1 /* 2 * Copyright (c) 2005, 2023, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "ci/ciFlatArrayKlass.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "gc/shared/collectedHeap.inline.hpp" 29 #include "gc/shared/tlab_globals.hpp" 30 #include "libadt/vectset.hpp" 31 #include "memory/universe.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/arraycopynode.hpp" 34 #include "opto/callnode.hpp" 35 #include "opto/castnode.hpp" 36 #include "opto/cfgnode.hpp" 37 #include "opto/compile.hpp" 38 #include "opto/convertnode.hpp" 39 #include "opto/graphKit.hpp" 40 #include "opto/inlinetypenode.hpp" 41 #include "opto/intrinsicnode.hpp" 42 #include "opto/locknode.hpp" 43 #include "opto/loopnode.hpp" 44 #include "opto/macro.hpp" 45 #include "opto/memnode.hpp" 46 #include "opto/narrowptrnode.hpp" 47 #include "opto/node.hpp" 48 #include "opto/opaquenode.hpp" 49 #include "opto/phaseX.hpp" 50 #include "opto/rootnode.hpp" 51 #include "opto/runtime.hpp" 52 #include "opto/subnode.hpp" 53 #include "opto/subtypenode.hpp" 54 #include "opto/type.hpp" 55 #include "prims/jvmtiExport.hpp" 56 #include "runtime/continuation.hpp" 57 #include "runtime/sharedRuntime.hpp" 58 #include "runtime/stubRoutines.hpp" 59 #include "utilities/macros.hpp" 60 #include "utilities/powerOfTwo.hpp" 61 #if INCLUDE_G1GC 62 #include "gc/g1/g1ThreadLocalData.hpp" 63 #endif // INCLUDE_G1GC 64 65 66 // 67 // Replace any references to "oldref" in inputs to "use" with "newref". 68 // Returns the number of replacements made. 69 // 70 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { 71 int nreplacements = 0; 72 uint req = use->req(); 73 for (uint j = 0; j < use->len(); j++) { 74 Node *uin = use->in(j); 75 if (uin == oldref) { 76 if (j < req) 77 use->set_req(j, newref); 78 else 79 use->set_prec(j, newref); 80 nreplacements++; 81 } else if (j >= req && uin == nullptr) { 82 break; 83 } 84 } 85 return nreplacements; 86 } 87 88 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) { 89 Node* cmp; 90 if (mask != 0) { 91 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask))); 92 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits))); 93 } else { 94 cmp = word; 95 } 96 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne)); 97 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); 98 transform_later(iff); 99 100 // Fast path taken. 101 Node *fast_taken = transform_later(new IfFalseNode(iff)); 102 103 // Fast path not-taken, i.e. slow path 104 Node *slow_taken = transform_later(new IfTrueNode(iff)); 105 106 if (return_fast_path) { 107 region->init_req(edge, slow_taken); // Capture slow-control 108 return fast_taken; 109 } else { 110 region->init_req(edge, fast_taken); // Capture fast-control 111 return slow_taken; 112 } 113 } 114 115 //--------------------copy_predefined_input_for_runtime_call-------------------- 116 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { 117 // Set fixed predefined input arguments 118 call->init_req( TypeFunc::Control, ctrl ); 119 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); 120 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? 121 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); 122 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); 123 } 124 125 //------------------------------make_slow_call--------------------------------- 126 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, 127 address slow_call, const char* leaf_name, Node* slow_path, 128 Node* parm0, Node* parm1, Node* parm2) { 129 130 // Slow-path call 131 CallNode *call = leaf_name 132 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) 133 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM ); 134 135 // Slow path call has no side-effects, uses few values 136 copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); 137 if (parm0 != nullptr) call->init_req(TypeFunc::Parms+0, parm0); 138 if (parm1 != nullptr) call->init_req(TypeFunc::Parms+1, parm1); 139 if (parm2 != nullptr) call->init_req(TypeFunc::Parms+2, parm2); 140 call->copy_call_debug_info(&_igvn, oldcall); 141 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 142 _igvn.replace_node(oldcall, call); 143 transform_later(call); 144 145 return call; 146 } 147 148 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) { 149 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2(); 150 bs->eliminate_gc_barrier(&_igvn, p2x); 151 #ifndef PRODUCT 152 if (PrintOptoStatistics) { 153 Atomic::inc(&PhaseMacroExpand::_GC_barriers_removed_counter); 154 } 155 #endif 156 } 157 158 // Search for a memory operation for the specified memory slice. 159 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) { 160 Node *orig_mem = mem; 161 Node *alloc_mem = alloc->in(TypeFunc::Memory); 162 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr(); 163 while (true) { 164 if (mem == alloc_mem || mem == start_mem ) { 165 return mem; // hit one of our sentinels 166 } else if (mem->is_MergeMem()) { 167 mem = mem->as_MergeMem()->memory_at(alias_idx); 168 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { 169 Node *in = mem->in(0); 170 // we can safely skip over safepoints, calls, locks and membars because we 171 // already know that the object is safe to eliminate. 172 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { 173 return in; 174 } else if (in->is_Call()) { 175 CallNode *call = in->as_Call(); 176 if (call->may_modify(tinst, phase)) { 177 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape"); 178 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) { 179 return in; 180 } 181 } 182 mem = in->in(TypeFunc::Memory); 183 } else if (in->is_MemBar()) { 184 ArrayCopyNode* ac = nullptr; 185 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) { 186 if (ac != nullptr) { 187 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone"); 188 return ac; 189 } 190 } 191 mem = in->in(TypeFunc::Memory); 192 } else { 193 #ifdef ASSERT 194 in->dump(); 195 mem->dump(); 196 assert(false, "unexpected projection"); 197 #endif 198 } 199 } else if (mem->is_Store()) { 200 const TypePtr* atype = mem->as_Store()->adr_type(); 201 int adr_idx = phase->C->get_alias_index(atype); 202 if (adr_idx == alias_idx) { 203 assert(atype->isa_oopptr(), "address type must be oopptr"); 204 int adr_offset = atype->flat_offset(); 205 uint adr_iid = atype->is_oopptr()->instance_id(); 206 // Array elements references have the same alias_idx 207 // but different offset and different instance_id. 208 if (adr_offset == offset && adr_iid == alloc->_idx) { 209 return mem; 210 } 211 } else { 212 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw"); 213 } 214 mem = mem->in(MemNode::Memory); 215 } else if (mem->is_ClearArray()) { 216 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) { 217 // Can not bypass initialization of the instance 218 // we are looking. 219 debug_only(intptr_t offset;) 220 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity"); 221 InitializeNode* init = alloc->as_Allocate()->initialization(); 222 // We are looking for stored value, return Initialize node 223 // or memory edge from Allocate node. 224 if (init != nullptr) { 225 return init; 226 } else { 227 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers). 228 } 229 } 230 // Otherwise skip it (the call updated 'mem' value). 231 } else if (mem->Opcode() == Op_SCMemProj) { 232 mem = mem->in(0); 233 Node* adr = nullptr; 234 if (mem->is_LoadStore()) { 235 adr = mem->in(MemNode::Address); 236 } else { 237 assert(mem->Opcode() == Op_EncodeISOArray || 238 mem->Opcode() == Op_StrCompressedCopy, "sanity"); 239 adr = mem->in(3); // Destination array 240 } 241 const TypePtr* atype = adr->bottom_type()->is_ptr(); 242 int adr_idx = phase->C->get_alias_index(atype); 243 if (adr_idx == alias_idx) { 244 DEBUG_ONLY(mem->dump();) 245 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 246 return nullptr; 247 } 248 mem = mem->in(MemNode::Memory); 249 } else if (mem->Opcode() == Op_StrInflatedCopy) { 250 Node* adr = mem->in(3); // Destination array 251 const TypePtr* atype = adr->bottom_type()->is_ptr(); 252 int adr_idx = phase->C->get_alias_index(atype); 253 if (adr_idx == alias_idx) { 254 DEBUG_ONLY(mem->dump();) 255 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field"); 256 return nullptr; 257 } 258 mem = mem->in(MemNode::Memory); 259 } else { 260 return mem; 261 } 262 assert(mem != orig_mem, "dead memory loop"); 263 } 264 } 265 266 // Generate loads from source of the arraycopy for fields of 267 // destination needed at a deoptimization point 268 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) { 269 BasicType bt = ft; 270 const Type *type = ftype; 271 if (ft == T_NARROWOOP) { 272 bt = T_OBJECT; 273 type = ftype->make_oopptr(); 274 } 275 Node* res = nullptr; 276 if (ac->is_clonebasic()) { 277 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination"); 278 Node* base = ac->in(ArrayCopyNode::Src); 279 Node* adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(offset))); 280 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset); 281 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem(); 282 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 283 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt); 284 } else { 285 if (ac->modifies(offset, offset, &_igvn, true)) { 286 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result"); 287 uint shift = exact_log2(type2aelembytes(bt)); 288 Node* src_pos = ac->in(ArrayCopyNode::SrcPos); 289 Node* dest_pos = ac->in(ArrayCopyNode::DestPos); 290 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int(); 291 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int(); 292 293 Node* adr = nullptr; 294 Node* base = ac->in(ArrayCopyNode::Src); 295 const TypeAryPtr* adr_type = _igvn.type(base)->is_aryptr(); 296 if (adr_type->is_flat()) { 297 shift = adr_type->flat_log_elem_size(); 298 } 299 if (src_pos_t->is_con() && dest_pos_t->is_con()) { 300 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset; 301 adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(off))); 302 adr_type = _igvn.type(adr)->is_aryptr(); 303 assert(adr_type == _igvn.type(base)->is_aryptr()->add_field_offset_and_offset(off), "incorrect address type"); 304 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) { 305 // Don't emit a new load from src if src == dst but try to get the value from memory instead 306 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type, alloc); 307 } 308 } else { 309 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) { 310 // Non constant offset in the array: we can't statically 311 // determine the value 312 return nullptr; 313 } 314 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 315 #ifdef _LP64 316 diff = _igvn.transform(new ConvI2LNode(diff)); 317 #endif 318 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift))); 319 320 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff)); 321 adr = _igvn.transform(new AddPNode(base, base, off)); 322 // In the case of a flat inline type array, each field has its 323 // own slice so we need to extract the field being accessed from 324 // the address computation 325 adr_type = adr_type->add_field_offset_and_offset(offset)->add_offset(Type::OffsetBot)->is_aryptr(); 326 adr = _igvn.transform(new CastPPNode(adr, adr_type)); 327 } 328 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem(); 329 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 330 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt); 331 } 332 } 333 if (res != nullptr) { 334 if (ftype->isa_narrowoop()) { 335 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes 336 assert(res->isa_DecodeN(), "should be narrow oop"); 337 res = _igvn.transform(new EncodePNode(res, ftype)); 338 } 339 return res; 340 } 341 return nullptr; 342 } 343 344 // 345 // Given a Memory Phi, compute a value Phi containing the values from stores 346 // on the input paths. 347 // Note: this function is recursive, its depth is limited by the "level" argument 348 // Returns the computed Phi, or null if it cannot compute it. 349 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) { 350 assert(mem->is_Phi(), "sanity"); 351 int alias_idx = C->get_alias_index(adr_t); 352 int offset = adr_t->flat_offset(); 353 int instance_id = adr_t->instance_id(); 354 355 // Check if an appropriate value phi already exists. 356 Node* region = mem->in(0); 357 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { 358 Node* phi = region->fast_out(k); 359 if (phi->is_Phi() && phi != mem && 360 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) { 361 return phi; 362 } 363 } 364 // Check if an appropriate new value phi already exists. 365 Node* new_phi = value_phis->find(mem->_idx); 366 if (new_phi != nullptr) 367 return new_phi; 368 369 if (level <= 0) { 370 return nullptr; // Give up: phi tree too deep 371 } 372 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 373 Node *alloc_mem = alloc->in(TypeFunc::Memory); 374 375 uint length = mem->req(); 376 GrowableArray <Node *> values(length, length, nullptr); 377 378 // create a new Phi for the value 379 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset); 380 transform_later(phi); 381 value_phis->push(phi, mem->_idx); 382 383 for (uint j = 1; j < length; j++) { 384 Node *in = mem->in(j); 385 if (in == nullptr || in->is_top()) { 386 values.at_put(j, in); 387 } else { 388 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); 389 if (val == start_mem || val == alloc_mem) { 390 // hit a sentinel, return appropriate 0 value 391 Node* default_value = alloc->in(AllocateNode::DefaultValue); 392 if (default_value != nullptr) { 393 values.at_put(j, default_value); 394 } else { 395 assert(alloc->in(AllocateNode::RawDefaultValue) == nullptr, "default value may not be null"); 396 values.at_put(j, _igvn.zerocon(ft)); 397 } 398 continue; 399 } 400 if (val->is_Initialize()) { 401 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 402 } 403 if (val == nullptr) { 404 return nullptr; // can't find a value on this path 405 } 406 if (val == mem) { 407 values.at_put(j, mem); 408 } else if (val->is_Store()) { 409 Node* n = val->in(MemNode::ValueIn); 410 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 411 n = bs->step_over_gc_barrier(n); 412 if (is_subword_type(ft)) { 413 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true); 414 } 415 values.at_put(j, n); 416 } else if(val->is_Proj() && val->in(0) == alloc) { 417 Node* default_value = alloc->in(AllocateNode::DefaultValue); 418 if (default_value != nullptr) { 419 values.at_put(j, default_value); 420 } else { 421 assert(alloc->in(AllocateNode::RawDefaultValue) == nullptr, "default value may not be null"); 422 values.at_put(j, _igvn.zerocon(ft)); 423 } 424 } else if (val->is_Phi()) { 425 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); 426 if (val == nullptr) { 427 return nullptr; 428 } 429 values.at_put(j, val); 430 } else if (val->Opcode() == Op_SCMemProj) { 431 assert(val->in(0)->is_LoadStore() || 432 val->in(0)->Opcode() == Op_EncodeISOArray || 433 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity"); 434 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 435 return nullptr; 436 } else if (val->is_ArrayCopy()) { 437 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc); 438 if (res == nullptr) { 439 return nullptr; 440 } 441 values.at_put(j, res); 442 } else { 443 DEBUG_ONLY( val->dump(); ) 444 assert(false, "unknown node on this path"); 445 return nullptr; // unknown node on this path 446 } 447 } 448 } 449 // Set Phi's inputs 450 for (uint j = 1; j < length; j++) { 451 if (values.at(j) == mem) { 452 phi->init_req(j, phi); 453 } else { 454 phi->init_req(j, values.at(j)); 455 } 456 } 457 return phi; 458 } 459 460 // Search the last value stored into the object's field. 461 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) { 462 assert(adr_t->is_known_instance_field(), "instance required"); 463 int instance_id = adr_t->instance_id(); 464 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 465 466 int alias_idx = C->get_alias_index(adr_t); 467 int offset = adr_t->flat_offset(); 468 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 469 Node *alloc_mem = alloc->in(TypeFunc::Memory); 470 VectorSet visited; 471 472 bool done = sfpt_mem == alloc_mem; 473 Node *mem = sfpt_mem; 474 while (!done) { 475 if (visited.test_set(mem->_idx)) { 476 return nullptr; // found a loop, give up 477 } 478 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 479 if (mem == start_mem || mem == alloc_mem) { 480 done = true; // hit a sentinel, return appropriate 0 value 481 } else if (mem->is_Initialize()) { 482 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 483 if (mem == nullptr) { 484 done = true; // Something went wrong. 485 } else if (mem->is_Store()) { 486 const TypePtr* atype = mem->as_Store()->adr_type(); 487 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 488 done = true; 489 } 490 } else if (mem->is_Store()) { 491 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 492 assert(atype != nullptr, "address type must be oopptr"); 493 assert(C->get_alias_index(atype) == alias_idx && 494 atype->is_known_instance_field() && atype->flat_offset() == offset && 495 atype->instance_id() == instance_id, "store is correct memory slice"); 496 done = true; 497 } else if (mem->is_Phi()) { 498 // try to find a phi's unique input 499 Node *unique_input = nullptr; 500 Node *top = C->top(); 501 for (uint i = 1; i < mem->req(); i++) { 502 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 503 if (n == nullptr || n == top || n == mem) { 504 continue; 505 } else if (unique_input == nullptr) { 506 unique_input = n; 507 } else if (unique_input != n) { 508 unique_input = top; 509 break; 510 } 511 } 512 if (unique_input != nullptr && unique_input != top) { 513 mem = unique_input; 514 } else { 515 done = true; 516 } 517 } else if (mem->is_ArrayCopy()) { 518 done = true; 519 } else { 520 DEBUG_ONLY( mem->dump(); ) 521 assert(false, "unexpected node"); 522 } 523 } 524 if (mem != nullptr) { 525 if (mem == start_mem || mem == alloc_mem) { 526 // hit a sentinel, return appropriate 0 value 527 Node* default_value = alloc->in(AllocateNode::DefaultValue); 528 if (default_value != nullptr) { 529 return default_value; 530 } 531 assert(alloc->in(AllocateNode::RawDefaultValue) == nullptr, "default value may not be null"); 532 return _igvn.zerocon(ft); 533 } else if (mem->is_Store()) { 534 Node* n = mem->in(MemNode::ValueIn); 535 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 536 n = bs->step_over_gc_barrier(n); 537 return n; 538 } else if (mem->is_Phi()) { 539 // attempt to produce a Phi reflecting the values on the input paths of the Phi 540 Node_Stack value_phis(8); 541 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 542 if (phi != nullptr) { 543 return phi; 544 } else { 545 // Kill all new Phis 546 while(value_phis.is_nonempty()) { 547 Node* n = value_phis.node(); 548 _igvn.replace_node(n, C->top()); 549 value_phis.pop(); 550 } 551 } 552 } else if (mem->is_ArrayCopy()) { 553 Node* ctl = mem->in(0); 554 Node* m = mem->in(TypeFunc::Memory); 555 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj()) { 556 // pin the loads in the uncommon trap path 557 ctl = sfpt_ctl; 558 m = sfpt_mem; 559 } 560 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc); 561 } 562 } 563 // Something went wrong. 564 return nullptr; 565 } 566 567 // Search the last value stored into the inline type's fields. 568 Node* PhaseMacroExpand::inline_type_from_mem(Node* mem, Node* ctl, ciInlineKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) { 569 // Subtract the offset of the first field to account for the missing oop header 570 offset -= vk->first_field_offset(); 571 // Create a new InlineTypeNode and retrieve the field values from memory 572 InlineTypeNode* vt = InlineTypeNode::make_uninitialized(_igvn, vk); 573 transform_later(vt); 574 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) { 575 ciType* field_type = vt->field_type(i); 576 int field_offset = offset + vt->field_offset(i); 577 Node* value = nullptr; 578 if (vt->field_is_flat(i)) { 579 value = inline_type_from_mem(mem, ctl, field_type->as_inline_klass(), adr_type, field_offset, alloc); 580 } else { 581 const Type* ft = Type::get_const_type(field_type); 582 BasicType bt = type2field[field_type->basic_type()]; 583 if (UseCompressedOops && !is_java_primitive(bt)) { 584 ft = ft->make_narrowoop(); 585 bt = T_NARROWOOP; 586 } 587 // Each inline type field has its own memory slice 588 adr_type = adr_type->with_field_offset(field_offset); 589 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc); 590 if (value != nullptr && ft->isa_narrowoop()) { 591 assert(UseCompressedOops, "unexpected narrow oop"); 592 if (value->is_EncodeP()) { 593 value = value->in(1); 594 } else { 595 value = transform_later(new DecodeNNode(value, value->get_ptr_type())); 596 } 597 } 598 } 599 if (value != nullptr) { 600 vt->set_field_value(i, value); 601 } else { 602 // We might have reached the TrackedInitializationLimit 603 return nullptr; 604 } 605 } 606 return vt; 607 } 608 609 // Check the possibility of scalar replacement. 610 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) { 611 // Scan the uses of the allocation to check for anything that would 612 // prevent us from eliminating it. 613 NOT_PRODUCT( const char* fail_eliminate = nullptr; ) 614 DEBUG_ONLY( Node* disq_node = nullptr; ) 615 bool can_eliminate = true; 616 bool reduce_merge_precheck = (safepoints == nullptr); 617 618 Unique_Node_List worklist; 619 Node* res = alloc->result_cast(); 620 const TypeOopPtr* res_type = nullptr; 621 if (res == nullptr) { 622 // All users were eliminated. 623 } else if (!res->is_CheckCastPP()) { 624 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 625 can_eliminate = false; 626 } else { 627 worklist.push(res); 628 res_type = igvn->type(res)->isa_oopptr(); 629 if (res_type == nullptr) { 630 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 631 can_eliminate = false; 632 } else if (res_type->isa_aryptr()) { 633 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 634 if (length < 0) { 635 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 636 can_eliminate = false; 637 } 638 } 639 } 640 641 while (can_eliminate && worklist.size() > 0) { 642 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2(); 643 res = worklist.pop(); 644 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); j < jmax && can_eliminate; j++) { 645 Node* use = res->fast_out(j); 646 647 if (use->is_AddP()) { 648 const TypePtr* addp_type = igvn->type(use)->is_ptr(); 649 int offset = addp_type->offset(); 650 651 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 652 NOT_PRODUCT(fail_eliminate = "Undefined field reference";) 653 can_eliminate = false; 654 break; 655 } 656 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 657 k < kmax && can_eliminate; k++) { 658 Node* n = use->fast_out(k); 659 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n) && !reduce_merge_precheck) { 660 DEBUG_ONLY(disq_node = n;) 661 if (n->is_Load() || n->is_LoadStore()) { 662 NOT_PRODUCT(fail_eliminate = "Field load";) 663 } else { 664 NOT_PRODUCT(fail_eliminate = "Not store field reference";) 665 } 666 can_eliminate = false; 667 } 668 } 669 } else if (use->is_ArrayCopy() && 670 (use->as_ArrayCopy()->is_clonebasic() || 671 use->as_ArrayCopy()->is_arraycopy_validated() || 672 use->as_ArrayCopy()->is_copyof_validated() || 673 use->as_ArrayCopy()->is_copyofrange_validated()) && 674 use->in(ArrayCopyNode::Dest) == res) { 675 // ok to eliminate 676 } else if (use->is_SafePoint()) { 677 SafePointNode* sfpt = use->as_SafePoint(); 678 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 679 // Object is passed as argument. 680 DEBUG_ONLY(disq_node = use;) 681 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 682 can_eliminate = false; 683 } 684 Node* sfptMem = sfpt->memory(); 685 if (sfptMem == nullptr || sfptMem->is_top()) { 686 DEBUG_ONLY(disq_node = use;) 687 NOT_PRODUCT(fail_eliminate = "null or TOP memory";) 688 can_eliminate = false; 689 } else if (!reduce_merge_precheck) { 690 if (res->is_Phi() && res->as_Phi()->can_be_inline_type()) { 691 // Can only eliminate allocation if the phi had been replaced by an InlineTypeNode before which did not happen. 692 // TODO 8325106 Why wasn't it replaced by an InlineTypeNode? 693 can_eliminate = false; 694 } 695 safepoints->append_if_missing(sfpt); 696 } 697 } else if (use->is_InlineType() && use->as_InlineType()->get_oop() == res) { 698 // Look at uses 699 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) { 700 Node* u = use->fast_out(k); 701 if (u->is_InlineType()) { 702 // Use in flat field can be eliminated 703 InlineTypeNode* vt = u->as_InlineType(); 704 for (uint i = 0; i < vt->field_count(); ++i) { 705 if (vt->field_value(i) == use && !vt->field_is_flat(i)) { 706 can_eliminate = false; // Use in non-flat field 707 break; 708 } 709 } 710 } else { 711 // Add other uses to the worklist to process individually 712 // TODO will be fixed by 8328470 713 worklist.push(use); 714 } 715 } 716 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) { 717 // Store to mark word of inline type larval buffer 718 assert(res_type->is_inlinetypeptr(), "Unexpected store to mark word"); 719 } else if (res_type->is_inlinetypeptr() && use->Opcode() == Op_MemBarRelease) { 720 // Inline type buffer allocations are followed by a membar 721 } else if (reduce_merge_precheck && (use->is_Phi() || use->is_EncodeP() || use->Opcode() == Op_MemBarRelease)) { 722 // Nothing to do 723 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 724 if (use->is_Phi()) { 725 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 726 NOT_PRODUCT(fail_eliminate = "Object is return value";) 727 } else { 728 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 729 } 730 DEBUG_ONLY(disq_node = use;) 731 } else { 732 if (use->Opcode() == Op_Return) { 733 NOT_PRODUCT(fail_eliminate = "Object is return value";) 734 } else { 735 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 736 } 737 DEBUG_ONLY(disq_node = use;) 738 } 739 can_eliminate = false; 740 } else { 741 assert(use->Opcode() == Op_CastP2X, "should be"); 742 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null"); 743 } 744 } 745 } 746 747 #ifndef PRODUCT 748 if (PrintEliminateAllocations && safepoints != nullptr) { 749 if (can_eliminate) { 750 tty->print("Scalar "); 751 if (res == nullptr) 752 alloc->dump(); 753 else 754 res->dump(); 755 } else { 756 tty->print("NotScalar (%s)", fail_eliminate); 757 if (res == nullptr) 758 alloc->dump(); 759 else 760 res->dump(); 761 #ifdef ASSERT 762 if (disq_node != nullptr) { 763 tty->print(" >>>> "); 764 disq_node->dump(); 765 } 766 #endif /*ASSERT*/ 767 } 768 } 769 770 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) { 771 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : ""); 772 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();) 773 } 774 #endif 775 return can_eliminate; 776 } 777 778 void PhaseMacroExpand::undo_previous_scalarizations(GrowableArray <SafePointNode *> safepoints_done, AllocateNode* alloc) { 779 Node* res = alloc->result_cast(); 780 int nfields = 0; 781 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result"); 782 783 if (res != nullptr) { 784 const TypeOopPtr* res_type = _igvn.type(res)->isa_oopptr(); 785 786 if (res_type->isa_instptr()) { 787 // find the fields of the class which will be needed for safepoint debug information 788 ciInstanceKlass* iklass = res_type->is_instptr()->instance_klass(); 789 nfields = iklass->nof_nonstatic_fields(); 790 } else { 791 // find the array's elements which will be needed for safepoint debug information 792 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 793 assert(nfields >= 0, "must be an array klass."); 794 } 795 } 796 797 // rollback processed safepoints 798 while (safepoints_done.length() > 0) { 799 SafePointNode* sfpt_done = safepoints_done.pop(); 800 // remove any extra entries we added to the safepoint 801 uint last = sfpt_done->req() - 1; 802 for (int k = 0; k < nfields; k++) { 803 sfpt_done->del_req(last--); 804 } 805 JVMState *jvms = sfpt_done->jvms(); 806 jvms->set_endoff(sfpt_done->req()); 807 // Now make a pass over the debug information replacing any references 808 // to SafePointScalarObjectNode with the allocated object. 809 int start = jvms->debug_start(); 810 int end = jvms->debug_end(); 811 for (int i = start; i < end; i++) { 812 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 813 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 814 if (scobj->first_index(jvms) == sfpt_done->req() && 815 scobj->n_fields() == (uint)nfields) { 816 assert(scobj->alloc() == alloc, "sanity"); 817 sfpt_done->set_req(i, res); 818 } 819 } 820 } 821 _igvn._worklist.push(sfpt_done); 822 } 823 } 824 825 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode *alloc, SafePointNode* sfpt, 826 Unique_Node_List* value_worklist) { 827 // Fields of scalar objs are referenced only at the end 828 // of regular debuginfo at the last (youngest) JVMS. 829 // Record relative start index. 830 ciInstanceKlass* iklass = nullptr; 831 BasicType basic_elem_type = T_ILLEGAL; 832 const Type* field_type = nullptr; 833 const TypeOopPtr* res_type = nullptr; 834 int nfields = 0; 835 int array_base = 0; 836 int element_size = 0; 837 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff()); 838 Node* res = alloc->result_cast(); 839 840 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result"); 841 assert(sfpt->jvms() != nullptr, "missed JVMS"); 842 843 if (res != nullptr) { // Could be null when there are no users 844 res_type = _igvn.type(res)->isa_oopptr(); 845 846 if (res_type->isa_instptr()) { 847 // find the fields of the class which will be needed for safepoint debug information 848 iklass = res_type->is_instptr()->instance_klass(); 849 nfields = iklass->nof_nonstatic_fields(); 850 } else { 851 // find the array's elements which will be needed for safepoint debug information 852 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 853 assert(nfields >= 0, "must be an array klass."); 854 basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type(); 855 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 856 element_size = type2aelembytes(basic_elem_type); 857 field_type = res_type->is_aryptr()->elem(); 858 if (res_type->is_flat()) { 859 // Flat inline type array 860 element_size = res_type->is_aryptr()->flat_elem_size(); 861 } 862 } 863 } 864 865 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, nfields); 866 sobj->init_req(0, C->root()); 867 transform_later(sobj); 868 869 // Scan object's fields adding an input to the safepoint for each field. 870 for (int j = 0; j < nfields; j++) { 871 intptr_t offset; 872 ciField* field = nullptr; 873 if (iklass != nullptr) { 874 field = iklass->nonstatic_field_at(j); 875 offset = field->offset_in_bytes(); 876 ciType* elem_type = field->type(); 877 basic_elem_type = field->layout_type(); 878 assert(!field->is_flat(), "flat inline type fields should not have safepoint uses"); 879 880 // The next code is taken from Parse::do_get_xxx(). 881 if (is_reference_type(basic_elem_type)) { 882 if (!elem_type->is_loaded()) { 883 field_type = TypeInstPtr::BOTTOM; 884 } else if (field != nullptr && field->is_static_constant()) { 885 ciObject* con = field->constant_value().as_object(); 886 // Do not "join" in the previous type; it doesn't add value, 887 // and may yield a vacuous result if the field is of interface type. 888 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 889 assert(field_type != nullptr, "field singleton type must be consistent"); 890 } else { 891 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 892 } 893 if (UseCompressedOops) { 894 field_type = field_type->make_narrowoop(); 895 basic_elem_type = T_NARROWOOP; 896 } 897 } else { 898 field_type = Type::get_const_basic_type(basic_elem_type); 899 } 900 } else { 901 offset = array_base + j * (intptr_t)element_size; 902 } 903 904 Node* field_val = nullptr; 905 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 906 if (res_type->is_flat()) { 907 ciInlineKlass* inline_klass = res_type->is_aryptr()->elem()->inline_klass(); 908 assert(inline_klass->flat_in_array(), "must be flat in array"); 909 field_val = inline_type_from_mem(sfpt->memory(), sfpt->control(), inline_klass, field_addr_type->isa_aryptr(), 0, alloc); 910 } else { 911 field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc); 912 } 913 914 // We weren't able to find a value for this field, 915 // give up on eliminating this allocation. 916 if (field_val == nullptr) { 917 uint last = sfpt->req() - 1; 918 for (int k = 0; k < j; k++) { 919 sfpt->del_req(last--); 920 } 921 _igvn._worklist.push(sfpt); 922 923 #ifndef PRODUCT 924 if (PrintEliminateAllocations) { 925 if (field != nullptr) { 926 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx); 927 field->print(); 928 int field_idx = C->get_alias_index(field_addr_type); 929 tty->print(" (alias_idx=%d)", field_idx); 930 } else { // Array's element 931 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, j); 932 } 933 tty->print(", which prevents elimination of: "); 934 if (res == nullptr) 935 alloc->dump(); 936 else 937 res->dump(); 938 } 939 #endif 940 941 return nullptr; 942 } 943 944 if (UseCompressedOops && field_type->isa_narrowoop()) { 945 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 946 // to be able scalar replace the allocation. 947 if (field_val->is_EncodeP()) { 948 field_val = field_val->in(1); 949 } else if (!field_val->is_InlineType()) { 950 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type())); 951 } 952 } 953 954 // Keep track of inline types to scalarize them later 955 if (field_val->is_InlineType()) { 956 value_worklist->push(field_val); 957 } else if (field_val->is_Phi()) { 958 PhiNode* phi = field_val->as_Phi(); 959 // Eagerly replace inline type phis now since we could be removing an inline type allocation where we must 960 // scalarize all its fields in safepoints. 961 field_val = phi->try_push_inline_types_down(&_igvn, true); 962 if (field_val->is_InlineType()) { 963 value_worklist->push(field_val); 964 } 965 } 966 sfpt->add_req(field_val); 967 } 968 969 sfpt->jvms()->set_endoff(sfpt->req()); 970 971 return sobj; 972 } 973 974 // Do scalar replacement. 975 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 976 GrowableArray <SafePointNode *> safepoints_done; 977 Node* res = alloc->result_cast(); 978 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result"); 979 const TypeOopPtr* res_type = nullptr; 980 if (res != nullptr) { // Could be null when there are no users 981 res_type = _igvn.type(res)->isa_oopptr(); 982 } 983 984 // Process the safepoint uses 985 assert(safepoints.length() == 0 || !res_type->is_inlinetypeptr(), "Inline type allocations should not have safepoint uses"); 986 Unique_Node_List value_worklist; 987 while (safepoints.length() > 0) { 988 SafePointNode* sfpt = safepoints.pop(); 989 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt, &value_worklist); 990 991 if (sobj == nullptr) { 992 undo_previous_scalarizations(safepoints_done, alloc); 993 return false; 994 } 995 996 // Now make a pass over the debug information replacing any references 997 // to the allocated object with "sobj" 998 JVMState *jvms = sfpt->jvms(); 999 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn); 1000 _igvn._worklist.push(sfpt); 1001 1002 // keep it for rollback 1003 safepoints_done.append_if_missing(sfpt); 1004 } 1005 // Scalarize inline types that were added to the safepoint. 1006 // Don't allow linking a constant oop (if available) for flat array elements 1007 // because Deoptimization::reassign_flat_array_elements needs field values. 1008 bool allow_oop = (res_type != nullptr) && !res_type->is_flat(); 1009 for (uint i = 0; i < value_worklist.size(); ++i) { 1010 InlineTypeNode* vt = value_worklist.at(i)->as_InlineType(); 1011 vt->make_scalar_in_safepoints(&_igvn, allow_oop); 1012 } 1013 return true; 1014 } 1015 1016 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) { 1017 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control); 1018 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory); 1019 if (ctl_proj != nullptr) { 1020 igvn.replace_node(ctl_proj, n->in(0)); 1021 } 1022 if (mem_proj != nullptr) { 1023 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory)); 1024 } 1025 } 1026 1027 // Process users of eliminated allocation. 1028 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc, bool inline_alloc) { 1029 Unique_Node_List worklist; 1030 Node* res = alloc->result_cast(); 1031 if (res != nullptr) { 1032 worklist.push(res); 1033 } 1034 while (worklist.size() > 0) { 1035 res = worklist.pop(); 1036 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 1037 Node *use = res->last_out(j); 1038 uint oc1 = res->outcnt(); 1039 1040 if (use->is_AddP()) { 1041 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 1042 Node *n = use->last_out(k); 1043 uint oc2 = use->outcnt(); 1044 if (n->is_Store()) { 1045 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); p < pmax; p++) { 1046 MemBarNode* mb = n->fast_out(p)->isa_MemBar(); 1047 if (mb != nullptr && mb->req() <= MemBarNode::Precedent && mb->in(MemBarNode::Precedent) == n) { 1048 // MemBarVolatiles should have been removed by MemBarNode::Ideal() for non-inline allocations 1049 assert(inline_alloc, "MemBarVolatile should be eliminated for non-escaping object"); 1050 mb->remove(&_igvn); 1051 } 1052 } 1053 _igvn.replace_node(n, n->in(MemNode::Memory)); 1054 } else { 1055 eliminate_gc_barrier(n); 1056 } 1057 k -= (oc2 - use->outcnt()); 1058 } 1059 _igvn.remove_dead_node(use); 1060 } else if (use->is_ArrayCopy()) { 1061 // Disconnect ArrayCopy node 1062 ArrayCopyNode* ac = use->as_ArrayCopy(); 1063 if (ac->is_clonebasic()) { 1064 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out(); 1065 disconnect_projections(ac, _igvn); 1066 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation"); 1067 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0); 1068 disconnect_projections(membar_before->as_MemBar(), _igvn); 1069 if (membar_after->is_MemBar()) { 1070 disconnect_projections(membar_after->as_MemBar(), _igvn); 1071 } 1072 } else { 1073 assert(ac->is_arraycopy_validated() || 1074 ac->is_copyof_validated() || 1075 ac->is_copyofrange_validated(), "unsupported"); 1076 CallProjections* callprojs = ac->extract_projections(true); 1077 1078 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O)); 1079 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory)); 1080 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control)); 1081 1082 // Set control to top. IGVN will remove the remaining projections 1083 ac->set_req(0, top()); 1084 ac->replace_edge(res, top(), &_igvn); 1085 1086 // Disconnect src right away: it can help find new 1087 // opportunities for allocation elimination 1088 Node* src = ac->in(ArrayCopyNode::Src); 1089 ac->replace_edge(src, top(), &_igvn); 1090 // src can be top at this point if src and dest of the 1091 // arraycopy were the same 1092 if (src->outcnt() == 0 && !src->is_top()) { 1093 _igvn.remove_dead_node(src); 1094 } 1095 } 1096 _igvn._worklist.push(ac); 1097 } else if (use->is_InlineType()) { 1098 assert(use->as_InlineType()->get_oop() == res, "unexpected inline type ptr use"); 1099 // Cut off oop input and remove known instance id from type 1100 _igvn.rehash_node_delayed(use); 1101 use->as_InlineType()->set_oop(_igvn, _igvn.zerocon(T_OBJECT)); 1102 const TypeOopPtr* toop = _igvn.type(use)->is_oopptr()->cast_to_instance_id(TypeOopPtr::InstanceBot); 1103 _igvn.set_type(use, toop); 1104 use->as_InlineType()->set_type(toop); 1105 // Process users 1106 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) { 1107 Node* u = use->fast_out(k); 1108 if (!u->is_InlineType()) { 1109 worklist.push(u); 1110 } 1111 } 1112 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) { 1113 // Store to mark word of inline type larval buffer 1114 assert(inline_alloc, "Unexpected store to mark word"); 1115 _igvn.replace_node(use, use->in(MemNode::Memory)); 1116 } else if (use->Opcode() == Op_MemBarRelease) { 1117 // Inline type buffer allocations are followed by a membar 1118 assert(inline_alloc, "Unexpected MemBarRelease"); 1119 use->as_MemBar()->remove(&_igvn); 1120 } else { 1121 eliminate_gc_barrier(use); 1122 } 1123 j -= (oc1 - res->outcnt()); 1124 } 1125 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 1126 _igvn.remove_dead_node(res); 1127 } 1128 1129 // 1130 // Process other users of allocation's projections 1131 // 1132 if (_callprojs->resproj[0] != nullptr && _callprojs->resproj[0]->outcnt() != 0) { 1133 // First disconnect stores captured by Initialize node. 1134 // If Initialize node is eliminated first in the following code, 1135 // it will kill such stores and DUIterator_Last will assert. 1136 for (DUIterator_Fast jmax, j = _callprojs->resproj[0]->fast_outs(jmax); j < jmax; j++) { 1137 Node* use = _callprojs->resproj[0]->fast_out(j); 1138 if (use->is_AddP()) { 1139 // raw memory addresses used only by the initialization 1140 _igvn.replace_node(use, C->top()); 1141 --j; --jmax; 1142 } 1143 } 1144 for (DUIterator_Last jmin, j = _callprojs->resproj[0]->last_outs(jmin); j >= jmin; ) { 1145 Node* use = _callprojs->resproj[0]->last_out(j); 1146 uint oc1 = _callprojs->resproj[0]->outcnt(); 1147 if (use->is_Initialize()) { 1148 // Eliminate Initialize node. 1149 InitializeNode *init = use->as_Initialize(); 1150 assert(init->outcnt() <= 2, "only a control and memory projection expected"); 1151 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control); 1152 if (ctrl_proj != nullptr) { 1153 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control)); 1154 #ifdef ASSERT 1155 // If the InitializeNode has no memory out, it will die, and tmp will become null 1156 Node* tmp = init->in(TypeFunc::Control); 1157 assert(tmp == nullptr || tmp == _callprojs->fallthrough_catchproj, "allocation control projection"); 1158 #endif 1159 } 1160 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory); 1161 if (mem_proj != nullptr) { 1162 Node *mem = init->in(TypeFunc::Memory); 1163 #ifdef ASSERT 1164 if (mem->is_MergeMem()) { 1165 assert(mem->in(TypeFunc::Memory) == _callprojs->fallthrough_memproj, "allocation memory projection"); 1166 } else { 1167 assert(mem == _callprojs->fallthrough_memproj, "allocation memory projection"); 1168 } 1169 #endif 1170 _igvn.replace_node(mem_proj, mem); 1171 } 1172 } else if (use->Opcode() == Op_MemBarStoreStore) { 1173 // Inline type buffer allocations are followed by a membar 1174 assert(inline_alloc, "Unexpected MemBarStoreStore"); 1175 use->as_MemBar()->remove(&_igvn); 1176 } else { 1177 assert(false, "only Initialize or AddP expected"); 1178 } 1179 j -= (oc1 - _callprojs->resproj[0]->outcnt()); 1180 } 1181 } 1182 if (_callprojs->fallthrough_catchproj != nullptr) { 1183 _igvn.replace_node(_callprojs->fallthrough_catchproj, alloc->in(TypeFunc::Control)); 1184 } 1185 if (_callprojs->fallthrough_memproj != nullptr) { 1186 _igvn.replace_node(_callprojs->fallthrough_memproj, alloc->in(TypeFunc::Memory)); 1187 } 1188 if (_callprojs->catchall_memproj != nullptr) { 1189 _igvn.replace_node(_callprojs->catchall_memproj, C->top()); 1190 } 1191 if (_callprojs->fallthrough_ioproj != nullptr) { 1192 _igvn.replace_node(_callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 1193 } 1194 if (_callprojs->catchall_ioproj != nullptr) { 1195 _igvn.replace_node(_callprojs->catchall_ioproj, C->top()); 1196 } 1197 if (_callprojs->catchall_catchproj != nullptr) { 1198 _igvn.replace_node(_callprojs->catchall_catchproj, C->top()); 1199 } 1200 } 1201 1202 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 1203 // If reallocation fails during deoptimization we'll pop all 1204 // interpreter frames for this compiled frame and that won't play 1205 // nice with JVMTI popframe. 1206 // We avoid this issue by eager reallocation when the popframe request 1207 // is received. 1208 if (!EliminateAllocations) { 1209 return false; 1210 } 1211 Node* klass = alloc->in(AllocateNode::KlassNode); 1212 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); 1213 1214 // Attempt to eliminate inline type buffer allocations 1215 // regardless of usage and escape/replaceable status. 1216 bool inline_alloc = tklass->isa_instklassptr() && 1217 tklass->is_instklassptr()->instance_klass()->is_inlinetype(); 1218 if (!alloc->_is_non_escaping && !inline_alloc) { 1219 return false; 1220 } 1221 // Eliminate boxing allocations which are not used 1222 // regardless scalar replaceable status. 1223 Node* res = alloc->result_cast(); 1224 bool boxing_alloc = (res == nullptr) && C->eliminate_boxing() && 1225 tklass->isa_instklassptr() && 1226 tklass->is_instklassptr()->instance_klass()->is_box_klass(); 1227 if (!alloc->_is_scalar_replaceable && !boxing_alloc && !inline_alloc) { 1228 return false; 1229 } 1230 1231 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/); 1232 1233 GrowableArray <SafePointNode *> safepoints; 1234 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) { 1235 return false; 1236 } 1237 1238 if (!alloc->_is_scalar_replaceable) { 1239 assert(res == nullptr || inline_alloc, "sanity"); 1240 // We can only eliminate allocation if all debug info references 1241 // are already replaced with SafePointScalarObject because 1242 // we can't search for a fields value without instance_id. 1243 if (safepoints.length() > 0) { 1244 assert(!inline_alloc, "Inline type allocations should not have safepoint uses"); 1245 return false; 1246 } 1247 } 1248 1249 if (!scalar_replacement(alloc, safepoints)) { 1250 return false; 1251 } 1252 1253 CompileLog* log = C->log(); 1254 if (log != nullptr) { 1255 log->head("eliminate_allocation type='%d'", 1256 log->identify(tklass->exact_klass())); 1257 JVMState* p = alloc->jvms(); 1258 while (p != nullptr) { 1259 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1260 p = p->caller(); 1261 } 1262 log->tail("eliminate_allocation"); 1263 } 1264 1265 process_users_of_allocation(alloc, inline_alloc); 1266 1267 #ifndef PRODUCT 1268 if (PrintEliminateAllocations) { 1269 if (alloc->is_AllocateArray()) 1270 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1271 else 1272 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1273 } 1274 #endif 1275 1276 return true; 1277 } 1278 1279 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) { 1280 // EA should remove all uses of non-escaping boxing node. 1281 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) { 1282 return false; 1283 } 1284 1285 assert(boxing->result_cast() == nullptr, "unexpected boxing node result"); 1286 1287 _callprojs = boxing->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/); 1288 1289 const TypeTuple* r = boxing->tf()->range_sig(); 1290 assert(r->cnt() > TypeFunc::Parms, "sanity"); 1291 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr(); 1292 assert(t != nullptr, "sanity"); 1293 1294 CompileLog* log = C->log(); 1295 if (log != nullptr) { 1296 log->head("eliminate_boxing type='%d'", 1297 log->identify(t->instance_klass())); 1298 JVMState* p = boxing->jvms(); 1299 while (p != nullptr) { 1300 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1301 p = p->caller(); 1302 } 1303 log->tail("eliminate_boxing"); 1304 } 1305 1306 process_users_of_allocation(boxing); 1307 1308 #ifndef PRODUCT 1309 if (PrintEliminateAllocations) { 1310 tty->print("++++ Eliminated: %d ", boxing->_idx); 1311 boxing->method()->print_short_name(tty); 1312 tty->cr(); 1313 } 1314 #endif 1315 1316 return true; 1317 } 1318 1319 1320 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 1321 Node* adr = basic_plus_adr(base, offset); 1322 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 1323 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered); 1324 transform_later(value); 1325 return value; 1326 } 1327 1328 1329 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 1330 Node* adr = basic_plus_adr(base, offset); 1331 mem = StoreNode::make(_igvn, ctl, mem, adr, nullptr, value, bt, MemNode::unordered); 1332 transform_later(mem); 1333 return mem; 1334 } 1335 1336 //============================================================================= 1337 // 1338 // A L L O C A T I O N 1339 // 1340 // Allocation attempts to be fast in the case of frequent small objects. 1341 // It breaks down like this: 1342 // 1343 // 1) Size in doublewords is computed. This is a constant for objects and 1344 // variable for most arrays. Doubleword units are used to avoid size 1345 // overflow of huge doubleword arrays. We need doublewords in the end for 1346 // rounding. 1347 // 1348 // 2) Size is checked for being 'too large'. Too-large allocations will go 1349 // the slow path into the VM. The slow path can throw any required 1350 // exceptions, and does all the special checks for very large arrays. The 1351 // size test can constant-fold away for objects. For objects with 1352 // finalizers it constant-folds the otherway: you always go slow with 1353 // finalizers. 1354 // 1355 // 3) If NOT using TLABs, this is the contended loop-back point. 1356 // Load-Locked the heap top. If using TLABs normal-load the heap top. 1357 // 1358 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 1359 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 1360 // "size*8" we always enter the VM, where "largish" is a constant picked small 1361 // enough that there's always space between the eden max and 4Gig (old space is 1362 // there so it's quite large) and large enough that the cost of entering the VM 1363 // is dwarfed by the cost to initialize the space. 1364 // 1365 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 1366 // down. If contended, repeat at step 3. If using TLABs normal-store 1367 // adjusted heap top back down; there is no contention. 1368 // 1369 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 1370 // fields. 1371 // 1372 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 1373 // oop flavor. 1374 // 1375 //============================================================================= 1376 // FastAllocateSizeLimit value is in DOUBLEWORDS. 1377 // Allocations bigger than this always go the slow route. 1378 // This value must be small enough that allocation attempts that need to 1379 // trigger exceptions go the slow route. Also, it must be small enough so 1380 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 1381 //=============================================================================j// 1382 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 1383 // The allocator will coalesce int->oop copies away. See comment in 1384 // coalesce.cpp about how this works. It depends critically on the exact 1385 // code shape produced here, so if you are changing this code shape 1386 // make sure the GC info for the heap-top is correct in and around the 1387 // slow-path call. 1388 // 1389 1390 void PhaseMacroExpand::expand_allocate_common( 1391 AllocateNode* alloc, // allocation node to be expanded 1392 Node* length, // array length for an array allocation 1393 const TypeFunc* slow_call_type, // Type of slow call 1394 address slow_call_address, // Address of slow call 1395 Node* valid_length_test // whether length is valid or not 1396 ) 1397 { 1398 Node* ctrl = alloc->in(TypeFunc::Control); 1399 Node* mem = alloc->in(TypeFunc::Memory); 1400 Node* i_o = alloc->in(TypeFunc::I_O); 1401 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 1402 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1403 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 1404 assert(ctrl != nullptr, "must have control"); 1405 1406 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 1407 // they will not be used if "always_slow" is set 1408 enum { slow_result_path = 1, fast_result_path = 2 }; 1409 Node *result_region = nullptr; 1410 Node *result_phi_rawmem = nullptr; 1411 Node *result_phi_rawoop = nullptr; 1412 Node *result_phi_i_o = nullptr; 1413 1414 // The initial slow comparison is a size check, the comparison 1415 // we want to do is a BoolTest::gt 1416 bool expand_fast_path = true; 1417 int tv = _igvn.find_int_con(initial_slow_test, -1); 1418 if (tv >= 0) { 1419 // InitialTest has constant result 1420 // 0 - can fit in TLAB 1421 // 1 - always too big or negative 1422 assert(tv <= 1, "0 or 1 if a constant"); 1423 expand_fast_path = (tv == 0); 1424 initial_slow_test = nullptr; 1425 } else { 1426 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 1427 } 1428 1429 if (!UseTLAB) { 1430 // Force slow-path allocation 1431 expand_fast_path = false; 1432 initial_slow_test = nullptr; 1433 } 1434 1435 bool allocation_has_use = (alloc->result_cast() != nullptr); 1436 if (!allocation_has_use) { 1437 InitializeNode* init = alloc->initialization(); 1438 if (init != nullptr) { 1439 init->remove(&_igvn); 1440 } 1441 if (expand_fast_path && (initial_slow_test == nullptr)) { 1442 // Remove allocation node and return. 1443 // Size is a non-negative constant -> no initial check needed -> directly to fast path. 1444 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left. 1445 #ifndef PRODUCT 1446 if (PrintEliminateAllocations) { 1447 tty->print("NotUsed "); 1448 Node* res = alloc->proj_out_or_null(TypeFunc::Parms); 1449 if (res != nullptr) { 1450 res->dump(); 1451 } else { 1452 alloc->dump(); 1453 } 1454 } 1455 #endif 1456 yank_alloc_node(alloc); 1457 return; 1458 } 1459 } 1460 1461 enum { too_big_or_final_path = 1, need_gc_path = 2 }; 1462 Node *slow_region = nullptr; 1463 Node *toobig_false = ctrl; 1464 1465 // generate the initial test if necessary 1466 if (initial_slow_test != nullptr ) { 1467 assert (expand_fast_path, "Only need test if there is a fast path"); 1468 slow_region = new RegionNode(3); 1469 1470 // Now make the initial failure test. Usually a too-big test but 1471 // might be a TRUE for finalizers or a fancy class check for 1472 // newInstance0. 1473 IfNode* toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1474 transform_later(toobig_iff); 1475 // Plug the failing-too-big test into the slow-path region 1476 Node* toobig_true = new IfTrueNode(toobig_iff); 1477 transform_later(toobig_true); 1478 slow_region ->init_req( too_big_or_final_path, toobig_true ); 1479 toobig_false = new IfFalseNode(toobig_iff); 1480 transform_later(toobig_false); 1481 } else { 1482 // No initial test, just fall into next case 1483 assert(allocation_has_use || !expand_fast_path, "Should already have been handled"); 1484 toobig_false = ctrl; 1485 debug_only(slow_region = NodeSentinel); 1486 } 1487 1488 // If we are here there are several possibilities 1489 // - expand_fast_path is false - then only a slow path is expanded. That's it. 1490 // no_initial_check means a constant allocation. 1491 // - If check always evaluates to false -> expand_fast_path is false (see above) 1492 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath) 1493 // if !allocation_has_use the fast path is empty 1494 // if !allocation_has_use && no_initial_check 1495 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all. 1496 // removed by yank_alloc_node above. 1497 1498 Node *slow_mem = mem; // save the current memory state for slow path 1499 // generate the fast allocation code unless we know that the initial test will always go slow 1500 if (expand_fast_path) { 1501 // Fast path modifies only raw memory. 1502 if (mem->is_MergeMem()) { 1503 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1504 } 1505 1506 // allocate the Region and Phi nodes for the result 1507 result_region = new RegionNode(3); 1508 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1509 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch 1510 1511 // Grab regular I/O before optional prefetch may change it. 1512 // Slow-path does no I/O so just set it to the original I/O. 1513 result_phi_i_o->init_req(slow_result_path, i_o); 1514 1515 // Name successful fast-path variables 1516 Node* fast_oop_ctrl; 1517 Node* fast_oop_rawmem; 1518 1519 if (allocation_has_use) { 1520 Node* needgc_ctrl = nullptr; 1521 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM); 1522 1523 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1524 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1525 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl, 1526 fast_oop_ctrl, fast_oop_rawmem, 1527 prefetch_lines); 1528 1529 if (initial_slow_test != nullptr) { 1530 // This completes all paths into the slow merge point 1531 slow_region->init_req(need_gc_path, needgc_ctrl); 1532 transform_later(slow_region); 1533 } else { 1534 // No initial slow path needed! 1535 // Just fall from the need-GC path straight into the VM call. 1536 slow_region = needgc_ctrl; 1537 } 1538 1539 InitializeNode* init = alloc->initialization(); 1540 fast_oop_rawmem = initialize_object(alloc, 1541 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1542 klass_node, length, size_in_bytes); 1543 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem); 1544 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem); 1545 1546 result_phi_rawoop->init_req(fast_result_path, fast_oop); 1547 } else { 1548 assert (initial_slow_test != nullptr, "sanity"); 1549 fast_oop_ctrl = toobig_false; 1550 fast_oop_rawmem = mem; 1551 transform_later(slow_region); 1552 } 1553 1554 // Plug in the successful fast-path into the result merge point 1555 result_region ->init_req(fast_result_path, fast_oop_ctrl); 1556 result_phi_i_o ->init_req(fast_result_path, i_o); 1557 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); 1558 } else { 1559 slow_region = ctrl; 1560 result_phi_i_o = i_o; // Rename it to use in the following code. 1561 } 1562 1563 // Generate slow-path call 1564 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address, 1565 OptoRuntime::stub_name(slow_call_address), 1566 TypePtr::BOTTOM); 1567 call->init_req(TypeFunc::Control, slow_region); 1568 call->init_req(TypeFunc::I_O, top()); // does no i/o 1569 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs 1570 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1571 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1572 1573 call->init_req(TypeFunc::Parms+0, klass_node); 1574 if (length != nullptr) { 1575 call->init_req(TypeFunc::Parms+1, length); 1576 } else { 1577 // Let the runtime know if this is a larval allocation 1578 call->init_req(TypeFunc::Parms+1, _igvn.intcon(alloc->_larval)); 1579 } 1580 1581 // Copy debug information and adjust JVMState information, then replace 1582 // allocate node with the call 1583 call->copy_call_debug_info(&_igvn, alloc); 1584 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify 1585 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough 1586 // path dies). 1587 if (valid_length_test != nullptr) { 1588 call->add_req(valid_length_test); 1589 } 1590 if (expand_fast_path) { 1591 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1592 } else { 1593 // Hook i_o projection to avoid its elimination during allocation 1594 // replacement (when only a slow call is generated). 1595 call->set_req(TypeFunc::I_O, result_phi_i_o); 1596 } 1597 _igvn.replace_node(alloc, call); 1598 transform_later(call); 1599 1600 // Identify the output projections from the allocate node and 1601 // adjust any references to them. 1602 // The control and io projections look like: 1603 // 1604 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1605 // Allocate Catch 1606 // ^---Proj(io) <-------+ ^---CatchProj(io) 1607 // 1608 // We are interested in the CatchProj nodes. 1609 // 1610 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/); 1611 1612 // An allocate node has separate memory projections for the uses on 1613 // the control and i_o paths. Replace the control memory projection with 1614 // result_phi_rawmem (unless we are only generating a slow call when 1615 // both memory projections are combined) 1616 if (expand_fast_path && _callprojs->fallthrough_memproj != nullptr) { 1617 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, result_phi_rawmem); 1618 } 1619 // Now change uses of catchall_memproj to use fallthrough_memproj and delete 1620 // catchall_memproj so we end up with a call that has only 1 memory projection. 1621 if (_callprojs->catchall_memproj != nullptr) { 1622 if (_callprojs->fallthrough_memproj == nullptr) { 1623 _callprojs->fallthrough_memproj = new ProjNode(call, TypeFunc::Memory); 1624 transform_later(_callprojs->fallthrough_memproj); 1625 } 1626 _igvn.replace_in_uses(_callprojs->catchall_memproj, _callprojs->fallthrough_memproj); 1627 _igvn.remove_dead_node(_callprojs->catchall_memproj); 1628 } 1629 1630 // An allocate node has separate i_o projections for the uses on the control 1631 // and i_o paths. Always replace the control i_o projection with result i_o 1632 // otherwise incoming i_o become dead when only a slow call is generated 1633 // (it is different from memory projections where both projections are 1634 // combined in such case). 1635 if (_callprojs->fallthrough_ioproj != nullptr) { 1636 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, result_phi_i_o); 1637 } 1638 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete 1639 // catchall_ioproj so we end up with a call that has only 1 i_o projection. 1640 if (_callprojs->catchall_ioproj != nullptr) { 1641 if (_callprojs->fallthrough_ioproj == nullptr) { 1642 _callprojs->fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O); 1643 transform_later(_callprojs->fallthrough_ioproj); 1644 } 1645 _igvn.replace_in_uses(_callprojs->catchall_ioproj, _callprojs->fallthrough_ioproj); 1646 _igvn.remove_dead_node(_callprojs->catchall_ioproj); 1647 } 1648 1649 // if we generated only a slow call, we are done 1650 if (!expand_fast_path) { 1651 // Now we can unhook i_o. 1652 if (result_phi_i_o->outcnt() > 1) { 1653 call->set_req(TypeFunc::I_O, top()); 1654 } else { 1655 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity"); 1656 // Case of new array with negative size known during compilation. 1657 // AllocateArrayNode::Ideal() optimization disconnect unreachable 1658 // following code since call to runtime will throw exception. 1659 // As result there will be no users of i_o after the call. 1660 // Leave i_o attached to this call to avoid problems in preceding graph. 1661 } 1662 return; 1663 } 1664 1665 if (_callprojs->fallthrough_catchproj != nullptr) { 1666 ctrl = _callprojs->fallthrough_catchproj->clone(); 1667 transform_later(ctrl); 1668 _igvn.replace_node(_callprojs->fallthrough_catchproj, result_region); 1669 } else { 1670 ctrl = top(); 1671 } 1672 Node *slow_result; 1673 if (_callprojs->resproj[0] == nullptr) { 1674 // no uses of the allocation result 1675 slow_result = top(); 1676 } else { 1677 slow_result = _callprojs->resproj[0]->clone(); 1678 transform_later(slow_result); 1679 _igvn.replace_node(_callprojs->resproj[0], result_phi_rawoop); 1680 } 1681 1682 // Plug slow-path into result merge point 1683 result_region->init_req( slow_result_path, ctrl); 1684 transform_later(result_region); 1685 if (allocation_has_use) { 1686 result_phi_rawoop->init_req(slow_result_path, slow_result); 1687 transform_later(result_phi_rawoop); 1688 } 1689 result_phi_rawmem->init_req(slow_result_path, _callprojs->fallthrough_memproj); 1690 transform_later(result_phi_rawmem); 1691 transform_later(result_phi_i_o); 1692 // This completes all paths into the result merge point 1693 } 1694 1695 // Remove alloc node that has no uses. 1696 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) { 1697 Node* ctrl = alloc->in(TypeFunc::Control); 1698 Node* mem = alloc->in(TypeFunc::Memory); 1699 Node* i_o = alloc->in(TypeFunc::I_O); 1700 1701 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/); 1702 if (_callprojs->resproj[0] != nullptr) { 1703 for (DUIterator_Fast imax, i = _callprojs->resproj[0]->fast_outs(imax); i < imax; i++) { 1704 Node* use = _callprojs->resproj[0]->fast_out(i); 1705 use->isa_MemBar()->remove(&_igvn); 1706 --imax; 1707 --i; // back up iterator 1708 } 1709 assert(_callprojs->resproj[0]->outcnt() == 0, "all uses must be deleted"); 1710 _igvn.remove_dead_node(_callprojs->resproj[0]); 1711 } 1712 if (_callprojs->fallthrough_catchproj != nullptr) { 1713 _igvn.replace_in_uses(_callprojs->fallthrough_catchproj, ctrl); 1714 _igvn.remove_dead_node(_callprojs->fallthrough_catchproj); 1715 } 1716 if (_callprojs->catchall_catchproj != nullptr) { 1717 _igvn.rehash_node_delayed(_callprojs->catchall_catchproj); 1718 _callprojs->catchall_catchproj->set_req(0, top()); 1719 } 1720 if (_callprojs->fallthrough_proj != nullptr) { 1721 Node* catchnode = _callprojs->fallthrough_proj->unique_ctrl_out(); 1722 _igvn.remove_dead_node(catchnode); 1723 _igvn.remove_dead_node(_callprojs->fallthrough_proj); 1724 } 1725 if (_callprojs->fallthrough_memproj != nullptr) { 1726 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, mem); 1727 _igvn.remove_dead_node(_callprojs->fallthrough_memproj); 1728 } 1729 if (_callprojs->fallthrough_ioproj != nullptr) { 1730 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, i_o); 1731 _igvn.remove_dead_node(_callprojs->fallthrough_ioproj); 1732 } 1733 if (_callprojs->catchall_memproj != nullptr) { 1734 _igvn.rehash_node_delayed(_callprojs->catchall_memproj); 1735 _callprojs->catchall_memproj->set_req(0, top()); 1736 } 1737 if (_callprojs->catchall_ioproj != nullptr) { 1738 _igvn.rehash_node_delayed(_callprojs->catchall_ioproj); 1739 _callprojs->catchall_ioproj->set_req(0, top()); 1740 } 1741 #ifndef PRODUCT 1742 if (PrintEliminateAllocations) { 1743 if (alloc->is_AllocateArray()) { 1744 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1745 } else { 1746 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1747 } 1748 } 1749 #endif 1750 _igvn.remove_dead_node(alloc); 1751 } 1752 1753 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init, 1754 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) { 1755 // If initialization is performed by an array copy, any required 1756 // MemBarStoreStore was already added. If the object does not 1757 // escape no need for a MemBarStoreStore. If the object does not 1758 // escape in its initializer and memory barrier (MemBarStoreStore or 1759 // stronger) is already added at exit of initializer, also no need 1760 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore 1761 // so that stores that initialize this object can't be reordered 1762 // with a subsequent store that makes this object accessible by 1763 // other threads. 1764 // Other threads include java threads and JVM internal threads 1765 // (for example concurrent GC threads). Current concurrent GC 1766 // implementation: G1 will not scan newly created object, 1767 // so it's safe to skip storestore barrier when allocation does 1768 // not escape. 1769 if (!alloc->does_not_escape_thread() && 1770 !alloc->is_allocation_MemBar_redundant() && 1771 (init == nullptr || !init->is_complete_with_arraycopy())) { 1772 if (init == nullptr || init->req() < InitializeNode::RawStores) { 1773 // No InitializeNode or no stores captured by zeroing 1774 // elimination. Simply add the MemBarStoreStore after object 1775 // initialization. 1776 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1777 transform_later(mb); 1778 1779 mb->init_req(TypeFunc::Memory, fast_oop_rawmem); 1780 mb->init_req(TypeFunc::Control, fast_oop_ctrl); 1781 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control); 1782 transform_later(fast_oop_ctrl); 1783 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory); 1784 transform_later(fast_oop_rawmem); 1785 } else { 1786 // Add the MemBarStoreStore after the InitializeNode so that 1787 // all stores performing the initialization that were moved 1788 // before the InitializeNode happen before the storestore 1789 // barrier. 1790 1791 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control); 1792 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory); 1793 1794 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1795 transform_later(mb); 1796 1797 Node* ctrl = new ProjNode(init, TypeFunc::Control); 1798 transform_later(ctrl); 1799 Node* mem = new ProjNode(init, TypeFunc::Memory); 1800 transform_later(mem); 1801 1802 // The MemBarStoreStore depends on control and memory coming 1803 // from the InitializeNode 1804 mb->init_req(TypeFunc::Memory, mem); 1805 mb->init_req(TypeFunc::Control, ctrl); 1806 1807 ctrl = new ProjNode(mb, TypeFunc::Control); 1808 transform_later(ctrl); 1809 mem = new ProjNode(mb, TypeFunc::Memory); 1810 transform_later(mem); 1811 1812 // All nodes that depended on the InitializeNode for control 1813 // and memory must now depend on the MemBarNode that itself 1814 // depends on the InitializeNode 1815 if (init_ctrl != nullptr) { 1816 _igvn.replace_node(init_ctrl, ctrl); 1817 } 1818 if (init_mem != nullptr) { 1819 _igvn.replace_node(init_mem, mem); 1820 } 1821 } 1822 } 1823 } 1824 1825 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop, 1826 Node*& ctrl, Node*& rawmem) { 1827 if (C->env()->dtrace_alloc_probes()) { 1828 // Slow-path call 1829 int size = TypeFunc::Parms + 2; 1830 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1831 CAST_FROM_FN_PTR(address, 1832 static_cast<int (*)(JavaThread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)), 1833 "dtrace_object_alloc", 1834 TypeRawPtr::BOTTOM); 1835 1836 // Get base of thread-local storage area 1837 Node* thread = new ThreadLocalNode(); 1838 transform_later(thread); 1839 1840 call->init_req(TypeFunc::Parms + 0, thread); 1841 call->init_req(TypeFunc::Parms + 1, oop); 1842 call->init_req(TypeFunc::Control, ctrl); 1843 call->init_req(TypeFunc::I_O , top()); // does no i/o 1844 call->init_req(TypeFunc::Memory , rawmem); 1845 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1846 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1847 transform_later(call); 1848 ctrl = new ProjNode(call, TypeFunc::Control); 1849 transform_later(ctrl); 1850 rawmem = new ProjNode(call, TypeFunc::Memory); 1851 transform_later(rawmem); 1852 } 1853 } 1854 1855 // Helper for PhaseMacroExpand::expand_allocate_common. 1856 // Initializes the newly-allocated storage. 1857 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1858 Node* control, Node* rawmem, Node* object, 1859 Node* klass_node, Node* length, 1860 Node* size_in_bytes) { 1861 InitializeNode* init = alloc->initialization(); 1862 // Store the klass & mark bits 1863 Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem); 1864 if (!mark_node->is_Con()) { 1865 transform_later(mark_node); 1866 } 1867 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type()); 1868 1869 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 1870 int header_size = alloc->minimum_header_size(); // conservatively small 1871 1872 // Array length 1873 if (length != nullptr) { // Arrays need length field 1874 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1875 // conservatively small header size: 1876 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1877 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases: 1878 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type(); 1879 if (is_reference_type(elem, true)) { 1880 elem = T_OBJECT; 1881 } 1882 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem)); 1883 } 1884 } 1885 1886 // Clear the object body, if necessary. 1887 if (init == nullptr) { 1888 // The init has somehow disappeared; be cautious and clear everything. 1889 // 1890 // This can happen if a node is allocated but an uncommon trap occurs 1891 // immediately. In this case, the Initialize gets associated with the 1892 // trap, and may be placed in a different (outer) loop, if the Allocate 1893 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1894 // there can be two Allocates to one Initialize. The answer in all these 1895 // edge cases is safety first. It is always safe to clear immediately 1896 // within an Allocate, and then (maybe or maybe not) clear some more later. 1897 if (!(UseTLAB && ZeroTLAB)) { 1898 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1899 alloc->in(AllocateNode::DefaultValue), 1900 alloc->in(AllocateNode::RawDefaultValue), 1901 header_size, size_in_bytes, 1902 &_igvn); 1903 } 1904 } else { 1905 if (!init->is_complete()) { 1906 // Try to win by zeroing only what the init does not store. 1907 // We can also try to do some peephole optimizations, 1908 // such as combining some adjacent subword stores. 1909 rawmem = init->complete_stores(control, rawmem, object, 1910 header_size, size_in_bytes, &_igvn); 1911 } 1912 // We have no more use for this link, since the AllocateNode goes away: 1913 init->set_req(InitializeNode::RawAddress, top()); 1914 // (If we keep the link, it just confuses the register allocator, 1915 // who thinks he sees a real use of the address by the membar.) 1916 } 1917 1918 return rawmem; 1919 } 1920 1921 // Generate prefetch instructions for next allocations. 1922 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1923 Node*& contended_phi_rawmem, 1924 Node* old_eden_top, Node* new_eden_top, 1925 intx lines) { 1926 enum { fall_in_path = 1, pf_path = 2 }; 1927 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1928 // Generate prefetch allocation with watermark check. 1929 // As an allocation hits the watermark, we will prefetch starting 1930 // at a "distance" away from watermark. 1931 1932 Node *pf_region = new RegionNode(3); 1933 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 1934 TypeRawPtr::BOTTOM ); 1935 // I/O is used for Prefetch 1936 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO ); 1937 1938 Node *thread = new ThreadLocalNode(); 1939 transform_later(thread); 1940 1941 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread, 1942 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1943 transform_later(eden_pf_adr); 1944 1945 Node *old_pf_wm = new LoadPNode(needgc_false, 1946 contended_phi_rawmem, eden_pf_adr, 1947 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, 1948 MemNode::unordered); 1949 transform_later(old_pf_wm); 1950 1951 // check against new_eden_top 1952 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm ); 1953 transform_later(need_pf_cmp); 1954 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge ); 1955 transform_later(need_pf_bol); 1956 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol, 1957 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1958 transform_later(need_pf_iff); 1959 1960 // true node, add prefetchdistance 1961 Node *need_pf_true = new IfTrueNode( need_pf_iff ); 1962 transform_later(need_pf_true); 1963 1964 Node *need_pf_false = new IfFalseNode( need_pf_iff ); 1965 transform_later(need_pf_false); 1966 1967 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm, 1968 _igvn.MakeConX(AllocatePrefetchDistance) ); 1969 transform_later(new_pf_wmt ); 1970 new_pf_wmt->set_req(0, need_pf_true); 1971 1972 Node *store_new_wmt = new StorePNode(need_pf_true, 1973 contended_phi_rawmem, eden_pf_adr, 1974 TypeRawPtr::BOTTOM, new_pf_wmt, 1975 MemNode::unordered); 1976 transform_later(store_new_wmt); 1977 1978 // adding prefetches 1979 pf_phi_abio->init_req( fall_in_path, i_o ); 1980 1981 Node *prefetch_adr; 1982 Node *prefetch; 1983 uint step_size = AllocatePrefetchStepSize; 1984 uint distance = 0; 1985 1986 for ( intx i = 0; i < lines; i++ ) { 1987 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt, 1988 _igvn.MakeConX(distance) ); 1989 transform_later(prefetch_adr); 1990 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1991 transform_later(prefetch); 1992 distance += step_size; 1993 i_o = prefetch; 1994 } 1995 pf_phi_abio->set_req( pf_path, i_o ); 1996 1997 pf_region->init_req( fall_in_path, need_pf_false ); 1998 pf_region->init_req( pf_path, need_pf_true ); 1999 2000 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 2001 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 2002 2003 transform_later(pf_region); 2004 transform_later(pf_phi_rawmem); 2005 transform_later(pf_phi_abio); 2006 2007 needgc_false = pf_region; 2008 contended_phi_rawmem = pf_phi_rawmem; 2009 i_o = pf_phi_abio; 2010 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { 2011 // Insert a prefetch instruction for each allocation. 2012 // This code is used to generate 1 prefetch instruction per cache line. 2013 2014 // Generate several prefetch instructions. 2015 uint step_size = AllocatePrefetchStepSize; 2016 uint distance = AllocatePrefetchDistance; 2017 2018 // Next cache address. 2019 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top, 2020 _igvn.MakeConX(step_size + distance)); 2021 transform_later(cache_adr); 2022 cache_adr = new CastP2XNode(needgc_false, cache_adr); 2023 transform_later(cache_adr); 2024 // Address is aligned to execute prefetch to the beginning of cache line size 2025 // (it is important when BIS instruction is used on SPARC as prefetch). 2026 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); 2027 cache_adr = new AndXNode(cache_adr, mask); 2028 transform_later(cache_adr); 2029 cache_adr = new CastX2PNode(cache_adr); 2030 transform_later(cache_adr); 2031 2032 // Prefetch 2033 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); 2034 prefetch->set_req(0, needgc_false); 2035 transform_later(prefetch); 2036 contended_phi_rawmem = prefetch; 2037 Node *prefetch_adr; 2038 distance = step_size; 2039 for ( intx i = 1; i < lines; i++ ) { 2040 prefetch_adr = new AddPNode( cache_adr, cache_adr, 2041 _igvn.MakeConX(distance) ); 2042 transform_later(prefetch_adr); 2043 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); 2044 transform_later(prefetch); 2045 distance += step_size; 2046 contended_phi_rawmem = prefetch; 2047 } 2048 } else if( AllocatePrefetchStyle > 0 ) { 2049 // Insert a prefetch for each allocation only on the fast-path 2050 Node *prefetch_adr; 2051 Node *prefetch; 2052 // Generate several prefetch instructions. 2053 uint step_size = AllocatePrefetchStepSize; 2054 uint distance = AllocatePrefetchDistance; 2055 for ( intx i = 0; i < lines; i++ ) { 2056 prefetch_adr = new AddPNode( old_eden_top, new_eden_top, 2057 _igvn.MakeConX(distance) ); 2058 transform_later(prefetch_adr); 2059 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 2060 // Do not let it float too high, since if eden_top == eden_end, 2061 // both might be null. 2062 if( i == 0 ) { // Set control for first prefetch, next follows it 2063 prefetch->init_req(0, needgc_false); 2064 } 2065 transform_later(prefetch); 2066 distance += step_size; 2067 i_o = prefetch; 2068 } 2069 } 2070 return i_o; 2071 } 2072 2073 2074 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 2075 expand_allocate_common(alloc, nullptr, 2076 OptoRuntime::new_instance_Type(), 2077 OptoRuntime::new_instance_Java(), nullptr); 2078 } 2079 2080 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 2081 Node* length = alloc->in(AllocateNode::ALength); 2082 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest); 2083 InitializeNode* init = alloc->initialization(); 2084 Node* klass_node = alloc->in(AllocateNode::KlassNode); 2085 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr(); 2086 address slow_call_address; // Address of slow call 2087 if (init != nullptr && init->is_complete_with_arraycopy() && 2088 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == nullptr) { 2089 // Don't zero type array during slow allocation in VM since 2090 // it will be initialized later by arraycopy in compiled code. 2091 slow_call_address = OptoRuntime::new_array_nozero_Java(); 2092 } else { 2093 slow_call_address = OptoRuntime::new_array_Java(); 2094 } 2095 expand_allocate_common(alloc, length, 2096 OptoRuntime::new_array_Type(), 2097 slow_call_address, valid_length_test); 2098 } 2099 2100 //-------------------mark_eliminated_box---------------------------------- 2101 // 2102 // During EA obj may point to several objects but after few ideal graph 2103 // transformations (CCP) it may point to only one non escaping object 2104 // (but still using phi), corresponding locks and unlocks will be marked 2105 // for elimination. Later obj could be replaced with a new node (new phi) 2106 // and which does not have escape information. And later after some graph 2107 // reshape other locks and unlocks (which were not marked for elimination 2108 // before) are connected to this new obj (phi) but they still will not be 2109 // marked for elimination since new obj has no escape information. 2110 // Mark all associated (same box and obj) lock and unlock nodes for 2111 // elimination if some of them marked already. 2112 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) { 2113 if (oldbox->as_BoxLock()->is_eliminated()) { 2114 return; // This BoxLock node was processed already. 2115 } 2116 // New implementation (EliminateNestedLocks) has separate BoxLock 2117 // node for each locked region so mark all associated locks/unlocks as 2118 // eliminated even if different objects are referenced in one locked region 2119 // (for example, OSR compilation of nested loop inside locked scope). 2120 if (EliminateNestedLocks || 2121 oldbox->as_BoxLock()->is_simple_lock_region(nullptr, obj, nullptr)) { 2122 // Box is used only in one lock region. Mark this box as eliminated. 2123 _igvn.hash_delete(oldbox); 2124 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value 2125 _igvn.hash_insert(oldbox); 2126 2127 for (uint i = 0; i < oldbox->outcnt(); i++) { 2128 Node* u = oldbox->raw_out(i); 2129 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) { 2130 AbstractLockNode* alock = u->as_AbstractLock(); 2131 // Check lock's box since box could be referenced by Lock's debug info. 2132 if (alock->box_node() == oldbox) { 2133 // Mark eliminated all related locks and unlocks. 2134 #ifdef ASSERT 2135 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4"); 2136 #endif 2137 alock->set_non_esc_obj(); 2138 } 2139 } 2140 } 2141 return; 2142 } 2143 2144 // Create new "eliminated" BoxLock node and use it in monitor debug info 2145 // instead of oldbox for the same object. 2146 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 2147 2148 // Note: BoxLock node is marked eliminated only here and it is used 2149 // to indicate that all associated lock and unlock nodes are marked 2150 // for elimination. 2151 newbox->set_eliminated(); 2152 transform_later(newbox); 2153 2154 // Replace old box node with new box for all users of the same object. 2155 for (uint i = 0; i < oldbox->outcnt();) { 2156 bool next_edge = true; 2157 2158 Node* u = oldbox->raw_out(i); 2159 if (u->is_AbstractLock()) { 2160 AbstractLockNode* alock = u->as_AbstractLock(); 2161 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) { 2162 // Replace Box and mark eliminated all related locks and unlocks. 2163 #ifdef ASSERT 2164 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5"); 2165 #endif 2166 alock->set_non_esc_obj(); 2167 _igvn.rehash_node_delayed(alock); 2168 alock->set_box_node(newbox); 2169 next_edge = false; 2170 } 2171 } 2172 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) { 2173 FastLockNode* flock = u->as_FastLock(); 2174 assert(flock->box_node() == oldbox, "sanity"); 2175 _igvn.rehash_node_delayed(flock); 2176 flock->set_box_node(newbox); 2177 next_edge = false; 2178 } 2179 2180 // Replace old box in monitor debug info. 2181 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 2182 SafePointNode* sfn = u->as_SafePoint(); 2183 JVMState* youngest_jvms = sfn->jvms(); 2184 int max_depth = youngest_jvms->depth(); 2185 for (int depth = 1; depth <= max_depth; depth++) { 2186 JVMState* jvms = youngest_jvms->of_depth(depth); 2187 int num_mon = jvms->nof_monitors(); 2188 // Loop over monitors 2189 for (int idx = 0; idx < num_mon; idx++) { 2190 Node* obj_node = sfn->monitor_obj(jvms, idx); 2191 Node* box_node = sfn->monitor_box(jvms, idx); 2192 if (box_node == oldbox && obj_node->eqv_uncast(obj)) { 2193 int j = jvms->monitor_box_offset(idx); 2194 _igvn.replace_input_of(u, j, newbox); 2195 next_edge = false; 2196 } 2197 } 2198 } 2199 } 2200 if (next_edge) i++; 2201 } 2202 } 2203 2204 //-----------------------mark_eliminated_locking_nodes----------------------- 2205 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { 2206 if (EliminateNestedLocks) { 2207 if (alock->is_nested()) { 2208 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity"); 2209 return; 2210 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened 2211 // Only Lock node has JVMState needed here. 2212 // Not that preceding claim is documented anywhere else. 2213 if (alock->jvms() != nullptr) { 2214 if (alock->as_Lock()->is_nested_lock_region()) { 2215 // Mark eliminated related nested locks and unlocks. 2216 Node* obj = alock->obj_node(); 2217 BoxLockNode* box_node = alock->box_node()->as_BoxLock(); 2218 assert(!box_node->is_eliminated(), "should not be marked yet"); 2219 // Note: BoxLock node is marked eliminated only here 2220 // and it is used to indicate that all associated lock 2221 // and unlock nodes are marked for elimination. 2222 box_node->set_eliminated(); // Box's hash is always NO_HASH here 2223 for (uint i = 0; i < box_node->outcnt(); i++) { 2224 Node* u = box_node->raw_out(i); 2225 if (u->is_AbstractLock()) { 2226 alock = u->as_AbstractLock(); 2227 if (alock->box_node() == box_node) { 2228 // Verify that this Box is referenced only by related locks. 2229 assert(alock->obj_node()->eqv_uncast(obj), ""); 2230 // Mark all related locks and unlocks. 2231 #ifdef ASSERT 2232 alock->log_lock_optimization(C, "eliminate_lock_set_nested"); 2233 #endif 2234 alock->set_nested(); 2235 } 2236 } 2237 } 2238 } else { 2239 #ifdef ASSERT 2240 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region"); 2241 if (C->log() != nullptr) 2242 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output 2243 #endif 2244 } 2245 } 2246 return; 2247 } 2248 // Process locks for non escaping object 2249 assert(alock->is_non_esc_obj(), ""); 2250 } // EliminateNestedLocks 2251 2252 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object 2253 // Look for all locks of this object and mark them and 2254 // corresponding BoxLock nodes as eliminated. 2255 Node* obj = alock->obj_node(); 2256 for (uint j = 0; j < obj->outcnt(); j++) { 2257 Node* o = obj->raw_out(j); 2258 if (o->is_AbstractLock() && 2259 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) { 2260 alock = o->as_AbstractLock(); 2261 Node* box = alock->box_node(); 2262 // Replace old box node with new eliminated box for all users 2263 // of the same object and mark related locks as eliminated. 2264 mark_eliminated_box(box, obj); 2265 } 2266 } 2267 } 2268 } 2269 2270 void PhaseMacroExpand::inline_type_guard(Node** ctrl, LockNode* lock) { 2271 Node* obj = lock->obj_node(); 2272 const TypePtr* obj_type = _igvn.type(obj)->make_ptr(); 2273 if (!obj_type->can_be_inline_type()) { 2274 return; 2275 } 2276 Node* mark = make_load(*ctrl, lock->memory(), obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2277 Node* value_mask = _igvn.MakeConX(markWord::inline_type_pattern); 2278 Node* is_value = _igvn.transform(new AndXNode(mark, value_mask)); 2279 Node* cmp = _igvn.transform(new CmpXNode(is_value, value_mask)); 2280 Node* bol = _igvn.transform(new BoolNode(cmp, BoolTest::eq)); 2281 Node* unc_ctrl = generate_slow_guard(ctrl, bol, nullptr); 2282 2283 int trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_class_check, Deoptimization::Action_none); 2284 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2285 const TypePtr* no_memory_effects = nullptr; 2286 CallNode* unc = new CallStaticJavaNode(OptoRuntime::uncommon_trap_Type(), call_addr, "uncommon_trap", 2287 no_memory_effects); 2288 unc->init_req(TypeFunc::Control, unc_ctrl); 2289 unc->init_req(TypeFunc::I_O, lock->i_o()); 2290 unc->init_req(TypeFunc::Memory, lock->memory()); 2291 unc->init_req(TypeFunc::FramePtr, lock->in(TypeFunc::FramePtr)); 2292 unc->init_req(TypeFunc::ReturnAdr, lock->in(TypeFunc::ReturnAdr)); 2293 unc->init_req(TypeFunc::Parms+0, _igvn.intcon(trap_request)); 2294 unc->set_cnt(PROB_UNLIKELY_MAG(4)); 2295 unc->copy_call_debug_info(&_igvn, lock); 2296 2297 assert(unc->peek_monitor_box() == lock->box_node(), "wrong monitor"); 2298 assert((obj_type->is_inlinetypeptr() && unc->peek_monitor_obj()->is_SafePointScalarObject()) || 2299 (obj->is_InlineType() && obj->in(1) == unc->peek_monitor_obj()) || 2300 (obj == unc->peek_monitor_obj()), "wrong monitor"); 2301 2302 // pop monitor and push obj back on stack: we trap before the monitorenter 2303 unc->pop_monitor(); 2304 unc->grow_stack(unc->jvms(), 1); 2305 unc->set_stack(unc->jvms(), unc->jvms()->stk_size()-1, obj); 2306 _igvn.register_new_node_with_optimizer(unc); 2307 2308 unc_ctrl = _igvn.transform(new ProjNode(unc, TypeFunc::Control)); 2309 Node* halt = _igvn.transform(new HaltNode(unc_ctrl, lock->in(TypeFunc::FramePtr), "monitor enter on inline type")); 2310 _igvn.add_input_to(C->root(), halt); 2311 } 2312 2313 // we have determined that this lock/unlock can be eliminated, we simply 2314 // eliminate the node without expanding it. 2315 // 2316 // Note: The membar's associated with the lock/unlock are currently not 2317 // eliminated. This should be investigated as a future enhancement. 2318 // 2319 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 2320 2321 if (!alock->is_eliminated()) { 2322 return false; 2323 } 2324 #ifdef ASSERT 2325 if (!alock->is_coarsened()) { 2326 // Check that new "eliminated" BoxLock node is created. 2327 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 2328 assert(oldbox->is_eliminated(), "should be done already"); 2329 } 2330 #endif 2331 2332 alock->log_lock_optimization(C, "eliminate_lock"); 2333 2334 #ifndef PRODUCT 2335 if (PrintEliminateLocks) { 2336 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string()); 2337 } 2338 #endif 2339 2340 Node* mem = alock->in(TypeFunc::Memory); 2341 Node* ctrl = alock->in(TypeFunc::Control); 2342 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null"); 2343 2344 _callprojs = alock->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/); 2345 // There are 2 projections from the lock. The lock node will 2346 // be deleted when its last use is subsumed below. 2347 assert(alock->outcnt() == 2 && 2348 _callprojs->fallthrough_proj != nullptr && 2349 _callprojs->fallthrough_memproj != nullptr, 2350 "Unexpected projections from Lock/Unlock"); 2351 2352 Node* fallthroughproj = _callprojs->fallthrough_proj; 2353 Node* memproj_fallthrough = _callprojs->fallthrough_memproj; 2354 2355 // The memory projection from a lock/unlock is RawMem 2356 // The input to a Lock is merged memory, so extract its RawMem input 2357 // (unless the MergeMem has been optimized away.) 2358 if (alock->is_Lock()) { 2359 // Deoptimize and re-execute if object is an inline type 2360 inline_type_guard(&ctrl, alock->as_Lock()); 2361 2362 // Search for MemBarAcquireLock node and delete it also. 2363 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 2364 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, ""); 2365 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 2366 Node* memproj = membar->proj_out(TypeFunc::Memory); 2367 _igvn.replace_node(ctrlproj, fallthroughproj); 2368 _igvn.replace_node(memproj, memproj_fallthrough); 2369 2370 // Delete FastLock node also if this Lock node is unique user 2371 // (a loop peeling may clone a Lock node). 2372 Node* flock = alock->as_Lock()->fastlock_node(); 2373 if (flock->outcnt() == 1) { 2374 assert(flock->unique_out() == alock, "sanity"); 2375 _igvn.replace_node(flock, top()); 2376 } 2377 } 2378 2379 // Search for MemBarReleaseLock node and delete it also. 2380 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) { 2381 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 2382 assert(membar->Opcode() == Op_MemBarReleaseLock && 2383 mem->is_Proj() && membar == mem->in(0), ""); 2384 _igvn.replace_node(fallthroughproj, ctrl); 2385 _igvn.replace_node(memproj_fallthrough, mem); 2386 fallthroughproj = ctrl; 2387 memproj_fallthrough = mem; 2388 ctrl = membar->in(TypeFunc::Control); 2389 mem = membar->in(TypeFunc::Memory); 2390 } 2391 2392 _igvn.replace_node(fallthroughproj, ctrl); 2393 _igvn.replace_node(memproj_fallthrough, mem); 2394 return true; 2395 } 2396 2397 2398 //------------------------------expand_lock_node---------------------- 2399 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 2400 2401 Node* ctrl = lock->in(TypeFunc::Control); 2402 Node* mem = lock->in(TypeFunc::Memory); 2403 Node* obj = lock->obj_node(); 2404 Node* box = lock->box_node(); 2405 Node* flock = lock->fastlock_node(); 2406 2407 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2408 2409 // Make the merge point 2410 Node *region; 2411 Node *mem_phi; 2412 Node *slow_path; 2413 2414 region = new RegionNode(3); 2415 // create a Phi for the memory state 2416 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2417 2418 // Optimize test; set region slot 2 2419 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); 2420 mem_phi->init_req(2, mem); 2421 2422 // Deoptimize and re-execute if object is an inline type 2423 inline_type_guard(&slow_path, lock); 2424 2425 // Make slow path call 2426 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), 2427 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path, 2428 obj, box, nullptr); 2429 2430 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/); 2431 2432 // Slow path can only throw asynchronous exceptions, which are always 2433 // de-opted. So the compiler thinks the slow-call can never throw an 2434 // exception. If it DOES throw an exception we would need the debug 2435 // info removed first (since if it throws there is no monitor). 2436 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr && 2437 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock"); 2438 2439 // Capture slow path 2440 // disconnect fall-through projection from call and create a new one 2441 // hook up users of fall-through projection to region 2442 Node *slow_ctrl = _callprojs->fallthrough_proj->clone(); 2443 transform_later(slow_ctrl); 2444 _igvn.hash_delete(_callprojs->fallthrough_proj); 2445 _callprojs->fallthrough_proj->disconnect_inputs(C); 2446 region->init_req(1, slow_ctrl); 2447 // region inputs are now complete 2448 transform_later(region); 2449 _igvn.replace_node(_callprojs->fallthrough_proj, region); 2450 2451 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory)); 2452 2453 mem_phi->init_req(1, memproj); 2454 2455 transform_later(mem_phi); 2456 2457 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi); 2458 } 2459 2460 //------------------------------expand_unlock_node---------------------- 2461 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 2462 2463 Node* ctrl = unlock->in(TypeFunc::Control); 2464 Node* mem = unlock->in(TypeFunc::Memory); 2465 Node* obj = unlock->obj_node(); 2466 Node* box = unlock->box_node(); 2467 2468 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2469 2470 // No need for a null check on unlock 2471 2472 // Make the merge point 2473 Node *region; 2474 Node *mem_phi; 2475 2476 region = new RegionNode(3); 2477 // create a Phi for the memory state 2478 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2479 2480 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box ); 2481 funlock = transform_later( funlock )->as_FastUnlock(); 2482 // Optimize test; set region slot 2 2483 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); 2484 Node *thread = transform_later(new ThreadLocalNode()); 2485 2486 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), 2487 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), 2488 "complete_monitor_unlocking_C", slow_path, obj, box, thread); 2489 2490 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/); 2491 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr && 2492 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock"); 2493 2494 // No exceptions for unlocking 2495 // Capture slow path 2496 // disconnect fall-through projection from call and create a new one 2497 // hook up users of fall-through projection to region 2498 Node *slow_ctrl = _callprojs->fallthrough_proj->clone(); 2499 transform_later(slow_ctrl); 2500 _igvn.hash_delete(_callprojs->fallthrough_proj); 2501 _callprojs->fallthrough_proj->disconnect_inputs(C); 2502 region->init_req(1, slow_ctrl); 2503 // region inputs are now complete 2504 transform_later(region); 2505 _igvn.replace_node(_callprojs->fallthrough_proj, region); 2506 2507 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) ); 2508 mem_phi->init_req(1, memproj ); 2509 mem_phi->init_req(2, mem); 2510 transform_later(mem_phi); 2511 2512 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi); 2513 } 2514 2515 // An inline type might be returned from the call but we don't know its 2516 // type. Either we get a buffered inline type (and nothing needs to be done) 2517 // or one of the values being returned is the klass of the inline type 2518 // and we need to allocate an inline type instance of that type and 2519 // initialize it with other values being returned. In that case, we 2520 // first try a fast path allocation and initialize the value with the 2521 // inline klass's pack handler or we fall back to a runtime call. 2522 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) { 2523 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call"); 2524 Node* ret = call->proj_out_or_null(TypeFunc::Parms); 2525 if (ret == nullptr) { 2526 return; 2527 } 2528 const TypeFunc* tf = call->_tf; 2529 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc(); 2530 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain); 2531 call->_tf = new_tf; 2532 // Make sure the change of type is applied before projections are processed by igvn 2533 _igvn.set_type(call, call->Value(&_igvn)); 2534 _igvn.set_type(ret, ret->Value(&_igvn)); 2535 2536 // Before any new projection is added: 2537 CallProjections* projs = call->extract_projections(true, true); 2538 2539 // Create temporary hook nodes that will be replaced below. 2540 // Add an input to prevent hook nodes from being dead. 2541 Node* ctl = new Node(call); 2542 Node* mem = new Node(ctl); 2543 Node* io = new Node(ctl); 2544 Node* ex_ctl = new Node(ctl); 2545 Node* ex_mem = new Node(ctl); 2546 Node* ex_io = new Node(ctl); 2547 Node* res = new Node(ctl); 2548 2549 // Allocate a new buffered inline type only if a new one is not returned 2550 Node* cast = transform_later(new CastP2XNode(ctl, res)); 2551 Node* mask = MakeConX(0x1); 2552 Node* masked = transform_later(new AndXNode(cast, mask)); 2553 Node* cmp = transform_later(new CmpXNode(masked, mask)); 2554 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq)); 2555 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN); 2556 transform_later(allocation_iff); 2557 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff)); 2558 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff)); 2559 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM)); 2560 2561 // Try to allocate a new buffered inline instance either from TLAB or eden space 2562 Node* needgc_ctrl = nullptr; // needgc means slowcase, i.e. allocation failed 2563 CallLeafNoFPNode* handler_call; 2564 const bool alloc_in_place = UseTLAB; 2565 if (alloc_in_place) { 2566 Node* fast_oop_ctrl = nullptr; 2567 Node* fast_oop_rawmem = nullptr; 2568 Node* mask2 = MakeConX(-2); 2569 Node* masked2 = transform_later(new AndXNode(cast, mask2)); 2570 Node* rawklassptr = transform_later(new CastX2PNode(masked2)); 2571 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeInstKlassPtr::OBJECT_OR_NULL)); 2572 Node* layout_val = make_load(nullptr, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT); 2573 Node* size_in_bytes = ConvI2X(layout_val); 2574 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2575 Node* fast_oop = bs->obj_allocate(this, mem, allocation_ctl, size_in_bytes, io, needgc_ctrl, 2576 fast_oop_ctrl, fast_oop_rawmem, 2577 AllocateInstancePrefetchLines); 2578 // Allocation succeed, initialize buffered inline instance header firstly, 2579 // and then initialize its fields with an inline class specific handler 2580 Node* mark_node = makecon(TypeRawPtr::make((address)markWord::inline_type_prototype().value())); 2581 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 2582 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 2583 if (UseCompressedClassPointers) { 2584 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT); 2585 } 2586 Node* fixed_block = make_load(fast_oop_ctrl, fast_oop_rawmem, klass_node, in_bytes(InstanceKlass::adr_inlineklass_fixed_block_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 2587 Node* pack_handler = make_load(fast_oop_ctrl, fast_oop_rawmem, fixed_block, in_bytes(InlineKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 2588 handler_call = new CallLeafNoFPNode(OptoRuntime::pack_inline_type_Type(), 2589 nullptr, 2590 "pack handler", 2591 TypeRawPtr::BOTTOM); 2592 handler_call->init_req(TypeFunc::Control, fast_oop_ctrl); 2593 handler_call->init_req(TypeFunc::Memory, fast_oop_rawmem); 2594 handler_call->init_req(TypeFunc::I_O, top()); 2595 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr)); 2596 handler_call->init_req(TypeFunc::ReturnAdr, top()); 2597 handler_call->init_req(TypeFunc::Parms, pack_handler); 2598 handler_call->init_req(TypeFunc::Parms+1, fast_oop); 2599 } else { 2600 needgc_ctrl = allocation_ctl; 2601 } 2602 2603 // Allocation failed, fall back to a runtime call 2604 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_inline_type_fields_Type(), 2605 StubRoutines::store_inline_type_fields_to_buf(), 2606 "store_inline_type_fields", 2607 TypePtr::BOTTOM); 2608 slow_call->init_req(TypeFunc::Control, needgc_ctrl); 2609 slow_call->init_req(TypeFunc::Memory, mem); 2610 slow_call->init_req(TypeFunc::I_O, io); 2611 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr)); 2612 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr)); 2613 slow_call->init_req(TypeFunc::Parms, res); 2614 2615 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control)); 2616 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory)); 2617 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O)); 2618 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms)); 2619 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2)); 2620 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci)); 2621 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci)); 2622 2623 Node* ex_r = new RegionNode(3); 2624 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM); 2625 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO); 2626 ex_r->init_req(1, slow_excp); 2627 ex_mem_phi->init_req(1, slow_mem); 2628 ex_io_phi->init_req(1, slow_io); 2629 ex_r->init_req(2, ex_ctl); 2630 ex_mem_phi->init_req(2, ex_mem); 2631 ex_io_phi->init_req(2, ex_io); 2632 transform_later(ex_r); 2633 transform_later(ex_mem_phi); 2634 transform_later(ex_io_phi); 2635 2636 // We don't know how many values are returned. This assumes the 2637 // worst case, that all available registers are used. 2638 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) { 2639 if (domain->field_at(i) == Type::HALF) { 2640 slow_call->init_req(i, top()); 2641 if (alloc_in_place) { 2642 handler_call->init_req(i+1, top()); 2643 } 2644 continue; 2645 } 2646 Node* proj = transform_later(new ProjNode(call, i)); 2647 slow_call->init_req(i, proj); 2648 if (alloc_in_place) { 2649 handler_call->init_req(i+1, proj); 2650 } 2651 } 2652 // We can safepoint at that new call 2653 slow_call->copy_call_debug_info(&_igvn, call); 2654 transform_later(slow_call); 2655 if (alloc_in_place) { 2656 transform_later(handler_call); 2657 } 2658 2659 Node* fast_ctl = nullptr; 2660 Node* fast_res = nullptr; 2661 MergeMemNode* fast_mem = nullptr; 2662 if (alloc_in_place) { 2663 fast_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control)); 2664 Node* rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory)); 2665 fast_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms)); 2666 fast_mem = MergeMemNode::make(mem); 2667 fast_mem->set_memory_at(Compile::AliasIdxRaw, rawmem); 2668 transform_later(fast_mem); 2669 } 2670 2671 Node* r = new RegionNode(alloc_in_place ? 4 : 3); 2672 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM); 2673 Node* io_phi = new PhiNode(r, Type::ABIO); 2674 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM); 2675 r->init_req(1, no_allocation_ctl); 2676 mem_phi->init_req(1, mem); 2677 io_phi->init_req(1, io); 2678 res_phi->init_req(1, no_allocation_res); 2679 r->init_req(2, slow_norm); 2680 mem_phi->init_req(2, slow_mem); 2681 io_phi->init_req(2, slow_io); 2682 res_phi->init_req(2, slow_res); 2683 if (alloc_in_place) { 2684 r->init_req(3, fast_ctl); 2685 mem_phi->init_req(3, fast_mem); 2686 io_phi->init_req(3, io); 2687 res_phi->init_req(3, fast_res); 2688 } 2689 transform_later(r); 2690 transform_later(mem_phi); 2691 transform_later(io_phi); 2692 transform_later(res_phi); 2693 2694 // Do not let stores that initialize this buffer be reordered with a subsequent 2695 // store that would make this buffer accessible by other threads. 2696 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 2697 transform_later(mb); 2698 mb->init_req(TypeFunc::Memory, mem_phi); 2699 mb->init_req(TypeFunc::Control, r); 2700 r = new ProjNode(mb, TypeFunc::Control); 2701 transform_later(r); 2702 mem_phi = new ProjNode(mb, TypeFunc::Memory); 2703 transform_later(mem_phi); 2704 2705 assert(projs->nb_resproj == 1, "unexpected number of results"); 2706 _igvn.replace_in_uses(projs->fallthrough_catchproj, r); 2707 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi); 2708 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi); 2709 _igvn.replace_in_uses(projs->resproj[0], res_phi); 2710 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r); 2711 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi); 2712 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi); 2713 // The CatchNode should not use the ex_io_phi. Re-connect it to the catchall_ioproj. 2714 Node* cn = projs->fallthrough_catchproj->in(0); 2715 _igvn.replace_input_of(cn, 1, projs->catchall_ioproj); 2716 2717 _igvn.replace_node(ctl, projs->fallthrough_catchproj); 2718 _igvn.replace_node(mem, projs->fallthrough_memproj); 2719 _igvn.replace_node(io, projs->fallthrough_ioproj); 2720 _igvn.replace_node(res, projs->resproj[0]); 2721 _igvn.replace_node(ex_ctl, projs->catchall_catchproj); 2722 _igvn.replace_node(ex_mem, projs->catchall_memproj); 2723 _igvn.replace_node(ex_io, projs->catchall_ioproj); 2724 } 2725 2726 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) { 2727 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned"); 2728 Node* bol = check->unique_out(); 2729 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass); 2730 Node* superklass = check->in(SubTypeCheckNode::SuperKlass); 2731 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node"); 2732 2733 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) { 2734 Node* iff = bol->last_out(i); 2735 assert(iff->is_If(), "where's the if?"); 2736 2737 if (iff->in(0)->is_top()) { 2738 _igvn.replace_input_of(iff, 1, C->top()); 2739 continue; 2740 } 2741 2742 Node* iftrue = iff->as_If()->proj_out(1); 2743 Node* iffalse = iff->as_If()->proj_out(0); 2744 Node* ctrl = iff->in(0); 2745 2746 Node* subklass = nullptr; 2747 if (_igvn.type(obj_or_subklass)->isa_klassptr()) { 2748 subklass = obj_or_subklass; 2749 } else { 2750 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes()); 2751 subklass = _igvn.transform(LoadKlassNode::make(_igvn, nullptr, C->immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT)); 2752 } 2753 2754 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci()); 2755 2756 _igvn.replace_input_of(iff, 0, C->top()); 2757 _igvn.replace_node(iftrue, not_subtype_ctrl); 2758 _igvn.replace_node(iffalse, ctrl); 2759 } 2760 _igvn.replace_node(check, C->top()); 2761 } 2762 2763 // FlatArrayCheckNode (array1 array2 ...) is expanded into: 2764 // 2765 // long mark = array1.mark | array2.mark | ...; 2766 // long locked_bit = markWord::unlocked_value & array1.mark & array2.mark & ...; 2767 // if (locked_bit == 0) { 2768 // // One array is locked, load prototype header from the klass 2769 // mark = array1.klass.proto | array2.klass.proto | ... 2770 // } 2771 // if ((mark & markWord::flat_array_bit_in_place) == 0) { 2772 // ... 2773 // } 2774 void PhaseMacroExpand::expand_flatarraycheck_node(FlatArrayCheckNode* check) { 2775 bool array_inputs = _igvn.type(check->in(FlatArrayCheckNode::ArrayOrKlass))->isa_oopptr() != nullptr; 2776 if (UseArrayMarkWordCheck && array_inputs) { 2777 Node* mark = MakeConX(0); 2778 Node* locked_bit = MakeConX(markWord::unlocked_value); 2779 Node* mem = check->in(FlatArrayCheckNode::Memory); 2780 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) { 2781 Node* ary = check->in(i); 2782 const TypeOopPtr* t = _igvn.type(ary)->isa_oopptr(); 2783 assert(t != nullptr, "Mixing array and klass inputs"); 2784 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out"); 2785 Node* mark_adr = basic_plus_adr(ary, oopDesc::mark_offset_in_bytes()); 2786 Node* mark_load = _igvn.transform(LoadNode::make(_igvn, nullptr, mem, mark_adr, mark_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered)); 2787 mark = _igvn.transform(new OrXNode(mark, mark_load)); 2788 locked_bit = _igvn.transform(new AndXNode(locked_bit, mark_load)); 2789 } 2790 assert(!mark->is_Con(), "Should have been optimized out"); 2791 Node* cmp = _igvn.transform(new CmpXNode(locked_bit, MakeConX(0))); 2792 Node* is_unlocked = _igvn.transform(new BoolNode(cmp, BoolTest::ne)); 2793 2794 // BoolNode might be shared, replace each if user 2795 Node* old_bol = check->unique_out(); 2796 assert(old_bol->is_Bool() && old_bol->as_Bool()->_test._test == BoolTest::ne, "unexpected condition"); 2797 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) { 2798 IfNode* old_iff = old_bol->last_out(i)->as_If(); 2799 Node* ctrl = old_iff->in(0); 2800 RegionNode* region = new RegionNode(3); 2801 Node* mark_phi = new PhiNode(region, TypeX_X); 2802 2803 // Check if array is unlocked 2804 IfNode* iff = _igvn.transform(new IfNode(ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If(); 2805 2806 // Unlocked: Use bits from mark word 2807 region->init_req(1, _igvn.transform(new IfTrueNode(iff))); 2808 mark_phi->init_req(1, mark); 2809 2810 // Locked: Load prototype header from klass 2811 ctrl = _igvn.transform(new IfFalseNode(iff)); 2812 Node* proto = MakeConX(0); 2813 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) { 2814 Node* ary = check->in(i); 2815 // Make loads control dependent to make sure they are only executed if array is locked 2816 Node* klass_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes()); 2817 Node* klass = _igvn.transform(LoadKlassNode::make(_igvn, ctrl, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT)); 2818 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset())); 2819 Node* proto_load = _igvn.transform(LoadNode::make(_igvn, ctrl, C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered)); 2820 proto = _igvn.transform(new OrXNode(proto, proto_load)); 2821 } 2822 region->init_req(2, ctrl); 2823 mark_phi->init_req(2, proto); 2824 2825 // Check if flat array bits are set 2826 Node* mask = MakeConX(markWord::flat_array_bit_in_place); 2827 Node* masked = _igvn.transform(new AndXNode(_igvn.transform(mark_phi), mask)); 2828 cmp = _igvn.transform(new CmpXNode(masked, MakeConX(0))); 2829 Node* is_not_flat = _igvn.transform(new BoolNode(cmp, BoolTest::eq)); 2830 2831 ctrl = _igvn.transform(region); 2832 iff = _igvn.transform(new IfNode(ctrl, is_not_flat, PROB_MAX, COUNT_UNKNOWN))->as_If(); 2833 _igvn.replace_node(old_iff, iff); 2834 } 2835 _igvn.replace_node(check, C->top()); 2836 } else { 2837 // Fall back to layout helper check 2838 Node* lhs = intcon(0); 2839 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) { 2840 Node* array_or_klass = check->in(i); 2841 Node* klass = nullptr; 2842 const TypePtr* t = _igvn.type(array_or_klass)->is_ptr(); 2843 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out"); 2844 if (t->isa_oopptr() != nullptr) { 2845 Node* klass_adr = basic_plus_adr(array_or_klass, oopDesc::klass_offset_in_bytes()); 2846 klass = transform_later(LoadKlassNode::make(_igvn, nullptr, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT)); 2847 } else { 2848 assert(t->isa_klassptr(), "Unexpected input type"); 2849 klass = array_or_klass; 2850 } 2851 Node* lh_addr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset())); 2852 Node* lh_val = _igvn.transform(LoadNode::make(_igvn, nullptr, C->immutable_memory(), lh_addr, lh_addr->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered)); 2853 lhs = _igvn.transform(new OrINode(lhs, lh_val)); 2854 } 2855 Node* masked = transform_later(new AndINode(lhs, intcon(Klass::_lh_array_tag_flat_value_bit_inplace))); 2856 Node* cmp = transform_later(new CmpINode(masked, intcon(0))); 2857 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq)); 2858 Node* m2b = transform_later(new Conv2BNode(masked)); 2859 // The matcher expects the input to If nodes to be produced by a Bool(CmpI..) 2860 // pattern, but the input to other potential users (e.g. Phi) to be some 2861 // other pattern (e.g. a Conv2B node, possibly idealized as a CMoveI). 2862 Node* old_bol = check->unique_out(); 2863 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) { 2864 Node* user = old_bol->last_out(i); 2865 for (uint j = 0; j < user->req(); j++) { 2866 Node* n = user->in(j); 2867 if (n == old_bol) { 2868 _igvn.replace_input_of(user, j, user->is_If() ? bol : m2b); 2869 } 2870 } 2871 } 2872 _igvn.replace_node(check, C->top()); 2873 } 2874 } 2875 2876 //---------------------------eliminate_macro_nodes---------------------- 2877 // Eliminate scalar replaced allocations and associated locks. 2878 void PhaseMacroExpand::eliminate_macro_nodes() { 2879 if (C->macro_count() == 0) 2880 return; 2881 NOT_PRODUCT(int membar_before = count_MemBar(C);) 2882 2883 // Before elimination may re-mark (change to Nested or NonEscObj) 2884 // all associated (same box and obj) lock and unlock nodes. 2885 int cnt = C->macro_count(); 2886 for (int i=0; i < cnt; i++) { 2887 Node *n = C->macro_node(i); 2888 if (n->is_AbstractLock()) { // Lock and Unlock nodes 2889 mark_eliminated_locking_nodes(n->as_AbstractLock()); 2890 } 2891 } 2892 // Re-marking may break consistency of Coarsened locks. 2893 if (!C->coarsened_locks_consistent()) { 2894 return; // recompile without Coarsened locks if broken 2895 } 2896 2897 // First, attempt to eliminate locks 2898 bool progress = true; 2899 while (progress) { 2900 progress = false; 2901 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once 2902 Node* n = C->macro_node(i - 1); 2903 bool success = false; 2904 DEBUG_ONLY(int old_macro_count = C->macro_count();) 2905 if (n->is_AbstractLock()) { 2906 success = eliminate_locking_node(n->as_AbstractLock()); 2907 #ifndef PRODUCT 2908 if (success && PrintOptoStatistics) { 2909 Atomic::inc(&PhaseMacroExpand::_monitor_objects_removed_counter); 2910 } 2911 #endif 2912 } 2913 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2914 progress = progress || success; 2915 } 2916 } 2917 // Next, attempt to eliminate allocations 2918 _has_locks = false; 2919 progress = true; 2920 while (progress) { 2921 progress = false; 2922 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once 2923 Node* n = C->macro_node(i - 1); 2924 bool success = false; 2925 DEBUG_ONLY(int old_macro_count = C->macro_count();) 2926 switch (n->class_id()) { 2927 case Node::Class_Allocate: 2928 case Node::Class_AllocateArray: 2929 success = eliminate_allocate_node(n->as_Allocate()); 2930 #ifndef PRODUCT 2931 if (success && PrintOptoStatistics) { 2932 Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter); 2933 } 2934 #endif 2935 break; 2936 case Node::Class_CallStaticJava: { 2937 CallStaticJavaNode* call = n->as_CallStaticJava(); 2938 if (!call->method()->is_method_handle_intrinsic()) { 2939 success = eliminate_boxing_node(n->as_CallStaticJava()); 2940 } 2941 break; 2942 } 2943 case Node::Class_Lock: 2944 case Node::Class_Unlock: 2945 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 2946 _has_locks = true; 2947 break; 2948 case Node::Class_ArrayCopy: 2949 break; 2950 case Node::Class_OuterStripMinedLoop: 2951 break; 2952 case Node::Class_SubTypeCheck: 2953 break; 2954 case Node::Class_Opaque1: 2955 break; 2956 case Node::Class_FlatArrayCheck: 2957 break; 2958 default: 2959 assert(n->Opcode() == Op_LoopLimit || 2960 n->Opcode() == Op_Opaque3 || 2961 n->Opcode() == Op_Opaque4 || 2962 n->Opcode() == Op_MaxL || 2963 n->Opcode() == Op_MinL || 2964 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n), 2965 "unknown node type in macro list"); 2966 } 2967 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2968 progress = progress || success; 2969 } 2970 } 2971 #ifndef PRODUCT 2972 if (PrintOptoStatistics) { 2973 int membar_after = count_MemBar(C); 2974 Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after); 2975 } 2976 #endif 2977 } 2978 2979 //------------------------------expand_macro_nodes---------------------- 2980 // Returns true if a failure occurred. 2981 bool PhaseMacroExpand::expand_macro_nodes() { 2982 // Last attempt to eliminate macro nodes. 2983 eliminate_macro_nodes(); 2984 if (C->failing()) return true; 2985 2986 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. 2987 bool progress = true; 2988 while (progress) { 2989 progress = false; 2990 for (int i = C->macro_count(); i > 0; i--) { 2991 Node* n = C->macro_node(i-1); 2992 bool success = false; 2993 DEBUG_ONLY(int old_macro_count = C->macro_count();) 2994 if (n->Opcode() == Op_LoopLimit) { 2995 // Remove it from macro list and put on IGVN worklist to optimize. 2996 C->remove_macro_node(n); 2997 _igvn._worklist.push(n); 2998 success = true; 2999 } else if (n->Opcode() == Op_CallStaticJava) { 3000 CallStaticJavaNode* call = n->as_CallStaticJava(); 3001 if (!call->method()->is_method_handle_intrinsic()) { 3002 // Remove it from macro list and put on IGVN worklist to optimize. 3003 C->remove_macro_node(n); 3004 _igvn._worklist.push(n); 3005 success = true; 3006 } 3007 } else if (n->is_Opaque1()) { 3008 _igvn.replace_node(n, n->in(1)); 3009 success = true; 3010 #if INCLUDE_RTM_OPT 3011 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) { 3012 assert(C->profile_rtm(), "should be used only in rtm deoptimization code"); 3013 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), ""); 3014 Node* cmp = n->unique_out(); 3015 #ifdef ASSERT 3016 // Validate graph. 3017 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), ""); 3018 BoolNode* bol = cmp->unique_out()->as_Bool(); 3019 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() && 3020 (bol->_test._test == BoolTest::ne), ""); 3021 IfNode* ifn = bol->unique_out()->as_If(); 3022 assert((ifn->outcnt() == 2) && 3023 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != nullptr, ""); 3024 #endif 3025 Node* repl = n->in(1); 3026 if (!_has_locks) { 3027 // Remove RTM state check if there are no locks in the code. 3028 // Replace input to compare the same value. 3029 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1); 3030 } 3031 _igvn.replace_node(n, repl); 3032 success = true; 3033 #endif 3034 } else if (n->Opcode() == Op_Opaque4) { 3035 // With Opaque4 nodes, the expectation is that the test of input 1 3036 // is always equal to the constant value of input 2. So we can 3037 // remove the Opaque4 and replace it by input 2. In debug builds, 3038 // leave the non constant test in instead to sanity check that it 3039 // never fails (if it does, that subgraph was constructed so, at 3040 // runtime, a Halt node is executed). 3041 #ifdef ASSERT 3042 _igvn.replace_node(n, n->in(1)); 3043 #else 3044 _igvn.replace_node(n, n->in(2)); 3045 #endif 3046 success = true; 3047 } else if (n->Opcode() == Op_OuterStripMinedLoop) { 3048 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn); 3049 C->remove_macro_node(n); 3050 success = true; 3051 } else if (n->Opcode() == Op_MaxL) { 3052 // Since MaxL and MinL are not implemented in the backend, we expand them to 3053 // a CMoveL construct now. At least until here, the type could be computed 3054 // precisely. CMoveL is not so smart, but we can give it at least the best 3055 // type we know abouot n now. 3056 Node* repl = MaxNode::signed_max(n->in(1), n->in(2), _igvn.type(n), _igvn); 3057 _igvn.replace_node(n, repl); 3058 success = true; 3059 } else if (n->Opcode() == Op_MinL) { 3060 Node* repl = MaxNode::signed_min(n->in(1), n->in(2), _igvn.type(n), _igvn); 3061 _igvn.replace_node(n, repl); 3062 success = true; 3063 } 3064 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list"); 3065 progress = progress || success; 3066 } 3067 } 3068 3069 // Clean up the graph so we're less likely to hit the maximum node 3070 // limit 3071 _igvn.set_delay_transform(false); 3072 _igvn.optimize(); 3073 if (C->failing()) return true; 3074 _igvn.set_delay_transform(true); 3075 3076 3077 // Because we run IGVN after each expansion, some macro nodes may go 3078 // dead and be removed from the list as we iterate over it. Move 3079 // Allocate nodes (processed in a second pass) at the beginning of 3080 // the list and then iterate from the last element of the list until 3081 // an Allocate node is seen. This is robust to random deletion in 3082 // the list due to nodes going dead. 3083 C->sort_macro_nodes(); 3084 3085 // expand arraycopy "macro" nodes first 3086 // For ReduceBulkZeroing, we must first process all arraycopy nodes 3087 // before the allocate nodes are expanded. 3088 while (C->macro_count() > 0) { 3089 int macro_count = C->macro_count(); 3090 Node * n = C->macro_node(macro_count-1); 3091 assert(n->is_macro(), "only macro nodes expected here"); 3092 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) { 3093 // node is unreachable, so don't try to expand it 3094 C->remove_macro_node(n); 3095 continue; 3096 } 3097 if (n->is_Allocate()) { 3098 break; 3099 } 3100 // Make sure expansion will not cause node limit to be exceeded. 3101 // Worst case is a macro node gets expanded into about 200 nodes. 3102 // Allow 50% more for optimization. 3103 if (C->check_node_count(300, "out of nodes before macro expansion")) { 3104 return true; 3105 } 3106 3107 DEBUG_ONLY(int old_macro_count = C->macro_count();) 3108 switch (n->class_id()) { 3109 case Node::Class_Lock: 3110 expand_lock_node(n->as_Lock()); 3111 break; 3112 case Node::Class_Unlock: 3113 expand_unlock_node(n->as_Unlock()); 3114 break; 3115 case Node::Class_ArrayCopy: 3116 expand_arraycopy_node(n->as_ArrayCopy()); 3117 break; 3118 case Node::Class_SubTypeCheck: 3119 expand_subtypecheck_node(n->as_SubTypeCheck()); 3120 break; 3121 case Node::Class_CallStaticJava: 3122 expand_mh_intrinsic_return(n->as_CallStaticJava()); 3123 C->remove_macro_node(n); 3124 break; 3125 case Node::Class_FlatArrayCheck: 3126 expand_flatarraycheck_node(n->as_FlatArrayCheck()); 3127 break; 3128 default: 3129 assert(false, "unknown node type in macro list"); 3130 } 3131 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list"); 3132 if (C->failing()) return true; 3133 3134 // Clean up the graph so we're less likely to hit the maximum node 3135 // limit 3136 _igvn.set_delay_transform(false); 3137 _igvn.optimize(); 3138 if (C->failing()) return true; 3139 _igvn.set_delay_transform(true); 3140 } 3141 3142 // All nodes except Allocate nodes are expanded now. There could be 3143 // new optimization opportunities (such as folding newly created 3144 // load from a just allocated object). Run IGVN. 3145 3146 // expand "macro" nodes 3147 // nodes are removed from the macro list as they are processed 3148 while (C->macro_count() > 0) { 3149 int macro_count = C->macro_count(); 3150 Node * n = C->macro_node(macro_count-1); 3151 assert(n->is_macro(), "only macro nodes expected here"); 3152 if (_igvn.type(n) == Type::TOP || (n->in(0) != nullptr && n->in(0)->is_top())) { 3153 // node is unreachable, so don't try to expand it 3154 C->remove_macro_node(n); 3155 continue; 3156 } 3157 // Make sure expansion will not cause node limit to be exceeded. 3158 // Worst case is a macro node gets expanded into about 200 nodes. 3159 // Allow 50% more for optimization. 3160 if (C->check_node_count(300, "out of nodes before macro expansion")) { 3161 return true; 3162 } 3163 switch (n->class_id()) { 3164 case Node::Class_Allocate: 3165 expand_allocate(n->as_Allocate()); 3166 break; 3167 case Node::Class_AllocateArray: 3168 expand_allocate_array(n->as_AllocateArray()); 3169 break; 3170 default: 3171 assert(false, "unknown node type in macro list"); 3172 } 3173 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 3174 if (C->failing()) return true; 3175 3176 // Clean up the graph so we're less likely to hit the maximum node 3177 // limit 3178 _igvn.set_delay_transform(false); 3179 _igvn.optimize(); 3180 if (C->failing()) return true; 3181 _igvn.set_delay_transform(true); 3182 } 3183 3184 _igvn.set_delay_transform(false); 3185 return false; 3186 } 3187 3188 #ifndef PRODUCT 3189 int PhaseMacroExpand::_objs_scalar_replaced_counter = 0; 3190 int PhaseMacroExpand::_monitor_objects_removed_counter = 0; 3191 int PhaseMacroExpand::_GC_barriers_removed_counter = 0; 3192 int PhaseMacroExpand::_memory_barriers_removed_counter = 0; 3193 3194 void PhaseMacroExpand::print_statistics() { 3195 tty->print("Objects scalar replaced = %d, ", Atomic::load(&_objs_scalar_replaced_counter)); 3196 tty->print("Monitor objects removed = %d, ", Atomic::load(&_monitor_objects_removed_counter)); 3197 tty->print("GC barriers removed = %d, ", Atomic::load(&_GC_barriers_removed_counter)); 3198 tty->print_cr("Memory barriers removed = %d", Atomic::load(&_memory_barriers_removed_counter)); 3199 } 3200 3201 int PhaseMacroExpand::count_MemBar(Compile *C) { 3202 if (!PrintOptoStatistics) { 3203 return 0; 3204 } 3205 Unique_Node_List ideal_nodes; 3206 int total = 0; 3207 ideal_nodes.map(C->live_nodes(), nullptr); 3208 ideal_nodes.push(C->root()); 3209 for (uint next = 0; next < ideal_nodes.size(); ++next) { 3210 Node* n = ideal_nodes.at(next); 3211 if (n->is_MemBar()) { 3212 total++; 3213 } 3214 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3215 Node* m = n->fast_out(i); 3216 ideal_nodes.push(m); 3217 } 3218 } 3219 return total; 3220 } 3221 #endif