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 "code/aotCodeCache.hpp"
  28 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
  29 #include "gc/shenandoah/mode/shenandoahMode.hpp"
  30 #include "gc/shenandoah/shenandoahBarrierSet.hpp"
  31 #include "gc/shenandoah/shenandoahBarrierSetAssembler.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       assert(!saved_regs.contains(rscratch1), "Sanity: about to clobber rscratch1");
  65       assert(!saved_regs.contains(rscratch2), "Sanity: about to clobber rscratch2");
  66       Address gc_state(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
  67       __ ldrb(rscratch1, gc_state);
  68       if (ShenandoahSATBBarrier && dest_uninitialized) {
  69         __ tbz(rscratch1, ShenandoahHeap::HAS_FORWARDED_BITPOS, done);
  70       } else {
  71         __ mov(rscratch2, ShenandoahHeap::HAS_FORWARDED | ShenandoahHeap::MARKING);
  72         __ tst(rscratch1, rscratch2);
  73         __ br(Assembler::EQ, done);
  74       }
  75 
  76       __ push_call_clobbered_registers();
  77       // If arguments are not in proper places, shuffle them.
  78       // Doing this via the stack is the most straight-forward way to avoid
  79       // accidentally smashing any register.
  80       if (c_rarg0 != src || c_rarg1 != dst || c_rarg2 != count) {
  81         __ push(RegSet::of(src), sp);
  82         __ push(RegSet::of(dst), sp);
  83         __ push(RegSet::of(count), sp);
  84         __ pop(RegSet::of(c_rarg2), sp);
  85         __ pop(RegSet::of(c_rarg1), sp);
  86         __ pop(RegSet::of(c_rarg0), sp);
  87       }
  88       address target = nullptr;
  89       if (UseCompressedOops) {
  90         target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::arraycopy_barrier_narrow_oop);
  91       } else {
  92         target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::arraycopy_barrier_oop);
  93       }
  94       __ call_VM_leaf(target, 3);
  95       __ pop_call_clobbered_registers();
  96       __ bind(done);
  97     }
  98   }
  99 }
 100 
 101 void ShenandoahBarrierSetAssembler::arraycopy_epilogue(MacroAssembler* masm, DecoratorSet decorators, bool is_oop,
 102                                                        Register start, Register count, Register tmp) {
 103   if (ShenandoahCardBarrier && is_oop) {
 104     gen_write_ref_array_post_barrier(masm, decorators, start, count, tmp);
 105   }
 106 }
 107 
 108 void ShenandoahBarrierSetAssembler::satb_barrier(MacroAssembler* masm,
 109                                                  Register obj,
 110                                                  Register pre_val,
 111                                                  Register thread,
 112                                                  Register tmp1,
 113                                                  Register tmp2) {
 114   assert(ShenandoahSATBBarrier, "Should be checked by caller");
 115   assert(thread == rthread, "must be");
 116 
 117   Label done;
 118   Label runtime;
 119 
 120   assert_different_registers(obj, pre_val, tmp1, tmp2);
 121   assert(pre_val != noreg && tmp1 != noreg && tmp2 != noreg, "expecting a register");
 122 
 123   Address index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
 124   Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
 125 
 126   // Is marking active?
 127   Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 128   __ ldrb(tmp1, gc_state);
 129   __ tbz(tmp1, ShenandoahHeap::MARKING_BITPOS, done);
 130 
 131   // Do we need to load the previous value?
 132   if (obj != noreg) {
 133     if (UseCompressedOops) {
 134       __ ldrw(pre_val, Address(obj, 0));
 135       __ decode_heap_oop(pre_val);
 136     } else {
 137       __ ldr(pre_val, Address(obj, 0));
 138     }
 139   }
 140 
 141   // Is the previous value null?
 142   __ cbz(pre_val, done);
 143 
 144   // Can we store original value in the thread's buffer?
 145   // Is index == 0?
 146   // (The index field is typed as size_t.)
 147 
 148   __ ldr(tmp1, index);                      // tmp := *index_adr
 149   __ cbz(tmp1, runtime);                    // tmp == 0?
 150                                         // If yes, goto runtime
 151 
 152   __ sub(tmp1, tmp1, wordSize);             // tmp := tmp - wordSize
 153   __ str(tmp1, index);                      // *index_adr := tmp
 154   __ ldr(tmp2, buffer);
 155   __ add(tmp1, tmp1, tmp2);                 // tmp := tmp + *buffer_adr
 156 
 157   // Record the previous value
 158   __ str(pre_val, Address(tmp1, 0));
 159   __ b(done);
 160 
 161   __ bind(runtime);
 162 
 163   // Slow-path call
 164   __ enter(/* strip_ret_addr = */ true);
 165   __ push_call_clobbered_registers();
 166   if (c_rarg0 != pre_val) {
 167     __ mov(c_rarg0, pre_val);
 168   }
 169   // Calling with super_call_VM_leaf with c_rarg0 bypasses interpreter checks and avoids any moves.
 170   __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_pre), c_rarg0);
 171   __ pop_call_clobbered_registers();
 172   __ leave();
 173 
 174   __ bind(done);
 175 }
 176 
 177 void ShenandoahBarrierSetAssembler::load_reference_barrier(MacroAssembler* masm, Register dst, Address load_addr, DecoratorSet decorators) {
 178   assert(ShenandoahLoadRefBarrier, "Should be enabled");
 179   assert(dst != rscratch2, "need rscratch2");
 180   assert_different_registers(load_addr.base(), load_addr.index(), rscratch1, rscratch2);
 181 
 182   bool is_strong  = ShenandoahBarrierSet::is_strong_access(decorators);
 183   bool is_weak    = ShenandoahBarrierSet::is_weak_access(decorators);
 184   bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
 185   bool is_native  = ShenandoahBarrierSet::is_native_access(decorators);
 186   bool is_narrow  = UseCompressedOops && !is_native;
 187 
 188   Label heap_stable, not_cset;
 189   Address gc_state(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 190   __ ldrb(rscratch2, gc_state);
 191 
 192   // Check for heap stability
 193   if (is_strong) {
 194     __ tbz(rscratch2, ShenandoahHeap::HAS_FORWARDED_BITPOS, heap_stable);
 195   } else {
 196     Label lrb;
 197     __ tbnz(rscratch2, ShenandoahHeap::WEAK_ROOTS_BITPOS, lrb);
 198     __ tbz(rscratch2, ShenandoahHeap::HAS_FORWARDED_BITPOS, heap_stable);
 199     __ bind(lrb);
 200   }
 201 
 202   // use r1 for load address
 203   Register result_dst = dst;
 204   if (dst == r1) {
 205     __ mov(rscratch1, dst);
 206     dst = rscratch1;
 207   }
 208 
 209   // Save r0 and r1, unless it is an output register
 210   RegSet to_save = RegSet::of(r0, r1) - result_dst;
 211   __ push(to_save, sp);
 212   __ lea(r1, load_addr);
 213   __ mov(r0, dst);
 214 
 215   // Test for in-cset
 216   if (is_strong) {
 217 #if INCLUDE_CDS
 218     if (AOTCodeCache::is_on_for_dump()) {
 219       __ lea(rscratch2, ExternalAddress(AOTRuntimeConstants::cset_base_address()));
 220       __ ldr(rscratch2, Address(rscratch2));
 221       __ lea(rscratch1, ExternalAddress(AOTRuntimeConstants::grain_shift_address()));
 222       __ ldrw(rscratch1, Address(rscratch1));
 223       __ lsrv(rscratch1, r0, rscratch1);
 224     } else
 225 #endif
 226     {
 227       __ mov(rscratch2, ShenandoahHeap::in_cset_fast_test_addr());
 228       __ lsr(rscratch1, r0, ShenandoahHeapRegion::region_size_bytes_shift_jint());
 229     }
 230     __ ldrb(rscratch2, Address(rscratch2, rscratch1));
 231     __ tbz(rscratch2, 0, not_cset);
 232   }
 233 
 234   // Slow-path call
 235   __ enter(/* strip_ret_addr = */ true);
 236   __ push_call_clobbered_registers();
 237   address target = nullptr;
 238   if (is_strong) {
 239     if (is_narrow) {
 240       target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow);
 241     } else {
 242       target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong);
 243     }
 244   } else if (is_weak) {
 245     if (is_narrow) {
 246       target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow);
 247     } else {
 248       target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak);
 249     }
 250   } else {
 251     assert(is_phantom, "only remaining strength");
 252     assert(!is_narrow, "phantom access cannot be narrow");
 253     target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom);
 254   }
 255   // Calling with super_call_VM_leaf with c_rarg0/1 bypasses interpreter checks and avoids any moves.
 256   __ super_call_VM_leaf(target, c_rarg0, c_rarg1);
 257   __ mov(rscratch1, r0);
 258   __ pop_call_clobbered_registers();
 259   __ mov(r0, rscratch1);
 260   __ leave();
 261 
 262   __ bind(not_cset);
 263 
 264   __ mov(result_dst, r0);
 265   __ pop(to_save, sp);
 266 
 267   __ bind(heap_stable);
 268 }
 269 
 270 //
 271 // Arguments:
 272 //
 273 // Inputs:
 274 //   src:        oop location to load from, might be clobbered
 275 //
 276 // Output:
 277 //   dst:        oop loaded from src location
 278 //
 279 // Kill:
 280 //   rscratch1 (scratch reg)
 281 //
 282 // Alias:
 283 //   dst: rscratch1 (might use rscratch1 as temporary output register to avoid clobbering src)
 284 //
 285 void ShenandoahBarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 286                                             Register dst, Address src, Register tmp1, Register tmp2) {
 287   // 1: non-reference load, no additional barrier is needed
 288   if (!is_reference_type(type)) {
 289     BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp2);
 290     return;
 291   }
 292 
 293   // 2: load a reference from src location and apply LRB if needed
 294   if (ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) {
 295     Register result_dst = dst;
 296 
 297     // Preserve src location for LRB
 298     if (dst == src.base() || dst == src.index()) {
 299       dst = rscratch1;
 300     }
 301     assert_different_registers(dst, src.base(), src.index());
 302 
 303     BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp2);
 304 
 305     load_reference_barrier(masm, dst, src, decorators);
 306 
 307     if (dst != result_dst) {
 308       __ mov(result_dst, dst);
 309       dst = result_dst;
 310     }
 311   } else {
 312     BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp2);
 313   }
 314 
 315   // 3: apply keep-alive barrier if needed
 316   if (ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) {
 317     satb_barrier(masm /* masm */,
 318                  noreg /* obj */,
 319                  dst /* pre_val */,
 320                  rthread /* thread */,
 321                  tmp1 /* tmp1 */,
 322                  tmp2 /* tmp2 */);
 323   }
 324 }
 325 
 326 void ShenandoahBarrierSetAssembler::card_barrier(MacroAssembler* masm, Register obj, Register tmp1, Register tmp2) {
 327   assert(ShenandoahCardBarrier, "Should have been checked by caller");
 328   assert_different_registers(obj, tmp1, tmp2);
 329   assert(CardTable::dirty_card_val() == 0, "must be");
 330 
 331   __ lsr(obj, obj, CardTable::card_shift());
 332 
 333   Address curr_ct_holder_addr(rthread, in_bytes(ShenandoahThreadLocalData::card_table_offset()));
 334   __ ldr(tmp1, curr_ct_holder_addr);
 335 
 336   if (UseCondCardMark) {
 337     Label L_already_dirty;
 338     __ ldrb(tmp2, Address(obj, tmp1));
 339     __ cbz(tmp2, L_already_dirty);
 340     __ strb(zr, Address(obj, tmp1));
 341     __ bind(L_already_dirty);
 342   } else {
 343     __ strb(zr, Address(obj, tmp1));
 344   }
 345 }
 346 
 347 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 348                                              Address dst, Register val, Register tmp1, Register tmp2, Register tmp3) {
 349   // 1: non-reference types require no barriers
 350   if (!is_reference_type(type)) {
 351     BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2, tmp3);
 352     return;
 353   }
 354 
 355   // Flatten object address right away for simplicity: likely needed by barriers
 356   if (dst.index() == noreg && dst.offset() == 0) {
 357     if (dst.base() != tmp3) {
 358       __ mov(tmp3, dst.base());
 359     }
 360   } else {
 361     __ lea(tmp3, dst);
 362   }
 363 
 364   // 2: pre-barrier: SATB needs the previous value
 365   if (ShenandoahBarrierSet::need_satb_barrier(decorators, type)) {
 366     satb_barrier(masm,
 367                  tmp3 /* obj */,
 368                  tmp2 /* pre_val */,
 369                  rthread /* thread */,
 370                  tmp1 /* tmp */,
 371                  rscratch1 /* tmp2 */);
 372   }
 373 
 374   // Store!
 375   BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp3, 0), val, noreg, noreg, noreg);
 376 
 377   // 3: post-barrier: card barrier needs store address
 378   bool storing_non_null = (val != noreg);
 379   if (ShenandoahBarrierSet::need_card_barrier(decorators, type) && storing_non_null) {
 380     card_barrier(masm, tmp3, tmp1, tmp2);
 381   }
 382 }
 383 
 384 void ShenandoahBarrierSetAssembler::try_resolve_jobject_in_native(MacroAssembler* masm, Register jni_env,
 385                                                                   Register obj, Register tmp, Label& slowpath) {
 386   Label done;
 387   // Resolve jobject
 388   BarrierSetAssembler::try_resolve_jobject_in_native(masm, jni_env, obj, tmp, slowpath);
 389 
 390   // Check for null.
 391   __ cbz(obj, done);
 392 
 393   assert(obj != rscratch2, "need rscratch2");
 394   Address gc_state(jni_env, ShenandoahThreadLocalData::gc_state_offset() - JavaThread::jni_environment_offset());
 395   __ lea(rscratch2, gc_state);
 396   __ ldrb(rscratch2, Address(rscratch2));
 397 
 398   // Check for heap in evacuation phase
 399   __ tbnz(rscratch2, ShenandoahHeap::EVACUATION_BITPOS, slowpath);
 400 
 401   __ bind(done);
 402 }
 403 
 404 void ShenandoahBarrierSetAssembler::try_peek_weak_handle_in_nmethod(MacroAssembler* masm, Register weak_handle, Register obj,
 405                                                                     Register tmp, Label& slow_path) {
 406   assert_different_registers(weak_handle, tmp, noreg);
 407   assert_different_registers(obj, tmp, noreg);
 408 
 409   Label done;
 410 
 411   // Peek weak handle using the standard implementation.
 412   BarrierSetAssembler::try_peek_weak_handle_in_nmethod(masm, weak_handle, obj, tmp, slow_path);
 413 
 414   // Check if the reference is null, and if it is, take the fast path.
 415   __ cbz(obj, done);
 416 
 417   Address gc_state(rthread, ShenandoahThreadLocalData::gc_state_offset());
 418   __ lea(tmp, gc_state);
 419   __ ldrb(tmp, __ legitimize_address(gc_state, 1, tmp));
 420 
 421   // Check if the heap is under weak-reference/roots processing, in
 422   // which case we need to take the slow path.
 423   __ tbnz(tmp, ShenandoahHeap::WEAK_ROOTS_BITPOS, slow_path);
 424   __ bind(done);
 425 }
 426 
 427 void ShenandoahBarrierSetAssembler::check_oop(MacroAssembler* masm, Register obj, Register tmp1, Register tmp2, Label& L_error) {
 428   assert_different_registers(obj, tmp1, tmp2);
 429   // Check if the oop is in the right area of memory
 430 #if INCLUDE_CDS
 431   if (AOTCodeCache::is_on_for_dump()) {
 432     __ lea(tmp2, ExternalAddress(AOTRuntimeConstants::verify_oop_mask_address()));
 433     __ ldr(tmp2, Address(tmp2));
 434     __ andr(tmp1, obj, tmp2);
 435     __ lea(tmp2, ExternalAddress(AOTRuntimeConstants::verify_oop_bits_address()));
 436     __ ldr(tmp2, Address(tmp2));
 437   } else
 438 #endif
 439   {
 440     __ mov(tmp2, (intptr_t) Universe::verify_oop_mask());
 441     __ andr(tmp1, obj, tmp2);
 442     __ mov(tmp2, (intptr_t) Universe::verify_oop_bits());
 443   }
 444   // Compare tmp1 and tmp2.  We don't use a compare
 445   // instruction here because the flags register is live.
 446   __ eor(tmp1, tmp1, tmp2);
 447   __ cbnz(tmp1, L_error);
 448 
 449   // This routine is sometimes called before applying GC barriers.
 450   // With +COH, loading the klass may end up loading forwarding pointer instead.
 451   Label L_skip;
 452   if (UseCompactObjectHeaders) {
 453     Address gc_state(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 454     __ ldrb(tmp1, gc_state);
 455     __ tbnz(tmp1, ShenandoahHeap::HAS_FORWARDED_BITPOS, L_skip);
 456   }
 457 
 458   // Make sure klass is 'reasonable', which is not zero.
 459   __ load_narrow_klass(tmp1, obj);
 460   __ cbz(tmp1, L_error);
 461   __ bind(L_skip);
 462 }
 463 
 464 void ShenandoahBarrierSetAssembler::gen_write_ref_array_post_barrier(MacroAssembler* masm, DecoratorSet decorators,
 465                                                                      Register start, Register count, Register scratch) {
 466   assert(ShenandoahCardBarrier, "Should have been checked by caller");
 467 
 468   Label L_loop, L_done;
 469   const Register end = count;
 470 
 471   // Zero count? Nothing to do.
 472   __ cbz(count, L_done);
 473 
 474   // end = start + count << LogBytesPerHeapOop
 475   // last element address to make inclusive
 476   __ lea(end, Address(start, count, Address::lsl(LogBytesPerHeapOop)));
 477   __ sub(end, end, BytesPerHeapOop);
 478   __ lsr(start, start, CardTable::card_shift());
 479   __ lsr(end, end, CardTable::card_shift());
 480 
 481   // number of bytes to copy
 482   __ sub(count, end, start);
 483 
 484   Address curr_ct_holder_addr(rthread, in_bytes(ShenandoahThreadLocalData::card_table_offset()));
 485   __ ldr(scratch, curr_ct_holder_addr);
 486   __ add(start, start, scratch);
 487   __ bind(L_loop);
 488   __ strb(zr, Address(start, count));
 489   __ subs(count, count, 1);
 490   __ br(Assembler::GE, L_loop);
 491   __ bind(L_done);
 492 }
 493 
 494 #undef __
 495 
 496 #ifdef COMPILER1
 497 
 498 #define __ ce->masm()->
 499 
 500 void ShenandoahBarrierSetAssembler::keepalive_barrier_c1_stub(LIR_Assembler* ce, ShenandoahKeepaliveBarrierStub* stub) {
 501   __ bind(*stub->entry());
 502 
 503   ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
 504 
 505   Register obj = stub->obj()->as_register();
 506 
 507   if (stub->do_load()) {
 508     ce->mem2reg(stub->addr(), stub->obj(), T_OBJECT, lir_patch_none, nullptr, /* wide = */ false);
 509   }
 510   __ cbz(obj, *stub->continuation());
 511   ce->store_parameter(obj, 0);
 512   __ far_call(RuntimeAddress(bs->keepalive_barrier_stub()));
 513   __ b(*stub->continuation());
 514 }
 515 
 516 void ShenandoahBarrierSetAssembler::load_reference_barrier_c1_stub(LIR_Assembler* ce, ShenandoahLoadReferenceBarrierStub* stub) {
 517   __ bind(*stub->entry());
 518 
 519   ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
 520 
 521   Register obj = stub->obj()->as_register();
 522   Register addr = stub->addr()->as_pointer_register();
 523   Register slow_result = stub->slow_result()->as_register();
 524   assert_different_registers(obj, addr, slow_result);
 525   assert(slow_result == r0, "C1 must know about our slow call result register");
 526 
 527   ce->store_parameter(obj, 0);
 528   ce->store_parameter(addr, 1);
 529   __ far_call(RuntimeAddress(bs->load_reference_barrier_stub(stub->decorators())));
 530   if (obj != slow_result) {
 531     __ mov(obj, slow_result);
 532   }
 533 
 534   __ b(*stub->continuation());
 535 }
 536 
 537 #undef __
 538 
 539 #define __ sasm->
 540 
 541 void ShenandoahBarrierSetAssembler::keepalive_barrier_c1_runtime_stub(StubAssembler* sasm) {
 542   __ prologue("shenandoah_keepalive_barrier", false);
 543   const Register tmp_obj = r0;
 544   const Register tmp1 = r1;
 545   const Register tmp2 = r2;
 546   __ push(RegSet::of(tmp1, tmp2, tmp_obj), sp);
 547   __ load_parameter(0, tmp_obj);
 548   satb_barrier(sasm, noreg, tmp_obj, rthread, tmp1, tmp2);
 549   __ pop(RegSet::of(tmp1, tmp2, tmp_obj), sp);
 550   __ epilogue();
 551 }
 552 
 553 void ShenandoahBarrierSetAssembler::load_reference_barrier_c1_runtime_stub(StubAssembler* sasm, DecoratorSet decorators) {
 554   __ prologue("shenandoah_load_reference_barrier", false);
 555   const Register tmp_obj = r0;
 556   const Register tmp_addr = r1;
 557   __ push(RegSet::of(tmp_addr), sp);
 558   __ load_parameter(0, tmp_obj);
 559   __ load_parameter(1, tmp_addr);
 560   load_reference_barrier(sasm, tmp_obj, Address(tmp_addr, 0), decorators);
 561   __ pop(RegSet::of(tmp_addr), sp);
 562   __ epilogue();
 563 }
 564 
 565 #undef __
 566 
 567 #endif // COMPILER1
 568 
 569 #ifdef COMPILER2
 570 
 571 #undef __
 572 #define __ masm->
 573 
 574 
 575 void ShenandoahBarrierSetAssembler::load_c2(const MachNode* node, MacroAssembler* masm, Register dst, Address src, Register tmp1, Register tmp2, bool is_narrow, bool is_acquire) {
 576   // Do the actual load. This load is the candidate for implicit null check, and MUST come first.
 577   if (is_narrow) {
 578     if (is_acquire) {
 579       assert(src.getMode() == Address::base_plus_offset && src.offset() == 0,
 580           "is_acquire path requires address to be base-only");
 581       __ ldarw(dst, src.base());
 582     } else {
 583       __ ldrw(dst, src);
 584     }
 585   } else {
 586     if (is_acquire) {
 587       assert(src.getMode() == Address::base_plus_offset && src.offset() == 0,
 588           "is_acquire path requires address to be base-only");
 589       __ ldar(dst, src.base());
 590     } else {
 591       __ ldr(dst, src);
 592     }
 593   }
 594 
 595   ShenandoahBarrierStubC2::load_post(masm, node, dst, src, tmp1, tmp2, is_narrow);
 596 }
 597 
 598 void ShenandoahBarrierSetAssembler::store_c2(const MachNode* node, MacroAssembler* masm, Address dst, bool dst_narrow,
 599     Register src, bool src_narrow, Register tmp1, Register tmp2, Register tmp3, bool is_volatile) {
 600 
 601   ShenandoahBarrierStubC2::store_pre(masm, node, dst, tmp1, tmp2, tmp3, dst_narrow);
 602 
 603   // Do the actual store
 604   if (dst_narrow) {
 605     if (!src_narrow) {
 606       // Need to encode into rscratch, because we cannot clobber src.
 607       if ((node->barrier_data() & ShenandoahBitNotNull) == 0) {
 608         __ encode_heap_oop(tmp2, src);
 609       } else {
 610         __ encode_heap_oop_not_null(tmp2, src);
 611       }
 612       src = tmp2;
 613     }
 614 
 615     if (is_volatile) {
 616       assert(dst.getMode() == Address::base_plus_offset && dst.offset() == 0,
 617           "is_acquire path requires address to be base-only");
 618       __ stlrw(src, dst.base());
 619     } else {
 620       __ strw(src, dst);
 621     }
 622   } else {
 623     if (is_volatile) {
 624       assert(dst.getMode() == Address::base_plus_offset && dst.offset() == 0,
 625           "is_acquire path requires address to be base-only");
 626       __ stlr(src, dst.base());
 627     } else {
 628       __ str(src, dst);
 629     }
 630   }
 631 
 632   ShenandoahBarrierStubC2::store_post(masm, node, dst, tmp2, tmp3);
 633 }
 634 
 635 void ShenandoahBarrierSetAssembler::compare_and_set_c2(const MachNode* node, MacroAssembler* masm, Register res, Register addr,
 636     Register oldval, Register newval, Register tmp1, Register tmp2, Register tmp3, bool exchange, bool narrow, bool weak, bool acquire) {
 637   Assembler::operand_size op_size = narrow ? Assembler::word : Assembler::xword;
 638 
 639   ShenandoahBarrierStubC2::load_store_pre(masm, node, addr, tmp1, tmp2, tmp3, narrow);
 640 
 641   atomic_memory_order order = acquire ? memory_order_seq_cst : memory_order_release;
 642 
 643   // CAS!
 644   if (weak) {
 645     __ cmpxchg_weak(addr, oldval, newval, op_size, order, exchange ? res : noreg);
 646   } else {
 647     __ cmpxchg(addr, oldval, newval, op_size, order, exchange ? res : noreg);
 648   }
 649 
 650   // If we need a boolean result out of CAS, set the flag appropriately and promote the result.
 651   if (!exchange) {
 652     assert(res != noreg, "need result register");
 653     __ cset(res, Assembler::EQ);
 654   }
 655 
 656   ShenandoahBarrierStubC2::load_store_post(masm, node, Address(addr, 0), tmp2, tmp3);
 657 }
 658 
 659 void ShenandoahBarrierSetAssembler::get_and_set_c2(const MachNode* node, MacroAssembler* masm, Register preval,
 660     Register newval, Register addr, Register tmp1, Register tmp2, Register tmp3, bool is_acquire) {
 661   bool is_narrow = node->bottom_type()->isa_narrowoop();
 662 
 663   ShenandoahBarrierStubC2::load_store_pre(masm, node, addr, tmp1, tmp2, tmp3, is_narrow);
 664 
 665   if (is_narrow) {
 666     if (is_acquire) {
 667       __ atomic_xchgalw(preval, newval, addr);
 668     } else {
 669       __ atomic_xchgw(preval, newval, addr);
 670     }
 671   } else {
 672     if (is_acquire) {
 673       __ atomic_xchgal(preval, newval, addr);
 674     } else {
 675       __ atomic_xchg(preval, newval, addr);
 676     }
 677   }
 678 
 679   ShenandoahBarrierStubC2::load_store_post(masm, node, Address(addr, 0), tmp2, tmp3);
 680 }
 681 
 682 #undef __
 683 #define __ masm.
 684 
 685 void ShenandoahBarrierStubC2::cardtable(MacroAssembler& masm, Address address, Register tmp1, Register tmp2) {
 686   assert(CardTable::dirty_card_val() == 0, "must be");
 687   Assembler::InlineSkippedInstructionsCounter skip_counter(&masm);
 688 
 689   // tmp1 = card table base (holder)
 690   Address curr_ct_holder_addr(rthread, in_bytes(ShenandoahThreadLocalData::card_table_offset()));
 691   __ ldr(tmp1, curr_ct_holder_addr);
 692 
 693   // tmp2 = effective address
 694   __ lea(tmp2, address);
 695 
 696   // tmp2 = &card_table[ addr >> CardTable::card_shift() ] ; card index
 697   __ add(tmp2, tmp1, tmp2, Assembler::LSR, CardTable::card_shift());
 698 
 699   if (UseCondCardMark) {
 700     Label L_already_dirty;
 701     __ ldrb(tmp1, Address(tmp2));
 702     __ cbz(tmp1, L_already_dirty);
 703     __ strb(zr, Address(tmp2));
 704     __ bind(L_already_dirty);
 705   } else {
 706     __ strb(zr, Address(tmp2));
 707   }
 708 }
 709 
 710 void ShenandoahBarrierStubC2::enter_if_gc_state(MacroAssembler& masm, const char test_state, Register tmp) {
 711   Assembler::InlineSkippedInstructionsCounter skip_counter(&masm);
 712   PhaseOutput* const output = Compile::current()->output();
 713   Address gc_state_fast(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_fast_array_offset(test_state)));
 714 
 715   // We piggyback on scratch_emit_size mode to compute the slowpath stub size.
 716   // We'll use that information to decide whether we need a far jump to the
 717   // stub entry point or not. In scratch_emit_size mode we don't bind entry()
 718   // because otherwise it will be rebound when we later emit the instructions
 719   // for real.
 720   if (_needs_far_jump) {
 721     __ ldrb(tmp, gc_state_fast);
 722     __ cbz(tmp, *continuation());
 723     __ b(output->in_scratch_emit_size() ? *continuation() : *entry());
 724   } else {
 725     __ ldrb(tmp, gc_state_fast);
 726     __ cbnz(tmp, output->in_scratch_emit_size() ? *continuation() : *entry());
 727   }
 728 
 729   // This is were the slowpath stub will return to or the code above will
 730   // jump to if the checks are false
 731   __ bind(*continuation());
 732 }
 733 
 734 void ShenandoahBarrierStubC2::emit_code(MacroAssembler& masm) {
 735   Assembler::InlineSkippedInstructionsCounter skip_counter(&masm);
 736   assert(_needs_keep_alive_barrier || _needs_load_ref_barrier, "Why are you here?");
 737   PhaseOutput* const output = Compile::current()->output();
 738 
 739   // We piggyback on scratch_emit_size mode to compute the slowpath stub size.
 740   // We'll use that information to decide whether we need a far jump to the
 741   // stub entry point or not. In scratch_emit_size mode we don't bind entry()
 742   // because otherwise it will be rebound when we later emit the instructions
 743   // for real.
 744   if (!output->in_scratch_emit_size()) {
 745     __ bind(*entry());
 746   }
 747 
 748   // If we need to load ourselves, do it here.
 749   if (_do_load) {
 750     if (_narrow) {
 751       __ ldrw(_obj, _addr);
 752     } else {
 753       __ ldr(_obj, _addr);
 754     }
 755   }
 756 
 757   // If the object is null, there is no point in applying barriers.
 758   maybe_far_jump_if_zero(masm, _obj);
 759 
 760   // We need to make sure that loads done by callers survive across slow-path calls.
 761   // For self-loads, we need to care about the case when both KA and LRB are enabled (rare).
 762   bool needs_both_barriers = _needs_keep_alive_barrier && _needs_load_ref_barrier;
 763   if (!_do_load || needs_both_barriers) {
 764     preserve(_obj);
 765   }
 766 
 767   // Go for barriers. Barriers can return straight to continuation, as long
 768   // as another barrier is not needed and we can reach the fastpath.
 769   if (needs_both_barriers) {
 770     // The Load match rule in the .ad file may have legitimized the load
 771     // address using a TEMP register and in that case we need to explicitly
 772     // preserve them here, because the RA does not consider TEMP as live-in,
 773     // and the KA runtime call may clobber them and cause a crash on the
 774     // subsequent LRB stub.
 775     if (_addr.base() != noreg) {
 776       preserve(_addr.base());
 777     }
 778     if (_addr.index() != noreg) {
 779       preserve(_addr.index());
 780     }
 781     keepalive(masm, nullptr);
 782     lrb(masm);
 783   } else if (_needs_keep_alive_barrier) {
 784     keepalive(masm, continuation());
 785   } else if (_needs_load_ref_barrier) {
 786     lrb(masm);
 787   } else {
 788     ShouldNotReachHere();
 789   }
 790 }
 791 
 792 void ShenandoahBarrierStubC2::maybe_far_jump_if_zero(MacroAssembler& masm, Register reg) {
 793   if (_needs_far_jump) {
 794     Label L_short_jump;
 795     __ cbnz(reg, L_short_jump);
 796     __ b(*continuation());
 797     __ bind(L_short_jump);
 798   } else {
 799     __ cbz(reg, *continuation());
 800   }
 801 }
 802 
 803 void ShenandoahBarrierStubC2::keepalive(MacroAssembler& masm, Label* L_done) {
 804   Address gcstate(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_fast_array_offset(ShenandoahHeap::MARKING)));
 805   Address index(rthread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
 806   Address buffer(rthread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
 807   Label L_through, L_slowpath;
 808 
 809   // If another barrier is enabled as well, do a runtime check for a specific barrier.
 810   if (_needs_load_ref_barrier) {
 811     assert(L_done == nullptr, "L_done is always null when _needs_load_ref_barrier is true");
 812     __ ldrb(_tmp1, gcstate);
 813     __ cbz(_tmp1, L_through);
 814   }
 815 
 816   // Fast-path: put object into buffer.
 817   // If buffer is already full, go slow.
 818   __ ldr(_tmp1, index);
 819   __ cbz(_tmp1, L_slowpath);
 820   __ sub(_tmp1, _tmp1, wordSize);
 821   __ str(_tmp1, index);
 822   __ ldr(_tmp2, buffer);
 823 
 824   // Store the object in queue.
 825   // If object is narrow, we need to decode it before inserting.
 826   if (_narrow) {
 827     __ add(_tmp2, _tmp2, _tmp1);
 828     __ decode_heap_oop_not_null(_tmp1, _obj);
 829     __ str(_tmp1, Address(_tmp2));
 830   } else {
 831     // Buffer is 64-bit address, must be in base register.
 832     __ str(_obj, Address(_tmp2, _tmp1));
 833   }
 834 
 835   // Fast-path exits here.
 836   if (L_done != nullptr) {
 837     __ b(*L_done);
 838   } else {
 839     __ b(L_through);
 840   }
 841 
 842   // Slow-path: call runtime to handle.
 843   __ bind(L_slowpath);
 844 
 845   {
 846     SaveLiveRegisters slr(&masm, this);
 847 
 848     // Go to runtime and handle the rest there.
 849     __ mov(c_rarg0, _obj);
 850     __ lea(lr, RuntimeAddress(keepalive_runtime_entry_addr()));
 851     __ blr(lr);
 852   }
 853   if (L_done != nullptr) {
 854     __ b(*L_done);
 855   } else {
 856     __ bind(L_through);
 857   }
 858 }
 859 
 860 void ShenandoahBarrierStubC2::lrb(MacroAssembler& masm) {
 861   Label L_slow;
 862 
 863   // If another barrier is enabled as well, do a runtime check for a specific barrier.
 864   if (_needs_keep_alive_barrier) {
 865     char state_to_check = ShenandoahHeap::HAS_FORWARDED | (_needs_load_ref_weak_barrier ? ShenandoahHeap::WEAK_ROOTS : 0);
 866     Address gc_state_fast(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_fast_array_offset(state_to_check)));
 867     __ ldrb(_tmp1, gc_state_fast);
 868     maybe_far_jump_if_zero(masm, _tmp1);
 869   }
 870 
 871   // If weak references are being processed, weak/phantom loads need to go slow,
 872   // regardless of their cset status.
 873   if (_needs_load_ref_weak_barrier) {
 874     Address gc_state_fast(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_fast_array_offset(ShenandoahHeap::WEAK_ROOTS)));
 875     __ ldrb(_tmp1, gc_state_fast);
 876     __ cbnz(_tmp1, L_slow);
 877   }
 878 
 879   // Cset-check. Fall-through to slow if in collection set.
 880   bool is_aot = AOTCodeCache::is_on_for_dump();
 881   if (!is_aot) {
 882     __ mov(_tmp1, ShenandoahHeap::in_cset_fast_test_addr());
 883     if (_narrow) {
 884       __ decode_heap_oop_not_null(_tmp2, _obj);
 885       __ add(_tmp1, _tmp1, _tmp2, Assembler::LSR, ShenandoahHeapRegion::region_size_bytes_shift_jint());
 886     } else {
 887       __ add(_tmp1, _tmp1, _obj, Assembler::LSR, ShenandoahHeapRegion::region_size_bytes_shift_jint());
 888     }
 889   } else {
 890     // Generating AOT code, pull the cset bitmap and region shift from AOT table.
 891     if (_narrow) {
 892       __ decode_heap_oop_not_null(_tmp1, _obj);
 893     } else {
 894       __ mov(_tmp1, _obj);
 895     }
 896     __ lea(_tmp2, ExternalAddress(AOTRuntimeConstants::grain_shift_address()));
 897     __ ldrw(_tmp2, Address(_tmp2));
 898     __ lsrv(_tmp2, _tmp1, _tmp2);
 899     __ lea(_tmp1, ExternalAddress(AOTRuntimeConstants::cset_base_address()));
 900     __ ldr(_tmp1, Address(_tmp1));
 901     __ add(_tmp1, _tmp1, _tmp2);
 902   }
 903   __ ldrb(_tmp1, Address(_tmp1, 0));
 904   maybe_far_jump_if_zero(masm, _tmp1);
 905 
 906   // Slow path
 907   __ bind(L_slow);
 908 
 909   // Obj is the result, need to temporarily stop preserving it.
 910   bool is_obj_preserved = is_preserved(_obj);
 911   if (is_obj_preserved) {
 912     dont_preserve(_obj);
 913   }
 914   {
 915     SaveLiveRegisters slr(&masm, this);
 916 
 917     // Shuffle in the arguments. The end result should be:
 918     //   c_rarg0 <-- obj
 919     //   c_rarg1 <-- lea(addr)
 920     if (c_rarg0 == _obj) {
 921       __ lea(c_rarg1, _addr);
 922     } else if (c_rarg1 == _obj) {
 923       __ mov(_tmp1, c_rarg1);
 924       __ lea(c_rarg1, _addr);
 925       __ mov(c_rarg0, _tmp1);
 926     } else {
 927       assert_different_registers(c_rarg1, _obj);
 928       __ lea(c_rarg1, _addr);
 929       __ mov(c_rarg0, _obj);
 930     }
 931 
 932     // Go to runtime and handle the rest there.
 933     __ lea(lr, RuntimeAddress(lrb_runtime_entry_addr()));
 934     __ blr(lr);
 935 
 936     // Save the result where needed. Narrow entries return narrowOop (32 bits)
 937     // and AAPCS does not guarantee the upper 32 bits of x0 are zero.
 938     if (_narrow) {
 939       __ movw(_obj, r0);
 940     } else if (_obj != r0) {
 941       __ mov(_obj, r0);
 942     }
 943   }
 944   if (is_obj_preserved) {
 945     preserve(_obj);
 946   }
 947 
 948   __ b(*continuation());
 949 }
 950 
 951 int ShenandoahBarrierStubC2::available_gp_registers() {
 952   Unimplemented(); // Not used
 953   return 0;
 954 }
 955 
 956 bool ShenandoahBarrierStubC2::is_special_register(Register r) {
 957   Unimplemented(); // Not used
 958   return true;
 959 }
 960 
 961 static ShenandoahBarrierSetC2State* barrier_set_state() {
 962   return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
 963 }
 964 
 965 static int get_stub_size(ShenandoahBarrierStubC2* stub) {
 966   PhaseOutput* const output = Compile::current()->output();
 967   assert(output->in_scratch_emit_size(), "only used when in scratch_emit_size.");
 968   BufferBlob* const blob = output->scratch_buffer_blob();
 969   CodeBuffer cb(blob->content_begin(), (address)output->scratch_locs_memory() - blob->content_begin());
 970   MacroAssembler masm(&cb);
 971   stub->emit_code(masm);
 972   return cb.insts_size();
 973 }
 974 
 975 void ShenandoahBarrierStubC2::post_init() {
 976   // If we are in scratch emit mode we assume worst case, and force the use of
 977   // far branches.
 978   PhaseOutput* const output = Compile::current()->output();
 979   ShenandoahBarrierSetC2State* state = barrier_set_state();
 980   if (output->in_scratch_emit_size()) {
 981     state->inc_stubs_current_total_size(get_stub_size(this));
 982     _needs_far_jump = true;
 983     return;
 984   }
 985 
 986   // The logic implemented in this stub only uses short jumps (cbz, cbnz) if
 987   // the aggregation of all relevant code sections of a method is less than 1MB
 988   // - 2KB. We could be more aggressive and try and compute the distance
 989   // between the fastpath branch and the stub entry but in practice not many
 990   // methods reach the 1MB size.
 991   const BufferSizingData* sizing = output->buffer_sizing_data();
 992   const int code_size = sizing->_code + state->stubs_current_total_size();
 993 
 994   // Maximum backward range is 1M. Maximum forward reach is 1M - 4bytes.
 995   // Subtract 2K to be ultra conservative.
 996   const int cond_branch_max_reach = (int)(1*M - 2*K);
 997   _needs_far_jump = code_size >= cond_branch_max_reach;
 998 }
 999 
1000 #endif // COMPILER2