I'm looking at the dmd source now...
The compression is done in the backend, file cgobj.c
The conditions are:
#define LIBIDMAX 128
if (len > LIBIDMAX)
{
// Attempt to compress the name
name2 = id_compress(name, len);
// snip
if (len2 > LIBIDMAX) // still too long
{
/* Form md5 digest of the name and store it in the
* last 32 bytes of the name.
*/
// snip impl, open the source to see specific details
/******************************************
* Compress an identifier.
* Format: if ASCII, then it's just the char
* if high bit set, then it's a length/offset pair
* Returns:
* malloc'd compressed identifier
*/
char *id_compress(char *id, int idlen)
{
The implementation, same source file, looks like it compresses by
looking for longest duplicate strings and then removes them,
using the offset instead.
The reason I snipped the implementations here is the backend is
under a more restrictive license so I don't want to get into
copying that. But with just what I've said here combined with
guess+check against dmd's output it might be enough to do a clean
room implementation.
Or if Walter can give us permission to copy/paste this into a D
file we could use id directly.
