This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]

Re: Bootstrap failure with enable-checking in _Unwind_GetGR (Zack's function.c changes broken?)


On Sat, Sep 01, 2001 at 12:27:47PM +0200, Andreas Jaeger wrote:
> 
> make bootstrap fails:
...
> /cvs/gcc/gcc/unwind-dw2.c: In function `_Unwind_GetGR':
> /cvs/gcc/gcc/unwind-dw2.c:1182: Internal error: Segmentation fault
> Please submit a full bug report,
> with preprocessed source if appropriate.

That was an interesting one.

output_inline_function calls rest_of_compilation.  As its last action,
rest_of_compilation calls free_after_compilation and then
ggc_collect.  free_after_compilation destroys the current function
structure, causing cfun->eh to be garbage.  It clears cfun before
calling ggc_collect, but it doesn't clear DECL_SAVED_INSNS, and the
garbage collector tries to look through DECL_SAVED_INSNS now.

output_inline_function _does_ clear DECL_SAVED_INSNS, which is why I
didn't catch this; without always-collect, the only way this bug can
trip is if the ggc_collect call right at the end of rest_of_compilation
happens to trigger a collection.

Here's a patch.  Would you mind testing?  If I try to do an
always-collect build it'll take three days.  It has run to the end of
stage1, and I'm doing the C testsuite, which should only take the rest
of today.

The function.c change may not actually be necessary; I think we are
protected by pushed contexts in any case where f->outer is non-null.
But this is the Right Thing.

zw

	* function.c (ggc_mark_struct_function): Follow f->outer.
	* toplev.c (rest_of_compilation): Clear DECL_INLINE and
	DECL_SAVED_INSNS here...
	* integrate.c (output_inline_function): ... not here.

===================================================================
Index: function.c
--- function.c	2001/08/31 22:22:01	1.303
+++ function.c	2001/09/01 17:33:04
@@ -7703,6 +7703,8 @@ ggc_mark_struct_function (f)
     ggc_mark_rtvec ((rtvec) f->original_arg_vector);
   if (f->original_decl_initial)
     ggc_mark_tree (f->original_decl_initial);
+  if (f->outer)
+    ggc_mark_struct_function (f->outer);
 }
 
 /* Called once, at initialization, to initialize function.c.  */
===================================================================
Index: integrate.c
--- integrate.c	2001/08/31 22:22:02	1.159
+++ integrate.c	2001/09/01 17:33:05
@@ -2899,13 +2899,10 @@ output_inline_function (fndecl)
      before it gets mangled by optimization.  */
   (*debug_hooks->outlining_inline_function) (fndecl);
 
-  /* Compile this function all the way down to assembly code.  */
+  /* Compile this function all the way down to assembly code.  As a
+     side effect this destroys the saved RTL representation, but
+     that's okay, because we don't need to inline this anymore.  */
   rest_of_compilation (fndecl);
-
-  /* We can't inline this anymore; rest_of_compilation destroyed the
-     data structures describing the function.  */
-  DECL_INLINE (fndecl) = 0;
-  DECL_SAVED_INSNS (fndecl) = 0;
 
   cfun = old_cfun;
   current_function_decl = old_cfun ? old_cfun->decl : 0;
===================================================================
Index: toplev.c
--- toplev.c	2001/08/22 14:35:48	1.509
+++ toplev.c	2001/09/01 17:33:06
@@ -3773,7 +3773,14 @@ rest_of_compilation (decl)
   /* We're done with this function.  Free up memory if we can.  */
   free_after_parsing (cfun);
   if (! DECL_DEFER_OUTPUT (decl))
-    free_after_compilation (cfun);
+    {
+      free_after_compilation (cfun);
+
+      /* Clear integrate.c's pointer to the cfun structure we just
+	 destroyed.  */
+      DECL_INLINE (decl) = 0;
+      DECL_SAVED_INSNS (decl) = 0;
+    }
   cfun = 0;
 
   ggc_collect ();


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]