OK, here's another thread related to COW and std.string. Currently when a
function in std.string needs to copy (the C in COW) it allocates the new
string from the GC. How about adding an optional trailing parameter for a
scratch char[] for the result if it fits in the scratch? Basically what
typically happens is that instead of a line line "result = new char[len]"
you instead initialize result to the scratch array and then say
"result.length = len;". For example, currently std.string.join looks rougly
like
char[] join(char[][] words, char[] sep)
{
  char[] result;
  if (words.length)
  {
    [ compute length of result in len ]
    result = new char[len];
    [ fill result ]
  }
  return result;
}

and instead if join took an optional trailing scratch array it could be written as

char[] join(char[][] words, char[] sep, char[] scratch = null)
{
  char[] result;
  if (words.length)
  {
    [ compute length of result in len ]
    result = scratch;
    result.length = len;
    [ fill result ]
  }
  return result;
}

I haven't done any actual performance testing of such a modified std.string. The function tolower etc also can take a scratch array to reduce the GC load but those would still suffer from the performance hit of copying the strings to the scratch (or whereever the result ends up).