RFR: 8365229: ARM32: c2i_no_clinit_check_entry assert failed after JDK-8364269 [v4]
Boris Ulasevich
bulasevich at openjdk.org
Wed Aug 13 18:42:14 UTC 2025
On Wed, 13 Aug 2025 11:16:59 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:
>> When recording adapter entries, we record _offsets_, not the actual addresses:
>>
>>
>> entry_offset[3] = handler->get_c2i_no_clinit_check_entry() - i2c_entry;
>>
>>
>> Every platform except ARM32 and Zero have all these entries set up, so offset are always sane. But those two platforms set up `nullptr` as `c2i_no_clinit_check_entry()`, because clinit barriers are unimplemented. So the new assert added in [JDK-8364269](https://bugs.openjdk.org/browse/JDK-8364269) fails encountering effectively `nullptr - i2c_entry` "garbage".
>>
>> This PR is the second least horrible (IMO) fix for this: relaxing assert by checking that "out of range" values are actually wrapping around back to `0`/`nullptr`. Had to do it in unsigned ints to avoid UB. For the affected platforms, we do not actually access this problematic/garbage entry offset, since we are always checking if clinit barriers are enabled. So the assert is the only place where it matters.
>>
>> The least horrible solution would be storing the actual `address`-es instead of `int` offsets. But that likely has footprint implications.
>>
>> Additional testing:
>> - [x] Linux x86_64 server fastdebug, `runtime/cds` still works
>> - [x] Linux ARM32 server fastdebug, `java -version` now works
>> - [x] Linux x86_64 zero fastdebug, `make bootcycle-images` now works
>
> Aleksey Shipilev has updated the pull request incrementally with one additional commit since the last revision:
>
> Handling Zero crash as well
Marked as reviewed by bulasevich (Committer).
-------------
PR Review: https://git.openjdk.org/jdk/pull/26746#pullrequestreview-3117248287
More information about the hotspot-compiler-dev
mailing list