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(); 370 if (adr_type->isa_aryptr()) { 371 // In the case of a flattened value type array, each field has its 372 // own slice so we need to extract the field being accessed from 373 // the address computation 374 adr_type = adr_type->is_aryptr()->add_field_offset_and_offset(offset); 375 adr = _igvn.transform(new CastPPNode(adr, adr_type)); 376 } else { 377 adr_type = adr_type->add_offset(offset); 378 } 379 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned); 380 } 381 } 382 if (res != NULL) { 383 res = _igvn.transform(res); 384 if (ftype->isa_narrowoop()) { 385 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes 386 assert(res->isa_DecodeN(), "should be narrow oop"); 387 res = _igvn.transform(new EncodePNode(res, ftype)); 388 } 389 return res; 390 } 391 return NULL; 392 } 393 394 // 395 // Given a Memory Phi, compute a value Phi containing the values from stores 396 // on the input paths. 397 // Note: this function is recursive, its depth is limited by the "level" argument 398 // Returns the computed Phi, or NULL if it cannot compute it. 399 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) { 400 assert(mem->is_Phi(), "sanity"); 401 int alias_idx = C->get_alias_index(adr_t); 402 int offset = adr_t->flattened_offset(); 403 int instance_id = adr_t->instance_id(); 404 405 // Check if an appropriate value phi already exists. 406 Node* region = mem->in(0); 407 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { 408 Node* phi = region->fast_out(k); 409 if (phi->is_Phi() && phi != mem && 410 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) { 411 return phi; 412 } 413 } 414 // Check if an appropriate new value phi already exists. 415 Node* new_phi = value_phis->find(mem->_idx); 416 if (new_phi != NULL) 417 return new_phi; 418 419 if (level <= 0) { 420 return NULL; // Give up: phi tree too deep 421 } 422 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 423 Node *alloc_mem = alloc->in(TypeFunc::Memory); 424 425 uint length = mem->req(); 426 GrowableArray <Node *> values(length, length, NULL, false); 427 428 // create a new Phi for the value 429 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset); 430 transform_later(phi); 431 value_phis->push(phi, mem->_idx); 432 433 for (uint j = 1; j < length; j++) { 434 Node *in = mem->in(j); 435 if (in == NULL || in->is_top()) { 436 values.at_put(j, in); 437 } else { 438 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); 439 if (val == start_mem || val == alloc_mem) { 440 // hit a sentinel, return appropriate 0 value 441 Node* default_value = alloc->in(AllocateNode::DefaultValue); 442 if (default_value != NULL) { 443 values.at_put(j, default_value); 444 } else { 445 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null"); 446 values.at_put(j, _igvn.zerocon(ft)); 447 } 448 continue; 449 } 450 if (val->is_Initialize()) { 451 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 452 } 453 if (val == NULL) { 454 return NULL; // can't find a value on this path 455 } 456 if (val == mem) { 457 values.at_put(j, mem); 458 } else if (val->is_Store()) { 459 Node* n = val->in(MemNode::ValueIn); 460 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 461 n = bs->step_over_gc_barrier(n); 462 values.at_put(j, n); 463 } else if(val->is_Proj() && val->in(0) == alloc) { 464 Node* default_value = alloc->in(AllocateNode::DefaultValue); 465 if (default_value != NULL) { 466 values.at_put(j, default_value); 467 } else { 468 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null"); 469 values.at_put(j, _igvn.zerocon(ft)); 470 } 471 } else if (val->is_Phi()) { 472 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); 473 if (val == NULL) { 474 return NULL; 475 } 476 values.at_put(j, val); 477 } else if (val->Opcode() == Op_SCMemProj) { 478 assert(val->in(0)->is_LoadStore() || 479 val->in(0)->Opcode() == Op_EncodeISOArray || 480 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity"); 481 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 482 return NULL; 483 } else if (val->is_ArrayCopy()) { 484 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc); 485 if (res == NULL) { 486 return NULL; 487 } 488 values.at_put(j, res); 489 } else { 490 #ifdef ASSERT 491 val->dump(); 492 assert(false, "unknown node on this path"); 493 #endif 494 return NULL; // unknown node on this path 495 } 496 } 497 } 498 // Set Phi's inputs 499 for (uint j = 1; j < length; j++) { 500 if (values.at(j) == mem) { 501 phi->init_req(j, phi); 502 } else { 503 phi->init_req(j, values.at(j)); 504 } 505 } 506 return phi; 507 } 508 509 // Search the last value stored into the object's field. 510 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) { 511 assert(adr_t->is_known_instance_field(), "instance required"); 512 int instance_id = adr_t->instance_id(); 513 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 514 515 int alias_idx = C->get_alias_index(adr_t); 516 int offset = adr_t->flattened_offset(); 517 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 518 Node *alloc_mem = alloc->in(TypeFunc::Memory); 519 Arena *a = Thread::current()->resource_area(); 520 VectorSet visited(a); 521 522 bool done = sfpt_mem == alloc_mem; 523 Node *mem = sfpt_mem; 524 while (!done) { 525 if (visited.test_set(mem->_idx)) { 526 return NULL; // found a loop, give up 527 } 528 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 529 if (mem == start_mem || mem == alloc_mem) { 530 done = true; // hit a sentinel, return appropriate 0 value 531 } else if (mem->is_Initialize()) { 532 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 533 if (mem == NULL) { 534 done = true; // Something went wrong. 535 } else if (mem->is_Store()) { 536 const TypePtr* atype = mem->as_Store()->adr_type(); 537 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 538 done = true; 539 } 540 } else if (mem->is_Store()) { 541 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 542 assert(atype != NULL, "address type must be oopptr"); 543 assert(C->get_alias_index(atype) == alias_idx && 544 atype->is_known_instance_field() && atype->flattened_offset() == offset && 545 atype->instance_id() == instance_id, "store is correct memory slice"); 546 done = true; 547 } else if (mem->is_Phi()) { 548 // try to find a phi's unique input 549 Node *unique_input = NULL; 550 Node *top = C->top(); 551 for (uint i = 1; i < mem->req(); i++) { 552 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 553 if (n == NULL || n == top || n == mem) { 554 continue; 555 } else if (unique_input == NULL) { 556 unique_input = n; 557 } else if (unique_input != n) { 558 unique_input = top; 559 break; 560 } 561 } 562 if (unique_input != NULL && unique_input != top) { 563 mem = unique_input; 564 } else { 565 done = true; 566 } 567 } else if (mem->is_ArrayCopy()) { 568 done = true; 569 } else { 570 assert(false, "unexpected node"); 571 } 572 } 573 if (mem != NULL) { 574 if (mem == start_mem || mem == alloc_mem) { 575 // hit a sentinel, return appropriate 0 value 576 Node* default_value = alloc->in(AllocateNode::DefaultValue); 577 if (default_value != NULL) { 578 return default_value; 579 } 580 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null"); 581 return _igvn.zerocon(ft); 582 } else if (mem->is_Store()) { 583 Node* n = mem->in(MemNode::ValueIn); 584 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 585 n = bs->step_over_gc_barrier(n); 586 return n; 587 } else if (mem->is_Phi()) { 588 // attempt to produce a Phi reflecting the values on the input paths of the Phi 589 Node_Stack value_phis(a, 8); 590 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 591 if (phi != NULL) { 592 return phi; 593 } else { 594 // Kill all new Phis 595 while(value_phis.is_nonempty()) { 596 Node* n = value_phis.node(); 597 _igvn.replace_node(n, C->top()); 598 value_phis.pop(); 599 } 600 } 601 } else if (mem->is_ArrayCopy()) { 602 Node* ctl = mem->in(0); 603 Node* m = mem->in(TypeFunc::Memory); 604 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) { 605 // pin the loads in the uncommon trap path 606 ctl = sfpt_ctl; 607 m = sfpt_mem; 608 } 609 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc); 610 } 611 } 612 // Something went wrong. 613 return NULL; 614 } 615 616 // Search the last value stored into the value type's fields. 617 Node* PhaseMacroExpand::value_type_from_mem(Node* mem, Node* ctl, ciValueKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) { 618 // Subtract the offset of the first field to account for the missing oop header 619 offset -= vk->first_field_offset(); 620 // Create a new ValueTypeNode and retrieve the field values from memory 621 ValueTypeNode* vt = ValueTypeNode::make_uninitialized(_igvn, vk)->as_ValueType(); 622 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) { 623 ciType* field_type = vt->field_type(i); 624 int field_offset = offset + vt->field_offset(i); 625 // Each value type field has its own memory slice 626 adr_type = adr_type->with_field_offset(field_offset); 627 Node* value = NULL; 628 if (vt->field_is_flattened(i)) { 629 value = value_type_from_mem(mem, ctl, field_type->as_value_klass(), adr_type, field_offset, alloc); 630 } else { 631 const Type* ft = Type::get_const_type(field_type); 632 BasicType bt = field_type->basic_type(); 633 if (UseCompressedOops && !is_java_primitive(bt)) { 634 ft = ft->make_narrowoop(); 635 bt = T_NARROWOOP; 636 } 637 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc); 638 if (value != NULL && ft->isa_narrowoop()) { 639 assert(UseCompressedOops, "unexpected narrow oop"); 640 value = transform_later(new DecodeNNode(value, value->get_ptr_type())); 641 } 642 } 643 if (value != NULL) { 644 vt->set_field_value(i, value); 645 } else { 646 // We might have reached the TrackedInitializationLimit 647 return NULL; 648 } 649 } 650 return transform_later(vt); 651 } 652 653 // Check the possibility of scalar replacement. 654 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 655 // Scan the uses of the allocation to check for anything that would 656 // prevent us from eliminating it. 657 NOT_PRODUCT( const char* fail_eliminate = NULL; ) 658 DEBUG_ONLY( Node* disq_node = NULL; ) 659 bool can_eliminate = true; 660 661 Node* res = alloc->result_cast(); 662 const TypeOopPtr* res_type = NULL; 663 if (res == NULL) { 664 // All users were eliminated. 665 } else if (!res->is_CheckCastPP()) { 666 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 667 can_eliminate = false; 668 } else { 669 res_type = _igvn.type(res)->isa_oopptr(); 670 if (res_type == NULL) { 671 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 672 can_eliminate = false; 673 } else if (res_type->isa_aryptr()) { 674 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 675 if (length < 0) { 676 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 677 can_eliminate = false; 678 } 679 } 680 } 681 682 if (can_eliminate && res != NULL) { 683 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 684 j < jmax && can_eliminate; j++) { 685 Node* use = res->fast_out(j); 686 687 if (use->is_AddP()) { 688 const TypePtr* addp_type = _igvn.type(use)->is_ptr(); 689 int offset = addp_type->offset(); 690 691 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 692 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) 693 can_eliminate = false; 694 break; 695 } 696 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 697 k < kmax && can_eliminate; k++) { 698 Node* n = use->fast_out(k); 699 if (!n->is_Store() && n->Opcode() != Op_CastP2X && 700 SHENANDOAHGC_ONLY((!UseShenandoahGC || !ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(n)) &&) 701 !(n->is_ArrayCopy() && 702 n->as_ArrayCopy()->is_clonebasic() && 703 n->in(ArrayCopyNode::Dest) == use)) { 704 DEBUG_ONLY(disq_node = n;) 705 if (n->is_Load() || n->is_LoadStore()) { 706 NOT_PRODUCT(fail_eliminate = "Field load";) 707 } else { 708 NOT_PRODUCT(fail_eliminate = "Not store field reference";) 709 } 710 can_eliminate = false; 711 } 712 } 713 } else if (use->is_ArrayCopy() && 714 (use->as_ArrayCopy()->is_arraycopy_validated() || 715 use->as_ArrayCopy()->is_copyof_validated() || 716 use->as_ArrayCopy()->is_copyofrange_validated()) && 717 use->in(ArrayCopyNode::Dest) == res) { 718 // ok to eliminate 719 } else if (use->is_SafePoint()) { 720 SafePointNode* sfpt = use->as_SafePoint(); 721 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 722 // Object is passed as argument. 723 DEBUG_ONLY(disq_node = use;) 724 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 725 can_eliminate = false; 726 } 727 Node* sfptMem = sfpt->memory(); 728 if (sfptMem == NULL || sfptMem->is_top()) { 729 DEBUG_ONLY(disq_node = use;) 730 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) 731 can_eliminate = false; 732 } else { 733 safepoints.append_if_missing(sfpt); 734 } 735 } else if (use->is_ValueType() && use->isa_ValueType()->get_oop() == res) { 736 // ok to eliminate 737 } else if (use->is_Store()) { 738 // store to mark work 739 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 740 if (use->is_Phi()) { 741 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 742 NOT_PRODUCT(fail_eliminate = "Object is return value";) 743 } else { 744 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 745 } 746 DEBUG_ONLY(disq_node = use;) 747 } else { 748 if (use->Opcode() == Op_Return) { 749 NOT_PRODUCT(fail_eliminate = "Object is return value";) 750 } else { 751 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 752 } 753 DEBUG_ONLY(disq_node = use;) 754 } 755 can_eliminate = false; 756 } else { 757 assert(use->Opcode() == Op_CastP2X, "should be"); 758 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null"); 759 } 760 } 761 } 762 763 #ifndef PRODUCT 764 if (PrintEliminateAllocations) { 765 if (can_eliminate) { 766 tty->print("Scalar "); 767 if (res == NULL) 768 alloc->dump(); 769 else 770 res->dump(); 771 } else if (alloc->_is_scalar_replaceable) { 772 tty->print("NotScalar (%s)", fail_eliminate); 773 if (res == NULL) 774 alloc->dump(); 775 else 776 res->dump(); 777 #ifdef ASSERT 778 if (disq_node != NULL) { 779 tty->print(" >>>> "); 780 disq_node->dump(); 781 } 782 #endif /*ASSERT*/ 783 } 784 } 785 #endif 786 return can_eliminate; 787 } 788 789 // Do scalar replacement. 790 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 791 GrowableArray <SafePointNode *> safepoints_done; 792 793 ciKlass* klass = NULL; 794 ciInstanceKlass* iklass = NULL; 795 int nfields = 0; 796 int array_base = 0; 797 int element_size = 0; 798 BasicType basic_elem_type = T_ILLEGAL; 799 ciType* elem_type = NULL; 800 801 Node* res = alloc->result_cast(); 802 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result"); 803 const TypeOopPtr* res_type = NULL; 804 if (res != NULL) { // Could be NULL when there are no users 805 res_type = _igvn.type(res)->isa_oopptr(); 806 } 807 808 if (res != NULL) { 809 klass = res_type->klass(); 810 if (res_type->isa_instptr()) { 811 // find the fields of the class which will be needed for safepoint debug information 812 assert(klass->is_instance_klass(), "must be an instance klass."); 813 iklass = klass->as_instance_klass(); 814 nfields = iklass->nof_nonstatic_fields(); 815 } else { 816 // find the array's elements which will be needed for safepoint debug information 817 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 818 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); 819 elem_type = klass->as_array_klass()->element_type(); 820 basic_elem_type = elem_type->basic_type(); 821 if (elem_type->is_valuetype() && !klass->is_value_array_klass()) { 822 assert(basic_elem_type == T_VALUETYPE, "unexpected element basic type"); 823 basic_elem_type = T_OBJECT; 824 } 825 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 826 element_size = type2aelembytes(basic_elem_type); 827 if (klass->is_value_array_klass()) { 828 // Flattened value type array 829 element_size = klass->as_value_array_klass()->element_byte_size(); 830 } 831 } 832 } 833 // 834 // Process the safepoint uses 835 // 836 Unique_Node_List value_worklist; 837 while (safepoints.length() > 0) { 838 SafePointNode* sfpt = safepoints.pop(); 839 Node* mem = sfpt->memory(); 840 Node* ctl = sfpt->control(); 841 assert(sfpt->jvms() != NULL, "missed JVMS"); 842 // Fields of scalar objs are referenced only at the end 843 // of regular debuginfo at the last (youngest) JVMS. 844 // Record relative start index. 845 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff()); 846 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, 847 #ifdef ASSERT 848 alloc, 849 #endif 850 first_ind, nfields); 851 sobj->init_req(0, C->root()); 852 transform_later(sobj); 853 854 // Scan object's fields adding an input to the safepoint for each field. 855 for (int j = 0; j < nfields; j++) { 856 intptr_t offset; 857 ciField* field = NULL; 858 if (iklass != NULL) { 859 field = iklass->nonstatic_field_at(j); 860 offset = field->offset(); 861 elem_type = field->type(); 862 basic_elem_type = field->layout_type(); 863 assert(!field->is_flattened(), "flattened value type fields should not have safepoint uses"); 864 } else { 865 offset = array_base + j * (intptr_t)element_size; 866 } 867 868 const Type *field_type; 869 // The next code is taken from Parse::do_get_xxx(). 870 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { 871 if (!elem_type->is_loaded()) { 872 field_type = TypeInstPtr::BOTTOM; 873 } else if (field != NULL && field->is_static_constant()) { 874 // This can happen if the constant oop is non-perm. 875 ciObject* con = field->constant_value().as_object(); 876 // Do not "join" in the previous type; it doesn't add value, 877 // and may yield a vacuous result if the field is of interface type. 878 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 879 assert(field_type != NULL, "field singleton type must be consistent"); 880 } else { 881 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 882 } 883 if (UseCompressedOops) { 884 field_type = field_type->make_narrowoop(); 885 basic_elem_type = T_NARROWOOP; 886 } 887 } else { 888 field_type = Type::get_const_basic_type(basic_elem_type); 889 } 890 891 Node* field_val = NULL; 892 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 893 if (klass->is_value_array_klass()) { 894 ciValueKlass* vk = elem_type->as_value_klass(); 895 assert(vk->flatten_array(), "must be flattened"); 896 field_val = value_type_from_mem(mem, ctl, vk, field_addr_type->isa_aryptr(), 0, alloc); 897 } else { 898 field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc); 899 } 900 if (field_val == NULL) { 901 // We weren't able to find a value for this field, 902 // give up on eliminating this allocation. 903 904 // Remove any extra entries we added to the safepoint. 905 uint last = sfpt->req() - 1; 906 for (int k = 0; k < j; k++) { 907 sfpt->del_req(last--); 908 } 909 _igvn._worklist.push(sfpt); 910 // rollback processed safepoints 911 while (safepoints_done.length() > 0) { 912 SafePointNode* sfpt_done = safepoints_done.pop(); 913 // remove any extra entries we added to the safepoint 914 last = sfpt_done->req() - 1; 915 for (int k = 0; k < nfields; k++) { 916 sfpt_done->del_req(last--); 917 } 918 JVMState *jvms = sfpt_done->jvms(); 919 jvms->set_endoff(sfpt_done->req()); 920 // Now make a pass over the debug information replacing any references 921 // to SafePointScalarObjectNode with the allocated object. 922 int start = jvms->debug_start(); 923 int end = jvms->debug_end(); 924 for (int i = start; i < end; i++) { 925 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 926 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 927 if (scobj->first_index(jvms) == sfpt_done->req() && 928 scobj->n_fields() == (uint)nfields) { 929 assert(scobj->alloc() == alloc, "sanity"); 930 sfpt_done->set_req(i, res); 931 } 932 } 933 } 934 _igvn._worklist.push(sfpt_done); 935 } 936 #ifndef PRODUCT 937 if (PrintEliminateAllocations) { 938 if (field != NULL) { 939 tty->print("=== At SafePoint node %d can't find value of Field: ", 940 sfpt->_idx); 941 field->print(); 942 int field_idx = C->get_alias_index(field_addr_type); 943 tty->print(" (alias_idx=%d)", field_idx); 944 } else { // Array's element 945 tty->print("=== At SafePoint node %d can't find value of array element [%d]", 946 sfpt->_idx, j); 947 } 948 tty->print(", which prevents elimination of: "); 949 if (res == NULL) 950 alloc->dump(); 951 else 952 res->dump(); 953 } 954 #endif 955 return false; 956 } 957 if (UseCompressedOops && field_type->isa_narrowoop()) { 958 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 959 // to be able scalar replace the allocation. 960 if (field_val->is_EncodeP()) { 961 field_val = field_val->in(1); 962 } else { 963 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type())); 964 } 965 } else if (field_val->is_ValueType()) { 966 // Keep track of value types to scalarize them later 967 value_worklist.push(field_val); 968 } 969 sfpt->add_req(field_val); 970 } 971 JVMState *jvms = sfpt->jvms(); 972 jvms->set_endoff(sfpt->req()); 973 // Now make a pass over the debug information replacing any references 974 // to the allocated object with "sobj" 975 int start = jvms->debug_start(); 976 int end = jvms->debug_end(); 977 sfpt->replace_edges_in_range(res, sobj, start, end); 978 _igvn._worklist.push(sfpt); 979 safepoints_done.append_if_missing(sfpt); // keep it for rollback 980 } 981 // Scalarize value types that were added to the safepoint 982 for (uint i = 0; i < value_worklist.size(); ++i) { 983 Node* vt = value_worklist.at(i); 984 vt->as_ValueType()->make_scalar_in_safepoints(&_igvn); 985 } 986 return true; 987 } 988 989 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) { 990 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control); 991 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory); 992 if (ctl_proj != NULL) { 993 igvn.replace_node(ctl_proj, n->in(0)); 994 } 995 if (mem_proj != NULL) { 996 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory)); 997 } 998 } 999 1000 // Process users of eliminated allocation. 1001 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) { 1002 Node* res = alloc->result_cast(); 1003 if (res != NULL) { 1004 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 1005 Node *use = res->last_out(j); 1006 uint oc1 = res->outcnt(); 1007 1008 if (use->is_AddP()) { 1009 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 1010 Node *n = use->last_out(k); 1011 uint oc2 = use->outcnt(); 1012 if (n->is_Store()) { 1013 #ifdef ASSERT 1014 // Verify that there is no dependent MemBarVolatile nodes, 1015 // they should be removed during IGVN, see MemBarNode::Ideal(). 1016 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); 1017 p < pmax; p++) { 1018 Node* mb = n->fast_out(p); 1019 assert(mb->is_Initialize() || !mb->is_MemBar() || 1020 mb->req() <= MemBarNode::Precedent || 1021 mb->in(MemBarNode::Precedent) != n, 1022 "MemBarVolatile should be eliminated for non-escaping object"); 1023 } 1024 #endif 1025 _igvn.replace_node(n, n->in(MemNode::Memory)); 1026 } else if (n->is_ArrayCopy()) { 1027 // Disconnect ArrayCopy node 1028 ArrayCopyNode* ac = n->as_ArrayCopy(); 1029 assert(ac->is_clonebasic(), "unexpected array copy kind"); 1030 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out(); 1031 disconnect_projections(ac, _igvn); 1032 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation"); 1033 Node* membar_before = alloc->in(0)->in(0); 1034 disconnect_projections(membar_before->as_MemBar(), _igvn); 1035 if (membar_after->is_MemBar()) { 1036 disconnect_projections(membar_after->as_MemBar(), _igvn); 1037 } 1038 } else { 1039 eliminate_gc_barrier(n); 1040 } 1041 k -= (oc2 - use->outcnt()); 1042 } 1043 _igvn.remove_dead_node(use); 1044 } else if (use->is_ArrayCopy()) { 1045 // Disconnect ArrayCopy node 1046 ArrayCopyNode* ac = use->as_ArrayCopy(); 1047 assert(ac->is_arraycopy_validated() || 1048 ac->is_copyof_validated() || 1049 ac->is_copyofrange_validated(), "unsupported"); 1050 CallProjections* callprojs = ac->extract_projections(true); 1051 1052 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O)); 1053 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory)); 1054 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control)); 1055 1056 // Set control to top. IGVN will remove the remaining projections 1057 ac->set_req(0, top()); 1058 ac->replace_edge(res, top()); 1059 1060 // Disconnect src right away: it can help find new 1061 // opportunities for allocation elimination 1062 Node* src = ac->in(ArrayCopyNode::Src); 1063 ac->replace_edge(src, top()); 1064 // src can be top at this point if src and dest of the 1065 // arraycopy were the same 1066 if (src->outcnt() == 0 && !src->is_top()) { 1067 _igvn.remove_dead_node(src); 1068 } 1069 1070 _igvn._worklist.push(ac); 1071 } else if (use->is_ValueType()) { 1072 assert(use->isa_ValueType()->get_oop() == res, "unexpected value type use"); 1073 _igvn.rehash_node_delayed(use); 1074 use->isa_ValueType()->set_oop(_igvn.zerocon(T_VALUETYPE)); 1075 } else if (use->is_Store()) { 1076 _igvn.replace_node(use, use->in(MemNode::Memory)); 1077 } else { 1078 eliminate_gc_barrier(use); 1079 } 1080 j -= (oc1 - res->outcnt()); 1081 } 1082 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 1083 _igvn.remove_dead_node(res); 1084 } 1085 1086 // 1087 // Process other users of allocation's projections 1088 // 1089 if (_resproj != NULL && _resproj->outcnt() != 0) { 1090 // First disconnect stores captured by Initialize node. 1091 // If Initialize node is eliminated first in the following code, 1092 // it will kill such stores and DUIterator_Last will assert. 1093 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) { 1094 Node *use = _resproj->fast_out(j); 1095 if (use->is_AddP()) { 1096 // raw memory addresses used only by the initialization 1097 _igvn.replace_node(use, C->top()); 1098 --j; --jmax; 1099 } 1100 } 1101 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { 1102 Node *use = _resproj->last_out(j); 1103 uint oc1 = _resproj->outcnt(); 1104 if (use->is_Initialize()) { 1105 // Eliminate Initialize node. 1106 InitializeNode *init = use->as_Initialize(); 1107 assert(init->outcnt() <= 2, "only a control and memory projection expected"); 1108 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control); 1109 if (ctrl_proj != NULL) { 1110 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection"); 1111 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj); 1112 } 1113 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory); 1114 if (mem_proj != NULL) { 1115 Node *mem = init->in(TypeFunc::Memory); 1116 #ifdef ASSERT 1117 if (mem->is_MergeMem()) { 1118 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection"); 1119 } else { 1120 assert(mem == _memproj_fallthrough, "allocation memory projection"); 1121 } 1122 #endif 1123 _igvn.replace_node(mem_proj, mem); 1124 } 1125 } else { 1126 assert(false, "only Initialize or AddP expected"); 1127 } 1128 j -= (oc1 - _resproj->outcnt()); 1129 } 1130 } 1131 if (_fallthroughcatchproj != NULL) { 1132 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); 1133 } 1134 if (_memproj_fallthrough != NULL) { 1135 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); 1136 } 1137 if (_memproj_catchall != NULL) { 1138 _igvn.replace_node(_memproj_catchall, C->top()); 1139 } 1140 if (_ioproj_fallthrough != NULL) { 1141 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); 1142 } 1143 if (_ioproj_catchall != NULL) { 1144 _igvn.replace_node(_ioproj_catchall, C->top()); 1145 } 1146 if (_catchallcatchproj != NULL) { 1147 _igvn.replace_node(_catchallcatchproj, C->top()); 1148 } 1149 } 1150 1151 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 1152 // Don't do scalar replacement if the frame can be popped by JVMTI: 1153 // if reallocation fails during deoptimization we'll pop all 1154 // interpreter frames for this compiled frame and that won't play 1155 // nice with JVMTI popframe. 1156 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) { 1157 return false; 1158 } 1159 Node* klass = alloc->in(AllocateNode::KlassNode); 1160 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); 1161 Node* res = alloc->result_cast(); 1162 // Eliminate boxing allocations which are not used 1163 // regardless scalar replacable status. 1164 bool boxing_alloc = C->eliminate_boxing() && 1165 tklass->klass()->is_instance_klass() && 1166 tklass->klass()->as_instance_klass()->is_box_klass(); 1167 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) { 1168 return false; 1169 } 1170 1171 extract_call_projections(alloc); 1172 1173 GrowableArray <SafePointNode *> safepoints; 1174 if (!can_eliminate_allocation(alloc, safepoints)) { 1175 return false; 1176 } 1177 1178 if (!alloc->_is_scalar_replaceable) { 1179 assert(res == NULL, "sanity"); 1180 // We can only eliminate allocation if all debug info references 1181 // are already replaced with SafePointScalarObject because 1182 // we can't search for a fields value without instance_id. 1183 if (safepoints.length() > 0) { 1184 return false; 1185 } 1186 } 1187 1188 if (!scalar_replacement(alloc, safepoints)) { 1189 return false; 1190 } 1191 1192 CompileLog* log = C->log(); 1193 if (log != NULL) { 1194 log->head("eliminate_allocation type='%d'", 1195 log->identify(tklass->klass())); 1196 JVMState* p = alloc->jvms(); 1197 while (p != NULL) { 1198 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1199 p = p->caller(); 1200 } 1201 log->tail("eliminate_allocation"); 1202 } 1203 1204 process_users_of_allocation(alloc); 1205 1206 #ifndef PRODUCT 1207 if (PrintEliminateAllocations) { 1208 if (alloc->is_AllocateArray()) 1209 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1210 else 1211 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1212 } 1213 #endif 1214 1215 return true; 1216 } 1217 1218 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) { 1219 // EA should remove all uses of non-escaping boxing node. 1220 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) { 1221 return false; 1222 } 1223 1224 assert(boxing->result_cast() == NULL, "unexpected boxing node result"); 1225 1226 extract_call_projections(boxing); 1227 1228 const TypeTuple* r = boxing->tf()->range_sig(); 1229 assert(r->cnt() > TypeFunc::Parms, "sanity"); 1230 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr(); 1231 assert(t != NULL, "sanity"); 1232 1233 CompileLog* log = C->log(); 1234 if (log != NULL) { 1235 log->head("eliminate_boxing type='%d'", 1236 log->identify(t->klass())); 1237 JVMState* p = boxing->jvms(); 1238 while (p != NULL) { 1239 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1240 p = p->caller(); 1241 } 1242 log->tail("eliminate_boxing"); 1243 } 1244 1245 process_users_of_allocation(boxing); 1246 1247 #ifndef PRODUCT 1248 if (PrintEliminateAllocations) { 1249 tty->print("++++ Eliminated: %d ", boxing->_idx); 1250 boxing->method()->print_short_name(tty); 1251 tty->cr(); 1252 } 1253 #endif 1254 1255 return true; 1256 } 1257 1258 //---------------------------set_eden_pointers------------------------- 1259 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { 1260 if (UseTLAB) { // Private allocation: load from TLS 1261 Node* thread = transform_later(new ThreadLocalNode()); 1262 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); 1263 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); 1264 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); 1265 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); 1266 } else { // Shared allocation: load from globals 1267 CollectedHeap* ch = Universe::heap(); 1268 address top_adr = (address)ch->top_addr(); 1269 address end_adr = (address)ch->end_addr(); 1270 eden_top_adr = makecon(TypeRawPtr::make(top_adr)); 1271 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); 1272 } 1273 } 1274 1275 1276 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 1277 Node* adr = basic_plus_adr(base, offset); 1278 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 1279 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered); 1280 transform_later(value); 1281 return value; 1282 } 1283 1284 1285 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 1286 Node* adr = basic_plus_adr(base, offset); 1287 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered); 1288 transform_later(mem); 1289 return mem; 1290 } 1291 1292 //============================================================================= 1293 // 1294 // A L L O C A T I O N 1295 // 1296 // Allocation attempts to be fast in the case of frequent small objects. 1297 // It breaks down like this: 1298 // 1299 // 1) Size in doublewords is computed. This is a constant for objects and 1300 // variable for most arrays. Doubleword units are used to avoid size 1301 // overflow of huge doubleword arrays. We need doublewords in the end for 1302 // rounding. 1303 // 1304 // 2) Size is checked for being 'too large'. Too-large allocations will go 1305 // the slow path into the VM. The slow path can throw any required 1306 // exceptions, and does all the special checks for very large arrays. The 1307 // size test can constant-fold away for objects. For objects with 1308 // finalizers it constant-folds the otherway: you always go slow with 1309 // finalizers. 1310 // 1311 // 3) If NOT using TLABs, this is the contended loop-back point. 1312 // Load-Locked the heap top. If using TLABs normal-load the heap top. 1313 // 1314 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 1315 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 1316 // "size*8" we always enter the VM, where "largish" is a constant picked small 1317 // enough that there's always space between the eden max and 4Gig (old space is 1318 // there so it's quite large) and large enough that the cost of entering the VM 1319 // is dwarfed by the cost to initialize the space. 1320 // 1321 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 1322 // down. If contended, repeat at step 3. If using TLABs normal-store 1323 // adjusted heap top back down; there is no contention. 1324 // 1325 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 1326 // fields. 1327 // 1328 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 1329 // oop flavor. 1330 // 1331 //============================================================================= 1332 // FastAllocateSizeLimit value is in DOUBLEWORDS. 1333 // Allocations bigger than this always go the slow route. 1334 // This value must be small enough that allocation attempts that need to 1335 // trigger exceptions go the slow route. Also, it must be small enough so 1336 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 1337 //=============================================================================j// 1338 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 1339 // The allocator will coalesce int->oop copies away. See comment in 1340 // coalesce.cpp about how this works. It depends critically on the exact 1341 // code shape produced here, so if you are changing this code shape 1342 // make sure the GC info for the heap-top is correct in and around the 1343 // slow-path call. 1344 // 1345 1346 void PhaseMacroExpand::expand_allocate_common( 1347 AllocateNode* alloc, // allocation node to be expanded 1348 Node* length, // array length for an array allocation 1349 const TypeFunc* slow_call_type, // Type of slow call 1350 address slow_call_address // Address of slow call 1351 ) 1352 { 1353 1354 Node* ctrl = alloc->in(TypeFunc::Control); 1355 Node* mem = alloc->in(TypeFunc::Memory); 1356 Node* i_o = alloc->in(TypeFunc::I_O); 1357 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 1358 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1359 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 1360 1361 assert(ctrl != NULL, "must have control"); 1362 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 1363 // they will not be used if "always_slow" is set 1364 enum { slow_result_path = 1, fast_result_path = 2 }; 1365 Node *result_region = NULL; 1366 Node *result_phi_rawmem = NULL; 1367 Node *result_phi_rawoop = NULL; 1368 Node *result_phi_i_o = NULL; 1369 1370 // The initial slow comparison is a size check, the comparison 1371 // we want to do is a BoolTest::gt 1372 bool always_slow = false; 1373 int tv = _igvn.find_int_con(initial_slow_test, -1); 1374 if (tv >= 0) { 1375 always_slow = (tv == 1); 1376 initial_slow_test = NULL; 1377 } else { 1378 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 1379 } 1380 1381 if (C->env()->dtrace_alloc_probes() || 1382 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) { 1383 // Force slow-path allocation 1384 always_slow = true; 1385 initial_slow_test = NULL; 1386 } 1387 1388 Node *slow_region = NULL; 1389 Node *toobig_false = ctrl; 1390 1391 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); 1392 // generate the initial test if necessary 1393 if (initial_slow_test != NULL ) { 1394 if (slow_region == NULL) { 1395 slow_region = new RegionNode(1); 1396 } 1397 // Now make the initial failure test. Usually a too-big test but 1398 // might be a TRUE for finalizers or a fancy class check for 1399 // newInstance0. 1400 IfNode* toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1401 transform_later(toobig_iff); 1402 // Plug the failing-too-big test into the slow-path region 1403 Node* toobig_true = new IfTrueNode(toobig_iff); 1404 transform_later(toobig_true); 1405 slow_region ->add_req(toobig_true); 1406 toobig_false = new IfFalseNode(toobig_iff); 1407 transform_later(toobig_false); 1408 } else { // No initial test, just fall into next case 1409 toobig_false = ctrl; 1410 } 1411 1412 Node *slow_mem = mem; // save the current memory state for slow path 1413 // generate the fast allocation code unless we know that the initial test will always go slow 1414 if (!always_slow) { 1415 // Fast path modifies only raw memory. 1416 if (mem->is_MergeMem()) { 1417 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1418 } 1419 1420 // allocate the Region and Phi nodes for the result 1421 result_region = new RegionNode(3); 1422 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1423 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM); 1424 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch 1425 1426 // Grab regular I/O before optional prefetch may change it. 1427 // Slow-path does no I/O so just set it to the original I/O. 1428 result_phi_i_o->init_req(slow_result_path, i_o); 1429 1430 Node* needgc_ctrl = NULL; 1431 // Name successful fast-path variables 1432 Node* fast_oop_ctrl; 1433 Node* fast_oop_rawmem; 1434 1435 intx prefetch_lines = length != NULL ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1436 1437 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1438 Node* fast_oop = bs->obj_allocate(this, ctrl, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl, 1439 fast_oop_ctrl, fast_oop_rawmem, 1440 prefetch_lines); 1441 1442 if (slow_region != NULL) { 1443 slow_region->add_req(needgc_ctrl); 1444 // This completes all paths into the slow merge point 1445 transform_later(slow_region); 1446 } else { 1447 // Just fall from the need-GC path straight into the VM call. 1448 slow_region = needgc_ctrl; 1449 } 1450 1451 InitializeNode* init = alloc->initialization(); 1452 fast_oop_rawmem = initialize_object(alloc, 1453 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1454 klass_node, length, size_in_bytes); 1455 1456 // If initialization is performed by an array copy, any required 1457 // MemBarStoreStore was already added. If the object does not 1458 // escape no need for a MemBarStoreStore. If the object does not 1459 // escape in its initializer and memory barrier (MemBarStoreStore or 1460 // stronger) is already added at exit of initializer, also no need 1461 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore 1462 // so that stores that initialize this object can't be reordered 1463 // with a subsequent store that makes this object accessible by 1464 // other threads. 1465 // Other threads include java threads and JVM internal threads 1466 // (for example concurrent GC threads). Current concurrent GC 1467 // implementation: CMS and G1 will not scan newly created object, 1468 // so it's safe to skip storestore barrier when allocation does 1469 // not escape. 1470 if (!alloc->does_not_escape_thread() && 1471 !alloc->is_allocation_MemBar_redundant() && 1472 (init == NULL || !init->is_complete_with_arraycopy())) { 1473 if (init == NULL || init->req() < InitializeNode::RawStores) { 1474 // No InitializeNode or no stores captured by zeroing 1475 // elimination. Simply add the MemBarStoreStore after object 1476 // initialization. 1477 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1478 transform_later(mb); 1479 1480 mb->init_req(TypeFunc::Memory, fast_oop_rawmem); 1481 mb->init_req(TypeFunc::Control, fast_oop_ctrl); 1482 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control); 1483 transform_later(fast_oop_ctrl); 1484 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory); 1485 transform_later(fast_oop_rawmem); 1486 } else { 1487 // Add the MemBarStoreStore after the InitializeNode so that 1488 // all stores performing the initialization that were moved 1489 // before the InitializeNode happen before the storestore 1490 // barrier. 1491 1492 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control); 1493 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory); 1494 1495 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1496 transform_later(mb); 1497 1498 Node* ctrl = new ProjNode(init,TypeFunc::Control); 1499 transform_later(ctrl); 1500 Node* mem = new ProjNode(init,TypeFunc::Memory); 1501 transform_later(mem); 1502 1503 // The MemBarStoreStore depends on control and memory coming 1504 // from the InitializeNode 1505 mb->init_req(TypeFunc::Memory, mem); 1506 mb->init_req(TypeFunc::Control, ctrl); 1507 1508 ctrl = new ProjNode(mb,TypeFunc::Control); 1509 transform_later(ctrl); 1510 mem = new ProjNode(mb,TypeFunc::Memory); 1511 transform_later(mem); 1512 1513 // All nodes that depended on the InitializeNode for control 1514 // and memory must now depend on the MemBarNode that itself 1515 // depends on the InitializeNode 1516 if (init_ctrl != NULL) { 1517 _igvn.replace_node(init_ctrl, ctrl); 1518 } 1519 if (init_mem != NULL) { 1520 _igvn.replace_node(init_mem, mem); 1521 } 1522 } 1523 } 1524 1525 if (C->env()->dtrace_extended_probes()) { 1526 // Slow-path call 1527 int size = TypeFunc::Parms + 2; 1528 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1529 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), 1530 "dtrace_object_alloc", 1531 TypeRawPtr::BOTTOM); 1532 1533 // Get base of thread-local storage area 1534 Node* thread = new ThreadLocalNode(); 1535 transform_later(thread); 1536 1537 call->init_req(TypeFunc::Parms+0, thread); 1538 call->init_req(TypeFunc::Parms+1, fast_oop); 1539 call->init_req(TypeFunc::Control, fast_oop_ctrl); 1540 call->init_req(TypeFunc::I_O , top()); // does no i/o 1541 call->init_req(TypeFunc::Memory , fast_oop_rawmem); 1542 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1543 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1544 transform_later(call); 1545 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control); 1546 transform_later(fast_oop_ctrl); 1547 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory); 1548 transform_later(fast_oop_rawmem); 1549 } 1550 1551 // Plug in the successful fast-path into the result merge point 1552 result_region ->init_req(fast_result_path, fast_oop_ctrl); 1553 result_phi_rawoop->init_req(fast_result_path, fast_oop); 1554 result_phi_i_o ->init_req(fast_result_path, i_o); 1555 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); 1556 } else { 1557 slow_region = ctrl; 1558 result_phi_i_o = i_o; // Rename it to use in the following code. 1559 } 1560 1561 // Generate slow-path call 1562 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address, 1563 OptoRuntime::stub_name(slow_call_address), 1564 alloc->jvms()->bci(), 1565 TypePtr::BOTTOM); 1566 call->init_req( TypeFunc::Control, slow_region ); 1567 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1568 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs 1569 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1570 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1571 1572 call->init_req(TypeFunc::Parms+0, klass_node); 1573 if (length != NULL) { 1574 call->init_req(TypeFunc::Parms+1, length); 1575 } 1576 1577 // Copy debug information and adjust JVMState information, then replace 1578 // allocate node with the call 1579 copy_call_debug_info((CallNode *) alloc, call); 1580 if (!always_slow) { 1581 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1582 } else { 1583 // Hook i_o projection to avoid its elimination during allocation 1584 // replacement (when only a slow call is generated). 1585 call->set_req(TypeFunc::I_O, result_phi_i_o); 1586 } 1587 _igvn.replace_node(alloc, call); 1588 transform_later(call); 1589 1590 // Identify the output projections from the allocate node and 1591 // adjust any references to them. 1592 // The control and io projections look like: 1593 // 1594 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1595 // Allocate Catch 1596 // ^---Proj(io) <-------+ ^---CatchProj(io) 1597 // 1598 // We are interested in the CatchProj nodes. 1599 // 1600 extract_call_projections(call); 1601 1602 // An allocate node has separate memory projections for the uses on 1603 // the control and i_o paths. Replace the control memory projection with 1604 // result_phi_rawmem (unless we are only generating a slow call when 1605 // both memory projections are combined) 1606 if (!always_slow && _memproj_fallthrough != NULL) { 1607 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { 1608 Node *use = _memproj_fallthrough->fast_out(i); 1609 _igvn.rehash_node_delayed(use); 1610 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); 1611 // back up iterator 1612 --i; 1613 } 1614 } 1615 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete 1616 // _memproj_catchall so we end up with a call that has only 1 memory projection. 1617 if (_memproj_catchall != NULL ) { 1618 if (_memproj_fallthrough == NULL) { 1619 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory); 1620 transform_later(_memproj_fallthrough); 1621 } 1622 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { 1623 Node *use = _memproj_catchall->fast_out(i); 1624 _igvn.rehash_node_delayed(use); 1625 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); 1626 // back up iterator 1627 --i; 1628 } 1629 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted"); 1630 _igvn.remove_dead_node(_memproj_catchall); 1631 } 1632 1633 // An allocate node has separate i_o projections for the uses on the control 1634 // and i_o paths. Always replace the control i_o projection with result i_o 1635 // otherwise incoming i_o become dead when only a slow call is generated 1636 // (it is different from memory projections where both projections are 1637 // combined in such case). 1638 if (_ioproj_fallthrough != NULL) { 1639 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { 1640 Node *use = _ioproj_fallthrough->fast_out(i); 1641 _igvn.rehash_node_delayed(use); 1642 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); 1643 // back up iterator 1644 --i; 1645 } 1646 } 1647 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete 1648 // _ioproj_catchall so we end up with a call that has only 1 i_o projection. 1649 if (_ioproj_catchall != NULL ) { 1650 if (_ioproj_fallthrough == NULL) { 1651 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O); 1652 transform_later(_ioproj_fallthrough); 1653 } 1654 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { 1655 Node *use = _ioproj_catchall->fast_out(i); 1656 _igvn.rehash_node_delayed(use); 1657 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); 1658 // back up iterator 1659 --i; 1660 } 1661 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted"); 1662 _igvn.remove_dead_node(_ioproj_catchall); 1663 } 1664 1665 // if we generated only a slow call, we are done 1666 if (always_slow) { 1667 // Now we can unhook i_o. 1668 if (result_phi_i_o->outcnt() > 1) { 1669 call->set_req(TypeFunc::I_O, top()); 1670 } else { 1671 assert(result_phi_i_o->unique_ctrl_out() == call, ""); 1672 // Case of new array with negative size known during compilation. 1673 // AllocateArrayNode::Ideal() optimization disconnect unreachable 1674 // following code since call to runtime will throw exception. 1675 // As result there will be no users of i_o after the call. 1676 // Leave i_o attached to this call to avoid problems in preceding graph. 1677 } 1678 return; 1679 } 1680 1681 if (_fallthroughcatchproj != NULL) { 1682 ctrl = _fallthroughcatchproj->clone(); 1683 transform_later(ctrl); 1684 _igvn.replace_node(_fallthroughcatchproj, result_region); 1685 } else { 1686 ctrl = top(); 1687 } 1688 Node *slow_result; 1689 if (_resproj == NULL) { 1690 // no uses of the allocation result 1691 slow_result = top(); 1692 } else { 1693 slow_result = _resproj->clone(); 1694 transform_later(slow_result); 1695 _igvn.replace_node(_resproj, result_phi_rawoop); 1696 } 1697 1698 // Plug slow-path into result merge point 1699 result_region ->init_req( slow_result_path, ctrl ); 1700 result_phi_rawoop->init_req( slow_result_path, slow_result); 1701 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); 1702 transform_later(result_region); 1703 transform_later(result_phi_rawoop); 1704 transform_later(result_phi_rawmem); 1705 transform_later(result_phi_i_o); 1706 // This completes all paths into the result merge point 1707 } 1708 1709 1710 // Helper for PhaseMacroExpand::expand_allocate_common. 1711 // Initializes the newly-allocated storage. 1712 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1713 Node* control, Node* rawmem, Node* object, 1714 Node* klass_node, Node* length, 1715 Node* size_in_bytes) { 1716 InitializeNode* init = alloc->initialization(); 1717 // Store the klass & mark bits 1718 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem, klass_node); 1719 if (!mark_node->is_Con()) { 1720 transform_later(mark_node); 1721 } 1722 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type()); 1723 1724 BasicType bt = T_METADATA; 1725 Node* metadata = klass_node; 1726 Node* properties = alloc->in(AllocateNode::StorageProperties); 1727 if (properties != NULL) { 1728 // Encode array storage properties into klass pointer 1729 assert(EnableValhalla, "array storage properties not supported"); 1730 if (UseCompressedClassPointers) { 1731 // Compress the klass pointer before inserting the storage properties value 1732 metadata = transform_later(new EncodePKlassNode(metadata, metadata->bottom_type()->make_narrowklass())); 1733 metadata = transform_later(new CastN2INode(metadata)); 1734 metadata = transform_later(new OrINode(metadata, transform_later(new ConvL2INode(properties)))); 1735 bt = T_INT; 1736 } else { 1737 metadata = transform_later(new CastP2XNode(NULL, metadata)); 1738 metadata = transform_later(new OrXNode(metadata, properties)); 1739 bt = T_LONG; 1740 } 1741 } 1742 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), metadata, bt); 1743 1744 int header_size = alloc->minimum_header_size(); // conservatively small 1745 1746 // Array length 1747 if (length != NULL) { // Arrays need length field 1748 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1749 // conservatively small header size: 1750 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1751 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1752 if (k->is_array_klass()) // we know the exact header size in most cases: 1753 header_size = Klass::layout_helper_header_size(k->layout_helper()); 1754 } 1755 1756 // Clear the object body, if necessary. 1757 if (init == NULL) { 1758 // The init has somehow disappeared; be cautious and clear everything. 1759 // 1760 // This can happen if a node is allocated but an uncommon trap occurs 1761 // immediately. In this case, the Initialize gets associated with the 1762 // trap, and may be placed in a different (outer) loop, if the Allocate 1763 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1764 // there can be two Allocates to one Initialize. The answer in all these 1765 // edge cases is safety first. It is always safe to clear immediately 1766 // within an Allocate, and then (maybe or maybe not) clear some more later. 1767 if (!(UseTLAB && ZeroTLAB)) { 1768 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1769 alloc->in(AllocateNode::DefaultValue), 1770 alloc->in(AllocateNode::RawDefaultValue), 1771 header_size, size_in_bytes, 1772 &_igvn); 1773 } 1774 } else { 1775 if (!init->is_complete()) { 1776 // Try to win by zeroing only what the init does not store. 1777 // We can also try to do some peephole optimizations, 1778 // such as combining some adjacent subword stores. 1779 rawmem = init->complete_stores(control, rawmem, object, 1780 header_size, size_in_bytes, &_igvn); 1781 } 1782 // We have no more use for this link, since the AllocateNode goes away: 1783 init->set_req(InitializeNode::RawAddress, top()); 1784 // (If we keep the link, it just confuses the register allocator, 1785 // who thinks he sees a real use of the address by the membar.) 1786 } 1787 1788 return rawmem; 1789 } 1790 1791 // Generate prefetch instructions for next allocations. 1792 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1793 Node*& contended_phi_rawmem, 1794 Node* old_eden_top, Node* new_eden_top, 1795 intx lines) { 1796 enum { fall_in_path = 1, pf_path = 2 }; 1797 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1798 // Generate prefetch allocation with watermark check. 1799 // As an allocation hits the watermark, we will prefetch starting 1800 // at a "distance" away from watermark. 1801 1802 Node *pf_region = new RegionNode(3); 1803 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 1804 TypeRawPtr::BOTTOM ); 1805 // I/O is used for Prefetch 1806 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO ); 1807 1808 Node *thread = new ThreadLocalNode(); 1809 transform_later(thread); 1810 1811 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread, 1812 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1813 transform_later(eden_pf_adr); 1814 1815 Node *old_pf_wm = new LoadPNode(needgc_false, 1816 contended_phi_rawmem, eden_pf_adr, 1817 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, 1818 MemNode::unordered); 1819 transform_later(old_pf_wm); 1820 1821 // check against new_eden_top 1822 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm ); 1823 transform_later(need_pf_cmp); 1824 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge ); 1825 transform_later(need_pf_bol); 1826 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol, 1827 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1828 transform_later(need_pf_iff); 1829 1830 // true node, add prefetchdistance 1831 Node *need_pf_true = new IfTrueNode( need_pf_iff ); 1832 transform_later(need_pf_true); 1833 1834 Node *need_pf_false = new IfFalseNode( need_pf_iff ); 1835 transform_later(need_pf_false); 1836 1837 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm, 1838 _igvn.MakeConX(AllocatePrefetchDistance) ); 1839 transform_later(new_pf_wmt ); 1840 new_pf_wmt->set_req(0, need_pf_true); 1841 1842 Node *store_new_wmt = new StorePNode(need_pf_true, 1843 contended_phi_rawmem, eden_pf_adr, 1844 TypeRawPtr::BOTTOM, new_pf_wmt, 1845 MemNode::unordered); 1846 transform_later(store_new_wmt); 1847 1848 // adding prefetches 1849 pf_phi_abio->init_req( fall_in_path, i_o ); 1850 1851 Node *prefetch_adr; 1852 Node *prefetch; 1853 uint step_size = AllocatePrefetchStepSize; 1854 uint distance = 0; 1855 1856 for ( intx i = 0; i < lines; i++ ) { 1857 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt, 1858 _igvn.MakeConX(distance) ); 1859 transform_later(prefetch_adr); 1860 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1861 transform_later(prefetch); 1862 distance += step_size; 1863 i_o = prefetch; 1864 } 1865 pf_phi_abio->set_req( pf_path, i_o ); 1866 1867 pf_region->init_req( fall_in_path, need_pf_false ); 1868 pf_region->init_req( pf_path, need_pf_true ); 1869 1870 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 1871 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 1872 1873 transform_later(pf_region); 1874 transform_later(pf_phi_rawmem); 1875 transform_later(pf_phi_abio); 1876 1877 needgc_false = pf_region; 1878 contended_phi_rawmem = pf_phi_rawmem; 1879 i_o = pf_phi_abio; 1880 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { 1881 // Insert a prefetch instruction for each allocation. 1882 // This code is used to generate 1 prefetch instruction per cache line. 1883 1884 // Generate several prefetch instructions. 1885 uint step_size = AllocatePrefetchStepSize; 1886 uint distance = AllocatePrefetchDistance; 1887 1888 // Next cache address. 1889 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top, 1890 _igvn.MakeConX(step_size + distance)); 1891 transform_later(cache_adr); 1892 cache_adr = new CastP2XNode(needgc_false, cache_adr); 1893 transform_later(cache_adr); 1894 // Address is aligned to execute prefetch to the beginning of cache line size 1895 // (it is important when BIS instruction is used on SPARC as prefetch). 1896 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); 1897 cache_adr = new AndXNode(cache_adr, mask); 1898 transform_later(cache_adr); 1899 cache_adr = new CastX2PNode(cache_adr); 1900 transform_later(cache_adr); 1901 1902 // Prefetch 1903 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); 1904 prefetch->set_req(0, needgc_false); 1905 transform_later(prefetch); 1906 contended_phi_rawmem = prefetch; 1907 Node *prefetch_adr; 1908 distance = step_size; 1909 for ( intx i = 1; i < lines; i++ ) { 1910 prefetch_adr = new AddPNode( cache_adr, cache_adr, 1911 _igvn.MakeConX(distance) ); 1912 transform_later(prefetch_adr); 1913 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); 1914 transform_later(prefetch); 1915 distance += step_size; 1916 contended_phi_rawmem = prefetch; 1917 } 1918 } else if( AllocatePrefetchStyle > 0 ) { 1919 // Insert a prefetch for each allocation only on the fast-path 1920 Node *prefetch_adr; 1921 Node *prefetch; 1922 // Generate several prefetch instructions. 1923 uint step_size = AllocatePrefetchStepSize; 1924 uint distance = AllocatePrefetchDistance; 1925 for ( intx i = 0; i < lines; i++ ) { 1926 prefetch_adr = new AddPNode( old_eden_top, new_eden_top, 1927 _igvn.MakeConX(distance) ); 1928 transform_later(prefetch_adr); 1929 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1930 // Do not let it float too high, since if eden_top == eden_end, 1931 // both might be null. 1932 if( i == 0 ) { // Set control for first prefetch, next follows it 1933 prefetch->init_req(0, needgc_false); 1934 } 1935 transform_later(prefetch); 1936 distance += step_size; 1937 i_o = prefetch; 1938 } 1939 } 1940 return i_o; 1941 } 1942 1943 1944 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 1945 expand_allocate_common(alloc, NULL, 1946 OptoRuntime::new_instance_Type(), 1947 OptoRuntime::new_instance_Java()); 1948 } 1949 1950 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 1951 Node* length = alloc->in(AllocateNode::ALength); 1952 InitializeNode* init = alloc->initialization(); 1953 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1954 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1955 address slow_call_address; // Address of slow call 1956 if (init != NULL && init->is_complete_with_arraycopy() && 1957 k->is_type_array_klass()) { 1958 // Don't zero type array during slow allocation in VM since 1959 // it will be initialized later by arraycopy in compiled code. 1960 slow_call_address = OptoRuntime::new_array_nozero_Java(); 1961 } else { 1962 slow_call_address = OptoRuntime::new_array_Java(); 1963 } 1964 expand_allocate_common(alloc, length, 1965 OptoRuntime::new_array_Type(), 1966 slow_call_address); 1967 } 1968 1969 //-------------------mark_eliminated_box---------------------------------- 1970 // 1971 // During EA obj may point to several objects but after few ideal graph 1972 // transformations (CCP) it may point to only one non escaping object 1973 // (but still using phi), corresponding locks and unlocks will be marked 1974 // for elimination. Later obj could be replaced with a new node (new phi) 1975 // and which does not have escape information. And later after some graph 1976 // reshape other locks and unlocks (which were not marked for elimination 1977 // before) are connected to this new obj (phi) but they still will not be 1978 // marked for elimination since new obj has no escape information. 1979 // Mark all associated (same box and obj) lock and unlock nodes for 1980 // elimination if some of them marked already. 1981 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) { 1982 if (oldbox->as_BoxLock()->is_eliminated()) 1983 return; // This BoxLock node was processed already. 1984 1985 // New implementation (EliminateNestedLocks) has separate BoxLock 1986 // node for each locked region so mark all associated locks/unlocks as 1987 // eliminated even if different objects are referenced in one locked region 1988 // (for example, OSR compilation of nested loop inside locked scope). 1989 if (EliminateNestedLocks || 1990 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) { 1991 // Box is used only in one lock region. Mark this box as eliminated. 1992 _igvn.hash_delete(oldbox); 1993 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value 1994 _igvn.hash_insert(oldbox); 1995 1996 for (uint i = 0; i < oldbox->outcnt(); i++) { 1997 Node* u = oldbox->raw_out(i); 1998 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) { 1999 AbstractLockNode* alock = u->as_AbstractLock(); 2000 // Check lock's box since box could be referenced by Lock's debug info. 2001 if (alock->box_node() == oldbox) { 2002 // Mark eliminated all related locks and unlocks. 2003 #ifdef ASSERT 2004 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4"); 2005 #endif 2006 alock->set_non_esc_obj(); 2007 } 2008 } 2009 } 2010 return; 2011 } 2012 2013 // Create new "eliminated" BoxLock node and use it in monitor debug info 2014 // instead of oldbox for the same object. 2015 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 2016 2017 // Note: BoxLock node is marked eliminated only here and it is used 2018 // to indicate that all associated lock and unlock nodes are marked 2019 // for elimination. 2020 newbox->set_eliminated(); 2021 transform_later(newbox); 2022 2023 // Replace old box node with new box for all users of the same object. 2024 for (uint i = 0; i < oldbox->outcnt();) { 2025 bool next_edge = true; 2026 2027 Node* u = oldbox->raw_out(i); 2028 if (u->is_AbstractLock()) { 2029 AbstractLockNode* alock = u->as_AbstractLock(); 2030 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) { 2031 // Replace Box and mark eliminated all related locks and unlocks. 2032 #ifdef ASSERT 2033 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5"); 2034 #endif 2035 alock->set_non_esc_obj(); 2036 _igvn.rehash_node_delayed(alock); 2037 alock->set_box_node(newbox); 2038 next_edge = false; 2039 } 2040 } 2041 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) { 2042 FastLockNode* flock = u->as_FastLock(); 2043 assert(flock->box_node() == oldbox, "sanity"); 2044 _igvn.rehash_node_delayed(flock); 2045 flock->set_box_node(newbox); 2046 next_edge = false; 2047 } 2048 2049 // Replace old box in monitor debug info. 2050 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 2051 SafePointNode* sfn = u->as_SafePoint(); 2052 JVMState* youngest_jvms = sfn->jvms(); 2053 int max_depth = youngest_jvms->depth(); 2054 for (int depth = 1; depth <= max_depth; depth++) { 2055 JVMState* jvms = youngest_jvms->of_depth(depth); 2056 int num_mon = jvms->nof_monitors(); 2057 // Loop over monitors 2058 for (int idx = 0; idx < num_mon; idx++) { 2059 Node* obj_node = sfn->monitor_obj(jvms, idx); 2060 Node* box_node = sfn->monitor_box(jvms, idx); 2061 if (box_node == oldbox && obj_node->eqv_uncast(obj)) { 2062 int j = jvms->monitor_box_offset(idx); 2063 _igvn.replace_input_of(u, j, newbox); 2064 next_edge = false; 2065 } 2066 } 2067 } 2068 } 2069 if (next_edge) i++; 2070 } 2071 } 2072 2073 //-----------------------mark_eliminated_locking_nodes----------------------- 2074 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { 2075 if (EliminateNestedLocks) { 2076 if (alock->is_nested()) { 2077 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity"); 2078 return; 2079 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened 2080 // Only Lock node has JVMState needed here. 2081 // Not that preceding claim is documented anywhere else. 2082 if (alock->jvms() != NULL) { 2083 if (alock->as_Lock()->is_nested_lock_region()) { 2084 // Mark eliminated related nested locks and unlocks. 2085 Node* obj = alock->obj_node(); 2086 BoxLockNode* box_node = alock->box_node()->as_BoxLock(); 2087 assert(!box_node->is_eliminated(), "should not be marked yet"); 2088 // Note: BoxLock node is marked eliminated only here 2089 // and it is used to indicate that all associated lock 2090 // and unlock nodes are marked for elimination. 2091 box_node->set_eliminated(); // Box's hash is always NO_HASH here 2092 for (uint i = 0; i < box_node->outcnt(); i++) { 2093 Node* u = box_node->raw_out(i); 2094 if (u->is_AbstractLock()) { 2095 alock = u->as_AbstractLock(); 2096 if (alock->box_node() == box_node) { 2097 // Verify that this Box is referenced only by related locks. 2098 assert(alock->obj_node()->eqv_uncast(obj), ""); 2099 // Mark all related locks and unlocks. 2100 #ifdef ASSERT 2101 alock->log_lock_optimization(C, "eliminate_lock_set_nested"); 2102 #endif 2103 alock->set_nested(); 2104 } 2105 } 2106 } 2107 } else { 2108 #ifdef ASSERT 2109 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region"); 2110 if (C->log() != NULL) 2111 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output 2112 #endif 2113 } 2114 } 2115 return; 2116 } 2117 // Process locks for non escaping object 2118 assert(alock->is_non_esc_obj(), ""); 2119 } // EliminateNestedLocks 2120 2121 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object 2122 // Look for all locks of this object and mark them and 2123 // corresponding BoxLock nodes as eliminated. 2124 Node* obj = alock->obj_node(); 2125 for (uint j = 0; j < obj->outcnt(); j++) { 2126 Node* o = obj->raw_out(j); 2127 if (o->is_AbstractLock() && 2128 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) { 2129 alock = o->as_AbstractLock(); 2130 Node* box = alock->box_node(); 2131 // Replace old box node with new eliminated box for all users 2132 // of the same object and mark related locks as eliminated. 2133 mark_eliminated_box(box, obj); 2134 } 2135 } 2136 } 2137 } 2138 2139 // we have determined that this lock/unlock can be eliminated, we simply 2140 // eliminate the node without expanding it. 2141 // 2142 // Note: The membar's associated with the lock/unlock are currently not 2143 // eliminated. This should be investigated as a future enhancement. 2144 // 2145 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 2146 2147 if (!alock->is_eliminated()) { 2148 return false; 2149 } 2150 #ifdef ASSERT 2151 const Type* obj_type = _igvn.type(alock->obj_node()); 2152 assert(!obj_type->isa_valuetype() && !obj_type->is_valuetypeptr(), "Eliminating lock on value type"); 2153 if (!alock->is_coarsened()) { 2154 // Check that new "eliminated" BoxLock node is created. 2155 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 2156 assert(oldbox->is_eliminated(), "should be done already"); 2157 } 2158 #endif 2159 2160 alock->log_lock_optimization(C, "eliminate_lock"); 2161 2162 #ifndef PRODUCT 2163 if (PrintEliminateLocks) { 2164 if (alock->is_Lock()) { 2165 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx); 2166 } else { 2167 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx); 2168 } 2169 } 2170 #endif 2171 2172 Node* mem = alock->in(TypeFunc::Memory); 2173 Node* ctrl = alock->in(TypeFunc::Control); 2174 guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL"); 2175 2176 extract_call_projections(alock); 2177 // There are 2 projections from the lock. The lock node will 2178 // be deleted when its last use is subsumed below. 2179 assert(alock->outcnt() == 2 && 2180 _fallthroughproj != NULL && 2181 _memproj_fallthrough != NULL, 2182 "Unexpected projections from Lock/Unlock"); 2183 2184 Node* fallthroughproj = _fallthroughproj; 2185 Node* memproj_fallthrough = _memproj_fallthrough; 2186 2187 // The memory projection from a lock/unlock is RawMem 2188 // The input to a Lock is merged memory, so extract its RawMem input 2189 // (unless the MergeMem has been optimized away.) 2190 if (alock->is_Lock()) { 2191 // Seach for MemBarAcquireLock node and delete it also. 2192 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 2193 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, ""); 2194 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 2195 Node* memproj = membar->proj_out(TypeFunc::Memory); 2196 _igvn.replace_node(ctrlproj, fallthroughproj); 2197 _igvn.replace_node(memproj, memproj_fallthrough); 2198 2199 // Delete FastLock node also if this Lock node is unique user 2200 // (a loop peeling may clone a Lock node). 2201 Node* flock = alock->as_Lock()->fastlock_node(); 2202 if (flock->outcnt() == 1) { 2203 assert(flock->unique_out() == alock, "sanity"); 2204 _igvn.replace_node(flock, top()); 2205 } 2206 } 2207 2208 // Seach for MemBarReleaseLock node and delete it also. 2209 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) { 2210 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 2211 assert(membar->Opcode() == Op_MemBarReleaseLock && 2212 mem->is_Proj() && membar == mem->in(0), ""); 2213 _igvn.replace_node(fallthroughproj, ctrl); 2214 _igvn.replace_node(memproj_fallthrough, mem); 2215 fallthroughproj = ctrl; 2216 memproj_fallthrough = mem; 2217 ctrl = membar->in(TypeFunc::Control); 2218 mem = membar->in(TypeFunc::Memory); 2219 } 2220 2221 _igvn.replace_node(fallthroughproj, ctrl); 2222 _igvn.replace_node(memproj_fallthrough, mem); 2223 return true; 2224 } 2225 2226 2227 //------------------------------expand_lock_node---------------------- 2228 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 2229 2230 Node* ctrl = lock->in(TypeFunc::Control); 2231 Node* mem = lock->in(TypeFunc::Memory); 2232 Node* obj = lock->obj_node(); 2233 Node* box = lock->box_node(); 2234 Node* flock = lock->fastlock_node(); 2235 2236 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2237 2238 // Make the merge point 2239 Node *region; 2240 Node *mem_phi; 2241 Node *slow_path; 2242 2243 if (UseOptoBiasInlining) { 2244 /* 2245 * See the full description in MacroAssembler::biased_locking_enter(). 2246 * 2247 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { 2248 * // The object is biased. 2249 * proto_node = klass->prototype_header; 2250 * o_node = thread | proto_node; 2251 * x_node = o_node ^ mark_word; 2252 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? 2253 * // Done. 2254 * } else { 2255 * if( (x_node & biased_lock_mask) != 0 ) { 2256 * // The klass's prototype header is no longer biased. 2257 * cas(&mark_word, mark_word, proto_node) 2258 * goto cas_lock; 2259 * } else { 2260 * // The klass's prototype header is still biased. 2261 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? 2262 * old = mark_word; 2263 * new = o_node; 2264 * } else { 2265 * // Different thread or anonymous biased. 2266 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); 2267 * new = thread | old; 2268 * } 2269 * // Try to rebias. 2270 * if( cas(&mark_word, old, new) == 0 ) { 2271 * // Done. 2272 * } else { 2273 * goto slow_path; // Failed. 2274 * } 2275 * } 2276 * } 2277 * } else { 2278 * // The object is not biased. 2279 * cas_lock: 2280 * if( FastLock(obj) == 0 ) { 2281 * // Done. 2282 * } else { 2283 * slow_path: 2284 * OptoRuntime::complete_monitor_locking_Java(obj); 2285 * } 2286 * } 2287 */ 2288 2289 region = new RegionNode(5); 2290 // create a Phi for the memory state 2291 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2292 2293 Node* fast_lock_region = new RegionNode(3); 2294 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); 2295 2296 // First, check mark word for the biased lock pattern. 2297 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2298 2299 // Get fast path - mark word has the biased lock pattern. 2300 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, 2301 markOopDesc::biased_lock_mask_in_place, 2302 markOopDesc::biased_lock_pattern, true); 2303 // fast_lock_region->in(1) is set to slow path. 2304 fast_lock_mem_phi->init_req(1, mem); 2305 2306 // Now check that the lock is biased to the current thread and has 2307 // the same epoch and bias as Klass::_prototype_header. 2308 2309 // Special-case a fresh allocation to avoid building nodes: 2310 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); 2311 if (klass_node == NULL) { 2312 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 2313 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr())); 2314 #ifdef _LP64 2315 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) { 2316 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity"); 2317 klass_node->in(1)->init_req(0, ctrl); 2318 } else 2319 #endif 2320 klass_node->init_req(0, ctrl); 2321 } 2322 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type()); 2323 2324 Node* thread = transform_later(new ThreadLocalNode()); 2325 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2326 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node)); 2327 Node* x_node = transform_later(new XorXNode(o_node, mark_node)); 2328 2329 // Get slow path - mark word does NOT match the value. 2330 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, 2331 (~markOopDesc::age_mask_in_place), 0); 2332 // region->in(3) is set to fast path - the object is biased to the current thread. 2333 mem_phi->init_req(3, mem); 2334 2335 2336 // Mark word does NOT match the value (thread | Klass::_prototype_header). 2337 2338 2339 // First, check biased pattern. 2340 // Get fast path - _prototype_header has the same biased lock pattern. 2341 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, 2342 markOopDesc::biased_lock_mask_in_place, 0, true); 2343 2344 not_biased_ctrl = fast_lock_region->in(2); // Slow path 2345 // fast_lock_region->in(2) - the prototype header is no longer biased 2346 // and we have to revoke the bias on this object. 2347 // We are going to try to reset the mark of this object to the prototype 2348 // value and fall through to the CAS-based locking scheme. 2349 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 2350 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr, 2351 proto_node, mark_node); 2352 transform_later(cas); 2353 Node* proj = transform_later(new SCMemProjNode(cas)); 2354 fast_lock_mem_phi->init_req(2, proj); 2355 2356 2357 // Second, check epoch bits. 2358 Node* rebiased_region = new RegionNode(3); 2359 Node* old_phi = new PhiNode( rebiased_region, TypeX_X); 2360 Node* new_phi = new PhiNode( rebiased_region, TypeX_X); 2361 2362 // Get slow path - mark word does NOT match epoch bits. 2363 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, 2364 markOopDesc::epoch_mask_in_place, 0); 2365 // The epoch of the current bias is not valid, attempt to rebias the object 2366 // toward the current thread. 2367 rebiased_region->init_req(2, epoch_ctrl); 2368 old_phi->init_req(2, mark_node); 2369 new_phi->init_req(2, o_node); 2370 2371 // rebiased_region->in(1) is set to fast path. 2372 // The epoch of the current bias is still valid but we know 2373 // nothing about the owner; it might be set or it might be clear. 2374 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place | 2375 markOopDesc::age_mask_in_place | 2376 markOopDesc::epoch_mask_in_place); 2377 Node* old = transform_later(new AndXNode(mark_node, cmask)); 2378 cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2379 Node* new_mark = transform_later(new OrXNode(cast_thread, old)); 2380 old_phi->init_req(1, old); 2381 new_phi->init_req(1, new_mark); 2382 2383 transform_later(rebiased_region); 2384 transform_later(old_phi); 2385 transform_later(new_phi); 2386 2387 // Try to acquire the bias of the object using an atomic operation. 2388 // If this fails we will go in to the runtime to revoke the object's bias. 2389 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi); 2390 transform_later(cas); 2391 proj = transform_later(new SCMemProjNode(cas)); 2392 2393 // Get slow path - Failed to CAS. 2394 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); 2395 mem_phi->init_req(4, proj); 2396 // region->in(4) is set to fast path - the object is rebiased to the current thread. 2397 2398 // Failed to CAS. 2399 slow_path = new RegionNode(3); 2400 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); 2401 2402 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control 2403 slow_mem->init_req(1, proj); 2404 2405 // Call CAS-based locking scheme (FastLock node). 2406 2407 transform_later(fast_lock_region); 2408 transform_later(fast_lock_mem_phi); 2409 2410 // Get slow path - FastLock failed to lock the object. 2411 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); 2412 mem_phi->init_req(2, fast_lock_mem_phi); 2413 // region->in(2) is set to fast path - the object is locked to the current thread. 2414 2415 slow_path->init_req(2, ctrl); // Capture slow-control 2416 slow_mem->init_req(2, fast_lock_mem_phi); 2417 2418 transform_later(slow_path); 2419 transform_later(slow_mem); 2420 // Reset lock's memory edge. 2421 lock->set_req(TypeFunc::Memory, slow_mem); 2422 2423 } else { 2424 region = new RegionNode(3); 2425 // create a Phi for the memory state 2426 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2427 2428 // Optimize test; set region slot 2 2429 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); 2430 mem_phi->init_req(2, mem); 2431 } 2432 2433 const TypeOopPtr* objptr = _igvn.type(obj)->make_oopptr(); 2434 if (objptr->can_be_value_type()) { 2435 // Deoptimize and re-execute if a value 2436 assert(EnableValhalla, "should only be used if value types are enabled"); 2437 Node* mark = make_load(slow_path, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2438 Node* value_mask = _igvn.MakeConX(markOopDesc::always_locked_pattern); 2439 Node* is_value = _igvn.transform(new AndXNode(mark, value_mask)); 2440 Node* cmp = _igvn.transform(new CmpXNode(is_value, value_mask)); 2441 Node* bol = _igvn.transform(new BoolNode(cmp, BoolTest::eq)); 2442 Node* unc_ctrl = generate_slow_guard(&slow_path, bol, NULL); 2443 2444 int trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_class_check, Deoptimization::Action_none); 2445 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2446 const TypePtr* no_memory_effects = NULL; 2447 JVMState* jvms = lock->jvms(); 2448 CallNode* unc = new CallStaticJavaNode(OptoRuntime::uncommon_trap_Type(), call_addr, "uncommon_trap", 2449 jvms->bci(), no_memory_effects); 2450 2451 unc->init_req(TypeFunc::Control, unc_ctrl); 2452 unc->init_req(TypeFunc::I_O, lock->i_o()); 2453 unc->init_req(TypeFunc::Memory, mem); // may gc ptrs 2454 unc->init_req(TypeFunc::FramePtr, lock->in(TypeFunc::FramePtr)); 2455 unc->init_req(TypeFunc::ReturnAdr, lock->in(TypeFunc::ReturnAdr)); 2456 unc->init_req(TypeFunc::Parms+0, _igvn.intcon(trap_request)); 2457 unc->set_cnt(PROB_UNLIKELY_MAG(4)); 2458 copy_call_debug_info(lock, unc); 2459 2460 assert(unc->peek_monitor_box() == box, "wrong monitor"); 2461 assert(unc->peek_monitor_obj() == obj, "wrong monitor"); 2462 2463 // pop monitor and push obj back on stack: we trap before the monitorenter 2464 unc->pop_monitor(); 2465 unc->grow_stack(unc->jvms(), 1); 2466 unc->set_stack(unc->jvms(), unc->jvms()->stk_size()-1, obj); 2467 2468 _igvn.register_new_node_with_optimizer(unc); 2469 2470 Node* ctrl = _igvn.transform(new ProjNode(unc, TypeFunc::Control)); 2471 Node* halt = _igvn.transform(new HaltNode(ctrl, lock->in(TypeFunc::FramePtr))); 2472 C->root()->add_req(halt); 2473 } 2474 2475 // Make slow path call 2476 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), 2477 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, 2478 obj, box, NULL); 2479 2480 extract_call_projections(call); 2481 2482 // Slow path can only throw asynchronous exceptions, which are always 2483 // de-opted. So the compiler thinks the slow-call can never throw an 2484 // exception. If it DOES throw an exception we would need the debug 2485 // info removed first (since if it throws there is no monitor). 2486 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2487 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2488 2489 // Capture slow path 2490 // disconnect fall-through projection from call and create a new one 2491 // hook up users of fall-through projection to region 2492 Node *slow_ctrl = _fallthroughproj->clone(); 2493 transform_later(slow_ctrl); 2494 _igvn.hash_delete(_fallthroughproj); 2495 _fallthroughproj->disconnect_inputs(NULL, C); 2496 region->init_req(1, slow_ctrl); 2497 // region inputs are now complete 2498 transform_later(region); 2499 _igvn.replace_node(_fallthroughproj, region); 2500 2501 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory)); 2502 mem_phi->init_req(1, memproj ); 2503 transform_later(mem_phi); 2504 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2505 } 2506 2507 //------------------------------expand_unlock_node---------------------- 2508 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 2509 2510 Node* ctrl = unlock->in(TypeFunc::Control); 2511 Node* mem = unlock->in(TypeFunc::Memory); 2512 Node* obj = unlock->obj_node(); 2513 Node* box = unlock->box_node(); 2514 2515 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2516 2517 // No need for a null check on unlock 2518 2519 // Make the merge point 2520 Node *region; 2521 Node *mem_phi; 2522 2523 if (UseOptoBiasInlining) { 2524 // Check for biased locking unlock case, which is a no-op. 2525 // See the full description in MacroAssembler::biased_locking_exit(). 2526 region = new RegionNode(4); 2527 // create a Phi for the memory state 2528 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2529 mem_phi->init_req(3, mem); 2530 2531 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2532 ctrl = opt_bits_test(ctrl, region, 3, mark_node, 2533 markOopDesc::biased_lock_mask_in_place, 2534 markOopDesc::biased_lock_pattern); 2535 } else { 2536 region = new RegionNode(3); 2537 // create a Phi for the memory state 2538 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2539 } 2540 2541 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box ); 2542 funlock = transform_later( funlock )->as_FastUnlock(); 2543 // Optimize test; set region slot 2 2544 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); 2545 Node *thread = transform_later(new ThreadLocalNode()); 2546 2547 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), 2548 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), 2549 "complete_monitor_unlocking_C", slow_path, obj, box, thread); 2550 2551 extract_call_projections(call); 2552 2553 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2554 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2555 2556 // No exceptions for unlocking 2557 // Capture slow path 2558 // disconnect fall-through projection from call and create a new one 2559 // hook up users of fall-through projection to region 2560 Node *slow_ctrl = _fallthroughproj->clone(); 2561 transform_later(slow_ctrl); 2562 _igvn.hash_delete(_fallthroughproj); 2563 _fallthroughproj->disconnect_inputs(NULL, C); 2564 region->init_req(1, slow_ctrl); 2565 // region inputs are now complete 2566 transform_later(region); 2567 _igvn.replace_node(_fallthroughproj, region); 2568 2569 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) ); 2570 mem_phi->init_req(1, memproj ); 2571 mem_phi->init_req(2, mem); 2572 transform_later(mem_phi); 2573 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2574 } 2575 2576 // A value type might be returned from the call but we don't know its 2577 // type. Either we get a buffered value (and nothing needs to be done) 2578 // or one of the values being returned is the klass of the value type 2579 // and we need to allocate a value type instance of that type and 2580 // initialize it with other values being returned. In that case, we 2581 // first try a fast path allocation and initialize the value with the 2582 // value klass's pack handler or we fall back to a runtime call. 2583 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) { 2584 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call"); 2585 Node* ret = call->proj_out_or_null(TypeFunc::Parms); 2586 if (ret == NULL) { 2587 return; 2588 } 2589 const TypeFunc* tf = call->_tf; 2590 const TypeTuple* domain = OptoRuntime::store_value_type_fields_Type()->domain_cc(); 2591 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain); 2592 call->_tf = new_tf; 2593 // Make sure the change of type is applied before projections are processed by igvn 2594 _igvn.set_type(call, call->Value(&_igvn)); 2595 _igvn.set_type(ret, ret->Value(&_igvn)); 2596 2597 // Before any new projection is added: 2598 CallProjections* projs = call->extract_projections(true, true); 2599 2600 Node* ctl = new Node(1); 2601 Node* mem = new Node(1); 2602 Node* io = new Node(1); 2603 Node* ex_ctl = new Node(1); 2604 Node* ex_mem = new Node(1); 2605 Node* ex_io = new Node(1); 2606 Node* res = new Node(1); 2607 2608 Node* cast = transform_later(new CastP2XNode(ctl, res)); 2609 Node* mask = MakeConX(0x1); 2610 Node* masked = transform_later(new AndXNode(cast, mask)); 2611 Node* cmp = transform_later(new CmpXNode(masked, mask)); 2612 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq)); 2613 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN); 2614 transform_later(allocation_iff); 2615 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff)); 2616 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff)); 2617 2618 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM)); 2619 2620 Node* mask2 = MakeConX(-2); 2621 Node* masked2 = transform_later(new AndXNode(cast, mask2)); 2622 Node* rawklassptr = transform_later(new CastX2PNode(masked2)); 2623 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeKlassPtr::OBJECT_OR_NULL)); 2624 2625 Node* slowpath_bol = NULL; 2626 Node* top_adr = NULL; 2627 Node* old_top = NULL; 2628 Node* new_top = NULL; 2629 if (UseTLAB) { 2630 Node* end_adr = NULL; 2631 set_eden_pointers(top_adr, end_adr); 2632 Node* end = make_load(ctl, mem, end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 2633 old_top = new LoadPNode(ctl, mem, top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered); 2634 transform_later(old_top); 2635 Node* layout_val = make_load(NULL, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT); 2636 Node* size_in_bytes = ConvI2X(layout_val); 2637 new_top = new AddPNode(top(), old_top, size_in_bytes); 2638 transform_later(new_top); 2639 Node* slowpath_cmp = new CmpPNode(new_top, end); 2640 transform_later(slowpath_cmp); 2641 slowpath_bol = new BoolNode(slowpath_cmp, BoolTest::ge); 2642 transform_later(slowpath_bol); 2643 } else { 2644 slowpath_bol = intcon(1); 2645 top_adr = top(); 2646 old_top = top(); 2647 new_top = top(); 2648 } 2649 IfNode* slowpath_iff = new IfNode(allocation_ctl, slowpath_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); 2650 transform_later(slowpath_iff); 2651 2652 Node* slowpath_true = new IfTrueNode(slowpath_iff); 2653 transform_later(slowpath_true); 2654 2655 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_value_type_fields_Type(), 2656 StubRoutines::store_value_type_fields_to_buf(), 2657 "store_value_type_fields", 2658 call->jvms()->bci(), 2659 TypePtr::BOTTOM); 2660 slow_call->init_req(TypeFunc::Control, slowpath_true); 2661 slow_call->init_req(TypeFunc::Memory, mem); 2662 slow_call->init_req(TypeFunc::I_O, io); 2663 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr)); 2664 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr)); 2665 slow_call->init_req(TypeFunc::Parms, res); 2666 2667 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control)); 2668 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory)); 2669 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O)); 2670 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms)); 2671 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2)); 2672 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci)); 2673 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci)); 2674 2675 Node* ex_r = new RegionNode(3); 2676 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM); 2677 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO); 2678 ex_r->init_req(1, slow_excp); 2679 ex_mem_phi->init_req(1, slow_mem); 2680 ex_io_phi->init_req(1, slow_io); 2681 ex_r->init_req(2, ex_ctl); 2682 ex_mem_phi->init_req(2, ex_mem); 2683 ex_io_phi->init_req(2, ex_io); 2684 2685 transform_later(ex_r); 2686 transform_later(ex_mem_phi); 2687 transform_later(ex_io_phi); 2688 2689 Node* slowpath_false = new IfFalseNode(slowpath_iff); 2690 transform_later(slowpath_false); 2691 Node* rawmem = new StorePNode(slowpath_false, mem, top_adr, TypeRawPtr::BOTTOM, new_top, MemNode::unordered); 2692 transform_later(rawmem); 2693 Node* mark_node = makecon(TypeRawPtr::make((address)markOopDesc::always_locked_prototype())); 2694 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 2695 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 2696 if (UseCompressedClassPointers) { 2697 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT); 2698 } 2699 Node* fixed_block = make_load(slowpath_false, rawmem, klass_node, in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 2700 Node* pack_handler = make_load(slowpath_false, rawmem, fixed_block, in_bytes(ValueKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 2701 2702 CallLeafNoFPNode* handler_call = new CallLeafNoFPNode(OptoRuntime::pack_value_type_Type(), 2703 NULL, 2704 "pack handler", 2705 TypeRawPtr::BOTTOM); 2706 handler_call->init_req(TypeFunc::Control, slowpath_false); 2707 handler_call->init_req(TypeFunc::Memory, rawmem); 2708 handler_call->init_req(TypeFunc::I_O, top()); 2709 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr)); 2710 handler_call->init_req(TypeFunc::ReturnAdr, top()); 2711 handler_call->init_req(TypeFunc::Parms, pack_handler); 2712 handler_call->init_req(TypeFunc::Parms+1, old_top); 2713 2714 // We don't know how many values are returned. This assumes the 2715 // worst case, that all available registers are used. 2716 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) { 2717 if (domain->field_at(i) == Type::HALF) { 2718 slow_call->init_req(i, top()); 2719 handler_call->init_req(i+1, top()); 2720 continue; 2721 } 2722 Node* proj = transform_later(new ProjNode(call, i)); 2723 slow_call->init_req(i, proj); 2724 handler_call->init_req(i+1, proj); 2725 } 2726 2727 // We can safepoint at that new call 2728 copy_call_debug_info(call, slow_call); 2729 transform_later(slow_call); 2730 transform_later(handler_call); 2731 2732 Node* handler_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control)); 2733 rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory)); 2734 Node* slowpath_false_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms)); 2735 2736 MergeMemNode* slowpath_false_mem = MergeMemNode::make(mem); 2737 slowpath_false_mem->set_memory_at(Compile::AliasIdxRaw, rawmem); 2738 transform_later(slowpath_false_mem); 2739 2740 Node* r = new RegionNode(4); 2741 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM); 2742 Node* io_phi = new PhiNode(r, Type::ABIO); 2743 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM); 2744 2745 r->init_req(1, no_allocation_ctl); 2746 mem_phi->init_req(1, mem); 2747 io_phi->init_req(1, io); 2748 res_phi->init_req(1, no_allocation_res); 2749 r->init_req(2, slow_norm); 2750 mem_phi->init_req(2, slow_mem); 2751 io_phi->init_req(2, slow_io); 2752 res_phi->init_req(2, slow_res); 2753 r->init_req(3, handler_ctl); 2754 mem_phi->init_req(3, slowpath_false_mem); 2755 io_phi->init_req(3, io); 2756 res_phi->init_req(3, slowpath_false_res); 2757 2758 transform_later(r); 2759 transform_later(mem_phi); 2760 transform_later(io_phi); 2761 transform_later(res_phi); 2762 2763 assert(projs->nb_resproj == 1, "unexpected number of results"); 2764 _igvn.replace_in_uses(projs->fallthrough_catchproj, r); 2765 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi); 2766 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi); 2767 _igvn.replace_in_uses(projs->resproj[0], res_phi); 2768 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r); 2769 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi); 2770 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi); 2771 2772 _igvn.replace_node(ctl, projs->fallthrough_catchproj); 2773 _igvn.replace_node(mem, projs->fallthrough_memproj); 2774 _igvn.replace_node(io, projs->fallthrough_ioproj); 2775 _igvn.replace_node(res, projs->resproj[0]); 2776 _igvn.replace_node(ex_ctl, projs->catchall_catchproj); 2777 _igvn.replace_node(ex_mem, projs->catchall_memproj); 2778 _igvn.replace_node(ex_io, projs->catchall_ioproj); 2779 } 2780 2781 //---------------------------eliminate_macro_nodes---------------------- 2782 // Eliminate scalar replaced allocations and associated locks. 2783 void PhaseMacroExpand::eliminate_macro_nodes() { 2784 if (C->macro_count() == 0) 2785 return; 2786 2787 // First, attempt to eliminate locks 2788 int cnt = C->macro_count(); 2789 for (int i=0; i < cnt; i++) { 2790 Node *n = C->macro_node(i); 2791 if (n->is_AbstractLock()) { // Lock and Unlock nodes 2792 // Before elimination mark all associated (same box and obj) 2793 // lock and unlock nodes. 2794 mark_eliminated_locking_nodes(n->as_AbstractLock()); 2795 } 2796 } 2797 bool progress = true; 2798 while (progress) { 2799 progress = false; 2800 for (int i = C->macro_count(); i > 0; i--) { 2801 Node * n = C->macro_node(i-1); 2802 bool success = false; 2803 debug_only(int old_macro_count = C->macro_count();); 2804 if (n->is_AbstractLock()) { 2805 success = eliminate_locking_node(n->as_AbstractLock()); 2806 } 2807 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2808 progress = progress || success; 2809 } 2810 } 2811 // Next, attempt to eliminate allocations 2812 _has_locks = false; 2813 progress = true; 2814 while (progress) { 2815 progress = false; 2816 for (int i = C->macro_count(); i > 0; i--) { 2817 Node * n = C->macro_node(i-1); 2818 bool success = false; 2819 debug_only(int old_macro_count = C->macro_count();); 2820 switch (n->class_id()) { 2821 case Node::Class_Allocate: 2822 case Node::Class_AllocateArray: 2823 success = eliminate_allocate_node(n->as_Allocate()); 2824 break; 2825 case Node::Class_CallStaticJava: { 2826 CallStaticJavaNode* call = n->as_CallStaticJava(); 2827 if (!call->method()->is_method_handle_intrinsic()) { 2828 success = eliminate_boxing_node(n->as_CallStaticJava()); 2829 } 2830 break; 2831 } 2832 case Node::Class_Lock: 2833 case Node::Class_Unlock: 2834 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 2835 _has_locks = true; 2836 break; 2837 case Node::Class_ArrayCopy: 2838 break; 2839 case Node::Class_OuterStripMinedLoop: 2840 break; 2841 default: 2842 assert(n->Opcode() == Op_LoopLimit || 2843 n->Opcode() == Op_Opaque1 || 2844 n->Opcode() == Op_Opaque2 || 2845 n->Opcode() == Op_Opaque3 || 2846 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n), 2847 "unknown node type in macro list"); 2848 } 2849 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2850 progress = progress || success; 2851 } 2852 } 2853 } 2854 2855 //------------------------------expand_macro_nodes---------------------- 2856 // Returns true if a failure occurred. 2857 bool PhaseMacroExpand::expand_macro_nodes() { 2858 // Last attempt to eliminate macro nodes. 2859 eliminate_macro_nodes(); 2860 2861 // Make sure expansion will not cause node limit to be exceeded. 2862 // Worst case is a macro node gets expanded into about 200 nodes. 2863 // Allow 50% more for optimization. 2864 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) ) 2865 return true; 2866 2867 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. 2868 bool progress = true; 2869 while (progress) { 2870 progress = false; 2871 for (int i = C->macro_count(); i > 0; i--) { 2872 Node * n = C->macro_node(i-1); 2873 bool success = false; 2874 debug_only(int old_macro_count = C->macro_count();); 2875 if (n->Opcode() == Op_LoopLimit) { 2876 // Remove it from macro list and put on IGVN worklist to optimize. 2877 C->remove_macro_node(n); 2878 _igvn._worklist.push(n); 2879 success = true; 2880 } else if (n->Opcode() == Op_CallStaticJava) { 2881 CallStaticJavaNode* call = n->as_CallStaticJava(); 2882 if (!call->method()->is_method_handle_intrinsic()) { 2883 // Remove it from macro list and put on IGVN worklist to optimize. 2884 C->remove_macro_node(n); 2885 _igvn._worklist.push(n); 2886 success = true; 2887 } 2888 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { 2889 _igvn.replace_node(n, n->in(1)); 2890 success = true; 2891 #if INCLUDE_RTM_OPT 2892 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) { 2893 assert(C->profile_rtm(), "should be used only in rtm deoptimization code"); 2894 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), ""); 2895 Node* cmp = n->unique_out(); 2896 #ifdef ASSERT 2897 // Validate graph. 2898 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), ""); 2899 BoolNode* bol = cmp->unique_out()->as_Bool(); 2900 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() && 2901 (bol->_test._test == BoolTest::ne), ""); 2902 IfNode* ifn = bol->unique_out()->as_If(); 2903 assert((ifn->outcnt() == 2) && 2904 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, ""); 2905 #endif 2906 Node* repl = n->in(1); 2907 if (!_has_locks) { 2908 // Remove RTM state check if there are no locks in the code. 2909 // Replace input to compare the same value. 2910 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1); 2911 } 2912 _igvn.replace_node(n, repl); 2913 success = true; 2914 #endif 2915 } else if (n->Opcode() == Op_OuterStripMinedLoop) { 2916 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn); 2917 C->remove_macro_node(n); 2918 success = true; 2919 } 2920 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2921 progress = progress || success; 2922 } 2923 } 2924 2925 // expand arraycopy "macro" nodes first 2926 // For ReduceBulkZeroing, we must first process all arraycopy nodes 2927 // before the allocate nodes are expanded. 2928 int macro_idx = C->macro_count() - 1; 2929 while (macro_idx >= 0) { 2930 Node * n = C->macro_node(macro_idx); 2931 assert(n->is_macro(), "only macro nodes expected here"); 2932 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) { 2933 // node is unreachable, so don't try to expand it 2934 C->remove_macro_node(n); 2935 } else if (n->is_ArrayCopy()){ 2936 int macro_count = C->macro_count(); 2937 expand_arraycopy_node(n->as_ArrayCopy()); 2938 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2939 } 2940 if (C->failing()) return true; 2941 macro_idx --; 2942 } 2943 2944 // expand "macro" nodes 2945 // nodes are removed from the macro list as they are processed 2946 while (C->macro_count() > 0) { 2947 int macro_count = C->macro_count(); 2948 Node * n = C->macro_node(macro_count-1); 2949 assert(n->is_macro(), "only macro nodes expected here"); 2950 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) { 2951 // node is unreachable, so don't try to expand it 2952 C->remove_macro_node(n); 2953 continue; 2954 } 2955 switch (n->class_id()) { 2956 case Node::Class_Allocate: 2957 expand_allocate(n->as_Allocate()); 2958 break; 2959 case Node::Class_AllocateArray: 2960 expand_allocate_array(n->as_AllocateArray()); 2961 break; 2962 case Node::Class_Lock: 2963 expand_lock_node(n->as_Lock()); 2964 break; 2965 case Node::Class_Unlock: 2966 expand_unlock_node(n->as_Unlock()); 2967 break; 2968 case Node::Class_CallStaticJava: 2969 expand_mh_intrinsic_return(n->as_CallStaticJava()); 2970 C->remove_macro_node(n); 2971 break; 2972 default: 2973 assert(false, "unknown node type in macro list"); 2974 } 2975 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2976 if (C->failing()) return true; 2977 } 2978 2979 _igvn.set_delay_transform(false); 2980 _igvn.optimize(); 2981 if (C->failing()) return true; 2982 return false; 2983 }