diff --git a/python/dune/perftool/pdelab/basis.py b/python/dune/perftool/pdelab/basis.py
index 0c7ab18a63f12b9e870992aad88143d65229ef67..0f5b7f5ee5c815cf9b964bd8bd7ef36e2b553d18 100644
--- a/python/dune/perftool/pdelab/basis.py
+++ b/python/dune/perftool/pdelab/basis.py
@@ -170,14 +170,13 @@ def evaluate_coefficient(visitor, element, name, container, restriction, index):
 
     lfs = name_lfs(element, restriction, index)
     basis = visitor.interface.pymbolic_basis(sub_element, restriction, 0, context='trial')
-    basisindex, _ = get_pymbolic_indices(basis)
-
+    basisindex = get_pymbolic_indices(basis)[:-1]
     if get_option("blockstructured"):
         from dune.perftool.blockstructured.argument import pymbolic_coefficient
         coeff = pymbolic_coefficient(container, lfs, sub_element, basisindex)
     else:
         from dune.perftool.pdelab.argument import pymbolic_coefficient
-        coeff = pymbolic_coefficient(container, lfs, basisindex)
+        coeff = pymbolic_coefficient(container, lfs, basisindex[0])
 
     assignee = Variable(name)
 
@@ -207,7 +206,7 @@ def evaluate_coefficient_gradient(visitor, element, name, container, restriction
 
     lfs = name_lfs(element, restriction, index)
     basis = visitor.interface.pymbolic_reference_gradient(sub_element, restriction, 0, context='trialgrad')
-    basisindex, _ = get_pymbolic_indices(basis)
+    basisindex = get_pymbolic_indices(basis)[:-1]
     from dune.perftool.tools import maybe_wrap_subscript
     basis = maybe_wrap_subscript(basis, Variable(dimindex))
 
@@ -216,7 +215,7 @@ def evaluate_coefficient_gradient(visitor, element, name, container, restriction
         coeff = pymbolic_coefficient(container, lfs, sub_element, basisindex)
     else:
         from dune.perftool.pdelab.argument import pymbolic_coefficient
-        coeff = pymbolic_coefficient(container, lfs, basisindex)
+        coeff = pymbolic_coefficient(container, lfs, basisindex[0])
 
     assignee = Subscript(Variable(name), (Variable(dimindex),))
 
diff --git a/python/dune/perftool/pdelab/localoperator.py b/python/dune/perftool/pdelab/localoperator.py
index 23c92da10c0945b6e4196b1ccde44f0cdbdb5c83..64a96ed3ff5c943a38edc791fc7a5925f05899cc 100644
--- a/python/dune/perftool/pdelab/localoperator.py
+++ b/python/dune/perftool/pdelab/localoperator.py
@@ -247,7 +247,7 @@ def determine_accumulation_space(info, number):
     if get_option("blockstructured"):
         from dune.perftool.blockstructured.tools import micro_index_to_macro_index
         from dune.perftool.blockstructured.spaces import lfs_inames
-        lfsi = micro_index_to_macro_index(subel, lfs_inames(subel, info.restriction, count=number)[0])
+        lfsi = micro_index_to_macro_index(subel, lfs_inames(subel, info.restriction, count=number))
     else:
         from dune.perftool.pdelab.spaces import lfs_inames
         lfsi = Variable(lfs_iname(subel, info.restriction, count=number))
@@ -551,6 +551,8 @@ def extract_kernel_from_cache(tag, wrap_in_cgen=True):
     from dune.perftool.loopy import heuristic_duplication
     kernel = heuristic_duplication(kernel)
 
+    from dune.perftool.blockstructured.vectorization import vectorize_micro_elements
+    kernel = vectorize_micro_elements(kernel)
     # Maybe apply vectorization strategies
     if get_option("vectorization_quadloop"):
         if get_option("sumfact"):
@@ -569,6 +571,7 @@ def extract_kernel_from_cache(tag, wrap_in_cgen=True):
     # Do the loopy preprocessing!
     kernel = preprocess_kernel(kernel)
 
+
     # *REALLY* ignore boostability. This is - so far - necessary due to a mystery bug.
     kernel = kernel.copy(instructions=[i.copy(boostable=False, boostable_into=frozenset()) for i in kernel.instructions])