Long-term maintenance of verifier

[Roger Riggs]

HI,

One way to detect getting out of sync, is to structure the test cases 
for the native verifier such that they can be run by the classfile 
verifier.  Especially use for for test cases developed for the latest 
spec changes.

$.02, Roger

On 6/6/25 9:31 PM, Chen Liang wrote:
> Thanks for the clarification. I totally agree with this decision of 
> staying on track of Option 2, as verification in the ClassFile API is 
> mostly a debug tool instead of a performance sensitive component.
>
> That said, I noted we are a bit out of sync - for example, JDK-8350029 
> is applied to runtime recently and it is clearly not in the ClassFile 
> API. On the mainline, lack of synchronization is not that big a deal 
> as we have only minor updates; however, project valhalla introduces 
> strict fields, which makes this synchronization more necessary, 
> especially that otherwise strict fields and early_larval_frame will be 
> regarded as verification errors.
>
> We can probably start our updating on mainline so we can help 
> valhalla. We currently can claim the CF verifier lacks any patch after 
> the initial 6/10/2022 publication in jdk-sandbox, but it is still 
> helpful to know which commit or which JDK release of verifier.cpp was 
> the copy based off - we can probably drop this 
> "up-to-date-to-tag/commit" info in our shadow so we can sync more easily.
>
> Adam, do you know about the exact fork version of the verifier code? 
> Once we have a point in history, it's quite easy to check the hotspot 
> commits and see if we need to update our copy accordingly.
>
> Regards, Chen
>
> On Fri, Jun 6, 2025 at 11:46 PM Brian Goetz <brian.goetz at oracle.com> 
> wrote:
>
>     We had some discussions about this when the verifier was
>     hand-translated from the C++ code.  At the time, some care was
>     taken to ensure that we _not_ refactor-as-we-go, to preserve some
>     structural similarity with the C++ code, so that deltas in the C++
>     code could be forward-ported.
>
>     We were initially skeptical that duplicating this code would work,
>     but it turned out to be quite effective.  Of course, long-term
>     maintenance has to be part of the story.
>
>     Going forward, I think we are still in the phase where the "keep
>     it synchronized" strategy (Adam's #2) can work.  We have to be
>     diligent to not make stylistic changes as we go, and keep up with
>     changes in the runtime.
>
>     Long term, I think there may be something we can do with #1, now
>     that Panama is finalized, but that is not enough. While calling
>     into the C++ code is now more practical, but that's not enough;
>     the runtime code is designed solely to support on-the-fly analysis
>     during class loading.  Some refactoring of the C++ API, undertaken
>     by the runtime team, would be needed to allow it to serve multiple
>     masters.
>
>     On 6/5/2025 10:01 PM, Chen Liang wrote:
>>     Hello,
>>     I have noted that the classfile API's copy of migrated verifier
>>     seems to naturally diverge from the c++ code: for example,
>>     JDK-8350029 that restricts invokespecial to not allow invoking
>>     arbitrary interface methods is not shadowed to the classfile
>>     verifier. This problem will only get more serious once strict
>>     fields are added. Meanwhile, people expect ClassFile.verify to be
>>     up-to-date with the runtime verifier.
>>
>>     What should we do to resolve this discrepancy? Should we have a
>>     separately maintained Java-based verifier implementing JVMS 4.10,
>>     or should we just increase our frequency of synchronizing with
>>     runtime?
>>
>>     Regards,
>>     Chen Liang
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/classfile-api-dev/attachments/20250609/a8ab272d/attachment.htm>