[9] RFR(S): 8134739: compiler/loopopts/superword/TestVectorizationWithInvariant crashes in loop opts

Tobias Hartmann tobias.hartmann at oracle.com
Wed Sep 9 11:56:29 UTC 2015


Roland asked (off-thread) why we still optimize the empty loop after replacing the PhiNode to collapse it. Here is a more detailed trace of events to clarify this. We have the following loops:

 Loop: N2433/N2436  limit_check predicated counted [0,int),+8 (4 iters)  pre
 Loop: N2484/N1546  counted [int,92),+16 (-2147483648 iters)  main
 Loop: N2388/N2391  counted [int,100),+8 (4 iters)  post

IdealLoopTree::iteration_split() first looks a the pre-loop and removes it because it's empty:


 Empty with zero trip guard       Loop: N2433/N2436  limit_check predicated counted [0,100),+8 (4 iters)  pre

We do not continue optimizing the pre-loop but bail out when returning from IdealLoopTree::policy_do_remove_empty_loop() and continue with the main-loop:

 SuperWord::transform_loop(IdealLoopTree* lpt)
 SuperWord::get_pre_loop_end(CountedLoopNode* cl)
 // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))

The SuperWordLoopUnrollAnalysis now looks at the pre-loop again and fails because the PhiNode was replaced.

I noticed that the output of IdealLoopTree::dump_head() sometimes contains negative trip counts ("-2147483648 iters"). I filed JDK-8135252.


On 07.09.2015 14:03, Tobias Hartmann wrote:
> Hi,
> please review the following patch.
> https://bugs.openjdk.java.net/browse/JDK-8134739
> http://cr.openjdk.java.net/~thartmann/8134739/webrev.00
> Problem:
> The VM crashes during loop optimizations in CountedLoopEndNode::loopnode() because CountedLoopEndNode::phi() does not return a PhiNode but a SubINode. The problem is that the phi (2432 PhiNode) was replaced by the final value of the last loop iteration (2881 SubINode) to collapse the empty loop in IdealLoopTree::policy_do_remove_empty_loop(). See before [1] and after [2].
> Solution:
> CountedLoopEndNode::phi() should check if the node really is a PhiNode and return NULL otherwise.
> Testing:
> - Failing test
> - JPRT
> Thanks,
> Tobias
> [1] https://bugs.openjdk.java.net/secure/attachment/52878/emptyLoop1.png
> [2] https://bugs.openjdk.java.net/secure/attachment/52879/emptyLoop2.png

More information about the hotspot-compiler-dev mailing list