diff --git a/lib/unicode_normalize/normalize.rb b/lib/unicode_normalize/normalize.rb
index e67fad187a..7872f8a0bc 100644
--- a/lib/unicode_normalize/normalize.rb
+++ b/lib/unicode_normalize/normalize.rb
@@ -82,16 +82,22 @@ module UnicodeNormalize  # :nodoc:
 
   ## Canonical Ordering
   def self.canonical_ordering_one(string)
-    sorting = string.each_char.collect { |c| [c, CLASS_TABLE[c]] }
-    (sorting.length-2).downto(0) do |i| # almost, but not exactly bubble sort
-      (0..i).each do |j|
-        later_class = sorting[j+1].last
-        if 0<later_class and later_class<sorting[j].last
-          sorting[j], sorting[j+1] = sorting[j+1], sorting[j]
-        end
+    result = ''
+    unordered = []
+    chars = string.chars
+    n = chars.size
+    chars.each_with_index do |char, i|
+      ccc = CLASS_TABLE[char]
+      if ccc == 0
+        unordered.sort!.each { result << chars[it % n] }
+        unordered.clear
+        result << char
+      else
+        unordered << ccc * n + i
       end
     end
-    return sorting.collect(&:first).join('')
+    unordered.sort!.each { result << chars[it % n] }
+    result
   end
 
   ## Normalization Forms for Patterns (not whole Strings)
diff --git a/test/test_unicode_normalize.rb b/test/test_unicode_normalize.rb
index 8789ed92d2..98f46551d3 100644
--- a/test/test_unicode_normalize.rb
+++ b/test/test_unicode_normalize.rb
@@ -209,4 +209,23 @@ class TestUnicodeNormalize
     assert_equal true, ascii_string.unicode_normalized?(:nfkc)
     assert_equal true, ascii_string.unicode_normalized?(:nfkd)
   end
+
+  def test_canonical_ordering
+    a = "\u03B1\u0313\u0300\u0345"
+    a_unordered1 = "\u03B1\u0345\u0313\u0300"
+    a_unordered2 = "\u03B1\u0313\u0345\u0300"
+    u1 = "U\u0308\u0304"
+    u2 = "U\u0304\u0308"
+    s = "s\u0323\u0307"
+    s_unordered = "s\u0307\u0323"
+    o = "\u{1611e}\u{1611e}\u{1611f}"
+    # Actual cases called through String#unicode_normalize
+    assert_equal(s + o, UnicodeNormalize.canonical_ordering_one(s_unordered + o))
+    assert_equal(a[1..], UnicodeNormalize.canonical_ordering_one(a_unordered1[1..]))
+    assert_equal(a[1..] + o, UnicodeNormalize.canonical_ordering_one(a_unordered2[1..] + o))
+    # Artificial cases
+    assert_equal(a + u1 + o + u2 + s, UnicodeNormalize.canonical_ordering_one(a + u1 + o + u2 + s))
+    assert_equal(s[1..] + a + a, UnicodeNormalize.canonical_ordering_one(s_unordered[1..] + a_unordered1 + a_unordered2))
+    assert_equal(o + s + u1 + a + o + a + u2 + o, UnicodeNormalize.canonical_ordering_one(o + s_unordered + u1 + a_unordered1 + o + a_unordered2 + u2 + o))
+  end
 end