `clang-gpu` is not actually clang, and it can break builds. We should
  prefer `clang` if it's available.

- [x] prefer compilers that have C compilers and prefer no name variation
  to variation.

* tests: add test for which()

Apple's gcc is really clang. We previously ignored it by default but
there was a regression in #17110.

Originally we checked for all clang versions with this, but I know of
none other than `gcc` on macos that actually do this, so limiting to
`apple-clang` should be ok.

- [x] Fix check for `apple-clang` in `gcc.py` to use version detection
  from `spack.compilers.apple_clang`

Ensure compilers installed from buildcache are registered.

The `spack-build-env.txt` file may contains many secrets, but the obvious one is the private signing key in `SPACK_SIGNING_KEY`. This file is nonetheless uploaded as a build artifact to gitlab. For anyone running CI on a public version of Gitlab this is a major security problem. Even for private Gitlab instances it can be very problematic.

Co-authored-by: Scott Wittenburg <scott.wittenburg@kitware.com>

* llvm-flang Only build offload code if cuda enabled

The current version executes `cmake(*args)` always as part of the post install.  If device offload is not part of the build, this results in referencing `args` without it being set and the error:

```
==> Error: UnboundLocalError: local variable 'args' referenced before assignment

```

Looking at prevoous version of `llvm-package.py` this whole routine appears to be only required for offload, some indent `cmake/make/install` to be under the `if`.

* Update package.py

Add comment

* clear mpi env vars

The error message was not updated when the behavior of Spack environments
was changed to not automatically activate the local environment in #17258.
The previous error message no longer makes sense.

When Spack installs a package, it stores repository package.py files
for it and all of its dependencies - any package with a Spack metadata
directory in its installation prefix.

It turns out this was too broad: this ends up including external
packages installed by Spack (e.g. installed by another Spack instance).
Currently Spack doesn't store the namespace properly for such packages,
so even though the package file could be fetched from the external,
Spack is unable to locate it.

This commit avoids the issue by skipping any attempt to locate and copy
from the package repository of externals, regardless of whether they
have a Spack repo directory.

On Cray platforms, we rely heavily on the module system to figure out
what targets, compilers, etc. are available. This unfortunately means
that we shell out to the `module` command as part of platform
initialization.

Because we run subcommands in a shell, we can get infinite recursion if
`setup-env.sh` and friends are in some init script like `.bashrc`.

This fixes the infinite loop by adding guards around `setup-env.sh`,
`setup-env.csh`, and `setup-env.fish`, to prevent recursive
initializations of Spack. This is safe because Spack never shells out to
itself, so we do not need it to be initialized in subshells.

- [x] add recursion guard around `setup-env.sh`
- [x] add recursion guard around `setup-env.csh`
- [x] add recursion guard around `setup-env.fish`

* fix binutils deptype for gcc

binutils needs to be a run dependency of gcc

* Fix gcc+binutils build on RHEL7+

static-libstdc++ is not available with system gcc.
Anyway, as it is for bootstraping, we do not really care depending on
a shared libstdc++.

Co-authored-by: Michael Kuhn <michael@ikkoku.de>