Add opt-in link-at-build-time option to fail fast on NoClassDefFound #6725
Conversation
oracle-contributor-agreement
bot
added
the
OCA Verified
zakkak
added
redhat-interest
and removed
OCA Verified
zakkak
force-pushed
the
2023-04-27-fix-6253
branch
from
0b1eb3b to
d225ddf
Compare
oracle-contributor-agreement
bot
added
the
OCA Verified
zakkak
force-pushed
the
2023-04-27-fix-6253
branch
from
d225ddf to
b0fc67f
Compare
There was a problem hiding this comment.
This patch looks good to me. It gives users who rely on linkAtBuildTime the option to bail out with an informative error while still preserving the default behaviour: i.e. continuing the analysis with an UnknownType but foregoing the opportunity to observe what was the cause of the linkage error. This change means that users who run into an unexplained linkage issue with the default setting will not be left in the dark. They can rerun with this flag set in order to identify what went wrong.
| Throwable cause = e.getCause(); | ||
| if (cause instanceof NoClassDefFoundError && linkAtBuildTime && LinkAtBuildTimeSupport.singleton().failFast()) { | ||
| reportUnresolvedElement("type", method.format("%H.%n(%P)"), e); | ||
| } else if (cause instanceof LinkageError) { |
There was a problem hiding this comment.
Why is this dropping the IllegalAccessError check?
There was a problem hiding this comment.
Because it is a subclass of LinkageError
Works around oracle#6253 till a proper JVCMI fix is available. Closes: oracle#6253
zakkak
force-pushed
the
2023-04-27-fix-6253
branch
from
b0fc67f to
d1555ac
Compare
|
As mentioned in #6253 (comment) this appears to no longer be needed. |
Hi @christianwimmer and @cstancu
this is a respin of #6507
I understand that a proper solution requires JVMCI changes and I am willing to work on it given some guidance. However, this is expected to take significant time to reach a release.
In the meantime, #6253 (originally reported in #4661) is a long standing issue for Quarkus users and we would really like a fix for it (even a work around like the one I propose). Note that by default Quarkus initializes and links all classes at build time, so this is not a corner case for Quarkus.
This PR works around #6253, by providing users the option to fail fast if a class is missing and they are linking at build time. The option is intentionally made an opt-in in order to not alter the current behavior till a proper JVCMI fix is available and it should be completely removed once we get the JVMCI fix in.
Ideally we would like this backported to 23.0 as well.
To try this out you may run:
Passing
-H:+LinkAtBuildTimeFailFastresults in the following output:while without it we get: