Optimize size of text_style using bit packing

Author: LocalSpook

doc/api.md

@@ -580,7 +580,7 @@ performance bottleneck.
 
 `fmt/color.h` provides support for terminal color and text style output.
 
-::: print(const text_style&, format_string<T...>, T&&...)
+::: print(text_style, format_string<T...>, T&&...)
 
 ::: fg(detail::color_type)
 

include/fmt/color.h

@@ -155,7 +155,7 @@ enum class color : uint32_t {
   white_smoke = 0xF5F5F5,              // rgb(245,245,245)
   yellow = 0xFFFF00,                   // rgb(255,255,0)
   yellow_green = 0x9ACD32              // rgb(154,205,50)
-};                                     // enum class color
+};  // enum class color
 
 enum class terminal_color : uint8_t {
   black = 30,
@@ -205,121 +205,155 @@ struct rgb {
 
 namespace detail {
 
-// color is a struct of either a rgb color or a terminal color.
+// a bit-packed variant of either an rgb color or a terminal color.
+// see text_style for the bit-packing scheme.
 struct color_type {
-  FMT_CONSTEXPR color_type() noexcept : is_rgb(), value{} {}
-  FMT_CONSTEXPR color_type(color rgb_color) noexcept : is_rgb(true), value{} {
-    value.rgb_color = static_cast<uint32_t>(rgb_color);
-  }
-  FMT_CONSTEXPR color_type(rgb rgb_color) noexcept : is_rgb(true), value{} {
-    value.rgb_color = (static_cast<uint32_t>(rgb_color.r) << 16) |
-                      (static_cast<uint32_t>(rgb_color.g) << 8) | rgb_color.b;
-  }
+  FMT_CONSTEXPR color_type() noexcept = default;
+  FMT_CONSTEXPR color_type(color rgb_color) noexcept
+      : value_(static_cast<uint32_t>(rgb_color) | (1 << 24)) {}
+  FMT_CONSTEXPR color_type(rgb rgb_color) noexcept
+      : color_type(static_cast<color>(
+            (static_cast<uint32_t>(rgb_color.r) << 16) |
+            (static_cast<uint32_t>(rgb_color.g) << 8) | rgb_color.b)) {}
   FMT_CONSTEXPR color_type(terminal_color term_color) noexcept
-      : is_rgb(), value{} {
-    value.term_color = static_cast<uint8_t>(term_color);
+      : value_(static_cast<uint32_t>(term_color) | (3 << 24)) {}
+
+  FMT_CONSTEXPR auto is_terminal_color() const noexcept -> bool {
+    return (value_ & (1 << 25)) != 0;
+  }
+
+  FMT_CONSTEXPR auto get_value() const noexcept -> uint32_t {
+    return value_ & 0xFFFFFF;
   }
-  bool is_rgb;
-  union color_union {
-    uint8_t term_color;
-    uint32_t rgb_color;
-  } value;
+
+  FMT_CONSTEXPR color_type(uint32_t value) noexcept : value_(value) {}
+
+  uint32_t value_{};
 };
 }  // namespace detail
 
 /// A text style consisting of foreground and background colors and emphasis.
 class text_style {
+  // The information is packed as follows:
+  // ┌──┐
+  // │ 0│─┐
+  // │..│ ├── foreground color value
+  // │23│─┘
+  // ├──┤
+  // │24│─┬── discriminator for the above value. 00 if unset, 01 if it's
+  // │25│─┘   an RGB color, or 11 if it's a terminal color (10 is unused)
+  // ├──┤
+  // │26│──── overflow bit, always zero (see below)
+  // ├──┤
+  // │27│─┐
+  // │..│ │
+  // │50│ │
+  // ├──┤ │
+  // │51│ ├── background color (same format as the foreground color)
+  // │52│ │
+  // ├──┤ │
+  // │53│─┘
+  // ├──┤
+  // │54│─┐
+  // │..│ ├── emphases
+  // │61│─┘
+  // ├──┤
+  // │62│─┬── unused
+  // │63│─┘
+  // └──┘
+  // The overflow bits are there to make operator|= efficient.
+  // When ORing, we must throw if, for either the foreground or background,
+  // one style specifies a terminal color and the other specifies any color
+  // (terminal or RGB); in other words, if one discriminator is 11 and the
+  // other is 11 or 01.
+  //
+  // We do that check by adding the styles. Consider what adding does to each
+  // possible pair of discriminators:
+  //    00 + 00 = 000
+  //    01 + 00 = 001
+  //    11 + 00 = 011
+  //    01 + 01 = 010
+  //    11 + 01 = 100 (!!)
+  //    11 + 11 = 110 (!!)
+  // In the last two cases, the ones we want to catch, the third bit——the
+  // overflow bit——is set. Bingo.
+  //
+  // We must take into account the possible carry bit from the bits
+  // before the discriminator. The only potentially problematic case is
+  // 11 + 00 = 011 (a carry bit would make it 100, not good!), but a carry
+  // bit is impossible in that case, because 00 (unset color) means the
+  // 24 bits that precede the discriminator are all zero.
+  //
+  // This test can be applied to both colors simultaneously.
+
  public:
   FMT_CONSTEXPR text_style(emphasis em = emphasis()) noexcept
-      : set_foreground_color(), set_background_color(), ems(em) {}
-
-  FMT_CONSTEXPR auto operator|=(const text_style& rhs) -> text_style& {
-    if (!set_foreground_color) {
-      set_foreground_color = rhs.set_foreground_color;
-      foreground_color = rhs.foreground_color;
-    } else if (rhs.set_foreground_color) {
-      if (!foreground_color.is_rgb || !rhs.foreground_color.is_rgb)
-        report_error("can't OR a terminal color");
-      foreground_color.value.rgb_color |= rhs.foreground_color.value.rgb_color;
-    }
+      : style_(static_cast<uint64_t>(em) << 54) {}
 
-    if (!set_background_color) {
-      set_background_color = rhs.set_background_color;
-      background_color = rhs.background_color;
-    } else if (rhs.set_background_color) {
-      if (!background_color.is_rgb || !rhs.background_color.is_rgb)
-        report_error("can't OR a terminal color");
-      background_color.value.rgb_color |= rhs.background_color.value.rgb_color;
-    }
-
-    ems = static_cast<emphasis>(static_cast<uint8_t>(ems) |
-                                static_cast<uint8_t>(rhs.ems));
+  FMT_CONSTEXPR auto operator|=(text_style rhs) -> text_style& {
+    if (((style_ + rhs.style_) & ((1ULL << 26) | (1ULL << 53))) != 0)
+      report_error("can't OR a terminal color");
+    style_ |= rhs.style_;
     return *this;
   }
 
-  friend FMT_CONSTEXPR auto operator|(text_style lhs, const text_style& rhs)
+  friend FMT_CONSTEXPR auto operator|(text_style lhs, text_style rhs)
       -> text_style {
     return lhs |= rhs;
   }
 
+  FMT_CONSTEXPR auto operator==(text_style rhs) const noexcept -> bool {
+    return style_ == rhs.style_;
+  }
+
+  FMT_CONSTEXPR auto operator!=(text_style rhs) const noexcept -> bool {
+    return !(*this == rhs);
+  }
+
   FMT_CONSTEXPR auto has_foreground() const noexcept -> bool {
-    return set_foreground_color;
+    return (style_ & (1 << 24)) != 0;
   }
   FMT_CONSTEXPR auto has_background() const noexcept -> bool {
-    return set_background_color;
+    return (style_ & (1ULL << 51)) != 0;
   }
   FMT_CONSTEXPR auto has_emphasis() const noexcept -> bool {
-    return static_cast<uint8_t>(ems) != 0;
+    return (style_ >> 54) != 0;
   }
   FMT_CONSTEXPR auto get_foreground() const noexcept -> detail::color_type {
     FMT_ASSERT(has_foreground(), "no foreground specified for this style");
-    return foreground_color;
+    return style_ & 0x3FFFFFF;
   }
   FMT_CONSTEXPR auto get_background() const noexcept -> detail::color_type {
     FMT_ASSERT(has_background(), "no background specified for this style");
-    return background_color;
+    return (style_ >> 27) & 0x3FFFFFF;
   }
   FMT_CONSTEXPR auto get_emphasis() const noexcept -> emphasis {
     FMT_ASSERT(has_emphasis(), "no emphasis specified for this style");
-    return ems;
+    return static_cast<emphasis>(style_ >> 54);
   }
 
  private:
-  FMT_CONSTEXPR text_style(bool is_foreground,
-                           detail::color_type text_color) noexcept
-      : set_foreground_color(), set_background_color(), ems() {
-    if (is_foreground) {
-      foreground_color = text_color;
-      set_foreground_color = true;
-    } else {
-      background_color = text_color;
-      set_background_color = true;
-    }
-  }
+  FMT_CONSTEXPR text_style(uint64_t style) noexcept : style_(style) {}
 
   friend FMT_CONSTEXPR auto fg(detail::color_type foreground) noexcept
       -> text_style;
 
   friend FMT_CONSTEXPR auto bg(detail::color_type background) noexcept
       -> text_style;
 
-  detail::color_type foreground_color;
-  detail::color_type background_color;
-  bool set_foreground_color;
-  bool set_background_color;
-  emphasis ems;
+  uint64_t style_{};
 };
 
 /// Creates a text style from the foreground (text) color.
 FMT_CONSTEXPR inline auto fg(detail::color_type foreground) noexcept
     -> text_style {
-  return text_style(true, foreground);
+  return foreground.value_ | (1ULL << 24);
 }
 
 /// Creates a text style from the background color.
 FMT_CONSTEXPR inline auto bg(detail::color_type background) noexcept
     -> text_style {
-  return text_style(false, background);
+  return (background.value_ | (1ULL << 24)) << 27;
 }
 
 FMT_CONSTEXPR inline auto operator|(emphasis lhs, emphasis rhs) noexcept
@@ -334,9 +368,9 @@ template <typename Char> struct ansi_color_escape {
                                   const char* esc) noexcept {
     // If we have a terminal color, we need to output another escape code
     // sequence.
-    if (!text_color.is_rgb) {
+    if (text_color.is_terminal_color()) {
       bool is_background = esc == string_view("\x1b[48;2;");
-      uint32_t value = text_color.value.term_color;
+      uint32_t value = text_color.get_value();
       // Background ASCII codes are the same as the foreground ones but with
       // 10 more.
       if (is_background) value += 10u;
@@ -360,7 +394,7 @@ template <typename Char> struct ansi_color_escape {
     for (int i = 0; i < 7; i++) {
       buffer[i] = static_cast<Char>(esc[i]);
     }
-    rgb color(text_color.value.rgb_color);
+    rgb color(text_color.get_value());
     to_esc(color.r, buffer + 7, ';');
     to_esc(color.g, buffer + 11, ';');
     to_esc(color.b, buffer + 15, 'm');
@@ -441,32 +475,26 @@ template <typename T> struct styled_arg : view {
 };
 
 template <typename Char>
-void vformat_to(buffer<Char>& buf, const text_style& ts,
-                basic_string_view<Char> fmt,
+void vformat_to(buffer<Char>& buf, text_style ts, basic_string_view<Char> fmt,
                 basic_format_args<buffered_context<Char>> args) {
-  bool has_style = false;
   if (ts.has_emphasis()) {
-    has_style = true;
     auto emphasis = make_emphasis<Char>(ts.get_emphasis());
     buf.append(emphasis.begin(), emphasis.end());
   }
   if (ts.has_foreground()) {
-    has_style = true;
     auto foreground = make_foreground_color<Char>(ts.get_foreground());
     buf.append(foreground.begin(), foreground.end());
   }
   if (ts.has_background()) {
-    has_style = true;
     auto background = make_background_color<Char>(ts.get_background());
     buf.append(background.begin(), background.end());
   }
   vformat_to(buf, fmt, args);
-  if (has_style) reset_color<Char>(buf);
+  if (ts != text_style{}) reset_color<Char>(buf);
 }
 }  // namespace detail
 
-inline void vprint(FILE* f, const text_style& ts, string_view fmt,
-                   format_args args) {
+inline void vprint(FILE* f, text_style ts, string_view fmt, format_args args) {
   auto buf = memory_buffer();
   detail::vformat_to(buf, ts, fmt, args);
   print(f, FMT_STRING("{}"), string_view(buf.begin(), buf.size()));
@@ -482,8 +510,7 @@ inline void vprint(FILE* f, const text_style& ts, string_view fmt,
  *                "Elapsed time: {0:.2f} seconds", 1.23);
  */
 template <typename... T>
-void print(FILE* f, const text_style& ts, format_string<T...> fmt,
-           T&&... args) {
+void print(FILE* f, text_style ts, format_string<T...> fmt, T&&... args) {
   vprint(f, ts, fmt.str, vargs<T...>{{args...}});
 }
 
@@ -497,11 +524,11 @@ void print(FILE* f, const text_style& ts, format_string<T...> fmt,
  *                "Elapsed time: {0:.2f} seconds", 1.23);
  */
 template <typename... T>
-void print(const text_style& ts, format_string<T...> fmt, T&&... args) {
+void print(text_style ts, format_string<T...> fmt, T&&... args) {
   return print(stdout, ts, fmt, std::forward<T>(args)...);
 }
 
-inline auto vformat(const text_style& ts, string_view fmt, format_args args)
+inline auto vformat(text_style ts, string_view fmt, format_args args)
     -> std::string {
   auto buf = memory_buffer();
   detail::vformat_to(buf, ts, fmt, args);
@@ -521,16 +548,16 @@ inline auto vformat(const text_style& ts, string_view fmt, format_args args)
  * ```
  */
 template <typename... T>
-inline auto format(const text_style& ts, format_string<T...> fmt, T&&... args)
+inline auto format(text_style ts, format_string<T...> fmt, T&&... args)
     -> std::string {
   return fmt::vformat(ts, fmt.str, vargs<T...>{{args...}});
 }
 
 /// Formats a string with the given text_style and writes the output to `out`.
 template <typename OutputIt,
           FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value)>
-auto vformat_to(OutputIt out, const text_style& ts, string_view fmt,
-                format_args args) -> OutputIt {
+auto vformat_to(OutputIt out, text_style ts, string_view fmt, format_args args)
+    -> OutputIt {
   auto&& buf = detail::get_buffer<char>(out);
   detail::vformat_to(buf, ts, fmt, args);
   return detail::get_iterator(buf, out);
@@ -548,8 +575,8 @@ auto vformat_to(OutputIt out, const text_style& ts, string_view fmt,
  */
 template <typename OutputIt, typename... T,
           FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value)>
-inline auto format_to(OutputIt out, const text_style& ts,
-                      format_string<T...> fmt, T&&... args) -> OutputIt {
+inline auto format_to(OutputIt out, text_style ts, format_string<T...> fmt,
+                      T&&... args) -> OutputIt {
   return vformat_to(out, ts, fmt.str, vargs<T...>{{args...}});
 }
 

include/fmt/xchar.h

@@ -322,27 +322,27 @@ template <typename... T> void println(wformat_string<T...> fmt, T&&... args) {
   return print(L"{}\n", fmt::format(fmt, std::forward<T>(args)...));
 }
 
-inline auto vformat(const text_style& ts, wstring_view fmt, wformat_args args)
+inline auto vformat(text_style ts, wstring_view fmt, wformat_args args)
     -> std::wstring {
   auto buf = wmemory_buffer();
   detail::vformat_to(buf, ts, fmt, args);
   return {buf.data(), buf.size()};
 }
 
 template <typename... T>
-inline auto format(const text_style& ts, wformat_string<T...> fmt, T&&... args)
+inline auto format(text_style ts, wformat_string<T...> fmt, T&&... args)
     -> std::wstring {
   return fmt::vformat(ts, fmt, fmt::make_wformat_args(args...));
 }
 
 template <typename... T>
-FMT_DEPRECATED void print(std::FILE* f, const text_style& ts,
-                          wformat_string<T...> fmt, const T&... args) {
+FMT_DEPRECATED void print(std::FILE* f, text_style ts, wformat_string<T...> fmt,
+                          const T&... args) {
   vprint(f, ts, fmt, fmt::make_wformat_args(args...));
 }
 
 template <typename... T>
-FMT_DEPRECATED void print(const text_style& ts, wformat_string<T...> fmt,
+FMT_DEPRECATED void print(text_style ts, wformat_string<T...> fmt,
                           const T&... args) {
   return print(stdout, ts, fmt, args...);
 }