Feature/language

subrepo:
  subdir:   "packages/kokkos"
  merged:   "953d7968e"
upstream:
  origin:   "git@github.com:kokkos/kokkos.git"
  branch:   "master"
  commit:   "953d7968e"
git-subrepo:
  version:  "0.4.1"
  origin:   "git@github.com:ingydotnet/git-subrepo.git"
  commit:   "a04d8c2"

Separate the notion of Nlr (Eucclhyd normals) of the calculation of
Cjr in 3D. It remain to implement its calculation in 2D to finish this
part.

This should allow a trivial implementation of the Eucclhyd scheme and
simplify coding off pressure boundary conditions.

This impacts the use of AcousticSolver since updated mesh is now
returned by AcousticSolver::computeNextStep

Note that the code of AcousticSolver is now quite crappy and will be
reworked

Closes #9

thanks to clang 10

The aim of this singleton is to store (and eventually compute) various
data associated to a given mesh: volumes, corner vectors, centroids,
isobarycenters.

These data will be stored as long as the mesh lives and will be
computed on demand

This will also allow to define really constant meshes (removing the
remaining mutable node coordinates ...)

Actually, classes are created but files remain to be split

Extracts dimension 1 treatment to  improve readability

- code needs a lot of clean-up

Note that VTK output does not support diamond cells and only deals
with pyramids based on quadrangles

- ItemValueUtils: min/max/sum functions require improvement or fixes
- Diamond meshes are built correctly in 2D and sequential
  - faces boundaries references lists are correctly translated
  - node boundaries (ref lists) are missing
  - 1D and 3D are to be done

This allows to free memory that is no more needed

catch of runtime error was not tested

Remove some unreachable code and add missing treatment for bool-tuples

Remove type cases since any argument can be converted to a string

parallel::allReduceMax, parallel::allReduceMin, parallel::allReduceSum
can no longer be called on `bool` data, this is checked at compile
time.

These functions would fail at runtime (according to MPI standard)

Work in progress [ci skip]

Add tests for
- ASTNodeAffectationExpressionBuilder and
- ASTNodeNaturalConversionChecker

Also fix a minor issue where tuple(N) could be built instead of
tuple(N) and vice-versa.

Finally rename `is_vector` traits to `is_std_vector` for readability

Backtrace is now only produced when '-p' option is provided

This gives better user information and allows to debug more simply
since exceptions are no more rethrown

Going further would require to be able to indicate the current line of
execution during a non exception crash

Also add missing unit tests for ASTNodeDatatypeBuilder

Tuples are only defined for simple types.
This means that defining tuples like

let t:(R*R);

is forbidden. This may change in future, but it does not seem a
limitations and probably leads to simpler syntax.

Now, when the used type is not simple, it is clearly indicated.

- bunch of fixes and simplifications
- add unit tests

Thanks to Pascal Havé (@hpwxf) for reporting it

- simplify a lot (mainly construction/usage)
- add many more checks (during compilation and execution)

Add missing CI tests

These functions set initial density, velocity and pressure

These sets are defined for 2D and 3D meshes

Previously one required that corners a and b were defined such that
a[i]<b[j]. It is no more the case.

Also check that `a[i]-b[j] != 0`. This avoids crashes which would be
difficult to track.

The mesh is defined by two opposite corners and the number of cells in
each direction.

This contributes to #7

- do not use anymore gcc extension for __VA_ARGS__

- check that parameters are always correct (string literal could be
unchecked under NDEBUG)

- when using NDEBUG, assert exception cannot be evaluated. Previously
result was not used but non trivial assertions such as
  `container.find(key) != container.end()`
could be evaluated. More specifically functions with potential
boundary effects.

- also, this allows to compile the code with the `-pedantic` flag.

This was buggy since used gmsh meshes were incorrectly oriented ...
Should check cells volume

Also fix tuple arguments when a single value is passed (error was
related to bad casting)

Content data type was altered when a tuple was defined.
Data type must not change!

An assertion has also been added to avoid this kind of problems.

It purposely diminish the ambiguity with the dataTypeName() function.

(a,b) or (c,d,e) are no longer interpreted as TinyVector<2> or
TinyVector<3>. They are only converted when the context is
appropriate.

This should have been done much earlier.
It consists in defining simply an operator<<

Also one can now define tuples (which are viewed in the mathematical
sense: these are lists of elements of the same type).

As a proof of concept one can now define lists of boundary conditions
objects or pass lists to builtin functions (namely to module
functions).

The idea is that tuples might not always convert to R^d.
This will allow to define generic tuples and eventually to deal with sets.

This is just a small step
- only symmetry are supported (btw theses are the only implemented in
the code ...)
- one requires 6 boundary conditions to be passed to the `glace` function

It is quite obvious that a bunch of checks is required

Also split BoudaryConditionDescriptor and BoundaryDescriptor files

It is a first step: it reproduces Pug's main. The only available
parameter is the mesh.

AST related files are now in the 'language/ast' directory

Displace AST*Printer and BuiltinFunctionEmbedder to 'utils'

Now modules are defined in their own directory (language/modules)

Compilation: function descriptor name was not named
Test failure: remove useless assertion which broke exception catching tests

- code is now generic (even if more testing remains).
- data types are now check accordingly to catch error in a clean way

In some cases (generally?) data types can only be checked at
runtime. This is the case for instance if function definition depends
on the dimension (which is likely to happen in many cases). Thus a
special procedure has been developed to indicate where the semantic
error occurs at runtime. (this is in some way related to issue #19)

- Rename FunctionAdapter -> PugsFunctionAdapter to improve readability
- PugsFunctionAdapter is now defined in its own file