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.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::storeval_barrier(MacroAssembler* masm, Register dst, Register tmp) {
 405   if (ShenandoahStoreValEnqueueBarrier) {
 406     storeval_barrier_impl(masm, dst, tmp);
 407   }
 408 }
 409 
 410 void ShenandoahBarrierSetAssembler::storeval_barrier_impl(MacroAssembler* masm, Register dst, Register tmp) {
 411   assert(ShenandoahStoreValEnqueueBarrier, "should be enabled");
 412 
 413   if (dst == noreg) return;
 414 
 415   if (ShenandoahStoreValEnqueueBarrier) {
 416     // The set of registers to be saved+restored is the same as in the write-barrier above.
 417     // Those are the commonly used registers in the interpreter.
 418     __ pusha();
 419     // __ push_callee_saved_registers();
 420     __ subptr(rsp, 2 * Interpreter::stackElementSize);
 421     __ movdbl(Address(rsp, 0), xmm0);
 422 
 423 #ifdef _LP64
 424     Register thread = r15_thread;
 425 #else
 426     Register thread = rcx;
 427     if (thread == dst || thread == tmp) {
 428       thread = rdi;
 429     }
 430     if (thread == dst || thread == tmp) {
 431       thread = rbx;
 432     }
 433     __ get_thread(thread);
 434 #endif
 435     assert_different_registers(dst, tmp, thread);
 436 
 437     satb_write_barrier_pre(masm, noreg, dst, thread, tmp, true, false);
 438     __ movdbl(xmm0, Address(rsp, 0));
 439     __ addptr(rsp, 2 * Interpreter::stackElementSize);
 440     //__ pop_callee_saved_registers();
 441     __ popa();
 442   }
 443 }
 444 
 445 void ShenandoahBarrierSetAssembler::load_reference_barrier(MacroAssembler* masm, Register dst) {
 446   if (ShenandoahLoadRefBarrier) {
 447     Label done;
 448     __ testptr(dst, dst);
 449     __ jcc(Assembler::zero, done);
 450     load_reference_barrier_not_null(masm, dst);
 451     __ bind(done);
 452   }
 453 }
 454 
 455 void ShenandoahBarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 456              Register dst, Address src, Register tmp1, Register tmp_thread) {
 457   bool on_oop = type == T_OBJECT || type == T_ARRAY;
 458   bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
 459   bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
 460   bool on_reference = on_weak || on_phantom;
 461    BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp_thread);
 462   if (on_oop) {
 463     load_reference_barrier(masm, dst);
 464 
 465     if (ShenandoahKeepAliveBarrier && on_reference) {
 466       const Register thread = NOT_LP64(tmp_thread) LP64_ONLY(r15_thread);
 467       assert_different_registers(dst, tmp1, tmp_thread);
 468       NOT_LP64(__ get_thread(thread));
 469       // Generate the SATB pre-barrier code to log the value of
 470       // the referent field in an SATB buffer.
 471       shenandoah_write_barrier_pre(masm /* masm */,
 472                                    noreg /* obj */,
 473                                    dst /* pre_val */,
 474                                    thread /* thread */,
 475                                    tmp1 /* tmp */,
 476                                    true /* tosca_live */,
 477                                    true /* expand_call */);
 478     }
 479   }
 480 }
 481 
 482 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
 483               Address dst, Register val, Register tmp1, Register tmp2) {
 484 
 485   bool on_oop = type == T_OBJECT || type == T_ARRAY;
 486   bool in_heap = (decorators & IN_HEAP) != 0;
 487   bool as_normal = (decorators & AS_NORMAL) != 0;
 488   if (on_oop && in_heap) {
 489     bool needs_pre_barrier = as_normal;
 490 
 491     Register tmp3 = LP64_ONLY(r8) NOT_LP64(rsi);
 492     Register rthread = LP64_ONLY(r15_thread) NOT_LP64(rcx);
 493     // flatten object address if needed
 494     // We do it regardless of precise because we need the registers
 495     if (dst.index() == noreg && dst.disp() == 0) {
 496       if (dst.base() != tmp1) {
 497         __ movptr(tmp1, dst.base());
 498       }
 499     } else {
 500       __ lea(tmp1, dst);
 501     }
 502 
 503     assert_different_registers(val, tmp1, tmp2, tmp3, rthread);
 504 
 505 #ifndef _LP64
 506     __ get_thread(rthread);
 507     InterpreterMacroAssembler *imasm = static_cast<InterpreterMacroAssembler*>(masm);
 508     imasm->save_bcp();
 509 #endif
 510 
 511     if (needs_pre_barrier) {
 512       shenandoah_write_barrier_pre(masm /*masm*/,
 513                                    tmp1 /* obj */,
 514                                    tmp2 /* pre_val */,
 515                                    rthread /* thread */,
 516                                    tmp3  /* tmp */,
 517                                    val != noreg /* tosca_live */,
 518                                    false /* expand_call */);
 519     }
 520     if (val == noreg) {
 521       BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
 522     } else {
 523       storeval_barrier(masm, val, tmp3);
 524       BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
 525     }
 526     NOT_LP64(imasm->restore_bcp());
 527   } else {
 528     BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2);
 529   }
 530 }
 531 
 532 // Special Shenandoah CAS implementation that handles false negatives
 533 // due to concurrent evacuation.
 534 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm,
 535                                                 Register res, Address addr, Register oldval, Register newval,
 536                                                 bool exchange, Register tmp1, Register tmp2) {
 537   assert(ShenandoahCASBarrier, "Should only be used when CAS barrier is enabled");
 538   assert(oldval == rax, "must be in rax for implicit use in cmpxchg");
 539 
 540   Label retry, done;
 541 
 542   // Remember oldval for retry logic below
 543 #ifdef _LP64
 544   if (UseCompressedOops) {
 545     __ movl(tmp1, oldval);
 546   } else
 547 #endif
 548   {
 549     __ movptr(tmp1, oldval);
 550   }
 551 
 552   // Step 1. Try to CAS with given arguments. If successful, then we are done,
 553   // and can safely return.
 554   if (os::is_MP()) __ lock();
 555 #ifdef _LP64
 556   if (UseCompressedOops) {
 557     __ cmpxchgl(newval, addr);
 558   } else
 559 #endif
 560   {
 561     __ cmpxchgptr(newval, addr);
 562   }
 563   __ jcc(Assembler::equal, done, true);
 564 
 565   // Step 2. CAS had failed. This may be a false negative.
 566   //
 567   // The trouble comes when we compare the to-space pointer with the from-space
 568   // pointer to the same object. To resolve this, it will suffice to resolve both
 569   // oldval and the value from memory -- this will give both to-space pointers.
 570   // If they mismatch, then it was a legitimate failure.
 571   //
 572 #ifdef _LP64
 573   if (UseCompressedOops) {
 574     __ decode_heap_oop(tmp1);
 575   }
 576 #endif
 577   resolve_forward_pointer(masm, tmp1);
 578 
 579 #ifdef _LP64
 580   if (UseCompressedOops) {
 581     __ movl(tmp2, oldval);
 582     __ decode_heap_oop(tmp2);
 583   } else
 584 #endif
 585   {
 586     __ movptr(tmp2, oldval);
 587   }
 588   resolve_forward_pointer(masm, tmp2);
 589 
 590   __ cmpptr(tmp1, tmp2);
 591   __ jcc(Assembler::notEqual, done, true);
 592 
 593   // Step 3. Try to CAS again with resolved to-space pointers.
 594   //
 595   // Corner case: it may happen that somebody stored the from-space pointer
 596   // to memory while we were preparing for retry. Therefore, we can fail again
 597   // on retry, and so need to do this in loop, always resolving the failure
 598   // witness.
 599   __ bind(retry);
 600   if (os::is_MP()) __ lock();
 601 #ifdef _LP64
 602   if (UseCompressedOops) {
 603     __ cmpxchgl(newval, addr);
 604   } else
 605 #endif
 606   {
 607     __ cmpxchgptr(newval, addr);
 608   }
 609   __ jcc(Assembler::equal, done, true);
 610 
 611 #ifdef _LP64
 612   if (UseCompressedOops) {
 613     __ movl(tmp2, oldval);
 614     __ decode_heap_oop(tmp2);
 615   } else
 616 #endif
 617   {
 618     __ movptr(tmp2, oldval);
 619   }
 620   resolve_forward_pointer(masm, tmp2);
 621 
 622   __ cmpptr(tmp1, tmp2);
 623   __ jcc(Assembler::equal, retry, true);
 624 
 625   // Step 4. If we need a boolean result out of CAS, check the flag again,
 626   // and promote the result. Note that we handle the flag from both the CAS
 627   // itself and from the retry loop.
 628   __ bind(done);
 629   if (!exchange) {
 630     assert(res != NULL, "need result register");
 631 #ifdef _LP64
 632     __ setb(Assembler::equal, res);
 633     __ movzbl(res, res);
 634 #else
 635     // Need something else to clean the result, because some registers
 636     // do not have byte encoding that movzbl wants. Cannot do the xor first,
 637     // because it modifies the flags.
 638     Label res_non_zero;
 639     __ movptr(res, 1);
 640     __ jcc(Assembler::equal, res_non_zero, true);
 641     __ xorptr(res, res);
 642     __ bind(res_non_zero);
 643 #endif
 644   }
 645 }
 646 
 647 void ShenandoahBarrierSetAssembler::save_vector_registers(MacroAssembler* masm) {
 648   int num_xmm_regs = LP64_ONLY(16) NOT_LP64(8);
 649   if (UseAVX > 2) {
 650     num_xmm_regs = LP64_ONLY(32) NOT_LP64(8);
 651   }
 652 
 653   if (UseSSE == 1)  {
 654     __ subptr(rsp, sizeof(jdouble)*8);
 655     for (int n = 0; n < 8; n++) {
 656       __ movflt(Address(rsp, n*sizeof(jdouble)), as_XMMRegister(n));
 657     }
 658   } else if (UseSSE >= 2)  {
 659     if (UseAVX > 2) {
 660       __ push(rbx);
 661       __ movl(rbx, 0xffff);
 662       __ kmovwl(k1, rbx);
 663       __ pop(rbx);
 664     }
 665 #ifdef COMPILER2
 666     if (MaxVectorSize > 16) {
 667       if(UseAVX > 2) {
 668         // Save upper half of ZMM registers
 669         __ subptr(rsp, 32*num_xmm_regs);
 670         for (int n = 0; n < num_xmm_regs; n++) {
 671           __ vextractf64x4_high(Address(rsp, n*32), as_XMMRegister(n));
 672         }
 673       }
 674       assert(UseAVX > 0, "256 bit vectors are supported only with AVX");
 675       // Save upper half of YMM registers
 676       __ subptr(rsp, 16*num_xmm_regs);
 677       for (int n = 0; n < num_xmm_regs; n++) {
 678         __ vextractf128_high(Address(rsp, n*16), as_XMMRegister(n));
 679       }
 680     }
 681 #endif
 682     // Save whole 128bit (16 bytes) XMM registers
 683     __ subptr(rsp, 16*num_xmm_regs);
 684 #ifdef _LP64
 685     if (VM_Version::supports_evex()) {
 686       for (int n = 0; n < num_xmm_regs; n++) {
 687         __ vextractf32x4(Address(rsp, n*16), as_XMMRegister(n), 0);
 688       }
 689     } else {
 690       for (int n = 0; n < num_xmm_regs; n++) {
 691         __ movdqu(Address(rsp, n*16), as_XMMRegister(n));
 692       }
 693     }
 694 #else
 695     for (int n = 0; n < num_xmm_regs; n++) {
 696       __ movdqu(Address(rsp, n*16), as_XMMRegister(n));
 697     }
 698 #endif
 699   }
 700 }
 701 
 702 void ShenandoahBarrierSetAssembler::restore_vector_registers(MacroAssembler* masm) {
 703   int num_xmm_regs = LP64_ONLY(16) NOT_LP64(8);
 704   if (UseAVX > 2) {
 705     num_xmm_regs = LP64_ONLY(32) NOT_LP64(8);
 706   }
 707   if (UseSSE == 1)  {
 708     for (int n = 0; n < 8; n++) {
 709       __ movflt(as_XMMRegister(n), Address(rsp, n*sizeof(jdouble)));
 710     }
 711     __ addptr(rsp, sizeof(jdouble)*8);
 712   } else if (UseSSE >= 2)  {
 713     // Restore whole 128bit (16 bytes) XMM registers
 714 #ifdef _LP64
 715     if (VM_Version::supports_evex()) {
 716       for (int n = 0; n < num_xmm_regs; n++) {
 717         __ vinsertf32x4(as_XMMRegister(n), as_XMMRegister(n), Address(rsp, n*16), 0);
 718       }
 719     } else {
 720       for (int n = 0; n < num_xmm_regs; n++) {
 721         __ movdqu(as_XMMRegister(n), Address(rsp, n*16));
 722       }
 723     }
 724 #else
 725     for (int n = 0; n < num_xmm_regs; n++) {
 726       __ movdqu(as_XMMRegister(n), Address(rsp, n*16));
 727     }
 728 #endif
 729     __ addptr(rsp, 16*num_xmm_regs);
 730 
 731 #ifdef COMPILER2
 732     if (MaxVectorSize > 16) {
 733       // Restore upper half of YMM registers.
 734       for (int n = 0; n < num_xmm_regs; n++) {
 735         __ vinsertf128_high(as_XMMRegister(n), Address(rsp, n*16));
 736       }
 737       __ addptr(rsp, 16*num_xmm_regs);
 738       if (UseAVX > 2) {
 739         for (int n = 0; n < num_xmm_regs; n++) {
 740           __ vinsertf64x4_high(as_XMMRegister(n), Address(rsp, n*32));
 741         }
 742         __ addptr(rsp, 32*num_xmm_regs);
 743       }
 744     }
 745 #endif
 746   }
 747 }
 748 
 749 #undef __
 750 
 751 #ifdef COMPILER1
 752 
 753 #define __ ce->masm()->
 754 
 755 void ShenandoahBarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, ShenandoahPreBarrierStub* stub) {
 756   ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
 757   // At this point we know that marking is in progress.
 758   // If do_load() is true then we have to emit the
 759   // load of the previous value; otherwise it has already
 760   // been loaded into _pre_val.
 761 
 762   __ bind(*stub->entry());
 763   assert(stub->pre_val()->is_register(), "Precondition.");
 764 
 765   Register pre_val_reg = stub->pre_val()->as_register();
 766 
 767   if (stub->do_load()) {
 768     ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/, false /*unaligned*/);
 769   }
 770 
 771   __ cmpptr(pre_val_reg, (int32_t)NULL_WORD);
 772   __ jcc(Assembler::equal, *stub->continuation());
 773   ce->store_parameter(stub->pre_val()->as_register(), 0);
 774   __ call(RuntimeAddress(bs->pre_barrier_c1_runtime_code_blob()->code_begin()));
 775   __ jmp(*stub->continuation());
 776 
 777 }
 778 
 779 void ShenandoahBarrierSetAssembler::gen_load_reference_barrier_stub(LIR_Assembler* ce, ShenandoahLoadReferenceBarrierStub* stub) {
 780   __ bind(*stub->entry());
 781 
 782   Label done;
 783   Register obj = stub->obj()->as_register();
 784   Register res = stub->result()->as_register();
 785 
 786   if (res != obj) {
 787     __ mov(res, obj);
 788   }
 789 
 790   // Check for null.
 791   __ testptr(res, res);
 792   __ jcc(Assembler::zero, done);
 793 
 794   load_reference_barrier_not_null(ce->masm(), res);
 795 
 796   __ bind(done);
 797   __ jmp(*stub->continuation());
 798 }
 799 
 800 #undef __
 801 
 802 #define __ sasm->
 803 
 804 void ShenandoahBarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) {
 805   __ prologue("shenandoah_pre_barrier", false);
 806   // arg0 : previous value of memory
 807 
 808   __ push(rax);
 809   __ push(rdx);
 810 
 811   const Register pre_val = rax;
 812   const Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
 813   const Register tmp = rdx;
 814 
 815   NOT_LP64(__ get_thread(thread);)
 816 
 817   Address queue_index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
 818   Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
 819 
 820   Label done;
 821   Label runtime;
 822 
 823   // Is SATB still active?
 824   Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 825   __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL);
 826   __ jcc(Assembler::zero, done);
 827 
 828   // Can we store original value in the thread's buffer?
 829 
 830   __ movptr(tmp, queue_index);
 831   __ testptr(tmp, tmp);
 832   __ jcc(Assembler::zero, runtime);
 833   __ subptr(tmp, wordSize);
 834   __ movptr(queue_index, tmp);
 835   __ addptr(tmp, buffer);
 836 
 837   // prev_val (rax)
 838   __ load_parameter(0, pre_val);
 839   __ movptr(Address(tmp, 0), pre_val);
 840   __ jmp(done);
 841 
 842   __ bind(runtime);
 843 
 844   __ save_live_registers_no_oop_map(true);
 845 
 846   // load the pre-value
 847   __ load_parameter(0, rcx);
 848   __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), rcx, thread);
 849 
 850   __ restore_live_registers(true);
 851 
 852   __ bind(done);
 853 
 854   __ pop(rdx);
 855   __ pop(rax);
 856 
 857   __ epilogue();
 858 }
 859 
 860 #undef __
 861 
 862 #endif // COMPILER1
 863 
 864 address ShenandoahBarrierSetAssembler::shenandoah_lrb() {
 865   assert(_shenandoah_lrb != NULL, "need load reference barrier stub");
 866   return _shenandoah_lrb;
 867 }
 868 
 869 #define __ cgen->assembler()->
 870 
 871 address ShenandoahBarrierSetAssembler::generate_shenandoah_lrb(StubCodeGenerator* cgen) {
 872   __ align(CodeEntryAlignment);
 873   StubCodeMark mark(cgen, "StubRoutines", "shenandoah_lrb");
 874   address start = __ pc();
 875 
 876   Label resolve_oop, slow_path;
 877 
 878   // We use RDI, which also serves as argument register for slow call.
 879   // RAX always holds the src object ptr, except after the slow call and
 880   // the cmpxchg, then it holds the result. R8/RBX is used as temporary register.
 881 
 882   Register tmp1 = rdi;
 883   Register tmp2 = LP64_ONLY(r8) NOT_LP64(rbx);
 884 
 885   __ push(tmp1);
 886   __ push(tmp2);
 887 
 888   // Check for object being in the collection set.
 889   // TODO: Can we use only 1 register here?
 890   // The source object arrives here in rax.
 891   // live: rax
 892   // live: tmp1
 893   __ mov(tmp1, rax);
 894   __ shrptr(tmp1, ShenandoahHeapRegion::region_size_bytes_shift_jint());
 895   // live: tmp2
 896   __ movptr(tmp2, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr());
 897   __ movbool(tmp2, Address(tmp2, tmp1, Address::times_1));
 898   // unlive: tmp1
 899   __ testbool(tmp2);
 900   // unlive: tmp2
 901   __ jccb(Assembler::notZero, resolve_oop);
 902 
 903   __ pop(tmp2);
 904   __ pop(tmp1);
 905   __ ret(0);
 906 
 907   __ bind(resolve_oop);
 908 
 909   __ movptr(tmp2, Address(rax, oopDesc::mark_offset_in_bytes()));
 910   // Test if both lowest bits are set. We trick it by negating the bits
 911   // then test for both bits clear.
 912   __ notptr(tmp2);
 913   __ testb(tmp2, markOopDesc::marked_value);
 914   __ jccb(Assembler::notZero, slow_path);
 915   // Clear both lower bits. It's still inverted, so set them, and then invert back.
 916   __ orptr(tmp2, markOopDesc::marked_value);
 917   __ notptr(tmp2);
 918   // At this point, tmp2 contains the decoded forwarding pointer.
 919   __ mov(rax, tmp2);
 920 
 921   __ pop(tmp2);
 922   __ pop(tmp1);
 923   __ ret(0);
 924 
 925   __ bind(slow_path);
 926 
 927   __ push(rcx);
 928   __ push(rdx);
 929   __ push(rdi);
 930   __ push(rsi);
 931 #ifdef _LP64
 932   __ push(r8);
 933   __ push(r9);
 934   __ push(r10);
 935   __ push(r11);
 936   __ push(r12);
 937   __ push(r13);
 938   __ push(r14);
 939   __ push(r15);
 940 #endif
 941 
 942   save_vector_registers(cgen->assembler());
 943   __ movptr(rdi, rax);
 944   __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier), rdi);
 945   restore_vector_registers(cgen->assembler());
 946 
 947 #ifdef _LP64
 948   __ pop(r15);
 949   __ pop(r14);
 950   __ pop(r13);
 951   __ pop(r12);
 952   __ pop(r11);
 953   __ pop(r10);
 954   __ pop(r9);
 955   __ pop(r8);
 956 #endif
 957   __ pop(rsi);
 958   __ pop(rdi);
 959   __ pop(rdx);
 960   __ pop(rcx);
 961 
 962   __ pop(tmp2);
 963   __ pop(tmp1);
 964   __ ret(0);
 965 
 966   return start;
 967 }
 968 
 969 #undef __
 970 
 971 void ShenandoahBarrierSetAssembler::barrier_stubs_init() {
 972   if (ShenandoahLoadRefBarrier) {
 973     int stub_code_size = 4096;
 974     ResourceMark rm;
 975     BufferBlob* bb = BufferBlob::create("shenandoah_barrier_stubs", stub_code_size);
 976     CodeBuffer buf(bb);
 977     StubCodeGenerator cgen(&buf);
 978     _shenandoah_lrb = generate_shenandoah_lrb(&cgen);
 979   }
 980 }