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/shenandoahHeap.inline.hpp"
  32 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
  33 #include "gc/shenandoah/shenandoahNMethod.inline.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 "nativeInst_aarch64.hpp"
  39 #include "runtime/javaThread.hpp"
  40 #include "runtime/sharedRuntime.hpp"
  41 #ifdef COMPILER1
  42 #include "c1/c1_LIRAssembler.hpp"
  43 #include "c1/c1_MacroAssembler.hpp"
  44 #include "gc/shenandoah/c1/shenandoahBarrierSetC1.hpp"
  45 #endif
  46 #ifdef COMPILER2
  47 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
  48 #include "opto/output.hpp"
  49 #endif
  50 
  51 #define __ masm->
  52 
  53 void ShenandoahBarrierSetAssembler::arraycopy_prologue(MacroAssembler* masm, DecoratorSet decorators, bool is_oop,
  54                                                        Register src, Register dst, Register count, RegSet saved_regs) {
  55   if (is_oop) {
  56     bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0;
  57     if ((ShenandoahSATBBarrier && !dest_uninitialized) || ShenandoahLoadRefBarrier) {
  58 
  59       Label done;
  60 
  61       // Avoid calling runtime if count == 0
  62       __ cbz(count, done);
  63 
  64       // Is GC active?
  65       assert(!saved_regs.contains(rscratch1), "Sanity: about to clobber rscratch1");
  66       assert(!saved_regs.contains(rscratch2), "Sanity: about to clobber rscratch2");
  67       Address gc_state(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
  68       __ ldrb(rscratch1, gc_state);
  69       if (ShenandoahSATBBarrier && dest_uninitialized) {
  70         __ tbz(rscratch1, ShenandoahHeap::HAS_FORWARDED_BITPOS, done);
  71       } else {
  72         __ mov(rscratch2, ShenandoahHeap::HAS_FORWARDED | ShenandoahHeap::MARKING);
  73         __ tst(rscratch1, rscratch2);
  74         __ br(Assembler::EQ, done);
  75       }
  76 
  77       __ push_call_clobbered_registers();
  78       // If arguments are not in proper places, shuffle them.
  79       // Doing this via the stack is the most straight-forward way to avoid
  80       // accidentally smashing any register.
  81       if (c_rarg0 != src || c_rarg1 != dst || c_rarg2 != count) {
  82         __ push(RegSet::of(src), sp);
  83         __ push(RegSet::of(dst), sp);
  84         __ push(RegSet::of(count), sp);
  85         __ pop(RegSet::of(c_rarg2), sp);
  86         __ pop(RegSet::of(c_rarg1), sp);
  87         __ pop(RegSet::of(c_rarg0), sp);
  88       }
  89       address target = nullptr;
  90       if (UseCompressedOops) {
  91         target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::arraycopy_barrier_narrow_oop);
  92       } else {
  93         target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::arraycopy_barrier_oop);
  94       }
  95       __ call_VM_leaf(target, 3);
  96       __ pop_call_clobbered_registers();
  97       __ bind(done);
  98     }
  99   }
 100 }
 101 
 102 void ShenandoahBarrierSetAssembler::arraycopy_epilogue(MacroAssembler* masm, DecoratorSet decorators, bool is_oop,
 103                                                        Register start, Register count, Register tmp) {
 104   if (ShenandoahCardBarrier && is_oop) {
 105     gen_write_ref_array_post_barrier(masm, decorators, start, count, tmp);
 106   }
 107 }
 108 
 109 void ShenandoahBarrierSetAssembler::satb_barrier(MacroAssembler* masm,
 110                                                  Register obj,
 111                                                  Register pre_val,
 112                                                  Register thread,
 113                                                  Register tmp1,
 114                                                  Register tmp2) {
 115   assert(ShenandoahSATBBarrier, "Should be checked by caller");
 116   assert(thread == rthread, "must be");
 117 
 118   Label done;
 119   Label runtime;
 120 
 121   assert_different_registers(obj, pre_val, tmp1, tmp2);
 122   assert(pre_val != noreg && tmp1 != noreg && tmp2 != noreg, "expecting a register");
 123 
 124   Address index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
 125   Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
 126 
 127   // Is marking active?
 128   Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 129   __ ldrb(tmp1, gc_state);
 130   __ tbz(tmp1, ShenandoahHeap::MARKING_BITPOS, done);
 131 
 132   // Do we need to load the previous value?
 133   if (obj != noreg) {
 134     if (UseCompressedOops) {
 135       __ ldrw(pre_val, Address(obj, 0));
 136       __ decode_heap_oop(pre_val);
 137     } else {
 138       __ ldr(pre_val, Address(obj, 0));
 139     }
 140   }
 141 
 142   // Is the previous value null?
 143   __ cbz(pre_val, done);
 144 
 145   // Can we store original value in the thread's buffer?
 146   // Is index == 0?
 147   // (The index field is typed as size_t.)
 148 
 149   __ ldr(tmp1, index);                      // tmp := *index_adr
 150   __ cbz(tmp1, runtime);                    // tmp == 0?
 151                                         // If yes, goto runtime
 152 
 153   __ sub(tmp1, tmp1, wordSize);             // tmp := tmp - wordSize
 154   __ str(tmp1, index);                      // *index_adr := tmp
 155   __ ldr(tmp2, buffer);
 156   __ add(tmp1, tmp1, tmp2);                 // tmp := tmp + *buffer_adr
 157 
 158   // Record the previous value
 159   __ str(pre_val, Address(tmp1, 0));
 160   __ b(done);
 161 
 162   __ bind(runtime);
 163 
 164   // Slow-path call
 165   __ enter(/* strip_ret_addr = */ true);
 166   __ push_call_clobbered_registers();
 167   if (c_rarg0 != pre_val) {
 168     __ mov(c_rarg0, pre_val);
 169   }
 170   // Calling with super_call_VM_leaf with c_rarg0 bypasses interpreter checks and avoids any moves.
 171   __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_pre), c_rarg0);
 172   __ pop_call_clobbered_registers();
 173   __ leave();
 174 
 175   __ bind(done);
 176 }
 177 
 178 void ShenandoahBarrierSetAssembler::load_reference_barrier(MacroAssembler* masm, Register dst, Address load_addr, DecoratorSet decorators) {
 179   assert(ShenandoahLoadRefBarrier, "Should be enabled");
 180   assert(dst != rscratch2, "need rscratch2");
 181   assert_different_registers(load_addr.base(), load_addr.index(), rscratch1, rscratch2);
 182 
 183   bool is_strong  = ShenandoahBarrierSet::is_strong_access(decorators);
 184   bool is_weak    = ShenandoahBarrierSet::is_weak_access(decorators);
 185   bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
 186   bool is_native  = ShenandoahBarrierSet::is_native_access(decorators);
 187   bool is_narrow  = UseCompressedOops && !is_native;
 188 
 189   Label heap_stable, not_cset;
 190   Address gc_state(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 191   __ ldrb(rscratch2, gc_state);
 192 
 193   // Check for heap stability
 194   if (is_strong) {
 195     __ tbz(rscratch2, ShenandoahHeap::HAS_FORWARDED_BITPOS, heap_stable);
 196   } else {
 197     Label lrb;
 198     __ tbnz(rscratch2, ShenandoahHeap::WEAK_ROOTS_BITPOS, lrb);
 199     __ tbz(rscratch2, ShenandoahHeap::HAS_FORWARDED_BITPOS, heap_stable);
 200     __ bind(lrb);
 201   }
 202 
 203   // use r1 for load address
 204   Register result_dst = dst;
 205   if (dst == r1) {
 206     __ mov(rscratch1, dst);
 207     dst = rscratch1;
 208   }
 209 
 210   // Save r0 and r1, unless it is an output register
 211   RegSet to_save = RegSet::of(r0, r1) - result_dst;
 212   __ push(to_save, sp);
 213   __ lea(r1, load_addr);
 214   __ mov(r0, dst);
 215 
 216   // Test for in-cset
 217   if (is_strong) {
 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       __ mov(rscratch2, ShenandoahHeap::in_cset_fast_test_addr());
 226       __ lsr(rscratch1, r0, ShenandoahHeapRegion::region_size_bytes_shift_jint());
 227     }
 228     __ ldrb(rscratch2, Address(rscratch2, rscratch1));
 229     __ tbz(rscratch2, 0, not_cset);
 230   }
 231 
 232   // Slow-path call
 233   __ enter(/* strip_ret_addr = */ true);
 234   __ push_call_clobbered_registers();
 235   address target = nullptr;
 236   if (is_strong) {
 237     if (is_narrow) {
 238       target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow);
 239     } else {
 240       target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong);
 241     }
 242   } else if (is_weak) {
 243     if (is_narrow) {
 244       target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow);
 245     } else {
 246       target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak);
 247     }
 248   } else {
 249     assert(is_phantom, "only remaining strength");
 250     assert(!is_narrow, "phantom access cannot be narrow");
 251     target = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom);
 252   }
 253   // Calling with super_call_VM_leaf with c_rarg0/1 bypasses interpreter checks and avoids any moves.
 254   __ super_call_VM_leaf(target, c_rarg0, c_rarg1);
 255   __ mov(rscratch1, r0);
 256   __ pop_call_clobbered_registers();
 257   __ mov(r0, rscratch1);
 258   __ leave();
 259 
 260   __ bind(not_cset);
 261 
 262   __ mov(result_dst, r0);
 263   __ pop(to_save, sp);
 264 
 265   __ bind(heap_stable);
 266 }
 267 
 268 //
 269 // Arguments:
 270 //
 271 // Inputs:
 272 //   src:        oop location to load from, might be clobbered
 273 //
 274 // Output:
 275 //   dst:        oop loaded from src location
 276 //
 277 // Kill:
 278 //   rscratch1 (scratch reg)
 279 //
 280 // Alias:
 281 //   dst: rscratch1 (might use rscratch1 as temporary output register to avoid clobbering src)
 282 //
 283 void ShenandoahBarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 284                                             Register dst, Address src, Register tmp1, Register tmp2) {
 285   // 1: non-reference load, no additional barrier is needed
 286   if (!is_reference_type(type)) {
 287     BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp2);
 288     return;
 289   }
 290 
 291   // 2: load a reference from src location and apply LRB if needed
 292   if (ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) {
 293     Register result_dst = dst;
 294 
 295     // Preserve src location for LRB
 296     if (dst == src.base() || dst == src.index()) {
 297       dst = rscratch1;
 298     }
 299     assert_different_registers(dst, src.base(), src.index());
 300 
 301     BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp2);
 302 
 303     load_reference_barrier(masm, dst, src, decorators);
 304 
 305     if (dst != result_dst) {
 306       __ mov(result_dst, dst);
 307       dst = result_dst;
 308     }
 309   } else {
 310     BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp2);
 311   }
 312 
 313   // 3: apply keep-alive barrier if needed
 314   if (ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) {
 315     satb_barrier(masm /* masm */,
 316                  noreg /* obj */,
 317                  dst /* pre_val */,
 318                  rthread /* thread */,
 319                  tmp1 /* tmp1 */,
 320                  tmp2 /* tmp2 */);
 321   }
 322 }
 323 
 324 void ShenandoahBarrierSetAssembler::card_barrier(MacroAssembler* masm, Register obj) {
 325   assert(ShenandoahCardBarrier, "Should have been checked by caller");
 326 
 327   __ lsr(obj, obj, CardTable::card_shift());
 328 
 329   assert(CardTable::dirty_card_val() == 0, "must be");
 330 
 331   Address curr_ct_holder_addr(rthread, in_bytes(ShenandoahThreadLocalData::card_table_offset()));
 332   __ ldr(rscratch1, curr_ct_holder_addr);
 333 
 334   if (UseCondCardMark) {
 335     Label L_already_dirty;
 336     __ ldrb(rscratch2, Address(obj, rscratch1));
 337     __ cbz(rscratch2, L_already_dirty);
 338     __ strb(zr, Address(obj, rscratch1));
 339     __ bind(L_already_dirty);
 340   } else {
 341     __ strb(zr, Address(obj, rscratch1));
 342   }
 343 }
 344 
 345 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 346                                              Address dst, Register val, Register tmp1, Register tmp2, Register tmp3) {
 347   // 1: non-reference types require no barriers
 348   if (!is_reference_type(type)) {
 349     BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2, tmp3);
 350     return;
 351   }
 352 
 353   // Flatten object address right away for simplicity: likely needed by barriers
 354   if (dst.index() == noreg && dst.offset() == 0) {
 355     if (dst.base() != tmp3) {
 356       __ mov(tmp3, dst.base());
 357     }
 358   } else {
 359     __ lea(tmp3, dst);
 360   }
 361 
 362   // 2: pre-barrier: SATB needs the previous value
 363   if (ShenandoahBarrierSet::need_satb_barrier(decorators, type)) {
 364     satb_barrier(masm,
 365                  tmp3 /* obj */,
 366                  tmp2 /* pre_val */,
 367                  rthread /* thread */,
 368                  tmp1 /* tmp */,
 369                  rscratch1 /* tmp2 */);
 370   }
 371 
 372   // Store!
 373   BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp3, 0), val, noreg, noreg, noreg);
 374 
 375   // 3: post-barrier: card barrier needs store address
 376   bool storing_non_null = (val != noreg);
 377   if (ShenandoahBarrierSet::need_card_barrier(decorators, type) && storing_non_null) {
 378     card_barrier(masm, tmp3);
 379   }
 380 }
 381 
 382 void ShenandoahBarrierSetAssembler::try_resolve_jobject_in_native(MacroAssembler* masm, Register jni_env,
 383                                                                   Register obj, Register tmp, Label& slowpath) {
 384   Label done;
 385   // Resolve jobject
 386   BarrierSetAssembler::try_resolve_jobject_in_native(masm, jni_env, obj, tmp, slowpath);
 387 
 388   // Check for null.
 389   __ cbz(obj, done);
 390 
 391   assert(obj != rscratch2, "need rscratch2");
 392   Address gc_state(jni_env, ShenandoahThreadLocalData::gc_state_offset() - JavaThread::jni_environment_offset());
 393   __ lea(rscratch2, gc_state);
 394   __ ldrb(rscratch2, Address(rscratch2));
 395 
 396   // Check for heap in evacuation phase
 397   __ tbnz(rscratch2, ShenandoahHeap::EVACUATION_BITPOS, slowpath);
 398 
 399   __ bind(done);
 400 }
 401 
 402 void ShenandoahBarrierSetAssembler::try_peek_weak_handle_in_nmethod(MacroAssembler* masm, Register weak_handle, Register obj,
 403                                                                     Register tmp, Label& slow_path) {
 404   assert_different_registers(weak_handle, tmp, noreg);
 405   assert_different_registers(obj, tmp, noreg);
 406 
 407   Label done;
 408 
 409   // Peek weak handle using the standard implementation.
 410   BarrierSetAssembler::try_peek_weak_handle_in_nmethod(masm, weak_handle, obj, tmp, slow_path);
 411 
 412   // Check if the reference is null, and if it is, take the fast path.
 413   __ cbz(obj, done);
 414 
 415   Address gc_state(rthread, ShenandoahThreadLocalData::gc_state_offset());
 416   __ lea(tmp, gc_state);
 417   __ ldrb(tmp, __ legitimize_address(gc_state, 1, tmp));
 418 
 419   // Check if the heap is under weak-reference/roots processing, in
 420   // which case we need to take the slow path.
 421   __ tbnz(tmp, ShenandoahHeap::WEAK_ROOTS_BITPOS, slow_path);
 422   __ bind(done);
 423 }
 424 
 425 void ShenandoahBarrierSetAssembler::check_oop(MacroAssembler* masm, Register obj, Register tmp1, Register tmp2, Label& L_error) {
 426   // Check if the oop is in the right area of memory
 427   __ mov(tmp2, (intptr_t) Universe::verify_oop_mask());
 428   __ andr(tmp1, obj, tmp2);
 429   __ mov(tmp2, (intptr_t) Universe::verify_oop_bits());
 430 
 431   // Compare tmp1 and tmp2.  We don't use a compare
 432   // instruction here because the flags register is live.
 433   __ eor(tmp1, tmp1, tmp2);
 434   __ cbnz(tmp1, L_error);
 435 
 436   // This routine is sometimes called before applying GC barriers.
 437   // With +COH, loading the klass may end up loading forwarding pointer instead.
 438   Label L_skip;
 439   if (UseCompactObjectHeaders) {
 440     Address gc_state(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 441     __ ldrb(tmp1, gc_state);
 442     __ tbnz(tmp1, ShenandoahHeap::HAS_FORWARDED_BITPOS, L_skip);
 443   }
 444 
 445   // Make sure klass is 'reasonable', which is not zero.
 446   __ load_narrow_klass(tmp1, obj);
 447   __ cbz(tmp1, L_error);
 448   __ bind(L_skip);
 449 }
 450 
 451 void ShenandoahBarrierSetAssembler::gen_write_ref_array_post_barrier(MacroAssembler* masm, DecoratorSet decorators,
 452                                                                      Register start, Register count, Register scratch) {
 453   assert(ShenandoahCardBarrier, "Should have been checked by caller");
 454 
 455   Label L_loop, L_done;
 456   const Register end = count;
 457 
 458   // Zero count? Nothing to do.
 459   __ cbz(count, L_done);
 460 
 461   // end = start + count << LogBytesPerHeapOop
 462   // last element address to make inclusive
 463   __ lea(end, Address(start, count, Address::lsl(LogBytesPerHeapOop)));
 464   __ sub(end, end, BytesPerHeapOop);
 465   __ lsr(start, start, CardTable::card_shift());
 466   __ lsr(end, end, CardTable::card_shift());
 467 
 468   // number of bytes to copy
 469   __ sub(count, end, start);
 470 
 471   Address curr_ct_holder_addr(rthread, in_bytes(ShenandoahThreadLocalData::card_table_offset()));
 472   __ ldr(scratch, curr_ct_holder_addr);
 473   __ add(start, start, scratch);
 474   __ bind(L_loop);
 475   __ strb(zr, Address(start, count));
 476   __ subs(count, count, 1);
 477   __ br(Assembler::GE, L_loop);
 478   __ bind(L_done);
 479 }
 480 
 481 #undef __
 482 
 483 address ShenandoahBarrierSetAssembler::parse_jump_address(address pc) {
 484   NativeInstruction* ni = nativeInstruction_at(pc);
 485   assert(ni->is_jump(), "Initial code version: GC barrier fastpath must be a jump");
 486   NativeJump* jmp = nativeJump_at(pc);
 487   return jmp->jump_destination();
 488 }
 489 
 490 static uint32_t encode_patchable_nop() {
 491   return 0xD503201F;
 492 }
 493 
 494 static uint32_t encode_patchable_jump(address pc, address target_pc) {
 495   int32_t disp = checked_cast<int32_t>((intptr_t)target_pc - (intptr_t)pc);
 496   int64_t imm26 = disp >> 2;
 497   assert(Assembler::is_simm(imm26, 26), "maximum offset is 128MiB");
 498   return 0x14000000 | (imm26 & 0x03FFFFFF);
 499 }
 500 
 501 void ShenandoahBarrierSetAssembler::insert_patchable_nop(address pc) {
 502   *((uint32_t*)pc) = encode_patchable_nop();
 503 }
 504 
 505 void ShenandoahBarrierSetAssembler::insert_patchable_jump(address pc, address target_pc) {
 506   *((uint32_t*)pc) = encode_patchable_jump(pc, target_pc);
 507 }
 508 
 509 bool ShenandoahBarrierSetAssembler::is_patchable_nop(address pc) {
 510   return *((uint32_t*)pc) == encode_patchable_nop();
 511 }
 512 
 513 bool ShenandoahBarrierSetAssembler::is_patchable_jump(address pc, address target_pc) {
 514   return *((uint32_t*)pc) == encode_patchable_jump(pc, target_pc);
 515 }
 516 
 517 #ifdef COMPILER1
 518 
 519 #define __ ce->masm()->
 520 
 521 void ShenandoahBarrierSetAssembler::keepalive_barrier_c1_stub(LIR_Assembler* ce, ShenandoahKeepaliveBarrierStub* stub) {
 522   __ bind(*stub->entry());
 523 
 524   ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
 525 
 526   Register obj = stub->obj()->as_register();
 527 
 528   if (stub->do_load()) {
 529     ce->mem2reg(stub->addr(), stub->obj(), T_OBJECT, lir_patch_none, nullptr, /* wide = */ false);
 530   }
 531   __ cbz(obj, *stub->continuation());
 532   ce->store_parameter(obj, 0);
 533   __ far_call(RuntimeAddress(bs->keepalive_barrier_stub()));
 534   __ b(*stub->continuation());
 535 }
 536 
 537 void ShenandoahBarrierSetAssembler::load_reference_barrier_c1_stub(LIR_Assembler* ce, ShenandoahLoadReferenceBarrierStub* stub) {
 538   __ bind(*stub->entry());
 539 
 540   ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
 541 
 542   Register obj = stub->obj()->as_register();
 543   Register addr = stub->addr()->as_pointer_register();
 544   Register slow_result = stub->slow_result()->as_register();
 545   assert_different_registers(obj, addr, slow_result);
 546   assert(slow_result == r0, "C1 must know about our slow call result register");
 547 
 548   ce->store_parameter(obj, 0);
 549   ce->store_parameter(addr, 1);
 550   __ far_call(RuntimeAddress(bs->load_reference_barrier_stub(stub->decorators())));
 551   if (obj != slow_result) {
 552     __ mov(obj, slow_result);
 553   }
 554 
 555   __ b(*stub->continuation());
 556 }
 557 
 558 #undef __
 559 
 560 #define __ sasm->
 561 
 562 void ShenandoahBarrierSetAssembler::keepalive_barrier_c1_runtime_stub(StubAssembler* sasm) {
 563   __ prologue("shenandoah_keepalive_barrier", false);
 564   const Register tmp_obj = r0;
 565   const Register tmp1 = r1;
 566   const Register tmp2 = r2;
 567   __ push(RegSet::of(tmp1, tmp2, tmp_obj), sp);
 568   __ load_parameter(0, tmp_obj);
 569   satb_barrier(sasm, noreg, tmp_obj, rthread, tmp1, tmp2);
 570   __ pop(RegSet::of(tmp1, tmp2, tmp_obj), sp);
 571   __ epilogue();
 572 }
 573 
 574 void ShenandoahBarrierSetAssembler::load_reference_barrier_c1_runtime_stub(StubAssembler* sasm, DecoratorSet decorators) {
 575   __ prologue("shenandoah_load_reference_barrier", false);
 576   const Register tmp_obj = r0;
 577   const Register tmp_addr = r1;
 578   __ push(RegSet::of(tmp_addr), sp);
 579   __ load_parameter(0, tmp_obj);
 580   __ load_parameter(1, tmp_addr);
 581   load_reference_barrier(sasm, tmp_obj, Address(tmp_addr, 0), decorators);
 582   __ pop(RegSet::of(tmp_addr), sp);
 583   __ epilogue();
 584 }
 585 
 586 #undef __
 587 
 588 #endif // COMPILER1
 589 
 590 #ifdef COMPILER2
 591 
 592 #undef __
 593 #define __ masm->
 594 
 595 
 596 void ShenandoahBarrierSetAssembler::load_c2(const MachNode* node, MacroAssembler* masm, Register dst, Address src, Register tmp1, Register tmp2, bool is_narrow, bool is_acquire) {
 597   // Do the actual load. This load is the candidate for implicit null check, and MUST come first.
 598   if (is_narrow) {
 599     if (is_acquire) {
 600       assert(src.getMode() == Address::base_plus_offset && src.offset() == 0,
 601           "is_acquire path requires address to be base-only");
 602       __ ldarw(dst, src.base());
 603     } else {
 604       __ ldrw(dst, src);
 605     }
 606   } else {
 607     if (is_acquire) {
 608       assert(src.getMode() == Address::base_plus_offset && src.offset() == 0,
 609           "is_acquire path requires address to be base-only");
 610       __ ldar(dst, src.base());
 611     } else {
 612       __ ldr(dst, src);
 613     }
 614   }
 615 
 616   ShenandoahBarrierStubC2::load_post(masm, node, dst, src, tmp1, tmp2, is_narrow);
 617 }
 618 
 619 void ShenandoahBarrierSetAssembler::store_c2(const MachNode* node, MacroAssembler* masm, Address dst, bool dst_narrow,
 620     Register src, bool src_narrow, Register tmp1, Register tmp2, Register tmp3, bool is_volatile) {
 621 
 622   ShenandoahBarrierStubC2::store_pre(masm, node, dst, tmp1, tmp2, tmp3, dst_narrow);
 623 
 624   // Do the actual store
 625   if (dst_narrow) {
 626     if (!src_narrow) {
 627       // Need to encode into rscratch, because we cannot clobber src.
 628       if ((node->barrier_data() & ShenandoahBitNotNull) == 0) {
 629         __ encode_heap_oop(tmp2, src);
 630       } else {
 631         __ encode_heap_oop_not_null(tmp2, src);
 632       }
 633       src = tmp2;
 634     }
 635 
 636     if (is_volatile) {
 637       assert(dst.getMode() == Address::base_plus_offset && dst.offset() == 0,
 638           "is_acquire path requires address to be base-only");
 639       __ stlrw(src, dst.base());
 640     } else {
 641       __ strw(src, dst);
 642     }
 643   } else {
 644     if (is_volatile) {
 645       assert(dst.getMode() == Address::base_plus_offset && dst.offset() == 0,
 646           "is_acquire path requires address to be base-only");
 647       __ stlr(src, dst.base());
 648     } else {
 649       __ str(src, dst);
 650     }
 651   }
 652 
 653   ShenandoahBarrierStubC2::store_post(masm, node, dst, tmp2, tmp3);
 654 }
 655 
 656 void ShenandoahBarrierSetAssembler::compare_and_set_c2(const MachNode* node, MacroAssembler* masm, Register res, Register addr,
 657     Register oldval, Register newval, Register tmp1, Register tmp2, Register tmp3, bool exchange, bool narrow, bool weak, bool acquire) {
 658   Assembler::operand_size op_size = narrow ? Assembler::word : Assembler::xword;
 659 
 660   ShenandoahBarrierStubC2::load_store_pre(masm, node, addr, tmp1, tmp2, tmp3, narrow);
 661 
 662   atomic_memory_order order = acquire ? memory_order_seq_cst : memory_order_release;
 663 
 664   // CAS!
 665   if (weak) {
 666     __ cmpxchg_weak(addr, oldval, newval, op_size, order, exchange ? res : noreg);
 667   } else {
 668     __ cmpxchg(addr, oldval, newval, op_size, order, exchange ? res : noreg);
 669   }
 670 
 671   // If we need a boolean result out of CAS, set the flag appropriately and promote the result.
 672   if (!exchange) {
 673     assert(res != noreg, "need result register");
 674     __ cset(res, Assembler::EQ);
 675   }
 676 
 677   ShenandoahBarrierStubC2::load_store_post(masm, node, Address(addr, 0), tmp2, tmp3);
 678 }
 679 
 680 void ShenandoahBarrierSetAssembler::get_and_set_c2(const MachNode* node, MacroAssembler* masm, Register preval,
 681     Register newval, Register addr, Register tmp1, Register tmp2, Register tmp3, bool is_acquire) {
 682   bool is_narrow = node->bottom_type()->isa_narrowoop();
 683 
 684   ShenandoahBarrierStubC2::load_store_pre(masm, node, addr, tmp1, tmp2, tmp3, is_narrow);
 685 
 686   if (is_narrow) {
 687     if (is_acquire) {
 688       __ atomic_xchgalw(preval, newval, addr);
 689     } else {
 690       __ atomic_xchgw(preval, newval, addr);
 691     }
 692   } else {
 693     if (is_acquire) {
 694       __ atomic_xchgal(preval, newval, addr);
 695     } else {
 696       __ atomic_xchg(preval, newval, addr);
 697     }
 698   }
 699 
 700   ShenandoahBarrierStubC2::load_store_post(masm, node, Address(addr, 0), tmp2, tmp3);
 701 }
 702 
 703 #undef __
 704 #define __ masm.
 705 
 706 void ShenandoahBarrierStubC2::cardtable(MacroAssembler& masm, Address address, Register tmp1, Register tmp2) {
 707   assert(CardTable::dirty_card_val() == 0, "must be");
 708   Assembler::InlineSkippedInstructionsCounter skip_counter(&masm);
 709 
 710   // tmp1 = card table base (holder)
 711   Address curr_ct_holder_addr(rthread, in_bytes(ShenandoahThreadLocalData::card_table_offset()));
 712   __ ldr(tmp1, curr_ct_holder_addr);
 713 
 714   // tmp2 = effective address
 715   __ lea(tmp2, address);
 716 
 717   // tmp2 = &card_table[ addr >> CardTable::card_shift() ] ; card index
 718   __ add(tmp2, tmp1, tmp2, Assembler::LSR, CardTable::card_shift());
 719 
 720   if (UseCondCardMark) {
 721     Label L_already_dirty;
 722     __ ldrb(tmp1, Address(tmp2));
 723     __ cbz(tmp1, L_already_dirty);
 724     __ strb(zr, Address(tmp2));
 725     __ bind(L_already_dirty);
 726   } else {
 727     __ strb(zr, Address(tmp2));
 728   }
 729 }
 730 
 731 void ShenandoahBarrierStubC2::patchable_jump(MacroAssembler& masm, const char gc_state, bool jump_when_state, Label* L_target) {
 732   PhaseOutput* const output = Compile::current()->output();
 733   if (output->in_scratch_emit_size()) {
 734     // Avoid binding L_target in scratch emits.
 735     // We know the patched check is exactly one instruction long.
 736     __ nop();
 737     return;
 738   }
 739 
 740   // Emit the unconditional branch in the first version of the method.
 741   // Let the rest of runtime figure out how to manage it.
 742   __ relocate(patchable_barrier_Relocation::spec(ShenandoahNMethod::encode_to_reloc(gc_state, jump_when_state)));
 743   __ b(*L_target);
 744 }
 745 
 746 void ShenandoahBarrierStubC2::enter_if_gc_state(MacroAssembler& masm, const char test_state) {
 747   Assembler::InlineSkippedInstructionsCounter skip_counter(&masm);
 748   patchable_jump_if_gc_state(masm, test_state, entry());
 749   __ bind(*continuation());
 750 }
 751 
 752 void ShenandoahBarrierStubC2::emit_code(MacroAssembler& masm) {
 753   Assembler::InlineSkippedInstructionsCounter skip_counter(&masm);
 754   assert(_needs_keep_alive_barrier || _needs_load_ref_barrier, "Why are you here?");
 755   PhaseOutput* const output = Compile::current()->output();
 756 
 757   // We piggyback on scratch_emit_size mode to compute the slowpath stub size.
 758   // We'll use that information to decide whether we need a far jump to the
 759   // stub entry point or not. In scratch_emit_size mode we don't bind entry()
 760   // because otherwise it will be rebound when we later emit the instructions
 761   // for real.
 762   if (!output->in_scratch_emit_size()) {
 763     __ bind(*entry());
 764   }
 765 
 766   // If we need to load ourselves, do it here.
 767   if (_do_load) {
 768     if (_narrow) {
 769       __ ldrw(_obj, _addr);
 770     } else {
 771       __ ldr(_obj, _addr);
 772     }
 773   }
 774 
 775   // If the object is null, there is no point in applying barriers.
 776   maybe_far_jump_if_zero(masm, _obj);
 777 
 778   // We need to make sure that loads done by callers survive across slow-path calls.
 779   // For self-loads, we need to care about the case when both KA and LRB are enabled (rare).
 780   bool needs_both_barriers = _needs_keep_alive_barrier && _needs_load_ref_barrier;
 781   if (!_do_load || needs_both_barriers) {
 782     preserve(_obj);
 783   }
 784 
 785   // Go for barriers. Barriers can return straight to continuation, as long
 786   // as another barrier is not needed and we can reach the fastpath.
 787   if (needs_both_barriers) {
 788     // The Load match rule in the .ad file may have legitimized the load
 789     // address using a TEMP register and in that case we need to explicitly
 790     // preserve them here, because the RA does not consider TEMP as live-in,
 791     // and the KA runtime call may clobber them and cause a crash on the
 792     // subsequent LRB stub.
 793     if (_addr.base() != noreg) {
 794       preserve(_addr.base());
 795     }
 796     if (_addr.index() != noreg) {
 797       preserve(_addr.index());
 798     }
 799     keepalive(masm, nullptr);
 800     lrb(masm);
 801   } else if (_needs_keep_alive_barrier) {
 802     keepalive(masm, continuation());
 803   } else if (_needs_load_ref_barrier) {
 804     lrb(masm);
 805   } else {
 806     ShouldNotReachHere();
 807   }
 808 }
 809 
 810 void ShenandoahBarrierStubC2::maybe_far_jump_if_zero(MacroAssembler& masm, Register reg) {
 811   if (_needs_far_jump) {
 812     Label L_short_jump;
 813     __ cbnz(reg, L_short_jump);
 814     __ b(*continuation());
 815     __ bind(L_short_jump);
 816   } else {
 817     __ cbz(reg, *continuation());
 818   }
 819 }
 820 
 821 void ShenandoahBarrierStubC2::keepalive(MacroAssembler& masm, Label* L_done) {
 822   Address index(rthread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
 823   Address buffer(rthread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
 824   Label L_through, L_slowpath;
 825 
 826   // If another barrier is enabled as well, do a check for a specific barrier.
 827   if (_needs_load_ref_barrier) {
 828     assert(L_done == nullptr, "Should be");
 829     char state_to_check = ShenandoahHeap::MARKING;
 830     patchable_jump_if_not_gc_state(masm, state_to_check, &L_through);
 831   }
 832 
 833   // Fast-path: put object into buffer.
 834   // If buffer is already full, go slow.
 835   __ ldr(_tmp1, index);
 836   __ cbz(_tmp1, L_slowpath);
 837   __ sub(_tmp1, _tmp1, wordSize);
 838   __ str(_tmp1, index);
 839   __ ldr(_tmp2, buffer);
 840 
 841   // Store the object in queue.
 842   // If object is narrow, we need to decode it before inserting.
 843   if (_narrow) {
 844     __ add(_tmp2, _tmp2, _tmp1);
 845     __ decode_heap_oop_not_null(_tmp1, _obj);
 846     __ str(_tmp1, Address(_tmp2));
 847   } else {
 848     // Buffer is 64-bit address, must be in base register.
 849     __ str(_obj, Address(_tmp2, _tmp1));
 850   }
 851 
 852   // Fast-path exits here.
 853   if (L_done != nullptr) {
 854     __ b(*L_done);
 855   } else {
 856     __ b(L_through);
 857   }
 858 
 859   // Slow-path: call runtime to handle.
 860   __ bind(L_slowpath);
 861 
 862   {
 863     SaveLiveRegisters slr(&masm, this);
 864 
 865     // Go to runtime and handle the rest there.
 866     __ mov(c_rarg0, _obj);
 867     __ lea(lr, RuntimeAddress(keepalive_runtime_entry_addr()));
 868     __ blr(lr);
 869   }
 870   if (L_done != nullptr) {
 871     __ b(*L_done);
 872   } else {
 873     __ bind(L_through);
 874   }
 875 }
 876 
 877 void ShenandoahBarrierStubC2::lrb(MacroAssembler& masm) {
 878   Label L_slow;
 879 
 880   // If another barrier is enabled as well, do a check for a specific barrier.
 881   if (_needs_keep_alive_barrier) {
 882     char state_to_check = ShenandoahHeap::HAS_FORWARDED | (_needs_load_ref_weak_barrier ? ShenandoahHeap::WEAK_ROOTS : 0);
 883     patchable_jump_if_not_gc_state(masm, state_to_check, continuation());
 884   }
 885 
 886   // If weak references are being processed, weak/phantom loads need to go slow,
 887   // regardless of their cset status.
 888   if (_needs_load_ref_weak_barrier) {
 889     char state_to_check = ShenandoahHeap::WEAK_ROOTS;
 890     patchable_jump_if_gc_state(masm, state_to_check, &L_slow);
 891   }
 892 
 893   // Cset-check. Fall-through to slow if in collection set.
 894   bool is_aot = AOTCodeCache::is_on_for_dump();
 895   if (!is_aot) {
 896     __ mov(_tmp1, ShenandoahHeap::in_cset_fast_test_addr());
 897     if (_narrow) {
 898       __ decode_heap_oop_not_null(_tmp2, _obj);
 899       __ add(_tmp1, _tmp1, _tmp2, Assembler::LSR, ShenandoahHeapRegion::region_size_bytes_shift_jint());
 900     } else {
 901       __ add(_tmp1, _tmp1, _obj, Assembler::LSR, ShenandoahHeapRegion::region_size_bytes_shift_jint());
 902     }
 903   } else {
 904     // Generating AOT code, pull the cset bitmap and region shift from AOT table.
 905     if (_narrow) {
 906       __ decode_heap_oop_not_null(_tmp1, _obj);
 907     } else {
 908       __ mov(_tmp1, _obj);
 909     }
 910     __ lea(_tmp2, ExternalAddress(AOTRuntimeConstants::grain_shift_address()));
 911     __ ldrw(_tmp2, Address(_tmp2));
 912     __ lsrv(_tmp2, _tmp1, _tmp2);
 913     __ lea(_tmp1, ExternalAddress(AOTRuntimeConstants::cset_base_address()));
 914     __ ldr(_tmp1, Address(_tmp1));
 915     __ add(_tmp1, _tmp1, _tmp2);
 916   }
 917   __ ldrb(_tmp1, Address(_tmp1, 0));
 918   maybe_far_jump_if_zero(masm, _tmp1);
 919 
 920   // Slow path
 921   __ bind(L_slow);
 922 
 923   // Obj is the result, need to temporarily stop preserving it.
 924   bool is_obj_preserved = is_preserved(_obj);
 925   if (is_obj_preserved) {
 926     dont_preserve(_obj);
 927   }
 928   {
 929     SaveLiveRegisters slr(&masm, this);
 930 
 931     // Shuffle in the arguments. The end result should be:
 932     //   c_rarg0 <-- obj
 933     //   c_rarg1 <-- lea(addr)
 934     if (c_rarg0 == _obj) {
 935       __ lea(c_rarg1, _addr);
 936     } else if (c_rarg1 == _obj) {
 937       __ mov(_tmp1, c_rarg1);
 938       __ lea(c_rarg1, _addr);
 939       __ mov(c_rarg0, _tmp1);
 940     } else {
 941       assert_different_registers(c_rarg1, _obj);
 942       __ lea(c_rarg1, _addr);
 943       __ mov(c_rarg0, _obj);
 944     }
 945 
 946     // Go to runtime and handle the rest there.
 947     __ lea(lr, RuntimeAddress(lrb_runtime_entry_addr()));
 948     __ blr(lr);
 949 
 950     // Save the result where needed. Narrow entries return narrowOop (32 bits)
 951     // and AAPCS does not guarantee the upper 32 bits of x0 are zero.
 952     if (_narrow) {
 953       __ movw(_obj, r0);
 954     } else if (_obj != r0) {
 955       __ mov(_obj, r0);
 956     }
 957   }
 958   if (is_obj_preserved) {
 959     preserve(_obj);
 960   }
 961 
 962   __ b(*continuation());
 963 }
 964 
 965 int ShenandoahBarrierStubC2::available_gp_registers() {
 966   Unimplemented(); // Not used
 967   return 0;
 968 }
 969 
 970 bool ShenandoahBarrierStubC2::is_special_register(Register r) {
 971   Unimplemented(); // Not used
 972   return true;
 973 }
 974 
 975 static ShenandoahBarrierSetC2State* barrier_set_state() {
 976   return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
 977 }
 978 
 979 static int get_stub_size(ShenandoahBarrierStubC2* stub) {
 980   PhaseOutput* const output = Compile::current()->output();
 981   assert(output->in_scratch_emit_size(), "only used when in scratch_emit_size.");
 982   BufferBlob* const blob = output->scratch_buffer_blob();
 983   CodeBuffer cb(blob->content_begin(), (address)output->scratch_locs_memory() - blob->content_begin());
 984   MacroAssembler masm(&cb);
 985   stub->emit_code(masm);
 986   return cb.insts_size();
 987 }
 988 
 989 void ShenandoahBarrierStubC2::post_init() {
 990   // If we are in scratch emit mode we assume worst case, and force the use of
 991   // far branches.
 992   PhaseOutput* const output = Compile::current()->output();
 993   ShenandoahBarrierSetC2State* state = barrier_set_state();
 994   if (output->in_scratch_emit_size()) {
 995     state->inc_stubs_current_total_size(get_stub_size(this));
 996     _needs_far_jump = true;
 997     return;
 998   }
 999 
1000   // The logic implemented in this stub only uses short jumps (cbz, cbnz) if
1001   // the aggregation of all relevant code sections of a method is less than 1MB
1002   // - 2KB. We could be more aggressive and try and compute the distance
1003   // between the fastpath branch and the stub entry but in practice not many
1004   // methods reach the 1MB size.
1005   const BufferSizingData* sizing = output->buffer_sizing_data();
1006   const int code_size = sizing->_code + state->stubs_current_total_size();
1007 
1008   // Maximum backward range is 1M. Maximum forward reach is 1M - 4bytes.
1009   // Subtract 2K to be ultra conservative.
1010   const int cond_branch_max_reach = (int)(1*M - 2*K);
1011   _needs_far_jump = code_size >= cond_branch_max_reach;
1012 }
1013 
1014 #endif // COMPILER2