RFR: 8311279: TestStressIGVNAndCCP.java failed with different IGVN traces for the same seed

[Tobias Hartmann]

On Wed, 5 Jul 2023 09:10:48 GMT, Christian Hagedorn <chagedorn at openjdk.org> wrote:

> In `TestStressIGVNAndCCP`, we are executing a JVM twice and only want to compile and then run a single method `sum()`. The expectation is that we get the same output with `-XX:+TraceIterativeGVN`. However, our testing found a case where this did not match. When looking at the diff, I've noticed that with one JVM `sum()` had compile id 1, while the other JVM used compile id 2 for `sum()`. This makes a difference for the `debug_idx` which is printed for a dead node as "`compile id * 10000000000 + node index`":
> 
> Compile id 1:
> 
> 80 Phi === _ _ _ [[ ]] [10000000080] ...
> 
> Compile id 2:
> 
> 80 Phi === _ _ _ [[ ]] [20000000080] ...
> 
> 
> I was not able to reproduce the original report but my suspicion is that one JVM additionally compiled a native method wrapper or a method handle intrinsic for some reason but the other one did not. This would explain the different compile id because we are should only compiling `sum()` with the given `CompileOnly` JVM flag. A native compilation can be triggered, for example, by passing additionally passing `-esa` with `CompileOnly`. We get compile id 3 for `sum()`:
> 
>      60 1 n java.lang.invoke.MethodHandle::invokeBasic()I (native)
>      60 2 n java.lang.invoke.MethodHandle::linkToSpecial(LL)I (native) (static)
>      69 3 b compiler.debug.TestStressIGVNAndCCP::sum (27 bytes)
> 
> 
> To fix this, I suggest to use the `-XX:+CICountNative` flag which uses a separate counter for native compilations. Then, we'll always get compile id 1 for `sum()`:
> 
>      50 1 n java.lang.invoke.MethodHandle::invokeBasic()I (native)
>      51 2 n java.lang.invoke.MethodHandle::linkToSpecial(LL)I (native) (static)
>      59 1 b compiler.debug.TestStressIGVNAndCCP::sum (27 bytes)
> 
> 
> This is the same approach as done in [JDK-8269342](https://bugs.openjdk.org/browse/JDK-8269342) to reliably crash with `-XX:CICrashAt=1` in the first non-native compilation.
> 
> Thanks,
> Christian

That looks good and trivial to me.

-------------

Marked as reviewed by thartmann (Reviewer).

PR Review: https://git.openjdk.org/jdk/pull/14771#pullrequestreview-1514186228