Testet on mistral with intel/15.0.6 and
/sw/rhel6-x64/hdf5/hdf5-1.8.16-threadsafe-intel14

Nicolai Stange authored Jun 29, 2016 and Hendryk Bockelmann committed Oct 06, 2016

Automake is capable of handling the uninstallations of files it knows about
by itself and thus, the uninstall-local target in src/Makefile is
superfluous. Remove it.

Nicolai Stange authored Jun 29, 2016 and Hendryk Bockelmann committed Oct 06, 2016

CLEANFILES is way easier than writing custom clean-local hooks. Employ it.

Nicolai Stange authored Jun 29, 2016 and Hendryk Bockelmann committed Oct 06, 2016

sct.h is generated at build time. Add it to BUILT_SOURCES to ensure that
it gets built before anything else unconditionally.

Nicolai Stange authored Jun 29, 2016 and Hendryk Bockelmann committed Oct 06, 2016

The sct.h header is generated at build time and depends on the
configuration. Do not include it in 'make dist' by putting it into
the nodist_pkginclude_HEADERS variable.

Furthermore, remove the *_SOURCES dependencies on sct.h from the programs
in tests/ -- the Automake manual explicitly says that this doesn't work
(section 9.4.1 Built Sources Example) and they cause sct.h to be included
in the 'make dist' tarball.

Nicolai Stange authored Jun 29, 2016 and Hendryk Bockelmann committed Oct 06, 2016

Currently, the way src/sct.h is generated forbids out-of-tree builds.

Fix this by giving the util/gen_public_header.pl script full paths
to the source headers, i.e. prefix them by $(srcdir).

For the tests directory, add -I $(top_builddir)/src to its CPPFLAGS in
order to make the tests compilations find the sct.h places in the build
directory now.

Nicolai Stange authored Jun 29, 2016 and Hendryk Bockelmann committed Oct 06, 2016

Currently, configure constructs a $SETUPTYPE string based on the selected
configuration and appends that to installation prefix as given by --prefix.

This behaviour is not only unusual, but limits the user to completely
determine the final installation path.

Abandon it. Namely,
- purge the --with-mpitype option from configure,
- do not construct a $SETUPTYPE string,
- purge any reference to $SETUPTYPE from the Makefile.am's.

Additionally, in order to avoid name clashs in the case of
--enable-fakelib, install the "real" sct headers into $pkgincludedir, e.g.
$prefix/include/sct, and the fake ones into a directory named "nosct"
thereunder. Rename the fake libsct.la to libnosct.la.

The above changes render the install-all make target, which reconfigures
and rebuilds in all combinations of w/ and w/o OpenMP and MPI,
disfunctional. Functionality of that kind is better preserved in some
user defined wrapper scripts around the build process anyway and thus,
purge the install-all make target.

Finally, adapt the pkgconfig/sct.pc.in to properly reflect the changes.

Nicolai Stange authored Jun 29, 2016 and Hendryk Bockelmann committed Oct 06, 2016

While the common build generated files are excluded by the toplevel
.gitignore, the ones specific to the individual subdirectories are not.

Create a .gitignore per Makefile.am and include every build generated file
specific to that Makefile.am.

Nicolai Stange authored Jun 29, 2016 and Hendryk Bockelmann committed Oct 06, 2016

The current state of the upstream git repo is that most files created
by autoreconf are tracked, but not all: the Makefile.in's are missing
since commit d5c14d99 ("cleanup git repo for autotools usage").
Thus, the build system as distributed by the repo isn't in a workable
state.

It is common practice to not track autogenerated files in VCS and so,
instead of adding the missing Makefile.in's to the repo, purge everything
else recoverable by a simple run of autoreconf.

Namely,
- git rm everything stemming from autoreconf,
- add everything generated by autoreconf to .gitignore,
- remove the check for the autotools' version from autgen.sh -- version
  requirements are set in configure.ac already.

In addition, add some files created by the configure and build process to
.gitignore

Fixes: d5c14d99 ("cleanup git repo for autotools usage")

Automake is capable of handling the uninstallations of files it knows about
by itself and thus, the uninstall-local target in src/Makefile is
superfluous. Remove it.

CLEANFILES is way easier than writing custom clean-local hooks. Employ it.

sct.h is generated at build time. Add it to BUILT_SOURCES to ensure that
it gets built before anything else unconditionally.

The sct.h header is generated at build time and depends on the
configuration. Do not include it in 'make dist' by putting it into
the nodist_pkginclude_HEADERS variable.

Furthermore, remove the *_SOURCES dependencies on sct.h from the programs
in tests/ -- the Automake manual explicitly says that this doesn't work
(section 9.4.1 Built Sources Example) and they cause sct.h to be included
in the 'make dist' tarball.

Currently, the way src/sct.h is generated forbids out-of-tree builds.

Fix this by giving the util/gen_public_header.pl script full paths
to the source headers, i.e. prefix them by $(srcdir).

For the tests directory, add -I $(top_builddir)/src to its CPPFLAGS in
order to make the tests compilations find the sct.h places in the build
directory now.

Currently, configure constructs a $SETUPTYPE string based on the selected
configuration and appends that to installation prefix as given by --prefix.

This behaviour is not only unusual, but limits the user to completely
determine the final installation path.

Abandon it. Namely,
- purge the --with-mpitype option from configure,
- do not construct a $SETUPTYPE string,
- purge any reference to $SETUPTYPE from the Makefile.am's.

Additionally, in order to avoid name clashs in the case of
--enable-fakelib, install the "real" sct headers into $pkgincludedir, e.g.
$prefix/include/sct, and the fake ones into a directory named "nosct"
thereunder. Rename the fake libsct.la to libnosct.la.

The above changes render the install-all make target, which reconfigures
and rebuilds in all combinations of w/ and w/o OpenMP and MPI,
disfunctional. Functionality of that kind is better preserved in some
user defined wrapper scripts around the build process anyway and thus,
purge the install-all make target.

Finally, adapt the pkgconfig/sct.pc.in to properly reflect the changes.

While the common build generated files are excluded by the toplevel
.gitignore, the ones specific to the individual subdirectories are not.

Create a .gitignore per Makefile.am and include every build generated file
specific to that Makefile.am.

The current state of the upstream git repo is that most files created
by autoreconf are tracked, but not all: the Makefile.in's are missing
since commit d5c14d99 ("cleanup git repo for autotools usage").
Thus, the build system as distributed by the repo isn't in a workable
state.

It is common practice to not track autogenerated files in VCS and so,
instead of adding the missing Makefile.in's to the repo, purge everything
else recoverable by a simple run of autoreconf.

Namely,
- git rm everything stemming from autoreconf,
- add everything generated by autoreconf to .gitignore,
- remove the check for the autotools' version from autgen.sh -- version
  requirements are set in configure.ac already.

In addition, add some files created by the configure and build process to
.gitignore

Fixes: d5c14d99 ("cleanup git repo for autotools usage")

Consider the two timer nestings

  a => b => c

and

  a => c

In case they're encounted in the order presented, the ->active_under
of c will be set to -1 (root), because update_timer_stack_atstart()
only checks whether the current ->active_under is on the callstack.
After the first nesting has been processed, c->active_under == b and b is
not on the callstack of a => c.

The outcome is different if the two nestings are processed in reversed
order, i.e. a => c first, followed by a => b => c: the final
c->active_under will be a in this case.

Fix this by not only checking whether the current element's ->active_under
is on the callstack, but if any of its common ancestors, i.e. any member
of the whole list linked by ->active_under can be found.

In fact, do not check the whole callstack, but take advantage of its
invariants and traverse only the list linked by ->active_under rooted at
its topmost element.

Since the sct_stats_type instances' chains of common ancestors are linked
via ->active_under through their timer indicies,
update_timer_stack_atstart() will not only need access to the current
sct_stats_type instance, but to all of them in order to be able to walk
those chains upwards.

Instead of only passing the current sct_stats_type instance to
update_timer_stack_atstart(), give it a pointer to the array containing all
of them.