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