1 /* 2 * Copyright (c) 2018, Red Hat, Inc. and/or its affiliates. 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/brooksPointer.hpp" 26 #include "gc/shenandoah/shenandoahBaseBarrierSetAssembler.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 ShenandoahBaseBarrierSetAssembler::_shenandoah_wb = NULL; 46 address ShenandoahBaseBarrierSetAssembler::_shenandoah_wb_C = NULL; 47 48 void ShenandoahBaseBarrierSetAssembler::arraycopy_prologue(MacroAssembler* masm, DecoratorSet decorators, BasicType type, 49 Register src, Register dst, Register count) { 50 51 bool checkcast = (decorators & ARRAYCOPY_CHECKCAST) != 0; 52 bool disjoint = (decorators & ARRAYCOPY_DISJOINT) != 0; 53 bool obj_int = type == T_OBJECT LP64_ONLY(&& UseCompressedOops); 54 bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0; 55 56 if (type == T_OBJECT || type == T_ARRAY) { 57 #ifdef _LP64 58 if (!checkcast && !obj_int) { 59 // Save count for barrier 60 __ movptr(r11, count); 61 } else if (disjoint && obj_int) { 62 // Save dst in r11 in the disjoint case 63 __ movq(r11, dst); 64 } 65 #else 66 if (disjoint) { 67 __ mov(rdx, dst); // save 'to' 68 } 69 #endif 70 71 if (!dest_uninitialized && !ShenandoahHeap::heap()->heuristics()->can_do_traversal_gc()) { 72 Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread); 73 #ifndef _LP64 74 __ push(thread); 75 __ get_thread(thread); 76 #endif 77 78 Label filtered; 79 Address in_progress(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset())); 80 // Is marking active? 81 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) { 82 __ cmpl(in_progress, 0); 83 } else { 84 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption"); 85 __ cmpb(in_progress, 0); 86 } 87 88 NOT_LP64(__ pop(thread);) 89 90 __ jcc(Assembler::equal, filtered); 91 92 __ pusha(); // push registers 93 #ifdef _LP64 94 if (count == c_rarg0) { 95 if (dst == c_rarg1) { 96 // exactly backwards!! 97 __ xchgptr(c_rarg1, c_rarg0); 98 } else { 99 __ movptr(c_rarg1, count); 100 __ movptr(c_rarg0, dst); 101 } 102 } else { 103 __ movptr(c_rarg0, dst); 104 __ movptr(c_rarg1, count); 105 } 106 if (UseCompressedOops) { 107 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_narrow_oop_entry), 2); 108 } else { 109 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry), 2); 110 } 111 #else 112 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry), 113 dst, count); 114 #endif 115 __ popa(); 116 __ bind(filtered); 117 } 118 } 119 120 } 121 122 void ShenandoahBaseBarrierSetAssembler::arraycopy_epilogue(MacroAssembler* masm, DecoratorSet decorators, BasicType type, 123 Register src, Register dst, Register count) { 124 bool checkcast = (decorators & ARRAYCOPY_CHECKCAST) != 0; 125 bool disjoint = (decorators & ARRAYCOPY_DISJOINT) != 0; 126 bool obj_int = type == T_OBJECT LP64_ONLY(&& UseCompressedOops); 127 Register tmp = rax; 128 129 if (type == T_OBJECT || type == T_ARRAY) { 130 #ifdef _LP64 131 if (!checkcast && !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 } else if (checkcast) { 138 tmp = rscratch1; 139 } 140 #else 141 if (disjoint) { 142 __ mov(dst, rdx); // restore 'to' 143 } 144 #endif 145 146 __ pusha(); // push registers (overkill) 147 #ifdef _LP64 148 if (c_rarg0 == count) { // On win64 c_rarg0 == rcx 149 assert_different_registers(c_rarg1, dst); 150 __ mov(c_rarg1, count); 151 __ mov(c_rarg0, dst); 152 } else { 153 assert_different_registers(c_rarg0, count); 154 __ mov(c_rarg0, dst); 155 __ mov(c_rarg1, count); 156 } 157 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry), 2); 158 #else 159 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry), 160 dst, count); 161 #endif 162 __ popa(); 163 } 164 } 165 166 void ShenandoahBaseBarrierSetAssembler::shenandoah_write_barrier_pre(MacroAssembler* masm, 167 Register obj, 168 Register pre_val, 169 Register thread, 170 Register tmp, 171 bool tosca_live, 172 bool expand_call) { 173 174 if (ShenandoahSATBBarrier) { 175 satb_write_barrier_pre(masm, obj, pre_val, thread, tmp, tosca_live, expand_call); 176 } 177 } 178 179 void ShenandoahBaseBarrierSetAssembler::satb_write_barrier_pre(MacroAssembler* masm, 180 Register obj, 181 Register pre_val, 182 Register thread, 183 Register tmp, 184 bool tosca_live, 185 bool expand_call) { 186 // If expand_call is true then we expand the call_VM_leaf macro 187 // directly to skip generating the check by 188 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp. 189 190 #ifdef _LP64 191 assert(thread == r15_thread, "must be"); 192 #endif // _LP64 193 194 Label done; 195 Label runtime; 196 197 assert(pre_val != noreg, "check this code"); 198 199 if (obj != noreg) { 200 assert_different_registers(obj, pre_val, tmp); 201 assert(pre_val != rax, "check this code"); 202 } 203 204 Address in_progress(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset())); 205 Address index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset())); 206 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())); 207 208 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset())); 209 __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL); 210 __ jcc(Assembler::zero, done); 211 212 // Do we need to load the previous value? 213 if (obj != noreg) { 214 __ load_heap_oop(pre_val, Address(obj, 0), noreg, noreg, AS_RAW); 215 } 216 217 // Is the previous value null? 218 __ cmpptr(pre_val, (int32_t) NULL_WORD); 219 __ jcc(Assembler::equal, done); 220 221 // Can we store original value in the thread's buffer? 222 // Is index == 0? 223 // (The index field is typed as size_t.) 224 225 __ movptr(tmp, index); // tmp := *index_adr 226 __ cmpptr(tmp, 0); // tmp == 0? 227 __ jcc(Assembler::equal, runtime); // If yes, goto runtime 228 229 __ subptr(tmp, wordSize); // tmp := tmp - wordSize 230 __ movptr(index, tmp); // *index_adr := tmp 231 __ addptr(tmp, buffer); // tmp := tmp + *buffer_adr 232 233 // Record the previous value 234 __ movptr(Address(tmp, 0), pre_val); 235 __ jmp(done); 236 237 __ bind(runtime); 238 // save the live input values 239 if(tosca_live) __ push(rax); 240 241 if (obj != noreg && obj != rax) 242 __ push(obj); 243 244 if (pre_val != rax) 245 __ push(pre_val); 246 247 // Calling the runtime using the regular call_VM_leaf mechanism generates 248 // code (generated by InterpreterMacroAssember::call_VM_leaf_base) 249 // that checks that the *(ebp+frame::interpreter_frame_last_sp) == NULL. 250 // 251 // If we care generating the pre-barrier without a frame (e.g. in the 252 // intrinsified Reference.get() routine) then ebp might be pointing to 253 // the caller frame and so this check will most likely fail at runtime. 254 // 255 // Expanding the call directly bypasses the generation of the check. 256 // So when we do not have have a full interpreter frame on the stack 257 // expand_call should be passed true. 258 259 NOT_LP64( __ push(thread); ) 260 261 #ifdef _LP64 262 // We move pre_val into c_rarg0 early, in order to avoid smashing it, should 263 // pre_val be c_rarg1 (where the call prologue would copy thread argument). 264 // Note: this should not accidentally smash thread, because thread is always r15. 265 assert(thread != c_rarg0, "smashed arg"); 266 if (c_rarg0 != pre_val) { 267 __ mov(c_rarg0, pre_val); 268 } 269 #endif 270 271 if (expand_call) { 272 LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); ) 273 #ifdef _LP64 274 if (c_rarg1 != thread) { 275 __ mov(c_rarg1, thread); 276 } 277 // Already moved pre_val into c_rarg0 above 278 #else 279 __ push(thread); 280 __ push(pre_val); 281 #endif 282 __ MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), 2); 283 } else { 284 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), LP64_ONLY(c_rarg0) NOT_LP64(pre_val), thread); 285 } 286 287 NOT_LP64( __ pop(thread); ) 288 289 // save the live input values 290 if (pre_val != rax) 291 __ pop(pre_val); 292 293 if (obj != noreg && obj != rax) 294 __ pop(obj); 295 296 if(tosca_live) __ pop(rax); 297 298 __ bind(done); 299 } 300 301 void ShenandoahBaseBarrierSetAssembler::read_barrier(MacroAssembler* masm, Register dst) { 302 if (ShenandoahReadBarrier) { 303 read_barrier_impl(masm, dst); 304 } 305 } 306 307 void ShenandoahBaseBarrierSetAssembler::read_barrier_impl(MacroAssembler* masm, Register dst) { 308 assert(UseShenandoahGC && (ShenandoahReadBarrier || ShenandoahStoreValReadBarrier || ShenandoahCASBarrier), "should be enabled"); 309 Label is_null; 310 __ testptr(dst, dst); 311 __ jcc(Assembler::zero, is_null); 312 read_barrier_not_null_impl(masm, dst); 313 __ bind(is_null); 314 } 315 316 void ShenandoahBaseBarrierSetAssembler::read_barrier_not_null(MacroAssembler* masm, Register dst) { 317 if (ShenandoahReadBarrier) { 318 read_barrier_not_null_impl(masm, dst); 319 } 320 } 321 322 void ShenandoahBaseBarrierSetAssembler::read_barrier_not_null_impl(MacroAssembler* masm, Register dst) { 323 assert(UseShenandoahGC && (ShenandoahReadBarrier || ShenandoahStoreValReadBarrier || ShenandoahCASBarrier), "should be enabled"); 324 __ movptr(dst, Address(dst, BrooksPointer::byte_offset())); 325 } 326 327 328 void ShenandoahBaseBarrierSetAssembler::write_barrier(MacroAssembler* masm, Register dst) { 329 if (ShenandoahWriteBarrier) { 330 write_barrier_impl(masm, dst); 331 } 332 } 333 334 void ShenandoahBaseBarrierSetAssembler::write_barrier_impl(MacroAssembler* masm, Register dst) { 335 assert(UseShenandoahGC && (ShenandoahWriteBarrier || ShenandoahStoreValEnqueueBarrier), "Should be enabled"); 336 #ifdef _LP64 337 Label done; 338 339 Address gc_state(r15_thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset())); 340 __ testb(gc_state, ShenandoahHeap::HAS_FORWARDED | ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL); 341 __ jccb(Assembler::zero, done); 342 343 // Heap is unstable, need to perform the read-barrier even if WB is inactive 344 if (ShenandoahWriteBarrierRB) { 345 read_barrier_not_null(masm, dst); 346 } 347 348 __ testb(gc_state, ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL); 349 __ jccb(Assembler::zero, done); 350 351 if (dst != rax) { 352 __ xchgptr(dst, rax); // Move obj into rax and save rax into obj. 353 } 354 355 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, ShenandoahBaseBarrierSetAssembler::shenandoah_wb()))); 356 357 if (dst != rax) { 358 __ xchgptr(rax, dst); // Swap back obj with rax. 359 } 360 361 __ bind(done); 362 #else 363 Unimplemented(); 364 #endif 365 } 366 367 void ShenandoahBaseBarrierSetAssembler::storeval_barrier(MacroAssembler* masm, Register dst, Register tmp) { 368 if (ShenandoahStoreValReadBarrier || ShenandoahStoreValEnqueueBarrier) { 369 storeval_barrier_impl(masm, dst, tmp); 370 } 371 } 372 373 void ShenandoahBaseBarrierSetAssembler::storeval_barrier_impl(MacroAssembler* masm, Register dst, Register tmp) { 374 assert(UseShenandoahGC && (ShenandoahStoreValReadBarrier || ShenandoahStoreValEnqueueBarrier), "should be enabled"); 375 376 if (dst == noreg) return; 377 378 #ifdef _LP64 379 if (ShenandoahStoreValEnqueueBarrier) { 380 Label is_null; 381 __ testptr(dst, dst); 382 __ jcc(Assembler::zero, is_null); 383 write_barrier_impl(masm, dst); 384 __ bind(is_null); 385 386 enqueue_barrier(masm, dst, tmp); 387 } 388 if (ShenandoahStoreValReadBarrier) { 389 read_barrier_impl(masm, dst); 390 } 391 #else 392 Unimplemented(); 393 #endif 394 } 395 396 void ShenandoahBaseBarrierSetAssembler::enqueue_barrier(MacroAssembler* masm, Register dst, Register tmp) { 397 if (ShenandoahStoreValEnqueueBarrier) { 398 // The set of registers to be saved+restored is the same as in the write-barrier above. 399 // Those are the commonly used registers in the interpreter. 400 __ pusha(); 401 // __ push_callee_saved_registers(); 402 __ subptr(rsp, 2 * Interpreter::stackElementSize); 403 __ movdbl(Address(rsp, 0), xmm0); 404 405 satb_write_barrier_pre(masm, noreg, dst, r15_thread, tmp, true, false); 406 __ movdbl(xmm0, Address(rsp, 0)); 407 __ addptr(rsp, 2 * Interpreter::stackElementSize); 408 //__ pop_callee_saved_registers(); 409 __ popa(); 410 } 411 } 412 413 void ShenandoahBaseBarrierSetAssembler::tlab_allocate(MacroAssembler* masm, 414 Register thread, Register obj, 415 Register var_size_in_bytes, 416 int con_size_in_bytes, 417 Register t1, Register t2, 418 Label& slow_case) { 419 assert_different_registers(obj, t1, t2); 420 assert_different_registers(obj, var_size_in_bytes, t1); 421 Register end = t2; 422 if (!thread->is_valid()) { 423 #ifdef _LP64 424 thread = r15_thread; 425 #else 426 assert(t1->is_valid(), "need temp reg"); 427 thread = t1; 428 __ get_thread(thread); 429 #endif 430 } 431 432 __ verify_tlab(); 433 434 __ movptr(obj, Address(thread, JavaThread::tlab_top_offset())); 435 if (var_size_in_bytes == noreg) { 436 __ lea(end, Address(obj, con_size_in_bytes + BrooksPointer::byte_size())); 437 } else { 438 __ addptr(var_size_in_bytes, BrooksPointer::byte_size()); 439 __ lea(end, Address(obj, var_size_in_bytes, Address::times_1)); 440 } 441 __ cmpptr(end, Address(thread, JavaThread::tlab_end_offset())); 442 __ jcc(Assembler::above, slow_case); 443 444 // update the tlab top pointer 445 __ movptr(Address(thread, JavaThread::tlab_top_offset()), end); 446 447 // Initialize brooks pointer 448 #ifdef _LP64 449 __ incrementq(obj, BrooksPointer::byte_size()); 450 #else 451 __ incrementl(obj, BrooksPointer::byte_size()); 452 #endif 453 __ movptr(Address(obj, BrooksPointer::byte_offset()), obj); 454 455 // recover var_size_in_bytes if necessary 456 if (var_size_in_bytes == end) { 457 __ subptr(var_size_in_bytes, obj); 458 } 459 __ verify_tlab(); 460 } 461 462 463 // Special Shenandoah CAS implementation that handles false negatives 464 // due to concurrent evacuation. 465 #ifndef _LP64 466 void ShenandoahBaseBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm, 467 Register res, Address addr, Register oldval, Register newval, 468 bool exchange, bool encode, Register tmp1, Register tmp2) { 469 // Shenandoah has no 32-bit version for this. 470 Unimplemented(); 471 } 472 #else 473 void ShenandoahBaseBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm, 474 Register res, Address addr, Register oldval, Register newval, 475 bool exchange, bool encode, Register tmp1, Register tmp2) { 476 if (!ShenandoahCASBarrier) { 477 #ifdef _LP64 478 if (UseCompressedOops) { 479 if (encode) { 480 __ encode_heap_oop(oldval); 481 __ mov(rscratch1, newval); 482 __ encode_heap_oop(rscratch1); 483 newval = rscratch1; 484 } 485 if (os::is_MP()) { 486 __ lock(); 487 } 488 // oldval (rax) is implicitly used by this instruction 489 __ cmpxchgl(newval, addr); 490 } else 491 #endif 492 { 493 if (os::is_MP()) { 494 __ lock(); 495 } 496 __ cmpxchgptr(newval, addr); 497 } 498 499 if (!exchange) { 500 assert(res != NULL, "need result register"); 501 __ setb(Assembler::equal, res); 502 __ movzbl(res, res); 503 } 504 return; 505 } 506 507 assert(ShenandoahCASBarrier, "Should only be used when CAS barrier is enabled"); 508 assert(oldval == rax, "must be in rax for implicit use in cmpxchg"); 509 510 Label retry, done; 511 512 // Apply storeval barrier to newval. 513 if (encode) { 514 storeval_barrier(masm, newval, tmp1); 515 } 516 517 if (UseCompressedOops) { 518 if (encode) { 519 __ encode_heap_oop(oldval); 520 __ mov(rscratch1, newval); 521 __ encode_heap_oop(rscratch1); 522 newval = rscratch1; 523 } 524 } 525 526 // Remember oldval for retry logic below 527 if (UseCompressedOops) { 528 __ movl(tmp1, oldval); 529 } else { 530 __ movptr(tmp1, oldval); 531 } 532 533 // Step 1. Try to CAS with given arguments. If successful, then we are done, 534 // and can safely return. 535 if (os::is_MP()) __ lock(); 536 if (UseCompressedOops) { 537 __ cmpxchgl(newval, addr); 538 } else { 539 __ cmpxchgptr(newval, addr); 540 } 541 __ jcc(Assembler::equal, done, true); 542 543 // Step 2. CAS had failed. This may be a false negative. 544 // 545 // The trouble comes when we compare the to-space pointer with the from-space 546 // pointer to the same object. To resolve this, it will suffice to read both 547 // oldval and the value from memory through the read barriers -- this will give 548 // both to-space pointers. If they mismatch, then it was a legitimate failure. 549 // 550 if (UseCompressedOops) { 551 __ decode_heap_oop(tmp1); 552 } 553 read_barrier_impl(masm, tmp1); 554 555 if (UseCompressedOops) { 556 __ movl(tmp2, oldval); 557 __ decode_heap_oop(tmp2); 558 } else { 559 __ movptr(tmp2, oldval); 560 } 561 read_barrier_impl(masm, tmp2); 562 563 __ cmpptr(tmp1, tmp2); 564 __ jcc(Assembler::notEqual, done, true); 565 566 // Step 3. Try to CAS again with resolved to-space pointers. 567 // 568 // Corner case: it may happen that somebody stored the from-space pointer 569 // to memory while we were preparing for retry. Therefore, we can fail again 570 // on retry, and so need to do this in loop, always re-reading the failure 571 // witness through the read barrier. 572 __ bind(retry); 573 if (os::is_MP()) __ lock(); 574 if (UseCompressedOops) { 575 __ cmpxchgl(newval, addr); 576 } else { 577 __ cmpxchgptr(newval, addr); 578 } 579 __ jcc(Assembler::equal, done, true); 580 581 if (UseCompressedOops) { 582 __ movl(tmp2, oldval); 583 __ decode_heap_oop(tmp2); 584 } else { 585 __ movptr(tmp2, oldval); 586 } 587 read_barrier_impl(masm, tmp2); 588 589 __ cmpptr(tmp1, tmp2); 590 __ jcc(Assembler::equal, retry, true); 591 592 // Step 4. If we need a boolean result out of CAS, check the flag again, 593 // and promote the result. Note that we handle the flag from both the CAS 594 // itself and from the retry loop. 595 __ bind(done); 596 if (!exchange) { 597 assert(res != NULL, "need result register"); 598 __ setb(Assembler::equal, res); 599 __ movzbl(res, res); 600 } 601 } 602 #endif // LP64 603 604 #ifdef COMPILER1 605 606 #undef __ 607 #define __ ce->masm()-> 608 609 void ShenandoahBaseBarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, ShenandoahPreBarrierStub* stub) { 610 ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1(); 611 // At this point we know that marking is in progress. 612 // If do_load() is true then we have to emit the 613 // load of the previous value; otherwise it has already 614 // been loaded into _pre_val. 615 616 __ bind(*stub->entry()); 617 assert(stub->pre_val()->is_register(), "Precondition."); 618 619 Register pre_val_reg = stub->pre_val()->as_register(); 620 621 if (stub->do_load()) { 622 ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/, false /*unaligned*/); 623 } 624 625 __ cmpptr(pre_val_reg, (int32_t)NULL_WORD); 626 __ jcc(Assembler::equal, *stub->continuation()); 627 ce->store_parameter(stub->pre_val()->as_register(), 0); 628 __ call(RuntimeAddress(bs->pre_barrier_c1_runtime_code_blob()->code_begin())); 629 __ jmp(*stub->continuation()); 630 631 } 632 633 void ShenandoahBaseBarrierSetAssembler::gen_write_barrier_stub(LIR_Assembler* ce, ShenandoahWriteBarrierStub* stub) { 634 __ bind(*stub->entry()); 635 636 Label done; 637 Register obj = stub->obj()->as_register(); 638 Register res = stub->result()->as_register(); 639 640 if (res != obj) { 641 __ mov(res, obj); 642 } 643 644 // Check for null. 645 if (stub->needs_null_check()) { 646 __ testptr(res, res); 647 __ jcc(Assembler::zero, done); 648 } 649 650 write_barrier(ce->masm(), res); 651 652 __ bind(done); 653 __ jmp(*stub->continuation()); 654 } 655 656 #undef __ 657 658 #define __ sasm-> 659 660 void ShenandoahBaseBarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) { 661 __ prologue("shenandoah_pre_barrier", false); 662 // arg0 : previous value of memory 663 664 __ push(rax); 665 __ push(rdx); 666 667 const Register pre_val = rax; 668 const Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread); 669 const Register tmp = rdx; 670 671 NOT_LP64(__ get_thread(thread);) 672 673 Address queue_index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset())); 674 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())); 675 676 Label done; 677 Label runtime; 678 679 // Is SATB still active? 680 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset())); 681 __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL); 682 __ jcc(Assembler::zero, done); 683 684 // Can we store original value in the thread's buffer? 685 686 __ movptr(tmp, queue_index); 687 __ testptr(tmp, tmp); 688 __ jcc(Assembler::zero, runtime); 689 __ subptr(tmp, wordSize); 690 __ movptr(queue_index, tmp); 691 __ addptr(tmp, buffer); 692 693 // prev_val (rax) 694 __ load_parameter(0, pre_val); 695 __ movptr(Address(tmp, 0), pre_val); 696 __ jmp(done); 697 698 __ bind(runtime); 699 700 __ save_live_registers_no_oop_map(true); 701 702 // load the pre-value 703 __ load_parameter(0, rcx); 704 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), rcx, thread); 705 706 __ restore_live_registers(true); 707 708 __ bind(done); 709 710 __ pop(rdx); 711 __ pop(rax); 712 713 __ epilogue(); 714 } 715 716 #undef __ 717 718 #endif // COMPILER1 719 720 address ShenandoahBaseBarrierSetAssembler::shenandoah_wb() { 721 assert(_shenandoah_wb != NULL, "need write barrier stub"); 722 return _shenandoah_wb; 723 } 724 725 address ShenandoahBaseBarrierSetAssembler::shenandoah_wb_C() { 726 assert(_shenandoah_wb_C != NULL, "need write barrier stub"); 727 return _shenandoah_wb_C; 728 } 729 730 #define __ cgen->assembler()-> 731 732 address ShenandoahBaseBarrierSetAssembler::generate_shenandoah_wb(StubCodeGenerator* cgen, bool c_abi, bool do_cset_test) { 733 __ align(CodeEntryAlignment); 734 StubCodeMark mark(cgen, "StubRoutines", "shenandoah_wb"); 735 address start = __ pc(); 736 737 #ifdef _LP64 738 Label not_done; 739 740 // We use RDI, which also serves as argument register for slow call. 741 // RAX always holds the src object ptr, except after the slow call and 742 // the cmpxchg, then it holds the result. 743 // R8 and RCX are used as temporary registers. 744 if (!c_abi) { 745 __ push(rdi); 746 __ push(r8); 747 } 748 749 // Check for object beeing in the collection set. 750 // TODO: Can we use only 1 register here? 751 // The source object arrives here in rax. 752 // live: rax 753 // live: rdi 754 if (!c_abi) { 755 __ mov(rdi, rax); 756 } else { 757 if (rax != c_rarg0) { 758 __ mov(rax, c_rarg0); 759 } 760 } 761 if (do_cset_test) { 762 __ shrptr(rdi, ShenandoahHeapRegion::region_size_bytes_shift_jint()); 763 // live: r8 764 __ movptr(r8, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr()); 765 __ movbool(r8, Address(r8, rdi, Address::times_1)); 766 // unlive: rdi 767 __ testbool(r8); 768 // unlive: r8 769 __ jccb(Assembler::notZero, not_done); 770 771 if (!c_abi) { 772 __ pop(r8); 773 __ pop(rdi); 774 } 775 __ ret(0); 776 777 __ bind(not_done); 778 } 779 780 if (!c_abi) { 781 __ push(rcx); 782 } 783 784 if (!c_abi) { 785 __ push(rdx); 786 __ push(rdi); 787 __ push(rsi); 788 __ push(r8); 789 __ push(r9); 790 __ push(r10); 791 __ push(r11); 792 __ push(r12); 793 __ push(r13); 794 __ push(r14); 795 __ push(r15); 796 } 797 __ save_vector_registers(); 798 __ movptr(rdi, rax); 799 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_JRT), rdi); 800 __ restore_vector_registers(); 801 if (!c_abi) { 802 __ pop(r15); 803 __ pop(r14); 804 __ pop(r13); 805 __ pop(r12); 806 __ pop(r11); 807 __ pop(r10); 808 __ pop(r9); 809 __ pop(r8); 810 __ pop(rsi); 811 __ pop(rdi); 812 __ pop(rdx); 813 814 __ pop(rcx); 815 __ pop(r8); 816 __ pop(rdi); 817 } 818 __ ret(0); 819 #else 820 ShouldNotReachHere(); 821 #endif 822 return start; 823 } 824 825 #undef __ 826 827 void ShenandoahBaseBarrierSetAssembler::barrier_stubs_init() { 828 if (ShenandoahWriteBarrier || ShenandoahStoreValEnqueueBarrier) { 829 int stub_code_size = 4096; 830 ResourceMark rm; 831 BufferBlob* bb = BufferBlob::create("shenandoah_barrier_stubs", stub_code_size); 832 CodeBuffer buf(bb); 833 StubCodeGenerator cgen(&buf); 834 _shenandoah_wb = generate_shenandoah_wb(&cgen, false, true); 835 _shenandoah_wb_C = generate_shenandoah_wb(&cgen, true, !ShenandoahWriteBarrierCsetTestInIR); 836 } 837 }