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