|
Evaluation
|
This is the basic stuff :
HeapWord*CollectedHeap::common_mem_allocate_noinit
ffffffff7c4bb978 HeapWord*CollectedHeap::common_mem_allocate_noinit(unsigned long,Thread*) (5, 10573c2b8, fffffffed5c00c40, ffffffff7ce31112, ffffffff7c8688e4,0) + 80
ffffffff7c4badb8 oopDesc*CollectedHeap::obj_allocate(KlassHandle,int,Thread*) (10573c5b8, 5, 10573c2b8, 1f, ffffffff7c8688e4, 0) + c0
ffffffff7c4b30c8 instanceOopDesc*instanceKlass::allocate_instance(Thread*) (ffffffff3300bfb8, 10573c2b8, 0, 0, 2, 1001ced58) + 4b8
ffffffff7cd07f40 void OptoRuntime::new_C(klassOopDesc*,JavaThread*) (ffffffff3300bfa8, 10573c2b8, 0, 0, 30, fffffffefaeba558) + 1a0
ffffffff38998bbc ???????? (ffffffff3300bfa8, fffffffefaeba558, 2, fffffffefaeba540, 2, 86)
# assert(top <= end, "pointers out of order")
#
# Error ID: /net/jdk/export/jpse01/hshen/J2SE/140/hotspot/src/share/vm/memory/collectedHeap.inline.hpp, 121 [ Patched ]
#
115 HeapWord* CollectedHeap::allocate_from_tlab(Thread* thread, size_t size) {
116 assert(UseTLAB, "should use UseTLAB");
117
118 HeapWord* top = thread->tlab().top();
119 HeapWord* end = thread->tlab().end();
120
121 assert(top <= end, "pointers out of order");
122
123 if (pointer_delta(end, top) >= size) {
124 // successful thread-local allocation
125 if (!ZeroTLAB) {
126 // need to clear individual objects
127 Memory::set_words(top, size);
128 }
129 // This addition is safe because we know that top is
130 // at least size below end, so the add can't wrap.
131 thread->tlab().set_top(top + size);
132
133 assert(thread->tlab().invariants(), "TLAB integrity violated");
134 return top;
135 }
136 // Otherwise...
137 return allocate_from_tlab_slow(thread, size);
138 }
This is the basic stuff in the bug report:
Awaiting core file /machine access... cycling home now..
xxxxx@xxxxx 2002-09-25
----- -----
xxxxx@xxxxx 2002-09-25
Hitting this assert in 1.4.1 is probably an occurrence of the following bug
4746263 JDK 1.4.1 dumps core during ECperf; fails debug assert top <= end
A fix for this is going into 1.4.1_01 and is available in HotSpot's current main/baseline
/net/balvenie/export/imgr_home/archive/main/baseline/20020920073822.mpal.c2_merge_20020919/product.tgz
----- -----
xxxxx@xxxxx 2002-09-27
Customer has confirmed that crashes no longer occur with the fix for 4746263 in 1.4.1_01
and would like the fix backported to the next 1.4.0_x as well.
===============================
xxxxx@xxxxx 2002-10-03
Unfortunately the following does not appear to be sufficient to
fix this in 1.4.0 :
------- memnode.hpp -------
*** sccs.HmayZJ Thu Oct 3 10:35:43 2002
--- memnode.hpp Mon Sep 30 12:33:33 2002
***************
*** 1,5 ****
#ifdef USE_PRAGMA_IDENT_HDR
! #pragma ident "%W% %E% %U% JVM"
#endif
/*
* Copyright 1991-2002 Sun Microsystems, Inc. All rights reserved.
--- 1,5 ----
#ifdef USE_PRAGMA_IDENT_HDR
! #pragma ident "@(#)memnode.hpp 1.87 02/09/30 12:31:36 JVM"
#endif
/*
* Copyright 1991-2002 Sun Microsystems, Inc. All rights reserved.
***************
*** 115,120 ****
--- 115,130 ----
virtual uint ideal_reg() const { return Op_RegI; }
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual int store_Opcode() const { return Op_StoreB; }
+ // depends_only_on_test is almost always true, and needs to be almost always
+ // true to enable key hoisting & commoning optimizations. However, for the
+ // special case of RawPtr loads from TLS top & end, the control edge carries
+ // the dependence preventing hoisting past a Safepoint instead of the memory
+ // edge. (An unfortunate consequence of having Safepoints not set Raw
+ // Memory; itself an unfortunate consequence of having Nodes which produce
+ // results (new raw memory state) inside of loops preventing all manner of
+ // other optimizations). Basically, it's ugly but so is the alternative.
+ // See comment in graphkit.cpp, around line 1923 GraphKit::allocate_heap.
+ virtual bool depends_only_on_test() const { return adr_type() != TypeRawPtr::BOTTOM; }
};
//------------------------------LoadCNode--------------------------------------
***************
*** 148,153 ****
--- 158,164 ----
: LoadINode(c,mem,adr,TypeAryPtr::RANGE,ti) {}
virtual int Opcode() const;
virtual const Type *Value( PhaseTransform *phase ) const;
+ virtual bool depends_only_on_test() const { return true; }
};
//------------------------------LoadLNode--------------------------------------
***************
*** 322,327 ****
--- 333,339 ----
: LoadPNode(c,mem,adr,TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM) {}
virtual int Opcode() const;
virtual int store_Opcode() const { return Op_StorePConditional; }
+ virtual bool depends_only_on_test() const { return true; }
};
//------------------------------LoadLLockedNode---------------------------------
------- loopopts.cpp -------
*** sccs.pcai0J Thu Oct 3 10:36:19 2002
--- loopopts.cpp Mon Sep 30 12:33:33 2002
***************
*** 1,5 ****
#ifdef USE_PRAGMA_IDENT_SRC
! #pragma ident "%W% %E% %U% JVM"
#endif
//
// Copyright 1997-2002 Sun Microsystems, Inc. All rights reserved.
--- 1,5 ----
#ifdef USE_PRAGMA_IDENT_SRC
! #pragma ident "@(#)loopopts.cpp 1.162 02/09/30 12:31:22 JVM"
#endif
//
// Copyright 1997-2002 Sun Microsystems, Inc. All rights reserved.
***************
*** 629,638 ****
// If trying to do a 'Split-If' at the loop head, it is only
// profitable if the cmp folds up on BOTH paths. Otherwise we
// risk peeling a loop forever.
- // CNC - Disabled for now.
- if( n_ctrl->is_Loop() )
- policy = 999; // Policy requires BOTH paths to win
// Split compare 'n' through the merge point if it is profitable
Node *phi = split_thru_phi( n, n_ctrl, policy );
if( !phi ) return;
--- 629,647 ----
// If trying to do a 'Split-If' at the loop head, it is only
// profitable if the cmp folds up on BOTH paths. Otherwise we
// risk peeling a loop forever.
+ // CNC - Disabled for now. Requires careful handling of loop
+ // body selection for the cloned code. Also, make sure we check
+ // for any input path not being in the same loop as n_ctrl. For
+ // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
+ // because the alternative loop entry points won't be converted
+ // into LoopNodes.
+ IdealLoopTree *n_loop = get_loop(n_ctrl);
+ for( uint j = 1; j < n_ctrl->req(); j++ )
+ if( get_loop(n_ctrl->in(j)) != n_loop )
+ return;
+
+
// Split compare 'n' through the merge point if it is profitable
Node *phi = split_thru_phi( n, n_ctrl, policy );
if( !phi ) return;
Any suggestions?
Heres the stack/info after the above with the same assert:
# Java VM: Java HotSpot(TM) 64-Bit Server VM (1.4.0+4753265-TEST+20020930.152537
+chrisph-debug mixed mode)
#
# assert(top <= end, "pointers out of order")
#
# Error ID: /net/altair.east/terra/space5/chrisph/4746263/build/src/share/vm/mem
ory/collectedHeap.inline.hpp, 121 [ Patched ]
#
# Problematic Thread: prio=5 tid=0x1059d8228 nid=0x38 runnable
#
Dumping core....
core '/t3-1/gatherer/g29/Gatherer/core' of 15088: /c1t1/tal/jre/14003e/bin/sparcv9/java -server -showversion -Xms1280m -
----------------- lwp# 57 / thread# 56 --------------------
...
ffffffff7c4bd800 HeapWord*CollectedHeap::common_mem_allocate_noinit(unsigned long,Thread*) (5, 1059d8228, fffffffeba000c40, ffffffff7ce3c692, ffffffff7c86b164, 0) + 80
ffffffff7c4bcc40 oopDesc*CollectedHeap::obj_allocate(KlassHandle,int,Thread*) (1004dd608, 5, 1059d8228, 1f, ffffffff7c86b164, 0) + c0
ffffffff7c4b4f5c instanceOopDesc*instanceKlass::allocate_instance(Thread*) (ffffffff3300bfb8, 1059d8228, 0, 0, 2, 1001ced98) + 4c4
ffffffff7cd0ac38 void OptoRuntime::new_C(klassOopDesc*,JavaThread*) (ffffffff3300bfa8, 1059d8228, 0, 0, 37, fffffffeda724390) + 1a0
xxxxx@xxxxx 2002-10-03
Hmm, I seemed to have botched the backport... I am moving the code
from LoadBNode to LoadPNode in memnodes.hpp, we'll see if that makes the
fix better.
xxxxx@xxxxx 2002-10-03
OK... The customer confirms the fix [See Suggested Fix.]
xxxxx@xxxxx 2002-10-04
|
