Change to how locations are passed around, for the sake of generational

GC. The issue being solved here is the accuracy of the gc_set function.
The existing impelmentation is too conservative. It has no generation
information about the memory location being stored, and so it assumes
the worst: that it is a location in the middle of a gen 1 object.
This is sub-optimal, creating unacceptable pressure against the
checkobj array and, worse, as a consequence causing unreachable gen 0
objects to be tenured into gen 1.

To solve this problem, we replace "val *" pointers with a structure
of type "loc" which keeps track of the object too, which lets us
discover the generation.

I tried another approach: using just a pointer with a bitfield
indicating the generation.  This turned out to have a serious issue:
such a bitfield goes stale when the object is moved to a different
generation. The object holding the memory location is in gen 1, but the
annotated pointer still indicates gen 0. The gc_set function then
makes the wrong decision, and premature reclamation takes place.

* combi.c (perm_init_common, comb_gen_fun_common,
rcomb_gen_fun_common, rcomb_list_gen_fun): Update to new interfaces
for managing mutation.

* debug.c (debug): Update to new interfaces for managing mutation.
Avoid loc variable name.

* eval.c (env_fbind, env_fbind): Update to new interfaces
for managing mutation.
(lookup_var_l, dwim_loc): Return loc type and update to new interfaces.
(apply_frob_args, op_modplace, op_dohash, transform_op, mapcarv,
mappendv, repeat_infinite_func, repeat_times_func): Update to new
interfaces for managing mutation.

* eval.h (lookup_var_l): Declaration updated.

* filter.c (trie_add, trie_compress, trie_compress_intrinsic,
* build_filter, built_filter_from_list, filter_init): Update to new
* interfaces.

* gc.c (gc_set): Rewritten to use loc type which provides the exact
generation. We do not need the in_malloc_range hack any more, since
we have the backpointer to the object.
(gc_push): Take loc rather than raw pointer.

* gc.h (gc_set, gc_push): Declarations updated.

* hash.c (struct hash): The acons* functions use loc instead
of val * now.
(hash_equal_op, copy_hash, gethash_c, inhash, gethash_n, pushhash,

Change to how locations are passed around, for the sake of generational
GC. The issue being solved here is the accuracy of the gc_set function.
The existing impelmentation is too conservative. It has no generation
information about the memory location being stored, and so it assumes
the worst: that it is a location in the middle of a gen 1 object.
This is sub-optimal, creating unacceptable pressure against the
checkobj array and, worse, as a consequence causing unreachable gen 0
objects to be tenured into gen 1.

To solve this problem, we replace "val *" pointers with a structure
of type "loc" which keeps track of the object too, which lets us
discover the generation.

I tried another approach: using just a pointer with a bitfield
indicating the generation.  This turned out to have a serious issue:
such a bitfield goes stale when the object is moved to a different
generation. The object holding the memory location is in gen 1, but the
annotated pointer still indicates gen 0. The gc_set function then
makes the wrong decision, and premature reclamation takes place.

* combi.c (perm_init_common, comb_gen_fun_common,
rcomb_gen_fun_common, rcomb_list_gen_fun): Update to new interfaces
for managing mutation.

* debug.c (debug): Update to new interfaces for managing mutation.
Avoid loc variable name.

* eval.c (env_fbind, env_fbind): Update to new interfaces
for managing mutation.
(lookup_var_l, dwim_loc): Return loc type and update to new interfaces.
(apply_frob_args, op_modplace, op_dohash, transform_op, mapcarv,
mappendv, repeat_infinite_func, repeat_times_func): Update to new
interfaces for managing mutation.

* eval.h (lookup_var_l): Declaration updated.

* filter.c (trie_add, trie_compress, trie_compress_intrinsic,
* build_filter, built_filter_from_list, filter_init): Update to new
* interfaces.

* gc.c (gc_set): Rewritten to use loc type which provides the exact
generation. We do not need the in_malloc_range hack any more, since
we have the backpointer to the object.
(gc_push): Take loc rather than raw pointer.

* gc.h (gc_set, gc_push): Declarations updated.

* hash.c (struct hash): The acons* functions use loc instead
of val * now.
(hash_equal_op, copy_hash, gethash_c, inhash, gethash_n, pushhash,

1 files changed, 6 insertions, 6 deletions

diff --git a/parser.y b/parser.y
index 2909944f..951427c1 100644
--- a/parser.y
+++ b/parser.y
@@ -1221,8 +1221,8 @@ static val expand_meta(val form, val menv)
     list_collect_decl (out, ptail);
 
     for (; consp(form); form = cdr(form)) {
-      val *nptail = list_collect(ptail, expand_meta(car(form), menv));
-      rlcp(*ptail, form);
+      loc nptail = list_collect(ptail, expand_meta(car(form), menv));
+      rlcp(deref(ptail), form);
       ptail = nptail;
     }
 
@@ -1234,10 +1234,10 @@ static val expand_meta(val form, val menv)
 
 val rlset(val form, val info)
 {
-  val cell = gethash_c(form_to_ln_hash, form, 0);
-  val *place = cdr_l(cell);
-  if (nilp(*place))
-    set(*place, info);
+  val cell = gethash_c(form_to_ln_hash, form, nulloc);
+  loc place = cdr_l(cell);
+  if (nilp(deref(place)))
+    set(place, info);
   return form;
 }