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 "opto/valuetypenode.hpp" 50 #include "runtime/sharedRuntime.hpp" 51 #if INCLUDE_G1GC 52 #include "gc/g1/g1ThreadLocalData.hpp" 53 #endif // INCLUDE_G1GC 54 55 56 // 57 // Replace any references to "oldref" in inputs to "use" with "newref". 58 // Returns the number of replacements made. 59 // 60 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { 61 int nreplacements = 0; 62 uint req = use->req(); 63 for (uint j = 0; j < use->len(); j++) { 64 Node *uin = use->in(j); 65 if (uin == oldref) { 66 if (j < req) 67 use->set_req(j, newref); 68 else 69 use->set_prec(j, newref); 70 nreplacements++; 71 } else if (j >= req && uin == NULL) { 72 break; 73 } 74 } 75 return nreplacements; 76 } 77 78 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) { 79 // Copy debug information and adjust JVMState information 80 uint old_dbg_start = oldcall->tf()->domain_sig()->cnt(); 81 uint new_dbg_start = newcall->tf()->domain_sig()->cnt(); 82 int jvms_adj = new_dbg_start - old_dbg_start; 83 assert (new_dbg_start == newcall->req(), "argument count mismatch"); 84 85 // SafePointScalarObject node could be referenced several times in debug info. 86 // Use Dict to record cloned nodes. 87 Dict* sosn_map = new Dict(cmpkey,hashkey); 88 for (uint i = old_dbg_start; i < oldcall->req(); i++) { 89 Node* old_in = oldcall->in(i); 90 // Clone old SafePointScalarObjectNodes, adjusting their field contents. 91 if (old_in != NULL && old_in->is_SafePointScalarObject()) { 92 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject(); 93 uint old_unique = C->unique(); 94 Node* new_in = old_sosn->clone(sosn_map); 95 if (old_unique != C->unique()) { // New node? 96 new_in->set_req(0, C->root()); // reset control edge 97 new_in = transform_later(new_in); // Register new node. 98 } 99 old_in = new_in; 100 } 101 newcall->add_req(old_in); 102 } 103 104 // JVMS may be shared so clone it before we modify it 105 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL); 106 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) { 107 jvms->set_map(newcall); 108 jvms->set_locoff(jvms->locoff()+jvms_adj); 109 jvms->set_stkoff(jvms->stkoff()+jvms_adj); 110 jvms->set_monoff(jvms->monoff()+jvms_adj); 111 jvms->set_scloff(jvms->scloff()+jvms_adj); 112 jvms->set_endoff(jvms->endoff()+jvms_adj); 113 } 114 } 115 116 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) { 117 Node* cmp; 118 if (mask != 0) { 119 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask))); 120 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits))); 121 } else { 122 cmp = word; 123 } 124 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne)); 125 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); 126 transform_later(iff); 127 128 // Fast path taken. 129 Node *fast_taken = transform_later(new IfFalseNode(iff)); 130 131 // Fast path not-taken, i.e. slow path 132 Node *slow_taken = transform_later(new IfTrueNode(iff)); 133 134 if (return_fast_path) { 135 region->init_req(edge, slow_taken); // Capture slow-control 136 return fast_taken; 137 } else { 138 region->init_req(edge, fast_taken); // Capture fast-control 139 return slow_taken; 140 } 141 } 142 143 //--------------------copy_predefined_input_for_runtime_call-------------------- 144 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { 145 // Set fixed predefined input arguments 146 call->init_req( TypeFunc::Control, ctrl ); 147 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); 148 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? 149 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); 150 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); 151 } 152 153 //------------------------------make_slow_call--------------------------------- 154 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, 155 address slow_call, const char* leaf_name, Node* slow_path, 156 Node* parm0, Node* parm1, Node* parm2) { 157 158 // Slow-path call 159 CallNode *call = leaf_name 160 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) 161 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM ); 162 163 // Slow path call has no side-effects, uses few values 164 copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); 165 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0); 166 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1); 167 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2); 168 copy_call_debug_info(oldcall, call); 169 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 170 _igvn.replace_node(oldcall, call); 171 transform_later(call); 172 173 return call; 174 } 175 176 void PhaseMacroExpand::extract_call_projections(CallNode *call) { 177 _fallthroughproj = NULL; 178 _fallthroughcatchproj = NULL; 179 _ioproj_fallthrough = NULL; 180 _ioproj_catchall = NULL; 181 _catchallcatchproj = NULL; 182 _memproj_fallthrough = NULL; 183 _memproj_catchall = NULL; 184 _resproj = NULL; 185 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) { 186 ProjNode *pn = call->fast_out(i)->as_Proj(); 187 switch (pn->_con) { 188 case TypeFunc::Control: 189 { 190 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj 191 _fallthroughproj = pn; 192 DUIterator_Fast jmax, j = pn->fast_outs(jmax); 193 const Node *cn = pn->fast_out(j); 194 if (cn->is_Catch()) { 195 ProjNode *cpn = NULL; 196 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { 197 cpn = cn->fast_out(k)->as_Proj(); 198 assert(cpn->is_CatchProj(), "must be a CatchProjNode"); 199 if (cpn->_con == CatchProjNode::fall_through_index) 200 _fallthroughcatchproj = cpn; 201 else { 202 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); 203 _catchallcatchproj = cpn; 204 } 205 } 206 } 207 break; 208 } 209 case TypeFunc::I_O: 210 if (pn->_is_io_use) 211 _ioproj_catchall = pn; 212 else 213 _ioproj_fallthrough = pn; 214 break; 215 case TypeFunc::Memory: 216 if (pn->_is_io_use) 217 _memproj_catchall = pn; 218 else 219 _memproj_fallthrough = pn; 220 break; 221 case TypeFunc::Parms: 222 _resproj = pn; 223 break; 224 default: 225 assert(false, "unexpected projection from allocation node."); 226 } 227 } 228 229 } 230 231 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) { 232 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2(); 233 bs->eliminate_gc_barrier(this, p2x); 234 } 235 236 // Search for a memory operation for the specified memory slice. 237 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) { 238 Node *orig_mem = mem; 239 Node *alloc_mem = alloc->in(TypeFunc::Memory); 240 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr(); 241 while (true) { 242 if (mem == alloc_mem || mem == start_mem ) { 243 return mem; // hit one of our sentinels 244 } else if (mem->is_MergeMem()) { 245 mem = mem->as_MergeMem()->memory_at(alias_idx); 246 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { 247 Node *in = mem->in(0); 248 // we can safely skip over safepoints, calls, locks and membars because we 249 // already know that the object is safe to eliminate. 250 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { 251 return in; 252 } else if (in->is_Call()) { 253 CallNode *call = in->as_Call(); 254 if (call->may_modify(tinst, phase)) { 255 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape"); 256 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) { 257 return in; 258 } 259 } 260 mem = in->in(TypeFunc::Memory); 261 } else if (in->is_MemBar()) { 262 ArrayCopyNode* ac = NULL; 263 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) { 264 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone"); 265 return ac; 266 } 267 mem = in->in(TypeFunc::Memory); 268 } else { 269 assert(false, "unexpected projection"); 270 } 271 } else if (mem->is_Store()) { 272 const TypePtr* atype = mem->as_Store()->adr_type(); 273 int adr_idx = phase->C->get_alias_index(atype); 274 if (adr_idx == alias_idx) { 275 assert(atype->isa_oopptr(), "address type must be oopptr"); 276 int adr_offset = atype->flattened_offset(); 277 uint adr_iid = atype->is_oopptr()->instance_id(); 278 // Array elements references have the same alias_idx 279 // but different offset and different instance_id. 280 if (adr_offset == offset && adr_iid == alloc->_idx) 281 return mem; 282 } else { 283 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw"); 284 } 285 mem = mem->in(MemNode::Memory); 286 } else if (mem->is_ClearArray()) { 287 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) { 288 // Can not bypass initialization of the instance 289 // we are looking. 290 debug_only(intptr_t offset;) 291 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity"); 292 InitializeNode* init = alloc->as_Allocate()->initialization(); 293 // We are looking for stored value, return Initialize node 294 // or memory edge from Allocate node. 295 if (init != NULL) 296 return init; 297 else 298 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers). 299 } 300 // Otherwise skip it (the call updated 'mem' value). 301 } else if (mem->Opcode() == Op_SCMemProj) { 302 mem = mem->in(0); 303 Node* adr = NULL; 304 if (mem->is_LoadStore()) { 305 adr = mem->in(MemNode::Address); 306 } else { 307 assert(mem->Opcode() == Op_EncodeISOArray || 308 mem->Opcode() == Op_StrCompressedCopy, "sanity"); 309 adr = mem->in(3); // Destination array 310 } 311 const TypePtr* atype = adr->bottom_type()->is_ptr(); 312 int adr_idx = phase->C->get_alias_index(atype); 313 if (adr_idx == alias_idx) { 314 DEBUG_ONLY(mem->dump();) 315 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 316 return NULL; 317 } 318 mem = mem->in(MemNode::Memory); 319 } else if (mem->Opcode() == Op_StrInflatedCopy) { 320 Node* adr = mem->in(3); // Destination array 321 const TypePtr* atype = adr->bottom_type()->is_ptr(); 322 int adr_idx = phase->C->get_alias_index(atype); 323 if (adr_idx == alias_idx) { 324 DEBUG_ONLY(mem->dump();) 325 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field"); 326 return NULL; 327 } 328 mem = mem->in(MemNode::Memory); 329 } else { 330 return mem; 331 } 332 assert(mem != orig_mem, "dead memory loop"); 333 } 334 } 335 336 // Generate loads from source of the arraycopy for fields of 337 // destination needed at a deoptimization point 338 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) { 339 BasicType bt = ft; 340 const Type *type = ftype; 341 if (ft == T_NARROWOOP) { 342 bt = T_OBJECT; 343 type = ftype->make_oopptr(); 344 } 345 Node* res = NULL; 346 if (ac->is_clonebasic()) { 347 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base); 348 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset))); 349 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset); 350 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned); 351 } else { 352 if (ac->modifies(offset, offset, &_igvn, true)) { 353 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result"); 354 uint shift = exact_log2(type2aelembytes(bt)); 355 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 356 #ifdef _LP64 357 diff = _igvn.transform(new ConvI2LNode(diff)); 358 #endif 359 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift))); 360 361 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff)); 362 Node* base = ac->in(ArrayCopyNode::Src); 363 Node* adr = _igvn.transform(new AddPNode(base, base, off)); 364 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset); 365 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned); 366 } 367 } 368 if (res != NULL) { 369 res = _igvn.transform(res); 370 if (ftype->isa_narrowoop()) { 371 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes 372 res = _igvn.transform(new EncodePNode(res, ftype)); 373 } 374 return res; 375 } 376 return NULL; 377 } 378 379 // 380 // Given a Memory Phi, compute a value Phi containing the values from stores 381 // on the input paths. 382 // Note: this function is recursive, its depth is limited by the "level" argument 383 // Returns the computed Phi, or NULL if it cannot compute it. 384 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) { 385 assert(mem->is_Phi(), "sanity"); 386 int alias_idx = C->get_alias_index(adr_t); 387 int offset = adr_t->flattened_offset(); 388 int instance_id = adr_t->instance_id(); 389 390 // Check if an appropriate value phi already exists. 391 Node* region = mem->in(0); 392 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { 393 Node* phi = region->fast_out(k); 394 if (phi->is_Phi() && phi != mem && 395 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) { 396 return phi; 397 } 398 } 399 // Check if an appropriate new value phi already exists. 400 Node* new_phi = value_phis->find(mem->_idx); 401 if (new_phi != NULL) 402 return new_phi; 403 404 if (level <= 0) { 405 return NULL; // Give up: phi tree too deep 406 } 407 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 408 Node *alloc_mem = alloc->in(TypeFunc::Memory); 409 410 uint length = mem->req(); 411 GrowableArray <Node *> values(length, length, NULL, false); 412 413 // create a new Phi for the value 414 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset); 415 transform_later(phi); 416 value_phis->push(phi, mem->_idx); 417 418 for (uint j = 1; j < length; j++) { 419 Node *in = mem->in(j); 420 if (in == NULL || in->is_top()) { 421 values.at_put(j, in); 422 } else { 423 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); 424 if (val == start_mem || val == alloc_mem) { 425 // hit a sentinel, return appropriate 0 value 426 values.at_put(j, _igvn.zerocon(ft)); 427 continue; 428 } 429 if (val->is_Initialize()) { 430 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 431 } 432 if (val == NULL) { 433 return NULL; // can't find a value on this path 434 } 435 if (val == mem) { 436 values.at_put(j, mem); 437 } else if (val->is_Store()) { 438 values.at_put(j, val->in(MemNode::ValueIn)); 439 } else if(val->is_Proj() && val->in(0) == alloc) { 440 values.at_put(j, _igvn.zerocon(ft)); 441 } else if (val->is_Phi()) { 442 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); 443 if (val == NULL) { 444 return NULL; 445 } 446 values.at_put(j, val); 447 } else if (val->Opcode() == Op_SCMemProj) { 448 assert(val->in(0)->is_LoadStore() || 449 val->in(0)->Opcode() == Op_EncodeISOArray || 450 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity"); 451 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 452 return NULL; 453 } else if (val->is_ArrayCopy()) { 454 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc); 455 if (res == NULL) { 456 return NULL; 457 } 458 values.at_put(j, res); 459 } else { 460 #ifdef ASSERT 461 val->dump(); 462 assert(false, "unknown node on this path"); 463 #endif 464 return NULL; // unknown node on this path 465 } 466 } 467 } 468 // Set Phi's inputs 469 for (uint j = 1; j < length; j++) { 470 if (values.at(j) == mem) { 471 phi->init_req(j, phi); 472 } else { 473 phi->init_req(j, values.at(j)); 474 } 475 } 476 return phi; 477 } 478 479 // Search the last value stored into the object's field. 480 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) { 481 assert(adr_t->is_known_instance_field(), "instance required"); 482 int instance_id = adr_t->instance_id(); 483 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 484 485 int alias_idx = C->get_alias_index(adr_t); 486 int offset = adr_t->flattened_offset(); 487 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 488 Node *alloc_ctrl = alloc->in(TypeFunc::Control); 489 Node *alloc_mem = alloc->in(TypeFunc::Memory); 490 Arena *a = Thread::current()->resource_area(); 491 VectorSet visited(a); 492 493 bool done = sfpt_mem == alloc_mem; 494 Node *mem = sfpt_mem; 495 while (!done) { 496 if (visited.test_set(mem->_idx)) { 497 return NULL; // found a loop, give up 498 } 499 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 500 if (mem == start_mem || mem == alloc_mem) { 501 done = true; // hit a sentinel, return appropriate 0 value 502 } else if (mem->is_Initialize()) { 503 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 504 if (mem == NULL) { 505 done = true; // Something went wrong. 506 } else if (mem->is_Store()) { 507 const TypePtr* atype = mem->as_Store()->adr_type(); 508 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 509 done = true; 510 } 511 } else if (mem->is_Store()) { 512 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 513 assert(atype != NULL, "address type must be oopptr"); 514 assert(C->get_alias_index(atype) == alias_idx && 515 atype->is_known_instance_field() && atype->flattened_offset() == offset && 516 atype->instance_id() == instance_id, "store is correct memory slice"); 517 done = true; 518 } else if (mem->is_Phi()) { 519 // try to find a phi's unique input 520 Node *unique_input = NULL; 521 Node *top = C->top(); 522 for (uint i = 1; i < mem->req(); i++) { 523 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 524 if (n == NULL || n == top || n == mem) { 525 continue; 526 } else if (unique_input == NULL) { 527 unique_input = n; 528 } else if (unique_input != n) { 529 unique_input = top; 530 break; 531 } 532 } 533 if (unique_input != NULL && unique_input != top) { 534 mem = unique_input; 535 } else { 536 done = true; 537 } 538 } else if (mem->is_ArrayCopy()) { 539 done = true; 540 } else { 541 assert(false, "unexpected node"); 542 } 543 } 544 if (mem != NULL) { 545 if (mem == start_mem || mem == alloc_mem) { 546 // hit a sentinel, return appropriate 0 value 547 return _igvn.zerocon(ft); 548 } else if (mem->is_Store()) { 549 return mem->in(MemNode::ValueIn); 550 } else if (mem->is_Phi()) { 551 // attempt to produce a Phi reflecting the values on the input paths of the Phi 552 Node_Stack value_phis(a, 8); 553 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 554 if (phi != NULL) { 555 return phi; 556 } else { 557 // Kill all new Phis 558 while(value_phis.is_nonempty()) { 559 Node* n = value_phis.node(); 560 _igvn.replace_node(n, C->top()); 561 value_phis.pop(); 562 } 563 } 564 } else if (mem->is_ArrayCopy()) { 565 Node* ctl = mem->in(0); 566 Node* m = mem->in(TypeFunc::Memory); 567 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) { 568 // pin the loads in the uncommon trap path 569 ctl = sfpt_ctl; 570 m = sfpt_mem; 571 } 572 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc); 573 } 574 } 575 // Something went wrong. 576 return NULL; 577 } 578 579 // Search the last value stored into the value type's fields. 580 Node* PhaseMacroExpand::value_type_from_mem(Node* mem, Node* ctl, ciValueKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) { 581 // Subtract the offset of the first field to account for the missing oop header 582 offset -= vk->first_field_offset(); 583 // Create a new ValueTypeNode and retrieve the field values from memory 584 ValueTypeNode* vt = ValueTypeNode::make_uninitialized(_igvn, vk)->as_ValueType(); 585 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) { 586 ciType* field_type = vt->field_type(i); 587 int field_offset = offset + vt->field_offset(i); 588 // Each value type field has its own memory slice 589 adr_type = adr_type->with_field_offset(field_offset); 590 Node* value = NULL; 591 if (vt->field_is_flattened(i)) { 592 value = value_type_from_mem(mem, ctl, field_type->as_value_klass(), adr_type, field_offset, alloc); 593 } else { 594 const Type* ft = Type::get_const_type(field_type); 595 BasicType bt = field_type->basic_type(); 596 if (UseCompressedOops && !is_java_primitive(bt)) { 597 ft = ft->make_narrowoop(); 598 bt = T_NARROWOOP; 599 } 600 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc); 601 if (value != NULL && ft->isa_narrowoop()) { 602 assert(UseCompressedOops, "unexpected narrow oop"); 603 value = transform_later(new DecodeNNode(value, value->get_ptr_type())); 604 } 605 } 606 if (value != NULL) { 607 vt->set_field_value(i, value); 608 } else { 609 // We might have reached the TrackedInitializationLimit 610 return NULL; 611 } 612 } 613 return vt; 614 } 615 616 // Check the possibility of scalar replacement. 617 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 618 // Scan the uses of the allocation to check for anything that would 619 // prevent us from eliminating it. 620 NOT_PRODUCT( const char* fail_eliminate = NULL; ) 621 DEBUG_ONLY( Node* disq_node = NULL; ) 622 bool can_eliminate = true; 623 624 Node* res = alloc->result_cast(); 625 const TypeOopPtr* res_type = NULL; 626 if (res == NULL) { 627 // All users were eliminated. 628 } else if (!res->is_CheckCastPP()) { 629 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 630 can_eliminate = false; 631 } else { 632 res_type = _igvn.type(res)->isa_oopptr(); 633 if (res_type == NULL) { 634 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 635 can_eliminate = false; 636 } else if (res_type->isa_aryptr()) { 637 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 638 if (length < 0) { 639 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 640 can_eliminate = false; 641 } 642 } 643 } 644 645 if (can_eliminate && res != NULL) { 646 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 647 j < jmax && can_eliminate; j++) { 648 Node* use = res->fast_out(j); 649 650 if (use->is_AddP()) { 651 const TypePtr* addp_type = _igvn.type(use)->is_ptr(); 652 int offset = addp_type->offset(); 653 654 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 655 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) 656 can_eliminate = false; 657 break; 658 } 659 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 660 k < kmax && can_eliminate; k++) { 661 Node* n = use->fast_out(k); 662 if (!n->is_Store() && n->Opcode() != Op_CastP2X && 663 !(n->is_ArrayCopy() && 664 n->as_ArrayCopy()->is_clonebasic() && 665 n->in(ArrayCopyNode::Dest) == use)) { 666 DEBUG_ONLY(disq_node = n;) 667 if (n->is_Load() || n->is_LoadStore()) { 668 NOT_PRODUCT(fail_eliminate = "Field load";) 669 } else { 670 NOT_PRODUCT(fail_eliminate = "Not store field reference";) 671 } 672 can_eliminate = false; 673 } 674 } 675 } else if (use->is_ArrayCopy() && 676 (use->as_ArrayCopy()->is_arraycopy_validated() || 677 use->as_ArrayCopy()->is_copyof_validated() || 678 use->as_ArrayCopy()->is_copyofrange_validated()) && 679 use->in(ArrayCopyNode::Dest) == res) { 680 // ok to eliminate 681 } else if (use->is_SafePoint()) { 682 SafePointNode* sfpt = use->as_SafePoint(); 683 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 684 // Object is passed as argument. 685 DEBUG_ONLY(disq_node = use;) 686 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 687 can_eliminate = false; 688 } 689 Node* sfptMem = sfpt->memory(); 690 if (sfptMem == NULL || sfptMem->is_top()) { 691 DEBUG_ONLY(disq_node = use;) 692 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) 693 can_eliminate = false; 694 } else { 695 safepoints.append_if_missing(sfpt); 696 } 697 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 698 if (use->is_Phi()) { 699 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 700 NOT_PRODUCT(fail_eliminate = "Object is return value";) 701 } else { 702 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 703 } 704 DEBUG_ONLY(disq_node = use;) 705 } else { 706 if (use->Opcode() == Op_Return) { 707 NOT_PRODUCT(fail_eliminate = "Object is return value";) 708 } else { 709 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 710 } 711 DEBUG_ONLY(disq_node = use;) 712 } 713 can_eliminate = false; 714 } 715 } 716 } 717 718 #ifndef PRODUCT 719 if (PrintEliminateAllocations) { 720 if (can_eliminate) { 721 tty->print("Scalar "); 722 if (res == NULL) 723 alloc->dump(); 724 else 725 res->dump(); 726 } else if (alloc->_is_scalar_replaceable) { 727 tty->print("NotScalar (%s)", fail_eliminate); 728 if (res == NULL) 729 alloc->dump(); 730 else 731 res->dump(); 732 #ifdef ASSERT 733 if (disq_node != NULL) { 734 tty->print(" >>>> "); 735 disq_node->dump(); 736 } 737 #endif /*ASSERT*/ 738 } 739 } 740 #endif 741 return can_eliminate; 742 } 743 744 // Do scalar replacement. 745 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 746 GrowableArray <SafePointNode *> safepoints_done; 747 748 ciKlass* klass = NULL; 749 ciInstanceKlass* iklass = NULL; 750 int nfields = 0; 751 int array_base = 0; 752 int element_size = 0; 753 BasicType basic_elem_type = T_ILLEGAL; 754 ciType* elem_type = NULL; 755 756 Node* res = alloc->result_cast(); 757 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result"); 758 const TypeOopPtr* res_type = NULL; 759 if (res != NULL) { // Could be NULL when there are no users 760 res_type = _igvn.type(res)->isa_oopptr(); 761 } 762 763 if (res != NULL) { 764 klass = res_type->klass(); 765 if (res_type->isa_instptr()) { 766 // find the fields of the class which will be needed for safepoint debug information 767 assert(klass->is_instance_klass(), "must be an instance klass."); 768 iklass = klass->as_instance_klass(); 769 nfields = iklass->nof_nonstatic_fields(); 770 } else { 771 // find the array's elements which will be needed for safepoint debug information 772 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 773 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); 774 elem_type = klass->as_array_klass()->element_type(); 775 basic_elem_type = elem_type->basic_type(); 776 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 777 element_size = type2aelembytes(basic_elem_type); 778 if (klass->is_value_array_klass()) { 779 // Flattened value type array 780 element_size = klass->as_value_array_klass()->element_byte_size(); 781 } 782 } 783 } 784 // 785 // Process the safepoint uses 786 // 787 Unique_Node_List value_worklist; 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 assert(!field->is_flattened(), "flattened value type fields should not have safepoint uses"); 815 } else { 816 offset = array_base + j * (intptr_t)element_size; 817 } 818 819 const Type *field_type; 820 // The next code is taken from Parse::do_get_xxx(). 821 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { 822 if (!elem_type->is_loaded()) { 823 field_type = TypeInstPtr::BOTTOM; 824 } else if (field != NULL && field->is_static_constant()) { 825 // This can happen if the constant oop is non-perm. 826 ciObject* con = field->constant_value().as_object(); 827 // Do not "join" in the previous type; it doesn't add value, 828 // and may yield a vacuous result if the field is of interface type. 829 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 830 assert(field_type != NULL, "field singleton type must be consistent"); 831 } else { 832 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 833 } 834 if (UseCompressedOops) { 835 field_type = field_type->make_narrowoop(); 836 basic_elem_type = T_NARROWOOP; 837 } 838 } else { 839 field_type = Type::get_const_basic_type(basic_elem_type); 840 } 841 842 Node* field_val = NULL; 843 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 844 if (klass->is_value_array_klass()) { 845 ciValueKlass* vk = elem_type->as_value_klass(); 846 assert(vk->flatten_array(), "must be flattened"); 847 field_val = value_type_from_mem(mem, ctl, vk, field_addr_type->isa_aryptr(), 0, alloc); 848 } else { 849 field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc); 850 } 851 if (field_val == NULL) { 852 // We weren't able to find a value for this field, 853 // give up on eliminating this allocation. 854 855 // Remove any extra entries we added to the safepoint. 856 uint last = sfpt->req() - 1; 857 for (int k = 0; k < j; k++) { 858 sfpt->del_req(last--); 859 } 860 _igvn._worklist.push(sfpt); 861 // rollback processed safepoints 862 while (safepoints_done.length() > 0) { 863 SafePointNode* sfpt_done = safepoints_done.pop(); 864 // remove any extra entries we added to the safepoint 865 last = sfpt_done->req() - 1; 866 for (int k = 0; k < nfields; k++) { 867 sfpt_done->del_req(last--); 868 } 869 JVMState *jvms = sfpt_done->jvms(); 870 jvms->set_endoff(sfpt_done->req()); 871 // Now make a pass over the debug information replacing any references 872 // to SafePointScalarObjectNode with the allocated object. 873 int start = jvms->debug_start(); 874 int end = jvms->debug_end(); 875 for (int i = start; i < end; i++) { 876 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 877 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 878 if (scobj->first_index(jvms) == sfpt_done->req() && 879 scobj->n_fields() == (uint)nfields) { 880 assert(scobj->alloc() == alloc, "sanity"); 881 sfpt_done->set_req(i, res); 882 } 883 } 884 } 885 _igvn._worklist.push(sfpt_done); 886 } 887 #ifndef PRODUCT 888 if (PrintEliminateAllocations) { 889 if (field != NULL) { 890 tty->print("=== At SafePoint node %d can't find value of Field: ", 891 sfpt->_idx); 892 field->print(); 893 int field_idx = C->get_alias_index(field_addr_type); 894 tty->print(" (alias_idx=%d)", field_idx); 895 } else { // Array's element 896 tty->print("=== At SafePoint node %d can't find value of array element [%d]", 897 sfpt->_idx, j); 898 } 899 tty->print(", which prevents elimination of: "); 900 if (res == NULL) 901 alloc->dump(); 902 else 903 res->dump(); 904 } 905 #endif 906 return false; 907 } 908 if (UseCompressedOops && field_type->isa_narrowoop()) { 909 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 910 // to be able scalar replace the allocation. 911 if (field_val->is_EncodeP()) { 912 field_val = field_val->in(1); 913 } else { 914 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type())); 915 } 916 } else if (field_val->is_ValueType()) { 917 // Keep track of value types to scalarize them later 918 value_worklist.push(field_val); 919 } 920 sfpt->add_req(field_val); 921 } 922 JVMState *jvms = sfpt->jvms(); 923 jvms->set_endoff(sfpt->req()); 924 // Now make a pass over the debug information replacing any references 925 // to the allocated object with "sobj" 926 int start = jvms->debug_start(); 927 int end = jvms->debug_end(); 928 sfpt->replace_edges_in_range(res, sobj, start, end); 929 _igvn._worklist.push(sfpt); 930 safepoints_done.append_if_missing(sfpt); // keep it for rollback 931 } 932 // Scalarize value types that were added to the safepoint 933 for (uint i = 0; i < value_worklist.size(); ++i) { 934 Node* vt = value_worklist.at(i); 935 vt->as_ValueType()->make_scalar_in_safepoints(C->root(), &_igvn); 936 } 937 return true; 938 } 939 940 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) { 941 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control); 942 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory); 943 if (ctl_proj != NULL) { 944 igvn.replace_node(ctl_proj, n->in(0)); 945 } 946 if (mem_proj != NULL) { 947 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory)); 948 } 949 } 950 951 // Process users of eliminated allocation. 952 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) { 953 Node* res = alloc->result_cast(); 954 if (res != NULL) { 955 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 956 Node *use = res->last_out(j); 957 uint oc1 = res->outcnt(); 958 959 if (use->is_AddP()) { 960 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 961 Node *n = use->last_out(k); 962 uint oc2 = use->outcnt(); 963 if (n->is_Store()) { 964 #ifdef ASSERT 965 // Verify that there is no dependent MemBarVolatile nodes, 966 // they should be removed during IGVN, see MemBarNode::Ideal(). 967 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); 968 p < pmax; p++) { 969 Node* mb = n->fast_out(p); 970 assert(mb->is_Initialize() || !mb->is_MemBar() || 971 mb->req() <= MemBarNode::Precedent || 972 mb->in(MemBarNode::Precedent) != n, 973 "MemBarVolatile should be eliminated for non-escaping object"); 974 } 975 #endif 976 _igvn.replace_node(n, n->in(MemNode::Memory)); 977 } else if (n->is_ArrayCopy()) { 978 // Disconnect ArrayCopy node 979 ArrayCopyNode* ac = n->as_ArrayCopy(); 980 assert(ac->is_clonebasic(), "unexpected array copy kind"); 981 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out(); 982 disconnect_projections(ac, _igvn); 983 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation"); 984 Node* membar_before = alloc->in(0)->in(0); 985 disconnect_projections(membar_before->as_MemBar(), _igvn); 986 if (membar_after->is_MemBar()) { 987 disconnect_projections(membar_after->as_MemBar(), _igvn); 988 } 989 } else { 990 eliminate_gc_barrier(n); 991 } 992 k -= (oc2 - use->outcnt()); 993 } 994 } else if (use->is_ArrayCopy()) { 995 // Disconnect ArrayCopy node 996 ArrayCopyNode* ac = use->as_ArrayCopy(); 997 assert(ac->is_arraycopy_validated() || 998 ac->is_copyof_validated() || 999 ac->is_copyofrange_validated(), "unsupported"); 1000 CallProjections* callprojs = ac->extract_projections(true); 1001 1002 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O)); 1003 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory)); 1004 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control)); 1005 1006 // Set control to top. IGVN will remove the remaining projections 1007 ac->set_req(0, top()); 1008 ac->replace_edge(res, top()); 1009 1010 // Disconnect src right away: it can help find new 1011 // opportunities for allocation elimination 1012 Node* src = ac->in(ArrayCopyNode::Src); 1013 ac->replace_edge(src, top()); 1014 // src can be top at this point if src and dest of the 1015 // arraycopy were the same 1016 if (src->outcnt() == 0 && !src->is_top()) { 1017 _igvn.remove_dead_node(src); 1018 } 1019 1020 _igvn._worklist.push(ac); 1021 } else { 1022 eliminate_gc_barrier(use); 1023 } 1024 j -= (oc1 - res->outcnt()); 1025 } 1026 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 1027 _igvn.remove_dead_node(res); 1028 } 1029 1030 // 1031 // Process other users of allocation's projections 1032 // 1033 if (_resproj != NULL && _resproj->outcnt() != 0) { 1034 // First disconnect stores captured by Initialize node. 1035 // If Initialize node is eliminated first in the following code, 1036 // it will kill such stores and DUIterator_Last will assert. 1037 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) { 1038 Node *use = _resproj->fast_out(j); 1039 if (use->is_AddP()) { 1040 // raw memory addresses used only by the initialization 1041 _igvn.replace_node(use, C->top()); 1042 --j; --jmax; 1043 } 1044 } 1045 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { 1046 Node *use = _resproj->last_out(j); 1047 uint oc1 = _resproj->outcnt(); 1048 if (use->is_Initialize()) { 1049 // Eliminate Initialize node. 1050 InitializeNode *init = use->as_Initialize(); 1051 assert(init->outcnt() <= 2, "only a control and memory projection expected"); 1052 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control); 1053 if (ctrl_proj != NULL) { 1054 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection"); 1055 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj); 1056 } 1057 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory); 1058 if (mem_proj != NULL) { 1059 Node *mem = init->in(TypeFunc::Memory); 1060 #ifdef ASSERT 1061 if (mem->is_MergeMem()) { 1062 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection"); 1063 } else { 1064 assert(mem == _memproj_fallthrough, "allocation memory projection"); 1065 } 1066 #endif 1067 _igvn.replace_node(mem_proj, mem); 1068 } 1069 } else { 1070 assert(false, "only Initialize or AddP expected"); 1071 } 1072 j -= (oc1 - _resproj->outcnt()); 1073 } 1074 } 1075 if (_fallthroughcatchproj != NULL) { 1076 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); 1077 } 1078 if (_memproj_fallthrough != NULL) { 1079 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); 1080 } 1081 if (_memproj_catchall != NULL) { 1082 _igvn.replace_node(_memproj_catchall, C->top()); 1083 } 1084 if (_ioproj_fallthrough != NULL) { 1085 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); 1086 } 1087 if (_ioproj_catchall != NULL) { 1088 _igvn.replace_node(_ioproj_catchall, C->top()); 1089 } 1090 if (_catchallcatchproj != NULL) { 1091 _igvn.replace_node(_catchallcatchproj, C->top()); 1092 } 1093 } 1094 1095 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 1096 // Don't do scalar replacement if the frame can be popped by JVMTI: 1097 // if reallocation fails during deoptimization we'll pop all 1098 // interpreter frames for this compiled frame and that won't play 1099 // nice with JVMTI popframe. 1100 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) { 1101 return false; 1102 } 1103 Node* klass = alloc->in(AllocateNode::KlassNode); 1104 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); 1105 Node* res = alloc->result_cast(); 1106 // Eliminate boxing allocations which are not used 1107 // regardless scalar replacable status. 1108 bool boxing_alloc = C->eliminate_boxing() && 1109 tklass->klass()->is_instance_klass() && 1110 tklass->klass()->as_instance_klass()->is_box_klass(); 1111 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) { 1112 return false; 1113 } 1114 1115 extract_call_projections(alloc); 1116 1117 GrowableArray <SafePointNode *> safepoints; 1118 if (!can_eliminate_allocation(alloc, safepoints)) { 1119 return false; 1120 } 1121 1122 if (!alloc->_is_scalar_replaceable) { 1123 assert(res == NULL, "sanity"); 1124 // We can only eliminate allocation if all debug info references 1125 // are already replaced with SafePointScalarObject because 1126 // we can't search for a fields value without instance_id. 1127 if (safepoints.length() > 0) { 1128 return false; 1129 } 1130 } 1131 1132 if (!scalar_replacement(alloc, safepoints)) { 1133 return false; 1134 } 1135 1136 CompileLog* log = C->log(); 1137 if (log != NULL) { 1138 log->head("eliminate_allocation type='%d'", 1139 log->identify(tklass->klass())); 1140 JVMState* p = alloc->jvms(); 1141 while (p != NULL) { 1142 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1143 p = p->caller(); 1144 } 1145 log->tail("eliminate_allocation"); 1146 } 1147 1148 process_users_of_allocation(alloc); 1149 1150 #ifndef PRODUCT 1151 if (PrintEliminateAllocations) { 1152 if (alloc->is_AllocateArray()) 1153 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1154 else 1155 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1156 } 1157 #endif 1158 1159 return true; 1160 } 1161 1162 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) { 1163 // EA should remove all uses of non-escaping boxing node. 1164 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) { 1165 return false; 1166 } 1167 1168 assert(boxing->result_cast() == NULL, "unexpected boxing node result"); 1169 1170 extract_call_projections(boxing); 1171 1172 const TypeTuple* r = boxing->tf()->range_sig(); 1173 assert(r->cnt() > TypeFunc::Parms, "sanity"); 1174 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr(); 1175 assert(t != NULL, "sanity"); 1176 1177 CompileLog* log = C->log(); 1178 if (log != NULL) { 1179 log->head("eliminate_boxing type='%d'", 1180 log->identify(t->klass())); 1181 JVMState* p = boxing->jvms(); 1182 while (p != NULL) { 1183 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1184 p = p->caller(); 1185 } 1186 log->tail("eliminate_boxing"); 1187 } 1188 1189 process_users_of_allocation(boxing); 1190 1191 #ifndef PRODUCT 1192 if (PrintEliminateAllocations) { 1193 tty->print("++++ Eliminated: %d ", boxing->_idx); 1194 boxing->method()->print_short_name(tty); 1195 tty->cr(); 1196 } 1197 #endif 1198 1199 return true; 1200 } 1201 1202 //---------------------------set_eden_pointers------------------------- 1203 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { 1204 if (UseTLAB) { // Private allocation: load from TLS 1205 Node* thread = transform_later(new ThreadLocalNode()); 1206 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); 1207 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); 1208 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); 1209 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); 1210 } else { // Shared allocation: load from globals 1211 CollectedHeap* ch = Universe::heap(); 1212 address top_adr = (address)ch->top_addr(); 1213 address end_adr = (address)ch->end_addr(); 1214 eden_top_adr = makecon(TypeRawPtr::make(top_adr)); 1215 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); 1216 } 1217 } 1218 1219 1220 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 1221 Node* adr = basic_plus_adr(base, offset); 1222 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 1223 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered); 1224 transform_later(value); 1225 return value; 1226 } 1227 1228 1229 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 1230 Node* adr = basic_plus_adr(base, offset); 1231 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered); 1232 transform_later(mem); 1233 return mem; 1234 } 1235 1236 //============================================================================= 1237 // 1238 // A L L O C A T I O N 1239 // 1240 // Allocation attempts to be fast in the case of frequent small objects. 1241 // It breaks down like this: 1242 // 1243 // 1) Size in doublewords is computed. This is a constant for objects and 1244 // variable for most arrays. Doubleword units are used to avoid size 1245 // overflow of huge doubleword arrays. We need doublewords in the end for 1246 // rounding. 1247 // 1248 // 2) Size is checked for being 'too large'. Too-large allocations will go 1249 // the slow path into the VM. The slow path can throw any required 1250 // exceptions, and does all the special checks for very large arrays. The 1251 // size test can constant-fold away for objects. For objects with 1252 // finalizers it constant-folds the otherway: you always go slow with 1253 // finalizers. 1254 // 1255 // 3) If NOT using TLABs, this is the contended loop-back point. 1256 // Load-Locked the heap top. If using TLABs normal-load the heap top. 1257 // 1258 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 1259 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 1260 // "size*8" we always enter the VM, where "largish" is a constant picked small 1261 // enough that there's always space between the eden max and 4Gig (old space is 1262 // there so it's quite large) and large enough that the cost of entering the VM 1263 // is dwarfed by the cost to initialize the space. 1264 // 1265 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 1266 // down. If contended, repeat at step 3. If using TLABs normal-store 1267 // adjusted heap top back down; there is no contention. 1268 // 1269 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 1270 // fields. 1271 // 1272 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 1273 // oop flavor. 1274 // 1275 //============================================================================= 1276 // FastAllocateSizeLimit value is in DOUBLEWORDS. 1277 // Allocations bigger than this always go the slow route. 1278 // This value must be small enough that allocation attempts that need to 1279 // trigger exceptions go the slow route. Also, it must be small enough so 1280 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 1281 //=============================================================================j// 1282 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 1283 // The allocator will coalesce int->oop copies away. See comment in 1284 // coalesce.cpp about how this works. It depends critically on the exact 1285 // code shape produced here, so if you are changing this code shape 1286 // make sure the GC info for the heap-top is correct in and around the 1287 // slow-path call. 1288 // 1289 1290 void PhaseMacroExpand::expand_allocate_common( 1291 AllocateNode* alloc, // allocation node to be expanded 1292 Node* length, // array length for an array allocation 1293 const TypeFunc* slow_call_type, // Type of slow call 1294 address slow_call_address // Address of slow call 1295 ) 1296 { 1297 1298 Node* ctrl = alloc->in(TypeFunc::Control); 1299 Node* mem = alloc->in(TypeFunc::Memory); 1300 Node* i_o = alloc->in(TypeFunc::I_O); 1301 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 1302 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1303 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 1304 1305 assert(ctrl != NULL, "must have control"); 1306 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 1307 // they will not be used if "always_slow" is set 1308 enum { slow_result_path = 1, fast_result_path = 2 }; 1309 Node *result_region = NULL; 1310 Node *result_phi_rawmem = NULL; 1311 Node *result_phi_rawoop = NULL; 1312 Node *result_phi_i_o = NULL; 1313 1314 // The initial slow comparison is a size check, the comparison 1315 // we want to do is a BoolTest::gt 1316 bool always_slow = false; 1317 int tv = _igvn.find_int_con(initial_slow_test, -1); 1318 if (tv >= 0) { 1319 always_slow = (tv == 1); 1320 initial_slow_test = NULL; 1321 } else { 1322 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 1323 } 1324 1325 if (C->env()->dtrace_alloc_probes() || 1326 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) { 1327 // Force slow-path allocation 1328 always_slow = true; 1329 initial_slow_test = NULL; 1330 } 1331 1332 Node *slow_region = NULL; 1333 Node *toobig_false = ctrl; 1334 1335 if (!always_slow && (alloc->initialization() == NULL || alloc->initialization()->is_unknown_value())) { 1336 const TypeOopPtr* ary_type = _igvn.type(klass_node)->is_klassptr()->as_instance_type(); 1337 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr(); 1338 1339 ciKlass* elem_klass = NULL; 1340 if (ary_ptr != NULL && ary_ptr->klass() != NULL) { 1341 elem_klass = ary_ptr->klass()->as_array_klass()->element_klass(); 1342 } 1343 1344 if (elem_klass == NULL || (elem_klass->is_java_lang_Object() && !ary_ptr->klass_is_exact()) || elem_klass->is_valuetype()) { 1345 // If it's an array of values we must go to the slow path so it is 1346 // correctly initialized with default values. 1347 Node* fast_region = new RegionNode(3); 1348 Node* not_obj_array = ctrl; 1349 Node* obj_array = generate_object_array_guard(¬_obj_array, mem, klass_node, NULL); 1350 1351 fast_region->init_req(1, not_obj_array); 1352 slow_region = new RegionNode(1); 1353 Node* k_adr = basic_plus_adr(klass_node, klass_node, in_bytes(ArrayKlass::element_klass_offset())); 1354 Node* elem_klass = LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), k_adr, TypeInstPtr::KLASS); 1355 transform_later(elem_klass); 1356 Node* flags = make_load(NULL, mem, elem_klass, in_bytes(Klass::access_flags_offset()), TypeInt::INT, T_INT); 1357 Node* is_value_elem = new AndINode(flags, intcon(JVM_ACC_VALUE)); 1358 transform_later(is_value_elem); 1359 Node* cmp = new CmpINode(is_value_elem, _igvn.intcon(0)); 1360 transform_later(cmp); 1361 Node* bol = new BoolNode(cmp, BoolTest::ne); 1362 transform_later(bol); 1363 IfNode* value_array_iff = new IfNode(obj_array, bol, PROB_MIN, COUNT_UNKNOWN); 1364 transform_later(value_array_iff); 1365 Node* value_array = new IfTrueNode(value_array_iff); 1366 transform_later(value_array); 1367 slow_region->add_req(value_array); 1368 Node* not_value_array = new IfFalseNode(value_array_iff); 1369 transform_later(not_value_array); 1370 fast_region->init_req(2, not_value_array); 1371 transform_later(fast_region); 1372 ctrl = fast_region; 1373 } 1374 alloc->initialization()->clear_unknown_value(); 1375 } 1376 1377 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); 1378 // generate the initial test if necessary 1379 if (initial_slow_test != NULL ) { 1380 if (slow_region == NULL) { 1381 slow_region = new RegionNode(1); 1382 } 1383 // Now make the initial failure test. Usually a too-big test but 1384 // might be a TRUE for finalizers or a fancy class check for 1385 // newInstance0. 1386 IfNode* toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1387 transform_later(toobig_iff); 1388 // Plug the failing-too-big test into the slow-path region 1389 Node* toobig_true = new IfTrueNode(toobig_iff); 1390 transform_later(toobig_true); 1391 slow_region ->add_req(toobig_true); 1392 toobig_false = new IfFalseNode(toobig_iff); 1393 transform_later(toobig_false); 1394 } else { // No initial test, just fall into next case 1395 toobig_false = ctrl; 1396 } 1397 1398 Node *slow_mem = mem; // save the current memory state for slow path 1399 // generate the fast allocation code unless we know that the initial test will always go slow 1400 if (!always_slow) { 1401 // Fast path modifies only raw memory. 1402 if (mem->is_MergeMem()) { 1403 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1404 } 1405 1406 Node* eden_top_adr; 1407 Node* eden_end_adr; 1408 1409 set_eden_pointers(eden_top_adr, eden_end_adr); 1410 1411 // Load Eden::end. Loop invariant and hoisted. 1412 // 1413 // Note: We set the control input on "eden_end" and "old_eden_top" when using 1414 // a TLAB to work around a bug where these values were being moved across 1415 // a safepoint. These are not oops, so they cannot be include in the oop 1416 // map, but they can be changed by a GC. The proper way to fix this would 1417 // be to set the raw memory state when generating a SafepointNode. However 1418 // this will require extensive changes to the loop optimization in order to 1419 // prevent a degradation of the optimization. 1420 // See comment in memnode.hpp, around line 227 in class LoadPNode. 1421 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 1422 1423 // allocate the Region and Phi nodes for the result 1424 result_region = new RegionNode(3); 1425 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1426 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM); 1427 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch 1428 1429 // We need a Region for the loop-back contended case. 1430 enum { fall_in_path = 1, contended_loopback_path = 2 }; 1431 Node *contended_region; 1432 Node *contended_phi_rawmem; 1433 if (UseTLAB) { 1434 contended_region = toobig_false; 1435 contended_phi_rawmem = mem; 1436 } else { 1437 contended_region = new RegionNode(3); 1438 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1439 // Now handle the passing-too-big test. We fall into the contended 1440 // loop-back merge point. 1441 contended_region ->init_req(fall_in_path, toobig_false); 1442 contended_phi_rawmem->init_req(fall_in_path, mem); 1443 transform_later(contended_region); 1444 transform_later(contended_phi_rawmem); 1445 } 1446 1447 // Load(-locked) the heap top. 1448 // See note above concerning the control input when using a TLAB 1449 Node *old_eden_top = UseTLAB 1450 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered) 1451 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire); 1452 1453 transform_later(old_eden_top); 1454 // Add to heap top to get a new heap top 1455 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes); 1456 transform_later(new_eden_top); 1457 // Check for needing a GC; compare against heap end 1458 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end); 1459 transform_later(needgc_cmp); 1460 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge); 1461 transform_later(needgc_bol); 1462 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); 1463 transform_later(needgc_iff); 1464 1465 // Plug the failing-heap-space-need-gc test into the slow-path region 1466 Node *needgc_true = new IfTrueNode(needgc_iff); 1467 transform_later(needgc_true); 1468 if (slow_region != NULL) { 1469 slow_region->add_req(needgc_true); 1470 // This completes all paths into the slow merge point 1471 transform_later(slow_region); 1472 } else { 1473 // Just fall from the need-GC path straight into the VM call. 1474 slow_region = needgc_true; 1475 } 1476 // No need for a GC. Setup for the Store-Conditional 1477 Node *needgc_false = new IfFalseNode(needgc_iff); 1478 transform_later(needgc_false); 1479 1480 // Grab regular I/O before optional prefetch may change it. 1481 // Slow-path does no I/O so just set it to the original I/O. 1482 result_phi_i_o->init_req(slow_result_path, i_o); 1483 1484 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem, 1485 old_eden_top, new_eden_top, length); 1486 1487 // Name successful fast-path variables 1488 Node* fast_oop = old_eden_top; 1489 Node* fast_oop_ctrl; 1490 Node* fast_oop_rawmem; 1491 1492 // Store (-conditional) the modified eden top back down. 1493 // StorePConditional produces flags for a test PLUS a modified raw 1494 // memory state. 1495 if (UseTLAB) { 1496 Node* store_eden_top = 1497 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1498 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered); 1499 transform_later(store_eden_top); 1500 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path 1501 fast_oop_rawmem = store_eden_top; 1502 } else { 1503 Node* store_eden_top = 1504 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1505 new_eden_top, fast_oop/*old_eden_top*/); 1506 transform_later(store_eden_top); 1507 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne); 1508 transform_later(contention_check); 1509 store_eden_top = new SCMemProjNode(store_eden_top); 1510 transform_later(store_eden_top); 1511 1512 // If not using TLABs, check to see if there was contention. 1513 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN); 1514 transform_later(contention_iff); 1515 Node *contention_true = new IfTrueNode(contention_iff); 1516 transform_later(contention_true); 1517 // If contention, loopback and try again. 1518 contended_region->init_req(contended_loopback_path, contention_true); 1519 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top); 1520 1521 // Fast-path succeeded with no contention! 1522 Node *contention_false = new IfFalseNode(contention_iff); 1523 transform_later(contention_false); 1524 fast_oop_ctrl = contention_false; 1525 1526 // Bump total allocated bytes for this thread 1527 Node* thread = new ThreadLocalNode(); 1528 transform_later(thread); 1529 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread, 1530 in_bytes(JavaThread::allocated_bytes_offset())); 1531 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1532 0, TypeLong::LONG, T_LONG); 1533 #ifdef _LP64 1534 Node* alloc_size = size_in_bytes; 1535 #else 1536 Node* alloc_size = new ConvI2LNode(size_in_bytes); 1537 transform_later(alloc_size); 1538 #endif 1539 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size); 1540 transform_later(new_alloc_bytes); 1541 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1542 0, new_alloc_bytes, T_LONG); 1543 } 1544 1545 InitializeNode* init = alloc->initialization(); 1546 fast_oop_rawmem = initialize_object(alloc, 1547 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1548 klass_node, length, size_in_bytes); 1549 1550 // If initialization is performed by an array copy, any required 1551 // MemBarStoreStore was already added. If the object does not 1552 // escape no need for a MemBarStoreStore. If the object does not 1553 // escape in its initializer and memory barrier (MemBarStoreStore or 1554 // stronger) is already added at exit of initializer, also no need 1555 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore 1556 // so that stores that initialize this object can't be reordered 1557 // with a subsequent store that makes this object accessible by 1558 // other threads. 1559 // Other threads include java threads and JVM internal threads 1560 // (for example concurrent GC threads). Current concurrent GC 1561 // implementation: CMS and G1 will not scan newly created object, 1562 // so it's safe to skip storestore barrier when allocation does 1563 // not escape. 1564 if (!alloc->does_not_escape_thread() && 1565 !alloc->is_allocation_MemBar_redundant() && 1566 (init == NULL || !init->is_complete_with_arraycopy())) { 1567 if (init == NULL || init->req() < InitializeNode::RawStores) { 1568 // No InitializeNode or no stores captured by zeroing 1569 // elimination. Simply add the MemBarStoreStore after object 1570 // initialization. 1571 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1572 transform_later(mb); 1573 1574 mb->init_req(TypeFunc::Memory, fast_oop_rawmem); 1575 mb->init_req(TypeFunc::Control, fast_oop_ctrl); 1576 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control); 1577 transform_later(fast_oop_ctrl); 1578 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory); 1579 transform_later(fast_oop_rawmem); 1580 } else { 1581 // Add the MemBarStoreStore after the InitializeNode so that 1582 // all stores performing the initialization that were moved 1583 // before the InitializeNode happen before the storestore 1584 // barrier. 1585 1586 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control); 1587 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory); 1588 1589 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1590 transform_later(mb); 1591 1592 Node* ctrl = new ProjNode(init,TypeFunc::Control); 1593 transform_later(ctrl); 1594 Node* mem = new ProjNode(init,TypeFunc::Memory); 1595 transform_later(mem); 1596 1597 // The MemBarStoreStore depends on control and memory coming 1598 // from the InitializeNode 1599 mb->init_req(TypeFunc::Memory, mem); 1600 mb->init_req(TypeFunc::Control, ctrl); 1601 1602 ctrl = new ProjNode(mb,TypeFunc::Control); 1603 transform_later(ctrl); 1604 mem = new ProjNode(mb,TypeFunc::Memory); 1605 transform_later(mem); 1606 1607 // All nodes that depended on the InitializeNode for control 1608 // and memory must now depend on the MemBarNode that itself 1609 // depends on the InitializeNode 1610 if (init_ctrl != NULL) { 1611 _igvn.replace_node(init_ctrl, ctrl); 1612 } 1613 if (init_mem != NULL) { 1614 _igvn.replace_node(init_mem, mem); 1615 } 1616 } 1617 } 1618 1619 if (C->env()->dtrace_extended_probes()) { 1620 // Slow-path call 1621 int size = TypeFunc::Parms + 2; 1622 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1623 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), 1624 "dtrace_object_alloc", 1625 TypeRawPtr::BOTTOM); 1626 1627 // Get base of thread-local storage area 1628 Node* thread = new ThreadLocalNode(); 1629 transform_later(thread); 1630 1631 call->init_req(TypeFunc::Parms+0, thread); 1632 call->init_req(TypeFunc::Parms+1, fast_oop); 1633 call->init_req(TypeFunc::Control, fast_oop_ctrl); 1634 call->init_req(TypeFunc::I_O , top()); // does no i/o 1635 call->init_req(TypeFunc::Memory , fast_oop_rawmem); 1636 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1637 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1638 transform_later(call); 1639 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control); 1640 transform_later(fast_oop_ctrl); 1641 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory); 1642 transform_later(fast_oop_rawmem); 1643 } 1644 1645 // Plug in the successful fast-path into the result merge point 1646 result_region ->init_req(fast_result_path, fast_oop_ctrl); 1647 result_phi_rawoop->init_req(fast_result_path, fast_oop); 1648 result_phi_i_o ->init_req(fast_result_path, i_o); 1649 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); 1650 } else { 1651 slow_region = ctrl; 1652 result_phi_i_o = i_o; // Rename it to use in the following code. 1653 } 1654 1655 // Generate slow-path call 1656 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address, 1657 OptoRuntime::stub_name(slow_call_address), 1658 alloc->jvms()->bci(), 1659 TypePtr::BOTTOM); 1660 call->init_req( TypeFunc::Control, slow_region ); 1661 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1662 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs 1663 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1664 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1665 1666 call->init_req(TypeFunc::Parms+0, klass_node); 1667 if (length != NULL) { 1668 call->init_req(TypeFunc::Parms+1, length); 1669 } 1670 1671 // Copy debug information and adjust JVMState information, then replace 1672 // allocate node with the call 1673 copy_call_debug_info((CallNode *) alloc, call); 1674 if (!always_slow) { 1675 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1676 } else { 1677 // Hook i_o projection to avoid its elimination during allocation 1678 // replacement (when only a slow call is generated). 1679 call->set_req(TypeFunc::I_O, result_phi_i_o); 1680 } 1681 _igvn.replace_node(alloc, call); 1682 transform_later(call); 1683 1684 // Identify the output projections from the allocate node and 1685 // adjust any references to them. 1686 // The control and io projections look like: 1687 // 1688 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1689 // Allocate Catch 1690 // ^---Proj(io) <-------+ ^---CatchProj(io) 1691 // 1692 // We are interested in the CatchProj nodes. 1693 // 1694 extract_call_projections(call); 1695 1696 // An allocate node has separate memory projections for the uses on 1697 // the control and i_o paths. Replace the control memory projection with 1698 // result_phi_rawmem (unless we are only generating a slow call when 1699 // both memory projections are combined) 1700 if (!always_slow && _memproj_fallthrough != NULL) { 1701 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { 1702 Node *use = _memproj_fallthrough->fast_out(i); 1703 _igvn.rehash_node_delayed(use); 1704 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); 1705 // back up iterator 1706 --i; 1707 } 1708 } 1709 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete 1710 // _memproj_catchall so we end up with a call that has only 1 memory projection. 1711 if (_memproj_catchall != NULL ) { 1712 if (_memproj_fallthrough == NULL) { 1713 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory); 1714 transform_later(_memproj_fallthrough); 1715 } 1716 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { 1717 Node *use = _memproj_catchall->fast_out(i); 1718 _igvn.rehash_node_delayed(use); 1719 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); 1720 // back up iterator 1721 --i; 1722 } 1723 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted"); 1724 _igvn.remove_dead_node(_memproj_catchall); 1725 } 1726 1727 // An allocate node has separate i_o projections for the uses on the control 1728 // and i_o paths. Always replace the control i_o projection with result i_o 1729 // otherwise incoming i_o become dead when only a slow call is generated 1730 // (it is different from memory projections where both projections are 1731 // combined in such case). 1732 if (_ioproj_fallthrough != NULL) { 1733 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { 1734 Node *use = _ioproj_fallthrough->fast_out(i); 1735 _igvn.rehash_node_delayed(use); 1736 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); 1737 // back up iterator 1738 --i; 1739 } 1740 } 1741 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete 1742 // _ioproj_catchall so we end up with a call that has only 1 i_o projection. 1743 if (_ioproj_catchall != NULL ) { 1744 if (_ioproj_fallthrough == NULL) { 1745 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O); 1746 transform_later(_ioproj_fallthrough); 1747 } 1748 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { 1749 Node *use = _ioproj_catchall->fast_out(i); 1750 _igvn.rehash_node_delayed(use); 1751 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); 1752 // back up iterator 1753 --i; 1754 } 1755 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted"); 1756 _igvn.remove_dead_node(_ioproj_catchall); 1757 } 1758 1759 // if we generated only a slow call, we are done 1760 if (always_slow) { 1761 // Now we can unhook i_o. 1762 if (result_phi_i_o->outcnt() > 1) { 1763 call->set_req(TypeFunc::I_O, top()); 1764 } else { 1765 assert(result_phi_i_o->unique_ctrl_out() == call, ""); 1766 // Case of new array with negative size known during compilation. 1767 // AllocateArrayNode::Ideal() optimization disconnect unreachable 1768 // following code since call to runtime will throw exception. 1769 // As result there will be no users of i_o after the call. 1770 // Leave i_o attached to this call to avoid problems in preceding graph. 1771 } 1772 return; 1773 } 1774 1775 1776 if (_fallthroughcatchproj != NULL) { 1777 ctrl = _fallthroughcatchproj->clone(); 1778 transform_later(ctrl); 1779 _igvn.replace_node(_fallthroughcatchproj, result_region); 1780 } else { 1781 ctrl = top(); 1782 } 1783 Node *slow_result; 1784 if (_resproj == NULL) { 1785 // no uses of the allocation result 1786 slow_result = top(); 1787 } else { 1788 slow_result = _resproj->clone(); 1789 transform_later(slow_result); 1790 _igvn.replace_node(_resproj, result_phi_rawoop); 1791 } 1792 1793 // Plug slow-path into result merge point 1794 result_region ->init_req( slow_result_path, ctrl ); 1795 result_phi_rawoop->init_req( slow_result_path, slow_result); 1796 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); 1797 transform_later(result_region); 1798 transform_later(result_phi_rawoop); 1799 transform_later(result_phi_rawmem); 1800 transform_later(result_phi_i_o); 1801 // This completes all paths into the result merge point 1802 } 1803 1804 1805 // Helper for PhaseMacroExpand::expand_allocate_common. 1806 // Initializes the newly-allocated storage. 1807 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1808 Node* control, Node* rawmem, Node* object, 1809 Node* klass_node, Node* length, 1810 Node* size_in_bytes) { 1811 InitializeNode* init = alloc->initialization(); 1812 // Store the klass & mark bits 1813 Node* mark_node = NULL; 1814 // For now only enable fast locking for non-array types 1815 if ((EnableValhalla || UseBiasedLocking) && length == NULL) { 1816 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 1817 } else { 1818 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); 1819 } 1820 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 1821 1822 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 1823 int header_size = alloc->minimum_header_size(); // conservatively small 1824 1825 // Array length 1826 if (length != NULL) { // Arrays need length field 1827 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1828 // conservatively small header size: 1829 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1830 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1831 if (k->is_array_klass()) // we know the exact header size in most cases: 1832 header_size = Klass::layout_helper_header_size(k->layout_helper()); 1833 } 1834 1835 // Clear the object body, if necessary. 1836 if (init == NULL) { 1837 // The init has somehow disappeared; be cautious and clear everything. 1838 // 1839 // This can happen if a node is allocated but an uncommon trap occurs 1840 // immediately. In this case, the Initialize gets associated with the 1841 // trap, and may be placed in a different (outer) loop, if the Allocate 1842 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1843 // there can be two Allocates to one Initialize. The answer in all these 1844 // edge cases is safety first. It is always safe to clear immediately 1845 // within an Allocate, and then (maybe or maybe not) clear some more later. 1846 if (!(UseTLAB && ZeroTLAB)) { 1847 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1848 header_size, size_in_bytes, 1849 &_igvn); 1850 } 1851 } else { 1852 if (!init->is_complete()) { 1853 // Try to win by zeroing only what the init does not store. 1854 // We can also try to do some peephole optimizations, 1855 // such as combining some adjacent subword stores. 1856 rawmem = init->complete_stores(control, rawmem, object, 1857 header_size, size_in_bytes, &_igvn); 1858 } 1859 // We have no more use for this link, since the AllocateNode goes away: 1860 init->set_req(InitializeNode::RawAddress, top()); 1861 // (If we keep the link, it just confuses the register allocator, 1862 // who thinks he sees a real use of the address by the membar.) 1863 } 1864 1865 return rawmem; 1866 } 1867 1868 // Generate prefetch instructions for next allocations. 1869 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1870 Node*& contended_phi_rawmem, 1871 Node* old_eden_top, Node* new_eden_top, 1872 Node* length) { 1873 enum { fall_in_path = 1, pf_path = 2 }; 1874 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1875 // Generate prefetch allocation with watermark check. 1876 // As an allocation hits the watermark, we will prefetch starting 1877 // at a "distance" away from watermark. 1878 1879 Node *pf_region = new RegionNode(3); 1880 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 1881 TypeRawPtr::BOTTOM ); 1882 // I/O is used for Prefetch 1883 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO ); 1884 1885 Node *thread = new ThreadLocalNode(); 1886 transform_later(thread); 1887 1888 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread, 1889 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1890 transform_later(eden_pf_adr); 1891 1892 Node *old_pf_wm = new LoadPNode(needgc_false, 1893 contended_phi_rawmem, eden_pf_adr, 1894 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, 1895 MemNode::unordered); 1896 transform_later(old_pf_wm); 1897 1898 // check against new_eden_top 1899 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm ); 1900 transform_later(need_pf_cmp); 1901 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge ); 1902 transform_later(need_pf_bol); 1903 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol, 1904 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1905 transform_later(need_pf_iff); 1906 1907 // true node, add prefetchdistance 1908 Node *need_pf_true = new IfTrueNode( need_pf_iff ); 1909 transform_later(need_pf_true); 1910 1911 Node *need_pf_false = new IfFalseNode( need_pf_iff ); 1912 transform_later(need_pf_false); 1913 1914 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm, 1915 _igvn.MakeConX(AllocatePrefetchDistance) ); 1916 transform_later(new_pf_wmt ); 1917 new_pf_wmt->set_req(0, need_pf_true); 1918 1919 Node *store_new_wmt = new StorePNode(need_pf_true, 1920 contended_phi_rawmem, eden_pf_adr, 1921 TypeRawPtr::BOTTOM, new_pf_wmt, 1922 MemNode::unordered); 1923 transform_later(store_new_wmt); 1924 1925 // adding prefetches 1926 pf_phi_abio->init_req( fall_in_path, i_o ); 1927 1928 Node *prefetch_adr; 1929 Node *prefetch; 1930 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1931 uint step_size = AllocatePrefetchStepSize; 1932 uint distance = 0; 1933 1934 for ( uint i = 0; i < lines; i++ ) { 1935 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt, 1936 _igvn.MakeConX(distance) ); 1937 transform_later(prefetch_adr); 1938 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1939 transform_later(prefetch); 1940 distance += step_size; 1941 i_o = prefetch; 1942 } 1943 pf_phi_abio->set_req( pf_path, i_o ); 1944 1945 pf_region->init_req( fall_in_path, need_pf_false ); 1946 pf_region->init_req( pf_path, need_pf_true ); 1947 1948 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 1949 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 1950 1951 transform_later(pf_region); 1952 transform_later(pf_phi_rawmem); 1953 transform_later(pf_phi_abio); 1954 1955 needgc_false = pf_region; 1956 contended_phi_rawmem = pf_phi_rawmem; 1957 i_o = pf_phi_abio; 1958 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { 1959 // Insert a prefetch instruction for each allocation. 1960 // This code is used to generate 1 prefetch instruction per cache line. 1961 1962 // Generate several prefetch instructions. 1963 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1964 uint step_size = AllocatePrefetchStepSize; 1965 uint distance = AllocatePrefetchDistance; 1966 1967 // Next cache address. 1968 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top, 1969 _igvn.MakeConX(step_size + distance)); 1970 transform_later(cache_adr); 1971 cache_adr = new CastP2XNode(needgc_false, cache_adr); 1972 transform_later(cache_adr); 1973 // Address is aligned to execute prefetch to the beginning of cache line size 1974 // (it is important when BIS instruction is used on SPARC as prefetch). 1975 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); 1976 cache_adr = new AndXNode(cache_adr, mask); 1977 transform_later(cache_adr); 1978 cache_adr = new CastX2PNode(cache_adr); 1979 transform_later(cache_adr); 1980 1981 // Prefetch 1982 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); 1983 prefetch->set_req(0, needgc_false); 1984 transform_later(prefetch); 1985 contended_phi_rawmem = prefetch; 1986 Node *prefetch_adr; 1987 distance = step_size; 1988 for ( uint i = 1; i < lines; i++ ) { 1989 prefetch_adr = new AddPNode( cache_adr, cache_adr, 1990 _igvn.MakeConX(distance) ); 1991 transform_later(prefetch_adr); 1992 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); 1993 transform_later(prefetch); 1994 distance += step_size; 1995 contended_phi_rawmem = prefetch; 1996 } 1997 } else if( AllocatePrefetchStyle > 0 ) { 1998 // Insert a prefetch for each allocation only on the fast-path 1999 Node *prefetch_adr; 2000 Node *prefetch; 2001 // Generate several prefetch instructions. 2002 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 2003 uint step_size = AllocatePrefetchStepSize; 2004 uint distance = AllocatePrefetchDistance; 2005 for ( uint i = 0; i < lines; i++ ) { 2006 prefetch_adr = new AddPNode( old_eden_top, new_eden_top, 2007 _igvn.MakeConX(distance) ); 2008 transform_later(prefetch_adr); 2009 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 2010 // Do not let it float too high, since if eden_top == eden_end, 2011 // both might be null. 2012 if( i == 0 ) { // Set control for first prefetch, next follows it 2013 prefetch->init_req(0, needgc_false); 2014 } 2015 transform_later(prefetch); 2016 distance += step_size; 2017 i_o = prefetch; 2018 } 2019 } 2020 return i_o; 2021 } 2022 2023 2024 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 2025 expand_allocate_common(alloc, NULL, 2026 OptoRuntime::new_instance_Type(), 2027 OptoRuntime::new_instance_Java()); 2028 } 2029 2030 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 2031 Node* length = alloc->in(AllocateNode::ALength); 2032 InitializeNode* init = alloc->initialization(); 2033 Node* klass_node = alloc->in(AllocateNode::KlassNode); 2034 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 2035 address slow_call_address; // Address of slow call 2036 if (init != NULL && init->is_complete_with_arraycopy() && 2037 k->is_type_array_klass()) { 2038 // Don't zero type array during slow allocation in VM since 2039 // it will be initialized later by arraycopy in compiled code. 2040 slow_call_address = OptoRuntime::new_array_nozero_Java(); 2041 } else { 2042 slow_call_address = OptoRuntime::new_array_Java(); 2043 } 2044 expand_allocate_common(alloc, length, 2045 OptoRuntime::new_array_Type(), 2046 slow_call_address); 2047 } 2048 2049 //-------------------mark_eliminated_box---------------------------------- 2050 // 2051 // During EA obj may point to several objects but after few ideal graph 2052 // transformations (CCP) it may point to only one non escaping object 2053 // (but still using phi), corresponding locks and unlocks will be marked 2054 // for elimination. Later obj could be replaced with a new node (new phi) 2055 // and which does not have escape information. And later after some graph 2056 // reshape other locks and unlocks (which were not marked for elimination 2057 // before) are connected to this new obj (phi) but they still will not be 2058 // marked for elimination since new obj has no escape information. 2059 // Mark all associated (same box and obj) lock and unlock nodes for 2060 // elimination if some of them marked already. 2061 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) { 2062 if (oldbox->as_BoxLock()->is_eliminated()) 2063 return; // This BoxLock node was processed already. 2064 2065 // New implementation (EliminateNestedLocks) has separate BoxLock 2066 // node for each locked region so mark all associated locks/unlocks as 2067 // eliminated even if different objects are referenced in one locked region 2068 // (for example, OSR compilation of nested loop inside locked scope). 2069 if (EliminateNestedLocks || 2070 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) { 2071 // Box is used only in one lock region. Mark this box as eliminated. 2072 _igvn.hash_delete(oldbox); 2073 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value 2074 _igvn.hash_insert(oldbox); 2075 2076 for (uint i = 0; i < oldbox->outcnt(); i++) { 2077 Node* u = oldbox->raw_out(i); 2078 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) { 2079 AbstractLockNode* alock = u->as_AbstractLock(); 2080 // Check lock's box since box could be referenced by Lock's debug info. 2081 if (alock->box_node() == oldbox) { 2082 // Mark eliminated all related locks and unlocks. 2083 #ifdef ASSERT 2084 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4"); 2085 #endif 2086 alock->set_non_esc_obj(); 2087 } 2088 } 2089 } 2090 return; 2091 } 2092 2093 // Create new "eliminated" BoxLock node and use it in monitor debug info 2094 // instead of oldbox for the same object. 2095 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 2096 2097 // Note: BoxLock node is marked eliminated only here and it is used 2098 // to indicate that all associated lock and unlock nodes are marked 2099 // for elimination. 2100 newbox->set_eliminated(); 2101 transform_later(newbox); 2102 2103 // Replace old box node with new box for all users of the same object. 2104 for (uint i = 0; i < oldbox->outcnt();) { 2105 bool next_edge = true; 2106 2107 Node* u = oldbox->raw_out(i); 2108 if (u->is_AbstractLock()) { 2109 AbstractLockNode* alock = u->as_AbstractLock(); 2110 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) { 2111 // Replace Box and mark eliminated all related locks and unlocks. 2112 #ifdef ASSERT 2113 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5"); 2114 #endif 2115 alock->set_non_esc_obj(); 2116 _igvn.rehash_node_delayed(alock); 2117 alock->set_box_node(newbox); 2118 next_edge = false; 2119 } 2120 } 2121 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) { 2122 FastLockNode* flock = u->as_FastLock(); 2123 assert(flock->box_node() == oldbox, "sanity"); 2124 _igvn.rehash_node_delayed(flock); 2125 flock->set_box_node(newbox); 2126 next_edge = false; 2127 } 2128 2129 // Replace old box in monitor debug info. 2130 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 2131 SafePointNode* sfn = u->as_SafePoint(); 2132 JVMState* youngest_jvms = sfn->jvms(); 2133 int max_depth = youngest_jvms->depth(); 2134 for (int depth = 1; depth <= max_depth; depth++) { 2135 JVMState* jvms = youngest_jvms->of_depth(depth); 2136 int num_mon = jvms->nof_monitors(); 2137 // Loop over monitors 2138 for (int idx = 0; idx < num_mon; idx++) { 2139 Node* obj_node = sfn->monitor_obj(jvms, idx); 2140 Node* box_node = sfn->monitor_box(jvms, idx); 2141 if (box_node == oldbox && obj_node->eqv_uncast(obj)) { 2142 int j = jvms->monitor_box_offset(idx); 2143 _igvn.replace_input_of(u, j, newbox); 2144 next_edge = false; 2145 } 2146 } 2147 } 2148 } 2149 if (next_edge) i++; 2150 } 2151 } 2152 2153 //-----------------------mark_eliminated_locking_nodes----------------------- 2154 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { 2155 if (EliminateNestedLocks) { 2156 if (alock->is_nested()) { 2157 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity"); 2158 return; 2159 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened 2160 // Only Lock node has JVMState needed here. 2161 // Not that preceding claim is documented anywhere else. 2162 if (alock->jvms() != NULL) { 2163 if (alock->as_Lock()->is_nested_lock_region()) { 2164 // Mark eliminated related nested locks and unlocks. 2165 Node* obj = alock->obj_node(); 2166 BoxLockNode* box_node = alock->box_node()->as_BoxLock(); 2167 assert(!box_node->is_eliminated(), "should not be marked yet"); 2168 // Note: BoxLock node is marked eliminated only here 2169 // and it is used to indicate that all associated lock 2170 // and unlock nodes are marked for elimination. 2171 box_node->set_eliminated(); // Box's hash is always NO_HASH here 2172 for (uint i = 0; i < box_node->outcnt(); i++) { 2173 Node* u = box_node->raw_out(i); 2174 if (u->is_AbstractLock()) { 2175 alock = u->as_AbstractLock(); 2176 if (alock->box_node() == box_node) { 2177 // Verify that this Box is referenced only by related locks. 2178 assert(alock->obj_node()->eqv_uncast(obj), ""); 2179 // Mark all related locks and unlocks. 2180 #ifdef ASSERT 2181 alock->log_lock_optimization(C, "eliminate_lock_set_nested"); 2182 #endif 2183 alock->set_nested(); 2184 } 2185 } 2186 } 2187 } else { 2188 #ifdef ASSERT 2189 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region"); 2190 if (C->log() != NULL) 2191 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output 2192 #endif 2193 } 2194 } 2195 return; 2196 } 2197 // Process locks for non escaping object 2198 assert(alock->is_non_esc_obj(), ""); 2199 } // EliminateNestedLocks 2200 2201 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object 2202 // Look for all locks of this object and mark them and 2203 // corresponding BoxLock nodes as eliminated. 2204 Node* obj = alock->obj_node(); 2205 for (uint j = 0; j < obj->outcnt(); j++) { 2206 Node* o = obj->raw_out(j); 2207 if (o->is_AbstractLock() && 2208 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) { 2209 alock = o->as_AbstractLock(); 2210 Node* box = alock->box_node(); 2211 // Replace old box node with new eliminated box for all users 2212 // of the same object and mark related locks as eliminated. 2213 mark_eliminated_box(box, obj); 2214 } 2215 } 2216 } 2217 } 2218 2219 // we have determined that this lock/unlock can be eliminated, we simply 2220 // eliminate the node without expanding it. 2221 // 2222 // Note: The membar's associated with the lock/unlock are currently not 2223 // eliminated. This should be investigated as a future enhancement. 2224 // 2225 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 2226 2227 if (!alock->is_eliminated()) { 2228 return false; 2229 } 2230 #ifdef ASSERT 2231 if (!alock->is_coarsened()) { 2232 // Check that new "eliminated" BoxLock node is created. 2233 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 2234 assert(oldbox->is_eliminated(), "should be done already"); 2235 } 2236 #endif 2237 2238 alock->log_lock_optimization(C, "eliminate_lock"); 2239 2240 #ifndef PRODUCT 2241 if (PrintEliminateLocks) { 2242 if (alock->is_Lock()) { 2243 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx); 2244 } else { 2245 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx); 2246 } 2247 } 2248 #endif 2249 2250 Node* mem = alock->in(TypeFunc::Memory); 2251 Node* ctrl = alock->in(TypeFunc::Control); 2252 2253 extract_call_projections(alock); 2254 // There are 2 projections from the lock. The lock node will 2255 // be deleted when its last use is subsumed below. 2256 assert(alock->outcnt() == 2 && 2257 _fallthroughproj != NULL && 2258 _memproj_fallthrough != NULL, 2259 "Unexpected projections from Lock/Unlock"); 2260 2261 Node* fallthroughproj = _fallthroughproj; 2262 Node* memproj_fallthrough = _memproj_fallthrough; 2263 2264 // The memory projection from a lock/unlock is RawMem 2265 // The input to a Lock is merged memory, so extract its RawMem input 2266 // (unless the MergeMem has been optimized away.) 2267 if (alock->is_Lock()) { 2268 // Seach for MemBarAcquireLock node and delete it also. 2269 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 2270 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, ""); 2271 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 2272 Node* memproj = membar->proj_out(TypeFunc::Memory); 2273 _igvn.replace_node(ctrlproj, fallthroughproj); 2274 _igvn.replace_node(memproj, memproj_fallthrough); 2275 2276 // Delete FastLock node also if this Lock node is unique user 2277 // (a loop peeling may clone a Lock node). 2278 Node* flock = alock->as_Lock()->fastlock_node(); 2279 if (flock->outcnt() == 1) { 2280 assert(flock->unique_out() == alock, "sanity"); 2281 _igvn.replace_node(flock, top()); 2282 } 2283 } 2284 2285 // Seach for MemBarReleaseLock node and delete it also. 2286 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() && 2287 ctrl->in(0)->is_MemBar()) { 2288 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 2289 assert(membar->Opcode() == Op_MemBarReleaseLock && 2290 mem->is_Proj() && membar == mem->in(0), ""); 2291 _igvn.replace_node(fallthroughproj, ctrl); 2292 _igvn.replace_node(memproj_fallthrough, mem); 2293 fallthroughproj = ctrl; 2294 memproj_fallthrough = mem; 2295 ctrl = membar->in(TypeFunc::Control); 2296 mem = membar->in(TypeFunc::Memory); 2297 } 2298 2299 _igvn.replace_node(fallthroughproj, ctrl); 2300 _igvn.replace_node(memproj_fallthrough, mem); 2301 return true; 2302 } 2303 2304 2305 //------------------------------expand_lock_node---------------------- 2306 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 2307 2308 Node* ctrl = lock->in(TypeFunc::Control); 2309 Node* mem = lock->in(TypeFunc::Memory); 2310 Node* obj = lock->obj_node(); 2311 Node* box = lock->box_node(); 2312 Node* flock = lock->fastlock_node(); 2313 2314 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2315 2316 // Make the merge point 2317 Node *region; 2318 Node *mem_phi; 2319 Node *slow_path; 2320 2321 if (UseOptoBiasInlining) { 2322 /* 2323 * See the full description in MacroAssembler::biased_locking_enter(). 2324 * 2325 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { 2326 * // The object is biased. 2327 * proto_node = klass->prototype_header; 2328 * o_node = thread | proto_node; 2329 * x_node = o_node ^ mark_word; 2330 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? 2331 * // Done. 2332 * } else { 2333 * if( (x_node & biased_lock_mask) != 0 ) { 2334 * // The klass's prototype header is no longer biased. 2335 * cas(&mark_word, mark_word, proto_node) 2336 * goto cas_lock; 2337 * } else { 2338 * // The klass's prototype header is still biased. 2339 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? 2340 * old = mark_word; 2341 * new = o_node; 2342 * } else { 2343 * // Different thread or anonymous biased. 2344 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); 2345 * new = thread | old; 2346 * } 2347 * // Try to rebias. 2348 * if( cas(&mark_word, old, new) == 0 ) { 2349 * // Done. 2350 * } else { 2351 * goto slow_path; // Failed. 2352 * } 2353 * } 2354 * } 2355 * } else { 2356 * // The object is not biased. 2357 * cas_lock: 2358 * if( FastLock(obj) == 0 ) { 2359 * // Done. 2360 * } else { 2361 * slow_path: 2362 * OptoRuntime::complete_monitor_locking_Java(obj); 2363 * } 2364 * } 2365 */ 2366 2367 region = new RegionNode(5); 2368 // create a Phi for the memory state 2369 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2370 2371 Node* fast_lock_region = new RegionNode(3); 2372 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); 2373 2374 // First, check mark word for the biased lock pattern. 2375 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2376 2377 // Get fast path - mark word has the biased lock pattern. 2378 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, 2379 markOopDesc::biased_lock_mask_in_place, 2380 markOopDesc::biased_lock_pattern, true); 2381 // fast_lock_region->in(1) is set to slow path. 2382 fast_lock_mem_phi->init_req(1, mem); 2383 2384 // Now check that the lock is biased to the current thread and has 2385 // the same epoch and bias as Klass::_prototype_header. 2386 2387 // Special-case a fresh allocation to avoid building nodes: 2388 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); 2389 if (klass_node == NULL) { 2390 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 2391 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr())); 2392 #ifdef _LP64 2393 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) { 2394 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity"); 2395 klass_node->in(1)->init_req(0, ctrl); 2396 } else 2397 #endif 2398 klass_node->init_req(0, ctrl); 2399 } 2400 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type()); 2401 2402 Node* thread = transform_later(new ThreadLocalNode()); 2403 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2404 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node)); 2405 Node* x_node = transform_later(new XorXNode(o_node, mark_node)); 2406 2407 // Get slow path - mark word does NOT match the value. 2408 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, 2409 (~markOopDesc::age_mask_in_place), 0); 2410 // region->in(3) is set to fast path - the object is biased to the current thread. 2411 mem_phi->init_req(3, mem); 2412 2413 2414 // Mark word does NOT match the value (thread | Klass::_prototype_header). 2415 2416 2417 // First, check biased pattern. 2418 // Get fast path - _prototype_header has the same biased lock pattern. 2419 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, 2420 markOopDesc::biased_lock_mask_in_place, 0, true); 2421 2422 not_biased_ctrl = fast_lock_region->in(2); // Slow path 2423 // fast_lock_region->in(2) - the prototype header is no longer biased 2424 // and we have to revoke the bias on this object. 2425 // We are going to try to reset the mark of this object to the prototype 2426 // value and fall through to the CAS-based locking scheme. 2427 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 2428 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr, 2429 proto_node, mark_node); 2430 transform_later(cas); 2431 Node* proj = transform_later(new SCMemProjNode(cas)); 2432 fast_lock_mem_phi->init_req(2, proj); 2433 2434 2435 // Second, check epoch bits. 2436 Node* rebiased_region = new RegionNode(3); 2437 Node* old_phi = new PhiNode( rebiased_region, TypeX_X); 2438 Node* new_phi = new PhiNode( rebiased_region, TypeX_X); 2439 2440 // Get slow path - mark word does NOT match epoch bits. 2441 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, 2442 markOopDesc::epoch_mask_in_place, 0); 2443 // The epoch of the current bias is not valid, attempt to rebias the object 2444 // toward the current thread. 2445 rebiased_region->init_req(2, epoch_ctrl); 2446 old_phi->init_req(2, mark_node); 2447 new_phi->init_req(2, o_node); 2448 2449 // rebiased_region->in(1) is set to fast path. 2450 // The epoch of the current bias is still valid but we know 2451 // nothing about the owner; it might be set or it might be clear. 2452 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place | 2453 markOopDesc::age_mask_in_place | 2454 markOopDesc::epoch_mask_in_place); 2455 Node* old = transform_later(new AndXNode(mark_node, cmask)); 2456 cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2457 Node* new_mark = transform_later(new OrXNode(cast_thread, old)); 2458 old_phi->init_req(1, old); 2459 new_phi->init_req(1, new_mark); 2460 2461 transform_later(rebiased_region); 2462 transform_later(old_phi); 2463 transform_later(new_phi); 2464 2465 // Try to acquire the bias of the object using an atomic operation. 2466 // If this fails we will go in to the runtime to revoke the object's bias. 2467 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi); 2468 transform_later(cas); 2469 proj = transform_later(new SCMemProjNode(cas)); 2470 2471 // Get slow path - Failed to CAS. 2472 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); 2473 mem_phi->init_req(4, proj); 2474 // region->in(4) is set to fast path - the object is rebiased to the current thread. 2475 2476 // Failed to CAS. 2477 slow_path = new RegionNode(3); 2478 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); 2479 2480 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control 2481 slow_mem->init_req(1, proj); 2482 2483 // Call CAS-based locking scheme (FastLock node). 2484 2485 transform_later(fast_lock_region); 2486 transform_later(fast_lock_mem_phi); 2487 2488 // Get slow path - FastLock failed to lock the object. 2489 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); 2490 mem_phi->init_req(2, fast_lock_mem_phi); 2491 // region->in(2) is set to fast path - the object is locked to the current thread. 2492 2493 slow_path->init_req(2, ctrl); // Capture slow-control 2494 slow_mem->init_req(2, fast_lock_mem_phi); 2495 2496 transform_later(slow_path); 2497 transform_later(slow_mem); 2498 // Reset lock's memory edge. 2499 lock->set_req(TypeFunc::Memory, slow_mem); 2500 2501 } else { 2502 region = new RegionNode(3); 2503 // create a Phi for the memory state 2504 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2505 2506 // Optimize test; set region slot 2 2507 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); 2508 mem_phi->init_req(2, mem); 2509 } 2510 2511 // Make slow path call 2512 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), 2513 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, 2514 obj, box, NULL); 2515 2516 extract_call_projections(call); 2517 2518 // Slow path can only throw asynchronous exceptions, which are always 2519 // de-opted. So the compiler thinks the slow-call can never throw an 2520 // exception. If it DOES throw an exception we would need the debug 2521 // info removed first (since if it throws there is no monitor). 2522 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2523 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2524 2525 // Capture slow path 2526 // disconnect fall-through projection from call and create a new one 2527 // hook up users of fall-through projection to region 2528 Node *slow_ctrl = _fallthroughproj->clone(); 2529 transform_later(slow_ctrl); 2530 _igvn.hash_delete(_fallthroughproj); 2531 _fallthroughproj->disconnect_inputs(NULL, C); 2532 region->init_req(1, slow_ctrl); 2533 // region inputs are now complete 2534 transform_later(region); 2535 _igvn.replace_node(_fallthroughproj, region); 2536 2537 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory)); 2538 mem_phi->init_req(1, memproj ); 2539 transform_later(mem_phi); 2540 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2541 } 2542 2543 //------------------------------expand_unlock_node---------------------- 2544 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 2545 2546 Node* ctrl = unlock->in(TypeFunc::Control); 2547 Node* mem = unlock->in(TypeFunc::Memory); 2548 Node* obj = unlock->obj_node(); 2549 Node* box = unlock->box_node(); 2550 2551 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2552 2553 // No need for a null check on unlock 2554 2555 // Make the merge point 2556 Node *region; 2557 Node *mem_phi; 2558 2559 if (UseOptoBiasInlining) { 2560 // Check for biased locking unlock case, which is a no-op. 2561 // See the full description in MacroAssembler::biased_locking_exit(). 2562 region = new RegionNode(4); 2563 // create a Phi for the memory state 2564 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2565 mem_phi->init_req(3, mem); 2566 2567 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2568 ctrl = opt_bits_test(ctrl, region, 3, mark_node, 2569 markOopDesc::biased_lock_mask_in_place, 2570 markOopDesc::biased_lock_pattern); 2571 } else { 2572 region = new RegionNode(3); 2573 // create a Phi for the memory state 2574 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2575 } 2576 2577 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box ); 2578 funlock = transform_later( funlock )->as_FastUnlock(); 2579 // Optimize test; set region slot 2 2580 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); 2581 Node *thread = transform_later(new ThreadLocalNode()); 2582 2583 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), 2584 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), 2585 "complete_monitor_unlocking_C", slow_path, obj, box, thread); 2586 2587 extract_call_projections(call); 2588 2589 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2590 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2591 2592 // No exceptions for unlocking 2593 // Capture slow path 2594 // disconnect fall-through projection from call and create a new one 2595 // hook up users of fall-through projection to region 2596 Node *slow_ctrl = _fallthroughproj->clone(); 2597 transform_later(slow_ctrl); 2598 _igvn.hash_delete(_fallthroughproj); 2599 _fallthroughproj->disconnect_inputs(NULL, C); 2600 region->init_req(1, slow_ctrl); 2601 // region inputs are now complete 2602 transform_later(region); 2603 _igvn.replace_node(_fallthroughproj, region); 2604 2605 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) ); 2606 mem_phi->init_req(1, memproj ); 2607 mem_phi->init_req(2, mem); 2608 transform_later(mem_phi); 2609 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2610 } 2611 2612 // A value type is returned from the call but we don't know its 2613 // type. Either we get a buffered value (and nothing needs to be done) 2614 // or one of the values being returned is the klass of the value type 2615 // and we need to allocate a value type instance of that type and 2616 // initialize it with other values being returned. In that case, we 2617 // first try a fast path allocation and initialize the value with the 2618 // value klass's pack handler or we fall back to a runtime call. 2619 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) { 2620 Node* ret = call->proj_out(TypeFunc::Parms); 2621 if (ret == NULL) { 2622 return; 2623 } 2624 // TODO fix this with the calling convention changes 2625 //assert(ret->bottom_type()->is_valuetypeptr()->is__Value(), "unexpected return type from MH intrinsic"); 2626 const TypeFunc* tf = call->_tf; 2627 const TypeTuple* domain = OptoRuntime::store_value_type_fields_Type()->domain_cc(); 2628 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain); 2629 call->_tf = new_tf; 2630 // Make sure the change of type is applied before projections are 2631 // processed by igvn 2632 _igvn.set_type(call, call->Value(&_igvn)); 2633 _igvn.set_type(ret, ret->Value(&_igvn)); 2634 2635 // Before any new projection is added: 2636 CallProjections* projs = call->extract_projections(true, true); 2637 2638 Node* ctl = new Node(1); 2639 Node* mem = new Node(1); 2640 Node* io = new Node(1); 2641 Node* ex_ctl = new Node(1); 2642 Node* ex_mem = new Node(1); 2643 Node* ex_io = new Node(1); 2644 Node* res = new Node(1); 2645 2646 Node* cast = transform_later(new CastP2XNode(ctl, res)); 2647 Node* mask = MakeConX(0x1); 2648 Node* masked = transform_later(new AndXNode(cast, mask)); 2649 Node* cmp = transform_later(new CmpXNode(masked, mask)); 2650 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq)); 2651 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN); 2652 transform_later(allocation_iff); 2653 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff)); 2654 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff)); 2655 2656 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::NOTNULL)); 2657 2658 Node* mask2 = MakeConX(-2); 2659 Node* masked2 = transform_later(new AndXNode(cast, mask2)); 2660 Node* rawklassptr = transform_later(new CastX2PNode(masked2)); 2661 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeKlassPtr::OBJECT_OR_NULL)); 2662 2663 Node* slowpath_bol = NULL; 2664 Node* top_adr = NULL; 2665 Node* old_top = NULL; 2666 Node* new_top = NULL; 2667 if (UseTLAB) { 2668 Node* end_adr = NULL; 2669 set_eden_pointers(top_adr, end_adr); 2670 Node* end = make_load(ctl, mem, end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 2671 old_top = new LoadPNode(ctl, mem, top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered); 2672 transform_later(old_top); 2673 Node* layout_val = make_load(NULL, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT); 2674 Node* size_in_bytes = ConvI2X(layout_val); 2675 new_top = new AddPNode(top(), old_top, size_in_bytes); 2676 transform_later(new_top); 2677 Node* slowpath_cmp = new CmpPNode(new_top, end); 2678 transform_later(slowpath_cmp); 2679 slowpath_bol = new BoolNode(slowpath_cmp, BoolTest::ge); 2680 transform_later(slowpath_bol); 2681 } else { 2682 slowpath_bol = intcon(1); 2683 top_adr = top(); 2684 old_top = top(); 2685 new_top = top(); 2686 } 2687 IfNode* slowpath_iff = new IfNode(allocation_ctl, slowpath_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); 2688 transform_later(slowpath_iff); 2689 2690 Node* slowpath_true = new IfTrueNode(slowpath_iff); 2691 transform_later(slowpath_true); 2692 2693 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_value_type_fields_Type(), 2694 StubRoutines::store_value_type_fields_to_buf(), 2695 "store_value_type_fields", 2696 call->jvms()->bci(), 2697 TypePtr::BOTTOM); 2698 slow_call->init_req(TypeFunc::Control, slowpath_true); 2699 slow_call->init_req(TypeFunc::Memory, mem); 2700 slow_call->init_req(TypeFunc::I_O, io); 2701 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr)); 2702 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr)); 2703 slow_call->init_req(TypeFunc::Parms, res); 2704 2705 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control)); 2706 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory)); 2707 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O)); 2708 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms)); 2709 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2)); 2710 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci)); 2711 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci)); 2712 2713 Node* ex_r = new RegionNode(3); 2714 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM); 2715 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO); 2716 ex_r->init_req(1, slow_excp); 2717 ex_mem_phi->init_req(1, slow_mem); 2718 ex_io_phi->init_req(1, slow_io); 2719 ex_r->init_req(2, ex_ctl); 2720 ex_mem_phi->init_req(2, ex_mem); 2721 ex_io_phi->init_req(2, ex_io); 2722 2723 transform_later(ex_r); 2724 transform_later(ex_mem_phi); 2725 transform_later(ex_io_phi); 2726 2727 Node* slowpath_false = new IfFalseNode(slowpath_iff); 2728 transform_later(slowpath_false); 2729 Node* rawmem = new StorePNode(slowpath_false, mem, top_adr, TypeRawPtr::BOTTOM, new_top, MemNode::unordered); 2730 transform_later(rawmem); 2731 Node* mark_node = mark_node = makecon(TypeRawPtr::make((address)markOopDesc::always_locked_prototype())); 2732 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 2733 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 2734 if (UseCompressedClassPointers) { 2735 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT); 2736 } 2737 Node* pack_handler = make_load(slowpath_false, rawmem, klass_node, in_bytes(ValueKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 2738 2739 CallLeafNoFPNode* handler_call = new CallLeafNoFPNode(OptoRuntime::pack_value_type_Type(), 2740 NULL, 2741 "pack handler", 2742 TypeRawPtr::BOTTOM); 2743 handler_call->init_req(TypeFunc::Control, slowpath_false); 2744 handler_call->init_req(TypeFunc::Memory, rawmem); 2745 handler_call->init_req(TypeFunc::I_O, top()); 2746 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr)); 2747 handler_call->init_req(TypeFunc::ReturnAdr, top()); 2748 handler_call->init_req(TypeFunc::Parms, pack_handler); 2749 handler_call->init_req(TypeFunc::Parms+1, old_top); 2750 2751 // We don't know how many values are returned. This assumes the 2752 // worst case, that all available registers are used. 2753 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) { 2754 if (domain->field_at(i) == Type::HALF) { 2755 slow_call->init_req(i, top()); 2756 handler_call->init_req(i+1, top()); 2757 continue; 2758 } 2759 Node* proj = transform_later(new ProjNode(call, i)); 2760 slow_call->init_req(i, proj); 2761 handler_call->init_req(i+1, proj); 2762 } 2763 2764 // We can safepoint at that new call 2765 copy_call_debug_info(call, slow_call); 2766 transform_later(slow_call); 2767 transform_later(handler_call); 2768 2769 Node* handler_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control)); 2770 rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory)); 2771 Node* slowpath_false_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms)); 2772 2773 MergeMemNode* slowpath_false_mem = MergeMemNode::make(mem); 2774 slowpath_false_mem->set_memory_at(Compile::AliasIdxRaw, rawmem); 2775 transform_later(slowpath_false_mem); 2776 2777 Node* r = new RegionNode(4); 2778 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM); 2779 Node* io_phi = new PhiNode(r, Type::ABIO); 2780 Node* res_phi = new PhiNode(r, ret->bottom_type()); 2781 2782 r->init_req(1, no_allocation_ctl); 2783 mem_phi->init_req(1, mem); 2784 io_phi->init_req(1, io); 2785 res_phi->init_req(1, no_allocation_res); 2786 r->init_req(2, slow_norm); 2787 mem_phi->init_req(2, slow_mem); 2788 io_phi->init_req(2, slow_io); 2789 res_phi->init_req(2, slow_res); 2790 r->init_req(3, handler_ctl); 2791 mem_phi->init_req(3, slowpath_false_mem); 2792 io_phi->init_req(3, io); 2793 res_phi->init_req(3, slowpath_false_res); 2794 2795 transform_later(r); 2796 transform_later(mem_phi); 2797 transform_later(io_phi); 2798 transform_later(res_phi); 2799 2800 assert(projs->nb_resproj == 1, "unexpected number of results"); 2801 _igvn.replace_in_uses(projs->fallthrough_catchproj, r); 2802 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi); 2803 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi); 2804 _igvn.replace_in_uses(projs->resproj[0], res_phi); 2805 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r); 2806 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi); 2807 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi); 2808 2809 _igvn.replace_node(ctl, projs->fallthrough_catchproj); 2810 _igvn.replace_node(mem, projs->fallthrough_memproj); 2811 _igvn.replace_node(io, projs->fallthrough_ioproj); 2812 _igvn.replace_node(res, projs->resproj[0]); 2813 _igvn.replace_node(ex_ctl, projs->catchall_catchproj); 2814 _igvn.replace_node(ex_mem, projs->catchall_memproj); 2815 _igvn.replace_node(ex_io, projs->catchall_ioproj); 2816 } 2817 2818 //---------------------------eliminate_macro_nodes---------------------- 2819 // Eliminate scalar replaced allocations and associated locks. 2820 void PhaseMacroExpand::eliminate_macro_nodes() { 2821 if (C->macro_count() == 0) 2822 return; 2823 2824 // First, attempt to eliminate locks 2825 int cnt = C->macro_count(); 2826 for (int i=0; i < cnt; i++) { 2827 Node *n = C->macro_node(i); 2828 if (n->is_AbstractLock()) { // Lock and Unlock nodes 2829 // Before elimination mark all associated (same box and obj) 2830 // lock and unlock nodes. 2831 mark_eliminated_locking_nodes(n->as_AbstractLock()); 2832 } 2833 } 2834 bool progress = true; 2835 while (progress) { 2836 progress = false; 2837 for (int i = C->macro_count(); i > 0; i--) { 2838 Node * n = C->macro_node(i-1); 2839 bool success = false; 2840 debug_only(int old_macro_count = C->macro_count();); 2841 if (n->is_AbstractLock()) { 2842 success = eliminate_locking_node(n->as_AbstractLock()); 2843 } 2844 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2845 progress = progress || success; 2846 } 2847 } 2848 // Next, attempt to eliminate allocations 2849 _has_locks = false; 2850 progress = true; 2851 while (progress) { 2852 progress = false; 2853 for (int i = C->macro_count(); i > 0; i--) { 2854 Node * n = C->macro_node(i-1); 2855 bool success = false; 2856 debug_only(int old_macro_count = C->macro_count();); 2857 switch (n->class_id()) { 2858 case Node::Class_Allocate: 2859 case Node::Class_AllocateArray: 2860 success = eliminate_allocate_node(n->as_Allocate()); 2861 break; 2862 case Node::Class_CallStaticJava: { 2863 CallStaticJavaNode* call = n->as_CallStaticJava(); 2864 if (!call->method()->is_method_handle_intrinsic()) { 2865 success = eliminate_boxing_node(n->as_CallStaticJava()); 2866 } 2867 break; 2868 } 2869 case Node::Class_Lock: 2870 case Node::Class_Unlock: 2871 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 2872 _has_locks = true; 2873 break; 2874 case Node::Class_ArrayCopy: 2875 break; 2876 case Node::Class_OuterStripMinedLoop: 2877 break; 2878 default: 2879 assert(n->Opcode() == Op_LoopLimit || 2880 n->Opcode() == Op_Opaque1 || 2881 n->Opcode() == Op_Opaque2 || 2882 n->Opcode() == Op_Opaque3, "unknown node type in macro list"); 2883 } 2884 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2885 progress = progress || success; 2886 } 2887 } 2888 } 2889 2890 //------------------------------expand_macro_nodes---------------------- 2891 // Returns true if a failure occurred. 2892 bool PhaseMacroExpand::expand_macro_nodes() { 2893 // Last attempt to eliminate macro nodes. 2894 eliminate_macro_nodes(); 2895 2896 // Make sure expansion will not cause node limit to be exceeded. 2897 // Worst case is a macro node gets expanded into about 200 nodes. 2898 // Allow 50% more for optimization. 2899 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) ) 2900 return true; 2901 2902 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. 2903 bool progress = true; 2904 while (progress) { 2905 progress = false; 2906 for (int i = C->macro_count(); i > 0; i--) { 2907 Node * n = C->macro_node(i-1); 2908 bool success = false; 2909 debug_only(int old_macro_count = C->macro_count();); 2910 if (n->Opcode() == Op_LoopLimit) { 2911 // Remove it from macro list and put on IGVN worklist to optimize. 2912 C->remove_macro_node(n); 2913 _igvn._worklist.push(n); 2914 success = true; 2915 } else if (n->Opcode() == Op_CallStaticJava) { 2916 CallStaticJavaNode* call = n->as_CallStaticJava(); 2917 if (!call->method()->is_method_handle_intrinsic()) { 2918 // Remove it from macro list and put on IGVN worklist to optimize. 2919 C->remove_macro_node(n); 2920 _igvn._worklist.push(n); 2921 success = true; 2922 } 2923 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { 2924 _igvn.replace_node(n, n->in(1)); 2925 success = true; 2926 #if INCLUDE_RTM_OPT 2927 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) { 2928 assert(C->profile_rtm(), "should be used only in rtm deoptimization code"); 2929 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), ""); 2930 Node* cmp = n->unique_out(); 2931 #ifdef ASSERT 2932 // Validate graph. 2933 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), ""); 2934 BoolNode* bol = cmp->unique_out()->as_Bool(); 2935 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() && 2936 (bol->_test._test == BoolTest::ne), ""); 2937 IfNode* ifn = bol->unique_out()->as_If(); 2938 assert((ifn->outcnt() == 2) && 2939 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, ""); 2940 #endif 2941 Node* repl = n->in(1); 2942 if (!_has_locks) { 2943 // Remove RTM state check if there are no locks in the code. 2944 // Replace input to compare the same value. 2945 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1); 2946 } 2947 _igvn.replace_node(n, repl); 2948 success = true; 2949 #endif 2950 } else if (n->Opcode() == Op_OuterStripMinedLoop) { 2951 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn); 2952 C->remove_macro_node(n); 2953 success = true; 2954 } 2955 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2956 progress = progress || success; 2957 } 2958 } 2959 2960 // expand arraycopy "macro" nodes first 2961 // For ReduceBulkZeroing, we must first process all arraycopy nodes 2962 // before the allocate nodes are expanded. 2963 int macro_idx = C->macro_count() - 1; 2964 while (macro_idx >= 0) { 2965 Node * n = C->macro_node(macro_idx); 2966 assert(n->is_macro(), "only macro nodes expected here"); 2967 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2968 // node is unreachable, so don't try to expand it 2969 C->remove_macro_node(n); 2970 } else if (n->is_ArrayCopy()){ 2971 int macro_count = C->macro_count(); 2972 expand_arraycopy_node(n->as_ArrayCopy()); 2973 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2974 } 2975 if (C->failing()) return true; 2976 macro_idx --; 2977 } 2978 2979 // expand "macro" nodes 2980 // nodes are removed from the macro list as they are processed 2981 while (C->macro_count() > 0) { 2982 int macro_count = C->macro_count(); 2983 Node * n = C->macro_node(macro_count-1); 2984 assert(n->is_macro(), "only macro nodes expected here"); 2985 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2986 // node is unreachable, so don't try to expand it 2987 C->remove_macro_node(n); 2988 continue; 2989 } 2990 switch (n->class_id()) { 2991 case Node::Class_Allocate: 2992 expand_allocate(n->as_Allocate()); 2993 break; 2994 case Node::Class_AllocateArray: 2995 expand_allocate_array(n->as_AllocateArray()); 2996 break; 2997 case Node::Class_Lock: 2998 expand_lock_node(n->as_Lock()); 2999 break; 3000 case Node::Class_Unlock: 3001 expand_unlock_node(n->as_Unlock()); 3002 break; 3003 case Node::Class_CallStaticJava: 3004 expand_mh_intrinsic_return(n->as_CallStaticJava()); 3005 C->remove_macro_node(n); 3006 break; 3007 default: 3008 assert(false, "unknown node type in macro list"); 3009 } 3010 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 3011 if (C->failing()) return true; 3012 } 3013 3014 _igvn.set_delay_transform(false); 3015 _igvn.optimize(); 3016 if (C->failing()) return true; 3017 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 3018 return bs->expand_macro_nodes(this); 3019 }