1 /* 2 * Copyright (c) 2015, 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/shenandoahHeap.hpp" 27 #include "gc/shenandoah/shenandoahHeapRegion.hpp" 28 #include "gc/shenandoah/shenandoahRuntime.hpp" 29 #include "opto/arraycopynode.hpp" 30 #include "opto/block.hpp" 31 #include "opto/callnode.hpp" 32 #include "opto/castnode.hpp" 33 #include "opto/movenode.hpp" 34 #include "opto/phaseX.hpp" 35 #include "opto/rootnode.hpp" 36 #include "opto/runtime.hpp" 37 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp" 38 #include "gc/shenandoah/c2/shenandoahSupport.hpp" 39 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp" 40 #include "opto/subnode.hpp" 41 42 Node* ShenandoahBarrierNode::skip_through_barrier(Node* n) { 43 if (n == NULL) { 44 return NULL; 45 } 46 if (n->Opcode() == Op_ShenandoahEnqueueBarrier) { 47 n = n->in(1); 48 } 49 50 if (n->is_ShenandoahBarrier()) { 51 return n->in(ValueIn); 52 } else if (n->is_Phi() && 53 n->req() == 3 && 54 n->in(1) != NULL && 55 n->in(1)->is_ShenandoahBarrier() && 56 n->in(2) != NULL && 57 n->in(2)->bottom_type() == TypePtr::NULL_PTR && 58 n->in(0) != NULL && 59 n->in(0)->in(1) != NULL && 60 n->in(0)->in(1)->is_IfProj() && 61 n->in(0)->in(2) != NULL && 62 n->in(0)->in(2)->is_IfProj() && 63 n->in(0)->in(1)->in(0) != NULL && 64 n->in(0)->in(1)->in(0) == n->in(0)->in(2)->in(0) && 65 n->in(1)->in(ValueIn)->Opcode() == Op_CastPP) { 66 Node* iff = n->in(0)->in(1)->in(0); 67 Node* res = n->in(1)->in(ValueIn)->in(1); 68 if (iff->is_If() && 69 iff->in(1) != NULL && 70 iff->in(1)->is_Bool() && 71 iff->in(1)->as_Bool()->_test._test == BoolTest::ne && 72 iff->in(1)->in(1) != NULL && 73 iff->in(1)->in(1)->Opcode() == Op_CmpP && 74 iff->in(1)->in(1)->in(1) != NULL && 75 iff->in(1)->in(1)->in(1) == res && 76 iff->in(1)->in(1)->in(2) != NULL && 77 iff->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) { 78 return res; 79 } 80 } 81 return n; 82 } 83 84 bool ShenandoahBarrierNode::needs_barrier(PhaseGVN* phase, ShenandoahBarrierNode* orig, Node* n, Node* rb_mem, bool allow_fromspace) { 85 Unique_Node_List visited; 86 return needs_barrier_impl(phase, orig, n, rb_mem, allow_fromspace, visited); 87 } 88 89 bool ShenandoahBarrierNode::needs_barrier_impl(PhaseGVN* phase, ShenandoahBarrierNode* orig, Node* n, Node* rb_mem, bool allow_fromspace, Unique_Node_List &visited) { 90 91 if (visited.member(n)) { 92 return false; // Been there. 93 } 94 visited.push(n); 95 96 if (n->is_Allocate()) { 97 // tty->print_cr("killed barrier for newly allocated object"); 98 return false; 99 } 100 101 if (n->is_CallJava() || n->Opcode() == Op_CallLeafNoFP) { 102 return true; 103 } 104 105 const Type* type = phase->type(n); 106 if (type == Type::TOP) { 107 return false; 108 } 109 if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) { 110 // tty->print_cr("killed barrier for NULL object"); 111 return false; 112 } 113 if (type->make_oopptr() && type->make_oopptr()->const_oop() != NULL) { 114 // tty->print_cr("killed barrier for constant object"); 115 return false; 116 } 117 118 if (ShenandoahOptimizeStableFinals) { 119 const TypeAryPtr* ary = type->isa_aryptr(); 120 if (ary && ary->is_stable() && allow_fromspace) { 121 return false; 122 } 123 } 124 125 if (n->is_CheckCastPP() || n->is_ConstraintCast() || n->Opcode() == Op_ShenandoahEnqueueBarrier) { 126 return needs_barrier_impl(phase, orig, n->in(1), rb_mem, allow_fromspace, visited); 127 } 128 if (n->is_Parm()) { 129 return true; 130 } 131 if (n->is_Proj()) { 132 return needs_barrier_impl(phase, orig, n->in(0), rb_mem, allow_fromspace, visited); 133 } 134 if (n->is_Phi()) { 135 bool need_barrier = false; 136 for (uint i = 1; i < n->req() && ! need_barrier; i++) { 137 Node* input = n->in(i); 138 if (input == NULL) { 139 need_barrier = true; // Phi not complete yet? 140 } else if (needs_barrier_impl(phase, orig, input, rb_mem, allow_fromspace, visited)) { 141 need_barrier = true; 142 } 143 } 144 return need_barrier; 145 } 146 if (n->is_CMove()) { 147 return needs_barrier_impl(phase, orig, n->in(CMoveNode::IfFalse), rb_mem, allow_fromspace, visited) || 148 needs_barrier_impl(phase, orig, n->in(CMoveNode::IfTrue ), rb_mem, allow_fromspace, visited); 149 } 150 if (n->Opcode() == Op_CreateEx) { 151 return true; 152 } 153 if (n->Opcode() == Op_ShenandoahWriteBarrier) { 154 // tty->print_cr("skipped barrier for chained write barrier object"); 155 return false; 156 } 157 if (n->Opcode() == Op_ShenandoahReadBarrier) { 158 if (rb_mem == n->in(Memory)) { 159 // tty->print_cr("Eliminated chained read barrier"); 160 return false; 161 } else { 162 return true; 163 } 164 } 165 166 if (n->Opcode() == Op_LoadP || 167 n->Opcode() == Op_LoadN || 168 n->Opcode() == Op_GetAndSetP || 169 n->Opcode() == Op_CompareAndExchangeP || 170 n->Opcode() == Op_ShenandoahCompareAndExchangeP || 171 n->Opcode() == Op_GetAndSetN || 172 n->Opcode() == Op_CompareAndExchangeN || 173 n->Opcode() == Op_ShenandoahCompareAndExchangeN) { 174 return true; 175 } 176 if (n->Opcode() == Op_DecodeN || 177 n->Opcode() == Op_EncodeP) { 178 return needs_barrier_impl(phase, orig, n->in(1), rb_mem, allow_fromspace, visited); 179 } 180 181 #ifdef ASSERT 182 tty->print("need barrier on?: "); n->dump(); 183 ShouldNotReachHere(); 184 #endif 185 return true; 186 } 187 188 bool ShenandoahReadBarrierNode::dominates_memory_rb_impl(PhaseGVN* phase, 189 Node* b1, 190 Node* b2, 191 Node* current, 192 bool linear) { 193 ResourceMark rm; 194 VectorSet visited(Thread::current()->resource_area()); 195 Node_Stack phis(0); 196 197 for(int i = 0; i < 10; i++) { 198 if (current == NULL) { 199 return false; 200 } else if (visited.test_set(current->_idx) || current->is_top() || current == b1) { 201 current = NULL; 202 while (phis.is_nonempty() && current == NULL) { 203 uint idx = phis.index(); 204 Node* phi = phis.node(); 205 if (idx >= phi->req()) { 206 phis.pop(); 207 } else { 208 current = phi->in(idx); 209 phis.set_index(idx+1); 210 } 211 } 212 if (current == NULL) { 213 return true; 214 } 215 } else if (current == phase->C->immutable_memory()) { 216 return false; 217 } else if (current->isa_Phi()) { 218 if (!linear) { 219 return false; 220 } 221 phis.push(current, 2); 222 current = current->in(1); 223 } else if (current->Opcode() == Op_ShenandoahWriteBarrier) { 224 const Type* in_type = current->bottom_type(); 225 const Type* this_type = b2->bottom_type(); 226 if (is_independent(in_type, this_type)) { 227 current = current->in(Memory); 228 } else { 229 return false; 230 } 231 } else if (current->Opcode() == Op_ShenandoahWBMemProj) { 232 current = current->in(0); 233 } else if (current->is_Proj()) { 234 current = current->in(0); 235 } else if (current->is_Call()) { 236 return false; // TODO: Maybe improve by looking at the call's memory effects? 237 } else if (current->is_MemBar()) { 238 return false; // TODO: Do we need to stop at *any* membar? 239 } else if (current->is_MergeMem()) { 240 // if (true) return false; 241 // tty->print_cr("current == mergemem: "); current->dump(); 242 const TypePtr* adr_type = brooks_pointer_type(phase->type(b2)); 243 uint alias_idx = phase->C->get_alias_index(adr_type); 244 current = current->as_MergeMem()->memory_at(alias_idx); 245 } else { 246 // tty->print_cr("what else can we see here:"); 247 #ifdef ASSERT 248 current->dump(); 249 #endif 250 ShouldNotReachHere(); 251 return false; 252 } 253 } 254 return false; 255 } 256 257 bool ShenandoahReadBarrierNode::is_independent(Node* mem) { 258 if (mem->is_Phi() || mem->is_Proj() || mem->is_MergeMem()) { 259 return true; 260 } else if (mem->Opcode() == Op_ShenandoahWriteBarrier) { 261 const Type* mem_type = mem->bottom_type(); 262 const Type* this_type = bottom_type(); 263 if (is_independent(mem_type, this_type)) { 264 return true; 265 } else { 266 return false; 267 } 268 } else if (mem->is_Call() || mem->is_MemBar()) { 269 return false; 270 } 271 #ifdef ASSERT 272 mem->dump(); 273 #endif 274 ShouldNotReachHere(); 275 return true; 276 } 277 278 279 bool ShenandoahReadBarrierNode::dominates_memory_rb(PhaseGVN* phase, Node* b1, Node* b2, bool linear) { 280 return dominates_memory_rb_impl(phase, b1->in(Memory), b2, b2->in(Memory), linear); 281 } 282 283 bool ShenandoahReadBarrierNode::is_independent(const Type* in_type, const Type* this_type) { 284 assert(in_type->isa_oopptr(), "expect oop ptr"); 285 assert(this_type->isa_oopptr(), "expect oop ptr"); 286 /* 287 if ((! in_type->isa_oopptr()) || (! this_type->isa_oopptr())) { 288 #ifdef ASSERT 289 tty->print_cr("not oopptr"); 290 tty->print("in: "); in_type->dump(); tty->print_cr(" "); 291 tty->print("this: "); this_type->dump(); tty->print_cr(" "); 292 #endif 293 return false; 294 } 295 */ 296 297 ciKlass* in_kls = in_type->is_oopptr()->klass(); 298 ciKlass* this_kls = this_type->is_oopptr()->klass(); 299 if (in_kls != NULL && this_kls != NULL && 300 in_kls->is_loaded() && this_kls->is_loaded() && 301 (!in_kls->is_subclass_of(this_kls)) && 302 (!this_kls->is_subclass_of(in_kls))) { 303 #ifdef ASSERT 304 // tty->print_cr("independent: "); 305 // tty->print("in: "); in_kls->print(); tty->print_cr(" "); 306 // tty->print("this: "); this_kls->print(); tty->print_cr(" "); 307 #endif 308 return true; 309 } 310 #ifdef ASSERT 311 // tty->print_cr("possibly dependend?"); 312 // tty->print("in: "); in_type->dump(); tty->print_cr(" "); 313 // tty->print("this: "); this_type->dump(); tty->print_cr(" "); 314 #endif 315 return false; 316 } 317 318 Node* ShenandoahReadBarrierNode::Ideal(PhaseGVN *phase, bool can_reshape) { 319 320 if (! can_reshape) { 321 return NULL; 322 } 323 324 if (in(Memory) == phase->C->immutable_memory()) return NULL; 325 326 // If memory input is a MergeMem, take the appropriate slice out of it. 327 Node* mem_in = in(Memory); 328 if (mem_in->isa_MergeMem()) { 329 const TypePtr* adr_type = brooks_pointer_type(bottom_type()); 330 uint alias_idx = phase->C->get_alias_index(adr_type); 331 mem_in = mem_in->as_MergeMem()->memory_at(alias_idx); 332 set_req(Memory, mem_in); 333 return this; 334 } 335 336 Node* input = in(Memory); 337 if (input->Opcode() == Op_ShenandoahWBMemProj) { 338 ResourceMark rm; 339 VectorSet seen(Thread::current()->resource_area()); 340 Node* n = in(Memory); 341 while (n->Opcode() == Op_ShenandoahWBMemProj && 342 n->in(0) != NULL && 343 n->in(0)->Opcode() == Op_ShenandoahWriteBarrier && 344 n->in(0)->in(Memory) != NULL) { 345 if (seen.test_set(n->_idx)) { 346 return NULL; // loop 347 } 348 n = n->in(0)->in(Memory); 349 } 350 351 Node* wb = input->in(0); 352 const Type* in_type = phase->type(wb); 353 // is_top() test not sufficient here: we can come here after CCP 354 // in a dead branch of the graph that has not yet been removed. 355 if (in_type == Type::TOP) return NULL; // Dead path. 356 assert(wb->Opcode() == Op_ShenandoahWriteBarrier, "expect write barrier"); 357 if (is_independent(in_type, _type)) { 358 phase->igvn_rehash_node_delayed(wb); 359 set_req(Memory, wb->in(Memory)); 360 if (can_reshape && input->outcnt() == 0) { 361 phase->is_IterGVN()->_worklist.push(input); 362 } 363 return this; 364 } 365 } 366 return NULL; 367 } 368 369 ShenandoahWriteBarrierNode::ShenandoahWriteBarrierNode(Compile* C, Node* ctrl, Node* mem, Node* obj) 370 : ShenandoahBarrierNode(ctrl, mem, obj, false) { 371 assert(UseShenandoahGC && ShenandoahWriteBarrier, "should be enabled"); 372 ShenandoahBarrierSetC2::bsc2()->state()->add_shenandoah_barrier(this); 373 } 374 375 376 Node* ShenandoahWriteBarrierNode::Identity(PhaseGVN* phase) { 377 assert(in(0) != NULL, "should have control"); 378 PhaseIterGVN* igvn = phase->is_IterGVN(); 379 Node* mem_in = in(Memory); 380 Node* mem_proj = NULL; 381 382 if (igvn != NULL) { 383 mem_proj = find_out_with(Op_ShenandoahWBMemProj); 384 if (mem_proj == NULL || mem_in == mem_proj) { 385 return this; 386 } 387 } 388 389 Node* replacement = Identity_impl(phase); 390 if (igvn != NULL) { 391 if (replacement != NULL && replacement != this) { 392 igvn->replace_node(mem_proj, mem_in); 393 } 394 } 395 return replacement; 396 } 397 398 399 Node* ShenandoahWriteBarrierNode::Ideal(PhaseGVN *phase, bool can_reshape) { 400 assert(in(0) != NULL, "should have control"); 401 if (!can_reshape) { 402 return NULL; 403 } 404 405 Node* mem_in = in(Memory); 406 407 if (mem_in->isa_MergeMem()) { 408 const TypePtr* adr_type = brooks_pointer_type(bottom_type()); 409 uint alias_idx = phase->C->get_alias_index(adr_type); 410 mem_in = mem_in->as_MergeMem()->memory_at(alias_idx); 411 set_req(Memory, mem_in); 412 return this; 413 } 414 415 Node* val = in(ValueIn); 416 if (val->is_ShenandoahBarrier()) { 417 set_req(ValueIn, val->in(ValueIn)); 418 return this; 419 } 420 421 return NULL; 422 } 423 424 bool ShenandoahWriteBarrierNode::expand(Compile* C, PhaseIterGVN& igvn, int& loop_opts_cnt) { 425 if (UseShenandoahGC) { 426 if (ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count() > 0 || (!ShenandoahWriteBarrier && ShenandoahStoreValEnqueueBarrier)) { 427 bool attempt_more_loopopts = ShenandoahLoopOptsAfterExpansion; 428 C->clear_major_progress(); 429 PhaseIdealLoop ideal_loop(igvn, LoopOptsShenandoahExpand); 430 if (C->failing()) return false; 431 PhaseIdealLoop::verify(igvn); 432 DEBUG_ONLY(ShenandoahBarrierNode::verify_raw_mem(C->root());) 433 if (attempt_more_loopopts) { 434 C->set_major_progress(); 435 if (!C->optimize_loops(loop_opts_cnt, igvn, LoopOptsShenandoahPostExpand)) { 436 return false; 437 } 438 C->clear_major_progress(); 439 } 440 } 441 } 442 return true; 443 } 444 445 bool ShenandoahWriteBarrierNode::is_heap_state_test(Node* iff, int mask) { 446 if (!UseShenandoahGC) { 447 return false; 448 } 449 assert(iff->is_If(), "bad input"); 450 if (iff->Opcode() != Op_If) { 451 return false; 452 } 453 Node* bol = iff->in(1); 454 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) { 455 return false; 456 } 457 Node* cmp = bol->in(1); 458 if (cmp->Opcode() != Op_CmpI) { 459 return false; 460 } 461 Node* in1 = cmp->in(1); 462 Node* in2 = cmp->in(2); 463 if (in2->find_int_con(-1) != 0) { 464 return false; 465 } 466 if (in1->Opcode() != Op_AndI) { 467 return false; 468 } 469 in2 = in1->in(2); 470 if (in2->find_int_con(-1) != mask) { 471 return false; 472 } 473 in1 = in1->in(1); 474 475 return is_gc_state_load(in1); 476 } 477 478 479 bool ShenandoahWriteBarrierNode::is_evacuation_in_progress_test(Node* iff) { 480 return is_heap_state_test(iff, ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL); 481 } 482 483 bool ShenandoahWriteBarrierNode::is_heap_stable_test(Node* iff) { 484 return is_heap_state_test(iff, ShenandoahHeap::HAS_FORWARDED); 485 } 486 487 bool ShenandoahWriteBarrierNode::is_gc_state_load(Node *n) { 488 if (!UseShenandoahGC) { 489 return false; 490 } 491 if (n->Opcode() != Op_LoadB) { 492 return false; 493 } 494 Node* addp = n->in(MemNode::Address); 495 if (!addp->is_AddP()) { 496 return false; 497 } 498 Node* base = addp->in(AddPNode::Address); 499 Node* off = addp->in(AddPNode::Offset); 500 if (base->Opcode() != Op_ThreadLocal) { 501 return false; 502 } 503 if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) { 504 return false; 505 } 506 return true; 507 } 508 509 bool ShenandoahWriteBarrierNode::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) { 510 assert(phase->is_dominator(stop, start), "bad inputs"); 511 ResourceMark rm; 512 Unique_Node_List wq; 513 wq.push(start); 514 for (uint next = 0; next < wq.size(); next++) { 515 Node *m = wq.at(next); 516 if (m == stop) { 517 continue; 518 } 519 if (m->is_SafePoint() && !m->is_CallLeaf()) { 520 return true; 521 } 522 if (m->is_Region()) { 523 for (uint i = 1; i < m->req(); i++) { 524 wq.push(m->in(i)); 525 } 526 } else { 527 wq.push(m->in(0)); 528 } 529 } 530 return false; 531 } 532 533 bool ShenandoahWriteBarrierNode::try_common_gc_state_load(Node *n, PhaseIdealLoop *phase) { 534 assert(is_gc_state_load(n), "inconsistent"); 535 Node* addp = n->in(MemNode::Address); 536 Node* dominator = NULL; 537 for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) { 538 Node* u = addp->fast_out(i); 539 assert(is_gc_state_load(u), "inconsistent"); 540 if (u != n && phase->is_dominator(u->in(0), n->in(0))) { 541 if (dominator == NULL) { 542 dominator = u; 543 } else { 544 if (phase->dom_depth(u->in(0)) < phase->dom_depth(dominator->in(0))) { 545 dominator = u; 546 } 547 } 548 } 549 } 550 if (dominator == NULL || has_safepoint_between(n->in(0), dominator->in(0), phase)) { 551 return false; 552 } 553 phase->igvn().replace_node(n, dominator); 554 555 return true; 556 } 557 558 Node* ShenandoahWriteBarrierNode::evacuation_in_progress_test_ctrl(Node* iff) { 559 assert(is_evacuation_in_progress_test(iff), "bad input"); 560 return iff->in(0); 561 } 562 563 bool ShenandoahBarrierNode::dominates_memory_impl(PhaseGVN* phase, 564 Node* b1, 565 Node* b2, 566 Node* current, 567 bool linear) { 568 ResourceMark rm; 569 VectorSet visited(Thread::current()->resource_area()); 570 Node_Stack phis(0); 571 572 573 for(int i = 0; i < 10; i++) { 574 if (current == NULL) { 575 return false; 576 } else if (visited.test_set(current->_idx) || current->is_top() || current == b1) { 577 current = NULL; 578 while (phis.is_nonempty() && current == NULL) { 579 uint idx = phis.index(); 580 Node* phi = phis.node(); 581 if (idx >= phi->req()) { 582 phis.pop(); 583 } else { 584 current = phi->in(idx); 585 phis.set_index(idx+1); 586 } 587 } 588 if (current == NULL) { 589 return true; 590 } 591 } else if (current == b2) { 592 return false; 593 } else if (current == phase->C->immutable_memory()) { 594 return false; 595 } else if (current->isa_Phi()) { 596 if (!linear) { 597 return false; 598 } 599 phis.push(current, 2); 600 current = current->in(1); 601 } else if (current->Opcode() == Op_ShenandoahWriteBarrier) { 602 current = current->in(Memory); 603 } else if (current->Opcode() == Op_ShenandoahWBMemProj) { 604 current = current->in(0); 605 } else if (current->is_Proj()) { 606 current = current->in(0); 607 } else if (current->is_Call()) { 608 current = current->in(TypeFunc::Memory); 609 } else if (current->is_MemBar()) { 610 current = current->in(TypeFunc::Memory); 611 } else if (current->is_MergeMem()) { 612 const TypePtr* adr_type = brooks_pointer_type(phase->type(b2)); 613 uint alias_idx = phase->C->get_alias_index(adr_type); 614 current = current->as_MergeMem()->memory_at(alias_idx); 615 } else { 616 #ifdef ASSERT 617 current->dump(); 618 #endif 619 ShouldNotReachHere(); 620 return false; 621 } 622 } 623 return false; 624 } 625 626 /** 627 * Determines if b1 dominates b2 through memory inputs. It returns true if: 628 * - b1 can be reached by following each branch in b2's memory input (through phis, etc) 629 * - or we get back to b2 (i.e. through a loop) without seeing b1 630 * In all other cases, (in particular, if we reach immutable_memory without having seen b1) 631 * we return false. 632 */ 633 bool ShenandoahBarrierNode::dominates_memory(PhaseGVN* phase, Node* b1, Node* b2, bool linear) { 634 return dominates_memory_impl(phase, b1, b2, b2->in(Memory), linear); 635 } 636 637 Node* ShenandoahBarrierNode::Identity_impl(PhaseGVN* phase) { 638 Node* n = in(ValueIn); 639 640 Node* rb_mem = Opcode() == Op_ShenandoahReadBarrier ? in(Memory) : NULL; 641 if (! needs_barrier(phase, this, n, rb_mem, _allow_fromspace)) { 642 return n; 643 } 644 645 // tty->print_cr("find sibling for: "); dump(2); 646 // Try to find a write barrier sibling with identical inputs that we can fold into. 647 for (DUIterator i = n->outs(); n->has_out(i); i++) { 648 Node* sibling = n->out(i); 649 if (sibling == this) { 650 continue; 651 } 652 /* 653 assert(sibling->Opcode() != Op_ShenandoahWriteBarrier || 654 Opcode() != Op_ShenandoahWriteBarrier || hash() == sibling->hash(), 655 "if this is a write barrier, then sibling can't be write barrier too"); 656 */ 657 if (sibling->Opcode() != Op_ShenandoahWriteBarrier) { 658 continue; 659 } 660 /* 661 if (sibling->outcnt() == 0) { 662 // Some dead node. 663 continue; 664 } 665 */ 666 assert(sibling->in(ValueIn) == in(ValueIn), "sanity"); 667 assert(sibling->Opcode() == Op_ShenandoahWriteBarrier, "sanity"); 668 // tty->print_cr("candidate: "); sibling->dump(); 669 670 if (dominates_memory(phase, sibling, this, phase->is_IterGVN() == NULL)) { 671 672 /* 673 tty->print_cr("matched barrier:"); 674 sibling->dump(); 675 tty->print_cr("for: "); 676 dump(); 677 */ 678 return sibling; 679 } 680 681 /* 682 tty->print_cr("couldn't match candidate:"); 683 sibling->dump(2); 684 */ 685 } 686 /* 687 tty->print_cr("couldn't match barrier to any:"); 688 dump(); 689 */ 690 return this; 691 } 692 693 #ifndef PRODUCT 694 void ShenandoahBarrierNode::dump_spec(outputStream *st) const { 695 const TypePtr* adr = adr_type(); 696 if (adr == NULL) { 697 return; 698 } 699 st->print(" @"); 700 adr->dump_on(st); 701 st->print(" ("); 702 Compile::current()->alias_type(adr)->adr_type()->dump_on(st); 703 st->print(") "); 704 } 705 #endif 706 707 Node* ShenandoahReadBarrierNode::Identity(PhaseGVN* phase) { 708 709 // if (true) return this; 710 711 // tty->print("optimizing rb: "); dump(); 712 Node* id = Identity_impl(phase); 713 714 if (id == this && phase->is_IterGVN()) { 715 Node* n = in(ValueIn); 716 // No success in super call. Try to combine identical read barriers. 717 for (DUIterator i = n->outs(); n->has_out(i); i++) { 718 Node* sibling = n->out(i); 719 if (sibling == this || sibling->Opcode() != Op_ShenandoahReadBarrier) { 720 continue; 721 } 722 assert(sibling->in(ValueIn) == in(ValueIn), "sanity"); 723 if (phase->is_IterGVN()->hash_find(sibling) && 724 sibling->bottom_type() == bottom_type() && 725 sibling->in(Control) == in(Control) && 726 dominates_memory_rb(phase, sibling, this, phase->is_IterGVN() == NULL)) { 727 /* 728 if (in(Memory) != sibling->in(Memory)) { 729 tty->print_cr("interesting rb-fold"); 730 dump(); 731 sibling->dump(); 732 } 733 */ 734 return sibling; 735 } 736 } 737 } 738 return id; 739 } 740 741 const Type* ShenandoahBarrierNode::Value(PhaseGVN* phase) const { 742 // Either input is TOP ==> the result is TOP 743 const Type *t1 = phase->type(in(Memory)); 744 if (t1 == Type::TOP) return Type::TOP; 745 const Type *t2 = phase->type(in(ValueIn)); 746 if( t2 == Type::TOP ) return Type::TOP; 747 748 if (t2 == TypePtr::NULL_PTR) { 749 return _type; 750 } 751 752 const Type* type = t2->is_oopptr()->cast_to_nonconst(); 753 return type; 754 } 755 756 uint ShenandoahBarrierNode::hash() const { 757 return TypeNode::hash() + _allow_fromspace; 758 } 759 760 uint ShenandoahBarrierNode::cmp(const Node& n) const { 761 return _allow_fromspace == ((ShenandoahBarrierNode&) n)._allow_fromspace 762 && TypeNode::cmp(n); 763 } 764 765 uint ShenandoahBarrierNode::size_of() const { 766 return sizeof(*this); 767 } 768 769 Node* ShenandoahWBMemProjNode::Identity(PhaseGVN* phase) { 770 771 Node* wb = in(0); 772 if (wb->is_top()) return phase->C->top(); // Dead path. 773 774 assert(wb->Opcode() == Op_ShenandoahWriteBarrier, "expect write barrier"); 775 PhaseIterGVN* igvn = phase->is_IterGVN(); 776 // We can't do the below unless the graph is fully constructed. 777 if (igvn == NULL) { 778 return this; 779 } 780 781 // If the mem projection has no barrier users, it's not needed anymore. 782 if (wb->outcnt() == 1) { 783 return wb->in(ShenandoahBarrierNode::Memory); 784 } 785 786 return this; 787 } 788 789 #ifdef ASSERT 790 bool ShenandoahBarrierNode::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) { 791 assert(phis.size() == 0, ""); 792 793 while (true) { 794 if (in->bottom_type() == TypePtr::NULL_PTR) { 795 if (trace) {tty->print_cr("NULL");} 796 } else if (!in->bottom_type()->make_ptr()->make_oopptr()) { 797 if (trace) {tty->print_cr("Non oop");} 798 } else if (t == ShenandoahLoad && ShenandoahOptimizeStableFinals && 799 in->bottom_type()->make_ptr()->isa_aryptr() && 800 in->bottom_type()->make_ptr()->is_aryptr()->is_stable()) { 801 if (trace) {tty->print_cr("Stable array load");} 802 } else { 803 if (in->is_ConstraintCast()) { 804 in = in->in(1); 805 continue; 806 } else if (in->is_AddP()) { 807 assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access"); 808 in = in->in(AddPNode::Address); 809 continue; 810 } else if (in->is_Con()) { 811 if (trace) {tty->print("Found constant"); in->dump();} 812 } else if (in->is_ShenandoahBarrier()) { 813 if (t == ShenandoahOopStore) { 814 if (in->Opcode() != Op_ShenandoahWriteBarrier) { 815 return false; 816 } 817 uint i = 0; 818 for (; i < phis.size(); i++) { 819 Node* n = phis.node_at(i); 820 if (n->Opcode() == Op_ShenandoahEnqueueBarrier) { 821 break; 822 } 823 } 824 if (i == phis.size()) { 825 return false; 826 } 827 } else if (t == ShenandoahStore && in->Opcode() != Op_ShenandoahWriteBarrier) { 828 return false; 829 } 830 barriers_used.push(in); 831 if (trace) {tty->print("Found barrier"); in->dump();} 832 } else if (in->Opcode() == Op_ShenandoahEnqueueBarrier) { 833 if (t != ShenandoahOopStore) { 834 return false; 835 } 836 if (trace) {tty->print("Found enqueue barrier"); in->dump();} 837 phis.push(in, in->req()); 838 in = in->in(1); 839 continue; 840 } else if (in->is_Proj() && in->in(0)->is_Allocate()) { 841 if (trace) {tty->print("Found alloc"); in->in(0)->dump();} 842 } else if (in->is_Phi()) { 843 if (!visited.test_set(in->_idx)) { 844 if (trace) {tty->print("Pushed phi:"); in->dump();} 845 phis.push(in, 2); 846 in = in->in(1); 847 continue; 848 } 849 if (trace) {tty->print("Already seen phi:"); in->dump();} 850 } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) { 851 if (!visited.test_set(in->_idx)) { 852 if (trace) {tty->print("Pushed cmovep:"); in->dump();} 853 phis.push(in, CMoveNode::IfTrue); 854 in = in->in(CMoveNode::IfFalse); 855 continue; 856 } 857 if (trace) {tty->print("Already seen cmovep:"); in->dump();} 858 } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) { 859 in = in->in(1); 860 continue; 861 } else { 862 return false; 863 } 864 } 865 bool cont = false; 866 while (phis.is_nonempty()) { 867 uint idx = phis.index(); 868 Node* phi = phis.node(); 869 if (idx >= phi->req()) { 870 if (trace) {tty->print("Popped phi:"); phi->dump();} 871 phis.pop(); 872 continue; 873 } 874 if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();} 875 in = phi->in(idx); 876 phis.set_index(idx+1); 877 cont = true; 878 break; 879 } 880 if (!cont) { 881 break; 882 } 883 } 884 return true; 885 } 886 887 void ShenandoahBarrierNode::report_verify_failure(const char *msg, Node *n1, Node *n2) { 888 if (n1 != NULL) { 889 n1->dump(+10); 890 } 891 if (n2 != NULL) { 892 n2->dump(+10); 893 } 894 fatal("%s", msg); 895 } 896 897 void ShenandoahBarrierNode::verify(RootNode* root) { 898 ResourceMark rm; 899 Unique_Node_List wq; 900 GrowableArray<Node*> barriers; 901 Unique_Node_List barriers_used; 902 Node_Stack phis(0); 903 VectorSet visited(Thread::current()->resource_area()); 904 const bool trace = false; 905 const bool verify_no_useless_barrier = false; 906 907 wq.push(root); 908 for (uint next = 0; next < wq.size(); next++) { 909 Node *n = wq.at(next); 910 if (n->is_Load()) { 911 const bool trace = false; 912 if (trace) {tty->print("Verifying"); n->dump();} 913 if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) { 914 if (trace) {tty->print_cr("Load range/klass");} 915 } else { 916 const TypePtr* adr_type = n->as_Load()->adr_type(); 917 918 if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) { 919 if (trace) {tty->print_cr("Mark load");} 920 } else if (adr_type->isa_instptr() && 921 adr_type->is_instptr()->klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) && 922 adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset) { 923 if (trace) {tty->print_cr("Reference.get()");} 924 } else { 925 bool verify = true; 926 if (adr_type->isa_instptr()) { 927 const TypeInstPtr* tinst = adr_type->is_instptr(); 928 ciKlass* k = tinst->klass(); 929 assert(k->is_instance_klass(), ""); 930 ciInstanceKlass* ik = (ciInstanceKlass*)k; 931 int offset = adr_type->offset(); 932 933 if ((ik->debug_final_field_at(offset) && ShenandoahOptimizeInstanceFinals) || 934 (ik->debug_stable_field_at(offset) && ShenandoahOptimizeStableFinals)) { 935 if (trace) {tty->print_cr("Final/stable");} 936 verify = false; 937 } else if (k == ciEnv::current()->Class_klass() && 938 tinst->const_oop() != NULL && 939 tinst->offset() >= (ik->size_helper() * wordSize)) { 940 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 941 ciField* field = k->get_field_by_offset(tinst->offset(), true); 942 if ((ShenandoahOptimizeStaticFinals && field->is_final()) || 943 (ShenandoahOptimizeStableFinals && field->is_stable())) { 944 verify = false; 945 } 946 } 947 } 948 949 if (verify && !ShenandoahBarrierNode::verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) { 950 report_verify_failure("Shenandoah verification: Load should have barriers", n); 951 } 952 } 953 } 954 } else if (n->is_Store()) { 955 const bool trace = false; 956 957 if (trace) {tty->print("Verifying"); n->dump();} 958 if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) { 959 Node* adr = n->in(MemNode::Address); 960 bool verify = true; 961 962 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) { 963 adr = adr->in(AddPNode::Address); 964 if (adr->is_AddP()) { 965 assert(adr->in(AddPNode::Base)->is_top(), ""); 966 adr = adr->in(AddPNode::Address); 967 if (adr->Opcode() == Op_LoadP && 968 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() && 969 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 970 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) { 971 if (trace) {tty->print_cr("SATB prebarrier");} 972 verify = false; 973 } 974 } 975 } 976 977 if (verify && !ShenandoahBarrierNode::verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahStoreValEnqueueBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) { 978 report_verify_failure("Shenandoah verification: Store should have barriers", n); 979 } 980 } 981 if (!ShenandoahBarrierNode::verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) { 982 report_verify_failure("Shenandoah verification: Store (address) should have barriers", n); 983 } 984 } else if (n->Opcode() == Op_CmpP) { 985 const bool trace = false; 986 987 Node* in1 = n->in(1); 988 Node* in2 = n->in(2); 989 if (in1->bottom_type()->isa_oopptr()) { 990 if (trace) {tty->print("Verifying"); n->dump();} 991 992 bool mark_inputs = false; 993 if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR || 994 (in1->is_Con() || in2->is_Con())) { 995 if (trace) {tty->print_cr("Comparison against a constant");} 996 mark_inputs = true; 997 } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) || 998 (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) { 999 if (trace) {tty->print_cr("Comparison with newly alloc'ed object");} 1000 mark_inputs = true; 1001 } else { 1002 assert(in2->bottom_type()->isa_oopptr(), ""); 1003 1004 if (!ShenandoahBarrierNode::verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) || 1005 !ShenandoahBarrierNode::verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) { 1006 report_verify_failure("Shenandoah verification: Cmp should have barriers", n); 1007 } 1008 } 1009 if (verify_no_useless_barrier && 1010 mark_inputs && 1011 (!ShenandoahBarrierNode::verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) || 1012 !ShenandoahBarrierNode::verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) { 1013 phis.clear(); 1014 visited.Reset(); 1015 } 1016 } 1017 } else if (n->is_LoadStore()) { 1018 if (n->in(MemNode::ValueIn)->bottom_type()->isa_ptr() && 1019 !ShenandoahBarrierNode::verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahLoad, trace, barriers_used)) { 1020 report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n); 1021 } 1022 1023 if (n->in(MemNode::Address)->bottom_type()->isa_oopptr() && !ShenandoahBarrierNode::verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) { 1024 report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n); 1025 } 1026 } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) { 1027 CallNode* call = n->as_Call(); 1028 1029 static struct { 1030 const char* name; 1031 struct { 1032 int pos; 1033 verify_type t; 1034 } args[6]; 1035 } calls[] = { 1036 "aescrypt_encryptBlock", 1037 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad }, 1038 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1039 "aescrypt_decryptBlock", 1040 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad }, 1041 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1042 "multiplyToLen", 1043 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { TypeFunc::Parms+4, ShenandoahStore }, 1044 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1045 "squareToLen", 1046 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { -1, ShenandoahNone}, 1047 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1048 "montgomery_multiply", 1049 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, 1050 { TypeFunc::Parms+6, ShenandoahStore }, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1051 "montgomery_square", 1052 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+5, ShenandoahStore }, 1053 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1054 "mulAdd", 1055 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, 1056 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1057 "vectorizedMismatch", 1058 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, 1059 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1060 "updateBytesCRC32", 1061 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone}, 1062 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1063 "updateBytesAdler32", 1064 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone}, 1065 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1066 "updateBytesCRC32C", 1067 { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad}, { -1, ShenandoahNone}, 1068 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1069 "counterMode_AESCrypt", 1070 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad }, 1071 { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } }, 1072 "cipherBlockChaining_encryptAESCrypt", 1073 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad }, 1074 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1075 "cipherBlockChaining_decryptAESCrypt", 1076 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad }, 1077 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1078 "shenandoah_clone_barrier", 1079 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone}, 1080 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1081 "ghash_processBlocks", 1082 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, 1083 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1084 "sha1_implCompress", 1085 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone }, 1086 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1087 "sha256_implCompress", 1088 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone }, 1089 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1090 "sha512_implCompress", 1091 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone }, 1092 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1093 "sha1_implCompressMB", 1094 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone }, 1095 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1096 "sha256_implCompressMB", 1097 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone }, 1098 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1099 "sha512_implCompressMB", 1100 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone }, 1101 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} }, 1102 }; 1103 1104 if (call->is_call_to_arraycopystub()) { 1105 Node* dest = NULL; 1106 const TypeTuple* args = n->as_Call()->_tf->domain(); 1107 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) { 1108 if (args->field_at(i)->isa_ptr()) { 1109 j++; 1110 if (j == 2) { 1111 dest = n->in(i); 1112 break; 1113 } 1114 } 1115 } 1116 if (!ShenandoahBarrierNode::verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) || 1117 !ShenandoahBarrierNode::verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) { 1118 report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n); 1119 } 1120 } else if (strlen(call->_name) > 5 && 1121 !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) { 1122 if (!ShenandoahBarrierNode::verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) { 1123 report_verify_failure("Shenandoah verification: _fill should have barriers", n); 1124 } 1125 } else if (!strcmp(call->_name, "shenandoah_wb_pre")) { 1126 // skip 1127 } else { 1128 const int calls_len = sizeof(calls) / sizeof(calls[0]); 1129 int i = 0; 1130 for (; i < calls_len; i++) { 1131 if (!strcmp(calls[i].name, call->_name)) { 1132 break; 1133 } 1134 } 1135 if (i != calls_len) { 1136 const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]); 1137 for (uint j = 0; j < args_len; j++) { 1138 int pos = calls[i].args[j].pos; 1139 if (pos == -1) { 1140 break; 1141 } 1142 if (!ShenandoahBarrierNode::verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) { 1143 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n); 1144 } 1145 } 1146 for (uint j = TypeFunc::Parms; j < call->req(); j++) { 1147 if (call->in(j)->bottom_type()->make_ptr() && 1148 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) { 1149 uint k = 0; 1150 for (; k < args_len && calls[i].args[k].pos != (int)j; k++); 1151 if (k == args_len) { 1152 fatal("arg %d for call %s not covered", j, call->_name); 1153 } 1154 } 1155 } 1156 } else { 1157 for (uint j = TypeFunc::Parms; j < call->req(); j++) { 1158 if (call->in(j)->bottom_type()->make_ptr() && 1159 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) { 1160 fatal("%s not covered", call->_name); 1161 } 1162 } 1163 } 1164 } 1165 } else if (n->is_ShenandoahBarrier()) { 1166 assert(!barriers.contains(n), ""); 1167 assert(n->Opcode() != Op_ShenandoahWriteBarrier || n->find_out_with(Op_ShenandoahWBMemProj) != NULL, "bad shenandoah write barrier"); 1168 assert(n->Opcode() != Op_ShenandoahWriteBarrier || n->outcnt() > 1, "bad shenandoah write barrier"); 1169 barriers.push(n); 1170 } else if (n->Opcode() == Op_ShenandoahEnqueueBarrier) { 1171 // skip 1172 } else if (n->is_AddP() 1173 || n->is_Phi() 1174 || n->is_ConstraintCast() 1175 || n->Opcode() == Op_Return 1176 || n->Opcode() == Op_CMoveP 1177 || n->Opcode() == Op_CMoveN 1178 || n->Opcode() == Op_Rethrow 1179 || n->is_MemBar() 1180 || n->Opcode() == Op_Conv2B 1181 || n->Opcode() == Op_SafePoint 1182 || n->is_CallJava() 1183 || n->Opcode() == Op_Unlock 1184 || n->Opcode() == Op_EncodeP 1185 || n->Opcode() == Op_DecodeN) { 1186 // nothing to do 1187 } else { 1188 static struct { 1189 int opcode; 1190 struct { 1191 int pos; 1192 verify_type t; 1193 } inputs[2]; 1194 } others[] = { 1195 Op_FastLock, 1196 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} }, 1197 Op_Lock, 1198 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone} }, 1199 Op_ArrayCopy, 1200 { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } }, 1201 Op_StrCompressedCopy, 1202 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } }, 1203 Op_StrInflatedCopy, 1204 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } }, 1205 Op_AryEq, 1206 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } }, 1207 Op_StrIndexOf, 1208 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } }, 1209 Op_StrComp, 1210 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } }, 1211 Op_StrEquals, 1212 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } }, 1213 Op_EncodeISOArray, 1214 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } }, 1215 Op_HasNegatives, 1216 { { 2, ShenandoahLoad }, { -1, ShenandoahNone} }, 1217 Op_CastP2X, 1218 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} }, 1219 Op_StrIndexOfChar, 1220 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } }, 1221 }; 1222 1223 const int others_len = sizeof(others) / sizeof(others[0]); 1224 int i = 0; 1225 for (; i < others_len; i++) { 1226 if (others[i].opcode == n->Opcode()) { 1227 break; 1228 } 1229 } 1230 uint stop = n->is_Call() ? n->as_Call()->tf()->domain()->cnt() : n->req(); 1231 if (i != others_len) { 1232 const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]); 1233 for (uint j = 0; j < inputs_len; j++) { 1234 int pos = others[i].inputs[j].pos; 1235 if (pos == -1) { 1236 break; 1237 } 1238 if (!ShenandoahBarrierNode::verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) { 1239 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n); 1240 } 1241 } 1242 for (uint j = 1; j < stop; j++) { 1243 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() && 1244 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) { 1245 uint k = 0; 1246 for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++); 1247 if (k == inputs_len) { 1248 fatal("arg %d for node %s not covered", j, n->Name()); 1249 } 1250 } 1251 } 1252 } else { 1253 for (uint j = 1; j < stop; j++) { 1254 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() && 1255 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) { 1256 fatal("%s not covered", n->Name()); 1257 } 1258 } 1259 } 1260 } 1261 1262 if (n->is_SafePoint()) { 1263 SafePointNode* sfpt = n->as_SafePoint(); 1264 if (verify_no_useless_barrier && sfpt->jvms() != NULL) { 1265 for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) { 1266 if (!ShenandoahBarrierNode::verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) { 1267 phis.clear(); 1268 visited.Reset(); 1269 } 1270 } 1271 } 1272 } 1273 for( uint i = 0; i < n->len(); ++i ) { 1274 Node *m = n->in(i); 1275 if (m == NULL) continue; 1276 1277 // In most cases, inputs should be known to be non null. If it's 1278 // not the case, it could be a missing cast_not_null() in an 1279 // intrinsic or support might be needed in AddPNode::Ideal() to 1280 // avoid a NULL+offset input. 1281 if (!(n->is_Phi() || 1282 (n->is_SafePoint() && (!n->is_CallRuntime() || !strcmp(n->as_Call()->_name, "shenandoah_wb_pre") || !strcmp(n->as_Call()->_name, "unsafe_arraycopy"))) || 1283 n->Opcode() == Op_CmpP || 1284 n->Opcode() == Op_CmpN || 1285 (n->Opcode() == Op_StoreP && i == StoreNode::ValueIn) || 1286 (n->Opcode() == Op_StoreN && i == StoreNode::ValueIn) || 1287 n->is_ConstraintCast() || 1288 n->Opcode() == Op_Return || 1289 n->Opcode() == Op_Conv2B || 1290 n->is_AddP() || 1291 n->Opcode() == Op_CMoveP || 1292 n->Opcode() == Op_CMoveN || 1293 n->Opcode() == Op_Rethrow || 1294 n->is_MemBar() || 1295 n->is_Mem() || 1296 n->Opcode() == Op_AryEq || 1297 n->Opcode() == Op_SCMemProj || 1298 n->Opcode() == Op_EncodeP || 1299 n->Opcode() == Op_DecodeN || 1300 n->Opcode() == Op_ShenandoahWriteBarrier || 1301 n->Opcode() == Op_ShenandoahWBMemProj || 1302 n->Opcode() == Op_ShenandoahEnqueueBarrier)) { 1303 if (m->bottom_type()->make_oopptr() && m->bottom_type()->make_oopptr()->meet(TypePtr::NULL_PTR) == m->bottom_type()) { 1304 report_verify_failure("Shenandoah verification: null input", n, m); 1305 } 1306 } 1307 1308 wq.push(m); 1309 } 1310 } 1311 1312 if (verify_no_useless_barrier) { 1313 for (int i = 0; i < barriers.length(); i++) { 1314 Node* n = barriers.at(i); 1315 if (!barriers_used.member(n)) { 1316 tty->print("XXX useless barrier"); n->dump(-2); 1317 ShouldNotReachHere(); 1318 } 1319 } 1320 } 1321 } 1322 #endif 1323 1324 bool ShenandoahBarrierNode::is_dominator_same_ctrl(Node*c, Node* d, Node* n, PhaseIdealLoop* phase) { 1325 // That both nodes have the same control is not sufficient to prove 1326 // domination, verify that there's no path from d to n 1327 ResourceMark rm; 1328 Unique_Node_List wq; 1329 wq.push(d); 1330 for (uint next = 0; next < wq.size(); next++) { 1331 Node *m = wq.at(next); 1332 if (m == n) { 1333 return false; 1334 } 1335 if (m->is_Phi() && m->in(0)->is_Loop()) { 1336 assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control"); 1337 } else { 1338 for (uint i = 0; i < m->req(); i++) { 1339 if (m->in(i) != NULL && phase->ctrl_or_self(m->in(i)) == c) { 1340 wq.push(m->in(i)); 1341 } 1342 } 1343 } 1344 } 1345 return true; 1346 } 1347 1348 bool ShenandoahBarrierNode::is_dominator(Node *d_c, Node *n_c, Node* d, Node* n, PhaseIdealLoop* phase) { 1349 if (d_c != n_c) { 1350 return phase->is_dominator(d_c, n_c); 1351 } 1352 return is_dominator_same_ctrl(d_c, d, n, phase); 1353 } 1354 1355 Node* next_mem(Node* mem, int alias) { 1356 Node* res = NULL; 1357 if (mem->is_Proj()) { 1358 res = mem->in(0); 1359 } else if (mem->is_SafePoint() || mem->is_MemBar()) { 1360 res = mem->in(TypeFunc::Memory); 1361 } else if (mem->is_Phi()) { 1362 res = mem->in(1); 1363 } else if (mem->is_ShenandoahBarrier()) { 1364 res = mem->in(ShenandoahBarrierNode::Memory); 1365 } else if (mem->is_MergeMem()) { 1366 res = mem->as_MergeMem()->memory_at(alias); 1367 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) { 1368 assert(alias = Compile::AliasIdxRaw, "following raw memory can't lead to a barrier"); 1369 res = mem->in(MemNode::Memory); 1370 } else { 1371 #ifdef ASSERT 1372 mem->dump(); 1373 #endif 1374 ShouldNotReachHere(); 1375 } 1376 return res; 1377 } 1378 1379 Node* ShenandoahBarrierNode::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) { 1380 Node* iffproj = NULL; 1381 while (c != dom) { 1382 Node* next = phase->idom(c); 1383 assert(next->unique_ctrl_out() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?"); 1384 if (c->is_Region()) { 1385 ResourceMark rm; 1386 Unique_Node_List wq; 1387 wq.push(c); 1388 for (uint i = 0; i < wq.size(); i++) { 1389 Node *n = wq.at(i); 1390 if (n == next) { 1391 continue; 1392 } 1393 if (n->is_Region()) { 1394 for (uint j = 1; j < n->req(); j++) { 1395 wq.push(n->in(j)); 1396 } 1397 } else { 1398 wq.push(n->in(0)); 1399 } 1400 } 1401 for (uint i = 0; i < wq.size(); i++) { 1402 Node *n = wq.at(i); 1403 assert(n->is_CFG(), ""); 1404 if (n->is_Multi()) { 1405 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1406 Node* u = n->fast_out(j); 1407 if (u->is_CFG()) { 1408 if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) { 1409 return NodeSentinel; 1410 } 1411 } 1412 } 1413 } 1414 } 1415 } else if (c->is_Proj()) { 1416 if (c->is_IfProj()) { 1417 if (c->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) != NULL) { 1418 // continue; 1419 } else { 1420 if (!allow_one_proj) { 1421 return NodeSentinel; 1422 } 1423 if (iffproj == NULL) { 1424 iffproj = c; 1425 } else { 1426 return NodeSentinel; 1427 } 1428 } 1429 } else if (c->Opcode() == Op_JumpProj) { 1430 return NodeSentinel; // unsupported 1431 } else if (c->Opcode() == Op_CatchProj) { 1432 return NodeSentinel; // unsupported 1433 } else if (c->Opcode() == Op_CProj && next->Opcode() == Op_NeverBranch) { 1434 return NodeSentinel; // unsupported 1435 } else { 1436 assert(next->unique_ctrl_out() == c, "unsupported branch pattern"); 1437 } 1438 } 1439 c = next; 1440 } 1441 return iffproj; 1442 } 1443 1444 #ifdef ASSERT 1445 void ShenandoahWriteBarrierNode::memory_dominates_all_paths_helper(Node* c, Node* rep_ctrl, Unique_Node_List& controls, PhaseIdealLoop* phase) { 1446 const bool trace = false; 1447 if (trace) { tty->print("X control is"); c->dump(); } 1448 1449 uint start = controls.size(); 1450 controls.push(c); 1451 for (uint i = start; i < controls.size(); i++) { 1452 Node *n = controls.at(i); 1453 1454 if (trace) { tty->print("X from"); n->dump(); } 1455 1456 if (n == rep_ctrl) { 1457 continue; 1458 } 1459 1460 if (n->is_Proj()) { 1461 Node* n_dom = n->in(0); 1462 IdealLoopTree* n_dom_loop = phase->get_loop(n_dom); 1463 if (n->is_IfProj() && n_dom->outcnt() == 2) { 1464 n_dom_loop = phase->get_loop(n_dom->as_If()->proj_out(n->as_Proj()->_con == 0 ? 1 : 0)); 1465 } 1466 if (n_dom_loop != phase->ltree_root()) { 1467 Node* tail = n_dom_loop->tail(); 1468 if (tail->is_Region()) { 1469 for (uint j = 1; j < tail->req(); j++) { 1470 if (phase->is_dominator(n_dom, tail->in(j)) && !phase->is_dominator(n, tail->in(j))) { 1471 assert(phase->is_dominator(rep_ctrl, tail->in(j)), "why are we here?"); 1472 // entering loop from below, mark backedge 1473 if (trace) { tty->print("X pushing backedge"); tail->in(j)->dump(); } 1474 controls.push(tail->in(j)); 1475 //assert(n->in(0) == n_dom, "strange flow control"); 1476 } 1477 } 1478 } else if (phase->get_loop(n) != n_dom_loop && phase->is_dominator(n_dom, tail)) { 1479 // entering loop from below, mark backedge 1480 if (trace) { tty->print("X pushing backedge"); tail->dump(); } 1481 controls.push(tail); 1482 //assert(n->in(0) == n_dom, "strange flow control"); 1483 } 1484 } 1485 } 1486 1487 if (n->is_Loop()) { 1488 Node* c = n->in(LoopNode::EntryControl); 1489 if (trace) { tty->print("X pushing"); c->dump(); } 1490 controls.push(c); 1491 } else if (n->is_Region()) { 1492 for (uint i = 1; i < n->req(); i++) { 1493 Node* c = n->in(i); 1494 if (trace) { tty->print("X pushing"); c->dump(); } 1495 controls.push(c); 1496 } 1497 } else { 1498 Node* c = n->in(0); 1499 if (trace) { tty->print("X pushing"); c->dump(); } 1500 controls.push(c); 1501 } 1502 } 1503 } 1504 1505 bool ShenandoahWriteBarrierNode::memory_dominates_all_paths(Node* mem, Node* rep_ctrl, int alias, PhaseIdealLoop* phase) { 1506 const bool trace = false; 1507 if (trace) { 1508 tty->print("XXX mem is"); mem->dump(); 1509 tty->print("XXX rep ctrl is"); rep_ctrl->dump(); 1510 tty->print_cr("XXX alias is %d", alias); 1511 } 1512 ResourceMark rm; 1513 Unique_Node_List wq; 1514 Unique_Node_List controls; 1515 wq.push(mem); 1516 for (uint next = 0; next < wq.size(); next++) { 1517 Node *nn = wq.at(next); 1518 if (trace) { tty->print("XX from mem"); nn->dump(); } 1519 assert(nn->bottom_type() == Type::MEMORY, "memory only"); 1520 1521 if (nn->is_Phi()) { 1522 Node* r = nn->in(0); 1523 for (DUIterator_Fast jmax, j = r->fast_outs(jmax); j < jmax; j++) { 1524 Node* u = r->fast_out(j); 1525 if (u->is_Phi() && u->bottom_type() == Type::MEMORY && u != nn && 1526 (u->adr_type() == TypePtr::BOTTOM || phase->C->get_alias_index(u->adr_type()) == alias)) { 1527 if (trace) { tty->print("XX Next mem (other phi)"); u->dump(); } 1528 wq.push(u); 1529 } 1530 } 1531 } 1532 1533 for (DUIterator_Fast imax, i = nn->fast_outs(imax); i < imax; i++) { 1534 Node* use = nn->fast_out(i); 1535 1536 if (trace) { tty->print("XX use %p", use->adr_type()); use->dump(); } 1537 if (use->is_CFG() && use->in(TypeFunc::Memory) == nn) { 1538 Node* c = use->in(0); 1539 if (phase->is_dominator(rep_ctrl, c)) { 1540 memory_dominates_all_paths_helper(c, rep_ctrl, controls, phase); 1541 } else if (use->is_CallStaticJava() && use->as_CallStaticJava()->uncommon_trap_request() != 0 && c->is_Region()) { 1542 Node* region = c; 1543 if (trace) { tty->print("XX unc region"); region->dump(); } 1544 for (uint j = 1; j < region->req(); j++) { 1545 if (phase->is_dominator(rep_ctrl, region->in(j))) { 1546 if (trace) { tty->print("XX unc follows"); region->in(j)->dump(); } 1547 memory_dominates_all_paths_helper(region->in(j), rep_ctrl, controls, phase); 1548 } 1549 } 1550 } 1551 //continue; 1552 } else if (use->is_Phi()) { 1553 assert(use->bottom_type() == Type::MEMORY, "bad phi"); 1554 if ((use->adr_type() == TypePtr::BOTTOM /*&& !shenandoah_has_alias_phi(C, use, alias)*/) || 1555 phase->C->get_alias_index(use->adr_type()) == alias) { 1556 for (uint j = 1; j < use->req(); j++) { 1557 if (use->in(j) == nn) { 1558 Node* c = use->in(0)->in(j); 1559 if (phase->is_dominator(rep_ctrl, c)) { 1560 memory_dominates_all_paths_helper(c, rep_ctrl, controls, phase); 1561 } 1562 } 1563 } 1564 } 1565 // continue; 1566 } 1567 1568 if (use->is_MergeMem()) { 1569 if (use->as_MergeMem()->memory_at(alias) == nn) { 1570 if (trace) { tty->print("XX Next mem"); use->dump(); } 1571 // follow the memory edges 1572 wq.push(use); 1573 } 1574 } else if (use->is_Phi()) { 1575 assert(use->bottom_type() == Type::MEMORY, "bad phi"); 1576 if ((use->adr_type() == TypePtr::BOTTOM /*&& !shenandoah_has_alias_phi(C, use, alias)*/) || 1577 phase->C->get_alias_index(use->adr_type()) == alias) { 1578 if (trace) { tty->print("XX Next mem"); use->dump(); } 1579 // follow the memory edges 1580 wq.push(use); 1581 } 1582 } else if (use->bottom_type() == Type::MEMORY && 1583 (use->adr_type() == TypePtr::BOTTOM || phase->C->get_alias_index(use->adr_type()) == alias)) { 1584 if (trace) { tty->print("XX Next mem"); use->dump(); } 1585 // follow the memory edges 1586 wq.push(use); 1587 } else if ((use->is_SafePoint() || use->is_MemBar()) && 1588 (use->adr_type() == TypePtr::BOTTOM || phase->C->get_alias_index(use->adr_type()) == alias)) { 1589 for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) { 1590 Node* u = use->fast_out(j); 1591 if (u->bottom_type() == Type::MEMORY) { 1592 if (trace) { tty->print("XX Next mem"); u->dump(); } 1593 // follow the memory edges 1594 wq.push(u); 1595 } 1596 } 1597 } else if (use->Opcode() == Op_ShenandoahWriteBarrier && phase->C->get_alias_index(use->adr_type()) == alias) { 1598 Node* m = use->find_out_with(Op_ShenandoahWBMemProj); 1599 if (m != NULL) { 1600 if (trace) { tty->print("XX Next mem"); m->dump(); } 1601 // follow the memory edges 1602 wq.push(m); 1603 } 1604 } 1605 } 1606 } 1607 1608 if (controls.size() == 0) { 1609 return false; 1610 } 1611 1612 for (uint i = 0; i < controls.size(); i++) { 1613 Node *n = controls.at(i); 1614 1615 if (trace) { tty->print("X checking"); n->dump(); } 1616 1617 if (n->unique_ctrl_out() != NULL) { 1618 continue; 1619 } 1620 1621 if (n->Opcode() == Op_NeverBranch) { 1622 Node* taken = n->as_Multi()->proj_out(0); 1623 if (!controls.member(taken)) { 1624 if (trace) { tty->print("X not seen"); taken->dump(); } 1625 return false; 1626 } 1627 continue; 1628 } 1629 1630 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1631 Node* u = n->fast_out(j); 1632 1633 if (u->is_CFG()) { 1634 if (!controls.member(u)) { 1635 if (u->is_Proj() && u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) { 1636 if (trace) { tty->print("X not seen but unc"); u->dump(); } 1637 } else { 1638 Node* c = u; 1639 do { 1640 c = c->unique_ctrl_out(); 1641 } while (c != NULL && c->is_Region()); 1642 if (c != NULL && c->Opcode() == Op_Halt) { 1643 if (trace) { tty->print("X not seen but halt"); c->dump(); } 1644 } else { 1645 if (trace) { tty->print("X not seen"); u->dump(); } 1646 return false; 1647 } 1648 } 1649 } else { 1650 if (trace) { tty->print("X seen"); u->dump(); } 1651 } 1652 } 1653 } 1654 } 1655 return true; 1656 } 1657 #endif 1658 1659 Node* ShenandoahBarrierNode::dom_mem(Node* mem, Node*& mem_ctrl, Node* n, Node* rep_ctrl, int alias, PhaseIdealLoop* phase) { 1660 ResourceMark rm; 1661 VectorSet wq(Thread::current()->resource_area()); 1662 wq.set(mem->_idx); 1663 mem_ctrl = phase->get_ctrl(mem); 1664 while (!is_dominator(mem_ctrl, rep_ctrl, mem, n, phase)) { 1665 mem = next_mem(mem, alias); 1666 if (wq.test_set(mem->_idx)) { 1667 return NULL; // hit an unexpected loop 1668 } 1669 mem_ctrl = phase->ctrl_or_self(mem); 1670 } 1671 if (mem->is_MergeMem()) { 1672 mem = mem->as_MergeMem()->memory_at(alias); 1673 mem_ctrl = phase->ctrl_or_self(mem); 1674 } 1675 return mem; 1676 } 1677 1678 Node* ShenandoahBarrierNode::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) { 1679 ResourceMark rm; 1680 VectorSet wq(Thread::current()->resource_area()); 1681 wq.set(mem->_idx); 1682 mem_ctrl = phase->ctrl_or_self(mem); 1683 while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) { 1684 mem = next_mem(mem, alias); 1685 if (wq.test_set(mem->_idx)) { 1686 return NULL; 1687 } 1688 mem_ctrl = phase->ctrl_or_self(mem); 1689 } 1690 if (mem->is_MergeMem()) { 1691 mem = mem->as_MergeMem()->memory_at(alias); 1692 mem_ctrl = phase->ctrl_or_self(mem); 1693 } 1694 return mem; 1695 } 1696 1697 static void disconnect_barrier_mem(Node* wb, PhaseIterGVN& igvn) { 1698 Node* mem_in = wb->in(ShenandoahBarrierNode::Memory); 1699 Node* proj = wb->find_out_with(Op_ShenandoahWBMemProj); 1700 1701 for (DUIterator_Last imin, i = proj->last_outs(imin); i >= imin; ) { 1702 Node* u = proj->last_out(i); 1703 igvn.rehash_node_delayed(u); 1704 int nb = u->replace_edge(proj, mem_in); 1705 assert(nb > 0, "no replacement?"); 1706 i -= nb; 1707 } 1708 } 1709 1710 Node* ShenandoahWriteBarrierNode::move_above_predicates(LoopNode* cl, Node* val_ctrl, PhaseIdealLoop* phase) { 1711 Node* entry = cl->skip_strip_mined(-1)->in(LoopNode::EntryControl); 1712 Node* above_pred = phase->skip_all_loop_predicates(entry); 1713 Node* ctrl = entry; 1714 while (ctrl != above_pred) { 1715 Node* next = ctrl->in(0); 1716 if (!phase->is_dominator(val_ctrl, next)) { 1717 break; 1718 } 1719 ctrl = next; 1720 } 1721 return ctrl; 1722 } 1723 1724 static MemoryGraphFixer* find_fixer(GrowableArray<MemoryGraphFixer*>& memory_graph_fixers, int alias) { 1725 for (int i = 0; i < memory_graph_fixers.length(); i++) { 1726 if (memory_graph_fixers.at(i)->alias() == alias) { 1727 return memory_graph_fixers.at(i); 1728 } 1729 } 1730 return NULL; 1731 } 1732 1733 static MemoryGraphFixer* create_fixer(GrowableArray<MemoryGraphFixer*>& memory_graph_fixers, int alias, PhaseIdealLoop* phase, bool include_lsm) { 1734 assert(find_fixer(memory_graph_fixers, alias) == NULL, "none should exist yet"); 1735 MemoryGraphFixer* fixer = new MemoryGraphFixer(alias, include_lsm, phase); 1736 memory_graph_fixers.push(fixer); 1737 return fixer; 1738 } 1739 1740 void ShenandoahWriteBarrierNode::try_move_before_loop_helper(LoopNode* cl, Node* val_ctrl, GrowableArray<MemoryGraphFixer*>& memory_graph_fixers, PhaseIdealLoop* phase, bool include_lsm, Unique_Node_List& uses) { 1741 assert(cl->is_Loop(), "bad control"); 1742 Node* ctrl = move_above_predicates(cl, val_ctrl, phase); 1743 Node* mem_ctrl = NULL; 1744 int alias = phase->C->get_alias_index(adr_type()); 1745 1746 MemoryGraphFixer* fixer = find_fixer(memory_graph_fixers, alias); 1747 if (fixer == NULL) { 1748 fixer = create_fixer(memory_graph_fixers, alias, phase, include_lsm); 1749 } 1750 1751 Node* proj = find_out_with(Op_ShenandoahWBMemProj); 1752 1753 fixer->remove(proj); 1754 Node* mem = fixer->find_mem(ctrl, NULL); 1755 1756 assert(!ShenandoahVerifyOptoBarriers || memory_dominates_all_paths(mem, ctrl, alias, phase), "can't fix the memory graph"); 1757 1758 phase->set_ctrl_and_loop(this, ctrl); 1759 phase->igvn().replace_input_of(this, Control, ctrl); 1760 1761 disconnect_barrier_mem(this, phase->igvn()); 1762 1763 phase->igvn().replace_input_of(this, Memory, mem); 1764 phase->set_ctrl_and_loop(proj, ctrl); 1765 1766 fixer->fix_mem(ctrl, ctrl, mem, mem, proj, uses); 1767 assert(proj->outcnt() > 0, "disconnected write barrier"); 1768 } 1769 1770 LoopNode* ShenandoahWriteBarrierNode::try_move_before_pre_loop(Node* c, Node* val_ctrl, PhaseIdealLoop* phase) { 1771 // A write barrier between a pre and main loop can get in the way of 1772 // vectorization. Move it above the pre loop if possible 1773 CountedLoopNode* cl = NULL; 1774 if (c->is_IfFalse() && 1775 c->in(0)->is_CountedLoopEnd()) { 1776 cl = c->in(0)->as_CountedLoopEnd()->loopnode(); 1777 } else if (c->is_IfProj() && 1778 c->in(0)->is_If() && 1779 c->in(0)->in(0)->is_IfFalse() && 1780 c->in(0)->in(0)->in(0)->is_CountedLoopEnd()) { 1781 cl = c->in(0)->in(0)->in(0)->as_CountedLoopEnd()->loopnode(); 1782 } 1783 if (cl != NULL && 1784 cl->is_pre_loop() && 1785 val_ctrl != cl && 1786 phase->is_dominator(val_ctrl, cl)) { 1787 return cl; 1788 } 1789 return NULL; 1790 } 1791 1792 void ShenandoahWriteBarrierNode::try_move_before_loop(GrowableArray<MemoryGraphFixer*>& memory_graph_fixers, PhaseIdealLoop* phase, bool include_lsm, Unique_Node_List& uses) { 1793 Node *n_ctrl = phase->get_ctrl(this); 1794 IdealLoopTree *n_loop = phase->get_loop(n_ctrl); 1795 Node* val = in(ValueIn); 1796 Node* val_ctrl = phase->get_ctrl(val); 1797 if (n_loop != phase->ltree_root() && !n_loop->_irreducible) { 1798 IdealLoopTree *val_loop = phase->get_loop(val_ctrl); 1799 Node* mem = in(Memory); 1800 IdealLoopTree *mem_loop = phase->get_loop(phase->get_ctrl(mem)); 1801 if (!n_loop->is_member(val_loop) && 1802 n_loop->is_member(mem_loop)) { 1803 Node* n_loop_head = n_loop->_head; 1804 1805 if (n_loop_head->is_Loop()) { 1806 LoopNode* loop = n_loop_head->as_Loop(); 1807 if (n_loop_head->is_CountedLoop() && n_loop_head->as_CountedLoop()->is_main_loop()) { 1808 LoopNode* res = try_move_before_pre_loop(n_loop_head->in(LoopNode::EntryControl), val_ctrl, phase); 1809 if (res != NULL) { 1810 loop = res; 1811 } 1812 } 1813 1814 try_move_before_loop_helper(loop, val_ctrl, memory_graph_fixers, phase, include_lsm, uses); 1815 } 1816 } 1817 } 1818 LoopNode* ctrl = try_move_before_pre_loop(in(0), val_ctrl, phase); 1819 if (ctrl != NULL) { 1820 try_move_before_loop_helper(ctrl, val_ctrl, memory_graph_fixers, phase, include_lsm, uses); 1821 } 1822 } 1823 1824 Node* ShenandoahWriteBarrierNode::would_subsume(ShenandoahBarrierNode* other, PhaseIdealLoop* phase) { 1825 Node* val = in(ValueIn); 1826 Node* val_ctrl = phase->get_ctrl(val); 1827 Node* other_mem = other->in(Memory); 1828 Node* other_ctrl = phase->get_ctrl(other); 1829 Node* this_ctrl = phase->get_ctrl(this); 1830 IdealLoopTree* this_loop = phase->get_loop(this_ctrl); 1831 IdealLoopTree* other_loop = phase->get_loop(other_ctrl); 1832 1833 Node* ctrl = phase->dom_lca(other_ctrl, this_ctrl); 1834 1835 if (ctrl->is_Proj() && 1836 ctrl->in(0)->is_Call() && 1837 ctrl->unique_ctrl_out() != NULL && 1838 ctrl->unique_ctrl_out()->Opcode() == Op_Catch && 1839 !phase->is_dominator(val_ctrl, ctrl->in(0)->in(0))) { 1840 return NULL; 1841 } 1842 1843 IdealLoopTree* loop = phase->get_loop(ctrl); 1844 1845 // We don't want to move a write barrier in a loop 1846 // If the LCA is in a inner loop, try a control out of loop if possible 1847 while (!loop->is_member(this_loop) && (other->Opcode() != Op_ShenandoahWriteBarrier || !loop->is_member(other_loop))) { 1848 ctrl = phase->idom(ctrl); 1849 if (ctrl->is_MultiBranch()) { 1850 ctrl = ctrl->in(0); 1851 } 1852 if (ctrl != val_ctrl && phase->is_dominator(ctrl, val_ctrl)) { 1853 return NULL; 1854 } 1855 loop = phase->get_loop(ctrl); 1856 } 1857 1858 if (ShenandoahDontIncreaseWBFreq) { 1859 Node* this_iffproj = no_branches(this_ctrl, ctrl, true, phase); 1860 if (other->Opcode() == Op_ShenandoahWriteBarrier) { 1861 Node* other_iffproj = no_branches(other_ctrl, ctrl, true, phase); 1862 if (other_iffproj == NULL || this_iffproj == NULL) { 1863 return ctrl; 1864 } else if (other_iffproj != NodeSentinel && this_iffproj != NodeSentinel && 1865 other_iffproj->in(0) == this_iffproj->in(0)) { 1866 return ctrl; 1867 } 1868 } else if (this_iffproj == NULL) { 1869 return ctrl; 1870 } 1871 return NULL; 1872 } 1873 1874 return ctrl; 1875 } 1876 1877 void ShenandoahWriteBarrierNode::optimize_before_expansion(PhaseIdealLoop* phase, GrowableArray<MemoryGraphFixer*> memory_graph_fixers, bool include_lsm) { 1878 bool progress = false; 1879 Unique_Node_List uses; 1880 do { 1881 progress = false; 1882 for (int i = 0; i < ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count(); i++) { 1883 ShenandoahWriteBarrierNode* wb = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barrier(i); 1884 1885 wb->try_move_before_loop(memory_graph_fixers, phase, include_lsm, uses); 1886 1887 Node* val = wb->in(ValueIn); 1888 1889 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) { 1890 Node* u = val->fast_out(j); 1891 if (u != wb && u->is_ShenandoahBarrier()) { 1892 Node* rep_ctrl = wb->would_subsume(u->as_ShenandoahBarrier(), phase); 1893 1894 if (rep_ctrl != NULL) { 1895 Node* other = u; 1896 Node* val_ctrl = phase->get_ctrl(val); 1897 if (rep_ctrl->is_Proj() && 1898 rep_ctrl->in(0)->is_Call() && 1899 rep_ctrl->unique_ctrl_out() != NULL && 1900 rep_ctrl->unique_ctrl_out()->Opcode() == Op_Catch) { 1901 rep_ctrl = rep_ctrl->in(0)->in(0); 1902 1903 assert(phase->is_dominator(val_ctrl, rep_ctrl), "bad control"); 1904 } else { 1905 LoopNode* c = ShenandoahWriteBarrierNode::try_move_before_pre_loop(rep_ctrl, val_ctrl, phase); 1906 if (c != NULL) { 1907 rep_ctrl = ShenandoahWriteBarrierNode::move_above_predicates(c, val_ctrl, phase); 1908 } else { 1909 while (rep_ctrl->is_IfProj()) { 1910 CallStaticJavaNode* unc = rep_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 1911 if (unc != NULL) { 1912 int req = unc->uncommon_trap_request(); 1913 Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req); 1914 if ((trap_reason == Deoptimization::Reason_loop_limit_check || 1915 trap_reason == Deoptimization::Reason_predicate || 1916 trap_reason == Deoptimization::Reason_profile_predicate) && 1917 phase->is_dominator(val_ctrl, rep_ctrl->in(0)->in(0))) { 1918 rep_ctrl = rep_ctrl->in(0)->in(0); 1919 continue; 1920 } 1921 } 1922 break; 1923 } 1924 } 1925 } 1926 1927 Node* wb_ctrl = phase->get_ctrl(wb); 1928 Node* other_ctrl = phase->get_ctrl(other); 1929 int alias = phase->C->get_alias_index(wb->adr_type()); 1930 MemoryGraphFixer* fixer = find_fixer(memory_graph_fixers, alias);; 1931 if (!is_dominator(wb_ctrl, other_ctrl, wb, other, phase)) { 1932 if (fixer == NULL) { 1933 fixer = create_fixer(memory_graph_fixers, alias, phase, include_lsm); 1934 } 1935 Node* mem = fixer->find_mem(rep_ctrl, phase->get_ctrl(other) == rep_ctrl ? other : NULL); 1936 1937 if (mem->has_out_with(Op_Lock) || mem->has_out_with(Op_Unlock)) { 1938 continue; 1939 } 1940 1941 Node* wb_proj = wb->find_out_with(Op_ShenandoahWBMemProj); 1942 fixer->remove(wb_proj); 1943 Node* mem_for_ctrl = fixer->find_mem(rep_ctrl, NULL); 1944 1945 if (wb->in(Memory) != mem) { 1946 disconnect_barrier_mem(wb, phase->igvn()); 1947 phase->igvn().replace_input_of(wb, Memory, mem); 1948 } 1949 if (rep_ctrl != wb_ctrl) { 1950 phase->set_ctrl_and_loop(wb, rep_ctrl); 1951 phase->igvn().replace_input_of(wb, Control, rep_ctrl); 1952 phase->set_ctrl_and_loop(wb_proj, rep_ctrl); 1953 progress = true; 1954 } 1955 1956 fixer->fix_mem(rep_ctrl, rep_ctrl, mem, mem_for_ctrl, wb_proj, uses); 1957 1958 assert(!ShenandoahVerifyOptoBarriers || ShenandoahWriteBarrierNode::memory_dominates_all_paths(mem, rep_ctrl, alias, phase), "can't fix the memory graph"); 1959 } 1960 1961 if (other->Opcode() == Op_ShenandoahWriteBarrier) { 1962 Node* other_proj = other->find_out_with(Op_ShenandoahWBMemProj); 1963 if (fixer != NULL) { 1964 fixer->remove(other_proj); 1965 } 1966 phase->igvn().replace_node(other_proj, other->in(Memory)); 1967 } 1968 phase->igvn().replace_node(other, wb); 1969 --j; --jmax; 1970 } 1971 } 1972 } 1973 } 1974 } while(progress); 1975 } 1976 1977 void ShenandoahReadBarrierNode::try_move(Node *n_ctrl, PhaseIdealLoop* phase) { 1978 Node* mem = in(MemNode::Memory); 1979 int alias = phase->C->get_alias_index(adr_type()); 1980 const bool trace = false; 1981 1982 #ifdef ASSERT 1983 if (trace) { tty->print("Trying to move mem of"); dump(); } 1984 #endif 1985 1986 Node* new_mem = mem; 1987 1988 ResourceMark rm; 1989 VectorSet seen(Thread::current()->resource_area()); 1990 Node_List phis; 1991 1992 for (;;) { 1993 #ifdef ASSERT 1994 if (trace) { tty->print("Looking for dominator from"); mem->dump(); } 1995 #endif 1996 if (mem->is_Proj() && mem->in(0)->is_Start()) { 1997 if (new_mem != in(MemNode::Memory)) { 1998 #ifdef ASSERT 1999 if (trace) { tty->print("XXX Setting mem to"); new_mem->dump(); tty->print(" for "); dump(); } 2000 #endif 2001 phase->igvn().replace_input_of(this, MemNode::Memory, new_mem); 2002 } 2003 return; 2004 } 2005 2006 Node* candidate = mem; 2007 do { 2008 if (!is_independent(mem)) { 2009 if (trace) { tty->print_cr("Not independent"); } 2010 if (new_mem != in(MemNode::Memory)) { 2011 #ifdef ASSERT 2012 if (trace) { tty->print("XXX Setting mem to"); new_mem->dump(); tty->print(" for "); dump(); } 2013 #endif 2014 phase->igvn().replace_input_of(this, MemNode::Memory, new_mem); 2015 } 2016 return; 2017 } 2018 if (seen.test_set(mem->_idx)) { 2019 if (trace) { tty->print_cr("Already seen"); } 2020 ShouldNotReachHere(); 2021 // Strange graph 2022 if (new_mem != in(MemNode::Memory)) { 2023 #ifdef ASSERT 2024 if (trace) { tty->print("XXX Setting mem to"); new_mem->dump(); tty->print(" for "); dump(); } 2025 #endif 2026 phase->igvn().replace_input_of(this, MemNode::Memory, new_mem); 2027 } 2028 return; 2029 } 2030 if (mem->is_Phi()) { 2031 phis.push(mem); 2032 } 2033 mem = next_mem(mem, alias); 2034 if (mem->bottom_type() == Type::MEMORY) { 2035 candidate = mem; 2036 } 2037 assert(is_dominator(phase->ctrl_or_self(mem), n_ctrl, mem, this, phase) == phase->is_dominator(phase->ctrl_or_self(mem), n_ctrl), "strange dominator"); 2038 #ifdef ASSERT 2039 if (trace) { tty->print("Next mem is"); mem->dump(); } 2040 #endif 2041 } while (mem->bottom_type() != Type::MEMORY || !phase->is_dominator(phase->ctrl_or_self(mem), n_ctrl)); 2042 2043 assert(mem->bottom_type() == Type::MEMORY, "bad mem"); 2044 2045 bool not_dom = false; 2046 for (uint i = 0; i < phis.size() && !not_dom; i++) { 2047 Node* nn = phis.at(i); 2048 2049 #ifdef ASSERT 2050 if (trace) { tty->print("Looking from phi"); nn->dump(); } 2051 #endif 2052 assert(nn->is_Phi(), "phis only"); 2053 for (uint j = 2; j < nn->req() && !not_dom; j++) { 2054 Node* m = nn->in(j); 2055 #ifdef ASSERT 2056 if (trace) { tty->print("Input %d is", j); m->dump(); } 2057 #endif 2058 while (m != mem && !seen.test_set(m->_idx)) { 2059 if (is_dominator(phase->ctrl_or_self(m), phase->ctrl_or_self(mem), m, mem, phase)) { 2060 not_dom = true; 2061 // Scheduling anomaly 2062 #ifdef ASSERT 2063 if (trace) { tty->print("Giving up"); m->dump(); } 2064 #endif 2065 break; 2066 } 2067 if (!is_independent(m)) { 2068 if (trace) { tty->print_cr("Not independent"); } 2069 if (new_mem != in(MemNode::Memory)) { 2070 #ifdef ASSERT 2071 if (trace) { tty->print("XXX Setting mem to"); new_mem->dump(); tty->print(" for "); dump(); } 2072 #endif 2073 phase->igvn().replace_input_of(this, MemNode::Memory, new_mem); 2074 } 2075 return; 2076 } 2077 if (m->is_Phi()) { 2078 phis.push(m); 2079 } 2080 m = next_mem(m, alias); 2081 #ifdef ASSERT 2082 if (trace) { tty->print("Next mem is"); m->dump(); } 2083 #endif 2084 } 2085 } 2086 } 2087 if (!not_dom) { 2088 new_mem = mem; 2089 phis.clear(); 2090 } else { 2091 seen.Clear(); 2092 } 2093 } 2094 } 2095 2096 CallStaticJavaNode* ShenandoahWriteBarrierNode::pin_and_expand_null_check(PhaseIterGVN& igvn) { 2097 Node* val = in(ValueIn); 2098 2099 const Type* val_t = igvn.type(val); 2100 2101 if (val_t->meet(TypePtr::NULL_PTR) != val_t && 2102 val->Opcode() == Op_CastPP && 2103 val->in(0) != NULL && 2104 val->in(0)->Opcode() == Op_IfTrue && 2105 val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) && 2106 val->in(0)->in(0)->is_If() && 2107 val->in(0)->in(0)->in(1)->Opcode() == Op_Bool && 2108 val->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne && 2109 val->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP && 2110 val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1) && 2111 val->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) { 2112 assert(val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1), ""); 2113 CallStaticJavaNode* unc = val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 2114 return unc; 2115 } 2116 return NULL; 2117 } 2118 2119 void ShenandoahWriteBarrierNode::pin_and_expand_move_barrier(PhaseIdealLoop* phase, GrowableArray<MemoryGraphFixer*>& memory_graph_fixers, Unique_Node_List& uses) { 2120 Node* unc = pin_and_expand_null_check(phase->igvn()); 2121 Node* val = in(ValueIn); 2122 2123 if (unc != NULL) { 2124 Node* ctrl = phase->get_ctrl(this); 2125 Node* unc_ctrl = val->in(0); 2126 2127 // Don't move write barrier in a loop 2128 IdealLoopTree* loop = phase->get_loop(ctrl); 2129 IdealLoopTree* unc_loop = phase->get_loop(unc_ctrl); 2130 2131 if (!unc_loop->is_member(loop)) { 2132 return; 2133 } 2134 2135 Node* branch = no_branches(ctrl, unc_ctrl, false, phase); 2136 assert(branch == NULL || branch == NodeSentinel, "was not looking for a branch"); 2137 if (branch == NodeSentinel) { 2138 return; 2139 } 2140 2141 2142 RegionNode* r = new RegionNode(3); 2143 IfNode* iff = unc_ctrl->in(0)->as_If(); 2144 2145 Node* ctrl_use = unc_ctrl->unique_ctrl_out(); 2146 Node* unc_ctrl_clone = unc_ctrl->clone(); 2147 phase->register_control(unc_ctrl_clone, loop, iff); 2148 Node* c = unc_ctrl_clone; 2149 Node* new_cast = clone_null_check(c, val, unc_ctrl_clone, r, 1, phase); 2150 2151 phase->igvn().replace_input_of(unc_ctrl, 0, c->in(0)); 2152 phase->set_idom(unc_ctrl, c->in(0), phase->dom_depth(unc_ctrl)); 2153 phase->lazy_replace(c, unc_ctrl); 2154 c = NULL;; 2155 phase->igvn().replace_input_of(val, 0, unc_ctrl_clone); 2156 phase->set_ctrl(val, unc_ctrl_clone); 2157 2158 IfNode* new_iff = new_cast->in(0)->in(0)->as_If(); 2159 fix_null_check(iff, unc, unc_ctrl_clone, r, uses, phase); 2160 Node* iff_proj = iff->proj_out(0); 2161 r->init_req(2, iff_proj); 2162 2163 Node* new_bol = new_iff->in(1)->clone(); 2164 Node* new_cmp = new_bol->in(1)->clone(); 2165 assert(new_cmp->Opcode() == Op_CmpP, "broken"); 2166 assert(new_cmp->in(1) == val->in(1), "broken"); 2167 new_bol->set_req(1, new_cmp); 2168 new_cmp->set_req(1, this); 2169 phase->register_new_node(new_bol, new_iff->in(0)); 2170 phase->register_new_node(new_cmp, new_iff->in(0)); 2171 phase->igvn().replace_input_of(new_iff, 1, new_bol); 2172 phase->igvn().replace_input_of(new_cast, 1, this); 2173 2174 for (DUIterator_Fast imax, i = this->fast_outs(imax); i < imax; i++) { 2175 Node* u = this->fast_out(i); 2176 if (u == new_cast || u->Opcode() == Op_ShenandoahWBMemProj || u == new_cmp) { 2177 continue; 2178 } 2179 phase->igvn().rehash_node_delayed(u); 2180 int nb = u->replace_edge(this, new_cast); 2181 assert(nb > 0, "no update?"); 2182 --i; imax -= nb; 2183 } 2184 2185 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) { 2186 Node* u = val->fast_out(i); 2187 if (u == this) { 2188 continue; 2189 } 2190 phase->igvn().rehash_node_delayed(u); 2191 int nb = u->replace_edge(val, new_cast); 2192 assert(nb > 0, "no update?"); 2193 --i; imax -= nb; 2194 } 2195 2196 Node* new_ctrl = unc_ctrl_clone; 2197 2198 int alias = phase->C->get_alias_index(adr_type()); 2199 MemoryGraphFixer* fixer = find_fixer(memory_graph_fixers, alias); 2200 if (fixer == NULL) { 2201 fixer = create_fixer(memory_graph_fixers, alias, phase, true); 2202 } 2203 2204 Node* proj = find_out_with(Op_ShenandoahWBMemProj); 2205 fixer->remove(proj); 2206 Node* mem = fixer->find_mem(new_ctrl, NULL); 2207 2208 if (in(Memory) != mem) { 2209 disconnect_barrier_mem(this, phase->igvn()); 2210 phase->igvn().replace_input_of(this, Memory, mem); 2211 } 2212 2213 phase->set_ctrl_and_loop(this, new_ctrl); 2214 phase->igvn().replace_input_of(this, Control, new_ctrl); 2215 phase->set_ctrl_and_loop(proj, new_ctrl); 2216 2217 fixer->fix_mem(new_ctrl, new_ctrl, mem, mem, proj, uses); 2218 } 2219 } 2220 2221 void ShenandoahWriteBarrierNode::pin_and_expand_helper(PhaseIdealLoop* phase) { 2222 Node* val = in(ValueIn); 2223 CallStaticJavaNode* unc = pin_and_expand_null_check(phase->igvn()); 2224 Node* rep = this; 2225 Node* ctrl = phase->get_ctrl(this); 2226 if (unc != NULL && val->in(0) == ctrl) { 2227 Node* unc_ctrl = val->in(0); 2228 IfNode* other_iff = unc_ctrl->unique_ctrl_out()->as_If(); 2229 ProjNode* other_unc_ctrl = other_iff->proj_out(1); 2230 Node* cast = NULL; 2231 for (DUIterator_Fast imax, i = other_unc_ctrl->fast_outs(imax); i < imax && cast == NULL; i++) { 2232 Node* u = other_unc_ctrl->fast_out(i); 2233 if (u->Opcode() == Op_CastPP && u->in(1) == this) { 2234 cast = u; 2235 } 2236 } 2237 assert(other_unc_ctrl->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) == unc, "broken"); 2238 rep = cast; 2239 } 2240 2241 // Replace all uses of barrier's input that are dominated by ctrl 2242 // with the value returned by the barrier: no need to keep both 2243 // live. 2244 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) { 2245 Node* u = val->fast_out(i); 2246 if (u != this) { 2247 if (u->is_Phi()) { 2248 int nb = 0; 2249 for (uint j = 1; j < u->req(); j++) { 2250 if (u->in(j) == val) { 2251 Node* c = u->in(0)->in(j); 2252 if (phase->is_dominator(ctrl, c)) { 2253 phase->igvn().replace_input_of(u, j, rep); 2254 nb++; 2255 } 2256 } 2257 } 2258 if (nb > 0) { 2259 imax -= nb; 2260 --i; 2261 } 2262 } else { 2263 Node* c = phase->ctrl_or_self(u); 2264 if (is_dominator(ctrl, c, this, u, phase)) { 2265 phase->igvn().rehash_node_delayed(u); 2266 int nb = u->replace_edge(val, rep); 2267 assert(nb > 0, "no update?"); 2268 --i, imax -= nb; 2269 } 2270 } 2271 } 2272 } 2273 } 2274 2275 Node* ShenandoahWriteBarrierNode::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) { 2276 Node* mem = NULL; 2277 Node* c = ctrl; 2278 do { 2279 if (c->is_Region()) { 2280 Node* phi_bottom = NULL; 2281 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == NULL; i++) { 2282 Node* u = c->fast_out(i); 2283 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) { 2284 if (u->adr_type() == TypePtr::BOTTOM) { 2285 mem = u; 2286 } 2287 } 2288 } 2289 } else { 2290 if (c->is_Call() && c->as_Call()->adr_type() != NULL) { 2291 CallProjections projs; 2292 c->as_Call()->extract_projections(&projs, true, false); 2293 if (projs.fallthrough_memproj != NULL) { 2294 if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) { 2295 if (projs.catchall_memproj == NULL) { 2296 mem = projs.fallthrough_memproj; 2297 } else { 2298 if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) { 2299 mem = projs.fallthrough_memproj; 2300 } else { 2301 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier"); 2302 mem = projs.catchall_memproj; 2303 } 2304 } 2305 } 2306 } else { 2307 Node* proj = c->as_Call()->proj_out(TypeFunc::Memory); 2308 if (proj != NULL && 2309 proj->adr_type() == TypePtr::BOTTOM) { 2310 mem = proj; 2311 } 2312 } 2313 } else { 2314 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) { 2315 Node* u = c->fast_out(i); 2316 if (u->is_Proj() && 2317 u->bottom_type() == Type::MEMORY && 2318 u->adr_type() == TypePtr::BOTTOM) { 2319 assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), ""); 2320 assert(mem == NULL, "only one proj"); 2321 mem = u; 2322 } 2323 } 2324 assert(!c->is_Call() || c->as_Call()->adr_type() != NULL || mem == NULL, "no mem projection expected"); 2325 } 2326 } 2327 c = phase->idom(c); 2328 } while (mem == NULL); 2329 return mem; 2330 } 2331 2332 void ShenandoahWriteBarrierNode::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) { 2333 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2334 Node* u = n->fast_out(i); 2335 if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) { 2336 uses.push(u); 2337 } 2338 } 2339 } 2340 2341 void ShenandoahWriteBarrierNode::test_heap_stable(Node* ctrl, Node* raw_mem, Node*& gc_state, Node*& heap_stable, 2342 Node*& heap_not_stable, PhaseIdealLoop* phase) { 2343 IdealLoopTree *loop = phase->get_loop(ctrl); 2344 Node* thread = new ThreadLocalNode(); 2345 phase->register_new_node(thread, ctrl); 2346 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())); 2347 phase->set_ctrl(offset, phase->C->root()); 2348 Node* gc_state_addr = new AddPNode(phase->C->top(), thread, offset); 2349 phase->register_new_node(gc_state_addr, ctrl); 2350 uint gc_state_idx = Compile::AliasIdxRaw; 2351 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument 2352 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx)); 2353 2354 gc_state = new LoadBNode(ctrl, raw_mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered); 2355 phase->register_new_node(gc_state, ctrl); 2356 Node* heap_stable_and = new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::HAS_FORWARDED)); 2357 phase->register_new_node(heap_stable_and, ctrl); 2358 Node* heap_stable_cmp = new CmpINode(heap_stable_and, phase->igvn().zerocon(T_INT)); 2359 phase->register_new_node(heap_stable_cmp, ctrl); 2360 Node* heap_stable_test = new BoolNode(heap_stable_cmp, BoolTest::ne); 2361 phase->register_new_node(heap_stable_test, ctrl); 2362 IfNode* heap_stable_iff = new IfNode(ctrl, heap_stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN); 2363 phase->register_control(heap_stable_iff, loop, ctrl); 2364 2365 heap_stable = new IfFalseNode(heap_stable_iff); 2366 phase->register_control(heap_stable, loop, heap_stable_iff); 2367 heap_not_stable = new IfTrueNode(heap_stable_iff); 2368 phase->register_control(heap_not_stable, loop, heap_stable_iff); 2369 2370 assert(is_heap_stable_test(heap_stable_iff), "Should match the shape"); 2371 } 2372 2373 2374 void ShenandoahWriteBarrierNode::test_evacuation_in_progress(Node* ctrl, Node* val, Node*& raw_mem, 2375 Node*& evac_in_progress, Node*& evac_not_in_progress, 2376 Node*& heap_stable, Node*& null_val, 2377 PhaseIdealLoop* phase) { 2378 IdealLoopTree *loop = phase->get_loop(ctrl); 2379 Node* heap_not_stable = NULL; 2380 Node* unused_gc_state = NULL; 2381 2382 test_heap_stable(ctrl, raw_mem, unused_gc_state, heap_stable, heap_not_stable, phase); 2383 2384 ctrl = heap_not_stable; 2385 2386 const Type* val_t = phase->igvn().type(val); 2387 2388 if (val_t->meet(TypePtr::NULL_PTR) == val_t) { 2389 Node* null_cmp = new CmpPNode(val, phase->igvn().zerocon(T_OBJECT)); 2390 phase->register_new_node(null_cmp, ctrl); 2391 Node* null_test = new BoolNode(null_cmp, BoolTest::ne); 2392 phase->register_new_node(null_test, ctrl); 2393 IfNode* null_iff = new IfNode(ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN); 2394 phase->register_control(null_iff, loop, ctrl); 2395 Node* not_null = new IfTrueNode(null_iff); 2396 phase->register_control(not_null, loop, null_iff); 2397 Node* null = new IfFalseNode(null_iff); 2398 phase->register_control(null, loop, null_iff); 2399 null_val = null; 2400 ctrl = not_null; 2401 } 2402 2403 Node* thread = new ThreadLocalNode(); 2404 phase->register_new_node(thread, ctrl); 2405 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())); 2406 phase->set_ctrl(offset, phase->C->root()); 2407 Node* gc_state_addr = new AddPNode(phase->C->top(), thread, offset); 2408 phase->register_new_node(gc_state_addr, ctrl); 2409 uint gc_state_idx = Compile::AliasIdxRaw; 2410 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument 2411 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx)); 2412 2413 Node* gc_state = new LoadBNode(ctrl, raw_mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered); 2414 phase->register_new_node(gc_state, ctrl); 2415 2416 Node* evacuation_in_progress = new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL)); 2417 phase->register_new_node(evacuation_in_progress, ctrl); 2418 Node* evacuation_in_progress_cmp = new CmpINode(evacuation_in_progress, phase->igvn().zerocon(T_INT)); 2419 phase->register_new_node(evacuation_in_progress_cmp, ctrl); 2420 Node* evacuation_in_progress_test = new BoolNode(evacuation_in_progress_cmp, BoolTest::ne); 2421 phase->register_new_node(evacuation_in_progress_test, ctrl); 2422 IfNode* evacuation_iff = new IfNode(ctrl, evacuation_in_progress_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN); 2423 phase->register_control(evacuation_iff, loop, ctrl); 2424 2425 assert(is_evacuation_in_progress_test(evacuation_iff), "Should match the shape"); 2426 assert(is_gc_state_load(gc_state), "Should match the shape"); 2427 2428 evac_not_in_progress = new IfFalseNode(evacuation_iff); 2429 phase->register_control(evac_not_in_progress, loop, evacuation_iff); 2430 evac_in_progress = new IfTrueNode(evacuation_iff); 2431 phase->register_control(evac_in_progress, loop, evacuation_iff); 2432 } 2433 2434 Node* ShenandoahWriteBarrierNode::clone_null_check(Node*& c, Node* val, Node* unc_ctrl, 2435 Node* unc_region, uint input, PhaseIdealLoop* phase) { 2436 IdealLoopTree *loop = phase->get_loop(c); 2437 Node* iff = unc_ctrl->in(0); 2438 assert(iff->is_If(), "broken"); 2439 Node* new_iff = iff->clone(); 2440 new_iff->set_req(0, c); 2441 phase->register_control(new_iff, loop, c); 2442 Node* iffalse = new IfFalseNode(new_iff->as_If()); 2443 phase->register_control(iffalse, loop, new_iff); 2444 Node* iftrue = new IfTrueNode(new_iff->as_If()); 2445 phase->register_control(iftrue, loop, new_iff); 2446 c = iftrue; 2447 const Type *t = phase->igvn().type(val); 2448 assert(val->Opcode() == Op_CastPP, "expect cast to non null here"); 2449 Node* uncasted_val = val->in(1); 2450 val = new CastPPNode(uncasted_val, t); 2451 val->init_req(0, c); 2452 phase->register_new_node(val, c); 2453 unc_region->init_req(input, iffalse); 2454 return val; 2455 } 2456 2457 void ShenandoahWriteBarrierNode::fix_null_check(Node* dom, Node* unc, Node* unc_ctrl, Node* unc_region, 2458 Unique_Node_List& uses, PhaseIdealLoop* phase) { 2459 IfNode* iff = unc_ctrl->in(0)->as_If(); 2460 Node* proj = iff->proj_out(0); 2461 assert(proj != unc_ctrl, "bad projection"); 2462 Node* use = proj->unique_ctrl_out(); 2463 2464 assert(use == unc || use->is_Region(), "what else?"); 2465 2466 uses.clear(); 2467 if (use == unc) { 2468 phase->set_idom(use, unc_region, phase->dom_depth(use)); 2469 for (uint i = 1; i < unc->req(); i++) { 2470 Node* n = unc->in(i); 2471 if (phase->has_ctrl(n) && phase->get_ctrl(n) == proj) { 2472 uses.push(n); 2473 } 2474 } 2475 } else { 2476 assert(use->is_Region(), "what else?"); 2477 uint idx = 1; 2478 for (; use->in(idx) != proj; idx++); 2479 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) { 2480 Node* u = use->fast_out(i); 2481 if (u->is_Phi() && phase->get_ctrl(u->in(idx)) == proj) { 2482 uses.push(u->in(idx)); 2483 } 2484 } 2485 } 2486 for(uint next = 0; next < uses.size(); next++ ) { 2487 Node *n = uses.at(next); 2488 assert(phase->get_ctrl(n) == proj, "bad control"); 2489 phase->set_ctrl_and_loop(n, unc_region); 2490 if (n->in(0) == proj) { 2491 phase->igvn().replace_input_of(n, 0, unc_region); 2492 } 2493 for (uint i = 0; i < n->req(); i++) { 2494 Node* m = n->in(i); 2495 if (m != NULL && phase->has_ctrl(m) && phase->get_ctrl(m) == proj) { 2496 uses.push(m); 2497 } 2498 } 2499 } 2500 2501 phase->igvn().rehash_node_delayed(use); 2502 int nb = use->replace_edge(proj, unc_region); 2503 assert(nb == 1, "only use expected"); 2504 phase->register_control(unc_region, phase->ltree_root(), dom); 2505 } 2506 2507 void ShenandoahWriteBarrierNode::evacuation_not_in_progress_null_check(Node*& c, Node*& val, Node* unc_ctrl, Node*& unc_region, PhaseIdealLoop* phase) { 2508 if (unc_ctrl != NULL) { 2509 // Clone the null check in this branch to allow implicit null check 2510 unc_region = new RegionNode(3); 2511 val = clone_null_check(c, val, unc_ctrl, unc_region, 1, phase); 2512 } 2513 } 2514 2515 void ShenandoahWriteBarrierNode::evacuation_not_in_progress(Node* c, Node* val, Node* unc_ctrl, Node* raw_mem, Node* wb_mem, Node* region, 2516 Node* val_phi, Node* mem_phi, Node* raw_mem_phi, Node*& unc_region, PhaseIdealLoop* phase) { 2517 evacuation_not_in_progress_null_check(c, val, unc_ctrl, unc_region, phase); 2518 region->init_req(1, c); 2519 if (ShenandoahWriteBarrierRB) { 2520 Node* rbfalse = new ShenandoahReadBarrierNode(c, wb_mem, val); 2521 phase->register_new_node(rbfalse, c); 2522 val_phi->init_req(1, rbfalse); 2523 } else { 2524 val_phi->init_req(1, val); 2525 } 2526 mem_phi->init_req(1, wb_mem); 2527 raw_mem_phi->init_req(1, raw_mem); 2528 } 2529 2530 void ShenandoahWriteBarrierNode::heap_stable(Node* c, Node* val, Node* unc_ctrl, Node* raw_mem, Node* wb_mem, Node* region, 2531 Node* val_phi, Node* mem_phi, Node* raw_mem_phi, Node* unc_region, PhaseIdealLoop* phase) { 2532 region->init_req(1, c); 2533 if (unc_ctrl != NULL) { 2534 val = val->in(1); 2535 } 2536 val_phi->init_req(1, val); 2537 mem_phi->init_req(1, wb_mem); 2538 raw_mem_phi->init_req(1, raw_mem); 2539 } 2540 2541 void ShenandoahWriteBarrierNode::evacuation_in_progress_null_check(Node*& c, Node*& val, Node* evacuation_iff, Node* unc, Node* unc_ctrl, 2542 Node* unc_region, Unique_Node_List& uses, PhaseIdealLoop* phase) { 2543 if (unc != NULL) { 2544 // Clone the null check in this branch to allow implicit null check 2545 val = clone_null_check(c, val, unc_ctrl, unc_region, 2, phase); 2546 2547 fix_null_check(evacuation_iff, unc, unc_ctrl, unc_region, uses, phase); 2548 2549 IfNode* iff = unc_ctrl->in(0)->as_If(); 2550 phase->igvn().replace_input_of(iff, 1, phase->igvn().intcon(1)); 2551 } 2552 } 2553 2554 void ShenandoahWriteBarrierNode::in_cset_fast_test(Node*& c, Node* rbtrue, Node* raw_mem, Node* wb_mem, Node* region, Node* val_phi, Node* mem_phi, 2555 Node* raw_mem_phi, PhaseIdealLoop* phase) { 2556 if (ShenandoahWriteBarrierCsetTestInIR) { 2557 IdealLoopTree *loop = phase->get_loop(c); 2558 Node* raw_rbtrue = new CastP2XNode(c, rbtrue); 2559 phase->register_new_node(raw_rbtrue, c); 2560 Node* cset_offset = new URShiftXNode(raw_rbtrue, phase->igvn().intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint())); 2561 phase->register_new_node(cset_offset, c); 2562 Node* in_cset_fast_test_base_addr = phase->igvn().makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr())); 2563 phase->set_ctrl(in_cset_fast_test_base_addr, phase->C->root()); 2564 Node* in_cset_fast_test_adr = new AddPNode(phase->C->top(), in_cset_fast_test_base_addr, cset_offset); 2565 phase->register_new_node(in_cset_fast_test_adr, c); 2566 uint in_cset_fast_test_idx = Compile::AliasIdxRaw; 2567 const TypePtr* in_cset_fast_test_adr_type = NULL; // debug-mode-only argument 2568 debug_only(in_cset_fast_test_adr_type = phase->C->get_adr_type(in_cset_fast_test_idx)); 2569 Node* in_cset_fast_test_load = new LoadBNode(c, raw_mem, in_cset_fast_test_adr, in_cset_fast_test_adr_type, TypeInt::BYTE, MemNode::unordered); 2570 phase->register_new_node(in_cset_fast_test_load, c); 2571 Node* in_cset_fast_test_cmp = new CmpINode(in_cset_fast_test_load, phase->igvn().zerocon(T_INT)); 2572 phase->register_new_node(in_cset_fast_test_cmp, c); 2573 Node* in_cset_fast_test_test = new BoolNode(in_cset_fast_test_cmp, BoolTest::ne); 2574 phase->register_new_node(in_cset_fast_test_test, c); 2575 IfNode* in_cset_fast_test_iff = new IfNode(c, in_cset_fast_test_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN); 2576 phase->register_control(in_cset_fast_test_iff, loop, c); 2577 2578 Node* in_cset_fast_test_success = new IfFalseNode(in_cset_fast_test_iff); 2579 phase->register_control(in_cset_fast_test_success, loop, in_cset_fast_test_iff); 2580 2581 region->init_req(3, in_cset_fast_test_success); 2582 val_phi->init_req(3, rbtrue); 2583 mem_phi->init_req(3, wb_mem); 2584 raw_mem_phi->init_req(3, raw_mem); 2585 2586 Node* in_cset_fast_test_failure = new IfTrueNode(in_cset_fast_test_iff); 2587 phase->register_control(in_cset_fast_test_failure, loop, in_cset_fast_test_iff); 2588 2589 c = in_cset_fast_test_failure; 2590 } 2591 } 2592 2593 void ShenandoahWriteBarrierNode::evacuation_in_progress(Node* c, Node* val, Node* evacuation_iff, Node* unc, Node* unc_ctrl, 2594 Node* raw_mem, Node* wb_mem, Node* region, Node* val_phi, Node* mem_phi, 2595 Node* raw_mem_phi, Node* unc_region, int alias, Unique_Node_List& uses, 2596 PhaseIdealLoop* phase) { 2597 evacuation_in_progress_null_check(c, val, evacuation_iff, unc, unc_ctrl, unc_region, uses, phase); 2598 2599 IdealLoopTree *loop = phase->get_loop(c); 2600 2601 // Important to perform resolve here, before doing cset check, because that would 2602 // capture forwarded objects we do not need to evacuate again. 2603 Node* rbtrue = new ShenandoahReadBarrierNode(c, wb_mem, val); 2604 phase->register_new_node(rbtrue, c); 2605 2606 in_cset_fast_test(c, rbtrue, raw_mem, wb_mem, region, val_phi, mem_phi, raw_mem_phi, phase); 2607 2608 // The slow path stub consumes and produces raw memory in addition 2609 // to the existing memory edges 2610 Node* base = find_bottom_mem(c, phase); 2611 2612 MergeMemNode* mm = MergeMemNode::make(base); 2613 mm->set_memory_at(alias, wb_mem); 2614 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem); 2615 phase->register_new_node(mm, c); 2616 2617 Node* call = new CallLeafNoFPNode(ShenandoahBarrierSetC2::shenandoah_write_barrier_Type(), ShenandoahBarrierSetAssembler::shenandoah_wb_C(), "shenandoah_write_barrier", TypeRawPtr::BOTTOM); 2618 call->init_req(TypeFunc::Control, c); 2619 call->init_req(TypeFunc::I_O, phase->C->top()); 2620 call->init_req(TypeFunc::Memory, mm); 2621 call->init_req(TypeFunc::FramePtr, phase->C->top()); 2622 call->init_req(TypeFunc::ReturnAdr, phase->C->top()); 2623 call->init_req(TypeFunc::Parms, rbtrue); 2624 phase->register_control(call, loop, c); 2625 Node* ctrl_proj = new ProjNode(call, TypeFunc::Control); 2626 phase->register_control(ctrl_proj, loop, call); 2627 Node* mem_proj = new ProjNode(call, TypeFunc::Memory); 2628 phase->register_new_node(mem_proj, call); 2629 Node* res_proj = new ProjNode(call, TypeFunc::Parms); 2630 phase->register_new_node(res_proj, call); 2631 Node* res = new CheckCastPPNode(ctrl_proj, res_proj, phase->igvn().type(val)->is_oopptr()->cast_to_nonconst()); 2632 phase->register_new_node(res, ctrl_proj); 2633 region->init_req(2, ctrl_proj); 2634 val_phi->init_req(2, res); 2635 mem_phi->init_req(2, mem_proj); 2636 raw_mem_phi->init_req(2, mem_proj); 2637 } 2638 2639 void ShenandoahWriteBarrierNode::fix_ctrl(Node* barrier, Node* region, const MemoryGraphFixer& fixer, Unique_Node_List& uses, Unique_Node_List& uses_to_ignore, uint last, PhaseIdealLoop* phase) { 2640 Node* ctrl = phase->get_ctrl(barrier); 2641 Node* init_raw_mem = fixer.find_mem(ctrl, barrier); 2642 2643 // Update the control of all nodes that should be after the 2644 // barrier control flow 2645 uses.clear(); 2646 // Every node that is control dependent on the barrier's input 2647 // control will be after the expanded barrier. The raw memory (if 2648 // its memory is control dependent on the barrier's input control) 2649 // must stay above the barrier. 2650 uses_to_ignore.clear(); 2651 if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) { 2652 uses_to_ignore.push(init_raw_mem); 2653 } 2654 for (uint next = 0; next < uses_to_ignore.size(); next++) { 2655 Node *n = uses_to_ignore.at(next); 2656 for (uint i = 0; i < n->req(); i++) { 2657 Node* in = n->in(i); 2658 if (in != NULL && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) { 2659 uses_to_ignore.push(in); 2660 } 2661 } 2662 } 2663 for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) { 2664 Node* u = ctrl->fast_out(i); 2665 if (u->_idx < last && 2666 u != barrier && 2667 !uses_to_ignore.member(u) && 2668 (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) && 2669 (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) { 2670 Node* old_c = phase->ctrl_or_self(u); 2671 Node* c = old_c; 2672 if (c != ctrl || 2673 is_dominator_same_ctrl(old_c, barrier, u, phase) || 2674 ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) { 2675 phase->igvn().rehash_node_delayed(u); 2676 int nb = u->replace_edge(ctrl, region); 2677 if (u->is_CFG()) { 2678 if (phase->idom(u) == ctrl) { 2679 phase->set_idom(u, region, phase->dom_depth(region)); 2680 } 2681 } else if (phase->get_ctrl(u) == ctrl) { 2682 assert(u != init_raw_mem, "should leave input raw mem above the barrier"); 2683 uses.push(u); 2684 } 2685 assert(nb == 1, "more than 1 ctrl input?"); 2686 --i, imax -= nb; 2687 } 2688 } 2689 } 2690 } 2691 2692 2693 void ShenandoahWriteBarrierNode::pin_and_expand(PhaseIdealLoop* phase) { 2694 Node_List enqueue_barriers; 2695 if (ShenandoahStoreValEnqueueBarrier) { 2696 Unique_Node_List wq; 2697 wq.push(phase->C->root()); 2698 for (uint i = 0; i < wq.size(); i++) { 2699 Node* n = wq.at(i); 2700 if (n->Opcode() == Op_ShenandoahEnqueueBarrier) { 2701 enqueue_barriers.push(n); 2702 } 2703 for (uint i = 0; i < n->req(); i++) { 2704 Node* in = n->in(i); 2705 if (in != NULL) { 2706 wq.push(in); 2707 } 2708 } 2709 } 2710 } 2711 2712 const bool trace = false; 2713 2714 // Collect raw memory state at CFG points in the entire graph and 2715 // record it in memory_nodes. Optimize the raw memory graph in the 2716 // process. Optimizing the memory graph also makes the memory graph 2717 // simpler. 2718 GrowableArray<MemoryGraphFixer*> memory_graph_fixers; 2719 2720 // Let's try to common write barriers again 2721 optimize_before_expansion(phase, memory_graph_fixers, true); 2722 2723 Unique_Node_List uses; 2724 for (int i = 0; i < ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count(); i++) { 2725 ShenandoahWriteBarrierNode* wb = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barrier(i); 2726 Node* ctrl = phase->get_ctrl(wb); 2727 2728 Node* val = wb->in(ValueIn); 2729 if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) { 2730 assert(is_dominator(phase->get_ctrl(val), ctrl->in(0)->in(0), val, ctrl->in(0), phase), "can't move"); 2731 phase->set_ctrl(wb, ctrl->in(0)->in(0)); 2732 } else if (ctrl->is_CallRuntime()) { 2733 assert(is_dominator(phase->get_ctrl(val), ctrl->in(0), val, ctrl, phase), "can't move"); 2734 phase->set_ctrl(wb, ctrl->in(0)); 2735 } 2736 2737 assert(wb->Opcode() == Op_ShenandoahWriteBarrier, "only for write barriers"); 2738 // Look for a null check that dominates this barrier and move the 2739 // barrier right after the null check to enable implicit null 2740 // checks 2741 wb->pin_and_expand_move_barrier(phase, memory_graph_fixers, uses); 2742 2743 wb->pin_and_expand_helper(phase); 2744 } 2745 2746 MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase); 2747 Unique_Node_List uses_to_ignore; 2748 for (uint i = 0; i < enqueue_barriers.size(); i++) { 2749 Node* barrier = enqueue_barriers.at(i); 2750 Node* pre_val = barrier->in(1); 2751 2752 if (phase->igvn().type(pre_val)->higher_equal(TypePtr::NULL_PTR)) { 2753 ShouldNotReachHere(); 2754 continue; 2755 } 2756 2757 Node* ctrl = phase->get_ctrl(barrier); 2758 2759 if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) { 2760 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0)->in(0), pre_val, ctrl->in(0), phase), "can't move"); 2761 ctrl = ctrl->in(0)->in(0); 2762 phase->set_ctrl(barrier, ctrl); 2763 } else if (ctrl->is_CallRuntime()) { 2764 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0), pre_val, ctrl, phase), "can't move"); 2765 ctrl = ctrl->in(0); 2766 phase->set_ctrl(barrier, ctrl); 2767 } 2768 2769 Node* init_ctrl = ctrl; 2770 IdealLoopTree* loop = phase->get_loop(ctrl); 2771 Node* raw_mem = fixer.find_mem(ctrl, barrier); 2772 Node* init_raw_mem = raw_mem; 2773 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL); 2774 Node* evac_in_progress = NULL; 2775 Node* evac_not_in_progress = NULL; 2776 Node* heap_stable = NULL; 2777 Node* null_val = NULL; 2778 uint last = phase->C->unique(); 2779 2780 Node* stable_test_region = new RegionNode(3); 2781 Node* stable_test_phi = PhiNode::make(stable_test_region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM); 2782 Node* unstable_region = new RegionNode(5); 2783 Node* unstable_phi = PhiNode::make(unstable_region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM); 2784 2785 test_evacuation_in_progress(ctrl, pre_val, raw_mem, evac_in_progress, evac_not_in_progress, heap_stable, null_val, phase); 2786 2787 stable_test_region->init_req(1, heap_stable); 2788 stable_test_region->init_req(2, unstable_region); 2789 stable_test_phi->init_req(1, raw_mem); 2790 stable_test_phi->init_req(2, unstable_phi); 2791 2792 if (null_val != NULL) { 2793 unstable_region->init_req(1, null_val); 2794 unstable_phi->init_req(1, raw_mem); 2795 } 2796 unstable_region->init_req(2, evac_not_in_progress); 2797 unstable_phi->init_req(2, raw_mem); 2798 2799 ctrl = evac_in_progress; 2800 2801 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()); 2802 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()); 2803 Node* thread = new ThreadLocalNode(); 2804 phase->register_new_node(thread, ctrl); 2805 Node* buffer_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(buffer_offset)); 2806 phase->register_new_node(buffer_adr, ctrl); 2807 Node* index_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(index_offset)); 2808 phase->register_new_node(index_adr, ctrl); 2809 2810 BasicType index_bt = TypeX_X->basic_type(); 2811 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 2812 const TypePtr* adr_type = TypeRawPtr::BOTTOM; 2813 Node* index = new LoadXNode(ctrl, raw_mem, index_adr, adr_type, TypeX_X, MemNode::unordered); 2814 phase->register_new_node(index, ctrl); 2815 Node* index_cmp = new CmpXNode(index, phase->igvn().MakeConX(0)); 2816 phase->register_new_node(index_cmp, ctrl); 2817 Node* index_test = new BoolNode(index_cmp, BoolTest::ne); 2818 phase->register_new_node(index_test, ctrl); 2819 IfNode* queue_full_iff = new IfNode(ctrl, index_test, PROB_LIKELY(0.999), COUNT_UNKNOWN); 2820 phase->register_control(queue_full_iff, loop, ctrl); 2821 Node* not_full = new IfTrueNode(queue_full_iff); 2822 phase->register_control(not_full, loop, queue_full_iff); 2823 Node* full = new IfFalseNode(queue_full_iff); 2824 phase->register_control(full, loop, queue_full_iff); 2825 2826 ctrl = not_full; 2827 2828 Node* next_index = new SubXNode(index, phase->igvn().MakeConX(sizeof(intptr_t))); 2829 phase->register_new_node(next_index, ctrl); 2830 2831 Node* buffer = new LoadPNode(ctrl, raw_mem, buffer_adr, adr_type, TypeRawPtr::NOTNULL, MemNode::unordered); 2832 phase->register_new_node(buffer, ctrl); 2833 Node *log_addr = new AddPNode(phase->C->top(), buffer, next_index); 2834 phase->register_new_node(log_addr, ctrl); 2835 Node* log_store = new StorePNode(ctrl, raw_mem, log_addr, adr_type, pre_val, MemNode::unordered); 2836 phase->register_new_node(log_store, ctrl); 2837 // update the index 2838 Node* index_update = new StoreXNode(ctrl, log_store, index_adr, adr_type, next_index, MemNode::unordered); 2839 phase->register_new_node(index_update, ctrl); 2840 2841 unstable_region->init_req(3, ctrl); 2842 unstable_phi->init_req(3, index_update); 2843 2844 ctrl = full; 2845 2846 Node* base = find_bottom_mem(ctrl, phase); 2847 2848 MergeMemNode* mm = MergeMemNode::make(base); 2849 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem); 2850 phase->register_new_node(mm, ctrl); 2851 2852 Node* call = new CallLeafNode(ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type(), CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", TypeRawPtr::BOTTOM); 2853 call->init_req(TypeFunc::Control, ctrl); 2854 call->init_req(TypeFunc::I_O, phase->C->top()); 2855 call->init_req(TypeFunc::Memory, mm); 2856 call->init_req(TypeFunc::FramePtr, phase->C->top()); 2857 call->init_req(TypeFunc::ReturnAdr, phase->C->top()); 2858 call->init_req(TypeFunc::Parms, pre_val); 2859 call->init_req(TypeFunc::Parms+1, thread); 2860 phase->register_control(call, loop, ctrl); 2861 2862 Node* ctrl_proj = new ProjNode(call, TypeFunc::Control); 2863 phase->register_control(ctrl_proj, loop, call); 2864 Node* mem_proj = new ProjNode(call, TypeFunc::Memory); 2865 phase->register_new_node(mem_proj, call); 2866 2867 unstable_region->init_req(4, ctrl_proj); 2868 unstable_phi->init_req(4, mem_proj); 2869 2870 phase->register_control(unstable_region, loop, null_val != NULL ? null_val->in(0) : evac_in_progress->in(0)); 2871 phase->register_new_node(unstable_phi, unstable_region); 2872 phase->register_control(stable_test_region, loop, heap_stable->in(0)); 2873 phase->register_new_node(stable_test_phi, stable_test_region); 2874 2875 fix_ctrl(barrier, stable_test_region, fixer, uses, uses_to_ignore, last, phase); 2876 for(uint next = 0; next < uses.size(); next++ ) { 2877 Node *n = uses.at(next); 2878 assert(phase->get_ctrl(n) == init_ctrl, "bad control"); 2879 assert(n != init_raw_mem, "should leave input raw mem above the barrier"); 2880 phase->set_ctrl(n, stable_test_region); 2881 follow_barrier_uses(n, init_ctrl, uses, phase); 2882 } 2883 fixer.fix_mem(init_ctrl, stable_test_region, init_raw_mem, raw_mem_for_ctrl, stable_test_phi, uses); 2884 2885 phase->igvn().replace_node(barrier, pre_val); 2886 } 2887 2888 for (int i = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count(); i > 0; i--) { 2889 int cnt = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count(); 2890 ShenandoahWriteBarrierNode* wb = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barrier(i-1); 2891 2892 uint last = phase->C->unique(); 2893 Node* ctrl = phase->get_ctrl(wb); 2894 2895 Node* raw_mem = fixer.find_mem(ctrl, wb); 2896 Node* init_raw_mem = raw_mem; 2897 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL); 2898 int alias = phase->C->get_alias_index(wb->adr_type()); 2899 Node* wb_mem = wb->in(Memory); 2900 Node* init_wb_mem = wb_mem; 2901 2902 Node* val = wb->in(ValueIn); 2903 Node* wbproj = wb->find_out_with(Op_ShenandoahWBMemProj); 2904 IdealLoopTree *loop = phase->get_loop(ctrl); 2905 2906 assert(val->Opcode() != Op_ShenandoahWriteBarrier, "No chain of write barriers"); 2907 2908 CallStaticJavaNode* unc = wb->pin_and_expand_null_check(phase->igvn()); 2909 Node* unc_ctrl = NULL; 2910 if (unc != NULL) { 2911 if (val->in(0) != ctrl) { 2912 unc = NULL; 2913 } else { 2914 unc_ctrl = val->in(0); 2915 } 2916 } 2917 2918 Node* uncasted_val = val; 2919 if (unc != NULL) { 2920 uncasted_val = val->in(1); 2921 } 2922 2923 Node* evac_in_progress = NULL; 2924 Node* evac_not_in_progress = NULL; 2925 Node* heap_stable_ctrl = NULL; 2926 Node* null_ctrl = NULL; 2927 test_evacuation_in_progress(ctrl, val, raw_mem, evac_in_progress, evac_not_in_progress, heap_stable_ctrl, null_ctrl, phase); 2928 IfNode* evacuation_iff = evac_in_progress->in(0)->as_If(); 2929 IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If(); 2930 2931 Node* evacuation_region = new RegionNode(5); 2932 Node* evacuation_val_phi = new PhiNode(evacuation_region, uncasted_val->bottom_type()->is_oopptr()->cast_to_nonconst()); 2933 Node* evacuation_mem_phi = PhiNode::make(evacuation_region, wb_mem, Type::MEMORY, phase->C->alias_type(wb->adr_type())->adr_type()); 2934 Node* evacuation_raw_mem_phi = PhiNode::make(evacuation_region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM); 2935 Node* region = new RegionNode(3); 2936 Node* val_phi = new PhiNode(region, uncasted_val->bottom_type()->is_oopptr()->cast_to_nonconst()); 2937 Node* mem_phi = PhiNode::make(region, wb_mem, Type::MEMORY, phase->C->alias_type(wb->adr_type())->adr_type()); 2938 Node* raw_mem_phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM); 2939 2940 if (null_ctrl != NULL) { 2941 evacuation_region->init_req(4, null_ctrl); 2942 evacuation_val_phi->init_req(4, phase->igvn().zerocon(T_OBJECT)); 2943 evacuation_mem_phi->init_req(4, wb_mem); 2944 evacuation_raw_mem_phi->init_req(4, raw_mem); 2945 } else { 2946 evacuation_region->del_req(4); 2947 evacuation_val_phi->del_req(4); 2948 evacuation_mem_phi->del_req(4); 2949 evacuation_raw_mem_phi->del_req(4); 2950 } 2951 2952 Node* unc_region = NULL; 2953 evacuation_not_in_progress(evac_not_in_progress, val, unc_ctrl, raw_mem, wb_mem, 2954 evacuation_region, evacuation_val_phi, evacuation_mem_phi, evacuation_raw_mem_phi, unc_region, 2955 phase); 2956 2957 heap_stable(heap_stable_ctrl, val, unc_ctrl, init_raw_mem, init_wb_mem, region, val_phi, mem_phi, raw_mem_phi, 2958 unc_region, phase); 2959 2960 evacuation_in_progress(evac_in_progress, val, evacuation_iff, unc, unc_ctrl, 2961 raw_mem, wb_mem, evacuation_region, evacuation_val_phi, evacuation_mem_phi, evacuation_raw_mem_phi, 2962 unc_region, alias, uses, 2963 phase); 2964 region->init_req(2, evacuation_region); 2965 val_phi->init_req(2, evacuation_val_phi); 2966 mem_phi->init_req(2, evacuation_mem_phi); 2967 raw_mem_phi->init_req(2, evacuation_raw_mem_phi); 2968 phase->register_control(evacuation_region, loop, evacuation_iff); 2969 phase->register_new_node(evacuation_val_phi, evacuation_region); 2970 phase->register_new_node(evacuation_mem_phi, evacuation_region); 2971 phase->register_new_node(evacuation_raw_mem_phi, evacuation_region); 2972 2973 phase->register_control(region, loop, heap_stable_iff); 2974 2975 Node* out_val = val_phi; 2976 phase->register_new_node(val_phi, region); 2977 phase->register_new_node(mem_phi, region); 2978 phase->register_new_node(raw_mem_phi, region); 2979 2980 fix_ctrl(wb, region, fixer, uses, uses_to_ignore, last, phase); 2981 2982 phase->igvn().replace_input_of(wbproj, 0, phase->C->top()); 2983 phase->lazy_replace(wbproj, mem_phi); 2984 2985 if (unc != NULL) { 2986 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) { 2987 Node* u = val->fast_out(i); 2988 Node* c = phase->ctrl_or_self(u); 2989 if (u != wb && (c != ctrl || is_dominator_same_ctrl(c, wb, u, phase))) { 2990 phase->igvn().rehash_node_delayed(u); 2991 int nb = u->replace_edge(val, out_val); 2992 --i, imax -= nb; 2993 } 2994 } 2995 if (val->outcnt() == 0) { 2996 phase->lazy_update(val, out_val); 2997 phase->igvn()._worklist.push(val); 2998 } 2999 } 3000 phase->lazy_replace(wb, out_val); 3001 3002 follow_barrier_uses(mem_phi, ctrl, uses, phase); 3003 follow_barrier_uses(out_val, ctrl, uses, phase); 3004 3005 for(uint next = 0; next < uses.size(); next++ ) { 3006 Node *n = uses.at(next); 3007 assert(phase->get_ctrl(n) == ctrl, "bad control"); 3008 assert(n != init_raw_mem, "should leave input raw mem above the barrier"); 3009 phase->set_ctrl(n, region); 3010 follow_barrier_uses(n, ctrl, uses, phase); 3011 } 3012 3013 // The slow path call produces memory: hook the raw memory phi 3014 // from the expanded write barrier with the rest of the graph 3015 // which may require adding memory phis at every post dominated 3016 // region and at enclosing loop heads. Use the memory state 3017 // collected in memory_nodes to fix the memory graph. Update that 3018 // memory state as we go. 3019 fixer.fix_mem(ctrl, region, init_raw_mem, raw_mem_for_ctrl, raw_mem_phi, uses); 3020 assert(ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count() == cnt - 1, "not replaced"); 3021 } 3022 3023 assert(ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count() == 0, "all write barrier nodes should have been replaced"); 3024 } 3025 3026 void ShenandoahWriteBarrierNode::move_evacuation_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) { 3027 // move test and its mem barriers out of the loop 3028 assert(is_evacuation_in_progress_test(iff), "inconsistent"); 3029 3030 IdealLoopTree *loop = phase->get_loop(iff); 3031 Node* loop_head = loop->_head; 3032 Node* entry_c = loop_head->in(LoopNode::EntryControl); 3033 3034 Node* load = iff->in(1)->in(1)->in(1)->in(1); 3035 assert(is_gc_state_load(load), "broken"); 3036 if (!phase->is_dominator(load->in(0), entry_c)) { 3037 Node* mem_ctrl = NULL; 3038 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase); 3039 phase->igvn().replace_input_of(load, MemNode::Memory, mem); 3040 phase->igvn().replace_input_of(load, 0, entry_c); 3041 phase->set_ctrl_and_loop(load, entry_c); 3042 } 3043 } 3044 3045 void ShenandoahWriteBarrierNode::move_heap_stable_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) { 3046 IdealLoopTree *loop = phase->get_loop(iff); 3047 Node* loop_head = loop->_head; 3048 Node* entry_c = loop_head->in(LoopNode::EntryControl); 3049 3050 Node* load = iff->in(1)->in(1)->in(1)->in(1); 3051 assert(is_gc_state_load(load), "broken"); 3052 if (!phase->is_dominator(load->in(0), entry_c)) { 3053 Node* mem_ctrl = NULL; 3054 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase); 3055 phase->igvn().replace_input_of(load, MemNode::Memory, mem); 3056 phase->igvn().replace_input_of(load, 0, entry_c); 3057 phase->set_ctrl_and_loop(load, entry_c); 3058 } 3059 } 3060 3061 void ShenandoahWriteBarrierNode::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) { 3062 assert(is_evacuation_in_progress_test(n) || is_heap_stable_test(n), "no other tests"); 3063 if (phase->identical_backtoback_ifs(n)) { 3064 Node* n_ctrl = is_evacuation_in_progress_test(n) ? ShenandoahWriteBarrierNode::evacuation_in_progress_test_ctrl(n) : n->in(0); 3065 if (phase->can_split_if(n_ctrl)) { 3066 IfNode* dom_if = phase->idom(n_ctrl)->as_If(); 3067 if (is_heap_stable_test(n)) { 3068 Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1); 3069 assert(is_gc_state_load(gc_state_load), "broken"); 3070 Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1); 3071 assert(is_gc_state_load(dom_gc_state_load), "broken"); 3072 if (gc_state_load != dom_gc_state_load) { 3073 phase->igvn().replace_node(gc_state_load, dom_gc_state_load); 3074 } 3075 } 3076 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 3077 Node* proj_true = dom_if->proj_out(1); 3078 Node* proj_false = dom_if->proj_out(0); 3079 Node* con_true = phase->igvn().makecon(TypeInt::ONE); 3080 Node* con_false = phase->igvn().makecon(TypeInt::ZERO); 3081 3082 for (uint i = 1; i < n_ctrl->req(); i++) { 3083 if (phase->is_dominator(proj_true, n_ctrl->in(i))) { 3084 bolphi->init_req(i, con_true); 3085 } else { 3086 assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 3087 bolphi->init_req(i, con_false); 3088 } 3089 } 3090 phase->register_new_node(bolphi, n_ctrl); 3091 phase->igvn().replace_input_of(n, 1, bolphi); 3092 phase->do_split_if(n); 3093 } 3094 } 3095 } 3096 3097 void ShenandoahWriteBarrierNode::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) { 3098 Node_List heap_stable_tests; 3099 Node_List evacuation_tests; 3100 Node_List gc_state_loads; 3101 3102 stack.push(phase->C->start(), 0); 3103 do { 3104 Node* n = stack.node(); 3105 uint i = stack.index(); 3106 3107 if (i < n->outcnt()) { 3108 Node* u = n->raw_out(i); 3109 stack.set_index(i+1); 3110 if (!visited.test_set(u->_idx)) { 3111 stack.push(u, 0); 3112 } 3113 } else { 3114 stack.pop(); 3115 if (n->is_If() && ShenandoahWriteBarrierNode::is_evacuation_in_progress_test(n)) { 3116 evacuation_tests.push(n); 3117 } 3118 if (ShenandoahCommonGCStateLoads && ShenandoahWriteBarrierNode::is_gc_state_load(n)) { 3119 gc_state_loads.push(n); 3120 } 3121 if (n->is_If() && ShenandoahWriteBarrierNode::is_heap_stable_test(n)) { 3122 heap_stable_tests.push(n); 3123 } 3124 } 3125 } while (stack.size() > 0); 3126 3127 bool progress; 3128 do { 3129 progress = false; 3130 for (uint i = 0; i < gc_state_loads.size(); i++) { 3131 Node* n = gc_state_loads.at(i); 3132 if (n->outcnt() != 0) { 3133 progress |= ShenandoahWriteBarrierNode::try_common_gc_state_load(n, phase); 3134 } 3135 } 3136 } while (progress); 3137 3138 for (uint i = 0; i < heap_stable_tests.size(); i++) { 3139 Node* n = heap_stable_tests.at(i); 3140 assert(is_heap_stable_test(n), "only evacuation test"); 3141 merge_back_to_back_tests(n, phase); 3142 } 3143 3144 if (!phase->C->major_progress()) { 3145 for (uint i = 0; i < evacuation_tests.size(); i++) { 3146 Node* n = evacuation_tests.at(i); 3147 assert(is_evacuation_in_progress_test(n), "only evacuation test"); 3148 merge_back_to_back_tests(n, phase); 3149 } 3150 } 3151 3152 if (!phase->C->major_progress()) { 3153 VectorSet seen(Thread::current()->resource_area()); 3154 for (uint i = 0; i < heap_stable_tests.size(); i++) { 3155 Node* n = heap_stable_tests.at(i); 3156 IdealLoopTree* loop = phase->get_loop(n); 3157 if (loop != phase->ltree_root() && 3158 loop->_child == NULL && 3159 !loop->_irreducible) { 3160 LoopNode* head = loop->_head->as_Loop(); 3161 if ((!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) && 3162 !seen.test_set(head->_idx) && 3163 loop->policy_unswitching(phase, true)) { 3164 IfNode* iff = phase->find_unswitching_candidate(loop, true); 3165 if (iff != NULL && (is_evacuation_in_progress_test(iff) || is_heap_stable_test(iff))) { 3166 if (head->is_strip_mined()) { 3167 head->verify_strip_mined(0); 3168 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop(); 3169 OuterStripMinedLoopEndNode* le = head->outer_loop_end(); 3170 Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl)); 3171 phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl)); 3172 Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt); 3173 phase->register_control(new_le, phase->get_loop(le), le->in(0)); 3174 phase->lazy_replace(outer, new_outer); 3175 phase->lazy_replace(le, new_le); 3176 head->clear_strip_mined(); 3177 } 3178 phase->do_unswitching(loop, old_new, true); 3179 } 3180 } 3181 } 3182 } 3183 } 3184 } 3185 3186 #ifdef ASSERT 3187 void ShenandoahBarrierNode::verify_raw_mem(RootNode* root) { 3188 const bool trace = false; 3189 ResourceMark rm; 3190 Unique_Node_List nodes; 3191 Unique_Node_List controls; 3192 Unique_Node_List memories; 3193 3194 nodes.push(root); 3195 for (uint next = 0; next < nodes.size(); next++) { 3196 Node *n = nodes.at(next); 3197 if (n->Opcode() == Op_CallLeafNoFP && 3198 ShenandoahBarrierSetAssembler::is_shenandoah_wb_C_call(n->as_Call()->entry_point())) { 3199 controls.push(n); 3200 if (trace) { tty->print("XXXXXX verifying"); n->dump(); } 3201 for (uint next2 = 0; next2 < controls.size(); next2++) { 3202 Node *m = controls.at(next2); 3203 if (!m->is_Loop() || controls.member(m->in(LoopNode::EntryControl)) || 1) { 3204 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3205 Node* u = m->fast_out(i); 3206 if (u->is_CFG() && !u->is_Root() && 3207 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1) && 3208 !(u->is_Region() && u->unique_ctrl_out()->Opcode() == Op_Halt)) { 3209 if (trace) { tty->print("XXXXXX pushing control"); u->dump(); } 3210 controls.push(u); 3211 } 3212 } 3213 } 3214 } 3215 memories.push(n->as_Call()->proj_out(TypeFunc::Memory)); 3216 for (uint next2 = 0; next2 < memories.size(); next2++) { 3217 Node *m = memories.at(next2); 3218 assert(m->bottom_type() == Type::MEMORY, ""); 3219 if (!m->is_Phi() || !m->in(0)->is_Loop() || controls.member(m->in(0)->in(LoopNode::EntryControl)) || 1) { 3220 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3221 Node* u = m->fast_out(i); 3222 if (u->bottom_type() == Type::MEMORY && (u->is_Mem() || u->is_ClearArray())) { 3223 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); } 3224 memories.push(u); 3225 } else if (u->is_LoadStore()) { 3226 if (trace) { tty->print("XXXXXX pushing memory"); u->find_out_with(Op_SCMemProj)->dump(); } 3227 memories.push(u->find_out_with(Op_SCMemProj)); 3228 } else if (u->is_MergeMem() && u->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == m) { 3229 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); } 3230 memories.push(u); 3231 } else if (u->is_Phi()) { 3232 assert(u->bottom_type() == Type::MEMORY, ""); 3233 if (u->adr_type() == TypeRawPtr::BOTTOM || u->adr_type() == TypePtr::BOTTOM) { 3234 assert(controls.member(u->in(0)), ""); 3235 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); } 3236 memories.push(u); 3237 } 3238 } else if (u->is_SafePoint() || u->is_MemBar()) { 3239 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) { 3240 Node* uu = u->fast_out(j); 3241 if (uu->bottom_type() == Type::MEMORY) { 3242 if (trace) { tty->print("XXXXXX pushing memory"); uu->dump(); } 3243 memories.push(uu); 3244 } 3245 } 3246 } 3247 } 3248 } 3249 } 3250 for (uint next2 = 0; next2 < controls.size(); next2++) { 3251 Node *m = controls.at(next2); 3252 if (m->is_Region()) { 3253 bool all_in = true; 3254 for (uint i = 1; i < m->req(); i++) { 3255 if (!controls.member(m->in(i))) { 3256 all_in = false; 3257 break; 3258 } 3259 } 3260 if (trace) { tty->print("XXX verifying %s", all_in ? "all in" : ""); m->dump(); } 3261 bool found_phi = false; 3262 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax && !found_phi; j++) { 3263 Node* u = m->fast_out(j); 3264 if (u->is_Phi() && memories.member(u)) { 3265 found_phi = true; 3266 for (uint i = 1; i < u->req() && found_phi; i++) { 3267 Node* k = u->in(i); 3268 if (memories.member(k) != controls.member(m->in(i))) { 3269 found_phi = false; 3270 } 3271 } 3272 } 3273 } 3274 assert(found_phi || all_in, ""); 3275 } 3276 } 3277 controls.clear(); 3278 memories.clear(); 3279 } 3280 for( uint i = 0; i < n->len(); ++i ) { 3281 Node *m = n->in(i); 3282 if (m != NULL) { 3283 nodes.push(m); 3284 } 3285 } 3286 } 3287 } 3288 #endif 3289 3290 const Type* ShenandoahEnqueueBarrierNode::bottom_type() const { 3291 if (in(1) == NULL || in(1)->is_top()) { 3292 return Type::TOP; 3293 } 3294 const Type* t = in(1)->bottom_type(); 3295 if (t == TypePtr::NULL_PTR) { 3296 return t; 3297 } 3298 return t->is_oopptr()->cast_to_nonconst(); 3299 } 3300 3301 const Type* ShenandoahEnqueueBarrierNode::Value(PhaseGVN* phase) const { 3302 if (in(1) == NULL) { 3303 return Type::TOP; 3304 } 3305 const Type* t = phase->type(in(1)); 3306 if (t == Type::TOP) { 3307 return Type::TOP; 3308 } 3309 if (t == TypePtr::NULL_PTR) { 3310 return t; 3311 } 3312 return t->is_oopptr()->cast_to_nonconst(); 3313 } 3314 3315 int ShenandoahEnqueueBarrierNode::needed(Node* n) { 3316 if (n == NULL || 3317 n->is_Allocate() || 3318 n->bottom_type() == TypePtr::NULL_PTR || 3319 n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL) { 3320 return NotNeeded; 3321 } 3322 if (n->is_Phi() || 3323 n->is_CMove()) { 3324 return MaybeNeeded; 3325 } 3326 return Needed; 3327 } 3328 3329 Node* ShenandoahEnqueueBarrierNode::next(Node* n) { 3330 for (;;) { 3331 if (n == NULL) { 3332 return n; 3333 } else if (n->bottom_type() == TypePtr::NULL_PTR) { 3334 return n; 3335 } else if (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL) { 3336 return n; 3337 } else if (n->is_ConstraintCast() || 3338 n->Opcode() == Op_DecodeN || 3339 n->Opcode() == Op_EncodeP) { 3340 n = n->in(1); 3341 } else if (n->is_Proj()) { 3342 n = n->in(0); 3343 } else { 3344 return n; 3345 } 3346 } 3347 ShouldNotReachHere(); 3348 return NULL; 3349 } 3350 3351 3352 Node* ShenandoahEnqueueBarrierNode::Identity(PhaseGVN* phase) { 3353 PhaseIterGVN* igvn = phase->is_IterGVN(); 3354 3355 Node* n = next(in(1)); 3356 3357 int cont = needed(n); 3358 3359 if (cont == NotNeeded) { 3360 return in(1); 3361 } else if (cont == MaybeNeeded) { 3362 if (igvn == NULL) { 3363 phase->record_for_igvn(this); 3364 return this; 3365 } else { 3366 ResourceMark rm; 3367 Unique_Node_List wq; 3368 uint wq_i = 0; 3369 3370 for (;;) { 3371 if (n->is_Phi()) { 3372 for (uint i = 1; i < n->req(); i++) { 3373 Node* m = n->in(i); 3374 if (m != NULL) { 3375 wq.push(m); 3376 } 3377 } 3378 } else { 3379 assert(n->is_CMove(), "nothing else here"); 3380 Node* m = n->in(CMoveNode::IfFalse); 3381 wq.push(m); 3382 m = n->in(CMoveNode::IfTrue); 3383 wq.push(m); 3384 } 3385 Node* orig_n = NULL; 3386 do { 3387 if (wq_i >= wq.size()) { 3388 return in(1); 3389 } 3390 n = wq.at(wq_i); 3391 wq_i++; 3392 orig_n = n; 3393 n = next(n); 3394 cont = needed(n); 3395 if (cont == Needed) { 3396 return this; 3397 } 3398 } while (cont != MaybeNeeded || (orig_n != n && wq.member(n))); 3399 } 3400 } 3401 } 3402 3403 return this; 3404 } 3405 3406 #ifdef ASSERT 3407 static bool has_never_branch(Node* root) { 3408 for (uint i = 1; i < root->req(); i++) { 3409 Node* in = root->in(i); 3410 if (in != NULL && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->Opcode() == Op_NeverBranch) { 3411 return true; 3412 } 3413 } 3414 return false; 3415 } 3416 #endif 3417 3418 void MemoryGraphFixer::collect_memory_nodes() { 3419 Node_Stack stack(0); 3420 VectorSet visited(Thread::current()->resource_area()); 3421 Node_List regions; 3422 3423 // Walk the raw memory graph and create a mapping from CFG node to 3424 // memory node. Exclude phis for now. 3425 stack.push(_phase->C->root(), 1); 3426 do { 3427 Node* n = stack.node(); 3428 int opc = n->Opcode(); 3429 uint i = stack.index(); 3430 if (i < n->req()) { 3431 Node* mem = NULL; 3432 if (opc == Op_Root) { 3433 Node* in = n->in(i); 3434 int in_opc = in->Opcode(); 3435 if (in_opc == Op_Return || in_opc == Op_Rethrow) { 3436 mem = in->in(TypeFunc::Memory); 3437 } else if (in_opc == Op_Halt) { 3438 if (!in->in(0)->is_Region()) { 3439 Node* proj = in->in(0); 3440 assert(proj->is_Proj(), ""); 3441 Node* in = proj->in(0); 3442 assert(in->is_CallStaticJava() || in->Opcode() == Op_NeverBranch || in->Opcode() == Op_Catch || proj->is_IfProj(), ""); 3443 if (in->is_CallStaticJava()) { 3444 mem = in->in(TypeFunc::Memory); 3445 } else if (in->Opcode() == Op_Catch) { 3446 Node* call = in->in(0)->in(0); 3447 assert(call->is_Call(), ""); 3448 mem = call->in(TypeFunc::Memory); 3449 } 3450 } 3451 } else { 3452 #ifdef ASSERT 3453 n->dump(); 3454 in->dump(); 3455 #endif 3456 ShouldNotReachHere(); 3457 } 3458 } else { 3459 assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, ""); 3460 assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, ""); 3461 mem = n->in(i); 3462 } 3463 i++; 3464 stack.set_index(i); 3465 if (mem == NULL) { 3466 continue; 3467 } 3468 for (;;) { 3469 if (visited.test_set(mem->_idx) || mem->is_Start()) { 3470 break; 3471 } 3472 if (mem->is_Phi()) { 3473 stack.push(mem, 2); 3474 mem = mem->in(1); 3475 } else if (mem->is_Proj()) { 3476 stack.push(mem, mem->req()); 3477 mem = mem->in(0); 3478 } else if (mem->is_SafePoint() || mem->is_MemBar()) { 3479 mem = mem->in(TypeFunc::Memory); 3480 } else if (mem->is_MergeMem()) { 3481 MergeMemNode* mm = mem->as_MergeMem(); 3482 mem = mm->memory_at(_alias); 3483 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) { 3484 assert(_alias == Compile::AliasIdxRaw, ""); 3485 stack.push(mem, mem->req()); 3486 mem = mem->in(MemNode::Memory); 3487 } else if (mem->Opcode() == Op_ShenandoahWriteBarrier) { 3488 assert(_alias != Compile::AliasIdxRaw, ""); 3489 mem = mem->in(ShenandoahBarrierNode::Memory); 3490 } else { 3491 #ifdef ASSERT 3492 mem->dump(); 3493 #endif 3494 ShouldNotReachHere(); 3495 } 3496 } 3497 } else { 3498 if (n->is_Phi()) { 3499 // Nothing 3500 } else if (!n->is_Root()) { 3501 Node* c = get_ctrl(n); 3502 _memory_nodes.map(c->_idx, n); 3503 } 3504 stack.pop(); 3505 } 3506 } while(stack.is_nonempty()); 3507 3508 // Iterate over CFG nodes in rpo and propagate memory state to 3509 // compute memory state at regions, creating new phis if needed. 3510 Node_List rpo_list; 3511 visited.Clear(); 3512 _phase->rpo(_phase->C->root(), stack, visited, rpo_list); 3513 Node* root = rpo_list.pop(); 3514 assert(root == _phase->C->root(), ""); 3515 3516 const bool trace = false; 3517 #ifdef ASSERT 3518 if (trace) { 3519 for (int i = rpo_list.size() - 1; i >= 0; i--) { 3520 Node* c = rpo_list.at(i); 3521 if (_memory_nodes[c->_idx] != NULL) { 3522 tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); 3523 } 3524 } 3525 } 3526 #endif 3527 uint last = _phase->C->unique(); 3528 3529 #ifdef ASSERT 3530 uint8_t max_depth = 0; 3531 for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) { 3532 IdealLoopTree* lpt = iter.current(); 3533 max_depth = MAX2(max_depth, lpt->_nest); 3534 } 3535 #endif 3536 3537 bool progress = true; 3538 int iteration = 0; 3539 Node_List dead_phis; 3540 while (progress) { 3541 progress = false; 3542 iteration++; 3543 assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop(), ""); 3544 if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); } 3545 IdealLoopTree* last_updated_ilt = NULL; 3546 for (int i = rpo_list.size() - 1; i >= 0; i--) { 3547 Node* c = rpo_list.at(i); 3548 3549 Node* prev_mem = _memory_nodes[c->_idx]; 3550 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) { 3551 Node* prev_region = regions[c->_idx]; 3552 Node* unique = NULL; 3553 for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) { 3554 Node* m = _memory_nodes[c->in(j)->_idx]; 3555 assert(m != NULL || (c->is_Loop() && j == LoopNode::LoopBackControl && iteration == 1) || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "expect memory state"); 3556 if (m != NULL) { 3557 if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) { 3558 assert(c->is_Loop() && j == LoopNode::LoopBackControl || _phase->C->has_irreducible_loop(), ""); 3559 // continue 3560 } else if (unique == NULL) { 3561 unique = m; 3562 } else if (m == unique) { 3563 // continue 3564 } else { 3565 unique = NodeSentinel; 3566 } 3567 } 3568 } 3569 assert(unique != NULL, "empty phi???"); 3570 if (unique != NodeSentinel) { 3571 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c) { 3572 dead_phis.push(prev_region); 3573 } 3574 regions.map(c->_idx, unique); 3575 } else { 3576 Node* phi = NULL; 3577 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) { 3578 phi = prev_region; 3579 for (uint k = 1; k < c->req(); k++) { 3580 Node* m = _memory_nodes[c->in(k)->_idx]; 3581 assert(m != NULL, "expect memory state"); 3582 phi->set_req(k, m); 3583 } 3584 } else { 3585 for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == NULL; j++) { 3586 Node* u = c->fast_out(j); 3587 if (u->is_Phi() && u->bottom_type() == Type::MEMORY && 3588 (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) { 3589 phi = u; 3590 for (uint k = 1; k < c->req() && phi != NULL; k++) { 3591 Node* m = _memory_nodes[c->in(k)->_idx]; 3592 assert(m != NULL, "expect memory state"); 3593 if (u->in(k) != m) { 3594 phi = NULL; 3595 } 3596 } 3597 } 3598 } 3599 if (phi == NULL) { 3600 phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias)); 3601 for (uint k = 1; k < c->req(); k++) { 3602 Node* m = _memory_nodes[c->in(k)->_idx]; 3603 assert(m != NULL, "expect memory state"); 3604 phi->init_req(k, m); 3605 } 3606 } 3607 } 3608 assert(phi != NULL, ""); 3609 regions.map(c->_idx, phi); 3610 } 3611 Node* current_region = regions[c->_idx]; 3612 if (current_region != prev_region) { 3613 progress = true; 3614 if (prev_region == prev_mem) { 3615 _memory_nodes.map(c->_idx, current_region); 3616 } 3617 } 3618 } else if (prev_mem == NULL || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) { 3619 Node* m = _memory_nodes[_phase->idom(c)->_idx]; 3620 assert(m != NULL, "expect memory state"); 3621 if (m != prev_mem) { 3622 _memory_nodes.map(c->_idx, m); 3623 progress = true; 3624 } 3625 } 3626 #ifdef ASSERT 3627 if (trace) { tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); } 3628 #endif 3629 } 3630 } 3631 3632 // Replace existing phi with computed memory state for that region 3633 // if different (could be a new phi or a dominating memory node if 3634 // that phi was found to be useless). 3635 while (dead_phis.size() > 0) { 3636 Node* n = dead_phis.pop(); 3637 n->replace_by(_phase->C->top()); 3638 n->destruct(); 3639 } 3640 for (int i = rpo_list.size() - 1; i >= 0; i--) { 3641 Node* c = rpo_list.at(i); 3642 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) { 3643 Node* n = regions[c->_idx]; 3644 if (n->is_Phi() && n->_idx >= last && n->in(0) == c) { 3645 _phase->register_new_node(n, c); 3646 } 3647 } 3648 } 3649 for (int i = rpo_list.size() - 1; i >= 0; i--) { 3650 Node* c = rpo_list.at(i); 3651 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) { 3652 Node* n = regions[c->_idx]; 3653 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) { 3654 Node* u = c->fast_out(i); 3655 if (u->is_Phi() && u->bottom_type() == Type::MEMORY && 3656 u != n) { 3657 if (u->adr_type() == TypePtr::BOTTOM) { 3658 fix_memory_uses(u, n, n, c); 3659 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) { 3660 _phase->lazy_replace(u, n); 3661 --i; --imax; 3662 } 3663 } 3664 } 3665 } 3666 } 3667 } 3668 3669 Node* MemoryGraphFixer::get_ctrl(Node* n) const { 3670 Node* c = _phase->get_ctrl(n); 3671 if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Call()) { 3672 assert(c == n->in(0), ""); 3673 CallNode* call = c->as_Call(); 3674 CallProjections projs; 3675 call->extract_projections(&projs, true, false); 3676 if (projs.catchall_memproj != NULL) { 3677 if (projs.fallthrough_memproj == n) { 3678 c = projs.fallthrough_catchproj; 3679 } else { 3680 assert(projs.catchall_memproj == n, ""); 3681 c = projs.catchall_catchproj; 3682 } 3683 } 3684 } 3685 return c; 3686 } 3687 3688 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const { 3689 if (_phase->has_ctrl(n)) 3690 return get_ctrl(n); 3691 else { 3692 assert (n->is_CFG(), "must be a CFG node"); 3693 return n; 3694 } 3695 } 3696 3697 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const { 3698 return m != NULL && get_ctrl(m) == c; 3699 } 3700 3701 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const { 3702 assert(n == NULL || _phase->ctrl_or_self(n) == ctrl, ""); 3703 Node* mem = _memory_nodes[ctrl->_idx]; 3704 Node* c = ctrl; 3705 while (!mem_is_valid(mem, c) && 3706 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem))) { 3707 c = _phase->idom(c); 3708 mem = _memory_nodes[c->_idx]; 3709 } 3710 if (n != NULL && mem_is_valid(mem, c)) { 3711 while (!ShenandoahWriteBarrierNode::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) { 3712 mem = next_mem(mem, _alias); 3713 } 3714 if (mem->is_MergeMem()) { 3715 mem = mem->as_MergeMem()->memory_at(_alias); 3716 } 3717 if (!mem_is_valid(mem, c)) { 3718 do { 3719 c = _phase->idom(c); 3720 mem = _memory_nodes[c->_idx]; 3721 } while (!mem_is_valid(mem, c) && 3722 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem))); 3723 } 3724 } 3725 assert(mem->bottom_type() == Type::MEMORY, ""); 3726 return mem; 3727 } 3728 3729 bool MemoryGraphFixer::has_mem_phi(Node* region) const { 3730 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 3731 Node* use = region->fast_out(i); 3732 if (use->is_Phi() && use->bottom_type() == Type::MEMORY && 3733 (_phase->C->get_alias_index(use->adr_type()) == _alias)) { 3734 return true; 3735 } 3736 } 3737 return false; 3738 } 3739 3740 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) { 3741 assert(_phase->ctrl_or_self(new_mem) == new_ctrl, ""); 3742 const bool trace = false; 3743 DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); }); 3744 DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); }); 3745 GrowableArray<Node*> phis; 3746 if (mem_for_ctrl != mem) { 3747 Node* old = mem_for_ctrl; 3748 Node* prev = NULL; 3749 while (old != mem) { 3750 prev = old; 3751 if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) { 3752 assert(_alias == Compile::AliasIdxRaw, ""); 3753 old = old->in(MemNode::Memory); 3754 } else if (old->Opcode() == Op_SCMemProj) { 3755 assert(_alias == Compile::AliasIdxRaw, ""); 3756 old = old->in(0); 3757 } else if (old->Opcode() == Op_ShenandoahWBMemProj) { 3758 assert(_alias != Compile::AliasIdxRaw, ""); 3759 old = old->in(0); 3760 } else if (old->Opcode() == Op_ShenandoahWriteBarrier) { 3761 assert(_alias != Compile::AliasIdxRaw, ""); 3762 old = old->in(ShenandoahBarrierNode::Memory); 3763 } else { 3764 ShouldNotReachHere(); 3765 } 3766 } 3767 assert(prev != NULL, ""); 3768 if (new_ctrl != ctrl) { 3769 _memory_nodes.map(ctrl->_idx, mem); 3770 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl); 3771 } 3772 uint input = prev->Opcode() == Op_ShenandoahWriteBarrier ? (uint)ShenandoahBarrierNode::Memory : (uint)MemNode::Memory; 3773 _phase->igvn().replace_input_of(prev, input, new_mem); 3774 } else { 3775 uses.clear(); 3776 _memory_nodes.map(new_ctrl->_idx, new_mem); 3777 uses.push(new_ctrl); 3778 for(uint next = 0; next < uses.size(); next++ ) { 3779 Node *n = uses.at(next); 3780 assert(n->is_CFG(), ""); 3781 DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); }); 3782 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3783 Node* u = n->fast_out(i); 3784 if (!u->is_Root() && u->is_CFG() && u != n) { 3785 Node* m = _memory_nodes[u->_idx]; 3786 if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) && 3787 !has_mem_phi(u) && 3788 u->unique_ctrl_out()->Opcode() != Op_Halt) { 3789 DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); }); 3790 DEBUG_ONLY(if (trace && m != NULL) { tty->print("ZZZ mem"); m->dump(); }); 3791 3792 if (!mem_is_valid(m, u) || !m->is_Phi()) { 3793 bool push = true; 3794 bool create_phi = true; 3795 if (_phase->is_dominator(new_ctrl, u)) { 3796 create_phi = false; 3797 } else if (!_phase->C->has_irreducible_loop()) { 3798 IdealLoopTree* loop = _phase->get_loop(ctrl); 3799 bool do_check = true; 3800 IdealLoopTree* l = loop; 3801 create_phi = false; 3802 while (l != _phase->ltree_root()) { 3803 if (_phase->is_dominator(l->_head, u) && _phase->is_dominator(_phase->idom(u), l->_head)) { 3804 create_phi = true; 3805 do_check = false; 3806 break; 3807 } 3808 l = l->_parent; 3809 } 3810 3811 if (do_check) { 3812 assert(!create_phi, ""); 3813 IdealLoopTree* u_loop = _phase->get_loop(u); 3814 if (u_loop != _phase->ltree_root() && u_loop->is_member(loop)) { 3815 Node* c = ctrl; 3816 while (!_phase->is_dominator(c, u_loop->tail())) { 3817 c = _phase->idom(c); 3818 } 3819 if (!_phase->is_dominator(c, u)) { 3820 do_check = false; 3821 } 3822 } 3823 } 3824 3825 if (do_check && _phase->is_dominator(_phase->idom(u), new_ctrl)) { 3826 create_phi = true; 3827 } 3828 } 3829 if (create_phi) { 3830 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias)); 3831 _phase->register_new_node(phi, u); 3832 phis.push(phi); 3833 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); }); 3834 if (!mem_is_valid(m, u)) { 3835 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); }); 3836 _memory_nodes.map(u->_idx, phi); 3837 } else { 3838 DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); }); 3839 for (;;) { 3840 assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj() || m->Opcode() == Op_ShenandoahWriteBarrier /*|| m->is_MergeMem()*/, ""); 3841 Node* next = NULL; 3842 if (m->is_Proj()) { 3843 next = m->in(0); 3844 } else if (m->is_Mem() || m->is_LoadStore()) { 3845 assert(_alias == Compile::AliasIdxRaw, ""); 3846 next = m->in(MemNode::Memory); 3847 } else { 3848 assert(_alias != Compile::AliasIdxRaw, ""); 3849 assert (m->Opcode() == Op_ShenandoahWriteBarrier, ""); 3850 next = m->in(ShenandoahBarrierNode::Memory); 3851 } 3852 if (_phase->get_ctrl(next) != u) { 3853 break; 3854 } 3855 if (next->is_MergeMem()) { 3856 assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, ""); 3857 break; 3858 } 3859 if (next->is_Phi()) { 3860 assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, ""); 3861 break; 3862 } 3863 m = next; 3864 } 3865 3866 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); }); 3867 assert(m->is_Mem() || m->is_LoadStore() || m->Opcode() == Op_ShenandoahWriteBarrier, ""); 3868 uint input = (m->is_Mem() || m->is_LoadStore()) ? (uint)MemNode::Memory : (uint)ShenandoahBarrierNode::Memory; 3869 _phase->igvn().replace_input_of(m, input, phi); 3870 push = false; 3871 } 3872 } else { 3873 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); }); 3874 } 3875 if (push) { 3876 uses.push(u); 3877 } 3878 } 3879 } else if (!mem_is_valid(m, u) && 3880 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1)) { 3881 uses.push(u); 3882 } 3883 } 3884 } 3885 } 3886 for (int i = 0; i < phis.length(); i++) { 3887 Node* n = phis.at(i); 3888 Node* r = n->in(0); 3889 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); }); 3890 for (uint j = 1; j < n->req(); j++) { 3891 Node* m = find_mem(r->in(j), NULL); 3892 _phase->igvn().replace_input_of(n, j, m); 3893 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); }); 3894 } 3895 } 3896 } 3897 uint last = _phase->C->unique(); 3898 MergeMemNode* mm = NULL; 3899 int alias = _alias; 3900 DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); }); 3901 for (DUIterator i = mem->outs(); mem->has_out(i); i++) { 3902 Node* u = mem->out(i); 3903 if (u->_idx < last) { 3904 if (u->is_Mem()) { 3905 if (_phase->C->get_alias_index(u->adr_type()) == alias) { 3906 Node* m = find_mem(_phase->get_ctrl(u), u); 3907 if (m != mem) { 3908 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); }); 3909 _phase->igvn().replace_input_of(u, MemNode::Memory, m); 3910 --i; 3911 } 3912 } 3913 } else if (u->is_MergeMem()) { 3914 MergeMemNode* u_mm = u->as_MergeMem(); 3915 if (u_mm->memory_at(alias) == mem) { 3916 MergeMemNode* newmm = NULL; 3917 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) { 3918 Node* uu = u->fast_out(j); 3919 assert(!uu->is_MergeMem(), "chain of MergeMems?"); 3920 if (uu->is_Phi()) { 3921 assert(uu->adr_type() == TypePtr::BOTTOM, ""); 3922 Node* region = uu->in(0); 3923 int nb = 0; 3924 for (uint k = 1; k < uu->req(); k++) { 3925 if (uu->in(k) == u) { 3926 Node* m = find_mem(region->in(k), NULL); 3927 if (m != mem) { 3928 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); }); 3929 if (newmm == NULL || 1) { 3930 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i); 3931 } 3932 if (newmm != u) { 3933 _phase->igvn().replace_input_of(uu, k, newmm); 3934 nb++; 3935 --jmax; 3936 } 3937 } 3938 } 3939 } 3940 if (nb > 0) { 3941 --j; 3942 } 3943 } else { 3944 Node* m = find_mem(_phase->ctrl_or_self(uu), uu); 3945 if (m != mem) { 3946 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); }); 3947 if (newmm == NULL || 1) { 3948 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i); 3949 } 3950 if (newmm != u) { 3951 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm); 3952 --j, --jmax; 3953 } 3954 } 3955 } 3956 } 3957 } 3958 } else if (u->is_Phi()) { 3959 assert(u->bottom_type() == Type::MEMORY, "what else?"); 3960 if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) { 3961 Node* region = u->in(0); 3962 bool replaced = false; 3963 for (uint j = 1; j < u->req(); j++) { 3964 if (u->in(j) == mem) { 3965 Node* m = find_mem(region->in(j), NULL); 3966 Node* nnew = m; 3967 if (m != mem) { 3968 if (u->adr_type() == TypePtr::BOTTOM) { 3969 mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m)); 3970 nnew = mm; 3971 } 3972 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); }); 3973 _phase->igvn().replace_input_of(u, j, nnew); 3974 replaced = true; 3975 } 3976 } 3977 } 3978 if (replaced) { 3979 --i; 3980 } 3981 } 3982 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) || 3983 u->adr_type() == NULL) { 3984 assert(u->adr_type() != NULL || 3985 u->Opcode() == Op_Rethrow || 3986 u->Opcode() == Op_Return || 3987 u->Opcode() == Op_SafePoint || 3988 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) || 3989 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) || 3990 u->Opcode() == Op_CallLeaf, ""); 3991 Node* m = find_mem(_phase->ctrl_or_self(u), u); 3992 if (m != mem) { 3993 mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m)); 3994 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm); 3995 --i; 3996 } 3997 } else if (_phase->C->get_alias_index(u->adr_type()) == alias) { 3998 Node* m = find_mem(_phase->ctrl_or_self(u), u); 3999 if (m != mem) { 4000 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); }); 4001 _phase->igvn().replace_input_of(u, u->find_edge(mem), m); 4002 --i; 4003 } 4004 } else if (u->adr_type() != TypePtr::BOTTOM && 4005 _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) { 4006 Node* m = find_mem(_phase->ctrl_or_self(u), u); 4007 assert(m != mem, ""); 4008 // u is on the wrong slice... 4009 assert(u->is_ClearArray(), ""); 4010 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); }); 4011 _phase->igvn().replace_input_of(u, u->find_edge(mem), m); 4012 --i; 4013 } 4014 } 4015 } 4016 #ifdef ASSERT 4017 assert(new_mem->outcnt() > 0, ""); 4018 for (int i = 0; i < phis.length(); i++) { 4019 Node* n = phis.at(i); 4020 assert(n->outcnt() > 0, "new phi must have uses now"); 4021 } 4022 #endif 4023 } 4024 4025 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const { 4026 MergeMemNode* mm = MergeMemNode::make(mem); 4027 mm->set_memory_at(_alias, rep_proj); 4028 _phase->register_new_node(mm, rep_ctrl); 4029 return mm; 4030 } 4031 4032 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const { 4033 MergeMemNode* newmm = NULL; 4034 MergeMemNode* u_mm = u->as_MergeMem(); 4035 Node* c = _phase->get_ctrl(u); 4036 if (_phase->is_dominator(c, rep_ctrl)) { 4037 c = rep_ctrl; 4038 } else { 4039 assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other"); 4040 } 4041 if (u->outcnt() == 1) { 4042 if (u->req() > (uint)_alias && u->in(_alias) == mem) { 4043 _phase->igvn().replace_input_of(u, _alias, rep_proj); 4044 --i; 4045 } else { 4046 _phase->igvn().rehash_node_delayed(u); 4047 u_mm->set_memory_at(_alias, rep_proj); 4048 } 4049 newmm = u_mm; 4050 _phase->set_ctrl_and_loop(u, c); 4051 } else { 4052 // can't simply clone u and then change one of its input because 4053 // it adds and then removes an edge which messes with the 4054 // DUIterator 4055 newmm = MergeMemNode::make(u_mm->base_memory()); 4056 for (uint j = 0; j < u->req(); j++) { 4057 if (j < newmm->req()) { 4058 if (j == (uint)_alias) { 4059 newmm->set_req(j, rep_proj); 4060 } else if (newmm->in(j) != u->in(j)) { 4061 newmm->set_req(j, u->in(j)); 4062 } 4063 } else if (j == (uint)_alias) { 4064 newmm->add_req(rep_proj); 4065 } else { 4066 newmm->add_req(u->in(j)); 4067 } 4068 } 4069 if ((uint)_alias >= u->req()) { 4070 newmm->set_memory_at(_alias, rep_proj); 4071 } 4072 _phase->register_new_node(newmm, c); 4073 } 4074 return newmm; 4075 } 4076 4077 bool MemoryGraphFixer::should_process_phi(Node* phi) const { 4078 if (phi->adr_type() == TypePtr::BOTTOM) { 4079 Node* region = phi->in(0); 4080 for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) { 4081 Node* uu = region->fast_out(j); 4082 if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) { 4083 return false; 4084 } 4085 } 4086 return true; 4087 } 4088 return _phase->C->get_alias_index(phi->adr_type()) == _alias; 4089 } 4090 4091 4092 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const { 4093 uint last = _phase-> C->unique(); 4094 MergeMemNode* mm = NULL; 4095 assert(mem->bottom_type() == Type::MEMORY, ""); 4096 for (DUIterator i = mem->outs(); mem->has_out(i); i++) { 4097 Node* u = mem->out(i); 4098 if (u != replacement && u->_idx < last) { 4099 if (u->is_ShenandoahBarrier() && _alias != Compile::AliasIdxRaw) { 4100 if (_phase->C->get_alias_index(u->adr_type()) == _alias && ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) { 4101 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj); 4102 assert(u->find_edge(mem) == -1, "only one edge"); 4103 --i; 4104 } 4105 } else if (u->is_Mem()) { 4106 if (_phase->C->get_alias_index(u->adr_type()) == _alias && ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) { 4107 assert(_alias == Compile::AliasIdxRaw , "only raw memory can lead to a memory operation"); 4108 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj); 4109 assert(u->find_edge(mem) == -1, "only one edge"); 4110 --i; 4111 } 4112 } else if (u->is_MergeMem()) { 4113 MergeMemNode* u_mm = u->as_MergeMem(); 4114 if (u_mm->memory_at(_alias) == mem) { 4115 MergeMemNode* newmm = NULL; 4116 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) { 4117 Node* uu = u->fast_out(j); 4118 assert(!uu->is_MergeMem(), "chain of MergeMems?"); 4119 if (uu->is_Phi()) { 4120 if (should_process_phi(uu)) { 4121 Node* region = uu->in(0); 4122 int nb = 0; 4123 for (uint k = 1; k < uu->req(); k++) { 4124 if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) { 4125 if (newmm == NULL) { 4126 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i); 4127 } 4128 if (newmm != u) { 4129 _phase->igvn().replace_input_of(uu, k, newmm); 4130 nb++; 4131 --jmax; 4132 } 4133 } 4134 } 4135 if (nb > 0) { 4136 --j; 4137 } 4138 } 4139 } else { 4140 if (rep_ctrl != uu && ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) { 4141 if (newmm == NULL) { 4142 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i); 4143 } 4144 if (newmm != u) { 4145 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm); 4146 --j, --jmax; 4147 } 4148 } 4149 } 4150 } 4151 } 4152 } else if (u->is_Phi()) { 4153 assert(u->bottom_type() == Type::MEMORY, "what else?"); 4154 Node* region = u->in(0); 4155 if (should_process_phi(u)) { 4156 bool replaced = false; 4157 for (uint j = 1; j < u->req(); j++) { 4158 if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) { 4159 Node* nnew = rep_proj; 4160 if (u->adr_type() == TypePtr::BOTTOM) { 4161 if (mm == NULL) { 4162 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl); 4163 } 4164 nnew = mm; 4165 } 4166 _phase->igvn().replace_input_of(u, j, nnew); 4167 replaced = true; 4168 } 4169 } 4170 if (replaced) { 4171 --i; 4172 } 4173 4174 } 4175 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) || 4176 u->adr_type() == NULL) { 4177 assert(u->adr_type() != NULL || 4178 u->Opcode() == Op_Rethrow || 4179 u->Opcode() == Op_Return || 4180 u->Opcode() == Op_SafePoint || 4181 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) || 4182 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) || 4183 u->Opcode() == Op_CallLeaf, ""); 4184 if (ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) { 4185 if (mm == NULL) { 4186 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl); 4187 } 4188 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm); 4189 --i; 4190 } 4191 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) { 4192 if (ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) { 4193 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj); 4194 --i; 4195 } 4196 } 4197 } 4198 } 4199 } 4200 4201 void MemoryGraphFixer::remove(Node* n) { 4202 assert(n->Opcode() == Op_ShenandoahWBMemProj, ""); 4203 Node* c = _phase->get_ctrl(n); 4204 Node* mem = find_mem(c, NULL); 4205 if (mem == n) { 4206 _memory_nodes.map(c->_idx, mem->in(0)->in(ShenandoahBarrierNode::Memory)); 4207 } 4208 } 4209 4210 static bool is_on_null_check_path(Block* b, Block* null_check_block) { 4211 if (null_check_block == NULL) { 4212 return false; 4213 } 4214 do { 4215 assert(null_check_block->_num_succs == 1, "only one succ on the path to unc"); 4216 if (b == null_check_block) { 4217 return true; 4218 } 4219 null_check_block = null_check_block->_succs[0]; 4220 } while(!null_check_block->head()->is_Root()); 4221 4222 return false; 4223 } 4224 4225 int PhaseCFG::replace_uses_with_shenandoah_barrier_helper(Node* n, Node* use, Node* val, Block* block, Block* null_check_block) { 4226 int nb = 0; 4227 Block* buse = get_block_for_node(use); 4228 if (is_on_null_check_path(buse, null_check_block)) { 4229 return 0; 4230 } 4231 if (use->is_Phi()) { 4232 for (uint j = 1; j < use->req(); j++) { 4233 if (use->in(j) == val) { 4234 Block* b = get_block_for_node(use->in(0)->in(j)); 4235 if ((block != b && block->dom_lca(b) == block) || 4236 block == b) { 4237 use->set_req(j, n); 4238 nb++; 4239 } 4240 } 4241 } 4242 } else { 4243 if ((block != buse && block->dom_lca(buse) == block) || 4244 (block == buse && !use->is_scheduled())) { 4245 // Let precedence edges alone (can confuse anti-dependence verification code) 4246 for (uint i = 0; i < use->req(); i++) { 4247 if (use->in(i) == val) { 4248 use->set_req(i, n); 4249 nb++; 4250 } 4251 } 4252 assert(nb > 0 || use->find_prec_edge(val) != -1, "no replacement?"); 4253 } 4254 } 4255 4256 return nb; 4257 } 4258 4259 void PhaseCFG::replace_uses_with_shenandoah_barrier(Node* n, Block* block, Node_List& worklist, GrowableArray<int>& ready_cnt, uint max_idx, uint& phi_cnt) { 4260 // Replace all uses of barrier's input that are dominated by the 4261 // barrier with the value returned by the barrier: no need to keep 4262 // both live. 4263 if (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_ShenandoahReadBarrier) { 4264 MachNullCheckNode* null_check = NULL; 4265 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && null_check == NULL; i++) { 4266 Node* use = n->fast_out(i); 4267 if (use->is_MachNullCheck()) { 4268 null_check = use->as_MachNullCheck(); 4269 } 4270 } 4271 Block* null_check_block = NULL; 4272 if (null_check != NULL) { 4273 Node* proj = null_check->find_out_with(Op_IfTrue); 4274 Node* head = proj->unique_out(); 4275 null_check_block = get_block_for_node(head); 4276 } 4277 4278 Node* val = n->in(ShenandoahBarrierNode::ValueIn); 4279 if (!val->bottom_type()->isa_narrowoop()) { 4280 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) { 4281 Node* use = val->fast_out(i); 4282 if (use != n) { 4283 int nb = replace_uses_with_shenandoah_barrier_helper(n, use, val, block, null_check_block); 4284 if (nb > 0) { 4285 --i; imax -= nb; 4286 } 4287 } 4288 } 4289 } else { 4290 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) { 4291 Node* u = val->fast_out(i); 4292 if (u->is_Mach() && u->as_Mach()->ideal_Opcode() == Op_DecodeN) { 4293 int projs = 0; 4294 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) { 4295 Node* uu = u->fast_out(j); 4296 assert(!uu->is_MachTemp(), ""); 4297 if (uu->is_MachProj() && uu->outcnt() == 0) { 4298 projs++; 4299 } else { 4300 int nb = replace_uses_with_shenandoah_barrier_helper(n, uu, u, block, null_check_block); 4301 if (nb > 0) { 4302 if (!u->is_scheduled()) { 4303 push_ready_nodes(n, uu, block, ready_cnt, worklist, max_idx, nb); 4304 } 4305 --j; jmax -= nb; 4306 } 4307 } 4308 } 4309 // The DecodeN may have gone dead 4310 if (u->outcnt() - projs == 0) { 4311 u->disconnect_inputs(NULL, C); 4312 Block* bu = get_block_for_node(u); 4313 unmap_node_from_block(u); 4314 if (bu == block) { 4315 if (u->is_scheduled()) { 4316 block->find_remove(u); 4317 phi_cnt--; 4318 } else { 4319 worklist.yank(u); 4320 block->remove_node(block->end_idx()-1); 4321 } 4322 } else { 4323 bu->find_remove(u); 4324 } 4325 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) { 4326 Node* uu = u->fast_out(j); 4327 assert(uu->is_MachProj() && uu->outcnt() == 0, ""); 4328 assert(bu == get_block_for_node(uu), ""); 4329 uu->disconnect_inputs(NULL, C); 4330 --j; --jmax; 4331 unmap_node_from_block(uu); 4332 if (bu == block) { 4333 if (u->is_scheduled()) { 4334 block->find_remove(uu); 4335 phi_cnt--; 4336 } else { 4337 worklist.yank(uu); 4338 block->remove_node(block->end_idx()-1); 4339 } 4340 } else { 4341 bu->find_remove(uu); 4342 } 4343 assert(uu->is_scheduled() == u->is_scheduled(), ""); 4344 } 4345 --i; --imax; 4346 } 4347 } 4348 } 4349 } 4350 } 4351 }