SUGGESTED FIX
--- src/share/vm/opto/memnode.cpp- Mon Dec 18 18:53:13 2006
+++ src/share/vm/opto/memnode.cpp Wed Jan 17 22:59:57 2007
@@ -1,10 +1,10 @@
#ifdef USE_PRAGMA_IDENT_SRC
-#pragma ident "@(#)memnode.cpp 1.218 06/12/13 16:09:23 JVM"
+#pragma ident "@(#)memnode.cpp 1.219 07/01/17 15:38:52 JVM"
#endif
/*
- * @(#)memnode.cpp 1.218 06/12/13
+ * @(#)memnode.cpp 1.219 07/01/17
*
* Copyright 2005 Sun Microsystems, Inc. All rights reserved.
* SUN PROPRIETARY/CONFIDENTIAL. Use is subject to license terms.
*/
@@ -894,13 +894,24 @@
Node* slow = phi->in(1);
if (slow->is_Proj()) {
CallNode* call = slow->in(0)->is_Call();
if (call && (call->entry_point() == OptoRuntime::new_typeArray_Java() ||
call->entry_point() == OptoRuntime::new_objArray_Java())) {
- return call->in(TypeFunc::Parms + 1);
+ Node* length = call->in(TypeFunc::Parms + 1);
+ Node* slow_ctrl = call->in(0);
+ if (!(length->is_Phi() &&
+ slow_ctrl->is_Region() &&
+ length->in(0) == slow_ctrl)) {
+ // As long as the length node is produced before the
+ // allocation diamond it must dominate the result of the
+ // allocation. If split_if or some other transformation
+ // turns the length into a phi for the slow path then the
+ // length can't be used.
+ return length;
}
}
+ }
}
}
return this;
}
|
EVALUATION
With +EliminateZeroing, LoadRangeNode::Identity() finds a substitute for the LoadRangeNode by looking past a PhiNode. If the returned node does not dominate the LoadRangeNode, and we can not be sure that all uses are dominated by the new def.
In most cases, the Identity() transformation works without incident. However, in this test case, the split-if optimization works to move the PhiNode that is the length on a slow allocation path to a point that no longer dominates the use. (Using -XX:LoopOptsCount=0 can be used as a workaround.)
In the failing case, the MachNode live range formed by such a transformation is spilled by the register allocator. Reg_split asserts when it can't find a reaching def for the use in question.
|
Java