On Sun, Jul 21, 2019 at 7:29 AM Paul Eggert <eggert@cs.ucla.edu> wrote:
> > I think we have the following options:
> >
> > 1. remove pure space entirely
> > 2. remove pure space, but leave PURE_P and CHECK_IMPURE as reminders
> > to do something about it.
> > 3. move pure space to BSS
> > 4. xmalloc() pure space, only when needed
> > 5. modify pdumper to mark and recognize pure objects
> > 6. do nothing and accept the wastefulness
>
> I suggest (1), since it will result in simpler code.

I think we should do (1) for now, since it simplifies the code enough
to introduce immutable objects "soon"; but until that time, we waste
more space on duplicate objects that we no longer know to be
immutable, so cannot merge.

I'm attaching a first patch that removes pure space, pinned symbols,
pinned objects, but keeps Fpurecopy (for hash consing), and doesn't
touch the Lisp codebase.

With this patch, I have:
-rw-r--r-- 2 pip pip 11102752 Jul 21 12:28 src/emacs.pdmp

before:
-rw-r--r-- 2 pip pip 10381464 Jul 21 12:29 src/emacs.pdmp

However, the (uncompressed) disk space requirement is about the same,
since the emacs binary is a lot smaller.

I think the next steps are to look at actual live memory usage (which
will increase due to the non-duplication of objects, but not by an
entire megabyte because some of that data is relocations), and GC
performance (no prediction here, it could improve or deteriorate).