Optimise guarded memcpy (#449)

* Improve testing of memcpy including adding perf test.

* Change remaining_bytes to be branch free.

Use reciprocal division followed by multiply to remove a branch.

src/test/measuretime.h

@@ -5,17 +5,40 @@
 #include <iostream>
 #include <sstream>
 
-class MeasureTime : public std::stringstream
+class MeasureTime
 {
+  std::stringstream ss;
   std::chrono::time_point<std::chrono::high_resolution_clock> start =
     std::chrono::high_resolution_clock::now();
 
+  bool quiet = false;
+
 public:
   ~MeasureTime()
   {
     auto finish = std::chrono::high_resolution_clock::now();
     auto diff = finish - start;
-    std::cout << str() << ": " << std::setw(12) << diff.count() << " ns"
-              << std::endl;
+    if (!quiet)
+    {
+      std::cout << ss.str() << ": " << std::setw(12) << diff.count() << " ns"
+                << std::endl;
+    }
+  }
+
+  MeasureTime(bool quiet = false) : quiet(quiet) {}
+
+  template<typename T>
+  MeasureTime& operator<<(const T& s)
+  {
+    ss << s;
+    start = std::chrono::high_resolution_clock::now();
+    return *this;
+  }
+
+  std::chrono::nanoseconds get_time()
+  {
+    auto finish = std::chrono::high_resolution_clock::now();
+    auto diff = finish - start;
+    return diff;
   }
 };