5717 – 1.067 regression: appending Unicode char to string broken

void main()
{
	string s, s2; 
	s = "Привет";
	foreach(c; s)
		s2 ~= c;
	assert(s == s2);
}
DMD now seems to consider each individual char a whole code point (as if it was automatically promoted to dchar).

Same problem happens on FreeBSD 8.2 with DMD 1.067 too.
But the problem does not happen with DMD 1.066.

I think this is a foreach problem.
Probably triggered by the fix to bug 4389.

It doesn't look like a foreach problem. This fails too:
void main()
{
 string s, s2; 
 s = "Привет";
 for (int i=0; i<s.length; i++)
 s2 ~= s[i];
 assert(s == s2);
}

(In reply to comment #3)
> It doesn't look like a foreach problem. This fails too:
Hmm. You're right. And yet it works fine on D2. 
It's inserting a call to _d_arrayappendcd, which means the append has been changed into char[] ~ dchar.

It was indeed caused by the fix to bug 4389, which wasn't tight enough. 
s~= c shouldn't turn c into a dchar, if both s and c are the same type. (ie, char[]~=char should go through unaltered). That leaves wchar[] ~ char, which I think is inevitably a mess if c is outside the ASCII range.
expression.c, line 8593. CatAssignExp::semantic()
 { // Append array
 e2 = e2->castTo(sc, e1->type);
 type = e1->type;
 e = this;
 }
 else if (tb1->ty == Tarray &&
 (tb1next->ty == Tchar || tb1next->ty == Twchar) &&
+ e2->type->ty != tb1next->ty &&
 e2->implicitConvTo(Type::tdchar)
 )
 { // Append dchar to char[] or wchar[]
 e2 = e2->castTo(sc, Type::tdchar);
 type = e1->type;
 e = this;
 /* Do not allow appending wchar to char[] because if wchar happens
 * to be a surrogate pair, nothing good can result.
 */

(In reply to comment #5)
I've tried Don's patch, it works good in my environment.
That's great.
Thank you.

https://github.com/D-Programming-Language/dmd/commit/19e819f6b9e71bc18bc5496ff9638ae7ade3e5ad
https://github.com/D-Programming-Language/dmd/commit/610064e7f74ce8258bc4a3eafde52b7311b56da7 

Thanks - not sure what the second commit has to do with it, though.