[!278] Enable extendability of the visitor for custom UFL nodes

Merge branch 'feature/allow-custom-ufl-types' into 'master'

ref:dominic/dune-perftool We do want code bloat in dune-perftool for the
specific needs of a given application. We therefore allow custom UFL nodes to
procide their own visitor implementation. This was done previously for the
very specific case of a MathFunction node, but it can be done with any node.

See merge request [dominic/dune-perftool!278]

  [dominic/dune-perftool!278]: gitlab.dune-project.org/dominic/dune-perftool/merge_requests/278

python/dune/perftool/ufl/visitor.py

@@ -77,7 +77,16 @@ class UFL2LoopyVisitor(ModifiedTerminalTracker):
 
         return self.call(o)
 
-    call = MultiFunction.__call__
+    def call(self, o):
+        # Allow downstream projects to provide custom nodes that provide their own visiting
+        # implementation. To do so, inherit from the UFL node which is closest to what you need
+        # and provide it with the following:
+        # * a _ufl_expr_reconstruct_ method to avoid preprocessing discarding the node
+        # * a visit method taking the visitor as the only argument
+        if hasattr(o, "visit"):
+            return o.visit(self)
+        else:
+            return MultiFunction.__call__(self, o)
 
     #
     # Form argument/coefficients handlers:
@@ -343,22 +352,6 @@ class UFL2LoopyVisitor(ModifiedTerminalTracker):
     def min_value(self, o):
         return self._minmax_impl(min, "min", tuple(self.call(op) for op in o.ufl_operands))
 
-    def math_function(self, o):
-        # MathFunction is a base class for unary functions. We use this to provide
-        # custom functions. Such a custom functions inherits from it and defines the
-        # following methods:
-        # * visit: This function is called from here to delegate the visiting process
-        #          to the user code. The only argument is this visitor instance.
-        # * derivative: It is called from UFL AD code to determine the derivative.
-        #               Upstream documentation indicates that FEniCS allows the same
-        #               (ab)use of the MathFunction node.
-        # Note that if the __init__ method of your function differs from MathFunction,
-        # you also need to implement the method _ufl_expr_reconstruct_
-        if hasattr(o, "visit"):
-            return o.visit(self)
-        else:
-            raise NotImplementedError("Function {} is not known to dune-perftool.".format(o._name))
-
     #
     # Handler for conditionals, use pymbolic base implementation
     #