On Wed, 14 Jun 2023 01:02:09 GMT, Alex Menkov <[email protected]> wrote:
> If virtual thread has frames in stackChunks, need to apply load barriers
> before processing nmethods (javaVFrame::locals() and
> javaVFrame::expressions() do it internally)
>
> Testing: tier1-tier5 (still in progress);
> 400 runs of VThreadStackRefTest.java test on linux-x64 and linux-x64-debug
> with "-XX:+UseZGC -Xcomp -XX:-TieredCompilation"
src/hotspot/share/prims/jvmtiTagMap.cpp line 2325:
> 2323: // Need to apply load barriers for unmounted vthreads.
> 2324: nmethod* nm = jvf->cb()->as_nmethod();
> 2325: nm->run_nmethod_entry_barrier();
This looks good in general.
However, I'm looking at the `stackChunkOopDesc::do_barriers0()` implementation
and wonder if a similar trick is needed for interpreter frames (to support
`Class Redefinition`):
template <stackChunkOopDesc::BarrierType barrier, ChunkFrames frame_kind,
typename RegisterMapT>
void stackChunkOopDesc::do_barriers0(const StackChunkFrameStream<frame_kind>&
f, const RegisterMapT* map) {
// We need to invoke the write barriers so as not to miss oops in old chunks
that haven't yet been concurrently scanned
assert (!f.is_done(), "");
if (f.is_interpreted()) {
Method* m = f.to_frame().interpreter_frame_method();
// Class redefinition support
m->record_gc_epoch();
} else if (f.is_compiled()) {
nmethod* nm = f.cb()->as_nmethod();
// The entry barrier takes care of having the right synchronization
// when keeping the nmethod alive during concurrent execution.
nm->run_nmethod_entry_barrier();
// There is no need to mark the Method, as class redefinition will walk the
// CodeCache, noting their Methods
}
. . .
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/14460#discussion_r1229273861
