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/shenandoahBarrierSetAssembler.hpp" 27 #include "gc/shenandoah/shenandoahConnectionMatrix.hpp" 28 #include "gc/shenandoah/shenandoahHeap.hpp" 29 #include "gc/shenandoah/shenandoahHeapRegion.hpp" 30 #include "gc/shenandoah/shenandoahHeuristics.hpp" 31 #include "gc/shenandoah/shenandoahRuntime.hpp" 32 #include "gc/shenandoah/shenandoahThreadLocalData.hpp" 33 #include "interpreter/interpreter.hpp" 34 #include "interpreter/interp_masm.hpp" 35 #include "runtime/sharedRuntime.hpp" 36 #include "runtime/thread.hpp" 37 #include "utilities/macros.hpp" 38 #ifdef COMPILER1 39 #include "c1/c1_LIRAssembler.hpp" 40 #include "c1/c1_MacroAssembler.hpp" 41 #include "gc/shenandoah/c1/shenandoahBarrierSetC1.hpp" 42 #endif 43 44 #define __ masm-> 45 46 address ShenandoahBarrierSetAssembler::_shenandoah_wb = NULL; 47 address ShenandoahBarrierSetAssembler::_shenandoah_wb_C = NULL; 48 49 void ShenandoahBarrierSetAssembler::arraycopy_prologue(MacroAssembler* masm, DecoratorSet decorators, BasicType type, 50 Register src, Register dst, Register count) { 51 52 bool checkcast = (decorators & ARRAYCOPY_CHECKCAST) != 0; 53 bool disjoint = (decorators & ARRAYCOPY_DISJOINT) != 0; 54 bool obj_int = type == T_OBJECT LP64_ONLY(&& UseCompressedOops); 55 bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0; 56 57 if (type == T_OBJECT || type == T_ARRAY) { 58 #ifdef _LP64 59 if (!checkcast && !obj_int) { 60 // Save count for barrier 61 __ movptr(r11, count); 62 } else if (disjoint && obj_int) { 63 // Save dst in r11 in the disjoint case 64 __ movq(r11, dst); 65 } 66 #else 67 if (disjoint) { 68 __ mov(rdx, dst); // save 'to' 69 } 70 #endif 71 72 if (!dest_uninitialized && !ShenandoahHeap::heap()->heuristics()->can_do_traversal_gc()) { 73 Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread); 74 #ifndef _LP64 75 __ push(thread); 76 __ get_thread(thread); 77 #endif 78 79 Label filtered; 80 Address in_progress(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset())); 81 // Is marking active? 82 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) { 83 __ cmpl(in_progress, 0); 84 } else { 85 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption"); 86 __ cmpb(in_progress, 0); 87 } 88 89 NOT_LP64(__ pop(thread);) 90 91 __ jcc(Assembler::equal, filtered); 92 93 __ pusha(); // push registers 94 #ifdef _LP64 95 if (count == c_rarg0) { 96 if (dst == c_rarg1) { 97 // exactly backwards!! 98 __ xchgptr(c_rarg1, c_rarg0); 99 } else { 100 __ movptr(c_rarg1, count); 101 __ movptr(c_rarg0, dst); 102 } 103 } else { 104 __ movptr(c_rarg0, dst); 105 __ movptr(c_rarg1, count); 106 } 107 if (UseCompressedOops) { 108 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_narrow_oop_entry), 2); 109 } else { 110 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry), 2); 111 } 112 #else 113 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry), 114 dst, count); 115 #endif 116 __ popa(); 117 __ bind(filtered); 118 } 119 } 120 121 } 122 123 void ShenandoahBarrierSetAssembler::arraycopy_epilogue(MacroAssembler* masm, DecoratorSet decorators, BasicType type, 124 Register src, Register dst, Register count) { 125 bool checkcast = (decorators & ARRAYCOPY_CHECKCAST) != 0; 126 bool disjoint = (decorators & ARRAYCOPY_DISJOINT) != 0; 127 bool obj_int = type == T_OBJECT LP64_ONLY(&& UseCompressedOops); 128 Register tmp = rax; 129 130 if (type == T_OBJECT || type == T_ARRAY) { 131 #ifdef _LP64 132 if (!checkcast && !obj_int) { 133 // Save count for barrier 134 count = r11; 135 } else if (disjoint && obj_int) { 136 // Use the saved dst in the disjoint case 137 dst = r11; 138 } else if (checkcast) { 139 tmp = rscratch1; 140 } 141 #else 142 if (disjoint) { 143 __ mov(dst, rdx); // restore 'to' 144 } 145 #endif 146 147 __ pusha(); // push registers (overkill) 148 #ifdef _LP64 149 if (c_rarg0 == count) { // On win64 c_rarg0 == rcx 150 assert_different_registers(c_rarg1, dst); 151 __ mov(c_rarg1, count); 152 __ mov(c_rarg0, dst); 153 } else { 154 assert_different_registers(c_rarg0, count); 155 __ mov(c_rarg0, dst); 156 __ mov(c_rarg1, count); 157 } 158 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry), 2); 159 #else 160 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry), 161 dst, count); 162 #endif 163 __ popa(); 164 } 165 } 166 167 void ShenandoahBarrierSetAssembler::shenandoah_write_barrier_pre(MacroAssembler* masm, 168 Register obj, 169 Register pre_val, 170 Register thread, 171 Register tmp, 172 bool tosca_live, 173 bool expand_call) { 174 175 if (ShenandoahSATBBarrier) { 176 satb_write_barrier_pre(masm, obj, pre_val, thread, tmp, tosca_live, expand_call); 177 } 178 } 179 180 void ShenandoahBarrierSetAssembler::satb_write_barrier_pre(MacroAssembler* masm, 181 Register obj, 182 Register pre_val, 183 Register thread, 184 Register tmp, 185 bool tosca_live, 186 bool expand_call) { 187 // If expand_call is true then we expand the call_VM_leaf macro 188 // directly to skip generating the check by 189 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp. 190 191 #ifdef _LP64 192 assert(thread == r15_thread, "must be"); 193 #endif // _LP64 194 195 Label done; 196 Label runtime; 197 198 assert(pre_val != noreg, "check this code"); 199 200 if (obj != noreg) { 201 assert_different_registers(obj, pre_val, tmp); 202 assert(pre_val != rax, "check this code"); 203 } 204 205 Address in_progress(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset())); 206 Address index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset())); 207 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())); 208 209 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset())); 210 __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL); 211 __ jcc(Assembler::zero, done); 212 213 // Do we need to load the previous value? 214 if (obj != noreg) { 215 __ load_heap_oop(pre_val, Address(obj, 0), noreg, noreg, AS_RAW); 216 } 217 218 // Is the previous value null? 219 __ cmpptr(pre_val, (int32_t) NULL_WORD); 220 __ jcc(Assembler::equal, done); 221 222 // Can we store original value in the thread's buffer? 223 // Is index == 0? 224 // (The index field is typed as size_t.) 225 226 __ movptr(tmp, index); // tmp := *index_adr 227 __ cmpptr(tmp, 0); // tmp == 0? 228 __ jcc(Assembler::equal, runtime); // If yes, goto runtime 229 230 __ subptr(tmp, wordSize); // tmp := tmp - wordSize 231 __ movptr(index, tmp); // *index_adr := tmp 232 __ addptr(tmp, buffer); // tmp := tmp + *buffer_adr 233 234 // Record the previous value 235 __ movptr(Address(tmp, 0), pre_val); 236 __ jmp(done); 237 238 __ bind(runtime); 239 // save the live input values 240 if(tosca_live) __ push(rax); 241 242 if (obj != noreg && obj != rax) 243 __ push(obj); 244 245 if (pre_val != rax) 246 __ push(pre_val); 247 248 // Calling the runtime using the regular call_VM_leaf mechanism generates 249 // code (generated by InterpreterMacroAssember::call_VM_leaf_base) 250 // that checks that the *(ebp+frame::interpreter_frame_last_sp) == NULL. 251 // 252 // If we care generating the pre-barrier without a frame (e.g. in the 253 // intrinsified Reference.get() routine) then ebp might be pointing to 254 // the caller frame and so this check will most likely fail at runtime. 255 // 256 // Expanding the call directly bypasses the generation of the check. 257 // So when we do not have have a full interpreter frame on the stack 258 // expand_call should be passed true. 259 260 NOT_LP64( __ push(thread); ) 261 262 #ifdef _LP64 263 // We move pre_val into c_rarg0 early, in order to avoid smashing it, should 264 // pre_val be c_rarg1 (where the call prologue would copy thread argument). 265 // Note: this should not accidentally smash thread, because thread is always r15. 266 assert(thread != c_rarg0, "smashed arg"); 267 if (c_rarg0 != pre_val) { 268 __ mov(c_rarg0, pre_val); 269 } 270 #endif 271 272 if (expand_call) { 273 LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); ) 274 #ifdef _LP64 275 if (c_rarg1 != thread) { 276 __ mov(c_rarg1, thread); 277 } 278 // Already moved pre_val into c_rarg0 above 279 #else 280 __ push(thread); 281 __ push(pre_val); 282 #endif 283 __ MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), 2); 284 } else { 285 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), LP64_ONLY(c_rarg0) NOT_LP64(pre_val), thread); 286 } 287 288 NOT_LP64( __ pop(thread); ) 289 290 // save the live input values 291 if (pre_val != rax) 292 __ pop(pre_val); 293 294 if (obj != noreg && obj != rax) 295 __ pop(obj); 296 297 if(tosca_live) __ pop(rax); 298 299 __ bind(done); 300 } 301 302 void ShenandoahBarrierSetAssembler::shenandoah_write_barrier_post(MacroAssembler* masm, 303 Register store_addr, 304 Register new_val, 305 Register thread, 306 Register tmp, 307 Register tmp2) { 308 assert(UseShenandoahGC, "why else should we be here?"); 309 310 if (! UseShenandoahMatrix) { 311 // No need for that barrier if not using matrix. 312 return; 313 } 314 315 Label done; 316 __ testptr(new_val, new_val); 317 __ jcc(Assembler::zero, done); 318 ShenandoahConnectionMatrix* matrix = ShenandoahHeap::heap()->connection_matrix(); 319 address matrix_addr = matrix->matrix_addr(); 320 __ movptr(rscratch1, (intptr_t) ShenandoahHeap::heap()->base()); 321 // Compute to-region index 322 __ movptr(tmp, new_val); 323 __ subptr(tmp, rscratch1); 324 __ shrptr(tmp, ShenandoahHeapRegion::region_size_bytes_shift_jint()); 325 // Compute from-region index 326 __ movptr(tmp2, store_addr); 327 __ subptr(tmp2, rscratch1); 328 __ shrptr(tmp2, ShenandoahHeapRegion::region_size_bytes_shift_jint()); 329 // Compute matrix index 330 __ imulptr(tmp, tmp, matrix->stride_jint()); 331 __ addptr(tmp, tmp2); 332 // Address is _matrix[to * stride + from] 333 __ movptr(rscratch1, (intptr_t) matrix_addr); 334 // Test if the element is already set. 335 __ cmpb(Address(rscratch1, tmp, Address::times_1), 0); 336 __ jcc(Assembler::notEqual, done); 337 // Store true, if not yet set. 338 __ movb(Address(rscratch1, tmp, Address::times_1), 1); 339 __ bind(done); 340 } 341 342 void ShenandoahBarrierSetAssembler::read_barrier(MacroAssembler* masm, Register dst) { 343 if (ShenandoahReadBarrier) { 344 read_barrier_impl(masm, dst); 345 } 346 } 347 348 void ShenandoahBarrierSetAssembler::read_barrier_impl(MacroAssembler* masm, Register dst) { 349 assert(UseShenandoahGC && (ShenandoahReadBarrier || ShenandoahStoreValReadBarrier), "should be enabled"); 350 Label is_null; 351 __ testptr(dst, dst); 352 __ jcc(Assembler::zero, is_null); 353 read_barrier_not_null_impl(masm, dst); 354 __ bind(is_null); 355 } 356 357 void ShenandoahBarrierSetAssembler::read_barrier_not_null(MacroAssembler* masm, Register dst) { 358 if (ShenandoahReadBarrier) { 359 read_barrier_not_null_impl(masm, dst); 360 } 361 } 362 363 void ShenandoahBarrierSetAssembler::read_barrier_not_null_impl(MacroAssembler* masm, Register dst) { 364 assert(UseShenandoahGC && (ShenandoahReadBarrier || ShenandoahStoreValReadBarrier), "should be enabled"); 365 __ movptr(dst, Address(dst, BrooksPointer::byte_offset())); 366 } 367 368 369 void ShenandoahBarrierSetAssembler::write_barrier(MacroAssembler* masm, Register dst) { 370 if (ShenandoahWriteBarrier) { 371 write_barrier_impl(masm, dst); 372 } 373 } 374 375 void ShenandoahBarrierSetAssembler::write_barrier_impl(MacroAssembler* masm, Register dst) { 376 assert(UseShenandoahGC && (ShenandoahWriteBarrier || ShenandoahStoreValEnqueueBarrier), "Should be enabled"); 377 #ifdef _LP64 378 Label done; 379 380 Address gc_state(r15_thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset())); 381 __ testb(gc_state, ShenandoahHeap::HAS_FORWARDED | ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL); 382 __ jccb(Assembler::zero, done); 383 384 // Heap is unstable, need to perform the read-barrier even if WB is inactive 385 if (ShenandoahWriteBarrierRB) { 386 read_barrier_not_null(masm, dst); 387 } 388 389 __ testb(gc_state, ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL); 390 __ jccb(Assembler::zero, done); 391 392 if (dst != rax) { 393 __ xchgptr(dst, rax); // Move obj into rax and save rax into obj. 394 } 395 396 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, ShenandoahBarrierSetAssembler::shenandoah_wb()))); 397 398 if (dst != rax) { 399 __ xchgptr(rax, dst); // Swap back obj with rax. 400 } 401 402 __ bind(done); 403 #else 404 Unimplemented(); 405 #endif 406 } 407 408 void ShenandoahBarrierSetAssembler::storeval_barrier(MacroAssembler* masm, Register dst, Register tmp) { 409 if (ShenandoahStoreValReadBarrier || ShenandoahStoreValEnqueueBarrier) { 410 storeval_barrier_impl(masm, dst, tmp); 411 } 412 } 413 414 void ShenandoahBarrierSetAssembler::storeval_barrier_impl(MacroAssembler* masm, Register dst, Register tmp) { 415 assert(UseShenandoahGC && (ShenandoahStoreValReadBarrier || ShenandoahStoreValEnqueueBarrier), "should be enabled"); 416 417 if (dst == noreg) return; 418 419 #ifdef _LP64 420 if (ShenandoahStoreValEnqueueBarrier) { 421 Label is_null; 422 __ testptr(dst, dst); 423 __ jcc(Assembler::zero, is_null); 424 write_barrier_impl(masm, dst); 425 __ bind(is_null); 426 427 // The set of registers to be saved+restored is the same as in the write-barrier above. 428 // Those are the commonly used registers in the interpreter. 429 __ pusha(); 430 // __ push_callee_saved_registers(); 431 __ subptr(rsp, 2 * Interpreter::stackElementSize); 432 __ movdbl(Address(rsp, 0), xmm0); 433 434 satb_write_barrier_pre(masm, noreg, dst, r15_thread, tmp, true, false); 435 __ movdbl(xmm0, Address(rsp, 0)); 436 __ addptr(rsp, 2 * Interpreter::stackElementSize); 437 //__ pop_callee_saved_registers(); 438 __ popa(); 439 } 440 if (ShenandoahStoreValReadBarrier) { 441 read_barrier_impl(masm, dst); 442 } 443 #else 444 Unimplemented(); 445 #endif 446 } 447 448 void ShenandoahBarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type, 449 Register dst, Address src, Register tmp1, Register tmp_thread) { 450 bool on_oop = type == T_OBJECT || type == T_ARRAY; 451 bool in_heap = (decorators & IN_HEAP) != 0; 452 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0; 453 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0; 454 bool on_reference = on_weak || on_phantom; 455 if (in_heap) { 456 read_barrier_not_null(masm, src.base()); 457 } 458 BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp_thread); 459 if (ShenandoahKeepAliveBarrier && on_oop && on_reference) { 460 const Register thread = NOT_LP64(tmp_thread) LP64_ONLY(r15_thread); 461 NOT_LP64(__ get_thread(thread)); 462 463 // Generate the SATB pre-barrier code to log the value of 464 // the referent field in an SATB buffer. 465 shenandoah_write_barrier_pre(masm /* masm */, 466 noreg /* obj */, 467 dst /* pre_val */, 468 thread /* thread */, 469 tmp1 /* tmp */, 470 true /* tosca_live */, 471 true /* expand_call */); 472 } 473 } 474 475 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type, 476 Address dst, Register val, Register tmp1, Register tmp2) { 477 478 bool in_heap = (decorators & IN_HEAP) != 0; 479 bool as_normal = (decorators & AS_NORMAL) != 0; 480 if (in_heap) { 481 write_barrier(masm, dst.base()); 482 } 483 if (type == T_OBJECT || type == T_ARRAY) { 484 bool needs_pre_barrier = as_normal; 485 bool needs_post_barrier = val != noreg && in_heap && UseShenandoahMatrix; 486 487 Register tmp3 = LP64_ONLY(r8) NOT_LP64(rsi); 488 Register rthread = LP64_ONLY(r15_thread) NOT_LP64(rcx); 489 // flatten object address if needed 490 // We do it regardless of precise because we need the registers 491 if (dst.index() == noreg && dst.disp() == 0) { 492 if (dst.base() != tmp1) { 493 __ movptr(tmp1, dst.base()); 494 } 495 } else { 496 __ lea(tmp1, dst); 497 } 498 499 #ifndef _LP64 500 InterpreterMacroAssembler *imasm = static_cast<InterpreterMacroAssembler*>(masm); 501 #endif 502 503 NOT_LP64(__ get_thread(rcx)); 504 NOT_LP64(imasm->save_bcp()); 505 506 if (needs_pre_barrier) { 507 shenandoah_write_barrier_pre(masm /*masm*/, 508 tmp1 /* obj */, 509 tmp2 /* pre_val */, 510 rthread /* thread */, 511 tmp3 /* tmp */, 512 val != noreg /* tosca_live */, 513 false /* expand_call */); 514 } 515 if (val == noreg) { 516 BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg); 517 } else { 518 storeval_barrier(masm, val, tmp3); 519 Register new_val = val; 520 if (needs_post_barrier) { 521 if (UseCompressedOops) { 522 new_val = tmp2; 523 __ movptr(new_val, val); 524 } 525 } 526 BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg); 527 528 if (needs_post_barrier) { 529 shenandoah_write_barrier_post(masm /*masm*/, 530 tmp1 /* store_adr */, 531 new_val /* new_val */, 532 rthread /* thread */, 533 tmp3 /* tmp */, 534 tmp2 /* tmp2 */); 535 } 536 } 537 NOT_LP64(imasm->restore_bcp()); 538 539 } else { 540 BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2); 541 } 542 } 543 544 #ifndef _LP64 545 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm, 546 Address obj1, jobject obj2) { 547 Unimplemented(); 548 } 549 550 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm, 551 Register obj1, jobject obj2) { 552 Unimplemented(); 553 } 554 #endif 555 556 557 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm, Register op1, Register op2) { 558 __ cmpptr(op1, op2); 559 if (ShenandoahAcmpBarrier) { 560 Label done; 561 __ jccb(Assembler::equal, done); 562 read_barrier(masm, op1); 563 read_barrier(masm, op2); 564 __ cmpptr(op1, op2); 565 __ bind(done); 566 } 567 } 568 569 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm, Register src1, Address src2) { 570 __ cmpptr(src1, src2); 571 if (ShenandoahAcmpBarrier) { 572 Label done; 573 __ jccb(Assembler::equal, done); 574 __ movptr(rscratch2, src2); 575 read_barrier(masm, src1); 576 read_barrier(masm, rscratch2); 577 __ cmpptr(src1, rscratch2); 578 __ bind(done); 579 } 580 } 581 582 void ShenandoahBarrierSetAssembler::tlab_allocate(MacroAssembler* masm, 583 Register thread, Register obj, 584 Register var_size_in_bytes, 585 int con_size_in_bytes, 586 Register t1, Register t2, 587 Label& slow_case) { 588 assert_different_registers(obj, t1, t2); 589 assert_different_registers(obj, var_size_in_bytes, t1); 590 Register end = t2; 591 if (!thread->is_valid()) { 592 #ifdef _LP64 593 thread = r15_thread; 594 #else 595 assert(t1->is_valid(), "need temp reg"); 596 thread = t1; 597 __ get_thread(thread); 598 #endif 599 } 600 601 __ verify_tlab(); 602 603 __ movptr(obj, Address(thread, JavaThread::tlab_top_offset())); 604 if (var_size_in_bytes == noreg) { 605 __ lea(end, Address(obj, con_size_in_bytes + BrooksPointer::byte_size())); 606 } else { 607 __ addptr(var_size_in_bytes, BrooksPointer::byte_size()); 608 __ lea(end, Address(obj, var_size_in_bytes, Address::times_1)); 609 } 610 __ cmpptr(end, Address(thread, JavaThread::tlab_end_offset())); 611 __ jcc(Assembler::above, slow_case); 612 613 // update the tlab top pointer 614 __ movptr(Address(thread, JavaThread::tlab_top_offset()), end); 615 616 // Initialize brooks pointer 617 #ifdef _LP64 618 __ incrementq(obj, BrooksPointer::byte_size()); 619 #else 620 __ incrementl(obj, BrooksPointer::byte_size()); 621 #endif 622 __ movptr(Address(obj, BrooksPointer::byte_offset()), obj); 623 624 // recover var_size_in_bytes if necessary 625 if (var_size_in_bytes == end) { 626 __ subptr(var_size_in_bytes, obj); 627 } 628 __ verify_tlab(); 629 } 630 631 void ShenandoahBarrierSetAssembler::resolve_for_read(MacroAssembler* masm, DecoratorSet decorators, Register obj) { 632 bool oop_not_null = (decorators & IS_NOT_NULL) != 0; 633 if (oop_not_null) { 634 read_barrier_not_null(masm, obj); 635 } else { 636 read_barrier(masm, obj); 637 } 638 } 639 640 void ShenandoahBarrierSetAssembler::resolve_for_write(MacroAssembler* masm, DecoratorSet decorators, Register obj) { 641 bool oop_not_null = (decorators & IS_NOT_NULL) != 0; 642 if (oop_not_null) { 643 write_barrier(masm, obj); 644 } else { 645 Label done; 646 __ testptr(obj, obj); 647 __ jcc(Assembler::zero, done); 648 write_barrier(masm, obj); 649 __ bind(done); 650 } 651 } 652 653 // Special Shenandoah CAS implementation that handles false negatives 654 // due to concurrent evacuation. 655 #ifndef _LP64 656 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm, DecoratorSet decorators, 657 Register res, Address addr, Register oldval, Register newval, 658 bool exchange, bool encode, Register tmp1, Register tmp2) { 659 // Shenandoah has no 32-bit version for this. 660 Unimplemented(); 661 } 662 #else 663 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm, DecoratorSet decorators, 664 Register res, Address addr, Register oldval, Register newval, 665 bool exchange, bool encode, Register tmp1, Register tmp2) { 666 667 if (!ShenandoahCASBarrier) { 668 BarrierSetAssembler::cmpxchg_oop(masm, decorators, res, addr, oldval, newval, exchange, encode, tmp1, tmp2); 669 return; 670 } 671 672 assert(ShenandoahCASBarrier, "Should only be used when CAS barrier is enabled"); 673 assert(oldval == rax, "must be in rax for implicit use in cmpxchg"); 674 675 Label retry, done; 676 677 // Apply storeval barrier to newval. 678 if (encode) { 679 storeval_barrier(masm, newval, tmp1); 680 } 681 682 if (UseCompressedOops) { 683 if (encode) { 684 __ encode_heap_oop(oldval); 685 __ mov(rscratch1, newval); 686 __ encode_heap_oop(rscratch1); 687 newval = rscratch1; 688 } 689 } 690 691 // Remember oldval for retry logic below 692 if (UseCompressedOops) { 693 __ movl(tmp1, oldval); 694 } else { 695 __ movptr(tmp1, oldval); 696 } 697 698 // Step 1. Try to CAS with given arguments. If successful, then we are done, 699 // and can safely return. 700 if (os::is_MP()) __ lock(); 701 if (UseCompressedOops) { 702 __ cmpxchgl(newval, addr); 703 } else { 704 __ cmpxchgptr(newval, addr); 705 } 706 __ jcc(Assembler::equal, done, true); 707 708 // Step 2. CAS had failed. This may be a false negative. 709 // 710 // The trouble comes when we compare the to-space pointer with the from-space 711 // pointer to the same object. To resolve this, it will suffice to read both 712 // oldval and the value from memory through the read barriers -- this will give 713 // both to-space pointers. If they mismatch, then it was a legitimate failure. 714 // 715 if (UseCompressedOops) { 716 __ decode_heap_oop(tmp1); 717 } 718 __ resolve_for_read(0, tmp1); 719 720 if (UseCompressedOops) { 721 __ movl(tmp2, oldval); 722 __ decode_heap_oop(tmp2); 723 } else { 724 __ movptr(tmp2, oldval); 725 } 726 __ resolve_for_read(0, tmp2); 727 728 __ cmpptr(tmp1, tmp2); 729 __ jcc(Assembler::notEqual, done, true); 730 731 // Step 3. Try to CAS again with resolved to-space pointers. 732 // 733 // Corner case: it may happen that somebody stored the from-space pointer 734 // to memory while we were preparing for retry. Therefore, we can fail again 735 // on retry, and so need to do this in loop, always re-reading the failure 736 // witness through the read barrier. 737 __ bind(retry); 738 if (os::is_MP()) __ lock(); 739 if (UseCompressedOops) { 740 __ cmpxchgl(newval, addr); 741 } else { 742 __ cmpxchgptr(newval, addr); 743 } 744 __ jcc(Assembler::equal, done, true); 745 746 if (UseCompressedOops) { 747 __ movl(tmp2, oldval); 748 __ decode_heap_oop(tmp2); 749 } else { 750 __ movptr(tmp2, oldval); 751 } 752 __ resolve_for_read(0, tmp2); 753 754 __ cmpptr(tmp1, tmp2); 755 __ jcc(Assembler::equal, retry, true); 756 757 // Step 4. If we need a boolean result out of CAS, check the flag again, 758 // and promote the result. Note that we handle the flag from both the CAS 759 // itself and from the retry loop. 760 __ bind(done); 761 if (!exchange) { 762 assert(res != NULL, "need result register"); 763 __ setb(Assembler::equal, res); 764 __ movzbl(res, res); 765 } 766 } 767 #endif // LP64 768 769 void ShenandoahBarrierSetAssembler::xchg_oop(MacroAssembler* masm, DecoratorSet decorators, 770 Register obj, Address addr, Register tmp) { 771 storeval_barrier(masm, obj, tmp); 772 BarrierSetAssembler::xchg_oop(masm, decorators, obj, addr, tmp); 773 } 774 775 #ifdef COMPILER1 776 777 #undef __ 778 #define __ ce->masm()-> 779 780 void ShenandoahBarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, ShenandoahPreBarrierStub* stub) { 781 ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1(); 782 // At this point we know that marking is in progress. 783 // If do_load() is true then we have to emit the 784 // load of the previous value; otherwise it has already 785 // been loaded into _pre_val. 786 787 __ bind(*stub->entry()); 788 assert(stub->pre_val()->is_register(), "Precondition."); 789 790 Register pre_val_reg = stub->pre_val()->as_register(); 791 792 if (stub->do_load()) { 793 ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/, false /*unaligned*/); 794 } 795 796 __ cmpptr(pre_val_reg, (int32_t)NULL_WORD); 797 __ jcc(Assembler::equal, *stub->continuation()); 798 ce->store_parameter(stub->pre_val()->as_register(), 0); 799 __ call(RuntimeAddress(bs->pre_barrier_c1_runtime_code_blob()->code_begin())); 800 __ jmp(*stub->continuation()); 801 802 } 803 804 void ShenandoahBarrierSetAssembler::gen_write_barrier_stub(LIR_Assembler* ce, ShenandoahWriteBarrierStub* stub) { 805 __ bind(*stub->entry()); 806 807 Label done; 808 Register obj = stub->obj()->as_register(); 809 Register res = stub->result()->as_register(); 810 811 if (res != obj) { 812 __ mov(res, obj); 813 } 814 815 // Check for null. 816 if (stub->needs_null_check()) { 817 __ testptr(res, res); 818 __ jcc(Assembler::zero, done); 819 } 820 821 write_barrier(ce->masm(), res); 822 823 __ bind(done); 824 __ jmp(*stub->continuation()); 825 } 826 827 #undef __ 828 829 #define __ sasm-> 830 831 void ShenandoahBarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) { 832 __ prologue("shenandoah_pre_barrier", false); 833 // arg0 : previous value of memory 834 835 __ push(rax); 836 __ push(rdx); 837 838 const Register pre_val = rax; 839 const Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread); 840 const Register tmp = rdx; 841 842 NOT_LP64(__ get_thread(thread);) 843 844 Address queue_index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset())); 845 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())); 846 847 Label done; 848 Label runtime; 849 850 // Is SATB still active? 851 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset())); 852 __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL); 853 __ jcc(Assembler::zero, done); 854 855 // Can we store original value in the thread's buffer? 856 857 __ movptr(tmp, queue_index); 858 __ testptr(tmp, tmp); 859 __ jcc(Assembler::zero, runtime); 860 __ subptr(tmp, wordSize); 861 __ movptr(queue_index, tmp); 862 __ addptr(tmp, buffer); 863 864 // prev_val (rax) 865 __ load_parameter(0, pre_val); 866 __ movptr(Address(tmp, 0), pre_val); 867 __ jmp(done); 868 869 __ bind(runtime); 870 871 __ save_live_registers_no_oop_map(true); 872 873 // load the pre-value 874 __ load_parameter(0, rcx); 875 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), rcx, thread); 876 877 __ restore_live_registers(true); 878 879 __ bind(done); 880 881 __ pop(rdx); 882 __ pop(rax); 883 884 __ epilogue(); 885 } 886 887 #undef __ 888 889 #endif // COMPILER1 890 891 address ShenandoahBarrierSetAssembler::shenandoah_wb() { 892 assert(_shenandoah_wb != NULL, "need write barrier stub"); 893 return _shenandoah_wb; 894 } 895 896 address ShenandoahBarrierSetAssembler::shenandoah_wb_C() { 897 assert(_shenandoah_wb_C != NULL, "need write barrier stub"); 898 return _shenandoah_wb_C; 899 } 900 901 #define __ cgen->assembler()-> 902 903 address ShenandoahBarrierSetAssembler::generate_shenandoah_wb(StubCodeGenerator* cgen, bool c_abi, bool do_cset_test) { 904 __ align(CodeEntryAlignment); 905 StubCodeMark mark(cgen, "StubRoutines", "shenandoah_wb"); 906 address start = __ pc(); 907 908 #ifdef _LP64 909 Label not_done; 910 911 // We use RDI, which also serves as argument register for slow call. 912 // RAX always holds the src object ptr, except after the slow call and 913 // the cmpxchg, then it holds the result. 914 // R8 and RCX are used as temporary registers. 915 if (!c_abi) { 916 __ push(rdi); 917 __ push(r8); 918 } 919 920 // Check for object beeing in the collection set. 921 // TODO: Can we use only 1 register here? 922 // The source object arrives here in rax. 923 // live: rax 924 // live: rdi 925 if (!c_abi) { 926 __ mov(rdi, rax); 927 } else { 928 if (rax != c_rarg0) { 929 __ mov(rax, c_rarg0); 930 } 931 } 932 if (do_cset_test) { 933 __ shrptr(rdi, ShenandoahHeapRegion::region_size_bytes_shift_jint()); 934 // live: r8 935 __ movptr(r8, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr()); 936 __ movbool(r8, Address(r8, rdi, Address::times_1)); 937 // unlive: rdi 938 __ testbool(r8); 939 // unlive: r8 940 __ jccb(Assembler::notZero, not_done); 941 942 if (!c_abi) { 943 __ pop(r8); 944 __ pop(rdi); 945 } 946 __ ret(0); 947 948 __ bind(not_done); 949 } 950 951 if (!c_abi) { 952 __ push(rcx); 953 } 954 955 if (!c_abi) { 956 __ push(rdx); 957 __ push(rdi); 958 __ push(rsi); 959 __ push(r8); 960 __ push(r9); 961 __ push(r10); 962 __ push(r11); 963 __ push(r12); 964 __ push(r13); 965 __ push(r14); 966 __ push(r15); 967 } 968 __ save_vector_registers(); 969 __ movptr(rdi, rax); 970 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_JRT), rdi); 971 __ restore_vector_registers(); 972 if (!c_abi) { 973 __ pop(r15); 974 __ pop(r14); 975 __ pop(r13); 976 __ pop(r12); 977 __ pop(r11); 978 __ pop(r10); 979 __ pop(r9); 980 __ pop(r8); 981 __ pop(rsi); 982 __ pop(rdi); 983 __ pop(rdx); 984 985 __ pop(rcx); 986 __ pop(r8); 987 __ pop(rdi); 988 } 989 __ ret(0); 990 #else 991 ShouldNotReachHere(); 992 #endif 993 return start; 994 } 995 996 #undef __ 997 998 void ShenandoahBarrierSetAssembler::barrier_stubs_init() { 999 if (ShenandoahWriteBarrier || ShenandoahStoreValEnqueueBarrier) { 1000 int stub_code_size = 4096; 1001 ResourceMark rm; 1002 BufferBlob* bb = BufferBlob::create("shenandoah_barrier_stubs", stub_code_size); 1003 CodeBuffer buf(bb); 1004 StubCodeGenerator cgen(&buf); 1005 _shenandoah_wb = generate_shenandoah_wb(&cgen, false, true); 1006 _shenandoah_wb_C = generate_shenandoah_wb(&cgen, true, !ShenandoahWriteBarrierCsetTestInIR); 1007 } 1008 }