diff -r 6390e7056d12 -r bfcdcd00e4fb src/hotspot/cpu/x86/gc/shenandoah/shenandoahBarrierSetAssembler_x86.cpp
--- a/src/hotspot/cpu/x86/gc/shenandoah/shenandoahBarrierSetAssembler_x86.cpp	Mon Aug 12 20:27:14 2019 +0200
+++ b/src/hotspot/cpu/x86/gc/shenandoah/shenandoahBarrierSetAssembler_x86.cpp	Mon Aug 12 23:37:48 2019 +0200
@@ -752,108 +752,6 @@
   }
 }
 
-void ShenandoahBarrierSetAssembler::save_vector_registers(MacroAssembler* masm) {
-  int num_xmm_regs = LP64_ONLY(16) NOT_LP64(8);
-  if (UseAVX > 2) {
-    num_xmm_regs = LP64_ONLY(32) NOT_LP64(8);
-  }
-
-  if (UseSSE == 1)  {
-    __ subptr(rsp, sizeof(jdouble)*8);
-    for (int n = 0; n < 8; n++) {
-      __ movflt(Address(rsp, n*sizeof(jdouble)), as_XMMRegister(n));
-    }
-  } else if (UseSSE >= 2)  {
-    if (UseAVX > 2) {
-      __ push(rbx);
-      __ movl(rbx, 0xffff);
-      __ kmovwl(k1, rbx);
-      __ pop(rbx);
-    }
-#ifdef COMPILER2
-    if (MaxVectorSize > 16) {
-      if(UseAVX > 2) {
-        // Save upper half of ZMM registers
-        __ subptr(rsp, 32*num_xmm_regs);
-        for (int n = 0; n < num_xmm_regs; n++) {
-          __ vextractf64x4_high(Address(rsp, n*32), as_XMMRegister(n));
-        }
-      }
-      assert(UseAVX > 0, "256 bit vectors are supported only with AVX");
-      // Save upper half of YMM registers
-      __ subptr(rsp, 16*num_xmm_regs);
-      for (int n = 0; n < num_xmm_regs; n++) {
-        __ vextractf128_high(Address(rsp, n*16), as_XMMRegister(n));
-      }
-    }
-#endif
-    // Save whole 128bit (16 bytes) XMM registers
-    __ subptr(rsp, 16*num_xmm_regs);
-#ifdef _LP64
-    if (VM_Version::supports_evex()) {
-      for (int n = 0; n < num_xmm_regs; n++) {
-        __ vextractf32x4(Address(rsp, n*16), as_XMMRegister(n), 0);
-      }
-    } else {
-      for (int n = 0; n < num_xmm_regs; n++) {
-        __ movdqu(Address(rsp, n*16), as_XMMRegister(n));
-      }
-    }
-#else
-    for (int n = 0; n < num_xmm_regs; n++) {
-      __ movdqu(Address(rsp, n*16), as_XMMRegister(n));
-    }
-#endif
-  }
-}
-
-void ShenandoahBarrierSetAssembler::restore_vector_registers(MacroAssembler* masm) {
-  int num_xmm_regs = LP64_ONLY(16) NOT_LP64(8);
-  if (UseAVX > 2) {
-    num_xmm_regs = LP64_ONLY(32) NOT_LP64(8);
-  }
-  if (UseSSE == 1)  {
-    for (int n = 0; n < 8; n++) {
-      __ movflt(as_XMMRegister(n), Address(rsp, n*sizeof(jdouble)));
-    }
-    __ addptr(rsp, sizeof(jdouble)*8);
-  } else if (UseSSE >= 2)  {
-    // Restore whole 128bit (16 bytes) XMM registers
-#ifdef _LP64
-    if (VM_Version::supports_evex()) {
-      for (int n = 0; n < num_xmm_regs; n++) {
-        __ vinsertf32x4(as_XMMRegister(n), as_XMMRegister(n), Address(rsp, n*16), 0);
-      }
-    } else {
-      for (int n = 0; n < num_xmm_regs; n++) {
-        __ movdqu(as_XMMRegister(n), Address(rsp, n*16));
-      }
-    }
-#else
-    for (int n = 0; n < num_xmm_regs; n++) {
-      __ movdqu(as_XMMRegister(n), Address(rsp, n*16));
-    }
-#endif
-    __ addptr(rsp, 16*num_xmm_regs);
-
-#ifdef COMPILER2
-    if (MaxVectorSize > 16) {
-      // Restore upper half of YMM registers.
-      for (int n = 0; n < num_xmm_regs; n++) {
-        __ vinsertf128_high(as_XMMRegister(n), Address(rsp, n*16));
-      }
-      __ addptr(rsp, 16*num_xmm_regs);
-      if (UseAVX > 2) {
-        for (int n = 0; n < num_xmm_regs; n++) {
-          __ vinsertf64x4_high(as_XMMRegister(n), Address(rsp, n*32));
-        }
-        __ addptr(rsp, 32*num_xmm_regs);
-      }
-    }
-#endif
-  }
-}
-
 #undef __
 
 #ifdef COMPILER1
@@ -885,11 +783,17 @@
 }
 
 void ShenandoahBarrierSetAssembler::gen_load_reference_barrier_stub(LIR_Assembler* ce, ShenandoahLoadReferenceBarrierStub* stub) {
+  ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
   __ bind(*stub->entry());
 
-  Label done;
   Register obj = stub->obj()->as_register();
   Register res = stub->result()->as_register();
+  Register tmp1 = stub->tmp1()->as_register();
+  Register tmp2 = stub->tmp2()->as_register();
+
+  Label slow_path;
+
+  assert(res == rax, "result must arrive in rax");
 
   if (res != obj) {
     __ mov(res, obj);
@@ -897,11 +801,35 @@
 
   // Check for null.
   __ testptr(res, res);
-  __ jcc(Assembler::zero, done);
+  __ jcc(Assembler::zero, *stub->continuation());
+
+  // Check for object being in the collection set.
+  __ mov(tmp1, res);
+  __ shrptr(tmp1, ShenandoahHeapRegion::region_size_bytes_shift_jint());
+  __ movptr(tmp2, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr());
+  __ movbool(tmp2, Address(tmp2, tmp1, Address::times_1));
+  __ testbool(tmp2);
+  __ jcc(Assembler::zero, *stub->continuation());
 
-  load_reference_barrier_not_null(ce->masm(), res);
+  // Test if object is resolved.
+  __ movptr(tmp1, Address(res, oopDesc::mark_offset_in_bytes()));
+  // Test if both lowest bits are set. We trick it by negating the bits
+  // then test for both bits clear.
+  __ notptr(tmp1);
+  __ testb(tmp1, markOopDesc::marked_value);
+  __ jccb(Assembler::notZero, slow_path);
+  // Clear both lower bits. It's still inverted, so set them, and then invert back.
+  __ orptr(tmp1, markOopDesc::marked_value);
+  __ notptr(tmp1);
+  // At this point, tmp1 contains the decoded forwarding pointer.
+  __ mov(res, tmp1);
 
-  __ bind(done);
+  __ jmp(*stub->continuation());
+
+  __ bind(slow_path);
+  ce->store_parameter(res, 0);
+  __ call(RuntimeAddress(bs->load_reference_barrier_rt_code_blob()->code_begin()));
+
   __ jmp(*stub->continuation());
 }
 
@@ -965,6 +893,18 @@
   __ epilogue();
 }
 
+void ShenandoahBarrierSetAssembler::generate_c1_load_reference_barrier_runtime_stub(StubAssembler* sasm) {
+  __ prologue("shenandoah_load_reference_barrier", false);
+  // arg0 : object to be resolved
+
+  __ save_live_registers_no_oop_map(true);
+  __ load_parameter(0, c_rarg0);
+  __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier), c_rarg0);
+  __ restore_live_registers_except_rax(true);
+
+  __ epilogue();
+}
+
 #undef __
 
 #endif // COMPILER1
@@ -981,11 +921,11 @@
   StubCodeMark mark(cgen, "StubRoutines", "shenandoah_lrb");
   address start = __ pc();
 
-  Label resolve_oop, slow_path;
+  Label resolve_oop, slow_path, done;
 
   // We use RDI, which also serves as argument register for slow call.
-  // RAX always holds the src object ptr, except after the slow call and
-  // the cmpxchg, then it holds the result. R8/RBX is used as temporary register.
+  // RAX always holds the src object ptr, except after the slow call,
+  // then it holds the result. R8/RBX is used as temporary register.
 
   Register tmp1 = rdi;
   Register tmp2 = LP64_ONLY(r8) NOT_LP64(rbx);
@@ -994,26 +934,18 @@
   __ push(tmp2);
 
   // Check for object being in the collection set.
-  // TODO: Can we use only 1 register here?
-  // The source object arrives here in rax.
-  // live: rax
-  // live: tmp1
   __ mov(tmp1, rax);
   __ shrptr(tmp1, ShenandoahHeapRegion::region_size_bytes_shift_jint());
-  // live: tmp2
   __ movptr(tmp2, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr());
   __ movbool(tmp2, Address(tmp2, tmp1, Address::times_1));
-  // unlive: tmp1
   __ testbool(tmp2);
-  // unlive: tmp2
   __ jccb(Assembler::notZero, resolve_oop);
-
   __ pop(tmp2);
   __ pop(tmp1);
   __ ret(0);
 
+  // Test if object is already resolved.
   __ bind(resolve_oop);
-
   __ movptr(tmp2, Address(rax, oopDesc::mark_offset_in_bytes()));
   // Test if both lowest bits are set. We trick it by negating the bits
   // then test for both bits clear.
@@ -1026,6 +958,7 @@
   // At this point, tmp2 contains the decoded forwarding pointer.
   __ mov(rax, tmp2);
 
+  __ bind(done);
   __ pop(tmp2);
   __ pop(tmp1);
   __ ret(0);
@@ -1047,10 +980,7 @@
   __ push(r15);
 #endif
 
-  save_vector_registers(cgen->assembler());
-  __ movptr(rdi, rax);
-  __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier), rdi);
-  restore_vector_registers(cgen->assembler());
+  __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier), rax);
 
 #ifdef _LP64
   __ pop(r15);