The `compiler/intrinsics/bmi` tests started to fail with new AOT changes merged in.
### AOT Warnings with Valhalla due to Adapter Hash Collision >From the log messages, I saw that we have the following warning message >printed with UL: '[0.020s][warning][aot,codecache,stubs] Saved blob's name 'IIIIIIIL' is different from the expected name 'LIIIIIII' This warning can be traced back to a collision in caching adapters with AOT where two adapters have the same hash. As a result, we need to drop one of the adapters in the AOT cache which is unfortunate but not a problem in correctness. This hash collision seems to be specific to Valhalla because we have a lot more adapters compared to mainline. There is currently an open RFE in mainline ([JDK-8364929](https://bugs.openjdk.org/browse/JDK-8364929)) which tries to improve the situation to give each adapters a unique id. ### Comparison of `Xint` and `Xcomp` Output Fails due to different Timestamps The `compiler/intrinsics/bmi` tests call `BMITestRunner` which will run the same test with `-Xint` and `-Xcomp` and then compare the output: https://github.com/openjdk/valhalla/blob/28b96c080081b8ebab2656d5dc0f904072d15e40/test/hotspot/jtreg/compiler/intrinsics/bmi/BMITestRunner.java#L87-L100 We see the AOT warning actually in both runs. The comparison fails because we have a different timestamp in the log for `-Xint` and `-Xcomp`: [0.020s][warning][aot,codecache,stubs] Saved blob's name 'IIIIIIIL' is different from the expected name 'LIIIIIII' [0.013s][warning][aot,codecache,stubs] Saved blob's name 'IIIIIIIL' is different from the expected name 'LIIIIIII' So, the fix would be to get rid of the timestamps. ### Getting Rid of Timestamps already Implemented but Wrongly However, we already have some special code that actually wants to remove timestamps and all AOT logs altogether (introduced with https://github.com/openjdk/valhalla/commit/3b32f6a8ec37338764d3e6713247ff96e49bf5b3 for some Graal specific failure) : https://github.com/openjdk/valhalla/blob/28b96c080081b8ebab2656d5dc0f904072d15e40/test/hotspot/jtreg/compiler/intrinsics/bmi/BMITestRunner.java#L125-L129 But this does not seem to work. I played around with `Xlog` and found that surprisingly this: -Xlog:all=warning:stdout:level,tags -Xlog:aot=off seems to be expanded to -Xlog:all=warning:stdout:level,tags -Xlog:aot=off:stdout:<default decorators> and the second `Xlog` overrides the decorators and reintroduces the timestamps again. ### Proposed Fix The fix is to use one `Xlog` like this: -Xlog:all=warning,aot=off:stdout:level,tags I tested tier5 where the tests failed with `-XX:+AOTClassLinking` and this solved the issue. ### Only Apply to Valhalla and not Mainline I'm not sure why the test was not failing in mainline anymore. Maybe it was too intermittent and thus not noticed again. We could also fix this bug in mainline but since it's only failing in Valhalla, I suggest to do a point fix here and not upstream it to mainline for now. Thanks, Christian ------------- Commit messages: - 8367548: compiler/intrinsics/bmi tests fail with -XX:+AOTClassLinking Changes: https://git.openjdk.org/valhalla/pull/1582/files Webrev: https://webrevs.openjdk.org/?repo=valhalla&pr=1582&range=00 Issue: https://bugs.openjdk.org/browse/JDK-8367548 Stats: 2 lines in 1 file changed: 0 ins; 1 del; 1 mod Patch: https://git.openjdk.org/valhalla/pull/1582.diff Fetch: git fetch https://git.openjdk.org/valhalla.git pull/1582/head:pull/1582 PR: https://git.openjdk.org/valhalla/pull/1582
