RubyGems - carray - Versions diffs - 3.0.4 → 3.0.6 - Mend

carray 3.0.4 → 3.0.6

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (155) hide show

checksums.yaml +4 -4
data/CHANGELOG.md +438 -0
data/README.md +1 -1
data/carray.gemspec +3 -2
data/ext/ca_array_pool.c +25 -1
data/ext/ca_axis_dispatch.c +19 -20
data/ext/ca_axis_group.c +1 -1
data/ext/ca_categorical_iterator.c +34 -27
data/ext/ca_compare.c +4 -4
data/ext/ca_composite_dispatch.c +9 -9
data/ext/ca_kernel_iterator.c +122 -237
data/ext/ca_kernel_iterator.h +522 -337
data/ext/ca_obj_array.c +69 -19
data/ext/ca_obj_bincmp.c +10 -1
data/ext/ca_obj_binop.c +11 -1
data/ext/ca_obj_bitarray.c +5 -1
data/ext/ca_obj_bitfield.c +60 -43
data/ext/ca_obj_byte_swap.c +4 -1
data/ext/ca_obj_const_string.c +15 -3
data/ext/ca_obj_face.c +44 -1
data/ext/ca_obj_fake.c +3 -1
data/ext/ca_obj_field.c +13 -3
data/ext/ca_obj_grid.c +219 -17
data/ext/ca_obj_meld.c +69 -113
data/ext/ca_obj_moncmp.c +9 -1
data/ext/ca_obj_monop.c +86 -12
data/ext/ca_obj_object.c +118 -32
data/ext/ca_obj_record.c +3 -0
data/ext/ca_obj_reduce.c +19 -1
data/ext/ca_obj_refer.c +80 -26
data/ext/ca_obj_remap.c +69 -10
data/ext/ca_obj_roll.c +27 -6
data/ext/ca_obj_select.c +38 -6
data/ext/ca_obj_select_axis.c +1 -0
data/ext/ca_obj_shift.c +32 -7
data/ext/ca_obj_stack.c +311 -111
data/ext/ca_obj_stride.c +159 -64
data/ext/ca_obj_string.c +27 -2
data/ext/ca_obj_tile.c +28 -7
data/ext/ca_obj_transpose.c +8 -4
data/ext/ca_obj_triop.c +12 -1
data/ext/ca_obj_window.c +80 -50
data/ext/ca_op_ipower.c +12 -12
data/ext/ca_sort_kernels.h +6 -3
data/ext/caframe_csv_reader.c +1 -1
data/ext/carray.h +35 -1
data/ext/carray_access.c +43 -67
data/ext/carray_address_basis.c +134 -0
data/ext/carray_attribute.c +32 -0
data/ext/carray_broadcast.c +3 -4
data/ext/carray_build_flags.h +1 -1
data/ext/carray_cast.c +251 -93
data/ext/carray_cast_func.rb +56 -26
data/ext/carray_conversion.c +11 -7
data/ext/carray_core.c +814 -72
data/ext/carray_count.c +6 -0
data/ext/carray_element.c +22 -13
data/ext/carray_factorize.c +20 -20
data/ext/carray_generate.c +8 -2
data/ext/carray_histogram.c +26 -4
data/ext/carray_index_classifier.c +4 -4
data/ext/carray_internal.h +159 -12
data/ext/carray_kernels_bincmp.c +2174 -170
data/ext/carray_kernels_binop.c +25653 -5487
data/ext/carray_kernels_init.c +8 -8
data/ext/carray_kernels_moncmp.c +96 -96
data/ext/carray_kernels_monop.c +12253 -5235
data/ext/carray_kernels_reduce_aggregate.c +6150 -2423
data/ext/carray_kernels_reduce_boolean.c +1533 -15
data/ext/carray_kernels_reduce_cumulative.c +2442 -882
data/ext/carray_kernels_reduce_extreme.c +5658 -1661
data/ext/carray_kernels_reduce_variance.c +2534 -402
data/ext/carray_kernels_scan.c +368 -368
data/ext/carray_kernels_search.c +4551 -604
data/ext/carray_kernels_sort.c +659 -751
data/ext/carray_kernels_triop.c +114 -114
data/ext/carray_key_windows.c +4 -0
data/ext/carray_lazy.c +47 -0
data/ext/carray_loop.c +1 -1
data/ext/carray_mask.c +175 -161
data/ext/carray_median_percentile.c +4 -4
data/ext/carray_memory_view.c +15 -9
data/ext/carray_operator.c +152 -49
data/ext/carray_order.c +230 -10
data/ext/carray_random.c +22 -19
data/ext/carray_scatter.c +24 -7
data/ext/carray_slab.c +128 -29
data/ext/carray_sort.c +224 -16
data/ext/carray_sort_kernel.c +5 -7
data/ext/carray_struct.c +35 -43
data/ext/carray_test.c +74 -19
data/ext/carray_time_text.c +6 -4
data/ext/carray_utils.c +63 -21
data/ext/extconf.rb +11 -0
data/ext/mkkernel.rb +546 -199
data/ext/ruby_carray.c +4 -1
data/ext/version.h +4 -4
data/lib/carray/attribute.rb +42 -5
data/lib/carray/autoload_carray.rb +3 -3
data/lib/carray/autoload_method_extension.rb +5 -3
data/lib/carray/axis_group.rb +13 -7
data/lib/carray/basics.rb +28 -8
data/lib/carray/block_iterator.rb +17 -3
data/lib/carray/categorical.rb +35 -23
data/lib/carray/categorical_iterator.rb +8 -8
data/lib/carray/complex.rb +2 -2
data/lib/carray/conditional.rb +3 -2
data/lib/carray/const_string.rb +10 -0
data/lib/carray/data_type_extension.rb +8 -1
data/lib/carray/fixlen_string.rb +41 -2
data/lib/carray/frame/column_verbs.rb +48 -16
data/lib/carray/frame/concat.rb +17 -3
data/lib/carray/frame/convert.rb +19 -0
data/lib/carray/frame/csv_parser.rb +32 -32
data/lib/carray/frame/describe.rb +4 -2
data/lib/carray/frame/frame.rb +405 -85
data/lib/carray/frame/group.rb +777 -64
data/lib/carray/frame/io.rb +212 -72
data/lib/carray/frame/join.rb +146 -37
data/lib/carray/frame/records.rb +3 -3
data/lib/carray/frame/reshape.rb +89 -11
data/lib/carray/frame/sort.rb +57 -9
data/lib/carray/frame/verbs.rb +384 -49
data/lib/carray/frame/windows.rb +44 -6
data/lib/carray/fuse_source.rb +31 -2
data/lib/carray/inspect.rb +7 -2
data/lib/carray/iterator.rb +32 -0
data/lib/carray/lazy.rb +52 -8
data/lib/carray/mask_gap_fill.rb +37 -17
data/lib/carray/meld_reduce.rb +1 -1
data/lib/carray/methods/choose.rb +3 -2
data/lib/carray/methods/is_in.rb +8 -1
data/lib/carray/methods/meshgrid.rb +3 -1
data/lib/carray/runtime.rb +47 -8
data/lib/carray/segment_iterator.rb +35 -16
data/lib/carray/serialize.rb +53 -19
data/lib/carray/stack.rb +33 -18
data/lib/carray/string.rb +8 -0
data/lib/carray/string_operation_extension.rb +13 -5
data/lib/carray/struct_builder.rb +5 -3
data/lib/carray/time.rb +112 -68
data/lib/carray/window_iterator.rb +16 -4
data/yard-stubs/ca_obj_array.rb +26 -1
data/yard-stubs/ca_obj_transpose.rb +5 -4
data/yard-stubs/carray_attribute.rb +3 -1
data/yard-stubs/carray_cast.rb +10 -5
data/yard-stubs/carray_conversion.rb +13 -3
data/yard-stubs/carray_element.rb +7 -6
data/yard-stubs/carray_mask.rb +28 -4
data/yard-stubs/carray_math.rb +45 -20
data/yard-stubs/carray_memory_view.rb +7 -0
data/yard-stubs/carray_order.rb +100 -4
data/yard-stubs/carray_scatter.rb +7 -5
data/yard-stubs/carray_test.rb +7 -4
metadata +18 -4

data/ext/carray_kernels_sort.c CHANGED Viewed

@@ -63,6 +63,13 @@ sort_index_ki_native_i8 (VALUE self, CArray *ca, int axis, int do_stable, int ma
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_index_pair_i8 *buf = ALLOCV_N(sort_index_pair_i8, hbuf, fiber_n);
+  sort_index_pair_i8 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(sort_index_pair_i8, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -77,14 +84,6 @@ sort_index_ki_native_i8 (VALUE self, CArray *ca, int axis, int do_stable, int ma
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_index_pair_i8 *buf = ALLOCV_N(sort_index_pair_i8, hbuf, fiber_n);
-  sort_index_pair_i8 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_index_pair_i8, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -101,13 +100,13 @@ sort_index_ki_native_i8 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -115,13 +114,13 @@ sort_index_ki_native_i8 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -129,7 +128,7 @@ sort_index_ki_native_i8 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -178,6 +177,13 @@ sort_index_ki_native_u8 (VALUE self, CArray *ca, int axis, int do_stable, int ma
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_index_pair_u8 *buf = ALLOCV_N(sort_index_pair_u8, hbuf, fiber_n);
+  sort_index_pair_u8 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(sort_index_pair_u8, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -192,14 +198,6 @@ sort_index_ki_native_u8 (VALUE self, CArray *ca, int axis, int do_stable, int ma
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_index_pair_u8 *buf = ALLOCV_N(sort_index_pair_u8, hbuf, fiber_n);
-  sort_index_pair_u8 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_index_pair_u8, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -216,13 +214,13 @@ sort_index_ki_native_u8 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -230,13 +228,13 @@ sort_index_ki_native_u8 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -244,7 +242,7 @@ sort_index_ki_native_u8 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -276,7 +274,7 @@ sort_index_ki_native_u8 (VALUE self, CArray *ca, int axis, int do_stable, int ma
 }
 typedef struct {
-  int16_t v;
+  ca_u_int16_t v;
   ca_size_t i;
 } sort_index_pair_i16;
@@ -293,6 +291,13 @@ sort_index_ki_native_i16 (VALUE self, CArray *ca, int axis, int do_stable, int m
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_index_pair_i16 *buf = ALLOCV_N(sort_index_pair_i16, hbuf, fiber_n);
+  sort_index_pair_i16 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(sort_index_pair_i16, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -307,14 +312,6 @@ sort_index_ki_native_i16 (VALUE self, CArray *ca, int axis, int do_stable, int m
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_index_pair_i16 *buf = ALLOCV_N(sort_index_pair_i16, hbuf, fiber_n);
-  sort_index_pair_i16 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_index_pair_i16, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -331,13 +328,13 @@ sort_index_ki_native_i16 (VALUE self, CArray *ca, int axis, int do_stable, int m
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -345,13 +342,13 @@ sort_index_ki_native_i16 (VALUE self, CArray *ca, int axis, int do_stable, int m
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -359,7 +356,7 @@ sort_index_ki_native_i16 (VALUE self, CArray *ca, int axis, int do_stable, int m
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -391,7 +388,7 @@ sort_index_ki_native_i16 (VALUE self, CArray *ca, int axis, int do_stable, int m
 }
 typedef struct {
-  uint16_t v;
+  ca_u_uint16_t v;
   ca_size_t i;
 } sort_index_pair_u16;
@@ -408,6 +405,13 @@ sort_index_ki_native_u16 (VALUE self, CArray *ca, int axis, int do_stable, int m
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_index_pair_u16 *buf = ALLOCV_N(sort_index_pair_u16, hbuf, fiber_n);
+  sort_index_pair_u16 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(sort_index_pair_u16, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -422,14 +426,6 @@ sort_index_ki_native_u16 (VALUE self, CArray *ca, int axis, int do_stable, int m
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_index_pair_u16 *buf = ALLOCV_N(sort_index_pair_u16, hbuf, fiber_n);
-  sort_index_pair_u16 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_index_pair_u16, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -446,13 +442,13 @@ sort_index_ki_native_u16 (VALUE self, CArray *ca, int axis, int do_stable, int m
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -460,13 +456,13 @@ sort_index_ki_native_u16 (VALUE self, CArray *ca, int axis, int do_stable, int m
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -474,7 +470,7 @@ sort_index_ki_native_u16 (VALUE self, CArray *ca, int axis, int do_stable, int m
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -506,7 +502,7 @@ sort_index_ki_native_u16 (VALUE self, CArray *ca, int axis, int do_stable, int m
 }
 typedef struct {
-  int32_t v;
+  ca_u_int32_t v;
   ca_size_t i;
 } sort_index_pair_i32;
@@ -523,6 +519,13 @@ sort_index_ki_native_i32 (VALUE self, CArray *ca, int axis, int do_stable, int m
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_index_pair_i32 *buf = ALLOCV_N(sort_index_pair_i32, hbuf, fiber_n);
+  sort_index_pair_i32 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(sort_index_pair_i32, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -537,14 +540,6 @@ sort_index_ki_native_i32 (VALUE self, CArray *ca, int axis, int do_stable, int m
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_index_pair_i32 *buf = ALLOCV_N(sort_index_pair_i32, hbuf, fiber_n);
-  sort_index_pair_i32 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_index_pair_i32, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -561,13 +556,13 @@ sort_index_ki_native_i32 (VALUE self, CArray *ca, int axis, int do_stable, int m
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -575,13 +570,13 @@ sort_index_ki_native_i32 (VALUE self, CArray *ca, int axis, int do_stable, int m
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -589,7 +584,7 @@ sort_index_ki_native_i32 (VALUE self, CArray *ca, int axis, int do_stable, int m
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -621,7 +616,7 @@ sort_index_ki_native_i32 (VALUE self, CArray *ca, int axis, int do_stable, int m
 }
 typedef struct {
-  uint32_t v;
+  ca_u_uint32_t v;
   ca_size_t i;
 } sort_index_pair_u32;
@@ -638,6 +633,13 @@ sort_index_ki_native_u32 (VALUE self, CArray *ca, int axis, int do_stable, int m
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_index_pair_u32 *buf = ALLOCV_N(sort_index_pair_u32, hbuf, fiber_n);
+  sort_index_pair_u32 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(sort_index_pair_u32, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -652,14 +654,6 @@ sort_index_ki_native_u32 (VALUE self, CArray *ca, int axis, int do_stable, int m
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_index_pair_u32 *buf = ALLOCV_N(sort_index_pair_u32, hbuf, fiber_n);
-  sort_index_pair_u32 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_index_pair_u32, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -676,13 +670,13 @@ sort_index_ki_native_u32 (VALUE self, CArray *ca, int axis, int do_stable, int m
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -690,13 +684,13 @@ sort_index_ki_native_u32 (VALUE self, CArray *ca, int axis, int do_stable, int m
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -704,7 +698,7 @@ sort_index_ki_native_u32 (VALUE self, CArray *ca, int axis, int do_stable, int m
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -736,7 +730,7 @@ sort_index_ki_native_u32 (VALUE self, CArray *ca, int axis, int do_stable, int m
 }
 typedef struct {
-  int64_t v;
+  ca_u_int64_t v;
   ca_size_t i;
 } sort_index_pair_i64;
@@ -753,6 +747,13 @@ sort_index_ki_native_i64 (VALUE self, CArray *ca, int axis, int do_stable, int m
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_index_pair_i64 *buf = ALLOCV_N(sort_index_pair_i64, hbuf, fiber_n);
+  sort_index_pair_i64 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(sort_index_pair_i64, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -767,14 +768,6 @@ sort_index_ki_native_i64 (VALUE self, CArray *ca, int axis, int do_stable, int m
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_index_pair_i64 *buf = ALLOCV_N(sort_index_pair_i64, hbuf, fiber_n);
-  sort_index_pair_i64 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_index_pair_i64, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -791,13 +784,13 @@ sort_index_ki_native_i64 (VALUE self, CArray *ca, int axis, int do_stable, int m
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -805,13 +798,13 @@ sort_index_ki_native_i64 (VALUE self, CArray *ca, int axis, int do_stable, int m
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -819,7 +812,7 @@ sort_index_ki_native_i64 (VALUE self, CArray *ca, int axis, int do_stable, int m
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -851,7 +844,7 @@ sort_index_ki_native_i64 (VALUE self, CArray *ca, int axis, int do_stable, int m
 }
 typedef struct {
-  uint64_t v;
+  ca_u_uint64_t v;
   ca_size_t i;
 } sort_index_pair_u64;
@@ -868,6 +861,13 @@ sort_index_ki_native_u64 (VALUE self, CArray *ca, int axis, int do_stable, int m
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_index_pair_u64 *buf = ALLOCV_N(sort_index_pair_u64, hbuf, fiber_n);
+  sort_index_pair_u64 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(sort_index_pair_u64, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -882,14 +882,6 @@ sort_index_ki_native_u64 (VALUE self, CArray *ca, int axis, int do_stable, int m
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_index_pair_u64 *buf = ALLOCV_N(sort_index_pair_u64, hbuf, fiber_n);
-  sort_index_pair_u64 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_index_pair_u64, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -906,13 +898,13 @@ sort_index_ki_native_u64 (VALUE self, CArray *ca, int axis, int do_stable, int m
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -920,13 +912,13 @@ sort_index_ki_native_u64 (VALUE self, CArray *ca, int axis, int do_stable, int m
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -934,7 +926,7 @@ sort_index_ki_native_u64 (VALUE self, CArray *ca, int axis, int do_stable, int m
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -966,7 +958,7 @@ sort_index_ki_native_u64 (VALUE self, CArray *ca, int axis, int do_stable, int m
 }
 typedef struct {
-  float v;
+  ca_u_float32_t v;
   ca_size_t i;
 } sort_index_pair_f32;
@@ -983,6 +975,10 @@ sort_index_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int m
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_index_pair_f32 *buf = ALLOCV_N(sort_index_pair_f32, hbuf, fiber_n);
+  sort_index_pair_f32 *aux = ALLOCV_N(sort_index_pair_f32, haux, fiber_n);
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -997,14 +993,6 @@ sort_index_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int m
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_index_pair_f32 *buf = ALLOCV_N(sort_index_pair_f32, hbuf, fiber_n);
-  sort_index_pair_f32 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_index_pair_f32, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -1021,13 +1009,13 @@ sort_index_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int m
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(float *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(float *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -1035,13 +1023,13 @@ sort_index_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int m
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(float *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(float *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -1049,7 +1037,7 @@ sort_index_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int m
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(float *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -1065,7 +1053,7 @@ sort_index_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int m
     /* P.3/P.4: NaN pre-partition for float argsort -- NaN pairs are
        pushed to the tail (= NaN-at-end policy, matches sort_copy
        AC8).  Only the finite slice is sorted. */
-    ca_size_t finite_n = ca_partition_nan_pair_f32((ca_pair_f32 *) (buf + sort_lo), (ca_size_t) sort_n);
+    ca_size_t finite_n = ca_partition_nan_pair_f32((ca_pair_f32 *) (buf + sort_lo), (ca_size_t) sort_n, (ca_pair_f32 *) aux);
     if ( do_stable ) {
       ca_sort_merge_pair_f32((ca_pair_f32 *) (buf + sort_lo), (ca_pair_f32 *) aux, finite_n);
     } else {
@@ -1085,7 +1073,7 @@ sort_index_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int m
 }
 typedef struct {
-  double v;
+  ca_u_float64_t v;
   ca_size_t i;
 } sort_index_pair_f64;
@@ -1102,6 +1090,10 @@ sort_index_ki_native_f64 (VALUE self, CArray *ca, int axis, int do_stable, int m
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_index_pair_f64 *buf = ALLOCV_N(sort_index_pair_f64, hbuf, fiber_n);
+  sort_index_pair_f64 *aux = ALLOCV_N(sort_index_pair_f64, haux, fiber_n);
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -1116,14 +1108,6 @@ sort_index_ki_native_f64 (VALUE self, CArray *ca, int axis, int do_stable, int m
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_index_pair_f64 *buf = ALLOCV_N(sort_index_pair_f64, hbuf, fiber_n);
-  sort_index_pair_f64 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_index_pair_f64, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -1140,13 +1124,13 @@ sort_index_ki_native_f64 (VALUE self, CArray *ca, int axis, int do_stable, int m
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(double *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(double *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -1154,13 +1138,13 @@ sort_index_ki_native_f64 (VALUE self, CArray *ca, int axis, int do_stable, int m
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(double *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(double *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -1168,7 +1152,7 @@ sort_index_ki_native_f64 (VALUE self, CArray *ca, int axis, int do_stable, int m
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(double *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -1184,7 +1168,7 @@ sort_index_ki_native_f64 (VALUE self, CArray *ca, int axis, int do_stable, int m
     /* P.3/P.4: NaN pre-partition for float argsort -- NaN pairs are
        pushed to the tail (= NaN-at-end policy, matches sort_copy
        AC8).  Only the finite slice is sorted. */
-    ca_size_t finite_n = ca_partition_nan_pair_f64((ca_pair_f64 *) (buf + sort_lo), (ca_size_t) sort_n);
+    ca_size_t finite_n = ca_partition_nan_pair_f64((ca_pair_f64 *) (buf + sort_lo), (ca_size_t) sort_n, (ca_pair_f64 *) aux);
     if ( do_stable ) {
       ca_sort_merge_pair_f64((ca_pair_f64 *) (buf + sort_lo), (ca_pair_f64 *) aux, finite_n);
     } else {
@@ -1221,6 +1205,13 @@ sort_index_ki_native_bool (VALUE self, CArray *ca, int axis, int do_stable, int
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_index_pair_bool *buf = ALLOCV_N(sort_index_pair_bool, hbuf, fiber_n);
+  sort_index_pair_bool *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(sort_index_pair_bool, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -1235,14 +1226,6 @@ sort_index_ki_native_bool (VALUE self, CArray *ca, int axis, int do_stable, int
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_index_pair_bool *buf = ALLOCV_N(sort_index_pair_bool, hbuf, fiber_n);
-  sort_index_pair_bool *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_index_pair_bool, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -1259,13 +1242,13 @@ sort_index_ki_native_bool (VALUE self, CArray *ca, int axis, int do_stable, int
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -1273,13 +1256,13 @@ sort_index_ki_native_bool (VALUE self, CArray *ca, int axis, int do_stable, int
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -1287,7 +1270,7 @@ sort_index_ki_native_bool (VALUE self, CArray *ca, int axis, int do_stable, int
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -1336,9 +1319,9 @@ sort_index_cmp_object (const void *a, const void *b)
   return 0;
 }
-/* The object lane calls Ruby for every cell, so a walk can raise part
-   way.  Its states live in the caller's frame, and ca_iter_ensure
-   finishes them however the body leaves. */
+/* The walks can raise part way (a Ruby call per cell in the object
+   lane, an allocation in any lane).  Their states live in the caller's
+   frame, and ca_iter_ensure finishes them however the body leaves. */
 typedef struct {
   ca_iter_state states[2];
   VALUE self;
@@ -1374,6 +1357,10 @@ sort_index_ki_native_object_body (VALUE arg)
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_index_pair_object *buf = ALLOCV_N(sort_index_pair_object, hbuf, fiber_n);
+  sort_index_pair_object *aux = ALLOCV_N(sort_index_pair_object, haux, fiber_n);
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
                              slab_axes, 1, 0);
@@ -1387,11 +1374,6 @@ sort_index_ki_native_object_body (VALUE arg)
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_index_pair_object *buf = ALLOCV_N(sort_index_pair_object, hbuf, fiber_n);
-  sort_index_pair_object *aux = ALLOCV_N(sort_index_pair_object, haux, fiber_n);
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -1408,13 +1390,13 @@ sort_index_ki_native_object_body (VALUE arg)
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -1422,13 +1404,13 @@ sort_index_ki_native_object_body (VALUE arg)
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -1436,7 +1418,7 @@ sort_index_ki_native_object_body (VALUE arg)
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -1475,7 +1457,7 @@ sort_index_ki_native_object (VALUE self, CArray *ca, int axis, int do_stable, in
   fr.axis = axis;
   fr.do_stable = do_stable;
   fr.masked_last = masked_last;
-  return ca_iter_ensure(2, fr.states, sort_index_ki_native_object_body, (VALUE) &fr);
+  return ca_iter_ensure_computing(ca, 2, fr.states, sort_index_ki_native_object_body, (VALUE) &fr);
 }
@@ -1511,6 +1493,10 @@ sort_index_ki_native_fixlen (VALUE self, CArray *ca, int axis, int do_stable, in
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_index_pair_fixlen *buf = ALLOCV_N(sort_index_pair_fixlen, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -1525,14 +1511,6 @@ sort_index_ki_native_fixlen (VALUE self, CArray *ca, int axis, int do_stable, in
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_index_pair_fixlen *buf = ALLOCV_N(sort_index_pair_fixlen, hbuf, fiber_n);
-  sort_index_pair_fixlen *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_index_pair_fixlen, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -1713,6 +1691,13 @@ sort_addr_ki_native_i8 (VALUE self, CArray *ca, int axis, int do_stable, int mas
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_addr_pair_i8 *buf = ALLOCV_N(sort_addr_pair_i8, hbuf, fiber_n);
+  sort_addr_pair_i8 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(sort_addr_pair_i8, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -1727,14 +1712,6 @@ sort_addr_ki_native_i8 (VALUE self, CArray *ca, int axis, int do_stable, int mas
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_addr_pair_i8 *buf = ALLOCV_N(sort_addr_pair_i8, hbuf, fiber_n);
-  sort_addr_pair_i8 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_addr_pair_i8, haux, fiber_n);
-  }
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -1778,13 +1755,13 @@ sort_addr_ki_native_i8 (VALUE self, CArray *ca, int axis, int do_stable, int mas
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -1792,13 +1769,13 @@ sort_addr_ki_native_i8 (VALUE self, CArray *ca, int axis, int do_stable, int mas
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -1806,7 +1783,7 @@ sort_addr_ki_native_i8 (VALUE self, CArray *ca, int axis, int do_stable, int mas
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -1860,6 +1837,13 @@ sort_addr_ki_native_u8 (VALUE self, CArray *ca, int axis, int do_stable, int mas
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_addr_pair_u8 *buf = ALLOCV_N(sort_addr_pair_u8, hbuf, fiber_n);
+  sort_addr_pair_u8 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(sort_addr_pair_u8, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -1874,14 +1858,6 @@ sort_addr_ki_native_u8 (VALUE self, CArray *ca, int axis, int do_stable, int mas
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_addr_pair_u8 *buf = ALLOCV_N(sort_addr_pair_u8, hbuf, fiber_n);
-  sort_addr_pair_u8 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_addr_pair_u8, haux, fiber_n);
-  }
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -1925,13 +1901,13 @@ sort_addr_ki_native_u8 (VALUE self, CArray *ca, int axis, int do_stable, int mas
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -1939,13 +1915,13 @@ sort_addr_ki_native_u8 (VALUE self, CArray *ca, int axis, int do_stable, int mas
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -1953,7 +1929,7 @@ sort_addr_ki_native_u8 (VALUE self, CArray *ca, int axis, int do_stable, int mas
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -1990,7 +1966,7 @@ sort_addr_ki_native_u8 (VALUE self, CArray *ca, int axis, int do_stable, int mas
 }
 typedef struct {
-  int16_t v;
+  ca_u_int16_t v;
   ca_size_t i;
 } sort_addr_pair_i16;
@@ -2007,6 +1983,13 @@ sort_addr_ki_native_i16 (VALUE self, CArray *ca, int axis, int do_stable, int ma
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_addr_pair_i16 *buf = ALLOCV_N(sort_addr_pair_i16, hbuf, fiber_n);
+  sort_addr_pair_i16 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(sort_addr_pair_i16, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -2021,14 +2004,6 @@ sort_addr_ki_native_i16 (VALUE self, CArray *ca, int axis, int do_stable, int ma
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_addr_pair_i16 *buf = ALLOCV_N(sort_addr_pair_i16, hbuf, fiber_n);
-  sort_addr_pair_i16 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_addr_pair_i16, haux, fiber_n);
-  }
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -2072,13 +2047,13 @@ sort_addr_ki_native_i16 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -2086,13 +2061,13 @@ sort_addr_ki_native_i16 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -2100,7 +2075,7 @@ sort_addr_ki_native_i16 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -2137,7 +2112,7 @@ sort_addr_ki_native_i16 (VALUE self, CArray *ca, int axis, int do_stable, int ma
 }
 typedef struct {
-  uint16_t v;
+  ca_u_uint16_t v;
   ca_size_t i;
 } sort_addr_pair_u16;
@@ -2154,6 +2129,13 @@ sort_addr_ki_native_u16 (VALUE self, CArray *ca, int axis, int do_stable, int ma
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_addr_pair_u16 *buf = ALLOCV_N(sort_addr_pair_u16, hbuf, fiber_n);
+  sort_addr_pair_u16 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(sort_addr_pair_u16, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -2168,14 +2150,6 @@ sort_addr_ki_native_u16 (VALUE self, CArray *ca, int axis, int do_stable, int ma
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_addr_pair_u16 *buf = ALLOCV_N(sort_addr_pair_u16, hbuf, fiber_n);
-  sort_addr_pair_u16 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_addr_pair_u16, haux, fiber_n);
-  }
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -2219,13 +2193,13 @@ sort_addr_ki_native_u16 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -2233,13 +2207,13 @@ sort_addr_ki_native_u16 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -2247,7 +2221,7 @@ sort_addr_ki_native_u16 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -2284,7 +2258,7 @@ sort_addr_ki_native_u16 (VALUE self, CArray *ca, int axis, int do_stable, int ma
 }
 typedef struct {
-  int32_t v;
+  ca_u_int32_t v;
   ca_size_t i;
 } sort_addr_pair_i32;
@@ -2301,6 +2275,13 @@ sort_addr_ki_native_i32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_addr_pair_i32 *buf = ALLOCV_N(sort_addr_pair_i32, hbuf, fiber_n);
+  sort_addr_pair_i32 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(sort_addr_pair_i32, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -2315,14 +2296,6 @@ sort_addr_ki_native_i32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_addr_pair_i32 *buf = ALLOCV_N(sort_addr_pair_i32, hbuf, fiber_n);
-  sort_addr_pair_i32 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_addr_pair_i32, haux, fiber_n);
-  }
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -2366,13 +2339,13 @@ sort_addr_ki_native_i32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -2380,13 +2353,13 @@ sort_addr_ki_native_i32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -2394,7 +2367,7 @@ sort_addr_ki_native_i32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -2431,7 +2404,7 @@ sort_addr_ki_native_i32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
 }
 typedef struct {
-  uint32_t v;
+  ca_u_uint32_t v;
   ca_size_t i;
 } sort_addr_pair_u32;
@@ -2448,6 +2421,13 @@ sort_addr_ki_native_u32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_addr_pair_u32 *buf = ALLOCV_N(sort_addr_pair_u32, hbuf, fiber_n);
+  sort_addr_pair_u32 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(sort_addr_pair_u32, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -2462,14 +2442,6 @@ sort_addr_ki_native_u32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_addr_pair_u32 *buf = ALLOCV_N(sort_addr_pair_u32, hbuf, fiber_n);
-  sort_addr_pair_u32 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_addr_pair_u32, haux, fiber_n);
-  }
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -2513,13 +2485,13 @@ sort_addr_ki_native_u32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -2527,13 +2499,13 @@ sort_addr_ki_native_u32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -2541,7 +2513,7 @@ sort_addr_ki_native_u32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -2578,7 +2550,7 @@ sort_addr_ki_native_u32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
 }
 typedef struct {
-  int64_t v;
+  ca_u_int64_t v;
   ca_size_t i;
 } sort_addr_pair_i64;
@@ -2595,6 +2567,13 @@ sort_addr_ki_native_i64 (VALUE self, CArray *ca, int axis, int do_stable, int ma
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_addr_pair_i64 *buf = ALLOCV_N(sort_addr_pair_i64, hbuf, fiber_n);
+  sort_addr_pair_i64 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(sort_addr_pair_i64, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -2609,14 +2588,6 @@ sort_addr_ki_native_i64 (VALUE self, CArray *ca, int axis, int do_stable, int ma
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_addr_pair_i64 *buf = ALLOCV_N(sort_addr_pair_i64, hbuf, fiber_n);
-  sort_addr_pair_i64 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_addr_pair_i64, haux, fiber_n);
-  }
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -2660,13 +2631,13 @@ sort_addr_ki_native_i64 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -2674,13 +2645,13 @@ sort_addr_ki_native_i64 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -2688,7 +2659,7 @@ sort_addr_ki_native_i64 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -2725,7 +2696,7 @@ sort_addr_ki_native_i64 (VALUE self, CArray *ca, int axis, int do_stable, int ma
 }
 typedef struct {
-  uint64_t v;
+  ca_u_uint64_t v;
   ca_size_t i;
 } sort_addr_pair_u64;
@@ -2742,6 +2713,13 @@ sort_addr_ki_native_u64 (VALUE self, CArray *ca, int axis, int do_stable, int ma
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_addr_pair_u64 *buf = ALLOCV_N(sort_addr_pair_u64, hbuf, fiber_n);
+  sort_addr_pair_u64 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(sort_addr_pair_u64, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -2756,14 +2734,6 @@ sort_addr_ki_native_u64 (VALUE self, CArray *ca, int axis, int do_stable, int ma
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_addr_pair_u64 *buf = ALLOCV_N(sort_addr_pair_u64, hbuf, fiber_n);
-  sort_addr_pair_u64 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_addr_pair_u64, haux, fiber_n);
-  }
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -2807,13 +2777,13 @@ sort_addr_ki_native_u64 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -2821,13 +2791,13 @@ sort_addr_ki_native_u64 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -2835,7 +2805,7 @@ sort_addr_ki_native_u64 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -2872,7 +2842,7 @@ sort_addr_ki_native_u64 (VALUE self, CArray *ca, int axis, int do_stable, int ma
 }
 typedef struct {
-  float v;
+  ca_u_float32_t v;
   ca_size_t i;
 } sort_addr_pair_f32;
@@ -2889,6 +2859,10 @@ sort_addr_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_addr_pair_f32 *buf = ALLOCV_N(sort_addr_pair_f32, hbuf, fiber_n);
+  sort_addr_pair_f32 *aux = ALLOCV_N(sort_addr_pair_f32, haux, fiber_n);
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -2903,14 +2877,6 @@ sort_addr_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_addr_pair_f32 *buf = ALLOCV_N(sort_addr_pair_f32, hbuf, fiber_n);
-  sort_addr_pair_f32 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_addr_pair_f32, haux, fiber_n);
-  }
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -2954,13 +2920,13 @@ sort_addr_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(float *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(float *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -2968,13 +2934,13 @@ sort_addr_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(float *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(float *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -2982,7 +2948,7 @@ sort_addr_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(float *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -2998,7 +2964,7 @@ sort_addr_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
     /* P.3/P.4: NaN pre-partition for float argsort -- NaN pairs are
        pushed to the tail (= NaN-at-end policy, matches sort_copy
        AC8).  Only the finite slice is sorted. */
-    ca_size_t finite_n = ca_partition_nan_pair_f32((ca_pair_f32 *) (buf + sort_lo), (ca_size_t) sort_n);
+    ca_size_t finite_n = ca_partition_nan_pair_f32((ca_pair_f32 *) (buf + sort_lo), (ca_size_t) sort_n, (ca_pair_f32 *) aux);
     if ( do_stable ) {
       ca_sort_merge_pair_f32((ca_pair_f32 *) (buf + sort_lo), (ca_pair_f32 *) aux, finite_n);
     } else {
@@ -3023,7 +2989,7 @@ sort_addr_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int ma
 }
 typedef struct {
-  double v;
+  ca_u_float64_t v;
   ca_size_t i;
 } sort_addr_pair_f64;
@@ -3040,6 +3006,10 @@ sort_addr_ki_native_f64 (VALUE self, CArray *ca, int axis, int do_stable, int ma
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_addr_pair_f64 *buf = ALLOCV_N(sort_addr_pair_f64, hbuf, fiber_n);
+  sort_addr_pair_f64 *aux = ALLOCV_N(sort_addr_pair_f64, haux, fiber_n);
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -3054,14 +3024,6 @@ sort_addr_ki_native_f64 (VALUE self, CArray *ca, int axis, int do_stable, int ma
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_addr_pair_f64 *buf = ALLOCV_N(sort_addr_pair_f64, hbuf, fiber_n);
-  sort_addr_pair_f64 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_addr_pair_f64, haux, fiber_n);
-  }
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -3105,13 +3067,13 @@ sort_addr_ki_native_f64 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(double *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(double *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -3119,13 +3081,13 @@ sort_addr_ki_native_f64 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(double *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(double *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -3133,7 +3095,7 @@ sort_addr_ki_native_f64 (VALUE self, CArray *ca, int axis, int do_stable, int ma
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(double *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -3149,7 +3111,7 @@ sort_addr_ki_native_f64 (VALUE self, CArray *ca, int axis, int do_stable, int ma
     /* P.3/P.4: NaN pre-partition for float argsort -- NaN pairs are
        pushed to the tail (= NaN-at-end policy, matches sort_copy
        AC8).  Only the finite slice is sorted. */
-    ca_size_t finite_n = ca_partition_nan_pair_f64((ca_pair_f64 *) (buf + sort_lo), (ca_size_t) sort_n);
+    ca_size_t finite_n = ca_partition_nan_pair_f64((ca_pair_f64 *) (buf + sort_lo), (ca_size_t) sort_n, (ca_pair_f64 *) aux);
     if ( do_stable ) {
       ca_sort_merge_pair_f64((ca_pair_f64 *) (buf + sort_lo), (ca_pair_f64 *) aux, finite_n);
     } else {
@@ -3191,6 +3153,13 @@ sort_addr_ki_native_bool (VALUE self, CArray *ca, int axis, int do_stable, int m
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_addr_pair_bool *buf = ALLOCV_N(sort_addr_pair_bool, hbuf, fiber_n);
+  sort_addr_pair_bool *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(sort_addr_pair_bool, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -3205,14 +3174,6 @@ sort_addr_ki_native_bool (VALUE self, CArray *ca, int axis, int do_stable, int m
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_addr_pair_bool *buf = ALLOCV_N(sort_addr_pair_bool, hbuf, fiber_n);
-  sort_addr_pair_bool *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_addr_pair_bool, haux, fiber_n);
-  }
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -3256,13 +3217,13 @@ sort_addr_ki_native_bool (VALUE self, CArray *ca, int axis, int do_stable, int m
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -3270,13 +3231,13 @@ sort_addr_ki_native_bool (VALUE self, CArray *ca, int axis, int do_stable, int m
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -3284,7 +3245,7 @@ sort_addr_ki_native_bool (VALUE self, CArray *ca, int axis, int do_stable, int m
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -3338,9 +3299,9 @@ sort_addr_cmp_object (const void *a, const void *b)
   return 0;
 }
-/* The object lane calls Ruby for every cell, so a walk can raise part
-   way.  Its states live in the caller's frame, and ca_iter_ensure
-   finishes them however the body leaves. */
+/* The walks can raise part way (a Ruby call per cell in the object
+   lane, an allocation in any lane).  Their states live in the caller's
+   frame, and ca_iter_ensure finishes them however the body leaves. */
 typedef struct {
   ca_iter_state states[2];
   VALUE self;
@@ -3376,6 +3337,10 @@ sort_addr_ki_native_object_body (VALUE arg)
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_addr_pair_object *buf = ALLOCV_N(sort_addr_pair_object, hbuf, fiber_n);
+  sort_addr_pair_object *aux = ALLOCV_N(sort_addr_pair_object, haux, fiber_n);
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
                              slab_axes, 1, 0);
@@ -3389,11 +3354,6 @@ sort_addr_ki_native_object_body (VALUE arg)
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_addr_pair_object *buf = ALLOCV_N(sort_addr_pair_object, hbuf, fiber_n);
-  sort_addr_pair_object *aux = ALLOCV_N(sort_addr_pair_object, haux, fiber_n);
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -3437,13 +3397,13 @@ sort_addr_ki_native_object_body (VALUE arg)
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -3451,13 +3411,13 @@ sort_addr_ki_native_object_body (VALUE arg)
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -3465,7 +3425,7 @@ sort_addr_ki_native_object_body (VALUE arg)
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -3509,7 +3469,7 @@ sort_addr_ki_native_object (VALUE self, CArray *ca, int axis, int do_stable, int
   fr.axis = axis;
   fr.do_stable = do_stable;
   fr.masked_last = masked_last;
-  return ca_iter_ensure(2, fr.states, sort_addr_ki_native_object_body, (VALUE) &fr);
+  return ca_iter_ensure_computing(ca, 2, fr.states, sort_addr_ki_native_object_body, (VALUE) &fr);
 }
@@ -3545,6 +3505,10 @@ sort_addr_ki_native_fixlen (VALUE self, CArray *ca, int axis, int do_stable, int
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  sort_addr_pair_fixlen *buf = ALLOCV_N(sort_addr_pair_fixlen, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -3559,14 +3523,6 @@ sort_addr_ki_native_fixlen (VALUE self, CArray *ca, int axis, int do_stable, int
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  sort_addr_pair_fixlen *buf = ALLOCV_N(sort_addr_pair_fixlen, hbuf, fiber_n);
-  sort_addr_pair_fixlen *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(sort_addr_pair_fixlen, haux, fiber_n);
-  }
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -3817,6 +3773,10 @@ partition_index_ki_native_i8 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_index_pair_i8 *buf = ALLOCV_N(partition_index_pair_i8, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -3831,11 +3791,6 @@ partition_index_ki_native_i8 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_index_pair_i8 *buf = ALLOCV_N(partition_index_pair_i8, hbuf, fiber_n);
-  (void) haux;
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -3852,13 +3807,13 @@ partition_index_ki_native_i8 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -3866,13 +3821,13 @@ partition_index_ki_native_i8 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -3880,7 +3835,7 @@ partition_index_ki_native_i8 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -3996,6 +3951,10 @@ partition_index_ki_native_u8 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_index_pair_u8 *buf = ALLOCV_N(partition_index_pair_u8, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -4010,11 +3969,6 @@ partition_index_ki_native_u8 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_index_pair_u8 *buf = ALLOCV_N(partition_index_pair_u8, hbuf, fiber_n);
-  (void) haux;
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -4031,13 +3985,13 @@ partition_index_ki_native_u8 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -4045,13 +3999,13 @@ partition_index_ki_native_u8 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -4059,7 +4013,7 @@ partition_index_ki_native_u8 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -4090,7 +4044,7 @@ partition_index_ki_native_u8 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
 }
 typedef struct {
-  int16_t v;
+  ca_u_int16_t v;
   ca_size_t i;
 } partition_index_pair_i16;
@@ -4175,6 +4129,10 @@ partition_index_ki_native_i16 (VALUE self, CArray *ca, int axis, ca_size_t kth,
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_index_pair_i16 *buf = ALLOCV_N(partition_index_pair_i16, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -4189,11 +4147,6 @@ partition_index_ki_native_i16 (VALUE self, CArray *ca, int axis, ca_size_t kth,
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_index_pair_i16 *buf = ALLOCV_N(partition_index_pair_i16, hbuf, fiber_n);
-  (void) haux;
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -4210,13 +4163,13 @@ partition_index_ki_native_i16 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -4224,13 +4177,13 @@ partition_index_ki_native_i16 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -4238,7 +4191,7 @@ partition_index_ki_native_i16 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -4269,7 +4222,7 @@ partition_index_ki_native_i16 (VALUE self, CArray *ca, int axis, ca_size_t kth,
 }
 typedef struct {
-  uint16_t v;
+  ca_u_uint16_t v;
   ca_size_t i;
 } partition_index_pair_u16;
@@ -4354,6 +4307,10 @@ partition_index_ki_native_u16 (VALUE self, CArray *ca, int axis, ca_size_t kth,
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_index_pair_u16 *buf = ALLOCV_N(partition_index_pair_u16, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -4368,11 +4325,6 @@ partition_index_ki_native_u16 (VALUE self, CArray *ca, int axis, ca_size_t kth,
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_index_pair_u16 *buf = ALLOCV_N(partition_index_pair_u16, hbuf, fiber_n);
-  (void) haux;
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -4389,13 +4341,13 @@ partition_index_ki_native_u16 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -4403,13 +4355,13 @@ partition_index_ki_native_u16 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -4417,7 +4369,7 @@ partition_index_ki_native_u16 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -4448,7 +4400,7 @@ partition_index_ki_native_u16 (VALUE self, CArray *ca, int axis, ca_size_t kth,
 }
 typedef struct {
-  int32_t v;
+  ca_u_int32_t v;
   ca_size_t i;
 } partition_index_pair_i32;
@@ -4533,6 +4485,10 @@ partition_index_ki_native_i32 (VALUE self, CArray *ca, int axis, ca_size_t kth,
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_index_pair_i32 *buf = ALLOCV_N(partition_index_pair_i32, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -4547,11 +4503,6 @@ partition_index_ki_native_i32 (VALUE self, CArray *ca, int axis, ca_size_t kth,
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_index_pair_i32 *buf = ALLOCV_N(partition_index_pair_i32, hbuf, fiber_n);
-  (void) haux;
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -4568,13 +4519,13 @@ partition_index_ki_native_i32 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -4582,13 +4533,13 @@ partition_index_ki_native_i32 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -4596,7 +4547,7 @@ partition_index_ki_native_i32 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -4627,7 +4578,7 @@ partition_index_ki_native_i32 (VALUE self, CArray *ca, int axis, ca_size_t kth,
 }
 typedef struct {
-  uint32_t v;
+  ca_u_uint32_t v;
   ca_size_t i;
 } partition_index_pair_u32;
@@ -4712,6 +4663,10 @@ partition_index_ki_native_u32 (VALUE self, CArray *ca, int axis, ca_size_t kth,
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_index_pair_u32 *buf = ALLOCV_N(partition_index_pair_u32, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -4726,11 +4681,6 @@ partition_index_ki_native_u32 (VALUE self, CArray *ca, int axis, ca_size_t kth,
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_index_pair_u32 *buf = ALLOCV_N(partition_index_pair_u32, hbuf, fiber_n);
-  (void) haux;
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -4747,13 +4697,13 @@ partition_index_ki_native_u32 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -4761,13 +4711,13 @@ partition_index_ki_native_u32 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -4775,7 +4725,7 @@ partition_index_ki_native_u32 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -4806,7 +4756,7 @@ partition_index_ki_native_u32 (VALUE self, CArray *ca, int axis, ca_size_t kth,
 }
 typedef struct {
-  int64_t v;
+  ca_u_int64_t v;
   ca_size_t i;
 } partition_index_pair_i64;
@@ -4891,6 +4841,10 @@ partition_index_ki_native_i64 (VALUE self, CArray *ca, int axis, ca_size_t kth,
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_index_pair_i64 *buf = ALLOCV_N(partition_index_pair_i64, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -4905,11 +4859,6 @@ partition_index_ki_native_i64 (VALUE self, CArray *ca, int axis, ca_size_t kth,
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_index_pair_i64 *buf = ALLOCV_N(partition_index_pair_i64, hbuf, fiber_n);
-  (void) haux;
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -4926,13 +4875,13 @@ partition_index_ki_native_i64 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -4940,13 +4889,13 @@ partition_index_ki_native_i64 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -4954,7 +4903,7 @@ partition_index_ki_native_i64 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -4985,7 +4934,7 @@ partition_index_ki_native_i64 (VALUE self, CArray *ca, int axis, ca_size_t kth,
 }
 typedef struct {
-  uint64_t v;
+  ca_u_uint64_t v;
   ca_size_t i;
 } partition_index_pair_u64;
@@ -5070,6 +5019,10 @@ partition_index_ki_native_u64 (VALUE self, CArray *ca, int axis, ca_size_t kth,
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_index_pair_u64 *buf = ALLOCV_N(partition_index_pair_u64, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -5084,11 +5037,6 @@ partition_index_ki_native_u64 (VALUE self, CArray *ca, int axis, ca_size_t kth,
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_index_pair_u64 *buf = ALLOCV_N(partition_index_pair_u64, hbuf, fiber_n);
-  (void) haux;
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -5105,13 +5053,13 @@ partition_index_ki_native_u64 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -5119,13 +5067,13 @@ partition_index_ki_native_u64 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -5133,7 +5081,7 @@ partition_index_ki_native_u64 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -5164,7 +5112,7 @@ partition_index_ki_native_u64 (VALUE self, CArray *ca, int axis, ca_size_t kth,
 }
 typedef struct {
-  float v;
+  ca_u_float32_t v;
   ca_size_t i;
 } partition_index_pair_f32;
@@ -5249,6 +5197,10 @@ partition_index_ki_native_f32 (VALUE self, CArray *ca, int axis, ca_size_t kth,
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_index_pair_f32 *buf = ALLOCV_N(partition_index_pair_f32, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -5263,11 +5215,6 @@ partition_index_ki_native_f32 (VALUE self, CArray *ca, int axis, ca_size_t kth,
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_index_pair_f32 *buf = ALLOCV_N(partition_index_pair_f32, hbuf, fiber_n);
-  (void) haux;
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -5284,13 +5231,13 @@ partition_index_ki_native_f32 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(float *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(float *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -5298,13 +5245,13 @@ partition_index_ki_native_f32 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(float *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(float *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -5312,7 +5259,7 @@ partition_index_ki_native_f32 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(float *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -5343,7 +5290,7 @@ partition_index_ki_native_f32 (VALUE self, CArray *ca, int axis, ca_size_t kth,
 }
 typedef struct {
-  double v;
+  ca_u_float64_t v;
   ca_size_t i;
 } partition_index_pair_f64;
@@ -5428,6 +5375,10 @@ partition_index_ki_native_f64 (VALUE self, CArray *ca, int axis, ca_size_t kth,
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_index_pair_f64 *buf = ALLOCV_N(partition_index_pair_f64, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -5442,11 +5393,6 @@ partition_index_ki_native_f64 (VALUE self, CArray *ca, int axis, ca_size_t kth,
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_index_pair_f64 *buf = ALLOCV_N(partition_index_pair_f64, hbuf, fiber_n);
-  (void) haux;
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -5463,13 +5409,13 @@ partition_index_ki_native_f64 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(double *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(double *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -5477,13 +5423,13 @@ partition_index_ki_native_f64 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(double *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(double *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -5491,7 +5437,7 @@ partition_index_ki_native_f64 (VALUE self, CArray *ca, int axis, ca_size_t kth,
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(double *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -5607,6 +5553,10 @@ partition_index_ki_native_bool (VALUE self, CArray *ca, int axis, ca_size_t kth,
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_index_pair_bool *buf = ALLOCV_N(partition_index_pair_bool, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -5621,11 +5571,6 @@ partition_index_ki_native_bool (VALUE self, CArray *ca, int axis, ca_size_t kth,
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_index_pair_bool *buf = ALLOCV_N(partition_index_pair_bool, hbuf, fiber_n);
-  (void) haux;
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -5642,13 +5587,13 @@ partition_index_ki_native_bool (VALUE self, CArray *ca, int axis, ca_size_t kth,
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -5656,13 +5601,13 @@ partition_index_ki_native_bool (VALUE self, CArray *ca, int axis, ca_size_t kth,
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -5670,7 +5615,7 @@ partition_index_ki_native_bool (VALUE self, CArray *ca, int axis, ca_size_t kth,
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -5773,9 +5718,9 @@ partition_index_quickselect_object (partition_index_pair_object *arr,
   }
 }
-/* The object lane calls Ruby for every cell, so a walk can raise part
-   way.  Its states live in the caller's frame, and ca_iter_ensure
-   finishes them however the body leaves. */
+/* The walks can raise part way (a Ruby call per cell in the object
+   lane, an allocation in any lane).  Their states live in the caller's
+   frame, and ca_iter_ensure finishes them however the body leaves. */
 typedef struct {
   ca_iter_state states[2];
   VALUE self;
@@ -5811,6 +5756,10 @@ partition_index_ki_native_object_body (VALUE arg)
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_index_pair_object *buf = ALLOCV_N(partition_index_pair_object, hbuf, fiber_n);
+  (void) haux;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
                              slab_axes, 1, 0);
@@ -5824,11 +5773,6 @@ partition_index_ki_native_object_body (VALUE arg)
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_index_pair_object *buf = ALLOCV_N(partition_index_pair_object, hbuf, fiber_n);
-  (void) haux;
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -5845,13 +5789,13 @@ partition_index_ki_native_object_body (VALUE arg)
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = k;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = k;
           n_masked++;
         }
@@ -5859,13 +5803,13 @@ partition_index_ki_native_object_body (VALUE arg)
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = k;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = k;
           n_valid++;
         }
@@ -5873,7 +5817,7 @@ partition_index_ki_native_object_body (VALUE arg)
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = k;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -5914,7 +5858,7 @@ partition_index_ki_native_object (VALUE self, CArray *ca, int axis, ca_size_t kt
   fr.axis = axis;
   fr.kth = kth;
   fr.masked_last = masked_last;
-  return ca_iter_ensure(2, fr.states, partition_index_ki_native_object_body, (VALUE) &fr);
+  return ca_iter_ensure_computing(ca, 2, fr.states, partition_index_ki_native_object_body, (VALUE) &fr);
 }
@@ -6005,6 +5949,10 @@ partition_index_ki_native_fixlen (VALUE self, CArray *ca, int axis, ca_size_t kt
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_index_pair_fixlen *buf = ALLOCV_N(partition_index_pair_fixlen, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -6019,11 +5967,6 @@ partition_index_ki_native_fixlen (VALUE self, CArray *ca, int axis, ca_size_t kt
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_index_pair_fixlen *buf = ALLOCV_N(partition_index_pair_fixlen, hbuf, fiber_n);
-  (void) haux;
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -6260,6 +6203,10 @@ partition_addr_ki_native_i8 (VALUE self, CArray *ca, int axis, ca_size_t kth, in
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_addr_pair_i8 *buf = ALLOCV_N(partition_addr_pair_i8, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -6274,11 +6221,6 @@ partition_addr_ki_native_i8 (VALUE self, CArray *ca, int axis, ca_size_t kth, in
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_addr_pair_i8 *buf = ALLOCV_N(partition_addr_pair_i8, hbuf, fiber_n);
-  (void) haux;
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -6322,13 +6264,13 @@ partition_addr_ki_native_i8 (VALUE self, CArray *ca, int axis, ca_size_t kth, in
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -6336,13 +6278,13 @@ partition_addr_ki_native_i8 (VALUE self, CArray *ca, int axis, ca_size_t kth, in
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -6350,7 +6292,7 @@ partition_addr_ki_native_i8 (VALUE self, CArray *ca, int axis, ca_size_t kth, in
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(int8_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -6471,6 +6413,10 @@ partition_addr_ki_native_u8 (VALUE self, CArray *ca, int axis, ca_size_t kth, in
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_addr_pair_u8 *buf = ALLOCV_N(partition_addr_pair_u8, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -6485,11 +6431,6 @@ partition_addr_ki_native_u8 (VALUE self, CArray *ca, int axis, ca_size_t kth, in
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_addr_pair_u8 *buf = ALLOCV_N(partition_addr_pair_u8, hbuf, fiber_n);
-  (void) haux;
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -6533,13 +6474,13 @@ partition_addr_ki_native_u8 (VALUE self, CArray *ca, int axis, ca_size_t kth, in
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -6547,13 +6488,13 @@ partition_addr_ki_native_u8 (VALUE self, CArray *ca, int axis, ca_size_t kth, in
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -6561,7 +6502,7 @@ partition_addr_ki_native_u8 (VALUE self, CArray *ca, int axis, ca_size_t kth, in
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(uint8_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -6597,7 +6538,7 @@ partition_addr_ki_native_u8 (VALUE self, CArray *ca, int axis, ca_size_t kth, in
 }
 typedef struct {
-  int16_t v;
+  ca_u_int16_t v;
   ca_size_t i;
 } partition_addr_pair_i16;
@@ -6682,6 +6623,10 @@ partition_addr_ki_native_i16 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_addr_pair_i16 *buf = ALLOCV_N(partition_addr_pair_i16, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -6696,11 +6641,6 @@ partition_addr_ki_native_i16 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_addr_pair_i16 *buf = ALLOCV_N(partition_addr_pair_i16, hbuf, fiber_n);
-  (void) haux;
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -6744,13 +6684,13 @@ partition_addr_ki_native_i16 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -6758,13 +6698,13 @@ partition_addr_ki_native_i16 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -6772,7 +6712,7 @@ partition_addr_ki_native_i16 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(int16_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -6808,7 +6748,7 @@ partition_addr_ki_native_i16 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
 }
 typedef struct {
-  uint16_t v;
+  ca_u_uint16_t v;
   ca_size_t i;
 } partition_addr_pair_u16;
@@ -6893,6 +6833,10 @@ partition_addr_ki_native_u16 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_addr_pair_u16 *buf = ALLOCV_N(partition_addr_pair_u16, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -6907,11 +6851,6 @@ partition_addr_ki_native_u16 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_addr_pair_u16 *buf = ALLOCV_N(partition_addr_pair_u16, hbuf, fiber_n);
-  (void) haux;
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -6955,13 +6894,13 @@ partition_addr_ki_native_u16 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -6969,13 +6908,13 @@ partition_addr_ki_native_u16 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -6983,7 +6922,7 @@ partition_addr_ki_native_u16 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(uint16_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -7019,7 +6958,7 @@ partition_addr_ki_native_u16 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
 }
 typedef struct {
-  int32_t v;
+  ca_u_int32_t v;
   ca_size_t i;
 } partition_addr_pair_i32;
@@ -7104,6 +7043,10 @@ partition_addr_ki_native_i32 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_addr_pair_i32 *buf = ALLOCV_N(partition_addr_pair_i32, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -7118,11 +7061,6 @@ partition_addr_ki_native_i32 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_addr_pair_i32 *buf = ALLOCV_N(partition_addr_pair_i32, hbuf, fiber_n);
-  (void) haux;
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -7166,13 +7104,13 @@ partition_addr_ki_native_i32 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -7180,13 +7118,13 @@ partition_addr_ki_native_i32 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -7194,7 +7132,7 @@ partition_addr_ki_native_i32 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(int32_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -7230,7 +7168,7 @@ partition_addr_ki_native_i32 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
 }
 typedef struct {
-  uint32_t v;
+  ca_u_uint32_t v;
   ca_size_t i;
 } partition_addr_pair_u32;
@@ -7315,6 +7253,10 @@ partition_addr_ki_native_u32 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_addr_pair_u32 *buf = ALLOCV_N(partition_addr_pair_u32, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -7329,11 +7271,6 @@ partition_addr_ki_native_u32 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_addr_pair_u32 *buf = ALLOCV_N(partition_addr_pair_u32, hbuf, fiber_n);
-  (void) haux;
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -7377,13 +7314,13 @@ partition_addr_ki_native_u32 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -7391,13 +7328,13 @@ partition_addr_ki_native_u32 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -7405,7 +7342,7 @@ partition_addr_ki_native_u32 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(uint32_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -7441,7 +7378,7 @@ partition_addr_ki_native_u32 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
 }
 typedef struct {
-  int64_t v;
+  ca_u_int64_t v;
   ca_size_t i;
 } partition_addr_pair_i64;
@@ -7526,6 +7463,10 @@ partition_addr_ki_native_i64 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_addr_pair_i64 *buf = ALLOCV_N(partition_addr_pair_i64, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -7540,11 +7481,6 @@ partition_addr_ki_native_i64 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_addr_pair_i64 *buf = ALLOCV_N(partition_addr_pair_i64, hbuf, fiber_n);
-  (void) haux;
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -7588,13 +7524,13 @@ partition_addr_ki_native_i64 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -7602,13 +7538,13 @@ partition_addr_ki_native_i64 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -7616,7 +7552,7 @@ partition_addr_ki_native_i64 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(int64_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -7652,7 +7588,7 @@ partition_addr_ki_native_i64 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
 }
 typedef struct {
-  uint64_t v;
+  ca_u_uint64_t v;
   ca_size_t i;
 } partition_addr_pair_u64;
@@ -7737,6 +7673,10 @@ partition_addr_ki_native_u64 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_addr_pair_u64 *buf = ALLOCV_N(partition_addr_pair_u64, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -7751,11 +7691,6 @@ partition_addr_ki_native_u64 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_addr_pair_u64 *buf = ALLOCV_N(partition_addr_pair_u64, hbuf, fiber_n);
-  (void) haux;
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -7799,13 +7734,13 @@ partition_addr_ki_native_u64 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -7813,13 +7748,13 @@ partition_addr_ki_native_u64 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -7827,7 +7762,7 @@ partition_addr_ki_native_u64 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(uint64_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -7863,7 +7798,7 @@ partition_addr_ki_native_u64 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
 }
 typedef struct {
-  float v;
+  ca_u_float32_t v;
   ca_size_t i;
 } partition_addr_pair_f32;
@@ -7948,6 +7883,10 @@ partition_addr_ki_native_f32 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_addr_pair_f32 *buf = ALLOCV_N(partition_addr_pair_f32, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -7962,11 +7901,6 @@ partition_addr_ki_native_f32 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_addr_pair_f32 *buf = ALLOCV_N(partition_addr_pair_f32, hbuf, fiber_n);
-  (void) haux;
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -8010,13 +7944,13 @@ partition_addr_ki_native_f32 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(float *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(float *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -8024,13 +7958,13 @@ partition_addr_ki_native_f32 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(float *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(float *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -8038,7 +7972,7 @@ partition_addr_ki_native_f32 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(float *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -8074,7 +8008,7 @@ partition_addr_ki_native_f32 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
 }
 typedef struct {
-  double v;
+  ca_u_float64_t v;
   ca_size_t i;
 } partition_addr_pair_f64;
@@ -8159,6 +8093,10 @@ partition_addr_ki_native_f64 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_addr_pair_f64 *buf = ALLOCV_N(partition_addr_pair_f64, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -8173,11 +8111,6 @@ partition_addr_ki_native_f64 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_addr_pair_f64 *buf = ALLOCV_N(partition_addr_pair_f64, hbuf, fiber_n);
-  (void) haux;
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -8221,13 +8154,13 @@ partition_addr_ki_native_f64 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(double *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(double *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -8235,13 +8168,13 @@ partition_addr_ki_native_f64 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(double *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(double *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -8249,7 +8182,7 @@ partition_addr_ki_native_f64 (VALUE self, CArray *ca, int axis, ca_size_t kth, i
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(double *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -8370,6 +8303,10 @@ partition_addr_ki_native_bool (VALUE self, CArray *ca, int axis, ca_size_t kth,
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_addr_pair_bool *buf = ALLOCV_N(partition_addr_pair_bool, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -8384,11 +8321,6 @@ partition_addr_ki_native_bool (VALUE self, CArray *ca, int axis, ca_size_t kth,
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_addr_pair_bool *buf = ALLOCV_N(partition_addr_pair_bool, hbuf, fiber_n);
-  (void) haux;
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -8432,13 +8364,13 @@ partition_addr_ki_native_bool (VALUE self, CArray *ca, int axis, ca_size_t kth,
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -8446,13 +8378,13 @@ partition_addr_ki_native_bool (VALUE self, CArray *ca, int axis, ca_size_t kth,
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -8460,7 +8392,7 @@ partition_addr_ki_native_bool (VALUE self, CArray *ca, int axis, ca_size_t kth,
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(boolean8_t *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -8568,9 +8500,9 @@ partition_addr_quickselect_object (partition_addr_pair_object *arr,
   }
 }
-/* The object lane calls Ruby for every cell, so a walk can raise part
-   way.  Its states live in the caller's frame, and ca_iter_ensure
-   finishes them however the body leaves. */
+/* The walks can raise part way (a Ruby call per cell in the object
+   lane, an allocation in any lane).  Their states live in the caller's
+   frame, and ca_iter_ensure finishes them however the body leaves. */
 typedef struct {
   ca_iter_state states[2];
   VALUE self;
@@ -8606,6 +8538,10 @@ partition_addr_ki_native_object_body (VALUE arg)
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_addr_pair_object *buf = ALLOCV_N(partition_addr_pair_object, hbuf, fiber_n);
+  (void) haux;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
                              slab_axes, 1, 0);
@@ -8619,11 +8555,6 @@ partition_addr_ki_native_object_body (VALUE arg)
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_addr_pair_object *buf = ALLOCV_N(partition_addr_pair_object, hbuf, fiber_n);
-  (void) haux;
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -8667,13 +8598,13 @@ partition_addr_ki_native_object_body (VALUE arg)
       if ( masked_last ) {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_valid].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[n_valid].v, pi + k * istride, sizeof(buf[n_valid].v));
           buf[n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_valid + n_masked].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[n_valid + n_masked].v, pi + k * istride, sizeof(buf[n_valid + n_masked].v));
           buf[n_valid + n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
@@ -8681,13 +8612,13 @@ partition_addr_ki_native_object_body (VALUE arg)
       } else {
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( ! mi[k * mstride] ) continue;
-          buf[n_masked].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[n_masked].v, pi + k * istride, sizeof(buf[n_masked].v));
           buf[n_masked].i = outer_off + k * axis_vstride;
           n_masked++;
         }
         for ( ca_size_t k = 0; k < fiber_n; k++ ) {
           if ( mi[k * mstride] ) continue;
-          buf[n_masked + n_valid].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[n_masked + n_valid].v, pi + k * istride, sizeof(buf[n_masked + n_valid].v));
           buf[n_masked + n_valid].i = outer_off + k * axis_vstride;
           n_valid++;
         }
@@ -8695,7 +8626,7 @@ partition_addr_ki_native_object_body (VALUE arg)
       }
     } else {
       for ( ca_size_t k = 0; k < fiber_n; k++ ) {
-          buf[k].v = *(VALUE *) (pi + k * istride);
+          memcpy(&buf[k].v, pi + k * istride, sizeof(buf[k].v));
         buf[k].i = outer_off + k * axis_vstride;
       }
       sort_lo = 0; sort_n = fiber_n;
@@ -8741,7 +8672,7 @@ partition_addr_ki_native_object (VALUE self, CArray *ca, int axis, ca_size_t kth
   fr.axis = axis;
   fr.kth = kth;
   fr.masked_last = masked_last;
-  return ca_iter_ensure(2, fr.states, partition_addr_ki_native_object_body, (VALUE) &fr);
+  return ca_iter_ensure_computing(ca, 2, fr.states, partition_addr_ki_native_object_body, (VALUE) &fr);
 }
@@ -8832,6 +8763,10 @@ partition_addr_ki_native_fixlen (VALUE self, CArray *ca, int axis, ca_size_t kth
      positions align between iterators. */
   int8_t slab_axes[CA_RANK_MAX];
   slab_axes[0] = (int8_t) axis;
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  partition_addr_pair_fixlen *buf = ALLOCV_N(partition_addr_pair_fixlen, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -8846,11 +8781,6 @@ partition_addr_ki_native_fixlen (VALUE self, CArray *ca, int axis, ca_size_t kth
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  partition_addr_pair_fixlen *buf = ALLOCV_N(partition_addr_pair_fixlen, hbuf, fiber_n);
-  (void) haux;
   /* :view_flat semantics: kernel computes view-flat address
      (= row-major position in ca's shape, 0..elements-1) for
      each cell.  Used to feed CARemap.idx directly (= sort
@@ -9034,6 +8964,13 @@ rank_index_ki_native_i8 (VALUE self, CArray *ca, int axis, int do_stable, int de
      next_slab_axes will then be non-NULL).  Created up-front
      when input has any mask. */
   if ( ca_has_mask(ca) ) ca_create_mask(co);
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  rank_index_pair_i8 *buf = ALLOCV_N(rank_index_pair_i8, hbuf, fiber_n);
+  rank_index_pair_i8 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(rank_index_pair_i8, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -9048,14 +8985,6 @@ rank_index_ki_native_i8 (VALUE self, CArray *ca, int axis, int do_stable, int de
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  rank_index_pair_i8 *buf = ALLOCV_N(rank_index_pair_i8, hbuf, fiber_n);
-  rank_index_pair_i8 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(rank_index_pair_i8, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -9068,7 +8997,7 @@ rank_index_ki_native_i8 (VALUE self, CArray *ca, int axis, int do_stable, int de
     ca_size_t k_unmasked = 0;
     for ( ca_size_t k = 0; k < fiber_n; k++ ) {
       if ( mi && mi[k * mstride] ) continue;
-      buf[k_unmasked].v = *(int8_t *) (pi + k * istride);
+      memcpy(&buf[k_unmasked].v, pi + k * istride, sizeof(buf[k_unmasked].v));
       buf[k_unmasked].i = k;
       k_unmasked++;
     }
@@ -9165,6 +9094,13 @@ rank_index_ki_native_u8 (VALUE self, CArray *ca, int axis, int do_stable, int de
      next_slab_axes will then be non-NULL).  Created up-front
      when input has any mask. */
   if ( ca_has_mask(ca) ) ca_create_mask(co);
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  rank_index_pair_u8 *buf = ALLOCV_N(rank_index_pair_u8, hbuf, fiber_n);
+  rank_index_pair_u8 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(rank_index_pair_u8, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -9179,14 +9115,6 @@ rank_index_ki_native_u8 (VALUE self, CArray *ca, int axis, int do_stable, int de
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  rank_index_pair_u8 *buf = ALLOCV_N(rank_index_pair_u8, hbuf, fiber_n);
-  rank_index_pair_u8 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(rank_index_pair_u8, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -9199,7 +9127,7 @@ rank_index_ki_native_u8 (VALUE self, CArray *ca, int axis, int do_stable, int de
     ca_size_t k_unmasked = 0;
     for ( ca_size_t k = 0; k < fiber_n; k++ ) {
       if ( mi && mi[k * mstride] ) continue;
-      buf[k_unmasked].v = *(uint8_t *) (pi + k * istride);
+      memcpy(&buf[k_unmasked].v, pi + k * istride, sizeof(buf[k_unmasked].v));
       buf[k_unmasked].i = k;
       k_unmasked++;
     }
@@ -9274,7 +9202,7 @@ rank_index_ki_native_u8 (VALUE self, CArray *ca, int axis, int do_stable, int de
 }
 typedef struct {
-  int16_t v;
+  ca_u_int16_t v;
   ca_size_t i;
 } rank_index_pair_i16;
@@ -9296,6 +9224,13 @@ rank_index_ki_native_i16 (VALUE self, CArray *ca, int axis, int do_stable, int d
      next_slab_axes will then be non-NULL).  Created up-front
      when input has any mask. */
   if ( ca_has_mask(ca) ) ca_create_mask(co);
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  rank_index_pair_i16 *buf = ALLOCV_N(rank_index_pair_i16, hbuf, fiber_n);
+  rank_index_pair_i16 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(rank_index_pair_i16, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -9310,14 +9245,6 @@ rank_index_ki_native_i16 (VALUE self, CArray *ca, int axis, int do_stable, int d
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  rank_index_pair_i16 *buf = ALLOCV_N(rank_index_pair_i16, hbuf, fiber_n);
-  rank_index_pair_i16 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(rank_index_pair_i16, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -9330,7 +9257,7 @@ rank_index_ki_native_i16 (VALUE self, CArray *ca, int axis, int do_stable, int d
     ca_size_t k_unmasked = 0;
     for ( ca_size_t k = 0; k < fiber_n; k++ ) {
       if ( mi && mi[k * mstride] ) continue;
-      buf[k_unmasked].v = *(int16_t *) (pi + k * istride);
+      memcpy(&buf[k_unmasked].v, pi + k * istride, sizeof(buf[k_unmasked].v));
       buf[k_unmasked].i = k;
       k_unmasked++;
     }
@@ -9405,7 +9332,7 @@ rank_index_ki_native_i16 (VALUE self, CArray *ca, int axis, int do_stable, int d
 }
 typedef struct {
-  uint16_t v;
+  ca_u_uint16_t v;
   ca_size_t i;
 } rank_index_pair_u16;
@@ -9427,6 +9354,13 @@ rank_index_ki_native_u16 (VALUE self, CArray *ca, int axis, int do_stable, int d
      next_slab_axes will then be non-NULL).  Created up-front
      when input has any mask. */
   if ( ca_has_mask(ca) ) ca_create_mask(co);
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  rank_index_pair_u16 *buf = ALLOCV_N(rank_index_pair_u16, hbuf, fiber_n);
+  rank_index_pair_u16 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(rank_index_pair_u16, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -9441,14 +9375,6 @@ rank_index_ki_native_u16 (VALUE self, CArray *ca, int axis, int do_stable, int d
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  rank_index_pair_u16 *buf = ALLOCV_N(rank_index_pair_u16, hbuf, fiber_n);
-  rank_index_pair_u16 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(rank_index_pair_u16, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -9461,7 +9387,7 @@ rank_index_ki_native_u16 (VALUE self, CArray *ca, int axis, int do_stable, int d
     ca_size_t k_unmasked = 0;
     for ( ca_size_t k = 0; k < fiber_n; k++ ) {
       if ( mi && mi[k * mstride] ) continue;
-      buf[k_unmasked].v = *(uint16_t *) (pi + k * istride);
+      memcpy(&buf[k_unmasked].v, pi + k * istride, sizeof(buf[k_unmasked].v));
       buf[k_unmasked].i = k;
       k_unmasked++;
     }
@@ -9536,7 +9462,7 @@ rank_index_ki_native_u16 (VALUE self, CArray *ca, int axis, int do_stable, int d
 }
 typedef struct {
-  int32_t v;
+  ca_u_int32_t v;
   ca_size_t i;
 } rank_index_pair_i32;
@@ -9558,6 +9484,13 @@ rank_index_ki_native_i32 (VALUE self, CArray *ca, int axis, int do_stable, int d
      next_slab_axes will then be non-NULL).  Created up-front
      when input has any mask. */
   if ( ca_has_mask(ca) ) ca_create_mask(co);
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  rank_index_pair_i32 *buf = ALLOCV_N(rank_index_pair_i32, hbuf, fiber_n);
+  rank_index_pair_i32 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(rank_index_pair_i32, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -9572,14 +9505,6 @@ rank_index_ki_native_i32 (VALUE self, CArray *ca, int axis, int do_stable, int d
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  rank_index_pair_i32 *buf = ALLOCV_N(rank_index_pair_i32, hbuf, fiber_n);
-  rank_index_pair_i32 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(rank_index_pair_i32, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -9592,7 +9517,7 @@ rank_index_ki_native_i32 (VALUE self, CArray *ca, int axis, int do_stable, int d
     ca_size_t k_unmasked = 0;
     for ( ca_size_t k = 0; k < fiber_n; k++ ) {
       if ( mi && mi[k * mstride] ) continue;
-      buf[k_unmasked].v = *(int32_t *) (pi + k * istride);
+      memcpy(&buf[k_unmasked].v, pi + k * istride, sizeof(buf[k_unmasked].v));
       buf[k_unmasked].i = k;
       k_unmasked++;
     }
@@ -9667,7 +9592,7 @@ rank_index_ki_native_i32 (VALUE self, CArray *ca, int axis, int do_stable, int d
 }
 typedef struct {
-  uint32_t v;
+  ca_u_uint32_t v;
   ca_size_t i;
 } rank_index_pair_u32;
@@ -9689,6 +9614,13 @@ rank_index_ki_native_u32 (VALUE self, CArray *ca, int axis, int do_stable, int d
      next_slab_axes will then be non-NULL).  Created up-front
      when input has any mask. */
   if ( ca_has_mask(ca) ) ca_create_mask(co);
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  rank_index_pair_u32 *buf = ALLOCV_N(rank_index_pair_u32, hbuf, fiber_n);
+  rank_index_pair_u32 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(rank_index_pair_u32, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -9703,14 +9635,6 @@ rank_index_ki_native_u32 (VALUE self, CArray *ca, int axis, int do_stable, int d
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  rank_index_pair_u32 *buf = ALLOCV_N(rank_index_pair_u32, hbuf, fiber_n);
-  rank_index_pair_u32 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(rank_index_pair_u32, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -9723,7 +9647,7 @@ rank_index_ki_native_u32 (VALUE self, CArray *ca, int axis, int do_stable, int d
     ca_size_t k_unmasked = 0;
     for ( ca_size_t k = 0; k < fiber_n; k++ ) {
       if ( mi && mi[k * mstride] ) continue;
-      buf[k_unmasked].v = *(uint32_t *) (pi + k * istride);
+      memcpy(&buf[k_unmasked].v, pi + k * istride, sizeof(buf[k_unmasked].v));
       buf[k_unmasked].i = k;
       k_unmasked++;
     }
@@ -9798,7 +9722,7 @@ rank_index_ki_native_u32 (VALUE self, CArray *ca, int axis, int do_stable, int d
 }
 typedef struct {
-  int64_t v;
+  ca_u_int64_t v;
   ca_size_t i;
 } rank_index_pair_i64;
@@ -9820,6 +9744,13 @@ rank_index_ki_native_i64 (VALUE self, CArray *ca, int axis, int do_stable, int d
      next_slab_axes will then be non-NULL).  Created up-front
      when input has any mask. */
   if ( ca_has_mask(ca) ) ca_create_mask(co);
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  rank_index_pair_i64 *buf = ALLOCV_N(rank_index_pair_i64, hbuf, fiber_n);
+  rank_index_pair_i64 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(rank_index_pair_i64, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -9834,14 +9765,6 @@ rank_index_ki_native_i64 (VALUE self, CArray *ca, int axis, int do_stable, int d
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  rank_index_pair_i64 *buf = ALLOCV_N(rank_index_pair_i64, hbuf, fiber_n);
-  rank_index_pair_i64 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(rank_index_pair_i64, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -9854,7 +9777,7 @@ rank_index_ki_native_i64 (VALUE self, CArray *ca, int axis, int do_stable, int d
     ca_size_t k_unmasked = 0;
     for ( ca_size_t k = 0; k < fiber_n; k++ ) {
       if ( mi && mi[k * mstride] ) continue;
-      buf[k_unmasked].v = *(int64_t *) (pi + k * istride);
+      memcpy(&buf[k_unmasked].v, pi + k * istride, sizeof(buf[k_unmasked].v));
       buf[k_unmasked].i = k;
       k_unmasked++;
     }
@@ -9929,7 +9852,7 @@ rank_index_ki_native_i64 (VALUE self, CArray *ca, int axis, int do_stable, int d
 }
 typedef struct {
-  uint64_t v;
+  ca_u_uint64_t v;
   ca_size_t i;
 } rank_index_pair_u64;
@@ -9951,6 +9874,13 @@ rank_index_ki_native_u64 (VALUE self, CArray *ca, int axis, int do_stable, int d
      next_slab_axes will then be non-NULL).  Created up-front
      when input has any mask. */
   if ( ca_has_mask(ca) ) ca_create_mask(co);
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  rank_index_pair_u64 *buf = ALLOCV_N(rank_index_pair_u64, hbuf, fiber_n);
+  rank_index_pair_u64 *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(rank_index_pair_u64, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -9965,14 +9895,6 @@ rank_index_ki_native_u64 (VALUE self, CArray *ca, int axis, int do_stable, int d
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  rank_index_pair_u64 *buf = ALLOCV_N(rank_index_pair_u64, hbuf, fiber_n);
-  rank_index_pair_u64 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(rank_index_pair_u64, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -9985,7 +9907,7 @@ rank_index_ki_native_u64 (VALUE self, CArray *ca, int axis, int do_stable, int d
     ca_size_t k_unmasked = 0;
     for ( ca_size_t k = 0; k < fiber_n; k++ ) {
       if ( mi && mi[k * mstride] ) continue;
-      buf[k_unmasked].v = *(uint64_t *) (pi + k * istride);
+      memcpy(&buf[k_unmasked].v, pi + k * istride, sizeof(buf[k_unmasked].v));
       buf[k_unmasked].i = k;
       k_unmasked++;
     }
@@ -10060,7 +9982,7 @@ rank_index_ki_native_u64 (VALUE self, CArray *ca, int axis, int do_stable, int d
 }
 typedef struct {
-  float v;
+  ca_u_float32_t v;
   ca_size_t i;
 } rank_index_pair_f32;
@@ -10082,6 +10004,10 @@ rank_index_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int d
      next_slab_axes will then be non-NULL).  Created up-front
      when input has any mask. */
   if ( ca_has_mask(ca) ) ca_create_mask(co);
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  rank_index_pair_f32 *buf = ALLOCV_N(rank_index_pair_f32, hbuf, fiber_n);
+  rank_index_pair_f32 *aux = ALLOCV_N(rank_index_pair_f32, haux, fiber_n);
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -10096,14 +10022,6 @@ rank_index_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int d
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  rank_index_pair_f32 *buf = ALLOCV_N(rank_index_pair_f32, hbuf, fiber_n);
-  rank_index_pair_f32 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(rank_index_pair_f32, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -10116,7 +10034,7 @@ rank_index_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int d
     ca_size_t k_unmasked = 0;
     for ( ca_size_t k = 0; k < fiber_n; k++ ) {
       if ( mi && mi[k * mstride] ) continue;
-      buf[k_unmasked].v = *(float *) (pi + k * istride);
+      memcpy(&buf[k_unmasked].v, pi + k * istride, sizeof(buf[k_unmasked].v));
       buf[k_unmasked].i = k;
       k_unmasked++;
     }
@@ -10130,7 +10048,7 @@ rank_index_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int d
     /* P.3/P.4: NaN pre-partition for float argsort -- NaN pairs are
        pushed to the tail (= NaN-at-end policy, matches sort_copy
        AC8).  Only the finite slice is sorted. */
-    ca_size_t finite_n = ca_partition_nan_pair_f32((ca_pair_f32 *) buf, (ca_size_t) k_unmasked);
+    ca_size_t finite_n = ca_partition_nan_pair_f32((ca_pair_f32 *) buf, (ca_size_t) k_unmasked, (ca_pair_f32 *) aux);
     if ( do_stable ) {
       ca_sort_merge_pair_f32((ca_pair_f32 *) buf, (ca_pair_f32 *) aux, finite_n);
     } else {
@@ -10197,7 +10115,7 @@ rank_index_ki_native_f32 (VALUE self, CArray *ca, int axis, int do_stable, int d
 }
 typedef struct {
-  double v;
+  ca_u_float64_t v;
   ca_size_t i;
 } rank_index_pair_f64;
@@ -10219,6 +10137,10 @@ rank_index_ki_native_f64 (VALUE self, CArray *ca, int axis, int do_stable, int d
      next_slab_axes will then be non-NULL).  Created up-front
      when input has any mask. */
   if ( ca_has_mask(ca) ) ca_create_mask(co);
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  rank_index_pair_f64 *buf = ALLOCV_N(rank_index_pair_f64, hbuf, fiber_n);
+  rank_index_pair_f64 *aux = ALLOCV_N(rank_index_pair_f64, haux, fiber_n);
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -10233,14 +10155,6 @@ rank_index_ki_native_f64 (VALUE self, CArray *ca, int axis, int do_stable, int d
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  rank_index_pair_f64 *buf = ALLOCV_N(rank_index_pair_f64, hbuf, fiber_n);
-  rank_index_pair_f64 *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(rank_index_pair_f64, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -10253,7 +10167,7 @@ rank_index_ki_native_f64 (VALUE self, CArray *ca, int axis, int do_stable, int d
     ca_size_t k_unmasked = 0;
     for ( ca_size_t k = 0; k < fiber_n; k++ ) {
       if ( mi && mi[k * mstride] ) continue;
-      buf[k_unmasked].v = *(double *) (pi + k * istride);
+      memcpy(&buf[k_unmasked].v, pi + k * istride, sizeof(buf[k_unmasked].v));
       buf[k_unmasked].i = k;
       k_unmasked++;
     }
@@ -10267,7 +10181,7 @@ rank_index_ki_native_f64 (VALUE self, CArray *ca, int axis, int do_stable, int d
     /* P.3/P.4: NaN pre-partition for float argsort -- NaN pairs are
        pushed to the tail (= NaN-at-end policy, matches sort_copy
        AC8).  Only the finite slice is sorted. */
-    ca_size_t finite_n = ca_partition_nan_pair_f64((ca_pair_f64 *) buf, (ca_size_t) k_unmasked);
+    ca_size_t finite_n = ca_partition_nan_pair_f64((ca_pair_f64 *) buf, (ca_size_t) k_unmasked, (ca_pair_f64 *) aux);
     if ( do_stable ) {
       ca_sort_merge_pair_f64((ca_pair_f64 *) buf, (ca_pair_f64 *) aux, finite_n);
     } else {
@@ -10356,6 +10270,13 @@ rank_index_ki_native_bool (VALUE self, CArray *ca, int axis, int do_stable, int
      next_slab_axes will then be non-NULL).  Created up-front
      when input has any mask. */
   if ( ca_has_mask(ca) ) ca_create_mask(co);
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  rank_index_pair_bool *buf = ALLOCV_N(rank_index_pair_bool, hbuf, fiber_n);
+  rank_index_pair_bool *aux = NULL;
+  if ( do_stable ) {
+    aux = ALLOCV_N(rank_index_pair_bool, haux, fiber_n);
+  }
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -10370,14 +10291,6 @@ rank_index_ki_native_bool (VALUE self, CArray *ca, int axis, int do_stable, int
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  rank_index_pair_bool *buf = ALLOCV_N(rank_index_pair_bool, hbuf, fiber_n);
-  rank_index_pair_bool *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(rank_index_pair_bool, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -10390,7 +10303,7 @@ rank_index_ki_native_bool (VALUE self, CArray *ca, int axis, int do_stable, int
     ca_size_t k_unmasked = 0;
     for ( ca_size_t k = 0; k < fiber_n; k++ ) {
       if ( mi && mi[k * mstride] ) continue;
-      buf[k_unmasked].v = *(boolean8_t *) (pi + k * istride);
+      memcpy(&buf[k_unmasked].v, pi + k * istride, sizeof(buf[k_unmasked].v));
       buf[k_unmasked].i = k;
       k_unmasked++;
     }
@@ -10482,9 +10395,9 @@ rank_index_cmp_object (const void *a, const void *b)
   return 0;
 }
-/* The object lane calls Ruby for every cell, so a walk can raise part
-   way.  Its states live in the caller's frame, and ca_iter_ensure
-   finishes them however the body leaves. */
+/* The walks can raise part way (a Ruby call per cell in the object
+   lane, an allocation in any lane).  Their states live in the caller's
+   frame, and ca_iter_ensure finishes them however the body leaves. */
 typedef struct {
   ca_iter_state states[2];
   VALUE self;
@@ -10525,6 +10438,10 @@ rank_index_ki_native_object_body (VALUE arg)
      next_slab_axes will then be non-NULL).  Created up-front
      when input has any mask. */
   if ( ca_has_mask(ca) ) ca_create_mask(co);
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  rank_index_pair_object *buf = ALLOCV_N(rank_index_pair_object, hbuf, fiber_n);
+  rank_index_pair_object *aux = ALLOCV_N(rank_index_pair_object, haux, fiber_n);
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
                              slab_axes, 1, 0);
@@ -10538,11 +10455,6 @@ rank_index_ki_native_object_body (VALUE arg)
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  rank_index_pair_object *buf = ALLOCV_N(rank_index_pair_object, hbuf, fiber_n);
-  rank_index_pair_object *aux = ALLOCV_N(rank_index_pair_object, haux, fiber_n);
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&
@@ -10555,7 +10467,7 @@ rank_index_ki_native_object_body (VALUE arg)
     ca_size_t k_unmasked = 0;
     for ( ca_size_t k = 0; k < fiber_n; k++ ) {
       if ( mi && mi[k * mstride] ) continue;
-      buf[k_unmasked].v = *(VALUE *) (pi + k * istride);
+      memcpy(&buf[k_unmasked].v, pi + k * istride, sizeof(buf[k_unmasked].v));
       buf[k_unmasked].i = k;
       k_unmasked++;
     }
@@ -10639,7 +10551,7 @@ rank_index_ki_native_object (VALUE self, CArray *ca, int axis, int do_stable, in
   fr.axis = axis;
   fr.do_stable = do_stable;
   fr.dense = dense;
-  return ca_iter_ensure(2, fr.states, rank_index_ki_native_object_body, (VALUE) &fr);
+  return ca_iter_ensure_computing(ca, 2, fr.states, rank_index_ki_native_object_body, (VALUE) &fr);
 }
@@ -10680,6 +10592,10 @@ rank_index_ki_native_fixlen (VALUE self, CArray *ca, int axis, int do_stable, in
      next_slab_axes will then be non-NULL).  Created up-front
      when input has any mask. */
   if ( ca_has_mask(ca) ) ca_create_mask(co);
+  ca_size_t fiber_n = ca->dim[axis];
+  volatile VALUE hbuf = 0, haux = 0;
+  rank_index_pair_fixlen *buf = ALLOCV_N(rank_index_pair_fixlen, hbuf, fiber_n);
+  (void) haux;
   ca_iter_state st_in, st_out;
   int rc;
   rc = ca_iter_state_init_l2(&st_in,  ca, CA_SLAB_AXES,
@@ -10694,14 +10610,6 @@ rank_index_ki_native_fixlen (VALUE self, CArray *ca, int axis, int do_stable, in
     rb_raise(rb_eRuntimeError, "%s: output init failed rc=%d", ca_calling_method_name(), rc);
   }
-  ca_size_t fiber_n = st_in.slab_elements;
-  volatile VALUE hbuf = 0, haux = 0;
-  rank_index_pair_fixlen *buf = ALLOCV_N(rank_index_pair_fixlen, hbuf, fiber_n);
-  rank_index_pair_fixlen *aux = NULL;
-  if ( do_stable ) {
-    aux = ALLOCV_N(rank_index_pair_fixlen, haux, fiber_n);
-  }
   char       *pi, *po;
   boolean8_t *mi, *mo;
   while ( ca_iter_state_next_slab_axes(&st_in,  &pi, &mi) &&