RFR: 8214936: assert(_needs_refill == 0) failed: Forgot to handle a failed IC transition requiring IC stubs
Erik Österlund
erik.osterlund at oracle.com
Fri Dec 7 12:03:09 UTC 2018
Hi,
Per requested a trivial change that the verifier types be available as
dummies doing nothing in product builds so that we don't need to wrap
their uses in DEBUG_ONLY, as Per dislikes using DEBUG_ONLY.
Full:
http://cr.openjdk.java.net/~eosterlund/8214936/webrev.01/
Incremental:
http://cr.openjdk.java.net/~eosterlund/8214936/webrev.00_01/
Thanks,
/Erik
On 2018-12-06 14:46, Erik Österlund wrote:
> Hi,
>
> In 8214338 I moved IC stub refilling out from IC stub transitions,
> making it the responsibility of the caller. To make sure we don't
> accidentally forget calling IC stub refill and retrying at the call
> site, I added some verification code that complains if the caller failed
> to do so.
>
> I have tracked down that what can happen is that we try to make an IC
> megamorphic, but fail due to requiring IC stubs to do so. In such a
> failure, we try setting the IC to clean, expecting that would also fail,
> due to still not having refilled the buffers yet, but actually succeeding.
>
> As a result of this, the assert fails because we did not refill the IC
> stubs and retry after the transition to megamorphic failed, due to
> subsequently succeeding to clean the IC.
>
> In this patch, I made sure we refill and retry instead as intended. I
> have also improved the verification to use a thread-local verification
> RAII object that makes it easier to not just see that something
> unexpected happened, but also where it happened.
>
> Bug:
> https://bugs.openjdk.java.net/browse/JDK-8214936
>
> Webrev:
> http://cr.openjdk.java.net/~eosterlund/8214936/webrev.00/
>
> Thanks,
> /Erik
More information about the hotspot-compiler-dev
mailing list