RFR: 8262017: C2: assert(n != __null) failed: Bad immediate dominator info. [v4]

[Roland Westrelin]

On Tue, 15 Jun 2021 07:31:19 GMT, Christian Hagedorn <chagedorn at openjdk.org> wrote:

>> The wrong dominance information is caused by a `ConvL2I` node whose input gets a wrong type which is out of the range of the `ConvL2I`. As a consequence, the node becomes top and some data nodes are folded away which eventually results in the dominance assertion failure.
>> 
>> The wrong type information can be traced back to the overflow/underflow handling in range check elimination. Originally, the code in `PhaseIdealLoop::add_constraint()` covered the following special case for the adjustment of the pre loop limit:
>> https://github.com/openjdk/jdk/blob/cd33abb136d415455090cef4fe6211d9e6940948/src/hotspot/share/opto/loopTransform.cpp#L2423-L2456
>> 
>> This code, however, was removed later by accident:
>> https://github.com/openjdk/jdk/commit/afd852ccb8f8d36c721963ba68de695ce3e0e23d#diff-6a59f91cb710d682247df87c75faf602f0ff9f87e2855ead1b80719704fbedffL2370-L2394
>> 
>> This was only revealed after enabling more precise type information for phi nodes in [JDK-8257813](https://bugs.openjdk.java.net/browse/JDK-8257813) after which the testcase starts to fail.
>> 
>> The proposed fix is straight forward to reintroduce the wrongly removed code and adapt it to long instead of int.
>> 
>> Thanks,
>> Christian
>
> Christian Hagedorn has updated the pull request incrementally with one additional commit since the last revision:
> 
>   cleanup & signed_max/min

src/hotspot/share/opto/loopTransform.cpp line 2351:

> 2349:   Node* inner_result_long;
> 2350:   if (is_positive_stride) {
> 2351:     inner_result_long = MaxNode::signed_max(limit, _igvn.longcon(min_jint), TypeLong::LONG, _igvn);

You still need to register the new nodes with PhaseIdealLoop (what you achieved before with register_new_node()). set_subtree_ctrl() is what you need I think.

-------------

PR: https://git.openjdk.java.net/jdk/pull/4362