RFR: 8309268: C2: "assert(in_bb(n)) failed: must be" after JDK-8306302 [v2]
Tobias Hartmann
thartmann at openjdk.org
Mon Jun 5 05:39:10 UTC 2023
On Fri, 2 Jun 2023 13:23:34 GMT, Emanuel Peter <epeter at openjdk.org> wrote:
>> This is the fix to a regression caused in the CMoveV fix JDK-8306302.
>>
>> I had implicitly assumed that all `Cmp` in the loop also have their `in(1)` inside the loop (`in_bb`). This is not always true, and hence we hit the assert.
>>
>> **Solution**
>>
>> However, we know that at least one of the two inputs of a `Cmp` must also be in the loop, else the `Cmp` would float outside the loop. So if `in(1)` is not in the loop then we can just pick `in(2)` for the `velt_type`.
>>
>> **Testing**
>> I added 2 regression tests that were provided in the bug, and also extended `TestVectorConditionalMove.java` (though it is currently problemlisted because of an IR framework bug). This extension also triggered the assert, and now properly vectorizes.
>>
>> I tested up to tier6 and stress testing. I ran the tests both with `TestVectorConditionalMove` problemlisted and without the problemlisting.
>> **Running... but so far all good**
>
> Emanuel Peter has updated the pull request incrementally with one additional commit since the last revision:
>
> Roberto's suggestions
Looks good to me too.
-------------
Marked as reviewed by thartmann (Reviewer).
PR Review: https://git.openjdk.org/jdk/pull/14268#pullrequestreview-1461709992
More information about the hotspot-compiler-dev
mailing list