1 /*
   2  * Copyright (c) 2026, Oracle and/or its affiliates. All rights reserved.
   3  * Copyright (c) 2018, 2022, Red Hat, Inc. All rights reserved.
   4  * Copyright Amazon.com Inc. or its affiliates. All Rights Reserved.
   5  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   6  *
   7  * This code is free software; you can redistribute it and/or modify it
   8  * under the terms of the GNU General Public License version 2 only, as
   9  * published by the Free Software Foundation.
  10  *
  11  * This code is distributed in the hope that it will be useful, but WITHOUT
  12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  14  * version 2 for more details (a copy is included in the LICENSE file that
  15  * accompanied this code).
  16  *
  17  * You should have received a copy of the GNU General Public License version
  18  * 2 along with this work; if not, write to the Free Software Foundation,
  19  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  20  *
  21  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  22  * or visit www.oracle.com if you need additional information or have any
  23  * questions.
  24  *
  25  */
  26 
  27 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
  28 #include "gc/shenandoah/mode/shenandoahMode.hpp"
  29 #include "gc/shenandoah/shenandoahBarrierSet.hpp"
  30 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
  31 #include "gc/shenandoah/shenandoahForwarding.hpp"
  32 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
  33 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
  34 #include "gc/shenandoah/shenandoahRuntime.hpp"
  35 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
  36 #include "interpreter/interp_masm.hpp"
  37 #include "interpreter/interpreter.hpp"
  38 #include "runtime/javaThread.hpp"
  39 #include "runtime/sharedRuntime.hpp"
  40 #ifdef COMPILER1
  41 #include "c1/c1_LIRAssembler.hpp"
  42 #include "c1/c1_MacroAssembler.hpp"
  43 #include "gc/shenandoah/c1/shenandoahBarrierSetC1.hpp"
  44 #endif
  45 #ifdef COMPILER2
  46 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
  47 #include "opto/output.hpp"
  48 #endif
  49 
  50 #define __ masm->
  51 
  52 void ShenandoahBarrierSetAssembler::arraycopy_prologue(MacroAssembler* masm, DecoratorSet decorators, bool is_oop,
  53                                                        Register src, Register dst, Register count, RegSet saved_regs) {
  54   if (is_oop) {
  55     bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0;
  56     if ((ShenandoahSATBBarrier && !dest_uninitialized) || ShenandoahLoadRefBarrier) {
  57 
  58       Label done;
  59 
  60       // Avoid calling runtime if count == 0
  61       __ cbz(count, done);
  62 
  63       // Is GC active?
  64       Address gc_state(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
  65       __ ldrb(rscratch1, gc_state);
  66       if (ShenandoahSATBBarrier && dest_uninitialized) {
  67         __ tbz(rscratch1, ShenandoahHeap::HAS_FORWARDED_BITPOS, done);
  68       } else {
  69         __ mov(rscratch2, ShenandoahHeap::HAS_FORWARDED | ShenandoahHeap::MARKING);
  70         __ tst(rscratch1, rscratch2);
  71         __ br(Assembler::EQ, done);
  72       }
  73 
  74       __ push(saved_regs, sp);
  75       if (UseCompressedOops) {
  76         __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::arraycopy_barrier_narrow_oop), src, dst, count);
  77       } else {
  78         __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::arraycopy_barrier_oop), src, dst, count);
  79       }
  80       __ pop(saved_regs, sp);
  81       __ bind(done);
  82     }
  83   }
  84 }
  85 
  86 void ShenandoahBarrierSetAssembler::arraycopy_epilogue(MacroAssembler* masm, DecoratorSet decorators, bool is_oop,
  87                                                        Register start, Register count, Register tmp) {
  88   if (ShenandoahCardBarrier && is_oop) {
  89     gen_write_ref_array_post_barrier(masm, decorators, start, count, tmp);
  90   }
  91 }
  92 
  93 void ShenandoahBarrierSetAssembler::satb_barrier(MacroAssembler* masm,
  94                                                  Register obj,
  95                                                  Register pre_val,
  96                                                  Register thread,
  97                                                  Register tmp1,
  98                                                  Register tmp2,
  99                                                  bool tosca_live,
 100                                                  bool expand_call) {
 101   assert(ShenandoahSATBBarrier, "Should be checked by caller");
 102 
 103   // If expand_call is true then we expand the call_VM_leaf macro
 104   // directly to skip generating the check by
 105   // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
 106 
 107   assert(thread == rthread, "must be");
 108 
 109   Label done;
 110   Label runtime;
 111 
 112   assert_different_registers(obj, pre_val, tmp1, tmp2);
 113   assert(pre_val != noreg && tmp1 != noreg && tmp2 != noreg, "expecting a register");
 114 
 115   Address index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
 116   Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
 117 
 118   // Is marking active?
 119   Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 120   __ ldrb(tmp1, gc_state);
 121   __ tbz(tmp1, ShenandoahHeap::MARKING_BITPOS, done);
 122 
 123   // Do we need to load the previous value?
 124   if (obj != noreg) {
 125     __ load_heap_oop(pre_val, Address(obj, 0), noreg, noreg, AS_RAW);
 126   }
 127 
 128   // Is the previous value null?
 129   __ cbz(pre_val, done);
 130 
 131   // Can we store original value in the thread's buffer?
 132   // Is index == 0?
 133   // (The index field is typed as size_t.)
 134 
 135   __ ldr(tmp1, index);                      // tmp := *index_adr
 136   __ cbz(tmp1, runtime);                    // tmp == 0?
 137                                         // If yes, goto runtime
 138 
 139   __ sub(tmp1, tmp1, wordSize);             // tmp := tmp - wordSize
 140   __ str(tmp1, index);                      // *index_adr := tmp
 141   __ ldr(tmp2, buffer);
 142   __ add(tmp1, tmp1, tmp2);                 // tmp := tmp + *buffer_adr
 143 
 144   // Record the previous value
 145   __ str(pre_val, Address(tmp1, 0));
 146   __ b(done);
 147 
 148   __ bind(runtime);
 149   // save the live input values
 150   RegSet saved = RegSet::of(pre_val);
 151   if (tosca_live) saved += RegSet::of(r0);
 152   if (obj != noreg) saved += RegSet::of(obj);
 153 
 154   __ push(saved, sp);
 155 
 156   // Calling the runtime using the regular call_VM_leaf mechanism generates
 157   // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
 158   // that checks that the *(rfp+frame::interpreter_frame_last_sp) == nullptr.
 159   //
 160   // If we care generating the pre-barrier without a frame (e.g. in the
 161   // intrinsified Reference.get() routine) then rfp might be pointing to
 162   // the caller frame and so this check will most likely fail at runtime.
 163   //
 164   // Expanding the call directly bypasses the generation of the check.
 165   // So when we do not have have a full interpreter frame on the stack
 166   // expand_call should be passed true.
 167 
 168   if (expand_call) {
 169     assert(pre_val != c_rarg1, "smashed arg");
 170     __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_pre), pre_val);
 171   } else {
 172     __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_pre), pre_val);
 173   }
 174 
 175   __ pop(saved, sp);
 176 
 177   __ bind(done);
 178 }
 179 
 180 void ShenandoahBarrierSetAssembler::resolve_forward_pointer(MacroAssembler* masm, Register dst, Register tmp) {
 181   assert(ShenandoahLoadRefBarrier || ShenandoahCASBarrier, "Should be enabled");
 182   Label is_null;
 183   __ cbz(dst, is_null);
 184   resolve_forward_pointer_not_null(masm, dst, tmp);
 185   __ bind(is_null);
 186 }
 187 
 188 // IMPORTANT: This must preserve all registers, even rscratch1 and rscratch2, except those explicitly
 189 // passed in.
 190 void ShenandoahBarrierSetAssembler::resolve_forward_pointer_not_null(MacroAssembler* masm, Register dst, Register tmp) {
 191   assert(ShenandoahLoadRefBarrier || ShenandoahCASBarrier, "Should be enabled");
 192   // The below loads the mark word, checks if the lowest two bits are
 193   // set, and if so, clear the lowest two bits and copy the result
 194   // to dst. Otherwise it leaves dst alone.
 195   // Implementing this is surprisingly awkward. I do it here by:
 196   // - Inverting the mark word
 197   // - Test lowest two bits == 0
 198   // - If so, set the lowest two bits
 199   // - Invert the result back, and copy to dst
 200 
 201   bool borrow_reg = (tmp == noreg);
 202   if (borrow_reg) {
 203     // No free registers available. Make one useful.
 204     tmp = rscratch1;
 205     if (tmp == dst) {
 206       tmp = rscratch2;
 207     }
 208     __ push(RegSet::of(tmp), sp);
 209   }
 210 
 211   assert_different_registers(tmp, dst);
 212 
 213   Label done;
 214   __ ldr(tmp, Address(dst, oopDesc::mark_offset_in_bytes()));
 215   __ eon(tmp, tmp, zr);
 216   __ ands(zr, tmp, markWord::lock_mask_in_place);
 217   __ br(Assembler::NE, done);
 218   __ orr(tmp, tmp, markWord::marked_value);
 219   __ eon(dst, tmp, zr);
 220   __ bind(done);
 221 
 222   if (borrow_reg) {
 223     __ pop(RegSet::of(tmp), sp);
 224   }
 225 }
 226 
 227 void ShenandoahBarrierSetAssembler::load_reference_barrier(MacroAssembler* masm, Register dst, Address load_addr, DecoratorSet decorators) {
 228   assert(ShenandoahLoadRefBarrier, "Should be enabled");
 229   assert(dst != rscratch2, "need rscratch2");
 230   assert_different_registers(load_addr.base(), load_addr.index(), rscratch1, rscratch2);
 231 
 232   bool is_strong  = ShenandoahBarrierSet::is_strong_access(decorators);
 233   bool is_weak    = ShenandoahBarrierSet::is_weak_access(decorators);
 234   bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
 235   bool is_native  = ShenandoahBarrierSet::is_native_access(decorators);
 236   bool is_narrow  = UseCompressedOops && !is_native;
 237 
 238   Label heap_stable, not_cset;
 239   __ enter(/*strip_ret_addr*/true);
 240   Address gc_state(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 241   __ ldrb(rscratch2, gc_state);
 242 
 243   // Check for heap stability
 244   if (is_strong) {
 245     __ tbz(rscratch2, ShenandoahHeap::HAS_FORWARDED_BITPOS, heap_stable);
 246   } else {
 247     Label lrb;
 248     __ tbnz(rscratch2, ShenandoahHeap::WEAK_ROOTS_BITPOS, lrb);
 249     __ tbz(rscratch2, ShenandoahHeap::HAS_FORWARDED_BITPOS, heap_stable);
 250     __ bind(lrb);
 251   }
 252 
 253   // use r1 for load address
 254   Register result_dst = dst;
 255   if (dst == r1) {
 256     __ mov(rscratch1, dst);
 257     dst = rscratch1;
 258   }
 259 
 260   // Save r0 and r1, unless it is an output register
 261   RegSet to_save = RegSet::of(r0, r1) - result_dst;
 262   __ push(to_save, sp);
 263   __ lea(r1, load_addr);
 264   __ mov(r0, dst);
 265 
 266   // Test for in-cset
 267   if (is_strong) {
 268     __ mov(rscratch2, ShenandoahHeap::in_cset_fast_test_addr());
 269     __ lsr(rscratch1, r0, ShenandoahHeapRegion::region_size_bytes_shift_jint());
 270     __ ldrb(rscratch2, Address(rscratch2, rscratch1));
 271     __ tbz(rscratch2, 0, not_cset);
 272   }
 273 
 274   __ push_call_clobbered_registers();
 275   if (is_strong) {
 276     if (is_narrow) {
 277       __ mov(lr, CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow));
 278     } else {
 279       __ mov(lr, CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong));
 280     }
 281   } else if (is_weak) {
 282     if (is_narrow) {
 283       __ mov(lr, CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow));
 284     } else {
 285       __ mov(lr, CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak));
 286     }
 287   } else {
 288     assert(is_phantom, "only remaining strength");
 289     assert(!is_narrow, "phantom access cannot be narrow");
 290     // AOT saved adapters need relocation for this call.
 291     __ lea(lr, RuntimeAddress(CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom)));
 292   }
 293   __ blr(lr);
 294   __ mov(rscratch1, r0);
 295   __ pop_call_clobbered_registers();
 296   __ mov(r0, rscratch1);
 297 
 298   __ bind(not_cset);
 299 
 300   __ mov(result_dst, r0);
 301   __ pop(to_save, sp);
 302 
 303   __ bind(heap_stable);
 304   __ leave();
 305 }
 306 
 307 //
 308 // Arguments:
 309 //
 310 // Inputs:
 311 //   src:        oop location to load from, might be clobbered
 312 //
 313 // Output:
 314 //   dst:        oop loaded from src location
 315 //
 316 // Kill:
 317 //   rscratch1 (scratch reg)
 318 //
 319 // Alias:
 320 //   dst: rscratch1 (might use rscratch1 as temporary output register to avoid clobbering src)
 321 //
 322 void ShenandoahBarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 323                                             Register dst, Address src, Register tmp1, Register tmp2) {
 324   // 1: non-reference load, no additional barrier is needed
 325   if (!is_reference_type(type)) {
 326     BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp2);
 327     return;
 328   }
 329 
 330   // 2: load a reference from src location and apply LRB if needed
 331   if (ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) {
 332     Register result_dst = dst;
 333 
 334     // Preserve src location for LRB
 335     if (dst == src.base() || dst == src.index()) {
 336       dst = rscratch1;
 337     }
 338     assert_different_registers(dst, src.base(), src.index());
 339 
 340     BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp2);
 341 
 342     load_reference_barrier(masm, dst, src, decorators);
 343 
 344     if (dst != result_dst) {
 345       __ mov(result_dst, dst);
 346       dst = result_dst;
 347     }
 348   } else {
 349     BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp2);
 350   }
 351 
 352   // 3: apply keep-alive barrier if needed
 353   if (ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) {
 354     __ enter(/*strip_ret_addr*/true);
 355     __ push_call_clobbered_registers();
 356     satb_barrier(masm /* masm */,
 357                  noreg /* obj */,
 358                  dst /* pre_val */,
 359                  rthread /* thread */,
 360                  tmp1 /* tmp1 */,
 361                  tmp2 /* tmp2 */,
 362                  true /* tosca_live */,
 363                  true /* expand_call */);
 364     __ pop_call_clobbered_registers();
 365     __ leave();
 366   }
 367 }
 368 
 369 void ShenandoahBarrierSetAssembler::card_barrier(MacroAssembler* masm, Register obj) {
 370   assert(ShenandoahCardBarrier, "Should have been checked by caller");
 371 
 372   __ lsr(obj, obj, CardTable::card_shift());
 373 
 374   assert(CardTable::dirty_card_val() == 0, "must be");
 375 
 376   Address curr_ct_holder_addr(rthread, in_bytes(ShenandoahThreadLocalData::card_table_offset()));
 377   __ ldr(rscratch1, curr_ct_holder_addr);
 378 
 379   if (UseCondCardMark) {
 380     Label L_already_dirty;
 381     __ ldrb(rscratch2, Address(obj, rscratch1));
 382     __ cbz(rscratch2, L_already_dirty);
 383     __ strb(zr, Address(obj, rscratch1));
 384     __ bind(L_already_dirty);
 385   } else {
 386     __ strb(zr, Address(obj, rscratch1));
 387   }
 388 }
 389 
 390 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 391                                              Address dst, Register val, Register tmp1, Register tmp2, Register tmp3) {
 392   // 1: non-reference types require no barriers
 393   if (!is_reference_type(type)) {
 394     BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2, tmp3);
 395     return;
 396   }
 397 
 398   // Flatten object address right away for simplicity: likely needed by barriers
 399   if (dst.index() == noreg && dst.offset() == 0) {
 400     if (dst.base() != tmp3) {
 401       __ mov(tmp3, dst.base());
 402     }
 403   } else {
 404     __ lea(tmp3, dst);
 405   }
 406 
 407   bool storing_non_null = (val != noreg);
 408 
 409   // 2: pre-barrier: SATB needs the previous value
 410   if (ShenandoahBarrierSet::need_satb_barrier(decorators, type)) {
 411     satb_barrier(masm,
 412                  tmp3 /* obj */,
 413                  tmp2 /* pre_val */,
 414                  rthread /* thread */,
 415                  tmp1 /* tmp */,
 416                  rscratch1 /* tmp2 */,
 417                  storing_non_null /* tosca_live */,
 418                  false /* expand_call */);
 419   }
 420 
 421   // Store!
 422   BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp3, 0), val, noreg, noreg, noreg);
 423 
 424   // 3: post-barrier: card barrier needs store address
 425   if (ShenandoahBarrierSet::need_card_barrier(decorators, type) && storing_non_null) {
 426     card_barrier(masm, tmp3);
 427   }
 428 }
 429 
 430 void ShenandoahBarrierSetAssembler::try_resolve_jobject_in_native(MacroAssembler* masm, Register jni_env,
 431                                                                   Register obj, Register tmp, Label& slowpath) {
 432   Label done;
 433   // Resolve jobject
 434   BarrierSetAssembler::try_resolve_jobject_in_native(masm, jni_env, obj, tmp, slowpath);
 435 
 436   // Check for null.
 437   __ cbz(obj, done);
 438 
 439   assert(obj != rscratch2, "need rscratch2");
 440   Address gc_state(jni_env, ShenandoahThreadLocalData::gc_state_offset() - JavaThread::jni_environment_offset());
 441   __ lea(rscratch2, gc_state);
 442   __ ldrb(rscratch2, Address(rscratch2));
 443 
 444   // Check for heap in evacuation phase
 445   __ tbnz(rscratch2, ShenandoahHeap::EVACUATION_BITPOS, slowpath);
 446 
 447   __ bind(done);
 448 }
 449 
 450 #ifdef COMPILER2
 451 void ShenandoahBarrierSetAssembler::try_resolve_weak_handle_in_c2(MacroAssembler* masm, Register obj,
 452                                                                   Register tmp, Label& slow_path) {
 453   assert_different_registers(obj, tmp);
 454 
 455   Label done;
 456 
 457   // Resolve weak handle using the standard implementation.
 458   BarrierSetAssembler::try_resolve_weak_handle_in_c2(masm, obj, tmp, slow_path);
 459 
 460   // Check if the reference is null, and if it is, take the fast path.
 461   __ cbz(obj, done);
 462 
 463   Address gc_state(rthread, ShenandoahThreadLocalData::gc_state_offset());
 464   __ lea(tmp, gc_state);
 465   __ ldrb(tmp, __ legitimize_address(gc_state, 1, tmp));
 466 
 467   // Check if the heap is under weak-reference/roots processing, in
 468   // which case we need to take the slow path.
 469   __ tbnz(tmp, ShenandoahHeap::WEAK_ROOTS_BITPOS, slow_path);
 470   __ bind(done);
 471 }
 472 #endif
 473 
 474 // Special Shenandoah CAS implementation that handles false negatives due
 475 // to concurrent evacuation.  The service is more complex than a
 476 // traditional CAS operation because the CAS operation is intended to
 477 // succeed if the reference at addr exactly matches expected or if the
 478 // reference at addr holds a pointer to a from-space object that has
 479 // been relocated to the location named by expected.  There are two
 480 // races that must be addressed:
 481 //  a) A parallel thread may mutate the contents of addr so that it points
 482 //     to a different object.  In this case, the CAS operation should fail.
 483 //  b) A parallel thread may heal the contents of addr, replacing a
 484 //     from-space pointer held in addr with the to-space pointer
 485 //     representing the new location of the object.
 486 // Upon entry to cmpxchg_oop, it is assured that new_val equals null
 487 // or it refers to an object that is not being evacuated out of
 488 // from-space, or it refers to the to-space version of an object that
 489 // is being evacuated out of from-space.
 490 //
 491 // By default the value held in the result register following execution
 492 // of the generated code sequence is 0 to indicate failure of CAS,
 493 // non-zero to indicate success. If is_cae, the result is the value most
 494 // recently fetched from addr rather than a boolean success indicator.
 495 //
 496 // Clobbers rscratch1, rscratch2
 497 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm,
 498                                                 Register addr,
 499                                                 Register expected,
 500                                                 Register new_val,
 501                                                 bool acquire, bool release,
 502                                                 bool is_cae,
 503                                                 Register result) {
 504   Register tmp1 = rscratch1;
 505   Register tmp2 = rscratch2;
 506   bool is_narrow = UseCompressedOops;
 507   Assembler::operand_size size = is_narrow ? Assembler::word : Assembler::xword;
 508 
 509   assert_different_registers(addr, expected, tmp1, tmp2);
 510   assert_different_registers(addr, new_val,  tmp1, tmp2);
 511 
 512   Label step4, done;
 513 
 514   // There are two ways to reach this label.  Initial entry into the
 515   // cmpxchg_oop code expansion starts at step1 (which is equivalent
 516   // to label step4).  Additionally, in the rare case that four steps
 517   // are required to perform the requested operation, the fourth step
 518   // is the same as the first.  On a second pass through step 1,
 519   // control may flow through step 2 on its way to failure.  It will
 520   // not flow from step 2 to step 3 since we are assured that the
 521   // memory at addr no longer holds a from-space pointer.
 522   //
 523   // The comments that immediately follow the step4 label apply only
 524   // to the case in which control reaches this label by branch from
 525   // step 3.
 526 
 527   __ bind (step4);
 528 
 529   // Step 4. CAS has failed because the value most recently fetched
 530   // from addr is no longer the from-space pointer held in tmp2.  If a
 531   // different thread replaced the in-memory value with its equivalent
 532   // to-space pointer, then CAS may still be able to succeed.  The
 533   // value held in the expected register has not changed.
 534   //
 535   // It is extremely rare we reach this point.  For this reason, the
 536   // implementation opts for smaller rather than potentially faster
 537   // code.  Ultimately, smaller code for this rare case most likely
 538   // delivers higher overall throughput by enabling improved icache
 539   // performance.
 540 
 541   // Step 1. Fast-path.
 542   //
 543   // Try to CAS with given arguments.  If successful, then we are done.
 544   //
 545   // No label required for step 1.
 546 
 547   __ cmpxchg(addr, expected, new_val, size, acquire, release, false, tmp2);
 548   // EQ flag set iff success.  tmp2 holds value fetched.
 549 
 550   // If expected equals null but tmp2 does not equal null, the
 551   // following branches to done to report failure of CAS.  If both
 552   // expected and tmp2 equal null, the following branches to done to
 553   // report success of CAS.  There's no need for a special test of
 554   // expected equal to null.
 555 
 556   __ br(Assembler::EQ, done);
 557   // if CAS failed, fall through to step 2
 558 
 559   // Step 2. CAS has failed because the value held at addr does not
 560   // match expected.  This may be a false negative because the value fetched
 561   // from addr (now held in tmp2) may be a from-space pointer to the
 562   // original copy of same object referenced by to-space pointer expected.
 563   //
 564   // To resolve this, it suffices to find the forward pointer associated
 565   // with fetched value.  If this matches expected, retry CAS with new
 566   // parameters.  If this mismatches, then we have a legitimate
 567   // failure, and we're done.
 568   //
 569   // No need for step2 label.
 570 
 571   // overwrite tmp1 with from-space pointer fetched from memory
 572   __ mov(tmp1, tmp2);
 573 
 574   if (is_narrow) {
 575     // Decode tmp1 in order to resolve its forward pointer
 576     __ decode_heap_oop(tmp1, tmp1);
 577   }
 578   resolve_forward_pointer(masm, tmp1);
 579   // Encode tmp1 to compare against expected.
 580   __ encode_heap_oop(tmp1, tmp1);
 581 
 582   // Does forwarded value of fetched from-space pointer match original
 583   // value of expected?  If tmp1 holds null, this comparison will fail
 584   // because we know from step1 that expected is not null.  There is
 585   // no need for a separate test for tmp1 (the value originally held
 586   // in memory) equal to null.
 587   __ cmp(tmp1, expected);
 588 
 589   // If not, then the failure was legitimate and we're done.
 590   // Branching to done with NE condition denotes failure.
 591   __ br(Assembler::NE, done);
 592 
 593   // Fall through to step 3.  No need for step3 label.
 594 
 595   // Step 3.  We've confirmed that the value originally held in memory
 596   // (now held in tmp2) pointed to from-space version of original
 597   // expected value.  Try the CAS again with the from-space expected
 598   // value.  If it now succeeds, we're good.
 599   //
 600   // Note: tmp2 holds encoded from-space pointer that matches to-space
 601   // object residing at expected.  tmp2 is the new "expected".
 602 
 603   // Note that macro implementation of __cmpxchg cannot use same register
 604   // tmp2 for result and expected since it overwrites result before it
 605   // compares result with expected.
 606   __ cmpxchg(addr, tmp2, new_val, size, acquire, release, false, noreg);
 607   // EQ flag set iff success.  tmp2 holds value fetched, tmp1 (rscratch1) clobbered.
 608 
 609   // If fetched value did not equal the new expected, this could
 610   // still be a false negative because some other thread may have
 611   // newly overwritten the memory value with its to-space equivalent.
 612   __ br(Assembler::NE, step4);
 613 
 614   if (is_cae) {
 615     // We're falling through to done to indicate success.  Success
 616     // with is_cae is denoted by returning the value of expected as
 617     // result.
 618     __ mov(tmp2, expected);
 619   }
 620 
 621   __ bind(done);
 622   // At entry to done, the Z (EQ) flag is on iff if the CAS
 623   // operation was successful.  Additionally, if is_cae, tmp2 holds
 624   // the value most recently fetched from addr. In this case, success
 625   // is denoted by tmp2 matching expected.
 626 
 627   if (is_cae) {
 628     __ mov(result, tmp2);
 629   } else {
 630     __ cset(result, Assembler::EQ);
 631   }
 632 }
 633 
 634 #ifdef COMPILER2
 635 void ShenandoahBarrierStubC2::gc_state_check_c2(MacroAssembler* masm, Register gcstate, const unsigned char test_state, ShenandoahBarrierStubC2* slow_stub) {
 636   if (ShenandoahGCStateCheckRemove) {
 637     // Unrealistic: remove all barrier fastpath checks.
 638   } else if (ShenandoahGCStateCheckHotpatch) {
 639     // In the ideal world, we would hot-patch the branch to slow stub with a single
 640     // (unconditional) jump or nop, based on our current GC state.
 641     __ nop();
 642   } else {
 643     int bit_to_check = ShenandoahThreadLocalData::gc_state_to_fast_bit(test_state);
 644     Address gc_state_fast(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_fast_offset()));
 645     __ ldrb(gcstate, gc_state_fast);
 646     if (slow_stub->_test_and_branch_reachable) {
 647       __ tbnz(gcstate, bit_to_check, *slow_stub->entry());
 648     } else {
 649       __ tbz(gcstate, bit_to_check, *slow_stub->continuation());
 650       __ b(*slow_stub->entry());
 651     }
 652 
 653     // This is were the slowpath stub will return to or the code above will
 654     // jump to if the checks are false
 655     __ bind(*slow_stub->continuation());
 656   }
 657 }
 658 
 659 bool needs_acquiring_load_exclusive(const MachNode *n) {
 660   assert(n->is_CAS(true), "expecting a compare and swap");
 661   if (n->is_CAS(false)) {
 662     assert(n->has_trailing_membar(), "expected trailing membar");
 663   } else {
 664     return n->has_trailing_membar();
 665   }
 666 
 667   // so we can just return true here
 668   return true;
 669 }
 670 
 671 void ShenandoahBarrierSetAssembler::compare_and_set_c2(const MachNode* node, MacroAssembler* masm, Register res, Register addr,
 672     Register oldval, Register newval, Register tmp, bool exchange, bool maybe_null, bool narrow, bool weak) {
 673   bool acquire = needs_acquiring_load_exclusive(node);
 674   Assembler::operand_size op_size = narrow ? Assembler::word : Assembler::xword;
 675 
 676   // Pre-barrier covers several things:
 677   //  a. Avoids false positives from CAS encountering to-space memory values.
 678   //  b. Satisfies the need for LRB for the CAE result.
 679   //  c. Records old value for the sake of SATB.
 680   //
 681   // (a) and (b) are covered because load barrier does memory location fixup.
 682   // (c) is covered by KA on the current memory value.
 683   if (ShenandoahBarrierStubC2::needs_slow_barrier(node)) {
 684     ShenandoahBarrierStubC2* const stub = ShenandoahBarrierStubC2::create(node, tmp, addr, narrow, /* do_load: */ true, __ offset());
 685     char check = 0;
 686     check |= ShenandoahBarrierStubC2::needs_keep_alive_barrier(node) ? ShenandoahHeap::MARKING : 0;
 687     check |= ShenandoahBarrierStubC2::needs_load_ref_barrier(node)   ? ShenandoahHeap::HAS_FORWARDED : 0;
 688     assert(!ShenandoahBarrierStubC2::needs_load_ref_barrier_weak(node), "Not supported for CAS");
 689     ShenandoahBarrierStubC2::gc_state_check_c2(masm, rscratch1, check, stub);
 690   }
 691 
 692   // CAS!
 693   __ cmpxchg(addr, oldval, newval, op_size, acquire, /* release */ true, weak, exchange ? res : noreg);
 694 
 695   // If we need a boolean result out of CAS, set the flag appropriately and promote the result.
 696   if (!exchange) {
 697     assert(res != noreg, "need result register");
 698     __ cset(res, Assembler::EQ);
 699   }
 700 
 701   // Post-barrier deals with card updates.
 702   card_barrier_c2(node, masm, Address(addr, 0));
 703 }
 704 
 705 void ShenandoahBarrierSetAssembler::get_and_set_c2(const MachNode* node, MacroAssembler* masm, Register preval,
 706     Register newval, Register addr, Register tmp) {
 707   bool acquire = needs_acquiring_load_exclusive(node);
 708   bool narrow = node->bottom_type()->isa_narrowoop();
 709 
 710   // Pre-barrier covers several things:
 711   //  a. Satisfies the need for LRB for the GAS result.
 712   //  b. Records old value for the sake of SATB.
 713   //
 714   // (a) is covered because load barrier does memory location fixup.
 715   // (b) is covered by KA on the current memory value.
 716   if (ShenandoahBarrierStubC2::needs_slow_barrier(node)) {
 717     ShenandoahBarrierStubC2* const stub = ShenandoahBarrierStubC2::create(node, tmp, addr, narrow, /* do_load: */ true, __ offset());
 718     char check = 0;
 719     check |= ShenandoahBarrierStubC2::needs_keep_alive_barrier(node) ? ShenandoahHeap::MARKING : 0;
 720     check |= ShenandoahBarrierStubC2::needs_load_ref_barrier(node)   ? ShenandoahHeap::HAS_FORWARDED : 0;
 721     assert(!ShenandoahBarrierStubC2::needs_load_ref_barrier_weak(node), "Not supported for GAS");
 722     ShenandoahBarrierStubC2::gc_state_check_c2(masm, rscratch1, check, stub);
 723   }
 724 
 725   if (narrow) {
 726     if (acquire) {
 727       __ atomic_xchgalw(preval, newval, addr);
 728     } else {
 729       __ atomic_xchgw(preval, newval, addr);
 730     }
 731   } else {
 732     if (acquire) {
 733       __ atomic_xchgal(preval, newval, addr);
 734     } else {
 735       __ atomic_xchg(preval, newval, addr);
 736     }
 737   }
 738 
 739   // Post-barrier deals with card updates.
 740   card_barrier_c2(node, masm, Address(addr, 0));
 741 }
 742 
 743 void ShenandoahBarrierSetAssembler::store_c2(const MachNode* node, MacroAssembler* masm, Address dst, bool dst_narrow,
 744     Register src, bool src_narrow, Register tmp) {
 745 
 746   // Pre-barrier: SATB, keep-alive the current memory value.
 747   if (ShenandoahBarrierStubC2::needs_slow_barrier(node)) {
 748     assert(!ShenandoahBarrierStubC2::needs_load_ref_barrier(node), "Should not be required for stores");
 749     ShenandoahBarrierStubC2* const stub = ShenandoahBarrierStubC2::create(node, tmp, dst, dst_narrow, /* do_load: */ true, __ offset());
 750     ShenandoahBarrierStubC2::gc_state_check_c2(masm, rscratch1, ShenandoahHeap::MARKING, stub);
 751   }
 752 
 753   // Do the actual store
 754   bool is_volatile = node->has_trailing_membar();
 755   if (dst_narrow) {
 756     if (!src_narrow) {
 757       // Need to encode into rscratch, because we cannot clobber src.
 758       // TODO: Maybe there is a matcher way to test that src is unused after this?
 759       __ mov(rscratch1, src);
 760       if (ShenandoahBarrierStubC2::src_not_null(node)) {
 761         __ encode_heap_oop_not_null(rscratch1);
 762       } else {
 763         __ encode_heap_oop(rscratch1);
 764       }
 765       src = rscratch1;
 766     }
 767 
 768     if (is_volatile) {
 769       __ stlrw(src, dst.base());
 770     } else {
 771       __ strw(src, dst);
 772     }
 773   } else {
 774     if (is_volatile) {
 775       __ stlr(src, dst.base());
 776     } else {
 777       __ str(src, dst);
 778     }
 779   }
 780 
 781   // Post-barrier: card updates.
 782   card_barrier_c2(node, masm, dst);
 783 }
 784 
 785 void ShenandoahBarrierSetAssembler::load_c2(const MachNode* node, MacroAssembler* masm, Register dst, Address src) {
 786   bool acquire = node->memory_order() == MemNode::MemOrd::acquire;
 787   bool narrow = node->bottom_type()->isa_narrowoop();
 788 
 789   // Do the actual load. This load is the candidate for implicit null check, and MUST come first.
 790   if (narrow) {
 791     if (acquire) {
 792       __ ldarw(dst, src.base());
 793     } else {
 794       __ ldrw(dst, src);
 795     }
 796   } else {
 797     if (acquire) {
 798       __ ldar(dst, src.base());
 799     } else {
 800       __ ldr(dst, src);
 801     }
 802   }
 803 
 804   // Post-barrier: LRB / KA / weak-root processing.
 805   if (ShenandoahBarrierStubC2::needs_slow_barrier(node)) {
 806     ShenandoahBarrierStubC2* const stub = ShenandoahBarrierStubC2::create(node, dst, src, narrow, /* do_load: */ false, __ offset());
 807     char check = 0;
 808     check |= ShenandoahBarrierStubC2::needs_keep_alive_barrier(node)    ? ShenandoahHeap::MARKING : 0;
 809     check |= ShenandoahBarrierStubC2::needs_load_ref_barrier(node)      ? ShenandoahHeap::HAS_FORWARDED : 0;
 810     check |= ShenandoahBarrierStubC2::needs_load_ref_barrier_weak(node) ? ShenandoahHeap::WEAK_ROOTS : 0;
 811     ShenandoahBarrierStubC2::gc_state_check_c2(masm, rscratch1, check, stub);
 812   }
 813 }
 814 
 815 void ShenandoahBarrierSetAssembler::card_barrier_c2(const MachNode* node, MacroAssembler* masm, Address address) {
 816   if (ShenandoahSkipBarriers || (node->barrier_data() & ShenandoahBitCardMark) == 0) {
 817     return;
 818   }
 819 
 820   assert(CardTable::dirty_card_val() == 0, "must be");
 821   Assembler::InlineSkippedInstructionsCounter skip_counter(masm);
 822 
 823   // rscratch2 = addr >> CardTable::card_shift()
 824   __ lea(rscratch2, address);
 825   __ lsr(rscratch2, rscratch2, CardTable::card_shift());
 826 
 827   // rscratch1 = card table base (holder)
 828   Address curr_ct_holder_addr(rthread, in_bytes(ShenandoahThreadLocalData::card_table_offset()));
 829   __ ldr(rscratch1, curr_ct_holder_addr);
 830 
 831   // rscratch2 = &card_table[card_index]
 832   __ add(rscratch2, rscratch1, rscratch2);
 833 
 834   if (UseCondCardMark) {
 835     Label L_already_dirty;
 836     __ ldrb(rscratch1, Address(rscratch2));
 837     __ cbz(rscratch1, L_already_dirty);
 838     __ strb(zr, Address(rscratch2));
 839     __ bind(L_already_dirty);
 840   } else {
 841     __ strb(zr, Address(rscratch2));
 842   }
 843 }
 844 #undef __
 845 #define __ masm.
 846 
 847 // Only handles forward branch jumps, target_offset >= branch_offset
 848 // FIXME: copied verbatim from ZGC, duplicated code.
 849 static bool aarch64_test_and_branch_reachable(int branch_offset, int target_offset) {
 850   assert(branch_offset >= 0, "branch to stub offsets must be positive");
 851   assert(target_offset >= 0, "offset in stubs section must be positive");
 852   assert(target_offset >= branch_offset, "forward branches only, branch_offset -> target_offset");
 853 
 854   const int test_and_branch_delta_limit = 32 * K;
 855 
 856   const int test_and_branch_to_trampoline_delta = target_offset - branch_offset;
 857 
 858   return test_and_branch_to_trampoline_delta < test_and_branch_delta_limit;
 859 }
 860 
 861 ShenandoahBarrierStubC2::ShenandoahBarrierStubC2(const MachNode* node, Register obj, Address addr, bool narrow, bool do_load, int offset) :
 862   BarrierStubC2(node),
 863   _obj(obj),
 864   _addr(addr),
 865   _do_load(do_load),
 866   _narrow(narrow),
 867   _maybe_null(!src_not_null(node)),
 868   _needs_load_ref_barrier(needs_load_ref_barrier(node)),
 869   _needs_load_ref_weak_barrier(needs_load_ref_barrier_weak(node)),
 870   _needs_keep_alive_barrier(needs_keep_alive_barrier(node)),
 871   _fastpath_branch_offset(offset),
 872   _test_and_branch_reachable(),
 873   _skip_trampoline(),
 874   _test_and_branch_reachable_entry() {
 875 
 876   assert(!ShenandoahSkipBarriers, "Do not touch stubs when disabled");
 877   assert(!_narrow || is_heap_access(node), "Only heap accesses can be narrow");
 878 
 879   // If we are in scratch emit mode we assume worse case by leaving
 880   // _test_and_branch_reachable false.
 881   PhaseOutput* const output = Compile::current()->output();
 882   if (output->in_scratch_emit_size()) {
 883     return;
 884   }
 885 
 886   // Assume that each trampoline is one single instruction and that the stubs
 887   // will follow immediatelly after the _code section. Therefore, we are
 888   // checking if the distance between the fastpath branch and the
 889   // trampoline/entry of the current Stub is less than 32K.
 890   const int code_size = output->buffer_sizing_data()->_code;
 891   const int trampoline_offset = trampoline_stubs_count() * NativeInstruction::instruction_size;
 892   _test_and_branch_reachable = aarch64_test_and_branch_reachable(_fastpath_branch_offset, code_size + trampoline_offset);
 893   if (_test_and_branch_reachable) {
 894     inc_trampoline_stubs_count();
 895   }
 896 }
 897 
 898 void ShenandoahBarrierStubC2::emit_code(MacroAssembler& masm) {
 899   // If we reach here with _skip_trampoline set it means that earlier we
 900   // emitted a trampoline to this stub and now we need to emit the actual stub.
 901   if (_skip_trampoline) {
 902     emit_code_actual(masm);
 903     return;
 904   }
 905   _skip_trampoline = true;
 906 
 907   // The fastpath executes two branch instructions to reach this stub, let's
 908   // just emit the stub here and not add a third one.
 909   if (!_test_and_branch_reachable) {
 910     // By registering the stub again, after setting _skip_trampoline, we'll
 911     // effectivelly cause the stub to be emitted the next time ::emit_code is
 912     // called.
 913     ShenandoahBarrierStubC2::register_stub(this);
 914     return;
 915   }
 916 
 917   // This is entry point when coming from fastpath, IFF it's able to reach here
 918   // with a test and branch instruction, otherwise the entry is
 919   // ShenandoahBarrierStubC2::entry();
 920   const int target_offset = __ offset();
 921   __ bind(_test_and_branch_reachable_entry);
 922 
 923 #ifdef ASSERT
 924   // Current assumption is that the barrier stubs are the first stubs emitted
 925   // after the actual code
 926   PhaseOutput* const output = Compile::current()->output();
 927   assert(stubs_start_offset() <= output->buffer_sizing_data()->_code, "stubs are assumed to be emitted directly after code and code_size is a hard limit on where it can start");
 928   assert(aarch64_test_and_branch_reachable(_fastpath_branch_offset, target_offset), "trampoline should be reachable");
 929 #endif
 930 
 931   // Next fastpath branch's offset is unknown, but it's > current _fastpath_branch_offset
 932   const int next_branch_offset = _fastpath_branch_offset + NativeInstruction::instruction_size;
 933 
 934   // If emitting the current stub directly does not interfere with emission of
 935   // the next potential trampoline then do it to avoid executing additional
 936   // branch when coming from fastpath.
 937   if (aarch64_test_and_branch_reachable(next_branch_offset, target_offset + get_stub_size())) {
 938     emit_code_actual(masm);
 939   } else {
 940     __ b(*BarrierStubC2::entry());
 941     // By registering the stub again, after setting _skip_trampoline to true,
 942     // we'll effectivelly cause the stub to be emitted the next time
 943     // ::emit_code is called.
 944     ShenandoahBarrierStubC2::register_stub(this);
 945   }
 946 }
 947 
 948 Label* ShenandoahBarrierStubC2::entry() {
 949   if (_test_and_branch_reachable) {
 950     return &_test_and_branch_reachable_entry;
 951   }
 952   return BarrierStubC2::entry();
 953 }
 954 
 955 int ShenandoahBarrierStubC2::get_stub_size() {
 956   PhaseOutput* const output = Compile::current()->output();
 957   assert(!output->in_scratch_emit_size(), "only used when emitting stubs");
 958   BufferBlob* const blob = output->scratch_buffer_blob();
 959   CodeBuffer cb(blob->content_begin(), (address)output->scratch_locs_memory() - blob->content_begin());
 960   MacroAssembler masm(&cb);
 961   output->set_in_scratch_emit_size(true);
 962   emit_code_actual(masm);
 963   output->set_in_scratch_emit_size(false);
 964   return cb.insts_size();
 965 }
 966 
 967 void ShenandoahBarrierStubC2::emit_code_actual(MacroAssembler& masm) {
 968   assert(_needs_keep_alive_barrier || _needs_load_ref_barrier, "Why are you here?");
 969 
 970   // Stub entry
 971   if (!Compile::current()->output()->in_scratch_emit_size()) {
 972     __ bind(*BarrierStubC2::entry());
 973   }
 974 
 975   // If we need to load ourselves, do it here.
 976   if (_do_load) {
 977     // This does the load and the decode if necessary
 978     __ load_heap_oop(_obj, _addr, noreg, noreg, AS_RAW);
 979   } else if (_narrow) {
 980     // If object is narrow, we need to decode it first: barrier checks need full oops.
 981     if (_maybe_null) {
 982       __ decode_heap_oop(_obj);
 983     } else {
 984       __ decode_heap_oop_not_null(_obj);
 985     }
 986   }
 987 
 988   if (_do_load || _maybe_null) {
 989     __ cbz(_obj, *continuation());
 990   }
 991 
 992   keepalive(&masm, _obj, rscratch1, rscratch2);
 993 
 994   lrb(&masm, _obj, _addr, noreg);
 995 
 996   // If object is narrow, we need to encode it before exiting.
 997   // For encoding, dst can only turn null if we are dealing with weak loads.
 998   // Otherwise, we have already null-checked. We can skip all this if we performed
 999   // the load ourselves, which means the value is not used by caller.
1000   if (_narrow && !_do_load) {
1001     if (_needs_load_ref_weak_barrier) {
1002       __ encode_heap_oop(_obj);
1003     } else {
1004       __ encode_heap_oop_not_null(_obj);
1005     }
1006   }
1007 
1008   // Go back to fast path
1009   __ b(*continuation());
1010 }
1011 
1012 #undef __
1013 #define __ masm->
1014 
1015 void ShenandoahBarrierStubC2::keepalive(MacroAssembler* masm, Register obj, Register tmp1, Register tmp2) {
1016   Address index(rthread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
1017   Address buffer(rthread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
1018   Label L_runtime;
1019   Label L_done;
1020 
1021   // The node doesn't even need keepalive barrier, just don't check anything else
1022   if (!_needs_keep_alive_barrier) {
1023     return ;
1024   }
1025 
1026   // If both LRB and KeepAlive barriers are required (rare), do a runtime check
1027   // for enabled barrier.
1028   if (_needs_load_ref_barrier) {
1029     Address gcs_addr(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
1030     __ ldrb(rscratch1, gcs_addr);
1031     __ tbz(rscratch1, ShenandoahHeap::MARKING_BITPOS, L_done);
1032   }
1033 
1034   // If buffer is full, call into runtime.
1035   __ ldr(tmp1, index);
1036   __ cbz(tmp1, L_runtime);
1037 
1038   // The buffer is not full, store value into it.
1039   __ sub(tmp1, tmp1, wordSize);
1040   __ str(tmp1, index);
1041   __ ldr(tmp2, buffer);
1042   __ str(obj, Address(tmp2, tmp1));
1043   __ b(L_done);
1044 
1045   // Runtime call
1046   __ bind(L_runtime);
1047 
1048   preserve(obj);
1049   {
1050     SaveLiveRegisters save_registers(masm, this);
1051     __ mov(c_rarg0, obj);
1052     __ mov(tmp1, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_pre));
1053     __ blr(tmp1);
1054   }
1055 
1056   __ bind(L_done);
1057 }
1058 
1059 void ShenandoahBarrierStubC2::lrb(MacroAssembler* masm, Register obj, Address addr, Register tmp) {
1060   Label L_done;
1061 
1062   // The node doesn't even need LRB barrier, just don't check anything else
1063   if (!_needs_load_ref_barrier) {
1064     return ;
1065   }
1066 
1067   if ((_node->barrier_data() & ShenandoahBitStrong) != 0) {
1068     // If both LRB and KeepAlive barriers are required (rare), do a runtime
1069     // check for enabled barrier.
1070     if (_needs_keep_alive_barrier) {
1071       Address gcs_addr(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
1072       __ ldrb(rscratch1, gcs_addr);
1073       if (_needs_load_ref_weak_barrier) {
1074         __ orr(rscratch1, rscratch1, rscratch1, Assembler::LSR, ShenandoahHeap::WEAK_ROOTS_BITPOS);
1075       }
1076       __ tbz(rscratch1, ShenandoahHeap::HAS_FORWARDED_BITPOS, L_done);
1077     }
1078 
1079     // Weak/phantom loads always need to go to runtime. For strong refs we
1080     // check if the object in cset, if they are not, then we are done with LRB.
1081     __ mov(rscratch2, ShenandoahHeap::in_cset_fast_test_addr());
1082     __ lsr(rscratch1, obj, ShenandoahHeapRegion::region_size_bytes_shift_jint());
1083     __ ldrb(rscratch2, Address(rscratch2, rscratch1));
1084     __ cbz(rscratch2, L_done);
1085   }
1086 
1087   dont_preserve(obj);
1088   {
1089     SaveLiveRegisters save_registers(masm, this);
1090 
1091     // Shuffle in the arguments. The end result should be:
1092     //   c_rarg0 <-- obj
1093     //   c_rarg1 <-- lea(addr)
1094     if (c_rarg0 == obj) {
1095       __ lea(c_rarg1, addr);
1096     } else if (c_rarg1 == obj) {
1097       // Set up arguments in reverse, and then flip them
1098       __ lea(c_rarg0, addr);
1099       // flip them
1100       __ mov(rscratch1, c_rarg0);
1101       __ mov(c_rarg0, c_rarg1);
1102       __ mov(c_rarg1, rscratch1);
1103     } else {
1104       assert_different_registers(c_rarg1, obj);
1105       __ lea(c_rarg1, addr);
1106       __ mov(c_rarg0, obj);
1107     }
1108 
1109     // Get address of runtime LRB entry and call it
1110     __ mov(rscratch1, lrb_runtime_entry_addr());
1111     __ blr(rscratch1);
1112 
1113     // If we loaded the object in the stub it means we don't need to return it
1114     // to fastpath, so no need to make this mov.
1115     if (!_do_load) {
1116       __ mov(obj, r0);
1117     }
1118   }
1119 
1120   __ bind(L_done);
1121 }
1122 
1123 #endif // COMPILER2
1124 
1125 void ShenandoahBarrierSetAssembler::gen_write_ref_array_post_barrier(MacroAssembler* masm, DecoratorSet decorators,
1126                                                                      Register start, Register count, Register scratch) {
1127   assert(ShenandoahCardBarrier, "Should have been checked by caller");
1128 
1129   Label L_loop, L_done;
1130   const Register end = count;
1131 
1132   // Zero count? Nothing to do.
1133   __ cbz(count, L_done);
1134 
1135   // end = start + count << LogBytesPerHeapOop
1136   // last element address to make inclusive
1137   __ lea(end, Address(start, count, Address::lsl(LogBytesPerHeapOop)));
1138   __ sub(end, end, BytesPerHeapOop);
1139   __ lsr(start, start, CardTable::card_shift());
1140   __ lsr(end, end, CardTable::card_shift());
1141 
1142   // number of bytes to copy
1143   __ sub(count, end, start);
1144 
1145   Address curr_ct_holder_addr(rthread, in_bytes(ShenandoahThreadLocalData::card_table_offset()));
1146   __ ldr(scratch, curr_ct_holder_addr);
1147   __ add(start, start, scratch);
1148   __ bind(L_loop);
1149   __ strb(zr, Address(start, count));
1150   __ subs(count, count, 1);
1151   __ br(Assembler::GE, L_loop);
1152   __ bind(L_done);
1153 }
1154 
1155 #undef __
1156 
1157 #ifdef COMPILER1
1158 
1159 #define __ ce->masm()->
1160 
1161 void ShenandoahBarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, ShenandoahPreBarrierStub* stub) {
1162   ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
1163   // At this point we know that marking is in progress.
1164   // If do_load() is true then we have to emit the
1165   // load of the previous value; otherwise it has already
1166   // been loaded into _pre_val.
1167 
1168   __ bind(*stub->entry());
1169 
1170   assert(stub->pre_val()->is_register(), "Precondition.");
1171 
1172   Register pre_val_reg = stub->pre_val()->as_register();
1173 
1174   if (stub->do_load()) {
1175     ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/);
1176   }
1177   __ cbz(pre_val_reg, *stub->continuation());
1178   ce->store_parameter(stub->pre_val()->as_register(), 0);
1179   __ far_call(RuntimeAddress(bs->pre_barrier_c1_runtime_code_blob()->code_begin()));
1180   __ b(*stub->continuation());
1181 }
1182 
1183 void ShenandoahBarrierSetAssembler::gen_load_reference_barrier_stub(LIR_Assembler* ce, ShenandoahLoadReferenceBarrierStub* stub) {
1184   ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
1185   __ bind(*stub->entry());
1186 
1187   DecoratorSet decorators = stub->decorators();
1188   bool is_strong  = ShenandoahBarrierSet::is_strong_access(decorators);
1189   bool is_weak    = ShenandoahBarrierSet::is_weak_access(decorators);
1190   bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
1191   bool is_native  = ShenandoahBarrierSet::is_native_access(decorators);
1192 
1193   Register obj = stub->obj()->as_register();
1194   Register res = stub->result()->as_register();
1195   Register addr = stub->addr()->as_pointer_register();
1196   Register tmp1 = stub->tmp1()->as_register();
1197   Register tmp2 = stub->tmp2()->as_register();
1198 
1199   assert(res == r0, "result must arrive in r0");
1200 
1201   if (res != obj) {
1202     __ mov(res, obj);
1203   }
1204 
1205   if (is_strong) {
1206     // Check for object in cset.
1207     __ mov(tmp2, ShenandoahHeap::in_cset_fast_test_addr());
1208     __ lsr(tmp1, res, ShenandoahHeapRegion::region_size_bytes_shift_jint());
1209     __ ldrb(tmp2, Address(tmp2, tmp1));
1210     __ cbz(tmp2, *stub->continuation());
1211   }
1212 
1213   ce->store_parameter(res, 0);
1214   ce->store_parameter(addr, 1);
1215   if (is_strong) {
1216     if (is_native) {
1217       __ far_call(RuntimeAddress(bs->load_reference_barrier_strong_native_rt_code_blob()->code_begin()));
1218     } else {
1219       __ far_call(RuntimeAddress(bs->load_reference_barrier_strong_rt_code_blob()->code_begin()));
1220     }
1221   } else if (is_weak) {
1222     __ far_call(RuntimeAddress(bs->load_reference_barrier_weak_rt_code_blob()->code_begin()));
1223   } else {
1224     assert(is_phantom, "only remaining strength");
1225     __ far_call(RuntimeAddress(bs->load_reference_barrier_phantom_rt_code_blob()->code_begin()));
1226   }
1227 
1228   __ b(*stub->continuation());
1229 }
1230 
1231 #undef __
1232 
1233 #define __ sasm->
1234 
1235 void ShenandoahBarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) {
1236   __ prologue("shenandoah_pre_barrier", false);
1237 
1238   // arg0 : previous value of memory
1239 
1240   BarrierSet* bs = BarrierSet::barrier_set();
1241 
1242   const Register pre_val = r0;
1243   const Register thread = rthread;
1244   const Register tmp = rscratch1;
1245 
1246   Address queue_index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
1247   Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
1248 
1249   Label done;
1250   Label runtime;
1251 
1252   // Is marking still active?
1253   Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
1254   __ ldrb(tmp, gc_state);
1255   __ tbz(tmp, ShenandoahHeap::MARKING_BITPOS, done);
1256 
1257   // Can we store original value in the thread's buffer?
1258   __ ldr(tmp, queue_index);
1259   __ cbz(tmp, runtime);
1260 
1261   __ sub(tmp, tmp, wordSize);
1262   __ str(tmp, queue_index);
1263   __ ldr(rscratch2, buffer);
1264   __ add(tmp, tmp, rscratch2);
1265   __ load_parameter(0, rscratch2);
1266   __ str(rscratch2, Address(tmp, 0));
1267   __ b(done);
1268 
1269   __ bind(runtime);
1270   __ push_call_clobbered_registers();
1271   __ load_parameter(0, pre_val);
1272   __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_pre), pre_val);
1273   __ pop_call_clobbered_registers();
1274   __ bind(done);
1275 
1276   __ epilogue();
1277 }
1278 
1279 void ShenandoahBarrierSetAssembler::generate_c1_load_reference_barrier_runtime_stub(StubAssembler* sasm, DecoratorSet decorators) {
1280   __ prologue("shenandoah_load_reference_barrier", false);
1281   // arg0 : object to be resolved
1282 
1283   __ push_call_clobbered_registers();
1284   __ load_parameter(0, r0);
1285   __ load_parameter(1, r1);
1286 
1287   bool is_strong  = ShenandoahBarrierSet::is_strong_access(decorators);
1288   bool is_weak    = ShenandoahBarrierSet::is_weak_access(decorators);
1289   bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
1290   bool is_native  = ShenandoahBarrierSet::is_native_access(decorators);
1291   if (is_strong) {
1292     if (is_native) {
1293       __ mov(lr, CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong));
1294     } else {
1295       if (UseCompressedOops) {
1296         __ mov(lr, CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow));
1297       } else {
1298         __ mov(lr, CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong));
1299       }
1300     }
1301   } else if (is_weak) {
1302     assert(!is_native, "weak must not be called off-heap");
1303     if (UseCompressedOops) {
1304       __ mov(lr, CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow));
1305     } else {
1306       __ mov(lr, CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak));
1307     }
1308   } else {
1309     assert(is_phantom, "only remaining strength");
1310     assert(is_native, "phantom must only be called off-heap");
1311     __ mov(lr, CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom));
1312   }
1313   __ blr(lr);
1314   __ mov(rscratch1, r0);
1315   __ pop_call_clobbered_registers();
1316   __ mov(r0, rscratch1);
1317 
1318   __ epilogue();
1319 }
1320 
1321 #undef __
1322 
1323 #endif // COMPILER1