RFR: JDK-8272574: C2: assert(false) failed: Bad graph detected in build_loop_late [v6]

[Christian Hagedorn]

On Mon, 23 Aug 2021 02:27:53 GMT, 王超 <github.com+25214855+casparcwang at openjdk.org> wrote:

>> Current loop predication will promote nodes(with a dependency on a later control node) to the insertion point which dominates the later control node.
>> 
>> In the following example, loopPrediction will promote node 434 to the outer loop(predicted insert point is right after node 424), and it depends on control node 207.  But node 424 dominates node 207, which means after the promotion, the cloned nodes have a control dependency on a later control node, which leads to a bad graph.
>> 
>> ![image](https://user-images.githubusercontent.com/25214855/129720970-ff65b8f4-8bef-401d-8590-54aca6de470e.png)
>> 
>> ![image](https://user-images.githubusercontent.com/25214855/129721369-4c61222b-7305-4522-9a37-e3e6e2138aa9.png)
>
> 王超 has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Add comments to make the code more readable

Otherwise, looks good! Thanks for making these changes.

I tested your code through tier1-7 over the weekend. Test results looked good!

src/hotspot/share/opto/loopPredicate.cpp line 674:

> 672:     if (phase->get_loop(predicate_proj) == phase->get_loop(offset_ctrl) &&
> 673:         phase->is_dominator(predicate_proj, offset_ctrl)) {
> 674:       assert(false, "cyclic dependency will lead to missed range check elimination");

Nit: I suggest to replace "will lead to missed" with "prevents".

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

Marked as reviewed by chagedorn (Reviewer).

PR: https://git.openjdk.java.net/jdk/pull/5142