1 /* 2 * Copyright (c) 2005, 2016, 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 "libadt/vectset.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/arraycopynode.hpp" 30 #include "opto/callnode.hpp" 31 #include "opto/castnode.hpp" 32 #include "opto/cfgnode.hpp" 33 #include "opto/compile.hpp" 34 #include "opto/convertnode.hpp" 35 #include "opto/graphKit.hpp" 36 #include "opto/locknode.hpp" 37 #include "opto/loopnode.hpp" 38 #include "opto/macro.hpp" 39 #include "opto/memnode.hpp" 40 #include "opto/narrowptrnode.hpp" 41 #include "opto/node.hpp" 42 #include "opto/opaquenode.hpp" 43 #include "opto/phaseX.hpp" 44 #include "opto/rootnode.hpp" 45 #include "opto/runtime.hpp" 46 #include "opto/subnode.hpp" 47 #include "opto/type.hpp" 48 #include "runtime/sharedRuntime.hpp" 49 50 51 // 52 // Replace any references to "oldref" in inputs to "use" with "newref". 53 // Returns the number of replacements made. 54 // 55 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { 56 int nreplacements = 0; 57 uint req = use->req(); 58 for (uint j = 0; j < use->len(); j++) { 59 Node *uin = use->in(j); 60 if (uin == oldref) { 61 if (j < req) 62 use->set_req(j, newref); 63 else 64 use->set_prec(j, newref); 65 nreplacements++; 66 } else if (j >= req && uin == NULL) { 67 break; 68 } 69 } 70 return nreplacements; 71 } 72 73 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) { 74 // Copy debug information and adjust JVMState information 75 uint old_dbg_start = oldcall->tf()->domain()->cnt(); 76 uint new_dbg_start = newcall->tf()->domain()->cnt(); 77 int jvms_adj = new_dbg_start - old_dbg_start; 78 assert (new_dbg_start == newcall->req(), "argument count mismatch"); 79 80 // SafePointScalarObject node could be referenced several times in debug info. 81 // Use Dict to record cloned nodes. 82 Dict* sosn_map = new Dict(cmpkey,hashkey); 83 for (uint i = old_dbg_start; i < oldcall->req(); i++) { 84 Node* old_in = oldcall->in(i); 85 // Clone old SafePointScalarObjectNodes, adjusting their field contents. 86 if (old_in != NULL && old_in->is_SafePointScalarObject()) { 87 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject(); 88 uint old_unique = C->unique(); 89 Node* new_in = old_sosn->clone(sosn_map); 90 if (old_unique != C->unique()) { // New node? 91 new_in->set_req(0, C->root()); // reset control edge 92 new_in = transform_later(new_in); // Register new node. 93 } 94 old_in = new_in; 95 } 96 newcall->add_req(old_in); 97 } 98 99 // JVMS may be shared so clone it before we modify it 100 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL); 101 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) { 102 jvms->set_map(newcall); 103 jvms->set_locoff(jvms->locoff()+jvms_adj); 104 jvms->set_stkoff(jvms->stkoff()+jvms_adj); 105 jvms->set_monoff(jvms->monoff()+jvms_adj); 106 jvms->set_scloff(jvms->scloff()+jvms_adj); 107 jvms->set_endoff(jvms->endoff()+jvms_adj); 108 } 109 } 110 111 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) { 112 Node* cmp; 113 if (mask != 0) { 114 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask))); 115 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits))); 116 } else { 117 cmp = word; 118 } 119 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne)); 120 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); 121 transform_later(iff); 122 123 // Fast path taken. 124 Node *fast_taken = transform_later(new IfFalseNode(iff)); 125 126 // Fast path not-taken, i.e. slow path 127 Node *slow_taken = transform_later(new IfTrueNode(iff)); 128 129 if (return_fast_path) { 130 region->init_req(edge, slow_taken); // Capture slow-control 131 return fast_taken; 132 } else { 133 region->init_req(edge, fast_taken); // Capture fast-control 134 return slow_taken; 135 } 136 } 137 138 //--------------------copy_predefined_input_for_runtime_call-------------------- 139 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { 140 // Set fixed predefined input arguments 141 call->init_req( TypeFunc::Control, ctrl ); 142 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); 143 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? 144 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); 145 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); 146 } 147 148 //------------------------------make_slow_call--------------------------------- 149 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, 150 address slow_call, const char* leaf_name, Node* slow_path, 151 Node* parm0, Node* parm1, Node* parm2) { 152 153 // Slow-path call 154 CallNode *call = leaf_name 155 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) 156 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM ); 157 158 // Slow path call has no side-effects, uses few values 159 copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); 160 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0); 161 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1); 162 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2); 163 copy_call_debug_info(oldcall, call); 164 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 165 _igvn.replace_node(oldcall, call); 166 transform_later(call); 167 168 return call; 169 } 170 171 void PhaseMacroExpand::extract_call_projections(CallNode *call) { 172 _fallthroughproj = NULL; 173 _fallthroughcatchproj = NULL; 174 _ioproj_fallthrough = NULL; 175 _ioproj_catchall = NULL; 176 _catchallcatchproj = NULL; 177 _memproj_fallthrough = NULL; 178 _memproj_catchall = NULL; 179 _resproj = NULL; 180 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) { 181 ProjNode *pn = call->fast_out(i)->as_Proj(); 182 switch (pn->_con) { 183 case TypeFunc::Control: 184 { 185 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj 186 _fallthroughproj = pn; 187 DUIterator_Fast jmax, j = pn->fast_outs(jmax); 188 const Node *cn = pn->fast_out(j); 189 if (cn->is_Catch()) { 190 ProjNode *cpn = NULL; 191 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { 192 cpn = cn->fast_out(k)->as_Proj(); 193 assert(cpn->is_CatchProj(), "must be a CatchProjNode"); 194 if (cpn->_con == CatchProjNode::fall_through_index) 195 _fallthroughcatchproj = cpn; 196 else { 197 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); 198 _catchallcatchproj = cpn; 199 } 200 } 201 } 202 break; 203 } 204 case TypeFunc::I_O: 205 if (pn->_is_io_use) 206 _ioproj_catchall = pn; 207 else 208 _ioproj_fallthrough = pn; 209 break; 210 case TypeFunc::Memory: 211 if (pn->_is_io_use) 212 _memproj_catchall = pn; 213 else 214 _memproj_fallthrough = pn; 215 break; 216 case TypeFunc::Parms: 217 _resproj = pn; 218 break; 219 default: 220 assert(false, "unexpected projection from allocation node."); 221 } 222 } 223 224 } 225 226 // Eliminate a card mark sequence. p2x is a ConvP2XNode 227 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) { 228 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required"); 229 if (!UseG1GC) { 230 // vanilla/CMS post barrier 231 Node *shift = p2x->unique_out(); 232 Node *addp = shift->unique_out(); 233 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) { 234 Node *mem = addp->last_out(j); 235 if (UseCondCardMark && mem->is_Load()) { 236 assert(mem->Opcode() == Op_LoadB, "unexpected code shape"); 237 // The load is checking if the card has been written so 238 // replace it with zero to fold the test. 239 _igvn.replace_node(mem, intcon(0)); 240 continue; 241 } 242 assert(mem->is_Store(), "store required"); 243 _igvn.replace_node(mem, mem->in(MemNode::Memory)); 244 } 245 } else { 246 // G1 pre/post barriers 247 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes"); 248 // It could be only one user, URShift node, in Object.clone() intrinsic 249 // but the new allocation is passed to arraycopy stub and it could not 250 // be scalar replaced. So we don't check the case. 251 252 // An other case of only one user (Xor) is when the value check for NULL 253 // in G1 post barrier is folded after CCP so the code which used URShift 254 // is removed. 255 256 // Take Region node before eliminating post barrier since it also 257 // eliminates CastP2X node when it has only one user. 258 Node* this_region = p2x->in(0); 259 assert(this_region != NULL, ""); 260 261 // Remove G1 post barrier. 262 263 // Search for CastP2X->Xor->URShift->Cmp path which 264 // checks if the store done to a different from the value's region. 265 // And replace Cmp with #0 (false) to collapse G1 post barrier. 266 Node* xorx = p2x->find_out_with(Op_XorX); 267 if (xorx != NULL) { 268 Node* shift = xorx->unique_out(); 269 Node* cmpx = shift->unique_out(); 270 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() && 271 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne, 272 "missing region check in G1 post barrier"); 273 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 274 275 // Remove G1 pre barrier. 276 277 // Search "if (marking != 0)" check and set it to "false". 278 // There is no G1 pre barrier if previous stored value is NULL 279 // (for example, after initialization). 280 if (this_region->is_Region() && this_region->req() == 3) { 281 int ind = 1; 282 if (!this_region->in(ind)->is_IfFalse()) { 283 ind = 2; 284 } 285 if (this_region->in(ind)->is_IfFalse()) { 286 Node* bol = this_region->in(ind)->in(0)->in(1); 287 assert(bol->is_Bool(), ""); 288 cmpx = bol->in(1); 289 if (bol->as_Bool()->_test._test == BoolTest::ne && 290 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) && 291 cmpx->in(1)->is_Load()) { 292 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address); 293 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + 294 SATBMarkQueue::byte_offset_of_active()); 295 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() && 296 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 297 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) { 298 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 299 } 300 } 301 } 302 } 303 } else { 304 assert(!GraphKit::use_ReduceInitialCardMarks(), "can only happen with card marking"); 305 // This is a G1 post barrier emitted by the Object.clone() intrinsic. 306 // Search for the CastP2X->URShiftX->AddP->LoadB->Cmp path which checks if the card 307 // is marked as young_gen and replace the Cmp with 0 (false) to collapse the barrier. 308 Node* shift = p2x->find_out_with(Op_URShiftX); 309 assert(shift != NULL, "missing G1 post barrier"); 310 Node* addp = shift->unique_out(); 311 Node* load = addp->find_out_with(Op_LoadB); 312 assert(load != NULL, "missing G1 post barrier"); 313 Node* cmpx = load->unique_out(); 314 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() && 315 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne, 316 "missing card value check in G1 post barrier"); 317 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 318 // There is no G1 pre barrier in this case 319 } 320 // Now CastP2X can be removed since it is used only on dead path 321 // which currently still alive until igvn optimize it. 322 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, ""); 323 _igvn.replace_node(p2x, top()); 324 } 325 } 326 327 // Search for a memory operation for the specified memory slice. 328 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) { 329 Node *orig_mem = mem; 330 Node *alloc_mem = alloc->in(TypeFunc::Memory); 331 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr(); 332 while (true) { 333 if (mem == alloc_mem || mem == start_mem ) { 334 return mem; // hit one of our sentinels 335 } else if (mem->is_MergeMem()) { 336 mem = mem->as_MergeMem()->memory_at(alias_idx); 337 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { 338 Node *in = mem->in(0); 339 // we can safely skip over safepoints, calls, locks and membars because we 340 // already know that the object is safe to eliminate. 341 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { 342 return in; 343 } else if (in->is_Call()) { 344 CallNode *call = in->as_Call(); 345 if (call->may_modify(tinst, phase)) { 346 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape"); 347 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) { 348 return in; 349 } 350 } 351 mem = in->in(TypeFunc::Memory); 352 } else if (in->is_MemBar()) { 353 ArrayCopyNode* ac = NULL; 354 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) { 355 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone"); 356 return ac; 357 } 358 mem = in->in(TypeFunc::Memory); 359 } else { 360 assert(false, "unexpected projection"); 361 } 362 } else if (mem->is_Store()) { 363 const TypePtr* atype = mem->as_Store()->adr_type(); 364 int adr_idx = phase->C->get_alias_index(atype); 365 if (adr_idx == alias_idx) { 366 assert(atype->isa_oopptr(), "address type must be oopptr"); 367 int adr_offset = atype->offset(); 368 uint adr_iid = atype->is_oopptr()->instance_id(); 369 // Array elements references have the same alias_idx 370 // but different offset and different instance_id. 371 if (adr_offset == offset && adr_iid == alloc->_idx) 372 return mem; 373 } else { 374 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw"); 375 } 376 mem = mem->in(MemNode::Memory); 377 } else if (mem->is_ClearArray()) { 378 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) { 379 // Can not bypass initialization of the instance 380 // we are looking. 381 debug_only(intptr_t offset;) 382 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity"); 383 InitializeNode* init = alloc->as_Allocate()->initialization(); 384 // We are looking for stored value, return Initialize node 385 // or memory edge from Allocate node. 386 if (init != NULL) 387 return init; 388 else 389 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers). 390 } 391 // Otherwise skip it (the call updated 'mem' value). 392 } else if (mem->Opcode() == Op_SCMemProj) { 393 mem = mem->in(0); 394 Node* adr = NULL; 395 if (mem->is_LoadStore()) { 396 adr = mem->in(MemNode::Address); 397 } else { 398 assert(mem->Opcode() == Op_EncodeISOArray || 399 mem->Opcode() == Op_StrCompressedCopy, "sanity"); 400 adr = mem->in(3); // Destination array 401 } 402 const TypePtr* atype = adr->bottom_type()->is_ptr(); 403 int adr_idx = phase->C->get_alias_index(atype); 404 if (adr_idx == alias_idx) { 405 DEBUG_ONLY(mem->dump();) 406 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 407 return NULL; 408 } 409 mem = mem->in(MemNode::Memory); 410 } else if (mem->Opcode() == Op_StrInflatedCopy) { 411 Node* adr = mem->in(3); // Destination array 412 const TypePtr* atype = adr->bottom_type()->is_ptr(); 413 int adr_idx = phase->C->get_alias_index(atype); 414 if (adr_idx == alias_idx) { 415 DEBUG_ONLY(mem->dump();) 416 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field"); 417 return NULL; 418 } 419 mem = mem->in(MemNode::Memory); 420 } else { 421 return mem; 422 } 423 assert(mem != orig_mem, "dead memory loop"); 424 } 425 } 426 427 // Generate loads from source of the arraycopy for fields of 428 // destination needed at a deoptimization point 429 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) { 430 BasicType bt = ft; 431 const Type *type = ftype; 432 if (ft == T_NARROWOOP) { 433 bt = T_OBJECT; 434 type = ftype->make_oopptr(); 435 } 436 Node* res = NULL; 437 if (ac->is_clonebasic()) { 438 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base); 439 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset))); 440 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset); 441 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned); 442 } else { 443 if (ac->modifies(offset, offset, &_igvn, true)) { 444 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result"); 445 uint shift = exact_log2(type2aelembytes(bt)); 446 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 447 #ifdef _LP64 448 diff = _igvn.transform(new ConvI2LNode(diff)); 449 #endif 450 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift))); 451 452 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff)); 453 Node* base = ac->in(ArrayCopyNode::Src); 454 Node* adr = _igvn.transform(new AddPNode(base, base, off)); 455 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset); 456 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned); 457 } 458 } 459 if (res != NULL) { 460 res = _igvn.transform(res); 461 if (ftype->isa_narrowoop()) { 462 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes 463 res = _igvn.transform(new EncodePNode(res, ftype)); 464 } 465 return res; 466 } 467 return NULL; 468 } 469 470 // 471 // Given a Memory Phi, compute a value Phi containing the values from stores 472 // on the input paths. 473 // Note: this function is recursive, its depth is limited by the "level" argument 474 // Returns the computed Phi, or NULL if it cannot compute it. 475 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) { 476 assert(mem->is_Phi(), "sanity"); 477 int alias_idx = C->get_alias_index(adr_t); 478 int offset = adr_t->offset(); 479 int instance_id = adr_t->instance_id(); 480 481 // Check if an appropriate value phi already exists. 482 Node* region = mem->in(0); 483 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { 484 Node* phi = region->fast_out(k); 485 if (phi->is_Phi() && phi != mem && 486 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) { 487 return phi; 488 } 489 } 490 // Check if an appropriate new value phi already exists. 491 Node* new_phi = value_phis->find(mem->_idx); 492 if (new_phi != NULL) 493 return new_phi; 494 495 if (level <= 0) { 496 return NULL; // Give up: phi tree too deep 497 } 498 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 499 Node *alloc_mem = alloc->in(TypeFunc::Memory); 500 501 uint length = mem->req(); 502 GrowableArray <Node *> values(length, length, NULL, false); 503 504 // create a new Phi for the value 505 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset); 506 transform_later(phi); 507 value_phis->push(phi, mem->_idx); 508 509 for (uint j = 1; j < length; j++) { 510 Node *in = mem->in(j); 511 if (in == NULL || in->is_top()) { 512 values.at_put(j, in); 513 } else { 514 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); 515 if (val == start_mem || val == alloc_mem) { 516 // hit a sentinel, return appropriate 0 value 517 values.at_put(j, _igvn.zerocon(ft)); 518 continue; 519 } 520 if (val->is_Initialize()) { 521 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 522 } 523 if (val == NULL) { 524 return NULL; // can't find a value on this path 525 } 526 if (val == mem) { 527 values.at_put(j, mem); 528 } else if (val->is_Store()) { 529 values.at_put(j, val->in(MemNode::ValueIn)); 530 } else if(val->is_Proj() && val->in(0) == alloc) { 531 values.at_put(j, _igvn.zerocon(ft)); 532 } else if (val->is_Phi()) { 533 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); 534 if (val == NULL) { 535 return NULL; 536 } 537 values.at_put(j, val); 538 } else if (val->Opcode() == Op_SCMemProj) { 539 assert(val->in(0)->is_LoadStore() || 540 val->in(0)->Opcode() == Op_EncodeISOArray || 541 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity"); 542 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 543 return NULL; 544 } else if (val->is_ArrayCopy()) { 545 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc); 546 if (res == NULL) { 547 return NULL; 548 } 549 values.at_put(j, res); 550 } else { 551 #ifdef ASSERT 552 val->dump(); 553 assert(false, "unknown node on this path"); 554 #endif 555 return NULL; // unknown node on this path 556 } 557 } 558 } 559 // Set Phi's inputs 560 for (uint j = 1; j < length; j++) { 561 if (values.at(j) == mem) { 562 phi->init_req(j, phi); 563 } else { 564 phi->init_req(j, values.at(j)); 565 } 566 } 567 return phi; 568 } 569 570 // Search the last value stored into the object's field. 571 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) { 572 assert(adr_t->is_known_instance_field(), "instance required"); 573 int instance_id = adr_t->instance_id(); 574 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 575 576 int alias_idx = C->get_alias_index(adr_t); 577 int offset = adr_t->offset(); 578 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 579 Node *alloc_ctrl = alloc->in(TypeFunc::Control); 580 Node *alloc_mem = alloc->in(TypeFunc::Memory); 581 Arena *a = Thread::current()->resource_area(); 582 VectorSet visited(a); 583 584 585 bool done = sfpt_mem == alloc_mem; 586 Node *mem = sfpt_mem; 587 while (!done) { 588 if (visited.test_set(mem->_idx)) { 589 return NULL; // found a loop, give up 590 } 591 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 592 if (mem == start_mem || mem == alloc_mem) { 593 done = true; // hit a sentinel, return appropriate 0 value 594 } else if (mem->is_Initialize()) { 595 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 596 if (mem == NULL) { 597 done = true; // Something go wrong. 598 } else if (mem->is_Store()) { 599 const TypePtr* atype = mem->as_Store()->adr_type(); 600 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 601 done = true; 602 } 603 } else if (mem->is_Store()) { 604 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 605 assert(atype != NULL, "address type must be oopptr"); 606 assert(C->get_alias_index(atype) == alias_idx && 607 atype->is_known_instance_field() && atype->offset() == offset && 608 atype->instance_id() == instance_id, "store is correct memory slice"); 609 done = true; 610 } else if (mem->is_Phi()) { 611 // try to find a phi's unique input 612 Node *unique_input = NULL; 613 Node *top = C->top(); 614 for (uint i = 1; i < mem->req(); i++) { 615 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 616 if (n == NULL || n == top || n == mem) { 617 continue; 618 } else if (unique_input == NULL) { 619 unique_input = n; 620 } else if (unique_input != n) { 621 unique_input = top; 622 break; 623 } 624 } 625 if (unique_input != NULL && unique_input != top) { 626 mem = unique_input; 627 } else { 628 done = true; 629 } 630 } else if (mem->is_ArrayCopy()) { 631 done = true; 632 } else { 633 assert(false, "unexpected node"); 634 } 635 } 636 if (mem != NULL) { 637 if (mem == start_mem || mem == alloc_mem) { 638 // hit a sentinel, return appropriate 0 value 639 return _igvn.zerocon(ft); 640 } else if (mem->is_Store()) { 641 return mem->in(MemNode::ValueIn); 642 } else if (mem->is_Phi()) { 643 // attempt to produce a Phi reflecting the values on the input paths of the Phi 644 Node_Stack value_phis(a, 8); 645 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 646 if (phi != NULL) { 647 return phi; 648 } else { 649 // Kill all new Phis 650 while(value_phis.is_nonempty()) { 651 Node* n = value_phis.node(); 652 _igvn.replace_node(n, C->top()); 653 value_phis.pop(); 654 } 655 } 656 } else if (mem->is_ArrayCopy()) { 657 Node* ctl = mem->in(0); 658 Node* m = mem->in(TypeFunc::Memory); 659 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) { 660 // pin the loads in the uncommon trap path 661 ctl = sfpt_ctl; 662 m = sfpt_mem; 663 } 664 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc); 665 } 666 } 667 // Something go wrong. 668 return NULL; 669 } 670 671 // Check the possibility of scalar replacement. 672 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 673 // Scan the uses of the allocation to check for anything that would 674 // prevent us from eliminating it. 675 NOT_PRODUCT( const char* fail_eliminate = NULL; ) 676 DEBUG_ONLY( Node* disq_node = NULL; ) 677 bool can_eliminate = true; 678 679 Node* res = alloc->result_cast(); 680 const TypeOopPtr* res_type = NULL; 681 if (res == NULL) { 682 // All users were eliminated. 683 } else if (!res->is_CheckCastPP()) { 684 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 685 can_eliminate = false; 686 } else { 687 res_type = _igvn.type(res)->isa_oopptr(); 688 if (res_type == NULL) { 689 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 690 can_eliminate = false; 691 } else if (res_type->isa_aryptr()) { 692 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 693 if (length < 0) { 694 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 695 can_eliminate = false; 696 } 697 } 698 } 699 700 if (can_eliminate && res != NULL) { 701 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 702 j < jmax && can_eliminate; j++) { 703 Node* use = res->fast_out(j); 704 705 if (use->is_AddP()) { 706 const TypePtr* addp_type = _igvn.type(use)->is_ptr(); 707 int offset = addp_type->offset(); 708 709 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 710 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) 711 can_eliminate = false; 712 break; 713 } 714 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 715 k < kmax && can_eliminate; k++) { 716 Node* n = use->fast_out(k); 717 if (!n->is_Store() && n->Opcode() != Op_CastP2X && 718 !(n->is_ArrayCopy() && 719 n->as_ArrayCopy()->is_clonebasic() && 720 n->in(ArrayCopyNode::Dest) == use)) { 721 DEBUG_ONLY(disq_node = n;) 722 if (n->is_Load() || n->is_LoadStore()) { 723 NOT_PRODUCT(fail_eliminate = "Field load";) 724 } else { 725 NOT_PRODUCT(fail_eliminate = "Not store field referrence";) 726 } 727 can_eliminate = false; 728 } 729 } 730 } else if (use->is_ArrayCopy() && 731 (use->as_ArrayCopy()->is_arraycopy_validated() || 732 use->as_ArrayCopy()->is_copyof_validated() || 733 use->as_ArrayCopy()->is_copyofrange_validated()) && 734 use->in(ArrayCopyNode::Dest) == res) { 735 // ok to eliminate 736 } else if (use->is_SafePoint()) { 737 SafePointNode* sfpt = use->as_SafePoint(); 738 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 739 // Object is passed as argument. 740 DEBUG_ONLY(disq_node = use;) 741 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 742 can_eliminate = false; 743 } 744 Node* sfptMem = sfpt->memory(); 745 if (sfptMem == NULL || sfptMem->is_top()) { 746 DEBUG_ONLY(disq_node = use;) 747 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) 748 can_eliminate = false; 749 } else { 750 safepoints.append_if_missing(sfpt); 751 } 752 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 753 if (use->is_Phi()) { 754 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 755 NOT_PRODUCT(fail_eliminate = "Object is return value";) 756 } else { 757 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 758 } 759 DEBUG_ONLY(disq_node = use;) 760 } else { 761 if (use->Opcode() == Op_Return) { 762 NOT_PRODUCT(fail_eliminate = "Object is return value";) 763 }else { 764 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 765 } 766 DEBUG_ONLY(disq_node = use;) 767 } 768 can_eliminate = false; 769 } 770 } 771 } 772 773 #ifndef PRODUCT 774 if (PrintEliminateAllocations) { 775 if (can_eliminate) { 776 tty->print("Scalar "); 777 if (res == NULL) 778 alloc->dump(); 779 else 780 res->dump(); 781 } else if (alloc->_is_scalar_replaceable) { 782 tty->print("NotScalar (%s)", fail_eliminate); 783 if (res == NULL) 784 alloc->dump(); 785 else 786 res->dump(); 787 #ifdef ASSERT 788 if (disq_node != NULL) { 789 tty->print(" >>>> "); 790 disq_node->dump(); 791 } 792 #endif /*ASSERT*/ 793 } 794 } 795 #endif 796 return can_eliminate; 797 } 798 799 // Do scalar replacement. 800 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 801 GrowableArray <SafePointNode *> safepoints_done; 802 803 ciKlass* klass = NULL; 804 ciInstanceKlass* iklass = NULL; 805 int nfields = 0; 806 int array_base = 0; 807 int element_size = 0; 808 BasicType basic_elem_type = T_ILLEGAL; 809 ciType* elem_type = NULL; 810 811 Node* res = alloc->result_cast(); 812 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result"); 813 const TypeOopPtr* res_type = NULL; 814 if (res != NULL) { // Could be NULL when there are no users 815 res_type = _igvn.type(res)->isa_oopptr(); 816 } 817 818 if (res != NULL) { 819 klass = res_type->klass(); 820 if (res_type->isa_instptr()) { 821 // find the fields of the class which will be needed for safepoint debug information 822 assert(klass->is_instance_klass(), "must be an instance klass."); 823 iklass = klass->as_instance_klass(); 824 nfields = iklass->nof_nonstatic_fields(); 825 } else { 826 // find the array's elements which will be needed for safepoint debug information 827 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 828 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); 829 elem_type = klass->as_array_klass()->element_type(); 830 basic_elem_type = elem_type->basic_type(); 831 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 832 element_size = type2aelembytes(basic_elem_type); 833 } 834 } 835 // 836 // Process the safepoint uses 837 // 838 while (safepoints.length() > 0) { 839 SafePointNode* sfpt = safepoints.pop(); 840 Node* mem = sfpt->memory(); 841 Node* ctl = sfpt->control(); 842 assert(sfpt->jvms() != NULL, "missed JVMS"); 843 // Fields of scalar objs are referenced only at the end 844 // of regular debuginfo at the last (youngest) JVMS. 845 // Record relative start index. 846 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff()); 847 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, 848 #ifdef ASSERT 849 alloc, 850 #endif 851 first_ind, nfields); 852 sobj->init_req(0, C->root()); 853 transform_later(sobj); 854 855 // Scan object's fields adding an input to the safepoint for each field. 856 for (int j = 0; j < nfields; j++) { 857 intptr_t offset; 858 ciField* field = NULL; 859 if (iklass != NULL) { 860 field = iklass->nonstatic_field_at(j); 861 offset = field->offset(); 862 elem_type = field->type(); 863 basic_elem_type = field->layout_type(); 864 } else { 865 offset = array_base + j * (intptr_t)element_size; 866 } 867 868 const Type *field_type; 869 // The next code is taken from Parse::do_get_xxx(). 870 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { 871 if (!elem_type->is_loaded()) { 872 field_type = TypeInstPtr::BOTTOM; 873 } else if (field != NULL && field->is_static_constant()) { 874 // This can happen if the constant oop is non-perm. 875 ciObject* con = field->constant_value().as_object(); 876 // Do not "join" in the previous type; it doesn't add value, 877 // and may yield a vacuous result if the field is of interface type. 878 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 879 assert(field_type != NULL, "field singleton type must be consistent"); 880 } else { 881 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 882 } 883 if (UseCompressedOops) { 884 field_type = field_type->make_narrowoop(); 885 basic_elem_type = T_NARROWOOP; 886 } 887 } else { 888 field_type = Type::get_const_basic_type(basic_elem_type); 889 } 890 891 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 892 893 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc); 894 if (field_val == NULL) { 895 // We weren't able to find a value for this field, 896 // give up on eliminating this allocation. 897 898 // Remove any extra entries we added to the safepoint. 899 uint last = sfpt->req() - 1; 900 for (int k = 0; k < j; k++) { 901 sfpt->del_req(last--); 902 } 903 _igvn._worklist.push(sfpt); 904 // rollback processed safepoints 905 while (safepoints_done.length() > 0) { 906 SafePointNode* sfpt_done = safepoints_done.pop(); 907 // remove any extra entries we added to the safepoint 908 last = sfpt_done->req() - 1; 909 for (int k = 0; k < nfields; k++) { 910 sfpt_done->del_req(last--); 911 } 912 JVMState *jvms = sfpt_done->jvms(); 913 jvms->set_endoff(sfpt_done->req()); 914 // Now make a pass over the debug information replacing any references 915 // to SafePointScalarObjectNode with the allocated object. 916 int start = jvms->debug_start(); 917 int end = jvms->debug_end(); 918 for (int i = start; i < end; i++) { 919 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 920 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 921 if (scobj->first_index(jvms) == sfpt_done->req() && 922 scobj->n_fields() == (uint)nfields) { 923 assert(scobj->alloc() == alloc, "sanity"); 924 sfpt_done->set_req(i, res); 925 } 926 } 927 } 928 _igvn._worklist.push(sfpt_done); 929 } 930 #ifndef PRODUCT 931 if (PrintEliminateAllocations) { 932 if (field != NULL) { 933 tty->print("=== At SafePoint node %d can't find value of Field: ", 934 sfpt->_idx); 935 field->print(); 936 int field_idx = C->get_alias_index(field_addr_type); 937 tty->print(" (alias_idx=%d)", field_idx); 938 } else { // Array's element 939 tty->print("=== At SafePoint node %d can't find value of array element [%d]", 940 sfpt->_idx, j); 941 } 942 tty->print(", which prevents elimination of: "); 943 if (res == NULL) 944 alloc->dump(); 945 else 946 res->dump(); 947 } 948 #endif 949 return false; 950 } 951 if (UseCompressedOops && field_type->isa_narrowoop()) { 952 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 953 // to be able scalar replace the allocation. 954 if (field_val->is_EncodeP()) { 955 field_val = field_val->in(1); 956 } else { 957 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type())); 958 } 959 } 960 sfpt->add_req(field_val); 961 } 962 JVMState *jvms = sfpt->jvms(); 963 jvms->set_endoff(sfpt->req()); 964 // Now make a pass over the debug information replacing any references 965 // to the allocated object with "sobj" 966 int start = jvms->debug_start(); 967 int end = jvms->debug_end(); 968 sfpt->replace_edges_in_range(res, sobj, start, end); 969 _igvn._worklist.push(sfpt); 970 safepoints_done.append_if_missing(sfpt); // keep it for rollback 971 } 972 return true; 973 } 974 975 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) { 976 Node* ctl_proj = n->proj_out(TypeFunc::Control); 977 Node* mem_proj = n->proj_out(TypeFunc::Memory); 978 if (ctl_proj != NULL) { 979 igvn.replace_node(ctl_proj, n->in(0)); 980 } 981 if (mem_proj != NULL) { 982 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory)); 983 } 984 } 985 986 // Process users of eliminated allocation. 987 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) { 988 Node* res = alloc->result_cast(); 989 if (res != NULL) { 990 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 991 Node *use = res->last_out(j); 992 uint oc1 = res->outcnt(); 993 994 if (use->is_AddP()) { 995 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 996 Node *n = use->last_out(k); 997 uint oc2 = use->outcnt(); 998 if (n->is_Store()) { 999 #ifdef ASSERT 1000 // Verify that there is no dependent MemBarVolatile nodes, 1001 // they should be removed during IGVN, see MemBarNode::Ideal(). 1002 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); 1003 p < pmax; p++) { 1004 Node* mb = n->fast_out(p); 1005 assert(mb->is_Initialize() || !mb->is_MemBar() || 1006 mb->req() <= MemBarNode::Precedent || 1007 mb->in(MemBarNode::Precedent) != n, 1008 "MemBarVolatile should be eliminated for non-escaping object"); 1009 } 1010 #endif 1011 _igvn.replace_node(n, n->in(MemNode::Memory)); 1012 } else if (n->is_ArrayCopy()) { 1013 // Disconnect ArrayCopy node 1014 ArrayCopyNode* ac = n->as_ArrayCopy(); 1015 assert(ac->is_clonebasic(), "unexpected array copy kind"); 1016 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out(); 1017 disconnect_projections(ac, _igvn); 1018 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation"); 1019 Node* membar_before = alloc->in(0)->in(0); 1020 disconnect_projections(membar_before->as_MemBar(), _igvn); 1021 if (membar_after->is_MemBar()) { 1022 disconnect_projections(membar_after->as_MemBar(), _igvn); 1023 } 1024 } else { 1025 eliminate_card_mark(n); 1026 } 1027 k -= (oc2 - use->outcnt()); 1028 } 1029 } else if (use->is_ArrayCopy()) { 1030 // Disconnect ArrayCopy node 1031 ArrayCopyNode* ac = use->as_ArrayCopy(); 1032 assert(ac->is_arraycopy_validated() || 1033 ac->is_copyof_validated() || 1034 ac->is_copyofrange_validated(), "unsupported"); 1035 CallProjections callprojs; 1036 ac->extract_projections(&callprojs, true); 1037 1038 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O)); 1039 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory)); 1040 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control)); 1041 1042 // Set control to top. IGVN will remove the remaining projections 1043 ac->set_req(0, top()); 1044 ac->replace_edge(res, top()); 1045 1046 // Disconnect src right away: it can help find new 1047 // opportunities for allocation elimination 1048 Node* src = ac->in(ArrayCopyNode::Src); 1049 ac->replace_edge(src, top()); 1050 if (src->outcnt() == 0) { 1051 _igvn.remove_dead_node(src); 1052 } 1053 1054 _igvn._worklist.push(ac); 1055 } else { 1056 eliminate_card_mark(use); 1057 } 1058 j -= (oc1 - res->outcnt()); 1059 } 1060 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 1061 _igvn.remove_dead_node(res); 1062 } 1063 1064 // 1065 // Process other users of allocation's projections 1066 // 1067 if (_resproj != NULL && _resproj->outcnt() != 0) { 1068 // First disconnect stores captured by Initialize node. 1069 // If Initialize node is eliminated first in the following code, 1070 // it will kill such stores and DUIterator_Last will assert. 1071 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) { 1072 Node *use = _resproj->fast_out(j); 1073 if (use->is_AddP()) { 1074 // raw memory addresses used only by the initialization 1075 _igvn.replace_node(use, C->top()); 1076 --j; --jmax; 1077 } 1078 } 1079 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { 1080 Node *use = _resproj->last_out(j); 1081 uint oc1 = _resproj->outcnt(); 1082 if (use->is_Initialize()) { 1083 // Eliminate Initialize node. 1084 InitializeNode *init = use->as_Initialize(); 1085 assert(init->outcnt() <= 2, "only a control and memory projection expected"); 1086 Node *ctrl_proj = init->proj_out(TypeFunc::Control); 1087 if (ctrl_proj != NULL) { 1088 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection"); 1089 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj); 1090 } 1091 Node *mem_proj = init->proj_out(TypeFunc::Memory); 1092 if (mem_proj != NULL) { 1093 Node *mem = init->in(TypeFunc::Memory); 1094 #ifdef ASSERT 1095 if (mem->is_MergeMem()) { 1096 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection"); 1097 } else { 1098 assert(mem == _memproj_fallthrough, "allocation memory projection"); 1099 } 1100 #endif 1101 _igvn.replace_node(mem_proj, mem); 1102 } 1103 } else { 1104 assert(false, "only Initialize or AddP expected"); 1105 } 1106 j -= (oc1 - _resproj->outcnt()); 1107 } 1108 } 1109 if (_fallthroughcatchproj != NULL) { 1110 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); 1111 } 1112 if (_memproj_fallthrough != NULL) { 1113 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); 1114 } 1115 if (_memproj_catchall != NULL) { 1116 _igvn.replace_node(_memproj_catchall, C->top()); 1117 } 1118 if (_ioproj_fallthrough != NULL) { 1119 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); 1120 } 1121 if (_ioproj_catchall != NULL) { 1122 _igvn.replace_node(_ioproj_catchall, C->top()); 1123 } 1124 if (_catchallcatchproj != NULL) { 1125 _igvn.replace_node(_catchallcatchproj, C->top()); 1126 } 1127 } 1128 1129 void PhaseMacroExpand::conditional_sample(Node *should_sample, 1130 BoolTest::mask test, 1131 float probability, 1132 CallLeafNode *call, 1133 Node *thread, 1134 Node **fast_oop_ctrl, 1135 Node **fast_oop_rawmem, 1136 Node **fast_oop, 1137 Node *size_in_bytes, 1138 Node *in_node) { 1139 Node* sample_cmp = new CmpXNode(should_sample, _igvn.MakeConX(0)); 1140 transform_later(sample_cmp); 1141 1142 Node *sample_bool = new BoolNode(sample_cmp, test); 1143 transform_later(sample_bool); 1144 1145 IfNode *sample_if = new IfNode(*fast_oop_ctrl, 1146 sample_bool, 1147 probability, 1148 COUNT_UNKNOWN); 1149 transform_later(sample_if); 1150 1151 // Slow-path call to sample 1152 Node *sample_true = new IfTrueNode(sample_if); 1153 transform_later(sample_true); 1154 1155 // Fast path to no sample 1156 Node *sample_false = new IfFalseNode(sample_if); 1157 transform_later(sample_false); 1158 1159 // Create postdominators for both the control and data flow paths. 1160 Node *sample_region = new RegionNode(3); 1161 Node *sample_phi_rawmem = new PhiNode(sample_region, 1162 Type::MEMORY, 1163 TypeRawPtr::BOTTOM); 1164 1165 sample_region->init_req(1, sample_false); 1166 sample_phi_rawmem->init_req(1, *fast_oop_rawmem); 1167 1168 // Invoke the sampling method on the slow path. 1169 int size = TypeFunc::Parms + 2; 1170 1171 call->init_req(TypeFunc::Parms+0, thread); 1172 call->init_req(TypeFunc::Parms+1, *fast_oop); 1173 call->init_req(TypeFunc::Parms+2, size_in_bytes); 1174 #ifdef _LP64 1175 // The size is TypeX, so in a 64-bit JVM this a long, and we need 1176 // // a second, dummy argument (an idiosyncracy of C2). 1177 call->init_req(TypeFunc::Parms+3, C->top()); 1178 #endif 1179 call->init_req( TypeFunc::Control, sample_true); 1180 call->init_req( TypeFunc::I_O , top()); // does no i/o 1181 call->init_req( TypeFunc::Memory , *fast_oop_rawmem ); 1182 call->init_req( TypeFunc::ReturnAdr, in_node->in(TypeFunc::ReturnAdr)); 1183 call->init_req( TypeFunc::FramePtr, in_node->in(TypeFunc::FramePtr)); 1184 transform_later(call); 1185 Node *sample_oop_rawmem = new ProjNode(call, TypeFunc::Memory); 1186 transform_later(sample_oop_rawmem); 1187 1188 // Tie the slow path to the postdominating node. 1189 sample_region->init_req(2, sample_true); 1190 sample_phi_rawmem->init_req(2, sample_oop_rawmem); 1191 transform_later(sample_region); 1192 1193 *fast_oop_ctrl = sample_region; 1194 *fast_oop_rawmem = sample_phi_rawmem; 1195 transform_later(*fast_oop_rawmem); 1196 } 1197 1198 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 1199 // Don't do scalar replacement if the frame can be popped by JVMTI: 1200 // if reallocation fails during deoptimization we'll pop all 1201 // interpreter frames for this compiled frame and that won't play 1202 // nice with JVMTI popframe. 1203 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) { 1204 return false; 1205 } 1206 Node* klass = alloc->in(AllocateNode::KlassNode); 1207 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); 1208 Node* res = alloc->result_cast(); 1209 // Eliminate boxing allocations which are not used 1210 // regardless scalar replacable status. 1211 bool boxing_alloc = C->eliminate_boxing() && 1212 tklass->klass()->is_instance_klass() && 1213 tklass->klass()->as_instance_klass()->is_box_klass(); 1214 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) { 1215 return false; 1216 } 1217 1218 extract_call_projections(alloc); 1219 1220 GrowableArray <SafePointNode *> safepoints; 1221 if (!can_eliminate_allocation(alloc, safepoints)) { 1222 return false; 1223 } 1224 1225 if (!alloc->_is_scalar_replaceable) { 1226 assert(res == NULL, "sanity"); 1227 // We can only eliminate allocation if all debug info references 1228 // are already replaced with SafePointScalarObject because 1229 // we can't search for a fields value without instance_id. 1230 if (safepoints.length() > 0) { 1231 return false; 1232 } 1233 } 1234 1235 if (!scalar_replacement(alloc, safepoints)) { 1236 return false; 1237 } 1238 1239 CompileLog* log = C->log(); 1240 if (log != NULL) { 1241 log->head("eliminate_allocation type='%d'", 1242 log->identify(tklass->klass())); 1243 JVMState* p = alloc->jvms(); 1244 while (p != NULL) { 1245 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1246 p = p->caller(); 1247 } 1248 log->tail("eliminate_allocation"); 1249 } 1250 1251 process_users_of_allocation(alloc); 1252 1253 #ifndef PRODUCT 1254 if (PrintEliminateAllocations) { 1255 if (alloc->is_AllocateArray()) 1256 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1257 else 1258 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1259 } 1260 #endif 1261 1262 return true; 1263 } 1264 1265 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) { 1266 // EA should remove all uses of non-escaping boxing node. 1267 if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) { 1268 return false; 1269 } 1270 1271 assert(boxing->result_cast() == NULL, "unexpected boxing node result"); 1272 1273 extract_call_projections(boxing); 1274 1275 const TypeTuple* r = boxing->tf()->range(); 1276 assert(r->cnt() > TypeFunc::Parms, "sanity"); 1277 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr(); 1278 assert(t != NULL, "sanity"); 1279 1280 CompileLog* log = C->log(); 1281 if (log != NULL) { 1282 log->head("eliminate_boxing type='%d'", 1283 log->identify(t->klass())); 1284 JVMState* p = boxing->jvms(); 1285 while (p != NULL) { 1286 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1287 p = p->caller(); 1288 } 1289 log->tail("eliminate_boxing"); 1290 } 1291 1292 process_users_of_allocation(boxing); 1293 1294 #ifndef PRODUCT 1295 if (PrintEliminateAllocations) { 1296 tty->print("++++ Eliminated: %d ", boxing->_idx); 1297 boxing->method()->print_short_name(tty); 1298 tty->cr(); 1299 } 1300 #endif 1301 1302 return true; 1303 } 1304 1305 //---------------------------set_eden_pointers------------------------- 1306 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { 1307 if (UseTLAB) { // Private allocation: load from TLS 1308 Node* thread = transform_later(new ThreadLocalNode()); 1309 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); 1310 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); 1311 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); 1312 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); 1313 } else { // Shared allocation: load from globals 1314 CollectedHeap* ch = Universe::heap(); 1315 address top_adr = (address)ch->top_addr(); 1316 address end_adr = (address)ch->end_addr(); 1317 eden_top_adr = makecon(TypeRawPtr::make(top_adr)); 1318 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); 1319 } 1320 } 1321 1322 1323 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 1324 Node* adr = basic_plus_adr(base, offset); 1325 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 1326 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered); 1327 transform_later(value); 1328 return value; 1329 } 1330 1331 1332 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 1333 Node* adr = basic_plus_adr(base, offset); 1334 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered); 1335 transform_later(mem); 1336 return mem; 1337 } 1338 1339 //============================================================================= 1340 // 1341 // A L L O C A T I O N 1342 // 1343 // Allocation attempts to be fast in the case of frequent small objects. 1344 // It breaks down like this: 1345 // 1346 // 1) Size in doublewords is computed. This is a constant for objects and 1347 // variable for most arrays. Doubleword units are used to avoid size 1348 // overflow of huge doubleword arrays. We need doublewords in the end for 1349 // rounding. 1350 // 1351 // 2) Size is checked for being 'too large'. Too-large allocations will go 1352 // the slow path into the VM. The slow path can throw any required 1353 // exceptions, and does all the special checks for very large arrays. The 1354 // size test can constant-fold away for objects. For objects with 1355 // finalizers it constant-folds the otherway: you always go slow with 1356 // finalizers. 1357 // 1358 // 3) If NOT using TLABs, this is the contended loop-back point. 1359 // Load-Locked the heap top. If using TLABs normal-load the heap top. 1360 // 1361 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 1362 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 1363 // "size*8" we always enter the VM, where "largish" is a constant picked small 1364 // enough that there's always space between the eden max and 4Gig (old space is 1365 // there so it's quite large) and large enough that the cost of entering the VM 1366 // is dwarfed by the cost to initialize the space. 1367 // 1368 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 1369 // down. If contended, repeat at step 3. If using TLABs normal-store 1370 // adjusted heap top back down; there is no contention. 1371 // 1372 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 1373 // fields. 1374 // 1375 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 1376 // oop flavor. 1377 // 1378 //============================================================================= 1379 // FastAllocateSizeLimit value is in DOUBLEWORDS. 1380 // Allocations bigger than this always go the slow route. 1381 // This value must be small enough that allocation attempts that need to 1382 // trigger exceptions go the slow route. Also, it must be small enough so 1383 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 1384 //=============================================================================j// 1385 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 1386 // The allocator will coalesce int->oop copies away. See comment in 1387 // coalesce.cpp about how this works. It depends critically on the exact 1388 // code shape produced here, so if you are changing this code shape 1389 // make sure the GC info for the heap-top is correct in and around the 1390 // slow-path call. 1391 // 1392 1393 void PhaseMacroExpand::expand_allocate_common( 1394 AllocateNode* alloc, // allocation node to be expanded 1395 Node* length, // array length for an array allocation 1396 const TypeFunc* slow_call_type, // Type of slow call 1397 address slow_call_address // Address of slow call 1398 ) 1399 { 1400 1401 Node* ctrl = alloc->in(TypeFunc::Control); 1402 Node* mem = alloc->in(TypeFunc::Memory); 1403 Node* i_o = alloc->in(TypeFunc::I_O); 1404 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 1405 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1406 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 1407 1408 assert(ctrl != NULL, "must have control"); 1409 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 1410 // they will not be used if "always_slow" is set 1411 enum { slow_result_path = 1, fast_result_path = 2 }; 1412 Node *result_region = NULL; 1413 Node *result_phi_rawmem = NULL; 1414 Node *result_phi_rawoop = NULL; 1415 Node *result_phi_i_o = NULL; 1416 1417 // The initial slow comparison is a size check, the comparison 1418 // we want to do is a BoolTest::gt 1419 bool always_slow = false; 1420 int tv = _igvn.find_int_con(initial_slow_test, -1); 1421 if (tv >= 0) { 1422 always_slow = (tv == 1); 1423 initial_slow_test = NULL; 1424 } else { 1425 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 1426 } 1427 1428 if (C->env()->dtrace_alloc_probes() || 1429 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc())) { 1430 // Force slow-path allocation 1431 always_slow = true; 1432 initial_slow_test = NULL; 1433 } 1434 1435 1436 enum { too_big_or_final_path = 1, need_gc_path = 2 }; 1437 Node *slow_region = NULL; 1438 Node *toobig_false = ctrl; 1439 1440 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); 1441 // generate the initial test if necessary 1442 if (initial_slow_test != NULL ) { 1443 slow_region = new RegionNode(3); 1444 1445 // Now make the initial failure test. Usually a too-big test but 1446 // might be a TRUE for finalizers or a fancy class check for 1447 // newInstance0. 1448 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1449 transform_later(toobig_iff); 1450 // Plug the failing-too-big test into the slow-path region 1451 Node *toobig_true = new IfTrueNode( toobig_iff ); 1452 transform_later(toobig_true); 1453 slow_region ->init_req( too_big_or_final_path, toobig_true ); 1454 toobig_false = new IfFalseNode( toobig_iff ); 1455 transform_later(toobig_false); 1456 } else { // No initial test, just fall into next case 1457 toobig_false = ctrl; 1458 debug_only(slow_region = NodeSentinel); 1459 } 1460 1461 Node *slow_mem = mem; // save the current memory state for slow path 1462 // generate the fast allocation code unless we know that the initial test will always go slow 1463 if (!always_slow) { 1464 // Fast path modifies only raw memory. 1465 if (mem->is_MergeMem()) { 1466 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1467 } 1468 1469 Node* eden_top_adr; 1470 Node* eden_end_adr; 1471 1472 set_eden_pointers(eden_top_adr, eden_end_adr); 1473 1474 // Load Eden::end. Loop invariant and hoisted. 1475 // 1476 // Note: We set the control input on "eden_end" and "old_eden_top" when using 1477 // a TLAB to work around a bug where these values were being moved across 1478 // a safepoint. These are not oops, so they cannot be include in the oop 1479 // map, but they can be changed by a GC. The proper way to fix this would 1480 // be to set the raw memory state when generating a SafepointNode. However 1481 // this will require extensive changes to the loop optimization in order to 1482 // prevent a degradation of the optimization. 1483 // See comment in memnode.hpp, around line 227 in class LoadPNode. 1484 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 1485 1486 // allocate the Region and Phi nodes for the result 1487 result_region = new RegionNode(3); 1488 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1489 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM); 1490 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch 1491 1492 // We need a Region for the loop-back contended case. 1493 enum { fall_in_path = 1, contended_loopback_path = 2 }; 1494 Node *contended_region; 1495 Node *contended_phi_rawmem; 1496 if (UseTLAB) { 1497 contended_region = toobig_false; 1498 contended_phi_rawmem = mem; 1499 } else { 1500 contended_region = new RegionNode(3); 1501 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1502 // Now handle the passing-too-big test. We fall into the contended 1503 // loop-back merge point. 1504 contended_region ->init_req(fall_in_path, toobig_false); 1505 contended_phi_rawmem->init_req(fall_in_path, mem); 1506 transform_later(contended_region); 1507 transform_later(contended_phi_rawmem); 1508 } 1509 1510 // Load(-locked) the heap top. 1511 // See note above concerning the control input when using a TLAB 1512 Node *old_eden_top = UseTLAB 1513 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered) 1514 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire); 1515 1516 transform_later(old_eden_top); 1517 // Add to heap top to get a new heap top 1518 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes); 1519 transform_later(new_eden_top); 1520 // Check for needing a GC; compare against heap end 1521 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end); 1522 transform_later(needgc_cmp); 1523 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge); 1524 transform_later(needgc_bol); 1525 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); 1526 transform_later(needgc_iff); 1527 1528 // Plug the failing-heap-space-need-gc test into the slow-path region 1529 Node *needgc_true = new IfTrueNode(needgc_iff); 1530 transform_later(needgc_true); 1531 if (initial_slow_test) { 1532 slow_region->init_req(need_gc_path, needgc_true); 1533 // This completes all paths into the slow merge point 1534 transform_later(slow_region); 1535 } else { // No initial slow path needed! 1536 // Just fall from the need-GC path straight into the VM call. 1537 slow_region = needgc_true; 1538 } 1539 // No need for a GC. Setup for the Store-Conditional 1540 Node *needgc_false = new IfFalseNode(needgc_iff); 1541 transform_later(needgc_false); 1542 1543 // Grab regular I/O before optional prefetch may change it. 1544 // Slow-path does no I/O so just set it to the original I/O. 1545 result_phi_i_o->init_req(slow_result_path, i_o); 1546 1547 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem, 1548 old_eden_top, new_eden_top, length); 1549 1550 // Name successful fast-path variables 1551 Node* fast_oop = old_eden_top; 1552 Node* fast_oop_ctrl; 1553 Node* fast_oop_rawmem; 1554 1555 // Store (-conditional) the modified eden top back down. 1556 // StorePConditional produces flags for a test PLUS a modified raw 1557 // memory state. 1558 if (UseTLAB) { 1559 Node* store_eden_top = 1560 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1561 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered); 1562 transform_later(store_eden_top); 1563 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path 1564 fast_oop_rawmem = store_eden_top; 1565 } else { 1566 Node* store_eden_top = 1567 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1568 new_eden_top, fast_oop/*old_eden_top*/); 1569 transform_later(store_eden_top); 1570 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne); 1571 transform_later(contention_check); 1572 store_eden_top = new SCMemProjNode(store_eden_top); 1573 transform_later(store_eden_top); 1574 1575 // If not using TLABs, check to see if there was contention. 1576 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN); 1577 transform_later(contention_iff); 1578 Node *contention_true = new IfTrueNode(contention_iff); 1579 transform_later(contention_true); 1580 // If contention, loopback and try again. 1581 contended_region->init_req(contended_loopback_path, contention_true); 1582 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top); 1583 1584 // Fast-path succeeded with no contention! 1585 Node *contention_false = new IfFalseNode(contention_iff); 1586 transform_later(contention_false); 1587 fast_oop_ctrl = contention_false; 1588 1589 // Bump total allocated bytes for this thread 1590 Node* thread = new ThreadLocalNode(); 1591 transform_later(thread); 1592 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread, 1593 in_bytes(JavaThread::allocated_bytes_offset())); 1594 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1595 0, TypeLong::LONG, T_LONG); 1596 #ifdef _LP64 1597 Node* alloc_size = size_in_bytes; 1598 #else 1599 Node* alloc_size = new ConvI2LNode(size_in_bytes); 1600 transform_later(alloc_size); 1601 #endif 1602 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size); 1603 transform_later(new_alloc_bytes); 1604 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1605 0, new_alloc_bytes, T_LONG); 1606 } 1607 1608 InitializeNode* init = alloc->initialization(); 1609 fast_oop_rawmem = initialize_object(alloc, 1610 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1611 klass_node, length, size_in_bytes); 1612 1613 // If initialization is performed by an array copy, any required 1614 // MemBarStoreStore was already added. If the object does not 1615 // escape no need for a MemBarStoreStore. If the object does not 1616 // escape in its initializer and memory barrier (MemBarStoreStore or 1617 // stronger) is already added at exit of initializer, also no need 1618 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore 1619 // so that stores that initialize this object can't be reordered 1620 // with a subsequent store that makes this object accessible by 1621 // other threads. 1622 // Other threads include java threads and JVM internal threads 1623 // (for example concurrent GC threads). Current concurrent GC 1624 // implementation: CMS and G1 will not scan newly created object, 1625 // so it's safe to skip storestore barrier when allocation does 1626 // not escape. 1627 if (!alloc->does_not_escape_thread() && 1628 !alloc->is_allocation_MemBar_redundant() && 1629 (init == NULL || !init->is_complete_with_arraycopy())) { 1630 if (init == NULL || init->req() < InitializeNode::RawStores) { 1631 // No InitializeNode or no stores captured by zeroing 1632 // elimination. Simply add the MemBarStoreStore after object 1633 // initialization. 1634 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1635 transform_later(mb); 1636 1637 mb->init_req(TypeFunc::Memory, fast_oop_rawmem); 1638 mb->init_req(TypeFunc::Control, fast_oop_ctrl); 1639 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control); 1640 transform_later(fast_oop_ctrl); 1641 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory); 1642 transform_later(fast_oop_rawmem); 1643 } else { 1644 // Add the MemBarStoreStore after the InitializeNode so that 1645 // all stores performing the initialization that were moved 1646 // before the InitializeNode happen before the storestore 1647 // barrier. 1648 1649 Node* init_ctrl = init->proj_out(TypeFunc::Control); 1650 Node* init_mem = init->proj_out(TypeFunc::Memory); 1651 1652 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1653 transform_later(mb); 1654 1655 Node* ctrl = new ProjNode(init,TypeFunc::Control); 1656 transform_later(ctrl); 1657 Node* mem = new ProjNode(init,TypeFunc::Memory); 1658 transform_later(mem); 1659 1660 // The MemBarStoreStore depends on control and memory coming 1661 // from the InitializeNode 1662 mb->init_req(TypeFunc::Memory, mem); 1663 mb->init_req(TypeFunc::Control, ctrl); 1664 1665 ctrl = new ProjNode(mb,TypeFunc::Control); 1666 transform_later(ctrl); 1667 mem = new ProjNode(mb,TypeFunc::Memory); 1668 transform_later(mem); 1669 1670 // All nodes that depended on the InitializeNode for control 1671 // and memory must now depend on the MemBarNode that itself 1672 // depends on the InitializeNode 1673 if (init_ctrl != NULL) { 1674 _igvn.replace_node(init_ctrl, ctrl); 1675 } 1676 if (init_mem != NULL) { 1677 _igvn.replace_node(init_mem, mem); 1678 } 1679 } 1680 } 1681 1682 if (C->env()->dtrace_extended_probes()) { 1683 // Slow-path call 1684 int size = TypeFunc::Parms + 2; 1685 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1686 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), 1687 "dtrace_object_alloc", 1688 TypeRawPtr::BOTTOM); 1689 1690 // Get base of thread-local storage area 1691 Node* thread = new ThreadLocalNode(); 1692 transform_later(thread); 1693 1694 call->init_req(TypeFunc::Parms+0, thread); 1695 call->init_req(TypeFunc::Parms+1, fast_oop); 1696 call->init_req(TypeFunc::Control, fast_oop_ctrl); 1697 call->init_req(TypeFunc::I_O , top()); // does no i/o 1698 call->init_req(TypeFunc::Memory , fast_oop_rawmem); 1699 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1700 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1701 transform_later(call); 1702 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control); 1703 transform_later(fast_oop_ctrl); 1704 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory); 1705 transform_later(fast_oop_rawmem); 1706 } 1707 1708 if (HeapMonitoring::initialized()) { 1709 // Inlined version of HeapMonitoring::object_alloc_base 1710 // Get base of thread-local storage area 1711 Node* thread = new ThreadLocalNode(); 1712 transform_later(thread); 1713 1714 ByteSize sample_offset = JavaThread::bytes_until_sample_offset(); 1715 1716 // Do test to see if we should sample. 1717 // Get bytes_until_sample from thread local storage. 1718 Node *bytes_until_sample = make_load(fast_oop_ctrl, 1719 fast_oop_rawmem, 1720 thread, 1721 in_bytes(sample_offset), 1722 TypeX_X, 1723 TypeX_X->basic_type()); 1724 1725 // new_bytes_until_sample = bytes_until_sample - size_in_bytes 1726 Node *new_bytes_until_sample = 1727 new SubXNode(bytes_until_sample, size_in_bytes); 1728 transform_later(new_bytes_until_sample); 1729 1730 // bytes_until_sample = new_bytes_until_sample; 1731 fast_oop_rawmem = make_store(fast_oop_ctrl, 1732 fast_oop_rawmem, 1733 thread, 1734 in_bytes(sample_offset), 1735 new_bytes_until_sample, 1736 TypeX_X->basic_type()); 1737 1738 // Call to make if sampling succeeds 1739 CallLeafNode *call = new CallLeafNode( 1740 OptoRuntime::heap_object_alloc_Type(), 1741 CAST_FROM_FN_PTR(address, 1742 HeapMonitoring::object_alloc_do_sample), 1743 "object_alloc_do_sample", 1744 TypeRawPtr::BOTTOM); 1745 1746 // Copy debug information and adjust JVMState information: 1747 // We are copying the debug information because the allocation JVMS 1748 // might be modified/removed, etc. But also we want to preserve the JVMS 1749 // in case this node is from an inlined method. 1750 copy_call_debug_info((CallNode *) alloc, call); 1751 1752 // if (new_bytes_until_sample < 0) 1753 conditional_sample(new_bytes_until_sample, 1754 BoolTest::le, 1755 // Probability 1756 // ~1/10000 1757 PROB_UNLIKELY_MAG(4), 1758 call, 1759 thread, 1760 &fast_oop_ctrl, 1761 &fast_oop_rawmem, 1762 &fast_oop, 1763 size_in_bytes, 1764 alloc); 1765 } 1766 1767 // Plug in the successful fast-path into the result merge point 1768 result_region ->init_req(fast_result_path, fast_oop_ctrl); 1769 result_phi_rawoop->init_req(fast_result_path, fast_oop); 1770 result_phi_i_o ->init_req(fast_result_path, i_o); 1771 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); 1772 } else { 1773 slow_region = ctrl; 1774 result_phi_i_o = i_o; // Rename it to use in the following code. 1775 } 1776 1777 // Generate slow-path call 1778 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address, 1779 OptoRuntime::stub_name(slow_call_address), 1780 alloc->jvms()->bci(), 1781 TypePtr::BOTTOM); 1782 call->init_req( TypeFunc::Control, slow_region ); 1783 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1784 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs 1785 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1786 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1787 1788 call->init_req(TypeFunc::Parms+0, klass_node); 1789 if (length != NULL) { 1790 call->init_req(TypeFunc::Parms+1, length); 1791 } 1792 1793 // Copy debug information and adjust JVMState information, then replace 1794 // allocate node with the call 1795 copy_call_debug_info((CallNode *) alloc, call); 1796 if (!always_slow) { 1797 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1798 } else { 1799 // Hook i_o projection to avoid its elimination during allocation 1800 // replacement (when only a slow call is generated). 1801 call->set_req(TypeFunc::I_O, result_phi_i_o); 1802 } 1803 _igvn.replace_node(alloc, call); 1804 transform_later(call); 1805 1806 // Identify the output projections from the allocate node and 1807 // adjust any references to them. 1808 // The control and io projections look like: 1809 // 1810 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1811 // Allocate Catch 1812 // ^---Proj(io) <-------+ ^---CatchProj(io) 1813 // 1814 // We are interested in the CatchProj nodes. 1815 // 1816 extract_call_projections(call); 1817 1818 // An allocate node has separate memory projections for the uses on 1819 // the control and i_o paths. Replace the control memory projection with 1820 // result_phi_rawmem (unless we are only generating a slow call when 1821 // both memory projections are combined) 1822 if (!always_slow && _memproj_fallthrough != NULL) { 1823 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { 1824 Node *use = _memproj_fallthrough->fast_out(i); 1825 _igvn.rehash_node_delayed(use); 1826 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); 1827 // back up iterator 1828 --i; 1829 } 1830 } 1831 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete 1832 // _memproj_catchall so we end up with a call that has only 1 memory projection. 1833 if (_memproj_catchall != NULL ) { 1834 if (_memproj_fallthrough == NULL) { 1835 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory); 1836 transform_later(_memproj_fallthrough); 1837 } 1838 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { 1839 Node *use = _memproj_catchall->fast_out(i); 1840 _igvn.rehash_node_delayed(use); 1841 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); 1842 // back up iterator 1843 --i; 1844 } 1845 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted"); 1846 _igvn.remove_dead_node(_memproj_catchall); 1847 } 1848 1849 // An allocate node has separate i_o projections for the uses on the control 1850 // and i_o paths. Always replace the control i_o projection with result i_o 1851 // otherwise incoming i_o become dead when only a slow call is generated 1852 // (it is different from memory projections where both projections are 1853 // combined in such case). 1854 if (_ioproj_fallthrough != NULL) { 1855 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { 1856 Node *use = _ioproj_fallthrough->fast_out(i); 1857 _igvn.rehash_node_delayed(use); 1858 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); 1859 // back up iterator 1860 --i; 1861 } 1862 } 1863 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete 1864 // _ioproj_catchall so we end up with a call that has only 1 i_o projection. 1865 if (_ioproj_catchall != NULL ) { 1866 if (_ioproj_fallthrough == NULL) { 1867 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O); 1868 transform_later(_ioproj_fallthrough); 1869 } 1870 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { 1871 Node *use = _ioproj_catchall->fast_out(i); 1872 _igvn.rehash_node_delayed(use); 1873 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); 1874 // back up iterator 1875 --i; 1876 } 1877 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted"); 1878 _igvn.remove_dead_node(_ioproj_catchall); 1879 } 1880 1881 // if we generated only a slow call, we are done 1882 if (always_slow) { 1883 // Now we can unhook i_o. 1884 if (result_phi_i_o->outcnt() > 1) { 1885 call->set_req(TypeFunc::I_O, top()); 1886 } else { 1887 assert(result_phi_i_o->unique_ctrl_out() == call, ""); 1888 // Case of new array with negative size known during compilation. 1889 // AllocateArrayNode::Ideal() optimization disconnect unreachable 1890 // following code since call to runtime will throw exception. 1891 // As result there will be no users of i_o after the call. 1892 // Leave i_o attached to this call to avoid problems in preceding graph. 1893 } 1894 return; 1895 } 1896 1897 1898 if (_fallthroughcatchproj != NULL) { 1899 ctrl = _fallthroughcatchproj->clone(); 1900 transform_later(ctrl); 1901 _igvn.replace_node(_fallthroughcatchproj, result_region); 1902 } else { 1903 ctrl = top(); 1904 } 1905 Node *slow_result; 1906 if (_resproj == NULL) { 1907 // no uses of the allocation result 1908 slow_result = top(); 1909 } else { 1910 slow_result = _resproj->clone(); 1911 transform_later(slow_result); 1912 _igvn.replace_node(_resproj, result_phi_rawoop); 1913 } 1914 1915 // Plug slow-path into result merge point 1916 result_region ->init_req( slow_result_path, ctrl ); 1917 result_phi_rawoop->init_req( slow_result_path, slow_result); 1918 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); 1919 transform_later(result_region); 1920 transform_later(result_phi_rawoop); 1921 transform_later(result_phi_rawmem); 1922 transform_later(result_phi_i_o); 1923 // This completes all paths into the result merge point 1924 } 1925 1926 1927 // Helper for PhaseMacroExpand::expand_allocate_common. 1928 // Initializes the newly-allocated storage. 1929 Node* 1930 PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1931 Node* control, Node* rawmem, Node* object, 1932 Node* klass_node, Node* length, 1933 Node* size_in_bytes) { 1934 InitializeNode* init = alloc->initialization(); 1935 // Store the klass & mark bits 1936 Node* mark_node = NULL; 1937 // For now only enable fast locking for non-array types 1938 if (UseBiasedLocking && (length == NULL)) { 1939 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 1940 } else { 1941 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); 1942 } 1943 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 1944 1945 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 1946 int header_size = alloc->minimum_header_size(); // conservatively small 1947 1948 // Array length 1949 if (length != NULL) { // Arrays need length field 1950 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1951 // conservatively small header size: 1952 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1953 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1954 if (k->is_array_klass()) // we know the exact header size in most cases: 1955 header_size = Klass::layout_helper_header_size(k->layout_helper()); 1956 } 1957 1958 // Clear the object body, if necessary. 1959 if (init == NULL) { 1960 // The init has somehow disappeared; be cautious and clear everything. 1961 // 1962 // This can happen if a node is allocated but an uncommon trap occurs 1963 // immediately. In this case, the Initialize gets associated with the 1964 // trap, and may be placed in a different (outer) loop, if the Allocate 1965 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1966 // there can be two Allocates to one Initialize. The answer in all these 1967 // edge cases is safety first. It is always safe to clear immediately 1968 // within an Allocate, and then (maybe or maybe not) clear some more later. 1969 if (!(UseTLAB && ZeroTLAB)) { 1970 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1971 header_size, size_in_bytes, 1972 &_igvn); 1973 } 1974 } else { 1975 if (!init->is_complete()) { 1976 // Try to win by zeroing only what the init does not store. 1977 // We can also try to do some peephole optimizations, 1978 // such as combining some adjacent subword stores. 1979 rawmem = init->complete_stores(control, rawmem, object, 1980 header_size, size_in_bytes, &_igvn); 1981 } 1982 // We have no more use for this link, since the AllocateNode goes away: 1983 init->set_req(InitializeNode::RawAddress, top()); 1984 // (If we keep the link, it just confuses the register allocator, 1985 // who thinks he sees a real use of the address by the membar.) 1986 } 1987 1988 return rawmem; 1989 } 1990 1991 // Generate prefetch instructions for next allocations. 1992 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1993 Node*& contended_phi_rawmem, 1994 Node* old_eden_top, Node* new_eden_top, 1995 Node* length) { 1996 enum { fall_in_path = 1, pf_path = 2 }; 1997 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1998 // Generate prefetch allocation with watermark check. 1999 // As an allocation hits the watermark, we will prefetch starting 2000 // at a "distance" away from watermark. 2001 2002 Node *pf_region = new RegionNode(3); 2003 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 2004 TypeRawPtr::BOTTOM ); 2005 // I/O is used for Prefetch 2006 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO ); 2007 2008 Node *thread = new ThreadLocalNode(); 2009 transform_later(thread); 2010 2011 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread, 2012 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 2013 transform_later(eden_pf_adr); 2014 2015 Node *old_pf_wm = new LoadPNode(needgc_false, 2016 contended_phi_rawmem, eden_pf_adr, 2017 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, 2018 MemNode::unordered); 2019 transform_later(old_pf_wm); 2020 2021 // check against new_eden_top 2022 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm ); 2023 transform_later(need_pf_cmp); 2024 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge ); 2025 transform_later(need_pf_bol); 2026 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol, 2027 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 2028 transform_later(need_pf_iff); 2029 2030 // true node, add prefetchdistance 2031 Node *need_pf_true = new IfTrueNode( need_pf_iff ); 2032 transform_later(need_pf_true); 2033 2034 Node *need_pf_false = new IfFalseNode( need_pf_iff ); 2035 transform_later(need_pf_false); 2036 2037 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm, 2038 _igvn.MakeConX(AllocatePrefetchDistance) ); 2039 transform_later(new_pf_wmt ); 2040 new_pf_wmt->set_req(0, need_pf_true); 2041 2042 Node *store_new_wmt = new StorePNode(need_pf_true, 2043 contended_phi_rawmem, eden_pf_adr, 2044 TypeRawPtr::BOTTOM, new_pf_wmt, 2045 MemNode::unordered); 2046 transform_later(store_new_wmt); 2047 2048 // adding prefetches 2049 pf_phi_abio->init_req( fall_in_path, i_o ); 2050 2051 Node *prefetch_adr; 2052 Node *prefetch; 2053 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 2054 uint step_size = AllocatePrefetchStepSize; 2055 uint distance = 0; 2056 2057 for ( uint i = 0; i < lines; i++ ) { 2058 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt, 2059 _igvn.MakeConX(distance) ); 2060 transform_later(prefetch_adr); 2061 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 2062 transform_later(prefetch); 2063 distance += step_size; 2064 i_o = prefetch; 2065 } 2066 pf_phi_abio->set_req( pf_path, i_o ); 2067 2068 pf_region->init_req( fall_in_path, need_pf_false ); 2069 pf_region->init_req( pf_path, need_pf_true ); 2070 2071 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 2072 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 2073 2074 transform_later(pf_region); 2075 transform_later(pf_phi_rawmem); 2076 transform_later(pf_phi_abio); 2077 2078 needgc_false = pf_region; 2079 contended_phi_rawmem = pf_phi_rawmem; 2080 i_o = pf_phi_abio; 2081 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { 2082 // Insert a prefetch instruction for each allocation. 2083 // This code is used to generate 1 prefetch instruction per cache line. 2084 2085 // Generate several prefetch instructions. 2086 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 2087 uint step_size = AllocatePrefetchStepSize; 2088 uint distance = AllocatePrefetchDistance; 2089 2090 // Next cache address. 2091 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top, 2092 _igvn.MakeConX(step_size + distance)); 2093 transform_later(cache_adr); 2094 cache_adr = new CastP2XNode(needgc_false, cache_adr); 2095 transform_later(cache_adr); 2096 // Address is aligned to execute prefetch to the beginning of cache line size 2097 // (it is important when BIS instruction is used on SPARC as prefetch). 2098 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); 2099 cache_adr = new AndXNode(cache_adr, mask); 2100 transform_later(cache_adr); 2101 cache_adr = new CastX2PNode(cache_adr); 2102 transform_later(cache_adr); 2103 2104 // Prefetch 2105 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); 2106 prefetch->set_req(0, needgc_false); 2107 transform_later(prefetch); 2108 contended_phi_rawmem = prefetch; 2109 Node *prefetch_adr; 2110 distance = step_size; 2111 for ( uint i = 1; i < lines; i++ ) { 2112 prefetch_adr = new AddPNode( cache_adr, cache_adr, 2113 _igvn.MakeConX(distance) ); 2114 transform_later(prefetch_adr); 2115 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); 2116 transform_later(prefetch); 2117 distance += step_size; 2118 contended_phi_rawmem = prefetch; 2119 } 2120 } else if( AllocatePrefetchStyle > 0 ) { 2121 // Insert a prefetch for each allocation only on the fast-path 2122 Node *prefetch_adr; 2123 Node *prefetch; 2124 // Generate several prefetch instructions. 2125 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 2126 uint step_size = AllocatePrefetchStepSize; 2127 uint distance = AllocatePrefetchDistance; 2128 for ( uint i = 0; i < lines; i++ ) { 2129 prefetch_adr = new AddPNode( old_eden_top, new_eden_top, 2130 _igvn.MakeConX(distance) ); 2131 transform_later(prefetch_adr); 2132 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 2133 // Do not let it float too high, since if eden_top == eden_end, 2134 // both might be null. 2135 if( i == 0 ) { // Set control for first prefetch, next follows it 2136 prefetch->init_req(0, needgc_false); 2137 } 2138 transform_later(prefetch); 2139 distance += step_size; 2140 i_o = prefetch; 2141 } 2142 } 2143 return i_o; 2144 } 2145 2146 2147 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 2148 expand_allocate_common(alloc, NULL, 2149 OptoRuntime::new_instance_Type(), 2150 OptoRuntime::new_instance_Java()); 2151 } 2152 2153 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 2154 Node* length = alloc->in(AllocateNode::ALength); 2155 InitializeNode* init = alloc->initialization(); 2156 Node* klass_node = alloc->in(AllocateNode::KlassNode); 2157 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 2158 address slow_call_address; // Address of slow call 2159 if (init != NULL && init->is_complete_with_arraycopy() && 2160 k->is_type_array_klass()) { 2161 // Don't zero type array during slow allocation in VM since 2162 // it will be initialized later by arraycopy in compiled code. 2163 slow_call_address = OptoRuntime::new_array_nozero_Java(); 2164 } else { 2165 slow_call_address = OptoRuntime::new_array_Java(); 2166 } 2167 expand_allocate_common(alloc, length, 2168 OptoRuntime::new_array_Type(), 2169 slow_call_address); 2170 } 2171 2172 //-------------------mark_eliminated_box---------------------------------- 2173 // 2174 // During EA obj may point to several objects but after few ideal graph 2175 // transformations (CCP) it may point to only one non escaping object 2176 // (but still using phi), corresponding locks and unlocks will be marked 2177 // for elimination. Later obj could be replaced with a new node (new phi) 2178 // and which does not have escape information. And later after some graph 2179 // reshape other locks and unlocks (which were not marked for elimination 2180 // before) are connected to this new obj (phi) but they still will not be 2181 // marked for elimination since new obj has no escape information. 2182 // Mark all associated (same box and obj) lock and unlock nodes for 2183 // elimination if some of them marked already. 2184 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) { 2185 if (oldbox->as_BoxLock()->is_eliminated()) 2186 return; // This BoxLock node was processed already. 2187 2188 // New implementation (EliminateNestedLocks) has separate BoxLock 2189 // node for each locked region so mark all associated locks/unlocks as 2190 // eliminated even if different objects are referenced in one locked region 2191 // (for example, OSR compilation of nested loop inside locked scope). 2192 if (EliminateNestedLocks || 2193 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) { 2194 // Box is used only in one lock region. Mark this box as eliminated. 2195 _igvn.hash_delete(oldbox); 2196 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value 2197 _igvn.hash_insert(oldbox); 2198 2199 for (uint i = 0; i < oldbox->outcnt(); i++) { 2200 Node* u = oldbox->raw_out(i); 2201 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) { 2202 AbstractLockNode* alock = u->as_AbstractLock(); 2203 // Check lock's box since box could be referenced by Lock's debug info. 2204 if (alock->box_node() == oldbox) { 2205 // Mark eliminated all related locks and unlocks. 2206 #ifdef ASSERT 2207 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4"); 2208 #endif 2209 alock->set_non_esc_obj(); 2210 } 2211 } 2212 } 2213 return; 2214 } 2215 2216 // Create new "eliminated" BoxLock node and use it in monitor debug info 2217 // instead of oldbox for the same object. 2218 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 2219 2220 // Note: BoxLock node is marked eliminated only here and it is used 2221 // to indicate that all associated lock and unlock nodes are marked 2222 // for elimination. 2223 newbox->set_eliminated(); 2224 transform_later(newbox); 2225 2226 // Replace old box node with new box for all users of the same object. 2227 for (uint i = 0; i < oldbox->outcnt();) { 2228 bool next_edge = true; 2229 2230 Node* u = oldbox->raw_out(i); 2231 if (u->is_AbstractLock()) { 2232 AbstractLockNode* alock = u->as_AbstractLock(); 2233 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) { 2234 // Replace Box and mark eliminated all related locks and unlocks. 2235 #ifdef ASSERT 2236 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5"); 2237 #endif 2238 alock->set_non_esc_obj(); 2239 _igvn.rehash_node_delayed(alock); 2240 alock->set_box_node(newbox); 2241 next_edge = false; 2242 } 2243 } 2244 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) { 2245 FastLockNode* flock = u->as_FastLock(); 2246 assert(flock->box_node() == oldbox, "sanity"); 2247 _igvn.rehash_node_delayed(flock); 2248 flock->set_box_node(newbox); 2249 next_edge = false; 2250 } 2251 2252 // Replace old box in monitor debug info. 2253 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 2254 SafePointNode* sfn = u->as_SafePoint(); 2255 JVMState* youngest_jvms = sfn->jvms(); 2256 int max_depth = youngest_jvms->depth(); 2257 for (int depth = 1; depth <= max_depth; depth++) { 2258 JVMState* jvms = youngest_jvms->of_depth(depth); 2259 int num_mon = jvms->nof_monitors(); 2260 // Loop over monitors 2261 for (int idx = 0; idx < num_mon; idx++) { 2262 Node* obj_node = sfn->monitor_obj(jvms, idx); 2263 Node* box_node = sfn->monitor_box(jvms, idx); 2264 if (box_node == oldbox && obj_node->eqv_uncast(obj)) { 2265 int j = jvms->monitor_box_offset(idx); 2266 _igvn.replace_input_of(u, j, newbox); 2267 next_edge = false; 2268 } 2269 } 2270 } 2271 } 2272 if (next_edge) i++; 2273 } 2274 } 2275 2276 //-----------------------mark_eliminated_locking_nodes----------------------- 2277 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { 2278 if (EliminateNestedLocks) { 2279 if (alock->is_nested()) { 2280 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity"); 2281 return; 2282 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened 2283 // Only Lock node has JVMState needed here. 2284 // Not that preceding claim is documented anywhere else. 2285 if (alock->jvms() != NULL) { 2286 if (alock->as_Lock()->is_nested_lock_region()) { 2287 // Mark eliminated related nested locks and unlocks. 2288 Node* obj = alock->obj_node(); 2289 BoxLockNode* box_node = alock->box_node()->as_BoxLock(); 2290 assert(!box_node->is_eliminated(), "should not be marked yet"); 2291 // Note: BoxLock node is marked eliminated only here 2292 // and it is used to indicate that all associated lock 2293 // and unlock nodes are marked for elimination. 2294 box_node->set_eliminated(); // Box's hash is always NO_HASH here 2295 for (uint i = 0; i < box_node->outcnt(); i++) { 2296 Node* u = box_node->raw_out(i); 2297 if (u->is_AbstractLock()) { 2298 alock = u->as_AbstractLock(); 2299 if (alock->box_node() == box_node) { 2300 // Verify that this Box is referenced only by related locks. 2301 assert(alock->obj_node()->eqv_uncast(obj), ""); 2302 // Mark all related locks and unlocks. 2303 #ifdef ASSERT 2304 alock->log_lock_optimization(C, "eliminate_lock_set_nested"); 2305 #endif 2306 alock->set_nested(); 2307 } 2308 } 2309 } 2310 } else { 2311 #ifdef ASSERT 2312 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region"); 2313 if (C->log() != NULL) 2314 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output 2315 #endif 2316 } 2317 } 2318 return; 2319 } 2320 // Process locks for non escaping object 2321 assert(alock->is_non_esc_obj(), ""); 2322 } // EliminateNestedLocks 2323 2324 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object 2325 // Look for all locks of this object and mark them and 2326 // corresponding BoxLock nodes as eliminated. 2327 Node* obj = alock->obj_node(); 2328 for (uint j = 0; j < obj->outcnt(); j++) { 2329 Node* o = obj->raw_out(j); 2330 if (o->is_AbstractLock() && 2331 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) { 2332 alock = o->as_AbstractLock(); 2333 Node* box = alock->box_node(); 2334 // Replace old box node with new eliminated box for all users 2335 // of the same object and mark related locks as eliminated. 2336 mark_eliminated_box(box, obj); 2337 } 2338 } 2339 } 2340 } 2341 2342 // we have determined that this lock/unlock can be eliminated, we simply 2343 // eliminate the node without expanding it. 2344 // 2345 // Note: The membar's associated with the lock/unlock are currently not 2346 // eliminated. This should be investigated as a future enhancement. 2347 // 2348 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 2349 2350 if (!alock->is_eliminated()) { 2351 return false; 2352 } 2353 #ifdef ASSERT 2354 if (!alock->is_coarsened()) { 2355 // Check that new "eliminated" BoxLock node is created. 2356 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 2357 assert(oldbox->is_eliminated(), "should be done already"); 2358 } 2359 #endif 2360 2361 alock->log_lock_optimization(C, "eliminate_lock"); 2362 2363 #ifndef PRODUCT 2364 if (PrintEliminateLocks) { 2365 if (alock->is_Lock()) { 2366 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx); 2367 } else { 2368 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx); 2369 } 2370 } 2371 #endif 2372 2373 Node* mem = alock->in(TypeFunc::Memory); 2374 Node* ctrl = alock->in(TypeFunc::Control); 2375 2376 extract_call_projections(alock); 2377 // There are 2 projections from the lock. The lock node will 2378 // be deleted when its last use is subsumed below. 2379 assert(alock->outcnt() == 2 && 2380 _fallthroughproj != NULL && 2381 _memproj_fallthrough != NULL, 2382 "Unexpected projections from Lock/Unlock"); 2383 2384 Node* fallthroughproj = _fallthroughproj; 2385 Node* memproj_fallthrough = _memproj_fallthrough; 2386 2387 // The memory projection from a lock/unlock is RawMem 2388 // The input to a Lock is merged memory, so extract its RawMem input 2389 // (unless the MergeMem has been optimized away.) 2390 if (alock->is_Lock()) { 2391 // Seach for MemBarAcquireLock node and delete it also. 2392 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 2393 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, ""); 2394 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 2395 Node* memproj = membar->proj_out(TypeFunc::Memory); 2396 _igvn.replace_node(ctrlproj, fallthroughproj); 2397 _igvn.replace_node(memproj, memproj_fallthrough); 2398 2399 // Delete FastLock node also if this Lock node is unique user 2400 // (a loop peeling may clone a Lock node). 2401 Node* flock = alock->as_Lock()->fastlock_node(); 2402 if (flock->outcnt() == 1) { 2403 assert(flock->unique_out() == alock, "sanity"); 2404 _igvn.replace_node(flock, top()); 2405 } 2406 } 2407 2408 // Seach for MemBarReleaseLock node and delete it also. 2409 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() && 2410 ctrl->in(0)->is_MemBar()) { 2411 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 2412 assert(membar->Opcode() == Op_MemBarReleaseLock && 2413 mem->is_Proj() && membar == mem->in(0), ""); 2414 _igvn.replace_node(fallthroughproj, ctrl); 2415 _igvn.replace_node(memproj_fallthrough, mem); 2416 fallthroughproj = ctrl; 2417 memproj_fallthrough = mem; 2418 ctrl = membar->in(TypeFunc::Control); 2419 mem = membar->in(TypeFunc::Memory); 2420 } 2421 2422 _igvn.replace_node(fallthroughproj, ctrl); 2423 _igvn.replace_node(memproj_fallthrough, mem); 2424 return true; 2425 } 2426 2427 2428 //------------------------------expand_lock_node---------------------- 2429 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 2430 2431 Node* ctrl = lock->in(TypeFunc::Control); 2432 Node* mem = lock->in(TypeFunc::Memory); 2433 Node* obj = lock->obj_node(); 2434 Node* box = lock->box_node(); 2435 Node* flock = lock->fastlock_node(); 2436 2437 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2438 2439 // Make the merge point 2440 Node *region; 2441 Node *mem_phi; 2442 Node *slow_path; 2443 2444 if (UseOptoBiasInlining) { 2445 /* 2446 * See the full description in MacroAssembler::biased_locking_enter(). 2447 * 2448 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { 2449 * // The object is biased. 2450 * proto_node = klass->prototype_header; 2451 * o_node = thread | proto_node; 2452 * x_node = o_node ^ mark_word; 2453 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? 2454 * // Done. 2455 * } else { 2456 * if( (x_node & biased_lock_mask) != 0 ) { 2457 * // The klass's prototype header is no longer biased. 2458 * cas(&mark_word, mark_word, proto_node) 2459 * goto cas_lock; 2460 * } else { 2461 * // The klass's prototype header is still biased. 2462 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? 2463 * old = mark_word; 2464 * new = o_node; 2465 * } else { 2466 * // Different thread or anonymous biased. 2467 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); 2468 * new = thread | old; 2469 * } 2470 * // Try to rebias. 2471 * if( cas(&mark_word, old, new) == 0 ) { 2472 * // Done. 2473 * } else { 2474 * goto slow_path; // Failed. 2475 * } 2476 * } 2477 * } 2478 * } else { 2479 * // The object is not biased. 2480 * cas_lock: 2481 * if( FastLock(obj) == 0 ) { 2482 * // Done. 2483 * } else { 2484 * slow_path: 2485 * OptoRuntime::complete_monitor_locking_Java(obj); 2486 * } 2487 * } 2488 */ 2489 2490 region = new RegionNode(5); 2491 // create a Phi for the memory state 2492 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2493 2494 Node* fast_lock_region = new RegionNode(3); 2495 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); 2496 2497 // First, check mark word for the biased lock pattern. 2498 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2499 2500 // Get fast path - mark word has the biased lock pattern. 2501 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, 2502 markOopDesc::biased_lock_mask_in_place, 2503 markOopDesc::biased_lock_pattern, true); 2504 // fast_lock_region->in(1) is set to slow path. 2505 fast_lock_mem_phi->init_req(1, mem); 2506 2507 // Now check that the lock is biased to the current thread and has 2508 // the same epoch and bias as Klass::_prototype_header. 2509 2510 // Special-case a fresh allocation to avoid building nodes: 2511 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); 2512 if (klass_node == NULL) { 2513 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 2514 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr())); 2515 #ifdef _LP64 2516 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) { 2517 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity"); 2518 klass_node->in(1)->init_req(0, ctrl); 2519 } else 2520 #endif 2521 klass_node->init_req(0, ctrl); 2522 } 2523 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type()); 2524 2525 Node* thread = transform_later(new ThreadLocalNode()); 2526 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2527 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node)); 2528 Node* x_node = transform_later(new XorXNode(o_node, mark_node)); 2529 2530 // Get slow path - mark word does NOT match the value. 2531 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, 2532 (~markOopDesc::age_mask_in_place), 0); 2533 // region->in(3) is set to fast path - the object is biased to the current thread. 2534 mem_phi->init_req(3, mem); 2535 2536 2537 // Mark word does NOT match the value (thread | Klass::_prototype_header). 2538 2539 2540 // First, check biased pattern. 2541 // Get fast path - _prototype_header has the same biased lock pattern. 2542 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, 2543 markOopDesc::biased_lock_mask_in_place, 0, true); 2544 2545 not_biased_ctrl = fast_lock_region->in(2); // Slow path 2546 // fast_lock_region->in(2) - the prototype header is no longer biased 2547 // and we have to revoke the bias on this object. 2548 // We are going to try to reset the mark of this object to the prototype 2549 // value and fall through to the CAS-based locking scheme. 2550 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 2551 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr, 2552 proto_node, mark_node); 2553 transform_later(cas); 2554 Node* proj = transform_later(new SCMemProjNode(cas)); 2555 fast_lock_mem_phi->init_req(2, proj); 2556 2557 2558 // Second, check epoch bits. 2559 Node* rebiased_region = new RegionNode(3); 2560 Node* old_phi = new PhiNode( rebiased_region, TypeX_X); 2561 Node* new_phi = new PhiNode( rebiased_region, TypeX_X); 2562 2563 // Get slow path - mark word does NOT match epoch bits. 2564 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, 2565 markOopDesc::epoch_mask_in_place, 0); 2566 // The epoch of the current bias is not valid, attempt to rebias the object 2567 // toward the current thread. 2568 rebiased_region->init_req(2, epoch_ctrl); 2569 old_phi->init_req(2, mark_node); 2570 new_phi->init_req(2, o_node); 2571 2572 // rebiased_region->in(1) is set to fast path. 2573 // The epoch of the current bias is still valid but we know 2574 // nothing about the owner; it might be set or it might be clear. 2575 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place | 2576 markOopDesc::age_mask_in_place | 2577 markOopDesc::epoch_mask_in_place); 2578 Node* old = transform_later(new AndXNode(mark_node, cmask)); 2579 cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2580 Node* new_mark = transform_later(new OrXNode(cast_thread, old)); 2581 old_phi->init_req(1, old); 2582 new_phi->init_req(1, new_mark); 2583 2584 transform_later(rebiased_region); 2585 transform_later(old_phi); 2586 transform_later(new_phi); 2587 2588 // Try to acquire the bias of the object using an atomic operation. 2589 // If this fails we will go in to the runtime to revoke the object's bias. 2590 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi); 2591 transform_later(cas); 2592 proj = transform_later(new SCMemProjNode(cas)); 2593 2594 // Get slow path - Failed to CAS. 2595 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); 2596 mem_phi->init_req(4, proj); 2597 // region->in(4) is set to fast path - the object is rebiased to the current thread. 2598 2599 // Failed to CAS. 2600 slow_path = new RegionNode(3); 2601 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); 2602 2603 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control 2604 slow_mem->init_req(1, proj); 2605 2606 // Call CAS-based locking scheme (FastLock node). 2607 2608 transform_later(fast_lock_region); 2609 transform_later(fast_lock_mem_phi); 2610 2611 // Get slow path - FastLock failed to lock the object. 2612 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); 2613 mem_phi->init_req(2, fast_lock_mem_phi); 2614 // region->in(2) is set to fast path - the object is locked to the current thread. 2615 2616 slow_path->init_req(2, ctrl); // Capture slow-control 2617 slow_mem->init_req(2, fast_lock_mem_phi); 2618 2619 transform_later(slow_path); 2620 transform_later(slow_mem); 2621 // Reset lock's memory edge. 2622 lock->set_req(TypeFunc::Memory, slow_mem); 2623 2624 } else { 2625 region = new RegionNode(3); 2626 // create a Phi for the memory state 2627 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2628 2629 // Optimize test; set region slot 2 2630 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); 2631 mem_phi->init_req(2, mem); 2632 } 2633 2634 // Make slow path call 2635 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), 2636 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, 2637 obj, box, NULL); 2638 2639 extract_call_projections(call); 2640 2641 // Slow path can only throw asynchronous exceptions, which are always 2642 // de-opted. So the compiler thinks the slow-call can never throw an 2643 // exception. If it DOES throw an exception we would need the debug 2644 // info removed first (since if it throws there is no monitor). 2645 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2646 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2647 2648 // Capture slow path 2649 // disconnect fall-through projection from call and create a new one 2650 // hook up users of fall-through projection to region 2651 Node *slow_ctrl = _fallthroughproj->clone(); 2652 transform_later(slow_ctrl); 2653 _igvn.hash_delete(_fallthroughproj); 2654 _fallthroughproj->disconnect_inputs(NULL, C); 2655 region->init_req(1, slow_ctrl); 2656 // region inputs are now complete 2657 transform_later(region); 2658 _igvn.replace_node(_fallthroughproj, region); 2659 2660 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory)); 2661 mem_phi->init_req(1, memproj ); 2662 transform_later(mem_phi); 2663 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2664 } 2665 2666 //------------------------------expand_unlock_node---------------------- 2667 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 2668 2669 Node* ctrl = unlock->in(TypeFunc::Control); 2670 Node* mem = unlock->in(TypeFunc::Memory); 2671 Node* obj = unlock->obj_node(); 2672 Node* box = unlock->box_node(); 2673 2674 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2675 2676 // No need for a null check on unlock 2677 2678 // Make the merge point 2679 Node *region; 2680 Node *mem_phi; 2681 2682 if (UseOptoBiasInlining) { 2683 // Check for biased locking unlock case, which is a no-op. 2684 // See the full description in MacroAssembler::biased_locking_exit(). 2685 region = new RegionNode(4); 2686 // create a Phi for the memory state 2687 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2688 mem_phi->init_req(3, mem); 2689 2690 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2691 ctrl = opt_bits_test(ctrl, region, 3, mark_node, 2692 markOopDesc::biased_lock_mask_in_place, 2693 markOopDesc::biased_lock_pattern); 2694 } else { 2695 region = new RegionNode(3); 2696 // create a Phi for the memory state 2697 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2698 } 2699 2700 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box ); 2701 funlock = transform_later( funlock )->as_FastUnlock(); 2702 // Optimize test; set region slot 2 2703 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); 2704 Node *thread = transform_later(new ThreadLocalNode()); 2705 2706 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), 2707 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), 2708 "complete_monitor_unlocking_C", slow_path, obj, box, thread); 2709 2710 extract_call_projections(call); 2711 2712 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2713 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2714 2715 // No exceptions for unlocking 2716 // Capture slow path 2717 // disconnect fall-through projection from call and create a new one 2718 // hook up users of fall-through projection to region 2719 Node *slow_ctrl = _fallthroughproj->clone(); 2720 transform_later(slow_ctrl); 2721 _igvn.hash_delete(_fallthroughproj); 2722 _fallthroughproj->disconnect_inputs(NULL, C); 2723 region->init_req(1, slow_ctrl); 2724 // region inputs are now complete 2725 transform_later(region); 2726 _igvn.replace_node(_fallthroughproj, region); 2727 2728 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) ); 2729 mem_phi->init_req(1, memproj ); 2730 mem_phi->init_req(2, mem); 2731 transform_later(mem_phi); 2732 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2733 } 2734 2735 //---------------------------eliminate_macro_nodes---------------------- 2736 // Eliminate scalar replaced allocations and associated locks. 2737 void PhaseMacroExpand::eliminate_macro_nodes() { 2738 if (C->macro_count() == 0) 2739 return; 2740 2741 // First, attempt to eliminate locks 2742 int cnt = C->macro_count(); 2743 for (int i=0; i < cnt; i++) { 2744 Node *n = C->macro_node(i); 2745 if (n->is_AbstractLock()) { // Lock and Unlock nodes 2746 // Before elimination mark all associated (same box and obj) 2747 // lock and unlock nodes. 2748 mark_eliminated_locking_nodes(n->as_AbstractLock()); 2749 } 2750 } 2751 bool progress = true; 2752 while (progress) { 2753 progress = false; 2754 for (int i = C->macro_count(); i > 0; i--) { 2755 Node * n = C->macro_node(i-1); 2756 bool success = false; 2757 debug_only(int old_macro_count = C->macro_count();); 2758 if (n->is_AbstractLock()) { 2759 success = eliminate_locking_node(n->as_AbstractLock()); 2760 } 2761 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2762 progress = progress || success; 2763 } 2764 } 2765 // Next, attempt to eliminate allocations 2766 _has_locks = false; 2767 progress = true; 2768 while (progress) { 2769 progress = false; 2770 for (int i = C->macro_count(); i > 0; i--) { 2771 Node * n = C->macro_node(i-1); 2772 bool success = false; 2773 debug_only(int old_macro_count = C->macro_count();); 2774 switch (n->class_id()) { 2775 case Node::Class_Allocate: 2776 case Node::Class_AllocateArray: 2777 success = eliminate_allocate_node(n->as_Allocate()); 2778 break; 2779 case Node::Class_CallStaticJava: 2780 success = eliminate_boxing_node(n->as_CallStaticJava()); 2781 break; 2782 case Node::Class_Lock: 2783 case Node::Class_Unlock: 2784 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 2785 _has_locks = true; 2786 break; 2787 case Node::Class_ArrayCopy: 2788 break; 2789 default: 2790 assert(n->Opcode() == Op_LoopLimit || 2791 n->Opcode() == Op_Opaque1 || 2792 n->Opcode() == Op_Opaque2 || 2793 n->Opcode() == Op_Opaque3, "unknown node type in macro list"); 2794 } 2795 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2796 progress = progress || success; 2797 } 2798 } 2799 } 2800 2801 //------------------------------expand_macro_nodes---------------------- 2802 // Returns true if a failure occurred. 2803 bool PhaseMacroExpand::expand_macro_nodes() { 2804 // Last attempt to eliminate macro nodes. 2805 eliminate_macro_nodes(); 2806 2807 // Make sure expansion will not cause node limit to be exceeded. 2808 // Worst case is a macro node gets expanded into about 200 nodes. 2809 // Allow 50% more for optimization. 2810 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) ) 2811 return true; 2812 2813 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. 2814 bool progress = true; 2815 while (progress) { 2816 progress = false; 2817 for (int i = C->macro_count(); i > 0; i--) { 2818 Node * n = C->macro_node(i-1); 2819 bool success = false; 2820 debug_only(int old_macro_count = C->macro_count();); 2821 if (n->Opcode() == Op_LoopLimit) { 2822 // Remove it from macro list and put on IGVN worklist to optimize. 2823 C->remove_macro_node(n); 2824 _igvn._worklist.push(n); 2825 success = true; 2826 } else if (n->Opcode() == Op_CallStaticJava) { 2827 // Remove it from macro list and put on IGVN worklist to optimize. 2828 C->remove_macro_node(n); 2829 _igvn._worklist.push(n); 2830 success = true; 2831 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { 2832 _igvn.replace_node(n, n->in(1)); 2833 success = true; 2834 #if INCLUDE_RTM_OPT 2835 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) { 2836 assert(C->profile_rtm(), "should be used only in rtm deoptimization code"); 2837 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), ""); 2838 Node* cmp = n->unique_out(); 2839 #ifdef ASSERT 2840 // Validate graph. 2841 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), ""); 2842 BoolNode* bol = cmp->unique_out()->as_Bool(); 2843 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() && 2844 (bol->_test._test == BoolTest::ne), ""); 2845 IfNode* ifn = bol->unique_out()->as_If(); 2846 assert((ifn->outcnt() == 2) && 2847 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, ""); 2848 #endif 2849 Node* repl = n->in(1); 2850 if (!_has_locks) { 2851 // Remove RTM state check if there are no locks in the code. 2852 // Replace input to compare the same value. 2853 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1); 2854 } 2855 _igvn.replace_node(n, repl); 2856 success = true; 2857 #endif 2858 } 2859 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2860 progress = progress || success; 2861 } 2862 } 2863 2864 // expand arraycopy "macro" nodes first 2865 // For ReduceBulkZeroing, we must first process all arraycopy nodes 2866 // before the allocate nodes are expanded. 2867 int macro_idx = C->macro_count() - 1; 2868 while (macro_idx >= 0) { 2869 Node * n = C->macro_node(macro_idx); 2870 assert(n->is_macro(), "only macro nodes expected here"); 2871 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2872 // node is unreachable, so don't try to expand it 2873 C->remove_macro_node(n); 2874 } else if (n->is_ArrayCopy()){ 2875 int macro_count = C->macro_count(); 2876 expand_arraycopy_node(n->as_ArrayCopy()); 2877 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2878 } 2879 if (C->failing()) return true; 2880 macro_idx --; 2881 } 2882 2883 // expand "macro" nodes 2884 // nodes are removed from the macro list as they are processed 2885 while (C->macro_count() > 0) { 2886 int macro_count = C->macro_count(); 2887 Node * n = C->macro_node(macro_count-1); 2888 assert(n->is_macro(), "only macro nodes expected here"); 2889 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2890 // node is unreachable, so don't try to expand it 2891 C->remove_macro_node(n); 2892 continue; 2893 } 2894 switch (n->class_id()) { 2895 case Node::Class_Allocate: 2896 expand_allocate(n->as_Allocate()); 2897 break; 2898 case Node::Class_AllocateArray: 2899 expand_allocate_array(n->as_AllocateArray()); 2900 break; 2901 case Node::Class_Lock: 2902 expand_lock_node(n->as_Lock()); 2903 break; 2904 case Node::Class_Unlock: 2905 expand_unlock_node(n->as_Unlock()); 2906 break; 2907 default: 2908 assert(false, "unknown node type in macro list"); 2909 } 2910 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2911 if (C->failing()) return true; 2912 } 2913 2914 _igvn.set_delay_transform(false); 2915 _igvn.optimize(); 2916 if (C->failing()) return true; 2917 return false; 2918 }