src/hotspot/cpu/aarch64/gc/g1/g1BarrierSetAssembler_aarch64.cpp

@@ -130,10 +130,11 @@
    __ bind(done);
  }
  
  static void generate_queue_test_and_insertion(MacroAssembler* masm, ByteSize index_offset, ByteSize buffer_offset, Label& runtime,
                                                const Register thread, const Register value, const Register temp1, const Register temp2) {
+   assert_different_registers(value, temp1, temp2);
    // Can we store a value in the given thread's buffer?
    // (The index field is typed as size_t.)
    __ ldr(temp1, Address(thread, in_bytes(index_offset)));   // temp1 := *(index address)
    __ cbz(temp1, runtime);                                   // jump to runtime if index == 0 (full buffer)
    // The buffer is not full, store value into it.

@@ -203,11 +204,23 @@
    __ cbzw(tmp1, done);
    generate_pre_barrier_slow_path(masm, obj, pre_val, thread, tmp1, tmp2, done, runtime);
  
    __ bind(runtime);
  
-   __ push_call_clobbered_registers();
+   // save the live input values
+   RegSet saved = RegSet::of(pre_val);
+   FloatRegSet fsaved;
+ 
+   // Barriers might be emitted when converting between (scalarized) calling
+   // conventions for inline types. Save all argument registers before calling
+   // into the runtime.
+ 
+   // TODO 8366717 This came with 8284161: Implementation of Virtual Threads (Preview) later in May 2022
+   // Check if it's sufficient
+   //__ push_call_clobbered_registers();
+   assert_different_registers(rscratch1, pre_val); // push_CPU_state trashes rscratch1
+   __ push_CPU_state(true);
  
    // Calling the runtime using the regular call_VM_leaf mechanism generates
    // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
    // that checks that the *(rfp+frame::interpreter_frame_last_sp) == nullptr.
    //

@@ -224,11 +237,11 @@
      __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), pre_val, thread);
    } else {
      __ call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), pre_val, thread);
    }
  
-   __ pop_call_clobbered_registers();
+   __ pop_CPU_state(true);
  
    __ bind(done);
  
  }
  

@@ -365,44 +378,61 @@
    }
  }
  
  void G1BarrierSetAssembler::oop_store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
                                           Address dst, Register val, Register tmp1, Register tmp2, Register tmp3) {
+ 
+   bool in_heap = (decorators & IN_HEAP) != 0;
+   bool as_normal = (decorators & AS_NORMAL) != 0;
+   bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0;
+ 
+   bool needs_pre_barrier = as_normal && !dest_uninitialized;
+   bool needs_post_barrier = (val != noreg && in_heap);
+ 
+   assert_different_registers(val, tmp1, tmp2, tmp3);
+ 
    // flatten object address if needed
    if (dst.index() == noreg && dst.offset() == 0) {
      if (dst.base() != tmp3) {
        __ mov(tmp3, dst.base());
      }
    } else {
      __ lea(tmp3, dst);
    }
  
-   g1_write_barrier_pre(masm,
-                        tmp3 /* obj */,
-                        tmp2 /* pre_val */,
-                        rthread /* thread */,
-                        tmp1  /* tmp1 */,
-                        rscratch2  /* tmp2 */,
-                        val != noreg /* tosca_live */,
-                        false /* expand_call */);
+   if (needs_pre_barrier) {
+     g1_write_barrier_pre(masm,
+                          tmp3 /* obj */,
+                          tmp2 /* pre_val */,
+                          rthread /* thread */,
+                          tmp1  /* tmp1 */,
+                          rscratch2  /* tmp2 */,
+                          val != noreg /* tosca_live */,
+                          false /* expand_call */);
+   }
  
    if (val == noreg) {
      BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp3, 0), noreg, noreg, noreg, noreg);
    } else {
      // G1 barrier needs uncompressed oop for region cross check.
      Register new_val = val;
-     if (UseCompressedOops) {
-       new_val = rscratch2;
-       __ mov(new_val, val);
+     if (needs_post_barrier) {
+       if (UseCompressedOops) {
+         new_val = rscratch2;
+         __ mov(new_val, val);
+       }
      }
+ 
      BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp3, 0), val, noreg, noreg, noreg);
-     g1_write_barrier_post(masm,
-                           tmp3 /* store_adr */,
-                           new_val /* new_val */,
-                           rthread /* thread */,
-                           tmp1 /* tmp1 */,
-                           tmp2 /* tmp2 */);
+     if (needs_post_barrier) {
+       g1_write_barrier_post(masm,
+                             tmp3 /* store_adr */,
+                             new_val /* new_val */,
+                             rthread /* thread */,
+                             tmp1 /* tmp1 */,
+                             tmp2 /* tmp2 */);
+     }
    }
  
  }
  
  #ifdef COMPILER1