Speeding up write_significand()

include/fmt/format.h

@@ -1708,14 +1708,25 @@ inline auto write_significand(Char* out, UInt significand, int significand_size,
                               int integral_size, Char decimal_point) -> Char* {
   if (!decimal_point)
     return format_decimal(out, significand, significand_size).end;
-  auto end = format_decimal(out + 1, significand, significand_size).end;
-  if (integral_size == 1) {
-    out[0] = out[1];
+  out += significand_size + 1;
+  Char* end = out;
+  int floating_size = significand_size - integral_size;
+  for(int i = (floating_size + 1) / 2; i > 0; --i) {
+    out -= 2;
+    copy2(out, digits2(significand % 100));
+    significand /= 100;
+  }
+  if (floating_size % 2 != 0) {
+    Char tmp = *out;
+    *out = decimal_point;
+    *--out = tmp;
+    --integral_size;
   } else {
-    std::uninitialized_copy_n(out + 1, integral_size,
-                              make_checked(out, to_unsigned(integral_size)));
+    *--out = decimal_point;
   }
-  out[integral_size] = decimal_point;
+  if (integral_size > 0)
+    format_decimal(out - integral_size, significand, integral_size);
+
   return end;
 }