[RFC] proof-of-concept generational GC survives the bootstrap

[Dmitry Antipov]

This is a patch against bzr trunk 109851, ant it passes the full
bootstrap for me. It's also alive with basic editing operations,
but most probably will crash quite soon. There isn't too much
comments yet, so I tried to write some notices below.

* Object life cycle
  Each object contains 'struct gc_info' which is used to store
  additional object information. Allocation routines should call
  gengc_init to initialize 'gcinfo' slot of the object; it should
  specify correct GC_OBJ_xxx type and one of GC_NEW (if object will
  be immediately used) or GC_FREE (if the object will join a free
  list). If an object survives the GC, gengc_promote should be
  called to adjust object as GC_OLD; if an object dies, gengc_exit
  should be called for it.

* Write barrier
  The main routine is gengc_write_barrier; convenient wrappers
  are gengc_object_write_barrier and PTR_BARRIER/VECTOR_BARRIER
  macros. Inter-generational objects are stored in gengc_objects
  without duplications; to record special values, Lisp_Misc_Save_Value
  is (ab)used to hold either interval pointer or Lisp_Object address.
  The last one (via gengc_record_address) is used when an Lisp_Object
  slot's address is obtained; since we can't assume what object type
  will be written to it (and whether something will be written at all),
  we conservatively assume that GC_NEW object will be written to
  this slot of GC_OLD object, and so the slot is a subject for marking
  traversal during generational collection.

* Mark/sweep
  Collection scheduling is fairly simple: each odd collection is
  generational, and each even collection is full. During generational
  collection, only GC_NEW objects are marked with mark_object; the
  same applies to sweeping (gengc_sweep controls it); gengc_collect
  tries to collect some statistics about objects (set EMACS_GENGC_VERBOSE
  environment variable to see it).

Dmitry

gengc.patch
Description: Text document