The --enable-relax-coord-bound option was added in 1.8.0. NetCDF
4.6.3 enabled relax coords by default and required it from PNetCDF.
PNetCDF 1.10.0 enabled it by default but for versions between 1.8.0
and 1.10.0 it should be enabled explicitly.
* megadock: new package at 4.0.3
* megadock: remove CudaPackage redundancies
* megadock: move env vars to build targets, use cuda_arch variant
* megadock: flake8 fixes, more flexible mathlib
* Fix py-yt for Python 2 by requesting older py-ipython version.
* Add conflicts statement to py-ipython because version 7: is not
compatible with python@2.7
f242f5f removed spec.architecture.platform_os, which was used to test
an option in the 'm4' package. The attribute is now accessed as
'spec.architecture.os'.
* Add versions 3.27.{0,1,2} and 3.28.0
* Add url_for_version: the URL format includes the year so this
function maps version ranges to years
* Restore patches for versions removed in 84c547c: this commit
removed versions before 3.26.0 due to a security issue with FTS
support. Add a +fts variant (default true) that enables FTS
support along with a conflict for +fts with earlier versions.
* Starting with the upcoming 10.1 release, Dyninst changed the names of
some cmake args to camel case: LIBELF_* to LibElf_* and LIBDWARF_* to
LibDwarf_*. This change is needed now for @develop and will be used
for 10.1 when it is released.
* New versions of Dyninst now requires cmake version 3.4.0 or later.
* Restrict boost dependency version to < 1.70.0 to avoid build error
+ The additional patch enables a non-optimized random123 code path that will
work on ARM architectures. This solution is provided only for version 1.09 to
allow the most current version of Random123 to function on ARM architectures.
+ A more complete, long term solution will be the addition of a native ARM
intrinsics-based implementation of random123. This solution is being worked
on but it is several weeks or months away. Once the full implementation is
available it will be provided to the author/maintainer of Random123.
The default install for llvm should just be the common typical case, i.e.
support for local host and cpu architectures. Enablingsupport for the wide
array of auxiliary architectures should be explicit rather than implicit.
* Mesa should depend_on('glproto')
The mesa package refers to `GL/glproto.h`. On systems that don't have
the OS packages installed, this leads to failures during the build
[e.g. this comment in
01482](https://github.com/spack/spack/pull/10482#issuecomment-488786745).
This fixes it. Tested on a minimally provisioned CentOS 7.
* Constrain glproto prereq to when +glx
* mesa: make glproto a build only dep
* The 'mkl' variant implementation was not complete and was
preventing the package from building, so remove it.
* The slate.hh file has moved in the 'develop' version (which is
currently the only version of SLATE listed)
* FAODEL: Add FAODEL package to spack
FAODEL (Flexible, Asynchronous, Object Data-Exchange Libraries) is a collection
of software libraries that are used to implement different data management
services on high-performance computing (HPC) platforms. This project is part of
the Advanced Technology Development and Mitigation (ATDM) effort for NNSA's ASC
program at Sandia National Laboratories.
(cherry picked from commit 6afe045d25b5c000b612116a5695aebbec56b61a)
* FAODEL: Update FAODEL package
Dependencies: set minimum versions for boost, cmake and libfabric
Dependencies: add conflict for boost v1.59.0
Patch: add a patch for v1.1803.1
Patch: add a when= for the v1.1811.1 specific patches
Compiler: check for C++11 support in gcc
(cherry picked from commit fb6ed2b3cf20d7e4fff647cba5a70887a78f627f)
* FAODEL: Update FAODEL package
* replace previous maintainers with @tkordenbrock and @craigulmer
* default to shared libs to meet Spack policy
* increase min boost version to 1.60.0 and remove boost 1.59.0 conflict
* replace complex version check with conflict for gcc <= 4.8.0 (C++11 support required)
* set type=build for cmake
* enable googletest for all versions with type=build
* enable tests for for all versions except when MPI is disabled
* FAODEL: Update FAODEL package
* add version v1.1811.2
* limit comment line length
* arm-forge: add all available versions
* Add TODO for the version hell
* Fix linting
Looks like 'black' formatter isn't Spack friendly
* Fix versions ordering
* Remove obsolete package (allinea-forge → arm-forge)
* Set mpi compiler wrappers using CC/etc. variables as recommended
at https://spack.readthedocs.io/en/latest/packaging_guide.html#compiler-wrappers
rather than setting CMake variables like -DCMAKE_C_COMPILER
* Use convenience functionality provided by inherited CMakePackage to
export all compiler flags as CMake options
* Remove old Singularity versions, add version 3.1.1
* Recent versions of Singularity build with go rather than autotools:
update dependencies and build logic
* Move old Singularity versions to new singularity-legacy package
which uses the autotools build system
* Some binaries built by Singularity need to be run as root: include
a script that the user can run after the Spack Singularity install
to change these permissions.
* The Singularity go build expects to work with a custom stage
directory relative to GOPATH: override the stage phase to create
this expected path.
* Update Singularity install config to point to Spack-installed
makesquashfs dependency
Based on the LLVM documentation [1], Python is used to run the automated
test suite. Therefore is it always a dependency for LLVM. However, if
build without Python (~python), we limit it to a build time dependency.
Note that py-lit is not added as a spack dependency even though it is
available as a spack package. This is because it is already included
in llvm and llvm is difficult to configure using an external py-lit
(several CMake variables to set correctly). Additionally, having
py-lit as a spack dependency adds Python as a runtime dependency
for llvm even though it is not required at runtime.
[1] https://llvm.org/docs/GettingStarted.html#requirements
* extend Version class so that 2.0 > 1.develop > 1.1
* add concretization tests, with preferences and preferred version.
* add master, head, trunk as develop-like versions, develop > master > head > trunk
* update documentation on version comparison
* Version 1.0.0 was released and the master branch tracks the
latest release.
* Multiple issues with MacOS have been fixed. The static patch is
now unneeded.
* Add support for building Ginkgo with the full Block Jacobi
performance (compilation takes a long time).
* Do not build anything which is not useful for the installation.
Namely, documentation, benchmarks, examples and tests.
* mesa: re-writen to new meson build system
* xorg-server: Remove unneeded mesa dependency
* glx: provide the virtual GLX package
* Convert most "mesa" dependencies to use virtual gl and glx dependencies
* mesa: Switch to always building from the git repo instead of a tarball
* Use gl and glx version ranges instead of exact versions
* glx: provide a default package
* add auto to swr variant
* Revert "add auto to swr variant"
This reverts commit 2e7cd8a8aca07767c00657d0d09985ddd6306883.
* update mesa package to version 19.0
* set +glx by default only when running on linux
* fix style issues in geant4 package
* remove virtual package "egl"
margo package is a required dependency for unifycr but the package
was not available in spack. margo requires a recent version of argobots
that has the ABT_SCHED_BASIC_WAIT scheduler (rc1 does while previously
available versions of this package in spack did not have it).
Kahip does not build with clang (conflicts('%clang') in kahip package.
Because of this, openfoam-com %clang fails concretization. Therefore,
this changes the default openfoam-com variant for kahip to False.
* Remove -ffast-math optimization flag
* Add 'develop' version which points to git repo
* Depend on libint for all versions >= 3.0
* Build directory is different for versions >= 7 (including develop)
* Add "make test" as post install check
* Add versions 3.4.4 through 4.1.0 and 4.1.0-openvino
* Add conflict for gcc version >= 7 when building with +cuda
* Later version of protobuf is required for later versions of OpenCV
* Explicitly remove protobuf dependency with CMake option when not
building with +dnn
* root: new versions, variants and patches to bring in to line with FNAL's use.
* Support MySQL.
* Patches:
* Bigger stringbuf.
* Find MySQL even when it's called mariadb-c-client.
* Honor external provision of unuran.
* Properly handle the ROOT7 web GUI for 6.16/00.
* `memstat` variant is now default-OFF due to version-related constraints.
* Add missing dependencies.
* Propagate `python` to `libxml2`.
* Handle external FTGL dependency with X / OpenGL: depends on PR ####
* Force `SPACK_INCLUDE_DIRS` to avoid warning.
* New `gminimal` variant default ON per discussion with Javier Cervantes.
* Style fixes per flake8.
* Remove redundant `PATH_SUFFIX` per https://github.com/root-project/root/pull/1993#discussion_r276840329
* Comment patches per https://github.com/spack/spack/pull/11215#discussion_r276887199.
* Reflect the fact that MariaDB support patch is now upstream.
* Restore ncurses dependency for libtinfo.
Add zlib dependency to squashfs
On my CentOS system it was getting linked in (from Spack, no less)
fortuitously.
On a minimal Ubuntu system it failed until it was added as an explicit
dependency.
* ftgl: update to 2.1.3-rc5.
This updates to the last release (from 2008) with patches required to work with Spack and Root.
* Working 2.1.2 with/without docs, 2.1.3-rc5 without.
Allow specifying whether AVX2 or AVX512 should be enabled or disabled
on x86 targets.
As AVX2 hardware and toolchain support is quite ubiquitous by now,
AVX2 is enabled by default. Also AVX2 support is not disabled when
building the +virtual_machine variant.
AVX512 is not supported in older but still supported toolchains,
hardware is still expensive, and OpenBLAS AVX512 kernels still have
bugs. Thus AVX512 is disabled by default.
* Add Mercury version 1.0.1
* Remove cci and bmi dependency requirements: these are not currently
defined in Spack. If +bmi is enabled, the effect is that it may
use a system install of bmi
* Add a note that fabtests is deprecated and has been folded into the
libfabric package. Add the fabtests resource to the libfabric
package
* Add CMake option settings to Mercury based on variant settings
* Mercury: Set OFI testing protocol depending on how libfabric was
built
* Mercury: add version constraints to dependencies
Fixes#11226
MySQL 5.7.X and Boost 1.59.0 have a closer relationship than
MySQL >= 8.0 and their corresponding Boost versions: 5.7.X needs to
link against boost libraries which is indicated by marking boost as
a link dependency (removing the 'type' reverts to the default, which
includes link and build).
Starting with version 7.64.0, curl skips cookies specified with "-H"
(generic custom header specification) when following redirects, which
causes the "accept-license"-cookie for jdk to disappear. The result
is that jdk fails to download. This uses the "-b" option (which is
used specifically for configuring cookies) instead, which is more
permissive.
spec['dep'].libs.ld_flags can return system paths in some cases.
This is a targeted fix for the petsc package when retrieving
suite-sparse dependency libraries.
* added latest versions + malloc variant
* added a variant to activate different instrumentations
* added a variant to activate or not networking
* added a variant to activate or not cuda and tools
* enforce c++ standard and make it consistent with boost
* cflags and cxxflags from packages.yaml need to be passed into QMCPACK's CMake explictly for now.
* You need the ifcore library from the Intel compler to be manually linked in when you linked against a LAPACK provider that is not MKL.
* Update spec format to simpler syntax, maintain backwards compatibility
* Switch to new spec.format method throughout internals
* update package files for new format strings
* documentation and minor code cleanup. removed nonsensical variant sigils
* Initial commit for v2.0 of the CEED software suite.
* Update Nek packages and gslib
* Help spack concretize the hypre version for ceed-2.0.
* Fix nekcem install error
* Add support for gfortran v8 in nek5000 and nekcem.
* Split Nek5000 into Nek5000 and Nektools
* Get Nektools to build fine in Theta
* Fix travis failure: remove unused 'import numbers' from nek5000.
* Check for gfortran if it is wrapped
* Tweak the detection of gfortran in nek5000.
* Fix Nek packages to add -std=legacy when FC=gcc
* spack install ceed~petsc works fine on Theta
* Fix flake8 errors
* Fix more flake8 tests
* Fix an import issue
* Tweak the suite-sparse package to avoid interaction with existing system
installations of suite-sparse.
* petsc: update superlu-dist dependency
* Updates in the packages: occa, libceed, and ceed.
* In the libceed package, explicitly tell nvcc which host compiler to use.
* Fix python formatting.
* Simplify the test for gfortran in nek* packages.
* ceed: 2.0 uses petsc@3.11.0
* hpgmg-0.4; use from ceed@2.0.0
* Update the hypre dependency for ceed 2.0.
* Disable the superlu-dist dependency (through hypre) when using a
+quickbuild of ceed 2.0.
* petsc-3.11.0: add xlf fix
* nekcem: has a build dependency on Python 2.7+
* hpgmg: better setting of compiler options and use python for configure
* libceed: use v0.4 tag
* libceed: fix 0.4 release oops (pkgconfig version)
* Add a patch for magma-2.5.0 that brings it up the current 'master'.
* In the mfem package, install the examples, miniapps, and data under
$prefix/share/mfem.
* In the magma package, apply a patch to v2.5.0 that disables
magma_sparse - for testing purposes.
* In the magma package, link the 'magma' library with the
'nvToolsExt' library.
* In the magma package, update the 'magma-2.5.0.patch' with the latest
commits from the magma source repository. Also, remove the library
'nvToolsExt' from the 'magma-2.5.0-cmake.patch' - now it is not
needed.
* In the magma package, disable OpenMP when using v2.5.0 with the
IBM XL compiler.
Please enter the commit message for your changes. Lines starting
* In the mfem package, add version for the 'laghos-v2.0' tag; also,
prefix the versions `laghos-v*` with their respective development
version numbers -- this way they are properly ordered within spack
relative to the official numbered versions.
* petsc: add version 3.11.1 (#11179)
(cherry picked from commit 1eab6e3c86)
* ceed-2.0: use petsc-3.11.1
* this-is-so-dumb.f -> empty.f
* Add versions 1.4.0, 1.3.0, and 1.2.0
* Add dependencies and update constraints (e.g. version 1.4.0 of
precice requires a later version of boost)
* Add maintainers
A Spack hack for MKL usage was needed in QMCPACK prior to version 3.5
when non-Intel compilers were used. This Spack hack could have
undesirable side-effects and led to ugly code that was extra work to
maintain. For older versions of QMCPACK, we throw a conflict if you
want to use Intel MKL with the other compilers commonly found on x86.
Note that there is no impact to version of QMCPACK since 3.5.
* Fix bug when system perl is used.
If system perl is being used through packages.yaml there cannot be dependent modules.
The setup_dependent_packages function should not do anything in that case.
* Add versions 5.2.0 and 5.1.0
* Avoid adding version 6.0.0: this version has a new API that is
not compatible with older versions (and dependents may break with
it)
* Added comment block explaining the problems with version 6.x
* Replace old datumgrid resource (1.7) with new (1.8)
* Add speculative sqlite dependency for when a 6.x version is added
* add version 7.0.1
* replace md5 with sha256 for various versions.
* v7.0.1 fixes some environment hacking to properly find MPI and
system architecture
This adds a stub script for mpirun and other standard executables
when installing OpenMPI with slurm. The purpose is to make the
removal less of a surprise to administrators/users: it explains why
they were removed and how to restore them.
* adding package to install squashfs for singularity
* dont need os import
Signed-off-by: Vanessa Sochat <vsochat@stanford.edu>
* updating md5 to sha1 for squashfs package
Signed-off-by: Vanessa Sochat <vsochat@stanford.edu>
* two blank lines
Signed-off-by: Vanessa Sochat <vsochat@stanford.edu>
* Newer versions of IBM XL no longer define `__xlC__`, but define `__ibmxl__`
instead. This one-line patch fixes this problem in Random123.
* This patch was also provided to the Random123 maintiners
(Random123@DEShawResearch.com). I don't expect a new release from them for a
while, so I believe that this patch is the correct path forward (for now).
* initial work to make use of an 'upstream' spack installation: this uses the DB of the upstream installation to check if a package is installed
* need to query upstream dbs when adding new record to local db
* prevent reindexing upstream DBs
* set prefix on specs read from DB based on path stored in install record
* check that Spack does not install packages that are recorded as installed in an upstream db
* externals do not add their path to install records - need to use 'external_path' to get path of upstream externals
* views need to check for upstream installations when linking metadata
* package and spec now calculate upstream installation properties on-demand themselves rather than depending on concretization to set these properties up-front. The added tests for upstream installations don't work with this new strategy so they need to be updated
* only refresh modules for local specs (not those in upstream packages); optionally generate local module files for packages installed upstream
* when a user tries to locate a module file for a package installed upstream, tell them to use the upstream spack instance to locate it
* support recursive upstream databases (allow upstream databases to use their own upstream databases)
* separate upstream config into separate file with its own schema; each entry now also includes a name
* metadata_dir is no longer customizable on a per-instance basis for YamlDirectoryLayout
* treat metadata_dir as an instance variable but dont set it from kwargs; this follows several other hardcoded variables which must be consistent between upstream and downstream DBs. Also update DirectoryLayout.metadata_path to work entirely with Spec.prefix, since Spec.prefix is set from the DB when available (so metadata_path was duplicating that logic)
* Add latest libyogrt versions
* Add url_for_version to handle differing url formats for different
versions
* Add 'scheduler' variant to configure scheduler. Currently only
slurm can be installed with Spack. A value of 'system' indicates
to use the scheduler installed on the system
* py-prompt-toolkit: Add newer version for required for newer
py-ipython versions
* py-ipykernel:
* Add newer py-ipykernel version
* Depend on py-setuptools so a python egg isn't attempted to be
built
* Update dependency for newer py-ipykernel versions
* py-jupyter-console: Remove py-prompt-toolkit dependency since it is
picked up in py-ipython
* py-ipython:
* Add missing py-backcall dependency
* Adjust py-prompt-toolkit dependencies for newer versions of
ipython
* py-jupyter-notebook: Require newer version of py-ipykernel since
jupyter is broken with previous versions
* add cxxstd variant
* add CMake constraints based on platform/version
* add older versions
* update boost dependency version constraints (which are closely
tied to mysql version) and update boost cxxstd choice to be
the same as the cxxstd chosen for mysql
* add client-only support (including a patch for 5.5.x)
* record the mysql package as a provider of the mysql-client virtual
Improve management of the Fiber library and C++ standard support:
* Remove Fiber from list of libraries to build
* Improve variant management for Fiber; add variants for Context and
Coroutine libraries.
* Add known conflict with C++17 for boost < 1.63.0
* Remove C++ standard "default" option, which left the choice of
C++ standard to the compiler used to build boost
* record conflicts with compiler versions which don't provide
required c++ standard support
* add doxygen (build) dependency
* add note that range-v3 is header-only as of 0.3.6 and update
package description
We add new variants to handle readline vs libedit, a client-only
build and install, and bindings to TCL, Python and Perl. We also add
new versions and the ability to detect remote versions not otherwise
dealt with.
This avoids using a system-installed CUDA package. In the future a
variant can be added to allow using Spack-installed CUDA, but for
now CUDA support is always disabled.
* abinit: Fix building with hdf5/netcdf.
* gromacs: Fix attempt to build with cuda support when 'cuda=False'
If for some reason there's a cuda toolkit installed by other means,
(i.e. not by spack) cmake will still try to build with cuda support,
even though 'cuda=False' is the default of the spec.
* Revert "abinit: Fix building with hdf5/netcdf."
This reverts commit e16f725e37b91193fe519b1821446c76ab551928.
This should not be here.
* Draco: add variants
+ This package has many optional build dependencies that were not registered in
older versions of this recipe. I've added (and tested) this more complete
list of optional dependencies: parmetis, superlu-dist, qt.
* fix style issues
* hpcviewer: new package
Add binary package for hpcviewer and hpctraceviewer for the Rice
hpctoolkit on Linux x86_64, ppc64 and ppc64le.
* ibm-java: add property 'home' so that spec['java'].home will work.
* Flake
* More flake.
* Test that the version, machine type pair exists before using it in the
sha dictionaries so that 'spack info' doesn't crash on unsupported
configurations.
* mariadb-c-client is a new package using the distinct, LGPL,
MySQL-compatible client library from mariadb.com. It provides the
virtual package mariadb-client
* mariadb is recorded as a provider of the mariadb-client virtual
* The mysql-client virtual package is also added, and mariadb-c-client
is recorded as a provider for it
* Throw InstallError if more than one GPU architecture is passed to cuda_arch. Previous cuda_arch test was not actually working because comparison with none string was on the cuda_arch list instead of the first entry of the list.
* Removing redundant cuda_arch statement.
* New package: py-mysql-connector
* Fixed docstring
* 1. Determined that py-setuptools was not needed at all, so removed.
2. Added py-protobuf. Docs seem to imply that only C protobuf library is required; however, the Python setup.py says differently, and some Python code seems to reference protobuf too. I don't know why this worked for me, but it looks like including py-protobuf is the right thing to do.
* Applied solution detailed in:
https://github.com/mysql/mysql-connector-python/pull/9
Uncommenting this patch will make `error: option --single-version-externally-managed not recognized` reappear.
* Clean up / reorder lines
* flake8
qt currently falls back to bundled versions of sqlite, harfbuzz, pcre,
double-conversion and xcb. This adds the appropriate dependencies and
configure arguments. A new variant adds multibyte support to pcre and
pcre2, which is required by qt.
Additionally, newer versions of gcc (starting with @8.3.0) cause build
failures. This adds a patch to fix the problem.
The changes have been tested with all versions of qt currently available
in Spack. 5.2 and 5.3 do not build for reasons that seem to be unrelated
to these changes, though.
* Add binary package of the IBM Java SDK for big and little-endian
powerpc (power7, 8 and 9). The jdk and openjdk packages only install
on x86_64.
* Add ibm-java as a java provider
* The jdk and openjdk packages only install on x86_64. Add conflicts
for ppc64 and ppc64le to jdk and openjdk.
shmemrun and oshrun do not exist in OpenMPI v4.0.0
(ref: https://www.open-mpi.org/doc/v4.0/)
The Spack OpenMPI package was failing the install by trying to
remove them. This guards the removal of several scripts when
using the Slurm scheduler to handle the case where they don't exist.
It seems that this is actually a glibc problem and while 2.6.4 builds
without the patch on newer versions of gcc (@8:), it still sometimes
segfaults (as observed during the doxygen build).
* Add 'fiber' as a default library for boost
* Add autoconf/automake etc. dependencies to libseccomp package
* New package: brotli
* New package: editline
* Add brotli, editline, boost dependencies to Nix
Remove 2.6.3 as preferred version (but keep it available for
building). The latest version (currently 2.6.4) is now preferred
(according to Spack's defaults).
* Update dependencies for py-flake8 when building version 3.7.7
* Add FIXME comment for an example dependency constraint which causes
concretization to hang
* Add py-entrypoints version 0.3
* Add py-pycodestyle version 2.5.0
* Add libuv version 1.10.0
* CMake versions before 3.12.0 do not build with libuv version
1.25.0, so a constraint is added to build earlier versions of
CMake with libuv version 1.10.x
Update CPATH in setup_environment for Eigen, so that the
Spack-generated module for Eigen will help builds outside of Spack
use the appropriate include prefix for Eigen headers
(<install_prefix>/include/eigen3/ rather than <install_prefix>/include/)
Note that this only updates the run-time environment, rather than the
build-time environment, so Spack builds depending on Eigen that use
pkgconfig will not be confused by the presence of the Eigen include
directory in CPATH.
* Replace kim-api package with kim-api-v2, which has different
versions and removes the 'cmake_args' method
* Add openkim-models-v2 as an extension package
* libelf: move url to fossies.org
Libelf is no longer maintained and the original mr511.de web site no
longer exists. The final release, 0.8.13 from Nov 2009 is still
archived at fossies.org.
Fixes#10757.
* Change the homepage to the FSF entry.
New release of UnifyCR. Remove unsupported pre-1.0 versions
UnifyCR has had a few dependency changes to make the backend more
efficient. A new pre-release version has been released as well to
mark these changes. Prior versions are no longer supported.
Fixes#10769
This updates the .headers property to include header subdirectories
for Python and Eigen (as is recommended by these packages).
#10623 updated the default behavior of .headers.directories to
exclude subdirectories (since this can cause clashes with system
headers). This broke some packages which depended on the old behavior
of .headers.directories: for example if you had
<package-prefix>/include/subdir/ex1.h, .headers.directories would
include <package-prefix>/include/subdir.
The environment modules package has been updated to include
versions up to 4.0.0. The url of the package and the homepage
have been updated accordingly.
The `spack bootstrap` command now builds version 3.2.10 of
the environment-modules package, and will do until #10708
is fixed.
* Make use of new QE multi-valued HDF5 variant interface
* Remove explicit CUDA dependency (which is handled by inheriting
from CudaPackage)
* Update cuda_arch unset variant check
libbeagle compiles against CUDA by default but no there is no mention
of it in the package recipe. This PR adds explicit cuda paths and
variants, and fixes the target architecture as well (for those who
don't have compute_13)
Serial HDF5 in serial QE only works in develop version. This adds
a conflict to the QE package to prevent Spack from trying to build
specs which don't satisfy this constraint.
This restores the use of Package.headers when computing -I options
for building a package that was added in #8136 and reverted in
#10604. #8136 used utility logic that located all header files in
an installation prefix, and calculated the -I options as the
immediate roots containing those header files.
In some cases, for a package containing a directory structure like
prefix/
include/
ex1.h
subdir/
ex2.h
dependents may expect to include ex2.h relative to 'include', and
adding 'prefix/include/subdir' as a -I was causing errors,
in particular if ex2.h has the same name as a system header.
This updates header utility logic to by default return the base
"include" directory when it exists, rather than subdirectories.
It also makes it possible for package implementers to override
Package.headers to return the subdirectory when it is required
(for example with libxml2).
