[9] RFR(S): 8034216: assert(false) failed: infinite loop in PhaseIterGVN::optimize
Vladimir Kozlov
vladimir.kozlov at oracle.com
Thu Apr 10 17:11:49 UTC 2014
Albert,
You could do simple fix (--i) in original changes instead of adding new list. And actually you don't need to call
remove() because it will be removed by calling replace_node(), see remove_globally_dead_node():
+ for (uint i = 0; i < _worklist.size(); i++) {
+ Node* n = _worklist.at(i);
+
+ if (n->is_ConstraintCast()) {
+ Node* nn = n->Identity(this);
+ if (nn != n) {
+ replace_node(n, nn);
+ --i;
+ }
+ }
+ }
Thanks,
Vladimir
On 4/10/14 8:12 AM, Albert wrote:
> Hi Vladimir,
>
> Here is an updated version of the castPP node removal solution.
>
> http://cr.openjdk.java.net/~anoll/8034216/webrev.04/
> I executed specjvm98 with these changes on my local desktop and I could not see a performance regression.
>
> Best,
> Albert
>
> On 04/10/2014 07:46 AM, Albert wrote:
>> Hi Vladimir,
>>
>> thanks a lot for looking into this again. I will prepare a patch with which removes CastPP node (this was Roland's
>> idea). I will also run SpecJVM98 to see if there
>> is an impact on performance.
>>
>> Best,
>> Albert
>>
>>
>> On 04/09/2014 09:07 PM, Vladimir Kozlov wrote:
>>> Hi Albert,
>>>
>>> I am continue thinking about this problem.
>>> Your suggestion to process CastPP node first could be less intrusive than other solutions. It will do the same thing
>>> as current code if such nodes are first on worklist. It would be still nice to continue investigate the possibility
>>> to keep these nodes in graph after CCP but it should not be very high priority (file RFE).
>>>
>>> Note, you can't iterate the _worklist the way you do because remove() will put last element into current slot so you
>>> have to check it again.
>>> Also using remove(uint i) version will be faster.
>>>
>>> Thanks,
>>> Vladimir
>>>
>>> On 4/7/14 7:04 PM, Vladimir Kozlov wrote:
>>>> Hi Albert,
>>>>
>>>> Can you push the cleanup first under a different bug?
>>>> We may need to continue discussion how to fix the problem but I think
>>>> cleanup can be pushed now. Please, send a separate webrev for it under
>>>> new bug id.
>>>>
>>>> About the problem.
>>>>
>>>> It looks like 6916062 all over again :( See attached mail thread when I
>>>> did investigation:
>>>>
>>>> [217] CheckCastPP . . # java/util/ArrayList:NotNull*
>>>> [229] CheckCastPP . . #
>>>> java/util/Collections$UnmodifiableCollection:NotNull*
>>>> [292] Phi . [217] [229] # java/lang/Object:NotNull*
>>>> [11637] CastPP . [292] # _type = java/util/Collection* but
>>>> igvn.type(cast) = java/util/Collection:NotNull*
>>>> [11638] AddP . [11637][11637] + 8 # igvn.type(addp)
>>>> java/util/Collection:NotNull+8 *
>>>> [11639] LoadNKlass . . [11638]
>>>>
>>>> and we hit the assert in the 6916062 fix code:
>>>>
>>>> http://hg.openjdk.java.net/hsx/hsx19/baseline/rev/9a19ee0490e0
>>>>
>>>>
>>>> Types are converged during CCP phase when all CastPP nodes are still
>>>> present. Unfortunately, after CastPP nodes are removed, types could
>>>> degradate after your change:
>>>>
>>>> http://cr.openjdk.java.net/~anoll/8034216/webrev.01/
>>>>
>>>> In your case cached type for AddP (NotNull) will be replaced by input's
>>>> type because this is what AddPNode::Value() returns.
>>>>
>>>> I afraid this will affect types more than we want.
>>>>
>>>> Note, the only information about NotNull is located in the cached type
>>>> because corresponding CastPP was also removed.
>>>>
>>>>
>>>> Removing CastPP nodes first will replace cached type of AddP to even
>>>> more general j.l.Object (from Phi node).
>>>>
>>>> May be we should investigate if the next performance statement about
>>>> jvm98 is still true and solve it differently (in PhiNode::Ideal() for
>>>> example):
>>>>
>>>> // I tried to leave the CastPP's in. This makes the graph more accurate in
>>>> // some sense; we get to keep around the knowledge that an oop is not-null
>>>> // after some test. Alas, the CastPP's interfere with GVN (some values are
>>>> // the regular oop, some are the CastPP of the oop, all merge at Phi's
>>>> which
>>>> // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
>>>> // some of the more trivial cases in the optimizer. Removing more useless
>>>> // Phi's started allowing Loads to illegally float above null checks. I
>>>> gave
>>>> // up on this approach. CNC 10/20/2000
>>>>
>>>> Regards,
>>>> Vladimir
>>>>
>>>> On 4/4/14 8:09 AM, Albert wrote:
>>>>> Hi Vladimir,
>>>>>
>>>>> thanks for looking at this. Please see comments inline:
>>>>>
>>>>> On 04/04/2014 04:03 AM, Vladimir Kozlov wrote:
>>>>>> On 4/3/14 4:14 PM, Vladimir Kozlov wrote:
>>>>>>> Hi Albert,
>>>>>>>
>>>>>>> Thank you for doing this investigation. I hope it was good C2
>>>>>>> learning exercise :)
>>>>>>>
>>>>> Yes, I learned a lot. Many thanks to you and Roland for teaching me.
>>>>>
>>>>>>> I prefer the second solution, webrev.01. I don't like that we are
>>>>>>> loosing more concrete type with first solution.
>>>>>>
>>>>>> Actually it would be more correct and complete to have a solution
>>>>>> which does not put a load node on worklist again when there is no
>>>>>> progress with type. Even with webrev.01 we loose NotNull information.
>>>>> I have one general question here: Are types expected to converge at this
>>>>> point if everything would work as expected?
>>>>> Why do we loose NotNull information if types converge to
>>>>> java/util/Spliterator:NotNull+8 * ?
>>>>>
>>>>> I think that your suggestion, not pushing a node on the worklist if we
>>>>> don't make progress, will result in the type:
>>>>> java/util/Spliterator for the LoadKlass node - so the NotNull
>>>>> information would be lost in that case. If you think that
>>>>> this is the solution we should go for I'll look more into it.
>>>>>
>>>>> Here is a new webrev of the cleanup. I included all your suggestion. I
>>>>> deleted the two lines
>>>>> 1123 // Move users of node to worklist
>>>>> 1124 add_users_to_worklist( k );
>>>>>
>>>>> accidentally. Thanks for catching this.
>>>>> http://cr.openjdk.java.net/~anoll/8034216/webrev.03/
>>>>>
>>>>>>
>>>>>> Thanks,
>>>>>> Vladimir
>>>>>>
>>>>>>>
>>>>>>> It would be nice to detect such situations and bailout compilation in
>>>>>>> product VM and throw assert in debug VM.
>>>>>>>
>>>>>>> The clean up is nice, go for it. Note that you should put new methods
>>>>>>> under #ifndef PRODUCT instead of using
>>>>>>> PRODUCT_RETURN because they are called only in NOT_PRODUCT().
>>>>>>>
>>>>>>> Why you removed next code?:
>>>>>>>
>>>>>>> 1123 // Move users of node to worklist
>>>>>>> 1124 add_users_to_worklist( k );
>>>>>>>
>>>>>>> Next code should be under #ifdef ASSERT because TraceIterativeGVN is
>>>>>>> develop flag. So you don't need for NOT_PRODUCT()
>>>>>>> and DEBUG_ONLY() in the code:
>>>>>>>
>>>>>>> + if (TraceIterativeGVN && Verbose) {
>>>>>>> + tty->print(" Pop ");
>>>>>>> + NOT_PRODUCT(n->dump();)
>>>>>>> + DEBUG_ONLY(if( (num_processed++ % 100) == 0 )
>>>>>>> _worklist.print_set();)
>>>>>>> + }
>>>>>>>
>>>>>>> For next I also think is better to use #ifdef ASSERT:
>>>>>>>
>>>>>>> + DEBUG_ONLY(n->dump(4);)
>>>>>>> + DEBUG_ONLY(_worklist.dump();)
>>>>>>> + assert(false, "infinite loop in PhaseIterGVN::optimize");
>>>>>>>
>>>>>>> And #ifndef PRODUCT for next:
>>>>>>>
>>>>>>> + NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype));
>>>>>>> + NOT_PRODUCT(if (nn != n) verify_step((Node*) NULL);) //
>>>>>>> ignore n, it might be subsumed
>>>>>>>
>>>>>>>
>>>>>>> On 4/3/14 5:40 AM, Albert wrote:
>>>>>>>> Hi,
>>>>>>>>
>>>>>>>> could I get reviews for this patch?
>>>>>>>>
>>>>>>>> Bug:
>>>>>>>> ------
>>>>>>>> https://bugs.openjdk.java.net/browse/JDK-8034216
>>>>>>>>
>>>>>>>>
>>>>>>>> Problem:
>>>>>>>> ------------
>>>>>>>> The endless loop occurs, since there are 2 nodes of type LoadKlass in
>>>>>>>> the worklist that keep pushing themselves
>>>>>>>> onto the worklist. The current implementation can only handle 1 node
>>>>>>>> that keeps pushing itself on the worklist.
>>>>>>>> The endless loop occurs in the IGVN pass that comes after CCP phase.
>>>>>>>>
>>>>>>>> This piece of code in memnode.cpp is responsible for the endless loop:
>>>>>>>>
>>>>>>>> Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
>>>>>>>> ...
>>>>>>>> if (can_reshape && igvn != NULL &&
>>>>>>>> (igvn->_worklist.member(address) ||
>>>>>>>> igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) {
>>>>>>>> // The address's base and type may change when the address is
>>>>>>>> processed.
>>>>>>>> // Delay this mem node transformation until the address is
>>>>>>>> processed.
>>>>>>>> phase->is_IterGVN()->_worklist.push(this);
>>>>>>>> return NodeSentinel; // caller will return NULL
>>>>>>>> }
>>>>>>>> ...
>>>>>>>> }
>>>>>>>> As seen above, we keep pushing the current node onto the worklist if
>>>>>>>> t_adr != adr_type().
>>>>>>>> Unfortunately the types do not converge. See the following node dump:
>>>>>>>>
>>>>>>>> 5559 272 b 4 java.util.concurrent.CountedCompleter::exec (6 bytes)
>>>>>>>> 2054 CheckCastPP === 4086 2023 [[ 2098 2064 2064 ]]
>>>>>>>> #java/util/Spliterators$DoubleArraySpliterator:NotNull:exact *
>>>>>>>> Oop:java/util/Spliterators$DoubleArraySpliterator:NotNull:exact *
>>>>>>>> !jvms:
>>>>>>>> AbstractTask::compute @ bci:6 CountedCompleter::exec @ bci:-1
>>>>>>>> 2044 CheckCastPP === 4090 2023 [[ 2098 2077 2077 ]]
>>>>>>>> #java/util/ArrayList$ArrayListSpliterator:NotNull:exact *
>>>>>>>> Oop:java/util/ArrayList$ArrayListSpliterator:NotNull:exact * !jvms:
>>>>>>>> AbstractTask::compute @ bci:6 CountedCompleter::exec @ bci:-1
>>>>>>>> 2095 Region === 2095 2085 2072 [[ 2095 2120 2096 2097 2098 2099 ]]
>>>>>>>> !jvms: AbstractTask::compute @ bci:6 CountedCompleter::exec @ bci:-1
>>>>>>>> 2169 CmpL === _ 2099 2164 [[ 2170 ]] !jvms: AbstractTask::compute @
>>>>>>>> bci:29 CountedCompleter::exec @ bci:-1
>>>>>>>> 2121 IfTrue === 2120 [[ 2125 ]] #1 !jvms:
>>>>>>>> AbstractTask::getTargetSize
>>>>>>>> @ bci:8 AbstractTask::compute @ bci:14 CountedCompleter::exec @ bci:-1
>>>>>>>> 2122 IfFalse === 2120 [[ 2125 2162 ]] #0 !jvms:
>>>>>>>> AbstractTask::getTargetSize @ bci:8 AbstractTask::compute @ bci:14
>>>>>>>> CountedCompleter::exec @ bci:-1
>>>>>>>> 2098 Phi === 2095 2044 2054 [[ 2229 ]] #java/lang/Object:NotNull *
>>>>>>>> Oop:java/lang/Object:NotNull * !orig=2181 !jvms:
>>>>>>>> AbstractTask::compute @
>>>>>>>> bci:6 CountedCompleter::exec @ bci:-1
>>>>>>>> 2170 Bool === _ 2169 [[ 2171 ]] [le] !jvms: AbstractTask::compute @
>>>>>>>> bci:29 CountedCompleter::exec @ bci:-1
>>>>>>>> 2125 Region === 2125 2122 2121 [[ 2125 2171 2163 2164 ]] !jvms:
>>>>>>>> AbstractTask::getTargetSize @ bci:24 AbstractTask::compute @ bci:14
>>>>>>>> CountedCompleter::exec @ bci:-1
>>>>>>>> 22 ConL === 0 [[ 24 2037 81 2230 288 ]] #long:8
>>>>>>>> 2229 CastPP === 2173 2098 [[ 2240 2240 2230 2230 2237 ]]
>>>>>>>> #java/util/Spliterator * Interface:java/util/Spliterator * !jvms:
>>>>>>>> AbstractTask::compute @ bci:33 CountedCompleter::exec @ bci:-1
>>>>>>>> 3 Start === 3 0 [[ 3 5 6 7 8 9 10 ]] #{0:control, 1:abIO, 2:memory,
>>>>>>>> 3:rawptr:BotPTR, 4:return_address,
>>>>>>>> 5:java/util/concurrent/CountedCompleter:NotNull *}
>>>>>>>> 2171 If === 2125 2170 [[ 2172 2173 ]] P=0.500000, C=-1.000000 !jvms:
>>>>>>>> AbstractTask::compute @ bci:29 CountedCompleter::exec @ bci:-1
>>>>>>>> 2230 AddP === _ 2229 2229 22 [[ 2231 ]]
>>>>>>>> Interface:java/util/Spliterator+8 * !jvms: AbstractTask::compute @
>>>>>>>> bci:33 CountedCompleter::exec @ bci:-1
>>>>>>>> 7 Parm === 3 [[ 103 289 110 123 25 82 2231 2077 47 2064 2057 2038 65
>>>>>>>> 2031 75 2023 ]] Memory Memory: @BotPTR *+bot, idx=Bot; !jvms:
>>>>>>>> CountedCompleter::exec @ bci:-1
>>>>>>>> 2173 IfFalse === 2171 [[ 4079 2231 2229 ]] #0 !orig=[2195] !jvms:
>>>>>>>> AbstractTask::compute @ bci:29 CountedCompleter::exec @ bci:-1
>>>>>>>> 2231 LoadKlass === 2173 7 2230 [[ 2232 ]] @java/lang/Object+8 *,
>>>>>>>> idx=4; #klass java/util/Spliterator: 0x0000000002252c98 *
>>>>>>>> Interface:klass java/util/Spliterator: 0x0000000002252c98 * !jvms
>>>>>>>>
>>>>>>>> t_adr: is the cached type of the Address input (2230 AddP):
>>>>>>>> java/util/Spliterator:NotNull
>>>>>>>> adr_type(): is the bottom type of the address input (2230 AddP):
>>>>>>>> java/util/Spliterator
>>>>>>>>
>>>>>>>> The bottom_type of the AddP node is determined by the bottom type of
>>>>>>>> it's address input node ( 2229 CastPP).
>>>>>>>> The bottom type of the CastPP node is set in the CCP phase:
>>>>>>>>
>>>>>>>> Node *PhaseCCP::transform_once( Node *n ) {
>>>>>>>> ...
>>>>>>>> Node *nn = n->Ideal_DU_postCCP(this);
>>>>>>>> ...
>>>>>>>> }
>>>>>>>>
>>>>>>>> In the concrete execution, the CCP phase sets the bottom type of the
>>>>>>>> castPP node equal to the type of its input (in(1)).
>>>>>>>> in(1) of the castPP node is a phi-node. Since the two types are equal,
>>>>>>>> the castPP node is redundant and could be removed
>>>>>>>> in the subsequent IGVN phase. However, the castPP node is not removed,
>>>>>>>> since another node in the worklist (a redundant
>>>>>>>> region node) is processed prior to the castPP node. Removing the
>>>>>>>> redundant region node results in a change of in(1) of the
>>>>>>>> castPP node. As a result, the type of in(1) of the castPP node changes
>>>>>>>> as well and the castPP node is not removed. Not
>>>>>>>> removing the castPP node makes the types not converge as expected.
>>>>>>>>
>>>>>>>>
>>>>>>>> Solution:
>>>>>>>> -----------
>>>>>>>> Add a pass after CCP that removes the redundant castPP node. As a
>>>>>>>> result, the types converge to
>>>>>>>> java/lang/Object:NotNull+8 * and the endless loop does not occur.
>>>>>>>>
>>>>>>>> Here is the webrev:
>>>>>>>> http://cr.openjdk.java.net/~anoll/8034216/webrev.00/
>>>>>>>>
>>>>>>>>
>>>>>>>> Alternative solution:
>>>>>>>> --------------------------
>>>>>>>> Instead of adding the extra pass over the IR, we could raise the
>>>>>>>> bottom
>>>>>>>> type of a type node
>>>>>>>> in IGVN. As a result, the types converge to
>>>>>>>> java/util/Spliterator:NotNull+8 * and the endless loop
>>>>>>>> does not occur.
>>>>>>>>
>>>>>>>> Here is the webrev:
>>>>>>>> http://cr.openjdk.java.net/~anoll/8034216/webrev.01/
>>>>>>>>
>>>>>>>>
>>>>>>>> Cleanup:
>>>>>>>> ------------
>>>>>>>> In addition, I have worked on refactoring/cleaning up the affected
>>>>>>>> code.
>>>>>>>> This code cleanup comes
>>>>>>>> in a separate webrev to make it easier to figure out the actual
>>>>>>>> changes
>>>>>>>> that fix the bug. The code
>>>>>>>> cleanup is optional.
>>>>>>>>
>>>>>>>> Here is the webrev:
>>>>>>>> http://cr.openjdk.java.net/~anoll/8034216/webrev.02/
>>>>>>>>
>>>>>>>>
>>>>>>>> Testing:
>>>>>>>> ----------
>>>>>>>> Both solutions successfully passed JPRT and CTW (rt.jar).
>>>>>>>>
>>>>>>>> Many thanks in advance,
>>>>>>>> Albert
>>>>>
>>
>
More information about the hotspot-compiler-dev
mailing list