RFR: JDK-8296764 NMT: reduce loads in os::malloc

[Ioi Lam]

On Thu, 10 Nov 2022 06:11:28 GMT, Thomas Stuefe <stuefe at openjdk.org> wrote:

> We read the NMT level at every ps::malloc/realloc/free etc, even if NMT is off. That is unavoidable. But we also read a second variable that shadows the NMT level `_nmt_was_initialized`. That state is synonymous with NMT level "unknown," so there is no need for this second load.
> 
> Furthermore, NMT level is volatile. That had been necessary in older times since we used to shut down NMT on resource exhaustion. But since [JDK-8256844](https://bugs.openjdk.org/browse/JDK-8256844), NMT level is fixed at initialization, so we can remove the volatile specifier.
> 
> Patch 
> - removes `_nmt_was_initialized` and redirects it to `MemTracker::is_initialized()`, which gets inlined to a load of NMT level. 
> - makes NMT level non-volatile
> - removes _is_nmt_env_valid, which had been dead code

src/hotspot/share/services/nmtPreInit.cpp line 182:

> 180: //  in post-init, no modifications to the lookup table are possible.
> 181: void NMTPreInit::pre_to_post() {
> 182:   assert(MemTracker::is_initialized() == false, "just once");

Maybe this should be `!MemTracker::is_initialized()` to be consistent with the rest of the code?

test/hotspot/gtest/nmt/test_nmtpreinit.cpp line 116:

> 114:   }
> 115:   void test_post() {
> 116:     assert(!MemTracker::is_initialized() == false,

Hmm, I am confused by the double negation :-)

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

PR: https://git.openjdk.org/jdk/pull/11080