in order to improve error messages further we introduce a new member
variable prefix_. Sometimes it is important to pass subtrees around.
In this case it is difficult to interpret error messages, as the user
doesn't know in which subtree the problem occured. Now a Parametertree
knows its prefix and can tell it to the user. This way the report
method is simplified and the throws are now more readable.

If an iterator facade (like entity iterators) wants to allow the
embedded implementation to return either an (internally stored)
reference or a temporary object and expose these two behaviors to enable
performance optimizations, operator->() needs special handling: If the
implementation returns a reference, operator->() in the facade can
simply return the address of the referenced object, but if the returned
object is a temporary, we need to capture and store it in a helper
object to make sure it outlives the member access.

This patch adds a little helper function that tansparently handles both
cases.

The ForLoop TMP used to put a number of restrictions on the arguments to
apply(), mainly due to the limited capabilities of C++01 wrt. to
argument forwarding.

But now we have C++11, so we can throw away all that awkward code and
replace it with a single forwarding function that supports arbitrary
numbers of arbitrary combinations of lvalues and rvalues - all hail
variadic templates and perfect forwarding!

If LAPACK is not available, it return 77 indicating skipped test

Like we do for Autotools.
This is only supported for CMake 3.0 or newer. Oder versions ignore
the property and report the test as failed.

g++-4.5 was not tested.
g++-4.6 is ok with lambdas.

These check whether std::call_once() works and provide a helpful error message
if it does not.  They should be used in any code that uses std::call_once().

Call once may not work if one forgets to _link_ with -pthread (or similar
options).  The nasty think about this is that linking still succeeds, so this
can only be detected at run time.  We cannot (in general) run compiled
programs during configure, since we may be cross-compiling, so whatever we get
from configure is at best a guess.

Even if configure detects the necessary flags correctly, there may still be
errors in the build system such that the executable is linked without them.

The reason to provide such a facility is that the bug is quite difficult to
debug when it appears.  The error message is mangled due to a different bug,
and is quite unhelpful anyway.  And then there is the fun with weak symbols...

This specialization was missing and made it impossible to parse bitsets.
The implementation uses the specialization for bool. It was not possible
to reuse the parseRange method, because a bitset doesnt have begin() and end().

While the following is legal c++11

  array<T,n> foo = { t1, t2, ...}

older compilers (namely gcc-4.4) issue a warning on missing braces.

This reverts commit d26e5ed3.
We have to wait until our minimal required compiler is GCC 4.7
or newer.
Thanks to Carsten for the heads-up.

The signature changed in C++11 which we require now.
To get more details see the mailinglost in October 2012.

It should be std::endl.
Courtesy of GCC 5-svn, it refuses to compile these.

GCC warns about it.

While gcc supports them (as extension) runtime sized arrays
are not standard c++. They are not even contained in c++14
but postponed to a later TS.

Add "@relatesalso null_deleter", this includes the documentation for the
function with null_deleter but also leaves a standalone version.  Document
what is so special about the shared_ptr the function returns.  Mention the
#include needed to get the function.

If it found an option, it now skips over the option's argument for further
processing.  Also, it throws an exception when the last command line argument
is an option (that option cannot have an argument).

"C", no matter what locale is currently set.

[parametertree][locale][fs1528] Check ParameterTree::get() with a locale using "," as the decimal seperator.

ParameterTrees are usually read from configuration files.  The format of the
configuration file should not depend on the locale.

For the commandline parser it would likewise be surprising if "progname -param
0.5" stopped working in a german locale.

Not long ago these kinds of errors would not even generate a diagnosis due to
FS#1527.  However, the locale of the program is "C" by default, so probably
not many poeple are actually affected by this problem.  It can happen however
that other libraries linked into the same program set the locale from the
environment.

"Makefile.in", "*.o", etc. need only be listed in the toplevel .gitignore, the
rules are applied recursively.

Test programs etc. should be listed as "/program" in their directory's
.gitignore, so they are not accidentially ignored in a lower level directory.

ParameterTree::get("key", def).

These overloads were using atoi() resp. atof(), effectively ignoring junk at
the end of values.  This led to the situation where a floating point value
"0.5" could be parsed as an int "0" without any diagnostic.

assert() works only in debug builds (without -DNDEBUG).  Instead, implement
our own check_assert().

I took the formula from

 http://en.wikipedia.org/wiki/Eigenvalue_algorithm  (retrieved lated August 2014).

Wikipedia claims to have taken it from

 Smith, Oliver K. (April 1961), Eigenvalues of a symmetric 3 × 3 matrix,
 Communications of the ACM 4 (4): 168, doi:10.1145/355578.366316

I did not check stability of efficiency of this, but no problems related to
either are mentioned in the literature.  The main reason I need the closed-form
expression is to be able to pipe it into an algorithmic differentiation system.

Rip out the complete test for eigenValuesNonSym and replace it by the Rosser matrix test from fieldmatrixtest.cc

The old test was complete nonsense: an implementation of some sort of strange
undocumented shape function set that happened to need eigenvalues for something.
Lots of undocumented code, and no checks at all whether the values obtained
by eigenValuesNonSym actually make sense.

Only, for the particular choice of parameters used, eigenValuesNonSym was never
called.  Yes, you read this correctly: these hundreds of lines of code didn't
actually test anything.

The new test is copied form fieldmatrixtest.cc and adapted for DynamicMatrix
types.  It computes eigenvalues of a given 8x8 matrix and compares with the known
results.  Unfortunately, the test matrix is only symmetric.  Tests with more
general matrices are left for another day.

Rather than assuming the caller has set it to the correct size already.
Setting the size in the method eigenValuesNonSym itself makes code
more robust.  Plus, the time needed for that should be negligible
compared to the actual eigenvalue computations.