[16] RFR(S): C2: assert(no_dead_loop) failed: dead loop detected

[Tobias Hartmann]

Hi Christian,

nice analysis, looks good to me!

Small typo in cfgnode.cpp:2239 "reference" -> "references"

Best regards,
Tobias

On 30.07.20 15:15, Christian Hagedorn wrote:
> Hi
> 
> Please review the following patch:
> https://bugs.openjdk.java.net/browse/JDK-8249605
> http://cr.openjdk.java.net/~chagedorn/8249605/webrev.00/
> 
> There is a dead memory loop detected during IGVN. In the testcase, many nodes are dying during IGVN
> because they are not reachable anymore. In this process, a (not yet dead) memory phi node (150 Phi)
> with two inputs is processed (see [1]):
> (1) 289 MergeMem, whose base memory is 150 Phi and has one slice for 274 StoreD which is again an
> output of 150 Phi
> (2) 356 MergeMem, whose base memory is top (i.e. dead and would be removed when IGVN processes this
> node)
> 
> In PhiNode::Ideal, we check if a phi node is part of a dead loop where all its inputs reference
> itself over a MergeMemNode input whose base memory is the phi node itself again [2]. However, in
> this check we do not account for dead MergeMemNodes (like the input 356 MergeMem of 150 Phi).
> Therefore, we do not return top and apply the optimization [3] to replace 150 Phi by a new
> MergeMemNode (380 MergeMem) whose base memory is top and now has again one slice which is input and
> output to 274 StoreD [4]. This cycle is later detected and the assertion fails.
> 
> The fix accounts additionally for dead MergeMemNodes when trying to detect dead loops in
> PhiNode::Ideal to return top instead of a new dead MergeMemNode.
> 
> Best regards,
> Christian
> 
> 
> [1] https://bugs.openjdk.java.net/secure/attachment/89582/before_PhiNode_Ideal.png
> [2] http://hg.openjdk.java.net/jdk/jdk/file/8f7ede592c28/src/hotspot/share/opto/cfgnode.cpp#l2234
> [3] http://hg.openjdk.java.net/jdk/jdk/file/8f7ede592c28/src/hotspot/share/opto/cfgnode.cpp#l2246
> [4] https://bugs.openjdk.java.net/secure/attachment/89583/after_PhiNode_Ideal.png