Generalize optimization

1 files changed, 54 insertions, 9 deletions

diff --git a/include/unicode.h b/include/unicode.h
index 5774db7..8dedb19 100644
--- a/include/unicode.h
+++ b/include/unicode.h
@@ -548,6 +548,47 @@ namespace unicode {
   return it2 - it1;
  }
 
+ template<int value_size>
+ struct ConvertInputOptimizer {};
+
+ template<> struct ConvertInputOptimizer<1>
+ {
+  static const uint32_t ascii_mask { 0x80808080 };
+ };
+ 
+ template<int value_size>
+ struct ConvertOutputOptimizer {};
+
+ template<> struct ConvertOutputOptimizer<1>
+ {
+  template<typename input_value_type, class output_string_type, int code_units>
+  inline static void append(const input_value_type* addr, output_string_type& s)
+  {
+   s.append(reinterpret_cast<const typename output_string_type::value_type*>(addr), code_units);
+  }
+ };
+ 
+ template<int AccuSize, class ConvertInputOptimizer, class ConvertOutputOptimizer>
+ struct ArchitectureOptimizer {};
+
+ template<class ConvertInputOptimizer, class ConvertOutputOptimizer>
+ struct ArchitectureOptimizer<8, ConvertInputOptimizer, ConvertOutputOptimizer>
+ {
+  typedef ConvertInputOptimizer input_optimizer;
+  typedef ConvertOutputOptimizer output_optimizer;
+  typedef uint64_t accu_type;
+  static const size_t accu_size {8};
+  static const accu_type addr_mask {accu_size - 1};
+  static const accu_type ascii_mask { ((accu_type)input_optimizer::ascii_mask) << 32 | (accu_type)input_optimizer::ascii_mask };
+  static const accu_type ascii_value { 0ULL };
+  
+  template<typename input_value_type, class output_string_type>
+  inline static void append(const input_value_type* addr, output_string_type& s)
+  {
+   output_optimizer::template append<input_value_type, output_string_type, accu_size>(addr, s);
+  }
+ };
+
  // From and To are facets
  template<typename From, typename To, std::enable_if_t<std::is_empty<From>::value, bool> = true>
  typename To::string_type convert(const typename From::string_type& s)
@@ -557,17 +598,21 @@ namespace unicode {
   if constexpr(sizeof(typename From::string_type::value_type) == 1 &&
                sizeof(typename To::value_type) == 1 &&
                sizeof(size_t) >= 8) {
+   typedef ConvertInputOptimizer<sizeof(typename From::string_type::value_type)> input_optimizer;
+   typedef ConvertOutputOptimizer<sizeof(typename To::value_type)> output_optimizer;
+   typedef ArchitectureOptimizer<sizeof(size_t), input_optimizer, output_optimizer> arch_optimizer;
+
    auto begin{From::begin(s)};
    auto end{From::end(s)};
    auto back_inserter{To::back_inserter(result)};
-   auto addr{reinterpret_cast<const uint64_t*>(&s.data()[s.size() - input_distance(begin, end)])};
-   while (input_distance(begin, end) >= 8) {
-    if (((uintptr_t)(void*)addr & 7) == 0) {
-     while (input_distance(begin, end) >= 8) {
-      uint64_t data{*addr};
-      if ((data & 0x8080808080808080ULL) == 0ULL) {
-       result.append(reinterpret_cast<const typename To::value_type*>(addr), 8);
-       begin += 8;
+   auto addr{reinterpret_cast<const typename arch_optimizer::accu_type*>(&s.data()[s.size() - input_distance(begin, end)])};
+   while (input_distance(begin, end) >= arch_optimizer::accu_size) {
+    if (((uintptr_t)(void*)addr & arch_optimizer::addr_mask) == 0) {
+     while (input_distance(begin, end) >= arch_optimizer::accu_size) {
+      typename arch_optimizer::accu_type data{*addr};
+      if ((data & arch_optimizer::ascii_mask) == arch_optimizer::ascii_value) {
+       arch_optimizer::template append<typename From::string_type::value_type, typename To::string_type>(reinterpret_cast<const typename From::string_type::value_type*>(addr), result);
+       begin += arch_optimizer::accu_size;
        ++addr;
       } else {
        // just advance one code unit for now
@@ -579,7 +624,7 @@ namespace unicode {
     }
 
     // keep up after unaligned Non-ASCII code points
-    while (begin!= end && (uintptr_t)(void*)(addr = reinterpret_cast<const uint64_t*>(&s.data()[s.size() - input_distance(begin, end)])) & 7) {
+    while (begin!= end && (uintptr_t)(void*)(addr = reinterpret_cast<const typename arch_optimizer::accu_type*>(&s.data()[s.size() - input_distance(begin, end)])) & arch_optimizer::addr_mask) {
      back_inserter = *begin;
      ++begin;
     }