1 /*
   2  * Copyright (c) 2018, Red Hat, Inc. All rights reserved.
   3  *
   4  * This code is free software; you can redistribute it and/or modify it
   5  * under the terms of the GNU General Public License version 2 only, as
   6  * published by the Free Software Foundation.
   7  *
   8  * This code is distributed in the hope that it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  11  * version 2 for more details (a copy is included in the LICENSE file that
  12  * accompanied this code).
  13  *
  14  * You should have received a copy of the GNU General Public License version
  15  * 2 along with this work; if not, write to the Free Software Foundation,
  16  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  17  *
  18  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  19  * or visit www.oracle.com if you need additional information or have any
  20  * questions.
  21  *
  22  */
  23 
  24 #include "precompiled.hpp"
  25 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
  26 #include "gc/shenandoah/shenandoahHeap.hpp"
  27 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
  28 #include "gc/shenandoah/shenandoahHeuristics.hpp"
  29 #include "gc/shenandoah/shenandoahRuntime.hpp"
  30 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
  31 #include "interpreter/interpreter.hpp"
  32 #include "interpreter/interp_masm.hpp"
  33 #include "runtime/sharedRuntime.hpp"
  34 #include "runtime/thread.hpp"
  35 #ifdef COMPILER1
  36 #include "c1/c1_LIRAssembler.hpp"
  37 #include "c1/c1_MacroAssembler.hpp"
  38 #include "gc/shenandoah/c1/shenandoahBarrierSetC1.hpp"
  39 #endif
  40 
  41 #define __ masm->
  42 
  43 address ShenandoahBarrierSetAssembler::_shenandoah_lrb = NULL;
  44 
  45 void ShenandoahBarrierSetAssembler::arraycopy_prologue(MacroAssembler* masm, DecoratorSet decorators, bool is_oop,
  46                                                        Register addr, Register count, RegSet saved_regs) {
  47   if (is_oop) {
  48     bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0;
  49     if (ShenandoahSATBBarrier && !dest_uninitialized && !ShenandoahHeap::heap()->heuristics()->can_do_traversal_gc()) {
  50 
  51       Label done;
  52 
  53       // Avoid calling runtime if count == 0
  54       __ cbz(count, done);
  55 
  56       // Is marking active?
  57       Address gc_state(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
  58       __ ldrb(rscratch1, gc_state);
  59       __ tbz(rscratch1, ShenandoahHeap::MARKING_BITPOS, done);
  60 
  61       __ push(saved_regs, sp);
  62       if (count == c_rarg0) {
  63         if (addr == c_rarg1) {
  64           // exactly backwards!!
  65           __ mov(rscratch1, c_rarg0);
  66           __ mov(c_rarg0, c_rarg1);
  67           __ mov(c_rarg1, rscratch1);
  68         } else {
  69           __ mov(c_rarg1, count);
  70           __ mov(c_rarg0, addr);
  71         }
  72       } else {
  73         __ mov(c_rarg0, addr);
  74         __ mov(c_rarg1, count);
  75       }
  76       if (UseCompressedOops) {
  77         __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_narrow_oop_entry), 2);
  78       } else {
  79         __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry), 2);
  80       }
  81       __ pop(saved_regs, sp);
  82       __ bind(done);
  83     }
  84   }
  85 }
  86 
  87 void ShenandoahBarrierSetAssembler::arraycopy_epilogue(MacroAssembler* masm, DecoratorSet decorators, bool is_oop,
  88                                                        Register start, Register count, Register scratch, RegSet saved_regs) {
  89   if (is_oop) {
  90     __ push(saved_regs, sp);
  91     assert_different_registers(start, count, scratch);
  92     assert_different_registers(c_rarg0, count);
  93     __ mov(c_rarg0, start);
  94     __ mov(c_rarg1, count);
  95     __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry), 2);
  96     __ pop(saved_regs, sp);
  97   }
  98 }
  99 
 100 void ShenandoahBarrierSetAssembler::shenandoah_write_barrier_pre(MacroAssembler* masm,
 101                                                                  Register obj,
 102                                                                  Register pre_val,
 103                                                                  Register thread,
 104                                                                  Register tmp,
 105                                                                  bool tosca_live,
 106                                                                  bool expand_call) {
 107   if (ShenandoahSATBBarrier) {
 108     satb_write_barrier_pre(masm, obj, pre_val, thread, tmp, tosca_live, expand_call);
 109   }
 110 }
 111 
 112 void ShenandoahBarrierSetAssembler::satb_write_barrier_pre(MacroAssembler* masm,
 113                                                            Register obj,
 114                                                            Register pre_val,
 115                                                            Register thread,
 116                                                            Register tmp,
 117                                                            bool tosca_live,
 118                                                            bool expand_call) {
 119   // If expand_call is true then we expand the call_VM_leaf macro
 120   // directly to skip generating the check by
 121   // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
 122 
 123   assert(thread == rthread, "must be");
 124 
 125   Label done;
 126   Label runtime;
 127 
 128   assert_different_registers(obj, pre_val, tmp, rscratch1);
 129   assert(pre_val != noreg &&  tmp != noreg, "expecting a register");
 130 
 131   Address in_progress(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()));
 132   Address index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
 133   Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
 134 
 135   // Is marking active?
 136   if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
 137     __ ldrw(tmp, in_progress);
 138   } else {
 139     assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");
 140     __ ldrb(tmp, in_progress);
 141   }
 142   __ cbzw(tmp, done);
 143 
 144   // Do we need to load the previous value?
 145   if (obj != noreg) {
 146     __ load_heap_oop(pre_val, Address(obj, 0), noreg, noreg, AS_RAW);
 147   }
 148 
 149   // Is the previous value null?
 150   __ cbz(pre_val, done);
 151 
 152   // Can we store original value in the thread's buffer?
 153   // Is index == 0?
 154   // (The index field is typed as size_t.)
 155 
 156   __ ldr(tmp, index);                      // tmp := *index_adr
 157   __ cbz(tmp, runtime);                    // tmp == 0?
 158                                         // If yes, goto runtime
 159 
 160   __ sub(tmp, tmp, wordSize);              // tmp := tmp - wordSize
 161   __ str(tmp, index);                      // *index_adr := tmp
 162   __ ldr(rscratch1, buffer);
 163   __ add(tmp, tmp, rscratch1);             // tmp := tmp + *buffer_adr
 164 
 165   // Record the previous value
 166   __ str(pre_val, Address(tmp, 0));
 167   __ b(done);
 168 
 169   __ bind(runtime);
 170   // save the live input values
 171   RegSet saved = RegSet::of(pre_val);
 172   if (tosca_live) saved += RegSet::of(r0);
 173   if (obj != noreg) saved += RegSet::of(obj);
 174 
 175   __ push(saved, sp);
 176 
 177   // Calling the runtime using the regular call_VM_leaf mechanism generates
 178   // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
 179   // that checks that the *(rfp+frame::interpreter_frame_last_sp) == NULL.
 180   //
 181   // If we care generating the pre-barrier without a frame (e.g. in the
 182   // intrinsified Reference.get() routine) then ebp might be pointing to
 183   // the caller frame and so this check will most likely fail at runtime.
 184   //
 185   // Expanding the call directly bypasses the generation of the check.
 186   // So when we do not have have a full interpreter frame on the stack
 187   // expand_call should be passed true.
 188 
 189   if (expand_call) {
 190     assert(pre_val != c_rarg1, "smashed arg");
 191     __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), pre_val, thread);
 192   } else {
 193     __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), pre_val, thread);
 194   }
 195 
 196   __ pop(saved, sp);
 197 
 198   __ bind(done);
 199 }
 200 
 201 void ShenandoahBarrierSetAssembler::resolve_forward_pointer(MacroAssembler* masm, Register dst) {
 202   assert(ShenandoahLoadRefBarrier || ShenandoahCASBarrier, "Should be enabled");
 203   Label is_null;
 204   __ cbz(dst, is_null);
 205   resolve_forward_pointer_not_null(masm, dst);
 206   __ bind(is_null);
 207 }
 208 
 209 void ShenandoahBarrierSetAssembler::resolve_forward_pointer_not_null(MacroAssembler* masm, Register dst) {
 210   assert(ShenandoahLoadRefBarrier || ShenandoahCASBarrier, "Should be enabled");
 211   __ ldr(dst, Address(dst, ShenandoahBrooksPointer::byte_offset()));
 212 }
 213 
 214 void ShenandoahBarrierSetAssembler::load_reference_barrier_not_null(MacroAssembler* masm, Register dst, Register tmp) {
 215   assert(ShenandoahLoadRefBarrier, "Should be enabled");
 216   assert(dst != rscratch2, "need rscratch2");
 217   if (tmp == noreg) {
 218     assert(dst != rscratch1, "need rscratch1");
 219     tmp = rscratch1;
 220   }
 221 
 222   Label done;
 223 
 224   Address gc_state(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 225   __ ldrb(tmp, gc_state);
 226 
 227   // Check for heap stability
 228   __ mov(rscratch2, ShenandoahHeap::HAS_FORWARDED | ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL);
 229   __ tst(tmp, rscratch2);
 230   __ br(Assembler::EQ, done);
 231 
 232   // Heap is unstable, need to perform the resolve even if LRB is inactive
 233   resolve_forward_pointer_not_null(masm, dst);
 234 
 235   // Check for evacuation-in-progress and jump to LRB slow-path if needed
 236   __ mov(rscratch2, ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL);
 237   __ tst(tmp, rscratch2);
 238   __ br(Assembler::EQ, done);
 239 
 240   RegSet to_save = RegSet::of(r0);
 241   if (dst != r0) {
 242     __ push(to_save, sp);
 243     __ mov(r0, dst);
 244   }
 245 
 246   __ far_call(RuntimeAddress(CAST_FROM_FN_PTR(address, ShenandoahBarrierSetAssembler::shenandoah_lrb())));
 247 
 248   if (dst != r0) {
 249     __ mov(dst, r0);
 250     __ pop(to_save, sp);
 251   }
 252 
 253   __ bind(done);
 254 }
 255 
 256 void ShenandoahBarrierSetAssembler::storeval_barrier(MacroAssembler* masm, Register dst, Register tmp) {
 257   if (ShenandoahStoreValEnqueueBarrier) {
 258     // Save possibly live regs.
 259     RegSet live_regs = RegSet::range(r0, r4) - dst;
 260     __ push(live_regs, sp);
 261     __ strd(v0, __ pre(sp, 2 * -wordSize));
 262 
 263     satb_write_barrier_pre(masm, noreg, dst, rthread, tmp, true, false);
 264 
 265     // Restore possibly live regs.
 266     __ ldrd(v0, __ post(sp, 2 * wordSize));
 267     __ pop(live_regs, sp);
 268   }
 269 }
 270 
 271 void ShenandoahBarrierSetAssembler::load_reference_barrier(MacroAssembler* masm, Register dst, Register tmp) {
 272   if (ShenandoahLoadRefBarrier) {
 273     Label is_null;
 274     __ cbz(dst, is_null);
 275     load_reference_barrier_not_null(masm, dst, tmp);
 276     __ bind(is_null);
 277   }
 278 }
 279 
 280 void ShenandoahBarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 281                                             Register dst, Address src, Register tmp1, Register tmp_thread) {
 282   bool on_oop = type == T_OBJECT || type == T_ARRAY;
 283   bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
 284   bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
 285   bool on_reference = on_weak || on_phantom;
 286 
 287   BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp_thread);
 288   if (on_oop) {
 289     load_reference_barrier(masm, dst, tmp1);
 290 
 291     if (ShenandoahKeepAliveBarrier && on_reference) {
 292       __ enter();
 293       satb_write_barrier_pre(masm /* masm */,
 294                              noreg /* obj */,
 295                              dst /* pre_val */,
 296                              rthread /* thread */,
 297                              tmp1 /* tmp */,
 298                              true /* tosca_live */,
 299                              true /* expand_call */);
 300       __ leave();
 301     }
 302   }
 303 }
 304 
 305 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 306                                              Address dst, Register val, Register tmp1, Register tmp2) {
 307   bool on_oop = type == T_OBJECT || type == T_ARRAY;
 308   if (!on_oop) {
 309     BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2);
 310     return;
 311   }
 312 
 313   // flatten object address if needed
 314   if (dst.index() == noreg && dst.offset() == 0) {
 315     if (dst.base() != r3) {
 316       __ mov(r3, dst.base());
 317     }
 318   } else {
 319     __ lea(r3, dst);
 320   }
 321 
 322   shenandoah_write_barrier_pre(masm,
 323                                r3 /* obj */,
 324                                tmp2 /* pre_val */,
 325                                rthread /* thread */,
 326                                tmp1  /* tmp */,
 327                                val != noreg /* tosca_live */,
 328                                false /* expand_call */);
 329 
 330   if (val == noreg) {
 331     BarrierSetAssembler::store_at(masm, decorators, type, Address(r3, 0), noreg, noreg, noreg);
 332   } else {
 333     storeval_barrier(masm, val, tmp1);
 334     // G1 barrier needs uncompressed oop for region cross check.
 335     Register new_val = val;
 336     if (UseCompressedOops) {
 337       new_val = rscratch2;
 338       __ mov(new_val, val);
 339     }
 340     BarrierSetAssembler::store_at(masm, decorators, type, Address(r3, 0), val, noreg, noreg);
 341   }
 342 
 343 }
 344 
 345 void ShenandoahBarrierSetAssembler::tlab_allocate(MacroAssembler* masm, Register obj,
 346                                                   Register var_size_in_bytes,
 347                                                   int con_size_in_bytes,
 348                                                   Register t1,
 349                                                   Register t2,
 350                                                   Label& slow_case) {
 351 
 352   assert_different_registers(obj, t2);
 353   assert_different_registers(obj, var_size_in_bytes);
 354   Register end = t2;
 355 
 356   __ ldr(obj, Address(rthread, JavaThread::tlab_top_offset()));
 357   if (var_size_in_bytes == noreg) {
 358     __ lea(end, Address(obj, (int) (con_size_in_bytes + ShenandoahBrooksPointer::byte_size())));
 359   } else {
 360     __ add(var_size_in_bytes, var_size_in_bytes, ShenandoahBrooksPointer::byte_size());
 361     __ lea(end, Address(obj, var_size_in_bytes));
 362   }
 363   __ ldr(rscratch1, Address(rthread, JavaThread::tlab_end_offset()));
 364   __ cmp(end, rscratch1);
 365   __ br(Assembler::HI, slow_case);
 366 
 367   // update the tlab top pointer
 368   __ str(end, Address(rthread, JavaThread::tlab_top_offset()));
 369 
 370   __ add(obj, obj, ShenandoahBrooksPointer::byte_size());
 371   __ str(obj, Address(obj, ShenandoahBrooksPointer::byte_offset()));
 372 
 373   // recover var_size_in_bytes if necessary
 374   if (var_size_in_bytes == end) {
 375     __ sub(var_size_in_bytes, var_size_in_bytes, obj);
 376   }
 377 }
 378 
 379 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm, Register addr, Register expected, Register new_val,
 380                                                 bool acquire, bool release, bool weak, bool is_cae,
 381                                                 Register result) {
 382   Register tmp1 = rscratch1;
 383   Register tmp2 = rscratch2;
 384   bool is_narrow = UseCompressedOops;
 385   Assembler::operand_size size = is_narrow ? Assembler::word : Assembler::xword;
 386 
 387   assert_different_registers(addr, expected, new_val, tmp1, tmp2);
 388 
 389   Label retry, done, fail;
 390 
 391   // CAS, using LL/SC pair.
 392   __ bind(retry);
 393   __ load_exclusive(tmp1, addr, size, acquire);
 394   if (is_narrow) {
 395     __ cmpw(tmp1, expected);
 396   } else {
 397     __ cmp(tmp1, expected);
 398   }
 399   __ br(Assembler::NE, fail);
 400   __ store_exclusive(tmp2, new_val, addr, size, release);
 401   if (weak) {
 402     __ cmpw(tmp2, 0u); // If the store fails, return NE to our caller
 403   } else {
 404     __ cbnzw(tmp2, retry);
 405   }
 406   __ b(done);
 407 
 408  __  bind(fail);
 409   // Check if rb(expected)==rb(tmp1)
 410   // Shuffle registers so that we have memory value ready for next expected.
 411   __ mov(tmp2, expected);
 412   __ mov(expected, tmp1);
 413   if (is_narrow) {
 414     __ decode_heap_oop(tmp1, tmp1);
 415     __ decode_heap_oop(tmp2, tmp2);
 416   }
 417   resolve_forward_pointer(masm, tmp1);
 418   resolve_forward_pointer(masm, tmp2);
 419   __ cmp(tmp1, tmp2);
 420   // Retry with expected now being the value we just loaded from addr.
 421   __ br(Assembler::EQ, retry);
 422   if (is_cae && is_narrow) {
 423     // For cmp-and-exchange and narrow oops, we need to restore
 424     // the compressed old-value. We moved it to 'expected' a few lines up.
 425     __ mov(tmp1, expected);
 426   }
 427   __ bind(done);
 428 
 429   if (is_cae) {
 430     __ mov(result, tmp1);
 431   } else {
 432     __ cset(result, Assembler::EQ);
 433   }
 434 }
 435 
 436 #ifdef COMPILER1
 437 
 438 #undef __
 439 #define __ ce->masm()->
 440 
 441 void ShenandoahBarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, ShenandoahPreBarrierStub* stub) {
 442   ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
 443   // At this point we know that marking is in progress.
 444   // If do_load() is true then we have to emit the
 445   // load of the previous value; otherwise it has already
 446   // been loaded into _pre_val.
 447 
 448   __ bind(*stub->entry());
 449 
 450   assert(stub->pre_val()->is_register(), "Precondition.");
 451 
 452   Register pre_val_reg = stub->pre_val()->as_register();
 453 
 454   if (stub->do_load()) {
 455     ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/, false /*unaligned*/);
 456   }
 457   __ cbz(pre_val_reg, *stub->continuation());
 458   ce->store_parameter(stub->pre_val()->as_register(), 0);
 459   __ far_call(RuntimeAddress(bs->pre_barrier_c1_runtime_code_blob()->code_begin()));
 460   __ b(*stub->continuation());
 461 }
 462 
 463 void ShenandoahBarrierSetAssembler::gen_load_reference_barrier_stub(LIR_Assembler* ce, ShenandoahLoadReferenceBarrierStub* stub) {
 464 
 465   Register obj = stub->obj()->as_register();
 466   Register res = stub->result()->as_register();
 467 
 468   Label done;
 469 
 470   __ bind(*stub->entry());
 471 
 472   if (res != obj) {
 473     __ mov(res, obj);
 474   }
 475   // Check for null.
 476   if (stub->needs_null_check()) {
 477     __ cbz(res, done);
 478   }
 479 
 480   load_reference_barrier_not_null(ce->masm(), res, rscratch1);
 481 
 482   __ bind(done);
 483   __ b(*stub->continuation());
 484 }
 485 
 486 #undef __
 487 
 488 #define __ sasm->
 489 
 490 void ShenandoahBarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) {
 491   __ prologue("shenandoah_pre_barrier", false);
 492 
 493   // arg0 : previous value of memory
 494 
 495   BarrierSet* bs = BarrierSet::barrier_set();
 496 
 497   const Register pre_val = r0;
 498   const Register thread = rthread;
 499   const Register tmp = rscratch1;
 500 
 501   Address queue_index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
 502   Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
 503 
 504   Label done;
 505   Label runtime;
 506 
 507   // Is marking still active?
 508   Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 509   __ ldrb(tmp, gc_state);
 510   __ mov(rscratch2, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL);
 511   __ tst(tmp, rscratch2);
 512   __ br(Assembler::EQ, done);
 513 
 514   // Can we store original value in the thread's buffer?
 515   __ ldr(tmp, queue_index);
 516   __ cbz(tmp, runtime);
 517 
 518   __ sub(tmp, tmp, wordSize);
 519   __ str(tmp, queue_index);
 520   __ ldr(rscratch2, buffer);
 521   __ add(tmp, tmp, rscratch2);
 522   __ load_parameter(0, rscratch2);
 523   __ str(rscratch2, Address(tmp, 0));
 524   __ b(done);
 525 
 526   __ bind(runtime);
 527   __ push_call_clobbered_registers();
 528   __ load_parameter(0, pre_val);
 529   __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), pre_val, thread);
 530   __ pop_call_clobbered_registers();
 531   __ bind(done);
 532 
 533   __ epilogue();
 534 }
 535 
 536 #undef __
 537 
 538 #endif // COMPILER1
 539 
 540 address ShenandoahBarrierSetAssembler::shenandoah_lrb() {
 541   assert(_shenandoah_lrb != NULL, "need load reference barrier stub");
 542   return _shenandoah_lrb;
 543 }
 544 
 545 #define __ cgen->assembler()->
 546 
 547 // Shenandoah load reference barrier.
 548 //
 549 // Input:
 550 //   r0: OOP to evacuate.  Not null.
 551 //
 552 // Output:
 553 //   r0: Pointer to evacuated OOP.
 554 //
 555 // Trash rscratch1, rscratch2.  Preserve everything else.
 556 address ShenandoahBarrierSetAssembler::generate_shenandoah_lrb(StubCodeGenerator* cgen) {
 557 
 558   __ align(6);
 559   StubCodeMark mark(cgen, "StubRoutines", "shenandoah_lrb");
 560   address start = __ pc();
 561 
 562   Label work;
 563   __ mov(rscratch2, ShenandoahHeap::in_cset_fast_test_addr());
 564   __ lsr(rscratch1, r0, ShenandoahHeapRegion::region_size_bytes_shift_jint());
 565   __ ldrb(rscratch2, Address(rscratch2, rscratch1));
 566   __ tbnz(rscratch2, 0, work);
 567   __ ret(lr);
 568   __ bind(work);
 569 
 570   Register obj = r0;
 571 
 572   __ enter(); // required for proper stackwalking of RuntimeStub frame
 573 
 574   __ push_call_clobbered_registers();
 575 
 576   __ mov(lr, CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_JRT));
 577   __ blrt(lr, 1, 0, MacroAssembler::ret_type_integral);
 578   __ mov(rscratch1, obj);
 579   __ pop_call_clobbered_registers();
 580   __ mov(obj, rscratch1);
 581 
 582   __ leave(); // required for proper stackwalking of RuntimeStub frame
 583   __ ret(lr);
 584 
 585   return start;
 586 }
 587 
 588 #undef __
 589 
 590 void ShenandoahBarrierSetAssembler::barrier_stubs_init() {
 591   if (ShenandoahLoadRefBarrier) {
 592     int stub_code_size = 2048;
 593     ResourceMark rm;
 594     BufferBlob* bb = BufferBlob::create("shenandoah_barrier_stubs", stub_code_size);
 595     CodeBuffer buf(bb);
 596     StubCodeGenerator cgen(&buf);
 597     _shenandoah_lrb = generate_shenandoah_lrb(&cgen);
 598   }
 599 }