From: Christophe Rhodes <csr21@cam.ac.uk>
Date: Fri, 23 Jan 2004 16:52:28 +0000 (+0000)
Subject: 0.8.7.21:
X-Git-Url: http://repo.macrolet.net/gitweb/?a=commitdiff_plain;h=60e88923a7610fa2cf6d5917707c149578776da2;p=sbcl.git

0.8.7.21:
	Dan's GC fragmentation fix
	... & BOXED_REGION
	... it builds, builds itself, doesn't break on PFD's tests;
		that's good enough for me :-)
	... I think I've summarized the "bug fix" right in NEWS, but
		do correct if not.
---

diff --git a/NEWS b/NEWS
index 176d94e..ee020d9 100644
--- a/NEWS
+++ b/NEWS
@@ -2237,15 +2237,19 @@ changes in sbcl-0.8.8 relative to sbcl-0.8.7:
   * bug fix: DECODE-UNIVERSAL-TIME now accepts timezone arguments with
     second-resolution: integer multiples of 1/3600 between -24 and 24.
     (thanks to Vincent Arkesteijn)
+  * bug fix: functions =, /=, <, <=, >, >= did not check the argument
+    type when called with 1 argument; PEEK-CHAR checked type of
+    PEEK-TYPE only after having read first character from a
+    stream.  (reported by Peter Graves)
+  * bug fix: the garbage collector now has much better locality
+    behaviour, and in particular no longer treats all memory as being
+    exhausted when many small objects point to each other in a deeply
+    recursive manner.
   * optimization: implemented multiplication as a modular
     (UNSIGNED-BYTE 32) operation on the x86 backend.
   * optimization: SEARCH on simple-base-strings can now be open-coded.
     (see also contrib/compiler-extras.lisp for inspiration for
     teaching the compiler about the Boyer-Moore algorithm).
-  * bug fix: functions =, /=, <, <=, >, >= did not check the argument
-    type when called with 1 argument; PEEK-CHAR checked type of
-    PEEK-TYPE only after having read first character from a
-    stream. (reported by Peter Graves)
   * fixed some bugs revealed by Paul Dietz' test suite:
     ** in stack analysis liveness information is propagated from
        non-local entry points.
diff --git a/src/runtime/gencgc.c b/src/runtime/gencgc.c
index e350702..5914f6f 100644
--- a/src/runtime/gencgc.c
+++ b/src/runtime/gencgc.c
@@ -966,16 +966,6 @@ gc_find_freeish_pages(int *restart_page_ptr, int nbytes, int unboxed)
 		    (unboxed ? UNBOXED_PAGE : BOXED_PAGE)) &&
 		   (page_table[first_page].large_object == 0) &&
 		   (page_table[first_page].gen == gc_alloc_generation) &&
-		   /* FIXME: Why?  Please tell me why?  Removal of
-		      this test, which restricts opening an allocation
-		      region on a partially-used page to the nursery
-		      generation, causes more or less instant heap
-		      corruption on forms such as 
-                        (loop repeat 2 
-                              do (compile nil '(lambda (x) x)) 
-                              do (gc :full t))
-		      -- CSR, 2003-01-14 */
-		   (gc_alloc_generation == 0) &&
 		   (page_table[first_page].bytes_used < (PAGE_BYTES-32)) &&
 		   (page_table[first_page].write_protected == 0) &&
 		   (page_table[first_page].dont_move == 0)) {
@@ -2763,7 +2753,7 @@ scavenge_newspace_generation_one_scan(int generation)
 	   generation));
     for (i = 0; i < last_free_page; i++) {
 	/* note that this skips over open regions when it encounters them */
-	if ((page_table[i].allocated == BOXED_PAGE)
+	if ((page_table[i].allocated & BOXED_PAGE)
 	    && (page_table[i].bytes_used != 0)
 	    && (page_table[i].gen == generation)
 	    && ((page_table[i].write_protected == 0)
diff --git a/version.lisp-expr b/version.lisp-expr
index cfd6bd6..e847bcc 100644
--- a/version.lisp-expr
+++ b/version.lisp-expr
@@ -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".)
-"0.8.7.20"
+"0.8.7.21"