RFR: Fix Minimal VM build failures

[Aleksey Shipilev]

On Mon, 27 Sep 2021 20:23:42 GMT, Chris Plummer <cjplummer at openjdk.org> wrote:

>> Minimal VM configs have no C2 and no JVMTI. New Loom code fails the Minimal VM builds because of this. See for example GHA runs: https://github.com/shipilev/loom/runs/3717293969?check_suite_focus=true
>> 
>> C2 needs just a little bit of protection when reaching to `DerivedPointersTable`.
>> 
>> JVMTI needs a bit more work. `jvmtiExport` and `jvmtiThreadState` headers are usually included even with JVMTI turned off, but `.cpp` would not be compiled without JVMTI. Therefore, non-trivial implementations should go into `.cpp`. Plus, some of the paths that call the actual methods declared but not implemented without JVMTI should be protected with `INCLUDE_JVMTI`.
>> 
>> Additional testing:
>>  - [x] Linux x86_64 server, `tier1_loom` still passes (includes JVMTI tests)
>>  - [x] Linux x86_64 minimal now builds, runs `tier1_loom` (cannot pass JVMTI tests without JVMTI)
>
> src/hotspot/share/prims/jvmtiThreadState.hpp line 99:
> 
>> 97:   static void finish_VTMT(jthread vthread, int callsite_tag);
>> 98:   static int  VTMT_disable_count();
>> 99:   static int  VTMT_count();
> 
> I'm not sure what is being fixed here.

As stated in PR: "jvmtiExport and jvmtiThreadState headers are usually included even with JVMTI turned off, but .cpp would not be compiled without JVMTI. Therefore, non-trivial implementations should go into .cpp." You cannot put the accessor bodies here, because the `_VTMT_count` and `_VTMT_disable_count` symbols would be inaccessible without `.cpp` compiled, and you'll get linker errors.

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

PR: https://git.openjdk.java.net/loom/pull/67