diff --git a/boxtree/tools.py b/boxtree/tools.py
index 73039bd3c3adf4020cf20c4dde8246692220e9a1..ee472de002506b6b9e9d8c34a00ee9a0e6ffd4ad 100644
--- a/boxtree/tools.py
+++ b/boxtree/tools.py
@@ -27,7 +27,7 @@ import numpy as np
 from pytools import Record, memoize_method
 import pyopencl as cl
 import pyopencl.array  # noqa
-from pyopencl.tools import first_arg_dependent_memoize_nested, dtype_to_c_struct
+from pyopencl.tools import dtype_to_c_struct
 from mako.template import Template
 from pytools.obj_array import make_obj_array
 
@@ -106,8 +106,7 @@ def make_surface_particle_array(queue, nparticles, dims, dtype, seed=15):
     import loopy as lp
 
     if dims == 2:
-        @first_arg_dependent_memoize_nested
-        def get_2d_knl(context, dtype):
+        def get_2d_knl(dtype):
             knl = lp.make_kernel(
                 "{[i]: 0<=i<n}",
                 """
@@ -123,9 +122,9 @@ def make_surface_particle_array(queue, nparticles, dims, dtype, seed=15):
 
             knl = lp.split_iname(knl, "i", 128, outer_tag="g.0", inner_tag="l.0")
 
-            return lp.CompiledKernel(context, knl)
+            return knl
 
-        evt, result = get_2d_knl(queue.context, dtype)(queue, n=nparticles)
+        evt, result = get_2d_knl(dtype)(queue, n=nparticles)
 
         result = [x.ravel() for x in result]
 
@@ -133,8 +132,7 @@ def make_surface_particle_array(queue, nparticles, dims, dtype, seed=15):
     elif dims == 3:
         n = int(nparticles**0.5)
 
-        @first_arg_dependent_memoize_nested
-        def get_3d_knl(context, dtype):
+        def get_3d_knl(dtype):
             knl = lp.make_kernel(
                 "{[i,j]: 0<=i,j<n}",
                 """
@@ -152,9 +150,9 @@ def make_surface_particle_array(queue, nparticles, dims, dtype, seed=15):
             knl = lp.split_iname(knl, "i", 16, outer_tag="g.1", inner_tag="l.1")
             knl = lp.split_iname(knl, "j", 16, outer_tag="g.0", inner_tag="l.0")
 
-            return lp.CompiledKernel(context, knl)
+            return knl
 
-        evt, result = get_3d_knl(queue.context, dtype)(queue, n=n)
+        evt, result = get_3d_knl(dtype)(queue, n=n)
 
         result = [x.ravel() for x in result]
 
@@ -169,8 +167,7 @@ def make_uniform_particle_array(queue, nparticles, dims, dtype, seed=15):
     if dims == 2:
         n = int(nparticles**0.5)
 
-        @first_arg_dependent_memoize_nested
-        def get_2d_knl(context, dtype):
+        def get_2d_knl(dtype):
             knl = lp.make_kernel(
                 "{[i,j]: 0<=i,j<n}",
                 """
@@ -190,9 +187,9 @@ def make_uniform_particle_array(queue, nparticles, dims, dtype, seed=15):
             knl = lp.split_iname(knl, "i", 16, outer_tag="g.1", inner_tag="l.1")
             knl = lp.split_iname(knl, "j", 16, outer_tag="g.0", inner_tag="l.0")
 
-            return lp.CompiledKernel(context, knl)
+            return knl
 
-        evt, result = get_2d_knl(queue.context, dtype)(queue, n=n)
+        evt, result = get_2d_knl(dtype)(queue, n=n)
 
         result = [x.ravel() for x in result]
 
@@ -200,8 +197,7 @@ def make_uniform_particle_array(queue, nparticles, dims, dtype, seed=15):
     elif dims == 3:
         n = int(nparticles**(1/3))
 
-        @first_arg_dependent_memoize_nested
-        def get_3d_knl(context, dtype):
+        def get_3d_knl(dtype):
             knl = lp.make_kernel(
                 "{[i,j,k]: 0<=i,j,k<n}",
                 """
@@ -233,9 +229,9 @@ def make_uniform_particle_array(queue, nparticles, dims, dtype, seed=15):
             knl = lp.split_iname(knl, "j", 16, outer_tag="g.1", inner_tag="l.1")
             knl = lp.split_iname(knl, "k", 16, outer_tag="g.0", inner_tag="l.0")
 
-            return lp.CompiledKernel(context, knl)
+            return knl
 
-        evt, result = get_3d_knl(queue.context, dtype)(queue, n=n)
+        evt, result = get_3d_knl(dtype)(queue, n=n)
 
         result = [x.ravel() for x in result]
 
diff --git a/boxtree/tree_build.py b/boxtree/tree_build.py
index 7a014cdd27c3dd4b274e0d60d7a5411bcb260e5c..de2140a141b06562117449cd513a2d42f5e87893 100644
--- a/boxtree/tree_build.py
+++ b/boxtree/tree_build.py
@@ -744,10 +744,12 @@ class TreeBuilder(object):
 
                 split_box_ids = my_realloc_nocopy(split_box_ids)
 
-                # FIXME: This can be reused across iterations, saving the cost
-                # of the particle scan. Right now there isn't a reallocator
-                # written for it.
-                box_morton_bin_counts, evt = my_realloc_zeros_nocopy(
+                # *Most*, but not *all* of the values in this array are
+                # rewritten when the morton scan is redone. Specifically,
+                # only the box morton bin counts of boxes on the level
+                # currently being processed are written-but we need to
+                # retain the box morton bin counts from the higher levels.
+                box_morton_bin_counts, evt = my_realloc_zeros(
                         box_morton_bin_counts)
                 resize_events.append(evt)
 
@@ -1030,6 +1032,48 @@ class TreeBuilder(object):
             level += 1
             have_oversize_split_box.fill(0)
 
+            # {{{ check that nonchild part of box_morton_bin_counts is consistent
+
+            if debug and 0:
+                h_box_morton_bin_counts = box_morton_bin_counts.get()
+                h_box_srcntgt_counts_cumul = box_srcntgt_counts_cumul.get()
+                h_box_child_ids = tuple(bci.get() for bci in box_child_ids)
+
+                has_mismatch = False
+                for ibox in range(level_start_box_nrs[-1]):
+                    is_leaf = all(bci[ibox] == 0 for bci in h_box_child_ids)
+                    if is_leaf:
+                        # nonchild count only found in box_info kernel
+                        continue
+
+                    if h_box_srcntgt_counts_cumul[ibox] == 0:
+                        # empty boxes don't have box_morton_bin_counts written
+                        continue
+
+                    kid_sum = sum(
+                            h_box_srcntgt_counts_cumul[bci[ibox]]
+                            for bci in h_box_child_ids
+                            if bci[ibox] != 0)
+
+                    if (
+                            h_box_srcntgt_counts_cumul[ibox]
+                            !=
+                            h_box_morton_bin_counts[ibox]["nonchild_srcntgts"]
+                            + kid_sum):
+                        print("MISMATCH", level, ibox)
+                        has_mismatch = True
+
+                assert not has_mismatch
+                print("LEVEL %d OK" % level)
+
+                # Cannot delete in Py 2.7: referred to from nested scope.
+                h_box_srcntgt_counts_cumul = None
+
+                del h_box_morton_bin_counts
+                del h_box_child_ids
+
+            # }}}
+
         end_time = time()
         elapsed = end_time-start_time
         npasses = level+1
@@ -1065,8 +1109,16 @@ class TreeBuilder(object):
             del highest_possibly_split_box_nr
 
             if debug:
-                assert (box_srcntgt_counts_nonchild.get()
-                        <= box_srcntgt_counts_cumul.get()[:nboxes]).all()
+                h_box_srcntgt_counts_nonchild = box_srcntgt_counts_nonchild.get()
+                h_box_srcntgt_counts_cumul = box_srcntgt_counts_cumul.get()
+
+                assert (h_box_srcntgt_counts_nonchild
+                        <= h_box_srcntgt_counts_cumul[:nboxes]).all()
+
+                del h_box_srcntgt_counts_nonchild
+
+                # Cannot delete in Py 2.7: referred to from nested scope.
+                h_box_srcntgt_counts_cumul = None
 
         # }}}
 
diff --git a/boxtree/tree_build_kernels.py b/boxtree/tree_build_kernels.py
index d59a9f6e1f42b9492c1f5cb3320dc5b8cc9440b8..b2468993689c75a8f0d8e336e89b3d4becbed051 100644
--- a/boxtree/tree_build_kernels.py
+++ b/boxtree/tree_build_kernels.py
@@ -1031,9 +1031,6 @@ SOURCE_AND_TARGET_INDEX_FINDER = ElementwiseTemplate(
                     target_nr + 1 - (particle_id_t) is_source
                     - box_start_target_nr;
             }
-        %elif srcntgts_have_extent:
-            box_source_counts_nonchild[box_id] = 0;
-            box_target_counts_nonchild[box_id] = 0;
         %endif
 
         // {{{ last particle for this or the parents' boxes? update counts
diff --git a/test/test_tree.py b/test/test_tree.py
index 0857787724a13cc43535d41cba7b713c0c6c16f5..371e500fe1412ec7405c8196f8464edc9e295707 100644
--- a/test/test_tree.py
+++ b/test/test_tree.py
@@ -462,6 +462,9 @@ def test_extent_tree(ctx_getter, dims, do_plot=False):
     targets = make_normal_particle_array(queue, ntargets, dims, dtype,
             seed=19)
 
+    refine_weights = cl.array.zeros(queue, nsources+ntargets, np.int32)
+    refine_weights[:nsources] = 1
+
     from pyopencl.clrandom import PhiloxGenerator
     rng = PhiloxGenerator(queue.context, seed=13)
     source_radii = 2**rng.uniform(queue, nsources, dtype=dtype,
@@ -474,13 +477,38 @@ def test_extent_tree(ctx_getter, dims, do_plot=False):
 
     queue.finish()
     dev_tree, _ = tb(queue, sources, targets=targets,
-            source_radii=source_radii, target_radii=target_radii,
-            max_particles_in_box=10, debug=True)
+            source_radii=source_radii,
+            target_radii=target_radii,
+
+            refine_weights=refine_weights,
+            max_leaf_refine_weight=20,
+
+            #max_particles_in_box=10,
+
+            # Set artificially small, to exercise the reallocation code.
+            nboxes_guess=10,
+
+            debug=True,
+            stick_out_factor=0)
 
     logger.info("transfer tree, check orderings")
 
     tree = dev_tree.get(queue=queue)
 
+    if do_plot:
+        import matplotlib.pyplot as pt
+        pt.plot(sources[0].get(), sources[1].get(), "rx")
+        pt.plot(targets[0].get(), targets[1].get(), "g+")
+
+        from boxtree.visualization import TreePlotter
+        plotter = TreePlotter(tree)
+        plotter.draw_tree(fill=False, edgecolor="black", zorder=10)
+        plotter.draw_box_numbers()
+        plotter.set_bounding_box()
+
+        pt.gca().set_aspect("equal", "datalim")
+        pt.show()
+
     sorted_sources = np.array(list(tree.sources))
     sorted_targets = np.array(list(tree.targets))
     sorted_source_radii = tree.source_radii
@@ -490,6 +518,7 @@ def test_extent_tree(ctx_getter, dims, do_plot=False):
     unsorted_targets = np.array([pi.get() for pi in targets])
     unsorted_source_radii = source_radii.get()
     unsorted_target_radii = target_radii.get()
+
     assert (sorted_sources
             == unsorted_sources[:, tree.user_source_ids]).all()
     assert (sorted_source_radii
@@ -511,6 +540,20 @@ def test_extent_tree(ctx_getter, dims, do_plot=False):
 
     # {{{ check sources, targets
 
+    assert np.sum(tree.box_source_counts_nonchild) == nsources
+    assert np.sum(tree.box_target_counts_nonchild) == ntargets
+
+    for ibox in range(tree.nboxes):
+        kid_sum = sum(
+                    tree.box_target_counts_cumul[ichild_box]
+                    for ichild_box in tree.box_child_ids[:, ibox]
+                    if ichild_box != 0)
+        assert (
+                tree.box_target_counts_cumul[ibox]
+                ==
+                tree.box_target_counts_nonchild[ibox]
+                + kid_sum), ibox
+
     for ibox in range(tree.nboxes):
         extent_low, extent_high = tree.get_box_extent(ibox)
 
@@ -862,7 +905,10 @@ def test_level_restriction(ctx_getter, dims, skip_prune, lookbehind, do_plot=Fal
     queue.finish()
     tree_dev, _ = tb(queue, particles, kind="adaptive-level-restricted",
                      max_particles_in_box=30, debug=True,
-                     skip_prune=skip_prune, lr_lookbehind=lookbehind)
+                     skip_prune=skip_prune, lr_lookbehind=lookbehind,
+
+                     # Artificially low to exercise reallocation code
+                     nboxes_guess=10)
 
     def find_neighbors(leaf_box_centers, leaf_box_radii):
         # We use an area query with a ball that is slightly larger than