gencgc: Reduce conservatism for pointers to unboxed pages.

  * Because return addresses are unmarked, unboxed interior
pointers must be allowed to pin if they point within a code object.
Assuming that code objects are only allocated to CODE_PAGE_FLAG
pages, preserve_pointer() will allow any code_page_p() true page
to be pinned by any pointer to within its body.

  * But code_page_p() was broken, being an alternate version of
page_allocated_p() in implementation, with no warning or comment.
The net effect is extra conservatism: Any pointer to allocated
space will pin whatever page it points to.

  * Code pages aren't always allocated to CODE_PAGE_FLAG pages,
for a couple of reasons.  Until we can maintain such an invariant,
we can only use the weaker version (which DOES hold): code objects
are always allocated to boxed pages (either BOXED_PAGE_FLAG or
CODE_PAGE_FLAG, never FREE_PAGE_FLAG or UNBOXED_PAGE_FLAG).

  * Reduce conservatism in the GC by making code_page_p() delegate
to page_boxed_p(), thus tightening up the test for pinning unboxed
pages.

  * Also leave the "correct" logic for code_page_p() in place but
disabled (and make it actually BE correct), and add a comment
explaining part of what's going on.

diff --git a/NEWS b/NEWS
index 75355ee..5977d64 100644 (file)
--- a/NEWS
+++ b/NEWS
@@ -14,6 +14,10 @@ changes relative to sbcl-1.1.14:
   * enhancement: sb-ext:save-lisp-and-die on Windows now accepts
     :application-type argument, which can be :console or :gui. :gui allows
     having GUI applications without an automatically appearing console window.
+  * enhancement: reduced conservativism on GENCGC platforms:
+    conservative roots that point to unboxed pages must be tagged
+    pointers to the start of a valid-looking object, not merely point
+    to within the allocated part of the page, in order to pin the page.
   * bug fix: Windows applications without the console window no longer misbehave.
     (patch by Wilfredo Velazquez, lp#1256034).
   * bug fix: modular arithmetic optimizations do not stumble on dead branches

diff --git a/src/runtime/gencgc.c b/src/runtime/gencgc.c
index 5d218b1..77ec797 100644 (file)
--- a/src/runtime/gencgc.c
+++ b/src/runtime/gencgc.c
@@ -201,7 +201,18 @@ static inline boolean page_boxed_p(page_index_t page) {
 }
 
 static inline boolean code_page_p(page_index_t page) {
-    return (page_table[page].allocated & CODE_PAGE_FLAG);
+    /* This is used by the conservative pinning logic to determine if
+     * a page can contain code objects.  Ideally, we'd be able to
+     * check the page allocation flag to see if it is CODE_PAGE_FLAG,
+     * but this turns out not to be reliable (in fact, badly
+     * unreliable) at the moment.  On the upside, all code objects are
+     * boxed objects, so we can simply re-use the boxed_page_p() logic
+     * for a tighter result than merely "is this page allocated". */
+#if 0
+    return (page_table[page].allocated & CODE_PAGE_FLAG) == CODE_PAGE_FLAG;
+#else
+    return page_boxed_p(page);
+#endif
 }
 
 static inline boolean page_boxed_no_region_p(page_index_t page) {