8369616: JavaFrameAnchor on RISC-V has unnecessary barriers and wrong store order in MacroAssembler

Reviewed-by: fyang, luhenry

Author: feilongjiang

src/hotspot/cpu/riscv/javaFrameAnchor_riscv.hpp

@@ -39,25 +39,23 @@
   //  3 - restoring an old state (javaCalls)
 
   void clear(void) {
+    // No hardware barriers are necessary. All members are volatile and the profiler
+    // is run from a signal handler and the only observer is the thread its running on.
+
     // clearing _last_Java_sp must be first
     _last_Java_sp = nullptr;
-    OrderAccess::release();
     _last_Java_fp = nullptr;
     _last_Java_pc = nullptr;
   }
 
   void copy(JavaFrameAnchor* src) {
-    // In order to make sure the transition state is valid for "this"
+    // No hardware barriers are necessary. All members are volatile and the profiler
+    // is run from a signal handler and the only observer is the thread its running on.
+
     // We must clear _last_Java_sp before copying the rest of the new data
-    //
-    // Hack Alert: Temporary bugfix for 4717480/4721647
-    // To act like previous version (pd_cache_state) don't null _last_Java_sp
-    // unless the value is changing
-    //
     assert(src != nullptr, "Src should not be null.");
     if (_last_Java_sp != src->_last_Java_sp) {
       _last_Java_sp = nullptr;
-      OrderAccess::release();
     }
     _last_Java_fp = src->_last_Java_fp;
     _last_Java_pc = src->_last_Java_pc;

src/hotspot/cpu/riscv/macroAssembler_riscv.cpp

@@ -360,12 +360,14 @@ void MacroAssembler::set_last_Java_frame(Register last_java_sp,
     last_java_sp = esp;
   }
 
-  sd(last_java_sp, Address(xthread, JavaThread::last_Java_sp_offset()));
-
   // last_java_fp is optional
   if (last_java_fp->is_valid()) {
     sd(last_java_fp, Address(xthread, JavaThread::last_Java_fp_offset()));
   }
+
+  // We must set sp last.
+  sd(last_java_sp, Address(xthread, JavaThread::last_Java_sp_offset()));
+
 }
 
 void MacroAssembler::set_last_Java_frame(Register last_java_sp,