carray 3.0.2 → 3.0.4

data/ext/carray_cast_func.rb

@@ -1,4 +1,5 @@
 data_type = []
+type_const = {}
 %w[
   CA_FIXLEN
   CA_BOOLEAN
@@ -20,6 +21,7 @@ data_type = []
 ].each_with_index do |name, i|
   eval("#{name} = #{i}")
   data_type.push(i)
+  type_const[i] = name
 end
 
 ctype = {
@@ -107,8 +109,8 @@ FLOAT = [
   CA_FLOAT64  ,
 ]
 
-# object -> int/float cast: a parse failure masks the output cell (UNDEF)
-# instead of a silent 0.0 (float) or a raise (int).  These parsers report
+# object -> int/float/complex cast: a parse failure masks the output cell
+# (UNDEF) instead of raising.  These parsers report
 # failure through an out-param whose C type is `oktype`.
 obj2cval_ok = {
   CA_INT8     => "rb_obj2long_ok",
@@ -121,6 +123,8 @@ obj2cval_ok = {
   CA_UINT64   => "rb_obj2ull_ok",
   CA_FLOAT32  => "ca_obj2dbl_ok",
   CA_FLOAT64  => "ca_obj2dbl_ok",
+  CA_CMPLX64  => "ca_obj2cmplx_ok",
+  CA_CMPLX128 => "ca_obj2cmplx_ok",
 }
 
 oktype = {
@@ -134,6 +138,8 @@ oktype = {
   CA_UINT64   => "unsigned long long",
   CA_FLOAT32  => "double",
   CA_FLOAT64  => "double",
+  CA_CMPLX64  => "double complex",
+  CA_CMPLX128 => "double complex",
 }
 
 BOOLEAN = [
@@ -172,6 +178,7 @@ puts %{
 ---------------------------------------------------------------------------- */
 
 #include "carray.h"
+#include "carray_internal.h"   /* ca_decimal_string_to / _store */
 
 static void __attribute__((noreturn))
 ca_cast_not_implemented(ca_size_t n, CArray *a1, void *ptr1, CArray *a2, void *ptr2, boolean8_t *m)
@@ -295,7 +302,8 @@ puts
                  *p2 = q;
                }
                else {
-                 rb_raise(rb_eRuntimeError, "out of range to cast to boolean (0 or 1)");
+                 rb_raise(rb_eRangeError, "%s: out of range to cast to boolean (0 or 1)",
+                          ca_calling_method_name());
                }
              }
              p1++; p2++; m++;
@@ -308,7 +316,8 @@ puts
                *p2 = q;
              }
              else {
-               rb_raise(rb_eRuntimeError, "out of range to cast to boolean (0 or 1)");
+               rb_raise(rb_eRangeError, "%s: out of range to cast to boolean (0 or 1)",
+                          ca_calling_method_name());
              }
              p1++; p2++; 
            }
@@ -327,7 +336,34 @@ puts
     ctype1 = ctype[type1]
     ctype2 = ctype[type2]
     CA_CAST_TABLE[type1][type2] = "ca_cast_#{ctype1}_#{ctype2}"
-    puts <<-END_DEF  .gsub(/^ {6}/, '')
+    # float / complex -> integer: a NaN has no integer and C leaves its
+    # conversion undefined.  A cell that is not masked refuses it before
+    # anything is written.  #to_type masks NaN cells first (UNDEF), so
+    # only a store, a view and a single cell reach the raise.
+    nan_check = ""
+    if (FLOAT+[CA_CMPLX64, CA_CMPLX128]).include?(type1) && INTEGER.include?(type2)
+      isnan = if [CA_CMPLX64, CA_CMPLX128].include?(type1)
+                "( isnan(creal(q1[k])) || isnan(cimag(q1[k])) )"
+              else
+                "isnan(q1[k])"
+              end
+      nan_check = <<-END_CHECK.gsub(/^ {6}/, '')
+         for (k=0; k<n; k++) {
+           if ( ( ! m || ! m[k] ) && #{isnan} ) {
+             rb_raise(rb_eFloatDomainError, "%s: NaN cannot be converted to %s",
+                      ca_calling_method_name(), ca_type_name[#{type_const[type2]}]);
+           }
+         }
+      END_CHECK
+    end
+    # int64 / uint64 -> float32 / cmplx64 goes through double, as a Ruby
+    # Integer stored into a float32 cell does (Integer#to_f).  Converting in
+    # one step rounds once where the compiler keeps the loop scalar and twice
+    # where it vectorises it through double, so the answer for a value past
+    # 2**53 would depend on its place in the array.
+    via = ( [CA_INT64, CA_UINT64].include?(type1) &&
+            [CA_FLOAT32, CA_CMPLX64].include?(type2) ) ? "(double) " : ""
+    puts <<-END_DEF.gsub(/^ {6}/, '').sub("@@NAN_CHECK@@\n", nan_check).gsub("@@VIA@@", via)
       static void
       ca_cast_#{ctype1}_#{ctype2}(ca_size_t n, CArray *a1, void *ptr1, CArray *a2, void *ptr2, boolean8_t *m)
       {
@@ -336,6 +372,7 @@ puts
          #{ctype2} *q2 = ptr2;
          #{ctype2} *p2 = q2;
          ca_size_t k;
+      @@NAN_CHECK@@
          if ( m ) {
            boolean8_t *pm = m;
            for (k=0; k<n; k++) {
@@ -343,7 +380,7 @@ puts
              if ( ! *pm ) { 
                p1 = q1 + k;
                p2 = q2 + k;
-               *p2 = (#{ctype2}) *p1;
+               *p2 = (#{ctype2}) @@VIA@@*p1;
              }
            }
          }
@@ -351,7 +388,7 @@ puts
            for (k=0; k<n; k++) {
              p1 = q1 + k;
              p2 = q2 + k;
-             *p2 = (#{ctype2}) *p1;
+             *p2 = (#{ctype2}) @@VIA@@*p1;
            }
          }
          return;
@@ -456,17 +493,38 @@ OBJECT.each do |type1|
     ctype1 = ctype[type1]
     ctype2 = ctype[type2]
     CA_CAST_TABLE[type1][type2] = "ca_cast_#{ctype1}_#{ctype2}"
-    if (INTEGER+FLOAT).include?(type2)
-      # object -> int/float: parse failure -> masked (UNDEF) output cell.
-      # The mask-aware loop runs only when a mask buffer is supplied (m);
-      # #to_type always supplies one for these targets so the whole
-      # user-facing object->numeric surface gets the masking semantics.
-      # With no mask buffer (internal single-cell / view paths) the legacy
-      # value parser keeps its old behaviour (lenient 0.0 float / raising
-      # int), leaving those paths byte-identical.
+    if (INTEGER+FLOAT+[CA_CMPLX64, CA_CMPLX128]).include?(type2)
+      # object -> int/float/complex: parse failure -> masked (UNDEF) output
+      # cell.  The mask-aware loop runs only when a mask buffer is supplied
+      # (m); #to_type always supplies one for these targets.  With no mask
+      # buffer (a store, single-cell and view paths) the value parser
+      # raises for a cell it cannot read.
       conv_ok = obj2cval_ok[type2]
       conv    = obj2cval[type2]
       oktype2 = oktype[type2]
+      # A String cell of an integer or float target is read as a decimal
+      # number (ca_decimal_string_to / _store in carray_cast.c), not as a
+      # Ruby literal. Complex keeps its own reader.
+      if (INTEGER+FLOAT).include?(type2)
+        type_name = "CA_" + ctype2.sub(/_t\z/, "").upcase
+        read_ok   = "( RB_TYPE_P(*p1, T_STRING) ? " \
+                    "ca_decimal_string_to(*p1, #{type_name}, p2) : " \
+                    "( #{conv_ok}(*p1, &tmp) ? ( *p2 = (#{ctype2}) tmp, 1 ) : 0 ) )"
+        # A store refuses an integer the cell cannot hold
+        # (ca_obj2integer_fitting); to_type, above, keeps the C conversion.
+        store_conv = if INTEGER.include?(type2)
+                       num = [CA_UINT8, CA_UINT16, CA_UINT32, CA_UINT64].include?(type2) ? "rb_num2ull" : "NUM2LL"
+                       "#{num}(ca_obj2integer_fitting(*p1, #{type_name}))"
+                     else
+                       "#{conv}(*p1)"
+                     end
+        read      = "if ( RB_TYPE_P(*p1, T_STRING) ) { " \
+                    "ca_decimal_string_store(*p1, #{type_name}, p2); } " \
+                    "else { *p2 = (#{ctype2}) #{store_conv}; }"
+      else
+        read_ok   = "( #{conv_ok}(*p1, &tmp) ? ( *p2 = (#{ctype2}) tmp, 1 ) : 0 )"
+        read      = "*p2 = (#{ctype2}) #{conv}(*p1);"
+      end
       puts <<-END_DEF  .gsub(/^ {6}/, '')
         static void
         ca_cast_#{ctype1}_#{ctype2}(ca_size_t n, CArray *a1, void *ptr1, CArray *a2, void *ptr2, boolean8_t *m)
@@ -476,20 +534,15 @@ OBJECT.each do |type1|
            #{oktype2} tmp;
            if ( m ) {
              while ( n-- ) {
-               if ( !*m ) {
-                 if ( #{conv_ok}(*p1, &tmp) ) {
-                   *p2 = (#{ctype2}) tmp;
-                 }
-                 else {
-                   *m = 1;
-                   *p2 = 0;
-                 }
+               if ( !*m && ! #{read_ok} ) {
+                 *m = 1;
+                 *p2 = 0;
                }
                p1++; p2++; m++;
              }
            }
            else {
-             while ( n-- ) { *p2 = (#{ctype2}) #{conv}(*p1); p1++; p2++; }
+             while ( n-- ) { #{read} p1++; p2++; }
            }
            return;
         }

data/ext/carray_class.c

@@ -6,6 +6,7 @@
 ---------------------------------------------------------------------------- */
 
 #include "carray.h"
+#include "carray_internal.h"   /* CA_CHECK_*_GIVEN */
 
 /* ------------------------------------------------------------------- */
 /* Symbol cache for data_type code → Symbol lookup.
@@ -25,7 +26,7 @@ ID ca_data_type_sym[CA_NTYPE];
 VALUE
 rb_ca_data_type_to_sym (int8_t data_type)
 {
-  CA_CHECK_DATA_TYPE(data_type);
+  CA_CHECK_DATA_TYPE_GIVEN(data_type);
   return ID2SYM(ca_data_type_sym[data_type]);
 }
 
@@ -69,7 +70,7 @@ static VALUE
 rb_ca_s_data_type_name (VALUE klass, VALUE type)
 {
   int8_t data_type = rb_ca_guess_type(type);
-  CA_CHECK_DATA_TYPE(data_type);
+  CA_CHECK_DATA_TYPE_GIVEN(data_type);
   return rb_str_new2(ca_type_name[data_type]);
 }
 
@@ -84,7 +85,7 @@ static VALUE
 rb_ca_s_data_type_code (VALUE klass, VALUE type)
 {
   int8_t data_type = rb_ca_guess_type(type);
-  CA_CHECK_DATA_TYPE(data_type);
+  CA_CHECK_DATA_TYPE_GIVEN(data_type);
   return INT2NUM(data_type);
 }
 

data/ext/carray_conversion.c

@@ -12,8 +12,11 @@
 ---------------------------------------------------------------------------- */
 
 #include "carray.h"
+#include "carray_internal.h"   /* ca_attach_window */
 
 #include "ruby/io.h"
+#include "ruby/util.h"            /* ruby_strtod */
+#include <math.h>
 
 /* to_a dispatch:
      numeric / boolean data type + entity or attach-able view -> fast path
@@ -175,19 +178,34 @@ rb_ca_to_a_loop_universal (VALUE self, int32_t level, ca_size_t *idx, VALUE ary)
  * fast path, so all view kinds land in the data type-direct leaf loop when
  * the data type is numeric or boolean; CA_OBJECT / CA_FIXLEN fall through
  * to the universal fetch_index walk. */
+/* The universal walk reads each cell through the fetch path, which
+   decodes -- an object cell, a Face's element -- and can raise, so it runs
+   inside an attach window.  The fast path only copies numbers out. */
+typedef struct {
+  VALUE self;
+  VALUE ary;
+} ca_to_a_ctx_t;
+
+static VALUE
+ca_to_a_universal (VALUE arg)
+{
+  ca_to_a_ctx_t *c = (ca_to_a_ctx_t *) arg;
+  ca_size_t idx[CA_RANK_MAX];
+  rb_ca_to_a_loop_universal(c->self, 0, idx, c->ary);
+  return Qnil;
+}
+
 VALUE
 rb_ca_to_a (VALUE self)
 {
   volatile VALUE ary;
   CArray *ca;
-  ca_size_t idx[CA_RANK_MAX];
   int data_type, fast_eligible;
 
   TypedData_Get_Struct(self, CArray, &carray_data_type, ca);
   ary = rb_ca_is_empty(self) ? rb_ary_new() : rb_ary_new2(ca->dim[0]);
-  ca_attach(ca);
 
-  /* After ca_attach, ca->ptr is a contig row-major buffer (alias for
+  /* Once attached, ca->ptr is a contig row-major buffer (alias for
      contig views, materialised otherwise) and ca->mask matches, so the
      fast path handles every view kind.  CA_OBJECT / CA_FIXLEN stay on
      the universal path because their ptr is VALUE / byte string and
@@ -210,16 +228,21 @@ rb_ca_to_a (VALUE self)
                       || data_type == CA_BOOLEAN);
 
   if ( fast_eligible ) {
-    int has_mask = (ca->mask != NULL);
-    const boolean8_t *mask_base = has_mask
-        ? (const boolean8_t *) ca->mask->ptr : NULL;
+    int has_mask;
+    const boolean8_t *mask_base;
+    ca_attach(ca);   /* window: nothing raises inside */
+    has_mask  = (ca->mask != NULL);
+    mask_base = has_mask ? (const boolean8_t *) ca->mask->ptr : NULL;
     to_a_fast_recurse(ca, 0, ca->ptr, mask_base, ary,
                       data_type, ca->bytes, has_mask);
+    ca_detach(ca);
   }
   else if ( ca->elements > 0 ) {
-    rb_ca_to_a_loop_universal(self, 0, idx, ary);
+    ca_to_a_ctx_t c;
+    c.self = self;
+    c.ary  = ary;
+    ca_attach_window(1, &ca, "r", ca_to_a_universal, (VALUE) &c);
   }
-  ca_detach(ca);
   return ary;
 }
 
@@ -228,65 +251,44 @@ rb_ca_to_a (VALUE self)
 #undef TO_A_DISPATCH_LEAF
 #undef TO_A_DISPATCH_LEAF_MASKED
 
-/* CArray#convert(data_type=nil, bytes: nil) { |elem| ... } — map every cell
- * through the block, into a new array whose data type is chosen by argv via
- * CArray#template (shape is inherited from self).
- *
- * Masked cells skip the block; if the block returns CA_UNDEF, the
- * corresponding output cell is masked instead of stored.  Both self
- * and the template result are attached symmetrically (R3-compliant) so
- * an overridden #template returning a view stays on the in-ptr path. */
+/* convert yields inside the window over self and the output, so the loop
+   runs through ca_attach_window.  An output mask made part way (the first
+   UNDEF from the block) is attached inside the loop, and closed by an
+   ensure of its own. */
+
+typedef struct {
+  VALUE    self;
+  VALUE    obj;
+  CArray  *ca;
+  CArray  *co;
+  int      has_mask;
+  int      co_mask_attached;
+} ca_convert_ctx_t;
+
 static VALUE
-rb_ca_convert (int argc, VALUE *argv, VALUE self)
+ca_convert_loop (VALUE arg)
 {
-  volatile VALUE obj;
-  CArray *ca, *co;
-  ca_size_t i, n, sbytes, dbytes;
-  int has_mask;
-  const boolean8_t *src_mask = NULL;
+  ca_convert_ctx_t *c = (ca_convert_ctx_t *) arg;
+  CArray *ca = c->ca, *co = c->co;
+  ca_size_t i, n = ca->elements;
+  ca_size_t sbytes = ca->bytes, dbytes = co->bytes;
+  const char *src_ptr = ca->ptr;
+  char *dst_ptr = co->ptr;
   boolean8_t *dst_mask = NULL;
-  const char *src_ptr;
-  char *dst_ptr;
 
-  rb_need_block();
-
-  obj = rb_apply(self, rb_intern("template"), rb_ary_new4(argc, argv));
-
-  TypedData_Get_Struct(self, CArray, &carray_data_type, ca);
-  TypedData_Get_Struct(obj, CArray, &carray_data_type, co);
-
-  n = ca->elements;
-  has_mask = ca_has_mask(ca);
-  sbytes = ca->bytes;
-  dbytes = co->bytes;
-
-  /* Attach every storage touched by ptr (self, obj, and either side's
-     mask) — parent → child order per the R4 attach protocol. */
-  ca_attach(ca);
-  ca_attach(co);
-  if ( has_mask ) {
-    ca_attach(ca->mask);
-    if ( ! co->mask ) {
-      ca_create_mask(co);
-    }
-    ca_attach(co->mask);
-    src_mask = (const boolean8_t *) ca->mask->ptr;
+  if ( c->has_mask ) {
+    const boolean8_t *src_mask = (const boolean8_t *) ca->mask->ptr;
     dst_mask = (boolean8_t *) co->mask->ptr;
     memset(dst_mask, 0, (size_t) n);
-  }
-  src_ptr = ca->ptr;
-  dst_ptr = co->ptr;
-
-  if ( has_mask ) {
     for (i = 0; i < n; i++) {
       if ( ! src_mask[i] ) {
-        VALUE val = rb_ca_ptr2obj(self, (void *) (src_ptr + i * sbytes));
+        VALUE val = rb_ca_ptr2obj(c->self, (void *) (src_ptr + i * sbytes));
         VALUE ret = rb_yield(val);
         if ( ret == CA_UNDEF ) {
           dst_mask[i] = 1;
         }
         else {
-          rb_ca_obj2ptr(obj, ret, dst_ptr + i * dbytes);
+          rb_ca_obj2ptr(c->obj, ret, dst_ptr + i * dbytes);
         }
       }
       else {
@@ -296,33 +298,78 @@ rb_ca_convert (int argc, VALUE *argv, VALUE self)
   }
   else {
     for (i = 0; i < n; i++) {
-      VALUE val = rb_ca_ptr2obj(self, (void *) (src_ptr + i * sbytes));
+      VALUE val = rb_ca_ptr2obj(c->self, (void *) (src_ptr + i * sbytes));
       VALUE ret = rb_yield(val);
       if ( ret == CA_UNDEF ) {
         if ( ! co->mask ) {
           ca_create_mask(co);
-          ca_attach(co->mask);
+          ca_attach(co->mask);   /* window: closed by the ensure in ca_convert_body */
+          c->co_mask_attached = 1;
           dst_mask = (boolean8_t *) co->mask->ptr;
           memset(dst_mask, 0, (size_t) n);
         }
         dst_mask[i] = 1;
       }
       else {
-        rb_ca_obj2ptr(obj, ret, dst_ptr + i * dbytes);
+        rb_ca_obj2ptr(c->obj, ret, dst_ptr + i * dbytes);
       }
     }
   }
+  return Qnil;
+}
 
-  if ( co->mask ) {
-    ca_sync(co->mask);
-    ca_detach(co->mask);
+static VALUE
+ca_convert_close_mask (VALUE arg)
+{
+  ca_convert_ctx_t *c = (ca_convert_ctx_t *) arg;
+  if ( c->co_mask_attached ) {
+    c->co_mask_attached = 0;
+    ca_sync_detach(c->co->mask);
   }
-  if ( has_mask ) {
-    ca_detach(ca->mask);
+  return Qnil;
+}
+
+static VALUE
+ca_convert_body (VALUE arg)
+{
+  return rb_ensure(ca_convert_loop, arg, ca_convert_close_mask, arg);
+}
+
+/* CArray#convert(data_type=nil, bytes: nil) { |elem| ... } — map every cell
+ * through the block, into a new array whose data type is chosen by argv via
+ * CArray#template (shape is inherited from self).
+ *
+ * Masked cells skip the block; if the block returns CA_UNDEF, the
+ * corresponding output cell is masked instead of stored.  Both self
+ * and the template result are attached symmetrically (R3-compliant) so
+ * an overridden #template returning a view stays on the in-ptr path. */
+static VALUE
+rb_ca_convert (int argc, VALUE *argv, VALUE self)
+{
+  volatile VALUE obj;
+  ca_convert_ctx_t c;
+  CArray *list[2];
+
+  rb_need_block();
+
+  obj = rb_apply(self, rb_intern("template"), rb_ary_new4(argc, argv));
+
+  TypedData_Get_Struct(self, CArray, &carray_data_type, c.ca);
+  TypedData_Get_Struct(obj, CArray, &carray_data_type, c.co);
+  c.self             = self;
+  c.obj              = obj;
+  c.has_mask         = ca_has_mask(c.ca);
+  c.co_mask_attached = 0;
+
+  /* Attaching an array attaches its mask, so the output's mask is made
+     before the window opens when the source already has one. */
+  if ( c.has_mask && ! c.co->mask ) {
+    ca_create_mask(c.co);
   }
-  ca_sync(co);
-  ca_detach(co);
-  ca_detach(ca);
+
+  list[0] = c.ca;
+  list[1] = c.co;
+  ca_attach_window(2, list, "rw", ca_convert_body, (VALUE) &c);
 
   return obj;
 }
@@ -395,7 +442,7 @@ rb_ca_dump_binary (int argc, VALUE *argv, VALUE self)
       rb_funcall(io, rb_intern("write"), 1, buf);
     }
     else {
-      rb_raise(rb_eRuntimeError, "IO like object should have 'write' method");
+      rb_raise(rb_eTypeError, "IO like object should have 'write' method");
     }
   }
 
@@ -422,6 +469,7 @@ rb_ca_load_binary (VALUE self, VALUE io)
 {
   CArray *ca;
 
+  rb_ca_modify(self);
   TypedData_Get_Struct(self, CArray, &carray_data_type, ca);
 
   if ( ca_is_object_type(ca) ) {
@@ -431,11 +479,17 @@ rb_ca_load_binary (VALUE self, VALUE io)
   switch ( TYPE(io) ) {
   case T_STRING:
     if ( ca_length(ca) != RSTRING_LEN(io) ) {
-      rb_raise(rb_eRuntimeError,
-               "data size mismatch (got %" PRId64 " bytes, expected %" PRId64 ")",
-               (ca_size_t) RSTRING_LEN(io), (ca_size_t) ca_length(ca));
+      rb_raise(rb_eArgError,
+               "%s: data size mismatch (got %" PRId64 " bytes, expected %" PRId64 ")",
+               ca_calling_method_name(), (ca_size_t) RSTRING_LEN(io), (ca_size_t) ca_length(ca));
+    }
+    if ( ca->data_type == CA_BOOLEAN ) {
+      io = rb_str_dup(io);
+      rb_str_modify(io);
+      ca_boolean_normalize_bytes(RSTRING_PTR(io), ca->elements);
     }
     ca_sync_data(ca, RSTRING_PTR(io));
+    RB_GC_GUARD(io);
     return self;
     break;
   default:
@@ -447,13 +501,13 @@ rb_ca_load_binary (VALUE self, VALUE io)
         const ca_size_t CHUNK = 65536;
         volatile VALUE scratch = rb_str_buf_new(CHUNK < total ? CHUNK : total);
         ca_size_t off = 0;
-        ca_allocate(ca);
+        ca_allocate(ca);   /* window: an entity, whose allocate holds nothing */
         while ( off < total ) {
           ca_size_t want = (total - off < CHUNK) ? (total - off) : CHUNK;
           VALUE r = rb_funcall(io, id_read, 2, SIZE2NUM(want), scratch);
           if ( NIL_P(r) || RSTRING_LEN(r) != want ) {
             ca_detach(ca);
-            rb_raise(rb_eRuntimeError,
+            rb_raise(rb_eArgError,
                      "short read from IO at offset %" PRId64 " "
                      "(got %" PRId64 " bytes, expected %" PRId64 ")",
                      (ca_size_t) off,
@@ -463,15 +517,17 @@ rb_ca_load_binary (VALUE self, VALUE io)
           memcpy(ca->ptr + off, RSTRING_PTR(r), want);
           off += want;
         }
-        ca_sync(ca);
-        ca_detach(ca);
+        if ( ca->data_type == CA_BOOLEAN ) {
+          ca_boolean_normalize_bytes(ca->ptr, ca->elements);
+        }
+        ca_sync_detach(ca);
       }
       else {
         /* ca_sync_data scatters in one pass, so buffer the full read
            and recurse via T_STRING. */
         volatile VALUE buf = rb_funcall(io, id_read, 1, SIZE2NUM(total));
         if ( NIL_P(buf) || RSTRING_LEN(buf) != total ) {
-          rb_raise(rb_eRuntimeError,
+          rb_raise(rb_eArgError,
                    "short read from IO (got %" PRId64 " bytes, expected %" PRId64 ")",
                    (ca_size_t) (NIL_P(buf) ? 0 : RSTRING_LEN(buf)),
                    (ca_size_t) total);
@@ -481,13 +537,89 @@ rb_ca_load_binary (VALUE self, VALUE io)
       return self;
     }
     else {
-      rb_raise(rb_eRuntimeError, "IO like object should have 'read' method");
+      rb_raise(rb_eTypeError, "IO like object should have 'read' method");
     }
   }
 
   return self;
 }
 
+/* The double nearest to the shortest decimal that reads back as x, the
+   float32 value, as from_csv reads text into float32: strtod, then the cast
+   to float.  Its digits are those of the decimal (at most 9, which a double
+   holds), so Float#to_s writes them. */
+static double
+ca_double_of_shortest_decimal_of_float32 (float x)
+{
+  char buf[32];
+  int lo = 1, hi = 9;
+  double best;
+  if ( isnan(x) || isinf(x) ) {
+    return (double) x;
+  }
+  /* Nine significant digits always read back; search for the fewest.  More
+     digits round closer to x, so the digit counts that read back are those
+     from some count up. */
+  best = 0.0;
+  while ( lo < hi ) {
+    int mid = (lo + hi) / 2;
+    double d;
+    snprintf(buf, sizeof(buf), "%.*e", mid - 1, (double) x);
+    d = ruby_strtod(buf, NULL);
+    if ( (float) d == x ) {
+      best = d;
+      hi = mid;
+    }
+    else {
+      lo = mid + 1;
+    }
+  }
+  if ( lo == 9 ) {
+    snprintf(buf, sizeof(buf), "%.8e", (double) x);
+    best = ruby_strtod(buf, NULL);
+  }
+  return best;
+}
+
+/* float32 -> float64 and cmplx64 -> cmplx128, each value replaced by the
+   double its shortest decimal reads as, for to_csv to write that decimal.
+   The mask is kept. */
+static VALUE
+rb_ca_widen_keeping_shortest_decimal (VALUE self)
+{
+  volatile VALUE out, holder = 0;
+  CArray *ca, *co;
+  ca_size_t i, n;
+  int8_t type;
+
+  GetCArray(self, ca);
+  if ( ca->data_type == CA_FLOAT32 ) {
+    type = CA_FLOAT64;
+  }
+  else if ( ca->data_type == CA_CMPLX64 ) {
+    type = CA_CMPLX128;
+  }
+  else {
+    rb_raise(rb_eCADataTypeError, "__widen_keeping_shortest_decimal__ takes float32 or cmplx64");
+  }
+  n = ca->elements * (ca->data_type == CA_CMPLX64 ? 2 : 1);
+  out = rb_carray_new_safe(type, ca->ndim, ca->dim, 0, NULL);
+  GetCArray(out, co);
+  {
+    float *src = ALLOCV_N(float, holder, n > 0 ? n : 1);
+    double *dst = (double *) co->ptr;
+    ca_copy_data(ca, (char *) src);
+    for (i = 0; i < n; i++) {
+      dst[i] = ca_double_of_shortest_decimal_of_float32(src[i]);
+    }
+    ALLOCV_END(holder);
+  }
+  if ( ca_has_mask(ca) ) {
+    ca_copy_mask_overlay(co, co->elements, 1, ca);
+  }
+  return out;
+}
+
 /* [MOVED] str_format (2.0-era per-element formatter) is replaced by the Ruby
    CArray#format / CArray.format surface in lib/carray/methods/string_format.rb,
    which yields a CAString. */
@@ -504,6 +636,7 @@ Init_carray_conversion (void)
   rb_define_method(rb_cCArray, "dump_binary", rb_ca_dump_binary, -1);
   rb_define_method(rb_cCArray, "to_s", rb_ca_to_s, 0);
   rb_define_method(rb_cCArray, "load_binary", rb_ca_load_binary, 1);
+  rb_define_method(rb_cCArray, "__widen_keeping_shortest_decimal__", rb_ca_widen_keeping_shortest_decimal, 0);
 
   /* CAREFUL: do not bind #to_ary — implicit array coercion in Ruby core
      (splat / multiple assignment / Array()) then triggers on any CArray