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