1 /* 2 * Copyright (c) 2005, 2019, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "gc/shared/collectedHeap.inline.hpp" 28 #include "libadt/vectset.hpp" 29 #include "memory/universe.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/arraycopynode.hpp" 32 #include "opto/callnode.hpp" 33 #include "opto/castnode.hpp" 34 #include "opto/cfgnode.hpp" 35 #include "opto/compile.hpp" 36 #include "opto/convertnode.hpp" 37 #include "opto/graphKit.hpp" 38 #include "opto/locknode.hpp" 39 #include "opto/loopnode.hpp" 40 #include "opto/macro.hpp" 41 #include "opto/memnode.hpp" 42 #include "opto/narrowptrnode.hpp" 43 #include "opto/node.hpp" 44 #include "opto/opaquenode.hpp" 45 #include "opto/phaseX.hpp" 46 #include "opto/rootnode.hpp" 47 #include "opto/runtime.hpp" 48 #include "opto/subnode.hpp" 49 #include "opto/type.hpp" 50 #include "opto/valuetypenode.hpp" 51 #include "runtime/sharedRuntime.hpp" 52 #include "utilities/macros.hpp" 53 #if INCLUDE_G1GC 54 #include "gc/g1/g1ThreadLocalData.hpp" 55 #endif // INCLUDE_G1GC 56 #if INCLUDE_SHENANDOAHGC 57 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp" 58 #endif 59 60 61 // 62 // Replace any references to "oldref" in inputs to "use" with "newref". 63 // Returns the number of replacements made. 64 // 65 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { 66 int nreplacements = 0; 67 uint req = use->req(); 68 for (uint j = 0; j < use->len(); j++) { 69 Node *uin = use->in(j); 70 if (uin == oldref) { 71 if (j < req) 72 use->set_req(j, newref); 73 else 74 use->set_prec(j, newref); 75 nreplacements++; 76 } else if (j >= req && uin == NULL) { 77 break; 78 } 79 } 80 return nreplacements; 81 } 82 83 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) { 84 Node* cmp; 85 if (mask != 0) { 86 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask))); 87 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits))); 88 } else { 89 cmp = word; 90 } 91 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne)); 92 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); 93 transform_later(iff); 94 95 // Fast path taken. 96 Node *fast_taken = transform_later(new IfFalseNode(iff)); 97 98 // Fast path not-taken, i.e. slow path 99 Node *slow_taken = transform_later(new IfTrueNode(iff)); 100 101 if (return_fast_path) { 102 region->init_req(edge, slow_taken); // Capture slow-control 103 return fast_taken; 104 } else { 105 region->init_req(edge, fast_taken); // Capture fast-control 106 return slow_taken; 107 } 108 } 109 110 //--------------------copy_predefined_input_for_runtime_call-------------------- 111 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { 112 // Set fixed predefined input arguments 113 call->init_req( TypeFunc::Control, ctrl ); 114 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); 115 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? 116 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); 117 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); 118 } 119 120 //------------------------------make_slow_call--------------------------------- 121 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, 122 address slow_call, const char* leaf_name, Node* slow_path, 123 Node* parm0, Node* parm1, Node* parm2) { 124 125 // Slow-path call 126 CallNode *call = leaf_name 127 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) 128 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM ); 129 130 // Slow path call has no side-effects, uses few values 131 copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); 132 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0); 133 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1); 134 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2); 135 call->copy_call_debug_info(&_igvn, oldcall); 136 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 137 _igvn.replace_node(oldcall, call); 138 transform_later(call); 139 140 return call; 141 } 142 143 void PhaseMacroExpand::extract_call_projections(CallNode *call) { 144 _fallthroughproj = NULL; 145 _fallthroughcatchproj = NULL; 146 _ioproj_fallthrough = NULL; 147 _ioproj_catchall = NULL; 148 _catchallcatchproj = NULL; 149 _memproj_fallthrough = NULL; 150 _memproj_catchall = NULL; 151 _resproj = NULL; 152 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) { 153 ProjNode *pn = call->fast_out(i)->as_Proj(); 154 switch (pn->_con) { 155 case TypeFunc::Control: 156 { 157 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj 158 _fallthroughproj = pn; 159 DUIterator_Fast jmax, j = pn->fast_outs(jmax); 160 const Node *cn = pn->fast_out(j); 161 if (cn->is_Catch()) { 162 ProjNode *cpn = NULL; 163 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { 164 cpn = cn->fast_out(k)->as_Proj(); 165 assert(cpn->is_CatchProj(), "must be a CatchProjNode"); 166 if (cpn->_con == CatchProjNode::fall_through_index) 167 _fallthroughcatchproj = cpn; 168 else { 169 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); 170 _catchallcatchproj = cpn; 171 } 172 } 173 } 174 break; 175 } 176 case TypeFunc::I_O: 177 if (pn->_is_io_use) 178 _ioproj_catchall = pn; 179 else 180 _ioproj_fallthrough = pn; 181 break; 182 case TypeFunc::Memory: 183 if (pn->_is_io_use) 184 _memproj_catchall = pn; 185 else 186 _memproj_fallthrough = pn; 187 break; 188 case TypeFunc::Parms: 189 _resproj = pn; 190 break; 191 default: 192 assert(false, "unexpected projection from allocation node."); 193 } 194 } 195 196 } 197 198 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) { 199 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2(); 200 bs->eliminate_gc_barrier(this, p2x); 201 } 202 203 // Search for a memory operation for the specified memory slice. 204 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) { 205 Node *orig_mem = mem; 206 Node *alloc_mem = alloc->in(TypeFunc::Memory); 207 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr(); 208 while (true) { 209 if (mem == alloc_mem || mem == start_mem ) { 210 return mem; // hit one of our sentinels 211 } else if (mem->is_MergeMem()) { 212 mem = mem->as_MergeMem()->memory_at(alias_idx); 213 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { 214 Node *in = mem->in(0); 215 // we can safely skip over safepoints, calls, locks and membars because we 216 // already know that the object is safe to eliminate. 217 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { 218 return in; 219 } else if (in->is_Call()) { 220 CallNode *call = in->as_Call(); 221 if (call->may_modify(tinst, phase)) { 222 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape"); 223 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) { 224 return in; 225 } 226 } 227 mem = in->in(TypeFunc::Memory); 228 } else if (in->is_MemBar()) { 229 ArrayCopyNode* ac = NULL; 230 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) { 231 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone"); 232 return ac; 233 } 234 mem = in->in(TypeFunc::Memory); 235 } else { 236 assert(false, "unexpected projection"); 237 } 238 } else if (mem->is_Store()) { 239 const TypePtr* atype = mem->as_Store()->adr_type(); 240 int adr_idx = phase->C->get_alias_index(atype); 241 if (adr_idx == alias_idx) { 242 assert(atype->isa_oopptr(), "address type must be oopptr"); 243 int adr_offset = atype->flattened_offset(); 244 uint adr_iid = atype->is_oopptr()->instance_id(); 245 // Array elements references have the same alias_idx 246 // but different offset and different instance_id. 247 if (adr_offset == offset && adr_iid == alloc->_idx) 248 return mem; 249 } else { 250 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw"); 251 } 252 mem = mem->in(MemNode::Memory); 253 } else if (mem->is_ClearArray()) { 254 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) { 255 // Can not bypass initialization of the instance 256 // we are looking. 257 debug_only(intptr_t offset;) 258 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity"); 259 InitializeNode* init = alloc->as_Allocate()->initialization(); 260 // We are looking for stored value, return Initialize node 261 // or memory edge from Allocate node. 262 if (init != NULL) 263 return init; 264 else 265 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers). 266 } 267 // Otherwise skip it (the call updated 'mem' value). 268 } else if (mem->Opcode() == Op_SCMemProj) { 269 mem = mem->in(0); 270 Node* adr = NULL; 271 if (mem->is_LoadStore()) { 272 adr = mem->in(MemNode::Address); 273 } else { 274 assert(mem->Opcode() == Op_EncodeISOArray || 275 mem->Opcode() == Op_StrCompressedCopy, "sanity"); 276 adr = mem->in(3); // Destination array 277 } 278 const TypePtr* atype = adr->bottom_type()->is_ptr(); 279 int adr_idx = phase->C->get_alias_index(atype); 280 if (adr_idx == alias_idx) { 281 DEBUG_ONLY(mem->dump();) 282 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 283 return NULL; 284 } 285 mem = mem->in(MemNode::Memory); 286 } else if (mem->Opcode() == Op_StrInflatedCopy) { 287 Node* adr = mem->in(3); // Destination array 288 const TypePtr* atype = adr->bottom_type()->is_ptr(); 289 int adr_idx = phase->C->get_alias_index(atype); 290 if (adr_idx == alias_idx) { 291 DEBUG_ONLY(mem->dump();) 292 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field"); 293 return NULL; 294 } 295 mem = mem->in(MemNode::Memory); 296 } else { 297 return mem; 298 } 299 assert(mem != orig_mem, "dead memory loop"); 300 } 301 } 302 303 // Generate loads from source of the arraycopy for fields of 304 // destination needed at a deoptimization point 305 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) { 306 BasicType bt = ft; 307 const Type *type = ftype; 308 if (ft == T_NARROWOOP) { 309 bt = T_OBJECT; 310 type = ftype->make_oopptr(); 311 } 312 Node* res = NULL; 313 if (ac->is_clonebasic()) { 314 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination"); 315 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base); 316 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset))); 317 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset); 318 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::UnknownControl); 319 } else { 320 if (ac->modifies(offset, offset, &_igvn, true)) { 321 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result"); 322 uint shift = exact_log2(type2aelembytes(bt)); 323 Node* src_pos = ac->in(ArrayCopyNode::SrcPos); 324 Node* dest_pos = ac->in(ArrayCopyNode::DestPos); 325 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int(); 326 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int(); 327 328 Node* adr = NULL; 329 Node* base = ac->in(ArrayCopyNode::Src); 330 const TypePtr* adr_type = _igvn.type(base)->is_ptr(); 331 assert(adr_type->isa_aryptr(), "only arrays here"); 332 if (src_pos_t->is_con() && dest_pos_t->is_con()) { 333 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset; 334 adr = _igvn.transform(new AddPNode(base, base, MakeConX(off))); 335 adr_type = _igvn.type(adr)->is_ptr(); 336 assert(adr_type == _igvn.type(base)->is_aryptr()->add_field_offset_and_offset(off), "incorrect address type"); 337 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) { 338 // Don't emit a new load from src if src == dst but try to get the value from memory instead 339 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc); 340 } 341 } else { 342 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) { 343 // Non constant offset in the array: we can't statically 344 // determine the value 345 return NULL; 346 } 347 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 348 #ifdef _LP64 349 diff = _igvn.transform(new ConvI2LNode(diff)); 350 #endif 351 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift))); 352 353 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff)); 354 adr = _igvn.transform(new AddPNode(base, base, off)); 355 // In the case of a flattened value type array, each field has its 356 // own slice so we need to extract the field being accessed from 357 // the address computation 358 adr_type = adr_type->is_aryptr()->add_field_offset_and_offset(offset)->add_offset(Type::OffsetBot); 359 adr = _igvn.transform(new CastPPNode(adr, adr_type)); 360 } 361 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::UnknownControl); 362 } 363 } 364 if (res != NULL) { 365 res = _igvn.transform(res); 366 if (ftype->isa_narrowoop()) { 367 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes 368 assert(res->isa_DecodeN(), "should be narrow oop"); 369 res = _igvn.transform(new EncodePNode(res, ftype)); 370 } 371 return res; 372 } 373 return NULL; 374 } 375 376 // 377 // Given a Memory Phi, compute a value Phi containing the values from stores 378 // on the input paths. 379 // Note: this function is recursive, its depth is limited by the "level" argument 380 // Returns the computed Phi, or NULL if it cannot compute it. 381 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) { 382 assert(mem->is_Phi(), "sanity"); 383 int alias_idx = C->get_alias_index(adr_t); 384 int offset = adr_t->flattened_offset(); 385 int instance_id = adr_t->instance_id(); 386 387 // Check if an appropriate value phi already exists. 388 Node* region = mem->in(0); 389 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { 390 Node* phi = region->fast_out(k); 391 if (phi->is_Phi() && phi != mem && 392 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) { 393 return phi; 394 } 395 } 396 // Check if an appropriate new value phi already exists. 397 Node* new_phi = value_phis->find(mem->_idx); 398 if (new_phi != NULL) 399 return new_phi; 400 401 if (level <= 0) { 402 return NULL; // Give up: phi tree too deep 403 } 404 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 405 Node *alloc_mem = alloc->in(TypeFunc::Memory); 406 407 uint length = mem->req(); 408 GrowableArray <Node *> values(length, length, NULL, false); 409 410 // create a new Phi for the value 411 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset); 412 transform_later(phi); 413 value_phis->push(phi, mem->_idx); 414 415 for (uint j = 1; j < length; j++) { 416 Node *in = mem->in(j); 417 if (in == NULL || in->is_top()) { 418 values.at_put(j, in); 419 } else { 420 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); 421 if (val == start_mem || val == alloc_mem) { 422 // hit a sentinel, return appropriate 0 value 423 Node* default_value = alloc->in(AllocateNode::DefaultValue); 424 if (default_value != NULL) { 425 values.at_put(j, default_value); 426 } else { 427 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null"); 428 values.at_put(j, _igvn.zerocon(ft)); 429 } 430 continue; 431 } 432 if (val->is_Initialize()) { 433 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 434 } 435 if (val == NULL) { 436 return NULL; // can't find a value on this path 437 } 438 if (val == mem) { 439 values.at_put(j, mem); 440 } else if (val->is_Store()) { 441 Node* n = val->in(MemNode::ValueIn); 442 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 443 n = bs->step_over_gc_barrier(n); 444 values.at_put(j, n); 445 } else if(val->is_Proj() && val->in(0) == alloc) { 446 Node* default_value = alloc->in(AllocateNode::DefaultValue); 447 if (default_value != NULL) { 448 values.at_put(j, default_value); 449 } else { 450 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null"); 451 values.at_put(j, _igvn.zerocon(ft)); 452 } 453 } else if (val->is_Phi()) { 454 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); 455 if (val == NULL) { 456 return NULL; 457 } 458 values.at_put(j, val); 459 } else if (val->Opcode() == Op_SCMemProj) { 460 assert(val->in(0)->is_LoadStore() || 461 val->in(0)->Opcode() == Op_EncodeISOArray || 462 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity"); 463 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 464 return NULL; 465 } else if (val->is_ArrayCopy()) { 466 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc); 467 if (res == NULL) { 468 return NULL; 469 } 470 values.at_put(j, res); 471 } else { 472 #ifdef ASSERT 473 val->dump(); 474 assert(false, "unknown node on this path"); 475 #endif 476 return NULL; // unknown node on this path 477 } 478 } 479 } 480 // Set Phi's inputs 481 for (uint j = 1; j < length; j++) { 482 if (values.at(j) == mem) { 483 phi->init_req(j, phi); 484 } else { 485 phi->init_req(j, values.at(j)); 486 } 487 } 488 return phi; 489 } 490 491 // Search the last value stored into the object's field. 492 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) { 493 assert(adr_t->is_known_instance_field(), "instance required"); 494 int instance_id = adr_t->instance_id(); 495 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 496 497 int alias_idx = C->get_alias_index(adr_t); 498 int offset = adr_t->flattened_offset(); 499 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 500 Node *alloc_mem = alloc->in(TypeFunc::Memory); 501 Arena *a = Thread::current()->resource_area(); 502 VectorSet visited(a); 503 504 bool done = sfpt_mem == alloc_mem; 505 Node *mem = sfpt_mem; 506 while (!done) { 507 if (visited.test_set(mem->_idx)) { 508 return NULL; // found a loop, give up 509 } 510 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 511 if (mem == start_mem || mem == alloc_mem) { 512 done = true; // hit a sentinel, return appropriate 0 value 513 } else if (mem->is_Initialize()) { 514 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 515 if (mem == NULL) { 516 done = true; // Something went wrong. 517 } else if (mem->is_Store()) { 518 const TypePtr* atype = mem->as_Store()->adr_type(); 519 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 520 done = true; 521 } 522 } else if (mem->is_Store()) { 523 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 524 assert(atype != NULL, "address type must be oopptr"); 525 assert(C->get_alias_index(atype) == alias_idx && 526 atype->is_known_instance_field() && atype->flattened_offset() == offset && 527 atype->instance_id() == instance_id, "store is correct memory slice"); 528 done = true; 529 } else if (mem->is_Phi()) { 530 // try to find a phi's unique input 531 Node *unique_input = NULL; 532 Node *top = C->top(); 533 for (uint i = 1; i < mem->req(); i++) { 534 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 535 if (n == NULL || n == top || n == mem) { 536 continue; 537 } else if (unique_input == NULL) { 538 unique_input = n; 539 } else if (unique_input != n) { 540 unique_input = top; 541 break; 542 } 543 } 544 if (unique_input != NULL && unique_input != top) { 545 mem = unique_input; 546 } else { 547 done = true; 548 } 549 } else if (mem->is_ArrayCopy()) { 550 done = true; 551 } else { 552 assert(false, "unexpected node"); 553 } 554 } 555 if (mem != NULL) { 556 if (mem == start_mem || mem == alloc_mem) { 557 // hit a sentinel, return appropriate 0 value 558 Node* default_value = alloc->in(AllocateNode::DefaultValue); 559 if (default_value != NULL) { 560 return default_value; 561 } 562 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null"); 563 return _igvn.zerocon(ft); 564 } else if (mem->is_Store()) { 565 Node* n = mem->in(MemNode::ValueIn); 566 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 567 n = bs->step_over_gc_barrier(n); 568 return n; 569 } else if (mem->is_Phi()) { 570 // attempt to produce a Phi reflecting the values on the input paths of the Phi 571 Node_Stack value_phis(a, 8); 572 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 573 if (phi != NULL) { 574 return phi; 575 } else { 576 // Kill all new Phis 577 while(value_phis.is_nonempty()) { 578 Node* n = value_phis.node(); 579 _igvn.replace_node(n, C->top()); 580 value_phis.pop(); 581 } 582 } 583 } else if (mem->is_ArrayCopy()) { 584 Node* ctl = mem->in(0); 585 Node* m = mem->in(TypeFunc::Memory); 586 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) { 587 // pin the loads in the uncommon trap path 588 ctl = sfpt_ctl; 589 m = sfpt_mem; 590 } 591 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc); 592 } 593 } 594 // Something went wrong. 595 return NULL; 596 } 597 598 // Search the last value stored into the value type's fields. 599 Node* PhaseMacroExpand::value_type_from_mem(Node* mem, Node* ctl, ciValueKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) { 600 // Subtract the offset of the first field to account for the missing oop header 601 offset -= vk->first_field_offset(); 602 // Create a new ValueTypeNode and retrieve the field values from memory 603 ValueTypeNode* vt = ValueTypeNode::make_uninitialized(_igvn, vk)->as_ValueType(); 604 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) { 605 ciType* field_type = vt->field_type(i); 606 int field_offset = offset + vt->field_offset(i); 607 // Each value type field has its own memory slice 608 adr_type = adr_type->with_field_offset(field_offset); 609 Node* value = NULL; 610 if (vt->field_is_flattened(i)) { 611 value = value_type_from_mem(mem, ctl, field_type->as_value_klass(), adr_type, field_offset, alloc); 612 } else { 613 const Type* ft = Type::get_const_type(field_type); 614 BasicType bt = field_type->basic_type(); 615 if (UseCompressedOops && !is_java_primitive(bt)) { 616 ft = ft->make_narrowoop(); 617 bt = T_NARROWOOP; 618 } 619 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc); 620 if (value != NULL && ft->isa_narrowoop()) { 621 assert(UseCompressedOops, "unexpected narrow oop"); 622 value = transform_later(new DecodeNNode(value, value->get_ptr_type())); 623 } 624 } 625 if (value != NULL) { 626 vt->set_field_value(i, value); 627 } else { 628 // We might have reached the TrackedInitializationLimit 629 return NULL; 630 } 631 } 632 return transform_later(vt); 633 } 634 635 // Check the possibility of scalar replacement. 636 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 637 // Scan the uses of the allocation to check for anything that would 638 // prevent us from eliminating it. 639 NOT_PRODUCT( const char* fail_eliminate = NULL; ) 640 DEBUG_ONLY( Node* disq_node = NULL; ) 641 bool can_eliminate = true; 642 643 Node* res = alloc->result_cast(); 644 const TypeOopPtr* res_type = NULL; 645 if (res == NULL) { 646 // All users were eliminated. 647 } else if (!res->is_CheckCastPP()) { 648 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 649 can_eliminate = false; 650 } else { 651 res_type = _igvn.type(res)->isa_oopptr(); 652 if (res_type == NULL) { 653 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 654 can_eliminate = false; 655 } else if (res_type->isa_aryptr()) { 656 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 657 if (length < 0) { 658 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 659 can_eliminate = false; 660 } 661 } 662 } 663 664 if (can_eliminate && res != NULL) { 665 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 666 j < jmax && can_eliminate; j++) { 667 Node* use = res->fast_out(j); 668 669 if (use->is_AddP()) { 670 const TypePtr* addp_type = _igvn.type(use)->is_ptr(); 671 int offset = addp_type->offset(); 672 673 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 674 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) 675 can_eliminate = false; 676 break; 677 } 678 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 679 k < kmax && can_eliminate; k++) { 680 Node* n = use->fast_out(k); 681 if (!n->is_Store() && n->Opcode() != Op_CastP2X && 682 SHENANDOAHGC_ONLY((!UseShenandoahGC || !ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(n)) &&) 683 !(n->is_ArrayCopy() && 684 n->as_ArrayCopy()->is_clonebasic() && 685 n->in(ArrayCopyNode::Dest) == use)) { 686 DEBUG_ONLY(disq_node = n;) 687 if (n->is_Load() || n->is_LoadStore()) { 688 NOT_PRODUCT(fail_eliminate = "Field load";) 689 } else { 690 NOT_PRODUCT(fail_eliminate = "Not store field reference";) 691 } 692 can_eliminate = false; 693 } 694 } 695 } else if (use->is_ArrayCopy() && 696 (use->as_ArrayCopy()->is_arraycopy_validated() || 697 use->as_ArrayCopy()->is_copyof_validated() || 698 use->as_ArrayCopy()->is_copyofrange_validated()) && 699 use->in(ArrayCopyNode::Dest) == res) { 700 // ok to eliminate 701 } else if (use->is_SafePoint()) { 702 SafePointNode* sfpt = use->as_SafePoint(); 703 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 704 // Object is passed as argument. 705 DEBUG_ONLY(disq_node = use;) 706 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 707 can_eliminate = false; 708 } 709 Node* sfptMem = sfpt->memory(); 710 if (sfptMem == NULL || sfptMem->is_top()) { 711 DEBUG_ONLY(disq_node = use;) 712 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) 713 can_eliminate = false; 714 } else { 715 safepoints.append_if_missing(sfpt); 716 } 717 } else if (use->is_ValueType() && use->isa_ValueType()->get_oop() == res) { 718 // ok to eliminate 719 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) { 720 // store to mark work 721 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 722 if (use->is_Phi()) { 723 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 724 NOT_PRODUCT(fail_eliminate = "Object is return value";) 725 } else { 726 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 727 } 728 DEBUG_ONLY(disq_node = use;) 729 } else { 730 if (use->Opcode() == Op_Return) { 731 NOT_PRODUCT(fail_eliminate = "Object is return value";) 732 } else { 733 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 734 } 735 DEBUG_ONLY(disq_node = use;) 736 } 737 can_eliminate = false; 738 } else { 739 assert(use->Opcode() == Op_CastP2X, "should be"); 740 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null"); 741 } 742 } 743 } 744 745 #ifndef PRODUCT 746 if (PrintEliminateAllocations) { 747 if (can_eliminate) { 748 tty->print("Scalar "); 749 if (res == NULL) 750 alloc->dump(); 751 else 752 res->dump(); 753 } else if (alloc->_is_scalar_replaceable) { 754 tty->print("NotScalar (%s)", fail_eliminate); 755 if (res == NULL) 756 alloc->dump(); 757 else 758 res->dump(); 759 #ifdef ASSERT 760 if (disq_node != NULL) { 761 tty->print(" >>>> "); 762 disq_node->dump(); 763 } 764 #endif /*ASSERT*/ 765 } 766 } 767 #endif 768 return can_eliminate; 769 } 770 771 // Do scalar replacement. 772 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 773 GrowableArray <SafePointNode *> safepoints_done; 774 775 ciKlass* klass = NULL; 776 ciInstanceKlass* iklass = NULL; 777 int nfields = 0; 778 int array_base = 0; 779 int element_size = 0; 780 BasicType basic_elem_type = T_ILLEGAL; 781 ciType* elem_type = NULL; 782 783 Node* res = alloc->result_cast(); 784 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result"); 785 const TypeOopPtr* res_type = NULL; 786 if (res != NULL) { // Could be NULL when there are no users 787 res_type = _igvn.type(res)->isa_oopptr(); 788 } 789 790 if (res != NULL) { 791 klass = res_type->klass(); 792 if (res_type->isa_instptr()) { 793 // find the fields of the class which will be needed for safepoint debug information 794 assert(klass->is_instance_klass(), "must be an instance klass."); 795 iklass = klass->as_instance_klass(); 796 nfields = iklass->nof_nonstatic_fields(); 797 } else { 798 // find the array's elements which will be needed for safepoint debug information 799 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 800 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); 801 elem_type = klass->as_array_klass()->element_type(); 802 basic_elem_type = elem_type->basic_type(); 803 if (elem_type->is_valuetype() && !klass->is_value_array_klass()) { 804 assert(basic_elem_type == T_VALUETYPE, "unexpected element basic type"); 805 basic_elem_type = T_OBJECT; 806 } 807 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 808 element_size = type2aelembytes(basic_elem_type); 809 if (klass->is_value_array_klass()) { 810 // Flattened value type array 811 element_size = klass->as_value_array_klass()->element_byte_size(); 812 } 813 } 814 } 815 // 816 // Process the safepoint uses 817 // 818 Unique_Node_List value_worklist; 819 while (safepoints.length() > 0) { 820 SafePointNode* sfpt = safepoints.pop(); 821 Node* mem = sfpt->memory(); 822 Node* ctl = sfpt->control(); 823 assert(sfpt->jvms() != NULL, "missed JVMS"); 824 // Fields of scalar objs are referenced only at the end 825 // of regular debuginfo at the last (youngest) JVMS. 826 // Record relative start index. 827 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff()); 828 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, 829 #ifdef ASSERT 830 alloc, 831 #endif 832 first_ind, nfields); 833 sobj->init_req(0, C->root()); 834 transform_later(sobj); 835 836 // Scan object's fields adding an input to the safepoint for each field. 837 for (int j = 0; j < nfields; j++) { 838 intptr_t offset; 839 ciField* field = NULL; 840 if (iklass != NULL) { 841 field = iklass->nonstatic_field_at(j); 842 offset = field->offset(); 843 elem_type = field->type(); 844 basic_elem_type = field->layout_type(); 845 assert(!field->is_flattened(), "flattened value type fields should not have safepoint uses"); 846 } else { 847 offset = array_base + j * (intptr_t)element_size; 848 } 849 850 const Type *field_type; 851 // The next code is taken from Parse::do_get_xxx(). 852 if (is_reference_type(basic_elem_type)) { 853 if (!elem_type->is_loaded()) { 854 field_type = TypeInstPtr::BOTTOM; 855 } else if (field != NULL && field->is_static_constant()) { 856 // This can happen if the constant oop is non-perm. 857 ciObject* con = field->constant_value().as_object(); 858 // Do not "join" in the previous type; it doesn't add value, 859 // and may yield a vacuous result if the field is of interface type. 860 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 861 assert(field_type != NULL, "field singleton type must be consistent"); 862 } else { 863 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 864 } 865 if (UseCompressedOops) { 866 field_type = field_type->make_narrowoop(); 867 basic_elem_type = T_NARROWOOP; 868 } 869 } else { 870 field_type = Type::get_const_basic_type(basic_elem_type); 871 } 872 873 Node* field_val = NULL; 874 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 875 if (klass->is_value_array_klass()) { 876 ciValueKlass* vk = elem_type->as_value_klass(); 877 assert(vk->flatten_array(), "must be flattened"); 878 field_val = value_type_from_mem(mem, ctl, vk, field_addr_type->isa_aryptr(), 0, alloc); 879 } else { 880 field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc); 881 } 882 if (field_val == NULL) { 883 // We weren't able to find a value for this field, 884 // give up on eliminating this allocation. 885 886 // Remove any extra entries we added to the safepoint. 887 uint last = sfpt->req() - 1; 888 for (int k = 0; k < j; k++) { 889 sfpt->del_req(last--); 890 } 891 _igvn._worklist.push(sfpt); 892 // rollback processed safepoints 893 while (safepoints_done.length() > 0) { 894 SafePointNode* sfpt_done = safepoints_done.pop(); 895 // remove any extra entries we added to the safepoint 896 last = sfpt_done->req() - 1; 897 for (int k = 0; k < nfields; k++) { 898 sfpt_done->del_req(last--); 899 } 900 JVMState *jvms = sfpt_done->jvms(); 901 jvms->set_endoff(sfpt_done->req()); 902 // Now make a pass over the debug information replacing any references 903 // to SafePointScalarObjectNode with the allocated object. 904 int start = jvms->debug_start(); 905 int end = jvms->debug_end(); 906 for (int i = start; i < end; i++) { 907 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 908 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 909 if (scobj->first_index(jvms) == sfpt_done->req() && 910 scobj->n_fields() == (uint)nfields) { 911 assert(scobj->alloc() == alloc, "sanity"); 912 sfpt_done->set_req(i, res); 913 } 914 } 915 } 916 _igvn._worklist.push(sfpt_done); 917 } 918 #ifndef PRODUCT 919 if (PrintEliminateAllocations) { 920 if (field != NULL) { 921 tty->print("=== At SafePoint node %d can't find value of Field: ", 922 sfpt->_idx); 923 field->print(); 924 int field_idx = C->get_alias_index(field_addr_type); 925 tty->print(" (alias_idx=%d)", field_idx); 926 } else { // Array's element 927 tty->print("=== At SafePoint node %d can't find value of array element [%d]", 928 sfpt->_idx, j); 929 } 930 tty->print(", which prevents elimination of: "); 931 if (res == NULL) 932 alloc->dump(); 933 else 934 res->dump(); 935 } 936 #endif 937 return false; 938 } 939 if (field_val->is_ValueType()) { 940 // Keep track of value types to scalarize them later 941 value_worklist.push(field_val); 942 } else if (UseCompressedOops && field_type->isa_narrowoop()) { 943 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 944 // to be able scalar replace the allocation. 945 if (field_val->is_EncodeP()) { 946 field_val = field_val->in(1); 947 } else { 948 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type())); 949 } 950 } 951 sfpt->add_req(field_val); 952 } 953 JVMState *jvms = sfpt->jvms(); 954 jvms->set_endoff(sfpt->req()); 955 // Now make a pass over the debug information replacing any references 956 // to the allocated object with "sobj" 957 int start = jvms->debug_start(); 958 int end = jvms->debug_end(); 959 sfpt->replace_edges_in_range(res, sobj, start, end); 960 _igvn._worklist.push(sfpt); 961 safepoints_done.append_if_missing(sfpt); // keep it for rollback 962 } 963 // Scalarize value types that were added to the safepoint 964 for (uint i = 0; i < value_worklist.size(); ++i) { 965 Node* vt = value_worklist.at(i); 966 vt->as_ValueType()->make_scalar_in_safepoints(&_igvn); 967 } 968 return true; 969 } 970 971 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) { 972 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control); 973 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory); 974 if (ctl_proj != NULL) { 975 igvn.replace_node(ctl_proj, n->in(0)); 976 } 977 if (mem_proj != NULL) { 978 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory)); 979 } 980 } 981 982 // Process users of eliminated allocation. 983 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) { 984 Node* res = alloc->result_cast(); 985 if (res != NULL) { 986 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 987 Node *use = res->last_out(j); 988 uint oc1 = res->outcnt(); 989 990 if (use->is_AddP()) { 991 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 992 Node *n = use->last_out(k); 993 uint oc2 = use->outcnt(); 994 if (n->is_Store()) { 995 #ifdef ASSERT 996 // Verify that there is no dependent MemBarVolatile nodes, 997 // they should be removed during IGVN, see MemBarNode::Ideal(). 998 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); 999 p < pmax; p++) { 1000 Node* mb = n->fast_out(p); 1001 assert(mb->is_Initialize() || !mb->is_MemBar() || 1002 mb->req() <= MemBarNode::Precedent || 1003 mb->in(MemBarNode::Precedent) != n, 1004 "MemBarVolatile should be eliminated for non-escaping object"); 1005 } 1006 #endif 1007 _igvn.replace_node(n, n->in(MemNode::Memory)); 1008 } else if (n->is_ArrayCopy()) { 1009 // Disconnect ArrayCopy node 1010 ArrayCopyNode* ac = n->as_ArrayCopy(); 1011 assert(ac->is_clonebasic(), "unexpected array copy kind"); 1012 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out(); 1013 disconnect_projections(ac, _igvn); 1014 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation"); 1015 Node* membar_before = alloc->in(0)->in(0); 1016 disconnect_projections(membar_before->as_MemBar(), _igvn); 1017 if (membar_after->is_MemBar()) { 1018 disconnect_projections(membar_after->as_MemBar(), _igvn); 1019 } 1020 } else { 1021 eliminate_gc_barrier(n); 1022 } 1023 k -= (oc2 - use->outcnt()); 1024 } 1025 _igvn.remove_dead_node(use); 1026 } else if (use->is_ArrayCopy()) { 1027 // Disconnect ArrayCopy node 1028 ArrayCopyNode* ac = use->as_ArrayCopy(); 1029 assert(ac->is_arraycopy_validated() || 1030 ac->is_copyof_validated() || 1031 ac->is_copyofrange_validated(), "unsupported"); 1032 CallProjections* callprojs = ac->extract_projections(true); 1033 1034 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O)); 1035 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory)); 1036 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control)); 1037 1038 // Set control to top. IGVN will remove the remaining projections 1039 ac->set_req(0, top()); 1040 ac->replace_edge(res, top()); 1041 1042 // Disconnect src right away: it can help find new 1043 // opportunities for allocation elimination 1044 Node* src = ac->in(ArrayCopyNode::Src); 1045 ac->replace_edge(src, top()); 1046 // src can be top at this point if src and dest of the 1047 // arraycopy were the same 1048 if (src->outcnt() == 0 && !src->is_top()) { 1049 _igvn.remove_dead_node(src); 1050 } 1051 1052 _igvn._worklist.push(ac); 1053 } else if (use->is_ValueType()) { 1054 assert(use->isa_ValueType()->get_oop() == res, "unexpected value type use"); 1055 _igvn.rehash_node_delayed(use); 1056 use->isa_ValueType()->set_oop(_igvn.zerocon(T_VALUETYPE)); 1057 } else if (use->is_Store()) { 1058 _igvn.replace_node(use, use->in(MemNode::Memory)); 1059 } else { 1060 eliminate_gc_barrier(use); 1061 } 1062 j -= (oc1 - res->outcnt()); 1063 } 1064 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 1065 _igvn.remove_dead_node(res); 1066 } 1067 1068 // 1069 // Process other users of allocation's projections 1070 // 1071 if (_resproj != NULL && _resproj->outcnt() != 0) { 1072 // First disconnect stores captured by Initialize node. 1073 // If Initialize node is eliminated first in the following code, 1074 // it will kill such stores and DUIterator_Last will assert. 1075 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) { 1076 Node *use = _resproj->fast_out(j); 1077 if (use->is_AddP()) { 1078 // raw memory addresses used only by the initialization 1079 _igvn.replace_node(use, C->top()); 1080 --j; --jmax; 1081 } 1082 } 1083 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { 1084 Node *use = _resproj->last_out(j); 1085 uint oc1 = _resproj->outcnt(); 1086 if (use->is_Initialize()) { 1087 // Eliminate Initialize node. 1088 InitializeNode *init = use->as_Initialize(); 1089 assert(init->outcnt() <= 2, "only a control and memory projection expected"); 1090 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control); 1091 if (ctrl_proj != NULL) { 1092 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control)); 1093 #ifdef ASSERT 1094 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1095 Node* tmp = init->in(TypeFunc::Control); 1096 while (bs->is_gc_barrier_node(tmp)) { 1097 Node* tmp2 = bs->step_over_gc_barrier_ctrl(tmp); 1098 assert(tmp != tmp2, "Must make progress"); 1099 tmp = tmp2; 1100 } 1101 assert(tmp == _fallthroughcatchproj, "allocation control projection"); 1102 #endif 1103 } 1104 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory); 1105 if (mem_proj != NULL) { 1106 Node *mem = init->in(TypeFunc::Memory); 1107 #ifdef ASSERT 1108 if (mem->is_MergeMem()) { 1109 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection"); 1110 } else { 1111 assert(mem == _memproj_fallthrough, "allocation memory projection"); 1112 } 1113 #endif 1114 _igvn.replace_node(mem_proj, mem); 1115 } 1116 } else { 1117 assert(false, "only Initialize or AddP expected"); 1118 } 1119 j -= (oc1 - _resproj->outcnt()); 1120 } 1121 } 1122 if (_fallthroughcatchproj != NULL) { 1123 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); 1124 } 1125 if (_memproj_fallthrough != NULL) { 1126 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); 1127 } 1128 if (_memproj_catchall != NULL) { 1129 _igvn.replace_node(_memproj_catchall, C->top()); 1130 } 1131 if (_ioproj_fallthrough != NULL) { 1132 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); 1133 } 1134 if (_ioproj_catchall != NULL) { 1135 _igvn.replace_node(_ioproj_catchall, C->top()); 1136 } 1137 if (_catchallcatchproj != NULL) { 1138 _igvn.replace_node(_catchallcatchproj, C->top()); 1139 } 1140 } 1141 1142 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 1143 // Don't do scalar replacement if the frame can be popped by JVMTI: 1144 // if reallocation fails during deoptimization we'll pop all 1145 // interpreter frames for this compiled frame and that won't play 1146 // nice with JVMTI popframe. 1147 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) { 1148 return false; 1149 } 1150 Node* klass = alloc->in(AllocateNode::KlassNode); 1151 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); 1152 Node* res = alloc->result_cast(); 1153 // Eliminate boxing allocations which are not used 1154 // regardless scalar replacable status. 1155 bool boxing_alloc = C->eliminate_boxing() && 1156 tklass->klass()->is_instance_klass() && 1157 tklass->klass()->as_instance_klass()->is_box_klass(); 1158 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) { 1159 return false; 1160 } 1161 1162 extract_call_projections(alloc); 1163 1164 GrowableArray <SafePointNode *> safepoints; 1165 if (!can_eliminate_allocation(alloc, safepoints)) { 1166 return false; 1167 } 1168 1169 if (!alloc->_is_scalar_replaceable) { 1170 assert(res == NULL, "sanity"); 1171 // We can only eliminate allocation if all debug info references 1172 // are already replaced with SafePointScalarObject because 1173 // we can't search for a fields value without instance_id. 1174 if (safepoints.length() > 0) { 1175 return false; 1176 } 1177 } 1178 1179 if (!scalar_replacement(alloc, safepoints)) { 1180 return false; 1181 } 1182 1183 CompileLog* log = C->log(); 1184 if (log != NULL) { 1185 log->head("eliminate_allocation type='%d'", 1186 log->identify(tklass->klass())); 1187 JVMState* p = alloc->jvms(); 1188 while (p != NULL) { 1189 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1190 p = p->caller(); 1191 } 1192 log->tail("eliminate_allocation"); 1193 } 1194 1195 process_users_of_allocation(alloc); 1196 1197 #ifndef PRODUCT 1198 if (PrintEliminateAllocations) { 1199 if (alloc->is_AllocateArray()) 1200 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1201 else 1202 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1203 } 1204 #endif 1205 1206 return true; 1207 } 1208 1209 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) { 1210 // EA should remove all uses of non-escaping boxing node. 1211 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) { 1212 return false; 1213 } 1214 1215 assert(boxing->result_cast() == NULL, "unexpected boxing node result"); 1216 1217 extract_call_projections(boxing); 1218 1219 const TypeTuple* r = boxing->tf()->range_sig(); 1220 assert(r->cnt() > TypeFunc::Parms, "sanity"); 1221 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr(); 1222 assert(t != NULL, "sanity"); 1223 1224 CompileLog* log = C->log(); 1225 if (log != NULL) { 1226 log->head("eliminate_boxing type='%d'", 1227 log->identify(t->klass())); 1228 JVMState* p = boxing->jvms(); 1229 while (p != NULL) { 1230 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1231 p = p->caller(); 1232 } 1233 log->tail("eliminate_boxing"); 1234 } 1235 1236 process_users_of_allocation(boxing); 1237 1238 #ifndef PRODUCT 1239 if (PrintEliminateAllocations) { 1240 tty->print("++++ Eliminated: %d ", boxing->_idx); 1241 boxing->method()->print_short_name(tty); 1242 tty->cr(); 1243 } 1244 #endif 1245 1246 return true; 1247 } 1248 1249 //---------------------------set_eden_pointers------------------------- 1250 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { 1251 if (UseTLAB) { // Private allocation: load from TLS 1252 Node* thread = transform_later(new ThreadLocalNode()); 1253 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); 1254 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); 1255 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); 1256 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); 1257 } else { // Shared allocation: load from globals 1258 CollectedHeap* ch = Universe::heap(); 1259 address top_adr = (address)ch->top_addr(); 1260 address end_adr = (address)ch->end_addr(); 1261 eden_top_adr = makecon(TypeRawPtr::make(top_adr)); 1262 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); 1263 } 1264 } 1265 1266 1267 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 1268 Node* adr = basic_plus_adr(base, offset); 1269 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 1270 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered); 1271 transform_later(value); 1272 return value; 1273 } 1274 1275 1276 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 1277 Node* adr = basic_plus_adr(base, offset); 1278 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered); 1279 transform_later(mem); 1280 return mem; 1281 } 1282 1283 //============================================================================= 1284 // 1285 // A L L O C A T I O N 1286 // 1287 // Allocation attempts to be fast in the case of frequent small objects. 1288 // It breaks down like this: 1289 // 1290 // 1) Size in doublewords is computed. This is a constant for objects and 1291 // variable for most arrays. Doubleword units are used to avoid size 1292 // overflow of huge doubleword arrays. We need doublewords in the end for 1293 // rounding. 1294 // 1295 // 2) Size is checked for being 'too large'. Too-large allocations will go 1296 // the slow path into the VM. The slow path can throw any required 1297 // exceptions, and does all the special checks for very large arrays. The 1298 // size test can constant-fold away for objects. For objects with 1299 // finalizers it constant-folds the otherway: you always go slow with 1300 // finalizers. 1301 // 1302 // 3) If NOT using TLABs, this is the contended loop-back point. 1303 // Load-Locked the heap top. If using TLABs normal-load the heap top. 1304 // 1305 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 1306 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 1307 // "size*8" we always enter the VM, where "largish" is a constant picked small 1308 // enough that there's always space between the eden max and 4Gig (old space is 1309 // there so it's quite large) and large enough that the cost of entering the VM 1310 // is dwarfed by the cost to initialize the space. 1311 // 1312 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 1313 // down. If contended, repeat at step 3. If using TLABs normal-store 1314 // adjusted heap top back down; there is no contention. 1315 // 1316 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 1317 // fields. 1318 // 1319 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 1320 // oop flavor. 1321 // 1322 //============================================================================= 1323 // FastAllocateSizeLimit value is in DOUBLEWORDS. 1324 // Allocations bigger than this always go the slow route. 1325 // This value must be small enough that allocation attempts that need to 1326 // trigger exceptions go the slow route. Also, it must be small enough so 1327 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 1328 //=============================================================================j// 1329 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 1330 // The allocator will coalesce int->oop copies away. See comment in 1331 // coalesce.cpp about how this works. It depends critically on the exact 1332 // code shape produced here, so if you are changing this code shape 1333 // make sure the GC info for the heap-top is correct in and around the 1334 // slow-path call. 1335 // 1336 1337 void PhaseMacroExpand::expand_allocate_common( 1338 AllocateNode* alloc, // allocation node to be expanded 1339 Node* length, // array length for an array allocation 1340 const TypeFunc* slow_call_type, // Type of slow call 1341 address slow_call_address // Address of slow call 1342 ) 1343 { 1344 1345 Node* ctrl = alloc->in(TypeFunc::Control); 1346 Node* mem = alloc->in(TypeFunc::Memory); 1347 Node* i_o = alloc->in(TypeFunc::I_O); 1348 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 1349 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1350 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 1351 1352 assert(ctrl != NULL, "must have control"); 1353 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 1354 // they will not be used if "always_slow" is set 1355 enum { slow_result_path = 1, fast_result_path = 2 }; 1356 Node *result_region = NULL; 1357 Node *result_phi_rawmem = NULL; 1358 Node *result_phi_rawoop = NULL; 1359 Node *result_phi_i_o = NULL; 1360 1361 // The initial slow comparison is a size check, the comparison 1362 // we want to do is a BoolTest::gt 1363 bool always_slow = false; 1364 int tv = _igvn.find_int_con(initial_slow_test, -1); 1365 if (tv >= 0) { 1366 always_slow = (tv == 1); 1367 initial_slow_test = NULL; 1368 } else { 1369 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 1370 } 1371 1372 if (C->env()->dtrace_alloc_probes() || 1373 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) { 1374 // Force slow-path allocation 1375 always_slow = true; 1376 initial_slow_test = NULL; 1377 } 1378 1379 Node *slow_region = NULL; 1380 Node *toobig_false = ctrl; 1381 1382 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); 1383 // generate the initial test if necessary 1384 if (initial_slow_test != NULL ) { 1385 if (slow_region == NULL) { 1386 slow_region = new RegionNode(1); 1387 } 1388 // Now make the initial failure test. Usually a too-big test but 1389 // might be a TRUE for finalizers or a fancy class check for 1390 // newInstance0. 1391 IfNode* toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1392 transform_later(toobig_iff); 1393 // Plug the failing-too-big test into the slow-path region 1394 Node* toobig_true = new IfTrueNode(toobig_iff); 1395 transform_later(toobig_true); 1396 slow_region ->add_req(toobig_true); 1397 toobig_false = new IfFalseNode(toobig_iff); 1398 transform_later(toobig_false); 1399 } else { // No initial test, just fall into next case 1400 toobig_false = ctrl; 1401 } 1402 1403 Node *slow_mem = mem; // save the current memory state for slow path 1404 // generate the fast allocation code unless we know that the initial test will always go slow 1405 if (!always_slow) { 1406 // Fast path modifies only raw memory. 1407 if (mem->is_MergeMem()) { 1408 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1409 } 1410 1411 // allocate the Region and Phi nodes for the result 1412 result_region = new RegionNode(3); 1413 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1414 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM); 1415 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch 1416 1417 // Grab regular I/O before optional prefetch may change it. 1418 // Slow-path does no I/O so just set it to the original I/O. 1419 result_phi_i_o->init_req(slow_result_path, i_o); 1420 1421 Node* needgc_ctrl = NULL; 1422 // Name successful fast-path variables 1423 Node* fast_oop_ctrl; 1424 Node* fast_oop_rawmem; 1425 1426 intx prefetch_lines = length != NULL ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1427 1428 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1429 Node* fast_oop = bs->obj_allocate(this, ctrl, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl, 1430 fast_oop_ctrl, fast_oop_rawmem, 1431 prefetch_lines); 1432 1433 if (slow_region != NULL) { 1434 slow_region->add_req(needgc_ctrl); 1435 // This completes all paths into the slow merge point 1436 transform_later(slow_region); 1437 } else { 1438 // Just fall from the need-GC path straight into the VM call. 1439 slow_region = needgc_ctrl; 1440 } 1441 1442 InitializeNode* init = alloc->initialization(); 1443 fast_oop_rawmem = initialize_object(alloc, 1444 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1445 klass_node, length, size_in_bytes); 1446 1447 // If initialization is performed by an array copy, any required 1448 // MemBarStoreStore was already added. If the object does not 1449 // escape no need for a MemBarStoreStore. If the object does not 1450 // escape in its initializer and memory barrier (MemBarStoreStore or 1451 // stronger) is already added at exit of initializer, also no need 1452 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore 1453 // so that stores that initialize this object can't be reordered 1454 // with a subsequent store that makes this object accessible by 1455 // other threads. 1456 // Other threads include java threads and JVM internal threads 1457 // (for example concurrent GC threads). Current concurrent GC 1458 // implementation: G1 will not scan newly created object, 1459 // so it's safe to skip storestore barrier when allocation does 1460 // not escape. 1461 if (!alloc->does_not_escape_thread() && 1462 !alloc->is_allocation_MemBar_redundant() && 1463 (init == NULL || !init->is_complete_with_arraycopy())) { 1464 if (init == NULL || init->req() < InitializeNode::RawStores) { 1465 // No InitializeNode or no stores captured by zeroing 1466 // elimination. Simply add the MemBarStoreStore after object 1467 // initialization. 1468 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1469 transform_later(mb); 1470 1471 mb->init_req(TypeFunc::Memory, fast_oop_rawmem); 1472 mb->init_req(TypeFunc::Control, fast_oop_ctrl); 1473 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control); 1474 transform_later(fast_oop_ctrl); 1475 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory); 1476 transform_later(fast_oop_rawmem); 1477 } else { 1478 // Add the MemBarStoreStore after the InitializeNode so that 1479 // all stores performing the initialization that were moved 1480 // before the InitializeNode happen before the storestore 1481 // barrier. 1482 1483 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control); 1484 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory); 1485 1486 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1487 transform_later(mb); 1488 1489 Node* ctrl = new ProjNode(init,TypeFunc::Control); 1490 transform_later(ctrl); 1491 Node* mem = new ProjNode(init,TypeFunc::Memory); 1492 transform_later(mem); 1493 1494 // The MemBarStoreStore depends on control and memory coming 1495 // from the InitializeNode 1496 mb->init_req(TypeFunc::Memory, mem); 1497 mb->init_req(TypeFunc::Control, ctrl); 1498 1499 ctrl = new ProjNode(mb,TypeFunc::Control); 1500 transform_later(ctrl); 1501 mem = new ProjNode(mb,TypeFunc::Memory); 1502 transform_later(mem); 1503 1504 // All nodes that depended on the InitializeNode for control 1505 // and memory must now depend on the MemBarNode that itself 1506 // depends on the InitializeNode 1507 if (init_ctrl != NULL) { 1508 _igvn.replace_node(init_ctrl, ctrl); 1509 } 1510 if (init_mem != NULL) { 1511 _igvn.replace_node(init_mem, mem); 1512 } 1513 } 1514 } 1515 1516 if (C->env()->dtrace_extended_probes()) { 1517 // Slow-path call 1518 int size = TypeFunc::Parms + 2; 1519 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1520 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), 1521 "dtrace_object_alloc", 1522 TypeRawPtr::BOTTOM); 1523 1524 // Get base of thread-local storage area 1525 Node* thread = new ThreadLocalNode(); 1526 transform_later(thread); 1527 1528 call->init_req(TypeFunc::Parms+0, thread); 1529 call->init_req(TypeFunc::Parms+1, fast_oop); 1530 call->init_req(TypeFunc::Control, fast_oop_ctrl); 1531 call->init_req(TypeFunc::I_O , top()); // does no i/o 1532 call->init_req(TypeFunc::Memory , fast_oop_rawmem); 1533 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1534 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1535 transform_later(call); 1536 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control); 1537 transform_later(fast_oop_ctrl); 1538 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory); 1539 transform_later(fast_oop_rawmem); 1540 } 1541 1542 // Plug in the successful fast-path into the result merge point 1543 result_region ->init_req(fast_result_path, fast_oop_ctrl); 1544 result_phi_rawoop->init_req(fast_result_path, fast_oop); 1545 result_phi_i_o ->init_req(fast_result_path, i_o); 1546 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); 1547 } else { 1548 slow_region = ctrl; 1549 result_phi_i_o = i_o; // Rename it to use in the following code. 1550 } 1551 1552 // Generate slow-path call 1553 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address, 1554 OptoRuntime::stub_name(slow_call_address), 1555 alloc->jvms()->bci(), 1556 TypePtr::BOTTOM); 1557 call->init_req( TypeFunc::Control, slow_region ); 1558 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1559 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs 1560 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1561 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1562 1563 call->init_req(TypeFunc::Parms+0, klass_node); 1564 if (length != NULL) { 1565 call->init_req(TypeFunc::Parms+1, length); 1566 } 1567 1568 // Copy debug information and adjust JVMState information, then replace 1569 // allocate node with the call 1570 call->copy_call_debug_info(&_igvn, alloc); 1571 if (!always_slow) { 1572 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1573 } else { 1574 // Hook i_o projection to avoid its elimination during allocation 1575 // replacement (when only a slow call is generated). 1576 call->set_req(TypeFunc::I_O, result_phi_i_o); 1577 } 1578 _igvn.replace_node(alloc, call); 1579 transform_later(call); 1580 1581 // Identify the output projections from the allocate node and 1582 // adjust any references to them. 1583 // The control and io projections look like: 1584 // 1585 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1586 // Allocate Catch 1587 // ^---Proj(io) <-------+ ^---CatchProj(io) 1588 // 1589 // We are interested in the CatchProj nodes. 1590 // 1591 extract_call_projections(call); 1592 1593 // An allocate node has separate memory projections for the uses on 1594 // the control and i_o paths. Replace the control memory projection with 1595 // result_phi_rawmem (unless we are only generating a slow call when 1596 // both memory projections are combined) 1597 if (!always_slow && _memproj_fallthrough != NULL) { 1598 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { 1599 Node *use = _memproj_fallthrough->fast_out(i); 1600 _igvn.rehash_node_delayed(use); 1601 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); 1602 // back up iterator 1603 --i; 1604 } 1605 } 1606 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete 1607 // _memproj_catchall so we end up with a call that has only 1 memory projection. 1608 if (_memproj_catchall != NULL ) { 1609 if (_memproj_fallthrough == NULL) { 1610 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory); 1611 transform_later(_memproj_fallthrough); 1612 } 1613 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { 1614 Node *use = _memproj_catchall->fast_out(i); 1615 _igvn.rehash_node_delayed(use); 1616 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); 1617 // back up iterator 1618 --i; 1619 } 1620 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted"); 1621 _igvn.remove_dead_node(_memproj_catchall); 1622 } 1623 1624 // An allocate node has separate i_o projections for the uses on the control 1625 // and i_o paths. Always replace the control i_o projection with result i_o 1626 // otherwise incoming i_o become dead when only a slow call is generated 1627 // (it is different from memory projections where both projections are 1628 // combined in such case). 1629 if (_ioproj_fallthrough != NULL) { 1630 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { 1631 Node *use = _ioproj_fallthrough->fast_out(i); 1632 _igvn.rehash_node_delayed(use); 1633 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); 1634 // back up iterator 1635 --i; 1636 } 1637 } 1638 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete 1639 // _ioproj_catchall so we end up with a call that has only 1 i_o projection. 1640 if (_ioproj_catchall != NULL ) { 1641 if (_ioproj_fallthrough == NULL) { 1642 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O); 1643 transform_later(_ioproj_fallthrough); 1644 } 1645 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { 1646 Node *use = _ioproj_catchall->fast_out(i); 1647 _igvn.rehash_node_delayed(use); 1648 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); 1649 // back up iterator 1650 --i; 1651 } 1652 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted"); 1653 _igvn.remove_dead_node(_ioproj_catchall); 1654 } 1655 1656 // if we generated only a slow call, we are done 1657 if (always_slow) { 1658 // Now we can unhook i_o. 1659 if (result_phi_i_o->outcnt() > 1) { 1660 call->set_req(TypeFunc::I_O, top()); 1661 } else { 1662 assert(result_phi_i_o->unique_ctrl_out() == call, ""); 1663 // Case of new array with negative size known during compilation. 1664 // AllocateArrayNode::Ideal() optimization disconnect unreachable 1665 // following code since call to runtime will throw exception. 1666 // As result there will be no users of i_o after the call. 1667 // Leave i_o attached to this call to avoid problems in preceding graph. 1668 } 1669 return; 1670 } 1671 1672 if (_fallthroughcatchproj != NULL) { 1673 ctrl = _fallthroughcatchproj->clone(); 1674 transform_later(ctrl); 1675 _igvn.replace_node(_fallthroughcatchproj, result_region); 1676 } else { 1677 ctrl = top(); 1678 } 1679 Node *slow_result; 1680 if (_resproj == NULL) { 1681 // no uses of the allocation result 1682 slow_result = top(); 1683 } else { 1684 slow_result = _resproj->clone(); 1685 transform_later(slow_result); 1686 _igvn.replace_node(_resproj, result_phi_rawoop); 1687 } 1688 1689 // Plug slow-path into result merge point 1690 result_region ->init_req( slow_result_path, ctrl ); 1691 result_phi_rawoop->init_req( slow_result_path, slow_result); 1692 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); 1693 transform_later(result_region); 1694 transform_later(result_phi_rawoop); 1695 transform_later(result_phi_rawmem); 1696 transform_later(result_phi_i_o); 1697 // This completes all paths into the result merge point 1698 } 1699 1700 1701 // Helper for PhaseMacroExpand::expand_allocate_common. 1702 // Initializes the newly-allocated storage. 1703 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1704 Node* control, Node* rawmem, Node* object, 1705 Node* klass_node, Node* length, 1706 Node* size_in_bytes) { 1707 InitializeNode* init = alloc->initialization(); 1708 // Store the klass & mark bits 1709 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem); 1710 if (!mark_node->is_Con()) { 1711 transform_later(mark_node); 1712 } 1713 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type()); 1714 1715 BasicType bt = T_METADATA; 1716 Node* metadata = klass_node; 1717 Node* properties = alloc->in(AllocateNode::StorageProperties); 1718 if (properties != NULL) { 1719 // Encode array storage properties into klass pointer 1720 assert(EnableValhalla, "array storage properties not supported"); 1721 if (UseCompressedClassPointers) { 1722 // Compress the klass pointer before inserting the storage properties value 1723 metadata = transform_later(new EncodePKlassNode(metadata, metadata->bottom_type()->make_narrowklass())); 1724 metadata = transform_later(new CastN2INode(metadata)); 1725 metadata = transform_later(new OrINode(metadata, transform_later(new ConvL2INode(properties)))); 1726 bt = T_INT; 1727 } else { 1728 metadata = transform_later(new CastP2XNode(NULL, metadata)); 1729 metadata = transform_later(new OrXNode(metadata, properties)); 1730 bt = T_LONG; 1731 } 1732 } 1733 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), metadata, bt); 1734 1735 int header_size = alloc->minimum_header_size(); // conservatively small 1736 1737 // Array length 1738 if (length != NULL) { // Arrays need length field 1739 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1740 // conservatively small header size: 1741 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1742 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1743 if (k->is_array_klass()) // we know the exact header size in most cases: 1744 header_size = Klass::layout_helper_header_size(k->layout_helper()); 1745 } 1746 1747 // Clear the object body, if necessary. 1748 if (init == NULL) { 1749 // The init has somehow disappeared; be cautious and clear everything. 1750 // 1751 // This can happen if a node is allocated but an uncommon trap occurs 1752 // immediately. In this case, the Initialize gets associated with the 1753 // trap, and may be placed in a different (outer) loop, if the Allocate 1754 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1755 // there can be two Allocates to one Initialize. The answer in all these 1756 // edge cases is safety first. It is always safe to clear immediately 1757 // within an Allocate, and then (maybe or maybe not) clear some more later. 1758 if (!(UseTLAB && ZeroTLAB)) { 1759 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1760 alloc->in(AllocateNode::DefaultValue), 1761 alloc->in(AllocateNode::RawDefaultValue), 1762 header_size, size_in_bytes, 1763 &_igvn); 1764 } 1765 } else { 1766 if (!init->is_complete()) { 1767 // Try to win by zeroing only what the init does not store. 1768 // We can also try to do some peephole optimizations, 1769 // such as combining some adjacent subword stores. 1770 rawmem = init->complete_stores(control, rawmem, object, 1771 header_size, size_in_bytes, &_igvn); 1772 } 1773 // We have no more use for this link, since the AllocateNode goes away: 1774 init->set_req(InitializeNode::RawAddress, top()); 1775 // (If we keep the link, it just confuses the register allocator, 1776 // who thinks he sees a real use of the address by the membar.) 1777 } 1778 1779 return rawmem; 1780 } 1781 1782 // Generate prefetch instructions for next allocations. 1783 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1784 Node*& contended_phi_rawmem, 1785 Node* old_eden_top, Node* new_eden_top, 1786 intx lines) { 1787 enum { fall_in_path = 1, pf_path = 2 }; 1788 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1789 // Generate prefetch allocation with watermark check. 1790 // As an allocation hits the watermark, we will prefetch starting 1791 // at a "distance" away from watermark. 1792 1793 Node *pf_region = new RegionNode(3); 1794 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 1795 TypeRawPtr::BOTTOM ); 1796 // I/O is used for Prefetch 1797 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO ); 1798 1799 Node *thread = new ThreadLocalNode(); 1800 transform_later(thread); 1801 1802 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread, 1803 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1804 transform_later(eden_pf_adr); 1805 1806 Node *old_pf_wm = new LoadPNode(needgc_false, 1807 contended_phi_rawmem, eden_pf_adr, 1808 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, 1809 MemNode::unordered); 1810 transform_later(old_pf_wm); 1811 1812 // check against new_eden_top 1813 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm ); 1814 transform_later(need_pf_cmp); 1815 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge ); 1816 transform_later(need_pf_bol); 1817 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol, 1818 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1819 transform_later(need_pf_iff); 1820 1821 // true node, add prefetchdistance 1822 Node *need_pf_true = new IfTrueNode( need_pf_iff ); 1823 transform_later(need_pf_true); 1824 1825 Node *need_pf_false = new IfFalseNode( need_pf_iff ); 1826 transform_later(need_pf_false); 1827 1828 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm, 1829 _igvn.MakeConX(AllocatePrefetchDistance) ); 1830 transform_later(new_pf_wmt ); 1831 new_pf_wmt->set_req(0, need_pf_true); 1832 1833 Node *store_new_wmt = new StorePNode(need_pf_true, 1834 contended_phi_rawmem, eden_pf_adr, 1835 TypeRawPtr::BOTTOM, new_pf_wmt, 1836 MemNode::unordered); 1837 transform_later(store_new_wmt); 1838 1839 // adding prefetches 1840 pf_phi_abio->init_req( fall_in_path, i_o ); 1841 1842 Node *prefetch_adr; 1843 Node *prefetch; 1844 uint step_size = AllocatePrefetchStepSize; 1845 uint distance = 0; 1846 1847 for ( intx i = 0; i < lines; i++ ) { 1848 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt, 1849 _igvn.MakeConX(distance) ); 1850 transform_later(prefetch_adr); 1851 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1852 transform_later(prefetch); 1853 distance += step_size; 1854 i_o = prefetch; 1855 } 1856 pf_phi_abio->set_req( pf_path, i_o ); 1857 1858 pf_region->init_req( fall_in_path, need_pf_false ); 1859 pf_region->init_req( pf_path, need_pf_true ); 1860 1861 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 1862 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 1863 1864 transform_later(pf_region); 1865 transform_later(pf_phi_rawmem); 1866 transform_later(pf_phi_abio); 1867 1868 needgc_false = pf_region; 1869 contended_phi_rawmem = pf_phi_rawmem; 1870 i_o = pf_phi_abio; 1871 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { 1872 // Insert a prefetch instruction for each allocation. 1873 // This code is used to generate 1 prefetch instruction per cache line. 1874 1875 // Generate several prefetch instructions. 1876 uint step_size = AllocatePrefetchStepSize; 1877 uint distance = AllocatePrefetchDistance; 1878 1879 // Next cache address. 1880 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top, 1881 _igvn.MakeConX(step_size + distance)); 1882 transform_later(cache_adr); 1883 cache_adr = new CastP2XNode(needgc_false, cache_adr); 1884 transform_later(cache_adr); 1885 // Address is aligned to execute prefetch to the beginning of cache line size 1886 // (it is important when BIS instruction is used on SPARC as prefetch). 1887 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); 1888 cache_adr = new AndXNode(cache_adr, mask); 1889 transform_later(cache_adr); 1890 cache_adr = new CastX2PNode(cache_adr); 1891 transform_later(cache_adr); 1892 1893 // Prefetch 1894 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); 1895 prefetch->set_req(0, needgc_false); 1896 transform_later(prefetch); 1897 contended_phi_rawmem = prefetch; 1898 Node *prefetch_adr; 1899 distance = step_size; 1900 for ( intx i = 1; i < lines; i++ ) { 1901 prefetch_adr = new AddPNode( cache_adr, cache_adr, 1902 _igvn.MakeConX(distance) ); 1903 transform_later(prefetch_adr); 1904 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); 1905 transform_later(prefetch); 1906 distance += step_size; 1907 contended_phi_rawmem = prefetch; 1908 } 1909 } else if( AllocatePrefetchStyle > 0 ) { 1910 // Insert a prefetch for each allocation only on the fast-path 1911 Node *prefetch_adr; 1912 Node *prefetch; 1913 // Generate several prefetch instructions. 1914 uint step_size = AllocatePrefetchStepSize; 1915 uint distance = AllocatePrefetchDistance; 1916 for ( intx i = 0; i < lines; i++ ) { 1917 prefetch_adr = new AddPNode( old_eden_top, new_eden_top, 1918 _igvn.MakeConX(distance) ); 1919 transform_later(prefetch_adr); 1920 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1921 // Do not let it float too high, since if eden_top == eden_end, 1922 // both might be null. 1923 if( i == 0 ) { // Set control for first prefetch, next follows it 1924 prefetch->init_req(0, needgc_false); 1925 } 1926 transform_later(prefetch); 1927 distance += step_size; 1928 i_o = prefetch; 1929 } 1930 } 1931 return i_o; 1932 } 1933 1934 1935 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 1936 expand_allocate_common(alloc, NULL, 1937 OptoRuntime::new_instance_Type(), 1938 OptoRuntime::new_instance_Java()); 1939 } 1940 1941 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 1942 Node* length = alloc->in(AllocateNode::ALength); 1943 InitializeNode* init = alloc->initialization(); 1944 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1945 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1946 address slow_call_address; // Address of slow call 1947 if (init != NULL && init->is_complete_with_arraycopy() && 1948 k->is_type_array_klass()) { 1949 // Don't zero type array during slow allocation in VM since 1950 // it will be initialized later by arraycopy in compiled code. 1951 slow_call_address = OptoRuntime::new_array_nozero_Java(); 1952 } else { 1953 slow_call_address = OptoRuntime::new_array_Java(); 1954 } 1955 expand_allocate_common(alloc, length, 1956 OptoRuntime::new_array_Type(), 1957 slow_call_address); 1958 } 1959 1960 //-------------------mark_eliminated_box---------------------------------- 1961 // 1962 // During EA obj may point to several objects but after few ideal graph 1963 // transformations (CCP) it may point to only one non escaping object 1964 // (but still using phi), corresponding locks and unlocks will be marked 1965 // for elimination. Later obj could be replaced with a new node (new phi) 1966 // and which does not have escape information. And later after some graph 1967 // reshape other locks and unlocks (which were not marked for elimination 1968 // before) are connected to this new obj (phi) but they still will not be 1969 // marked for elimination since new obj has no escape information. 1970 // Mark all associated (same box and obj) lock and unlock nodes for 1971 // elimination if some of them marked already. 1972 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) { 1973 if (oldbox->as_BoxLock()->is_eliminated()) 1974 return; // This BoxLock node was processed already. 1975 1976 // New implementation (EliminateNestedLocks) has separate BoxLock 1977 // node for each locked region so mark all associated locks/unlocks as 1978 // eliminated even if different objects are referenced in one locked region 1979 // (for example, OSR compilation of nested loop inside locked scope). 1980 if (EliminateNestedLocks || 1981 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) { 1982 // Box is used only in one lock region. Mark this box as eliminated. 1983 _igvn.hash_delete(oldbox); 1984 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value 1985 _igvn.hash_insert(oldbox); 1986 1987 for (uint i = 0; i < oldbox->outcnt(); i++) { 1988 Node* u = oldbox->raw_out(i); 1989 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) { 1990 AbstractLockNode* alock = u->as_AbstractLock(); 1991 // Check lock's box since box could be referenced by Lock's debug info. 1992 if (alock->box_node() == oldbox) { 1993 // Mark eliminated all related locks and unlocks. 1994 #ifdef ASSERT 1995 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4"); 1996 #endif 1997 alock->set_non_esc_obj(); 1998 } 1999 } 2000 } 2001 return; 2002 } 2003 2004 // Create new "eliminated" BoxLock node and use it in monitor debug info 2005 // instead of oldbox for the same object. 2006 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 2007 2008 // Note: BoxLock node is marked eliminated only here and it is used 2009 // to indicate that all associated lock and unlock nodes are marked 2010 // for elimination. 2011 newbox->set_eliminated(); 2012 transform_later(newbox); 2013 2014 // Replace old box node with new box for all users of the same object. 2015 for (uint i = 0; i < oldbox->outcnt();) { 2016 bool next_edge = true; 2017 2018 Node* u = oldbox->raw_out(i); 2019 if (u->is_AbstractLock()) { 2020 AbstractLockNode* alock = u->as_AbstractLock(); 2021 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) { 2022 // Replace Box and mark eliminated all related locks and unlocks. 2023 #ifdef ASSERT 2024 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5"); 2025 #endif 2026 alock->set_non_esc_obj(); 2027 _igvn.rehash_node_delayed(alock); 2028 alock->set_box_node(newbox); 2029 next_edge = false; 2030 } 2031 } 2032 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) { 2033 FastLockNode* flock = u->as_FastLock(); 2034 assert(flock->box_node() == oldbox, "sanity"); 2035 _igvn.rehash_node_delayed(flock); 2036 flock->set_box_node(newbox); 2037 next_edge = false; 2038 } 2039 2040 // Replace old box in monitor debug info. 2041 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 2042 SafePointNode* sfn = u->as_SafePoint(); 2043 JVMState* youngest_jvms = sfn->jvms(); 2044 int max_depth = youngest_jvms->depth(); 2045 for (int depth = 1; depth <= max_depth; depth++) { 2046 JVMState* jvms = youngest_jvms->of_depth(depth); 2047 int num_mon = jvms->nof_monitors(); 2048 // Loop over monitors 2049 for (int idx = 0; idx < num_mon; idx++) { 2050 Node* obj_node = sfn->monitor_obj(jvms, idx); 2051 Node* box_node = sfn->monitor_box(jvms, idx); 2052 if (box_node == oldbox && obj_node->eqv_uncast(obj)) { 2053 int j = jvms->monitor_box_offset(idx); 2054 _igvn.replace_input_of(u, j, newbox); 2055 next_edge = false; 2056 } 2057 } 2058 } 2059 } 2060 if (next_edge) i++; 2061 } 2062 } 2063 2064 //-----------------------mark_eliminated_locking_nodes----------------------- 2065 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { 2066 if (EliminateNestedLocks) { 2067 if (alock->is_nested()) { 2068 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity"); 2069 return; 2070 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened 2071 // Only Lock node has JVMState needed here. 2072 // Not that preceding claim is documented anywhere else. 2073 if (alock->jvms() != NULL) { 2074 if (alock->as_Lock()->is_nested_lock_region()) { 2075 // Mark eliminated related nested locks and unlocks. 2076 Node* obj = alock->obj_node(); 2077 BoxLockNode* box_node = alock->box_node()->as_BoxLock(); 2078 assert(!box_node->is_eliminated(), "should not be marked yet"); 2079 // Note: BoxLock node is marked eliminated only here 2080 // and it is used to indicate that all associated lock 2081 // and unlock nodes are marked for elimination. 2082 box_node->set_eliminated(); // Box's hash is always NO_HASH here 2083 for (uint i = 0; i < box_node->outcnt(); i++) { 2084 Node* u = box_node->raw_out(i); 2085 if (u->is_AbstractLock()) { 2086 alock = u->as_AbstractLock(); 2087 if (alock->box_node() == box_node) { 2088 // Verify that this Box is referenced only by related locks. 2089 assert(alock->obj_node()->eqv_uncast(obj), ""); 2090 // Mark all related locks and unlocks. 2091 #ifdef ASSERT 2092 alock->log_lock_optimization(C, "eliminate_lock_set_nested"); 2093 #endif 2094 alock->set_nested(); 2095 } 2096 } 2097 } 2098 } else { 2099 #ifdef ASSERT 2100 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region"); 2101 if (C->log() != NULL) 2102 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output 2103 #endif 2104 } 2105 } 2106 return; 2107 } 2108 // Process locks for non escaping object 2109 assert(alock->is_non_esc_obj(), ""); 2110 } // EliminateNestedLocks 2111 2112 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object 2113 // Look for all locks of this object and mark them and 2114 // corresponding BoxLock nodes as eliminated. 2115 Node* obj = alock->obj_node(); 2116 for (uint j = 0; j < obj->outcnt(); j++) { 2117 Node* o = obj->raw_out(j); 2118 if (o->is_AbstractLock() && 2119 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) { 2120 alock = o->as_AbstractLock(); 2121 Node* box = alock->box_node(); 2122 // Replace old box node with new eliminated box for all users 2123 // of the same object and mark related locks as eliminated. 2124 mark_eliminated_box(box, obj); 2125 } 2126 } 2127 } 2128 } 2129 2130 // we have determined that this lock/unlock can be eliminated, we simply 2131 // eliminate the node without expanding it. 2132 // 2133 // Note: The membar's associated with the lock/unlock are currently not 2134 // eliminated. This should be investigated as a future enhancement. 2135 // 2136 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 2137 2138 if (!alock->is_eliminated()) { 2139 return false; 2140 } 2141 #ifdef ASSERT 2142 const Type* obj_type = _igvn.type(alock->obj_node()); 2143 assert(!obj_type->isa_valuetype() && !obj_type->is_valuetypeptr(), "Eliminating lock on value type"); 2144 if (!alock->is_coarsened()) { 2145 // Check that new "eliminated" BoxLock node is created. 2146 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 2147 assert(oldbox->is_eliminated(), "should be done already"); 2148 } 2149 #endif 2150 2151 alock->log_lock_optimization(C, "eliminate_lock"); 2152 2153 #ifndef PRODUCT 2154 if (PrintEliminateLocks) { 2155 if (alock->is_Lock()) { 2156 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx); 2157 } else { 2158 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx); 2159 } 2160 } 2161 #endif 2162 2163 Node* mem = alock->in(TypeFunc::Memory); 2164 Node* ctrl = alock->in(TypeFunc::Control); 2165 guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL"); 2166 2167 extract_call_projections(alock); 2168 // There are 2 projections from the lock. The lock node will 2169 // be deleted when its last use is subsumed below. 2170 assert(alock->outcnt() == 2 && 2171 _fallthroughproj != NULL && 2172 _memproj_fallthrough != NULL, 2173 "Unexpected projections from Lock/Unlock"); 2174 2175 Node* fallthroughproj = _fallthroughproj; 2176 Node* memproj_fallthrough = _memproj_fallthrough; 2177 2178 // The memory projection from a lock/unlock is RawMem 2179 // The input to a Lock is merged memory, so extract its RawMem input 2180 // (unless the MergeMem has been optimized away.) 2181 if (alock->is_Lock()) { 2182 // Seach for MemBarAcquireLock node and delete it also. 2183 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 2184 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, ""); 2185 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 2186 Node* memproj = membar->proj_out(TypeFunc::Memory); 2187 _igvn.replace_node(ctrlproj, fallthroughproj); 2188 _igvn.replace_node(memproj, memproj_fallthrough); 2189 2190 // Delete FastLock node also if this Lock node is unique user 2191 // (a loop peeling may clone a Lock node). 2192 Node* flock = alock->as_Lock()->fastlock_node(); 2193 if (flock->outcnt() == 1) { 2194 assert(flock->unique_out() == alock, "sanity"); 2195 _igvn.replace_node(flock, top()); 2196 } 2197 } 2198 2199 // Seach for MemBarReleaseLock node and delete it also. 2200 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) { 2201 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 2202 assert(membar->Opcode() == Op_MemBarReleaseLock && 2203 mem->is_Proj() && membar == mem->in(0), ""); 2204 _igvn.replace_node(fallthroughproj, ctrl); 2205 _igvn.replace_node(memproj_fallthrough, mem); 2206 fallthroughproj = ctrl; 2207 memproj_fallthrough = mem; 2208 ctrl = membar->in(TypeFunc::Control); 2209 mem = membar->in(TypeFunc::Memory); 2210 } 2211 2212 _igvn.replace_node(fallthroughproj, ctrl); 2213 _igvn.replace_node(memproj_fallthrough, mem); 2214 return true; 2215 } 2216 2217 2218 //------------------------------expand_lock_node---------------------- 2219 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 2220 2221 Node* ctrl = lock->in(TypeFunc::Control); 2222 Node* mem = lock->in(TypeFunc::Memory); 2223 Node* obj = lock->obj_node(); 2224 Node* box = lock->box_node(); 2225 Node* flock = lock->fastlock_node(); 2226 2227 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2228 2229 // Make the merge point 2230 Node *region; 2231 Node *mem_phi; 2232 Node *slow_path; 2233 2234 if (UseOptoBiasInlining) { 2235 /* 2236 * See the full description in MacroAssembler::biased_locking_enter(). 2237 * 2238 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { 2239 * // The object is biased. 2240 * proto_node = klass->prototype_header; 2241 * o_node = thread | proto_node; 2242 * x_node = o_node ^ mark_word; 2243 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? 2244 * // Done. 2245 * } else { 2246 * if( (x_node & biased_lock_mask) != 0 ) { 2247 * // The klass's prototype header is no longer biased. 2248 * cas(&mark_word, mark_word, proto_node) 2249 * goto cas_lock; 2250 * } else { 2251 * // The klass's prototype header is still biased. 2252 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? 2253 * old = mark_word; 2254 * new = o_node; 2255 * } else { 2256 * // Different thread or anonymous biased. 2257 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); 2258 * new = thread | old; 2259 * } 2260 * // Try to rebias. 2261 * if( cas(&mark_word, old, new) == 0 ) { 2262 * // Done. 2263 * } else { 2264 * goto slow_path; // Failed. 2265 * } 2266 * } 2267 * } 2268 * } else { 2269 * // The object is not biased. 2270 * cas_lock: 2271 * if( FastLock(obj) == 0 ) { 2272 * // Done. 2273 * } else { 2274 * slow_path: 2275 * OptoRuntime::complete_monitor_locking_Java(obj); 2276 * } 2277 * } 2278 */ 2279 2280 region = new RegionNode(5); 2281 // create a Phi for the memory state 2282 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2283 2284 Node* fast_lock_region = new RegionNode(3); 2285 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); 2286 2287 // First, check mark word for the biased lock pattern. 2288 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2289 2290 // Get fast path - mark word has the biased lock pattern. 2291 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, 2292 markWord::biased_lock_mask_in_place, 2293 markWord::biased_lock_pattern, true); 2294 // fast_lock_region->in(1) is set to slow path. 2295 fast_lock_mem_phi->init_req(1, mem); 2296 2297 // Now check that the lock is biased to the current thread and has 2298 // the same epoch and bias as Klass::_prototype_header. 2299 2300 // Special-case a fresh allocation to avoid building nodes: 2301 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); 2302 if (klass_node == NULL) { 2303 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 2304 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr())); 2305 #ifdef _LP64 2306 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) { 2307 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity"); 2308 klass_node->in(1)->init_req(0, ctrl); 2309 } else 2310 #endif 2311 klass_node->init_req(0, ctrl); 2312 } 2313 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type()); 2314 2315 Node* thread = transform_later(new ThreadLocalNode()); 2316 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2317 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node)); 2318 Node* x_node = transform_later(new XorXNode(o_node, mark_node)); 2319 2320 // Get slow path - mark word does NOT match the value. 2321 STATIC_ASSERT(markWord::age_mask_in_place <= INT_MAX); 2322 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, 2323 (~(int)markWord::age_mask_in_place), 0); 2324 // region->in(3) is set to fast path - the object is biased to the current thread. 2325 mem_phi->init_req(3, mem); 2326 2327 2328 // Mark word does NOT match the value (thread | Klass::_prototype_header). 2329 2330 2331 // First, check biased pattern. 2332 // Get fast path - _prototype_header has the same biased lock pattern. 2333 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, 2334 markWord::biased_lock_mask_in_place, 0, true); 2335 2336 not_biased_ctrl = fast_lock_region->in(2); // Slow path 2337 // fast_lock_region->in(2) - the prototype header is no longer biased 2338 // and we have to revoke the bias on this object. 2339 // We are going to try to reset the mark of this object to the prototype 2340 // value and fall through to the CAS-based locking scheme. 2341 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 2342 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr, 2343 proto_node, mark_node); 2344 transform_later(cas); 2345 Node* proj = transform_later(new SCMemProjNode(cas)); 2346 fast_lock_mem_phi->init_req(2, proj); 2347 2348 2349 // Second, check epoch bits. 2350 Node* rebiased_region = new RegionNode(3); 2351 Node* old_phi = new PhiNode( rebiased_region, TypeX_X); 2352 Node* new_phi = new PhiNode( rebiased_region, TypeX_X); 2353 2354 // Get slow path - mark word does NOT match epoch bits. 2355 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, 2356 markWord::epoch_mask_in_place, 0); 2357 // The epoch of the current bias is not valid, attempt to rebias the object 2358 // toward the current thread. 2359 rebiased_region->init_req(2, epoch_ctrl); 2360 old_phi->init_req(2, mark_node); 2361 new_phi->init_req(2, o_node); 2362 2363 // rebiased_region->in(1) is set to fast path. 2364 // The epoch of the current bias is still valid but we know 2365 // nothing about the owner; it might be set or it might be clear. 2366 Node* cmask = MakeConX(markWord::biased_lock_mask_in_place | 2367 markWord::age_mask_in_place | 2368 markWord::epoch_mask_in_place); 2369 Node* old = transform_later(new AndXNode(mark_node, cmask)); 2370 cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2371 Node* new_mark = transform_later(new OrXNode(cast_thread, old)); 2372 old_phi->init_req(1, old); 2373 new_phi->init_req(1, new_mark); 2374 2375 transform_later(rebiased_region); 2376 transform_later(old_phi); 2377 transform_later(new_phi); 2378 2379 // Try to acquire the bias of the object using an atomic operation. 2380 // If this fails we will go in to the runtime to revoke the object's bias. 2381 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi); 2382 transform_later(cas); 2383 proj = transform_later(new SCMemProjNode(cas)); 2384 2385 // Get slow path - Failed to CAS. 2386 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); 2387 mem_phi->init_req(4, proj); 2388 // region->in(4) is set to fast path - the object is rebiased to the current thread. 2389 2390 // Failed to CAS. 2391 slow_path = new RegionNode(3); 2392 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); 2393 2394 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control 2395 slow_mem->init_req(1, proj); 2396 2397 // Call CAS-based locking scheme (FastLock node). 2398 2399 transform_later(fast_lock_region); 2400 transform_later(fast_lock_mem_phi); 2401 2402 // Get slow path - FastLock failed to lock the object. 2403 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); 2404 mem_phi->init_req(2, fast_lock_mem_phi); 2405 // region->in(2) is set to fast path - the object is locked to the current thread. 2406 2407 slow_path->init_req(2, ctrl); // Capture slow-control 2408 slow_mem->init_req(2, fast_lock_mem_phi); 2409 2410 transform_later(slow_path); 2411 transform_later(slow_mem); 2412 // Reset lock's memory edge. 2413 lock->set_req(TypeFunc::Memory, slow_mem); 2414 2415 } else { 2416 region = new RegionNode(3); 2417 // create a Phi for the memory state 2418 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2419 2420 // Optimize test; set region slot 2 2421 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); 2422 mem_phi->init_req(2, mem); 2423 } 2424 2425 const TypeOopPtr* objptr = _igvn.type(obj)->make_oopptr(); 2426 if (objptr->can_be_value_type()) { 2427 // Deoptimize and re-execute if a value 2428 assert(EnableValhalla, "should only be used if value types are enabled"); 2429 Node* mark = make_load(slow_path, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2430 Node* value_mask = _igvn.MakeConX(markWord::always_locked_pattern); 2431 Node* is_value = _igvn.transform(new AndXNode(mark, value_mask)); 2432 Node* cmp = _igvn.transform(new CmpXNode(is_value, value_mask)); 2433 Node* bol = _igvn.transform(new BoolNode(cmp, BoolTest::eq)); 2434 Node* unc_ctrl = generate_slow_guard(&slow_path, bol, NULL); 2435 2436 int trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_class_check, Deoptimization::Action_none); 2437 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2438 const TypePtr* no_memory_effects = NULL; 2439 JVMState* jvms = lock->jvms(); 2440 CallNode* unc = new CallStaticJavaNode(OptoRuntime::uncommon_trap_Type(), call_addr, "uncommon_trap", 2441 jvms->bci(), no_memory_effects); 2442 2443 unc->init_req(TypeFunc::Control, unc_ctrl); 2444 unc->init_req(TypeFunc::I_O, lock->i_o()); 2445 unc->init_req(TypeFunc::Memory, mem); // may gc ptrs 2446 unc->init_req(TypeFunc::FramePtr, lock->in(TypeFunc::FramePtr)); 2447 unc->init_req(TypeFunc::ReturnAdr, lock->in(TypeFunc::ReturnAdr)); 2448 unc->init_req(TypeFunc::Parms+0, _igvn.intcon(trap_request)); 2449 unc->set_cnt(PROB_UNLIKELY_MAG(4)); 2450 unc->copy_call_debug_info(&_igvn, lock); 2451 2452 assert(unc->peek_monitor_box() == box, "wrong monitor"); 2453 assert(unc->peek_monitor_obj() == obj, "wrong monitor"); 2454 2455 // pop monitor and push obj back on stack: we trap before the monitorenter 2456 unc->pop_monitor(); 2457 unc->grow_stack(unc->jvms(), 1); 2458 unc->set_stack(unc->jvms(), unc->jvms()->stk_size()-1, obj); 2459 2460 _igvn.register_new_node_with_optimizer(unc); 2461 2462 Node* ctrl = _igvn.transform(new ProjNode(unc, TypeFunc::Control)); 2463 Node* halt = _igvn.transform(new HaltNode(ctrl, lock->in(TypeFunc::FramePtr), "monitor enter on value-type")); 2464 C->root()->add_req(halt); 2465 } 2466 2467 // Make slow path call 2468 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), 2469 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, 2470 obj, box, NULL); 2471 2472 extract_call_projections(call); 2473 2474 // Slow path can only throw asynchronous exceptions, which are always 2475 // de-opted. So the compiler thinks the slow-call can never throw an 2476 // exception. If it DOES throw an exception we would need the debug 2477 // info removed first (since if it throws there is no monitor). 2478 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2479 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2480 2481 // Capture slow path 2482 // disconnect fall-through projection from call and create a new one 2483 // hook up users of fall-through projection to region 2484 Node *slow_ctrl = _fallthroughproj->clone(); 2485 transform_later(slow_ctrl); 2486 _igvn.hash_delete(_fallthroughproj); 2487 _fallthroughproj->disconnect_inputs(NULL, C); 2488 region->init_req(1, slow_ctrl); 2489 // region inputs are now complete 2490 transform_later(region); 2491 _igvn.replace_node(_fallthroughproj, region); 2492 2493 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory)); 2494 mem_phi->init_req(1, memproj ); 2495 transform_later(mem_phi); 2496 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2497 } 2498 2499 //------------------------------expand_unlock_node---------------------- 2500 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 2501 2502 Node* ctrl = unlock->in(TypeFunc::Control); 2503 Node* mem = unlock->in(TypeFunc::Memory); 2504 Node* obj = unlock->obj_node(); 2505 Node* box = unlock->box_node(); 2506 2507 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2508 2509 // No need for a null check on unlock 2510 2511 // Make the merge point 2512 Node *region; 2513 Node *mem_phi; 2514 2515 if (UseOptoBiasInlining) { 2516 // Check for biased locking unlock case, which is a no-op. 2517 // See the full description in MacroAssembler::biased_locking_exit(). 2518 region = new RegionNode(4); 2519 // create a Phi for the memory state 2520 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2521 mem_phi->init_req(3, mem); 2522 2523 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2524 ctrl = opt_bits_test(ctrl, region, 3, mark_node, 2525 markWord::biased_lock_mask_in_place, 2526 markWord::biased_lock_pattern); 2527 } else { 2528 region = new RegionNode(3); 2529 // create a Phi for the memory state 2530 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2531 } 2532 2533 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box ); 2534 funlock = transform_later( funlock )->as_FastUnlock(); 2535 // Optimize test; set region slot 2 2536 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); 2537 Node *thread = transform_later(new ThreadLocalNode()); 2538 2539 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), 2540 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), 2541 "complete_monitor_unlocking_C", slow_path, obj, box, thread); 2542 2543 extract_call_projections(call); 2544 2545 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2546 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2547 2548 // No exceptions for unlocking 2549 // Capture slow path 2550 // disconnect fall-through projection from call and create a new one 2551 // hook up users of fall-through projection to region 2552 Node *slow_ctrl = _fallthroughproj->clone(); 2553 transform_later(slow_ctrl); 2554 _igvn.hash_delete(_fallthroughproj); 2555 _fallthroughproj->disconnect_inputs(NULL, C); 2556 region->init_req(1, slow_ctrl); 2557 // region inputs are now complete 2558 transform_later(region); 2559 _igvn.replace_node(_fallthroughproj, region); 2560 2561 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) ); 2562 mem_phi->init_req(1, memproj ); 2563 mem_phi->init_req(2, mem); 2564 transform_later(mem_phi); 2565 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2566 } 2567 2568 // A value type might be returned from the call but we don't know its 2569 // type. Either we get a buffered value (and nothing needs to be done) 2570 // or one of the values being returned is the klass of the value type 2571 // and we need to allocate a value type instance of that type and 2572 // initialize it with other values being returned. In that case, we 2573 // first try a fast path allocation and initialize the value with the 2574 // value klass's pack handler or we fall back to a runtime call. 2575 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) { 2576 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call"); 2577 Node* ret = call->proj_out_or_null(TypeFunc::Parms); 2578 if (ret == NULL) { 2579 return; 2580 } 2581 const TypeFunc* tf = call->_tf; 2582 const TypeTuple* domain = OptoRuntime::store_value_type_fields_Type()->domain_cc(); 2583 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain); 2584 call->_tf = new_tf; 2585 // Make sure the change of type is applied before projections are processed by igvn 2586 _igvn.set_type(call, call->Value(&_igvn)); 2587 _igvn.set_type(ret, ret->Value(&_igvn)); 2588 2589 // Before any new projection is added: 2590 CallProjections* projs = call->extract_projections(true, true); 2591 2592 Node* ctl = new Node(1); 2593 Node* mem = new Node(1); 2594 Node* io = new Node(1); 2595 Node* ex_ctl = new Node(1); 2596 Node* ex_mem = new Node(1); 2597 Node* ex_io = new Node(1); 2598 Node* res = new Node(1); 2599 2600 Node* cast = transform_later(new CastP2XNode(ctl, res)); 2601 Node* mask = MakeConX(0x1); 2602 Node* masked = transform_later(new AndXNode(cast, mask)); 2603 Node* cmp = transform_later(new CmpXNode(masked, mask)); 2604 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq)); 2605 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN); 2606 transform_later(allocation_iff); 2607 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff)); 2608 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff)); 2609 2610 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM)); 2611 2612 Node* mask2 = MakeConX(-2); 2613 Node* masked2 = transform_later(new AndXNode(cast, mask2)); 2614 Node* rawklassptr = transform_later(new CastX2PNode(masked2)); 2615 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeKlassPtr::OBJECT_OR_NULL)); 2616 2617 Node* slowpath_bol = NULL; 2618 Node* top_adr = NULL; 2619 Node* old_top = NULL; 2620 Node* new_top = NULL; 2621 if (UseTLAB) { 2622 Node* end_adr = NULL; 2623 set_eden_pointers(top_adr, end_adr); 2624 Node* end = make_load(ctl, mem, end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 2625 old_top = new LoadPNode(ctl, mem, top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered); 2626 transform_later(old_top); 2627 Node* layout_val = make_load(NULL, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT); 2628 Node* size_in_bytes = ConvI2X(layout_val); 2629 new_top = new AddPNode(top(), old_top, size_in_bytes); 2630 transform_later(new_top); 2631 Node* slowpath_cmp = new CmpPNode(new_top, end); 2632 transform_later(slowpath_cmp); 2633 slowpath_bol = new BoolNode(slowpath_cmp, BoolTest::ge); 2634 transform_later(slowpath_bol); 2635 } else { 2636 slowpath_bol = intcon(1); 2637 top_adr = top(); 2638 old_top = top(); 2639 new_top = top(); 2640 } 2641 IfNode* slowpath_iff = new IfNode(allocation_ctl, slowpath_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); 2642 transform_later(slowpath_iff); 2643 2644 Node* slowpath_true = new IfTrueNode(slowpath_iff); 2645 transform_later(slowpath_true); 2646 2647 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_value_type_fields_Type(), 2648 StubRoutines::store_value_type_fields_to_buf(), 2649 "store_value_type_fields", 2650 call->jvms()->bci(), 2651 TypePtr::BOTTOM); 2652 slow_call->init_req(TypeFunc::Control, slowpath_true); 2653 slow_call->init_req(TypeFunc::Memory, mem); 2654 slow_call->init_req(TypeFunc::I_O, io); 2655 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr)); 2656 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr)); 2657 slow_call->init_req(TypeFunc::Parms, res); 2658 2659 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control)); 2660 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory)); 2661 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O)); 2662 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms)); 2663 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2)); 2664 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci)); 2665 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci)); 2666 2667 Node* ex_r = new RegionNode(3); 2668 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM); 2669 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO); 2670 ex_r->init_req(1, slow_excp); 2671 ex_mem_phi->init_req(1, slow_mem); 2672 ex_io_phi->init_req(1, slow_io); 2673 ex_r->init_req(2, ex_ctl); 2674 ex_mem_phi->init_req(2, ex_mem); 2675 ex_io_phi->init_req(2, ex_io); 2676 2677 transform_later(ex_r); 2678 transform_later(ex_mem_phi); 2679 transform_later(ex_io_phi); 2680 2681 Node* slowpath_false = new IfFalseNode(slowpath_iff); 2682 transform_later(slowpath_false); 2683 Node* rawmem = new StorePNode(slowpath_false, mem, top_adr, TypeRawPtr::BOTTOM, new_top, MemNode::unordered); 2684 transform_later(rawmem); 2685 Node* mark_node = makecon(TypeRawPtr::make((address)markWord::always_locked_prototype().value())); 2686 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 2687 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 2688 if (UseCompressedClassPointers) { 2689 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT); 2690 } 2691 Node* fixed_block = make_load(slowpath_false, rawmem, klass_node, in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 2692 Node* pack_handler = make_load(slowpath_false, rawmem, fixed_block, in_bytes(ValueKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 2693 2694 CallLeafNoFPNode* handler_call = new CallLeafNoFPNode(OptoRuntime::pack_value_type_Type(), 2695 NULL, 2696 "pack handler", 2697 TypeRawPtr::BOTTOM); 2698 handler_call->init_req(TypeFunc::Control, slowpath_false); 2699 handler_call->init_req(TypeFunc::Memory, rawmem); 2700 handler_call->init_req(TypeFunc::I_O, top()); 2701 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr)); 2702 handler_call->init_req(TypeFunc::ReturnAdr, top()); 2703 handler_call->init_req(TypeFunc::Parms, pack_handler); 2704 handler_call->init_req(TypeFunc::Parms+1, old_top); 2705 2706 // We don't know how many values are returned. This assumes the 2707 // worst case, that all available registers are used. 2708 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) { 2709 if (domain->field_at(i) == Type::HALF) { 2710 slow_call->init_req(i, top()); 2711 handler_call->init_req(i+1, top()); 2712 continue; 2713 } 2714 Node* proj = transform_later(new ProjNode(call, i)); 2715 slow_call->init_req(i, proj); 2716 handler_call->init_req(i+1, proj); 2717 } 2718 2719 // We can safepoint at that new call 2720 slow_call->copy_call_debug_info(&_igvn, call); 2721 transform_later(slow_call); 2722 transform_later(handler_call); 2723 2724 Node* handler_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control)); 2725 rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory)); 2726 Node* slowpath_false_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms)); 2727 2728 MergeMemNode* slowpath_false_mem = MergeMemNode::make(mem); 2729 slowpath_false_mem->set_memory_at(Compile::AliasIdxRaw, rawmem); 2730 transform_later(slowpath_false_mem); 2731 2732 Node* r = new RegionNode(4); 2733 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM); 2734 Node* io_phi = new PhiNode(r, Type::ABIO); 2735 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM); 2736 2737 r->init_req(1, no_allocation_ctl); 2738 mem_phi->init_req(1, mem); 2739 io_phi->init_req(1, io); 2740 res_phi->init_req(1, no_allocation_res); 2741 r->init_req(2, slow_norm); 2742 mem_phi->init_req(2, slow_mem); 2743 io_phi->init_req(2, slow_io); 2744 res_phi->init_req(2, slow_res); 2745 r->init_req(3, handler_ctl); 2746 mem_phi->init_req(3, slowpath_false_mem); 2747 io_phi->init_req(3, io); 2748 res_phi->init_req(3, slowpath_false_res); 2749 2750 transform_later(r); 2751 transform_later(mem_phi); 2752 transform_later(io_phi); 2753 transform_later(res_phi); 2754 2755 assert(projs->nb_resproj == 1, "unexpected number of results"); 2756 _igvn.replace_in_uses(projs->fallthrough_catchproj, r); 2757 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi); 2758 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi); 2759 _igvn.replace_in_uses(projs->resproj[0], res_phi); 2760 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r); 2761 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi); 2762 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi); 2763 2764 _igvn.replace_node(ctl, projs->fallthrough_catchproj); 2765 _igvn.replace_node(mem, projs->fallthrough_memproj); 2766 _igvn.replace_node(io, projs->fallthrough_ioproj); 2767 _igvn.replace_node(res, projs->resproj[0]); 2768 _igvn.replace_node(ex_ctl, projs->catchall_catchproj); 2769 _igvn.replace_node(ex_mem, projs->catchall_memproj); 2770 _igvn.replace_node(ex_io, projs->catchall_ioproj); 2771 } 2772 2773 //---------------------------eliminate_macro_nodes---------------------- 2774 // Eliminate scalar replaced allocations and associated locks. 2775 void PhaseMacroExpand::eliminate_macro_nodes() { 2776 if (C->macro_count() == 0) 2777 return; 2778 2779 // First, attempt to eliminate locks 2780 int cnt = C->macro_count(); 2781 for (int i=0; i < cnt; i++) { 2782 Node *n = C->macro_node(i); 2783 if (n->is_AbstractLock()) { // Lock and Unlock nodes 2784 // Before elimination mark all associated (same box and obj) 2785 // lock and unlock nodes. 2786 mark_eliminated_locking_nodes(n->as_AbstractLock()); 2787 } 2788 } 2789 bool progress = true; 2790 while (progress) { 2791 progress = false; 2792 for (int i = C->macro_count(); i > 0; i--) { 2793 Node * n = C->macro_node(i-1); 2794 bool success = false; 2795 debug_only(int old_macro_count = C->macro_count();); 2796 if (n->is_AbstractLock()) { 2797 success = eliminate_locking_node(n->as_AbstractLock()); 2798 } 2799 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2800 progress = progress || success; 2801 } 2802 } 2803 // Next, attempt to eliminate allocations 2804 _has_locks = false; 2805 progress = true; 2806 while (progress) { 2807 progress = false; 2808 for (int i = C->macro_count(); i > 0; i--) { 2809 Node * n = C->macro_node(i-1); 2810 bool success = false; 2811 debug_only(int old_macro_count = C->macro_count();); 2812 switch (n->class_id()) { 2813 case Node::Class_Allocate: 2814 case Node::Class_AllocateArray: 2815 success = eliminate_allocate_node(n->as_Allocate()); 2816 break; 2817 case Node::Class_CallStaticJava: { 2818 CallStaticJavaNode* call = n->as_CallStaticJava(); 2819 if (!call->method()->is_method_handle_intrinsic()) { 2820 success = eliminate_boxing_node(n->as_CallStaticJava()); 2821 } 2822 break; 2823 } 2824 case Node::Class_Lock: 2825 case Node::Class_Unlock: 2826 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 2827 _has_locks = true; 2828 break; 2829 case Node::Class_ArrayCopy: 2830 break; 2831 case Node::Class_OuterStripMinedLoop: 2832 break; 2833 default: 2834 assert(n->Opcode() == Op_LoopLimit || 2835 n->Opcode() == Op_Opaque1 || 2836 n->Opcode() == Op_Opaque2 || 2837 n->Opcode() == Op_Opaque3 || 2838 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n), 2839 "unknown node type in macro list"); 2840 } 2841 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2842 progress = progress || success; 2843 } 2844 } 2845 } 2846 2847 //------------------------------expand_macro_nodes---------------------- 2848 // Returns true if a failure occurred. 2849 bool PhaseMacroExpand::expand_macro_nodes() { 2850 // Last attempt to eliminate macro nodes. 2851 eliminate_macro_nodes(); 2852 2853 // Make sure expansion will not cause node limit to be exceeded. 2854 // Worst case is a macro node gets expanded into about 200 nodes. 2855 // Allow 50% more for optimization. 2856 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) ) 2857 return true; 2858 2859 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. 2860 bool progress = true; 2861 while (progress) { 2862 progress = false; 2863 for (int i = C->macro_count(); i > 0; i--) { 2864 Node* n = C->macro_node(i-1); 2865 bool success = false; 2866 debug_only(int old_macro_count = C->macro_count();); 2867 if (n->Opcode() == Op_LoopLimit) { 2868 // Remove it from macro list and put on IGVN worklist to optimize. 2869 C->remove_macro_node(n); 2870 _igvn._worklist.push(n); 2871 success = true; 2872 } else if (n->Opcode() == Op_CallStaticJava) { 2873 CallStaticJavaNode* call = n->as_CallStaticJava(); 2874 if (!call->method()->is_method_handle_intrinsic()) { 2875 // Remove it from macro list and put on IGVN worklist to optimize. 2876 C->remove_macro_node(n); 2877 _igvn._worklist.push(n); 2878 success = true; 2879 } 2880 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { 2881 _igvn.replace_node(n, n->in(1)); 2882 success = true; 2883 #if INCLUDE_RTM_OPT 2884 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) { 2885 assert(C->profile_rtm(), "should be used only in rtm deoptimization code"); 2886 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), ""); 2887 Node* cmp = n->unique_out(); 2888 #ifdef ASSERT 2889 // Validate graph. 2890 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), ""); 2891 BoolNode* bol = cmp->unique_out()->as_Bool(); 2892 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() && 2893 (bol->_test._test == BoolTest::ne), ""); 2894 IfNode* ifn = bol->unique_out()->as_If(); 2895 assert((ifn->outcnt() == 2) && 2896 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, ""); 2897 #endif 2898 Node* repl = n->in(1); 2899 if (!_has_locks) { 2900 // Remove RTM state check if there are no locks in the code. 2901 // Replace input to compare the same value. 2902 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1); 2903 } 2904 _igvn.replace_node(n, repl); 2905 success = true; 2906 #endif 2907 } else if (n->Opcode() == Op_OuterStripMinedLoop) { 2908 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn); 2909 C->remove_macro_node(n); 2910 success = true; 2911 } 2912 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list"); 2913 progress = progress || success; 2914 } 2915 } 2916 2917 // expand arraycopy "macro" nodes first 2918 // For ReduceBulkZeroing, we must first process all arraycopy nodes 2919 // before the allocate nodes are expanded. 2920 for (int i = C->macro_count(); i > 0; i--) { 2921 Node* n = C->macro_node(i-1); 2922 assert(n->is_macro(), "only macro nodes expected here"); 2923 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) { 2924 // node is unreachable, so don't try to expand it 2925 C->remove_macro_node(n); 2926 continue; 2927 } 2928 debug_only(int old_macro_count = C->macro_count();); 2929 switch (n->class_id()) { 2930 case Node::Class_Lock: 2931 expand_lock_node(n->as_Lock()); 2932 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list"); 2933 break; 2934 case Node::Class_Unlock: 2935 expand_unlock_node(n->as_Unlock()); 2936 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list"); 2937 break; 2938 case Node::Class_ArrayCopy: 2939 expand_arraycopy_node(n->as_ArrayCopy()); 2940 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list"); 2941 break; 2942 } 2943 if (C->failing()) return true; 2944 } 2945 2946 // All nodes except Allocate nodes are expanded now. There could be 2947 // new optimization opportunities (such as folding newly created 2948 // load from a just allocated object). Run IGVN. 2949 _igvn.set_delay_transform(false); 2950 _igvn.optimize(); 2951 if (C->failing()) return true; 2952 2953 _igvn.set_delay_transform(true); 2954 2955 // expand "macro" nodes 2956 // nodes are removed from the macro list as they are processed 2957 while (C->macro_count() > 0) { 2958 int macro_count = C->macro_count(); 2959 Node * n = C->macro_node(macro_count-1); 2960 assert(n->is_macro(), "only macro nodes expected here"); 2961 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) { 2962 // node is unreachable, so don't try to expand it 2963 C->remove_macro_node(n); 2964 continue; 2965 } 2966 switch (n->class_id()) { 2967 case Node::Class_Allocate: 2968 expand_allocate(n->as_Allocate()); 2969 break; 2970 case Node::Class_AllocateArray: 2971 expand_allocate_array(n->as_AllocateArray()); 2972 break; 2973 case Node::Class_CallStaticJava: 2974 expand_mh_intrinsic_return(n->as_CallStaticJava()); 2975 C->remove_macro_node(n); 2976 break; 2977 default: 2978 assert(false, "unknown node type in macro list"); 2979 } 2980 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2981 if (C->failing()) return true; 2982 } 2983 2984 _igvn.set_delay_transform(false); 2985 _igvn.optimize(); 2986 if (C->failing()) return true; 2987 return false; 2988 }