Updated mangling specs

[Daveed Vandevoorde]

Martin von Loewis wrote:
> 
> > It's definitely the latter interpretation.  The intent was to express the
> > "inverted A" quantifier, though it probably would have been clearer to
> > say "for all" instead of "for any".
> 
> I trust you that this was indeed the intent. However, it seems that
> the only reasonable implementation is to encode all expressions as
> they appear in the source. If that is done, what can be saved by not
> having to tell functionally-equivalent templates apart? Using an
> expression encoding, 'I+10' and 'I+1+2+3+4' would turn out to be
> different, so why explicitly mention that an implementation could
> consider them equivalent?

You can either mangle the sequence of tokens after normalizing the
parameter tokens (e.g., "J+10" and "L+10" are definitely equivalent
if L and J are both first template parameters of the nontype kind),
or you can construct a mangled representation of the tree structure
representing the expression.  In the latter case, you might have
constant-folded nondependent subexpressions.

The EDG front end does the former, but HP aC++ does the latter.

	// File 1:
	template<int I> void f(double (&)[I+1+2]) {}
	template void f<1>(double (&)[4]);  // (1)

	// File 2:
	template<int I> void f(double (&)[I+3]) {}
	template void f<1>(double (&)[4]);  // Links up with (1) on aC++
	                                    // but not with EDG

	Daveed