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