RFR: 8319793: C2 compilation fails with "Bad graph detected in build_loop_late" after JDK-8279888 [v7]

[Christian Hagedorn]

On Fri, 15 Dec 2023 08:41:20 GMT, Roland Westrelin <roland at openjdk.org> wrote:

>> Range check smearing and range check predication make an array access
>> dependent on 2 (or more in the case of RC smearing) conditions. As a
>> consequence, if a range check can be eliminated because there's an
>> identical dominating range check, the control dependent nodes that
>> could float and become dependent on the dominating range check cannot
>> be allowed to float because there's a risk that they would then bypass
>> one of the checks that make the access legal.
>> 
>> `IfNode::dominated_by()` and `PhaseIdealLoop::dominated_by()` have
>> logic to prevent this: nodes that are control dependent on a range
>> check or predicate are not allowed to float. This is however not
>> sufficient as demonstrated by the test cases.
>> 
>> In `TestArrayAccessAboveRCAfterSmearingOrPredication.testRangeCheckSmearing()`:
>> 
>> 
>>                 v += array[i];
>>                 if (flag2) {
>>                     if (flag3) {
>>                         field = 0x42;
>>                     }
>>                 }
>>                 if (flagField == 1) {
>>                     v += array[i];
>>                 }
>> 
>> 
>> The range check for the second `array[i]` load is replaced by the
>> dominating range check for the first `array[i]` but because the second
>> `array[i]` load could really be dependent on multiple range checks (in
>> case smearing happened which is not the case here), c2 doesn't allow
>> the second `array[i]` to float when the second range check is
>> removed. The second `array[i]` is then control dependent on:
>> 
>> 
>>                 if (flagField == 1) {
>> 
>> 
>> which is next found to be dominated by the same test:
>> 
>> 
>>         if (flag == 1) {
>> 
>> 
>> and is removed. However nothing in `dominated_by()` treats node
>> dependent on tests that are not range check or predicates
>> specially. So the second `array[i]` is allowed to float and become
>> dependent on:
>> 
>> 
>>         if (flag == 1) {
>> 
>> 
>> which is above the range check for that access. The test method in its
>> last invocation is passed an index for the array access that's widely
>> out of range. The array load happens before the range check and
>> crashes the VM. `testLoopPredication()` is a similar test where array
>> loads become dependent on predicates and end up above range checks.
>> 
>> `TestArrayAccessCastIIAboveRC.java` is the test case from the bug
>> where for similar reasons a range check `CastII` ends up above its
>> range check, becomes top because its input becomes some integer that
>> conflicts with its...
>
> Roland Westrelin has updated the pull request incrementally with two additional commits since the last revision:
> 
>  - Revert "Update src/hotspot/share/opto/castnode.hpp"
>    
>    This reverts commit 356c91cca911ed486f9f87f3eff53ce21e1e3ec9.
>  - Revert "Update src/hotspot/share/opto/memnode.hpp"
>    
>    This reverts commit bdb731ea562f314f44d327f7243ef5cf9ad40b2e.

Otherwise, looks good!

src/hotspot/share/opto/loopopts.cpp line 360:

> 358:         // dependent nodes end up at the lowest/nearest dominating check in the graph. To ensure that these Loads/Casts
> 359:         // do not float above any of the dominating checks (even when the lowest dominating check is later replaced by
> 360:         // yet another dominating check), we need to pin them at the lowest dominating check.

Should we also add this updated comment to `ifnode.cpp:569` and `ifnode.cpp:1536`?

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

Marked as reviewed by chagedorn (Reviewer).

PR Review: https://git.openjdk.org/jdk/pull/16886#pullrequestreview-1783911844
PR Review Comment: https://git.openjdk.org/jdk/pull/16886#discussion_r1427944258