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