RFR: 8319932: [JVMCI] class unloading related tests can fail on libgraal

[David Holmes]

On Wed, 28 Feb 2024 12:45:17 GMT, Doug Simon <dnsimon at openjdk.org> wrote:

> This PR addresses a race in `runtime/ClassInitErrors/InitExceptionUnloadTest.java` that is exposed by libgraal which sets the [default value for `TypeProfileWidth` to 8](https://github.com/openjdk/jdk/blob/eb4b6fa6212380c9b6d3c94f5aa3384f12a0c125/src/hotspot/share/compiler/compilerDefinitions.cpp#L429).
> 
> The problem is that `InitExceptionUnloadTest$SpecialException` is loaded by a [custom class loader](https://github.com/openjdk/jdk/blob/eb4b6fa6212380c9b6d3c94f5aa3384f12a0c125/test/hotspot/jtreg/runtime/ClassInitErrors/InitExceptionUnloadTest.java#L106) and appears in the profile of `String.valueOf` as shown by `-XX:+PrintMethodData`:
> 
> static java.lang.String::valueOf(Ljava/lang/Object;)Ljava/lang/String;
>   interpreter_invocation_count: 957
>   invocation_counter: 957
>   backedge_counter: 0
>   decompile_count: 0
>   mdo size: 496 bytes
> 
>    0 nofast_aload_0
>    1 ifnonnull 10
>   0 bci: 1 BranchData taken(698) displacement(56)
>                                     not taken(0)
>    4 fast_aldc_w "null"
>    7 goto 14
>   32 bci: 7 JumpData taken(0) displacement(168)
>   10 nofast_aload_0
>   11 invokevirtual 1023 <java/lang/Object.toString()Ljava/lang/String;>
>   56 bci: 11 VirtualCallData count(2) entries(8)
>                                     'java/lang/String'(470 0.67)
>                                     'java/lang/StackTraceElement'(204 0.29)
>                                     'java/lang/Error'(1 0.00)
>                                     'java/lang/NoClassDefFoundError'(6 0.01)
>                                     'java/lang/ExceptionInInitializerError'(8 0.01)
>                                     'InitExceptionUnloadTest$SpecialException'(2 0.00)          <===== unloadable type in profile
>                                     'java/lang/OutOfMemoryError'(2 0.00)
> 
> 
> When this method is being compiled at tier 3, C1 makes a strong GC root for each type in the profile. Only once the compilation completes, are these strong roots released.
> 
> The test currently triggers class unloading via a [whitebox full GC](https://github.com/openjdk/jdk/blob/eb4b6fa6212380c9b6d3c94f5aa3384f12a0c125/test/lib/jdk/test/lib/classloader/ClassUnloadCommon.java#L67) and then immediately expects the classes to be unloaded. If the above C1 compilation is still in progress, the classes may not yet be unloaded.
> 
> This PR modifies the test to add a retry loop around the unloading trigger and the check for unloaded classes.

Changes requested by dholmes (Reviewer).

test/hotspot/jtreg/runtime/ClassInitErrors/InitExceptionUnloadTest.java line 131:

> 129:         int attempt = 0;
> 130:         while (!loadedClasses.isEmpty() && attempt < 20) {
> 131:             ClassUnloadCommon.triggerUnloading();

This sort of re-try logic is what is supposed to be encapsulated by `ClassUnloadCommon.triggerUnloading`, or perhaps more accurately the underlying `WhiteBox.fullGC()`. Individual tests should not have to know about such things.

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

PR Review: https://git.openjdk.org/jdk/pull/18044#pullrequestreview-1908024160
PR Review Comment: https://git.openjdk.org/jdk/pull/18044#discussion_r1507051110