1 /*
   2  * Copyright (c) 2026, Oracle and/or its affiliates. All rights reserved.
   3  * Copyright (c) 2018, 2021, 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/shenandoahHeap.inline.hpp"
  32 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
  33 #include "gc/shenandoah/shenandoahRuntime.hpp"
  34 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
  35 #include "interpreter/interpreter.hpp"
  36 #include "runtime/javaThread.hpp"
  37 #include "runtime/sharedRuntime.hpp"
  38 #include "utilities/macros.hpp"
  39 #ifdef COMPILER1
  40 #include "c1/c1_LIRAssembler.hpp"
  41 #include "c1/c1_MacroAssembler.hpp"
  42 #include "gc/shenandoah/c1/shenandoahBarrierSetC1.hpp"
  43 #endif
  44 #ifdef COMPILER2
  45 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
  46 #endif
  47 
  48 #define __ masm->
  49 
  50 void ShenandoahBarrierSetAssembler::arraycopy_prologue(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
  51                                                        Register src, Register dst, Register count) {
  52 
  53   bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0;
  54 
  55   if (is_reference_type(type)) {
  56     if (ShenandoahCardBarrier) {
  57       bool checkcast = (decorators & ARRAYCOPY_CHECKCAST) != 0;
  58       bool disjoint = (decorators & ARRAYCOPY_DISJOINT) != 0;
  59       bool obj_int = (type == T_OBJECT) && UseCompressedOops;
  60 
  61       // We need to save the original element count because the array copy stub
  62       // will destroy the value and we need it for the card marking barrier.
  63       if (!checkcast) {
  64         if (!obj_int) {
  65           // Save count for barrier
  66           __ movptr(r11, count);
  67         } else if (disjoint) {
  68           // Save dst in r11 in the disjoint case
  69           __ movq(r11, dst);
  70         }
  71       }
  72     }
  73 
  74     if ((ShenandoahSATBBarrier && !dest_uninitialized) || ShenandoahLoadRefBarrier) {
  75       Register thread = r15_thread;
  76       assert_different_registers(src, dst, count, thread);
  77 
  78       Label L_done;
  79       // Short-circuit if count == 0.
  80       __ testptr(count, count);
  81       __ jcc(Assembler::zero, L_done);
  82 
  83       // Avoid runtime call when not active.
  84       Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
  85       int flags;
  86       if (ShenandoahSATBBarrier && dest_uninitialized) {
  87         flags = ShenandoahHeap::HAS_FORWARDED;
  88       } else {
  89         flags = ShenandoahHeap::HAS_FORWARDED | ShenandoahHeap::MARKING;
  90       }
  91       __ testb(gc_state, flags);
  92       __ jcc(Assembler::zero, L_done);
  93 
  94       __ push_call_clobbered_registers(/* save_fpu = */ false);
  95       // If arguments are not in proper places, shuffle them.
  96       // Doing this via the stack is the most straight-forward way to avoid
  97       // accidentally smashing any register.
  98       if (c_rarg0 != src || c_rarg1 != dst || c_rarg2 != count) {
  99         __ push(src);
 100         __ push(dst);
 101         __ push(count);
 102         __ pop(c_rarg2);
 103         __ pop(c_rarg1);
 104         __ pop(c_rarg0);
 105       }
 106       address target = nullptr;
 107       if (UseCompressedOops) {
 108         target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::arraycopy_barrier_narrow_oop);
 109       } else {
 110         target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::arraycopy_barrier_oop);
 111       }
 112       __ call_VM_leaf(target, 3);
 113 
 114       __ pop_call_clobbered_registers(/* restore_fpu = */ false);
 115 
 116       __ bind(L_done);
 117     }
 118   }
 119 
 120 }
 121 
 122 void ShenandoahBarrierSetAssembler::arraycopy_epilogue(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 123                                                        Register src, Register dst, Register count) {
 124 
 125   if (ShenandoahCardBarrier && is_reference_type(type)) {
 126     bool checkcast = (decorators & ARRAYCOPY_CHECKCAST) != 0;
 127     bool disjoint = (decorators & ARRAYCOPY_DISJOINT) != 0;
 128     bool obj_int = (type == T_OBJECT) && UseCompressedOops;
 129     Register tmp = rax;
 130 
 131     if (!checkcast) {
 132       if (!obj_int) {
 133         // Save count for barrier
 134         count = r11;
 135       } else if (disjoint) {
 136         // Use the saved dst in the disjoint case
 137         dst = r11;
 138       }
 139     } else {
 140       tmp = rscratch1;
 141     }
 142     gen_write_ref_array_post_barrier(masm, decorators, dst, count, tmp);
 143   }
 144 }
 145 
 146 void ShenandoahBarrierSetAssembler::satb_barrier(MacroAssembler* masm,
 147                                                  Register obj,
 148                                                  Register pre_val,
 149                                                  Register tmp) {
 150   assert(ShenandoahSATBBarrier, "Should be checked by caller");
 151   const Register thread = r15_thread;
 152 
 153   Label done;
 154   Label runtime;
 155 
 156   assert(pre_val != noreg, "check this code");
 157   assert_different_registers(obj, pre_val, tmp);
 158 
 159   Address index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
 160   Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
 161 
 162   Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 163   __ testb(gc_state, ShenandoahHeap::MARKING);
 164   __ jcc(Assembler::zero, done);
 165 
 166   // Do we need to load the previous value?
 167   if (obj != noreg) {
 168     if (UseCompressedOops) {
 169       __ movl(pre_val, Address(obj, 0));
 170       __ decode_heap_oop(pre_val);
 171     } else {
 172       __ movq(pre_val, Address(obj, 0));
 173     }
 174   }
 175 
 176   // Is the previous value null?
 177   __ cmpptr(pre_val, NULL_WORD);
 178   __ jcc(Assembler::equal, done);
 179 
 180   // Can we store original value in the thread's buffer?
 181   // Is index == 0?
 182   // (The index field is typed as size_t.)
 183 
 184   __ movptr(tmp, index);                   // tmp := *index_adr
 185   __ cmpptr(tmp, 0);                       // tmp == 0?
 186   __ jcc(Assembler::equal, runtime);       // If yes, goto runtime
 187 
 188   __ subptr(tmp, wordSize);                // tmp := tmp - wordSize
 189   __ movptr(index, tmp);                   // *index_adr := tmp
 190   __ addptr(tmp, buffer);                  // tmp := tmp + *buffer_adr
 191 
 192   // Record the previous value
 193   __ movptr(Address(tmp, 0), pre_val);
 194   __ jmp(done);
 195 
 196   __ bind(runtime);
 197 
 198   // Slow-path call.
 199   // Some paths can be reached from the c2i adapter with live fp arguments in registers.
 200   __ enter();
 201   __ push_call_clobbered_registers(/* save_fpu = */ true);
 202 
 203   assert(thread != c_rarg0, "smashed arg");
 204   if (c_rarg0 != pre_val) {
 205     __ mov(c_rarg0, pre_val);
 206   }
 207 
 208   // Calling with super_call_VM_leaf with c_rarg0 bypasses interpreter checks and avoids any moves.
 209   __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_pre), c_rarg0);
 210 
 211   __ pop_call_clobbered_registers(/* restore_fpu = */ true);
 212   __ leave();
 213 
 214   __ bind(done);
 215 }
 216 
 217 void ShenandoahBarrierSetAssembler::load_reference_barrier(MacroAssembler* masm, Register dst, Address src, DecoratorSet decorators) {
 218   assert(ShenandoahLoadRefBarrier, "Should be enabled");
 219 
 220   bool is_strong  = ShenandoahBarrierSet::is_strong_access(decorators);
 221   bool is_weak    = ShenandoahBarrierSet::is_weak_access(decorators);
 222   bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
 223   bool is_native  = ShenandoahBarrierSet::is_native_access(decorators);
 224   bool is_narrow  = UseCompressedOops && !is_native;
 225 
 226   Label heap_stable, not_cset;
 227 
 228   __ block_comment("load_reference_barrier { ");
 229 
 230   // Check if GC is active
 231   Register thread = r15_thread;
 232 
 233   Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 234   int flags = ShenandoahHeap::HAS_FORWARDED;
 235   if (!is_strong) {
 236     flags |= ShenandoahHeap::WEAK_ROOTS;
 237   }
 238   __ testb(gc_state, flags);
 239   __ jcc(Assembler::zero, heap_stable);
 240 
 241   Register tmp1 = noreg, tmp2 = noreg;
 242   if (is_strong) {
 243     // Test for object in cset
 244     // Allocate temporary registers
 245     for (int i = 0; i < Register::available_gp_registers(); i++) {
 246       Register r = as_Register(i);
 247       if (r != rsp && r != rbp && r != rcx && r != dst && r != src.base() && r != src.index() ) {
 248         if (tmp1 == noreg) {
 249           tmp1 = r;
 250         } else {
 251           tmp2 = r;
 252           break;
 253         }
 254       }
 255     }
 256     assert(tmp1 != noreg, "tmp1 allocated");
 257     assert(tmp2 != noreg, "tmp2 allocated");
 258     assert_different_registers(tmp1, tmp2, src.base(), src.index());
 259     assert_different_registers(tmp1, tmp2, dst);
 260 
 261     __ push(tmp1);
 262     __ push(tmp2);
 263 
 264     // Optimized cset-test
 265     __ movptr(tmp1, dst);
 266     if (AOTCodeCache::is_on_for_dump()) {
 267       assert_different_registers(tmp1, tmp2, rcx);
 268       __ lea(tmp2, ExternalAddress(AOTRuntimeConstants::grain_shift_address()));
 269       __ push(rcx);
 270       __ movb(rcx, Address(tmp2));
 271       __ shrptr(tmp1);
 272       __ pop(rcx);
 273       __ lea(tmp2, ExternalAddress(AOTRuntimeConstants::cset_base_address()));
 274       __ movptr(tmp2, Address(tmp2));
 275     } else {
 276       __ shrptr(tmp1, ShenandoahHeapRegion::region_size_bytes_shift_jint());
 277       __ movptr(tmp2, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr());
 278     }
 279     __ movbool(tmp1, Address(tmp1, tmp2, Address::times_1));
 280     __ testbool(tmp1);
 281     __ jcc(Assembler::zero, not_cset);
 282   }
 283 
 284   // Slow-path call.
 285   // Save registers that can be clobbered by call.
 286   // Some paths can be reached from the c2i adapter with live fp arguments in registers.
 287   __ enter();
 288   if (dst != rax) {
 289     __ push(rax);
 290   }
 291   __ push_call_clobbered_registers_except(rax, /* save_fpu = */ true);
 292 
 293   // Shuffle registers such that dst is in c_rarg0 and addr in c_rarg1.
 294   if (dst == c_rarg1) {
 295     __ lea(c_rarg0, src);
 296     __ xchgptr(c_rarg1, c_rarg0);
 297   } else {
 298     __ lea(c_rarg1, src);
 299     __ movptr(c_rarg0, dst);
 300   }
 301 
 302   address target = nullptr;
 303   if (is_strong) {
 304     if (is_narrow) {
 305       target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow);
 306     } else {
 307       target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong);
 308     }
 309   } else if (is_weak) {
 310     if (is_narrow) {
 311       target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow);
 312     } else {
 313       target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak);
 314     }
 315   } else {
 316     assert(is_phantom, "only remaining strength");
 317     assert(!is_narrow, "phantom access cannot be narrow");
 318     target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom);
 319   }
 320 
 321   // Calling with super_call_VM_leaf with c_rarg0/1 bypasses interpreter checks and avoids any moves.
 322   __ super_call_VM_leaf(target, c_rarg0, c_rarg1);
 323   __ pop_call_clobbered_registers_except(rax, /* restore_fpu = */ true);
 324   if (dst != rax) {
 325     __ movptr(dst, rax);
 326     __ pop(rax);
 327   }
 328   __ leave();
 329 
 330   __ bind(not_cset);
 331 
 332   if  (is_strong) {
 333     __ pop(tmp2);
 334     __ pop(tmp1);
 335   }
 336 
 337   __ bind(heap_stable);
 338 
 339   __ block_comment("} load_reference_barrier");
 340 }
 341 
 342 //
 343 // Arguments:
 344 //
 345 // Inputs:
 346 //   src:        oop location, might be clobbered
 347 //   tmp1:       scratch register, might not be valid.
 348 //
 349 // Output:
 350 //   dst:        oop loaded from src location
 351 //
 352 // Kill:
 353 //   tmp1 (if it is valid)
 354 //
 355 void ShenandoahBarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 356              Register dst, Address src, Register tmp1) {
 357   // 1: non-reference load, no additional barrier is needed
 358   if (!is_reference_type(type)) {
 359     BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1);
 360     return;
 361   }
 362 
 363   assert((decorators & ON_UNKNOWN_OOP_REF) == 0, "Not expected");
 364 
 365   // 2: load a reference from src location and apply LRB if needed
 366   if (ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) {
 367     Register result_dst = dst;
 368     bool use_tmp1_for_dst = false;
 369 
 370     // Preserve src location for LRB
 371     if (dst == src.base() || dst == src.index()) {
 372     // Use tmp1 for dst if possible, as it is not used in BarrierAssembler::load_at()
 373       if (tmp1->is_valid() && tmp1 != src.base() && tmp1 != src.index()) {
 374         dst = tmp1;
 375         use_tmp1_for_dst = true;
 376       } else {
 377         dst = rdi;
 378         __ push(dst);
 379       }
 380       assert_different_registers(dst, src.base(), src.index());
 381     }
 382 
 383     BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1);
 384 
 385     load_reference_barrier(masm, dst, src, decorators);
 386 
 387     // Move loaded oop to final destination
 388     if (dst != result_dst) {
 389       __ movptr(result_dst, dst);
 390 
 391       if (!use_tmp1_for_dst) {
 392         __ pop(dst);
 393       }
 394 
 395       dst = result_dst;
 396     }
 397   } else {
 398     BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1);
 399   }
 400 
 401   // 3: apply keep-alive barrier if needed
 402   if (ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) {
 403     satb_barrier(masm /* masm */,
 404                  noreg /* obj */,
 405                  dst /* pre_val */,
 406                  tmp1 /* tmp */);
 407   }
 408 }
 409 
 410 void ShenandoahBarrierSetAssembler::card_barrier(MacroAssembler* masm, Register obj) {
 411   assert(ShenandoahCardBarrier, "Should have been checked by caller");
 412 
 413   // Does a store check for the oop in register obj. The content of
 414   // register obj is destroyed afterwards.
 415   __ shrptr(obj, CardTable::card_shift());
 416 
 417   // We'll use this register as the TLS base address and also later on
 418   // to hold the byte_map_base.
 419   Register thread = r15_thread;
 420   Register tmp = rscratch1;
 421 
 422   Address curr_ct_holder_addr(thread, in_bytes(ShenandoahThreadLocalData::card_table_offset()));
 423   __ movptr(tmp, curr_ct_holder_addr);
 424   Address card_addr(tmp, obj, Address::times_1);
 425 
 426   int dirty = CardTable::dirty_card_val();
 427   if (UseCondCardMark) {
 428     Label L_already_dirty;
 429     __ cmpb(card_addr, dirty);
 430     __ jccb(Assembler::equal, L_already_dirty);
 431     __ movb(card_addr, dirty);
 432     __ bind(L_already_dirty);
 433   } else {
 434     __ movb(card_addr, dirty);
 435   }
 436 }
 437 
 438 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 439               Address dst, Register val, Register tmp1, Register tmp2, Register tmp3) {
 440 
 441   // 1: non-reference types require no barriers
 442   if (!is_reference_type(type)) {
 443     BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2, tmp3);
 444     return;
 445   }
 446 
 447   // Flatten object address right away for simplicity: likely needed by barriers
 448   assert_different_registers(val, tmp1, tmp2, tmp3, r15_thread);
 449   if (dst.index() == noreg && dst.disp() == 0) {
 450     if (dst.base() != tmp1) {
 451       __ movptr(tmp1, dst.base());
 452     }
 453   } else {
 454     __ lea(tmp1, dst);
 455   }
 456 
 457   // 2: pre-barrier: SATB needs the previous value
 458   if (ShenandoahBarrierSet::need_satb_barrier(decorators, type)) {
 459     satb_barrier(masm,
 460                  tmp1 /* obj */,
 461                  tmp2 /* pre_val */,
 462                  tmp3 /* tmp */);
 463   }
 464 
 465   // Store!
 466   BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg, noreg);
 467 
 468   // 3: post-barrier: card barrier needs store address
 469   bool storing_non_null = (val != noreg);
 470   if (ShenandoahBarrierSet::need_card_barrier(decorators, type) && storing_non_null) {
 471     card_barrier(masm, tmp1);
 472   }
 473 }
 474 
 475 void ShenandoahBarrierSetAssembler::try_resolve_jobject_in_native(MacroAssembler* masm, Register jni_env,
 476                                                                   Register obj, Register tmp, Label& slowpath) {
 477   Label done;
 478   // Resolve jobject
 479   BarrierSetAssembler::try_resolve_jobject_in_native(masm, jni_env, obj, tmp, slowpath);
 480 
 481   // Check for null.
 482   __ testptr(obj, obj);
 483   __ jcc(Assembler::zero, done);
 484 
 485   Address gc_state(jni_env, ShenandoahThreadLocalData::gc_state_offset() - JavaThread::jni_environment_offset());
 486   __ testb(gc_state, ShenandoahHeap::EVACUATION);
 487   __ jccb(Assembler::notZero, slowpath);
 488   __ bind(done);
 489 }
 490 
 491 void ShenandoahBarrierSetAssembler::try_peek_weak_handle_in_nmethod(MacroAssembler* masm, Register weak_handle, Register obj, Label& slowpath) {
 492   Label done;
 493 
 494   // Peek weak handle using the standard implementation.
 495   BarrierSetAssembler::try_peek_weak_handle_in_nmethod(masm, weak_handle, obj, slowpath);
 496 
 497   // Check if the reference is null, and if it is, take the fast path.
 498   __ testptr(obj, obj);
 499   __ jcc(Assembler::zero, done);
 500 
 501   Address gc_state(r15_thread, ShenandoahThreadLocalData::gc_state_offset());
 502 
 503   // Check if the heap is under weak-reference/roots processing, in
 504   // which case we need to take the slow path.
 505   __ testb(gc_state, ShenandoahHeap::WEAK_ROOTS);
 506   __ jcc(Assembler::notZero, slowpath);
 507   __ bind(done);
 508 }
 509 
 510 void ShenandoahBarrierSetAssembler::check_oop(MacroAssembler* masm, Register obj, Register tmp1, Register tmp2, Label& L_error) {
 511   // Check if the oop is in the right area of memory
 512   __ movptr(tmp1, obj);
 513   __ movptr(tmp2, (intptr_t) Universe::verify_oop_mask());
 514   __ andptr(tmp1, tmp2);
 515   __ movptr(tmp2, (intptr_t) Universe::verify_oop_bits());
 516   __ cmpptr(tmp1, tmp2);
 517   __ jcc(Assembler::notZero, L_error);
 518 
 519   // This routine is sometimes called before applying GC barriers.
 520   // With +COH, loading the klass may end up loading forwarding pointer instead.
 521   Label L_skip;
 522   if (UseCompactObjectHeaders) {
 523     Address gc_state(r15_thread, ShenandoahThreadLocalData::gc_state_offset());
 524     __ testb(gc_state, ShenandoahHeap::HAS_FORWARDED);
 525     __ jcc(Assembler::notZero, L_skip);
 526   }
 527 
 528   // Make sure klass is 'reasonable', which is not zero.
 529   __ load_narrow_klass(tmp1, obj);
 530   __ testl(tmp1, tmp1);
 531   __ jcc(Assembler::zero, L_error);
 532 
 533   __ bind(L_skip);
 534 }
 535 
 536 #ifdef PRODUCT
 537 #define BLOCK_COMMENT(str) /* nothing */
 538 #else
 539 #define BLOCK_COMMENT(str) __ block_comment(str)
 540 #endif
 541 
 542 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
 543 
 544 #define TIMES_OOP (UseCompressedOops ? Address::times_4 : Address::times_8)
 545 
 546 void ShenandoahBarrierSetAssembler::gen_write_ref_array_post_barrier(MacroAssembler* masm, DecoratorSet decorators,
 547                                                                      Register addr, Register count,
 548                                                                      Register tmp) {
 549   assert(ShenandoahCardBarrier, "Should have been checked by caller");
 550 
 551   Label L_loop, L_done;
 552   const Register end = count;
 553   assert_different_registers(addr, end);
 554 
 555   // Zero count? Nothing to do.
 556   __ testl(count, count);
 557   __ jccb(Assembler::zero, L_done);
 558 
 559   const Register thread = r15_thread;
 560   Address curr_ct_holder_addr(thread, in_bytes(ShenandoahThreadLocalData::card_table_offset()));
 561   __ movptr(tmp, curr_ct_holder_addr);
 562 
 563   __ leaq(end, Address(addr, count, TIMES_OOP, 0));  // end == addr+count*oop_size
 564   __ subptr(end, BytesPerHeapOop); // end - 1 to make inclusive
 565   __ shrptr(addr, CardTable::card_shift());
 566   __ shrptr(end, CardTable::card_shift());
 567   __ subptr(end, addr); // end --> cards count
 568 
 569   __ addptr(addr, tmp);
 570 
 571   __ BIND(L_loop);
 572   __ movb(Address(addr, count, Address::times_1), 0);
 573   __ decrement(count);
 574   __ jccb(Assembler::greaterEqual, L_loop);
 575 
 576   __ BIND(L_done);
 577 }
 578 
 579 #undef __
 580 
 581 #ifdef COMPILER1
 582 
 583 #define __ ce->masm()->
 584 
 585 void ShenandoahBarrierSetAssembler::keepalive_barrier_c1_stub(LIR_Assembler* ce, ShenandoahKeepaliveBarrierStub* stub) {
 586   __ bind(*stub->entry());
 587 
 588   ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
 589 
 590   Register obj = stub->obj()->as_register();
 591 
 592   if (stub->do_load()) {
 593     ce->mem2reg(stub->addr(), stub->obj(), T_OBJECT, lir_patch_none, nullptr, /* wide = */ false);
 594   }
 595   __ cmpptr(obj, NULL_WORD);
 596   __ jcc(Assembler::equal, *stub->continuation());
 597 
 598   ce->store_parameter(obj, 0);
 599   __ call(RuntimeAddress(bs->keepalive_barrier_stub()));
 600   __ jmp(*stub->continuation());
 601 }
 602 
 603 void ShenandoahBarrierSetAssembler::load_reference_barrier_c1_stub(LIR_Assembler* ce, ShenandoahLoadReferenceBarrierStub* stub) {
 604   __ bind(*stub->entry());
 605 
 606   ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
 607 
 608   Register obj = stub->obj()->as_register();
 609   Register addr = stub->addr()->as_pointer_register();
 610   Register slow_result = stub->slow_result()->as_register();
 611   assert_different_registers(obj, addr, slow_result);
 612   assert(slow_result == rax, "C1 must know about our slow call result register");
 613 
 614   ce->store_parameter(obj, 0);
 615   ce->store_parameter(addr, 1);
 616   __ call(RuntimeAddress(bs->load_reference_barrier_stub(stub->decorators())));
 617   if (obj != slow_result) {
 618     __ mov(obj, slow_result);
 619   }
 620 
 621   __ jmp(*stub->continuation());
 622 }
 623 
 624 #undef __
 625 
 626 #define __ sasm->
 627 
 628 void ShenandoahBarrierSetAssembler::keepalive_barrier_c1_runtime_stub(StubAssembler* sasm) {
 629   __ prologue("shenandoah_keepalive_barrier", false);
 630   const Register tmp_obj = rax;
 631   const Register tmp = rdx;
 632   __ push(tmp);
 633   __ push(tmp_obj);
 634   __ load_parameter(0, tmp_obj);
 635   satb_barrier(sasm, noreg, tmp_obj, tmp);
 636   __ pop(tmp_obj);
 637   __ pop(tmp);
 638   __ epilogue();
 639 }
 640 
 641 void ShenandoahBarrierSetAssembler::load_reference_barrier_c1_runtime_stub(StubAssembler* sasm, DecoratorSet decorators) {
 642   __ prologue("shenandoah_load_reference_barrier", false);
 643   const Register tmp_obj = rax;
 644   const Register tmp_addr = rdx;
 645   __ push(tmp_addr);
 646   __ load_parameter(0, tmp_obj);
 647   __ load_parameter(1, tmp_addr);
 648   load_reference_barrier(sasm, tmp_obj, Address(tmp_addr, 0), decorators);
 649   __ pop(tmp_addr);
 650   __ epilogue();
 651 }
 652 
 653 #undef __
 654 
 655 #endif // COMPILER1
 656 
 657 #ifdef COMPILER2
 658 
 659 #undef __
 660 #define __ masm->
 661 
 662 void ShenandoahBarrierSetAssembler::load_c2(const MachNode* node, MacroAssembler* masm, Register dst, Address src, bool narrow) {
 663   // Do the actual load. This load is the candidate for implicit null check, and MUST come first.
 664   if (narrow) {
 665     __ movl(dst, src);
 666   } else {
 667     __ movq(dst, src);
 668   }
 669 
 670   ShenandoahBarrierStubC2::load_post(masm, node, dst, src, noreg, noreg, narrow);
 671 }
 672 
 673 void ShenandoahBarrierSetAssembler::store_c2(const MachNode* node, MacroAssembler* masm,
 674                                              Address dst, bool dst_narrow,
 675                                              Register src, bool src_narrow,
 676                                              Register tmp) {
 677 
 678   ShenandoahBarrierStubC2::store_pre(masm, node, dst, tmp, noreg, noreg, dst_narrow);
 679 
 680   // Need to encode into tmp, because we cannot clobber src.
 681   if (dst_narrow && !src_narrow) {
 682     __ movq(tmp, src);
 683     if ((node->barrier_data() & ShenandoahBitNotNull) == 0) {
 684       __ encode_heap_oop(tmp);
 685     } else {
 686       __ encode_heap_oop_not_null(tmp);
 687     }
 688     src = tmp;
 689   }
 690 
 691   // Do the actual store
 692   if (dst_narrow) {
 693     __ movl(dst, src);
 694   } else {
 695     __ movq(dst, src);
 696   }
 697 
 698   ShenandoahBarrierStubC2::store_post(masm, node, dst, tmp, noreg);
 699 }
 700 
 701 void ShenandoahBarrierSetAssembler::compare_and_set_c2(const MachNode* node, MacroAssembler* masm,
 702                                                        Register res, Address addr,
 703                                                        Register oldval, Register newval, Register tmp,
 704                                                        bool narrow) {
 705 
 706   assert(oldval == rax, "must be in rax for implicit use in cmpxchg");
 707 
 708   // Oldval and newval cannot be clobbered by aliasing with tmp.
 709   assert_different_registers(oldval, tmp);
 710   assert_different_registers(newval, tmp);
 711 
 712   ShenandoahBarrierStubC2::load_store_pre(masm, node, addr, tmp, noreg, noreg, narrow);
 713 
 714   // CAS!
 715   __ lock();
 716   if (narrow) {
 717     __ cmpxchgl(newval, addr);
 718   } else {
 719     __ cmpxchgptr(newval, addr);
 720   }
 721 
 722   // If we need a boolean result out of CAS, set the flag appropriately and promote the result.
 723   if (res != noreg) {
 724     __ setcc(Assembler::equal, res);
 725   }
 726 
 727   ShenandoahBarrierStubC2::load_store_post(masm, node, addr, tmp, noreg);
 728 }
 729 
 730 void ShenandoahBarrierSetAssembler::get_and_set_c2(const MachNode* node, MacroAssembler* masm, Register newval, Address addr, Register tmp, bool narrow) {
 731   assert_different_registers(newval, tmp);
 732 
 733   ShenandoahBarrierStubC2::load_store_pre(masm, node, addr, tmp, noreg, noreg, narrow);
 734 
 735   if (narrow) {
 736     __ xchgl(newval, addr);
 737   } else {
 738     __ xchgq(newval, addr);
 739   }
 740 
 741   ShenandoahBarrierStubC2::load_store_post(masm, node, addr, tmp, noreg);
 742 }
 743 
 744 #undef __
 745 #define __ masm.
 746 
 747 void ShenandoahBarrierStubC2::cardtable(MacroAssembler& masm, Address addr, Register tmp1, Register tmp2) {
 748   Assembler::InlineSkippedInstructionsCounter skip_counter(&masm);
 749 
 750   __ lea(tmp1, addr);
 751   __ shrptr(tmp1, CardTable::card_shift());
 752   __ addptr(tmp1, Address(r15_thread, in_bytes(ShenandoahThreadLocalData::card_table_offset())));
 753   Address card_address(tmp1, 0);
 754 
 755   assert(CardTable::dirty_card_val() == 0, "Encoding assumption");
 756   Label L_done;
 757   if (UseCondCardMark) {
 758     __ cmpb(card_address, 0);
 759     __ jccb(Assembler::equal, L_done);
 760   }
 761   if (UseCompressedOops && CompressedOops::base() == nullptr) {
 762     __ movb(card_address, r12);
 763   } else {
 764     __ movb(card_address, 0);
 765   }
 766   __ bind(L_done);
 767 }
 768 
 769 void ShenandoahBarrierStubC2::enter_if_gc_state(MacroAssembler& masm, const char test_state, Register tmp) {
 770   Assembler::InlineSkippedInstructionsCounter skip_counter(&masm);
 771 
 772   Address gc_state_fast(r15_thread, in_bytes(ShenandoahThreadLocalData::gc_state_fast_array_offset(test_state)));
 773   __ cmpb(gc_state_fast, 0);
 774   __ jcc(Assembler::notEqual, *entry());
 775   __ bind(*continuation());
 776 }
 777 
 778 void ShenandoahBarrierStubC2::emit_code(MacroAssembler& masm) {
 779   Assembler::InlineSkippedInstructionsCounter skip_counter(&masm);
 780   assert(_needs_keep_alive_barrier || _needs_load_ref_barrier, "Why are you here?");
 781 
 782   // On x86, there is a significant penalty with unaligned branch target, for example
 783   // when the target instruction straggles the fetch line. It makes (performance) sense
 784   // to spend some code size to align the target better.
 785   __ align(16);
 786   __ bind(*entry());
 787 
 788   // If we need to load ourselves, do it here.
 789   if (_do_load) {
 790     if (_narrow) {
 791       __ movl(_obj, _addr);
 792     } else {
 793       __ movq(_obj, _addr);
 794     }
 795   }
 796 
 797   // If the object is null, there is no point in applying barriers.
 798   maybe_far_jump_if_zero(masm, _obj);
 799 
 800   // We need to make sure that loads done by callers survive across slow-path calls.
 801   // For self-loads, we need to care about the case when both KA and LRB are enabled (rare).
 802   bool needs_both_barriers = _needs_keep_alive_barrier && _needs_load_ref_barrier;
 803   if (!_do_load || needs_both_barriers) {
 804     preserve(_obj);
 805   }
 806 
 807   // Go for barriers. Barriers can return straight to continuation, as long
 808   // as another barrier is not needed.
 809   if (needs_both_barriers) {
 810     keepalive(masm, nullptr);
 811     lrb(masm);
 812   } else if (_needs_keep_alive_barrier) {
 813     keepalive(masm, continuation());
 814   } else if (_needs_load_ref_barrier) {
 815     lrb(masm);
 816   } else {
 817     ShouldNotReachHere();
 818   }
 819 }
 820 
 821 void ShenandoahBarrierStubC2::keepalive(MacroAssembler& masm, Label* L_done) {
 822   Address gc_state_fast(r15_thread, in_bytes(ShenandoahThreadLocalData::gc_state_fast_array_offset(ShenandoahHeap::MARKING)));
 823   Address index(r15_thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
 824   Address buffer(r15_thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
 825 
 826   Label L_through, L_pop_and_slow;
 827 
 828   // If another barrier is enabled as well, do a runtime check for a specific barrier.
 829   if (_needs_load_ref_barrier) {
 830     assert(L_done == nullptr, "L_done is always null when _needs_load_ref_barrier is true");
 831     __ cmpb(gc_state_fast, 0);
 832     __ jcc(Assembler::equal, L_through);
 833   }
 834 
 835   // Need temp to work, allocate one now.
 836   bool tmp_live;
 837   Register tmp = select_temp_register(tmp_live);
 838   if (tmp_live) {
 839     __ push(tmp);
 840   }
 841 
 842   // Fast-path: put object into buffer.
 843   // If buffer is already full, go slow.
 844   __ movptr(tmp, index);
 845   __ subptr(tmp, wordSize);
 846   __ jccb(Assembler::below, L_pop_and_slow);
 847   __ movptr(index, tmp);
 848   __ addptr(tmp, buffer);
 849 
 850   // Store the object in queue.
 851   // If object is narrow, we need to decode it before inserting.
 852   // We can skip the re-encoding if we know that object is not preserved.
 853   if (_narrow) {
 854     __ decode_heap_oop_not_null(_obj);
 855   }
 856   __ movptr(Address(tmp, 0), _obj);
 857   if (_narrow && is_preserved(_obj)) {
 858     __ encode_heap_oop_not_null(_obj);
 859   }
 860 
 861   // Fast-path exits here.
 862   if (tmp_live) {
 863     __ pop(tmp);
 864   }
 865 
 866   if (L_done != nullptr) {
 867     __ jmp(*L_done);
 868   } else {
 869     __ jmp(L_through);
 870   }
 871 
 872   // Slow-path: call runtime to handle.
 873   // Need to pop tmp immediately for stack to remain aligned.
 874   __ bind(L_pop_and_slow);
 875   if (tmp_live) {
 876     __ pop(tmp);
 877   }
 878   {
 879     SaveLiveRegisters slr(&masm, this);
 880 
 881     // Shuffle in the arguments. The end result should be:
 882     //   c_rarg0 <-- obj
 883     if (c_rarg0 != _obj) {
 884       __ mov(c_rarg0, _obj);
 885     }
 886 
 887     // Go to runtime and handle the rest there.
 888     // Use rax as scratch, as it will be saved if live.
 889     __ call(RuntimeAddress(keepalive_runtime_entry_addr()), rax);
 890   }
 891   if (L_done != nullptr) {
 892     __ jmp(*L_done);
 893   } else {
 894     __ bind(L_through);
 895   }
 896 }
 897 
 898 void ShenandoahBarrierStubC2::lrb(MacroAssembler& masm) {
 899   Label L_pop_and_slow, L_slow;
 900 
 901   // If another barrier is enabled as well, do a runtime check for a specific barrier.
 902   if (_needs_keep_alive_barrier) {
 903     char state_to_check = ShenandoahHeap::HAS_FORWARDED | (_needs_load_ref_weak_barrier ? ShenandoahHeap::WEAK_ROOTS : 0);
 904     Address gc_state_fast(r15_thread, in_bytes(ShenandoahThreadLocalData::gc_state_fast_array_offset(state_to_check)));
 905     __ cmpb(gc_state_fast, 0);
 906     __ jcc(Assembler::equal, *continuation());
 907   }
 908 
 909   // If weak references are being processed, weak/phantom loads need to go slow,
 910   // regardless of their cset status.
 911   if (_needs_load_ref_weak_barrier) {
 912     Address gc_state_fast(r15_thread, in_bytes(ShenandoahThreadLocalData::gc_state_fast_array_offset(ShenandoahHeap::WEAK_ROOTS)));
 913     __ cmpb(gc_state_fast, 0);
 914     __ jccb(Assembler::notEqual, L_slow);
 915   }
 916 
 917   bool is_aot = AOTCodeCache::is_on_for_dump();
 918 
 919   // Need temp to work, allocate one now.
 920   bool tmp_live;
 921   Register tmp = select_temp_register(tmp_live, /* skip_reg1 = */ is_aot ? rcx : noreg);
 922   if (tmp_live) {
 923     __ push(tmp);
 924   }
 925 
 926   // Compute the cset bitmap index
 927   if (_narrow) {
 928     __ decode_heap_oop_not_null(tmp, _obj);
 929   } else {
 930     __ movptr(tmp, _obj);
 931   }
 932 
 933   Address cset_addr_arg;
 934   intptr_t cset_addr = reinterpret_cast<intptr_t>(ShenandoahHeap::in_cset_fast_test_addr());
 935   if (!is_aot && cset_addr < INT32_MAX) {
 936     // Cset bitmap is at easily encodeable address. Just use it as displacement.
 937     __ shrptr(tmp, ShenandoahHeapRegion::region_size_bytes_shift_jint());
 938     cset_addr_arg = Address(tmp, checked_cast<int>(cset_addr));
 939   } else {
 940     bool tmp2_live;
 941     Register tmp2 = select_temp_register(tmp2_live, /* skip_reg1 = */ tmp, /* skip_reg2 = */ is_aot ? rcx : noreg);
 942     if (tmp2_live) {
 943       __ push(tmp2);
 944     }
 945     if (is_aot) {
 946       // Generating AOT code, pull the cset bitmap and region shift from AOT table.
 947       assert_different_registers(tmp, tmp2, rcx);
 948       __ push(rcx);
 949       __ lea(rcx, ExternalAddress(AOTRuntimeConstants::grain_shift_address()));
 950       __ movl(rcx, Address(rcx));
 951       __ shrptr(tmp);
 952       __ pop(rcx);
 953       __ lea(tmp2, ExternalAddress(AOTRuntimeConstants::cset_base_address()));
 954       __ addptr(tmp, Address(tmp2));
 955     } else {
 956       // Cset bitmap is far away. Add its address fully.
 957       __ shrptr(tmp, ShenandoahHeapRegion::region_size_bytes_shift_jint());
 958       __ movptr(tmp2, cset_addr);
 959       __ addptr(tmp, tmp2);
 960     }
 961     if (tmp2_live) {
 962       __ pop(tmp2);
 963     }
 964     cset_addr_arg = Address(tmp, 0);
 965   }
 966 
 967   // Cset-check. Fall-through to slow if in collection set.
 968   __ cmpb(cset_addr_arg, 0);
 969   if (tmp_live) {
 970     __ jccb(Assembler::notEqual, L_pop_and_slow);
 971     __ pop(tmp);
 972     __ jmp(*continuation());
 973   } else {
 974     // Nothing else to do, jump back
 975     __ jcc(Assembler::equal, *continuation());
 976   }
 977 
 978   // Slow path
 979   __ bind(L_pop_and_slow);
 980   // Need to pop tmp immediately for stack to remain aligned.
 981   if (tmp_live) {
 982     __ pop(tmp);
 983   }
 984   __ bind(L_slow);
 985 
 986   // Obj is the result, need to temporarily stop preserving it.
 987   bool is_obj_preserved = is_preserved(_obj);
 988   if (is_obj_preserved) {
 989     dont_preserve(_obj);
 990   }
 991   {
 992     SaveLiveRegisters slr(&masm, this);
 993 
 994     assert_different_registers(rax, c_rarg0, c_rarg1);
 995 
 996     // Shuffle in the arguments. The end result should be:
 997     //   c_rarg0 <-- obj
 998     //   c_rarg1 <-- lea(addr)
 999     if (_obj == c_rarg0) {
1000       __ lea(c_rarg1, _addr);
1001     } else if (_obj == c_rarg1) {
1002       // Set up arguments in reverse, and then flip them
1003       __ lea(c_rarg0, _addr);
1004       __ xchgptr(c_rarg0, c_rarg1);
1005     } else {
1006       assert_different_registers(_obj, c_rarg0, c_rarg1);
1007       __ lea(c_rarg1, _addr);
1008       __ movptr(c_rarg0, _obj);
1009     }
1010 
1011     // Go to runtime and handle the rest there.
1012     // Use rax as scratch, as it will be clobbered by result anyway.
1013     __ call(RuntimeAddress(lrb_runtime_entry_addr()), rax);
1014 
1015     // Save the result where needed.
1016     if (_narrow) {
1017       __ movl(_obj, rax);
1018     } else if (_obj != rax) {
1019       __ movptr(_obj, rax);
1020     }
1021   }
1022   if (is_obj_preserved) {
1023     preserve(_obj);
1024   }
1025 
1026   __ jmp(*continuation());
1027 }
1028 
1029 int ShenandoahBarrierStubC2::available_gp_registers() {
1030   return Register::available_gp_registers();
1031 }
1032 
1033 bool ShenandoahBarrierStubC2::is_special_register(Register r) {
1034   return r == rsp || r == rbp || r == r12_heapbase || r == r15_thread;
1035 }
1036 
1037 void ShenandoahBarrierStubC2::post_init() {
1038   // Do nothing.
1039 }
1040 
1041 void ShenandoahBarrierStubC2::maybe_far_jump_if_zero(MacroAssembler& masm, Register reg) {
1042   if (_narrow) {
1043     __ testl(reg, reg);
1044   } else {
1045     __ testq(reg, reg);
1046   }
1047   __ jcc(Assembler::zero, *continuation());
1048 }
1049 
1050 #endif // COMPILER2