Merge

src/cpu/sparc/vm/assembler_sparc.cpp

@@ -4104,7 +4104,7 @@ void MacroAssembler::tlab_refill(Label& retry, Label& try_eden, Label& slow_case
 
   ld_ptr(G2_thread, in_bytes(JavaThread::tlab_start_offset()), t1);
   sub(top, t1, t1); // size of tlab's allocated portion
-  incr_allocated_bytes(t1, 0, t2);
+  incr_allocated_bytes(t1, t2, t3);
 
   // refill the tlab with an eden allocation
   bind(do_refill);
@@ -4138,19 +4138,14 @@ void MacroAssembler::tlab_refill(Label& retry, Label& try_eden, Label& slow_case
   delayed()->nop();
 }
 
-void MacroAssembler::incr_allocated_bytes(Register var_size_in_bytes,
-                                          int con_size_in_bytes,
-                                          Register t1) {
+void MacroAssembler::incr_allocated_bytes(RegisterOrConstant size_in_bytes,
+                                          Register t1, Register t2) {
   // Bump total bytes allocated by this thread
   assert(t1->is_global(), "must be global reg"); // so all 64 bits are saved on a context switch
-  assert_different_registers(var_size_in_bytes, t1);
+  assert_different_registers(size_in_bytes.register_or_noreg(), t1, t2);
   // v8 support has gone the way of the dodo
   ldx(G2_thread, in_bytes(JavaThread::allocated_bytes_offset()), t1);
-  if (var_size_in_bytes->is_valid()) {
-    add(t1, var_size_in_bytes, t1);
-  } else {
-    add(t1, con_size_in_bytes, t1);
-  }
+  add(t1, ensure_simm13_or_reg(size_in_bytes, t2), t1);
   stx(t1, G2_thread, in_bytes(JavaThread::allocated_bytes_offset()));
 }
 

src/cpu/sparc/vm/assembler_sparc.hpp

@@ -2389,7 +2389,8 @@ public:
     Label&   slow_case                 // continuation point if fast allocation fails
   );
   void tlab_refill(Label& retry_tlab, Label& try_eden, Label& slow_case);
-  void incr_allocated_bytes(Register var_size_in_bytes, int con_size_in_bytes, Register t1);
+  void incr_allocated_bytes(RegisterOrConstant size_in_bytes,
+                            Register t1, Register t2);
 
   // interface method calling
   void lookup_interface_method(Register recv_klass,

src/cpu/sparc/vm/c1_MacroAssembler_sparc.cpp

@@ -170,11 +170,13 @@ void C1_MacroAssembler::try_allocate(
   Register t2,                         // temp register
   Label&   slow_case                   // continuation point if fast allocation fails
 ) {
+  RegisterOrConstant size_in_bytes = var_size_in_bytes->is_valid()
+    ? RegisterOrConstant(var_size_in_bytes) : RegisterOrConstant(con_size_in_bytes);
   if (UseTLAB) {
     tlab_allocate(obj, var_size_in_bytes, con_size_in_bytes, t1, slow_case);
   } else {
     eden_allocate(obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case);
-    incr_allocated_bytes(var_size_in_bytes, con_size_in_bytes, t1);
+    incr_allocated_bytes(size_in_bytes, t1, t2);
   }
 }
 

src/cpu/sparc/vm/c1_Runtime1_sparc.cpp

@@ -461,7 +461,7 @@ OopMapSet* Runtime1::generate_code_for(StubID id, StubAssembler* sasm) {
           // get the instance size
           __ ld(G5_klass, klassOopDesc::header_size() * HeapWordSize + Klass::layout_helper_offset_in_bytes(), G1_obj_size);
           __ eden_allocate(O0_obj, G1_obj_size, 0, G3_t1, G4_t2, slow_path);
-          __ incr_allocated_bytes(G1_obj_size, 0, G3_t1);
+          __ incr_allocated_bytes(G1_obj_size, G3_t1, G4_t2);
 
           __ initialize_object(O0_obj, G5_klass, G1_obj_size, 0, G3_t1, G4_t2);
           __ verify_oop(O0_obj);
@@ -577,7 +577,7 @@ OopMapSet* Runtime1::generate_code_for(StubID id, StubAssembler* sasm) {
           __ and3(G1_arr_size, ~MinObjAlignmentInBytesMask, G1_arr_size);
 
           __ eden_allocate(O0_obj, G1_arr_size, 0, G3_t1, O1_t2, slow_path);  // preserves G1_arr_size
-          __ incr_allocated_bytes(G1_arr_size, 0, G3_t1);
+          __ incr_allocated_bytes(G1_arr_size, G3_t1, O1_t2);
 
           __ initialize_header(O0_obj, G5_klass, G4_length, G3_t1, O1_t2);
           __ ldub(klass_lh, G3_t1, klass_lh_header_size_offset);

src/cpu/sparc/vm/templateTable_sparc.cpp

@@ -3447,7 +3447,8 @@ void TemplateTable::_new() {
     __ delayed()->nop();
 
     // bump total bytes allocated by this thread
-    __ incr_allocated_bytes(Roffset, 0, G1_scratch);
+    // RoldTopValue and RtopAddr are dead, so can use G1 and G3
+    __ incr_allocated_bytes(Roffset, G1_scratch, G3_scratch);
   }
 
   if (UseTLAB || Universe::heap()->supports_inline_contig_alloc()) {