* libguile/finalizers.c (async_gc_finalizer):
(scm_i_register_async_gc_callback): Replace "weak gc callback"
mechanism with "async gc callback" mechanism. Very similar but the
new API is designed to be called a bounded number of times, to avoid
running afoul of libgc heuristics.
* libguile/weak-list.h: New internal header.
* libguile/Makefile.am (noinst_HEADERS): Add weak-list.h.
* libguile/weak-set.c (vacuum_all_weak_sets):
(scm_c_make_weak_set, scm_init_weak_set):
* libguile/weak-table.c (vacuum_all_weak_tables):
(scm_c_make_weak_table, scm_init_weak_table): Arrange to vacuum all
weak sets from a single async GC callback, and likewise for weak
tables.
Thanks to Ludovic Courtès for tracking this bug down!
* libguile/weak-table.c (GC_move_disappearing_link)
(move_disappearing_links): Factor out stub, as in previous weak-set
commit.
(resize_table): As in weak-set commit, finalizers do not run within
allocators.
(do_vacuum_weak_table): Add comment about need for trylock.
(scm_c_make_weak_table): Use the new scm_i_register_weak_gc_callback.
Previously this was the case because of a mistyped check for a
"HAVE_GC_TABLE_START_CALLBACK" symbol, whereas the symbol was actually
HAVE_GC_SET_START_CALLBACK. Anyway, it's better to call this after GC
anyway, as vacuuming might need to allocate weak links (on pre-7.3
libgc without GC_move_disappearing_link) or reallocate the backing
vector.
* libguile/weak-set.c (resize_set): When releasing a lock temporarily to do
an allocation, re-lock it afterwards, instead of unlocking again.
* libguile/weak-table.c (resize_table): Likewise.
* libguile/weak-table.h:
* libguile/weak-table.c (scm_weak_table_putq_x)
(scm_weak_table_remq_x, scm_weak_table_clear_x)
(scm_weak_table_for_each): Declare these as returning void instead of
SCM.
* libguile/hashtab.c (scm_hashq_set_x, scm_hashq_remove_x)
(scm_hashv_set_x, scm_hashv_remove_x)
(scm_hash_set_x, scm_hash_remove_x)
(scm_hashx_set_x, scm_hashx_remove_x):
(scm_hash_for_each): For weak tables, have the set! functions return
the values, as they used to do. Have remove! functions return #f,
indicating the lack of a handle. Shim around for-each to return
unspecified, even though that wasn't yet a problem.
* test-suite/tests/weaks.test: Add a test.
* libguile/weak-set.c:
* libguile/weak-table.c: Attempt to avoid bouncing between sizes when
growing the vector causes elements to be removed from the set/table.
* libguile/weak-set.c (make_weak_set):
* libguile/weak-table.c (make_weak_table): If we have a threaded Guile,
* keep a weak set (table) of weak sets (tables). Use this and the
pthread_atfork mechanism to lock and unlock weak sets and weak tables
during a fork().
* libguile/weak-set.h (scm_weak_set_prehistory): New internal API.
* libguile/init.c: Add call to scm_weak_set_prehistory().
* libguile/weak-set.c (resize_set):
* libguile/weak-table.c (resize_table): Drop the set/table lock while
allocating the new vector. Fixes a bug in which a finalizer could
recursively try to grab the port table lock.
* libguile/ports.c (scm_putc, scm_puts):
* libguile/ports.h (scm_putc_unlocked, scm_puts_unlocked): Separate into
_unlocked and locked variants. Change all callers to use the
_unlocked versions.
* libguile/tags.h (SCM_HEAP_OBJECT_BASE): New macro. Given a SCM,
returns a pointer to the start of its memory area on the heap.
* libguile/bytevectors.c:
* libguile/fluids.c:
* libguile/foreign.c:
* libguile/gc.h:
* libguile/guardians.c:
* libguile/numbers.h:
* libguile/ports.c:
* libguile/smob.c:
* libguile/struct.c:
* libguile/weak-set.c:
* libguile/weak-table.c:
* libguile/weak-vector.c: Use it.
* libguile/tags.h (SCM_UNPACK_POINTER, SCM_PACK_POINTER): New macros.
The old SCM2PTR and PTR2SCM were defined in such a way that
round-tripping through a pointer could lose precision, even in the
case in which you weren't interested in actually dereferencing the
pointer, it was simply that you needed to plumb a SCM through APIs
that take pointers. These new macros are more like SCM_PACK and
SCM_UNPACK, but for pointer types. The bit representation of the
pointer should be the same as the scm_t_bits representation.
* libguile/gc.h (PTR2SCM, SCM2PTR): Remove support for (old) UNICOS
pointers. We are going to try tagging the SCM object itself in the
future, and I don't think that keeping this support is worth its
cost. It probably doesn't work anyway.
* libguile/backtrace.c:
* libguile/bytevectors.c:
* libguile/continuations.c:
* libguile/fluids.c:
* libguile/foreign.c:
* libguile/gc.h:
* libguile/guardians.c:
* libguile/hashtab.c:
* libguile/load.c:
* libguile/numbers.c:
* libguile/ports.c:
* libguile/smob.c:
* libguile/strings.c:
* libguile/symbols.c:
* libguile/vm.c:
* libguile/weak-set.c:
* libguile/weak-table.c:
* libguile/weak-vector.c: Update many sites to use the new macros.
* libguile/hashtab.c:
* libguile/hashtab.h: Reimplement the weak hash table implementation in
terms of weak tables. All is well except for the horrific hack for
hashx tables.
* libguile/weak-table.h:
* libguile/weak-table.c (scm_make_weak_key_hash_table)
(scm_make_weak_value_hash_table, scm_make_doubly_weak_hash_table)
(scm_weak_key_hash_table_p, scm_weak_value_hash_table_p)
(scm_doubly_weak_hash_table_p): Move these definitions here.
* libguile/weak-table.c:
* libguile/weak-table.h: New files, implementing open-addressed weak
hash tables, similar to the implementation of weak sets. This will
let us remove weak pairs.
* libguile.h:
* libguile/Makefile.am:
* libguile/evalext.c:
* libguile/gc.c:
* libguile/init.c:
* libguile/print.c:
* libguile/tags.h: Update all the pieces for the new files and tc7.
