1 /* 2 * Copyright (c) 2005, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "gc/shared/collectedHeap.inline.hpp" 28 #include "libadt/vectset.hpp" 29 #include "opto/addnode.hpp" 30 #include "opto/arraycopynode.hpp" 31 #include "opto/callnode.hpp" 32 #include "opto/castnode.hpp" 33 #include "opto/cfgnode.hpp" 34 #include "opto/compile.hpp" 35 #include "opto/convertnode.hpp" 36 #include "opto/graphKit.hpp" 37 #include "opto/locknode.hpp" 38 #include "opto/loopnode.hpp" 39 #include "opto/macro.hpp" 40 #include "opto/memnode.hpp" 41 #include "opto/narrowptrnode.hpp" 42 #include "opto/node.hpp" 43 #include "opto/opaquenode.hpp" 44 #include "opto/phaseX.hpp" 45 #include "opto/rootnode.hpp" 46 #include "opto/runtime.hpp" 47 #include "opto/subnode.hpp" 48 #include "opto/type.hpp" 49 #include "runtime/sharedRuntime.hpp" 50 #if INCLUDE_G1GC 51 #include "gc/g1/g1ThreadLocalData.hpp" 52 #endif // INCLUDE_G1GC 53 54 55 // 56 // Replace any references to "oldref" in inputs to "use" with "newref". 57 // Returns the number of replacements made. 58 // 59 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { 60 int nreplacements = 0; 61 uint req = use->req(); 62 for (uint j = 0; j < use->len(); j++) { 63 Node *uin = use->in(j); 64 if (uin == oldref) { 65 if (j < req) 66 use->set_req(j, newref); 67 else 68 use->set_prec(j, newref); 69 nreplacements++; 70 } else if (j >= req && uin == NULL) { 71 break; 72 } 73 } 74 return nreplacements; 75 } 76 77 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) { 78 // Copy debug information and adjust JVMState information 79 uint old_dbg_start = oldcall->tf()->domain()->cnt(); 80 uint new_dbg_start = newcall->tf()->domain()->cnt(); 81 int jvms_adj = new_dbg_start - old_dbg_start; 82 assert (new_dbg_start == newcall->req(), "argument count mismatch"); 83 84 // SafePointScalarObject node could be referenced several times in debug info. 85 // Use Dict to record cloned nodes. 86 Dict* sosn_map = new Dict(cmpkey,hashkey); 87 for (uint i = old_dbg_start; i < oldcall->req(); i++) { 88 Node* old_in = oldcall->in(i); 89 // Clone old SafePointScalarObjectNodes, adjusting their field contents. 90 if (old_in != NULL && old_in->is_SafePointScalarObject()) { 91 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject(); 92 uint old_unique = C->unique(); 93 Node* new_in = old_sosn->clone(sosn_map); 94 if (old_unique != C->unique()) { // New node? 95 new_in->set_req(0, C->root()); // reset control edge 96 new_in = transform_later(new_in); // Register new node. 97 } 98 old_in = new_in; 99 } 100 newcall->add_req(old_in); 101 } 102 103 // JVMS may be shared so clone it before we modify it 104 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL); 105 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) { 106 jvms->set_map(newcall); 107 jvms->set_locoff(jvms->locoff()+jvms_adj); 108 jvms->set_stkoff(jvms->stkoff()+jvms_adj); 109 jvms->set_monoff(jvms->monoff()+jvms_adj); 110 jvms->set_scloff(jvms->scloff()+jvms_adj); 111 jvms->set_endoff(jvms->endoff()+jvms_adj); 112 } 113 } 114 115 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) { 116 Node* cmp; 117 if (mask != 0) { 118 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask))); 119 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits))); 120 } else { 121 cmp = word; 122 } 123 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne)); 124 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); 125 transform_later(iff); 126 127 // Fast path taken. 128 Node *fast_taken = transform_later(new IfFalseNode(iff)); 129 130 // Fast path not-taken, i.e. slow path 131 Node *slow_taken = transform_later(new IfTrueNode(iff)); 132 133 if (return_fast_path) { 134 region->init_req(edge, slow_taken); // Capture slow-control 135 return fast_taken; 136 } else { 137 region->init_req(edge, fast_taken); // Capture fast-control 138 return slow_taken; 139 } 140 } 141 142 //--------------------copy_predefined_input_for_runtime_call-------------------- 143 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { 144 // Set fixed predefined input arguments 145 call->init_req( TypeFunc::Control, ctrl ); 146 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); 147 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? 148 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); 149 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); 150 } 151 152 //------------------------------make_slow_call--------------------------------- 153 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, 154 address slow_call, const char* leaf_name, Node* slow_path, 155 Node* parm0, Node* parm1, Node* parm2) { 156 157 // Slow-path call 158 CallNode *call = leaf_name 159 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) 160 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM ); 161 162 // Slow path call has no side-effects, uses few values 163 copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); 164 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0); 165 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1); 166 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2); 167 copy_call_debug_info(oldcall, call); 168 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 169 _igvn.replace_node(oldcall, call); 170 transform_later(call); 171 172 return call; 173 } 174 175 void PhaseMacroExpand::extract_call_projections(CallNode *call) { 176 _fallthroughproj = NULL; 177 _fallthroughcatchproj = NULL; 178 _ioproj_fallthrough = NULL; 179 _ioproj_catchall = NULL; 180 _catchallcatchproj = NULL; 181 _memproj_fallthrough = NULL; 182 _memproj_catchall = NULL; 183 _resproj = NULL; 184 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) { 185 ProjNode *pn = call->fast_out(i)->as_Proj(); 186 switch (pn->_con) { 187 case TypeFunc::Control: 188 { 189 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj 190 _fallthroughproj = pn; 191 DUIterator_Fast jmax, j = pn->fast_outs(jmax); 192 const Node *cn = pn->fast_out(j); 193 if (cn->is_Catch()) { 194 ProjNode *cpn = NULL; 195 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { 196 cpn = cn->fast_out(k)->as_Proj(); 197 assert(cpn->is_CatchProj(), "must be a CatchProjNode"); 198 if (cpn->_con == CatchProjNode::fall_through_index) 199 _fallthroughcatchproj = cpn; 200 else { 201 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); 202 _catchallcatchproj = cpn; 203 } 204 } 205 } 206 break; 207 } 208 case TypeFunc::I_O: 209 if (pn->_is_io_use) 210 _ioproj_catchall = pn; 211 else 212 _ioproj_fallthrough = pn; 213 break; 214 case TypeFunc::Memory: 215 if (pn->_is_io_use) 216 _memproj_catchall = pn; 217 else 218 _memproj_fallthrough = pn; 219 break; 220 case TypeFunc::Parms: 221 _resproj = pn; 222 break; 223 default: 224 assert(false, "unexpected projection from allocation node."); 225 } 226 } 227 228 } 229 230 // Eliminate a card mark sequence. p2x is a ConvP2XNode 231 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) { 232 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required"); 233 if (!UseG1GC) { 234 // vanilla/CMS post barrier 235 Node *shift = p2x->unique_out(); 236 Node *addp = shift->unique_out(); 237 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) { 238 Node *mem = addp->last_out(j); 239 if (UseCondCardMark && mem->is_Load()) { 240 assert(mem->Opcode() == Op_LoadB, "unexpected code shape"); 241 // The load is checking if the card has been written so 242 // replace it with zero to fold the test. 243 _igvn.replace_node(mem, intcon(0)); 244 continue; 245 } 246 assert(mem->is_Store(), "store required"); 247 _igvn.replace_node(mem, mem->in(MemNode::Memory)); 248 } 249 } 250 #if INCLUDE_G1GC 251 else { 252 // G1 pre/post barriers 253 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes"); 254 // It could be only one user, URShift node, in Object.clone() intrinsic 255 // but the new allocation is passed to arraycopy stub and it could not 256 // be scalar replaced. So we don't check the case. 257 258 // An other case of only one user (Xor) is when the value check for NULL 259 // in G1 post barrier is folded after CCP so the code which used URShift 260 // is removed. 261 262 // Take Region node before eliminating post barrier since it also 263 // eliminates CastP2X node when it has only one user. 264 Node* this_region = p2x->in(0); 265 assert(this_region != NULL, ""); 266 267 // Remove G1 post barrier. 268 269 // Search for CastP2X->Xor->URShift->Cmp path which 270 // checks if the store done to a different from the value's region. 271 // And replace Cmp with #0 (false) to collapse G1 post barrier. 272 Node* xorx = p2x->find_out_with(Op_XorX); 273 if (xorx != NULL) { 274 Node* shift = xorx->unique_out(); 275 Node* cmpx = shift->unique_out(); 276 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() && 277 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne, 278 "missing region check in G1 post barrier"); 279 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 280 281 // Remove G1 pre barrier. 282 283 // Search "if (marking != 0)" check and set it to "false". 284 // There is no G1 pre barrier if previous stored value is NULL 285 // (for example, after initialization). 286 if (this_region->is_Region() && this_region->req() == 3) { 287 int ind = 1; 288 if (!this_region->in(ind)->is_IfFalse()) { 289 ind = 2; 290 } 291 if (this_region->in(ind)->is_IfFalse() && 292 this_region->in(ind)->in(0)->Opcode() == Op_If) { 293 Node* bol = this_region->in(ind)->in(0)->in(1); 294 assert(bol->is_Bool(), ""); 295 cmpx = bol->in(1); 296 if (bol->as_Bool()->_test._test == BoolTest::ne && 297 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) && 298 cmpx->in(1)->is_Load()) { 299 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address); 300 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()); 301 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() && 302 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 303 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) { 304 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 305 } 306 } 307 } 308 } 309 } else { 310 assert(!GraphKit::use_ReduceInitialCardMarks(), "can only happen with card marking"); 311 // This is a G1 post barrier emitted by the Object.clone() intrinsic. 312 // Search for the CastP2X->URShiftX->AddP->LoadB->Cmp path which checks if the card 313 // is marked as young_gen and replace the Cmp with 0 (false) to collapse the barrier. 314 Node* shift = p2x->find_out_with(Op_URShiftX); 315 assert(shift != NULL, "missing G1 post barrier"); 316 Node* addp = shift->unique_out(); 317 Node* load = addp->find_out_with(Op_LoadB); 318 assert(load != NULL, "missing G1 post barrier"); 319 Node* cmpx = load->unique_out(); 320 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() && 321 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne, 322 "missing card value check in G1 post barrier"); 323 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 324 // There is no G1 pre barrier in this case 325 } 326 // Now CastP2X can be removed since it is used only on dead path 327 // which currently still alive until igvn optimize it. 328 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, ""); 329 _igvn.replace_node(p2x, top()); 330 } 331 #endif // INCLUDE_G1GC 332 } 333 334 // Search for a memory operation for the specified memory slice. 335 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) { 336 Node *orig_mem = mem; 337 Node *alloc_mem = alloc->in(TypeFunc::Memory); 338 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr(); 339 while (true) { 340 if (mem == alloc_mem || mem == start_mem ) { 341 return mem; // hit one of our sentinels 342 } else if (mem->is_MergeMem()) { 343 mem = mem->as_MergeMem()->memory_at(alias_idx); 344 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { 345 Node *in = mem->in(0); 346 // we can safely skip over safepoints, calls, locks and membars because we 347 // already know that the object is safe to eliminate. 348 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { 349 return in; 350 } else if (in->is_Call()) { 351 CallNode *call = in->as_Call(); 352 if (call->may_modify(tinst, phase)) { 353 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape"); 354 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) { 355 return in; 356 } 357 } 358 mem = in->in(TypeFunc::Memory); 359 } else if (in->is_MemBar()) { 360 ArrayCopyNode* ac = NULL; 361 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) { 362 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone"); 363 return ac; 364 } 365 mem = in->in(TypeFunc::Memory); 366 } else { 367 assert(false, "unexpected projection"); 368 } 369 } else if (mem->is_Store()) { 370 const TypePtr* atype = mem->as_Store()->adr_type(); 371 int adr_idx = phase->C->get_alias_index(atype); 372 if (adr_idx == alias_idx) { 373 assert(atype->isa_oopptr(), "address type must be oopptr"); 374 int adr_offset = atype->offset(); 375 uint adr_iid = atype->is_oopptr()->instance_id(); 376 // Array elements references have the same alias_idx 377 // but different offset and different instance_id. 378 if (adr_offset == offset && adr_iid == alloc->_idx) 379 return mem; 380 } else { 381 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw"); 382 } 383 mem = mem->in(MemNode::Memory); 384 } else if (mem->is_ClearArray()) { 385 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) { 386 // Can not bypass initialization of the instance 387 // we are looking. 388 debug_only(intptr_t offset;) 389 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity"); 390 InitializeNode* init = alloc->as_Allocate()->initialization(); 391 // We are looking for stored value, return Initialize node 392 // or memory edge from Allocate node. 393 if (init != NULL) 394 return init; 395 else 396 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers). 397 } 398 // Otherwise skip it (the call updated 'mem' value). 399 } else if (mem->Opcode() == Op_SCMemProj) { 400 mem = mem->in(0); 401 Node* adr = NULL; 402 if (mem->is_LoadStore()) { 403 adr = mem->in(MemNode::Address); 404 } else { 405 assert(mem->Opcode() == Op_EncodeISOArray || 406 mem->Opcode() == Op_StrCompressedCopy, "sanity"); 407 adr = mem->in(3); // Destination array 408 } 409 const TypePtr* atype = adr->bottom_type()->is_ptr(); 410 int adr_idx = phase->C->get_alias_index(atype); 411 if (adr_idx == alias_idx) { 412 DEBUG_ONLY(mem->dump();) 413 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 414 return NULL; 415 } 416 mem = mem->in(MemNode::Memory); 417 } else if (mem->Opcode() == Op_StrInflatedCopy) { 418 Node* adr = mem->in(3); // Destination array 419 const TypePtr* atype = adr->bottom_type()->is_ptr(); 420 int adr_idx = phase->C->get_alias_index(atype); 421 if (adr_idx == alias_idx) { 422 DEBUG_ONLY(mem->dump();) 423 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field"); 424 return NULL; 425 } 426 mem = mem->in(MemNode::Memory); 427 } else { 428 return mem; 429 } 430 assert(mem != orig_mem, "dead memory loop"); 431 } 432 } 433 434 // Generate loads from source of the arraycopy for fields of 435 // destination needed at a deoptimization point 436 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) { 437 BasicType bt = ft; 438 const Type *type = ftype; 439 if (ft == T_NARROWOOP) { 440 bt = T_OBJECT; 441 type = ftype->make_oopptr(); 442 } 443 Node* res = NULL; 444 if (ac->is_clonebasic()) { 445 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base); 446 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset))); 447 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset); 448 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned); 449 } else { 450 if (ac->modifies(offset, offset, &_igvn, true)) { 451 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result"); 452 uint shift = exact_log2(type2aelembytes(bt)); 453 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 454 #ifdef _LP64 455 diff = _igvn.transform(new ConvI2LNode(diff)); 456 #endif 457 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift))); 458 459 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff)); 460 Node* base = ac->in(ArrayCopyNode::Src); 461 Node* adr = _igvn.transform(new AddPNode(base, base, off)); 462 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset); 463 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned); 464 } 465 } 466 if (res != NULL) { 467 res = _igvn.transform(res); 468 if (ftype->isa_narrowoop()) { 469 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes 470 res = _igvn.transform(new EncodePNode(res, ftype)); 471 } 472 return res; 473 } 474 return NULL; 475 } 476 477 // 478 // Given a Memory Phi, compute a value Phi containing the values from stores 479 // on the input paths. 480 // Note: this function is recursive, its depth is limited by the "level" argument 481 // Returns the computed Phi, or NULL if it cannot compute it. 482 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) { 483 assert(mem->is_Phi(), "sanity"); 484 int alias_idx = C->get_alias_index(adr_t); 485 int offset = adr_t->offset(); 486 int instance_id = adr_t->instance_id(); 487 488 // Check if an appropriate value phi already exists. 489 Node* region = mem->in(0); 490 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { 491 Node* phi = region->fast_out(k); 492 if (phi->is_Phi() && phi != mem && 493 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) { 494 return phi; 495 } 496 } 497 // Check if an appropriate new value phi already exists. 498 Node* new_phi = value_phis->find(mem->_idx); 499 if (new_phi != NULL) 500 return new_phi; 501 502 if (level <= 0) { 503 return NULL; // Give up: phi tree too deep 504 } 505 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 506 Node *alloc_mem = alloc->in(TypeFunc::Memory); 507 508 uint length = mem->req(); 509 GrowableArray <Node *> values(length, length, NULL, false); 510 511 // create a new Phi for the value 512 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset); 513 transform_later(phi); 514 value_phis->push(phi, mem->_idx); 515 516 for (uint j = 1; j < length; j++) { 517 Node *in = mem->in(j); 518 if (in == NULL || in->is_top()) { 519 values.at_put(j, in); 520 } else { 521 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); 522 if (val == start_mem || val == alloc_mem) { 523 // hit a sentinel, return appropriate 0 value 524 values.at_put(j, _igvn.zerocon(ft)); 525 continue; 526 } 527 if (val->is_Initialize()) { 528 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 529 } 530 if (val == NULL) { 531 return NULL; // can't find a value on this path 532 } 533 if (val == mem) { 534 values.at_put(j, mem); 535 } else if (val->is_Store()) { 536 values.at_put(j, val->in(MemNode::ValueIn)); 537 } else if(val->is_Proj() && val->in(0) == alloc) { 538 values.at_put(j, _igvn.zerocon(ft)); 539 } else if (val->is_Phi()) { 540 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); 541 if (val == NULL) { 542 return NULL; 543 } 544 values.at_put(j, val); 545 } else if (val->Opcode() == Op_SCMemProj) { 546 assert(val->in(0)->is_LoadStore() || 547 val->in(0)->Opcode() == Op_EncodeISOArray || 548 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity"); 549 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 550 return NULL; 551 } else if (val->is_ArrayCopy()) { 552 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc); 553 if (res == NULL) { 554 return NULL; 555 } 556 values.at_put(j, res); 557 } else { 558 #ifdef ASSERT 559 val->dump(); 560 assert(false, "unknown node on this path"); 561 #endif 562 return NULL; // unknown node on this path 563 } 564 } 565 } 566 // Set Phi's inputs 567 for (uint j = 1; j < length; j++) { 568 if (values.at(j) == mem) { 569 phi->init_req(j, phi); 570 } else { 571 phi->init_req(j, values.at(j)); 572 } 573 } 574 return phi; 575 } 576 577 // Search the last value stored into the object's field. 578 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) { 579 assert(adr_t->is_known_instance_field(), "instance required"); 580 int instance_id = adr_t->instance_id(); 581 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 582 583 int alias_idx = C->get_alias_index(adr_t); 584 int offset = adr_t->offset(); 585 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); 586 Node *alloc_ctrl = alloc->in(TypeFunc::Control); 587 Node *alloc_mem = alloc->in(TypeFunc::Memory); 588 Arena *a = Thread::current()->resource_area(); 589 VectorSet visited(a); 590 591 592 bool done = sfpt_mem == alloc_mem; 593 Node *mem = sfpt_mem; 594 while (!done) { 595 if (visited.test_set(mem->_idx)) { 596 return NULL; // found a loop, give up 597 } 598 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 599 if (mem == start_mem || mem == alloc_mem) { 600 done = true; // hit a sentinel, return appropriate 0 value 601 } else if (mem->is_Initialize()) { 602 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 603 if (mem == NULL) { 604 done = true; // Something go wrong. 605 } else if (mem->is_Store()) { 606 const TypePtr* atype = mem->as_Store()->adr_type(); 607 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 608 done = true; 609 } 610 } else if (mem->is_Store()) { 611 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 612 assert(atype != NULL, "address type must be oopptr"); 613 assert(C->get_alias_index(atype) == alias_idx && 614 atype->is_known_instance_field() && atype->offset() == offset && 615 atype->instance_id() == instance_id, "store is correct memory slice"); 616 done = true; 617 } else if (mem->is_Phi()) { 618 // try to find a phi's unique input 619 Node *unique_input = NULL; 620 Node *top = C->top(); 621 for (uint i = 1; i < mem->req(); i++) { 622 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 623 if (n == NULL || n == top || n == mem) { 624 continue; 625 } else if (unique_input == NULL) { 626 unique_input = n; 627 } else if (unique_input != n) { 628 unique_input = top; 629 break; 630 } 631 } 632 if (unique_input != NULL && unique_input != top) { 633 mem = unique_input; 634 } else { 635 done = true; 636 } 637 } else if (mem->is_ArrayCopy()) { 638 done = true; 639 } else { 640 assert(false, "unexpected node"); 641 } 642 } 643 if (mem != NULL) { 644 if (mem == start_mem || mem == alloc_mem) { 645 // hit a sentinel, return appropriate 0 value 646 return _igvn.zerocon(ft); 647 } else if (mem->is_Store()) { 648 return mem->in(MemNode::ValueIn); 649 } else if (mem->is_Phi()) { 650 // attempt to produce a Phi reflecting the values on the input paths of the Phi 651 Node_Stack value_phis(a, 8); 652 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 653 if (phi != NULL) { 654 return phi; 655 } else { 656 // Kill all new Phis 657 while(value_phis.is_nonempty()) { 658 Node* n = value_phis.node(); 659 _igvn.replace_node(n, C->top()); 660 value_phis.pop(); 661 } 662 } 663 } else if (mem->is_ArrayCopy()) { 664 Node* ctl = mem->in(0); 665 Node* m = mem->in(TypeFunc::Memory); 666 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) { 667 // pin the loads in the uncommon trap path 668 ctl = sfpt_ctl; 669 m = sfpt_mem; 670 } 671 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc); 672 } 673 } 674 // Something go wrong. 675 return NULL; 676 } 677 678 // Check the possibility of scalar replacement. 679 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 680 // Scan the uses of the allocation to check for anything that would 681 // prevent us from eliminating it. 682 NOT_PRODUCT( const char* fail_eliminate = NULL; ) 683 DEBUG_ONLY( Node* disq_node = NULL; ) 684 bool can_eliminate = true; 685 686 Node* res = alloc->result_cast(); 687 const TypeOopPtr* res_type = NULL; 688 if (res == NULL) { 689 // All users were eliminated. 690 } else if (!res->is_CheckCastPP()) { 691 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 692 can_eliminate = false; 693 } else { 694 res_type = _igvn.type(res)->isa_oopptr(); 695 if (res_type == NULL) { 696 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 697 can_eliminate = false; 698 } else if (res_type->isa_aryptr()) { 699 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 700 if (length < 0) { 701 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 702 can_eliminate = false; 703 } 704 } 705 } 706 707 if (can_eliminate && res != NULL) { 708 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 709 j < jmax && can_eliminate; j++) { 710 Node* use = res->fast_out(j); 711 712 if (use->is_AddP()) { 713 const TypePtr* addp_type = _igvn.type(use)->is_ptr(); 714 int offset = addp_type->offset(); 715 716 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 717 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) 718 can_eliminate = false; 719 break; 720 } 721 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 722 k < kmax && can_eliminate; k++) { 723 Node* n = use->fast_out(k); 724 if (!n->is_Store() && n->Opcode() != Op_CastP2X && 725 !(n->is_ArrayCopy() && 726 n->as_ArrayCopy()->is_clonebasic() && 727 n->in(ArrayCopyNode::Dest) == use)) { 728 DEBUG_ONLY(disq_node = n;) 729 if (n->is_Load() || n->is_LoadStore()) { 730 NOT_PRODUCT(fail_eliminate = "Field load";) 731 } else { 732 NOT_PRODUCT(fail_eliminate = "Not store field referrence";) 733 } 734 can_eliminate = false; 735 } 736 } 737 } else if (use->is_ArrayCopy() && 738 (use->as_ArrayCopy()->is_arraycopy_validated() || 739 use->as_ArrayCopy()->is_copyof_validated() || 740 use->as_ArrayCopy()->is_copyofrange_validated()) && 741 use->in(ArrayCopyNode::Dest) == res) { 742 // ok to eliminate 743 } else if (use->is_SafePoint()) { 744 SafePointNode* sfpt = use->as_SafePoint(); 745 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 746 // Object is passed as argument. 747 DEBUG_ONLY(disq_node = use;) 748 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 749 can_eliminate = false; 750 } 751 Node* sfptMem = sfpt->memory(); 752 if (sfptMem == NULL || sfptMem->is_top()) { 753 DEBUG_ONLY(disq_node = use;) 754 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) 755 can_eliminate = false; 756 } else { 757 safepoints.append_if_missing(sfpt); 758 } 759 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 760 if (use->is_Phi()) { 761 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 762 NOT_PRODUCT(fail_eliminate = "Object is return value";) 763 } else { 764 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 765 } 766 DEBUG_ONLY(disq_node = use;) 767 } else { 768 if (use->Opcode() == Op_Return) { 769 NOT_PRODUCT(fail_eliminate = "Object is return value";) 770 }else { 771 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 772 } 773 DEBUG_ONLY(disq_node = use;) 774 } 775 can_eliminate = false; 776 } 777 } 778 } 779 780 #ifndef PRODUCT 781 if (PrintEliminateAllocations) { 782 if (can_eliminate) { 783 tty->print("Scalar "); 784 if (res == NULL) 785 alloc->dump(); 786 else 787 res->dump(); 788 } else if (alloc->_is_scalar_replaceable) { 789 tty->print("NotScalar (%s)", fail_eliminate); 790 if (res == NULL) 791 alloc->dump(); 792 else 793 res->dump(); 794 #ifdef ASSERT 795 if (disq_node != NULL) { 796 tty->print(" >>>> "); 797 disq_node->dump(); 798 } 799 #endif /*ASSERT*/ 800 } 801 } 802 #endif 803 return can_eliminate; 804 } 805 806 // Do scalar replacement. 807 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 808 GrowableArray <SafePointNode *> safepoints_done; 809 810 ciKlass* klass = NULL; 811 ciInstanceKlass* iklass = NULL; 812 int nfields = 0; 813 int array_base = 0; 814 int element_size = 0; 815 BasicType basic_elem_type = T_ILLEGAL; 816 ciType* elem_type = NULL; 817 818 Node* res = alloc->result_cast(); 819 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result"); 820 const TypeOopPtr* res_type = NULL; 821 if (res != NULL) { // Could be NULL when there are no users 822 res_type = _igvn.type(res)->isa_oopptr(); 823 } 824 825 if (res != NULL) { 826 klass = res_type->klass(); 827 if (res_type->isa_instptr()) { 828 // find the fields of the class which will be needed for safepoint debug information 829 assert(klass->is_instance_klass(), "must be an instance klass."); 830 iklass = klass->as_instance_klass(); 831 nfields = iklass->nof_nonstatic_fields(); 832 } else { 833 // find the array's elements which will be needed for safepoint debug information 834 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 835 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); 836 elem_type = klass->as_array_klass()->element_type(); 837 basic_elem_type = elem_type->basic_type(); 838 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 839 element_size = type2aelembytes(basic_elem_type); 840 } 841 } 842 // 843 // Process the safepoint uses 844 // 845 while (safepoints.length() > 0) { 846 SafePointNode* sfpt = safepoints.pop(); 847 Node* mem = sfpt->memory(); 848 Node* ctl = sfpt->control(); 849 assert(sfpt->jvms() != NULL, "missed JVMS"); 850 // Fields of scalar objs are referenced only at the end 851 // of regular debuginfo at the last (youngest) JVMS. 852 // Record relative start index. 853 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff()); 854 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, 855 #ifdef ASSERT 856 alloc, 857 #endif 858 first_ind, nfields); 859 sobj->init_req(0, C->root()); 860 transform_later(sobj); 861 862 // Scan object's fields adding an input to the safepoint for each field. 863 for (int j = 0; j < nfields; j++) { 864 intptr_t offset; 865 ciField* field = NULL; 866 if (iklass != NULL) { 867 field = iklass->nonstatic_field_at(j); 868 offset = field->offset(); 869 elem_type = field->type(); 870 basic_elem_type = field->layout_type(); 871 } else { 872 offset = array_base + j * (intptr_t)element_size; 873 } 874 875 const Type *field_type; 876 // The next code is taken from Parse::do_get_xxx(). 877 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { 878 if (!elem_type->is_loaded()) { 879 field_type = TypeInstPtr::BOTTOM; 880 } else if (field != NULL && field->is_static_constant()) { 881 // This can happen if the constant oop is non-perm. 882 ciObject* con = field->constant_value().as_object(); 883 // Do not "join" in the previous type; it doesn't add value, 884 // and may yield a vacuous result if the field is of interface type. 885 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 886 assert(field_type != NULL, "field singleton type must be consistent"); 887 } else { 888 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 889 } 890 if (UseCompressedOops) { 891 field_type = field_type->make_narrowoop(); 892 basic_elem_type = T_NARROWOOP; 893 } 894 } else { 895 field_type = Type::get_const_basic_type(basic_elem_type); 896 } 897 898 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 899 900 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc); 901 if (field_val == NULL) { 902 // We weren't able to find a value for this field, 903 // give up on eliminating this allocation. 904 905 // Remove any extra entries we added to the safepoint. 906 uint last = sfpt->req() - 1; 907 for (int k = 0; k < j; k++) { 908 sfpt->del_req(last--); 909 } 910 _igvn._worklist.push(sfpt); 911 // rollback processed safepoints 912 while (safepoints_done.length() > 0) { 913 SafePointNode* sfpt_done = safepoints_done.pop(); 914 // remove any extra entries we added to the safepoint 915 last = sfpt_done->req() - 1; 916 for (int k = 0; k < nfields; k++) { 917 sfpt_done->del_req(last--); 918 } 919 JVMState *jvms = sfpt_done->jvms(); 920 jvms->set_endoff(sfpt_done->req()); 921 // Now make a pass over the debug information replacing any references 922 // to SafePointScalarObjectNode with the allocated object. 923 int start = jvms->debug_start(); 924 int end = jvms->debug_end(); 925 for (int i = start; i < end; i++) { 926 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 927 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 928 if (scobj->first_index(jvms) == sfpt_done->req() && 929 scobj->n_fields() == (uint)nfields) { 930 assert(scobj->alloc() == alloc, "sanity"); 931 sfpt_done->set_req(i, res); 932 } 933 } 934 } 935 _igvn._worklist.push(sfpt_done); 936 } 937 #ifndef PRODUCT 938 if (PrintEliminateAllocations) { 939 if (field != NULL) { 940 tty->print("=== At SafePoint node %d can't find value of Field: ", 941 sfpt->_idx); 942 field->print(); 943 int field_idx = C->get_alias_index(field_addr_type); 944 tty->print(" (alias_idx=%d)", field_idx); 945 } else { // Array's element 946 tty->print("=== At SafePoint node %d can't find value of array element [%d]", 947 sfpt->_idx, j); 948 } 949 tty->print(", which prevents elimination of: "); 950 if (res == NULL) 951 alloc->dump(); 952 else 953 res->dump(); 954 } 955 #endif 956 return false; 957 } 958 if (UseCompressedOops && field_type->isa_narrowoop()) { 959 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 960 // to be able scalar replace the allocation. 961 if (field_val->is_EncodeP()) { 962 field_val = field_val->in(1); 963 } else { 964 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type())); 965 } 966 } 967 sfpt->add_req(field_val); 968 } 969 JVMState *jvms = sfpt->jvms(); 970 jvms->set_endoff(sfpt->req()); 971 // Now make a pass over the debug information replacing any references 972 // to the allocated object with "sobj" 973 int start = jvms->debug_start(); 974 int end = jvms->debug_end(); 975 sfpt->replace_edges_in_range(res, sobj, start, end); 976 _igvn._worklist.push(sfpt); 977 safepoints_done.append_if_missing(sfpt); // keep it for rollback 978 } 979 return true; 980 } 981 982 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) { 983 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control); 984 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory); 985 if (ctl_proj != NULL) { 986 igvn.replace_node(ctl_proj, n->in(0)); 987 } 988 if (mem_proj != NULL) { 989 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory)); 990 } 991 } 992 993 // Process users of eliminated allocation. 994 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) { 995 Node* res = alloc->result_cast(); 996 if (res != NULL) { 997 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 998 Node *use = res->last_out(j); 999 uint oc1 = res->outcnt(); 1000 1001 if (use->is_AddP()) { 1002 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 1003 Node *n = use->last_out(k); 1004 uint oc2 = use->outcnt(); 1005 if (n->is_Store()) { 1006 #ifdef ASSERT 1007 // Verify that there is no dependent MemBarVolatile nodes, 1008 // they should be removed during IGVN, see MemBarNode::Ideal(). 1009 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); 1010 p < pmax; p++) { 1011 Node* mb = n->fast_out(p); 1012 assert(mb->is_Initialize() || !mb->is_MemBar() || 1013 mb->req() <= MemBarNode::Precedent || 1014 mb->in(MemBarNode::Precedent) != n, 1015 "MemBarVolatile should be eliminated for non-escaping object"); 1016 } 1017 #endif 1018 _igvn.replace_node(n, n->in(MemNode::Memory)); 1019 } else if (n->is_ArrayCopy()) { 1020 // Disconnect ArrayCopy node 1021 ArrayCopyNode* ac = n->as_ArrayCopy(); 1022 assert(ac->is_clonebasic(), "unexpected array copy kind"); 1023 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out(); 1024 disconnect_projections(ac, _igvn); 1025 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation"); 1026 Node* membar_before = alloc->in(0)->in(0); 1027 disconnect_projections(membar_before->as_MemBar(), _igvn); 1028 if (membar_after->is_MemBar()) { 1029 disconnect_projections(membar_after->as_MemBar(), _igvn); 1030 } 1031 } else { 1032 eliminate_card_mark(n); 1033 } 1034 k -= (oc2 - use->outcnt()); 1035 } 1036 } else if (use->is_ArrayCopy()) { 1037 // Disconnect ArrayCopy node 1038 ArrayCopyNode* ac = use->as_ArrayCopy(); 1039 assert(ac->is_arraycopy_validated() || 1040 ac->is_copyof_validated() || 1041 ac->is_copyofrange_validated(), "unsupported"); 1042 CallProjections callprojs; 1043 ac->extract_projections(&callprojs, true); 1044 1045 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O)); 1046 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory)); 1047 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control)); 1048 1049 // Set control to top. IGVN will remove the remaining projections 1050 ac->set_req(0, top()); 1051 ac->replace_edge(res, top()); 1052 1053 // Disconnect src right away: it can help find new 1054 // opportunities for allocation elimination 1055 Node* src = ac->in(ArrayCopyNode::Src); 1056 ac->replace_edge(src, top()); 1057 // src can be top at this point if src and dest of the 1058 // arraycopy were the same 1059 if (src->outcnt() == 0 && !src->is_top()) { 1060 _igvn.remove_dead_node(src); 1061 } 1062 1063 _igvn._worklist.push(ac); 1064 } else { 1065 eliminate_card_mark(use); 1066 } 1067 j -= (oc1 - res->outcnt()); 1068 } 1069 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 1070 _igvn.remove_dead_node(res); 1071 } 1072 1073 // 1074 // Process other users of allocation's projections 1075 // 1076 if (_resproj != NULL && _resproj->outcnt() != 0) { 1077 // First disconnect stores captured by Initialize node. 1078 // If Initialize node is eliminated first in the following code, 1079 // it will kill such stores and DUIterator_Last will assert. 1080 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) { 1081 Node *use = _resproj->fast_out(j); 1082 if (use->is_AddP()) { 1083 // raw memory addresses used only by the initialization 1084 _igvn.replace_node(use, C->top()); 1085 --j; --jmax; 1086 } 1087 } 1088 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { 1089 Node *use = _resproj->last_out(j); 1090 uint oc1 = _resproj->outcnt(); 1091 if (use->is_Initialize()) { 1092 // Eliminate Initialize node. 1093 InitializeNode *init = use->as_Initialize(); 1094 assert(init->outcnt() <= 2, "only a control and memory projection expected"); 1095 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control); 1096 if (ctrl_proj != NULL) { 1097 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection"); 1098 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj); 1099 } 1100 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory); 1101 if (mem_proj != NULL) { 1102 Node *mem = init->in(TypeFunc::Memory); 1103 #ifdef ASSERT 1104 if (mem->is_MergeMem()) { 1105 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection"); 1106 } else { 1107 assert(mem == _memproj_fallthrough, "allocation memory projection"); 1108 } 1109 #endif 1110 _igvn.replace_node(mem_proj, mem); 1111 } 1112 } else { 1113 assert(false, "only Initialize or AddP expected"); 1114 } 1115 j -= (oc1 - _resproj->outcnt()); 1116 } 1117 } 1118 if (_fallthroughcatchproj != NULL) { 1119 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); 1120 } 1121 if (_memproj_fallthrough != NULL) { 1122 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); 1123 } 1124 if (_memproj_catchall != NULL) { 1125 _igvn.replace_node(_memproj_catchall, C->top()); 1126 } 1127 if (_ioproj_fallthrough != NULL) { 1128 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); 1129 } 1130 if (_ioproj_catchall != NULL) { 1131 _igvn.replace_node(_ioproj_catchall, C->top()); 1132 } 1133 if (_catchallcatchproj != NULL) { 1134 _igvn.replace_node(_catchallcatchproj, C->top()); 1135 } 1136 } 1137 1138 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 1139 // Don't do scalar replacement if the frame can be popped by JVMTI: 1140 // if reallocation fails during deoptimization we'll pop all 1141 // interpreter frames for this compiled frame and that won't play 1142 // nice with JVMTI popframe. 1143 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) { 1144 return false; 1145 } 1146 Node* klass = alloc->in(AllocateNode::KlassNode); 1147 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); 1148 Node* res = alloc->result_cast(); 1149 // Eliminate boxing allocations which are not used 1150 // regardless scalar replacable status. 1151 bool boxing_alloc = C->eliminate_boxing() && 1152 tklass->klass()->is_instance_klass() && 1153 tklass->klass()->as_instance_klass()->is_box_klass(); 1154 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) { 1155 return false; 1156 } 1157 1158 extract_call_projections(alloc); 1159 1160 GrowableArray <SafePointNode *> safepoints; 1161 if (!can_eliminate_allocation(alloc, safepoints)) { 1162 return false; 1163 } 1164 1165 if (!alloc->_is_scalar_replaceable) { 1166 assert(res == NULL, "sanity"); 1167 // We can only eliminate allocation if all debug info references 1168 // are already replaced with SafePointScalarObject because 1169 // we can't search for a fields value without instance_id. 1170 if (safepoints.length() > 0) { 1171 return false; 1172 } 1173 } 1174 1175 if (!scalar_replacement(alloc, safepoints)) { 1176 return false; 1177 } 1178 1179 CompileLog* log = C->log(); 1180 if (log != NULL) { 1181 log->head("eliminate_allocation type='%d'", 1182 log->identify(tklass->klass())); 1183 JVMState* p = alloc->jvms(); 1184 while (p != NULL) { 1185 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1186 p = p->caller(); 1187 } 1188 log->tail("eliminate_allocation"); 1189 } 1190 1191 process_users_of_allocation(alloc); 1192 1193 #ifndef PRODUCT 1194 if (PrintEliminateAllocations) { 1195 if (alloc->is_AllocateArray()) 1196 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1197 else 1198 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1199 } 1200 #endif 1201 1202 return true; 1203 } 1204 1205 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) { 1206 // EA should remove all uses of non-escaping boxing node. 1207 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) { 1208 return false; 1209 } 1210 1211 assert(boxing->result_cast() == NULL, "unexpected boxing node result"); 1212 1213 extract_call_projections(boxing); 1214 1215 const TypeTuple* r = boxing->tf()->range(); 1216 assert(r->cnt() > TypeFunc::Parms, "sanity"); 1217 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr(); 1218 assert(t != NULL, "sanity"); 1219 1220 CompileLog* log = C->log(); 1221 if (log != NULL) { 1222 log->head("eliminate_boxing type='%d'", 1223 log->identify(t->klass())); 1224 JVMState* p = boxing->jvms(); 1225 while (p != NULL) { 1226 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1227 p = p->caller(); 1228 } 1229 log->tail("eliminate_boxing"); 1230 } 1231 1232 process_users_of_allocation(boxing); 1233 1234 #ifndef PRODUCT 1235 if (PrintEliminateAllocations) { 1236 tty->print("++++ Eliminated: %d ", boxing->_idx); 1237 boxing->method()->print_short_name(tty); 1238 tty->cr(); 1239 } 1240 #endif 1241 1242 return true; 1243 } 1244 1245 //---------------------------set_eden_pointers------------------------- 1246 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { 1247 if (UseTLAB) { // Private allocation: load from TLS 1248 Node* thread = transform_later(new ThreadLocalNode()); 1249 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); 1250 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); 1251 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); 1252 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); 1253 } else { // Shared allocation: load from globals 1254 CollectedHeap* ch = Universe::heap(); 1255 address top_adr = (address)ch->top_addr(); 1256 address end_adr = (address)ch->end_addr(); 1257 eden_top_adr = makecon(TypeRawPtr::make(top_adr)); 1258 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); 1259 } 1260 } 1261 1262 1263 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 1264 Node* adr = basic_plus_adr(base, offset); 1265 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 1266 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered); 1267 transform_later(value); 1268 return value; 1269 } 1270 1271 1272 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 1273 Node* adr = basic_plus_adr(base, offset); 1274 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered); 1275 transform_later(mem); 1276 return mem; 1277 } 1278 1279 //============================================================================= 1280 // 1281 // A L L O C A T I O N 1282 // 1283 // Allocation attempts to be fast in the case of frequent small objects. 1284 // It breaks down like this: 1285 // 1286 // 1) Size in doublewords is computed. This is a constant for objects and 1287 // variable for most arrays. Doubleword units are used to avoid size 1288 // overflow of huge doubleword arrays. We need doublewords in the end for 1289 // rounding. 1290 // 1291 // 2) Size is checked for being 'too large'. Too-large allocations will go 1292 // the slow path into the VM. The slow path can throw any required 1293 // exceptions, and does all the special checks for very large arrays. The 1294 // size test can constant-fold away for objects. For objects with 1295 // finalizers it constant-folds the otherway: you always go slow with 1296 // finalizers. 1297 // 1298 // 3) If NOT using TLABs, this is the contended loop-back point. 1299 // Load-Locked the heap top. If using TLABs normal-load the heap top. 1300 // 1301 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 1302 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 1303 // "size*8" we always enter the VM, where "largish" is a constant picked small 1304 // enough that there's always space between the eden max and 4Gig (old space is 1305 // there so it's quite large) and large enough that the cost of entering the VM 1306 // is dwarfed by the cost to initialize the space. 1307 // 1308 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 1309 // down. If contended, repeat at step 3. If using TLABs normal-store 1310 // adjusted heap top back down; there is no contention. 1311 // 1312 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 1313 // fields. 1314 // 1315 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 1316 // oop flavor. 1317 // 1318 //============================================================================= 1319 // FastAllocateSizeLimit value is in DOUBLEWORDS. 1320 // Allocations bigger than this always go the slow route. 1321 // This value must be small enough that allocation attempts that need to 1322 // trigger exceptions go the slow route. Also, it must be small enough so 1323 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 1324 //=============================================================================j// 1325 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 1326 // The allocator will coalesce int->oop copies away. See comment in 1327 // coalesce.cpp about how this works. It depends critically on the exact 1328 // code shape produced here, so if you are changing this code shape 1329 // make sure the GC info for the heap-top is correct in and around the 1330 // slow-path call. 1331 // 1332 1333 void PhaseMacroExpand::expand_allocate_common( 1334 AllocateNode* alloc, // allocation node to be expanded 1335 Node* length, // array length for an array allocation 1336 const TypeFunc* slow_call_type, // Type of slow call 1337 address slow_call_address // Address of slow call 1338 ) 1339 { 1340 1341 Node* ctrl = alloc->in(TypeFunc::Control); 1342 Node* mem = alloc->in(TypeFunc::Memory); 1343 Node* i_o = alloc->in(TypeFunc::I_O); 1344 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 1345 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1346 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 1347 1348 assert(ctrl != NULL, "must have control"); 1349 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 1350 // they will not be used if "always_slow" is set 1351 enum { slow_result_path = 1, fast_result_path = 2 }; 1352 Node *result_region = NULL; 1353 Node *result_phi_rawmem = NULL; 1354 Node *result_phi_rawoop = NULL; 1355 Node *result_phi_i_o = NULL; 1356 1357 // The initial slow comparison is a size check, the comparison 1358 // we want to do is a BoolTest::gt 1359 bool always_slow = false; 1360 int tv = _igvn.find_int_con(initial_slow_test, -1); 1361 if (tv >= 0) { 1362 always_slow = (tv == 1); 1363 initial_slow_test = NULL; 1364 } else { 1365 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 1366 } 1367 1368 if (C->env()->dtrace_alloc_probes() || 1369 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) { 1370 // Force slow-path allocation 1371 always_slow = true; 1372 initial_slow_test = NULL; 1373 } 1374 1375 1376 enum { too_big_or_final_path = 1, need_gc_path = 2 }; 1377 Node *slow_region = NULL; 1378 Node *toobig_false = ctrl; 1379 1380 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); 1381 // generate the initial test if necessary 1382 if (initial_slow_test != NULL ) { 1383 slow_region = new RegionNode(3); 1384 1385 // Now make the initial failure test. Usually a too-big test but 1386 // might be a TRUE for finalizers or a fancy class check for 1387 // newInstance0. 1388 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1389 transform_later(toobig_iff); 1390 // Plug the failing-too-big test into the slow-path region 1391 Node *toobig_true = new IfTrueNode( toobig_iff ); 1392 transform_later(toobig_true); 1393 slow_region ->init_req( too_big_or_final_path, toobig_true ); 1394 toobig_false = new IfFalseNode( toobig_iff ); 1395 transform_later(toobig_false); 1396 } else { // No initial test, just fall into next case 1397 toobig_false = ctrl; 1398 debug_only(slow_region = NodeSentinel); 1399 } 1400 1401 Node *slow_mem = mem; // save the current memory state for slow path 1402 // generate the fast allocation code unless we know that the initial test will always go slow 1403 if (!always_slow) { 1404 // Fast path modifies only raw memory. 1405 if (mem->is_MergeMem()) { 1406 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1407 } 1408 1409 Node* eden_top_adr; 1410 Node* eden_end_adr; 1411 1412 set_eden_pointers(eden_top_adr, eden_end_adr); 1413 1414 // Load Eden::end. Loop invariant and hoisted. 1415 // 1416 // Note: We set the control input on "eden_end" and "old_eden_top" when using 1417 // a TLAB to work around a bug where these values were being moved across 1418 // a safepoint. These are not oops, so they cannot be include in the oop 1419 // map, but they can be changed by a GC. The proper way to fix this would 1420 // be to set the raw memory state when generating a SafepointNode. However 1421 // this will require extensive changes to the loop optimization in order to 1422 // prevent a degradation of the optimization. 1423 // See comment in memnode.hpp, around line 227 in class LoadPNode. 1424 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 1425 1426 // allocate the Region and Phi nodes for the result 1427 result_region = new RegionNode(3); 1428 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1429 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM); 1430 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch 1431 1432 // We need a Region for the loop-back contended case. 1433 enum { fall_in_path = 1, contended_loopback_path = 2 }; 1434 Node *contended_region; 1435 Node *contended_phi_rawmem; 1436 if (UseTLAB) { 1437 contended_region = toobig_false; 1438 contended_phi_rawmem = mem; 1439 } else { 1440 contended_region = new RegionNode(3); 1441 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1442 // Now handle the passing-too-big test. We fall into the contended 1443 // loop-back merge point. 1444 contended_region ->init_req(fall_in_path, toobig_false); 1445 contended_phi_rawmem->init_req(fall_in_path, mem); 1446 transform_later(contended_region); 1447 transform_later(contended_phi_rawmem); 1448 } 1449 1450 // Load(-locked) the heap top. 1451 // See note above concerning the control input when using a TLAB 1452 Node *old_eden_top = UseTLAB 1453 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered) 1454 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire); 1455 1456 transform_later(old_eden_top); 1457 // Add to heap top to get a new heap top 1458 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes); 1459 transform_later(new_eden_top); 1460 // Check for needing a GC; compare against heap end 1461 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end); 1462 transform_later(needgc_cmp); 1463 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge); 1464 transform_later(needgc_bol); 1465 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); 1466 transform_later(needgc_iff); 1467 1468 // Plug the failing-heap-space-need-gc test into the slow-path region 1469 Node *needgc_true = new IfTrueNode(needgc_iff); 1470 transform_later(needgc_true); 1471 if (initial_slow_test) { 1472 slow_region->init_req(need_gc_path, needgc_true); 1473 // This completes all paths into the slow merge point 1474 transform_later(slow_region); 1475 } else { // No initial slow path needed! 1476 // Just fall from the need-GC path straight into the VM call. 1477 slow_region = needgc_true; 1478 } 1479 // No need for a GC. Setup for the Store-Conditional 1480 Node *needgc_false = new IfFalseNode(needgc_iff); 1481 transform_later(needgc_false); 1482 1483 // Grab regular I/O before optional prefetch may change it. 1484 // Slow-path does no I/O so just set it to the original I/O. 1485 result_phi_i_o->init_req(slow_result_path, i_o); 1486 1487 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem, 1488 old_eden_top, new_eden_top, length); 1489 1490 // Name successful fast-path variables 1491 Node* fast_oop = old_eden_top; 1492 Node* fast_oop_ctrl; 1493 Node* fast_oop_rawmem; 1494 1495 // Store (-conditional) the modified eden top back down. 1496 // StorePConditional produces flags for a test PLUS a modified raw 1497 // memory state. 1498 if (UseTLAB) { 1499 Node* store_eden_top = 1500 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1501 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered); 1502 transform_later(store_eden_top); 1503 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path 1504 fast_oop_rawmem = store_eden_top; 1505 } else { 1506 Node* store_eden_top = 1507 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1508 new_eden_top, fast_oop/*old_eden_top*/); 1509 transform_later(store_eden_top); 1510 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne); 1511 transform_later(contention_check); 1512 store_eden_top = new SCMemProjNode(store_eden_top); 1513 transform_later(store_eden_top); 1514 1515 // If not using TLABs, check to see if there was contention. 1516 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN); 1517 transform_later(contention_iff); 1518 Node *contention_true = new IfTrueNode(contention_iff); 1519 transform_later(contention_true); 1520 // If contention, loopback and try again. 1521 contended_region->init_req(contended_loopback_path, contention_true); 1522 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top); 1523 1524 // Fast-path succeeded with no contention! 1525 Node *contention_false = new IfFalseNode(contention_iff); 1526 transform_later(contention_false); 1527 fast_oop_ctrl = contention_false; 1528 1529 // Bump total allocated bytes for this thread 1530 Node* thread = new ThreadLocalNode(); 1531 transform_later(thread); 1532 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread, 1533 in_bytes(JavaThread::allocated_bytes_offset())); 1534 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1535 0, TypeLong::LONG, T_LONG); 1536 #ifdef _LP64 1537 Node* alloc_size = size_in_bytes; 1538 #else 1539 Node* alloc_size = new ConvI2LNode(size_in_bytes); 1540 transform_later(alloc_size); 1541 #endif 1542 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size); 1543 transform_later(new_alloc_bytes); 1544 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1545 0, new_alloc_bytes, T_LONG); 1546 } 1547 1548 InitializeNode* init = alloc->initialization(); 1549 fast_oop_rawmem = initialize_object(alloc, 1550 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1551 klass_node, length, size_in_bytes); 1552 1553 // If initialization is performed by an array copy, any required 1554 // MemBarStoreStore was already added. If the object does not 1555 // escape no need for a MemBarStoreStore. If the object does not 1556 // escape in its initializer and memory barrier (MemBarStoreStore or 1557 // stronger) is already added at exit of initializer, also no need 1558 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore 1559 // so that stores that initialize this object can't be reordered 1560 // with a subsequent store that makes this object accessible by 1561 // other threads. 1562 // Other threads include java threads and JVM internal threads 1563 // (for example concurrent GC threads). Current concurrent GC 1564 // implementation: CMS and G1 will not scan newly created object, 1565 // so it's safe to skip storestore barrier when allocation does 1566 // not escape. 1567 if (!alloc->does_not_escape_thread() && 1568 !alloc->is_allocation_MemBar_redundant() && 1569 (init == NULL || !init->is_complete_with_arraycopy())) { 1570 if (init == NULL || init->req() < InitializeNode::RawStores) { 1571 // No InitializeNode or no stores captured by zeroing 1572 // elimination. Simply add the MemBarStoreStore after object 1573 // initialization. 1574 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1575 transform_later(mb); 1576 1577 mb->init_req(TypeFunc::Memory, fast_oop_rawmem); 1578 mb->init_req(TypeFunc::Control, fast_oop_ctrl); 1579 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control); 1580 transform_later(fast_oop_ctrl); 1581 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory); 1582 transform_later(fast_oop_rawmem); 1583 } else { 1584 // Add the MemBarStoreStore after the InitializeNode so that 1585 // all stores performing the initialization that were moved 1586 // before the InitializeNode happen before the storestore 1587 // barrier. 1588 1589 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control); 1590 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory); 1591 1592 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1593 transform_later(mb); 1594 1595 Node* ctrl = new ProjNode(init,TypeFunc::Control); 1596 transform_later(ctrl); 1597 Node* mem = new ProjNode(init,TypeFunc::Memory); 1598 transform_later(mem); 1599 1600 // The MemBarStoreStore depends on control and memory coming 1601 // from the InitializeNode 1602 mb->init_req(TypeFunc::Memory, mem); 1603 mb->init_req(TypeFunc::Control, ctrl); 1604 1605 ctrl = new ProjNode(mb,TypeFunc::Control); 1606 transform_later(ctrl); 1607 mem = new ProjNode(mb,TypeFunc::Memory); 1608 transform_later(mem); 1609 1610 // All nodes that depended on the InitializeNode for control 1611 // and memory must now depend on the MemBarNode that itself 1612 // depends on the InitializeNode 1613 if (init_ctrl != NULL) { 1614 _igvn.replace_node(init_ctrl, ctrl); 1615 } 1616 if (init_mem != NULL) { 1617 _igvn.replace_node(init_mem, mem); 1618 } 1619 } 1620 } 1621 1622 if (C->env()->dtrace_extended_probes()) { 1623 // Slow-path call 1624 int size = TypeFunc::Parms + 2; 1625 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1626 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), 1627 "dtrace_object_alloc", 1628 TypeRawPtr::BOTTOM); 1629 1630 // Get base of thread-local storage area 1631 Node* thread = new ThreadLocalNode(); 1632 transform_later(thread); 1633 1634 call->init_req(TypeFunc::Parms+0, thread); 1635 call->init_req(TypeFunc::Parms+1, fast_oop); 1636 call->init_req(TypeFunc::Control, fast_oop_ctrl); 1637 call->init_req(TypeFunc::I_O , top()); // does no i/o 1638 call->init_req(TypeFunc::Memory , fast_oop_rawmem); 1639 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1640 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1641 transform_later(call); 1642 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control); 1643 transform_later(fast_oop_ctrl); 1644 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory); 1645 transform_later(fast_oop_rawmem); 1646 } 1647 1648 // Plug in the successful fast-path into the result merge point 1649 result_region ->init_req(fast_result_path, fast_oop_ctrl); 1650 result_phi_rawoop->init_req(fast_result_path, fast_oop); 1651 result_phi_i_o ->init_req(fast_result_path, i_o); 1652 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); 1653 } else { 1654 slow_region = ctrl; 1655 result_phi_i_o = i_o; // Rename it to use in the following code. 1656 } 1657 1658 // Generate slow-path call 1659 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address, 1660 OptoRuntime::stub_name(slow_call_address), 1661 alloc->jvms()->bci(), 1662 TypePtr::BOTTOM); 1663 call->init_req( TypeFunc::Control, slow_region ); 1664 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1665 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs 1666 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1667 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1668 1669 call->init_req(TypeFunc::Parms+0, klass_node); 1670 if (length != NULL) { 1671 call->init_req(TypeFunc::Parms+1, length); 1672 } 1673 1674 // Copy debug information and adjust JVMState information, then replace 1675 // allocate node with the call 1676 copy_call_debug_info((CallNode *) alloc, call); 1677 if (!always_slow) { 1678 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1679 } else { 1680 // Hook i_o projection to avoid its elimination during allocation 1681 // replacement (when only a slow call is generated). 1682 call->set_req(TypeFunc::I_O, result_phi_i_o); 1683 } 1684 _igvn.replace_node(alloc, call); 1685 transform_later(call); 1686 1687 // Identify the output projections from the allocate node and 1688 // adjust any references to them. 1689 // The control and io projections look like: 1690 // 1691 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1692 // Allocate Catch 1693 // ^---Proj(io) <-------+ ^---CatchProj(io) 1694 // 1695 // We are interested in the CatchProj nodes. 1696 // 1697 extract_call_projections(call); 1698 1699 // An allocate node has separate memory projections for the uses on 1700 // the control and i_o paths. Replace the control memory projection with 1701 // result_phi_rawmem (unless we are only generating a slow call when 1702 // both memory projections are combined) 1703 if (!always_slow && _memproj_fallthrough != NULL) { 1704 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { 1705 Node *use = _memproj_fallthrough->fast_out(i); 1706 _igvn.rehash_node_delayed(use); 1707 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); 1708 // back up iterator 1709 --i; 1710 } 1711 } 1712 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete 1713 // _memproj_catchall so we end up with a call that has only 1 memory projection. 1714 if (_memproj_catchall != NULL ) { 1715 if (_memproj_fallthrough == NULL) { 1716 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory); 1717 transform_later(_memproj_fallthrough); 1718 } 1719 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { 1720 Node *use = _memproj_catchall->fast_out(i); 1721 _igvn.rehash_node_delayed(use); 1722 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); 1723 // back up iterator 1724 --i; 1725 } 1726 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted"); 1727 _igvn.remove_dead_node(_memproj_catchall); 1728 } 1729 1730 // An allocate node has separate i_o projections for the uses on the control 1731 // and i_o paths. Always replace the control i_o projection with result i_o 1732 // otherwise incoming i_o become dead when only a slow call is generated 1733 // (it is different from memory projections where both projections are 1734 // combined in such case). 1735 if (_ioproj_fallthrough != NULL) { 1736 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { 1737 Node *use = _ioproj_fallthrough->fast_out(i); 1738 _igvn.rehash_node_delayed(use); 1739 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); 1740 // back up iterator 1741 --i; 1742 } 1743 } 1744 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete 1745 // _ioproj_catchall so we end up with a call that has only 1 i_o projection. 1746 if (_ioproj_catchall != NULL ) { 1747 if (_ioproj_fallthrough == NULL) { 1748 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O); 1749 transform_later(_ioproj_fallthrough); 1750 } 1751 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { 1752 Node *use = _ioproj_catchall->fast_out(i); 1753 _igvn.rehash_node_delayed(use); 1754 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); 1755 // back up iterator 1756 --i; 1757 } 1758 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted"); 1759 _igvn.remove_dead_node(_ioproj_catchall); 1760 } 1761 1762 // if we generated only a slow call, we are done 1763 if (always_slow) { 1764 // Now we can unhook i_o. 1765 if (result_phi_i_o->outcnt() > 1) { 1766 call->set_req(TypeFunc::I_O, top()); 1767 } else { 1768 assert(result_phi_i_o->unique_ctrl_out() == call, ""); 1769 // Case of new array with negative size known during compilation. 1770 // AllocateArrayNode::Ideal() optimization disconnect unreachable 1771 // following code since call to runtime will throw exception. 1772 // As result there will be no users of i_o after the call. 1773 // Leave i_o attached to this call to avoid problems in preceding graph. 1774 } 1775 return; 1776 } 1777 1778 1779 if (_fallthroughcatchproj != NULL) { 1780 ctrl = _fallthroughcatchproj->clone(); 1781 transform_later(ctrl); 1782 _igvn.replace_node(_fallthroughcatchproj, result_region); 1783 } else { 1784 ctrl = top(); 1785 } 1786 Node *slow_result; 1787 if (_resproj == NULL) { 1788 // no uses of the allocation result 1789 slow_result = top(); 1790 } else { 1791 slow_result = _resproj->clone(); 1792 transform_later(slow_result); 1793 _igvn.replace_node(_resproj, result_phi_rawoop); 1794 } 1795 1796 // Plug slow-path into result merge point 1797 result_region ->init_req( slow_result_path, ctrl ); 1798 result_phi_rawoop->init_req( slow_result_path, slow_result); 1799 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); 1800 transform_later(result_region); 1801 transform_later(result_phi_rawoop); 1802 transform_later(result_phi_rawmem); 1803 transform_later(result_phi_i_o); 1804 // This completes all paths into the result merge point 1805 } 1806 1807 1808 // Helper for PhaseMacroExpand::expand_allocate_common. 1809 // Initializes the newly-allocated storage. 1810 Node* 1811 PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1812 Node* control, Node* rawmem, Node* object, 1813 Node* klass_node, Node* length, 1814 Node* size_in_bytes) { 1815 InitializeNode* init = alloc->initialization(); 1816 // Store the klass & mark bits 1817 Node* mark_node = NULL; 1818 // For now only enable fast locking for non-array types 1819 if (UseBiasedLocking && (length == NULL)) { 1820 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 1821 } else { 1822 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); 1823 } 1824 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 1825 1826 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 1827 int header_size = alloc->minimum_header_size(); // conservatively small 1828 1829 // Array length 1830 if (length != NULL) { // Arrays need length field 1831 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1832 // conservatively small header size: 1833 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1834 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1835 if (k->is_array_klass()) // we know the exact header size in most cases: 1836 header_size = Klass::layout_helper_header_size(k->layout_helper()); 1837 } 1838 1839 // Clear the object body, if necessary. 1840 if (init == NULL) { 1841 // The init has somehow disappeared; be cautious and clear everything. 1842 // 1843 // This can happen if a node is allocated but an uncommon trap occurs 1844 // immediately. In this case, the Initialize gets associated with the 1845 // trap, and may be placed in a different (outer) loop, if the Allocate 1846 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1847 // there can be two Allocates to one Initialize. The answer in all these 1848 // edge cases is safety first. It is always safe to clear immediately 1849 // within an Allocate, and then (maybe or maybe not) clear some more later. 1850 if (!(UseTLAB && ZeroTLAB)) { 1851 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1852 header_size, size_in_bytes, 1853 &_igvn); 1854 } 1855 } else { 1856 if (!init->is_complete()) { 1857 // Try to win by zeroing only what the init does not store. 1858 // We can also try to do some peephole optimizations, 1859 // such as combining some adjacent subword stores. 1860 rawmem = init->complete_stores(control, rawmem, object, 1861 header_size, size_in_bytes, &_igvn); 1862 } 1863 // We have no more use for this link, since the AllocateNode goes away: 1864 init->set_req(InitializeNode::RawAddress, top()); 1865 // (If we keep the link, it just confuses the register allocator, 1866 // who thinks he sees a real use of the address by the membar.) 1867 } 1868 1869 return rawmem; 1870 } 1871 1872 // Generate prefetch instructions for next allocations. 1873 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1874 Node*& contended_phi_rawmem, 1875 Node* old_eden_top, Node* new_eden_top, 1876 Node* length) { 1877 enum { fall_in_path = 1, pf_path = 2 }; 1878 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1879 // Generate prefetch allocation with watermark check. 1880 // As an allocation hits the watermark, we will prefetch starting 1881 // at a "distance" away from watermark. 1882 1883 Node *pf_region = new RegionNode(3); 1884 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 1885 TypeRawPtr::BOTTOM ); 1886 // I/O is used for Prefetch 1887 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO ); 1888 1889 Node *thread = new ThreadLocalNode(); 1890 transform_later(thread); 1891 1892 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread, 1893 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1894 transform_later(eden_pf_adr); 1895 1896 Node *old_pf_wm = new LoadPNode(needgc_false, 1897 contended_phi_rawmem, eden_pf_adr, 1898 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, 1899 MemNode::unordered); 1900 transform_later(old_pf_wm); 1901 1902 // check against new_eden_top 1903 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm ); 1904 transform_later(need_pf_cmp); 1905 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge ); 1906 transform_later(need_pf_bol); 1907 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol, 1908 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1909 transform_later(need_pf_iff); 1910 1911 // true node, add prefetchdistance 1912 Node *need_pf_true = new IfTrueNode( need_pf_iff ); 1913 transform_later(need_pf_true); 1914 1915 Node *need_pf_false = new IfFalseNode( need_pf_iff ); 1916 transform_later(need_pf_false); 1917 1918 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm, 1919 _igvn.MakeConX(AllocatePrefetchDistance) ); 1920 transform_later(new_pf_wmt ); 1921 new_pf_wmt->set_req(0, need_pf_true); 1922 1923 Node *store_new_wmt = new StorePNode(need_pf_true, 1924 contended_phi_rawmem, eden_pf_adr, 1925 TypeRawPtr::BOTTOM, new_pf_wmt, 1926 MemNode::unordered); 1927 transform_later(store_new_wmt); 1928 1929 // adding prefetches 1930 pf_phi_abio->init_req( fall_in_path, i_o ); 1931 1932 Node *prefetch_adr; 1933 Node *prefetch; 1934 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1935 uint step_size = AllocatePrefetchStepSize; 1936 uint distance = 0; 1937 1938 for ( uint i = 0; i < lines; i++ ) { 1939 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt, 1940 _igvn.MakeConX(distance) ); 1941 transform_later(prefetch_adr); 1942 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1943 transform_later(prefetch); 1944 distance += step_size; 1945 i_o = prefetch; 1946 } 1947 pf_phi_abio->set_req( pf_path, i_o ); 1948 1949 pf_region->init_req( fall_in_path, need_pf_false ); 1950 pf_region->init_req( pf_path, need_pf_true ); 1951 1952 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 1953 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 1954 1955 transform_later(pf_region); 1956 transform_later(pf_phi_rawmem); 1957 transform_later(pf_phi_abio); 1958 1959 needgc_false = pf_region; 1960 contended_phi_rawmem = pf_phi_rawmem; 1961 i_o = pf_phi_abio; 1962 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { 1963 // Insert a prefetch instruction for each allocation. 1964 // This code is used to generate 1 prefetch instruction per cache line. 1965 1966 // Generate several prefetch instructions. 1967 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1968 uint step_size = AllocatePrefetchStepSize; 1969 uint distance = AllocatePrefetchDistance; 1970 1971 // Next cache address. 1972 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top, 1973 _igvn.MakeConX(step_size + distance)); 1974 transform_later(cache_adr); 1975 cache_adr = new CastP2XNode(needgc_false, cache_adr); 1976 transform_later(cache_adr); 1977 // Address is aligned to execute prefetch to the beginning of cache line size 1978 // (it is important when BIS instruction is used on SPARC as prefetch). 1979 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); 1980 cache_adr = new AndXNode(cache_adr, mask); 1981 transform_later(cache_adr); 1982 cache_adr = new CastX2PNode(cache_adr); 1983 transform_later(cache_adr); 1984 1985 // Prefetch 1986 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); 1987 prefetch->set_req(0, needgc_false); 1988 transform_later(prefetch); 1989 contended_phi_rawmem = prefetch; 1990 Node *prefetch_adr; 1991 distance = step_size; 1992 for ( uint i = 1; i < lines; i++ ) { 1993 prefetch_adr = new AddPNode( cache_adr, cache_adr, 1994 _igvn.MakeConX(distance) ); 1995 transform_later(prefetch_adr); 1996 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); 1997 transform_later(prefetch); 1998 distance += step_size; 1999 contended_phi_rawmem = prefetch; 2000 } 2001 } else if( AllocatePrefetchStyle > 0 ) { 2002 // Insert a prefetch for each allocation only on the fast-path 2003 Node *prefetch_adr; 2004 Node *prefetch; 2005 // Generate several prefetch instructions. 2006 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 2007 uint step_size = AllocatePrefetchStepSize; 2008 uint distance = AllocatePrefetchDistance; 2009 for ( uint i = 0; i < lines; i++ ) { 2010 prefetch_adr = new AddPNode( old_eden_top, new_eden_top, 2011 _igvn.MakeConX(distance) ); 2012 transform_later(prefetch_adr); 2013 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 2014 // Do not let it float too high, since if eden_top == eden_end, 2015 // both might be null. 2016 if( i == 0 ) { // Set control for first prefetch, next follows it 2017 prefetch->init_req(0, needgc_false); 2018 } 2019 transform_later(prefetch); 2020 distance += step_size; 2021 i_o = prefetch; 2022 } 2023 } 2024 return i_o; 2025 } 2026 2027 2028 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 2029 expand_allocate_common(alloc, NULL, 2030 OptoRuntime::new_instance_Type(), 2031 OptoRuntime::new_instance_Java()); 2032 } 2033 2034 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 2035 Node* length = alloc->in(AllocateNode::ALength); 2036 InitializeNode* init = alloc->initialization(); 2037 Node* klass_node = alloc->in(AllocateNode::KlassNode); 2038 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 2039 address slow_call_address; // Address of slow call 2040 if (init != NULL && init->is_complete_with_arraycopy() && 2041 k->is_type_array_klass()) { 2042 // Don't zero type array during slow allocation in VM since 2043 // it will be initialized later by arraycopy in compiled code. 2044 slow_call_address = OptoRuntime::new_array_nozero_Java(); 2045 } else { 2046 slow_call_address = OptoRuntime::new_array_Java(); 2047 } 2048 expand_allocate_common(alloc, length, 2049 OptoRuntime::new_array_Type(), 2050 slow_call_address); 2051 } 2052 2053 //-------------------mark_eliminated_box---------------------------------- 2054 // 2055 // During EA obj may point to several objects but after few ideal graph 2056 // transformations (CCP) it may point to only one non escaping object 2057 // (but still using phi), corresponding locks and unlocks will be marked 2058 // for elimination. Later obj could be replaced with a new node (new phi) 2059 // and which does not have escape information. And later after some graph 2060 // reshape other locks and unlocks (which were not marked for elimination 2061 // before) are connected to this new obj (phi) but they still will not be 2062 // marked for elimination since new obj has no escape information. 2063 // Mark all associated (same box and obj) lock and unlock nodes for 2064 // elimination if some of them marked already. 2065 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) { 2066 if (oldbox->as_BoxLock()->is_eliminated()) 2067 return; // This BoxLock node was processed already. 2068 2069 // New implementation (EliminateNestedLocks) has separate BoxLock 2070 // node for each locked region so mark all associated locks/unlocks as 2071 // eliminated even if different objects are referenced in one locked region 2072 // (for example, OSR compilation of nested loop inside locked scope). 2073 if (EliminateNestedLocks || 2074 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) { 2075 // Box is used only in one lock region. Mark this box as eliminated. 2076 _igvn.hash_delete(oldbox); 2077 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value 2078 _igvn.hash_insert(oldbox); 2079 2080 for (uint i = 0; i < oldbox->outcnt(); i++) { 2081 Node* u = oldbox->raw_out(i); 2082 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) { 2083 AbstractLockNode* alock = u->as_AbstractLock(); 2084 // Check lock's box since box could be referenced by Lock's debug info. 2085 if (alock->box_node() == oldbox) { 2086 // Mark eliminated all related locks and unlocks. 2087 #ifdef ASSERT 2088 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4"); 2089 #endif 2090 alock->set_non_esc_obj(); 2091 } 2092 } 2093 } 2094 return; 2095 } 2096 2097 // Create new "eliminated" BoxLock node and use it in monitor debug info 2098 // instead of oldbox for the same object. 2099 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 2100 2101 // Note: BoxLock node is marked eliminated only here and it is used 2102 // to indicate that all associated lock and unlock nodes are marked 2103 // for elimination. 2104 newbox->set_eliminated(); 2105 transform_later(newbox); 2106 2107 // Replace old box node with new box for all users of the same object. 2108 for (uint i = 0; i < oldbox->outcnt();) { 2109 bool next_edge = true; 2110 2111 Node* u = oldbox->raw_out(i); 2112 if (u->is_AbstractLock()) { 2113 AbstractLockNode* alock = u->as_AbstractLock(); 2114 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) { 2115 // Replace Box and mark eliminated all related locks and unlocks. 2116 #ifdef ASSERT 2117 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5"); 2118 #endif 2119 alock->set_non_esc_obj(); 2120 _igvn.rehash_node_delayed(alock); 2121 alock->set_box_node(newbox); 2122 next_edge = false; 2123 } 2124 } 2125 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) { 2126 FastLockNode* flock = u->as_FastLock(); 2127 assert(flock->box_node() == oldbox, "sanity"); 2128 _igvn.rehash_node_delayed(flock); 2129 flock->set_box_node(newbox); 2130 next_edge = false; 2131 } 2132 2133 // Replace old box in monitor debug info. 2134 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 2135 SafePointNode* sfn = u->as_SafePoint(); 2136 JVMState* youngest_jvms = sfn->jvms(); 2137 int max_depth = youngest_jvms->depth(); 2138 for (int depth = 1; depth <= max_depth; depth++) { 2139 JVMState* jvms = youngest_jvms->of_depth(depth); 2140 int num_mon = jvms->nof_monitors(); 2141 // Loop over monitors 2142 for (int idx = 0; idx < num_mon; idx++) { 2143 Node* obj_node = sfn->monitor_obj(jvms, idx); 2144 Node* box_node = sfn->monitor_box(jvms, idx); 2145 if (box_node == oldbox && obj_node->eqv_uncast(obj)) { 2146 int j = jvms->monitor_box_offset(idx); 2147 _igvn.replace_input_of(u, j, newbox); 2148 next_edge = false; 2149 } 2150 } 2151 } 2152 } 2153 if (next_edge) i++; 2154 } 2155 } 2156 2157 //-----------------------mark_eliminated_locking_nodes----------------------- 2158 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { 2159 if (EliminateNestedLocks) { 2160 if (alock->is_nested()) { 2161 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity"); 2162 return; 2163 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened 2164 // Only Lock node has JVMState needed here. 2165 // Not that preceding claim is documented anywhere else. 2166 if (alock->jvms() != NULL) { 2167 if (alock->as_Lock()->is_nested_lock_region()) { 2168 // Mark eliminated related nested locks and unlocks. 2169 Node* obj = alock->obj_node(); 2170 BoxLockNode* box_node = alock->box_node()->as_BoxLock(); 2171 assert(!box_node->is_eliminated(), "should not be marked yet"); 2172 // Note: BoxLock node is marked eliminated only here 2173 // and it is used to indicate that all associated lock 2174 // and unlock nodes are marked for elimination. 2175 box_node->set_eliminated(); // Box's hash is always NO_HASH here 2176 for (uint i = 0; i < box_node->outcnt(); i++) { 2177 Node* u = box_node->raw_out(i); 2178 if (u->is_AbstractLock()) { 2179 alock = u->as_AbstractLock(); 2180 if (alock->box_node() == box_node) { 2181 // Verify that this Box is referenced only by related locks. 2182 assert(alock->obj_node()->eqv_uncast(obj), ""); 2183 // Mark all related locks and unlocks. 2184 #ifdef ASSERT 2185 alock->log_lock_optimization(C, "eliminate_lock_set_nested"); 2186 #endif 2187 alock->set_nested(); 2188 } 2189 } 2190 } 2191 } else { 2192 #ifdef ASSERT 2193 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region"); 2194 if (C->log() != NULL) 2195 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output 2196 #endif 2197 } 2198 } 2199 return; 2200 } 2201 // Process locks for non escaping object 2202 assert(alock->is_non_esc_obj(), ""); 2203 } // EliminateNestedLocks 2204 2205 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object 2206 // Look for all locks of this object and mark them and 2207 // corresponding BoxLock nodes as eliminated. 2208 Node* obj = alock->obj_node(); 2209 for (uint j = 0; j < obj->outcnt(); j++) { 2210 Node* o = obj->raw_out(j); 2211 if (o->is_AbstractLock() && 2212 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) { 2213 alock = o->as_AbstractLock(); 2214 Node* box = alock->box_node(); 2215 // Replace old box node with new eliminated box for all users 2216 // of the same object and mark related locks as eliminated. 2217 mark_eliminated_box(box, obj); 2218 } 2219 } 2220 } 2221 } 2222 2223 // we have determined that this lock/unlock can be eliminated, we simply 2224 // eliminate the node without expanding it. 2225 // 2226 // Note: The membar's associated with the lock/unlock are currently not 2227 // eliminated. This should be investigated as a future enhancement. 2228 // 2229 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 2230 2231 if (!alock->is_eliminated()) { 2232 return false; 2233 } 2234 #ifdef ASSERT 2235 if (!alock->is_coarsened()) { 2236 // Check that new "eliminated" BoxLock node is created. 2237 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 2238 assert(oldbox->is_eliminated(), "should be done already"); 2239 } 2240 #endif 2241 2242 alock->log_lock_optimization(C, "eliminate_lock"); 2243 2244 #ifndef PRODUCT 2245 if (PrintEliminateLocks) { 2246 if (alock->is_Lock()) { 2247 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx); 2248 } else { 2249 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx); 2250 } 2251 } 2252 #endif 2253 2254 Node* mem = alock->in(TypeFunc::Memory); 2255 Node* ctrl = alock->in(TypeFunc::Control); 2256 2257 extract_call_projections(alock); 2258 // There are 2 projections from the lock. The lock node will 2259 // be deleted when its last use is subsumed below. 2260 assert(alock->outcnt() == 2 && 2261 _fallthroughproj != NULL && 2262 _memproj_fallthrough != NULL, 2263 "Unexpected projections from Lock/Unlock"); 2264 2265 Node* fallthroughproj = _fallthroughproj; 2266 Node* memproj_fallthrough = _memproj_fallthrough; 2267 2268 // The memory projection from a lock/unlock is RawMem 2269 // The input to a Lock is merged memory, so extract its RawMem input 2270 // (unless the MergeMem has been optimized away.) 2271 if (alock->is_Lock()) { 2272 // Seach for MemBarAcquireLock node and delete it also. 2273 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 2274 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, ""); 2275 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 2276 Node* memproj = membar->proj_out(TypeFunc::Memory); 2277 _igvn.replace_node(ctrlproj, fallthroughproj); 2278 _igvn.replace_node(memproj, memproj_fallthrough); 2279 2280 // Delete FastLock node also if this Lock node is unique user 2281 // (a loop peeling may clone a Lock node). 2282 Node* flock = alock->as_Lock()->fastlock_node(); 2283 if (flock->outcnt() == 1) { 2284 assert(flock->unique_out() == alock, "sanity"); 2285 _igvn.replace_node(flock, top()); 2286 } 2287 } 2288 2289 // Seach for MemBarReleaseLock node and delete it also. 2290 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() && 2291 ctrl->in(0)->is_MemBar()) { 2292 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 2293 assert(membar->Opcode() == Op_MemBarReleaseLock && 2294 mem->is_Proj() && membar == mem->in(0), ""); 2295 _igvn.replace_node(fallthroughproj, ctrl); 2296 _igvn.replace_node(memproj_fallthrough, mem); 2297 fallthroughproj = ctrl; 2298 memproj_fallthrough = mem; 2299 ctrl = membar->in(TypeFunc::Control); 2300 mem = membar->in(TypeFunc::Memory); 2301 } 2302 2303 _igvn.replace_node(fallthroughproj, ctrl); 2304 _igvn.replace_node(memproj_fallthrough, mem); 2305 return true; 2306 } 2307 2308 2309 //------------------------------expand_lock_node---------------------- 2310 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 2311 2312 Node* ctrl = lock->in(TypeFunc::Control); 2313 Node* mem = lock->in(TypeFunc::Memory); 2314 Node* obj = lock->obj_node(); 2315 Node* box = lock->box_node(); 2316 Node* flock = lock->fastlock_node(); 2317 2318 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2319 2320 // Make the merge point 2321 Node *region; 2322 Node *mem_phi; 2323 Node *slow_path; 2324 2325 if (UseOptoBiasInlining) { 2326 /* 2327 * See the full description in MacroAssembler::biased_locking_enter(). 2328 * 2329 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { 2330 * // The object is biased. 2331 * proto_node = klass->prototype_header; 2332 * o_node = thread | proto_node; 2333 * x_node = o_node ^ mark_word; 2334 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? 2335 * // Done. 2336 * } else { 2337 * if( (x_node & biased_lock_mask) != 0 ) { 2338 * // The klass's prototype header is no longer biased. 2339 * cas(&mark_word, mark_word, proto_node) 2340 * goto cas_lock; 2341 * } else { 2342 * // The klass's prototype header is still biased. 2343 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? 2344 * old = mark_word; 2345 * new = o_node; 2346 * } else { 2347 * // Different thread or anonymous biased. 2348 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); 2349 * new = thread | old; 2350 * } 2351 * // Try to rebias. 2352 * if( cas(&mark_word, old, new) == 0 ) { 2353 * // Done. 2354 * } else { 2355 * goto slow_path; // Failed. 2356 * } 2357 * } 2358 * } 2359 * } else { 2360 * // The object is not biased. 2361 * cas_lock: 2362 * if( FastLock(obj) == 0 ) { 2363 * // Done. 2364 * } else { 2365 * slow_path: 2366 * OptoRuntime::complete_monitor_locking_Java(obj); 2367 * } 2368 * } 2369 */ 2370 2371 region = new RegionNode(5); 2372 // create a Phi for the memory state 2373 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2374 2375 Node* fast_lock_region = new RegionNode(3); 2376 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); 2377 2378 // First, check mark word for the biased lock pattern. 2379 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2380 2381 // Get fast path - mark word has the biased lock pattern. 2382 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, 2383 markOopDesc::biased_lock_mask_in_place, 2384 markOopDesc::biased_lock_pattern, true); 2385 // fast_lock_region->in(1) is set to slow path. 2386 fast_lock_mem_phi->init_req(1, mem); 2387 2388 // Now check that the lock is biased to the current thread and has 2389 // the same epoch and bias as Klass::_prototype_header. 2390 2391 // Special-case a fresh allocation to avoid building nodes: 2392 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); 2393 if (klass_node == NULL) { 2394 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 2395 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr())); 2396 #ifdef _LP64 2397 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) { 2398 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity"); 2399 klass_node->in(1)->init_req(0, ctrl); 2400 } else 2401 #endif 2402 klass_node->init_req(0, ctrl); 2403 } 2404 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type()); 2405 2406 Node* thread = transform_later(new ThreadLocalNode()); 2407 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2408 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node)); 2409 Node* x_node = transform_later(new XorXNode(o_node, mark_node)); 2410 2411 // Get slow path - mark word does NOT match the value. 2412 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, 2413 (~markOopDesc::age_mask_in_place), 0); 2414 // region->in(3) is set to fast path - the object is biased to the current thread. 2415 mem_phi->init_req(3, mem); 2416 2417 2418 // Mark word does NOT match the value (thread | Klass::_prototype_header). 2419 2420 2421 // First, check biased pattern. 2422 // Get fast path - _prototype_header has the same biased lock pattern. 2423 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, 2424 markOopDesc::biased_lock_mask_in_place, 0, true); 2425 2426 not_biased_ctrl = fast_lock_region->in(2); // Slow path 2427 // fast_lock_region->in(2) - the prototype header is no longer biased 2428 // and we have to revoke the bias on this object. 2429 // We are going to try to reset the mark of this object to the prototype 2430 // value and fall through to the CAS-based locking scheme. 2431 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 2432 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr, 2433 proto_node, mark_node); 2434 transform_later(cas); 2435 Node* proj = transform_later(new SCMemProjNode(cas)); 2436 fast_lock_mem_phi->init_req(2, proj); 2437 2438 2439 // Second, check epoch bits. 2440 Node* rebiased_region = new RegionNode(3); 2441 Node* old_phi = new PhiNode( rebiased_region, TypeX_X); 2442 Node* new_phi = new PhiNode( rebiased_region, TypeX_X); 2443 2444 // Get slow path - mark word does NOT match epoch bits. 2445 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, 2446 markOopDesc::epoch_mask_in_place, 0); 2447 // The epoch of the current bias is not valid, attempt to rebias the object 2448 // toward the current thread. 2449 rebiased_region->init_req(2, epoch_ctrl); 2450 old_phi->init_req(2, mark_node); 2451 new_phi->init_req(2, o_node); 2452 2453 // rebiased_region->in(1) is set to fast path. 2454 // The epoch of the current bias is still valid but we know 2455 // nothing about the owner; it might be set or it might be clear. 2456 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place | 2457 markOopDesc::age_mask_in_place | 2458 markOopDesc::epoch_mask_in_place); 2459 Node* old = transform_later(new AndXNode(mark_node, cmask)); 2460 cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2461 Node* new_mark = transform_later(new OrXNode(cast_thread, old)); 2462 old_phi->init_req(1, old); 2463 new_phi->init_req(1, new_mark); 2464 2465 transform_later(rebiased_region); 2466 transform_later(old_phi); 2467 transform_later(new_phi); 2468 2469 // Try to acquire the bias of the object using an atomic operation. 2470 // If this fails we will go in to the runtime to revoke the object's bias. 2471 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi); 2472 transform_later(cas); 2473 proj = transform_later(new SCMemProjNode(cas)); 2474 2475 // Get slow path - Failed to CAS. 2476 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); 2477 mem_phi->init_req(4, proj); 2478 // region->in(4) is set to fast path - the object is rebiased to the current thread. 2479 2480 // Failed to CAS. 2481 slow_path = new RegionNode(3); 2482 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); 2483 2484 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control 2485 slow_mem->init_req(1, proj); 2486 2487 // Call CAS-based locking scheme (FastLock node). 2488 2489 transform_later(fast_lock_region); 2490 transform_later(fast_lock_mem_phi); 2491 2492 // Get slow path - FastLock failed to lock the object. 2493 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); 2494 mem_phi->init_req(2, fast_lock_mem_phi); 2495 // region->in(2) is set to fast path - the object is locked to the current thread. 2496 2497 slow_path->init_req(2, ctrl); // Capture slow-control 2498 slow_mem->init_req(2, fast_lock_mem_phi); 2499 2500 transform_later(slow_path); 2501 transform_later(slow_mem); 2502 // Reset lock's memory edge. 2503 lock->set_req(TypeFunc::Memory, slow_mem); 2504 2505 } else { 2506 region = new RegionNode(3); 2507 // create a Phi for the memory state 2508 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2509 2510 // Optimize test; set region slot 2 2511 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); 2512 mem_phi->init_req(2, mem); 2513 } 2514 2515 // Make slow path call 2516 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), 2517 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, 2518 obj, box, NULL); 2519 2520 extract_call_projections(call); 2521 2522 // Slow path can only throw asynchronous exceptions, which are always 2523 // de-opted. So the compiler thinks the slow-call can never throw an 2524 // exception. If it DOES throw an exception we would need the debug 2525 // info removed first (since if it throws there is no monitor). 2526 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2527 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2528 2529 // Capture slow path 2530 // disconnect fall-through projection from call and create a new one 2531 // hook up users of fall-through projection to region 2532 Node *slow_ctrl = _fallthroughproj->clone(); 2533 transform_later(slow_ctrl); 2534 _igvn.hash_delete(_fallthroughproj); 2535 _fallthroughproj->disconnect_inputs(NULL, C); 2536 region->init_req(1, slow_ctrl); 2537 // region inputs are now complete 2538 transform_later(region); 2539 _igvn.replace_node(_fallthroughproj, region); 2540 2541 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory)); 2542 mem_phi->init_req(1, memproj ); 2543 transform_later(mem_phi); 2544 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2545 } 2546 2547 //------------------------------expand_unlock_node---------------------- 2548 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 2549 2550 Node* ctrl = unlock->in(TypeFunc::Control); 2551 Node* mem = unlock->in(TypeFunc::Memory); 2552 Node* obj = unlock->obj_node(); 2553 Node* box = unlock->box_node(); 2554 2555 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2556 2557 // No need for a null check on unlock 2558 2559 // Make the merge point 2560 Node *region; 2561 Node *mem_phi; 2562 2563 if (UseOptoBiasInlining) { 2564 // Check for biased locking unlock case, which is a no-op. 2565 // See the full description in MacroAssembler::biased_locking_exit(). 2566 region = new RegionNode(4); 2567 // create a Phi for the memory state 2568 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2569 mem_phi->init_req(3, mem); 2570 2571 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2572 ctrl = opt_bits_test(ctrl, region, 3, mark_node, 2573 markOopDesc::biased_lock_mask_in_place, 2574 markOopDesc::biased_lock_pattern); 2575 } else { 2576 region = new RegionNode(3); 2577 // create a Phi for the memory state 2578 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2579 } 2580 2581 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box ); 2582 funlock = transform_later( funlock )->as_FastUnlock(); 2583 // Optimize test; set region slot 2 2584 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); 2585 Node *thread = transform_later(new ThreadLocalNode()); 2586 2587 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), 2588 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), 2589 "complete_monitor_unlocking_C", slow_path, obj, box, thread); 2590 2591 extract_call_projections(call); 2592 2593 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2594 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2595 2596 // No exceptions for unlocking 2597 // Capture slow path 2598 // disconnect fall-through projection from call and create a new one 2599 // hook up users of fall-through projection to region 2600 Node *slow_ctrl = _fallthroughproj->clone(); 2601 transform_later(slow_ctrl); 2602 _igvn.hash_delete(_fallthroughproj); 2603 _fallthroughproj->disconnect_inputs(NULL, C); 2604 region->init_req(1, slow_ctrl); 2605 // region inputs are now complete 2606 transform_later(region); 2607 _igvn.replace_node(_fallthroughproj, region); 2608 2609 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) ); 2610 mem_phi->init_req(1, memproj ); 2611 mem_phi->init_req(2, mem); 2612 transform_later(mem_phi); 2613 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2614 } 2615 2616 //---------------------------eliminate_macro_nodes---------------------- 2617 // Eliminate scalar replaced allocations and associated locks. 2618 void PhaseMacroExpand::eliminate_macro_nodes() { 2619 if (C->macro_count() == 0) 2620 return; 2621 2622 // First, attempt to eliminate locks 2623 int cnt = C->macro_count(); 2624 for (int i=0; i < cnt; i++) { 2625 Node *n = C->macro_node(i); 2626 if (n->is_AbstractLock()) { // Lock and Unlock nodes 2627 // Before elimination mark all associated (same box and obj) 2628 // lock and unlock nodes. 2629 mark_eliminated_locking_nodes(n->as_AbstractLock()); 2630 } 2631 } 2632 bool progress = true; 2633 while (progress) { 2634 progress = false; 2635 for (int i = C->macro_count(); i > 0; i--) { 2636 Node * n = C->macro_node(i-1); 2637 bool success = false; 2638 debug_only(int old_macro_count = C->macro_count();); 2639 if (n->is_AbstractLock()) { 2640 success = eliminate_locking_node(n->as_AbstractLock()); 2641 } 2642 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2643 progress = progress || success; 2644 } 2645 } 2646 // Next, attempt to eliminate allocations 2647 _has_locks = false; 2648 progress = true; 2649 while (progress) { 2650 progress = false; 2651 for (int i = C->macro_count(); i > 0; i--) { 2652 Node * n = C->macro_node(i-1); 2653 bool success = false; 2654 debug_only(int old_macro_count = C->macro_count();); 2655 switch (n->class_id()) { 2656 case Node::Class_Allocate: 2657 case Node::Class_AllocateArray: 2658 success = eliminate_allocate_node(n->as_Allocate()); 2659 break; 2660 case Node::Class_CallStaticJava: 2661 success = eliminate_boxing_node(n->as_CallStaticJava()); 2662 break; 2663 case Node::Class_Lock: 2664 case Node::Class_Unlock: 2665 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 2666 _has_locks = true; 2667 break; 2668 case Node::Class_ArrayCopy: 2669 break; 2670 case Node::Class_OuterStripMinedLoop: 2671 break; 2672 default: 2673 assert(n->Opcode() == Op_LoopLimit || 2674 n->Opcode() == Op_Opaque1 || 2675 n->Opcode() == Op_Opaque2 || 2676 n->Opcode() == Op_Opaque3, "unknown node type in macro list"); 2677 } 2678 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2679 progress = progress || success; 2680 } 2681 } 2682 } 2683 2684 //------------------------------expand_macro_nodes---------------------- 2685 // Returns true if a failure occurred. 2686 bool PhaseMacroExpand::expand_macro_nodes() { 2687 // Last attempt to eliminate macro nodes. 2688 eliminate_macro_nodes(); 2689 2690 // Make sure expansion will not cause node limit to be exceeded. 2691 // Worst case is a macro node gets expanded into about 200 nodes. 2692 // Allow 50% more for optimization. 2693 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) ) 2694 return true; 2695 2696 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. 2697 bool progress = true; 2698 while (progress) { 2699 progress = false; 2700 for (int i = C->macro_count(); i > 0; i--) { 2701 Node * n = C->macro_node(i-1); 2702 bool success = false; 2703 debug_only(int old_macro_count = C->macro_count();); 2704 if (n->Opcode() == Op_LoopLimit) { 2705 // Remove it from macro list and put on IGVN worklist to optimize. 2706 C->remove_macro_node(n); 2707 _igvn._worklist.push(n); 2708 success = true; 2709 } else if (n->Opcode() == Op_CallStaticJava) { 2710 // Remove it from macro list and put on IGVN worklist to optimize. 2711 C->remove_macro_node(n); 2712 _igvn._worklist.push(n); 2713 success = true; 2714 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { 2715 _igvn.replace_node(n, n->in(1)); 2716 success = true; 2717 #if INCLUDE_RTM_OPT 2718 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) { 2719 assert(C->profile_rtm(), "should be used only in rtm deoptimization code"); 2720 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), ""); 2721 Node* cmp = n->unique_out(); 2722 #ifdef ASSERT 2723 // Validate graph. 2724 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), ""); 2725 BoolNode* bol = cmp->unique_out()->as_Bool(); 2726 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() && 2727 (bol->_test._test == BoolTest::ne), ""); 2728 IfNode* ifn = bol->unique_out()->as_If(); 2729 assert((ifn->outcnt() == 2) && 2730 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, ""); 2731 #endif 2732 Node* repl = n->in(1); 2733 if (!_has_locks) { 2734 // Remove RTM state check if there are no locks in the code. 2735 // Replace input to compare the same value. 2736 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1); 2737 } 2738 _igvn.replace_node(n, repl); 2739 success = true; 2740 #endif 2741 } else if (n->Opcode() == Op_OuterStripMinedLoop) { 2742 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn); 2743 C->remove_macro_node(n); 2744 success = true; 2745 } 2746 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2747 progress = progress || success; 2748 } 2749 } 2750 2751 // expand arraycopy "macro" nodes first 2752 // For ReduceBulkZeroing, we must first process all arraycopy nodes 2753 // before the allocate nodes are expanded. 2754 int macro_idx = C->macro_count() - 1; 2755 while (macro_idx >= 0) { 2756 Node * n = C->macro_node(macro_idx); 2757 assert(n->is_macro(), "only macro nodes expected here"); 2758 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2759 // node is unreachable, so don't try to expand it 2760 C->remove_macro_node(n); 2761 } else if (n->is_ArrayCopy()){ 2762 int macro_count = C->macro_count(); 2763 expand_arraycopy_node(n->as_ArrayCopy()); 2764 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2765 } 2766 if (C->failing()) return true; 2767 macro_idx --; 2768 } 2769 2770 // expand "macro" nodes 2771 // nodes are removed from the macro list as they are processed 2772 while (C->macro_count() > 0) { 2773 int macro_count = C->macro_count(); 2774 Node * n = C->macro_node(macro_count-1); 2775 assert(n->is_macro(), "only macro nodes expected here"); 2776 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2777 // node is unreachable, so don't try to expand it 2778 C->remove_macro_node(n); 2779 continue; 2780 } 2781 switch (n->class_id()) { 2782 case Node::Class_Allocate: 2783 expand_allocate(n->as_Allocate()); 2784 break; 2785 case Node::Class_AllocateArray: 2786 expand_allocate_array(n->as_AllocateArray()); 2787 break; 2788 case Node::Class_Lock: 2789 expand_lock_node(n->as_Lock()); 2790 break; 2791 case Node::Class_Unlock: 2792 expand_unlock_node(n->as_Unlock()); 2793 break; 2794 default: 2795 assert(false, "unknown node type in macro list"); 2796 } 2797 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2798 if (C->failing()) return true; 2799 } 2800 2801 _igvn.set_delay_transform(false); 2802 _igvn.optimize(); 2803 if (C->failing()) return true; 2804 return false; 2805 }