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