Re: [cp-patches] Patch: FYI: speed up split-for-gcj

[Mark Wielaard]

Hi Tom,

On Sat, 2006-04-01 at 12:11 -0700, Tom Tromey wrote:
> >>>>> "Mark" == Mark Wielaard <[EMAIL PROTECTED]> writes:
> 
> Mark> Unfortunately it seems to trigger some gcj bug with the svn trunk
> Mark> development version on builder (see below). I checked against 4.0.3 and
> Mark> 4.1 and they both seem fine. I am not 100% sure it is caused by this
> Mark> patch (still building a 4.2 version myself, but I need to go for today,
> Mark> so might not be able to check definitely till Monday).
> 
> I think this crash predates the split-for-gcj change by several hours.

Seems you are right. I tested against gcc/java revisions 112102, 112499
and 112540 which seem to compile things fine. Revision 112574 seems to
cause the failure, so it was probably this patch:

2006-03-30  Andrew Haley  <[EMAIL PROTECTED]>

        PR java/26858
        * lang.c (java_attribute_table): New.
        (LANG_HOOKS_ATTRIBUTE_TABLE): Define.
        * expr.c (build_field_ref): Add a null pointer check for all
        fields of offset > 4k.  Don't do so for accesses via the this
        pointer, which we know can never be null.
        * class.c (build_java_method_type): Mark arg 1 of all nonstatic
        methods nonnull.

After generating the list files in classpath/lib you can trigger the
compile error by:

$ /usr/local/gcc42/bin/gcj -Wno-deprecated --encoding=UTF-8
--bootclasspath ''
--classpath 
..:../vm/reference:..:../external/w3c_dom:../external/sax:../external/relaxngDatatype:.:
 -C -d . -MD -MF lists/java-lang.deps -MT lists/java-lang.stamp -MP 
@lists/java-lang.list
../java/lang/Class.java: In class 'java.lang.Class':
../java/lang/Class.java: In constructor
'(java.lang.Object,java.security.ProtectionDomain)':
../java/lang/Class.java:141: internal compiler error: Segmentation fault
Please submit a full bug report,
with preprocessed source if appropriate.
See <URL:http://gcc.gnu.org/bugs.html> for instructions.

It seems to be triggered by the assignment of this.pd to some other
Class pd field. Strangely enough the following workaround seems to not
trigger the compile failure:

diff -u -r1.43 Class.java
--- java/lang/Class.java        18 Dec 2005 13:03:48 -0000      1.43
+++ java/lang/Class.java        2 Apr 2006 20:27:40 -0000
@@ -138,7 +138,7 @@
     // saves the VM from having to worry about protection domains for array
     // classes.
     if (pd == null && isArray())
-      this.pd = getComponentType().pd;
+      this.pd = getComponentType().internalGetProtectionDomain();
     else
       this.pd = pd;
   }
@@ -1162,6 +1162,11 @@
     if (sm != null)
       sm.checkPermission(new RuntimePermission("getProtectionDomain"));

+    return internalGetProtectionDomain();
+  }
+
+  private ProtectionDomain internalGetProtectionDomain()
+  {
     return pd == null ? StaticData.unknownProtectionDomain : pd;
   }

So, if all else fails we can use that as workaround. But it would be
nice to know why it fails.

Mark

signature.asc
Description: This is a digitally signed message part