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 Node* n = val->in(MemNode::ValueIn); 438 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 439 n = bs->step_over_gc_barrier(n); 440 values.at_put(j, n); 441 } else if(val->is_Proj() && val->in(0) == alloc) { 442 values.at_put(j, _igvn.zerocon(ft)); 443 } else if (val->is_Phi()) { 444 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); 445 if (val == NULL) { 446 return NULL; 447 } 448 values.at_put(j, val); 449 } else if (val->Opcode() == Op_SCMemProj) { 450 assert(val->in(0)->is_LoadStore() || 451 val->in(0)->Opcode() == Op_EncodeISOArray || 452 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity"); 453 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 454 return NULL; 455 } else if (val->is_ArrayCopy()) { 456 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc); 457 if (res == NULL) { 458 return NULL; 459 } 460 values.at_put(j, res); 461 } else { 462 #ifdef ASSERT 463 val->dump(); 464 assert(false, "unknown node on this path"); 465 #endif 466 return NULL; // unknown node on this path 467 } 468 } 469 } 470 // Set Phi's inputs 471 for (uint j = 1; j < length; j++) { 472 if (values.at(j) == mem) { 473 phi->init_req(j, phi); 474 } else { 475 phi->init_req(j, values.at(j)); 476 } 477 } 478 return phi; 479 } 480 481 // Search the last value stored into the object's field. 482 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) { 483 assert(adr_t->is_known_instance_field(), "instance required"); 484 int instance_id = adr_t->instance_id(); 485 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 486 487 int alias_idx = C->get_alias_index(adr_t); 488 int offset = adr_t->offset(); 489 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 490 Node *alloc_ctrl = alloc->in(TypeFunc::Control); 491 Node *alloc_mem = alloc->in(TypeFunc::Memory); 492 Arena *a = Thread::current()->resource_area(); 493 VectorSet visited(a); 494 495 496 bool done = sfpt_mem == alloc_mem; 497 Node *mem = sfpt_mem; 498 while (!done) { 499 if (visited.test_set(mem->_idx)) { 500 return NULL; // found a loop, give up 501 } 502 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 503 if (mem == start_mem || mem == alloc_mem) { 504 done = true; // hit a sentinel, return appropriate 0 value 505 } else if (mem->is_Initialize()) { 506 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 507 if (mem == NULL) { 508 done = true; // Something go wrong. 509 } else if (mem->is_Store()) { 510 const TypePtr* atype = mem->as_Store()->adr_type(); 511 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 512 done = true; 513 } 514 } else if (mem->is_Store()) { 515 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 516 assert(atype != NULL, "address type must be oopptr"); 517 assert(C->get_alias_index(atype) == alias_idx && 518 atype->is_known_instance_field() && atype->offset() == offset && 519 atype->instance_id() == instance_id, "store is correct memory slice"); 520 done = true; 521 } else if (mem->is_Phi()) { 522 // try to find a phi's unique input 523 Node *unique_input = NULL; 524 Node *top = C->top(); 525 for (uint i = 1; i < mem->req(); i++) { 526 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 527 if (n == NULL || n == top || n == mem) { 528 continue; 529 } else if (unique_input == NULL) { 530 unique_input = n; 531 } else if (unique_input != n) { 532 unique_input = top; 533 break; 534 } 535 } 536 if (unique_input != NULL && unique_input != top) { 537 mem = unique_input; 538 } else { 539 done = true; 540 } 541 } else if (mem->is_ArrayCopy()) { 542 done = true; 543 } else { 544 assert(false, "unexpected node"); 545 } 546 } 547 if (mem != NULL) { 548 if (mem == start_mem || mem == alloc_mem) { 549 // hit a sentinel, return appropriate 0 value 550 return _igvn.zerocon(ft); 551 } else if (mem->is_Store()) { 552 Node* n = mem->in(MemNode::ValueIn); 553 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 554 n = bs->step_over_gc_barrier(n); 555 return n; 556 } else if (mem->is_Phi()) { 557 // attempt to produce a Phi reflecting the values on the input paths of the Phi 558 Node_Stack value_phis(a, 8); 559 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 560 if (phi != NULL) { 561 return phi; 562 } else { 563 // Kill all new Phis 564 while(value_phis.is_nonempty()) { 565 Node* n = value_phis.node(); 566 _igvn.replace_node(n, C->top()); 567 value_phis.pop(); 568 } 569 } 570 } else if (mem->is_ArrayCopy()) { 571 Node* ctl = mem->in(0); 572 Node* m = mem->in(TypeFunc::Memory); 573 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) { 574 // pin the loads in the uncommon trap path 575 ctl = sfpt_ctl; 576 m = sfpt_mem; 577 } 578 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc); 579 } 580 } 581 // Something go wrong. 582 return NULL; 583 } 584 585 // Check the possibility of scalar replacement. 586 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 587 // Scan the uses of the allocation to check for anything that would 588 // prevent us from eliminating it. 589 NOT_PRODUCT( const char* fail_eliminate = NULL; ) 590 DEBUG_ONLY( Node* disq_node = NULL; ) 591 bool can_eliminate = true; 592 593 Node* res = alloc->result_cast(); 594 const TypeOopPtr* res_type = NULL; 595 if (res == NULL) { 596 // All users were eliminated. 597 } else if (!res->is_CheckCastPP()) { 598 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 599 can_eliminate = false; 600 } else { 601 res_type = _igvn.type(res)->isa_oopptr(); 602 if (res_type == NULL) { 603 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 604 can_eliminate = false; 605 } else if (res_type->isa_aryptr()) { 606 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 607 if (length < 0) { 608 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 609 can_eliminate = false; 610 } 611 } 612 } 613 614 if (can_eliminate && res != NULL) { 615 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 616 j < jmax && can_eliminate; j++) { 617 Node* use = res->fast_out(j); 618 619 if (use->is_AddP()) { 620 const TypePtr* addp_type = _igvn.type(use)->is_ptr(); 621 int offset = addp_type->offset(); 622 623 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 624 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) 625 can_eliminate = false; 626 break; 627 } 628 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 629 k < kmax && can_eliminate; k++) { 630 Node* n = use->fast_out(k); 631 if (!n->is_Store() && n->Opcode() != Op_CastP2X && 632 !(n->is_ArrayCopy() && 633 n->as_ArrayCopy()->is_clonebasic() && 634 n->in(ArrayCopyNode::Dest) == use)) { 635 DEBUG_ONLY(disq_node = n;) 636 if (n->is_Load() || n->is_LoadStore()) { 637 NOT_PRODUCT(fail_eliminate = "Field load";) 638 } else { 639 NOT_PRODUCT(fail_eliminate = "Not store field referrence";) 640 } 641 can_eliminate = false; 642 } 643 } 644 } else if (use->is_ArrayCopy() && 645 (use->as_ArrayCopy()->is_arraycopy_validated() || 646 use->as_ArrayCopy()->is_copyof_validated() || 647 use->as_ArrayCopy()->is_copyofrange_validated()) && 648 use->in(ArrayCopyNode::Dest) == res) { 649 // ok to eliminate 650 } else if (use->is_SafePoint()) { 651 SafePointNode* sfpt = use->as_SafePoint(); 652 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 653 // Object is passed as argument. 654 DEBUG_ONLY(disq_node = use;) 655 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 656 can_eliminate = false; 657 } 658 Node* sfptMem = sfpt->memory(); 659 if (sfptMem == NULL || sfptMem->is_top()) { 660 DEBUG_ONLY(disq_node = use;) 661 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) 662 can_eliminate = false; 663 } else { 664 safepoints.append_if_missing(sfpt); 665 } 666 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 667 if (use->is_Phi()) { 668 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 669 NOT_PRODUCT(fail_eliminate = "Object is return value";) 670 } else { 671 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 672 } 673 DEBUG_ONLY(disq_node = use;) 674 } else { 675 if (use->Opcode() == Op_Return) { 676 NOT_PRODUCT(fail_eliminate = "Object is return value";) 677 }else { 678 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 679 } 680 DEBUG_ONLY(disq_node = use;) 681 } 682 can_eliminate = false; 683 } 684 } 685 } 686 687 #ifndef PRODUCT 688 if (PrintEliminateAllocations) { 689 if (can_eliminate) { 690 tty->print("Scalar "); 691 if (res == NULL) 692 alloc->dump(); 693 else 694 res->dump(); 695 } else if (alloc->_is_scalar_replaceable) { 696 tty->print("NotScalar (%s)", fail_eliminate); 697 if (res == NULL) 698 alloc->dump(); 699 else 700 res->dump(); 701 #ifdef ASSERT 702 if (disq_node != NULL) { 703 tty->print(" >>>> "); 704 disq_node->dump(); 705 } 706 #endif /*ASSERT*/ 707 } 708 } 709 #endif 710 return can_eliminate; 711 } 712 713 // Do scalar replacement. 714 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 715 GrowableArray <SafePointNode *> safepoints_done; 716 717 ciKlass* klass = NULL; 718 ciInstanceKlass* iklass = NULL; 719 int nfields = 0; 720 int array_base = 0; 721 int element_size = 0; 722 BasicType basic_elem_type = T_ILLEGAL; 723 ciType* elem_type = NULL; 724 725 Node* res = alloc->result_cast(); 726 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result"); 727 const TypeOopPtr* res_type = NULL; 728 if (res != NULL) { // Could be NULL when there are no users 729 res_type = _igvn.type(res)->isa_oopptr(); 730 } 731 732 if (res != NULL) { 733 klass = res_type->klass(); 734 if (res_type->isa_instptr()) { 735 // find the fields of the class which will be needed for safepoint debug information 736 assert(klass->is_instance_klass(), "must be an instance klass."); 737 iklass = klass->as_instance_klass(); 738 nfields = iklass->nof_nonstatic_fields(); 739 } else { 740 // find the array's elements which will be needed for safepoint debug information 741 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 742 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); 743 elem_type = klass->as_array_klass()->element_type(); 744 basic_elem_type = elem_type->basic_type(); 745 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 746 element_size = type2aelembytes(basic_elem_type); 747 } 748 } 749 // 750 // Process the safepoint uses 751 // 752 while (safepoints.length() > 0) { 753 SafePointNode* sfpt = safepoints.pop(); 754 Node* mem = sfpt->memory(); 755 Node* ctl = sfpt->control(); 756 assert(sfpt->jvms() != NULL, "missed JVMS"); 757 // Fields of scalar objs are referenced only at the end 758 // of regular debuginfo at the last (youngest) JVMS. 759 // Record relative start index. 760 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff()); 761 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, 762 #ifdef ASSERT 763 alloc, 764 #endif 765 first_ind, nfields); 766 sobj->init_req(0, C->root()); 767 transform_later(sobj); 768 769 // Scan object's fields adding an input to the safepoint for each field. 770 for (int j = 0; j < nfields; j++) { 771 intptr_t offset; 772 ciField* field = NULL; 773 if (iklass != NULL) { 774 field = iklass->nonstatic_field_at(j); 775 offset = field->offset(); 776 elem_type = field->type(); 777 basic_elem_type = field->layout_type(); 778 } else { 779 offset = array_base + j * (intptr_t)element_size; 780 } 781 782 const Type *field_type; 783 // The next code is taken from Parse::do_get_xxx(). 784 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { 785 if (!elem_type->is_loaded()) { 786 field_type = TypeInstPtr::BOTTOM; 787 } else if (field != NULL && field->is_static_constant()) { 788 // This can happen if the constant oop is non-perm. 789 ciObject* con = field->constant_value().as_object(); 790 // Do not "join" in the previous type; it doesn't add value, 791 // and may yield a vacuous result if the field is of interface type. 792 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 793 assert(field_type != NULL, "field singleton type must be consistent"); 794 } else { 795 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 796 } 797 if (UseCompressedOops) { 798 field_type = field_type->make_narrowoop(); 799 basic_elem_type = T_NARROWOOP; 800 } 801 } else { 802 field_type = Type::get_const_basic_type(basic_elem_type); 803 } 804 805 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 806 807 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc); 808 if (field_val == NULL) { 809 // We weren't able to find a value for this field, 810 // give up on eliminating this allocation. 811 812 // Remove any extra entries we added to the safepoint. 813 uint last = sfpt->req() - 1; 814 for (int k = 0; k < j; k++) { 815 sfpt->del_req(last--); 816 } 817 _igvn._worklist.push(sfpt); 818 // rollback processed safepoints 819 while (safepoints_done.length() > 0) { 820 SafePointNode* sfpt_done = safepoints_done.pop(); 821 // remove any extra entries we added to the safepoint 822 last = sfpt_done->req() - 1; 823 for (int k = 0; k < nfields; k++) { 824 sfpt_done->del_req(last--); 825 } 826 JVMState *jvms = sfpt_done->jvms(); 827 jvms->set_endoff(sfpt_done->req()); 828 // Now make a pass over the debug information replacing any references 829 // to SafePointScalarObjectNode with the allocated object. 830 int start = jvms->debug_start(); 831 int end = jvms->debug_end(); 832 for (int i = start; i < end; i++) { 833 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 834 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 835 if (scobj->first_index(jvms) == sfpt_done->req() && 836 scobj->n_fields() == (uint)nfields) { 837 assert(scobj->alloc() == alloc, "sanity"); 838 sfpt_done->set_req(i, res); 839 } 840 } 841 } 842 _igvn._worklist.push(sfpt_done); 843 } 844 #ifndef PRODUCT 845 if (PrintEliminateAllocations) { 846 if (field != NULL) { 847 tty->print("=== At SafePoint node %d can't find value of Field: ", 848 sfpt->_idx); 849 field->print(); 850 int field_idx = C->get_alias_index(field_addr_type); 851 tty->print(" (alias_idx=%d)", field_idx); 852 } else { // Array's element 853 tty->print("=== At SafePoint node %d can't find value of array element [%d]", 854 sfpt->_idx, j); 855 } 856 tty->print(", which prevents elimination of: "); 857 if (res == NULL) 858 alloc->dump(); 859 else 860 res->dump(); 861 } 862 #endif 863 return false; 864 } 865 if (UseCompressedOops && field_type->isa_narrowoop()) { 866 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 867 // to be able scalar replace the allocation. 868 if (field_val->is_EncodeP()) { 869 field_val = field_val->in(1); 870 } else { 871 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type())); 872 } 873 } 874 sfpt->add_req(field_val); 875 } 876 JVMState *jvms = sfpt->jvms(); 877 jvms->set_endoff(sfpt->req()); 878 // Now make a pass over the debug information replacing any references 879 // to the allocated object with "sobj" 880 int start = jvms->debug_start(); 881 int end = jvms->debug_end(); 882 sfpt->replace_edges_in_range(res, sobj, start, end); 883 _igvn._worklist.push(sfpt); 884 safepoints_done.append_if_missing(sfpt); // keep it for rollback 885 } 886 return true; 887 } 888 889 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) { 890 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control); 891 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory); 892 if (ctl_proj != NULL) { 893 igvn.replace_node(ctl_proj, n->in(0)); 894 } 895 if (mem_proj != NULL) { 896 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory)); 897 } 898 } 899 900 // Process users of eliminated allocation. 901 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) { 902 Node* res = alloc->result_cast(); 903 if (res != NULL) { 904 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 905 Node *use = res->last_out(j); 906 uint oc1 = res->outcnt(); 907 908 if (use->is_AddP()) { 909 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 910 Node *n = use->last_out(k); 911 uint oc2 = use->outcnt(); 912 if (n->is_Store()) { 913 #ifdef ASSERT 914 // Verify that there is no dependent MemBarVolatile nodes, 915 // they should be removed during IGVN, see MemBarNode::Ideal(). 916 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); 917 p < pmax; p++) { 918 Node* mb = n->fast_out(p); 919 assert(mb->is_Initialize() || !mb->is_MemBar() || 920 mb->req() <= MemBarNode::Precedent || 921 mb->in(MemBarNode::Precedent) != n, 922 "MemBarVolatile should be eliminated for non-escaping object"); 923 } 924 #endif 925 _igvn.replace_node(n, n->in(MemNode::Memory)); 926 } else if (n->is_ArrayCopy()) { 927 // Disconnect ArrayCopy node 928 ArrayCopyNode* ac = n->as_ArrayCopy(); 929 assert(ac->is_clonebasic(), "unexpected array copy kind"); 930 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out(); 931 disconnect_projections(ac, _igvn); 932 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation"); 933 Node* membar_before = alloc->in(0)->in(0); 934 disconnect_projections(membar_before->as_MemBar(), _igvn); 935 if (membar_after->is_MemBar()) { 936 disconnect_projections(membar_after->as_MemBar(), _igvn); 937 } 938 } else { 939 eliminate_gc_barrier(n); 940 } 941 k -= (oc2 - use->outcnt()); 942 } 943 _igvn.remove_dead_node(use); 944 } else if (use->is_ArrayCopy()) { 945 // Disconnect ArrayCopy node 946 ArrayCopyNode* ac = use->as_ArrayCopy(); 947 assert(ac->is_arraycopy_validated() || 948 ac->is_copyof_validated() || 949 ac->is_copyofrange_validated(), "unsupported"); 950 CallProjections callprojs; 951 ac->extract_projections(&callprojs, true); 952 953 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O)); 954 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory)); 955 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control)); 956 957 // Set control to top. IGVN will remove the remaining projections 958 ac->set_req(0, top()); 959 ac->replace_edge(res, top()); 960 961 // Disconnect src right away: it can help find new 962 // opportunities for allocation elimination 963 Node* src = ac->in(ArrayCopyNode::Src); 964 ac->replace_edge(src, top()); 965 // src can be top at this point if src and dest of the 966 // arraycopy were the same 967 if (src->outcnt() == 0 && !src->is_top()) { 968 _igvn.remove_dead_node(src); 969 } 970 971 _igvn._worklist.push(ac); 972 } else { 973 eliminate_gc_barrier(use); 974 } 975 j -= (oc1 - res->outcnt()); 976 } 977 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 978 _igvn.remove_dead_node(res); 979 } 980 981 // 982 // Process other users of allocation's projections 983 // 984 if (_resproj != NULL && _resproj->outcnt() != 0) { 985 // First disconnect stores captured by Initialize node. 986 // If Initialize node is eliminated first in the following code, 987 // it will kill such stores and DUIterator_Last will assert. 988 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) { 989 Node *use = _resproj->fast_out(j); 990 if (use->is_AddP()) { 991 // raw memory addresses used only by the initialization 992 _igvn.replace_node(use, C->top()); 993 --j; --jmax; 994 } 995 } 996 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { 997 Node *use = _resproj->last_out(j); 998 uint oc1 = _resproj->outcnt(); 999 if (use->is_Initialize()) { 1000 // Eliminate Initialize node. 1001 InitializeNode *init = use->as_Initialize(); 1002 assert(init->outcnt() <= 2, "only a control and memory projection expected"); 1003 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control); 1004 if (ctrl_proj != NULL) { 1005 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection"); 1006 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj); 1007 } 1008 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory); 1009 if (mem_proj != NULL) { 1010 Node *mem = init->in(TypeFunc::Memory); 1011 #ifdef ASSERT 1012 if (mem->is_MergeMem()) { 1013 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection"); 1014 } else { 1015 assert(mem == _memproj_fallthrough, "allocation memory projection"); 1016 } 1017 #endif 1018 _igvn.replace_node(mem_proj, mem); 1019 } 1020 } else { 1021 assert(false, "only Initialize or AddP expected"); 1022 } 1023 j -= (oc1 - _resproj->outcnt()); 1024 } 1025 } 1026 if (_fallthroughcatchproj != NULL) { 1027 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); 1028 } 1029 if (_memproj_fallthrough != NULL) { 1030 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); 1031 } 1032 if (_memproj_catchall != NULL) { 1033 _igvn.replace_node(_memproj_catchall, C->top()); 1034 } 1035 if (_ioproj_fallthrough != NULL) { 1036 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); 1037 } 1038 if (_ioproj_catchall != NULL) { 1039 _igvn.replace_node(_ioproj_catchall, C->top()); 1040 } 1041 if (_catchallcatchproj != NULL) { 1042 _igvn.replace_node(_catchallcatchproj, C->top()); 1043 } 1044 } 1045 1046 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 1047 // Don't do scalar replacement if the frame can be popped by JVMTI: 1048 // if reallocation fails during deoptimization we'll pop all 1049 // interpreter frames for this compiled frame and that won't play 1050 // nice with JVMTI popframe. 1051 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) { 1052 return false; 1053 } 1054 Node* klass = alloc->in(AllocateNode::KlassNode); 1055 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); 1056 Node* res = alloc->result_cast(); 1057 // Eliminate boxing allocations which are not used 1058 // regardless scalar replacable status. 1059 bool boxing_alloc = C->eliminate_boxing() && 1060 tklass->klass()->is_instance_klass() && 1061 tklass->klass()->as_instance_klass()->is_box_klass(); 1062 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) { 1063 return false; 1064 } 1065 1066 extract_call_projections(alloc); 1067 1068 GrowableArray <SafePointNode *> safepoints; 1069 if (!can_eliminate_allocation(alloc, safepoints)) { 1070 return false; 1071 } 1072 1073 if (!alloc->_is_scalar_replaceable) { 1074 assert(res == NULL, "sanity"); 1075 // We can only eliminate allocation if all debug info references 1076 // are already replaced with SafePointScalarObject because 1077 // we can't search for a fields value without instance_id. 1078 if (safepoints.length() > 0) { 1079 return false; 1080 } 1081 } 1082 1083 if (!scalar_replacement(alloc, safepoints)) { 1084 return false; 1085 } 1086 1087 CompileLog* log = C->log(); 1088 if (log != NULL) { 1089 log->head("eliminate_allocation type='%d'", 1090 log->identify(tklass->klass())); 1091 JVMState* p = alloc->jvms(); 1092 while (p != NULL) { 1093 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1094 p = p->caller(); 1095 } 1096 log->tail("eliminate_allocation"); 1097 } 1098 1099 process_users_of_allocation(alloc); 1100 1101 #ifndef PRODUCT 1102 if (PrintEliminateAllocations) { 1103 if (alloc->is_AllocateArray()) 1104 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1105 else 1106 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1107 } 1108 #endif 1109 1110 return true; 1111 } 1112 1113 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) { 1114 // EA should remove all uses of non-escaping boxing node. 1115 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) { 1116 return false; 1117 } 1118 1119 assert(boxing->result_cast() == NULL, "unexpected boxing node result"); 1120 1121 extract_call_projections(boxing); 1122 1123 const TypeTuple* r = boxing->tf()->range(); 1124 assert(r->cnt() > TypeFunc::Parms, "sanity"); 1125 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr(); 1126 assert(t != NULL, "sanity"); 1127 1128 CompileLog* log = C->log(); 1129 if (log != NULL) { 1130 log->head("eliminate_boxing type='%d'", 1131 log->identify(t->klass())); 1132 JVMState* p = boxing->jvms(); 1133 while (p != NULL) { 1134 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1135 p = p->caller(); 1136 } 1137 log->tail("eliminate_boxing"); 1138 } 1139 1140 process_users_of_allocation(boxing); 1141 1142 #ifndef PRODUCT 1143 if (PrintEliminateAllocations) { 1144 tty->print("++++ Eliminated: %d ", boxing->_idx); 1145 boxing->method()->print_short_name(tty); 1146 tty->cr(); 1147 } 1148 #endif 1149 1150 return true; 1151 } 1152 1153 //---------------------------set_eden_pointers------------------------- 1154 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { 1155 if (UseTLAB) { // Private allocation: load from TLS 1156 Node* thread = transform_later(new ThreadLocalNode()); 1157 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); 1158 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); 1159 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); 1160 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); 1161 } else { // Shared allocation: load from globals 1162 CollectedHeap* ch = Universe::heap(); 1163 address top_adr = (address)ch->top_addr(); 1164 address end_adr = (address)ch->end_addr(); 1165 eden_top_adr = makecon(TypeRawPtr::make(top_adr)); 1166 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); 1167 } 1168 } 1169 1170 1171 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 1172 Node* adr = basic_plus_adr(base, offset); 1173 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 1174 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered); 1175 transform_later(value); 1176 return value; 1177 } 1178 1179 1180 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 1181 Node* adr = basic_plus_adr(base, offset); 1182 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered); 1183 transform_later(mem); 1184 return mem; 1185 } 1186 1187 //============================================================================= 1188 // 1189 // A L L O C A T I O N 1190 // 1191 // Allocation attempts to be fast in the case of frequent small objects. 1192 // It breaks down like this: 1193 // 1194 // 1) Size in doublewords is computed. This is a constant for objects and 1195 // variable for most arrays. Doubleword units are used to avoid size 1196 // overflow of huge doubleword arrays. We need doublewords in the end for 1197 // rounding. 1198 // 1199 // 2) Size is checked for being 'too large'. Too-large allocations will go 1200 // the slow path into the VM. The slow path can throw any required 1201 // exceptions, and does all the special checks for very large arrays. The 1202 // size test can constant-fold away for objects. For objects with 1203 // finalizers it constant-folds the otherway: you always go slow with 1204 // finalizers. 1205 // 1206 // 3) If NOT using TLABs, this is the contended loop-back point. 1207 // Load-Locked the heap top. If using TLABs normal-load the heap top. 1208 // 1209 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 1210 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 1211 // "size*8" we always enter the VM, where "largish" is a constant picked small 1212 // enough that there's always space between the eden max and 4Gig (old space is 1213 // there so it's quite large) and large enough that the cost of entering the VM 1214 // is dwarfed by the cost to initialize the space. 1215 // 1216 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 1217 // down. If contended, repeat at step 3. If using TLABs normal-store 1218 // adjusted heap top back down; there is no contention. 1219 // 1220 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 1221 // fields. 1222 // 1223 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 1224 // oop flavor. 1225 // 1226 //============================================================================= 1227 // FastAllocateSizeLimit value is in DOUBLEWORDS. 1228 // Allocations bigger than this always go the slow route. 1229 // This value must be small enough that allocation attempts that need to 1230 // trigger exceptions go the slow route. Also, it must be small enough so 1231 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 1232 //=============================================================================j// 1233 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 1234 // The allocator will coalesce int->oop copies away. See comment in 1235 // coalesce.cpp about how this works. It depends critically on the exact 1236 // code shape produced here, so if you are changing this code shape 1237 // make sure the GC info for the heap-top is correct in and around the 1238 // slow-path call. 1239 // 1240 1241 void PhaseMacroExpand::expand_allocate_common( 1242 AllocateNode* alloc, // allocation node to be expanded 1243 Node* length, // array length for an array allocation 1244 const TypeFunc* slow_call_type, // Type of slow call 1245 address slow_call_address // Address of slow call 1246 ) 1247 { 1248 1249 Node* ctrl = alloc->in(TypeFunc::Control); 1250 Node* mem = alloc->in(TypeFunc::Memory); 1251 Node* i_o = alloc->in(TypeFunc::I_O); 1252 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 1253 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1254 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 1255 1256 assert(ctrl != NULL, "must have control"); 1257 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 1258 // they will not be used if "always_slow" is set 1259 enum { slow_result_path = 1, fast_result_path = 2 }; 1260 Node *result_region = NULL; 1261 Node *result_phi_rawmem = NULL; 1262 Node *result_phi_rawoop = NULL; 1263 Node *result_phi_i_o = NULL; 1264 1265 // The initial slow comparison is a size check, the comparison 1266 // we want to do is a BoolTest::gt 1267 bool always_slow = false; 1268 int tv = _igvn.find_int_con(initial_slow_test, -1); 1269 if (tv >= 0) { 1270 always_slow = (tv == 1); 1271 initial_slow_test = NULL; 1272 } else { 1273 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 1274 } 1275 1276 if (C->env()->dtrace_alloc_probes() || 1277 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) { 1278 // Force slow-path allocation 1279 always_slow = true; 1280 initial_slow_test = NULL; 1281 } 1282 1283 1284 enum { too_big_or_final_path = 1, need_gc_path = 2 }; 1285 Node *slow_region = NULL; 1286 Node *toobig_false = ctrl; 1287 1288 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); 1289 // generate the initial test if necessary 1290 if (initial_slow_test != NULL ) { 1291 slow_region = new RegionNode(3); 1292 1293 // Now make the initial failure test. Usually a too-big test but 1294 // might be a TRUE for finalizers or a fancy class check for 1295 // newInstance0. 1296 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1297 transform_later(toobig_iff); 1298 // Plug the failing-too-big test into the slow-path region 1299 Node *toobig_true = new IfTrueNode( toobig_iff ); 1300 transform_later(toobig_true); 1301 slow_region ->init_req( too_big_or_final_path, toobig_true ); 1302 toobig_false = new IfFalseNode( toobig_iff ); 1303 transform_later(toobig_false); 1304 } else { // No initial test, just fall into next case 1305 toobig_false = ctrl; 1306 debug_only(slow_region = NodeSentinel); 1307 } 1308 1309 Node *slow_mem = mem; // save the current memory state for slow path 1310 // generate the fast allocation code unless we know that the initial test will always go slow 1311 if (!always_slow) { 1312 // Fast path modifies only raw memory. 1313 if (mem->is_MergeMem()) { 1314 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1315 } 1316 1317 // allocate the Region and Phi nodes for the result 1318 result_region = new RegionNode(3); 1319 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1320 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM); 1321 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch 1322 1323 // Grab regular I/O before optional prefetch may change it. 1324 // Slow-path does no I/O so just set it to the original I/O. 1325 result_phi_i_o->init_req(slow_result_path, i_o); 1326 1327 Node* needgc_ctrl = NULL; 1328 // Name successful fast-path variables 1329 Node* fast_oop_ctrl; 1330 Node* fast_oop_rawmem; 1331 1332 intx prefetch_lines = length != NULL ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1333 1334 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1335 Node* fast_oop = bs->obj_allocate(this, ctrl, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl, 1336 fast_oop_ctrl, fast_oop_rawmem, 1337 prefetch_lines); 1338 1339 if (initial_slow_test) { 1340 slow_region->init_req(need_gc_path, needgc_ctrl); 1341 // This completes all paths into the slow merge point 1342 transform_later(slow_region); 1343 } else { // No initial slow path needed! 1344 // Just fall from the need-GC path straight into the VM call. 1345 slow_region = needgc_ctrl; 1346 } 1347 1348 InitializeNode* init = alloc->initialization(); 1349 fast_oop_rawmem = initialize_object(alloc, 1350 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1351 klass_node, length, size_in_bytes); 1352 1353 // If initialization is performed by an array copy, any required 1354 // MemBarStoreStore was already added. If the object does not 1355 // escape no need for a MemBarStoreStore. If the object does not 1356 // escape in its initializer and memory barrier (MemBarStoreStore or 1357 // stronger) is already added at exit of initializer, also no need 1358 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore 1359 // so that stores that initialize this object can't be reordered 1360 // with a subsequent store that makes this object accessible by 1361 // other threads. 1362 // Other threads include java threads and JVM internal threads 1363 // (for example concurrent GC threads). Current concurrent GC 1364 // implementation: CMS and G1 will not scan newly created object, 1365 // so it's safe to skip storestore barrier when allocation does 1366 // not escape. 1367 if (!alloc->does_not_escape_thread() && 1368 !alloc->is_allocation_MemBar_redundant() && 1369 (init == NULL || !init->is_complete_with_arraycopy())) { 1370 if (init == NULL || init->req() < InitializeNode::RawStores) { 1371 // No InitializeNode or no stores captured by zeroing 1372 // elimination. Simply add the MemBarStoreStore after object 1373 // initialization. 1374 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1375 transform_later(mb); 1376 1377 mb->init_req(TypeFunc::Memory, fast_oop_rawmem); 1378 mb->init_req(TypeFunc::Control, fast_oop_ctrl); 1379 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control); 1380 transform_later(fast_oop_ctrl); 1381 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory); 1382 transform_later(fast_oop_rawmem); 1383 } else { 1384 // Add the MemBarStoreStore after the InitializeNode so that 1385 // all stores performing the initialization that were moved 1386 // before the InitializeNode happen before the storestore 1387 // barrier. 1388 1389 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control); 1390 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory); 1391 1392 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1393 transform_later(mb); 1394 1395 Node* ctrl = new ProjNode(init,TypeFunc::Control); 1396 transform_later(ctrl); 1397 Node* mem = new ProjNode(init,TypeFunc::Memory); 1398 transform_later(mem); 1399 1400 // The MemBarStoreStore depends on control and memory coming 1401 // from the InitializeNode 1402 mb->init_req(TypeFunc::Memory, mem); 1403 mb->init_req(TypeFunc::Control, ctrl); 1404 1405 ctrl = new ProjNode(mb,TypeFunc::Control); 1406 transform_later(ctrl); 1407 mem = new ProjNode(mb,TypeFunc::Memory); 1408 transform_later(mem); 1409 1410 // All nodes that depended on the InitializeNode for control 1411 // and memory must now depend on the MemBarNode that itself 1412 // depends on the InitializeNode 1413 if (init_ctrl != NULL) { 1414 _igvn.replace_node(init_ctrl, ctrl); 1415 } 1416 if (init_mem != NULL) { 1417 _igvn.replace_node(init_mem, mem); 1418 } 1419 } 1420 } 1421 1422 if (C->env()->dtrace_extended_probes()) { 1423 // Slow-path call 1424 int size = TypeFunc::Parms + 2; 1425 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1426 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), 1427 "dtrace_object_alloc", 1428 TypeRawPtr::BOTTOM); 1429 1430 // Get base of thread-local storage area 1431 Node* thread = new ThreadLocalNode(); 1432 transform_later(thread); 1433 1434 call->init_req(TypeFunc::Parms+0, thread); 1435 call->init_req(TypeFunc::Parms+1, fast_oop); 1436 call->init_req(TypeFunc::Control, fast_oop_ctrl); 1437 call->init_req(TypeFunc::I_O , top()); // does no i/o 1438 call->init_req(TypeFunc::Memory , fast_oop_rawmem); 1439 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1440 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1441 transform_later(call); 1442 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control); 1443 transform_later(fast_oop_ctrl); 1444 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory); 1445 transform_later(fast_oop_rawmem); 1446 } 1447 1448 // Plug in the successful fast-path into the result merge point 1449 result_region ->init_req(fast_result_path, fast_oop_ctrl); 1450 result_phi_rawoop->init_req(fast_result_path, fast_oop); 1451 result_phi_i_o ->init_req(fast_result_path, i_o); 1452 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); 1453 } else { 1454 slow_region = ctrl; 1455 result_phi_i_o = i_o; // Rename it to use in the following code. 1456 } 1457 1458 // Generate slow-path call 1459 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address, 1460 OptoRuntime::stub_name(slow_call_address), 1461 alloc->jvms()->bci(), 1462 TypePtr::BOTTOM); 1463 call->init_req( TypeFunc::Control, slow_region ); 1464 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1465 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs 1466 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1467 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1468 1469 call->init_req(TypeFunc::Parms+0, klass_node); 1470 if (length != NULL) { 1471 call->init_req(TypeFunc::Parms+1, length); 1472 } 1473 1474 // Copy debug information and adjust JVMState information, then replace 1475 // allocate node with the call 1476 copy_call_debug_info((CallNode *) alloc, call); 1477 if (!always_slow) { 1478 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1479 } else { 1480 // Hook i_o projection to avoid its elimination during allocation 1481 // replacement (when only a slow call is generated). 1482 call->set_req(TypeFunc::I_O, result_phi_i_o); 1483 } 1484 _igvn.replace_node(alloc, call); 1485 transform_later(call); 1486 1487 // Identify the output projections from the allocate node and 1488 // adjust any references to them. 1489 // The control and io projections look like: 1490 // 1491 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1492 // Allocate Catch 1493 // ^---Proj(io) <-------+ ^---CatchProj(io) 1494 // 1495 // We are interested in the CatchProj nodes. 1496 // 1497 extract_call_projections(call); 1498 1499 // An allocate node has separate memory projections for the uses on 1500 // the control and i_o paths. Replace the control memory projection with 1501 // result_phi_rawmem (unless we are only generating a slow call when 1502 // both memory projections are combined) 1503 if (!always_slow && _memproj_fallthrough != NULL) { 1504 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { 1505 Node *use = _memproj_fallthrough->fast_out(i); 1506 _igvn.rehash_node_delayed(use); 1507 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); 1508 // back up iterator 1509 --i; 1510 } 1511 } 1512 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete 1513 // _memproj_catchall so we end up with a call that has only 1 memory projection. 1514 if (_memproj_catchall != NULL ) { 1515 if (_memproj_fallthrough == NULL) { 1516 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory); 1517 transform_later(_memproj_fallthrough); 1518 } 1519 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { 1520 Node *use = _memproj_catchall->fast_out(i); 1521 _igvn.rehash_node_delayed(use); 1522 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); 1523 // back up iterator 1524 --i; 1525 } 1526 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted"); 1527 _igvn.remove_dead_node(_memproj_catchall); 1528 } 1529 1530 // An allocate node has separate i_o projections for the uses on the control 1531 // and i_o paths. Always replace the control i_o projection with result i_o 1532 // otherwise incoming i_o become dead when only a slow call is generated 1533 // (it is different from memory projections where both projections are 1534 // combined in such case). 1535 if (_ioproj_fallthrough != NULL) { 1536 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { 1537 Node *use = _ioproj_fallthrough->fast_out(i); 1538 _igvn.rehash_node_delayed(use); 1539 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); 1540 // back up iterator 1541 --i; 1542 } 1543 } 1544 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete 1545 // _ioproj_catchall so we end up with a call that has only 1 i_o projection. 1546 if (_ioproj_catchall != NULL ) { 1547 if (_ioproj_fallthrough == NULL) { 1548 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O); 1549 transform_later(_ioproj_fallthrough); 1550 } 1551 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { 1552 Node *use = _ioproj_catchall->fast_out(i); 1553 _igvn.rehash_node_delayed(use); 1554 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); 1555 // back up iterator 1556 --i; 1557 } 1558 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted"); 1559 _igvn.remove_dead_node(_ioproj_catchall); 1560 } 1561 1562 // if we generated only a slow call, we are done 1563 if (always_slow) { 1564 // Now we can unhook i_o. 1565 if (result_phi_i_o->outcnt() > 1) { 1566 call->set_req(TypeFunc::I_O, top()); 1567 } else { 1568 assert(result_phi_i_o->unique_ctrl_out() == call, ""); 1569 // Case of new array with negative size known during compilation. 1570 // AllocateArrayNode::Ideal() optimization disconnect unreachable 1571 // following code since call to runtime will throw exception. 1572 // As result there will be no users of i_o after the call. 1573 // Leave i_o attached to this call to avoid problems in preceding graph. 1574 } 1575 return; 1576 } 1577 1578 1579 if (_fallthroughcatchproj != NULL) { 1580 ctrl = _fallthroughcatchproj->clone(); 1581 transform_later(ctrl); 1582 _igvn.replace_node(_fallthroughcatchproj, result_region); 1583 } else { 1584 ctrl = top(); 1585 } 1586 Node *slow_result; 1587 if (_resproj == NULL) { 1588 // no uses of the allocation result 1589 slow_result = top(); 1590 } else { 1591 slow_result = _resproj->clone(); 1592 transform_later(slow_result); 1593 _igvn.replace_node(_resproj, result_phi_rawoop); 1594 } 1595 1596 // Plug slow-path into result merge point 1597 result_region ->init_req( slow_result_path, ctrl ); 1598 result_phi_rawoop->init_req( slow_result_path, slow_result); 1599 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); 1600 transform_later(result_region); 1601 transform_later(result_phi_rawoop); 1602 transform_later(result_phi_rawmem); 1603 transform_later(result_phi_i_o); 1604 // This completes all paths into the result merge point 1605 } 1606 1607 1608 // Helper for PhaseMacroExpand::expand_allocate_common. 1609 // Initializes the newly-allocated storage. 1610 Node* 1611 PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1612 Node* control, Node* rawmem, Node* object, 1613 Node* klass_node, Node* length, 1614 Node* size_in_bytes) { 1615 InitializeNode* init = alloc->initialization(); 1616 // Store the klass & mark bits 1617 Node* mark_node = NULL; 1618 // For now only enable fast locking for non-array types 1619 if (UseBiasedLocking && (length == NULL)) { 1620 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 1621 } else { 1622 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); 1623 } 1624 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 1625 1626 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 1627 int header_size = alloc->minimum_header_size(); // conservatively small 1628 1629 // Array length 1630 if (length != NULL) { // Arrays need length field 1631 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1632 // conservatively small header size: 1633 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1634 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1635 if (k->is_array_klass()) // we know the exact header size in most cases: 1636 header_size = Klass::layout_helper_header_size(k->layout_helper()); 1637 } 1638 1639 // Clear the object body, if necessary. 1640 if (init == NULL) { 1641 // The init has somehow disappeared; be cautious and clear everything. 1642 // 1643 // This can happen if a node is allocated but an uncommon trap occurs 1644 // immediately. In this case, the Initialize gets associated with the 1645 // trap, and may be placed in a different (outer) loop, if the Allocate 1646 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1647 // there can be two Allocates to one Initialize. The answer in all these 1648 // edge cases is safety first. It is always safe to clear immediately 1649 // within an Allocate, and then (maybe or maybe not) clear some more later. 1650 if (!(UseTLAB && ZeroTLAB)) { 1651 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1652 header_size, size_in_bytes, 1653 &_igvn); 1654 } 1655 } else { 1656 if (!init->is_complete()) { 1657 // Try to win by zeroing only what the init does not store. 1658 // We can also try to do some peephole optimizations, 1659 // such as combining some adjacent subword stores. 1660 rawmem = init->complete_stores(control, rawmem, object, 1661 header_size, size_in_bytes, &_igvn); 1662 } 1663 // We have no more use for this link, since the AllocateNode goes away: 1664 init->set_req(InitializeNode::RawAddress, top()); 1665 // (If we keep the link, it just confuses the register allocator, 1666 // who thinks he sees a real use of the address by the membar.) 1667 } 1668 1669 return rawmem; 1670 } 1671 1672 // Generate prefetch instructions for next allocations. 1673 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1674 Node*& contended_phi_rawmem, 1675 Node* old_eden_top, Node* new_eden_top, 1676 intx lines) { 1677 enum { fall_in_path = 1, pf_path = 2 }; 1678 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1679 // Generate prefetch allocation with watermark check. 1680 // As an allocation hits the watermark, we will prefetch starting 1681 // at a "distance" away from watermark. 1682 1683 Node *pf_region = new RegionNode(3); 1684 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 1685 TypeRawPtr::BOTTOM ); 1686 // I/O is used for Prefetch 1687 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO ); 1688 1689 Node *thread = new ThreadLocalNode(); 1690 transform_later(thread); 1691 1692 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread, 1693 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1694 transform_later(eden_pf_adr); 1695 1696 Node *old_pf_wm = new LoadPNode(needgc_false, 1697 contended_phi_rawmem, eden_pf_adr, 1698 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, 1699 MemNode::unordered); 1700 transform_later(old_pf_wm); 1701 1702 // check against new_eden_top 1703 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm ); 1704 transform_later(need_pf_cmp); 1705 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge ); 1706 transform_later(need_pf_bol); 1707 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol, 1708 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1709 transform_later(need_pf_iff); 1710 1711 // true node, add prefetchdistance 1712 Node *need_pf_true = new IfTrueNode( need_pf_iff ); 1713 transform_later(need_pf_true); 1714 1715 Node *need_pf_false = new IfFalseNode( need_pf_iff ); 1716 transform_later(need_pf_false); 1717 1718 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm, 1719 _igvn.MakeConX(AllocatePrefetchDistance) ); 1720 transform_later(new_pf_wmt ); 1721 new_pf_wmt->set_req(0, need_pf_true); 1722 1723 Node *store_new_wmt = new StorePNode(need_pf_true, 1724 contended_phi_rawmem, eden_pf_adr, 1725 TypeRawPtr::BOTTOM, new_pf_wmt, 1726 MemNode::unordered); 1727 transform_later(store_new_wmt); 1728 1729 // adding prefetches 1730 pf_phi_abio->init_req( fall_in_path, i_o ); 1731 1732 Node *prefetch_adr; 1733 Node *prefetch; 1734 uint step_size = AllocatePrefetchStepSize; 1735 uint distance = 0; 1736 1737 for ( intx i = 0; i < lines; i++ ) { 1738 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt, 1739 _igvn.MakeConX(distance) ); 1740 transform_later(prefetch_adr); 1741 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1742 transform_later(prefetch); 1743 distance += step_size; 1744 i_o = prefetch; 1745 } 1746 pf_phi_abio->set_req( pf_path, i_o ); 1747 1748 pf_region->init_req( fall_in_path, need_pf_false ); 1749 pf_region->init_req( pf_path, need_pf_true ); 1750 1751 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 1752 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 1753 1754 transform_later(pf_region); 1755 transform_later(pf_phi_rawmem); 1756 transform_later(pf_phi_abio); 1757 1758 needgc_false = pf_region; 1759 contended_phi_rawmem = pf_phi_rawmem; 1760 i_o = pf_phi_abio; 1761 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { 1762 // Insert a prefetch instruction for each allocation. 1763 // This code is used to generate 1 prefetch instruction per cache line. 1764 1765 // Generate several prefetch instructions. 1766 uint step_size = AllocatePrefetchStepSize; 1767 uint distance = AllocatePrefetchDistance; 1768 1769 // Next cache address. 1770 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top, 1771 _igvn.MakeConX(step_size + distance)); 1772 transform_later(cache_adr); 1773 cache_adr = new CastP2XNode(needgc_false, cache_adr); 1774 transform_later(cache_adr); 1775 // Address is aligned to execute prefetch to the beginning of cache line size 1776 // (it is important when BIS instruction is used on SPARC as prefetch). 1777 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); 1778 cache_adr = new AndXNode(cache_adr, mask); 1779 transform_later(cache_adr); 1780 cache_adr = new CastX2PNode(cache_adr); 1781 transform_later(cache_adr); 1782 1783 // Prefetch 1784 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); 1785 prefetch->set_req(0, needgc_false); 1786 transform_later(prefetch); 1787 contended_phi_rawmem = prefetch; 1788 Node *prefetch_adr; 1789 distance = step_size; 1790 for ( intx i = 1; i < lines; i++ ) { 1791 prefetch_adr = new AddPNode( cache_adr, cache_adr, 1792 _igvn.MakeConX(distance) ); 1793 transform_later(prefetch_adr); 1794 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); 1795 transform_later(prefetch); 1796 distance += step_size; 1797 contended_phi_rawmem = prefetch; 1798 } 1799 } else if( AllocatePrefetchStyle > 0 ) { 1800 // Insert a prefetch for each allocation only on the fast-path 1801 Node *prefetch_adr; 1802 Node *prefetch; 1803 // Generate several prefetch instructions. 1804 uint step_size = AllocatePrefetchStepSize; 1805 uint distance = AllocatePrefetchDistance; 1806 for ( intx i = 0; i < lines; i++ ) { 1807 prefetch_adr = new AddPNode( old_eden_top, new_eden_top, 1808 _igvn.MakeConX(distance) ); 1809 transform_later(prefetch_adr); 1810 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1811 // Do not let it float too high, since if eden_top == eden_end, 1812 // both might be null. 1813 if( i == 0 ) { // Set control for first prefetch, next follows it 1814 prefetch->init_req(0, needgc_false); 1815 } 1816 transform_later(prefetch); 1817 distance += step_size; 1818 i_o = prefetch; 1819 } 1820 } 1821 return i_o; 1822 } 1823 1824 1825 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 1826 expand_allocate_common(alloc, NULL, 1827 OptoRuntime::new_instance_Type(), 1828 OptoRuntime::new_instance_Java()); 1829 } 1830 1831 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 1832 Node* length = alloc->in(AllocateNode::ALength); 1833 InitializeNode* init = alloc->initialization(); 1834 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1835 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1836 address slow_call_address; // Address of slow call 1837 if (init != NULL && init->is_complete_with_arraycopy() && 1838 k->is_type_array_klass()) { 1839 // Don't zero type array during slow allocation in VM since 1840 // it will be initialized later by arraycopy in compiled code. 1841 slow_call_address = OptoRuntime::new_array_nozero_Java(); 1842 } else { 1843 slow_call_address = OptoRuntime::new_array_Java(); 1844 } 1845 expand_allocate_common(alloc, length, 1846 OptoRuntime::new_array_Type(), 1847 slow_call_address); 1848 } 1849 1850 //-------------------mark_eliminated_box---------------------------------- 1851 // 1852 // During EA obj may point to several objects but after few ideal graph 1853 // transformations (CCP) it may point to only one non escaping object 1854 // (but still using phi), corresponding locks and unlocks will be marked 1855 // for elimination. Later obj could be replaced with a new node (new phi) 1856 // and which does not have escape information. And later after some graph 1857 // reshape other locks and unlocks (which were not marked for elimination 1858 // before) are connected to this new obj (phi) but they still will not be 1859 // marked for elimination since new obj has no escape information. 1860 // Mark all associated (same box and obj) lock and unlock nodes for 1861 // elimination if some of them marked already. 1862 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) { 1863 if (oldbox->as_BoxLock()->is_eliminated()) 1864 return; // This BoxLock node was processed already. 1865 1866 // New implementation (EliminateNestedLocks) has separate BoxLock 1867 // node for each locked region so mark all associated locks/unlocks as 1868 // eliminated even if different objects are referenced in one locked region 1869 // (for example, OSR compilation of nested loop inside locked scope). 1870 if (EliminateNestedLocks || 1871 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) { 1872 // Box is used only in one lock region. Mark this box as eliminated. 1873 _igvn.hash_delete(oldbox); 1874 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value 1875 _igvn.hash_insert(oldbox); 1876 1877 for (uint i = 0; i < oldbox->outcnt(); i++) { 1878 Node* u = oldbox->raw_out(i); 1879 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) { 1880 AbstractLockNode* alock = u->as_AbstractLock(); 1881 // Check lock's box since box could be referenced by Lock's debug info. 1882 if (alock->box_node() == oldbox) { 1883 // Mark eliminated all related locks and unlocks. 1884 #ifdef ASSERT 1885 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4"); 1886 #endif 1887 alock->set_non_esc_obj(); 1888 } 1889 } 1890 } 1891 return; 1892 } 1893 1894 // Create new "eliminated" BoxLock node and use it in monitor debug info 1895 // instead of oldbox for the same object. 1896 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 1897 1898 // Note: BoxLock node is marked eliminated only here and it is used 1899 // to indicate that all associated lock and unlock nodes are marked 1900 // for elimination. 1901 newbox->set_eliminated(); 1902 transform_later(newbox); 1903 1904 // Replace old box node with new box for all users of the same object. 1905 for (uint i = 0; i < oldbox->outcnt();) { 1906 bool next_edge = true; 1907 1908 Node* u = oldbox->raw_out(i); 1909 if (u->is_AbstractLock()) { 1910 AbstractLockNode* alock = u->as_AbstractLock(); 1911 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) { 1912 // Replace Box and mark eliminated all related locks and unlocks. 1913 #ifdef ASSERT 1914 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5"); 1915 #endif 1916 alock->set_non_esc_obj(); 1917 _igvn.rehash_node_delayed(alock); 1918 alock->set_box_node(newbox); 1919 next_edge = false; 1920 } 1921 } 1922 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) { 1923 FastLockNode* flock = u->as_FastLock(); 1924 assert(flock->box_node() == oldbox, "sanity"); 1925 _igvn.rehash_node_delayed(flock); 1926 flock->set_box_node(newbox); 1927 next_edge = false; 1928 } 1929 1930 // Replace old box in monitor debug info. 1931 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 1932 SafePointNode* sfn = u->as_SafePoint(); 1933 JVMState* youngest_jvms = sfn->jvms(); 1934 int max_depth = youngest_jvms->depth(); 1935 for (int depth = 1; depth <= max_depth; depth++) { 1936 JVMState* jvms = youngest_jvms->of_depth(depth); 1937 int num_mon = jvms->nof_monitors(); 1938 // Loop over monitors 1939 for (int idx = 0; idx < num_mon; idx++) { 1940 Node* obj_node = sfn->monitor_obj(jvms, idx); 1941 Node* box_node = sfn->monitor_box(jvms, idx); 1942 if (box_node == oldbox && obj_node->eqv_uncast(obj)) { 1943 int j = jvms->monitor_box_offset(idx); 1944 _igvn.replace_input_of(u, j, newbox); 1945 next_edge = false; 1946 } 1947 } 1948 } 1949 } 1950 if (next_edge) i++; 1951 } 1952 } 1953 1954 //-----------------------mark_eliminated_locking_nodes----------------------- 1955 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { 1956 if (EliminateNestedLocks) { 1957 if (alock->is_nested()) { 1958 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity"); 1959 return; 1960 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened 1961 // Only Lock node has JVMState needed here. 1962 // Not that preceding claim is documented anywhere else. 1963 if (alock->jvms() != NULL) { 1964 if (alock->as_Lock()->is_nested_lock_region()) { 1965 // Mark eliminated related nested locks and unlocks. 1966 Node* obj = alock->obj_node(); 1967 BoxLockNode* box_node = alock->box_node()->as_BoxLock(); 1968 assert(!box_node->is_eliminated(), "should not be marked yet"); 1969 // Note: BoxLock node is marked eliminated only here 1970 // and it is used to indicate that all associated lock 1971 // and unlock nodes are marked for elimination. 1972 box_node->set_eliminated(); // Box's hash is always NO_HASH here 1973 for (uint i = 0; i < box_node->outcnt(); i++) { 1974 Node* u = box_node->raw_out(i); 1975 if (u->is_AbstractLock()) { 1976 alock = u->as_AbstractLock(); 1977 if (alock->box_node() == box_node) { 1978 // Verify that this Box is referenced only by related locks. 1979 assert(alock->obj_node()->eqv_uncast(obj), ""); 1980 // Mark all related locks and unlocks. 1981 #ifdef ASSERT 1982 alock->log_lock_optimization(C, "eliminate_lock_set_nested"); 1983 #endif 1984 alock->set_nested(); 1985 } 1986 } 1987 } 1988 } else { 1989 #ifdef ASSERT 1990 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region"); 1991 if (C->log() != NULL) 1992 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output 1993 #endif 1994 } 1995 } 1996 return; 1997 } 1998 // Process locks for non escaping object 1999 assert(alock->is_non_esc_obj(), ""); 2000 } // EliminateNestedLocks 2001 2002 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object 2003 // Look for all locks of this object and mark them and 2004 // corresponding BoxLock nodes as eliminated. 2005 Node* obj = alock->obj_node(); 2006 for (uint j = 0; j < obj->outcnt(); j++) { 2007 Node* o = obj->raw_out(j); 2008 if (o->is_AbstractLock() && 2009 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) { 2010 alock = o->as_AbstractLock(); 2011 Node* box = alock->box_node(); 2012 // Replace old box node with new eliminated box for all users 2013 // of the same object and mark related locks as eliminated. 2014 mark_eliminated_box(box, obj); 2015 } 2016 } 2017 } 2018 } 2019 2020 // we have determined that this lock/unlock can be eliminated, we simply 2021 // eliminate the node without expanding it. 2022 // 2023 // Note: The membar's associated with the lock/unlock are currently not 2024 // eliminated. This should be investigated as a future enhancement. 2025 // 2026 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 2027 2028 if (!alock->is_eliminated()) { 2029 return false; 2030 } 2031 #ifdef ASSERT 2032 if (!alock->is_coarsened()) { 2033 // Check that new "eliminated" BoxLock node is created. 2034 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 2035 assert(oldbox->is_eliminated(), "should be done already"); 2036 } 2037 #endif 2038 2039 alock->log_lock_optimization(C, "eliminate_lock"); 2040 2041 #ifndef PRODUCT 2042 if (PrintEliminateLocks) { 2043 if (alock->is_Lock()) { 2044 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx); 2045 } else { 2046 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx); 2047 } 2048 } 2049 #endif 2050 2051 Node* mem = alock->in(TypeFunc::Memory); 2052 Node* ctrl = alock->in(TypeFunc::Control); 2053 guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL"); 2054 2055 extract_call_projections(alock); 2056 // There are 2 projections from the lock. The lock node will 2057 // be deleted when its last use is subsumed below. 2058 assert(alock->outcnt() == 2 && 2059 _fallthroughproj != NULL && 2060 _memproj_fallthrough != NULL, 2061 "Unexpected projections from Lock/Unlock"); 2062 2063 Node* fallthroughproj = _fallthroughproj; 2064 Node* memproj_fallthrough = _memproj_fallthrough; 2065 2066 // The memory projection from a lock/unlock is RawMem 2067 // The input to a Lock is merged memory, so extract its RawMem input 2068 // (unless the MergeMem has been optimized away.) 2069 if (alock->is_Lock()) { 2070 // Seach for MemBarAcquireLock node and delete it also. 2071 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 2072 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, ""); 2073 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 2074 Node* memproj = membar->proj_out(TypeFunc::Memory); 2075 _igvn.replace_node(ctrlproj, fallthroughproj); 2076 _igvn.replace_node(memproj, memproj_fallthrough); 2077 2078 // Delete FastLock node also if this Lock node is unique user 2079 // (a loop peeling may clone a Lock node). 2080 Node* flock = alock->as_Lock()->fastlock_node(); 2081 if (flock->outcnt() == 1) { 2082 assert(flock->unique_out() == alock, "sanity"); 2083 _igvn.replace_node(flock, top()); 2084 } 2085 } 2086 2087 // Seach for MemBarReleaseLock node and delete it also. 2088 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) { 2089 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 2090 assert(membar->Opcode() == Op_MemBarReleaseLock && 2091 mem->is_Proj() && membar == mem->in(0), ""); 2092 _igvn.replace_node(fallthroughproj, ctrl); 2093 _igvn.replace_node(memproj_fallthrough, mem); 2094 fallthroughproj = ctrl; 2095 memproj_fallthrough = mem; 2096 ctrl = membar->in(TypeFunc::Control); 2097 mem = membar->in(TypeFunc::Memory); 2098 } 2099 2100 _igvn.replace_node(fallthroughproj, ctrl); 2101 _igvn.replace_node(memproj_fallthrough, mem); 2102 return true; 2103 } 2104 2105 2106 //------------------------------expand_lock_node---------------------- 2107 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 2108 2109 Node* ctrl = lock->in(TypeFunc::Control); 2110 Node* mem = lock->in(TypeFunc::Memory); 2111 Node* obj = lock->obj_node(); 2112 Node* box = lock->box_node(); 2113 Node* flock = lock->fastlock_node(); 2114 2115 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2116 2117 // Make the merge point 2118 Node *region; 2119 Node *mem_phi; 2120 Node *slow_path; 2121 2122 if (UseOptoBiasInlining) { 2123 /* 2124 * See the full description in MacroAssembler::biased_locking_enter(). 2125 * 2126 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { 2127 * // The object is biased. 2128 * proto_node = klass->prototype_header; 2129 * o_node = thread | proto_node; 2130 * x_node = o_node ^ mark_word; 2131 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? 2132 * // Done. 2133 * } else { 2134 * if( (x_node & biased_lock_mask) != 0 ) { 2135 * // The klass's prototype header is no longer biased. 2136 * cas(&mark_word, mark_word, proto_node) 2137 * goto cas_lock; 2138 * } else { 2139 * // The klass's prototype header is still biased. 2140 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? 2141 * old = mark_word; 2142 * new = o_node; 2143 * } else { 2144 * // Different thread or anonymous biased. 2145 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); 2146 * new = thread | old; 2147 * } 2148 * // Try to rebias. 2149 * if( cas(&mark_word, old, new) == 0 ) { 2150 * // Done. 2151 * } else { 2152 * goto slow_path; // Failed. 2153 * } 2154 * } 2155 * } 2156 * } else { 2157 * // The object is not biased. 2158 * cas_lock: 2159 * if( FastLock(obj) == 0 ) { 2160 * // Done. 2161 * } else { 2162 * slow_path: 2163 * OptoRuntime::complete_monitor_locking_Java(obj); 2164 * } 2165 * } 2166 */ 2167 2168 region = new RegionNode(5); 2169 // create a Phi for the memory state 2170 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2171 2172 Node* fast_lock_region = new RegionNode(3); 2173 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); 2174 2175 // First, check mark word for the biased lock pattern. 2176 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2177 2178 // Get fast path - mark word has the biased lock pattern. 2179 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, 2180 markOopDesc::biased_lock_mask_in_place, 2181 markOopDesc::biased_lock_pattern, true); 2182 // fast_lock_region->in(1) is set to slow path. 2183 fast_lock_mem_phi->init_req(1, mem); 2184 2185 // Now check that the lock is biased to the current thread and has 2186 // the same epoch and bias as Klass::_prototype_header. 2187 2188 // Special-case a fresh allocation to avoid building nodes: 2189 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); 2190 if (klass_node == NULL) { 2191 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 2192 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr())); 2193 #ifdef _LP64 2194 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) { 2195 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity"); 2196 klass_node->in(1)->init_req(0, ctrl); 2197 } else 2198 #endif 2199 klass_node->init_req(0, ctrl); 2200 } 2201 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type()); 2202 2203 Node* thread = transform_later(new ThreadLocalNode()); 2204 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2205 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node)); 2206 Node* x_node = transform_later(new XorXNode(o_node, mark_node)); 2207 2208 // Get slow path - mark word does NOT match the value. 2209 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, 2210 (~markOopDesc::age_mask_in_place), 0); 2211 // region->in(3) is set to fast path - the object is biased to the current thread. 2212 mem_phi->init_req(3, mem); 2213 2214 2215 // Mark word does NOT match the value (thread | Klass::_prototype_header). 2216 2217 2218 // First, check biased pattern. 2219 // Get fast path - _prototype_header has the same biased lock pattern. 2220 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, 2221 markOopDesc::biased_lock_mask_in_place, 0, true); 2222 2223 not_biased_ctrl = fast_lock_region->in(2); // Slow path 2224 // fast_lock_region->in(2) - the prototype header is no longer biased 2225 // and we have to revoke the bias on this object. 2226 // We are going to try to reset the mark of this object to the prototype 2227 // value and fall through to the CAS-based locking scheme. 2228 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 2229 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr, 2230 proto_node, mark_node); 2231 transform_later(cas); 2232 Node* proj = transform_later(new SCMemProjNode(cas)); 2233 fast_lock_mem_phi->init_req(2, proj); 2234 2235 2236 // Second, check epoch bits. 2237 Node* rebiased_region = new RegionNode(3); 2238 Node* old_phi = new PhiNode( rebiased_region, TypeX_X); 2239 Node* new_phi = new PhiNode( rebiased_region, TypeX_X); 2240 2241 // Get slow path - mark word does NOT match epoch bits. 2242 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, 2243 markOopDesc::epoch_mask_in_place, 0); 2244 // The epoch of the current bias is not valid, attempt to rebias the object 2245 // toward the current thread. 2246 rebiased_region->init_req(2, epoch_ctrl); 2247 old_phi->init_req(2, mark_node); 2248 new_phi->init_req(2, o_node); 2249 2250 // rebiased_region->in(1) is set to fast path. 2251 // The epoch of the current bias is still valid but we know 2252 // nothing about the owner; it might be set or it might be clear. 2253 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place | 2254 markOopDesc::age_mask_in_place | 2255 markOopDesc::epoch_mask_in_place); 2256 Node* old = transform_later(new AndXNode(mark_node, cmask)); 2257 cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2258 Node* new_mark = transform_later(new OrXNode(cast_thread, old)); 2259 old_phi->init_req(1, old); 2260 new_phi->init_req(1, new_mark); 2261 2262 transform_later(rebiased_region); 2263 transform_later(old_phi); 2264 transform_later(new_phi); 2265 2266 // Try to acquire the bias of the object using an atomic operation. 2267 // If this fails we will go in to the runtime to revoke the object's bias. 2268 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi); 2269 transform_later(cas); 2270 proj = transform_later(new SCMemProjNode(cas)); 2271 2272 // Get slow path - Failed to CAS. 2273 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); 2274 mem_phi->init_req(4, proj); 2275 // region->in(4) is set to fast path - the object is rebiased to the current thread. 2276 2277 // Failed to CAS. 2278 slow_path = new RegionNode(3); 2279 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); 2280 2281 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control 2282 slow_mem->init_req(1, proj); 2283 2284 // Call CAS-based locking scheme (FastLock node). 2285 2286 transform_later(fast_lock_region); 2287 transform_later(fast_lock_mem_phi); 2288 2289 // Get slow path - FastLock failed to lock the object. 2290 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); 2291 mem_phi->init_req(2, fast_lock_mem_phi); 2292 // region->in(2) is set to fast path - the object is locked to the current thread. 2293 2294 slow_path->init_req(2, ctrl); // Capture slow-control 2295 slow_mem->init_req(2, fast_lock_mem_phi); 2296 2297 transform_later(slow_path); 2298 transform_later(slow_mem); 2299 // Reset lock's memory edge. 2300 lock->set_req(TypeFunc::Memory, slow_mem); 2301 2302 } else { 2303 region = new RegionNode(3); 2304 // create a Phi for the memory state 2305 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2306 2307 // Optimize test; set region slot 2 2308 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); 2309 mem_phi->init_req(2, mem); 2310 } 2311 2312 // Make slow path call 2313 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), 2314 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, 2315 obj, box, NULL); 2316 2317 extract_call_projections(call); 2318 2319 // Slow path can only throw asynchronous exceptions, which are always 2320 // de-opted. So the compiler thinks the slow-call can never throw an 2321 // exception. If it DOES throw an exception we would need the debug 2322 // info removed first (since if it throws there is no monitor). 2323 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2324 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2325 2326 // Capture slow path 2327 // disconnect fall-through projection from call and create a new one 2328 // hook up users of fall-through projection to region 2329 Node *slow_ctrl = _fallthroughproj->clone(); 2330 transform_later(slow_ctrl); 2331 _igvn.hash_delete(_fallthroughproj); 2332 _fallthroughproj->disconnect_inputs(NULL, C); 2333 region->init_req(1, slow_ctrl); 2334 // region inputs are now complete 2335 transform_later(region); 2336 _igvn.replace_node(_fallthroughproj, region); 2337 2338 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory)); 2339 mem_phi->init_req(1, memproj ); 2340 transform_later(mem_phi); 2341 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2342 } 2343 2344 //------------------------------expand_unlock_node---------------------- 2345 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 2346 2347 Node* ctrl = unlock->in(TypeFunc::Control); 2348 Node* mem = unlock->in(TypeFunc::Memory); 2349 Node* obj = unlock->obj_node(); 2350 Node* box = unlock->box_node(); 2351 2352 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2353 2354 // No need for a null check on unlock 2355 2356 // Make the merge point 2357 Node *region; 2358 Node *mem_phi; 2359 2360 if (UseOptoBiasInlining) { 2361 // Check for biased locking unlock case, which is a no-op. 2362 // See the full description in MacroAssembler::biased_locking_exit(). 2363 region = new RegionNode(4); 2364 // create a Phi for the memory state 2365 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2366 mem_phi->init_req(3, mem); 2367 2368 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2369 ctrl = opt_bits_test(ctrl, region, 3, mark_node, 2370 markOopDesc::biased_lock_mask_in_place, 2371 markOopDesc::biased_lock_pattern); 2372 } else { 2373 region = new RegionNode(3); 2374 // create a Phi for the memory state 2375 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2376 } 2377 2378 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box ); 2379 funlock = transform_later( funlock )->as_FastUnlock(); 2380 // Optimize test; set region slot 2 2381 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); 2382 Node *thread = transform_later(new ThreadLocalNode()); 2383 2384 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), 2385 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), 2386 "complete_monitor_unlocking_C", slow_path, obj, box, thread); 2387 2388 extract_call_projections(call); 2389 2390 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2391 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2392 2393 // No exceptions for unlocking 2394 // Capture slow path 2395 // disconnect fall-through projection from call and create a new one 2396 // hook up users of fall-through projection to region 2397 Node *slow_ctrl = _fallthroughproj->clone(); 2398 transform_later(slow_ctrl); 2399 _igvn.hash_delete(_fallthroughproj); 2400 _fallthroughproj->disconnect_inputs(NULL, C); 2401 region->init_req(1, slow_ctrl); 2402 // region inputs are now complete 2403 transform_later(region); 2404 _igvn.replace_node(_fallthroughproj, region); 2405 2406 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) ); 2407 mem_phi->init_req(1, memproj ); 2408 mem_phi->init_req(2, mem); 2409 transform_later(mem_phi); 2410 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2411 } 2412 2413 //---------------------------eliminate_macro_nodes---------------------- 2414 // Eliminate scalar replaced allocations and associated locks. 2415 void PhaseMacroExpand::eliminate_macro_nodes() { 2416 if (C->macro_count() == 0) 2417 return; 2418 2419 // First, attempt to eliminate locks 2420 int cnt = C->macro_count(); 2421 for (int i=0; i < cnt; i++) { 2422 Node *n = C->macro_node(i); 2423 if (n->is_AbstractLock()) { // Lock and Unlock nodes 2424 // Before elimination mark all associated (same box and obj) 2425 // lock and unlock nodes. 2426 mark_eliminated_locking_nodes(n->as_AbstractLock()); 2427 } 2428 } 2429 bool progress = true; 2430 while (progress) { 2431 progress = false; 2432 for (int i = C->macro_count(); i > 0; i--) { 2433 Node * n = C->macro_node(i-1); 2434 bool success = false; 2435 debug_only(int old_macro_count = C->macro_count();); 2436 if (n->is_AbstractLock()) { 2437 success = eliminate_locking_node(n->as_AbstractLock()); 2438 } 2439 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2440 progress = progress || success; 2441 } 2442 } 2443 // Next, attempt to eliminate allocations 2444 _has_locks = false; 2445 progress = true; 2446 while (progress) { 2447 progress = false; 2448 for (int i = C->macro_count(); i > 0; i--) { 2449 Node * n = C->macro_node(i-1); 2450 bool success = false; 2451 debug_only(int old_macro_count = C->macro_count();); 2452 switch (n->class_id()) { 2453 case Node::Class_Allocate: 2454 case Node::Class_AllocateArray: 2455 success = eliminate_allocate_node(n->as_Allocate()); 2456 break; 2457 case Node::Class_CallStaticJava: 2458 success = eliminate_boxing_node(n->as_CallStaticJava()); 2459 break; 2460 case Node::Class_Lock: 2461 case Node::Class_Unlock: 2462 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 2463 _has_locks = true; 2464 break; 2465 case Node::Class_ArrayCopy: 2466 break; 2467 case Node::Class_OuterStripMinedLoop: 2468 break; 2469 default: 2470 assert(n->Opcode() == Op_LoopLimit || 2471 n->Opcode() == Op_Opaque1 || 2472 n->Opcode() == Op_Opaque2 || 2473 n->Opcode() == Op_Opaque3 || 2474 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n), 2475 "unknown node type in macro list"); 2476 } 2477 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2478 progress = progress || success; 2479 } 2480 } 2481 } 2482 2483 //------------------------------expand_macro_nodes---------------------- 2484 // Returns true if a failure occurred. 2485 bool PhaseMacroExpand::expand_macro_nodes() { 2486 // Last attempt to eliminate macro nodes. 2487 eliminate_macro_nodes(); 2488 2489 // Make sure expansion will not cause node limit to be exceeded. 2490 // Worst case is a macro node gets expanded into about 200 nodes. 2491 // Allow 50% more for optimization. 2492 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) ) 2493 return true; 2494 2495 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. 2496 bool progress = true; 2497 while (progress) { 2498 progress = false; 2499 for (int i = C->macro_count(); i > 0; i--) { 2500 Node * n = C->macro_node(i-1); 2501 bool success = false; 2502 debug_only(int old_macro_count = C->macro_count();); 2503 if (n->Opcode() == Op_LoopLimit) { 2504 // Remove it from macro list and put on IGVN worklist to optimize. 2505 C->remove_macro_node(n); 2506 _igvn._worklist.push(n); 2507 success = true; 2508 } else if (n->Opcode() == Op_CallStaticJava) { 2509 // Remove it from macro list and put on IGVN worklist to optimize. 2510 C->remove_macro_node(n); 2511 _igvn._worklist.push(n); 2512 success = true; 2513 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { 2514 _igvn.replace_node(n, n->in(1)); 2515 success = true; 2516 #if INCLUDE_RTM_OPT 2517 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) { 2518 assert(C->profile_rtm(), "should be used only in rtm deoptimization code"); 2519 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), ""); 2520 Node* cmp = n->unique_out(); 2521 #ifdef ASSERT 2522 // Validate graph. 2523 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), ""); 2524 BoolNode* bol = cmp->unique_out()->as_Bool(); 2525 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() && 2526 (bol->_test._test == BoolTest::ne), ""); 2527 IfNode* ifn = bol->unique_out()->as_If(); 2528 assert((ifn->outcnt() == 2) && 2529 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, ""); 2530 #endif 2531 Node* repl = n->in(1); 2532 if (!_has_locks) { 2533 // Remove RTM state check if there are no locks in the code. 2534 // Replace input to compare the same value. 2535 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1); 2536 } 2537 _igvn.replace_node(n, repl); 2538 success = true; 2539 #endif 2540 } else if (n->Opcode() == Op_OuterStripMinedLoop) { 2541 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn); 2542 C->remove_macro_node(n); 2543 success = true; 2544 } 2545 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2546 progress = progress || success; 2547 } 2548 } 2549 2550 // expand arraycopy "macro" nodes first 2551 // For ReduceBulkZeroing, we must first process all arraycopy nodes 2552 // before the allocate nodes are expanded. 2553 int macro_idx = C->macro_count() - 1; 2554 while (macro_idx >= 0) { 2555 Node * n = C->macro_node(macro_idx); 2556 assert(n->is_macro(), "only macro nodes expected here"); 2557 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) { 2558 // node is unreachable, so don't try to expand it 2559 C->remove_macro_node(n); 2560 } else if (n->is_ArrayCopy()){ 2561 int macro_count = C->macro_count(); 2562 expand_arraycopy_node(n->as_ArrayCopy()); 2563 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2564 } 2565 if (C->failing()) return true; 2566 macro_idx --; 2567 } 2568 2569 // expand "macro" nodes 2570 // nodes are removed from the macro list as they are processed 2571 while (C->macro_count() > 0) { 2572 int macro_count = C->macro_count(); 2573 Node * n = C->macro_node(macro_count-1); 2574 assert(n->is_macro(), "only macro nodes expected here"); 2575 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) { 2576 // node is unreachable, so don't try to expand it 2577 C->remove_macro_node(n); 2578 continue; 2579 } 2580 switch (n->class_id()) { 2581 case Node::Class_Allocate: 2582 expand_allocate(n->as_Allocate()); 2583 break; 2584 case Node::Class_AllocateArray: 2585 expand_allocate_array(n->as_AllocateArray()); 2586 break; 2587 case Node::Class_Lock: 2588 expand_lock_node(n->as_Lock()); 2589 break; 2590 case Node::Class_Unlock: 2591 expand_unlock_node(n->as_Unlock()); 2592 break; 2593 default: 2594 assert(false, "unknown node type in macro list"); 2595 } 2596 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2597 if (C->failing()) return true; 2598 } 2599 2600 _igvn.set_delay_transform(false); 2601 _igvn.optimize(); 2602 if (C->failing()) return true; 2603 return false; 2604 }