1 /*
   2  * Copyright (c) 2018, 2019, Red Hat, Inc. All rights reserved.
   3  *
   4  * This code is free software; you can redistribute it and/or modify it
   5  * under the terms of the GNU General Public License version 2 only, as
   6  * published by the Free Software Foundation.
   7  *
   8  * This code is distributed in the hope that it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  11  * version 2 for more details (a copy is included in the LICENSE file that
  12  * accompanied this code).
  13  *
  14  * You should have received a copy of the GNU General Public License version
  15  * 2 along with this work; if not, write to the Free Software Foundation,
  16  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  17  *
  18  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  19  * or visit www.oracle.com if you need additional information or have any
  20  * questions.
  21  *
  22  */
  23 
  24 #include "precompiled.hpp"
  25 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
  26 #include "gc/shenandoah/shenandoahForwarding.hpp"
  27 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
  28 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
  29 #include "gc/shenandoah/shenandoahHeuristics.hpp"
  30 #include "gc/shenandoah/shenandoahRuntime.hpp"
  31 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
  32 #include "interpreter/interpreter.hpp"
  33 #include "interpreter/interp_masm.hpp"
  34 #include "runtime/sharedRuntime.hpp"
  35 #include "runtime/thread.hpp"
  36 #include "utilities/macros.hpp"
  37 #ifdef COMPILER1
  38 #include "c1/c1_LIRAssembler.hpp"
  39 #include "c1/c1_MacroAssembler.hpp"
  40 #include "gc/shenandoah/c1/shenandoahBarrierSetC1.hpp"
  41 #endif
  42 
  43 #define __ masm->
  44 
  45 address ShenandoahBarrierSetAssembler::_shenandoah_lrb = NULL;
  46 
  47 void ShenandoahBarrierSetAssembler::arraycopy_prologue(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
  48                                                        Register src, Register dst, Register count) {
  49 
  50   bool checkcast = (decorators & ARRAYCOPY_CHECKCAST) != 0;
  51   bool disjoint = (decorators & ARRAYCOPY_DISJOINT) != 0;
  52   bool obj_int = type == T_OBJECT LP64_ONLY(&& UseCompressedOops);
  53   bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0;
  54 
  55   if (type == T_OBJECT || type == T_ARRAY) {
  56 #ifdef _LP64
  57     if (!checkcast) {
  58       if (!obj_int) {
  59         // Save count for barrier
  60         __ movptr(r11, count);
  61       } else if (disjoint) {
  62         // Save dst in r11 in the disjoint case
  63         __ movq(r11, dst);
  64       }
  65     }
  66 #else
  67     if (disjoint) {
  68       __ mov(rdx, dst);          // save 'to'
  69     }
  70 #endif
  71 
  72     if (ShenandoahSATBBarrier && !dest_uninitialized) {
  73       Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
  74       assert_different_registers(dst, count, thread); // we don't care about src here?
  75 #ifndef _LP64
  76       __ push(thread);
  77       __ get_thread(thread);
  78 #endif
  79 
  80       Label done;
  81       // Short-circuit if count == 0.
  82       __ testptr(count, count);
  83       __ jcc(Assembler::zero, done);
  84 
  85       // Avoid runtime call when not marking.
  86       Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
  87       __ testb(gc_state, ShenandoahHeap::MARKING);
  88       __ jcc(Assembler::zero, done);
  89 
  90       __ pusha();                      // push registers
  91 #ifdef _LP64
  92       if (count == c_rarg0) {
  93         if (dst == c_rarg1) {
  94           // exactly backwards!!
  95           __ xchgptr(c_rarg1, c_rarg0);
  96         } else {
  97           __ movptr(c_rarg1, count);
  98           __ movptr(c_rarg0, dst);
  99         }
 100       } else {
 101         __ movptr(c_rarg0, dst);
 102         __ movptr(c_rarg1, count);
 103       }
 104       if (UseCompressedOops) {
 105         __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_narrow_oop_entry), 2);
 106       } else {
 107         __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry), 2);
 108       }
 109 #else
 110       __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry),
 111                       dst, count);
 112 #endif
 113       __ popa();
 114       __ bind(done);
 115       NOT_LP64(__ pop(thread);)
 116     }
 117   }
 118 
 119 }
 120 
 121 void ShenandoahBarrierSetAssembler::arraycopy_epilogue(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 122                                                        Register src, Register dst, Register count) {
 123   bool checkcast = (decorators & ARRAYCOPY_CHECKCAST) != 0;
 124   bool disjoint = (decorators & ARRAYCOPY_DISJOINT) != 0;
 125   bool obj_int = type == T_OBJECT LP64_ONLY(&& UseCompressedOops);
 126   Register tmp = rax;
 127 
 128   if (type == T_OBJECT || type == T_ARRAY) {
 129 #ifdef _LP64
 130     if (!checkcast) {
 131       if (!obj_int) {
 132         // Save count for barrier
 133         count = r11;
 134       } else if (disjoint && obj_int) {
 135         // Use the saved dst in the disjoint case
 136         dst = r11;
 137       }
 138     } else {
 139       tmp = rscratch1;
 140     }
 141 #else
 142     if (disjoint) {
 143       __ mov(dst, rdx); // restore 'to'
 144     }
 145 #endif
 146 
 147     Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
 148     assert_different_registers(dst, thread); // do we care about src at all here?
 149 
 150 #ifndef _LP64
 151     __ push(thread);
 152     __ get_thread(thread);
 153 #endif
 154 
 155     // Short-circuit if count == 0.
 156     Label done;
 157     __ testptr(count, count);
 158     __ jcc(Assembler::zero, done);
 159 
 160     // Skip runtime call if no forwarded objects.
 161     Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 162     __ testb(gc_state, ShenandoahHeap::UPDATEREFS);
 163     __ jcc(Assembler::zero, done);
 164 
 165     __ pusha();             // push registers (overkill)
 166 #ifdef _LP64
 167     if (c_rarg0 == count) { // On win64 c_rarg0 == rcx
 168       assert_different_registers(c_rarg1, dst);
 169       __ mov(c_rarg1, count);
 170       __ mov(c_rarg0, dst);
 171     } else {
 172       assert_different_registers(c_rarg0, count);
 173       __ mov(c_rarg0, dst);
 174       __ mov(c_rarg1, count);
 175     }
 176     __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry), 2);
 177 #else
 178     __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry),
 179                     dst, count);
 180 #endif
 181     __ popa();
 182 
 183     __ bind(done);
 184     NOT_LP64(__ pop(thread);)
 185   }
 186 }
 187 
 188 void ShenandoahBarrierSetAssembler::shenandoah_write_barrier_pre(MacroAssembler* masm,
 189                                                                  Register obj,
 190                                                                  Register pre_val,
 191                                                                  Register thread,
 192                                                                  Register tmp,
 193                                                                  bool tosca_live,
 194                                                                  bool expand_call) {
 195 
 196   if (ShenandoahSATBBarrier) {
 197     satb_write_barrier_pre(masm, obj, pre_val, thread, tmp, tosca_live, expand_call);
 198   }
 199 }
 200 
 201 void ShenandoahBarrierSetAssembler::satb_write_barrier_pre(MacroAssembler* masm,
 202                                                            Register obj,
 203                                                            Register pre_val,
 204                                                            Register thread,
 205                                                            Register tmp,
 206                                                            bool tosca_live,
 207                                                            bool expand_call) {
 208   // If expand_call is true then we expand the call_VM_leaf macro
 209   // directly to skip generating the check by
 210   // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
 211 
 212 #ifdef _LP64
 213   assert(thread == r15_thread, "must be");
 214 #endif // _LP64
 215 
 216   Label done;
 217   Label runtime;
 218 
 219   assert(pre_val != noreg, "check this code");
 220 
 221   if (obj != noreg) {
 222     assert_different_registers(obj, pre_val, tmp);
 223     assert(pre_val != rax, "check this code");
 224   }
 225 
 226   Address in_progress(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()));
 227   Address index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
 228   Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
 229 
 230   Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 231   __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL);
 232   __ jcc(Assembler::zero, done);
 233 
 234   // Do we need to load the previous value?
 235   if (obj != noreg) {
 236     __ load_heap_oop(pre_val, Address(obj, 0), noreg, noreg, AS_RAW);
 237   }
 238 
 239   // Is the previous value null?
 240   __ cmpptr(pre_val, (int32_t) NULL_WORD);
 241   __ jcc(Assembler::equal, done);
 242 
 243   // Can we store original value in the thread's buffer?
 244   // Is index == 0?
 245   // (The index field is typed as size_t.)
 246 
 247   __ movptr(tmp, index);                   // tmp := *index_adr
 248   __ cmpptr(tmp, 0);                       // tmp == 0?
 249   __ jcc(Assembler::equal, runtime);       // If yes, goto runtime
 250 
 251   __ subptr(tmp, wordSize);                // tmp := tmp - wordSize
 252   __ movptr(index, tmp);                   // *index_adr := tmp
 253   __ addptr(tmp, buffer);                  // tmp := tmp + *buffer_adr
 254 
 255   // Record the previous value
 256   __ movptr(Address(tmp, 0), pre_val);
 257   __ jmp(done);
 258 
 259   __ bind(runtime);
 260   // save the live input values
 261   if(tosca_live) __ push(rax);
 262 
 263   if (obj != noreg && obj != rax)
 264     __ push(obj);
 265 
 266   if (pre_val != rax)
 267     __ push(pre_val);
 268 
 269   // Calling the runtime using the regular call_VM_leaf mechanism generates
 270   // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
 271   // that checks that the *(ebp+frame::interpreter_frame_last_sp) == NULL.
 272   //
 273   // If we care generating the pre-barrier without a frame (e.g. in the
 274   // intrinsified Reference.get() routine) then ebp might be pointing to
 275   // the caller frame and so this check will most likely fail at runtime.
 276   //
 277   // Expanding the call directly bypasses the generation of the check.
 278   // So when we do not have have a full interpreter frame on the stack
 279   // expand_call should be passed true.
 280 
 281   NOT_LP64( __ push(thread); )
 282 
 283 #ifdef _LP64
 284   // We move pre_val into c_rarg0 early, in order to avoid smashing it, should
 285   // pre_val be c_rarg1 (where the call prologue would copy thread argument).
 286   // Note: this should not accidentally smash thread, because thread is always r15.
 287   assert(thread != c_rarg0, "smashed arg");
 288   if (c_rarg0 != pre_val) {
 289     __ mov(c_rarg0, pre_val);
 290   }
 291 #endif
 292 
 293   if (expand_call) {
 294     LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); )
 295 #ifdef _LP64
 296     if (c_rarg1 != thread) {
 297       __ mov(c_rarg1, thread);
 298     }
 299     // Already moved pre_val into c_rarg0 above
 300 #else
 301     __ push(thread);
 302     __ push(pre_val);
 303 #endif
 304     __ MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), 2);
 305   } else {
 306     __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), LP64_ONLY(c_rarg0) NOT_LP64(pre_val), thread);
 307   }
 308 
 309   NOT_LP64( __ pop(thread); )
 310 
 311   // save the live input values
 312   if (pre_val != rax)
 313     __ pop(pre_val);
 314 
 315   if (obj != noreg && obj != rax)
 316     __ pop(obj);
 317 
 318   if(tosca_live) __ pop(rax);
 319 
 320   __ bind(done);
 321 }
 322 
 323 void ShenandoahBarrierSetAssembler::resolve_forward_pointer(MacroAssembler* masm, Register dst, Register tmp) {
 324   assert(ShenandoahCASBarrier, "should be enabled");
 325   Label is_null;
 326   __ testptr(dst, dst);
 327   __ jcc(Assembler::zero, is_null);
 328   resolve_forward_pointer_not_null(masm, dst, tmp);
 329   __ bind(is_null);
 330 }
 331 
 332 void ShenandoahBarrierSetAssembler::resolve_forward_pointer_not_null(MacroAssembler* masm, Register dst, Register tmp) {
 333   assert(ShenandoahCASBarrier || ShenandoahLoadRefBarrier, "should be enabled");
 334   // The below loads the mark word, checks if the lowest two bits are
 335   // set, and if so, clear the lowest two bits and copy the result
 336   // to dst. Otherwise it leaves dst alone.
 337   // Implementing this is surprisingly awkward. I do it here by:
 338   // - Inverting the mark word
 339   // - Test lowest two bits == 0
 340   // - If so, set the lowest two bits
 341   // - Invert the result back, and copy to dst
 342 
 343   bool borrow_reg = (tmp == noreg);
 344   if (borrow_reg) {
 345     // No free registers available. Make one useful.
 346     tmp = LP64_ONLY(rscratch1) NOT_LP64(rdx);
 347     __ push(tmp);
 348   }
 349 
 350   Label done;
 351   __ movptr(tmp, Address(dst, oopDesc::mark_offset_in_bytes()));
 352   __ notptr(tmp);
 353   __ testb(tmp, markOopDesc::marked_value);
 354   __ jccb(Assembler::notZero, done);
 355   __ orptr(tmp, markOopDesc::marked_value);
 356   __ notptr(tmp);
 357   __ mov(dst, tmp);
 358   __ bind(done);
 359 
 360   if (borrow_reg) {
 361     __ pop(tmp);
 362   }
 363 }
 364 
 365 
 366 void ShenandoahBarrierSetAssembler::load_reference_barrier_not_null(MacroAssembler* masm, Register dst) {
 367   assert(ShenandoahLoadRefBarrier, "Should be enabled");
 368 
 369   Label done;
 370 
 371 #ifdef _LP64
 372   Register thread = r15_thread;
 373 #else
 374   Register thread = rcx;
 375   if (thread == dst) {
 376     thread = rbx;
 377   }
 378   __ push(thread);
 379   __ get_thread(thread);
 380 #endif
 381   assert_different_registers(dst, thread);
 382 
 383   Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 384   __ testb(gc_state, ShenandoahHeap::HAS_FORWARDED);
 385   __ jccb(Assembler::zero, done);
 386 
 387    if (dst != rax) {
 388      __ xchgptr(dst, rax); // Move obj into rax and save rax into obj.
 389    }
 390 
 391    __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, ShenandoahBarrierSetAssembler::shenandoah_lrb())));
 392 
 393    if (dst != rax) {
 394      __ xchgptr(rax, dst); // Swap back obj with rax.
 395    }
 396 
 397   __ bind(done);
 398 
 399 #ifndef _LP64
 400   __ pop(thread);
 401 #endif
 402 }
 403 
 404 void ShenandoahBarrierSetAssembler::load_reference_barrier_native(MacroAssembler* masm, Register dst) {
 405   if (!ShenandoahLoadRefBarrier) {
 406     return;
 407   }
 408 
 409   Label done;
 410   Label not_null;
 411   Label slow_path;
 412 
 413   // null check
 414   __ testptr(dst, dst);
 415   __ jcc(Assembler::notZero, not_null);
 416   __ jmp(done);
 417   __ bind(not_null);
 418 
 419 
 420 #ifdef _LP64
 421   Register thread = r15_thread;
 422 #else
 423   Register thread = rcx;
 424   if (thread == dst) {
 425     thread = rbx;
 426   }
 427   __ push(thread);
 428   __ get_thread(thread);
 429 #endif
 430   assert_different_registers(dst, thread);
 431 
 432   Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 433   __ testb(gc_state, ShenandoahHeap::EVACUATION);
 434 #ifndef _LP64
 435   __ pop(thread);
 436 #endif
 437   __ jccb(Assembler::notZero, slow_path);
 438   __ jmp(done);
 439   __ bind(slow_path);
 440 
 441   if (dst != rax) {
 442     __ xchgptr(dst, rax); // Move obj into rax and save rax into obj.
 443   }
 444   __ push(rcx);
 445   __ push(rdx);
 446   __ push(rdi);
 447   __ push(rsi);
 448 #ifdef _LP64
 449   __ push(r8);
 450   __ push(r9);
 451   __ push(r10);
 452   __ push(r11);
 453   __ push(r12);
 454   __ push(r13);
 455   __ push(r14);
 456   __ push(r15);
 457 #endif
 458 
 459   __ movptr(rdi, rax);
 460   __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_native), rdi);
 461 
 462 #ifdef _LP64
 463   __ pop(r15);
 464   __ pop(r14);
 465   __ pop(r13);
 466   __ pop(r12);
 467   __ pop(r11);
 468   __ pop(r10);
 469   __ pop(r9);
 470   __ pop(r8);
 471 #endif
 472   __ pop(rsi);
 473   __ pop(rdi);
 474   __ pop(rdx);
 475   __ pop(rcx);
 476 
 477   if (dst != rax) {
 478     __ xchgptr(rax, dst); // Swap back obj with rax.
 479   }
 480 
 481   __ bind(done);
 482 }
 483 
 484 void ShenandoahBarrierSetAssembler::storeval_barrier(MacroAssembler* masm, Register dst, Register tmp) {
 485   if (ShenandoahStoreValEnqueueBarrier) {
 486     storeval_barrier_impl(masm, dst, tmp);
 487   }
 488 }
 489 
 490 void ShenandoahBarrierSetAssembler::storeval_barrier_impl(MacroAssembler* masm, Register dst, Register tmp) {
 491   assert(ShenandoahStoreValEnqueueBarrier, "should be enabled");
 492 
 493   if (dst == noreg) return;
 494 
 495   if (ShenandoahStoreValEnqueueBarrier) {
 496     // The set of registers to be saved+restored is the same as in the write-barrier above.
 497     // Those are the commonly used registers in the interpreter.
 498     __ pusha();
 499     // __ push_callee_saved_registers();
 500     __ subptr(rsp, 2 * Interpreter::stackElementSize);
 501     __ movdbl(Address(rsp, 0), xmm0);
 502 
 503 #ifdef _LP64
 504     Register thread = r15_thread;
 505 #else
 506     Register thread = rcx;
 507     if (thread == dst || thread == tmp) {
 508       thread = rdi;
 509     }
 510     if (thread == dst || thread == tmp) {
 511       thread = rbx;
 512     }
 513     __ get_thread(thread);
 514 #endif
 515     assert_different_registers(dst, tmp, thread);
 516 
 517     satb_write_barrier_pre(masm, noreg, dst, thread, tmp, true, false);
 518     __ movdbl(xmm0, Address(rsp, 0));
 519     __ addptr(rsp, 2 * Interpreter::stackElementSize);
 520     //__ pop_callee_saved_registers();
 521     __ popa();
 522   }
 523 }
 524 
 525 void ShenandoahBarrierSetAssembler::load_reference_barrier(MacroAssembler* masm, Register dst) {
 526   if (ShenandoahLoadRefBarrier) {
 527     Label done;
 528     __ testptr(dst, dst);
 529     __ jcc(Assembler::zero, done);
 530     load_reference_barrier_not_null(masm, dst);
 531     __ bind(done);
 532   }
 533 }
 534 
 535 void ShenandoahBarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 536              Register dst, Address src, Register tmp1, Register tmp_thread) {
 537   bool on_oop = type == T_OBJECT || type == T_ARRAY;
 538   bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
 539   bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
 540   bool not_in_heap = (decorators & IN_NATIVE) != 0;
 541   bool on_reference = on_weak || on_phantom;
 542   bool keep_alive = (decorators & AS_NO_KEEPALIVE) == 0;
 543 
 544   BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp_thread);
 545   if (on_oop) {
 546     if (not_in_heap) {
 547       if (ShenandoahHeap::heap()->is_traversal_mode()) {
 548         load_reference_barrier(masm, dst);
 549         keep_alive = true;
 550       } else {
 551         load_reference_barrier_native(masm, dst);
 552       }
 553     } else {
 554       load_reference_barrier(masm, dst);
 555     }
 556 
 557     if (ShenandoahKeepAliveBarrier && on_reference && keep_alive) {
 558       const Register thread = NOT_LP64(tmp_thread) LP64_ONLY(r15_thread);
 559       assert_different_registers(dst, tmp1, tmp_thread);
 560       NOT_LP64(__ get_thread(thread));
 561       // Generate the SATB pre-barrier code to log the value of
 562       // the referent field in an SATB buffer.
 563       shenandoah_write_barrier_pre(masm /* masm */,
 564                                    noreg /* obj */,
 565                                    dst /* pre_val */,
 566                                    thread /* thread */,
 567                                    tmp1 /* tmp */,
 568                                    true /* tosca_live */,
 569                                    true /* expand_call */);
 570     }
 571   }
 572 }
 573 
 574 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 575               Address dst, Register val, Register tmp1, Register tmp2) {
 576 
 577   bool on_oop = type == T_OBJECT || type == T_ARRAY;
 578   bool in_heap = (decorators & IN_HEAP) != 0;
 579   bool as_normal = (decorators & AS_NORMAL) != 0;
 580   if (on_oop && in_heap) {
 581     bool needs_pre_barrier = as_normal;
 582 
 583     Register tmp3 = LP64_ONLY(r8) NOT_LP64(rsi);
 584     Register rthread = LP64_ONLY(r15_thread) NOT_LP64(rcx);
 585     // flatten object address if needed
 586     // We do it regardless of precise because we need the registers
 587     if (dst.index() == noreg && dst.disp() == 0) {
 588       if (dst.base() != tmp1) {
 589         __ movptr(tmp1, dst.base());
 590       }
 591     } else {
 592       __ lea(tmp1, dst);
 593     }
 594 
 595     assert_different_registers(val, tmp1, tmp2, tmp3, rthread);
 596 
 597 #ifndef _LP64
 598     __ get_thread(rthread);
 599     InterpreterMacroAssembler *imasm = static_cast<InterpreterMacroAssembler*>(masm);
 600     imasm->save_bcp();
 601 #endif
 602 
 603     if (needs_pre_barrier) {
 604       shenandoah_write_barrier_pre(masm /*masm*/,
 605                                    tmp1 /* obj */,
 606                                    tmp2 /* pre_val */,
 607                                    rthread /* thread */,
 608                                    tmp3  /* tmp */,
 609                                    val != noreg /* tosca_live */,
 610                                    false /* expand_call */);
 611     }
 612     if (val == noreg) {
 613       BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
 614     } else {
 615       storeval_barrier(masm, val, tmp3);
 616       BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
 617     }
 618     NOT_LP64(imasm->restore_bcp());
 619   } else {
 620     BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2);
 621   }
 622 }
 623 
 624 void ShenandoahBarrierSetAssembler::try_resolve_jobject_in_native(MacroAssembler* masm, Register jni_env,
 625                                                                   Register obj, Register tmp, Label& slowpath) {
 626   Label done;
 627   // Resolve jobject
 628   BarrierSetAssembler::try_resolve_jobject_in_native(masm, jni_env, obj, tmp, slowpath);
 629 
 630   // Check for null.
 631   __ testptr(obj, obj);
 632   __ jcc(Assembler::zero, done);
 633 
 634   Address gc_state(jni_env, ShenandoahThreadLocalData::gc_state_offset() - JavaThread::jni_environment_offset());
 635   __ testb(gc_state, ShenandoahHeap::EVACUATION);
 636   __ jccb(Assembler::notZero, slowpath);
 637   __ bind(done);
 638 }
 639 
 640 // Special Shenandoah CAS implementation that handles false negatives
 641 // due to concurrent evacuation.
 642 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm,
 643                                                 Register res, Address addr, Register oldval, Register newval,
 644                                                 bool exchange, Register tmp1, Register tmp2) {
 645   assert(ShenandoahCASBarrier, "Should only be used when CAS barrier is enabled");
 646   assert(oldval == rax, "must be in rax for implicit use in cmpxchg");
 647 
 648   Label retry, done;
 649 
 650   // Remember oldval for retry logic below
 651 #ifdef _LP64
 652   if (UseCompressedOops) {
 653     __ movl(tmp1, oldval);
 654   } else
 655 #endif
 656   {
 657     __ movptr(tmp1, oldval);
 658   }
 659 
 660   // Step 1. Try to CAS with given arguments. If successful, then we are done,
 661   // and can safely return.
 662   if (os::is_MP()) __ lock();
 663 #ifdef _LP64
 664   if (UseCompressedOops) {
 665     __ cmpxchgl(newval, addr);
 666   } else
 667 #endif
 668   {
 669     __ cmpxchgptr(newval, addr);
 670   }
 671   __ jcc(Assembler::equal, done, true);
 672 
 673   // Step 2. CAS had failed. This may be a false negative.
 674   //
 675   // The trouble comes when we compare the to-space pointer with the from-space
 676   // pointer to the same object. To resolve this, it will suffice to resolve both
 677   // oldval and the value from memory -- this will give both to-space pointers.
 678   // If they mismatch, then it was a legitimate failure.
 679   //
 680 #ifdef _LP64
 681   if (UseCompressedOops) {
 682     __ decode_heap_oop(tmp1);
 683   }
 684 #endif
 685   resolve_forward_pointer(masm, tmp1);
 686 
 687 #ifdef _LP64
 688   if (UseCompressedOops) {
 689     __ movl(tmp2, oldval);
 690     __ decode_heap_oop(tmp2);
 691   } else
 692 #endif
 693   {
 694     __ movptr(tmp2, oldval);
 695   }
 696   resolve_forward_pointer(masm, tmp2);
 697 
 698   __ cmpptr(tmp1, tmp2);
 699   __ jcc(Assembler::notEqual, done, true);
 700 
 701   // Step 3. Try to CAS again with resolved to-space pointers.
 702   //
 703   // Corner case: it may happen that somebody stored the from-space pointer
 704   // to memory while we were preparing for retry. Therefore, we can fail again
 705   // on retry, and so need to do this in loop, always resolving the failure
 706   // witness.
 707   __ bind(retry);
 708   if (os::is_MP()) __ lock();
 709 #ifdef _LP64
 710   if (UseCompressedOops) {
 711     __ cmpxchgl(newval, addr);
 712   } else
 713 #endif
 714   {
 715     __ cmpxchgptr(newval, addr);
 716   }
 717   __ jcc(Assembler::equal, done, true);
 718 
 719 #ifdef _LP64
 720   if (UseCompressedOops) {
 721     __ movl(tmp2, oldval);
 722     __ decode_heap_oop(tmp2);
 723   } else
 724 #endif
 725   {
 726     __ movptr(tmp2, oldval);
 727   }
 728   resolve_forward_pointer(masm, tmp2);
 729 
 730   __ cmpptr(tmp1, tmp2);
 731   __ jcc(Assembler::equal, retry, true);
 732 
 733   // Step 4. If we need a boolean result out of CAS, check the flag again,
 734   // and promote the result. Note that we handle the flag from both the CAS
 735   // itself and from the retry loop.
 736   __ bind(done);
 737   if (!exchange) {
 738     assert(res != NULL, "need result register");
 739 #ifdef _LP64
 740     __ setb(Assembler::equal, res);
 741     __ movzbl(res, res);
 742 #else
 743     // Need something else to clean the result, because some registers
 744     // do not have byte encoding that movzbl wants. Cannot do the xor first,
 745     // because it modifies the flags.
 746     Label res_non_zero;
 747     __ movptr(res, 1);
 748     __ jcc(Assembler::equal, res_non_zero, true);
 749     __ xorptr(res, res);
 750     __ bind(res_non_zero);
 751 #endif
 752   }
 753 }
 754 
 755 #undef __
 756 
 757 #ifdef COMPILER1
 758 
 759 #define __ ce->masm()->
 760 
 761 void ShenandoahBarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, ShenandoahPreBarrierStub* stub) {
 762   ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
 763   // At this point we know that marking is in progress.
 764   // If do_load() is true then we have to emit the
 765   // load of the previous value; otherwise it has already
 766   // been loaded into _pre_val.
 767 
 768   __ bind(*stub->entry());
 769   assert(stub->pre_val()->is_register(), "Precondition.");
 770 
 771   Register pre_val_reg = stub->pre_val()->as_register();
 772 
 773   if (stub->do_load()) {
 774     ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/, false /*unaligned*/);
 775   }
 776 
 777   __ cmpptr(pre_val_reg, (int32_t)NULL_WORD);
 778   __ jcc(Assembler::equal, *stub->continuation());
 779   ce->store_parameter(stub->pre_val()->as_register(), 0);
 780   __ call(RuntimeAddress(bs->pre_barrier_c1_runtime_code_blob()->code_begin()));
 781   __ jmp(*stub->continuation());
 782 
 783 }
 784 
 785 void ShenandoahBarrierSetAssembler::gen_load_reference_barrier_stub(LIR_Assembler* ce, ShenandoahLoadReferenceBarrierStub* stub) {
 786   ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
 787   __ bind(*stub->entry());
 788 
 789   Register obj = stub->obj()->as_register();
 790   Register res = stub->result()->as_register();
 791   Register tmp1 = stub->tmp1()->as_register();
 792   Register tmp2 = stub->tmp2()->as_register();
 793 
 794   Label slow_path;
 795 
 796   assert(res == rax, "result must arrive in rax");
 797 
 798   if (res != obj) {
 799     __ mov(res, obj);
 800   }
 801 
 802   // Check for null.
 803   __ testptr(res, res);
 804   __ jcc(Assembler::zero, *stub->continuation());
 805 
 806   // Check for object being in the collection set.
 807   __ mov(tmp1, res);
 808   __ shrptr(tmp1, ShenandoahHeapRegion::region_size_bytes_shift_jint());
 809   __ movptr(tmp2, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr());
 810 #ifdef _LP64
 811   __ movbool(tmp2, Address(tmp2, tmp1, Address::times_1));
 812   __ testbool(tmp2);
 813 #else
 814   // On x86_32, C1 register allocator can give us the register without 8-bit support.
 815   // Do the full-register access and test to avoid compilation failures.
 816   __ movptr(tmp2, Address(tmp2, tmp1, Address::times_1));
 817   __ testptr(tmp2, 0xFF);
 818 #endif
 819   __ jcc(Assembler::zero, *stub->continuation());
 820 
 821   // Test if object is resolved.
 822   __ movptr(tmp1, Address(res, oopDesc::mark_offset_in_bytes()));
 823   // Test if both lowest bits are set. We trick it by negating the bits
 824   // then test for both bits clear.
 825   __ notptr(tmp1);
 826 #ifdef _LP64
 827   __ testb(tmp1, markOopDesc::marked_value);
 828 #else
 829   // On x86_32, C1 register allocator can give us the register without 8-bit support.
 830   // Do the full-register access and test to avoid compilation failures.
 831   __ testptr(tmp1, markOopDesc::marked_value);
 832 #endif
 833   __ jccb(Assembler::notZero, slow_path);
 834   // Clear both lower bits. It's still inverted, so set them, and then invert back.
 835   __ orptr(tmp1, markOopDesc::marked_value);
 836   __ notptr(tmp1);
 837   // At this point, tmp1 contains the decoded forwarding pointer.
 838   __ mov(res, tmp1);
 839 
 840   __ jmp(*stub->continuation());
 841 
 842   __ bind(slow_path);
 843   ce->store_parameter(res, 0);
 844   __ call(RuntimeAddress(bs->load_reference_barrier_rt_code_blob()->code_begin()));
 845 
 846   __ jmp(*stub->continuation());
 847 }
 848 
 849 #undef __
 850 
 851 #define __ sasm->
 852 
 853 void ShenandoahBarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) {
 854   __ prologue("shenandoah_pre_barrier", false);
 855   // arg0 : previous value of memory
 856 
 857   __ push(rax);
 858   __ push(rdx);
 859 
 860   const Register pre_val = rax;
 861   const Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
 862   const Register tmp = rdx;
 863 
 864   NOT_LP64(__ get_thread(thread);)
 865 
 866   Address queue_index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
 867   Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
 868 
 869   Label done;
 870   Label runtime;
 871 
 872   // Is SATB still active?
 873   Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 874   __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL);
 875   __ jcc(Assembler::zero, done);
 876 
 877   // Can we store original value in the thread's buffer?
 878 
 879   __ movptr(tmp, queue_index);
 880   __ testptr(tmp, tmp);
 881   __ jcc(Assembler::zero, runtime);
 882   __ subptr(tmp, wordSize);
 883   __ movptr(queue_index, tmp);
 884   __ addptr(tmp, buffer);
 885 
 886   // prev_val (rax)
 887   __ load_parameter(0, pre_val);
 888   __ movptr(Address(tmp, 0), pre_val);
 889   __ jmp(done);
 890 
 891   __ bind(runtime);
 892 
 893   __ save_live_registers_no_oop_map(true);
 894 
 895   // load the pre-value
 896   __ load_parameter(0, rcx);
 897   __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), rcx, thread);
 898 
 899   __ restore_live_registers(true);
 900 
 901   __ bind(done);
 902 
 903   __ pop(rdx);
 904   __ pop(rax);
 905 
 906   __ epilogue();
 907 }
 908 
 909 void ShenandoahBarrierSetAssembler::generate_c1_load_reference_barrier_runtime_stub(StubAssembler* sasm) {
 910   __ prologue("shenandoah_load_reference_barrier", false);
 911   // arg0 : object to be resolved
 912 
 913   __ save_live_registers_no_oop_map(true);
 914   __ load_parameter(0, LP64_ONLY(c_rarg0) NOT_LP64(rax));
 915   __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier), LP64_ONLY(c_rarg0) NOT_LP64(rax));
 916   __ restore_live_registers_except_rax(true);
 917 
 918   __ epilogue();
 919 }
 920 
 921 #undef __
 922 
 923 #endif // COMPILER1
 924 
 925 address ShenandoahBarrierSetAssembler::shenandoah_lrb() {
 926   assert(_shenandoah_lrb != NULL, "need load reference barrier stub");
 927   return _shenandoah_lrb;
 928 }
 929 
 930 #define __ cgen->assembler()->
 931 
 932 address ShenandoahBarrierSetAssembler::generate_shenandoah_lrb(StubCodeGenerator* cgen) {
 933   __ align(CodeEntryAlignment);
 934   StubCodeMark mark(cgen, "StubRoutines", "shenandoah_lrb");
 935   address start = __ pc();
 936 
 937   Label resolve_oop, slow_path, done;
 938 
 939   // We use RDI, which also serves as argument register for slow call.
 940   // RAX always holds the src object ptr, except after the slow call,
 941   // then it holds the result. R8/RBX is used as temporary register.
 942 
 943   Register tmp1 = rdi;
 944   Register tmp2 = LP64_ONLY(r8) NOT_LP64(rbx);
 945 
 946   __ push(tmp1);
 947   __ push(tmp2);
 948 
 949   // Check for object being in the collection set.
 950   __ mov(tmp1, rax);
 951   __ shrptr(tmp1, ShenandoahHeapRegion::region_size_bytes_shift_jint());
 952   __ movptr(tmp2, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr());
 953   __ movbool(tmp2, Address(tmp2, tmp1, Address::times_1));
 954   __ testbool(tmp2);
 955   __ jccb(Assembler::notZero, resolve_oop);
 956   __ pop(tmp2);
 957   __ pop(tmp1);
 958   __ ret(0);
 959 
 960   // Test if object is already resolved.
 961   __ bind(resolve_oop);
 962   __ movptr(tmp2, Address(rax, oopDesc::mark_offset_in_bytes()));
 963   // Test if both lowest bits are set. We trick it by negating the bits
 964   // then test for both bits clear.
 965   __ notptr(tmp2);
 966   __ testb(tmp2, markOopDesc::marked_value);
 967   __ jccb(Assembler::notZero, slow_path);
 968   // Clear both lower bits. It's still inverted, so set them, and then invert back.
 969   __ orptr(tmp2, markOopDesc::marked_value);
 970   __ notptr(tmp2);
 971   // At this point, tmp2 contains the decoded forwarding pointer.
 972   __ mov(rax, tmp2);
 973 
 974   __ bind(done);
 975   __ pop(tmp2);
 976   __ pop(tmp1);
 977   __ ret(0);
 978 
 979   __ bind(slow_path);
 980 
 981   __ push(rcx);
 982   __ push(rdx);
 983   __ push(rdi);
 984   __ push(rsi);
 985 #ifdef _LP64
 986   __ push(r8);
 987   __ push(r9);
 988   __ push(r10);
 989   __ push(r11);
 990   __ push(r12);
 991   __ push(r13);
 992   __ push(r14);
 993   __ push(r15);
 994 #endif
 995 
 996   __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier), rax);
 997 
 998 #ifdef _LP64
 999   __ pop(r15);
1000   __ pop(r14);
1001   __ pop(r13);
1002   __ pop(r12);
1003   __ pop(r11);
1004   __ pop(r10);
1005   __ pop(r9);
1006   __ pop(r8);
1007 #endif
1008   __ pop(rsi);
1009   __ pop(rdi);
1010   __ pop(rdx);
1011   __ pop(rcx);
1012 
1013   __ pop(tmp2);
1014   __ pop(tmp1);
1015   __ ret(0);
1016 
1017   return start;
1018 }
1019 
1020 #undef __
1021 
1022 void ShenandoahBarrierSetAssembler::barrier_stubs_init() {
1023   if (ShenandoahLoadRefBarrier) {
1024     int stub_code_size = 4096;
1025     ResourceMark rm;
1026     BufferBlob* bb = BufferBlob::create("shenandoah_barrier_stubs", stub_code_size);
1027     CodeBuffer buf(bb);
1028     StubCodeGenerator cgen(&buf);
1029     _shenandoah_lrb = generate_shenandoah_lrb(&cgen);
1030   }
1031 }