1 /* 2 * Copyright (c) 2005, 2025, 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 "compiler/compileLog.hpp" 26 #include "gc/shared/collectedHeap.inline.hpp" 27 #include "gc/shared/tlab_globals.hpp" 28 #include "libadt/vectset.hpp" 29 #include "memory/universe.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/arraycopynode.hpp" 32 #include "opto/callnode.hpp" 33 #include "opto/castnode.hpp" 34 #include "opto/cfgnode.hpp" 35 #include "opto/compile.hpp" 36 #include "opto/convertnode.hpp" 37 #include "opto/graphKit.hpp" 38 #include "opto/intrinsicnode.hpp" 39 #include "opto/locknode.hpp" 40 #include "opto/loopnode.hpp" 41 #include "opto/macro.hpp" 42 #include "opto/memnode.hpp" 43 #include "opto/narrowptrnode.hpp" 44 #include "opto/node.hpp" 45 #include "opto/opaquenode.hpp" 46 #include "opto/phaseX.hpp" 47 #include "opto/rootnode.hpp" 48 #include "opto/runtime.hpp" 49 #include "opto/subnode.hpp" 50 #include "opto/subtypenode.hpp" 51 #include "opto/type.hpp" 52 #include "prims/jvmtiExport.hpp" 53 #include "runtime/continuation.hpp" 54 #include "runtime/sharedRuntime.hpp" 55 #include "utilities/macros.hpp" 56 #include "utilities/powerOfTwo.hpp" 57 #if INCLUDE_G1GC 58 #include "gc/g1/g1ThreadLocalData.hpp" 59 #endif // INCLUDE_G1GC 60 61 62 // 63 // Replace any references to "oldref" in inputs to "use" with "newref". 64 // Returns the number of replacements made. 65 // 66 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { 67 int nreplacements = 0; 68 uint req = use->req(); 69 for (uint j = 0; j < use->len(); j++) { 70 Node *uin = use->in(j); 71 if (uin == oldref) { 72 if (j < req) 73 use->set_req(j, newref); 74 else 75 use->set_prec(j, newref); 76 nreplacements++; 77 } else if (j >= req && uin == nullptr) { 78 break; 79 } 80 } 81 return nreplacements; 82 } 83 84 void PhaseMacroExpand::migrate_outs(Node *old, Node *target) { 85 assert(old != nullptr, "sanity"); 86 for (DUIterator_Fast imax, i = old->fast_outs(imax); i < imax; i++) { 87 Node* use = old->fast_out(i); 88 _igvn.rehash_node_delayed(use); 89 imax -= replace_input(use, old, target); 90 // back up iterator 91 --i; 92 } 93 assert(old->outcnt() == 0, "all uses must be deleted"); 94 } 95 96 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word) { 97 Node* cmp = word; 98 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne)); 99 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); 100 transform_later(iff); 101 102 // Fast path taken. 103 Node *fast_taken = transform_later(new IfFalseNode(iff)); 104 105 // Fast path not-taken, i.e. slow path 106 Node *slow_taken = transform_later(new IfTrueNode(iff)); 107 108 region->init_req(edge, fast_taken); // Capture fast-control 109 return slow_taken; 110 } 111 112 //--------------------copy_predefined_input_for_runtime_call-------------------- 113 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { 114 // Set fixed predefined input arguments 115 call->init_req( TypeFunc::Control, ctrl ); 116 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); 117 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? 118 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); 119 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); 120 } 121 122 //------------------------------make_slow_call--------------------------------- 123 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, 124 address slow_call, const char* leaf_name, Node* slow_path, 125 Node* parm0, Node* parm1, Node* parm2) { 126 127 // Slow-path call 128 CallNode *call = leaf_name 129 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) 130 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM ); 131 132 // Slow path call has no side-effects, uses few values 133 copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); 134 if (parm0 != nullptr) call->init_req(TypeFunc::Parms+0, parm0); 135 if (parm1 != nullptr) call->init_req(TypeFunc::Parms+1, parm1); 136 if (parm2 != nullptr) call->init_req(TypeFunc::Parms+2, parm2); 137 call->copy_call_debug_info(&_igvn, oldcall); 138 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 139 _igvn.replace_node(oldcall, call); 140 transform_later(call); 141 142 return call; 143 } 144 145 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) { 146 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2(); 147 bs->eliminate_gc_barrier(this, p2x); 148 #ifndef PRODUCT 149 if (PrintOptoStatistics) { 150 Atomic::inc(&PhaseMacroExpand::_GC_barriers_removed_counter); 151 } 152 #endif 153 } 154 155 // Search for a memory operation for the specified memory slice. 156 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) { 157 Node *orig_mem = mem; 158 Node *alloc_mem = alloc->as_Allocate()->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false); 159 assert(alloc_mem != nullptr, "Allocation without a memory projection."); 160 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr(); 161 while (true) { 162 if (mem == alloc_mem || mem == start_mem ) { 163 return mem; // hit one of our sentinels 164 } else if (mem->is_MergeMem()) { 165 mem = mem->as_MergeMem()->memory_at(alias_idx); 166 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { 167 Node *in = mem->in(0); 168 // we can safely skip over safepoints, calls, locks and membars because we 169 // already know that the object is safe to eliminate. 170 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { 171 return in; 172 } else if (in->is_Call()) { 173 CallNode *call = in->as_Call(); 174 if (call->may_modify(tinst, phase)) { 175 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape"); 176 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) { 177 return in; 178 } 179 } 180 mem = in->in(TypeFunc::Memory); 181 } else if (in->is_MemBar()) { 182 ArrayCopyNode* ac = nullptr; 183 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) { 184 if (ac != nullptr) { 185 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone"); 186 return ac; 187 } 188 } 189 mem = in->in(TypeFunc::Memory); 190 } else { 191 #ifdef ASSERT 192 in->dump(); 193 mem->dump(); 194 assert(false, "unexpected projection"); 195 #endif 196 } 197 } else if (mem->is_Store()) { 198 const TypePtr* atype = mem->as_Store()->adr_type(); 199 int adr_idx = phase->C->get_alias_index(atype); 200 if (adr_idx == alias_idx) { 201 assert(atype->isa_oopptr(), "address type must be oopptr"); 202 int adr_offset = atype->offset(); 203 uint adr_iid = atype->is_oopptr()->instance_id(); 204 // Array elements references have the same alias_idx 205 // but different offset and different instance_id. 206 if (adr_offset == offset && adr_iid == alloc->_idx) { 207 return mem; 208 } 209 } else { 210 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw"); 211 } 212 mem = mem->in(MemNode::Memory); 213 } else if (mem->is_ClearArray()) { 214 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) { 215 // Can not bypass initialization of the instance 216 // we are looking. 217 DEBUG_ONLY(intptr_t offset;) 218 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity"); 219 InitializeNode* init = alloc->as_Allocate()->initialization(); 220 // We are looking for stored value, return Initialize node 221 // or memory edge from Allocate node. 222 if (init != nullptr) { 223 return init; 224 } else { 225 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers). 226 } 227 } 228 // Otherwise skip it (the call updated 'mem' value). 229 } else if (mem->Opcode() == Op_SCMemProj) { 230 mem = mem->in(0); 231 Node* adr = nullptr; 232 if (mem->is_LoadStore()) { 233 adr = mem->in(MemNode::Address); 234 } else { 235 assert(mem->Opcode() == Op_EncodeISOArray || 236 mem->Opcode() == Op_StrCompressedCopy, "sanity"); 237 adr = mem->in(3); // Destination array 238 } 239 const TypePtr* atype = adr->bottom_type()->is_ptr(); 240 int adr_idx = phase->C->get_alias_index(atype); 241 if (adr_idx == alias_idx) { 242 DEBUG_ONLY(mem->dump();) 243 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 244 return nullptr; 245 } 246 mem = mem->in(MemNode::Memory); 247 } else if (mem->Opcode() == Op_StrInflatedCopy) { 248 Node* adr = mem->in(3); // Destination array 249 const TypePtr* atype = adr->bottom_type()->is_ptr(); 250 int adr_idx = phase->C->get_alias_index(atype); 251 if (adr_idx == alias_idx) { 252 DEBUG_ONLY(mem->dump();) 253 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field"); 254 return nullptr; 255 } 256 mem = mem->in(MemNode::Memory); 257 } else { 258 return mem; 259 } 260 assert(mem != orig_mem, "dead memory loop"); 261 } 262 } 263 264 // Generate loads from source of the arraycopy for fields of 265 // destination needed at a deoptimization point 266 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) { 267 BasicType bt = ft; 268 const Type *type = ftype; 269 if (ft == T_NARROWOOP) { 270 bt = T_OBJECT; 271 type = ftype->make_oopptr(); 272 } 273 Node* res = nullptr; 274 if (ac->is_clonebasic()) { 275 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination"); 276 Node* base = ac->in(ArrayCopyNode::Src); 277 Node* adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(offset))); 278 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset); 279 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem(); 280 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 281 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt); 282 } else { 283 if (ac->modifies(offset, offset, &_igvn, true)) { 284 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result"); 285 uint shift = exact_log2(type2aelembytes(bt)); 286 Node* src_pos = ac->in(ArrayCopyNode::SrcPos); 287 Node* dest_pos = ac->in(ArrayCopyNode::DestPos); 288 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int(); 289 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int(); 290 291 Node* adr = nullptr; 292 const TypePtr* adr_type = nullptr; 293 if (src_pos_t->is_con() && dest_pos_t->is_con()) { 294 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset; 295 Node* base = ac->in(ArrayCopyNode::Src); 296 adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(off))); 297 adr_type = _igvn.type(base)->is_ptr()->add_offset(off); 298 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) { 299 // Don't emit a new load from src if src == dst but try to get the value from memory instead 300 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc); 301 } 302 } else { 303 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 304 #ifdef _LP64 305 diff = _igvn.transform(new ConvI2LNode(diff)); 306 #endif 307 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift))); 308 309 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff)); 310 Node* base = ac->in(ArrayCopyNode::Src); 311 adr = _igvn.transform(new AddPNode(base, base, off)); 312 adr_type = _igvn.type(base)->is_ptr()->add_offset(Type::OffsetBot); 313 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) { 314 // Non constant offset in the array: we can't statically 315 // determine the value 316 return nullptr; 317 } 318 } 319 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem(); 320 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 321 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt); 322 } 323 } 324 if (res != nullptr) { 325 if (ftype->isa_narrowoop()) { 326 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes 327 res = _igvn.transform(new EncodePNode(res, ftype)); 328 } 329 return res; 330 } 331 return nullptr; 332 } 333 334 // 335 // Given a Memory Phi, compute a value Phi containing the values from stores 336 // on the input paths. 337 // Note: this function is recursive, its depth is limited by the "level" argument 338 // Returns the computed Phi, or null if it cannot compute it. 339 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) { 340 assert(mem->is_Phi(), "sanity"); 341 int alias_idx = C->get_alias_index(adr_t); 342 int offset = adr_t->offset(); 343 int instance_id = adr_t->instance_id(); 344 345 // Check if an appropriate value phi already exists. 346 Node* region = mem->in(0); 347 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { 348 Node* phi = region->fast_out(k); 349 if (phi->is_Phi() && phi != mem && 350 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) { 351 return phi; 352 } 353 } 354 // Check if an appropriate new value phi already exists. 355 Node* new_phi = value_phis->find(mem->_idx); 356 if (new_phi != nullptr) 357 return new_phi; 358 359 if (level <= 0) { 360 return nullptr; // Give up: phi tree too deep 361 } 362 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 363 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false); 364 assert(alloc_mem != nullptr, "Allocation without a memory projection."); 365 366 uint length = mem->req(); 367 GrowableArray <Node *> values(length, length, nullptr); 368 369 // create a new Phi for the value 370 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset); 371 transform_later(phi); 372 value_phis->push(phi, mem->_idx); 373 374 for (uint j = 1; j < length; j++) { 375 Node *in = mem->in(j); 376 if (in == nullptr || in->is_top()) { 377 values.at_put(j, in); 378 } else { 379 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); 380 if (val == start_mem || val == alloc_mem) { 381 // hit a sentinel, return appropriate 0 value 382 values.at_put(j, _igvn.zerocon(ft)); 383 continue; 384 } 385 if (val->is_Initialize()) { 386 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 387 } 388 if (val == nullptr) { 389 return nullptr; // can't find a value on this path 390 } 391 if (val == mem) { 392 values.at_put(j, mem); 393 } else if (val->is_Store()) { 394 Node* n = val->in(MemNode::ValueIn); 395 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 396 n = bs->step_over_gc_barrier(n); 397 if (is_subword_type(ft)) { 398 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true); 399 } 400 values.at_put(j, n); 401 } else if(val->is_Proj() && val->in(0) == alloc) { 402 values.at_put(j, _igvn.zerocon(ft)); 403 } else if (val->is_Phi()) { 404 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); 405 if (val == nullptr) { 406 return nullptr; 407 } 408 values.at_put(j, val); 409 } else if (val->Opcode() == Op_SCMemProj) { 410 assert(val->in(0)->is_LoadStore() || 411 val->in(0)->Opcode() == Op_EncodeISOArray || 412 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity"); 413 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 414 return nullptr; 415 } else if (val->is_ArrayCopy()) { 416 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc); 417 if (res == nullptr) { 418 return nullptr; 419 } 420 values.at_put(j, res); 421 } else if (val->is_top()) { 422 // This indicates that this path into the phi is dead. Top will eventually also propagate into the Region. 423 // IGVN will clean this up later. 424 values.at_put(j, val); 425 } else { 426 DEBUG_ONLY( val->dump(); ) 427 assert(false, "unknown node on this path"); 428 return nullptr; // unknown node on this path 429 } 430 } 431 } 432 // Set Phi's inputs 433 for (uint j = 1; j < length; j++) { 434 if (values.at(j) == mem) { 435 phi->init_req(j, phi); 436 } else { 437 phi->init_req(j, values.at(j)); 438 } 439 } 440 return phi; 441 } 442 443 // Search the last value stored into the object's field. 444 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) { 445 assert(adr_t->is_known_instance_field(), "instance required"); 446 int instance_id = adr_t->instance_id(); 447 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 448 449 int alias_idx = C->get_alias_index(adr_t); 450 int offset = adr_t->offset(); 451 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 452 Node *alloc_ctrl = alloc->in(TypeFunc::Control); 453 Node *alloc_mem = alloc->proj_out_or_null(TypeFunc::Memory, /*io_use:*/false); 454 assert(alloc_mem != nullptr, "Allocation without a memory projection."); 455 VectorSet visited; 456 457 bool done = sfpt_mem == alloc_mem; 458 Node *mem = sfpt_mem; 459 while (!done) { 460 if (visited.test_set(mem->_idx)) { 461 return nullptr; // found a loop, give up 462 } 463 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 464 if (mem == start_mem || mem == alloc_mem) { 465 done = true; // hit a sentinel, return appropriate 0 value 466 } else if (mem->is_Initialize()) { 467 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 468 if (mem == nullptr) { 469 done = true; // Something go wrong. 470 } else if (mem->is_Store()) { 471 const TypePtr* atype = mem->as_Store()->adr_type(); 472 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 473 done = true; 474 } 475 } else if (mem->is_Store()) { 476 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 477 assert(atype != nullptr, "address type must be oopptr"); 478 assert(C->get_alias_index(atype) == alias_idx && 479 atype->is_known_instance_field() && atype->offset() == offset && 480 atype->instance_id() == instance_id, "store is correct memory slice"); 481 done = true; 482 } else if (mem->is_Phi()) { 483 // try to find a phi's unique input 484 Node *unique_input = nullptr; 485 Node *top = C->top(); 486 for (uint i = 1; i < mem->req(); i++) { 487 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 488 if (n == nullptr || n == top || n == mem) { 489 continue; 490 } else if (unique_input == nullptr) { 491 unique_input = n; 492 } else if (unique_input != n) { 493 unique_input = top; 494 break; 495 } 496 } 497 if (unique_input != nullptr && unique_input != top) { 498 mem = unique_input; 499 } else { 500 done = true; 501 } 502 } else if (mem->is_ArrayCopy()) { 503 done = true; 504 } else { 505 DEBUG_ONLY( mem->dump(); ) 506 assert(false, "unexpected node"); 507 } 508 } 509 if (mem != nullptr) { 510 if (mem == start_mem || mem == alloc_mem) { 511 // hit a sentinel, return appropriate 0 value 512 return _igvn.zerocon(ft); 513 } else if (mem->is_Store()) { 514 Node* n = mem->in(MemNode::ValueIn); 515 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 516 n = bs->step_over_gc_barrier(n); 517 return n; 518 } else if (mem->is_Phi()) { 519 // attempt to produce a Phi reflecting the values on the input paths of the Phi 520 Node_Stack value_phis(8); 521 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 522 if (phi != nullptr) { 523 return phi; 524 } else { 525 // Kill all new Phis 526 while(value_phis.is_nonempty()) { 527 Node* n = value_phis.node(); 528 _igvn.replace_node(n, C->top()); 529 value_phis.pop(); 530 } 531 } 532 } else if (mem->is_ArrayCopy()) { 533 Node* ctl = mem->in(0); 534 Node* m = mem->in(TypeFunc::Memory); 535 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj()) { 536 // pin the loads in the uncommon trap path 537 ctl = sfpt_ctl; 538 m = sfpt_mem; 539 } 540 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc); 541 } 542 } 543 // Something go wrong. 544 return nullptr; 545 } 546 547 // Check the possibility of scalar replacement. 548 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) { 549 // Scan the uses of the allocation to check for anything that would 550 // prevent us from eliminating it. 551 NOT_PRODUCT( const char* fail_eliminate = nullptr; ) 552 DEBUG_ONLY( Node* disq_node = nullptr; ) 553 bool can_eliminate = true; 554 bool reduce_merge_precheck = (safepoints == nullptr); 555 556 Node* res = alloc->result_cast(); 557 const TypeOopPtr* res_type = nullptr; 558 if (res == nullptr) { 559 // All users were eliminated. 560 } else if (!res->is_CheckCastPP()) { 561 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 562 can_eliminate = false; 563 } else { 564 res_type = igvn->type(res)->isa_oopptr(); 565 if (res_type == nullptr) { 566 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 567 can_eliminate = false; 568 } else if (!res_type->klass_is_exact()) { 569 NOT_PRODUCT(fail_eliminate = "Not an exact type.";) 570 can_eliminate = false; 571 } else if (res_type->isa_aryptr()) { 572 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 573 if (length < 0) { 574 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 575 can_eliminate = false; 576 } 577 } 578 } 579 580 if (can_eliminate && res != nullptr) { 581 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2(); 582 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 583 j < jmax && can_eliminate; j++) { 584 Node* use = res->fast_out(j); 585 586 if (use->is_AddP()) { 587 const TypePtr* addp_type = igvn->type(use)->is_ptr(); 588 int offset = addp_type->offset(); 589 590 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 591 NOT_PRODUCT(fail_eliminate = "Undefined field reference";) 592 can_eliminate = false; 593 break; 594 } 595 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 596 k < kmax && can_eliminate; k++) { 597 Node* n = use->fast_out(k); 598 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) { 599 DEBUG_ONLY(disq_node = n;) 600 if (n->is_Load() || n->is_LoadStore()) { 601 NOT_PRODUCT(fail_eliminate = "Field load";) 602 } else { 603 NOT_PRODUCT(fail_eliminate = "Not store field reference";) 604 } 605 can_eliminate = false; 606 } 607 } 608 } else if (use->is_ArrayCopy() && 609 (use->as_ArrayCopy()->is_clonebasic() || 610 use->as_ArrayCopy()->is_arraycopy_validated() || 611 use->as_ArrayCopy()->is_copyof_validated() || 612 use->as_ArrayCopy()->is_copyofrange_validated()) && 613 use->in(ArrayCopyNode::Dest) == res) { 614 // ok to eliminate 615 } else if (use->is_SafePoint()) { 616 SafePointNode* sfpt = use->as_SafePoint(); 617 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 618 // Object is passed as argument. 619 DEBUG_ONLY(disq_node = use;) 620 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 621 can_eliminate = false; 622 } 623 Node* sfptMem = sfpt->memory(); 624 if (sfptMem == nullptr || sfptMem->is_top()) { 625 DEBUG_ONLY(disq_node = use;) 626 NOT_PRODUCT(fail_eliminate = "null or TOP memory";) 627 can_eliminate = false; 628 } else if (!reduce_merge_precheck) { 629 safepoints->append_if_missing(sfpt); 630 } 631 } else if (reduce_merge_precheck && 632 (use->is_Phi() || use->is_EncodeP() || 633 use->Opcode() == Op_MemBarRelease || 634 (UseStoreStoreForCtor && use->Opcode() == Op_MemBarStoreStore))) { 635 // Nothing to do 636 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 637 if (use->is_Phi()) { 638 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 639 NOT_PRODUCT(fail_eliminate = "Object is return value";) 640 } else { 641 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 642 } 643 DEBUG_ONLY(disq_node = use;) 644 } else { 645 if (use->Opcode() == Op_Return) { 646 NOT_PRODUCT(fail_eliminate = "Object is return value";) 647 } else { 648 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 649 } 650 DEBUG_ONLY(disq_node = use;) 651 } 652 can_eliminate = false; 653 } 654 } 655 } 656 657 #ifndef PRODUCT 658 if (PrintEliminateAllocations && safepoints != nullptr) { 659 if (can_eliminate) { 660 tty->print("Scalar "); 661 if (res == nullptr) 662 alloc->dump(); 663 else 664 res->dump(); 665 } else if (alloc->_is_scalar_replaceable) { 666 tty->print("NotScalar (%s)", fail_eliminate); 667 if (res == nullptr) 668 alloc->dump(); 669 else 670 res->dump(); 671 #ifdef ASSERT 672 if (disq_node != nullptr) { 673 tty->print(" >>>> "); 674 disq_node->dump(); 675 } 676 #endif /*ASSERT*/ 677 } 678 } 679 680 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) { 681 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : ""); 682 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();) 683 } 684 #endif 685 return can_eliminate; 686 } 687 688 void PhaseMacroExpand::undo_previous_scalarizations(GrowableArray <SafePointNode *> safepoints_done, AllocateNode* alloc) { 689 Node* res = alloc->result_cast(); 690 int nfields = 0; 691 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result"); 692 693 if (res != nullptr) { 694 const TypeOopPtr* res_type = _igvn.type(res)->isa_oopptr(); 695 696 if (res_type->isa_instptr()) { 697 // find the fields of the class which will be needed for safepoint debug information 698 ciInstanceKlass* iklass = res_type->is_instptr()->instance_klass(); 699 nfields = iklass->nof_nonstatic_fields(); 700 } else { 701 // find the array's elements which will be needed for safepoint debug information 702 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 703 assert(nfields >= 0, "must be an array klass."); 704 } 705 } 706 707 // rollback processed safepoints 708 while (safepoints_done.length() > 0) { 709 SafePointNode* sfpt_done = safepoints_done.pop(); 710 // remove any extra entries we added to the safepoint 711 uint last = sfpt_done->req() - 1; 712 for (int k = 0; k < nfields; k++) { 713 sfpt_done->del_req(last--); 714 } 715 JVMState *jvms = sfpt_done->jvms(); 716 jvms->set_endoff(sfpt_done->req()); 717 // Now make a pass over the debug information replacing any references 718 // to SafePointScalarObjectNode with the allocated object. 719 int start = jvms->debug_start(); 720 int end = jvms->debug_end(); 721 for (int i = start; i < end; i++) { 722 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 723 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 724 if (scobj->first_index(jvms) == sfpt_done->req() && 725 scobj->n_fields() == (uint)nfields) { 726 assert(scobj->alloc() == alloc, "sanity"); 727 sfpt_done->set_req(i, res); 728 } 729 } 730 } 731 _igvn._worklist.push(sfpt_done); 732 } 733 } 734 735 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode *alloc, SafePointNode* sfpt) { 736 // Fields of scalar objs are referenced only at the end 737 // of regular debuginfo at the last (youngest) JVMS. 738 // Record relative start index. 739 ciInstanceKlass* iklass = nullptr; 740 BasicType basic_elem_type = T_ILLEGAL; 741 const Type* field_type = nullptr; 742 const TypeOopPtr* res_type = nullptr; 743 int nfields = 0; 744 int array_base = 0; 745 int element_size = 0; 746 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff()); 747 Node* res = alloc->result_cast(); 748 749 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result"); 750 assert(sfpt->jvms() != nullptr, "missed JVMS"); 751 752 if (res != nullptr) { // Could be null when there are no users 753 res_type = _igvn.type(res)->isa_oopptr(); 754 755 if (res_type->isa_instptr()) { 756 // find the fields of the class which will be needed for safepoint debug information 757 iklass = res_type->is_instptr()->instance_klass(); 758 nfields = iklass->nof_nonstatic_fields(); 759 } else { 760 // find the array's elements which will be needed for safepoint debug information 761 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 762 assert(nfields >= 0, "must be an array klass."); 763 basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type(); 764 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 765 element_size = type2aelembytes(basic_elem_type); 766 field_type = res_type->is_aryptr()->elem(); 767 } 768 } 769 770 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, sfpt->jvms()->depth(), nfields); 771 sobj->init_req(0, C->root()); 772 transform_later(sobj); 773 774 // Scan object's fields adding an input to the safepoint for each field. 775 for (int j = 0; j < nfields; j++) { 776 intptr_t offset; 777 ciField* field = nullptr; 778 if (iklass != nullptr) { 779 field = iklass->nonstatic_field_at(j); 780 offset = field->offset_in_bytes(); 781 ciType* elem_type = field->type(); 782 basic_elem_type = field->layout_type(); 783 784 // The next code is taken from Parse::do_get_xxx(). 785 if (is_reference_type(basic_elem_type)) { 786 if (!elem_type->is_loaded()) { 787 field_type = TypeInstPtr::BOTTOM; 788 } else if (field != nullptr && field->is_static_constant()) { 789 ciObject* con = field->constant_value().as_object(); 790 // Do not "join" in the previous type; it doesn't add value, 791 // and may yield a vacuous result if the field is of interface type. 792 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 793 assert(field_type != nullptr, "field singleton type must be consistent"); 794 } else { 795 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 796 } 797 if (UseCompressedOops) { 798 field_type = field_type->make_narrowoop(); 799 basic_elem_type = T_NARROWOOP; 800 } 801 } else { 802 field_type = Type::get_const_basic_type(basic_elem_type); 803 } 804 } else { 805 offset = array_base + j * (intptr_t)element_size; 806 } 807 808 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 809 810 Node *field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc); 811 812 // We weren't able to find a value for this field, 813 // give up on eliminating this allocation. 814 if (field_val == nullptr) { 815 uint last = sfpt->req() - 1; 816 for (int k = 0; k < j; k++) { 817 sfpt->del_req(last--); 818 } 819 _igvn._worklist.push(sfpt); 820 821 #ifndef PRODUCT 822 if (PrintEliminateAllocations) { 823 if (field != nullptr) { 824 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx); 825 field->print(); 826 int field_idx = C->get_alias_index(field_addr_type); 827 tty->print(" (alias_idx=%d)", field_idx); 828 } else { // Array's element 829 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, j); 830 } 831 tty->print(", which prevents elimination of: "); 832 if (res == nullptr) 833 alloc->dump(); 834 else 835 res->dump(); 836 } 837 #endif 838 839 return nullptr; 840 } 841 842 if (UseCompressedOops && field_type->isa_narrowoop()) { 843 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 844 // to be able scalar replace the allocation. 845 if (field_val->is_EncodeP()) { 846 field_val = field_val->in(1); 847 } else { 848 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type())); 849 } 850 } 851 sfpt->add_req(field_val); 852 } 853 854 sfpt->jvms()->set_endoff(sfpt->req()); 855 856 return sobj; 857 } 858 859 // Do scalar replacement. 860 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 861 GrowableArray <SafePointNode *> safepoints_done; 862 Node* res = alloc->result_cast(); 863 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result"); 864 865 // Process the safepoint uses 866 while (safepoints.length() > 0) { 867 SafePointNode* sfpt = safepoints.pop(); 868 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt); 869 870 if (sobj == nullptr) { 871 undo_previous_scalarizations(safepoints_done, alloc); 872 return false; 873 } 874 875 // Now make a pass over the debug information replacing any references 876 // to the allocated object with "sobj" 877 JVMState *jvms = sfpt->jvms(); 878 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn); 879 _igvn._worklist.push(sfpt); 880 881 // keep it for rollback 882 safepoints_done.append_if_missing(sfpt); 883 } 884 885 return true; 886 } 887 888 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) { 889 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control); 890 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory); 891 if (ctl_proj != nullptr) { 892 igvn.replace_node(ctl_proj, n->in(0)); 893 } 894 if (mem_proj != nullptr) { 895 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory)); 896 } 897 } 898 899 // Process users of eliminated allocation. 900 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) { 901 Node* res = alloc->result_cast(); 902 if (res != nullptr) { 903 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 904 Node *use = res->last_out(j); 905 uint oc1 = res->outcnt(); 906 907 if (use->is_AddP()) { 908 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 909 Node *n = use->last_out(k); 910 uint oc2 = use->outcnt(); 911 if (n->is_Store()) { 912 #ifdef ASSERT 913 // Verify that there is no dependent MemBarVolatile nodes, 914 // they should be removed during IGVN, see MemBarNode::Ideal(). 915 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); 916 p < pmax; p++) { 917 Node* mb = n->fast_out(p); 918 assert(mb->is_Initialize() || !mb->is_MemBar() || 919 mb->req() <= MemBarNode::Precedent || 920 mb->in(MemBarNode::Precedent) != n, 921 "MemBarVolatile should be eliminated for non-escaping object"); 922 } 923 #endif 924 _igvn.replace_node(n, n->in(MemNode::Memory)); 925 } else { 926 eliminate_gc_barrier(n); 927 } 928 k -= (oc2 - use->outcnt()); 929 } 930 _igvn.remove_dead_node(use); 931 } else if (use->is_ArrayCopy()) { 932 // Disconnect ArrayCopy node 933 ArrayCopyNode* ac = use->as_ArrayCopy(); 934 if (ac->is_clonebasic()) { 935 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out(); 936 disconnect_projections(ac, _igvn); 937 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation"); 938 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0); 939 disconnect_projections(membar_before->as_MemBar(), _igvn); 940 if (membar_after->is_MemBar()) { 941 disconnect_projections(membar_after->as_MemBar(), _igvn); 942 } 943 } else { 944 assert(ac->is_arraycopy_validated() || 945 ac->is_copyof_validated() || 946 ac->is_copyofrange_validated(), "unsupported"); 947 CallProjections callprojs; 948 ac->extract_projections(&callprojs, true); 949 950 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O)); 951 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory)); 952 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control)); 953 954 // Set control to top. IGVN will remove the remaining projections 955 ac->set_req(0, top()); 956 ac->replace_edge(res, top(), &_igvn); 957 958 // Disconnect src right away: it can help find new 959 // opportunities for allocation elimination 960 Node* src = ac->in(ArrayCopyNode::Src); 961 ac->replace_edge(src, top(), &_igvn); 962 // src can be top at this point if src and dest of the 963 // arraycopy were the same 964 if (src->outcnt() == 0 && !src->is_top()) { 965 _igvn.remove_dead_node(src); 966 } 967 } 968 _igvn._worklist.push(ac); 969 } else { 970 eliminate_gc_barrier(use); 971 } 972 j -= (oc1 - res->outcnt()); 973 } 974 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 975 _igvn.remove_dead_node(res); 976 } 977 978 // 979 // Process other users of allocation's projections 980 // 981 if (_callprojs.resproj != nullptr && _callprojs.resproj->outcnt() != 0) { 982 // First disconnect stores captured by Initialize node. 983 // If Initialize node is eliminated first in the following code, 984 // it will kill such stores and DUIterator_Last will assert. 985 for (DUIterator_Fast jmax, j = _callprojs.resproj->fast_outs(jmax); j < jmax; j++) { 986 Node* use = _callprojs.resproj->fast_out(j); 987 if (use->is_AddP()) { 988 // raw memory addresses used only by the initialization 989 _igvn.replace_node(use, C->top()); 990 --j; --jmax; 991 } 992 } 993 for (DUIterator_Last jmin, j = _callprojs.resproj->last_outs(jmin); j >= jmin; ) { 994 Node* use = _callprojs.resproj->last_out(j); 995 uint oc1 = _callprojs.resproj->outcnt(); 996 if (use->is_Initialize()) { 997 // Eliminate Initialize node. 998 InitializeNode *init = use->as_Initialize(); 999 assert(init->outcnt() <= 2, "only a control and memory projection expected"); 1000 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control); 1001 if (ctrl_proj != nullptr) { 1002 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control)); 1003 #ifdef ASSERT 1004 // If the InitializeNode has no memory out, it will die, and tmp will become null 1005 Node* tmp = init->in(TypeFunc::Control); 1006 assert(tmp == nullptr || tmp == _callprojs.fallthrough_catchproj, "allocation control projection"); 1007 #endif 1008 } 1009 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory); 1010 if (mem_proj != nullptr) { 1011 Node *mem = init->in(TypeFunc::Memory); 1012 #ifdef ASSERT 1013 if (mem->is_MergeMem()) { 1014 assert(mem->in(TypeFunc::Memory) == _callprojs.fallthrough_memproj, "allocation memory projection"); 1015 } else { 1016 assert(mem == _callprojs.fallthrough_memproj, "allocation memory projection"); 1017 } 1018 #endif 1019 _igvn.replace_node(mem_proj, mem); 1020 } 1021 } else { 1022 assert(false, "only Initialize or AddP expected"); 1023 } 1024 j -= (oc1 - _callprojs.resproj->outcnt()); 1025 } 1026 } 1027 if (_callprojs.fallthrough_catchproj != nullptr) { 1028 _igvn.replace_node(_callprojs.fallthrough_catchproj, alloc->in(TypeFunc::Control)); 1029 } 1030 if (_callprojs.fallthrough_memproj != nullptr) { 1031 _igvn.replace_node(_callprojs.fallthrough_memproj, alloc->in(TypeFunc::Memory)); 1032 } 1033 if (_callprojs.catchall_memproj != nullptr) { 1034 _igvn.replace_node(_callprojs.catchall_memproj, C->top()); 1035 } 1036 if (_callprojs.fallthrough_ioproj != nullptr) { 1037 _igvn.replace_node(_callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 1038 } 1039 if (_callprojs.catchall_ioproj != nullptr) { 1040 _igvn.replace_node(_callprojs.catchall_ioproj, C->top()); 1041 } 1042 if (_callprojs.catchall_catchproj != nullptr) { 1043 _igvn.replace_node(_callprojs.catchall_catchproj, C->top()); 1044 } 1045 } 1046 1047 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 1048 // If reallocation fails during deoptimization we'll pop all 1049 // interpreter frames for this compiled frame and that won't play 1050 // nice with JVMTI popframe. 1051 // We avoid this issue by eager reallocation when the popframe request 1052 // is received. 1053 if (!EliminateAllocations || !alloc->_is_non_escaping) { 1054 return false; 1055 } 1056 Node* klass = alloc->in(AllocateNode::KlassNode); 1057 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); 1058 Node* res = alloc->result_cast(); 1059 // Eliminate boxing allocations which are not used 1060 // regardless scalar replaceable status. 1061 bool boxing_alloc = C->eliminate_boxing() && 1062 tklass->isa_instklassptr() && 1063 tklass->is_instklassptr()->instance_klass()->is_box_klass(); 1064 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != nullptr))) { 1065 return false; 1066 } 1067 1068 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/); 1069 1070 GrowableArray <SafePointNode *> safepoints; 1071 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) { 1072 return false; 1073 } 1074 1075 if (!alloc->_is_scalar_replaceable) { 1076 assert(res == nullptr, "sanity"); 1077 // We can only eliminate allocation if all debug info references 1078 // are already replaced with SafePointScalarObject because 1079 // we can't search for a fields value without instance_id. 1080 if (safepoints.length() > 0) { 1081 return false; 1082 } 1083 } 1084 1085 if (!scalar_replacement(alloc, safepoints)) { 1086 return false; 1087 } 1088 1089 CompileLog* log = C->log(); 1090 if (log != nullptr) { 1091 log->head("eliminate_allocation type='%d'", 1092 log->identify(tklass->exact_klass())); 1093 JVMState* p = alloc->jvms(); 1094 while (p != nullptr) { 1095 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1096 p = p->caller(); 1097 } 1098 log->tail("eliminate_allocation"); 1099 } 1100 1101 process_users_of_allocation(alloc); 1102 1103 #ifndef PRODUCT 1104 if (PrintEliminateAllocations) { 1105 if (alloc->is_AllocateArray()) 1106 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1107 else 1108 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1109 } 1110 #endif 1111 1112 return true; 1113 } 1114 1115 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) { 1116 // EA should remove all uses of non-escaping boxing node. 1117 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) { 1118 return false; 1119 } 1120 1121 assert(boxing->result_cast() == nullptr, "unexpected boxing node result"); 1122 1123 boxing->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/); 1124 1125 const TypeTuple* r = boxing->tf()->range(); 1126 assert(r->cnt() > TypeFunc::Parms, "sanity"); 1127 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr(); 1128 assert(t != nullptr, "sanity"); 1129 1130 CompileLog* log = C->log(); 1131 if (log != nullptr) { 1132 log->head("eliminate_boxing type='%d'", 1133 log->identify(t->instance_klass())); 1134 JVMState* p = boxing->jvms(); 1135 while (p != nullptr) { 1136 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1137 p = p->caller(); 1138 } 1139 log->tail("eliminate_boxing"); 1140 } 1141 1142 process_users_of_allocation(boxing); 1143 1144 #ifndef PRODUCT 1145 if (PrintEliminateAllocations) { 1146 tty->print("++++ Eliminated: %d ", boxing->_idx); 1147 boxing->method()->print_short_name(tty); 1148 tty->cr(); 1149 } 1150 #endif 1151 1152 return true; 1153 } 1154 1155 1156 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 1157 Node* adr = basic_plus_adr(base, offset); 1158 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 1159 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered); 1160 transform_later(value); 1161 return value; 1162 } 1163 1164 1165 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 1166 Node* adr = basic_plus_adr(base, offset); 1167 mem = StoreNode::make(_igvn, ctl, mem, adr, nullptr, value, bt, MemNode::unordered); 1168 transform_later(mem); 1169 return mem; 1170 } 1171 1172 //============================================================================= 1173 // 1174 // A L L O C A T I O N 1175 // 1176 // Allocation attempts to be fast in the case of frequent small objects. 1177 // It breaks down like this: 1178 // 1179 // 1) Size in doublewords is computed. This is a constant for objects and 1180 // variable for most arrays. Doubleword units are used to avoid size 1181 // overflow of huge doubleword arrays. We need doublewords in the end for 1182 // rounding. 1183 // 1184 // 2) Size is checked for being 'too large'. Too-large allocations will go 1185 // the slow path into the VM. The slow path can throw any required 1186 // exceptions, and does all the special checks for very large arrays. The 1187 // size test can constant-fold away for objects. For objects with 1188 // finalizers it constant-folds the otherway: you always go slow with 1189 // finalizers. 1190 // 1191 // 3) If NOT using TLABs, this is the contended loop-back point. 1192 // Load-Locked the heap top. If using TLABs normal-load the heap top. 1193 // 1194 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 1195 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 1196 // "size*8" we always enter the VM, where "largish" is a constant picked small 1197 // enough that there's always space between the eden max and 4Gig (old space is 1198 // there so it's quite large) and large enough that the cost of entering the VM 1199 // is dwarfed by the cost to initialize the space. 1200 // 1201 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 1202 // down. If contended, repeat at step 3. If using TLABs normal-store 1203 // adjusted heap top back down; there is no contention. 1204 // 1205 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 1206 // fields. 1207 // 1208 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 1209 // oop flavor. 1210 // 1211 //============================================================================= 1212 // FastAllocateSizeLimit value is in DOUBLEWORDS. 1213 // Allocations bigger than this always go the slow route. 1214 // This value must be small enough that allocation attempts that need to 1215 // trigger exceptions go the slow route. Also, it must be small enough so 1216 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 1217 //=============================================================================j// 1218 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 1219 // The allocator will coalesce int->oop copies away. See comment in 1220 // coalesce.cpp about how this works. It depends critically on the exact 1221 // code shape produced here, so if you are changing this code shape 1222 // make sure the GC info for the heap-top is correct in and around the 1223 // slow-path call. 1224 // 1225 1226 void PhaseMacroExpand::expand_allocate_common( 1227 AllocateNode* alloc, // allocation node to be expanded 1228 Node* length, // array length for an array allocation 1229 const TypeFunc* slow_call_type, // Type of slow call 1230 address slow_call_address, // Address of slow call 1231 Node* valid_length_test // whether length is valid or not 1232 ) 1233 { 1234 Node* ctrl = alloc->in(TypeFunc::Control); 1235 Node* mem = alloc->in(TypeFunc::Memory); 1236 Node* i_o = alloc->in(TypeFunc::I_O); 1237 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 1238 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1239 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 1240 assert(ctrl != nullptr, "must have control"); 1241 1242 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 1243 // they will not be used if "always_slow" is set 1244 enum { slow_result_path = 1, fast_result_path = 2 }; 1245 Node *result_region = nullptr; 1246 Node *result_phi_rawmem = nullptr; 1247 Node *result_phi_rawoop = nullptr; 1248 Node *result_phi_i_o = nullptr; 1249 1250 // The initial slow comparison is a size check, the comparison 1251 // we want to do is a BoolTest::gt 1252 bool expand_fast_path = true; 1253 int tv = _igvn.find_int_con(initial_slow_test, -1); 1254 if (tv >= 0) { 1255 // InitialTest has constant result 1256 // 0 - can fit in TLAB 1257 // 1 - always too big or negative 1258 assert(tv <= 1, "0 or 1 if a constant"); 1259 expand_fast_path = (tv == 0); 1260 initial_slow_test = nullptr; 1261 } else { 1262 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 1263 } 1264 1265 if (!UseTLAB) { 1266 // Force slow-path allocation 1267 expand_fast_path = false; 1268 initial_slow_test = nullptr; 1269 } 1270 1271 bool allocation_has_use = (alloc->result_cast() != nullptr); 1272 if (!allocation_has_use) { 1273 InitializeNode* init = alloc->initialization(); 1274 if (init != nullptr) { 1275 init->remove(&_igvn); 1276 } 1277 if (expand_fast_path && (initial_slow_test == nullptr)) { 1278 // Remove allocation node and return. 1279 // Size is a non-negative constant -> no initial check needed -> directly to fast path. 1280 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left. 1281 #ifndef PRODUCT 1282 if (PrintEliminateAllocations) { 1283 tty->print("NotUsed "); 1284 Node* res = alloc->proj_out_or_null(TypeFunc::Parms); 1285 if (res != nullptr) { 1286 res->dump(); 1287 } else { 1288 alloc->dump(); 1289 } 1290 } 1291 #endif 1292 yank_alloc_node(alloc); 1293 return; 1294 } 1295 } 1296 1297 enum { too_big_or_final_path = 1, need_gc_path = 2 }; 1298 Node *slow_region = nullptr; 1299 Node *toobig_false = ctrl; 1300 1301 // generate the initial test if necessary 1302 if (initial_slow_test != nullptr ) { 1303 assert (expand_fast_path, "Only need test if there is a fast path"); 1304 slow_region = new RegionNode(3); 1305 1306 // Now make the initial failure test. Usually a too-big test but 1307 // might be a TRUE for finalizers. 1308 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1309 transform_later(toobig_iff); 1310 // Plug the failing-too-big test into the slow-path region 1311 Node *toobig_true = new IfTrueNode( toobig_iff ); 1312 transform_later(toobig_true); 1313 slow_region ->init_req( too_big_or_final_path, toobig_true ); 1314 toobig_false = new IfFalseNode( toobig_iff ); 1315 transform_later(toobig_false); 1316 } else { 1317 // No initial test, just fall into next case 1318 assert(allocation_has_use || !expand_fast_path, "Should already have been handled"); 1319 toobig_false = ctrl; 1320 DEBUG_ONLY(slow_region = NodeSentinel); 1321 } 1322 1323 // If we are here there are several possibilities 1324 // - expand_fast_path is false - then only a slow path is expanded. That's it. 1325 // no_initial_check means a constant allocation. 1326 // - If check always evaluates to false -> expand_fast_path is false (see above) 1327 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath) 1328 // if !allocation_has_use the fast path is empty 1329 // if !allocation_has_use && no_initial_check 1330 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all. 1331 // removed by yank_alloc_node above. 1332 1333 Node *slow_mem = mem; // save the current memory state for slow path 1334 // generate the fast allocation code unless we know that the initial test will always go slow 1335 if (expand_fast_path) { 1336 // Fast path modifies only raw memory. 1337 if (mem->is_MergeMem()) { 1338 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1339 } 1340 1341 // allocate the Region and Phi nodes for the result 1342 result_region = new RegionNode(3); 1343 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1344 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch 1345 1346 // Grab regular I/O before optional prefetch may change it. 1347 // Slow-path does no I/O so just set it to the original I/O. 1348 result_phi_i_o->init_req(slow_result_path, i_o); 1349 1350 // Name successful fast-path variables 1351 Node* fast_oop_ctrl; 1352 Node* fast_oop_rawmem; 1353 if (allocation_has_use) { 1354 Node* needgc_ctrl = nullptr; 1355 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM); 1356 1357 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1358 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1359 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl, 1360 fast_oop_ctrl, fast_oop_rawmem, 1361 prefetch_lines); 1362 1363 if (initial_slow_test != nullptr) { 1364 // This completes all paths into the slow merge point 1365 slow_region->init_req(need_gc_path, needgc_ctrl); 1366 transform_later(slow_region); 1367 } else { 1368 // No initial slow path needed! 1369 // Just fall from the need-GC path straight into the VM call. 1370 slow_region = needgc_ctrl; 1371 } 1372 1373 InitializeNode* init = alloc->initialization(); 1374 fast_oop_rawmem = initialize_object(alloc, 1375 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1376 klass_node, length, size_in_bytes); 1377 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem); 1378 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem); 1379 1380 result_phi_rawoop->init_req(fast_result_path, fast_oop); 1381 } else { 1382 assert (initial_slow_test != nullptr, "sanity"); 1383 fast_oop_ctrl = toobig_false; 1384 fast_oop_rawmem = mem; 1385 transform_later(slow_region); 1386 } 1387 1388 // Plug in the successful fast-path into the result merge point 1389 result_region ->init_req(fast_result_path, fast_oop_ctrl); 1390 result_phi_i_o ->init_req(fast_result_path, i_o); 1391 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); 1392 } else { 1393 slow_region = ctrl; 1394 result_phi_i_o = i_o; // Rename it to use in the following code. 1395 } 1396 1397 // Generate slow-path call 1398 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address, 1399 OptoRuntime::stub_name(slow_call_address), 1400 TypePtr::BOTTOM); 1401 call->init_req(TypeFunc::Control, slow_region); 1402 call->init_req(TypeFunc::I_O, top()); // does no i/o 1403 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs 1404 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1405 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1406 1407 call->init_req(TypeFunc::Parms+0, klass_node); 1408 if (length != nullptr) { 1409 call->init_req(TypeFunc::Parms+1, length); 1410 } 1411 1412 // Copy debug information and adjust JVMState information, then replace 1413 // allocate node with the call 1414 call->copy_call_debug_info(&_igvn, alloc); 1415 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify 1416 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough 1417 // path dies). 1418 if (valid_length_test != nullptr) { 1419 call->add_req(valid_length_test); 1420 } 1421 if (expand_fast_path) { 1422 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1423 } else { 1424 // Hook i_o projection to avoid its elimination during allocation 1425 // replacement (when only a slow call is generated). 1426 call->set_req(TypeFunc::I_O, result_phi_i_o); 1427 } 1428 _igvn.replace_node(alloc, call); 1429 transform_later(call); 1430 1431 // Identify the output projections from the allocate node and 1432 // adjust any references to them. 1433 // The control and io projections look like: 1434 // 1435 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1436 // Allocate Catch 1437 // ^---Proj(io) <-------+ ^---CatchProj(io) 1438 // 1439 // We are interested in the CatchProj nodes. 1440 // 1441 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/); 1442 1443 // An allocate node has separate memory projections for the uses on 1444 // the control and i_o paths. Replace the control memory projection with 1445 // result_phi_rawmem (unless we are only generating a slow call when 1446 // both memory projections are combined) 1447 if (expand_fast_path && _callprojs.fallthrough_memproj != nullptr) { 1448 migrate_outs(_callprojs.fallthrough_memproj, result_phi_rawmem); 1449 } 1450 // Now change uses of catchall_memproj to use fallthrough_memproj and delete 1451 // catchall_memproj so we end up with a call that has only 1 memory projection. 1452 if (_callprojs.catchall_memproj != nullptr ) { 1453 if (_callprojs.fallthrough_memproj == nullptr) { 1454 _callprojs.fallthrough_memproj = new ProjNode(call, TypeFunc::Memory); 1455 transform_later(_callprojs.fallthrough_memproj); 1456 } 1457 migrate_outs(_callprojs.catchall_memproj, _callprojs.fallthrough_memproj); 1458 _igvn.remove_dead_node(_callprojs.catchall_memproj); 1459 } 1460 1461 // An allocate node has separate i_o projections for the uses on the control 1462 // and i_o paths. Always replace the control i_o projection with result i_o 1463 // otherwise incoming i_o become dead when only a slow call is generated 1464 // (it is different from memory projections where both projections are 1465 // combined in such case). 1466 if (_callprojs.fallthrough_ioproj != nullptr) { 1467 migrate_outs(_callprojs.fallthrough_ioproj, result_phi_i_o); 1468 } 1469 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete 1470 // catchall_ioproj so we end up with a call that has only 1 i_o projection. 1471 if (_callprojs.catchall_ioproj != nullptr ) { 1472 if (_callprojs.fallthrough_ioproj == nullptr) { 1473 _callprojs.fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O); 1474 transform_later(_callprojs.fallthrough_ioproj); 1475 } 1476 migrate_outs(_callprojs.catchall_ioproj, _callprojs.fallthrough_ioproj); 1477 _igvn.remove_dead_node(_callprojs.catchall_ioproj); 1478 } 1479 1480 // if we generated only a slow call, we are done 1481 if (!expand_fast_path) { 1482 // Now we can unhook i_o. 1483 if (result_phi_i_o->outcnt() > 1) { 1484 call->set_req(TypeFunc::I_O, top()); 1485 } else { 1486 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity"); 1487 // Case of new array with negative size known during compilation. 1488 // AllocateArrayNode::Ideal() optimization disconnect unreachable 1489 // following code since call to runtime will throw exception. 1490 // As result there will be no users of i_o after the call. 1491 // Leave i_o attached to this call to avoid problems in preceding graph. 1492 } 1493 return; 1494 } 1495 1496 if (_callprojs.fallthrough_catchproj != nullptr) { 1497 ctrl = _callprojs.fallthrough_catchproj->clone(); 1498 transform_later(ctrl); 1499 _igvn.replace_node(_callprojs.fallthrough_catchproj, result_region); 1500 } else { 1501 ctrl = top(); 1502 } 1503 Node *slow_result; 1504 if (_callprojs.resproj == nullptr) { 1505 // no uses of the allocation result 1506 slow_result = top(); 1507 } else { 1508 slow_result = _callprojs.resproj->clone(); 1509 transform_later(slow_result); 1510 _igvn.replace_node(_callprojs.resproj, result_phi_rawoop); 1511 } 1512 1513 // Plug slow-path into result merge point 1514 result_region->init_req( slow_result_path, ctrl); 1515 transform_later(result_region); 1516 if (allocation_has_use) { 1517 result_phi_rawoop->init_req(slow_result_path, slow_result); 1518 transform_later(result_phi_rawoop); 1519 } 1520 result_phi_rawmem->init_req(slow_result_path, _callprojs.fallthrough_memproj); 1521 transform_later(result_phi_rawmem); 1522 transform_later(result_phi_i_o); 1523 // This completes all paths into the result merge point 1524 } 1525 1526 // Remove alloc node that has no uses. 1527 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) { 1528 Node* ctrl = alloc->in(TypeFunc::Control); 1529 Node* mem = alloc->in(TypeFunc::Memory); 1530 Node* i_o = alloc->in(TypeFunc::I_O); 1531 1532 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/); 1533 if (_callprojs.resproj != nullptr) { 1534 for (DUIterator_Fast imax, i = _callprojs.resproj->fast_outs(imax); i < imax; i++) { 1535 Node* use = _callprojs.resproj->fast_out(i); 1536 use->isa_MemBar()->remove(&_igvn); 1537 --imax; 1538 --i; // back up iterator 1539 } 1540 assert(_callprojs.resproj->outcnt() == 0, "all uses must be deleted"); 1541 _igvn.remove_dead_node(_callprojs.resproj); 1542 } 1543 if (_callprojs.fallthrough_catchproj != nullptr) { 1544 migrate_outs(_callprojs.fallthrough_catchproj, ctrl); 1545 _igvn.remove_dead_node(_callprojs.fallthrough_catchproj); 1546 } 1547 if (_callprojs.catchall_catchproj != nullptr) { 1548 _igvn.rehash_node_delayed(_callprojs.catchall_catchproj); 1549 _callprojs.catchall_catchproj->set_req(0, top()); 1550 } 1551 if (_callprojs.fallthrough_proj != nullptr) { 1552 Node* catchnode = _callprojs.fallthrough_proj->unique_ctrl_out(); 1553 _igvn.remove_dead_node(catchnode); 1554 _igvn.remove_dead_node(_callprojs.fallthrough_proj); 1555 } 1556 if (_callprojs.fallthrough_memproj != nullptr) { 1557 migrate_outs(_callprojs.fallthrough_memproj, mem); 1558 _igvn.remove_dead_node(_callprojs.fallthrough_memproj); 1559 } 1560 if (_callprojs.fallthrough_ioproj != nullptr) { 1561 migrate_outs(_callprojs.fallthrough_ioproj, i_o); 1562 _igvn.remove_dead_node(_callprojs.fallthrough_ioproj); 1563 } 1564 if (_callprojs.catchall_memproj != nullptr) { 1565 _igvn.rehash_node_delayed(_callprojs.catchall_memproj); 1566 _callprojs.catchall_memproj->set_req(0, top()); 1567 } 1568 if (_callprojs.catchall_ioproj != nullptr) { 1569 _igvn.rehash_node_delayed(_callprojs.catchall_ioproj); 1570 _callprojs.catchall_ioproj->set_req(0, top()); 1571 } 1572 #ifndef PRODUCT 1573 if (PrintEliminateAllocations) { 1574 if (alloc->is_AllocateArray()) { 1575 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1576 } else { 1577 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1578 } 1579 } 1580 #endif 1581 _igvn.remove_dead_node(alloc); 1582 } 1583 1584 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init, 1585 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) { 1586 // If initialization is performed by an array copy, any required 1587 // MemBarStoreStore was already added. If the object does not 1588 // escape no need for a MemBarStoreStore. If the object does not 1589 // escape in its initializer and memory barrier (MemBarStoreStore or 1590 // stronger) is already added at exit of initializer, also no need 1591 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore 1592 // so that stores that initialize this object can't be reordered 1593 // with a subsequent store that makes this object accessible by 1594 // other threads. 1595 // Other threads include java threads and JVM internal threads 1596 // (for example concurrent GC threads). Current concurrent GC 1597 // implementation: G1 will not scan newly created object, 1598 // so it's safe to skip storestore barrier when allocation does 1599 // not escape. 1600 if (!alloc->does_not_escape_thread() && 1601 !alloc->is_allocation_MemBar_redundant() && 1602 (init == nullptr || !init->is_complete_with_arraycopy())) { 1603 if (init == nullptr || init->req() < InitializeNode::RawStores) { 1604 // No InitializeNode or no stores captured by zeroing 1605 // elimination. Simply add the MemBarStoreStore after object 1606 // initialization. 1607 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1608 transform_later(mb); 1609 1610 mb->init_req(TypeFunc::Memory, fast_oop_rawmem); 1611 mb->init_req(TypeFunc::Control, fast_oop_ctrl); 1612 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control); 1613 transform_later(fast_oop_ctrl); 1614 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory); 1615 transform_later(fast_oop_rawmem); 1616 } else { 1617 // Add the MemBarStoreStore after the InitializeNode so that 1618 // all stores performing the initialization that were moved 1619 // before the InitializeNode happen before the storestore 1620 // barrier. 1621 1622 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control); 1623 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory); 1624 1625 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1626 transform_later(mb); 1627 1628 Node* ctrl = new ProjNode(init, TypeFunc::Control); 1629 transform_later(ctrl); 1630 Node* mem = new ProjNode(init, TypeFunc::Memory); 1631 transform_later(mem); 1632 1633 // The MemBarStoreStore depends on control and memory coming 1634 // from the InitializeNode 1635 mb->init_req(TypeFunc::Memory, mem); 1636 mb->init_req(TypeFunc::Control, ctrl); 1637 1638 ctrl = new ProjNode(mb, TypeFunc::Control); 1639 transform_later(ctrl); 1640 mem = new ProjNode(mb, TypeFunc::Memory); 1641 transform_later(mem); 1642 1643 // All nodes that depended on the InitializeNode for control 1644 // and memory must now depend on the MemBarNode that itself 1645 // depends on the InitializeNode 1646 if (init_ctrl != nullptr) { 1647 _igvn.replace_node(init_ctrl, ctrl); 1648 } 1649 if (init_mem != nullptr) { 1650 _igvn.replace_node(init_mem, mem); 1651 } 1652 } 1653 } 1654 } 1655 1656 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop, 1657 Node*& ctrl, Node*& rawmem) { 1658 if (C->env()->dtrace_alloc_probes()) { 1659 // Slow-path call 1660 int size = TypeFunc::Parms + 2; 1661 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1662 CAST_FROM_FN_PTR(address, 1663 static_cast<int (*)(JavaThread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)), 1664 "dtrace_object_alloc", 1665 TypeRawPtr::BOTTOM); 1666 1667 // Get base of thread-local storage area 1668 Node* thread = new ThreadLocalNode(); 1669 transform_later(thread); 1670 1671 call->init_req(TypeFunc::Parms + 0, thread); 1672 call->init_req(TypeFunc::Parms + 1, oop); 1673 call->init_req(TypeFunc::Control, ctrl); 1674 call->init_req(TypeFunc::I_O , top()); // does no i/o 1675 call->init_req(TypeFunc::Memory , rawmem); 1676 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1677 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1678 transform_later(call); 1679 ctrl = new ProjNode(call, TypeFunc::Control); 1680 transform_later(ctrl); 1681 rawmem = new ProjNode(call, TypeFunc::Memory); 1682 transform_later(rawmem); 1683 } 1684 } 1685 1686 // Helper for PhaseMacroExpand::expand_allocate_common. 1687 // Initializes the newly-allocated storage. 1688 Node* 1689 PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1690 Node* control, Node* rawmem, Node* object, 1691 Node* klass_node, Node* length, 1692 Node* size_in_bytes) { 1693 InitializeNode* init = alloc->initialization(); 1694 // Store the klass & mark bits 1695 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem); 1696 if (!mark_node->is_Con()) { 1697 transform_later(mark_node); 1698 } 1699 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type()); 1700 1701 if (!UseCompactObjectHeaders) { 1702 rawmem = make_store(control, rawmem, object, Type::klass_offset(), klass_node, T_METADATA); 1703 } 1704 int header_size = alloc->minimum_header_size(); // conservatively small 1705 1706 // Array length 1707 if (length != nullptr) { // Arrays need length field 1708 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1709 // conservatively small header size: 1710 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1711 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases: 1712 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type(); 1713 if (is_reference_type(elem, true)) { 1714 elem = T_OBJECT; 1715 } 1716 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem)); 1717 } 1718 } 1719 1720 // Clear the object body, if necessary. 1721 if (init == nullptr) { 1722 // The init has somehow disappeared; be cautious and clear everything. 1723 // 1724 // This can happen if a node is allocated but an uncommon trap occurs 1725 // immediately. In this case, the Initialize gets associated with the 1726 // trap, and may be placed in a different (outer) loop, if the Allocate 1727 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1728 // there can be two Allocates to one Initialize. The answer in all these 1729 // edge cases is safety first. It is always safe to clear immediately 1730 // within an Allocate, and then (maybe or maybe not) clear some more later. 1731 if (!(UseTLAB && ZeroTLAB)) { 1732 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1733 header_size, size_in_bytes, 1734 &_igvn); 1735 } 1736 } else { 1737 if (!init->is_complete()) { 1738 // Try to win by zeroing only what the init does not store. 1739 // We can also try to do some peephole optimizations, 1740 // such as combining some adjacent subword stores. 1741 rawmem = init->complete_stores(control, rawmem, object, 1742 header_size, size_in_bytes, &_igvn); 1743 } 1744 // We have no more use for this link, since the AllocateNode goes away: 1745 init->set_req(InitializeNode::RawAddress, top()); 1746 // (If we keep the link, it just confuses the register allocator, 1747 // who thinks he sees a real use of the address by the membar.) 1748 } 1749 1750 return rawmem; 1751 } 1752 1753 // Generate prefetch instructions for next allocations. 1754 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1755 Node*& contended_phi_rawmem, 1756 Node* old_eden_top, Node* new_eden_top, 1757 intx lines) { 1758 enum { fall_in_path = 1, pf_path = 2 }; 1759 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1760 // Generate prefetch allocation with watermark check. 1761 // As an allocation hits the watermark, we will prefetch starting 1762 // at a "distance" away from watermark. 1763 1764 Node *pf_region = new RegionNode(3); 1765 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 1766 TypeRawPtr::BOTTOM ); 1767 // I/O is used for Prefetch 1768 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO ); 1769 1770 Node *thread = new ThreadLocalNode(); 1771 transform_later(thread); 1772 1773 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread, 1774 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1775 transform_later(eden_pf_adr); 1776 1777 Node *old_pf_wm = new LoadPNode(needgc_false, 1778 contended_phi_rawmem, eden_pf_adr, 1779 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, 1780 MemNode::unordered); 1781 transform_later(old_pf_wm); 1782 1783 // check against new_eden_top 1784 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm ); 1785 transform_later(need_pf_cmp); 1786 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge ); 1787 transform_later(need_pf_bol); 1788 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol, 1789 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1790 transform_later(need_pf_iff); 1791 1792 // true node, add prefetchdistance 1793 Node *need_pf_true = new IfTrueNode( need_pf_iff ); 1794 transform_later(need_pf_true); 1795 1796 Node *need_pf_false = new IfFalseNode( need_pf_iff ); 1797 transform_later(need_pf_false); 1798 1799 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm, 1800 _igvn.MakeConX(AllocatePrefetchDistance) ); 1801 transform_later(new_pf_wmt ); 1802 new_pf_wmt->set_req(0, need_pf_true); 1803 1804 Node *store_new_wmt = new StorePNode(need_pf_true, 1805 contended_phi_rawmem, eden_pf_adr, 1806 TypeRawPtr::BOTTOM, new_pf_wmt, 1807 MemNode::unordered); 1808 transform_later(store_new_wmt); 1809 1810 // adding prefetches 1811 pf_phi_abio->init_req( fall_in_path, i_o ); 1812 1813 Node *prefetch_adr; 1814 Node *prefetch; 1815 uint step_size = AllocatePrefetchStepSize; 1816 uint distance = 0; 1817 1818 for ( intx i = 0; i < lines; i++ ) { 1819 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt, 1820 _igvn.MakeConX(distance) ); 1821 transform_later(prefetch_adr); 1822 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1823 transform_later(prefetch); 1824 distance += step_size; 1825 i_o = prefetch; 1826 } 1827 pf_phi_abio->set_req( pf_path, i_o ); 1828 1829 pf_region->init_req( fall_in_path, need_pf_false ); 1830 pf_region->init_req( pf_path, need_pf_true ); 1831 1832 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 1833 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 1834 1835 transform_later(pf_region); 1836 transform_later(pf_phi_rawmem); 1837 transform_later(pf_phi_abio); 1838 1839 needgc_false = pf_region; 1840 contended_phi_rawmem = pf_phi_rawmem; 1841 i_o = pf_phi_abio; 1842 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { 1843 // Insert a prefetch instruction for each allocation. 1844 // This code is used to generate 1 prefetch instruction per cache line. 1845 1846 // Generate several prefetch instructions. 1847 uint step_size = AllocatePrefetchStepSize; 1848 uint distance = AllocatePrefetchDistance; 1849 1850 // Next cache address. 1851 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top, 1852 _igvn.MakeConX(step_size + distance)); 1853 transform_later(cache_adr); 1854 cache_adr = new CastP2XNode(needgc_false, cache_adr); 1855 transform_later(cache_adr); 1856 // Address is aligned to execute prefetch to the beginning of cache line size 1857 // (it is important when BIS instruction is used on SPARC as prefetch). 1858 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); 1859 cache_adr = new AndXNode(cache_adr, mask); 1860 transform_later(cache_adr); 1861 cache_adr = new CastX2PNode(cache_adr); 1862 transform_later(cache_adr); 1863 1864 // Prefetch 1865 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); 1866 prefetch->set_req(0, needgc_false); 1867 transform_later(prefetch); 1868 contended_phi_rawmem = prefetch; 1869 Node *prefetch_adr; 1870 distance = step_size; 1871 for ( intx i = 1; i < lines; i++ ) { 1872 prefetch_adr = new AddPNode( cache_adr, cache_adr, 1873 _igvn.MakeConX(distance) ); 1874 transform_later(prefetch_adr); 1875 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); 1876 transform_later(prefetch); 1877 distance += step_size; 1878 contended_phi_rawmem = prefetch; 1879 } 1880 } else if( AllocatePrefetchStyle > 0 ) { 1881 // Insert a prefetch for each allocation only on the fast-path 1882 Node *prefetch_adr; 1883 Node *prefetch; 1884 // Generate several prefetch instructions. 1885 uint step_size = AllocatePrefetchStepSize; 1886 uint distance = AllocatePrefetchDistance; 1887 for ( intx i = 0; i < lines; i++ ) { 1888 prefetch_adr = new AddPNode( old_eden_top, new_eden_top, 1889 _igvn.MakeConX(distance) ); 1890 transform_later(prefetch_adr); 1891 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1892 // Do not let it float too high, since if eden_top == eden_end, 1893 // both might be null. 1894 if( i == 0 ) { // Set control for first prefetch, next follows it 1895 prefetch->init_req(0, needgc_false); 1896 } 1897 transform_later(prefetch); 1898 distance += step_size; 1899 i_o = prefetch; 1900 } 1901 } 1902 return i_o; 1903 } 1904 1905 1906 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 1907 expand_allocate_common(alloc, nullptr, 1908 OptoRuntime::new_instance_Type(), 1909 OptoRuntime::new_instance_Java(), nullptr); 1910 } 1911 1912 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 1913 Node* length = alloc->in(AllocateNode::ALength); 1914 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest); 1915 InitializeNode* init = alloc->initialization(); 1916 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1917 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr(); 1918 address slow_call_address; // Address of slow call 1919 if (init != nullptr && init->is_complete_with_arraycopy() && 1920 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) { 1921 // Don't zero type array during slow allocation in VM since 1922 // it will be initialized later by arraycopy in compiled code. 1923 slow_call_address = OptoRuntime::new_array_nozero_Java(); 1924 } else { 1925 slow_call_address = OptoRuntime::new_array_Java(); 1926 } 1927 expand_allocate_common(alloc, length, 1928 OptoRuntime::new_array_Type(), 1929 slow_call_address, valid_length_test); 1930 } 1931 1932 //-------------------mark_eliminated_box---------------------------------- 1933 // 1934 // During EA obj may point to several objects but after few ideal graph 1935 // transformations (CCP) it may point to only one non escaping object 1936 // (but still using phi), corresponding locks and unlocks will be marked 1937 // for elimination. Later obj could be replaced with a new node (new phi) 1938 // and which does not have escape information. And later after some graph 1939 // reshape other locks and unlocks (which were not marked for elimination 1940 // before) are connected to this new obj (phi) but they still will not be 1941 // marked for elimination since new obj has no escape information. 1942 // Mark all associated (same box and obj) lock and unlock nodes for 1943 // elimination if some of them marked already. 1944 void PhaseMacroExpand::mark_eliminated_box(Node* box, Node* obj) { 1945 BoxLockNode* oldbox = box->as_BoxLock(); 1946 if (oldbox->is_eliminated()) { 1947 return; // This BoxLock node was processed already. 1948 } 1949 assert(!oldbox->is_unbalanced(), "this should not be called for unbalanced region"); 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 oldbox->set_local(); // This verifies correct state of BoxLock 1958 _igvn.hash_delete(oldbox); 1959 oldbox->set_eliminated(); // This changes box's hash value 1960 _igvn.hash_insert(oldbox); 1961 1962 for (uint i = 0; i < oldbox->outcnt(); i++) { 1963 Node* u = oldbox->raw_out(i); 1964 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) { 1965 AbstractLockNode* alock = u->as_AbstractLock(); 1966 // Check lock's box since box could be referenced by Lock's debug info. 1967 if (alock->box_node() == oldbox) { 1968 // Mark eliminated all related locks and unlocks. 1969 #ifdef ASSERT 1970 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4"); 1971 #endif 1972 alock->set_non_esc_obj(); 1973 } 1974 } 1975 } 1976 return; 1977 } 1978 1979 // Create new "eliminated" BoxLock node and use it in monitor debug info 1980 // instead of oldbox for the same object. 1981 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 1982 1983 // Note: BoxLock node is marked eliminated only here and it is used 1984 // to indicate that all associated lock and unlock nodes are marked 1985 // for elimination. 1986 newbox->set_local(); // This verifies correct state of BoxLock 1987 newbox->set_eliminated(); 1988 transform_later(newbox); 1989 1990 // Replace old box node with new box for all users of the same object. 1991 for (uint i = 0; i < oldbox->outcnt();) { 1992 bool next_edge = true; 1993 1994 Node* u = oldbox->raw_out(i); 1995 if (u->is_AbstractLock()) { 1996 AbstractLockNode* alock = u->as_AbstractLock(); 1997 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) { 1998 // Replace Box and mark eliminated all related locks and unlocks. 1999 #ifdef ASSERT 2000 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5"); 2001 #endif 2002 alock->set_non_esc_obj(); 2003 _igvn.rehash_node_delayed(alock); 2004 alock->set_box_node(newbox); 2005 next_edge = false; 2006 } 2007 } 2008 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) { 2009 FastLockNode* flock = u->as_FastLock(); 2010 assert(flock->box_node() == oldbox, "sanity"); 2011 _igvn.rehash_node_delayed(flock); 2012 flock->set_box_node(newbox); 2013 next_edge = false; 2014 } 2015 2016 // Replace old box in monitor debug info. 2017 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 2018 SafePointNode* sfn = u->as_SafePoint(); 2019 JVMState* youngest_jvms = sfn->jvms(); 2020 int max_depth = youngest_jvms->depth(); 2021 for (int depth = 1; depth <= max_depth; depth++) { 2022 JVMState* jvms = youngest_jvms->of_depth(depth); 2023 int num_mon = jvms->nof_monitors(); 2024 // Loop over monitors 2025 for (int idx = 0; idx < num_mon; idx++) { 2026 Node* obj_node = sfn->monitor_obj(jvms, idx); 2027 Node* box_node = sfn->monitor_box(jvms, idx); 2028 if (box_node == oldbox && obj_node->eqv_uncast(obj)) { 2029 int j = jvms->monitor_box_offset(idx); 2030 _igvn.replace_input_of(u, j, newbox); 2031 next_edge = false; 2032 } 2033 } 2034 } 2035 } 2036 if (next_edge) i++; 2037 } 2038 } 2039 2040 //-----------------------mark_eliminated_locking_nodes----------------------- 2041 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { 2042 if (!alock->is_balanced()) { 2043 return; // Can't do any more elimination for this locking region 2044 } 2045 if (EliminateNestedLocks) { 2046 if (alock->is_nested()) { 2047 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity"); 2048 return; 2049 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened 2050 // Only Lock node has JVMState needed here. 2051 // Not that preceding claim is documented anywhere else. 2052 if (alock->jvms() != nullptr) { 2053 if (alock->as_Lock()->is_nested_lock_region()) { 2054 // Mark eliminated related nested locks and unlocks. 2055 Node* obj = alock->obj_node(); 2056 BoxLockNode* box_node = alock->box_node()->as_BoxLock(); 2057 assert(!box_node->is_eliminated(), "should not be marked yet"); 2058 // Note: BoxLock node is marked eliminated only here 2059 // and it is used to indicate that all associated lock 2060 // and unlock nodes are marked for elimination. 2061 box_node->set_eliminated(); // Box's hash is always NO_HASH here 2062 for (uint i = 0; i < box_node->outcnt(); i++) { 2063 Node* u = box_node->raw_out(i); 2064 if (u->is_AbstractLock()) { 2065 alock = u->as_AbstractLock(); 2066 if (alock->box_node() == box_node) { 2067 // Verify that this Box is referenced only by related locks. 2068 assert(alock->obj_node()->eqv_uncast(obj), ""); 2069 // Mark all related locks and unlocks. 2070 #ifdef ASSERT 2071 alock->log_lock_optimization(C, "eliminate_lock_set_nested"); 2072 #endif 2073 alock->set_nested(); 2074 } 2075 } 2076 } 2077 } else { 2078 #ifdef ASSERT 2079 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region"); 2080 if (C->log() != nullptr) 2081 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output 2082 #endif 2083 } 2084 } 2085 return; 2086 } 2087 // Process locks for non escaping object 2088 assert(alock->is_non_esc_obj(), ""); 2089 } // EliminateNestedLocks 2090 2091 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object 2092 // Look for all locks of this object and mark them and 2093 // corresponding BoxLock nodes as eliminated. 2094 Node* obj = alock->obj_node(); 2095 for (uint j = 0; j < obj->outcnt(); j++) { 2096 Node* o = obj->raw_out(j); 2097 if (o->is_AbstractLock() && 2098 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) { 2099 alock = o->as_AbstractLock(); 2100 Node* box = alock->box_node(); 2101 // Replace old box node with new eliminated box for all users 2102 // of the same object and mark related locks as eliminated. 2103 mark_eliminated_box(box, obj); 2104 } 2105 } 2106 } 2107 } 2108 2109 // we have determined that this lock/unlock can be eliminated, we simply 2110 // eliminate the node without expanding it. 2111 // 2112 // Note: The membar's associated with the lock/unlock are currently not 2113 // eliminated. This should be investigated as a future enhancement. 2114 // 2115 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 2116 2117 if (!alock->is_eliminated()) { 2118 return false; 2119 } 2120 #ifdef ASSERT 2121 if (!alock->is_coarsened()) { 2122 // Check that new "eliminated" BoxLock node is created. 2123 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 2124 assert(oldbox->is_eliminated(), "should be done already"); 2125 } 2126 #endif 2127 2128 alock->log_lock_optimization(C, "eliminate_lock"); 2129 2130 #ifndef PRODUCT 2131 if (PrintEliminateLocks) { 2132 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string()); 2133 } 2134 #endif 2135 2136 Node* mem = alock->in(TypeFunc::Memory); 2137 Node* ctrl = alock->in(TypeFunc::Control); 2138 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null"); 2139 2140 alock->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/); 2141 // There are 2 projections from the lock. The lock node will 2142 // be deleted when its last use is subsumed below. 2143 assert(alock->outcnt() == 2 && 2144 _callprojs.fallthrough_proj != nullptr && 2145 _callprojs.fallthrough_memproj != nullptr, 2146 "Unexpected projections from Lock/Unlock"); 2147 2148 Node* fallthroughproj = _callprojs.fallthrough_proj; 2149 Node* memproj_fallthrough = _callprojs.fallthrough_memproj; 2150 2151 // The memory projection from a lock/unlock is RawMem 2152 // The input to a Lock is merged memory, so extract its RawMem input 2153 // (unless the MergeMem has been optimized away.) 2154 if (alock->is_Lock()) { 2155 // Search for MemBarAcquireLock node and delete it also. 2156 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 2157 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, ""); 2158 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 2159 Node* memproj = membar->proj_out(TypeFunc::Memory); 2160 _igvn.replace_node(ctrlproj, fallthroughproj); 2161 _igvn.replace_node(memproj, memproj_fallthrough); 2162 2163 // Delete FastLock node also if this Lock node is unique user 2164 // (a loop peeling may clone a Lock node). 2165 Node* flock = alock->as_Lock()->fastlock_node(); 2166 if (flock->outcnt() == 1) { 2167 assert(flock->unique_out() == alock, "sanity"); 2168 _igvn.replace_node(flock, top()); 2169 } 2170 } 2171 2172 // Search for MemBarReleaseLock node and delete it also. 2173 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) { 2174 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 2175 assert(membar->Opcode() == Op_MemBarReleaseLock && 2176 mem->is_Proj() && membar == mem->in(0), ""); 2177 _igvn.replace_node(fallthroughproj, ctrl); 2178 _igvn.replace_node(memproj_fallthrough, mem); 2179 fallthroughproj = ctrl; 2180 memproj_fallthrough = mem; 2181 ctrl = membar->in(TypeFunc::Control); 2182 mem = membar->in(TypeFunc::Memory); 2183 } 2184 2185 _igvn.replace_node(fallthroughproj, ctrl); 2186 _igvn.replace_node(memproj_fallthrough, mem); 2187 return true; 2188 } 2189 2190 2191 //------------------------------expand_lock_node---------------------- 2192 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 2193 2194 Node* ctrl = lock->in(TypeFunc::Control); 2195 Node* mem = lock->in(TypeFunc::Memory); 2196 Node* obj = lock->obj_node(); 2197 Node* box = lock->box_node(); 2198 Node* flock = lock->fastlock_node(); 2199 2200 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2201 2202 // Make the merge point 2203 Node *region; 2204 Node *mem_phi; 2205 Node *slow_path; 2206 2207 region = new RegionNode(3); 2208 // create a Phi for the memory state 2209 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2210 2211 // Optimize test; set region slot 2 2212 slow_path = opt_bits_test(ctrl, region, 2, flock); 2213 mem_phi->init_req(2, mem); 2214 2215 // Make slow path call 2216 CallNode* call = make_slow_call(lock, OptoRuntime::complete_monitor_enter_Type(), 2217 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path, 2218 obj, box, nullptr); 2219 2220 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/); 2221 2222 // Slow path can only throw asynchronous exceptions, which are always 2223 // de-opted. So the compiler thinks the slow-call can never throw an 2224 // exception. If it DOES throw an exception we would need the debug 2225 // info removed first (since if it throws there is no monitor). 2226 assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr && 2227 _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock"); 2228 2229 // Capture slow path 2230 // disconnect fall-through projection from call and create a new one 2231 // hook up users of fall-through projection to region 2232 Node *slow_ctrl = _callprojs.fallthrough_proj->clone(); 2233 transform_later(slow_ctrl); 2234 _igvn.hash_delete(_callprojs.fallthrough_proj); 2235 _callprojs.fallthrough_proj->disconnect_inputs(C); 2236 region->init_req(1, slow_ctrl); 2237 // region inputs are now complete 2238 transform_later(region); 2239 _igvn.replace_node(_callprojs.fallthrough_proj, region); 2240 2241 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory)); 2242 2243 mem_phi->init_req(1, memproj); 2244 2245 transform_later(mem_phi); 2246 2247 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi); 2248 } 2249 2250 //------------------------------expand_unlock_node---------------------- 2251 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 2252 2253 Node* ctrl = unlock->in(TypeFunc::Control); 2254 Node* mem = unlock->in(TypeFunc::Memory); 2255 Node* obj = unlock->obj_node(); 2256 Node* box = unlock->box_node(); 2257 2258 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2259 2260 // No need for a null check on unlock 2261 2262 // Make the merge point 2263 Node *region; 2264 Node *mem_phi; 2265 2266 region = new RegionNode(3); 2267 // create a Phi for the memory state 2268 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2269 2270 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box ); 2271 funlock = transform_later( funlock )->as_FastUnlock(); 2272 // Optimize test; set region slot 2 2273 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock); 2274 Node *thread = transform_later(new ThreadLocalNode()); 2275 2276 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), 2277 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), 2278 "complete_monitor_unlocking_C", slow_path, obj, box, thread); 2279 2280 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/); 2281 assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr && 2282 _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock"); 2283 2284 // No exceptions for unlocking 2285 // Capture slow path 2286 // disconnect fall-through projection from call and create a new one 2287 // hook up users of fall-through projection to region 2288 Node *slow_ctrl = _callprojs.fallthrough_proj->clone(); 2289 transform_later(slow_ctrl); 2290 _igvn.hash_delete(_callprojs.fallthrough_proj); 2291 _callprojs.fallthrough_proj->disconnect_inputs(C); 2292 region->init_req(1, slow_ctrl); 2293 // region inputs are now complete 2294 transform_later(region); 2295 _igvn.replace_node(_callprojs.fallthrough_proj, region); 2296 2297 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) ); 2298 mem_phi->init_req(1, memproj ); 2299 mem_phi->init_req(2, mem); 2300 transform_later(mem_phi); 2301 2302 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi); 2303 } 2304 2305 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) { 2306 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned"); 2307 Node* bol = check->unique_out(); 2308 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass); 2309 Node* superklass = check->in(SubTypeCheckNode::SuperKlass); 2310 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node"); 2311 2312 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) { 2313 Node* iff = bol->last_out(i); 2314 assert(iff->is_If(), "where's the if?"); 2315 2316 if (iff->in(0)->is_top()) { 2317 _igvn.replace_input_of(iff, 1, C->top()); 2318 continue; 2319 } 2320 2321 Node* iftrue = iff->as_If()->proj_out(1); 2322 Node* iffalse = iff->as_If()->proj_out(0); 2323 Node* ctrl = iff->in(0); 2324 2325 Node* subklass = nullptr; 2326 if (_igvn.type(obj_or_subklass)->isa_klassptr()) { 2327 subklass = obj_or_subklass; 2328 } else { 2329 Node* k_adr = basic_plus_adr(obj_or_subklass, Type::klass_offset()); 2330 subklass = _igvn.transform(LoadKlassNode::make(_igvn, C->immutable_memory(), k_adr, TypeInstPtr::KLASS)); 2331 } 2332 2333 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci()); 2334 2335 _igvn.replace_input_of(iff, 0, C->top()); 2336 _igvn.replace_node(iftrue, not_subtype_ctrl); 2337 _igvn.replace_node(iffalse, ctrl); 2338 } 2339 _igvn.replace_node(check, C->top()); 2340 } 2341 2342 // Perform refining of strip mined loop nodes in the macro nodes list. 2343 void PhaseMacroExpand::refine_strip_mined_loop_macro_nodes() { 2344 for (int i = C->macro_count(); i > 0; i--) { 2345 Node* n = C->macro_node(i - 1); 2346 if (n->is_OuterStripMinedLoop()) { 2347 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn); 2348 } 2349 } 2350 } 2351 2352 //---------------------------eliminate_macro_nodes---------------------- 2353 // Eliminate scalar replaced allocations and associated locks. 2354 void PhaseMacroExpand::eliminate_macro_nodes() { 2355 if (C->macro_count() == 0) 2356 return; 2357 2358 if (StressMacroElimination) { 2359 C->shuffle_macro_nodes(); 2360 } 2361 NOT_PRODUCT(int membar_before = count_MemBar(C);) 2362 2363 // Before elimination may re-mark (change to Nested or NonEscObj) 2364 // all associated (same box and obj) lock and unlock nodes. 2365 int cnt = C->macro_count(); 2366 for (int i=0; i < cnt; i++) { 2367 Node *n = C->macro_node(i); 2368 if (n->is_AbstractLock()) { // Lock and Unlock nodes 2369 mark_eliminated_locking_nodes(n->as_AbstractLock()); 2370 } 2371 } 2372 // Re-marking may break consistency of Coarsened locks. 2373 if (!C->coarsened_locks_consistent()) { 2374 return; // recompile without Coarsened locks if broken 2375 } else { 2376 // After coarsened locks are eliminated locking regions 2377 // become unbalanced. We should not execute any more 2378 // locks elimination optimizations on them. 2379 C->mark_unbalanced_boxes(); 2380 } 2381 2382 // First, attempt to eliminate locks 2383 bool progress = true; 2384 while (progress) { 2385 progress = false; 2386 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once 2387 Node* n = C->macro_node(i - 1); 2388 bool success = false; 2389 DEBUG_ONLY(int old_macro_count = C->macro_count();) 2390 if (n->is_AbstractLock()) { 2391 success = eliminate_locking_node(n->as_AbstractLock()); 2392 #ifndef PRODUCT 2393 if (success && PrintOptoStatistics) { 2394 Atomic::inc(&PhaseMacroExpand::_monitor_objects_removed_counter); 2395 } 2396 #endif 2397 } 2398 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2399 progress = progress || success; 2400 if (success) { 2401 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n); 2402 } 2403 } 2404 } 2405 // Next, attempt to eliminate allocations 2406 progress = true; 2407 while (progress) { 2408 progress = false; 2409 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once 2410 Node* n = C->macro_node(i - 1); 2411 bool success = false; 2412 DEBUG_ONLY(int old_macro_count = C->macro_count();) 2413 switch (n->class_id()) { 2414 case Node::Class_Allocate: 2415 case Node::Class_AllocateArray: 2416 success = eliminate_allocate_node(n->as_Allocate()); 2417 #ifndef PRODUCT 2418 if (success && PrintOptoStatistics) { 2419 Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter); 2420 } 2421 #endif 2422 break; 2423 case Node::Class_CallStaticJava: 2424 success = eliminate_boxing_node(n->as_CallStaticJava()); 2425 break; 2426 case Node::Class_Lock: 2427 case Node::Class_Unlock: 2428 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 2429 break; 2430 case Node::Class_ArrayCopy: 2431 break; 2432 case Node::Class_OuterStripMinedLoop: 2433 break; 2434 case Node::Class_SubTypeCheck: 2435 break; 2436 case Node::Class_Opaque1: 2437 break; 2438 default: 2439 assert(n->Opcode() == Op_LoopLimit || 2440 n->Opcode() == Op_ModD || 2441 n->Opcode() == Op_ModF || 2442 n->is_OpaqueNotNull() || 2443 n->is_OpaqueInitializedAssertionPredicate() || 2444 n->Opcode() == Op_MaxL || 2445 n->Opcode() == Op_MinL || 2446 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n), 2447 "unknown node type in macro list"); 2448 } 2449 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2450 progress = progress || success; 2451 if (success) { 2452 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n); 2453 } 2454 } 2455 } 2456 #ifndef PRODUCT 2457 if (PrintOptoStatistics) { 2458 int membar_after = count_MemBar(C); 2459 Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after); 2460 } 2461 #endif 2462 } 2463 2464 void PhaseMacroExpand::eliminate_opaque_looplimit_macro_nodes() { 2465 if (C->macro_count() == 0) { 2466 return; 2467 } 2468 refine_strip_mined_loop_macro_nodes(); 2469 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. 2470 bool progress = true; 2471 while (progress) { 2472 progress = false; 2473 for (int i = C->macro_count(); i > 0; i--) { 2474 Node* n = C->macro_node(i-1); 2475 bool success = false; 2476 DEBUG_ONLY(int old_macro_count = C->macro_count();) 2477 if (n->Opcode() == Op_LoopLimit) { 2478 // Remove it from macro list and put on IGVN worklist to optimize. 2479 C->remove_macro_node(n); 2480 _igvn._worklist.push(n); 2481 success = true; 2482 } else if (n->Opcode() == Op_CallStaticJava) { 2483 // Remove it from macro list and put on IGVN worklist to optimize. 2484 C->remove_macro_node(n); 2485 _igvn._worklist.push(n); 2486 success = true; 2487 } else if (n->is_Opaque1()) { 2488 _igvn.replace_node(n, n->in(1)); 2489 success = true; 2490 } else if (n->is_OpaqueNotNull()) { 2491 // Tests with OpaqueNotNull nodes are implicitly known to be true. Replace the node with true. In debug builds, 2492 // we leave the test in the graph to have an additional sanity check at runtime. If the test fails (i.e. a bug), 2493 // we will execute a Halt node. 2494 #ifdef ASSERT 2495 _igvn.replace_node(n, n->in(1)); 2496 #else 2497 _igvn.replace_node(n, _igvn.intcon(1)); 2498 #endif 2499 success = true; 2500 } else if (n->is_OpaqueInitializedAssertionPredicate()) { 2501 // Initialized Assertion Predicates must always evaluate to true. Therefore, we get rid of them in product 2502 // builds as they are useless. In debug builds we keep them as additional verification code. Even though 2503 // loop opts are already over, we want to keep Initialized Assertion Predicates alive as long as possible to 2504 // enable folding of dead control paths within which cast nodes become top after due to impossible types - 2505 // even after loop opts are over. Therefore, we delay the removal of these opaque nodes until now. 2506 #ifdef ASSERT 2507 _igvn.replace_node(n, n->in(1)); 2508 #else 2509 _igvn.replace_node(n, _igvn.intcon(1)); 2510 #endif // ASSERT 2511 } else if (n->Opcode() == Op_OuterStripMinedLoop) { 2512 C->remove_macro_node(n); 2513 success = true; 2514 } else if (n->Opcode() == Op_MaxL) { 2515 // Since MaxL and MinL are not implemented in the backend, we expand them to 2516 // a CMoveL construct now. At least until here, the type could be computed 2517 // precisely. CMoveL is not so smart, but we can give it at least the best 2518 // type we know abouot n now. 2519 Node* repl = MaxNode::signed_max(n->in(1), n->in(2), _igvn.type(n), _igvn); 2520 _igvn.replace_node(n, repl); 2521 success = true; 2522 } else if (n->Opcode() == Op_MinL) { 2523 Node* repl = MaxNode::signed_min(n->in(1), n->in(2), _igvn.type(n), _igvn); 2524 _igvn.replace_node(n, repl); 2525 success = true; 2526 } 2527 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list"); 2528 progress = progress || success; 2529 if (success) { 2530 C->print_method(PHASE_AFTER_MACRO_ELIMINATION_STEP, 5, n); 2531 } 2532 } 2533 } 2534 } 2535 2536 //------------------------------expand_macro_nodes---------------------- 2537 // Returns true if a failure occurred. 2538 bool PhaseMacroExpand::expand_macro_nodes() { 2539 if (StressMacroExpansion) { 2540 C->shuffle_macro_nodes(); 2541 } 2542 2543 // Clean up the graph so we're less likely to hit the maximum node 2544 // limit 2545 _igvn.set_delay_transform(false); 2546 _igvn.optimize(); 2547 if (C->failing()) return true; 2548 _igvn.set_delay_transform(true); 2549 2550 2551 // Because we run IGVN after each expansion, some macro nodes may go 2552 // dead and be removed from the list as we iterate over it. Move 2553 // Allocate nodes (processed in a second pass) at the beginning of 2554 // the list and then iterate from the last element of the list until 2555 // an Allocate node is seen. This is robust to random deletion in 2556 // the list due to nodes going dead. 2557 C->sort_macro_nodes(); 2558 2559 // expand arraycopy "macro" nodes first 2560 // For ReduceBulkZeroing, we must first process all arraycopy nodes 2561 // before the allocate nodes are expanded. 2562 while (C->macro_count() > 0) { 2563 int macro_count = C->macro_count(); 2564 Node * n = C->macro_node(macro_count-1); 2565 assert(n->is_macro(), "only macro nodes expected here"); 2566 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) { 2567 // node is unreachable, so don't try to expand it 2568 C->remove_macro_node(n); 2569 continue; 2570 } 2571 if (n->is_Allocate()) { 2572 break; 2573 } 2574 // Make sure expansion will not cause node limit to be exceeded. 2575 // Worst case is a macro node gets expanded into about 200 nodes. 2576 // Allow 50% more for optimization. 2577 if (C->check_node_count(300, "out of nodes before macro expansion")) { 2578 return true; 2579 } 2580 2581 DEBUG_ONLY(int old_macro_count = C->macro_count();) 2582 switch (n->class_id()) { 2583 case Node::Class_Lock: 2584 expand_lock_node(n->as_Lock()); 2585 break; 2586 case Node::Class_Unlock: 2587 expand_unlock_node(n->as_Unlock()); 2588 break; 2589 case Node::Class_ArrayCopy: 2590 expand_arraycopy_node(n->as_ArrayCopy()); 2591 break; 2592 case Node::Class_SubTypeCheck: 2593 expand_subtypecheck_node(n->as_SubTypeCheck()); 2594 break; 2595 default: 2596 switch (n->Opcode()) { 2597 case Op_ModD: 2598 case Op_ModF: { 2599 bool is_drem = n->Opcode() == Op_ModD; 2600 CallNode* mod_macro = n->as_Call(); 2601 CallNode* call = new CallLeafNode(mod_macro->tf(), 2602 is_drem ? CAST_FROM_FN_PTR(address, SharedRuntime::drem) 2603 : CAST_FROM_FN_PTR(address, SharedRuntime::frem), 2604 is_drem ? "drem" : "frem", TypeRawPtr::BOTTOM); 2605 call->init_req(TypeFunc::Control, mod_macro->in(TypeFunc::Control)); 2606 call->init_req(TypeFunc::I_O, mod_macro->in(TypeFunc::I_O)); 2607 call->init_req(TypeFunc::Memory, mod_macro->in(TypeFunc::Memory)); 2608 call->init_req(TypeFunc::ReturnAdr, mod_macro->in(TypeFunc::ReturnAdr)); 2609 call->init_req(TypeFunc::FramePtr, mod_macro->in(TypeFunc::FramePtr)); 2610 for (unsigned int i = 0; i < mod_macro->tf()->domain()->cnt() - TypeFunc::Parms; i++) { 2611 call->init_req(TypeFunc::Parms + i, mod_macro->in(TypeFunc::Parms + i)); 2612 } 2613 _igvn.replace_node(mod_macro, call); 2614 transform_later(call); 2615 break; 2616 } 2617 default: 2618 assert(false, "unknown node type in macro list"); 2619 } 2620 } 2621 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list"); 2622 if (C->failing()) return true; 2623 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n); 2624 2625 // Clean up the graph so we're less likely to hit the maximum node 2626 // limit 2627 _igvn.set_delay_transform(false); 2628 _igvn.optimize(); 2629 if (C->failing()) return true; 2630 _igvn.set_delay_transform(true); 2631 } 2632 2633 // All nodes except Allocate nodes are expanded now. There could be 2634 // new optimization opportunities (such as folding newly created 2635 // load from a just allocated object). Run IGVN. 2636 2637 // expand "macro" nodes 2638 // nodes are removed from the macro list as they are processed 2639 while (C->macro_count() > 0) { 2640 int macro_count = C->macro_count(); 2641 Node * n = C->macro_node(macro_count-1); 2642 assert(n->is_macro(), "only macro nodes expected here"); 2643 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) { 2644 // node is unreachable, so don't try to expand it 2645 C->remove_macro_node(n); 2646 continue; 2647 } 2648 // Make sure expansion will not cause node limit to be exceeded. 2649 // Worst case is a macro node gets expanded into about 200 nodes. 2650 // Allow 50% more for optimization. 2651 if (C->check_node_count(300, "out of nodes before macro expansion")) { 2652 return true; 2653 } 2654 switch (n->class_id()) { 2655 case Node::Class_Allocate: 2656 expand_allocate(n->as_Allocate()); 2657 break; 2658 case Node::Class_AllocateArray: 2659 expand_allocate_array(n->as_AllocateArray()); 2660 break; 2661 default: 2662 assert(false, "unknown node type in macro list"); 2663 } 2664 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2665 if (C->failing()) return true; 2666 C->print_method(PHASE_AFTER_MACRO_EXPANSION_STEP, 5, n); 2667 2668 // Clean up the graph so we're less likely to hit the maximum node 2669 // limit 2670 _igvn.set_delay_transform(false); 2671 _igvn.optimize(); 2672 if (C->failing()) return true; 2673 _igvn.set_delay_transform(true); 2674 } 2675 2676 _igvn.set_delay_transform(false); 2677 return false; 2678 } 2679 2680 #ifndef PRODUCT 2681 int PhaseMacroExpand::_objs_scalar_replaced_counter = 0; 2682 int PhaseMacroExpand::_monitor_objects_removed_counter = 0; 2683 int PhaseMacroExpand::_GC_barriers_removed_counter = 0; 2684 int PhaseMacroExpand::_memory_barriers_removed_counter = 0; 2685 2686 void PhaseMacroExpand::print_statistics() { 2687 tty->print("Objects scalar replaced = %d, ", Atomic::load(&_objs_scalar_replaced_counter)); 2688 tty->print("Monitor objects removed = %d, ", Atomic::load(&_monitor_objects_removed_counter)); 2689 tty->print("GC barriers removed = %d, ", Atomic::load(&_GC_barriers_removed_counter)); 2690 tty->print_cr("Memory barriers removed = %d", Atomic::load(&_memory_barriers_removed_counter)); 2691 } 2692 2693 int PhaseMacroExpand::count_MemBar(Compile *C) { 2694 if (!PrintOptoStatistics) { 2695 return 0; 2696 } 2697 Unique_Node_List ideal_nodes; 2698 int total = 0; 2699 ideal_nodes.map(C->live_nodes(), nullptr); 2700 ideal_nodes.push(C->root()); 2701 for (uint next = 0; next < ideal_nodes.size(); ++next) { 2702 Node* n = ideal_nodes.at(next); 2703 if (n->is_MemBar()) { 2704 total++; 2705 } 2706 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2707 Node* m = n->fast_out(i); 2708 ideal_nodes.push(m); 2709 } 2710 } 2711 return total; 2712 } 2713 #endif