RubyGems - carray - Versions diffs - 3.0.4 → 3.0.5 - Mend

carray 3.0.4 → 3.0.5

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 (111) hide show

checksums.yaml +4 -4
data/CHANGELOG.md +205 -0
data/carray.gemspec +1 -1
data/ext/ca_array_pool.c +25 -1
data/ext/ca_axis_dispatch.c +13 -20
data/ext/ca_categorical_iterator.c +27 -20
data/ext/ca_compare.c +4 -4
data/ext/ca_composite_dispatch.c +9 -9
data/ext/ca_kernel_iterator.c +6 -6
data/ext/ca_kernel_iterator.h +112 -83
data/ext/ca_obj_array.c +69 -19
data/ext/ca_obj_bincmp.c +1 -0
data/ext/ca_obj_binop.c +1 -0
data/ext/ca_obj_bitarray.c +5 -1
data/ext/ca_obj_bitfield.c +58 -43
data/ext/ca_obj_byte_swap.c +2 -1
data/ext/ca_obj_const_string.c +15 -3
data/ext/ca_obj_face.c +42 -1
data/ext/ca_obj_field.c +13 -3
data/ext/ca_obj_grid.c +52 -5
data/ext/ca_obj_monop.c +46 -10
data/ext/ca_obj_object.c +64 -12
data/ext/ca_obj_record.c +3 -0
data/ext/ca_obj_reduce.c +4 -0
data/ext/ca_obj_refer.c +18 -9
data/ext/ca_obj_roll.c +26 -6
data/ext/ca_obj_select.c +34 -5
data/ext/ca_obj_shift.c +31 -7
data/ext/ca_obj_stack.c +1 -0
data/ext/ca_obj_stride.c +28 -19
data/ext/ca_obj_string.c +11 -2
data/ext/ca_obj_tile.c +27 -7
data/ext/ca_obj_transpose.c +8 -4
data/ext/ca_obj_triop.c +3 -0
data/ext/ca_obj_window.c +79 -50
data/ext/ca_op_ipower.c +4 -4
data/ext/ca_sort_kernels.h +6 -3
data/ext/carray.h +33 -0
data/ext/carray_access.c +16 -3
data/ext/carray_attribute.c +24 -0
data/ext/carray_cast.c +243 -82
data/ext/carray_cast_func.rb +56 -26
data/ext/carray_conversion.c +2 -2
data/ext/carray_core.c +3 -0
data/ext/carray_count.c +6 -0
data/ext/carray_element.c +9 -4
data/ext/carray_factorize.c +20 -20
data/ext/carray_generate.c +8 -2
data/ext/carray_histogram.c +1 -1
data/ext/carray_index_classifier.c +4 -4
data/ext/carray_internal.h +12 -0
data/ext/carray_kernels_bincmp.c +168 -168
data/ext/carray_kernels_binop.c +900 -900
data/ext/carray_kernels_init.c +4 -4
data/ext/carray_kernels_moncmp.c +96 -96
data/ext/carray_kernels_monop.c +1984 -1984
data/ext/carray_kernels_reduce_aggregate.c +1733 -1733
data/ext/carray_kernels_reduce_cumulative.c +642 -642
data/ext/carray_kernels_reduce_extreme.c +1150 -1150
data/ext/carray_kernels_reduce_variance.c +258 -258
data/ext/carray_kernels_scan.c +128 -128
data/ext/carray_kernels_search.c +340 -340
data/ext/carray_kernels_sort.c +295 -304
data/ext/carray_kernels_triop.c +114 -114
data/ext/carray_key_windows.c +4 -0
data/ext/carray_mask.c +45 -22
data/ext/carray_memory_view.c +2 -2
data/ext/carray_operator.c +5 -5
data/ext/carray_order.c +227 -8
data/ext/carray_random.c +18 -18
data/ext/carray_scatter.c +24 -7
data/ext/carray_slab.c +32 -9
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_time_text.c +4 -2
data/ext/carray_utils.c +63 -21
data/ext/mkkernel.rb +210 -115
data/ext/version.h +4 -4
data/lib/carray/attribute.rb +17 -4
data/lib/carray/basics.rb +27 -7
data/lib/carray/categorical.rb +30 -16
data/lib/carray/categorical_iterator.rb +6 -6
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 +4 -3
data/lib/carray/frame/concat.rb +17 -3
data/lib/carray/frame/frame.rb +6 -7
data/lib/carray/frame/group.rb +100 -13
data/lib/carray/frame/io.rb +3 -1
data/lib/carray/fuse_source.rb +31 -2
data/lib/carray/lazy.rb +35 -0
data/lib/carray/methods/is_in.rb +8 -1
data/lib/carray/segment_iterator.rb +12 -8
data/lib/carray/serialize.rb +19 -2
data/lib/carray/stack.rb +15 -12
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 +107 -68
data/yard-stubs/ca_obj_transpose.rb +5 -4
data/yard-stubs/carray_attribute.rb +3 -1
data/yard-stubs/carray_cast.rb +2 -0
data/yard-stubs/carray_conversion.rb +5 -1
data/yard-stubs/carray_element.rb +7 -6
data/yard-stubs/carray_mask.rb +8 -3
data/yard-stubs/carray_math.rb +14 -4
data/yard-stubs/carray_order.rb +100 -4
data/yard-stubs/carray_scatter.rb +7 -5
metadata +1 -1

data/ext/carray_kernels_triop.c CHANGED Viewed

@@ -120,9 +120,9 @@ ca_triop_fma_int16_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  int16_t *q1 = (int16_t *) ptr1, *q2 = (int16_t *) ptr2,
-            *q3 = (int16_t *) ptr3, *q4 = (int16_t *) ptr4;
-  int16_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_int16_t *q1 = (ca_u_int16_t *) ptr1, *q2 = (ca_u_int16_t *) ptr2,
+            *q3 = (ca_u_int16_t *) ptr3, *q4 = (ca_u_int16_t *) ptr4;
+  ca_u_int16_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -154,9 +154,9 @@ ca_triop_fma_uint16_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  uint16_t *q1 = (uint16_t *) ptr1, *q2 = (uint16_t *) ptr2,
-            *q3 = (uint16_t *) ptr3, *q4 = (uint16_t *) ptr4;
-  uint16_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_uint16_t *q1 = (ca_u_uint16_t *) ptr1, *q2 = (ca_u_uint16_t *) ptr2,
+            *q3 = (ca_u_uint16_t *) ptr3, *q4 = (ca_u_uint16_t *) ptr4;
+  ca_u_uint16_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -188,9 +188,9 @@ ca_triop_fma_int32_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  int32_t *q1 = (int32_t *) ptr1, *q2 = (int32_t *) ptr2,
-            *q3 = (int32_t *) ptr3, *q4 = (int32_t *) ptr4;
-  int32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_int32_t *q1 = (ca_u_int32_t *) ptr1, *q2 = (ca_u_int32_t *) ptr2,
+            *q3 = (ca_u_int32_t *) ptr3, *q4 = (ca_u_int32_t *) ptr4;
+  ca_u_int32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -222,9 +222,9 @@ ca_triop_fma_uint32_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  uint32_t *q1 = (uint32_t *) ptr1, *q2 = (uint32_t *) ptr2,
-            *q3 = (uint32_t *) ptr3, *q4 = (uint32_t *) ptr4;
-  uint32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_uint32_t *q1 = (ca_u_uint32_t *) ptr1, *q2 = (ca_u_uint32_t *) ptr2,
+            *q3 = (ca_u_uint32_t *) ptr3, *q4 = (ca_u_uint32_t *) ptr4;
+  ca_u_uint32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -256,9 +256,9 @@ ca_triop_fma_int64_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  int64_t *q1 = (int64_t *) ptr1, *q2 = (int64_t *) ptr2,
-            *q3 = (int64_t *) ptr3, *q4 = (int64_t *) ptr4;
-  int64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_int64_t *q1 = (ca_u_int64_t *) ptr1, *q2 = (ca_u_int64_t *) ptr2,
+            *q3 = (ca_u_int64_t *) ptr3, *q4 = (ca_u_int64_t *) ptr4;
+  ca_u_int64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -290,9 +290,9 @@ ca_triop_fma_uint64_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  uint64_t *q1 = (uint64_t *) ptr1, *q2 = (uint64_t *) ptr2,
-            *q3 = (uint64_t *) ptr3, *q4 = (uint64_t *) ptr4;
-  uint64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_uint64_t *q1 = (ca_u_uint64_t *) ptr1, *q2 = (ca_u_uint64_t *) ptr2,
+            *q3 = (ca_u_uint64_t *) ptr3, *q4 = (ca_u_uint64_t *) ptr4;
+  ca_u_uint64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -324,9 +324,9 @@ ca_triop_fma_float32_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  float *q1 = (float *) ptr1, *q2 = (float *) ptr2,
-            *q3 = (float *) ptr3, *q4 = (float *) ptr4;
-  float *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_float32_t *q1 = (ca_u_float32_t *) ptr1, *q2 = (ca_u_float32_t *) ptr2,
+            *q3 = (ca_u_float32_t *) ptr3, *q4 = (ca_u_float32_t *) ptr4;
+  ca_u_float32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -358,9 +358,9 @@ ca_triop_fma_float64_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  double *q1 = (double *) ptr1, *q2 = (double *) ptr2,
-            *q3 = (double *) ptr3, *q4 = (double *) ptr4;
-  double *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_float64_t *q1 = (ca_u_float64_t *) ptr1, *q2 = (ca_u_float64_t *) ptr2,
+            *q3 = (ca_u_float64_t *) ptr3, *q4 = (ca_u_float64_t *) ptr4;
+  ca_u_float64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -392,9 +392,9 @@ ca_triop_fma_cmplx64_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  cmplx64_t *q1 = (cmplx64_t *) ptr1, *q2 = (cmplx64_t *) ptr2,
-            *q3 = (cmplx64_t *) ptr3, *q4 = (cmplx64_t *) ptr4;
-  cmplx64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_cmplx64_t *q1 = (ca_u_cmplx64_t *) ptr1, *q2 = (ca_u_cmplx64_t *) ptr2,
+            *q3 = (ca_u_cmplx64_t *) ptr3, *q4 = (ca_u_cmplx64_t *) ptr4;
+  ca_u_cmplx64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -426,9 +426,9 @@ ca_triop_fma_cmplx128_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  cmplx128_t *q1 = (cmplx128_t *) ptr1, *q2 = (cmplx128_t *) ptr2,
-            *q3 = (cmplx128_t *) ptr3, *q4 = (cmplx128_t *) ptr4;
-  cmplx128_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_cmplx128_t *q1 = (ca_u_cmplx128_t *) ptr1, *q2 = (ca_u_cmplx128_t *) ptr2,
+            *q3 = (ca_u_cmplx128_t *) ptr3, *q4 = (ca_u_cmplx128_t *) ptr4;
+  ca_u_cmplx128_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -591,9 +591,9 @@ ca_triop_fms_int16_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  int16_t *q1 = (int16_t *) ptr1, *q2 = (int16_t *) ptr2,
-            *q3 = (int16_t *) ptr3, *q4 = (int16_t *) ptr4;
-  int16_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_int16_t *q1 = (ca_u_int16_t *) ptr1, *q2 = (ca_u_int16_t *) ptr2,
+            *q3 = (ca_u_int16_t *) ptr3, *q4 = (ca_u_int16_t *) ptr4;
+  ca_u_int16_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -625,9 +625,9 @@ ca_triop_fms_uint16_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  uint16_t *q1 = (uint16_t *) ptr1, *q2 = (uint16_t *) ptr2,
-            *q3 = (uint16_t *) ptr3, *q4 = (uint16_t *) ptr4;
-  uint16_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_uint16_t *q1 = (ca_u_uint16_t *) ptr1, *q2 = (ca_u_uint16_t *) ptr2,
+            *q3 = (ca_u_uint16_t *) ptr3, *q4 = (ca_u_uint16_t *) ptr4;
+  ca_u_uint16_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -659,9 +659,9 @@ ca_triop_fms_int32_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  int32_t *q1 = (int32_t *) ptr1, *q2 = (int32_t *) ptr2,
-            *q3 = (int32_t *) ptr3, *q4 = (int32_t *) ptr4;
-  int32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_int32_t *q1 = (ca_u_int32_t *) ptr1, *q2 = (ca_u_int32_t *) ptr2,
+            *q3 = (ca_u_int32_t *) ptr3, *q4 = (ca_u_int32_t *) ptr4;
+  ca_u_int32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -693,9 +693,9 @@ ca_triop_fms_uint32_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  uint32_t *q1 = (uint32_t *) ptr1, *q2 = (uint32_t *) ptr2,
-            *q3 = (uint32_t *) ptr3, *q4 = (uint32_t *) ptr4;
-  uint32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_uint32_t *q1 = (ca_u_uint32_t *) ptr1, *q2 = (ca_u_uint32_t *) ptr2,
+            *q3 = (ca_u_uint32_t *) ptr3, *q4 = (ca_u_uint32_t *) ptr4;
+  ca_u_uint32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -727,9 +727,9 @@ ca_triop_fms_int64_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  int64_t *q1 = (int64_t *) ptr1, *q2 = (int64_t *) ptr2,
-            *q3 = (int64_t *) ptr3, *q4 = (int64_t *) ptr4;
-  int64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_int64_t *q1 = (ca_u_int64_t *) ptr1, *q2 = (ca_u_int64_t *) ptr2,
+            *q3 = (ca_u_int64_t *) ptr3, *q4 = (ca_u_int64_t *) ptr4;
+  ca_u_int64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -761,9 +761,9 @@ ca_triop_fms_uint64_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  uint64_t *q1 = (uint64_t *) ptr1, *q2 = (uint64_t *) ptr2,
-            *q3 = (uint64_t *) ptr3, *q4 = (uint64_t *) ptr4;
-  uint64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_uint64_t *q1 = (ca_u_uint64_t *) ptr1, *q2 = (ca_u_uint64_t *) ptr2,
+            *q3 = (ca_u_uint64_t *) ptr3, *q4 = (ca_u_uint64_t *) ptr4;
+  ca_u_uint64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -795,9 +795,9 @@ ca_triop_fms_float32_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  float *q1 = (float *) ptr1, *q2 = (float *) ptr2,
-            *q3 = (float *) ptr3, *q4 = (float *) ptr4;
-  float *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_float32_t *q1 = (ca_u_float32_t *) ptr1, *q2 = (ca_u_float32_t *) ptr2,
+            *q3 = (ca_u_float32_t *) ptr3, *q4 = (ca_u_float32_t *) ptr4;
+  ca_u_float32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -829,9 +829,9 @@ ca_triop_fms_float64_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  double *q1 = (double *) ptr1, *q2 = (double *) ptr2,
-            *q3 = (double *) ptr3, *q4 = (double *) ptr4;
-  double *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_float64_t *q1 = (ca_u_float64_t *) ptr1, *q2 = (ca_u_float64_t *) ptr2,
+            *q3 = (ca_u_float64_t *) ptr3, *q4 = (ca_u_float64_t *) ptr4;
+  ca_u_float64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -863,9 +863,9 @@ ca_triop_fms_cmplx64_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  cmplx64_t *q1 = (cmplx64_t *) ptr1, *q2 = (cmplx64_t *) ptr2,
-            *q3 = (cmplx64_t *) ptr3, *q4 = (cmplx64_t *) ptr4;
-  cmplx64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_cmplx64_t *q1 = (ca_u_cmplx64_t *) ptr1, *q2 = (ca_u_cmplx64_t *) ptr2,
+            *q3 = (ca_u_cmplx64_t *) ptr3, *q4 = (ca_u_cmplx64_t *) ptr4;
+  ca_u_cmplx64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -897,9 +897,9 @@ ca_triop_fms_cmplx128_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  cmplx128_t *q1 = (cmplx128_t *) ptr1, *q2 = (cmplx128_t *) ptr2,
-            *q3 = (cmplx128_t *) ptr3, *q4 = (cmplx128_t *) ptr4;
-  cmplx128_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_cmplx128_t *q1 = (ca_u_cmplx128_t *) ptr1, *q2 = (ca_u_cmplx128_t *) ptr2,
+            *q3 = (ca_u_cmplx128_t *) ptr3, *q4 = (ca_u_cmplx128_t *) ptr4;
+  ca_u_cmplx128_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -1062,9 +1062,9 @@ ca_triop_clip_int16_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  int16_t *q1 = (int16_t *) ptr1, *q2 = (int16_t *) ptr2,
-            *q3 = (int16_t *) ptr3, *q4 = (int16_t *) ptr4;
-  int16_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_int16_t *q1 = (ca_u_int16_t *) ptr1, *q2 = (ca_u_int16_t *) ptr2,
+            *q3 = (ca_u_int16_t *) ptr3, *q4 = (ca_u_int16_t *) ptr4;
+  ca_u_int16_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -1096,9 +1096,9 @@ ca_triop_clip_uint16_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  uint16_t *q1 = (uint16_t *) ptr1, *q2 = (uint16_t *) ptr2,
-            *q3 = (uint16_t *) ptr3, *q4 = (uint16_t *) ptr4;
-  uint16_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_uint16_t *q1 = (ca_u_uint16_t *) ptr1, *q2 = (ca_u_uint16_t *) ptr2,
+            *q3 = (ca_u_uint16_t *) ptr3, *q4 = (ca_u_uint16_t *) ptr4;
+  ca_u_uint16_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -1130,9 +1130,9 @@ ca_triop_clip_int32_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  int32_t *q1 = (int32_t *) ptr1, *q2 = (int32_t *) ptr2,
-            *q3 = (int32_t *) ptr3, *q4 = (int32_t *) ptr4;
-  int32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_int32_t *q1 = (ca_u_int32_t *) ptr1, *q2 = (ca_u_int32_t *) ptr2,
+            *q3 = (ca_u_int32_t *) ptr3, *q4 = (ca_u_int32_t *) ptr4;
+  ca_u_int32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -1164,9 +1164,9 @@ ca_triop_clip_uint32_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  uint32_t *q1 = (uint32_t *) ptr1, *q2 = (uint32_t *) ptr2,
-            *q3 = (uint32_t *) ptr3, *q4 = (uint32_t *) ptr4;
-  uint32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_uint32_t *q1 = (ca_u_uint32_t *) ptr1, *q2 = (ca_u_uint32_t *) ptr2,
+            *q3 = (ca_u_uint32_t *) ptr3, *q4 = (ca_u_uint32_t *) ptr4;
+  ca_u_uint32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -1198,9 +1198,9 @@ ca_triop_clip_int64_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  int64_t *q1 = (int64_t *) ptr1, *q2 = (int64_t *) ptr2,
-            *q3 = (int64_t *) ptr3, *q4 = (int64_t *) ptr4;
-  int64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_int64_t *q1 = (ca_u_int64_t *) ptr1, *q2 = (ca_u_int64_t *) ptr2,
+            *q3 = (ca_u_int64_t *) ptr3, *q4 = (ca_u_int64_t *) ptr4;
+  ca_u_int64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -1232,9 +1232,9 @@ ca_triop_clip_uint64_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  uint64_t *q1 = (uint64_t *) ptr1, *q2 = (uint64_t *) ptr2,
-            *q3 = (uint64_t *) ptr3, *q4 = (uint64_t *) ptr4;
-  uint64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_uint64_t *q1 = (ca_u_uint64_t *) ptr1, *q2 = (ca_u_uint64_t *) ptr2,
+            *q3 = (ca_u_uint64_t *) ptr3, *q4 = (ca_u_uint64_t *) ptr4;
+  ca_u_uint64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -1266,9 +1266,9 @@ ca_triop_clip_float32_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  float *q1 = (float *) ptr1, *q2 = (float *) ptr2,
-            *q3 = (float *) ptr3, *q4 = (float *) ptr4;
-  float *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_float32_t *q1 = (ca_u_float32_t *) ptr1, *q2 = (ca_u_float32_t *) ptr2,
+            *q3 = (ca_u_float32_t *) ptr3, *q4 = (ca_u_float32_t *) ptr4;
+  ca_u_float32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -1300,9 +1300,9 @@ ca_triop_clip_float64_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  double *q1 = (double *) ptr1, *q2 = (double *) ptr2,
-            *q3 = (double *) ptr3, *q4 = (double *) ptr4;
-  double *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_float64_t *q1 = (ca_u_float64_t *) ptr1, *q2 = (ca_u_float64_t *) ptr2,
+            *q3 = (ca_u_float64_t *) ptr3, *q4 = (ca_u_float64_t *) ptr4;
+  ca_u_float64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -1462,9 +1462,9 @@ ca_triop_select_int16_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  int16_t *q1 = (int16_t *) ptr1, *q2 = (int16_t *) ptr2,
-            *q3 = (int16_t *) ptr3, *q4 = (int16_t *) ptr4;
-  int16_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_int16_t *q1 = (ca_u_int16_t *) ptr1, *q2 = (ca_u_int16_t *) ptr2,
+            *q3 = (ca_u_int16_t *) ptr3, *q4 = (ca_u_int16_t *) ptr4;
+  ca_u_int16_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -1496,9 +1496,9 @@ ca_triop_select_uint16_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  uint16_t *q1 = (uint16_t *) ptr1, *q2 = (uint16_t *) ptr2,
-            *q3 = (uint16_t *) ptr3, *q4 = (uint16_t *) ptr4;
-  uint16_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_uint16_t *q1 = (ca_u_uint16_t *) ptr1, *q2 = (ca_u_uint16_t *) ptr2,
+            *q3 = (ca_u_uint16_t *) ptr3, *q4 = (ca_u_uint16_t *) ptr4;
+  ca_u_uint16_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -1530,9 +1530,9 @@ ca_triop_select_int32_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  int32_t *q1 = (int32_t *) ptr1, *q2 = (int32_t *) ptr2,
-            *q3 = (int32_t *) ptr3, *q4 = (int32_t *) ptr4;
-  int32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_int32_t *q1 = (ca_u_int32_t *) ptr1, *q2 = (ca_u_int32_t *) ptr2,
+            *q3 = (ca_u_int32_t *) ptr3, *q4 = (ca_u_int32_t *) ptr4;
+  ca_u_int32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -1564,9 +1564,9 @@ ca_triop_select_uint32_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  uint32_t *q1 = (uint32_t *) ptr1, *q2 = (uint32_t *) ptr2,
-            *q3 = (uint32_t *) ptr3, *q4 = (uint32_t *) ptr4;
-  uint32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_uint32_t *q1 = (ca_u_uint32_t *) ptr1, *q2 = (ca_u_uint32_t *) ptr2,
+            *q3 = (ca_u_uint32_t *) ptr3, *q4 = (ca_u_uint32_t *) ptr4;
+  ca_u_uint32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -1598,9 +1598,9 @@ ca_triop_select_int64_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  int64_t *q1 = (int64_t *) ptr1, *q2 = (int64_t *) ptr2,
-            *q3 = (int64_t *) ptr3, *q4 = (int64_t *) ptr4;
-  int64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_int64_t *q1 = (ca_u_int64_t *) ptr1, *q2 = (ca_u_int64_t *) ptr2,
+            *q3 = (ca_u_int64_t *) ptr3, *q4 = (ca_u_int64_t *) ptr4;
+  ca_u_int64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -1632,9 +1632,9 @@ ca_triop_select_uint64_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  uint64_t *q1 = (uint64_t *) ptr1, *q2 = (uint64_t *) ptr2,
-            *q3 = (uint64_t *) ptr3, *q4 = (uint64_t *) ptr4;
-  uint64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_uint64_t *q1 = (ca_u_uint64_t *) ptr1, *q2 = (ca_u_uint64_t *) ptr2,
+            *q3 = (ca_u_uint64_t *) ptr3, *q4 = (ca_u_uint64_t *) ptr4;
+  ca_u_uint64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -1666,9 +1666,9 @@ ca_triop_select_float32_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  float *q1 = (float *) ptr1, *q2 = (float *) ptr2,
-            *q3 = (float *) ptr3, *q4 = (float *) ptr4;
-  float *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_float32_t *q1 = (ca_u_float32_t *) ptr1, *q2 = (ca_u_float32_t *) ptr2,
+            *q3 = (ca_u_float32_t *) ptr3, *q4 = (ca_u_float32_t *) ptr4;
+  ca_u_float32_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -1700,9 +1700,9 @@ ca_triop_select_float64_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  double *q1 = (double *) ptr1, *q2 = (double *) ptr2,
-            *q3 = (double *) ptr3, *q4 = (double *) ptr4;
-  double *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_float64_t *q1 = (ca_u_float64_t *) ptr1, *q2 = (ca_u_float64_t *) ptr2,
+            *q3 = (ca_u_float64_t *) ptr3, *q4 = (ca_u_float64_t *) ptr4;
+  ca_u_float64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -1734,9 +1734,9 @@ ca_triop_select_cmplx64_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  cmplx64_t *q1 = (cmplx64_t *) ptr1, *q2 = (cmplx64_t *) ptr2,
-            *q3 = (cmplx64_t *) ptr3, *q4 = (cmplx64_t *) ptr4;
-  cmplx64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_cmplx64_t *q1 = (ca_u_cmplx64_t *) ptr1, *q2 = (ca_u_cmplx64_t *) ptr2,
+            *q3 = (ca_u_cmplx64_t *) ptr3, *q4 = (ca_u_cmplx64_t *) ptr4;
+  ca_u_cmplx64_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;
@@ -1768,9 +1768,9 @@ ca_triop_select_cmplx128_t (ca_size_t n, boolean8_t *m,
                                  char *ptr3, ca_size_t i3,
                                  char *ptr4, ca_size_t i4)
 {
-  cmplx128_t *q1 = (cmplx128_t *) ptr1, *q2 = (cmplx128_t *) ptr2,
-            *q3 = (cmplx128_t *) ptr3, *q4 = (cmplx128_t *) ptr4;
-  cmplx128_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
+  ca_u_cmplx128_t *q1 = (ca_u_cmplx128_t *) ptr1, *q2 = (ca_u_cmplx128_t *) ptr2,
+            *q3 = (ca_u_cmplx128_t *) ptr3, *q4 = (ca_u_cmplx128_t *) ptr4;
+  ca_u_cmplx128_t *p1 = q1, *p2 = q2, *p3 = q3, *p4 = q4;
   ca_size_t k;
   if ( m ) {
     boolean8_t *pm;

data/ext/carray_key_windows.c CHANGED Viewed

@@ -150,6 +150,10 @@ rb_ca_s_key_window_ranges (VALUE self, VALUE vkeys, VALUE vgroups,
        ( ck->data_type != CA_INT64 && ck->data_type != CA_FLOAT64 ) ) {
     rb_raise(rb_eArgError, "__find_key_window_row_ranges__: unexpected arguments");
   }
+  /* The completeness test takes the key difference modulo the interval. */
+  if ( !NIL_P(rinterval) && !(NUM2DBL(rinterval) > 0) ) {
+    rb_raise(rb_eArgError, "__find_key_window_row_ranges__: interval should be positive");
+  }
   n = ck->elements;
   grp = (const int64_t *) cg->ptr;
   nout = n;

data/ext/carray_mask.c CHANGED Viewed

@@ -37,16 +37,32 @@
 #define CA_MASK_WORD_BYTES 8
 #define CA_MASK_WORD_LSB   0x0101010101010101ULL
+/* A mask may start at any byte (a window or a block view offsets it), so a
+   word is read and written through memcpy, never through a uint64_t
+   pointer; the compiler still lowers each to a single load or store.  */
+static inline uint64_t
+ca_mask_word_load (const boolean8_t *p)
+{
+  uint64_t w;
+  memcpy(&w, p, sizeof(w));
+  return w;
+}
+static inline void
+ca_mask_word_store (boolean8_t *p, uint64_t w)
+{
+  memcpy(p, &w, sizeof(w));
+}
 /* ma[i] |= ms[i] for i in [0, n)  -- used by ca_copy_mask_overlay_n */
 static inline void
 ca_mask_word_or (boolean8_t *ma, const boolean8_t *ms, ca_size_t n)
 {
   ca_size_t       nw = n / CA_MASK_WORD_BYTES;
   ca_size_t       i;
-  uint64_t       *wa = (uint64_t *)       ma;
-  const uint64_t *ws = (const uint64_t *) ms;
   for (i = 0; i < nw; i++) {
-    wa[i] |= ws[i];
+    ca_mask_word_store(ma + i*CA_MASK_WORD_BYTES,
+                       ca_mask_word_load(ma + i*CA_MASK_WORD_BYTES) | ca_mask_word_load(ms + i*CA_MASK_WORD_BYTES));
   }
   for (i = nw * CA_MASK_WORD_BYTES; i < n; i++) {
     ma[i] |= ms[i];
@@ -62,11 +78,9 @@ ca_mask_word_and (boolean8_t       *md,
 {
   ca_size_t       nw = n / CA_MASK_WORD_BYTES;
   ca_size_t       i;
-  uint64_t       *wd = (uint64_t *)       md;
-  const uint64_t *wa = (const uint64_t *) ma;
-  const uint64_t *wb = (const uint64_t *) mb;
   for (i = 0; i < nw; i++) {
-    wd[i] = wa[i] & wb[i];
+    ca_mask_word_store(md + i*CA_MASK_WORD_BYTES,
+                       ca_mask_word_load(ma + i*CA_MASK_WORD_BYTES) & ca_mask_word_load(mb + i*CA_MASK_WORD_BYTES));
   }
   for (i = nw * CA_MASK_WORD_BYTES; i < n; i++) {
     md[i] = ma[i] & mb[i];
@@ -82,11 +96,9 @@ ca_mask_word_xor (boolean8_t       *md,
 {
   ca_size_t       nw = n / CA_MASK_WORD_BYTES;
   ca_size_t       i;
-  uint64_t       *wd = (uint64_t *)       md;
-  const uint64_t *wa = (const uint64_t *) ma;
-  const uint64_t *wb = (const uint64_t *) mb;
   for (i = 0; i < nw; i++) {
-    wd[i] = wa[i] ^ wb[i];
+    ca_mask_word_store(md + i*CA_MASK_WORD_BYTES,
+                       ca_mask_word_load(ma + i*CA_MASK_WORD_BYTES) ^ ca_mask_word_load(mb + i*CA_MASK_WORD_BYTES));
   }
   for (i = nw * CA_MASK_WORD_BYTES; i < n; i++) {
     md[i] = ma[i] ^ mb[i];
@@ -99,10 +111,8 @@ ca_mask_word_not (boolean8_t *md, const boolean8_t *ms, ca_size_t n)
 {
   ca_size_t       nw = n / CA_MASK_WORD_BYTES;
   ca_size_t       i;
-  uint64_t       *wd = (uint64_t *)       md;
-  const uint64_t *ws = (const uint64_t *) ms;
   for (i = 0; i < nw; i++) {
-    wd[i] = ws[i] ^ CA_MASK_WORD_LSB;
+    ca_mask_word_store(md + i*CA_MASK_WORD_BYTES, ca_mask_word_load(ms + i*CA_MASK_WORD_BYTES) ^ CA_MASK_WORD_LSB);
   }
   for (i = nw * CA_MASK_WORD_BYTES; i < n; i++) {
     md[i] = ms[i] ^ (boolean8_t) 1;
@@ -116,9 +126,8 @@ ca_mask_word_any (const boolean8_t *m, ca_size_t n)
 {
   ca_size_t       nw = n / CA_MASK_WORD_BYTES;
   ca_size_t       i;
-  const uint64_t *w  = (const uint64_t *) m;
   for (i = 0; i < nw; i++) {
-    if ( w[i] != 0 ) return 1;
+    if ( ca_mask_word_load(m + i*CA_MASK_WORD_BYTES) != 0 ) return 1;
   }
   for (i = nw * CA_MASK_WORD_BYTES; i < n; i++) {
     if ( m[i] ) return 1;
@@ -134,9 +143,8 @@ ca_mask_word_all_set (const boolean8_t *m, ca_size_t n)
 {
   ca_size_t       nw = n / CA_MASK_WORD_BYTES;
   ca_size_t       i;
-  const uint64_t *w  = (const uint64_t *) m;
   for (i = 0; i < nw; i++) {
-    if ( w[i] != CA_MASK_WORD_LSB ) return 0;
+    if ( ca_mask_word_load(m + i*CA_MASK_WORD_BYTES) != CA_MASK_WORD_LSB ) return 0;
   }
   for (i = nw * CA_MASK_WORD_BYTES; i < n; i++) {
     if ( ! m[i] ) return 0;
@@ -1872,6 +1880,19 @@ rb_ca_mask_where (int argc, VALUE *argv, VALUE self)
   return ca_mask_copy_where(self, argc, argv);
 }
+/* inherit_mask's sources are read cell by cell over self, so an array
+   source has self's number of cells (a scalar's one cell applies to all). */
+static void
+ca_inherit_mask_check_source (CArray *ca, CArray *cs)
+{
+  if ( ! ca_is_scalar(cs) && cs->elements != ca->elements ) {
+    rb_raise(rb_eArgError,
+             "%s: mask source has %lld elements, the receiver %lld",
+             ca_calling_method_name(),
+             (long long) cs->elements, (long long) ca->elements);
+  }
+}
 /* @overload inherit_mask (*others):
 (Masking, Destructive)
@@ -1895,6 +1916,7 @@ rb_ca_inherit_mask_method (int argc, VALUE *argv, VALUE self)
   for (i=0; i<argc; i++) {
     if ( rb_obj_is_carray(argv[i]) ) {
       TypedData_Get_Struct(argv[i], CArray, &carray_data_type, cs);
+      ca_inherit_mask_check_source(ca, cs);
       slist[i] = cs;
     }
     else {
@@ -1938,6 +1960,7 @@ rb_ca_inherit_mask (VALUE self, int n, ...)
     other = va_arg(rothers, VALUE);
     if ( rb_obj_is_carray(other) ) {
       TypedData_Get_Struct(other, CArray, &carray_data_type, cs);
+      ca_inherit_mask_check_source(ca, cs);
       slist[i] = cs;
     }
     else {
@@ -1976,6 +1999,7 @@ rb_ca_inherit_mask_replace_method (int argc, VALUE *argv, VALUE self)
   for (i=0; i<argc; i++) {
     if ( rb_obj_is_carray(argv[i]) ) {
       TypedData_Get_Struct(argv[i], CArray, &carray_data_type, cs);
+      ca_inherit_mask_check_source(ca, cs);
       slist[i] = cs;
     }
     else {
@@ -2019,6 +2043,7 @@ rb_ca_inherit_mask_replace (VALUE self, int n, ...)
     other = va_arg(rothers, VALUE);
     if ( rb_obj_is_carray(other) ) {
       TypedData_Get_Struct(other, CArray, &carray_data_type, cs);
+      ca_inherit_mask_check_source(ca, cs);
       slist[i] = cs;
     }
     else {
@@ -2266,11 +2291,9 @@ ca_binop_bit_or_i_boolean8_fast (ca_size_t n, boolean8_t *m,
        accumulating; here we need 3-arg). */
     ca_size_t       nw = n / CA_MASK_WORD_BYTES;
     ca_size_t       j;
-    uint64_t       *wd = (uint64_t *)       q3;
-    const uint64_t *wa = (const uint64_t *) q1;
-    const uint64_t *wb = (const uint64_t *) q2;
     for (j = 0; j < nw; j++) {
-      wd[j] = wa[j] | wb[j];
+      ca_mask_word_store(q3 + j*CA_MASK_WORD_BYTES,
+                         ca_mask_word_load(q1 + j*CA_MASK_WORD_BYTES) | ca_mask_word_load(q2 + j*CA_MASK_WORD_BYTES));
     }
     for (j = nw * CA_MASK_WORD_BYTES; j < n; j++) {
       q3[j] = q1[j] | q2[j];