1 /* 2 * Copyright (c) 2005, 2023, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "gc/shared/collectedHeap.inline.hpp" 28 #include "gc/shared/tlab_globals.hpp" 29 #include "libadt/vectset.hpp" 30 #include "memory/universe.hpp" 31 #include "opto/addnode.hpp" 32 #include "opto/arraycopynode.hpp" 33 #include "opto/callnode.hpp" 34 #include "opto/castnode.hpp" 35 #include "opto/cfgnode.hpp" 36 #include "opto/compile.hpp" 37 #include "opto/convertnode.hpp" 38 #include "opto/graphKit.hpp" 39 #include "opto/intrinsicnode.hpp" 40 #include "opto/locknode.hpp" 41 #include "opto/loopnode.hpp" 42 #include "opto/macro.hpp" 43 #include "opto/memnode.hpp" 44 #include "opto/narrowptrnode.hpp" 45 #include "opto/node.hpp" 46 #include "opto/opaquenode.hpp" 47 #include "opto/phaseX.hpp" 48 #include "opto/rootnode.hpp" 49 #include "opto/runtime.hpp" 50 #include "opto/subnode.hpp" 51 #include "opto/subtypenode.hpp" 52 #include "opto/type.hpp" 53 #include "prims/jvmtiExport.hpp" 54 #include "runtime/continuation.hpp" 55 #include "runtime/sharedRuntime.hpp" 56 #include "utilities/macros.hpp" 57 #include "utilities/powerOfTwo.hpp" 58 #if INCLUDE_G1GC 59 #include "gc/g1/g1ThreadLocalData.hpp" 60 #endif // INCLUDE_G1GC 61 62 63 // 64 // Replace any references to "oldref" in inputs to "use" with "newref". 65 // Returns the number of replacements made. 66 // 67 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { 68 int nreplacements = 0; 69 uint req = use->req(); 70 for (uint j = 0; j < use->len(); j++) { 71 Node *uin = use->in(j); 72 if (uin == oldref) { 73 if (j < req) 74 use->set_req(j, newref); 75 else 76 use->set_prec(j, newref); 77 nreplacements++; 78 } else if (j >= req && uin == nullptr) { 79 break; 80 } 81 } 82 return nreplacements; 83 } 84 85 void PhaseMacroExpand::migrate_outs(Node *old, Node *target) { 86 assert(old != nullptr, "sanity"); 87 for (DUIterator_Fast imax, i = old->fast_outs(imax); i < imax; i++) { 88 Node* use = old->fast_out(i); 89 _igvn.rehash_node_delayed(use); 90 imax -= replace_input(use, old, target); 91 // back up iterator 92 --i; 93 } 94 assert(old->outcnt() == 0, "all uses must be deleted"); 95 } 96 97 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) { 98 Node* cmp; 99 if (mask != 0) { 100 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask))); 101 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits))); 102 } else { 103 cmp = word; 104 } 105 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne)); 106 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); 107 transform_later(iff); 108 109 // Fast path taken. 110 Node *fast_taken = transform_later(new IfFalseNode(iff)); 111 112 // Fast path not-taken, i.e. slow path 113 Node *slow_taken = transform_later(new IfTrueNode(iff)); 114 115 if (return_fast_path) { 116 region->init_req(edge, slow_taken); // Capture slow-control 117 return fast_taken; 118 } else { 119 region->init_req(edge, fast_taken); // Capture fast-control 120 return slow_taken; 121 } 122 } 123 124 //--------------------copy_predefined_input_for_runtime_call-------------------- 125 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { 126 // Set fixed predefined input arguments 127 call->init_req( TypeFunc::Control, ctrl ); 128 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); 129 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? 130 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); 131 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); 132 } 133 134 //------------------------------make_slow_call--------------------------------- 135 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, 136 address slow_call, const char* leaf_name, Node* slow_path, 137 Node* parm0, Node* parm1, Node* parm2) { 138 139 // Slow-path call 140 CallNode *call = leaf_name 141 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) 142 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM ); 143 144 // Slow path call has no side-effects, uses few values 145 copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); 146 if (parm0 != nullptr) call->init_req(TypeFunc::Parms+0, parm0); 147 if (parm1 != nullptr) call->init_req(TypeFunc::Parms+1, parm1); 148 if (parm2 != nullptr) call->init_req(TypeFunc::Parms+2, parm2); 149 call->copy_call_debug_info(&_igvn, oldcall); 150 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 151 _igvn.replace_node(oldcall, call); 152 transform_later(call); 153 154 return call; 155 } 156 157 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) { 158 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2(); 159 bs->eliminate_gc_barrier(this, p2x); 160 #ifndef PRODUCT 161 if (PrintOptoStatistics) { 162 Atomic::inc(&PhaseMacroExpand::_GC_barriers_removed_counter); 163 } 164 #endif 165 } 166 167 // Search for a memory operation for the specified memory slice. 168 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) { 169 Node *orig_mem = mem; 170 Node *alloc_mem = alloc->in(TypeFunc::Memory); 171 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr(); 172 while (true) { 173 if (mem == alloc_mem || mem == start_mem ) { 174 return mem; // hit one of our sentinels 175 } else if (mem->is_MergeMem()) { 176 mem = mem->as_MergeMem()->memory_at(alias_idx); 177 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { 178 Node *in = mem->in(0); 179 // we can safely skip over safepoints, calls, locks and membars because we 180 // already know that the object is safe to eliminate. 181 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { 182 return in; 183 } else if (in->is_Call()) { 184 CallNode *call = in->as_Call(); 185 if (call->may_modify(tinst, phase)) { 186 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape"); 187 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) { 188 return in; 189 } 190 } 191 mem = in->in(TypeFunc::Memory); 192 } else if (in->is_MemBar()) { 193 ArrayCopyNode* ac = nullptr; 194 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) { 195 if (ac != nullptr) { 196 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone"); 197 return ac; 198 } 199 } 200 mem = in->in(TypeFunc::Memory); 201 } else { 202 #ifdef ASSERT 203 in->dump(); 204 mem->dump(); 205 assert(false, "unexpected projection"); 206 #endif 207 } 208 } else if (mem->is_Store()) { 209 const TypePtr* atype = mem->as_Store()->adr_type(); 210 int adr_idx = phase->C->get_alias_index(atype); 211 if (adr_idx == alias_idx) { 212 assert(atype->isa_oopptr(), "address type must be oopptr"); 213 int adr_offset = atype->offset(); 214 uint adr_iid = atype->is_oopptr()->instance_id(); 215 // Array elements references have the same alias_idx 216 // but different offset and different instance_id. 217 if (adr_offset == offset && adr_iid == alloc->_idx) { 218 return mem; 219 } 220 } else { 221 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw"); 222 } 223 mem = mem->in(MemNode::Memory); 224 } else if (mem->is_ClearArray()) { 225 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) { 226 // Can not bypass initialization of the instance 227 // we are looking. 228 debug_only(intptr_t offset;) 229 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity"); 230 InitializeNode* init = alloc->as_Allocate()->initialization(); 231 // We are looking for stored value, return Initialize node 232 // or memory edge from Allocate node. 233 if (init != nullptr) { 234 return init; 235 } else { 236 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers). 237 } 238 } 239 // Otherwise skip it (the call updated 'mem' value). 240 } else if (mem->Opcode() == Op_SCMemProj) { 241 mem = mem->in(0); 242 Node* adr = nullptr; 243 if (mem->is_LoadStore()) { 244 adr = mem->in(MemNode::Address); 245 } else { 246 assert(mem->Opcode() == Op_EncodeISOArray || 247 mem->Opcode() == Op_StrCompressedCopy, "sanity"); 248 adr = mem->in(3); // Destination array 249 } 250 const TypePtr* atype = adr->bottom_type()->is_ptr(); 251 int adr_idx = phase->C->get_alias_index(atype); 252 if (adr_idx == alias_idx) { 253 DEBUG_ONLY(mem->dump();) 254 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 255 return nullptr; 256 } 257 mem = mem->in(MemNode::Memory); 258 } else if (mem->Opcode() == Op_StrInflatedCopy) { 259 Node* adr = mem->in(3); // Destination array 260 const TypePtr* atype = adr->bottom_type()->is_ptr(); 261 int adr_idx = phase->C->get_alias_index(atype); 262 if (adr_idx == alias_idx) { 263 DEBUG_ONLY(mem->dump();) 264 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field"); 265 return nullptr; 266 } 267 mem = mem->in(MemNode::Memory); 268 } else { 269 return mem; 270 } 271 assert(mem != orig_mem, "dead memory loop"); 272 } 273 } 274 275 // Generate loads from source of the arraycopy for fields of 276 // destination needed at a deoptimization point 277 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) { 278 BasicType bt = ft; 279 const Type *type = ftype; 280 if (ft == T_NARROWOOP) { 281 bt = T_OBJECT; 282 type = ftype->make_oopptr(); 283 } 284 Node* res = nullptr; 285 if (ac->is_clonebasic()) { 286 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination"); 287 Node* base = ac->in(ArrayCopyNode::Src); 288 Node* adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(offset))); 289 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset); 290 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem(); 291 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 292 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt); 293 } else { 294 if (ac->modifies(offset, offset, &_igvn, true)) { 295 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result"); 296 uint shift = exact_log2(type2aelembytes(bt)); 297 Node* src_pos = ac->in(ArrayCopyNode::SrcPos); 298 Node* dest_pos = ac->in(ArrayCopyNode::DestPos); 299 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int(); 300 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int(); 301 302 Node* adr = nullptr; 303 const TypePtr* adr_type = nullptr; 304 if (src_pos_t->is_con() && dest_pos_t->is_con()) { 305 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset; 306 Node* base = ac->in(ArrayCopyNode::Src); 307 adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(off))); 308 adr_type = _igvn.type(base)->is_ptr()->add_offset(off); 309 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) { 310 // Don't emit a new load from src if src == dst but try to get the value from memory instead 311 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc); 312 } 313 } else { 314 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 315 #ifdef _LP64 316 diff = _igvn.transform(new ConvI2LNode(diff)); 317 #endif 318 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift))); 319 320 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff)); 321 Node* base = ac->in(ArrayCopyNode::Src); 322 adr = _igvn.transform(new AddPNode(base, base, off)); 323 adr_type = _igvn.type(base)->is_ptr()->add_offset(Type::OffsetBot); 324 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) { 325 // Non constant offset in the array: we can't statically 326 // determine the value 327 return nullptr; 328 } 329 } 330 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem(); 331 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 332 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt); 333 } 334 } 335 if (res != nullptr) { 336 if (ftype->isa_narrowoop()) { 337 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes 338 res = _igvn.transform(new EncodePNode(res, ftype)); 339 } 340 return res; 341 } 342 return nullptr; 343 } 344 345 // 346 // Given a Memory Phi, compute a value Phi containing the values from stores 347 // on the input paths. 348 // Note: this function is recursive, its depth is limited by the "level" argument 349 // Returns the computed Phi, or null if it cannot compute it. 350 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, AllocateNode *alloc, Node_Stack *value_phis, int level) { 351 assert(mem->is_Phi(), "sanity"); 352 int alias_idx = C->get_alias_index(adr_t); 353 int offset = adr_t->offset(); 354 int instance_id = adr_t->instance_id(); 355 356 // Check if an appropriate value phi already exists. 357 Node* region = mem->in(0); 358 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { 359 Node* phi = region->fast_out(k); 360 if (phi->is_Phi() && phi != mem && 361 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) { 362 return phi; 363 } 364 } 365 // Check if an appropriate new value phi already exists. 366 Node* new_phi = value_phis->find(mem->_idx); 367 if (new_phi != nullptr) 368 return new_phi; 369 370 if (level <= 0) { 371 return nullptr; // Give up: phi tree too deep 372 } 373 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 374 Node *alloc_mem = alloc->in(TypeFunc::Memory); 375 376 uint length = mem->req(); 377 GrowableArray <Node *> values(length, length, nullptr); 378 379 // create a new Phi for the value 380 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset); 381 transform_later(phi); 382 value_phis->push(phi, mem->_idx); 383 384 for (uint j = 1; j < length; j++) { 385 Node *in = mem->in(j); 386 if (in == nullptr || in->is_top()) { 387 values.at_put(j, in); 388 } else { 389 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); 390 if (val == start_mem || val == alloc_mem) { 391 // hit a sentinel, return appropriate 0 value 392 values.at_put(j, _igvn.zerocon(ft)); 393 continue; 394 } 395 if (val->is_Initialize()) { 396 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 397 } 398 if (val == nullptr) { 399 return nullptr; // can't find a value on this path 400 } 401 if (val == mem) { 402 values.at_put(j, mem); 403 } else if (val->is_Store()) { 404 Node* n = val->in(MemNode::ValueIn); 405 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 406 n = bs->step_over_gc_barrier(n); 407 if (is_subword_type(ft)) { 408 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true); 409 } 410 values.at_put(j, n); 411 } else if(val->is_Proj() && val->in(0) == alloc) { 412 values.at_put(j, _igvn.zerocon(ft)); 413 } else if (val->is_Phi()) { 414 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); 415 if (val == nullptr) { 416 return nullptr; 417 } 418 values.at_put(j, val); 419 } else if (val->Opcode() == Op_SCMemProj) { 420 assert(val->in(0)->is_LoadStore() || 421 val->in(0)->Opcode() == Op_EncodeISOArray || 422 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity"); 423 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 424 return nullptr; 425 } else if (val->is_ArrayCopy()) { 426 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc); 427 if (res == nullptr) { 428 return nullptr; 429 } 430 values.at_put(j, res); 431 } else { 432 DEBUG_ONLY( val->dump(); ) 433 assert(false, "unknown node on this path"); 434 return nullptr; // unknown node on this path 435 } 436 } 437 } 438 // Set Phi's inputs 439 for (uint j = 1; j < length; j++) { 440 if (values.at(j) == mem) { 441 phi->init_req(j, phi); 442 } else { 443 phi->init_req(j, values.at(j)); 444 } 445 } 446 return phi; 447 } 448 449 // Search the last value stored into the object's field. 450 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) { 451 assert(adr_t->is_known_instance_field(), "instance required"); 452 int instance_id = adr_t->instance_id(); 453 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 454 455 int alias_idx = C->get_alias_index(adr_t); 456 int offset = adr_t->offset(); 457 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 458 Node *alloc_ctrl = alloc->in(TypeFunc::Control); 459 Node *alloc_mem = alloc->in(TypeFunc::Memory); 460 VectorSet visited; 461 462 bool done = sfpt_mem == alloc_mem; 463 Node *mem = sfpt_mem; 464 while (!done) { 465 if (visited.test_set(mem->_idx)) { 466 return nullptr; // found a loop, give up 467 } 468 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 469 if (mem == start_mem || mem == alloc_mem) { 470 done = true; // hit a sentinel, return appropriate 0 value 471 } else if (mem->is_Initialize()) { 472 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 473 if (mem == nullptr) { 474 done = true; // Something go wrong. 475 } else if (mem->is_Store()) { 476 const TypePtr* atype = mem->as_Store()->adr_type(); 477 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 478 done = true; 479 } 480 } else if (mem->is_Store()) { 481 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 482 assert(atype != nullptr, "address type must be oopptr"); 483 assert(C->get_alias_index(atype) == alias_idx && 484 atype->is_known_instance_field() && atype->offset() == offset && 485 atype->instance_id() == instance_id, "store is correct memory slice"); 486 done = true; 487 } else if (mem->is_Phi()) { 488 // try to find a phi's unique input 489 Node *unique_input = nullptr; 490 Node *top = C->top(); 491 for (uint i = 1; i < mem->req(); i++) { 492 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 493 if (n == nullptr || n == top || n == mem) { 494 continue; 495 } else if (unique_input == nullptr) { 496 unique_input = n; 497 } else if (unique_input != n) { 498 unique_input = top; 499 break; 500 } 501 } 502 if (unique_input != nullptr && unique_input != top) { 503 mem = unique_input; 504 } else { 505 done = true; 506 } 507 } else if (mem->is_ArrayCopy()) { 508 done = true; 509 } else { 510 DEBUG_ONLY( mem->dump(); ) 511 assert(false, "unexpected node"); 512 } 513 } 514 if (mem != nullptr) { 515 if (mem == start_mem || mem == alloc_mem) { 516 // hit a sentinel, return appropriate 0 value 517 return _igvn.zerocon(ft); 518 } else if (mem->is_Store()) { 519 Node* n = mem->in(MemNode::ValueIn); 520 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 521 n = bs->step_over_gc_barrier(n); 522 return n; 523 } else if (mem->is_Phi()) { 524 // attempt to produce a Phi reflecting the values on the input paths of the Phi 525 Node_Stack value_phis(8); 526 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 527 if (phi != nullptr) { 528 return phi; 529 } else { 530 // Kill all new Phis 531 while(value_phis.is_nonempty()) { 532 Node* n = value_phis.node(); 533 _igvn.replace_node(n, C->top()); 534 value_phis.pop(); 535 } 536 } 537 } else if (mem->is_ArrayCopy()) { 538 Node* ctl = mem->in(0); 539 Node* m = mem->in(TypeFunc::Memory); 540 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj()) { 541 // pin the loads in the uncommon trap path 542 ctl = sfpt_ctl; 543 m = sfpt_mem; 544 } 545 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc); 546 } 547 } 548 // Something go wrong. 549 return nullptr; 550 } 551 552 // Check the possibility of scalar replacement. 553 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) { 554 // Scan the uses of the allocation to check for anything that would 555 // prevent us from eliminating it. 556 NOT_PRODUCT( const char* fail_eliminate = nullptr; ) 557 DEBUG_ONLY( Node* disq_node = nullptr; ) 558 bool can_eliminate = true; 559 bool reduce_merge_precheck = (safepoints == nullptr); 560 561 Node* res = alloc->result_cast(); 562 const TypeOopPtr* res_type = nullptr; 563 if (res == nullptr) { 564 // All users were eliminated. 565 } else if (!res->is_CheckCastPP()) { 566 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 567 can_eliminate = false; 568 } else { 569 res_type = igvn->type(res)->isa_oopptr(); 570 if (res_type == nullptr) { 571 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 572 can_eliminate = false; 573 } else if (res_type->isa_aryptr()) { 574 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 575 if (length < 0) { 576 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 577 can_eliminate = false; 578 } 579 } 580 } 581 582 if (can_eliminate && res != nullptr) { 583 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2(); 584 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 585 j < jmax && can_eliminate; j++) { 586 Node* use = res->fast_out(j); 587 588 if (use->is_AddP()) { 589 const TypePtr* addp_type = igvn->type(use)->is_ptr(); 590 int offset = addp_type->offset(); 591 592 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 593 NOT_PRODUCT(fail_eliminate = "Undefined field reference";) 594 can_eliminate = false; 595 break; 596 } 597 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 598 k < kmax && can_eliminate; k++) { 599 Node* n = use->fast_out(k); 600 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) { 601 DEBUG_ONLY(disq_node = n;) 602 if (n->is_Load() || n->is_LoadStore()) { 603 NOT_PRODUCT(fail_eliminate = "Field load";) 604 } else { 605 NOT_PRODUCT(fail_eliminate = "Not store field reference";) 606 } 607 can_eliminate = false; 608 } 609 } 610 } else if (use->is_ArrayCopy() && 611 (use->as_ArrayCopy()->is_clonebasic() || 612 use->as_ArrayCopy()->is_arraycopy_validated() || 613 use->as_ArrayCopy()->is_copyof_validated() || 614 use->as_ArrayCopy()->is_copyofrange_validated()) && 615 use->in(ArrayCopyNode::Dest) == res) { 616 // ok to eliminate 617 } else if (use->is_SafePoint()) { 618 SafePointNode* sfpt = use->as_SafePoint(); 619 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 620 // Object is passed as argument. 621 DEBUG_ONLY(disq_node = use;) 622 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 623 can_eliminate = false; 624 } 625 Node* sfptMem = sfpt->memory(); 626 if (sfptMem == nullptr || sfptMem->is_top()) { 627 DEBUG_ONLY(disq_node = use;) 628 NOT_PRODUCT(fail_eliminate = "null or TOP memory";) 629 can_eliminate = false; 630 } else if (!reduce_merge_precheck) { 631 safepoints->append_if_missing(sfpt); 632 } 633 } else if (reduce_merge_precheck && (use->is_Phi() || use->is_EncodeP() || use->Opcode() == Op_MemBarRelease)) { 634 // Nothing to do 635 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 636 if (use->is_Phi()) { 637 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 638 NOT_PRODUCT(fail_eliminate = "Object is return value";) 639 } else { 640 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 641 } 642 DEBUG_ONLY(disq_node = use;) 643 } else { 644 if (use->Opcode() == Op_Return) { 645 NOT_PRODUCT(fail_eliminate = "Object is return value";) 646 } else { 647 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 648 } 649 DEBUG_ONLY(disq_node = use;) 650 } 651 can_eliminate = false; 652 } 653 } 654 } 655 656 #ifndef PRODUCT 657 if (PrintEliminateAllocations && safepoints != nullptr) { 658 if (can_eliminate) { 659 tty->print("Scalar "); 660 if (res == nullptr) 661 alloc->dump(); 662 else 663 res->dump(); 664 } else if (alloc->_is_scalar_replaceable) { 665 tty->print("NotScalar (%s)", fail_eliminate); 666 if (res == nullptr) 667 alloc->dump(); 668 else 669 res->dump(); 670 #ifdef ASSERT 671 if (disq_node != nullptr) { 672 tty->print(" >>>> "); 673 disq_node->dump(); 674 } 675 #endif /*ASSERT*/ 676 } 677 } 678 679 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) { 680 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : ""); 681 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();) 682 } 683 #endif 684 return can_eliminate; 685 } 686 687 void PhaseMacroExpand::undo_previous_scalarizations(GrowableArray <SafePointNode *> safepoints_done, AllocateNode* alloc) { 688 Node* res = alloc->result_cast(); 689 int nfields = 0; 690 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result"); 691 692 if (res != nullptr) { 693 const TypeOopPtr* res_type = _igvn.type(res)->isa_oopptr(); 694 695 if (res_type->isa_instptr()) { 696 // find the fields of the class which will be needed for safepoint debug information 697 ciInstanceKlass* iklass = res_type->is_instptr()->instance_klass(); 698 nfields = iklass->nof_nonstatic_fields(); 699 } else { 700 // find the array's elements which will be needed for safepoint debug information 701 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 702 assert(nfields >= 0, "must be an array klass."); 703 } 704 } 705 706 // rollback processed safepoints 707 while (safepoints_done.length() > 0) { 708 SafePointNode* sfpt_done = safepoints_done.pop(); 709 // remove any extra entries we added to the safepoint 710 uint last = sfpt_done->req() - 1; 711 for (int k = 0; k < nfields; k++) { 712 sfpt_done->del_req(last--); 713 } 714 JVMState *jvms = sfpt_done->jvms(); 715 jvms->set_endoff(sfpt_done->req()); 716 // Now make a pass over the debug information replacing any references 717 // to SafePointScalarObjectNode with the allocated object. 718 int start = jvms->debug_start(); 719 int end = jvms->debug_end(); 720 for (int i = start; i < end; i++) { 721 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 722 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 723 if (scobj->first_index(jvms) == sfpt_done->req() && 724 scobj->n_fields() == (uint)nfields) { 725 assert(scobj->alloc() == alloc, "sanity"); 726 sfpt_done->set_req(i, res); 727 } 728 } 729 } 730 _igvn._worklist.push(sfpt_done); 731 } 732 } 733 734 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode *alloc, SafePointNode* sfpt) { 735 // Fields of scalar objs are referenced only at the end 736 // of regular debuginfo at the last (youngest) JVMS. 737 // Record relative start index. 738 ciInstanceKlass* iklass = nullptr; 739 BasicType basic_elem_type = T_ILLEGAL; 740 const Type* field_type = nullptr; 741 const TypeOopPtr* res_type = nullptr; 742 int nfields = 0; 743 int array_base = 0; 744 int element_size = 0; 745 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff()); 746 Node* res = alloc->result_cast(); 747 748 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result"); 749 assert(sfpt->jvms() != nullptr, "missed JVMS"); 750 751 if (res != nullptr) { // Could be null when there are no users 752 res_type = _igvn.type(res)->isa_oopptr(); 753 754 if (res_type->isa_instptr()) { 755 // find the fields of the class which will be needed for safepoint debug information 756 iklass = res_type->is_instptr()->instance_klass(); 757 nfields = iklass->nof_nonstatic_fields(); 758 } else { 759 // find the array's elements which will be needed for safepoint debug information 760 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 761 assert(nfields >= 0, "must be an array klass."); 762 basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type(); 763 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 764 element_size = type2aelembytes(basic_elem_type); 765 field_type = res_type->is_aryptr()->elem(); 766 } 767 } 768 769 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, nfields); 770 sobj->init_req(0, C->root()); 771 transform_later(sobj); 772 773 // Scan object's fields adding an input to the safepoint for each field. 774 for (int j = 0; j < nfields; j++) { 775 intptr_t offset; 776 ciField* field = nullptr; 777 if (iklass != nullptr) { 778 field = iklass->nonstatic_field_at(j); 779 offset = field->offset_in_bytes(); 780 ciType* elem_type = field->type(); 781 basic_elem_type = field->layout_type(); 782 783 // The next code is taken from Parse::do_get_xxx(). 784 if (is_reference_type(basic_elem_type)) { 785 if (!elem_type->is_loaded()) { 786 field_type = TypeInstPtr::BOTTOM; 787 } else if (field != nullptr && field->is_static_constant()) { 788 ciObject* con = field->constant_value().as_object(); 789 // Do not "join" in the previous type; it doesn't add value, 790 // and may yield a vacuous result if the field is of interface type. 791 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 792 assert(field_type != nullptr, "field singleton type must be consistent"); 793 } else { 794 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 795 } 796 if (UseCompressedOops) { 797 field_type = field_type->make_narrowoop(); 798 basic_elem_type = T_NARROWOOP; 799 } 800 } else { 801 field_type = Type::get_const_basic_type(basic_elem_type); 802 } 803 } else { 804 offset = array_base + j * (intptr_t)element_size; 805 } 806 807 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 808 809 Node *field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc); 810 811 // We weren't able to find a value for this field, 812 // give up on eliminating this allocation. 813 if (field_val == nullptr) { 814 uint last = sfpt->req() - 1; 815 for (int k = 0; k < j; k++) { 816 sfpt->del_req(last--); 817 } 818 _igvn._worklist.push(sfpt); 819 820 #ifndef PRODUCT 821 if (PrintEliminateAllocations) { 822 if (field != nullptr) { 823 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx); 824 field->print(); 825 int field_idx = C->get_alias_index(field_addr_type); 826 tty->print(" (alias_idx=%d)", field_idx); 827 } else { // Array's element 828 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, j); 829 } 830 tty->print(", which prevents elimination of: "); 831 if (res == nullptr) 832 alloc->dump(); 833 else 834 res->dump(); 835 } 836 #endif 837 838 return nullptr; 839 } 840 841 if (UseCompressedOops && field_type->isa_narrowoop()) { 842 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 843 // to be able scalar replace the allocation. 844 if (field_val->is_EncodeP()) { 845 field_val = field_val->in(1); 846 } else { 847 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type())); 848 } 849 } 850 sfpt->add_req(field_val); 851 } 852 853 sfpt->jvms()->set_endoff(sfpt->req()); 854 855 return sobj; 856 } 857 858 // Do scalar replacement. 859 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 860 GrowableArray <SafePointNode *> safepoints_done; 861 Node* res = alloc->result_cast(); 862 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result"); 863 864 // Process the safepoint uses 865 while (safepoints.length() > 0) { 866 SafePointNode* sfpt = safepoints.pop(); 867 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt); 868 869 if (sobj == nullptr) { 870 undo_previous_scalarizations(safepoints_done, alloc); 871 return false; 872 } 873 874 // Now make a pass over the debug information replacing any references 875 // to the allocated object with "sobj" 876 JVMState *jvms = sfpt->jvms(); 877 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn); 878 _igvn._worklist.push(sfpt); 879 880 // keep it for rollback 881 safepoints_done.append_if_missing(sfpt); 882 } 883 884 return true; 885 } 886 887 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) { 888 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control); 889 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory); 890 if (ctl_proj != nullptr) { 891 igvn.replace_node(ctl_proj, n->in(0)); 892 } 893 if (mem_proj != nullptr) { 894 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory)); 895 } 896 } 897 898 // Process users of eliminated allocation. 899 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) { 900 Node* res = alloc->result_cast(); 901 if (res != nullptr) { 902 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 903 Node *use = res->last_out(j); 904 uint oc1 = res->outcnt(); 905 906 if (use->is_AddP()) { 907 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 908 Node *n = use->last_out(k); 909 uint oc2 = use->outcnt(); 910 if (n->is_Store()) { 911 #ifdef ASSERT 912 // Verify that there is no dependent MemBarVolatile nodes, 913 // they should be removed during IGVN, see MemBarNode::Ideal(). 914 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); 915 p < pmax; p++) { 916 Node* mb = n->fast_out(p); 917 assert(mb->is_Initialize() || !mb->is_MemBar() || 918 mb->req() <= MemBarNode::Precedent || 919 mb->in(MemBarNode::Precedent) != n, 920 "MemBarVolatile should be eliminated for non-escaping object"); 921 } 922 #endif 923 _igvn.replace_node(n, n->in(MemNode::Memory)); 924 } else { 925 eliminate_gc_barrier(n); 926 } 927 k -= (oc2 - use->outcnt()); 928 } 929 _igvn.remove_dead_node(use); 930 } else if (use->is_ArrayCopy()) { 931 // Disconnect ArrayCopy node 932 ArrayCopyNode* ac = use->as_ArrayCopy(); 933 if (ac->is_clonebasic()) { 934 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out(); 935 disconnect_projections(ac, _igvn); 936 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation"); 937 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0); 938 disconnect_projections(membar_before->as_MemBar(), _igvn); 939 if (membar_after->is_MemBar()) { 940 disconnect_projections(membar_after->as_MemBar(), _igvn); 941 } 942 } else { 943 assert(ac->is_arraycopy_validated() || 944 ac->is_copyof_validated() || 945 ac->is_copyofrange_validated(), "unsupported"); 946 CallProjections callprojs; 947 ac->extract_projections(&callprojs, true); 948 949 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O)); 950 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory)); 951 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control)); 952 953 // Set control to top. IGVN will remove the remaining projections 954 ac->set_req(0, top()); 955 ac->replace_edge(res, top(), &_igvn); 956 957 // Disconnect src right away: it can help find new 958 // opportunities for allocation elimination 959 Node* src = ac->in(ArrayCopyNode::Src); 960 ac->replace_edge(src, top(), &_igvn); 961 // src can be top at this point if src and dest of the 962 // arraycopy were the same 963 if (src->outcnt() == 0 && !src->is_top()) { 964 _igvn.remove_dead_node(src); 965 } 966 } 967 _igvn._worklist.push(ac); 968 } else { 969 eliminate_gc_barrier(use); 970 } 971 j -= (oc1 - res->outcnt()); 972 } 973 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 974 _igvn.remove_dead_node(res); 975 } 976 977 // 978 // Process other users of allocation's projections 979 // 980 if (_callprojs.resproj != nullptr && _callprojs.resproj->outcnt() != 0) { 981 // First disconnect stores captured by Initialize node. 982 // If Initialize node is eliminated first in the following code, 983 // it will kill such stores and DUIterator_Last will assert. 984 for (DUIterator_Fast jmax, j = _callprojs.resproj->fast_outs(jmax); j < jmax; j++) { 985 Node* use = _callprojs.resproj->fast_out(j); 986 if (use->is_AddP()) { 987 // raw memory addresses used only by the initialization 988 _igvn.replace_node(use, C->top()); 989 --j; --jmax; 990 } 991 } 992 for (DUIterator_Last jmin, j = _callprojs.resproj->last_outs(jmin); j >= jmin; ) { 993 Node* use = _callprojs.resproj->last_out(j); 994 uint oc1 = _callprojs.resproj->outcnt(); 995 if (use->is_Initialize()) { 996 // Eliminate Initialize node. 997 InitializeNode *init = use->as_Initialize(); 998 assert(init->outcnt() <= 2, "only a control and memory projection expected"); 999 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control); 1000 if (ctrl_proj != nullptr) { 1001 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control)); 1002 #ifdef ASSERT 1003 // If the InitializeNode has no memory out, it will die, and tmp will become null 1004 Node* tmp = init->in(TypeFunc::Control); 1005 assert(tmp == nullptr || tmp == _callprojs.fallthrough_catchproj, "allocation control projection"); 1006 #endif 1007 } 1008 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory); 1009 if (mem_proj != nullptr) { 1010 Node *mem = init->in(TypeFunc::Memory); 1011 #ifdef ASSERT 1012 if (mem->is_MergeMem()) { 1013 assert(mem->in(TypeFunc::Memory) == _callprojs.fallthrough_memproj, "allocation memory projection"); 1014 } else { 1015 assert(mem == _callprojs.fallthrough_memproj, "allocation memory projection"); 1016 } 1017 #endif 1018 _igvn.replace_node(mem_proj, mem); 1019 } 1020 } else { 1021 assert(false, "only Initialize or AddP expected"); 1022 } 1023 j -= (oc1 - _callprojs.resproj->outcnt()); 1024 } 1025 } 1026 if (_callprojs.fallthrough_catchproj != nullptr) { 1027 _igvn.replace_node(_callprojs.fallthrough_catchproj, alloc->in(TypeFunc::Control)); 1028 } 1029 if (_callprojs.fallthrough_memproj != nullptr) { 1030 _igvn.replace_node(_callprojs.fallthrough_memproj, alloc->in(TypeFunc::Memory)); 1031 } 1032 if (_callprojs.catchall_memproj != nullptr) { 1033 _igvn.replace_node(_callprojs.catchall_memproj, C->top()); 1034 } 1035 if (_callprojs.fallthrough_ioproj != nullptr) { 1036 _igvn.replace_node(_callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 1037 } 1038 if (_callprojs.catchall_ioproj != nullptr) { 1039 _igvn.replace_node(_callprojs.catchall_ioproj, C->top()); 1040 } 1041 if (_callprojs.catchall_catchproj != nullptr) { 1042 _igvn.replace_node(_callprojs.catchall_catchproj, C->top()); 1043 } 1044 } 1045 1046 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 1047 // If reallocation fails during deoptimization we'll pop all 1048 // interpreter frames for this compiled frame and that won't play 1049 // nice with JVMTI popframe. 1050 // We avoid this issue by eager reallocation when the popframe request 1051 // is received. 1052 if (!EliminateAllocations || !alloc->_is_non_escaping) { 1053 return false; 1054 } 1055 Node* klass = alloc->in(AllocateNode::KlassNode); 1056 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); 1057 Node* res = alloc->result_cast(); 1058 // Eliminate boxing allocations which are not used 1059 // regardless scalar replaceable status. 1060 bool boxing_alloc = C->eliminate_boxing() && 1061 tklass->isa_instklassptr() && 1062 tklass->is_instklassptr()->instance_klass()->is_box_klass(); 1063 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != nullptr))) { 1064 return false; 1065 } 1066 1067 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/); 1068 1069 GrowableArray <SafePointNode *> safepoints; 1070 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) { 1071 return false; 1072 } 1073 1074 if (!alloc->_is_scalar_replaceable) { 1075 assert(res == nullptr, "sanity"); 1076 // We can only eliminate allocation if all debug info references 1077 // are already replaced with SafePointScalarObject because 1078 // we can't search for a fields value without instance_id. 1079 if (safepoints.length() > 0) { 1080 return false; 1081 } 1082 } 1083 1084 if (!scalar_replacement(alloc, safepoints)) { 1085 return false; 1086 } 1087 1088 CompileLog* log = C->log(); 1089 if (log != nullptr) { 1090 log->head("eliminate_allocation type='%d'", 1091 log->identify(tklass->exact_klass())); 1092 JVMState* p = alloc->jvms(); 1093 while (p != nullptr) { 1094 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1095 p = p->caller(); 1096 } 1097 log->tail("eliminate_allocation"); 1098 } 1099 1100 process_users_of_allocation(alloc); 1101 1102 #ifndef PRODUCT 1103 if (PrintEliminateAllocations) { 1104 if (alloc->is_AllocateArray()) 1105 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1106 else 1107 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1108 } 1109 #endif 1110 1111 return true; 1112 } 1113 1114 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) { 1115 // EA should remove all uses of non-escaping boxing node. 1116 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) { 1117 return false; 1118 } 1119 1120 assert(boxing->result_cast() == nullptr, "unexpected boxing node result"); 1121 1122 boxing->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/); 1123 1124 const TypeTuple* r = boxing->tf()->range(); 1125 assert(r->cnt() > TypeFunc::Parms, "sanity"); 1126 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr(); 1127 assert(t != nullptr, "sanity"); 1128 1129 CompileLog* log = C->log(); 1130 if (log != nullptr) { 1131 log->head("eliminate_boxing type='%d'", 1132 log->identify(t->instance_klass())); 1133 JVMState* p = boxing->jvms(); 1134 while (p != nullptr) { 1135 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1136 p = p->caller(); 1137 } 1138 log->tail("eliminate_boxing"); 1139 } 1140 1141 process_users_of_allocation(boxing); 1142 1143 #ifndef PRODUCT 1144 if (PrintEliminateAllocations) { 1145 tty->print("++++ Eliminated: %d ", boxing->_idx); 1146 boxing->method()->print_short_name(tty); 1147 tty->cr(); 1148 } 1149 #endif 1150 1151 return true; 1152 } 1153 1154 1155 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 1156 Node* adr = basic_plus_adr(base, offset); 1157 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 1158 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered); 1159 transform_later(value); 1160 return value; 1161 } 1162 1163 1164 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 1165 Node* adr = basic_plus_adr(base, offset); 1166 mem = StoreNode::make(_igvn, ctl, mem, adr, nullptr, value, bt, MemNode::unordered); 1167 transform_later(mem); 1168 return mem; 1169 } 1170 1171 //============================================================================= 1172 // 1173 // A L L O C A T I O N 1174 // 1175 // Allocation attempts to be fast in the case of frequent small objects. 1176 // It breaks down like this: 1177 // 1178 // 1) Size in doublewords is computed. This is a constant for objects and 1179 // variable for most arrays. Doubleword units are used to avoid size 1180 // overflow of huge doubleword arrays. We need doublewords in the end for 1181 // rounding. 1182 // 1183 // 2) Size is checked for being 'too large'. Too-large allocations will go 1184 // the slow path into the VM. The slow path can throw any required 1185 // exceptions, and does all the special checks for very large arrays. The 1186 // size test can constant-fold away for objects. For objects with 1187 // finalizers it constant-folds the otherway: you always go slow with 1188 // finalizers. 1189 // 1190 // 3) If NOT using TLABs, this is the contended loop-back point. 1191 // Load-Locked the heap top. If using TLABs normal-load the heap top. 1192 // 1193 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 1194 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 1195 // "size*8" we always enter the VM, where "largish" is a constant picked small 1196 // enough that there's always space between the eden max and 4Gig (old space is 1197 // there so it's quite large) and large enough that the cost of entering the VM 1198 // is dwarfed by the cost to initialize the space. 1199 // 1200 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 1201 // down. If contended, repeat at step 3. If using TLABs normal-store 1202 // adjusted heap top back down; there is no contention. 1203 // 1204 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 1205 // fields. 1206 // 1207 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 1208 // oop flavor. 1209 // 1210 //============================================================================= 1211 // FastAllocateSizeLimit value is in DOUBLEWORDS. 1212 // Allocations bigger than this always go the slow route. 1213 // This value must be small enough that allocation attempts that need to 1214 // trigger exceptions go the slow route. Also, it must be small enough so 1215 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 1216 //=============================================================================j// 1217 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 1218 // The allocator will coalesce int->oop copies away. See comment in 1219 // coalesce.cpp about how this works. It depends critically on the exact 1220 // code shape produced here, so if you are changing this code shape 1221 // make sure the GC info for the heap-top is correct in and around the 1222 // slow-path call. 1223 // 1224 1225 void PhaseMacroExpand::expand_allocate_common( 1226 AllocateNode* alloc, // allocation node to be expanded 1227 Node* length, // array length for an array allocation 1228 const TypeFunc* slow_call_type, // Type of slow call 1229 address slow_call_address, // Address of slow call 1230 Node* valid_length_test // whether length is valid or not 1231 ) 1232 { 1233 Node* ctrl = alloc->in(TypeFunc::Control); 1234 Node* mem = alloc->in(TypeFunc::Memory); 1235 Node* i_o = alloc->in(TypeFunc::I_O); 1236 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 1237 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1238 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 1239 assert(ctrl != nullptr, "must have control"); 1240 1241 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 1242 // they will not be used if "always_slow" is set 1243 enum { slow_result_path = 1, fast_result_path = 2 }; 1244 Node *result_region = nullptr; 1245 Node *result_phi_rawmem = nullptr; 1246 Node *result_phi_rawoop = nullptr; 1247 Node *result_phi_i_o = nullptr; 1248 1249 // The initial slow comparison is a size check, the comparison 1250 // we want to do is a BoolTest::gt 1251 bool expand_fast_path = true; 1252 int tv = _igvn.find_int_con(initial_slow_test, -1); 1253 if (tv >= 0) { 1254 // InitialTest has constant result 1255 // 0 - can fit in TLAB 1256 // 1 - always too big or negative 1257 assert(tv <= 1, "0 or 1 if a constant"); 1258 expand_fast_path = (tv == 0); 1259 initial_slow_test = nullptr; 1260 } else { 1261 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 1262 } 1263 1264 if (!UseTLAB) { 1265 // Force slow-path allocation 1266 expand_fast_path = false; 1267 initial_slow_test = nullptr; 1268 } 1269 1270 bool allocation_has_use = (alloc->result_cast() != nullptr); 1271 if (!allocation_has_use) { 1272 InitializeNode* init = alloc->initialization(); 1273 if (init != nullptr) { 1274 init->remove(&_igvn); 1275 } 1276 if (expand_fast_path && (initial_slow_test == nullptr)) { 1277 // Remove allocation node and return. 1278 // Size is a non-negative constant -> no initial check needed -> directly to fast path. 1279 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left. 1280 #ifndef PRODUCT 1281 if (PrintEliminateAllocations) { 1282 tty->print("NotUsed "); 1283 Node* res = alloc->proj_out_or_null(TypeFunc::Parms); 1284 if (res != nullptr) { 1285 res->dump(); 1286 } else { 1287 alloc->dump(); 1288 } 1289 } 1290 #endif 1291 yank_alloc_node(alloc); 1292 return; 1293 } 1294 } 1295 1296 enum { too_big_or_final_path = 1, need_gc_path = 2 }; 1297 Node *slow_region = nullptr; 1298 Node *toobig_false = ctrl; 1299 1300 if (PEAParanoid && alloc->materialized_cnt() > 0) { 1301 fatal("[PEA] Expanding obj#%d which has been materialized.", alloc->_idx); 1302 } 1303 1304 // generate the initial test if necessary 1305 if (initial_slow_test != nullptr ) { 1306 assert (expand_fast_path, "Only need test if there is a fast path"); 1307 slow_region = new RegionNode(3); 1308 1309 // Now make the initial failure test. Usually a too-big test but 1310 // might be a TRUE for finalizers or a fancy class check for 1311 // newInstance0. 1312 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1313 transform_later(toobig_iff); 1314 // Plug the failing-too-big test into the slow-path region 1315 Node *toobig_true = new IfTrueNode( toobig_iff ); 1316 transform_later(toobig_true); 1317 slow_region ->init_req( too_big_or_final_path, toobig_true ); 1318 toobig_false = new IfFalseNode( toobig_iff ); 1319 transform_later(toobig_false); 1320 } else { 1321 // No initial test, just fall into next case 1322 assert(allocation_has_use || !expand_fast_path, "Should already have been handled"); 1323 toobig_false = ctrl; 1324 debug_only(slow_region = NodeSentinel); 1325 } 1326 1327 // If we are here there are several possibilities 1328 // - expand_fast_path is false - then only a slow path is expanded. That's it. 1329 // no_initial_check means a constant allocation. 1330 // - If check always evaluates to false -> expand_fast_path is false (see above) 1331 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath) 1332 // if !allocation_has_use the fast path is empty 1333 // if !allocation_has_use && no_initial_check 1334 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all. 1335 // removed by yank_alloc_node above. 1336 1337 Node *slow_mem = mem; // save the current memory state for slow path 1338 // generate the fast allocation code unless we know that the initial test will always go slow 1339 if (expand_fast_path) { 1340 // Fast path modifies only raw memory. 1341 if (mem->is_MergeMem()) { 1342 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1343 } 1344 1345 // allocate the Region and Phi nodes for the result 1346 result_region = new RegionNode(3); 1347 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1348 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch 1349 1350 // Grab regular I/O before optional prefetch may change it. 1351 // Slow-path does no I/O so just set it to the original I/O. 1352 result_phi_i_o->init_req(slow_result_path, i_o); 1353 1354 // Name successful fast-path variables 1355 Node* fast_oop_ctrl; 1356 Node* fast_oop_rawmem; 1357 if (allocation_has_use) { 1358 Node* needgc_ctrl = nullptr; 1359 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM); 1360 1361 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1362 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1363 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl, 1364 fast_oop_ctrl, fast_oop_rawmem, 1365 prefetch_lines); 1366 1367 if (initial_slow_test != nullptr) { 1368 // This completes all paths into the slow merge point 1369 slow_region->init_req(need_gc_path, needgc_ctrl); 1370 transform_later(slow_region); 1371 } else { 1372 // No initial slow path needed! 1373 // Just fall from the need-GC path straight into the VM call. 1374 slow_region = needgc_ctrl; 1375 } 1376 1377 InitializeNode* init = alloc->initialization(); 1378 fast_oop_rawmem = initialize_object(alloc, 1379 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1380 klass_node, length, size_in_bytes); 1381 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem); 1382 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem); 1383 1384 result_phi_rawoop->init_req(fast_result_path, fast_oop); 1385 } else { 1386 assert (initial_slow_test != nullptr, "sanity"); 1387 fast_oop_ctrl = toobig_false; 1388 fast_oop_rawmem = mem; 1389 transform_later(slow_region); 1390 } 1391 1392 // Plug in the successful fast-path into the result merge point 1393 result_region ->init_req(fast_result_path, fast_oop_ctrl); 1394 result_phi_i_o ->init_req(fast_result_path, i_o); 1395 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); 1396 } else { 1397 slow_region = ctrl; 1398 result_phi_i_o = i_o; // Rename it to use in the following code. 1399 } 1400 1401 // Generate slow-path call 1402 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address, 1403 OptoRuntime::stub_name(slow_call_address), 1404 TypePtr::BOTTOM); 1405 call->init_req(TypeFunc::Control, slow_region); 1406 call->init_req(TypeFunc::I_O, top()); // does no i/o 1407 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs 1408 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1409 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1410 1411 call->init_req(TypeFunc::Parms+0, klass_node); 1412 if (length != nullptr) { 1413 call->init_req(TypeFunc::Parms+1, length); 1414 } 1415 1416 // Copy debug information and adjust JVMState information, then replace 1417 // allocate node with the call 1418 call->copy_call_debug_info(&_igvn, alloc); 1419 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify 1420 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough 1421 // path dies). 1422 if (valid_length_test != nullptr) { 1423 call->add_req(valid_length_test); 1424 } 1425 if (expand_fast_path) { 1426 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1427 } else { 1428 // Hook i_o projection to avoid its elimination during allocation 1429 // replacement (when only a slow call is generated). 1430 call->set_req(TypeFunc::I_O, result_phi_i_o); 1431 } 1432 _igvn.replace_node(alloc, call); 1433 transform_later(call); 1434 1435 // Identify the output projections from the allocate node and 1436 // adjust any references to them. 1437 // The control and io projections look like: 1438 // 1439 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1440 // Allocate Catch 1441 // ^---Proj(io) <-------+ ^---CatchProj(io) 1442 // 1443 // We are interested in the CatchProj nodes. 1444 // 1445 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/); 1446 1447 // An allocate node has separate memory projections for the uses on 1448 // the control and i_o paths. Replace the control memory projection with 1449 // result_phi_rawmem (unless we are only generating a slow call when 1450 // both memory projections are combined) 1451 if (expand_fast_path && _callprojs.fallthrough_memproj != nullptr) { 1452 migrate_outs(_callprojs.fallthrough_memproj, result_phi_rawmem); 1453 } 1454 // Now change uses of catchall_memproj to use fallthrough_memproj and delete 1455 // catchall_memproj so we end up with a call that has only 1 memory projection. 1456 if (_callprojs.catchall_memproj != nullptr ) { 1457 if (_callprojs.fallthrough_memproj == nullptr) { 1458 _callprojs.fallthrough_memproj = new ProjNode(call, TypeFunc::Memory); 1459 transform_later(_callprojs.fallthrough_memproj); 1460 } 1461 migrate_outs(_callprojs.catchall_memproj, _callprojs.fallthrough_memproj); 1462 _igvn.remove_dead_node(_callprojs.catchall_memproj); 1463 } 1464 1465 // An allocate node has separate i_o projections for the uses on the control 1466 // and i_o paths. Always replace the control i_o projection with result i_o 1467 // otherwise incoming i_o become dead when only a slow call is generated 1468 // (it is different from memory projections where both projections are 1469 // combined in such case). 1470 if (_callprojs.fallthrough_ioproj != nullptr) { 1471 migrate_outs(_callprojs.fallthrough_ioproj, result_phi_i_o); 1472 } 1473 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete 1474 // catchall_ioproj so we end up with a call that has only 1 i_o projection. 1475 if (_callprojs.catchall_ioproj != nullptr ) { 1476 if (_callprojs.fallthrough_ioproj == nullptr) { 1477 _callprojs.fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O); 1478 transform_later(_callprojs.fallthrough_ioproj); 1479 } 1480 migrate_outs(_callprojs.catchall_ioproj, _callprojs.fallthrough_ioproj); 1481 _igvn.remove_dead_node(_callprojs.catchall_ioproj); 1482 } 1483 1484 // if we generated only a slow call, we are done 1485 if (!expand_fast_path) { 1486 // Now we can unhook i_o. 1487 if (result_phi_i_o->outcnt() > 1) { 1488 call->set_req(TypeFunc::I_O, top()); 1489 } else { 1490 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity"); 1491 // Case of new array with negative size known during compilation. 1492 // AllocateArrayNode::Ideal() optimization disconnect unreachable 1493 // following code since call to runtime will throw exception. 1494 // As result there will be no users of i_o after the call. 1495 // Leave i_o attached to this call to avoid problems in preceding graph. 1496 } 1497 return; 1498 } 1499 1500 if (_callprojs.fallthrough_catchproj != nullptr) { 1501 ctrl = _callprojs.fallthrough_catchproj->clone(); 1502 transform_later(ctrl); 1503 _igvn.replace_node(_callprojs.fallthrough_catchproj, result_region); 1504 } else { 1505 ctrl = top(); 1506 } 1507 Node *slow_result; 1508 if (_callprojs.resproj == nullptr) { 1509 // no uses of the allocation result 1510 slow_result = top(); 1511 } else { 1512 slow_result = _callprojs.resproj->clone(); 1513 transform_later(slow_result); 1514 _igvn.replace_node(_callprojs.resproj, result_phi_rawoop); 1515 } 1516 1517 // Plug slow-path into result merge point 1518 result_region->init_req( slow_result_path, ctrl); 1519 transform_later(result_region); 1520 if (allocation_has_use) { 1521 result_phi_rawoop->init_req(slow_result_path, slow_result); 1522 transform_later(result_phi_rawoop); 1523 } 1524 result_phi_rawmem->init_req(slow_result_path, _callprojs.fallthrough_memproj); 1525 transform_later(result_phi_rawmem); 1526 transform_later(result_phi_i_o); 1527 // This completes all paths into the result merge point 1528 } 1529 1530 // Remove alloc node that has no uses. 1531 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) { 1532 Node* ctrl = alloc->in(TypeFunc::Control); 1533 Node* mem = alloc->in(TypeFunc::Memory); 1534 Node* i_o = alloc->in(TypeFunc::I_O); 1535 1536 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/); 1537 if (_callprojs.resproj != nullptr) { 1538 for (DUIterator_Fast imax, i = _callprojs.resproj->fast_outs(imax); i < imax; i++) { 1539 Node* use = _callprojs.resproj->fast_out(i); 1540 use->isa_MemBar()->remove(&_igvn); 1541 --imax; 1542 --i; // back up iterator 1543 } 1544 assert(_callprojs.resproj->outcnt() == 0, "all uses must be deleted"); 1545 _igvn.remove_dead_node(_callprojs.resproj); 1546 } 1547 if (_callprojs.fallthrough_catchproj != nullptr) { 1548 migrate_outs(_callprojs.fallthrough_catchproj, ctrl); 1549 _igvn.remove_dead_node(_callprojs.fallthrough_catchproj); 1550 } 1551 if (_callprojs.catchall_catchproj != nullptr) { 1552 _igvn.rehash_node_delayed(_callprojs.catchall_catchproj); 1553 _callprojs.catchall_catchproj->set_req(0, top()); 1554 } 1555 if (_callprojs.fallthrough_proj != nullptr) { 1556 Node* catchnode = _callprojs.fallthrough_proj->unique_ctrl_out(); 1557 _igvn.remove_dead_node(catchnode); 1558 _igvn.remove_dead_node(_callprojs.fallthrough_proj); 1559 } 1560 if (_callprojs.fallthrough_memproj != nullptr) { 1561 migrate_outs(_callprojs.fallthrough_memproj, mem); 1562 _igvn.remove_dead_node(_callprojs.fallthrough_memproj); 1563 } 1564 if (_callprojs.fallthrough_ioproj != nullptr) { 1565 migrate_outs(_callprojs.fallthrough_ioproj, i_o); 1566 _igvn.remove_dead_node(_callprojs.fallthrough_ioproj); 1567 } 1568 if (_callprojs.catchall_memproj != nullptr) { 1569 _igvn.rehash_node_delayed(_callprojs.catchall_memproj); 1570 _callprojs.catchall_memproj->set_req(0, top()); 1571 } 1572 if (_callprojs.catchall_ioproj != nullptr) { 1573 _igvn.rehash_node_delayed(_callprojs.catchall_ioproj); 1574 _callprojs.catchall_ioproj->set_req(0, top()); 1575 } 1576 #ifndef PRODUCT 1577 if (PrintEliminateAllocations) { 1578 if (alloc->is_AllocateArray()) { 1579 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1580 } else { 1581 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1582 } 1583 } 1584 #endif 1585 _igvn.remove_dead_node(alloc); 1586 } 1587 1588 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init, 1589 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) { 1590 // If initialization is performed by an array copy, any required 1591 // MemBarStoreStore was already added. If the object does not 1592 // escape no need for a MemBarStoreStore. If the object does not 1593 // escape in its initializer and memory barrier (MemBarStoreStore or 1594 // stronger) is already added at exit of initializer, also no need 1595 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore 1596 // so that stores that initialize this object can't be reordered 1597 // with a subsequent store that makes this object accessible by 1598 // other threads. 1599 // Other threads include java threads and JVM internal threads 1600 // (for example concurrent GC threads). Current concurrent GC 1601 // implementation: G1 will not scan newly created object, 1602 // so it's safe to skip storestore barrier when allocation does 1603 // not escape. 1604 if (!alloc->does_not_escape_thread() && 1605 !alloc->is_allocation_MemBar_redundant() && 1606 (init == nullptr || !init->is_complete_with_arraycopy())) { 1607 if (init == nullptr || init->req() < InitializeNode::RawStores) { 1608 // No InitializeNode or no stores captured by zeroing 1609 // elimination. Simply add the MemBarStoreStore after object 1610 // initialization. 1611 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1612 transform_later(mb); 1613 1614 mb->init_req(TypeFunc::Memory, fast_oop_rawmem); 1615 mb->init_req(TypeFunc::Control, fast_oop_ctrl); 1616 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control); 1617 transform_later(fast_oop_ctrl); 1618 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory); 1619 transform_later(fast_oop_rawmem); 1620 } else { 1621 // Add the MemBarStoreStore after the InitializeNode so that 1622 // all stores performing the initialization that were moved 1623 // before the InitializeNode happen before the storestore 1624 // barrier. 1625 1626 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control); 1627 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory); 1628 1629 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1630 transform_later(mb); 1631 1632 Node* ctrl = new ProjNode(init, TypeFunc::Control); 1633 transform_later(ctrl); 1634 Node* mem = new ProjNode(init, TypeFunc::Memory); 1635 transform_later(mem); 1636 1637 // The MemBarStoreStore depends on control and memory coming 1638 // from the InitializeNode 1639 mb->init_req(TypeFunc::Memory, mem); 1640 mb->init_req(TypeFunc::Control, ctrl); 1641 1642 ctrl = new ProjNode(mb, TypeFunc::Control); 1643 transform_later(ctrl); 1644 mem = new ProjNode(mb, TypeFunc::Memory); 1645 transform_later(mem); 1646 1647 // All nodes that depended on the InitializeNode for control 1648 // and memory must now depend on the MemBarNode that itself 1649 // depends on the InitializeNode 1650 if (init_ctrl != nullptr) { 1651 _igvn.replace_node(init_ctrl, ctrl); 1652 } 1653 if (init_mem != nullptr) { 1654 _igvn.replace_node(init_mem, mem); 1655 } 1656 } 1657 } 1658 } 1659 1660 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop, 1661 Node*& ctrl, Node*& rawmem) { 1662 if (C->env()->dtrace_alloc_probes()) { 1663 // Slow-path call 1664 int size = TypeFunc::Parms + 2; 1665 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1666 CAST_FROM_FN_PTR(address, 1667 static_cast<int (*)(JavaThread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)), 1668 "dtrace_object_alloc", 1669 TypeRawPtr::BOTTOM); 1670 1671 // Get base of thread-local storage area 1672 Node* thread = new ThreadLocalNode(); 1673 transform_later(thread); 1674 1675 call->init_req(TypeFunc::Parms + 0, thread); 1676 call->init_req(TypeFunc::Parms + 1, oop); 1677 call->init_req(TypeFunc::Control, ctrl); 1678 call->init_req(TypeFunc::I_O , top()); // does no i/o 1679 call->init_req(TypeFunc::Memory , rawmem); 1680 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1681 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1682 transform_later(call); 1683 ctrl = new ProjNode(call, TypeFunc::Control); 1684 transform_later(ctrl); 1685 rawmem = new ProjNode(call, TypeFunc::Memory); 1686 transform_later(rawmem); 1687 } 1688 } 1689 1690 // Helper for PhaseMacroExpand::expand_allocate_common. 1691 // Initializes the newly-allocated storage. 1692 Node* 1693 PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1694 Node* control, Node* rawmem, Node* object, 1695 Node* klass_node, Node* length, 1696 Node* size_in_bytes) { 1697 InitializeNode* init = alloc->initialization(); 1698 // Store the klass & mark bits 1699 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem); 1700 if (!mark_node->is_Con()) { 1701 transform_later(mark_node); 1702 } 1703 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type()); 1704 1705 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 1706 int header_size = alloc->minimum_header_size(); // conservatively small 1707 1708 // Array length 1709 if (length != nullptr) { // Arrays need length field 1710 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1711 // conservatively small header size: 1712 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1713 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases: 1714 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type(); 1715 if (is_reference_type(elem, true)) { 1716 elem = T_OBJECT; 1717 } 1718 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem)); 1719 } 1720 } 1721 1722 // Clear the object body, if necessary. 1723 if (init == nullptr) { 1724 // The init has somehow disappeared; be cautious and clear everything. 1725 // 1726 // This can happen if a node is allocated but an uncommon trap occurs 1727 // immediately. In this case, the Initialize gets associated with the 1728 // trap, and may be placed in a different (outer) loop, if the Allocate 1729 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1730 // there can be two Allocates to one Initialize. The answer in all these 1731 // edge cases is safety first. It is always safe to clear immediately 1732 // within an Allocate, and then (maybe or maybe not) clear some more later. 1733 if (!(UseTLAB && ZeroTLAB)) { 1734 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1735 header_size, size_in_bytes, 1736 &_igvn); 1737 } 1738 } else { 1739 if (!init->is_complete()) { 1740 // Try to win by zeroing only what the init does not store. 1741 // We can also try to do some peephole optimizations, 1742 // such as combining some adjacent subword stores. 1743 rawmem = init->complete_stores(control, rawmem, object, 1744 header_size, size_in_bytes, &_igvn); 1745 } 1746 // We have no more use for this link, since the AllocateNode goes away: 1747 init->set_req(InitializeNode::RawAddress, top()); 1748 // (If we keep the link, it just confuses the register allocator, 1749 // who thinks he sees a real use of the address by the membar.) 1750 } 1751 1752 return rawmem; 1753 } 1754 1755 // Generate prefetch instructions for next allocations. 1756 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1757 Node*& contended_phi_rawmem, 1758 Node* old_eden_top, Node* new_eden_top, 1759 intx lines) { 1760 enum { fall_in_path = 1, pf_path = 2 }; 1761 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1762 // Generate prefetch allocation with watermark check. 1763 // As an allocation hits the watermark, we will prefetch starting 1764 // at a "distance" away from watermark. 1765 1766 Node *pf_region = new RegionNode(3); 1767 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 1768 TypeRawPtr::BOTTOM ); 1769 // I/O is used for Prefetch 1770 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO ); 1771 1772 Node *thread = new ThreadLocalNode(); 1773 transform_later(thread); 1774 1775 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread, 1776 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1777 transform_later(eden_pf_adr); 1778 1779 Node *old_pf_wm = new LoadPNode(needgc_false, 1780 contended_phi_rawmem, eden_pf_adr, 1781 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, 1782 MemNode::unordered); 1783 transform_later(old_pf_wm); 1784 1785 // check against new_eden_top 1786 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm ); 1787 transform_later(need_pf_cmp); 1788 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge ); 1789 transform_later(need_pf_bol); 1790 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol, 1791 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1792 transform_later(need_pf_iff); 1793 1794 // true node, add prefetchdistance 1795 Node *need_pf_true = new IfTrueNode( need_pf_iff ); 1796 transform_later(need_pf_true); 1797 1798 Node *need_pf_false = new IfFalseNode( need_pf_iff ); 1799 transform_later(need_pf_false); 1800 1801 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm, 1802 _igvn.MakeConX(AllocatePrefetchDistance) ); 1803 transform_later(new_pf_wmt ); 1804 new_pf_wmt->set_req(0, need_pf_true); 1805 1806 Node *store_new_wmt = new StorePNode(need_pf_true, 1807 contended_phi_rawmem, eden_pf_adr, 1808 TypeRawPtr::BOTTOM, new_pf_wmt, 1809 MemNode::unordered); 1810 transform_later(store_new_wmt); 1811 1812 // adding prefetches 1813 pf_phi_abio->init_req( fall_in_path, i_o ); 1814 1815 Node *prefetch_adr; 1816 Node *prefetch; 1817 uint step_size = AllocatePrefetchStepSize; 1818 uint distance = 0; 1819 1820 for ( intx i = 0; i < lines; i++ ) { 1821 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt, 1822 _igvn.MakeConX(distance) ); 1823 transform_later(prefetch_adr); 1824 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1825 transform_later(prefetch); 1826 distance += step_size; 1827 i_o = prefetch; 1828 } 1829 pf_phi_abio->set_req( pf_path, i_o ); 1830 1831 pf_region->init_req( fall_in_path, need_pf_false ); 1832 pf_region->init_req( pf_path, need_pf_true ); 1833 1834 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 1835 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 1836 1837 transform_later(pf_region); 1838 transform_later(pf_phi_rawmem); 1839 transform_later(pf_phi_abio); 1840 1841 needgc_false = pf_region; 1842 contended_phi_rawmem = pf_phi_rawmem; 1843 i_o = pf_phi_abio; 1844 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { 1845 // Insert a prefetch instruction for each allocation. 1846 // This code is used to generate 1 prefetch instruction per cache line. 1847 1848 // Generate several prefetch instructions. 1849 uint step_size = AllocatePrefetchStepSize; 1850 uint distance = AllocatePrefetchDistance; 1851 1852 // Next cache address. 1853 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top, 1854 _igvn.MakeConX(step_size + distance)); 1855 transform_later(cache_adr); 1856 cache_adr = new CastP2XNode(needgc_false, cache_adr); 1857 transform_later(cache_adr); 1858 // Address is aligned to execute prefetch to the beginning of cache line size 1859 // (it is important when BIS instruction is used on SPARC as prefetch). 1860 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); 1861 cache_adr = new AndXNode(cache_adr, mask); 1862 transform_later(cache_adr); 1863 cache_adr = new CastX2PNode(cache_adr); 1864 transform_later(cache_adr); 1865 1866 // Prefetch 1867 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); 1868 prefetch->set_req(0, needgc_false); 1869 transform_later(prefetch); 1870 contended_phi_rawmem = prefetch; 1871 Node *prefetch_adr; 1872 distance = step_size; 1873 for ( intx i = 1; i < lines; i++ ) { 1874 prefetch_adr = new AddPNode( cache_adr, cache_adr, 1875 _igvn.MakeConX(distance) ); 1876 transform_later(prefetch_adr); 1877 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); 1878 transform_later(prefetch); 1879 distance += step_size; 1880 contended_phi_rawmem = prefetch; 1881 } 1882 } else if( AllocatePrefetchStyle > 0 ) { 1883 // Insert a prefetch for each allocation only on the fast-path 1884 Node *prefetch_adr; 1885 Node *prefetch; 1886 // Generate several prefetch instructions. 1887 uint step_size = AllocatePrefetchStepSize; 1888 uint distance = AllocatePrefetchDistance; 1889 for ( intx i = 0; i < lines; i++ ) { 1890 prefetch_adr = new AddPNode( old_eden_top, new_eden_top, 1891 _igvn.MakeConX(distance) ); 1892 transform_later(prefetch_adr); 1893 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1894 // Do not let it float too high, since if eden_top == eden_end, 1895 // both might be null. 1896 if( i == 0 ) { // Set control for first prefetch, next follows it 1897 prefetch->init_req(0, needgc_false); 1898 } 1899 transform_later(prefetch); 1900 distance += step_size; 1901 i_o = prefetch; 1902 } 1903 } 1904 return i_o; 1905 } 1906 1907 1908 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 1909 expand_allocate_common(alloc, nullptr, 1910 OptoRuntime::new_instance_Type(), 1911 OptoRuntime::new_instance_Java(), nullptr); 1912 } 1913 1914 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 1915 Node* length = alloc->in(AllocateNode::ALength); 1916 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest); 1917 InitializeNode* init = alloc->initialization(); 1918 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1919 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr(); 1920 address slow_call_address; // Address of slow call 1921 if (init != nullptr && init->is_complete_with_arraycopy() && 1922 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) { 1923 // Don't zero type array during slow allocation in VM since 1924 // it will be initialized later by arraycopy in compiled code. 1925 slow_call_address = OptoRuntime::new_array_nozero_Java(); 1926 } else { 1927 slow_call_address = OptoRuntime::new_array_Java(); 1928 } 1929 expand_allocate_common(alloc, length, 1930 OptoRuntime::new_array_Type(), 1931 slow_call_address, valid_length_test); 1932 } 1933 1934 //-------------------mark_eliminated_box---------------------------------- 1935 // 1936 // During EA obj may point to several objects but after few ideal graph 1937 // transformations (CCP) it may point to only one non escaping object 1938 // (but still using phi), corresponding locks and unlocks will be marked 1939 // for elimination. Later obj could be replaced with a new node (new phi) 1940 // and which does not have escape information. And later after some graph 1941 // reshape other locks and unlocks (which were not marked for elimination 1942 // before) are connected to this new obj (phi) but they still will not be 1943 // marked for elimination since new obj has no escape information. 1944 // Mark all associated (same box and obj) lock and unlock nodes for 1945 // elimination if some of them marked already. 1946 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) { 1947 if (oldbox->as_BoxLock()->is_eliminated()) { 1948 return; // This BoxLock node was processed already. 1949 } 1950 // New implementation (EliminateNestedLocks) has separate BoxLock 1951 // node for each locked region so mark all associated locks/unlocks as 1952 // eliminated even if different objects are referenced in one locked region 1953 // (for example, OSR compilation of nested loop inside locked scope). 1954 if (EliminateNestedLocks || 1955 oldbox->as_BoxLock()->is_simple_lock_region(nullptr, obj, nullptr)) { 1956 // Box is used only in one lock region. Mark this box as eliminated. 1957 int locks = 0; 1958 for (uint i = 0; i < oldbox->outcnt(); i++) { 1959 Node* u = oldbox->raw_out(i); 1960 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) { 1961 AbstractLockNode* alock = u->as_AbstractLock(); 1962 // Check lock's box since box could be referenced by Lock's debug info. 1963 if (alock->box_node() == oldbox) { 1964 locks++; 1965 1966 if (alock->obj_node() == obj) { 1967 // Mark eliminated all related locks and unlocks. 1968 #ifdef ASSERT 1969 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4"); 1970 #endif 1971 alock->set_non_esc_obj(); 1972 locks--; 1973 } 1974 } 1975 } 1976 } 1977 if (locks == 0) { 1978 _igvn.hash_delete(oldbox); 1979 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value 1980 _igvn.hash_insert(oldbox); 1981 } 1982 1983 return; 1984 } 1985 1986 // Create new "eliminated" BoxLock node and use it in monitor debug info 1987 // instead of oldbox for the same object. 1988 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 1989 1990 // Note: BoxLock node is marked eliminated only here and it is used 1991 // to indicate that all associated lock and unlock nodes are marked 1992 // for elimination. 1993 newbox->set_eliminated(); 1994 transform_later(newbox); 1995 1996 // Replace old box node with new box for all users of the same object. 1997 for (uint i = 0; i < oldbox->outcnt();) { 1998 bool next_edge = true; 1999 2000 Node* u = oldbox->raw_out(i); 2001 if (u->is_AbstractLock()) { 2002 AbstractLockNode* alock = u->as_AbstractLock(); 2003 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) { 2004 // Replace Box and mark eliminated all related locks and unlocks. 2005 #ifdef ASSERT 2006 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5"); 2007 #endif 2008 alock->set_non_esc_obj(); 2009 _igvn.rehash_node_delayed(alock); 2010 alock->set_box_node(newbox); 2011 next_edge = false; 2012 } 2013 } 2014 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) { 2015 FastLockNode* flock = u->as_FastLock(); 2016 assert(flock->box_node() == oldbox, "sanity"); 2017 _igvn.rehash_node_delayed(flock); 2018 flock->set_box_node(newbox); 2019 next_edge = false; 2020 } 2021 2022 // Replace old box in monitor debug info. 2023 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 2024 SafePointNode* sfn = u->as_SafePoint(); 2025 JVMState* youngest_jvms = sfn->jvms(); 2026 int max_depth = youngest_jvms->depth(); 2027 for (int depth = 1; depth <= max_depth; depth++) { 2028 JVMState* jvms = youngest_jvms->of_depth(depth); 2029 int num_mon = jvms->nof_monitors(); 2030 // Loop over monitors 2031 for (int idx = 0; idx < num_mon; idx++) { 2032 Node* obj_node = sfn->monitor_obj(jvms, idx); 2033 Node* box_node = sfn->monitor_box(jvms, idx); 2034 if (box_node == oldbox && obj_node->eqv_uncast(obj)) { 2035 int j = jvms->monitor_box_offset(idx); 2036 _igvn.replace_input_of(u, j, newbox); 2037 next_edge = false; 2038 } 2039 } 2040 } 2041 } 2042 if (next_edge) i++; 2043 } 2044 } 2045 2046 //-----------------------mark_eliminated_locking_nodes----------------------- 2047 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { 2048 if (EliminateNestedLocks) { 2049 if (alock->is_nested()) { 2050 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity"); 2051 return; 2052 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened 2053 // Only Lock node has JVMState needed here. 2054 // Not that preceding claim is documented anywhere else. 2055 if (alock->jvms() != nullptr) { 2056 if (alock->as_Lock()->is_nested_lock_region()) { 2057 // Mark eliminated related nested locks and unlocks. 2058 Node* obj = alock->obj_node(); 2059 BoxLockNode* box_node = alock->box_node()->as_BoxLock(); 2060 assert(!box_node->is_eliminated(), "should not be marked yet"); 2061 // Note: BoxLock node is marked eliminated only here 2062 // and it is used to indicate that all associated lock 2063 // and unlock nodes are marked for elimination. 2064 box_node->set_eliminated(); // Box's hash is always NO_HASH here 2065 for (uint i = 0; i < box_node->outcnt(); i++) { 2066 Node* u = box_node->raw_out(i); 2067 if (u->is_AbstractLock()) { 2068 alock = u->as_AbstractLock(); 2069 if (alock->box_node() == box_node) { 2070 // Verify that this Box is referenced only by related locks. 2071 assert(alock->obj_node()->eqv_uncast(obj), ""); 2072 // Mark all related locks and unlocks. 2073 #ifdef ASSERT 2074 alock->log_lock_optimization(C, "eliminate_lock_set_nested"); 2075 #endif 2076 alock->set_nested(); 2077 } 2078 } 2079 } 2080 } else { 2081 #ifdef ASSERT 2082 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region"); 2083 if (C->log() != nullptr) 2084 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output 2085 #endif 2086 } 2087 } 2088 return; 2089 } 2090 // Process locks for non escaping object 2091 assert(alock->is_non_esc_obj(), ""); 2092 } // EliminateNestedLocks 2093 2094 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object 2095 // Look for all locks of this object and mark them and 2096 // corresponding BoxLock nodes as eliminated. 2097 Node* obj = alock->obj_node(); 2098 for (uint j = 0; j < obj->outcnt(); j++) { 2099 Node* o = obj->raw_out(j); 2100 if (o->is_AbstractLock() && 2101 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) { 2102 alock = o->as_AbstractLock(); 2103 Node* box = alock->box_node(); 2104 // Replace old box node with new eliminated box for all users 2105 // of the same object and mark related locks as eliminated. 2106 mark_eliminated_box(box, obj); 2107 } 2108 } 2109 } 2110 } 2111 2112 // we have determined that this lock/unlock can be eliminated, we simply 2113 // eliminate the node without expanding it. 2114 // 2115 // Note: The membar's associated with the lock/unlock are currently not 2116 // eliminated. This should be investigated as a future enhancement. 2117 // 2118 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 2119 2120 if (!alock->is_eliminated()) { 2121 return false; 2122 } 2123 #ifdef ASSERT 2124 if (!alock->is_coarsened()) { 2125 // Check that new "eliminated" BoxLock node is created. 2126 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 2127 assert(oldbox->is_eliminated() || DoPartialEscapeAnalysis, "should be done already"); 2128 } 2129 #endif 2130 2131 alock->log_lock_optimization(C, "eliminate_lock"); 2132 2133 #ifndef PRODUCT 2134 if (PrintEliminateLocks) { 2135 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string()); 2136 } 2137 #endif 2138 2139 Node* mem = alock->in(TypeFunc::Memory); 2140 Node* ctrl = alock->in(TypeFunc::Control); 2141 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null"); 2142 2143 alock->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/); 2144 // There are 2 projections from the lock. The lock node will 2145 // be deleted when its last use is subsumed below. 2146 assert(alock->outcnt() == 2 && 2147 _callprojs.fallthrough_proj != nullptr && 2148 _callprojs.fallthrough_memproj != nullptr, 2149 "Unexpected projections from Lock/Unlock"); 2150 2151 Node* fallthroughproj = _callprojs.fallthrough_proj; 2152 Node* memproj_fallthrough = _callprojs.fallthrough_memproj; 2153 2154 // The memory projection from a lock/unlock is RawMem 2155 // The input to a Lock is merged memory, so extract its RawMem input 2156 // (unless the MergeMem has been optimized away.) 2157 if (alock->is_Lock()) { 2158 // Search for MemBarAcquireLock node and delete it also. 2159 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 2160 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, ""); 2161 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 2162 Node* memproj = membar->proj_out(TypeFunc::Memory); 2163 _igvn.replace_node(ctrlproj, fallthroughproj); 2164 _igvn.replace_node(memproj, memproj_fallthrough); 2165 2166 // Delete FastLock node also if this Lock node is unique user 2167 // (a loop peeling may clone a Lock node). 2168 Node* flock = alock->as_Lock()->fastlock_node(); 2169 if (flock->outcnt() == 1) { 2170 assert(flock->unique_out() == alock, "sanity"); 2171 _igvn.replace_node(flock, top()); 2172 } 2173 } 2174 2175 // Search for MemBarReleaseLock node and delete it also. 2176 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) { 2177 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 2178 assert(membar->Opcode() == Op_MemBarReleaseLock && 2179 mem->is_Proj() && membar == mem->in(0), ""); 2180 _igvn.replace_node(fallthroughproj, ctrl); 2181 _igvn.replace_node(memproj_fallthrough, mem); 2182 fallthroughproj = ctrl; 2183 memproj_fallthrough = mem; 2184 ctrl = membar->in(TypeFunc::Control); 2185 mem = membar->in(TypeFunc::Memory); 2186 } 2187 2188 _igvn.replace_node(fallthroughproj, ctrl); 2189 _igvn.replace_node(memproj_fallthrough, mem); 2190 return true; 2191 } 2192 2193 2194 //------------------------------expand_lock_node---------------------- 2195 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 2196 2197 Node* ctrl = lock->in(TypeFunc::Control); 2198 Node* mem = lock->in(TypeFunc::Memory); 2199 Node* obj = lock->obj_node(); 2200 Node* box = lock->box_node(); 2201 Node* flock = lock->fastlock_node(); 2202 2203 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2204 2205 // Make the merge point 2206 Node *region; 2207 Node *mem_phi; 2208 Node *slow_path; 2209 2210 region = new RegionNode(3); 2211 // create a Phi for the memory state 2212 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2213 2214 // Optimize test; set region slot 2 2215 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); 2216 mem_phi->init_req(2, mem); 2217 2218 // Make slow path call 2219 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), 2220 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path, 2221 obj, box, nullptr); 2222 2223 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/); 2224 2225 // Slow path can only throw asynchronous exceptions, which are always 2226 // de-opted. So the compiler thinks the slow-call can never throw an 2227 // exception. If it DOES throw an exception we would need the debug 2228 // info removed first (since if it throws there is no monitor). 2229 assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr && 2230 _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock"); 2231 2232 // Capture slow path 2233 // disconnect fall-through projection from call and create a new one 2234 // hook up users of fall-through projection to region 2235 Node *slow_ctrl = _callprojs.fallthrough_proj->clone(); 2236 transform_later(slow_ctrl); 2237 _igvn.hash_delete(_callprojs.fallthrough_proj); 2238 _callprojs.fallthrough_proj->disconnect_inputs(C); 2239 region->init_req(1, slow_ctrl); 2240 // region inputs are now complete 2241 transform_later(region); 2242 _igvn.replace_node(_callprojs.fallthrough_proj, region); 2243 2244 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory)); 2245 2246 mem_phi->init_req(1, memproj); 2247 2248 transform_later(mem_phi); 2249 2250 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi); 2251 } 2252 2253 //------------------------------expand_unlock_node---------------------- 2254 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 2255 2256 Node* ctrl = unlock->in(TypeFunc::Control); 2257 Node* mem = unlock->in(TypeFunc::Memory); 2258 Node* obj = unlock->obj_node(); 2259 Node* box = unlock->box_node(); 2260 2261 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2262 2263 // No need for a null check on unlock 2264 2265 // Make the merge point 2266 Node *region; 2267 Node *mem_phi; 2268 2269 region = new RegionNode(3); 2270 // create a Phi for the memory state 2271 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2272 2273 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box ); 2274 funlock = transform_later( funlock )->as_FastUnlock(); 2275 // Optimize test; set region slot 2 2276 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); 2277 Node *thread = transform_later(new ThreadLocalNode()); 2278 2279 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), 2280 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), 2281 "complete_monitor_unlocking_C", slow_path, obj, box, thread); 2282 2283 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/); 2284 assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr && 2285 _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock"); 2286 2287 // No exceptions for unlocking 2288 // Capture slow path 2289 // disconnect fall-through projection from call and create a new one 2290 // hook up users of fall-through projection to region 2291 Node *slow_ctrl = _callprojs.fallthrough_proj->clone(); 2292 transform_later(slow_ctrl); 2293 _igvn.hash_delete(_callprojs.fallthrough_proj); 2294 _callprojs.fallthrough_proj->disconnect_inputs(C); 2295 region->init_req(1, slow_ctrl); 2296 // region inputs are now complete 2297 transform_later(region); 2298 _igvn.replace_node(_callprojs.fallthrough_proj, region); 2299 2300 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) ); 2301 mem_phi->init_req(1, memproj ); 2302 mem_phi->init_req(2, mem); 2303 transform_later(mem_phi); 2304 2305 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi); 2306 } 2307 2308 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) { 2309 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned"); 2310 Node* bol = check->unique_out(); 2311 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass); 2312 Node* superklass = check->in(SubTypeCheckNode::SuperKlass); 2313 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node"); 2314 2315 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) { 2316 Node* iff = bol->last_out(i); 2317 assert(iff->is_If(), "where's the if?"); 2318 2319 if (iff->in(0)->is_top()) { 2320 _igvn.replace_input_of(iff, 1, C->top()); 2321 continue; 2322 } 2323 2324 Node* iftrue = iff->as_If()->proj_out(1); 2325 Node* iffalse = iff->as_If()->proj_out(0); 2326 Node* ctrl = iff->in(0); 2327 2328 Node* subklass = nullptr; 2329 if (_igvn.type(obj_or_subklass)->isa_klassptr()) { 2330 subklass = obj_or_subklass; 2331 } else { 2332 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes()); 2333 subklass = _igvn.transform(LoadKlassNode::make(_igvn, nullptr, C->immutable_memory(), k_adr, TypeInstPtr::KLASS)); 2334 } 2335 2336 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci()); 2337 2338 _igvn.replace_input_of(iff, 0, C->top()); 2339 _igvn.replace_node(iftrue, not_subtype_ctrl); 2340 _igvn.replace_node(iffalse, ctrl); 2341 } 2342 _igvn.replace_node(check, C->top()); 2343 } 2344 2345 //---------------------------eliminate_macro_nodes---------------------- 2346 // Eliminate scalar replaced allocations and associated locks. 2347 void PhaseMacroExpand::eliminate_macro_nodes() { 2348 if (C->macro_count() == 0) 2349 return; 2350 NOT_PRODUCT(int membar_before = count_MemBar(C);) 2351 2352 // Before elimination may re-mark (change to Nested or NonEscObj) 2353 // all associated (same box and obj) lock and unlock nodes. 2354 int cnt = C->macro_count(); 2355 for (int i=0; i < cnt; i++) { 2356 Node *n = C->macro_node(i); 2357 if (n->is_AbstractLock()) { // Lock and Unlock nodes 2358 mark_eliminated_locking_nodes(n->as_AbstractLock()); 2359 } 2360 } 2361 // Re-marking may break consistency of Coarsened locks. 2362 if (!C->coarsened_locks_consistent()) { 2363 return; // recompile without Coarsened locks if broken 2364 } 2365 2366 // First, attempt to eliminate locks 2367 bool progress = true; 2368 while (progress) { 2369 progress = false; 2370 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once 2371 Node* n = C->macro_node(i - 1); 2372 bool success = false; 2373 DEBUG_ONLY(int old_macro_count = C->macro_count();) 2374 if (n->is_AbstractLock()) { 2375 success = eliminate_locking_node(n->as_AbstractLock()); 2376 #ifndef PRODUCT 2377 if (success && PrintOptoStatistics) { 2378 Atomic::inc(&PhaseMacroExpand::_monitor_objects_removed_counter); 2379 } 2380 #endif 2381 } 2382 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2383 progress = progress || success; 2384 } 2385 } 2386 // Next, attempt to eliminate allocations 2387 _has_locks = false; 2388 progress = true; 2389 while (progress) { 2390 progress = false; 2391 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once 2392 Node* n = C->macro_node(i - 1); 2393 bool success = false; 2394 DEBUG_ONLY(int old_macro_count = C->macro_count();) 2395 switch (n->class_id()) { 2396 case Node::Class_Allocate: 2397 case Node::Class_AllocateArray: 2398 success = eliminate_allocate_node(n->as_Allocate()); 2399 #ifndef PRODUCT 2400 if (success && PrintOptoStatistics) { 2401 Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter); 2402 } 2403 #endif 2404 break; 2405 case Node::Class_CallStaticJava: 2406 success = eliminate_boxing_node(n->as_CallStaticJava()); 2407 break; 2408 case Node::Class_Lock: 2409 case Node::Class_Unlock: 2410 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 2411 _has_locks = true; 2412 break; 2413 case Node::Class_ArrayCopy: 2414 break; 2415 case Node::Class_OuterStripMinedLoop: 2416 break; 2417 case Node::Class_SubTypeCheck: 2418 break; 2419 case Node::Class_Opaque1: 2420 break; 2421 default: 2422 assert(n->Opcode() == Op_LoopLimit || 2423 n->Opcode() == Op_Opaque3 || 2424 n->Opcode() == Op_Opaque4 || 2425 n->Opcode() == Op_MaxL || 2426 n->Opcode() == Op_MinL || 2427 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n), 2428 "unknown node type in macro list"); 2429 } 2430 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2431 progress = progress || success; 2432 } 2433 } 2434 #ifndef PRODUCT 2435 if (PrintOptoStatistics) { 2436 int membar_after = count_MemBar(C); 2437 Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after); 2438 } 2439 #endif 2440 } 2441 2442 //------------------------------expand_macro_nodes---------------------- 2443 // Returns true if a failure occurred. 2444 bool PhaseMacroExpand::expand_macro_nodes() { 2445 // Last attempt to eliminate macro nodes. 2446 eliminate_macro_nodes(); 2447 if (C->failing()) return true; 2448 2449 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. 2450 bool progress = true; 2451 while (progress) { 2452 progress = false; 2453 for (int i = C->macro_count(); i > 0; i--) { 2454 Node* n = C->macro_node(i-1); 2455 bool success = false; 2456 DEBUG_ONLY(int old_macro_count = C->macro_count();) 2457 if (n->Opcode() == Op_LoopLimit) { 2458 // Remove it from macro list and put on IGVN worklist to optimize. 2459 C->remove_macro_node(n); 2460 _igvn._worklist.push(n); 2461 success = true; 2462 } else if (n->Opcode() == Op_CallStaticJava) { 2463 // Remove it from macro list and put on IGVN worklist to optimize. 2464 C->remove_macro_node(n); 2465 _igvn._worklist.push(n); 2466 success = true; 2467 } else if (n->is_Opaque1()) { 2468 _igvn.replace_node(n, n->in(1)); 2469 success = true; 2470 #if INCLUDE_RTM_OPT 2471 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) { 2472 assert(C->profile_rtm(), "should be used only in rtm deoptimization code"); 2473 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), ""); 2474 Node* cmp = n->unique_out(); 2475 #ifdef ASSERT 2476 // Validate graph. 2477 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), ""); 2478 BoolNode* bol = cmp->unique_out()->as_Bool(); 2479 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() && 2480 (bol->_test._test == BoolTest::ne), ""); 2481 IfNode* ifn = bol->unique_out()->as_If(); 2482 assert((ifn->outcnt() == 2) && 2483 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != nullptr, ""); 2484 #endif 2485 Node* repl = n->in(1); 2486 if (!_has_locks) { 2487 // Remove RTM state check if there are no locks in the code. 2488 // Replace input to compare the same value. 2489 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1); 2490 } 2491 _igvn.replace_node(n, repl); 2492 success = true; 2493 #endif 2494 } else if (n->Opcode() == Op_Opaque4) { 2495 // With Opaque4 nodes, the expectation is that the test of input 1 2496 // is always equal to the constant value of input 2. So we can 2497 // remove the Opaque4 and replace it by input 2. In debug builds, 2498 // leave the non constant test in instead to sanity check that it 2499 // never fails (if it does, that subgraph was constructed so, at 2500 // runtime, a Halt node is executed). 2501 #ifdef ASSERT 2502 _igvn.replace_node(n, n->in(1)); 2503 #else 2504 _igvn.replace_node(n, n->in(2)); 2505 #endif 2506 success = true; 2507 } else if (n->Opcode() == Op_OuterStripMinedLoop) { 2508 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn); 2509 C->remove_macro_node(n); 2510 success = true; 2511 } else if (n->Opcode() == Op_MaxL) { 2512 // Since MaxL and MinL are not implemented in the backend, we expand them to 2513 // a CMoveL construct now. At least until here, the type could be computed 2514 // precisely. CMoveL is not so smart, but we can give it at least the best 2515 // type we know abouot n now. 2516 Node* repl = MaxNode::signed_max(n->in(1), n->in(2), _igvn.type(n), _igvn); 2517 _igvn.replace_node(n, repl); 2518 success = true; 2519 } else if (n->Opcode() == Op_MinL) { 2520 Node* repl = MaxNode::signed_min(n->in(1), n->in(2), _igvn.type(n), _igvn); 2521 _igvn.replace_node(n, repl); 2522 success = true; 2523 } 2524 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list"); 2525 progress = progress || success; 2526 } 2527 } 2528 2529 // Clean up the graph so we're less likely to hit the maximum node 2530 // limit 2531 _igvn.set_delay_transform(false); 2532 _igvn.optimize(); 2533 if (C->failing()) return true; 2534 _igvn.set_delay_transform(true); 2535 2536 2537 // Because we run IGVN after each expansion, some macro nodes may go 2538 // dead and be removed from the list as we iterate over it. Move 2539 // Allocate nodes (processed in a second pass) at the beginning of 2540 // the list and then iterate from the last element of the list until 2541 // an Allocate node is seen. This is robust to random deletion in 2542 // the list due to nodes going dead. 2543 C->sort_macro_nodes(); 2544 2545 // expand arraycopy "macro" nodes first 2546 // For ReduceBulkZeroing, we must first process all arraycopy nodes 2547 // before the allocate nodes are expanded. 2548 while (C->macro_count() > 0) { 2549 int macro_count = C->macro_count(); 2550 Node * n = C->macro_node(macro_count-1); 2551 assert(n->is_macro(), "only macro nodes expected here"); 2552 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) { 2553 // node is unreachable, so don't try to expand it 2554 C->remove_macro_node(n); 2555 continue; 2556 } 2557 if (n->is_Allocate()) { 2558 break; 2559 } 2560 // Make sure expansion will not cause node limit to be exceeded. 2561 // Worst case is a macro node gets expanded into about 200 nodes. 2562 // Allow 50% more for optimization. 2563 if (C->check_node_count(300, "out of nodes before macro expansion")) { 2564 return true; 2565 } 2566 2567 DEBUG_ONLY(int old_macro_count = C->macro_count();) 2568 switch (n->class_id()) { 2569 case Node::Class_Lock: 2570 expand_lock_node(n->as_Lock()); 2571 break; 2572 case Node::Class_Unlock: 2573 expand_unlock_node(n->as_Unlock()); 2574 break; 2575 case Node::Class_ArrayCopy: 2576 expand_arraycopy_node(n->as_ArrayCopy()); 2577 break; 2578 case Node::Class_SubTypeCheck: 2579 expand_subtypecheck_node(n->as_SubTypeCheck()); 2580 break; 2581 default: 2582 assert(false, "unknown node type in macro list"); 2583 } 2584 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list"); 2585 if (C->failing()) return true; 2586 2587 // Clean up the graph so we're less likely to hit the maximum node 2588 // limit 2589 _igvn.set_delay_transform(false); 2590 _igvn.optimize(); 2591 if (C->failing()) return true; 2592 _igvn.set_delay_transform(true); 2593 } 2594 2595 // All nodes except Allocate nodes are expanded now. There could be 2596 // new optimization opportunities (such as folding newly created 2597 // load from a just allocated object). Run IGVN. 2598 2599 // expand "macro" nodes 2600 // nodes are removed from the macro list as they are processed 2601 while (C->macro_count() > 0) { 2602 int macro_count = C->macro_count(); 2603 Node * n = C->macro_node(macro_count-1); 2604 assert(n->is_macro(), "only macro nodes expected here"); 2605 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) { 2606 // node is unreachable, so don't try to expand it 2607 C->remove_macro_node(n); 2608 continue; 2609 } 2610 // Make sure expansion will not cause node limit to be exceeded. 2611 // Worst case is a macro node gets expanded into about 200 nodes. 2612 // Allow 50% more for optimization. 2613 if (C->check_node_count(300, "out of nodes before macro expansion")) { 2614 return true; 2615 } 2616 switch (n->class_id()) { 2617 case Node::Class_Allocate: 2618 expand_allocate(n->as_Allocate()); 2619 break; 2620 case Node::Class_AllocateArray: 2621 expand_allocate_array(n->as_AllocateArray()); 2622 break; 2623 default: 2624 assert(false, "unknown node type in macro list"); 2625 } 2626 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2627 if (C->failing()) return true; 2628 2629 // Clean up the graph so we're less likely to hit the maximum node 2630 // limit 2631 _igvn.set_delay_transform(false); 2632 _igvn.optimize(); 2633 if (C->failing()) return true; 2634 _igvn.set_delay_transform(true); 2635 } 2636 2637 _igvn.set_delay_transform(false); 2638 return false; 2639 } 2640 2641 #ifndef PRODUCT 2642 int PhaseMacroExpand::_objs_scalar_replaced_counter = 0; 2643 int PhaseMacroExpand::_monitor_objects_removed_counter = 0; 2644 int PhaseMacroExpand::_GC_barriers_removed_counter = 0; 2645 int PhaseMacroExpand::_memory_barriers_removed_counter = 0; 2646 2647 void PhaseMacroExpand::print_statistics() { 2648 tty->print("Objects scalar replaced = %d, ", Atomic::load(&_objs_scalar_replaced_counter)); 2649 tty->print("Monitor objects removed = %d, ", Atomic::load(&_monitor_objects_removed_counter)); 2650 tty->print("GC barriers removed = %d, ", Atomic::load(&_GC_barriers_removed_counter)); 2651 tty->print_cr("Memory barriers removed = %d", Atomic::load(&_memory_barriers_removed_counter)); 2652 } 2653 2654 int PhaseMacroExpand::count_MemBar(Compile *C) { 2655 if (!PrintOptoStatistics) { 2656 return 0; 2657 } 2658 Unique_Node_List ideal_nodes; 2659 int total = 0; 2660 ideal_nodes.map(C->live_nodes(), nullptr); 2661 ideal_nodes.push(C->root()); 2662 for (uint next = 0; next < ideal_nodes.size(); ++next) { 2663 Node* n = ideal_nodes.at(next); 2664 if (n->is_MemBar()) { 2665 total++; 2666 } 2667 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2668 Node* m = n->fast_out(i); 2669 ideal_nodes.push(m); 2670 } 2671 } 2672 return total; 2673 } 2674 #endif