0.pre8.85
authorDaniel Barlow <dan@telent.net>
Mon, 21 Apr 2003 21:06:24 +0000 (21:06 +0000)
committerDaniel Barlow <dan@telent.net>
Mon, 21 Apr 2003 21:06:24 +0000 (21:06 +0000)
Merge the definitely-maybe-gc branch:

Remove all the lispland logic that checks bytes-consed
thresholds to see if it's worth GCing: C does these checks
anyway.  SUB-GC now always GCs unless *GC-INHIBIT* says not to

Delete the notify before/after calls and notify-stream.
(Planned future change: pass the same information to
*{BEFORE,AFTER}-GC-HOOKS* instead)

*SOFT-HEAP-LIMIT* and MAYBE-GC are gone (unused).
*ALREADY-MAYBE-GCING* is also dead, replaced with a proper
mutex that will also prevent simultaneous GCs from multiple
threads.  This entails moving gc.lisp and purify.lisp after
target-thread in the build-order

(Unithread WITH-MUTEX now honours its WAIT-P argument, so
actually works in this situation)

16 files changed:
NEWS
build-order.lisp-expr
package-data-list.lisp-expr
src/code/gc.lisp
src/code/sysmacs.lisp
src/code/target-unithread.lisp
src/code/toplevel.lisp
src/compiler/alpha/parms.lisp
src/compiler/generic/genesis.lisp
src/compiler/hppa/parms.lisp
src/compiler/mips/parms.lisp
src/compiler/ppc/parms.lisp
src/compiler/sparc/parms.lisp
src/compiler/x86/parms.lisp
src/runtime/interrupt.c
version.lisp-expr

diff --git a/NEWS b/NEWS
index ee4c576..0ac4b8e 100644 (file)
--- a/NEWS
+++ b/NEWS
@@ -1662,6 +1662,11 @@ changes in sbcl-0.8.0 relative to sbcl-0.7.14
     compiled in by default: you need to add :SB-THREAD to the target
     features.  See the "Beyond ANSI" chapter of the manual for
     details.
+  * Garbage collection refactoring: user-visible change is that a
+    call to the GC function during WITHOUT-GCING will not do garbage
+    collection until the end of the WITHOUT-GCING.  If you were doing 
+    this you were probably losing anyway.
+
   * sb-aclrepl module improvements: an integrated inspector, added
     repl features, and a bug fix to :trace command.
   * fixed some bugs revealed by Paul Dietz' test suite:
index c31fc9b..7761ef2 100644 (file)
  ("src/code/misc")
 
  ("src/code/room"   :not-host)
- ("src/code/gc"     :not-host)
- ("src/code/purify" :not-host)
 
  ("src/code/stream"        :not-host)
  ("src/code/print"         :not-host)
  #!-sb-thread
  ("src/code/target-unithread"  :not-host)
  ;; defines SB!DI:DO-DEBUG-FUN-BLOCKS, needed by target-disassem.lisp
+ ("src/code/gc"     :not-host)
+ ("src/code/purify" :not-host)
  ("src/code/debug-int" :not-host)
 
  ;; target-only assemblerish stuff
index 72e86e4..12f5e8e 100644 (file)
@@ -1129,7 +1129,7 @@ is a good idea, but see SB-SYS re. blurring of boundaries."
              "%MAKE-INSTANCE"
             "MAKE-VALUE-CELL"
              "MAKE-VALUES-TYPE"
-             "MAYBE-GC" "MEMBER-TYPE" "MEMBER-TYPE-MEMBERS"
+             "MEMBER-TYPE" "MEMBER-TYPE-MEMBERS"
              "MEMBER-TYPE-P" "MERGE-BITS" "MODIFIED-NUMERIC-TYPE"
              "MUTATOR-SELF"
              "NAMED-TYPE" "NAMED-TYPE-NAME" "NAMED-TYPE-P"
@@ -1237,6 +1237,7 @@ is a good idea, but see SB-SYS re. blurring of boundaries."
              "SPECIFIER-TYPE" "STACK-REF"
              "STREAMLIKE" "STRINGABLE"
              "STRUCTURE-RAW-SLOT-TYPE-AND-SIZE"
+            "SUB-GC"
              "%INSTANCE-LENGTH"
              "%INSTANCE-REF" "%INSTANCE-SET"
              "SYSTEM-AREA-CLEAR"
index 7dfd23c..39d59a6 100644 (file)
@@ -180,24 +180,6 @@ and submit it as a patch."
    GET-INTERNAL-RUN-TIME)")
 (declaim (type index *gc-run-time*))
 
-;;; a limit to help catch programs which allocate too much memory,
-;;; since a hard heap overflow is so hard to recover from
-;;;
-;;; FIXME: Like *GC-TRIGGER*, this variable (1) should probably be
-;;; denominated in a larger unit than bytes and (2) should probably be
-;;; renamed so that it's clear from the name what unit it's
-;;; denominated in.
-(declaim (type (or unsigned-byte null) *soft-heap-limit*))
-(defvar *soft-heap-limit*
-  ;; As long as *GC-TRIGGER* is DECLAIMed as INDEX, we know that
-  ;; MOST-POSITIVE-FIXNUM is a hard limit on how much memory can be
-  ;; allocated. (Not necessarily *the* hard limit, which is fairly
-  ;; likely something like a Unix per-process limit that we don't know
-  ;; about, but a hard limit anyway.) And this gives us a reasonable
-  ;; conservative default for the soft limit...
-  (- most-positive-fixnum
-     *bytes-consed-between-gcs*))
-
 ;;;; The following specials are used to control when garbage
 ;;;; collection occurs.
 
@@ -229,50 +211,10 @@ and submit it as a patch."
 ;;; When >0, inhibits garbage collection.
 (defvar *gc-inhibit*) ; initialized in cold init
 
-;;; This flag is used to prevent recursive entry into the garbage
-;;; collector.
-(defvar *already-maybe-gcing*) ; initialized in cold init
-
-;;; When T, indicates that the dynamic usage has exceeded the value
-;;; *GC-TRIGGER*.
+;;; When T, indicates that a GC should have happened but did not due to 
+;;; *GC-INHIBIT*. 
 (defvar *need-to-collect-garbage* nil) ; initialized in cold init
 \f
-(defun default-gc-notify-before (notify-stream bytes-in-use)
-  (declare (type stream notify-stream))
-  (format
-   notify-stream
-   "~&; GC is beginning with ~:D bytes in use at internal runtime ~:D.~%"
-   bytes-in-use
-   (get-internal-run-time))
-  (finish-output notify-stream))
-(defparameter *gc-notify-before* #'default-gc-notify-before
-  #!+sb-doc
-  "The function bound to this variable is invoked before GC'ing (unless
-  *GC-NOTIFY-STREAM* is NIL) with the value of *GC-NOTIFY-STREAM* and
-  current amount of dynamic usage (in bytes). It should notify the
-  user that the system is going to GC.")
-
-(defun default-gc-notify-after (notify-stream
-                               bytes-retained
-                               bytes-freed
-                               new-trigger)
-  (declare (type stream notify-stream))
-  (format notify-stream
-         "~&; GC has finished with ~:D bytes in use (~:D bytes freed)~@
-           ; at internal runtime ~:D. The new GC trigger is ~:D bytes.~%"
-         bytes-retained
-         bytes-freed
-         (get-internal-run-time)
-         new-trigger)
-  (finish-output notify-stream))
-(defparameter *gc-notify-after* #'default-gc-notify-after
-  #!+sb-doc
-  "The function bound to this variable is invoked after GC'ing with the
-value of *GC-NOTIFY-STREAM*, the amount of dynamic usage (in bytes) now
-free, the number of bytes freed by the GC, and the new GC trigger
-threshold; or if *GC-NOTIFY-STREAM* is NIL, it's not invoked. The
-function should notify the user that the system has finished GC'ing.")
-\f
 ;;;; internal GC
 
 (sb!alien:define-alien-routine collect-garbage sb!alien:int
@@ -309,129 +251,61 @@ function should notify the user that the system has finished GC'ing.")
              (warn "(FUNCALL ~S~{ ~S~}) lost:~%~A" ',function ',args cond)
              nil))))
 
-;;; SUB-GC decides when and if to do a garbage collection. The FORCE-P
-;;; flags controls whether a GC should occur even if the dynamic usage
-;;; is not greater than *GC-TRIGGER*.
+;;; SUB-GC does a garbage collection.  This is called from three places:
+;;; (1) The C runtime will call here when it detects that we've consed 
+;;;     enough to exceed the gc trigger threshold
+;;; (2) The user may request a collection using GC, below
+;;; (3) At the end of a WITHOUT-GCING section, we are called if
+;;;     *NEED-TO-COLLECT-GARBAGE* is true
 ;;;
+;;; This is different from the behaviour in 0.7 and earlier: it no
+;;; longer decides whether to GC based on thresholds.  If you call
+;;; SUB-GC you will definitely get a GC either now or when the
+;;; WITHOUT-GCING is over
+
 ;;; For GENCGC all generations < GEN will be GC'ed.
 
-;;; XXX need (1) some kind of locking to ensure that only one thread
-;;; at a time is trying to GC, (2) to look at all these specials and
-;;; work out how much of this "do we really need to GC now?" stuff is
-;;; actually necessary: I think we actually end up GCing every time we
-;;; hit this code
-
-(defun sub-gc (&key force-p (gen 0))
-  (/show0 "entering SUB-GC")
-  (unless *already-maybe-gcing*
-    (let* ((*already-maybe-gcing* t)
-          (start-time (get-internal-run-time))
-          (pre-gc-dynamic-usage (dynamic-usage))
-          ;; Currently we only check *SOFT-HEAP-LIMIT* at GC time,
-          ;; not for every allocation. That makes it cheap to do,
-          ;; even if it is a little ugly.
-          (soft-heap-limit-exceeded? (and *soft-heap-limit*
-                                          (> pre-gc-dynamic-usage
-                                             *soft-heap-limit*)))
-          (*soft-heap-limit* (if soft-heap-limit-exceeded?
-                                 (+ pre-gc-dynamic-usage
-                                    *bytes-consed-between-gcs*)
-                                 *soft-heap-limit*)))
-      (when soft-heap-limit-exceeded?
-       (cerror "Continue with GC."
-               "soft heap limit exceeded (temporary new limit=~W)"
-               *soft-heap-limit*))
-      (when (and *gc-trigger* (> pre-gc-dynamic-usage *gc-trigger*))
-       (setf *need-to-collect-garbage* t))
-      (when (or force-p
-               (and *need-to-collect-garbage* (zerop *gc-inhibit*)))
-       ;; KLUDGE: Wow, we really mask interrupts all the time we're
-       ;; collecting garbage? That seems like a long time.. -- WHN 19991129
+(defvar *gc-mutex* (sb!thread:make-mutex :name "GC Mutex"))
+
+(defun sub-gc (&key (gen 0))
+  (when (sb!thread::mutex-value *gc-mutex*) (return-from sub-gc nil))
+  (sb!thread:with-mutex (*gc-mutex* :wait-p nil)
+    (let* ((start-time (get-internal-run-time)))
+      (setf *need-to-collect-garbage* t)
+      (when (zerop *gc-inhibit*)
        (without-interrupts
-        ;; FIXME: We probably shouldn't do this evil thing to
-        ;; *STANDARD-OUTPUT* in a binding which is wrapped around
-        ;; calls to user-settable GC hook functions.
-         (let ((*standard-output* *terminal-io*))
-           (when *gc-notify-stream*
-             (if (streamp *gc-notify-stream*)
-                 (carefully-funcall *gc-notify-before*
-                                    *gc-notify-stream*
-                                    pre-gc-dynamic-usage)
-                 (warn
-                  "*GC-NOTIFY-STREAM* is set, but not a STREAM -- ignored.")))
-           (dolist (hook *before-gc-hooks*)
-             (carefully-funcall hook))
-           (when *gc-trigger*
-             (clear-auto-gc-trigger))
-           (let* (;; We do DYNAMIC-USAGE once more here in order to
-                  ;; get a more accurate measurement of the space
-                  ;; actually freed, since the messing around, e.g.
-                  ;; GC-notify stuff, since the DYNAMIC-USAGE which
-                  ;; triggered GC could've done a fair amount of
-                  ;; consing.)
-                  (pre-internal-gc-dynamic-usage (dynamic-usage))
-                  (ignore-me (funcall *internal-gc* gen))
-                  (post-gc-dynamic-usage (dynamic-usage))
-                  (n-bytes-freed (- pre-internal-gc-dynamic-usage
-                                    post-gc-dynamic-usage))
-                  ;; In sbcl-0.6.12.39, the raw N-BYTES-FREED from
-                  ;; GENCGC could sometimes be substantially negative
-                  ;; (e.g. -5872). I haven't looked into what causes
-                  ;; that, but I suspect it has to do with
-                  ;; fluctuating inefficiency in the way that the
-                  ;; GENCGC packs things into page boundaries.
-                  ;; Bumping the raw result up to 0 is a little ugly,
-                  ;; but shouldn't be a problem, and it's even
-                  ;; possible to sort of justify it: the packing
-                  ;; inefficiency which has caused (DYNAMIC-USAGE) to
-                  ;; grow is effectively consing, or at least
-                  ;; overhead of consing, so it's sort of correct to
-                  ;; add it to the running total of consing. ("Man
-                  ;; isn't a rational animal, he's a rationalizing
-                  ;; animal.":-) -- WHN 2001-06-23
-                  (eff-n-bytes-freed (max 0 n-bytes-freed)))
-             (declare (ignore ignore-me))
-             (/show0 "got (DYNAMIC-USAGE) and EFF-N-BYTES-FREED")
-             (incf *n-bytes-freed-or-purified*
-                   eff-n-bytes-freed)
-             (/show0 "clearing *NEED-TO-COLLECT-GARBAGE*")
-             (setf *need-to-collect-garbage* nil)
-             (/show0 "calculating NEW-GC-TRIGGER")
-             (let ((new-gc-trigger (+ post-gc-dynamic-usage
-                                      *bytes-consed-between-gcs*)))
-               (/show0 "setting *GC-TRIGGER*")
-               (setf *gc-trigger* new-gc-trigger))
-             (/show0 "calling SET-AUTO-GC-TRIGGER")
-             (set-auto-gc-trigger *gc-trigger*)
-             (dolist (hook *after-gc-hooks*)
-               (/show0 "doing a hook from *AFTER-GC--HOOKS*")
-               ;; FIXME: This hook should be called with the same
-               ;; kind of information as *GC-NOTIFY-AFTER*. In
-               ;; particular, it would be nice for the hook function
-               ;; to be able to adjust *GC-TRIGGER* intelligently to
-               ;; e.g. 108% of total memory usage.
-               (carefully-funcall hook))
-             (when *gc-notify-stream*
-               (if (streamp *gc-notify-stream*)
-                   (carefully-funcall *gc-notify-after*
-                                      *gc-notify-stream*
-                                      post-gc-dynamic-usage
-                                      eff-n-bytes-freed
-                                      *gc-trigger*)
-                   (warn
-                    "*GC-NOTIFY-STREAM* is set, but not a stream -- ignored.")))))
-         (scrub-control-stack)))       ;XXX again?  we did this from C ...
-      (incf *gc-run-time* (- (get-internal-run-time)
-                            start-time))))
-  ;; FIXME: should probably return (VALUES), here and in RETURN-FROM
+        (dolist (hook  *before-gc-hooks*)  (carefully-funcall hook))
+        (when *gc-trigger*
+          (clear-auto-gc-trigger))
+        (let* ((pre-internal-gc-dynamic-usage (dynamic-usage))
+               (ignore-me (funcall *internal-gc* gen))
+               (post-gc-dynamic-usage (dynamic-usage))
+               (n-bytes-freed (- pre-internal-gc-dynamic-usage
+                                 post-gc-dynamic-usage))
+               ;; the raw N-BYTES-FREED from GENCGC can sometimes be
+               ;; substantially negative (e.g. -5872).  This is
+               ;; probably due to fluctuating inefficiency in the way
+               ;; that the GENCGC packs things into page boundaries.
+               ;; We bump the raw result up to 0: the space is
+               ;; allocated even if unusable, so should be counted
+               ;; for deciding when we've allocated enough to GC
+               ;; next.  ("Man isn't a rational animal, he's a
+               ;; rationalizing animal.":-) -- WHN 2001-06-23)
+               (eff-n-bytes-freed (max 0 n-bytes-freed)))
+          (declare (ignore ignore-me))
+          (incf *n-bytes-freed-or-purified*  eff-n-bytes-freed)
+          (setf *need-to-collect-garbage* nil)
+          (setf *gc-trigger*  (+ post-gc-dynamic-usage
+                                 *bytes-consed-between-gcs*))
+          (set-auto-gc-trigger *gc-trigger*)
+          (dolist (hook *after-gc-hooks*)
+            (carefully-funcall hook))))
+       (scrub-control-stack))       ;XXX again?  we did this from C ...
+      (incf *gc-run-time* (- (get-internal-run-time) start-time))))
   nil)
 
-;;; This routine is called by the allocation miscops to decide whether
-;;; a GC should occur. The argument, OBJECT, is the newly allocated
-;;; object which must be returned to the caller.
-(defun maybe-gc (&optional object)
-  (sub-gc)
-  object)
+
+
 
 ;;; This is the user-advertised garbage collection function.
 (defun gc (&key (gen 0) (full nil) &allow-other-keys)
@@ -441,7 +315,7 @@ function should notify the user that the system has finished GC'ing.")
   #!+(and sb-doc (not gencgc))
   "Initiate a garbage collection. GEN may be provided for compatibility with
   generational garbage collectors, but is ignored in this implementation."
-  (sub-gc :force-p t :gen (if full 6 gen)))
+  (sub-gc  :gen (if full 6 gen)))
 
 \f
 ;;;; auxiliary functions
index fae42e2..dc1358e 100644 (file)
@@ -34,7 +34,7 @@
       ,@body)
     (atomic-incf/symbol *gc-inhibit* -1)
     (when (and *need-to-collect-garbage* (zerop *gc-inhibit*))
-      (maybe-gc nil))))
+      (sub-gc))))
 
 \f
 ;;; EOF-OR-LOSE is a useful macro that handles EOF.
index c684a3c..e11ec9d 100644 (file)
      (setf old-value t1))))
 
 (defmacro with-mutex ((mutex &key value (wait-p t))  &body body)
-  (declare (ignore mutex value wait-p))
-  `(progn ,@body))
+  (cond ((not wait-p)
+        `(unless (mutex-value ,mutex)
+          (unwind-protect
+               (progn
+                 (setf (mutex-value ,mutex) (or ,value t))
+                 ,@body)
+            (setf (mutex-value ,mutex) nil))))
+       (t 
+        `(progn ,@body))))
 
 ;;; what's the best thing to do with these on unithread?
 #+NIl
index f245850..c05f092 100644 (file)
@@ -25,8 +25,7 @@
 ;;; specials initialized by !COLD-INIT
 
 ;;; FIXME: These could be converted to DEFVARs.
-(declaim (special *gc-inhibit* *already-maybe-gcing*
-                 *need-to-collect-garbage*
+(declaim (special *gc-inhibit* *need-to-collect-garbage*
                  *gc-notify-stream*
                  *before-gc-hooks* *after-gc-hooks*
                  #!+x86 *pseudo-atomic-atomic*
index 5bb2213..cb7da46 100644 (file)
     *posix-argv*
 
     ;; functions that the C code needs to call
-    maybe-gc
+    sub-gc
     sb!kernel::internal-error
     sb!kernel::control-stack-exhausted-error
     sb!di::handle-breakpoint
index 0de5f03..9bfeeaa 100644 (file)
   (macrolet ((frob (symbol)
               `(cold-set ',symbol
                          (cold-fdefinition-object (cold-intern ',symbol)))))
-    (frob maybe-gc)
+    (frob sub-gc)
     (frob internal-error)
     (frob sb!kernel::control-stack-exhausted-error)
     (frob sb!di::handle-breakpoint)
index 6be70c6..3624beb 100644 (file)
     *posix-argv*
     
     ;; Functions that the C code needs to call
-    sb!impl::maybe-gc
+    sb!impl::sub-gc
     sb!kernel::internal-error
     sb!kernel::control-stack-exhausted-error
     sb!di::handle-breakpoint
index 50d4d4e..971f901 100644 (file)
 
     *posix-argv*
 
-    sb!impl::maybe-gc
+    sb!impl::sub-gc
     sb!kernel::internal-error
     sb!kernel::control-stack-exhausted-error
     sb!di::handle-breakpoint
index 94332f3..d5f88c8 100644 (file)
     *posix-argv*
 
     ;; functions that the C code needs to call
-    sb!impl::maybe-gc
+    sb!impl::sub-gc
     sb!kernel::internal-error
     sb!kernel::control-stack-exhausted-error
     sb!di::handle-breakpoint
index 0b7ae33..4162357 100644 (file)
     sb!impl::*!initial-fdefn-objects*
 
     ;; functions that the C code needs to call
-    maybe-gc
+    sub-gc
     sb!kernel::internal-error
     sb!kernel::control-stack-exhausted-error
     sb!di::handle-breakpoint
index 4128ae9..f2c06a1 100644 (file)
 
     ;; functions that the C code needs to call.  When adding to this list,
     ;; also add a `frob' form in genesis.lisp finish-symbols.
-    maybe-gc
+    sub-gc
     sb!kernel::internal-error
     sb!kernel::control-stack-exhausted-error
     sb!di::handle-breakpoint
index efe76e3..5155c6a 100644 (file)
@@ -288,7 +288,7 @@ interrupt_handle_pending(os_context_t *context)
        {
            fake_foreign_function_call(context);
        }
-       funcall0(SymbolFunction(MAYBE_GC));
+       funcall0(SymbolFunction(SUB_GC));
 #ifndef __i386__
        if (were_in_lisp)
 #endif
@@ -606,7 +606,7 @@ interrupt_maybe_gc(int signal, siginfo_t *info, void *void_context)
        else {
            lispobj *old_free_space=current_dynamic_space;
            fake_foreign_function_call(context);
-           funcall0(SymbolFunction(MAYBE_GC));
+           funcall0(SymbolFunction(SUB_GC));
            undo_fake_foreign_function_call(context);
            if(current_dynamic_space==old_free_space) 
                /* MAYBE-GC (as the name suggest) might not.  If it
index b34ce86..f7ac900 100644 (file)
@@ -17,4 +17,4 @@
 ;;; checkins which aren't released. (And occasionally for internal
 ;;; versions, especially for internal versions off the main CVS
 ;;; branch, it gets hairier, e.g. "0.pre7.14.flaky4.13".)
-"0.pre8.84"
+"0.pre8.85"