1 /*
   2  * Copyright (c) 2018, 2021, Red Hat, Inc. All rights reserved.
   3  * Copyright (c) 2012, 2022 SAP SE. All rights reserved.
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This code is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License version 2 only, as
   8  * published by the Free Software Foundation.
   9  *
  10  * This code is distributed in the hope that it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  13  * version 2 for more details (a copy is included in the LICENSE file that
  14  * accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License version
  17  * 2 along with this work; if not, write to the Free Software Foundation,
  18  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  19  *
  20  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  21  * or visit www.oracle.com if you need additional information or have any
  22  * questions.
  23  *
  24  */
  25 
  26 #include "gc/shared/gcArguments.hpp"
  27 #include "gc/shared/gc_globals.hpp"
  28 #include "macroAssembler_ppc.hpp"
  29 #include "precompiled.hpp"
  30 #include "asm/macroAssembler.inline.hpp"
  31 #include "gc/shenandoah/shenandoahBarrierSet.hpp"
  32 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
  33 #include "gc/shenandoah/shenandoahForwarding.hpp"
  34 #include "gc/shenandoah/shenandoahHeap.hpp"
  35 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
  36 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
  37 #include "gc/shenandoah/shenandoahRuntime.hpp"
  38 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
  39 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
  40 #include "gc/shenandoah/mode/shenandoahMode.hpp"
  41 #include "interpreter/interpreter.hpp"
  42 #include "runtime/javaThread.hpp"
  43 #include "runtime/sharedRuntime.hpp"
  44 #include "utilities/globalDefinitions.hpp"
  45 #include "vm_version_ppc.hpp"
  46 #ifdef COMPILER1
  47 #include "c1/c1_LIRAssembler.hpp"
  48 #include "c1/c1_MacroAssembler.hpp"
  49 #include "gc/shenandoah/c1/shenandoahBarrierSetC1.hpp"
  50 #endif
  51 
  52 #define __ masm->
  53 
  54 void ShenandoahBarrierSetAssembler::satb_write_barrier(MacroAssembler *masm,
  55                                                        Register base, RegisterOrConstant ind_or_offs,
  56                                                        Register tmp1, Register tmp2, Register tmp3,
  57                                                        MacroAssembler::PreservationLevel preservation_level) {
  58   if (ShenandoahSATBBarrier) {
  59     __ block_comment("satb_write_barrier (shenandoahgc) {");
  60     satb_write_barrier_impl(masm, 0, base, ind_or_offs, tmp1, tmp2, tmp3, preservation_level);
  61     __ block_comment("} satb_write_barrier (shenandoahgc)");
  62   }
  63 }
  64 
  65 void ShenandoahBarrierSetAssembler::load_reference_barrier(MacroAssembler *masm, DecoratorSet decorators,
  66                                                            Register base, RegisterOrConstant ind_or_offs,
  67                                                            Register dst,
  68                                                            Register tmp1, Register tmp2,
  69                                                            MacroAssembler::PreservationLevel preservation_level) {
  70   if (ShenandoahLoadRefBarrier) {
  71     __ block_comment("load_reference_barrier (shenandoahgc) {");
  72     load_reference_barrier_impl(masm, decorators, base, ind_or_offs, dst, tmp1, tmp2, preservation_level);
  73     __ block_comment("} load_reference_barrier (shenandoahgc)");
  74   }
  75 }
  76 
  77 void ShenandoahBarrierSetAssembler::arraycopy_prologue(MacroAssembler *masm, DecoratorSet decorators, BasicType type,
  78                                                        Register src, Register dst, Register count,
  79                                                        Register preserve1, Register preserve2) {
  80   Register R11_tmp = R11_scratch1;
  81 
  82   assert_different_registers(src, dst, count, R11_tmp, noreg);
  83   if (preserve1 != noreg) {
  84     // Technically not required, but likely to indicate an error.
  85     assert_different_registers(preserve1, preserve2);
  86   }
  87 
  88   /* ==== Check whether barrier is required (optimizations) ==== */
  89   // Fast path: Component type of array is not a reference type.
  90   if (!is_reference_type(type)) {
  91     return;
  92   }
  93 
  94   bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0;
  95 
  96   // Fast path: No barrier required if for every barrier type, it is either disabled or would not store
  97   // any useful information.
  98   if ((!ShenandoahSATBBarrier || dest_uninitialized) && !ShenandoahLoadRefBarrier) {
  99     return;
 100   }
 101 
 102   __ block_comment("arraycopy_prologue (shenandoahgc) {");
 103   Label skip_prologue;
 104 
 105   // Fast path: Array is of length zero.
 106   __ cmpdi(CCR0, count, 0);
 107   __ beq(CCR0, skip_prologue);
 108 
 109   /* ==== Check whether barrier is required (gc state) ==== */
 110   __ lbz(R11_tmp, in_bytes(ShenandoahThreadLocalData::gc_state_offset()),
 111          R16_thread);
 112 
 113   // The set of garbage collection states requiring barriers depends on the available barrier types and the
 114   // type of the reference in question.
 115   // For instance, satb barriers may be skipped if it is certain that the overridden values are not relevant
 116   // for the garbage collector.
 117   const int required_states = ShenandoahSATBBarrier && dest_uninitialized
 118                               ? ShenandoahHeap::HAS_FORWARDED
 119                               : ShenandoahHeap::HAS_FORWARDED | ShenandoahHeap::MARKING;
 120 
 121   __ andi_(R11_tmp, R11_tmp, required_states);
 122   __ beq(CCR0, skip_prologue);
 123 
 124   /* ==== Invoke runtime ==== */
 125   // Save to-be-preserved registers.
 126   int highest_preserve_register_index = 0;
 127   {
 128     if (preserve1 != noreg && preserve1->is_volatile()) {
 129       __ std(preserve1, -BytesPerWord * ++highest_preserve_register_index, R1_SP);
 130     }
 131     if (preserve2 != noreg && preserve2 != preserve1 && preserve2->is_volatile()) {
 132       __ std(preserve2, -BytesPerWord * ++highest_preserve_register_index, R1_SP);
 133     }
 134 
 135     __ std(src, -BytesPerWord * ++highest_preserve_register_index, R1_SP);
 136     __ std(dst, -BytesPerWord * ++highest_preserve_register_index, R1_SP);
 137     __ std(count, -BytesPerWord * ++highest_preserve_register_index, R1_SP);
 138 
 139     __ save_LR_CR(R11_tmp);
 140     __ push_frame_reg_args(-BytesPerWord * highest_preserve_register_index,
 141                            R11_tmp);
 142   }
 143 
 144   // Invoke runtime.
 145   address jrt_address = nullptr;
 146   if (UseCompressedOops) {
 147     jrt_address = CAST_FROM_FN_PTR(address, ShenandoahRuntime::arraycopy_barrier_narrow_oop_entry);
 148   } else {
 149     jrt_address = CAST_FROM_FN_PTR(address, ShenandoahRuntime::arraycopy_barrier_oop_entry);
 150   }
 151   assert(jrt_address != nullptr, "jrt routine cannot be found");
 152 
 153   __ call_VM_leaf(jrt_address, src, dst, count);
 154 
 155   // Restore to-be-preserved registers.
 156   {
 157     __ pop_frame();
 158     __ restore_LR_CR(R11_tmp);
 159 
 160     __ ld(count, -BytesPerWord * highest_preserve_register_index--, R1_SP);
 161     __ ld(dst, -BytesPerWord * highest_preserve_register_index--, R1_SP);
 162     __ ld(src, -BytesPerWord * highest_preserve_register_index--, R1_SP);
 163 
 164     if (preserve2 != noreg && preserve2 != preserve1 && preserve2->is_volatile()) {
 165       __ ld(preserve2, -BytesPerWord * highest_preserve_register_index--, R1_SP);
 166     }
 167     if (preserve1 != noreg && preserve1->is_volatile()) {
 168       __ ld(preserve1, -BytesPerWord * highest_preserve_register_index--, R1_SP);
 169     }
 170   }
 171 
 172   __ bind(skip_prologue);
 173   __ block_comment("} arraycopy_prologue (shenandoahgc)");
 174 }
 175 
 176 void ShenandoahBarrierSetAssembler::arraycopy_epilogue(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 177                                                        Register dst, Register count,
 178                                                        Register preserve) {
 179   if (ShenandoahCardBarrier && is_reference_type(type)) {
 180     __ block_comment("arraycopy_epilogue (shenandoahgc) {");
 181     gen_write_ref_array_post_barrier(masm, decorators, dst, count, preserve);
 182     __ block_comment("} arraycopy_epilogue (shenandoahgc)");
 183   }
 184 }
 185 
 186 // The to-be-enqueued value can either be determined
 187 // - dynamically by passing the reference's address information (load mode) or
 188 // - statically by passing a register the value is stored in (preloaded mode)
 189 //   - for performance optimizations in cases where the previous value is known (currently not implemented) and
 190 //   - for incremental-update barriers.
 191 //
 192 // decorators:  The previous value's decorator set.
 193 //              In "load mode", the value must equal '0'.
 194 // base:        Base register of the reference's address (load mode).
 195 //              In "preloaded mode", the register must equal 'noreg'.
 196 // ind_or_offs: Index or offset of the reference's address (load mode).
 197 //              If 'base' equals 'noreg' (preloaded mode), the passed value is ignored.
 198 // pre_val:     Register holding the to-be-stored value (preloaded mode).
 199 //              In "load mode", this register acts as a temporary register and must
 200 //              thus not be 'noreg'.  In "preloaded mode", its content will be sustained.
 201 // tmp1/tmp2:   Temporary registers, one of which must be non-volatile in "preloaded mode".
 202 void ShenandoahBarrierSetAssembler::satb_write_barrier_impl(MacroAssembler *masm, DecoratorSet decorators,
 203                                                             Register base, RegisterOrConstant ind_or_offs,
 204                                                             Register pre_val,
 205                                                             Register tmp1, Register tmp2,
 206                                                             MacroAssembler::PreservationLevel preservation_level) {
 207   assert_different_registers(tmp1, tmp2, pre_val, noreg);
 208 
 209   Label skip_barrier;
 210 
 211   /* ==== Determine necessary runtime invocation preservation measures ==== */
 212   const bool needs_frame           = preservation_level >= MacroAssembler::PRESERVATION_FRAME_LR;
 213   const bool preserve_gp_registers = preservation_level >= MacroAssembler::PRESERVATION_FRAME_LR_GP_REGS;
 214   const bool preserve_fp_registers = preservation_level >= MacroAssembler::PRESERVATION_FRAME_LR_GP_FP_REGS;
 215 
 216   // Check whether marking is active.
 217   __ lbz(tmp1, in_bytes(ShenandoahThreadLocalData::gc_state_offset()), R16_thread);
 218 
 219   __ andi_(tmp1, tmp1, ShenandoahHeap::MARKING);
 220   __ beq(CCR0, skip_barrier);
 221 
 222   /* ==== Determine the reference's previous value ==== */
 223   bool preloaded_mode = base == noreg;
 224   Register pre_val_save = noreg;
 225 
 226   if (preloaded_mode) {
 227     // Previous value has been passed to the method, so it must not be determined manually.
 228     // In case 'pre_val' is a volatile register, it must be saved across the C-call
 229     // as callers may depend on its value.
 230     // Unless the general purposes registers are saved anyway, one of the temporary registers
 231     // (i.e., 'tmp1' and 'tmp2') is used to the preserve 'pre_val'.
 232     if (!preserve_gp_registers && pre_val->is_volatile()) {
 233       pre_val_save = !tmp1->is_volatile() ? tmp1 : tmp2;
 234       assert(!pre_val_save->is_volatile(), "at least one of the temporary registers must be non-volatile");
 235     }
 236 
 237     if ((decorators & IS_NOT_NULL) != 0) {
 238 #ifdef ASSERT
 239       __ cmpdi(CCR0, pre_val, 0);
 240       __ asm_assert_ne("null oop is not allowed");
 241 #endif // ASSERT
 242     } else {
 243       __ cmpdi(CCR0, pre_val, 0);
 244       __ beq(CCR0, skip_barrier);
 245     }
 246   } else {
 247     // Load from the reference address to determine the reference's current value (before the store is being performed).
 248     // Contrary to the given value in "preloaded mode", it is not necessary to preserve it.
 249     assert(decorators == 0, "decorator set must be empty");
 250     assert(base != noreg, "base must be a register");
 251     assert(!ind_or_offs.is_register() || ind_or_offs.as_register() != noreg, "ind_or_offs must be a register");
 252     if (UseCompressedOops) {
 253       __ lwz(pre_val, ind_or_offs, base);
 254     } else {
 255       __ ld(pre_val, ind_or_offs, base);
 256     }
 257 
 258     __ cmpdi(CCR0, pre_val, 0);
 259     __ beq(CCR0, skip_barrier);
 260 
 261     if (UseCompressedOops) {
 262       __ decode_heap_oop_not_null(pre_val);
 263     }
 264   }
 265 
 266   /* ==== Try to enqueue the to-be-stored value directly into thread's local SATB mark queue ==== */
 267   {
 268     Label runtime;
 269     Register Rbuffer = tmp1, Rindex = tmp2;
 270 
 271     // Check whether the queue has enough capacity to store another oop.
 272     // If not, jump to the runtime to commit the buffer and to allocate a new one.
 273     // (The buffer's index corresponds to the amount of remaining free space.)
 274     __ ld(Rindex, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()), R16_thread);
 275     __ cmpdi(CCR0, Rindex, 0);
 276     __ beq(CCR0, runtime); // If index == 0 (buffer is full), goto runtime.
 277 
 278     // Capacity suffices.  Decrement the queue's size by the size of one oop.
 279     // (The buffer is filled contrary to the heap's growing direction, i.e., it is filled downwards.)
 280     __ addi(Rindex, Rindex, -wordSize);
 281     __ std(Rindex, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()), R16_thread);
 282 
 283     // Enqueue the previous value and skip the invocation of the runtime.
 284     __ ld(Rbuffer, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()), R16_thread);
 285     __ stdx(pre_val, Rbuffer, Rindex);
 286     __ b(skip_barrier);
 287 
 288     __ bind(runtime);
 289   }
 290 
 291   /* ==== Invoke runtime to commit SATB mark queue to gc and allocate a new buffer ==== */
 292   // Save to-be-preserved registers.
 293   int nbytes_save = 0;
 294 
 295   if (needs_frame) {
 296     if (preserve_gp_registers) {
 297       nbytes_save = (preserve_fp_registers
 298                      ? MacroAssembler::num_volatile_gp_regs + MacroAssembler::num_volatile_fp_regs
 299                      : MacroAssembler::num_volatile_gp_regs) * BytesPerWord;
 300       __ save_volatile_gprs(R1_SP, -nbytes_save, preserve_fp_registers);
 301     }
 302 
 303     __ save_LR_CR(tmp1);
 304     __ push_frame_reg_args(nbytes_save, tmp2);
 305   }
 306 
 307   if (!preserve_gp_registers && preloaded_mode && pre_val->is_volatile()) {
 308     assert(pre_val_save != noreg, "nv_save must not be noreg");
 309 
 310     // 'pre_val' register must be saved manually unless general-purpose are preserved in general.
 311     __ mr(pre_val_save, pre_val);
 312   }
 313 
 314   // Invoke runtime.
 315   __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), pre_val, R16_thread);
 316 
 317   // Restore to-be-preserved registers.
 318   if (!preserve_gp_registers && preloaded_mode && pre_val->is_volatile()) {
 319     __ mr(pre_val, pre_val_save);
 320   }
 321 
 322   if (needs_frame) {
 323     __ pop_frame();
 324     __ restore_LR_CR(tmp1);
 325 
 326     if (preserve_gp_registers) {
 327       __ restore_volatile_gprs(R1_SP, -nbytes_save, preserve_fp_registers);
 328     }
 329   }
 330 
 331   __ bind(skip_barrier);
 332 }
 333 
 334 void ShenandoahBarrierSetAssembler::resolve_forward_pointer_not_null(MacroAssembler *masm, Register dst, Register tmp) {
 335   __ block_comment("resolve_forward_pointer_not_null (shenandoahgc) {");
 336 
 337   Register tmp1 = tmp,
 338            R0_tmp2 = R0;
 339   assert_different_registers(dst, tmp1, R0_tmp2, noreg);
 340 
 341   // If the object has been evacuated, the mark word layout is as follows:
 342   // | forwarding pointer (62-bit) | '11' (2-bit) |
 343 
 344   // The invariant that stack/thread pointers have the lowest two bits cleared permits retrieving
 345   // the forwarding pointer solely by inversing the lowest two bits.
 346   // This invariant follows inevitably from hotspot's minimal alignment.
 347   assert(markWord::marked_value <= (unsigned long) MinObjAlignmentInBytes,
 348          "marked value must not be higher than hotspot's minimal alignment");
 349 
 350   Label done;
 351 
 352   // Load the object's mark word.
 353   __ ld(tmp1, oopDesc::mark_offset_in_bytes(), dst);
 354 
 355   // Load the bit mask for the lock bits.
 356   __ li(R0_tmp2, markWord::lock_mask_in_place);
 357 
 358   // Check whether all bits matching the bit mask are set.
 359   // If that is the case, the object has been evacuated and the most significant bits form the forward pointer.
 360   __ andc_(R0_tmp2, R0_tmp2, tmp1);
 361 
 362   assert(markWord::lock_mask_in_place == markWord::marked_value,
 363          "marked value must equal the value obtained when all lock bits are being set");
 364   if (VM_Version::has_isel()) {
 365     __ xori(tmp1, tmp1, markWord::lock_mask_in_place);
 366     __ isel(dst, CCR0, Assembler::equal, false, tmp1);
 367   } else {
 368     __ bne(CCR0, done);
 369     __ xori(dst, tmp1, markWord::lock_mask_in_place);
 370   }
 371 
 372   __ bind(done);
 373   __ block_comment("} resolve_forward_pointer_not_null (shenandoahgc)");
 374 }
 375 
 376 // base:        Base register of the reference's address.
 377 // ind_or_offs: Index or offset of the reference's address (load mode).
 378 // dst:         Reference's address.  In case the object has been evacuated, this is the to-space version
 379 //              of that object.
 380 void ShenandoahBarrierSetAssembler::load_reference_barrier_impl(
 381     MacroAssembler *masm, DecoratorSet decorators,
 382     Register base, RegisterOrConstant ind_or_offs,
 383     Register dst,
 384     Register tmp1, Register tmp2,
 385     MacroAssembler::PreservationLevel preservation_level) {
 386   if (ind_or_offs.is_register()) {
 387     assert_different_registers(tmp1, tmp2, base, ind_or_offs.as_register(), dst, noreg);
 388   } else {
 389     assert_different_registers(tmp1, tmp2, base, dst, noreg);
 390   }
 391 
 392   Label skip_barrier;
 393 
 394   bool is_strong  = ShenandoahBarrierSet::is_strong_access(decorators);
 395   bool is_weak    = ShenandoahBarrierSet::is_weak_access(decorators);
 396   bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
 397   bool is_native  = ShenandoahBarrierSet::is_native_access(decorators);
 398   bool is_narrow  = UseCompressedOops && !is_native;
 399 
 400   /* ==== Check whether heap is stable ==== */
 401   __ lbz(tmp2, in_bytes(ShenandoahThreadLocalData::gc_state_offset()), R16_thread);
 402 
 403   if (is_strong) {
 404     // For strong references, the heap is considered stable if "has forwarded" is not active.
 405     __ andi_(tmp1, tmp2, ShenandoahHeap::HAS_FORWARDED | ShenandoahHeap::EVACUATION);
 406     __ beq(CCR0, skip_barrier);
 407 #ifdef ASSERT
 408     // "evacuation" -> (implies) "has forwarded".  If we reach this code, "has forwarded" must thus be set.
 409     __ andi_(tmp1, tmp1, ShenandoahHeap::HAS_FORWARDED);
 410     __ asm_assert_ne("'has forwarded' is missing");
 411 #endif // ASSERT
 412   } else {
 413     // For all non-strong references, the heap is considered stable if not any of "has forwarded",
 414     // "root set processing", and "weak reference processing" is active.
 415     // The additional phase conditions are in place to avoid the resurrection of weak references (see JDK-8266440).
 416     Label skip_fastpath;
 417     __ andi_(tmp1, tmp2, ShenandoahHeap::WEAK_ROOTS);
 418     __ bne(CCR0, skip_fastpath);
 419 
 420     __ andi_(tmp1, tmp2, ShenandoahHeap::HAS_FORWARDED | ShenandoahHeap::EVACUATION);
 421     __ beq(CCR0, skip_barrier);
 422 #ifdef ASSERT
 423     // "evacuation" -> (implies) "has forwarded".  If we reach this code, "has forwarded" must thus be set.
 424     __ andi_(tmp1, tmp1, ShenandoahHeap::HAS_FORWARDED);
 425     __ asm_assert_ne("'has forwarded' is missing");
 426 #endif // ASSERT
 427 
 428     __ bind(skip_fastpath);
 429   }
 430 
 431   /* ==== Check whether region is in collection set ==== */
 432   if (is_strong) {
 433     // Shenandoah stores metadata on regions in a continuous area of memory in which a single byte corresponds to
 434     // an entire region of the shenandoah heap.  At present, only the least significant bit is of significance
 435     // and indicates whether the region is part of the collection set.
 436     //
 437     // All regions are of the same size and are always aligned by a power of two.
 438     // Any address can thus be shifted by a fixed number of bits to retrieve the address prefix shared by
 439     // all objects within that region (region identification bits).
 440     //
 441     //  | unused bits | region identification bits | object identification bits |
 442     //  (Region size depends on a couple of criteria, such as page size, user-provided arguments and the max heap size.
 443     //   The number of object identification bits can thus not be determined at compile time.)
 444     //
 445     // -------------------------------------------------------  <--- cs (collection set) base address
 446     // | lost space due to heap space base address                   -> 'ShenandoahHeap::in_cset_fast_test_addr()'
 447     // | (region identification bits contain heap base offset)
 448     // |------------------------------------------------------  <--- cs base address + (heap_base >> region size shift)
 449     // | collection set in the proper                                -> shift: 'region_size_bytes_shift_jint()'
 450     // |
 451     // |------------------------------------------------------  <--- cs base address + (heap_base >> region size shift)
 452     //                                                                               + number of regions
 453     __ load_const_optimized(tmp2, ShenandoahHeap::in_cset_fast_test_addr(), tmp1);
 454     __ srdi(tmp1, dst, ShenandoahHeapRegion::region_size_bytes_shift_jint());
 455     __ lbzx(tmp2, tmp1, tmp2);
 456     __ andi_(tmp2, tmp2, 1);
 457     __ beq(CCR0, skip_barrier);
 458   }
 459 
 460   /* ==== Invoke runtime ==== */
 461   // Save to-be-preserved registers.
 462   int nbytes_save = 0;
 463 
 464   const bool needs_frame           = preservation_level >= MacroAssembler::PRESERVATION_FRAME_LR;
 465   const bool preserve_gp_registers = preservation_level >= MacroAssembler::PRESERVATION_FRAME_LR_GP_REGS;
 466   const bool preserve_fp_registers = preservation_level >= MacroAssembler::PRESERVATION_FRAME_LR_GP_FP_REGS;
 467 
 468   if (needs_frame) {
 469     if (preserve_gp_registers) {
 470       nbytes_save = (preserve_fp_registers
 471                      ? MacroAssembler::num_volatile_gp_regs + MacroAssembler::num_volatile_fp_regs
 472                      : MacroAssembler::num_volatile_gp_regs) * BytesPerWord;
 473       __ save_volatile_gprs(R1_SP, -nbytes_save, preserve_fp_registers);
 474     }
 475 
 476     __ save_LR_CR(tmp1);
 477     __ push_frame_reg_args(nbytes_save, tmp1);
 478   }
 479 
 480   // Calculate the reference's absolute address.
 481   __ add(R4_ARG2, ind_or_offs, base);
 482 
 483   // Invoke runtime.
 484   address jrt_address = nullptr;
 485 
 486   if (is_strong) {
 487     if (is_narrow) {
 488       jrt_address = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow);
 489     } else {
 490       jrt_address = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong);
 491     }
 492   } else if (is_weak) {
 493     if (is_narrow) {
 494       jrt_address = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow);
 495     } else {
 496       jrt_address = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak);
 497     }
 498   } else {
 499     assert(is_phantom, "only remaining strength");
 500     assert(!is_narrow, "phantom access cannot be narrow");
 501     jrt_address = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom);
 502   }
 503   assert(jrt_address != nullptr, "jrt routine cannot be found");
 504 
 505   __ call_VM_leaf(jrt_address, dst /* reference */, R4_ARG2 /* reference address */);
 506 
 507   // Restore to-be-preserved registers.
 508   if (preserve_gp_registers) {
 509     __ mr(R0, R3_RET);
 510   } else {
 511     __ mr_if_needed(dst, R3_RET);
 512   }
 513 
 514   if (needs_frame) {
 515     __ pop_frame();
 516     __ restore_LR_CR(tmp1);
 517 
 518     if (preserve_gp_registers) {
 519       __ restore_volatile_gprs(R1_SP, -nbytes_save, preserve_fp_registers);
 520       __ mr(dst, R0);
 521     }
 522   }
 523 
 524   __ bind(skip_barrier);
 525 }
 526 
 527 // base:           Base register of the reference's address.
 528 // ind_or_offs:    Index or offset of the reference's address.
 529 // L_handle_null:  An optional label that will be jumped to if the reference is null.
 530 void ShenandoahBarrierSetAssembler::load_at(
 531     MacroAssembler *masm, DecoratorSet decorators, BasicType type,
 532     Register base, RegisterOrConstant ind_or_offs, Register dst,
 533     Register tmp1, Register tmp2,
 534     MacroAssembler::PreservationLevel preservation_level, Label *L_handle_null) {
 535   // Register must not clash, except 'base' and 'dst'.
 536   if (ind_or_offs.is_register()) {
 537     if (base != noreg) {
 538       assert_different_registers(tmp1, tmp2, base, ind_or_offs.register_or_noreg(), R0, noreg);
 539     }
 540     assert_different_registers(tmp1, tmp2, dst, ind_or_offs.register_or_noreg(), R0, noreg);
 541   } else {
 542     if (base == noreg) {
 543       assert_different_registers(tmp1, tmp2, base, R0, noreg);
 544     }
 545     assert_different_registers(tmp1, tmp2, dst, R0, noreg);
 546   }
 547 
 548   /* ==== Apply load barrier, if required ==== */
 549   if (ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) {
 550     assert(is_reference_type(type), "need_load_reference_barrier must check whether type is a reference type");
 551 
 552     // If 'dst' clashes with either 'base' or 'ind_or_offs', use an intermediate result register
 553     // to keep the values of those alive until the load reference barrier is applied.
 554     Register intermediate_dst = (dst == base || (ind_or_offs.is_register() && dst == ind_or_offs.as_register()))
 555                                 ? tmp2
 556                                 : dst;
 557 
 558     BarrierSetAssembler::load_at(masm, decorators, type,
 559                                  base, ind_or_offs,
 560                                  intermediate_dst,
 561                                  tmp1, noreg,
 562                                  preservation_level, L_handle_null);
 563 
 564     load_reference_barrier(masm, decorators,
 565                            base, ind_or_offs,
 566                            intermediate_dst,
 567                            tmp1, R0,
 568                            preservation_level);
 569 
 570     __ mr_if_needed(dst, intermediate_dst);
 571   } else {
 572     BarrierSetAssembler::load_at(masm, decorators, type,
 573                                  base, ind_or_offs,
 574                                  dst,
 575                                  tmp1, tmp2,
 576                                  preservation_level, L_handle_null);
 577   }
 578 
 579   /* ==== Apply keep-alive barrier, if required (e.g., to inhibit weak reference resurrection) ==== */
 580   if (ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) {
 581     if (ShenandoahSATBBarrier) {
 582       __ block_comment("keep_alive_barrier (shenandoahgc) {");
 583       satb_write_barrier_impl(masm, 0, noreg, noreg, dst, tmp1, tmp2, preservation_level);
 584       __ block_comment("} keep_alive_barrier (shenandoahgc)");
 585     }
 586   }
 587 }
 588 
 589 void ShenandoahBarrierSetAssembler::store_check(MacroAssembler* masm, Register base, RegisterOrConstant ind_or_offs, Register tmp) {
 590   assert(ShenandoahCardBarrier, "Should have been checked by caller");
 591   assert_different_registers(base, tmp, R0);
 592 
 593   if (ind_or_offs.is_constant()) {
 594     __ add_const_optimized(base, base, ind_or_offs.as_constant(), tmp);
 595   } else {
 596     __ add(base, ind_or_offs.as_register(), base);
 597   }
 598 
 599   __ ld(tmp, in_bytes(ShenandoahThreadLocalData::card_table_offset()), R16_thread); /* tmp = *[R16_thread + card_table_offset] */
 600   __ srdi(base, base, CardTable::card_shift());
 601   __ li(R0, CardTable::dirty_card_val());
 602   __ stbx(R0, tmp, base);
 603 }
 604 
 605 // base:        Base register of the reference's address.
 606 // ind_or_offs: Index or offset of the reference's address.
 607 // val:         To-be-stored value/reference's new value.
 608 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler *masm, DecoratorSet decorators, BasicType type,
 609                                              Register base, RegisterOrConstant ind_or_offs, Register val,
 610                                              Register tmp1, Register tmp2, Register tmp3,
 611                                              MacroAssembler::PreservationLevel preservation_level) {
 612   if (is_reference_type(type)) {
 613     if (ShenandoahSATBBarrier) {
 614       satb_write_barrier(masm, base, ind_or_offs, tmp1, tmp2, tmp3, preservation_level);
 615     }
 616   }
 617 
 618   BarrierSetAssembler::store_at(masm, decorators, type,
 619                                 base, ind_or_offs,
 620                                 val,
 621                                 tmp1, tmp2, tmp3,
 622                                 preservation_level);
 623 
 624   // No need for post barrier if storing NULL
 625   if (ShenandoahCardBarrier && is_reference_type(type) && val != noreg) {
 626     store_check(masm, base, ind_or_offs, tmp1);
 627   }
 628 }
 629 
 630 void ShenandoahBarrierSetAssembler::try_resolve_jobject_in_native(MacroAssembler *masm,
 631                                                                   Register dst, Register jni_env, Register obj,
 632                                                                   Register tmp, Label &slowpath) {
 633   __ block_comment("try_resolve_jobject_in_native (shenandoahgc) {");
 634 
 635   assert_different_registers(jni_env, obj, tmp);
 636 
 637   Label done;
 638 
 639   // Fast path: Reference is null (JNI tags are zero for null pointers).
 640   __ cmpdi(CCR0, obj, 0);
 641   __ beq(CCR0, done);
 642 
 643   // Resolve jobject using standard implementation.
 644   BarrierSetAssembler::try_resolve_jobject_in_native(masm, dst, jni_env, obj, tmp, slowpath);
 645 
 646   // Check whether heap is stable.
 647   __ lbz(tmp,
 648          in_bytes(ShenandoahThreadLocalData::gc_state_offset() - JavaThread::jni_environment_offset()),
 649          jni_env);
 650 
 651   __ andi_(tmp, tmp, ShenandoahHeap::EVACUATION | ShenandoahHeap::HAS_FORWARDED);
 652   __ bne(CCR0, slowpath);
 653 
 654   __ bind(done);
 655   __ block_comment("} try_resolve_jobject_in_native (shenandoahgc)");
 656 }
 657 
 658 // Special shenandoah CAS implementation that handles false negatives due
 659 // to concurrent evacuation.  That is, the CAS operation is intended to succeed in
 660 // the following scenarios (success criteria):
 661 //  s1) The reference pointer ('base_addr') equals the expected ('expected') pointer.
 662 //  s2) The reference pointer refers to the from-space version of an already-evacuated
 663 //      object, whereas the expected pointer refers to the to-space version of the same object.
 664 // Situations in which the reference pointer refers to the to-space version of an object
 665 // and the expected pointer refers to the from-space version of the same object can not occur due to
 666 // shenandoah's strong to-space invariant.  This also implies that the reference stored in 'new_val'
 667 // can not refer to the from-space version of an already-evacuated object.
 668 //
 669 // To guarantee correct behavior in concurrent environments, two races must be addressed:
 670 //  r1) A concurrent thread may heal the reference pointer (i.e., it is no longer referring to the
 671 //      from-space version but to the to-space version of the object in question).
 672 //      In this case, the CAS operation should succeed.
 673 //  r2) A concurrent thread may mutate the reference (i.e., the reference pointer refers to an entirely different object).
 674 //      In this case, the CAS operation should fail.
 675 //
 676 // By default, the value held in the 'result' register is zero to indicate failure of CAS,
 677 // non-zero to indicate success.  If 'is_cae' is set, the result is the most recently fetched
 678 // value from 'base_addr' rather than a boolean success indicator.
 679 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler *masm, Register base_addr,
 680                                                 Register expected, Register new_val, Register tmp1, Register tmp2,
 681                                                 bool is_cae, Register result) {
 682   __ block_comment("cmpxchg_oop (shenandoahgc) {");
 683 
 684   assert_different_registers(base_addr, new_val, tmp1, tmp2, result, R0);
 685   assert_different_registers(base_addr, expected, tmp1, tmp2, result, R0);
 686 
 687   // Potential clash of 'success_flag' and 'tmp' is being accounted for.
 688   Register success_flag  = is_cae ? noreg  : result,
 689            current_value = is_cae ? result : tmp1,
 690            tmp           = is_cae ? tmp1   : result,
 691            initial_value = tmp2;
 692 
 693   Label done, step_four;
 694 
 695   __ bind(step_four);
 696 
 697   /* ==== Step 1 ("Standard" CAS) ==== */
 698   // Fast path: The values stored in 'expected' and 'base_addr' are equal.
 699   // Given that 'expected' must refer to the to-space object of an evacuated object (strong to-space invariant),
 700   // no special processing is required.
 701   if (UseCompressedOops) {
 702     __ cmpxchgw(CCR0, current_value, expected, new_val, base_addr, MacroAssembler::MemBarNone,
 703                 false, success_flag, true);
 704   } else {
 705     __ cmpxchgd(CCR0, current_value, expected, new_val, base_addr, MacroAssembler::MemBarNone,
 706                 false, success_flag, nullptr, true);
 707   }
 708 
 709   // Skip the rest of the barrier if the CAS operation succeeds immediately.
 710   // If it does not, the value stored at the address is either the from-space pointer of the
 711   // referenced object (success criteria s2)) or simply another object.
 712   __ beq(CCR0, done);
 713 
 714   /* ==== Step 2 (Null check) ==== */
 715   // The success criteria s2) cannot be matched with a null pointer
 716   // (null pointers cannot be subject to concurrent evacuation).  The failure of the CAS operation is thus legitimate.
 717   __ cmpdi(CCR0, current_value, 0);
 718   __ beq(CCR0, done);
 719 
 720   /* ==== Step 3 (reference pointer refers to from-space version; success criteria s2)) ==== */
 721   // To check whether the reference pointer refers to the from-space version, the forward
 722   // pointer of the object referred to by the reference is resolved and compared against the expected pointer.
 723   // If this check succeed, another CAS operation is issued with the from-space pointer being the expected pointer.
 724   //
 725   // Save the potential from-space pointer.
 726   __ mr(initial_value, current_value);
 727 
 728   // Resolve forward pointer.
 729   if (UseCompressedOops) { __ decode_heap_oop_not_null(current_value); }
 730   resolve_forward_pointer_not_null(masm, current_value, tmp);
 731   if (UseCompressedOops) { __ encode_heap_oop_not_null(current_value); }
 732 
 733   if (!is_cae) {
 734     // 'success_flag' was overwritten by call to 'resovle_forward_pointer_not_null'.
 735     // Load zero into register for the potential failure case.
 736     __ li(success_flag, 0);
 737   }
 738   __ cmpd(CCR0, current_value, expected);
 739   __ bne(CCR0, done);
 740 
 741   // Discard fetched value as it might be a reference to the from-space version of an object.
 742   if (UseCompressedOops) {
 743     __ cmpxchgw(CCR0, R0, initial_value, new_val, base_addr, MacroAssembler::MemBarNone,
 744                 false, success_flag);
 745   } else {
 746     __ cmpxchgd(CCR0, R0, initial_value, new_val, base_addr, MacroAssembler::MemBarNone,
 747                 false, success_flag);
 748   }
 749 
 750   /* ==== Step 4 (Retry CAS with to-space pointer (success criteria s2) under race r1)) ==== */
 751   // The reference pointer could have been healed whilst the previous CAS operation was being performed.
 752   // Another CAS operation must thus be issued with the to-space pointer being the expected pointer.
 753   // If that CAS operation fails as well, race r2) must have occurred, indicating that
 754   // the operation failure is legitimate.
 755   //
 756   // To keep the code's size small and thus improving cache (icache) performance, this highly
 757   // unlikely case should be handled by the smallest possible code.  Instead of emitting a third,
 758   // explicit CAS operation, the code jumps back and reuses the first CAS operation (step 1)
 759   // (passed arguments are identical).
 760   //
 761   // A failure of the CAS operation in step 1 would imply that the overall CAS operation is supposed
 762   // to fail.  Jumping back to step 1 requires, however, that step 2 and step 3 are re-executed as well.
 763   // It is thus important to ensure that a re-execution of those steps does not put program correctness
 764   // at risk:
 765   // - Step 2: Either terminates in failure (desired result) or falls through to step 3.
 766   // - Step 3: Terminates if the comparison between the forwarded, fetched pointer and the expected value
 767   //           fails.  Unless the reference has been updated in the meanwhile once again, this is
 768   //           guaranteed to be the case.
 769   //           In case of a concurrent update, the CAS would be retried again. This is legitimate
 770   //           in terms of program correctness (even though it is not desired).
 771   __ bne(CCR0, step_four);
 772 
 773   __ bind(done);
 774   __ block_comment("} cmpxchg_oop (shenandoahgc)");
 775 }
 776 
 777 void ShenandoahBarrierSetAssembler::gen_write_ref_array_post_barrier(MacroAssembler* masm, DecoratorSet decorators,
 778                                                                      Register addr, Register count, Register preserve) {
 779   assert(ShenandoahCardBarrier, "Should have been checked by caller");
 780 
 781   ShenandoahBarrierSet* bs = ShenandoahBarrierSet::barrier_set();
 782   CardTable* ct = bs->card_table();
 783   assert_different_registers(addr, count, R0);
 784 
 785   Label L_skip_loop, L_store_loop;
 786 
 787   __ sldi_(count, count, LogBytesPerHeapOop);
 788 
 789   // Zero length? Skip.
 790   __ beq(CCR0, L_skip_loop);
 791 
 792   __ addi(count, count, -BytesPerHeapOop);
 793   __ add(count, addr, count);
 794   // Use two shifts to clear out those low order two bits! (Cannot opt. into 1.)
 795   __ srdi(addr, addr, CardTable::card_shift());
 796   __ srdi(count, count, CardTable::card_shift());
 797   __ subf(count, addr, count);
 798   __ ld(R0, in_bytes(ShenandoahThreadLocalData::card_table_offset()), R16_thread);
 799   __ add(addr, addr, R0);
 800   __ addi(count, count, 1);
 801   __ li(R0, 0);
 802   __ mtctr(count);
 803 
 804   // Byte store loop
 805   __ bind(L_store_loop);
 806   __ stb(R0, 0, addr);
 807   __ addi(addr, addr, 1);
 808   __ bdnz(L_store_loop);
 809   __ bind(L_skip_loop);
 810 }
 811 
 812 #undef __
 813 
 814 #ifdef COMPILER1
 815 
 816 #define __ ce->masm()->
 817 
 818 void ShenandoahBarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler *ce, ShenandoahPreBarrierStub *stub) {
 819   __ block_comment("gen_pre_barrier_stub (shenandoahgc) {");
 820 
 821   ShenandoahBarrierSetC1 *bs = (ShenandoahBarrierSetC1*) BarrierSet::barrier_set()->barrier_set_c1();
 822   __ bind(*stub->entry());
 823 
 824   // GC status has already been verified by 'ShenandoahBarrierSetC1::pre_barrier'.
 825   // This stub is the slowpath of that function.
 826 
 827   assert(stub->pre_val()->is_register(), "pre_val must be a register");
 828   Register pre_val = stub->pre_val()->as_register();
 829 
 830   // If 'do_load()' returns false, the to-be-stored value is already available in 'stub->pre_val()'
 831   // ("preloaded mode" of the store barrier).
 832   if (stub->do_load()) {
 833     ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false);
 834   }
 835 
 836   // Fast path: Reference is null.
 837   __ cmpdi(CCR0, pre_val, 0);
 838   __ bc_far_optimized(Assembler::bcondCRbiIs1_bhintNoHint, __ bi0(CCR0, Assembler::equal), *stub->continuation());
 839 
 840   // Argument passing via the stack.
 841   __ std(pre_val, -8, R1_SP);
 842 
 843   __ load_const_optimized(R0, bs->pre_barrier_c1_runtime_code_blob()->code_begin());
 844   __ call_stub(R0);
 845 
 846   __ b(*stub->continuation());
 847   __ block_comment("} gen_pre_barrier_stub (shenandoahgc)");
 848 }
 849 
 850 void ShenandoahBarrierSetAssembler::gen_load_reference_barrier_stub(LIR_Assembler *ce,
 851                                                                     ShenandoahLoadReferenceBarrierStub *stub) {
 852   __ block_comment("gen_load_reference_barrier_stub (shenandoahgc) {");
 853 
 854   ShenandoahBarrierSetC1 *bs = (ShenandoahBarrierSetC1*) BarrierSet::barrier_set()->barrier_set_c1();
 855   __ bind(*stub->entry());
 856 
 857   Register obj  = stub->obj()->as_register();
 858   Register res  = stub->result()->as_register();
 859   Register addr = stub->addr()->as_pointer_register();
 860   Register tmp1 = stub->tmp1()->as_register();
 861   Register tmp2 = stub->tmp2()->as_register();
 862   assert_different_registers(addr, res, tmp1, tmp2);
 863 
 864 #ifdef ASSERT
 865   // Ensure that 'res' is 'R3_ARG1' and contains the same value as 'obj' to reduce the number of required
 866   // copy instructions.
 867   assert(R3_RET == res, "res must be r3");
 868   __ cmpd(CCR0, res, obj);
 869   __ asm_assert_eq("result register must contain the reference stored in obj");
 870 #endif
 871 
 872   DecoratorSet decorators = stub->decorators();
 873 
 874   /* ==== Check whether region is in collection set ==== */
 875   // GC status (unstable) has already been verified by 'ShenandoahBarrierSetC1::load_reference_barrier_impl'.
 876   // This stub is the slowpath of that function.
 877 
 878   bool is_strong  = ShenandoahBarrierSet::is_strong_access(decorators);
 879   bool is_weak    = ShenandoahBarrierSet::is_weak_access(decorators);
 880   bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
 881   bool is_native  = ShenandoahBarrierSet::is_native_access(decorators);
 882 
 883   if (is_strong) {
 884     // Check whether object is in collection set.
 885     __ load_const_optimized(tmp2, ShenandoahHeap::in_cset_fast_test_addr(), tmp1);
 886     __ srdi(tmp1, obj, ShenandoahHeapRegion::region_size_bytes_shift_jint());
 887     __ lbzx(tmp2, tmp1, tmp2);
 888 
 889     __ andi_(tmp2, tmp2, 1);
 890     __ bc_far_optimized(Assembler::bcondCRbiIs1_bhintNoHint, __ bi0(CCR0, Assembler::equal), *stub->continuation());
 891   }
 892 
 893   address blob_addr = nullptr;
 894 
 895   if (is_strong) {
 896     if (is_native) {
 897       blob_addr = bs->load_reference_barrier_strong_native_rt_code_blob()->code_begin();
 898     } else {
 899       blob_addr = bs->load_reference_barrier_strong_rt_code_blob()->code_begin();
 900     }
 901   } else if (is_weak) {
 902     blob_addr = bs->load_reference_barrier_weak_rt_code_blob()->code_begin();
 903   } else {
 904     assert(is_phantom, "only remaining strength");
 905     blob_addr = bs->load_reference_barrier_phantom_rt_code_blob()->code_begin();
 906   }
 907 
 908   assert(blob_addr != nullptr, "code blob cannot be found");
 909 
 910   // Argument passing via the stack.  'obj' is passed implicitly (as asserted above).
 911   __ std(addr, -8, R1_SP);
 912 
 913   __ load_const_optimized(tmp1, blob_addr, tmp2);
 914   __ call_stub(tmp1);
 915 
 916   // 'res' is 'R3_RET'.  The result is thus already in the correct register.
 917 
 918   __ b(*stub->continuation());
 919   __ block_comment("} gen_load_reference_barrier_stub (shenandoahgc)");
 920 }
 921 
 922 #undef __
 923 
 924 #define __ sasm->
 925 
 926 void ShenandoahBarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler *sasm) {
 927   __ block_comment("generate_c1_pre_barrier_runtime_stub (shenandoahgc) {");
 928 
 929   Label runtime, skip_barrier;
 930   BarrierSet *bs = BarrierSet::barrier_set();
 931 
 932   // Argument passing via the stack.
 933   const int caller_stack_slots = 3;
 934 
 935   Register R0_pre_val = R0;
 936   __ ld(R0, -8, R1_SP);
 937   Register R11_tmp1 = R11_scratch1;
 938   __ std(R11_tmp1, -16, R1_SP);
 939   Register R12_tmp2 = R12_scratch2;
 940   __ std(R12_tmp2, -24, R1_SP);
 941 
 942   /* ==== Check whether marking is active ==== */
 943   // Even though gc status was checked in 'ShenandoahBarrierSetAssembler::gen_pre_barrier_stub',
 944   // another check is required as a safepoint might have been reached in the meantime (JDK-8140588).
 945   __ lbz(R12_tmp2, in_bytes(ShenandoahThreadLocalData::gc_state_offset()), R16_thread);
 946 
 947   __ andi_(R12_tmp2, R12_tmp2, ShenandoahHeap::MARKING);
 948   __ beq(CCR0, skip_barrier);
 949 
 950   /* ==== Add previous value directly to thread-local SATB mark queue ==== */
 951   // Check queue's capacity.  Jump to runtime if no free slot is available.
 952   __ ld(R12_tmp2, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()), R16_thread);
 953   __ cmpdi(CCR0, R12_tmp2, 0);
 954   __ beq(CCR0, runtime);
 955 
 956   // Capacity suffices.  Decrement the queue's size by one slot (size of one oop).
 957   __ addi(R12_tmp2, R12_tmp2, -wordSize);
 958   __ std(R12_tmp2, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()), R16_thread);
 959 
 960   // Enqueue the previous value and skip the runtime invocation.
 961   __ ld(R11_tmp1, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()), R16_thread);
 962   __ stdx(R0_pre_val, R11_tmp1, R12_tmp2);
 963   __ b(skip_barrier);
 964 
 965   __ bind(runtime);
 966 
 967   /* ==== Invoke runtime to commit SATB mark queue to gc and allocate a new buffer ==== */
 968   // Save to-be-preserved registers.
 969   const int nbytes_save = (MacroAssembler::num_volatile_regs + caller_stack_slots) * BytesPerWord;
 970   __ save_volatile_gprs(R1_SP, -nbytes_save);
 971   __ save_LR_CR(R11_tmp1);
 972   __ push_frame_reg_args(nbytes_save, R11_tmp1);
 973 
 974   // Invoke runtime.
 975   __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), R0_pre_val, R16_thread);
 976 
 977   // Restore to-be-preserved registers.
 978   __ pop_frame();
 979   __ restore_LR_CR(R11_tmp1);
 980   __ restore_volatile_gprs(R1_SP, -nbytes_save);
 981 
 982   __ bind(skip_barrier);
 983 
 984   // Restore spilled registers.
 985   __ ld(R11_tmp1, -16, R1_SP);
 986   __ ld(R12_tmp2, -24, R1_SP);
 987 
 988   __ blr();
 989   __ block_comment("} generate_c1_pre_barrier_runtime_stub (shenandoahgc)");
 990 }
 991 
 992 void ShenandoahBarrierSetAssembler::generate_c1_load_reference_barrier_runtime_stub(StubAssembler *sasm,
 993                                                                                     DecoratorSet decorators) {
 994   __ block_comment("generate_c1_load_reference_barrier_runtime_stub (shenandoahgc) {");
 995 
 996   // Argument passing via the stack.
 997   const int caller_stack_slots = 1;
 998 
 999   // Save to-be-preserved registers.
1000   const int nbytes_save = (MacroAssembler::num_volatile_regs - 1 // 'R3_ARG1' is skipped
1001                            + caller_stack_slots) * BytesPerWord;
1002   __ save_volatile_gprs(R1_SP, -nbytes_save, true, false);
1003 
1004   // Load arguments from stack.
1005   // No load required, as assured by assertions in 'ShenandoahBarrierSetAssembler::gen_load_reference_barrier_stub'.
1006   Register R3_obj = R3_ARG1;
1007   Register R4_load_addr = R4_ARG2;
1008   __ ld(R4_load_addr, -8, R1_SP);
1009 
1010   Register R11_tmp = R11_scratch1;
1011 
1012   /* ==== Invoke runtime ==== */
1013   bool is_strong  = ShenandoahBarrierSet::is_strong_access(decorators);
1014   bool is_weak    = ShenandoahBarrierSet::is_weak_access(decorators);
1015   bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
1016   bool is_native  = ShenandoahBarrierSet::is_native_access(decorators);
1017 
1018   address jrt_address = nullptr;
1019 
1020   if (is_strong) {
1021     if (is_native) {
1022       jrt_address = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong);
1023     } else {
1024       if (UseCompressedOops) {
1025         jrt_address = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow);
1026       } else {
1027         jrt_address = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong);
1028       }
1029     }
1030   } else if (is_weak) {
1031     assert(!is_native, "weak load reference barrier must not be called off-heap");
1032     if (UseCompressedOops) {
1033       jrt_address = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow);
1034     } else {
1035       jrt_address = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak);
1036     }
1037   } else {
1038     assert(is_phantom, "reference type must be phantom");
1039     assert(is_native, "phantom load reference barrier must be called off-heap");
1040     jrt_address = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom);
1041   }
1042   assert(jrt_address != nullptr, "load reference barrier runtime routine cannot be found");
1043 
1044   __ save_LR_CR(R11_tmp);
1045   __ push_frame_reg_args(nbytes_save, R11_tmp);
1046 
1047   // Invoke runtime.  Arguments are already stored in the corresponding registers.
1048   __ call_VM_leaf(jrt_address, R3_obj, R4_load_addr);
1049 
1050   // Restore to-be-preserved registers.
1051   __ pop_frame();
1052   __ restore_LR_CR(R11_tmp);
1053   __ restore_volatile_gprs(R1_SP, -nbytes_save, true, false); // Skip 'R3_RET' register.
1054 
1055   __ blr();
1056   __ block_comment("} generate_c1_load_reference_barrier_runtime_stub (shenandoahgc)");
1057 }
1058 
1059 #undef __
1060 
1061 #endif // COMPILER1