1 /* 2 * Copyright (c) 2005, 2017, 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 "opto/valuetypenode.hpp" 49 #include "runtime/sharedRuntime.hpp" 50 51 52 // 53 // Replace any references to "oldref" in inputs to "use" with "newref". 54 // Returns the number of replacements made. 55 // 56 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { 57 int nreplacements = 0; 58 uint req = use->req(); 59 for (uint j = 0; j < use->len(); j++) { 60 Node *uin = use->in(j); 61 if (uin == oldref) { 62 if (j < req) 63 use->set_req(j, newref); 64 else 65 use->set_prec(j, newref); 66 nreplacements++; 67 } else if (j >= req && uin == NULL) { 68 break; 69 } 70 } 71 return nreplacements; 72 } 73 74 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) { 75 // Copy debug information and adjust JVMState information 76 uint old_dbg_start = oldcall->tf()->domain_sig()->cnt(); 77 uint new_dbg_start = newcall->tf()->domain_sig()->cnt(); 78 int jvms_adj = new_dbg_start - old_dbg_start; 79 assert (new_dbg_start == newcall->req(), "argument count mismatch"); 80 81 // SafePointScalarObject node could be referenced several times in debug info. 82 // Use Dict to record cloned nodes. 83 Dict* sosn_map = new Dict(cmpkey,hashkey); 84 for (uint i = old_dbg_start; i < oldcall->req(); i++) { 85 Node* old_in = oldcall->in(i); 86 // Clone old SafePointScalarObjectNodes, adjusting their field contents. 87 if (old_in != NULL && old_in->is_SafePointScalarObject()) { 88 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject(); 89 uint old_unique = C->unique(); 90 Node* new_in = old_sosn->clone(sosn_map); 91 if (old_unique != C->unique()) { // New node? 92 new_in->set_req(0, C->root()); // reset control edge 93 new_in = transform_later(new_in); // Register new node. 94 } 95 old_in = new_in; 96 } 97 newcall->add_req(old_in); 98 } 99 100 // JVMS may be shared so clone it before we modify it 101 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL); 102 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) { 103 jvms->set_map(newcall); 104 jvms->set_locoff(jvms->locoff()+jvms_adj); 105 jvms->set_stkoff(jvms->stkoff()+jvms_adj); 106 jvms->set_monoff(jvms->monoff()+jvms_adj); 107 jvms->set_scloff(jvms->scloff()+jvms_adj); 108 jvms->set_endoff(jvms->endoff()+jvms_adj); 109 } 110 } 111 112 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) { 113 Node* cmp; 114 if (mask != 0) { 115 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask))); 116 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits))); 117 } else { 118 cmp = word; 119 } 120 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne)); 121 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); 122 transform_later(iff); 123 124 // Fast path taken. 125 Node *fast_taken = transform_later(new IfFalseNode(iff)); 126 127 // Fast path not-taken, i.e. slow path 128 Node *slow_taken = transform_later(new IfTrueNode(iff)); 129 130 if (return_fast_path) { 131 region->init_req(edge, slow_taken); // Capture slow-control 132 return fast_taken; 133 } else { 134 region->init_req(edge, fast_taken); // Capture fast-control 135 return slow_taken; 136 } 137 } 138 139 //--------------------copy_predefined_input_for_runtime_call-------------------- 140 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { 141 // Set fixed predefined input arguments 142 call->init_req( TypeFunc::Control, ctrl ); 143 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); 144 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? 145 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); 146 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); 147 } 148 149 //------------------------------make_slow_call--------------------------------- 150 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, 151 address slow_call, const char* leaf_name, Node* slow_path, 152 Node* parm0, Node* parm1, Node* parm2) { 153 154 // Slow-path call 155 CallNode *call = leaf_name 156 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) 157 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM ); 158 159 // Slow path call has no side-effects, uses few values 160 copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); 161 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0); 162 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1); 163 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2); 164 copy_call_debug_info(oldcall, call); 165 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 166 _igvn.replace_node(oldcall, call); 167 transform_later(call); 168 169 return call; 170 } 171 172 void PhaseMacroExpand::extract_call_projections(CallNode *call) { 173 _fallthroughproj = NULL; 174 _fallthroughcatchproj = NULL; 175 _ioproj_fallthrough = NULL; 176 _ioproj_catchall = NULL; 177 _catchallcatchproj = NULL; 178 _memproj_fallthrough = NULL; 179 _memproj_catchall = NULL; 180 _resproj = NULL; 181 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) { 182 ProjNode *pn = call->fast_out(i)->as_Proj(); 183 switch (pn->_con) { 184 case TypeFunc::Control: 185 { 186 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj 187 _fallthroughproj = pn; 188 DUIterator_Fast jmax, j = pn->fast_outs(jmax); 189 const Node *cn = pn->fast_out(j); 190 if (cn->is_Catch()) { 191 ProjNode *cpn = NULL; 192 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { 193 cpn = cn->fast_out(k)->as_Proj(); 194 assert(cpn->is_CatchProj(), "must be a CatchProjNode"); 195 if (cpn->_con == CatchProjNode::fall_through_index) 196 _fallthroughcatchproj = cpn; 197 else { 198 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); 199 _catchallcatchproj = cpn; 200 } 201 } 202 } 203 break; 204 } 205 case TypeFunc::I_O: 206 if (pn->_is_io_use) 207 _ioproj_catchall = pn; 208 else 209 _ioproj_fallthrough = pn; 210 break; 211 case TypeFunc::Memory: 212 if (pn->_is_io_use) 213 _memproj_catchall = pn; 214 else 215 _memproj_fallthrough = pn; 216 break; 217 case TypeFunc::Parms: 218 _resproj = pn; 219 break; 220 default: 221 assert(false, "unexpected projection from allocation node."); 222 } 223 } 224 225 } 226 227 // Eliminate a card mark sequence. p2x is a ConvP2XNode 228 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) { 229 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required"); 230 if (!UseG1GC) { 231 // vanilla/CMS post barrier 232 Node *shift = p2x->unique_out(); 233 Node *addp = shift->unique_out(); 234 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) { 235 Node *mem = addp->last_out(j); 236 if (UseCondCardMark && mem->is_Load()) { 237 assert(mem->Opcode() == Op_LoadB, "unexpected code shape"); 238 // The load is checking if the card has been written so 239 // replace it with zero to fold the test. 240 _igvn.replace_node(mem, intcon(0)); 241 continue; 242 } 243 assert(mem->is_Store(), "store required"); 244 _igvn.replace_node(mem, mem->in(MemNode::Memory)); 245 } 246 } else { 247 // G1 pre/post barriers 248 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes"); 249 // It could be only one user, URShift node, in Object.clone() intrinsic 250 // but the new allocation is passed to arraycopy stub and it could not 251 // be scalar replaced. So we don't check the case. 252 253 // An other case of only one user (Xor) is when the value check for NULL 254 // in G1 post barrier is folded after CCP so the code which used URShift 255 // is removed. 256 257 // Take Region node before eliminating post barrier since it also 258 // eliminates CastP2X node when it has only one user. 259 Node* this_region = p2x->in(0); 260 assert(this_region != NULL, ""); 261 262 // Remove G1 post barrier. 263 264 // Search for CastP2X->Xor->URShift->Cmp path which 265 // checks if the store done to a different from the value's region. 266 // And replace Cmp with #0 (false) to collapse G1 post barrier. 267 Node* xorx = p2x->find_out_with(Op_XorX); 268 if (xorx != NULL) { 269 Node* shift = xorx->unique_out(); 270 Node* cmpx = shift->unique_out(); 271 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() && 272 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne, 273 "missing region check in G1 post barrier"); 274 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 275 276 // Remove G1 pre barrier. 277 278 // Search "if (marking != 0)" check and set it to "false". 279 // There is no G1 pre barrier if previous stored value is NULL 280 // (for example, after initialization). 281 if (this_region->is_Region() && this_region->req() == 3) { 282 int ind = 1; 283 if (!this_region->in(ind)->is_IfFalse()) { 284 ind = 2; 285 } 286 if (this_region->in(ind)->is_IfFalse()) { 287 Node* bol = this_region->in(ind)->in(0)->in(1); 288 assert(bol->is_Bool(), ""); 289 cmpx = bol->in(1); 290 if (bol->as_Bool()->_test._test == BoolTest::ne && 291 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) && 292 cmpx->in(1)->is_Load()) { 293 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address); 294 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + 295 SATBMarkQueue::byte_offset_of_active()); 296 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() && 297 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 298 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) { 299 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 300 } 301 } 302 } 303 } 304 } else { 305 assert(!GraphKit::use_ReduceInitialCardMarks(), "can only happen with card marking"); 306 // This is a G1 post barrier emitted by the Object.clone() intrinsic. 307 // Search for the CastP2X->URShiftX->AddP->LoadB->Cmp path which checks if the card 308 // is marked as young_gen and replace the Cmp with 0 (false) to collapse the barrier. 309 Node* shift = p2x->find_out_with(Op_URShiftX); 310 assert(shift != NULL, "missing G1 post barrier"); 311 Node* addp = shift->unique_out(); 312 Node* load = addp->find_out_with(Op_LoadB); 313 assert(load != NULL, "missing G1 post barrier"); 314 Node* cmpx = load->unique_out(); 315 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() && 316 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne, 317 "missing card value check in G1 post barrier"); 318 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 319 // There is no G1 pre barrier in this case 320 } 321 // Now CastP2X can be removed since it is used only on dead path 322 // which currently still alive until igvn optimize it. 323 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, ""); 324 _igvn.replace_node(p2x, top()); 325 } 326 } 327 328 // Search for a memory operation for the specified memory slice. 329 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) { 330 Node *orig_mem = mem; 331 Node *alloc_mem = alloc->in(TypeFunc::Memory); 332 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr(); 333 while (true) { 334 if (mem == alloc_mem || mem == start_mem ) { 335 return mem; // hit one of our sentinels 336 } else if (mem->is_MergeMem()) { 337 mem = mem->as_MergeMem()->memory_at(alias_idx); 338 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { 339 Node *in = mem->in(0); 340 // we can safely skip over safepoints, calls, locks and membars because we 341 // already know that the object is safe to eliminate. 342 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { 343 return in; 344 } else if (in->is_Call()) { 345 CallNode *call = in->as_Call(); 346 if (call->may_modify(tinst, phase)) { 347 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape"); 348 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) { 349 return in; 350 } 351 } 352 mem = in->in(TypeFunc::Memory); 353 } else if (in->is_MemBar()) { 354 ArrayCopyNode* ac = NULL; 355 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) { 356 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone"); 357 return ac; 358 } 359 mem = in->in(TypeFunc::Memory); 360 } else { 361 assert(false, "unexpected projection"); 362 } 363 } else if (mem->is_Store()) { 364 const TypePtr* atype = mem->as_Store()->adr_type(); 365 int adr_idx = phase->C->get_alias_index(atype); 366 if (adr_idx == alias_idx) { 367 assert(atype->isa_oopptr(), "address type must be oopptr"); 368 int adr_offset = atype->offset(); 369 uint adr_iid = atype->is_oopptr()->instance_id(); 370 // Array elements references have the same alias_idx 371 // but different offset and different instance_id. 372 if (adr_offset == offset && adr_iid == alloc->_idx) 373 return mem; 374 } else { 375 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw"); 376 } 377 mem = mem->in(MemNode::Memory); 378 } else if (mem->is_ClearArray()) { 379 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) { 380 // Can not bypass initialization of the instance 381 // we are looking. 382 debug_only(intptr_t offset;) 383 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity"); 384 InitializeNode* init = alloc->as_Allocate()->initialization(); 385 // We are looking for stored value, return Initialize node 386 // or memory edge from Allocate node. 387 if (init != NULL) 388 return init; 389 else 390 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers). 391 } 392 // Otherwise skip it (the call updated 'mem' value). 393 } else if (mem->Opcode() == Op_SCMemProj) { 394 mem = mem->in(0); 395 Node* adr = NULL; 396 if (mem->is_LoadStore()) { 397 adr = mem->in(MemNode::Address); 398 } else { 399 assert(mem->Opcode() == Op_EncodeISOArray || 400 mem->Opcode() == Op_StrCompressedCopy, "sanity"); 401 adr = mem->in(3); // Destination array 402 } 403 const TypePtr* atype = adr->bottom_type()->is_ptr(); 404 int adr_idx = phase->C->get_alias_index(atype); 405 if (adr_idx == alias_idx) { 406 DEBUG_ONLY(mem->dump();) 407 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 408 return NULL; 409 } 410 mem = mem->in(MemNode::Memory); 411 } else if (mem->Opcode() == Op_StrInflatedCopy) { 412 Node* adr = mem->in(3); // Destination array 413 const TypePtr* atype = adr->bottom_type()->is_ptr(); 414 int adr_idx = phase->C->get_alias_index(atype); 415 if (adr_idx == alias_idx) { 416 DEBUG_ONLY(mem->dump();) 417 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field"); 418 return NULL; 419 } 420 mem = mem->in(MemNode::Memory); 421 } else { 422 return mem; 423 } 424 assert(mem != orig_mem, "dead memory loop"); 425 } 426 } 427 428 // Generate loads from source of the arraycopy for fields of 429 // destination needed at a deoptimization point 430 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) { 431 BasicType bt = ft; 432 const Type *type = ftype; 433 if (ft == T_NARROWOOP) { 434 bt = T_OBJECT; 435 type = ftype->make_oopptr(); 436 } 437 Node* res = NULL; 438 if (ac->is_clonebasic()) { 439 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base); 440 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset))); 441 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset); 442 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned); 443 } else { 444 if (ac->modifies(offset, offset, &_igvn, true)) { 445 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result"); 446 uint shift = exact_log2(type2aelembytes(bt)); 447 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 448 #ifdef _LP64 449 diff = _igvn.transform(new ConvI2LNode(diff)); 450 #endif 451 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift))); 452 453 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff)); 454 Node* base = ac->in(ArrayCopyNode::Src); 455 Node* adr = _igvn.transform(new AddPNode(base, base, off)); 456 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset); 457 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned); 458 } 459 } 460 if (res != NULL) { 461 res = _igvn.transform(res); 462 if (ftype->isa_narrowoop()) { 463 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes 464 res = _igvn.transform(new EncodePNode(res, ftype)); 465 } 466 return res; 467 } 468 return NULL; 469 } 470 471 // 472 // Given a Memory Phi, compute a value Phi containing the values from stores 473 // on the input paths. 474 // Note: this function is recursive, its depth is limited by the "level" argument 475 // Returns the computed Phi, or NULL if it cannot compute it. 476 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) { 477 assert(mem->is_Phi(), "sanity"); 478 int alias_idx = C->get_alias_index(adr_t); 479 int offset = adr_t->offset(); 480 int instance_id = adr_t->instance_id(); 481 482 // Check if an appropriate value phi already exists. 483 Node* region = mem->in(0); 484 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { 485 Node* phi = region->fast_out(k); 486 if (phi->is_Phi() && phi != mem && 487 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) { 488 return phi; 489 } 490 } 491 // Check if an appropriate new value phi already exists. 492 Node* new_phi = value_phis->find(mem->_idx); 493 if (new_phi != NULL) 494 return new_phi; 495 496 if (level <= 0) { 497 return NULL; // Give up: phi tree too deep 498 } 499 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 500 Node *alloc_mem = alloc->in(TypeFunc::Memory); 501 502 uint length = mem->req(); 503 GrowableArray <Node *> values(length, length, NULL, false); 504 505 // create a new Phi for the value 506 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset); 507 transform_later(phi); 508 value_phis->push(phi, mem->_idx); 509 510 for (uint j = 1; j < length; j++) { 511 Node *in = mem->in(j); 512 if (in == NULL || in->is_top()) { 513 values.at_put(j, in); 514 } else { 515 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); 516 if (val == start_mem || val == alloc_mem) { 517 // hit a sentinel, return appropriate 0 value 518 values.at_put(j, _igvn.zerocon(ft)); 519 continue; 520 } 521 if (val->is_Initialize()) { 522 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 523 } 524 if (val == NULL) { 525 return NULL; // can't find a value on this path 526 } 527 if (val == mem) { 528 values.at_put(j, mem); 529 } else if (val->is_Store()) { 530 values.at_put(j, val->in(MemNode::ValueIn)); 531 } else if(val->is_Proj() && val->in(0) == alloc) { 532 values.at_put(j, _igvn.zerocon(ft)); 533 } else if (val->is_Phi()) { 534 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); 535 if (val == NULL) { 536 return NULL; 537 } 538 values.at_put(j, val); 539 } else if (val->Opcode() == Op_SCMemProj) { 540 assert(val->in(0)->is_LoadStore() || 541 val->in(0)->Opcode() == Op_EncodeISOArray || 542 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity"); 543 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 544 return NULL; 545 } else if (val->is_ArrayCopy()) { 546 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc); 547 if (res == NULL) { 548 return NULL; 549 } 550 values.at_put(j, res); 551 } else { 552 #ifdef ASSERT 553 val->dump(); 554 assert(false, "unknown node on this path"); 555 #endif 556 return NULL; // unknown node on this path 557 } 558 } 559 } 560 // Set Phi's inputs 561 for (uint j = 1; j < length; j++) { 562 if (values.at(j) == mem) { 563 phi->init_req(j, phi); 564 } else { 565 phi->init_req(j, values.at(j)); 566 } 567 } 568 return phi; 569 } 570 571 // Search the last value stored into the object's field. 572 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) { 573 assert(adr_t->is_known_instance_field(), "instance required"); 574 assert(ft != T_VALUETYPE, "should not be used for value type fields"); 575 int instance_id = adr_t->instance_id(); 576 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 577 578 int alias_idx = C->get_alias_index(adr_t); 579 int offset = adr_t->offset(); 580 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 581 Node *alloc_ctrl = alloc->in(TypeFunc::Control); 582 Node *alloc_mem = alloc->in(TypeFunc::Memory); 583 Arena *a = Thread::current()->resource_area(); 584 VectorSet visited(a); 585 586 bool done = sfpt_mem == alloc_mem; 587 Node *mem = sfpt_mem; 588 while (!done) { 589 if (visited.test_set(mem->_idx)) { 590 return NULL; // found a loop, give up 591 } 592 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 593 if (mem == start_mem || mem == alloc_mem) { 594 done = true; // hit a sentinel, return appropriate 0 value 595 } else if (mem->is_Initialize()) { 596 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 597 if (mem == NULL) { 598 done = true; // Something went wrong. 599 } else if (mem->is_Store()) { 600 const TypePtr* atype = mem->as_Store()->adr_type(); 601 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 602 done = true; 603 } 604 } else if (mem->is_Store()) { 605 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 606 assert(atype != NULL, "address type must be oopptr"); 607 assert(C->get_alias_index(atype) == alias_idx && 608 atype->is_known_instance_field() && atype->offset() == offset && 609 atype->instance_id() == instance_id, "store is correct memory slice"); 610 done = true; 611 } else if (mem->is_Phi()) { 612 // try to find a phi's unique input 613 Node *unique_input = NULL; 614 Node *top = C->top(); 615 for (uint i = 1; i < mem->req(); i++) { 616 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 617 if (n == NULL || n == top || n == mem) { 618 continue; 619 } else if (unique_input == NULL) { 620 unique_input = n; 621 } else if (unique_input != n) { 622 unique_input = top; 623 break; 624 } 625 } 626 if (unique_input != NULL && unique_input != top) { 627 mem = unique_input; 628 } else { 629 done = true; 630 } 631 } else if (mem->is_ArrayCopy()) { 632 done = true; 633 } else { 634 assert(false, "unexpected node"); 635 } 636 } 637 if (mem != NULL) { 638 if (mem == start_mem || mem == alloc_mem) { 639 // hit a sentinel, return appropriate 0 value 640 return _igvn.zerocon(ft); 641 } else if (mem->is_Store()) { 642 return mem->in(MemNode::ValueIn); 643 } else if (mem->is_Phi()) { 644 // attempt to produce a Phi reflecting the values on the input paths of the Phi 645 Node_Stack value_phis(a, 8); 646 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 647 if (phi != NULL) { 648 return phi; 649 } else { 650 // Kill all new Phis 651 while(value_phis.is_nonempty()) { 652 Node* n = value_phis.node(); 653 _igvn.replace_node(n, C->top()); 654 value_phis.pop(); 655 } 656 } 657 } else if (mem->is_ArrayCopy()) { 658 Node* ctl = mem->in(0); 659 Node* m = mem->in(TypeFunc::Memory); 660 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) { 661 // pin the loads in the uncommon trap path 662 ctl = sfpt_ctl; 663 m = sfpt_mem; 664 } 665 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc); 666 } 667 } 668 // Something went wrong. 669 return NULL; 670 } 671 672 // Search the last value stored into the value type's fields. 673 Node* PhaseMacroExpand::value_type_from_mem(Node* mem, Node* ctl, ciValueKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) { 674 // Subtract the offset of the first field to account for the missing oop header 675 offset -= vk->first_field_offset(); 676 // Create a new ValueTypeNode and retrieve the field values from memory 677 ValueTypeNode* vt = ValueTypeNode::make(_igvn, vk)->as_ValueType(); 678 for (int i = 0; i < vk->field_count(); ++i) { 679 ciType* field_type = vt->field_type(i); 680 int field_offset = offset + vt->field_offset(i); 681 // Each value type field has its own memory slice 682 adr_type = adr_type->with_field_offset(field_offset); 683 Node* value = NULL; 684 if (field_type->basic_type() == T_VALUETYPE) { 685 value = value_type_from_mem(mem, ctl, field_type->as_value_klass(), adr_type, field_offset, alloc); 686 } else { 687 value = value_from_mem(mem, ctl, field_type->basic_type(), Type::get_const_type(field_type), adr_type, alloc); 688 } 689 vt->set_field_value(i, value); 690 } 691 return vt; 692 } 693 694 // Check the possibility of scalar replacement. 695 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 696 // Scan the uses of the allocation to check for anything that would 697 // prevent us from eliminating it. 698 NOT_PRODUCT( const char* fail_eliminate = NULL; ) 699 DEBUG_ONLY( Node* disq_node = NULL; ) 700 bool can_eliminate = true; 701 702 Node* res = alloc->result_cast(); 703 const TypeOopPtr* res_type = NULL; 704 if (res == NULL) { 705 // All users were eliminated. 706 } else if (!res->is_CheckCastPP()) { 707 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 708 can_eliminate = false; 709 } else { 710 res_type = _igvn.type(res)->isa_oopptr(); 711 if (res_type == NULL) { 712 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 713 can_eliminate = false; 714 } else if (res_type->isa_aryptr()) { 715 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 716 if (length < 0) { 717 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 718 can_eliminate = false; 719 } 720 } 721 } 722 723 if (can_eliminate && res != NULL) { 724 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 725 j < jmax && can_eliminate; j++) { 726 Node* use = res->fast_out(j); 727 728 if (use->is_AddP()) { 729 const TypePtr* addp_type = _igvn.type(use)->is_ptr(); 730 int offset = addp_type->offset(); 731 732 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 733 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) 734 can_eliminate = false; 735 break; 736 } 737 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 738 k < kmax && can_eliminate; k++) { 739 Node* n = use->fast_out(k); 740 if (!n->is_Store() && n->Opcode() != Op_CastP2X && 741 !(n->is_ArrayCopy() && 742 n->as_ArrayCopy()->is_clonebasic() && 743 n->in(ArrayCopyNode::Dest) == use)) { 744 DEBUG_ONLY(disq_node = n;) 745 if (n->is_Load() || n->is_LoadStore()) { 746 NOT_PRODUCT(fail_eliminate = "Field load";) 747 } else { 748 NOT_PRODUCT(fail_eliminate = "Not store field reference";) 749 } 750 can_eliminate = false; 751 } 752 } 753 } else if (use->is_ArrayCopy() && 754 (use->as_ArrayCopy()->is_arraycopy_validated() || 755 use->as_ArrayCopy()->is_copyof_validated() || 756 use->as_ArrayCopy()->is_copyofrange_validated()) && 757 use->in(ArrayCopyNode::Dest) == res) { 758 // ok to eliminate 759 } else if (use->is_SafePoint()) { 760 SafePointNode* sfpt = use->as_SafePoint(); 761 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 762 // Object is passed as argument. 763 DEBUG_ONLY(disq_node = use;) 764 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 765 can_eliminate = false; 766 } 767 Node* sfptMem = sfpt->memory(); 768 if (sfptMem == NULL || sfptMem->is_top()) { 769 DEBUG_ONLY(disq_node = use;) 770 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) 771 can_eliminate = false; 772 } else { 773 safepoints.append_if_missing(sfpt); 774 } 775 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 776 if (use->is_Phi()) { 777 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 778 NOT_PRODUCT(fail_eliminate = "Object is return value";) 779 } else { 780 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 781 } 782 DEBUG_ONLY(disq_node = use;) 783 } else { 784 if (use->Opcode() == Op_Return) { 785 NOT_PRODUCT(fail_eliminate = "Object is return value";) 786 } else { 787 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 788 } 789 DEBUG_ONLY(disq_node = use;) 790 } 791 can_eliminate = false; 792 } 793 } 794 } 795 796 #ifndef PRODUCT 797 if (PrintEliminateAllocations) { 798 if (can_eliminate) { 799 tty->print("Scalar "); 800 if (res == NULL) 801 alloc->dump(); 802 else 803 res->dump(); 804 } else if (alloc->_is_scalar_replaceable) { 805 tty->print("NotScalar (%s)", fail_eliminate); 806 if (res == NULL) 807 alloc->dump(); 808 else 809 res->dump(); 810 #ifdef ASSERT 811 if (disq_node != NULL) { 812 tty->print(" >>>> "); 813 disq_node->dump(); 814 } 815 #endif /*ASSERT*/ 816 } 817 } 818 #endif 819 return can_eliminate; 820 } 821 822 // Do scalar replacement. 823 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 824 GrowableArray <SafePointNode *> safepoints_done; 825 826 ciKlass* klass = NULL; 827 ciInstanceKlass* iklass = NULL; 828 int nfields = 0; 829 int array_base = 0; 830 int element_size = 0; 831 BasicType basic_elem_type = T_ILLEGAL; 832 ciType* elem_type = NULL; 833 834 Node* res = alloc->result_cast(); 835 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result"); 836 const TypeOopPtr* res_type = NULL; 837 if (res != NULL) { // Could be NULL when there are no users 838 res_type = _igvn.type(res)->isa_oopptr(); 839 } 840 841 if (res != NULL) { 842 klass = res_type->klass(); 843 // Value types are only allocated on demand 844 if (res_type->isa_instptr() || res_type->isa_valuetypeptr()) { 845 // find the fields of the class which will be needed for safepoint debug information 846 assert(klass->is_instance_klass(), "must be an instance klass."); 847 iklass = klass->as_instance_klass(); 848 nfields = iklass->nof_nonstatic_fields(); 849 } else { 850 // find the array's elements which will be needed for safepoint debug information 851 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 852 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); 853 elem_type = klass->as_array_klass()->element_type(); 854 basic_elem_type = elem_type->basic_type(); 855 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 856 element_size = type2aelembytes(basic_elem_type); 857 } 858 } 859 // 860 // Process the safepoint uses 861 // 862 while (safepoints.length() > 0) { 863 SafePointNode* sfpt = safepoints.pop(); 864 Node* mem = sfpt->memory(); 865 Node* ctl = sfpt->control(); 866 assert(sfpt->jvms() != NULL, "missed JVMS"); 867 // Fields of scalar objs are referenced only at the end 868 // of regular debuginfo at the last (youngest) JVMS. 869 // Record relative start index. 870 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff()); 871 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, 872 #ifdef ASSERT 873 alloc, 874 #endif 875 first_ind, nfields); 876 sobj->init_req(0, C->root()); 877 transform_later(sobj); 878 879 // Scan object's fields adding an input to the safepoint for each field. 880 for (int j = 0; j < nfields; j++) { 881 intptr_t offset; 882 ciField* field = NULL; 883 if (iklass != NULL) { 884 field = iklass->nonstatic_field_at(j); 885 offset = field->offset(); 886 elem_type = field->type(); 887 basic_elem_type = field->layout_type(); 888 // Value type fields should not have safepoint uses 889 assert(basic_elem_type != T_VALUETYPE, "value type fields are flattened"); 890 } else { 891 offset = array_base + j * (intptr_t)element_size; 892 } 893 894 const Type *field_type; 895 // The next code is taken from Parse::do_get_xxx(). 896 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { 897 if (!elem_type->is_loaded()) { 898 field_type = TypeInstPtr::BOTTOM; 899 } else if (field != NULL && field->is_static_constant()) { 900 // This can happen if the constant oop is non-perm. 901 ciObject* con = field->constant_value().as_object(); 902 // Do not "join" in the previous type; it doesn't add value, 903 // and may yield a vacuous result if the field is of interface type. 904 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 905 assert(field_type != NULL, "field singleton type must be consistent"); 906 } else { 907 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 908 } 909 if (UseCompressedOops) { 910 field_type = field_type->make_narrowoop(); 911 basic_elem_type = T_NARROWOOP; 912 } 913 } else { 914 field_type = Type::get_const_basic_type(basic_elem_type); 915 } 916 917 Node* field_val = NULL; 918 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 919 if (klass->is_value_array_klass()) { 920 ciValueKlass* vk = elem_type->as_value_klass(); 921 assert(vk->flatten_array(), "must be flattened"); 922 field_val = value_type_from_mem(mem, ctl, vk, field_addr_type->isa_aryptr(), 0, alloc); 923 } else { 924 field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc); 925 } 926 if (field_val == NULL) { 927 // We weren't able to find a value for this field, 928 // give up on eliminating this allocation. 929 930 // Remove any extra entries we added to the safepoint. 931 uint last = sfpt->req() - 1; 932 for (int k = 0; k < j; k++) { 933 sfpt->del_req(last--); 934 } 935 _igvn._worklist.push(sfpt); 936 // rollback processed safepoints 937 while (safepoints_done.length() > 0) { 938 SafePointNode* sfpt_done = safepoints_done.pop(); 939 // remove any extra entries we added to the safepoint 940 last = sfpt_done->req() - 1; 941 for (int k = 0; k < nfields; k++) { 942 sfpt_done->del_req(last--); 943 } 944 JVMState *jvms = sfpt_done->jvms(); 945 jvms->set_endoff(sfpt_done->req()); 946 // Now make a pass over the debug information replacing any references 947 // to SafePointScalarObjectNode with the allocated object. 948 int start = jvms->debug_start(); 949 int end = jvms->debug_end(); 950 for (int i = start; i < end; i++) { 951 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 952 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 953 if (scobj->first_index(jvms) == sfpt_done->req() && 954 scobj->n_fields() == (uint)nfields) { 955 assert(scobj->alloc() == alloc, "sanity"); 956 sfpt_done->set_req(i, res); 957 } 958 } 959 } 960 _igvn._worklist.push(sfpt_done); 961 } 962 #ifndef PRODUCT 963 if (PrintEliminateAllocations) { 964 if (field != NULL) { 965 tty->print("=== At SafePoint node %d can't find value of Field: ", 966 sfpt->_idx); 967 field->print(); 968 int field_idx = C->get_alias_index(field_addr_type); 969 tty->print(" (alias_idx=%d)", field_idx); 970 } else { // Array's element 971 tty->print("=== At SafePoint node %d can't find value of array element [%d]", 972 sfpt->_idx, j); 973 } 974 tty->print(", which prevents elimination of: "); 975 if (res == NULL) 976 alloc->dump(); 977 else 978 res->dump(); 979 } 980 #endif 981 return false; 982 } 983 if (UseCompressedOops && field_type->isa_narrowoop()) { 984 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 985 // to be able scalar replace the allocation. 986 if (field_val->is_EncodeP()) { 987 field_val = field_val->in(1); 988 } else { 989 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type())); 990 } 991 } 992 sfpt->add_req(field_val); 993 } 994 JVMState *jvms = sfpt->jvms(); 995 jvms->set_endoff(sfpt->req()); 996 // Now make a pass over the debug information replacing any references 997 // to the allocated object with "sobj" 998 int start = jvms->debug_start(); 999 int end = jvms->debug_end(); 1000 sfpt->replace_edges_in_range(res, sobj, start, end); 1001 _igvn._worklist.push(sfpt); 1002 safepoints_done.append_if_missing(sfpt); // keep it for rollback 1003 } 1004 return true; 1005 } 1006 1007 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) { 1008 Node* ctl_proj = n->proj_out(TypeFunc::Control); 1009 Node* mem_proj = n->proj_out(TypeFunc::Memory); 1010 if (ctl_proj != NULL) { 1011 igvn.replace_node(ctl_proj, n->in(0)); 1012 } 1013 if (mem_proj != NULL) { 1014 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory)); 1015 } 1016 } 1017 1018 // Process users of eliminated allocation. 1019 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) { 1020 Node* res = alloc->result_cast(); 1021 if (res != NULL) { 1022 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 1023 Node *use = res->last_out(j); 1024 uint oc1 = res->outcnt(); 1025 1026 if (use->is_AddP()) { 1027 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 1028 Node *n = use->last_out(k); 1029 uint oc2 = use->outcnt(); 1030 if (n->is_Store()) { 1031 #ifdef ASSERT 1032 // Verify that there is no dependent MemBarVolatile nodes, 1033 // they should be removed during IGVN, see MemBarNode::Ideal(). 1034 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); 1035 p < pmax; p++) { 1036 Node* mb = n->fast_out(p); 1037 assert(mb->is_Initialize() || !mb->is_MemBar() || 1038 mb->req() <= MemBarNode::Precedent || 1039 mb->in(MemBarNode::Precedent) != n, 1040 "MemBarVolatile should be eliminated for non-escaping object"); 1041 } 1042 #endif 1043 _igvn.replace_node(n, n->in(MemNode::Memory)); 1044 } else if (n->is_ArrayCopy()) { 1045 // Disconnect ArrayCopy node 1046 ArrayCopyNode* ac = n->as_ArrayCopy(); 1047 assert(ac->is_clonebasic(), "unexpected array copy kind"); 1048 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out(); 1049 disconnect_projections(ac, _igvn); 1050 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation"); 1051 Node* membar_before = alloc->in(0)->in(0); 1052 disconnect_projections(membar_before->as_MemBar(), _igvn); 1053 if (membar_after->is_MemBar()) { 1054 disconnect_projections(membar_after->as_MemBar(), _igvn); 1055 } 1056 } else { 1057 eliminate_card_mark(n); 1058 } 1059 k -= (oc2 - use->outcnt()); 1060 } 1061 } else if (use->is_ArrayCopy()) { 1062 // Disconnect ArrayCopy node 1063 ArrayCopyNode* ac = use->as_ArrayCopy(); 1064 assert(ac->is_arraycopy_validated() || 1065 ac->is_copyof_validated() || 1066 ac->is_copyofrange_validated(), "unsupported"); 1067 CallProjections callprojs; 1068 ac->extract_projections(&callprojs, true); 1069 1070 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O)); 1071 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory)); 1072 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control)); 1073 1074 // Set control to top. IGVN will remove the remaining projections 1075 ac->set_req(0, top()); 1076 ac->replace_edge(res, top()); 1077 1078 // Disconnect src right away: it can help find new 1079 // opportunities for allocation elimination 1080 Node* src = ac->in(ArrayCopyNode::Src); 1081 ac->replace_edge(src, top()); 1082 // src can be top at this point if src and dest of the 1083 // arraycopy were the same 1084 if (src->outcnt() == 0 && !src->is_top()) { 1085 _igvn.remove_dead_node(src); 1086 } 1087 1088 _igvn._worklist.push(ac); 1089 } else { 1090 eliminate_card_mark(use); 1091 } 1092 j -= (oc1 - res->outcnt()); 1093 } 1094 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 1095 _igvn.remove_dead_node(res); 1096 } 1097 1098 // 1099 // Process other users of allocation's projections 1100 // 1101 if (_resproj != NULL && _resproj->outcnt() != 0) { 1102 // First disconnect stores captured by Initialize node. 1103 // If Initialize node is eliminated first in the following code, 1104 // it will kill such stores and DUIterator_Last will assert. 1105 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) { 1106 Node *use = _resproj->fast_out(j); 1107 if (use->is_AddP()) { 1108 // raw memory addresses used only by the initialization 1109 _igvn.replace_node(use, C->top()); 1110 --j; --jmax; 1111 } 1112 } 1113 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { 1114 Node *use = _resproj->last_out(j); 1115 uint oc1 = _resproj->outcnt(); 1116 if (use->is_Initialize()) { 1117 // Eliminate Initialize node. 1118 InitializeNode *init = use->as_Initialize(); 1119 assert(init->outcnt() <= 2, "only a control and memory projection expected"); 1120 Node *ctrl_proj = init->proj_out(TypeFunc::Control); 1121 if (ctrl_proj != NULL) { 1122 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection"); 1123 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj); 1124 } 1125 Node *mem_proj = init->proj_out(TypeFunc::Memory); 1126 if (mem_proj != NULL) { 1127 Node *mem = init->in(TypeFunc::Memory); 1128 #ifdef ASSERT 1129 if (mem->is_MergeMem()) { 1130 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection"); 1131 } else { 1132 assert(mem == _memproj_fallthrough, "allocation memory projection"); 1133 } 1134 #endif 1135 _igvn.replace_node(mem_proj, mem); 1136 } 1137 } else { 1138 assert(false, "only Initialize or AddP expected"); 1139 } 1140 j -= (oc1 - _resproj->outcnt()); 1141 } 1142 } 1143 if (_fallthroughcatchproj != NULL) { 1144 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); 1145 } 1146 if (_memproj_fallthrough != NULL) { 1147 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); 1148 } 1149 if (_memproj_catchall != NULL) { 1150 _igvn.replace_node(_memproj_catchall, C->top()); 1151 } 1152 if (_ioproj_fallthrough != NULL) { 1153 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); 1154 } 1155 if (_ioproj_catchall != NULL) { 1156 _igvn.replace_node(_ioproj_catchall, C->top()); 1157 } 1158 if (_catchallcatchproj != NULL) { 1159 _igvn.replace_node(_catchallcatchproj, C->top()); 1160 } 1161 } 1162 1163 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 1164 // Don't do scalar replacement if the frame can be popped by JVMTI: 1165 // if reallocation fails during deoptimization we'll pop all 1166 // interpreter frames for this compiled frame and that won't play 1167 // nice with JVMTI popframe. 1168 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) { 1169 return false; 1170 } 1171 Node* klass = alloc->in(AllocateNode::KlassNode); 1172 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); 1173 Node* res = alloc->result_cast(); 1174 // Eliminate boxing allocations which are not used 1175 // regardless scalar replacable status. 1176 bool boxing_alloc = C->eliminate_boxing() && 1177 tklass->klass()->is_instance_klass() && 1178 tklass->klass()->as_instance_klass()->is_box_klass(); 1179 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) { 1180 return false; 1181 } 1182 1183 extract_call_projections(alloc); 1184 1185 GrowableArray <SafePointNode *> safepoints; 1186 if (!can_eliminate_allocation(alloc, safepoints)) { 1187 return false; 1188 } 1189 1190 if (!alloc->_is_scalar_replaceable) { 1191 assert(res == NULL, "sanity"); 1192 // We can only eliminate allocation if all debug info references 1193 // are already replaced with SafePointScalarObject because 1194 // we can't search for a fields value without instance_id. 1195 if (safepoints.length() > 0) { 1196 return false; 1197 } 1198 } 1199 1200 if (!scalar_replacement(alloc, safepoints)) { 1201 return false; 1202 } 1203 1204 CompileLog* log = C->log(); 1205 if (log != NULL) { 1206 log->head("eliminate_allocation type='%d'", 1207 log->identify(tklass->klass())); 1208 JVMState* p = alloc->jvms(); 1209 while (p != NULL) { 1210 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1211 p = p->caller(); 1212 } 1213 log->tail("eliminate_allocation"); 1214 } 1215 1216 process_users_of_allocation(alloc); 1217 1218 #ifndef PRODUCT 1219 if (PrintEliminateAllocations) { 1220 if (alloc->is_AllocateArray()) 1221 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1222 else 1223 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1224 } 1225 #endif 1226 1227 return true; 1228 } 1229 1230 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) { 1231 // EA should remove all uses of non-escaping boxing node. 1232 if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) { 1233 return false; 1234 } 1235 1236 assert(boxing->result_cast() == NULL, "unexpected boxing node result"); 1237 1238 extract_call_projections(boxing); 1239 1240 const TypeTuple* r = boxing->tf()->range_sig(); 1241 assert(r->cnt() > TypeFunc::Parms, "sanity"); 1242 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr(); 1243 assert(t != NULL, "sanity"); 1244 1245 CompileLog* log = C->log(); 1246 if (log != NULL) { 1247 log->head("eliminate_boxing type='%d'", 1248 log->identify(t->klass())); 1249 JVMState* p = boxing->jvms(); 1250 while (p != NULL) { 1251 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1252 p = p->caller(); 1253 } 1254 log->tail("eliminate_boxing"); 1255 } 1256 1257 process_users_of_allocation(boxing); 1258 1259 #ifndef PRODUCT 1260 if (PrintEliminateAllocations) { 1261 tty->print("++++ Eliminated: %d ", boxing->_idx); 1262 boxing->method()->print_short_name(tty); 1263 tty->cr(); 1264 } 1265 #endif 1266 1267 return true; 1268 } 1269 1270 //---------------------------set_eden_pointers------------------------- 1271 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { 1272 if (UseTLAB) { // Private allocation: load from TLS 1273 Node* thread = transform_later(new ThreadLocalNode()); 1274 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); 1275 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); 1276 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); 1277 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); 1278 } else { // Shared allocation: load from globals 1279 CollectedHeap* ch = Universe::heap(); 1280 address top_adr = (address)ch->top_addr(); 1281 address end_adr = (address)ch->end_addr(); 1282 eden_top_adr = makecon(TypeRawPtr::make(top_adr)); 1283 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); 1284 } 1285 } 1286 1287 1288 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 1289 Node* adr = basic_plus_adr(base, offset); 1290 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 1291 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered); 1292 transform_later(value); 1293 return value; 1294 } 1295 1296 1297 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 1298 Node* adr = basic_plus_adr(base, offset); 1299 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered); 1300 transform_later(mem); 1301 return mem; 1302 } 1303 1304 //============================================================================= 1305 // 1306 // A L L O C A T I O N 1307 // 1308 // Allocation attempts to be fast in the case of frequent small objects. 1309 // It breaks down like this: 1310 // 1311 // 1) Size in doublewords is computed. This is a constant for objects and 1312 // variable for most arrays. Doubleword units are used to avoid size 1313 // overflow of huge doubleword arrays. We need doublewords in the end for 1314 // rounding. 1315 // 1316 // 2) Size is checked for being 'too large'. Too-large allocations will go 1317 // the slow path into the VM. The slow path can throw any required 1318 // exceptions, and does all the special checks for very large arrays. The 1319 // size test can constant-fold away for objects. For objects with 1320 // finalizers it constant-folds the otherway: you always go slow with 1321 // finalizers. 1322 // 1323 // 3) If NOT using TLABs, this is the contended loop-back point. 1324 // Load-Locked the heap top. If using TLABs normal-load the heap top. 1325 // 1326 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 1327 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 1328 // "size*8" we always enter the VM, where "largish" is a constant picked small 1329 // enough that there's always space between the eden max and 4Gig (old space is 1330 // there so it's quite large) and large enough that the cost of entering the VM 1331 // is dwarfed by the cost to initialize the space. 1332 // 1333 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 1334 // down. If contended, repeat at step 3. If using TLABs normal-store 1335 // adjusted heap top back down; there is no contention. 1336 // 1337 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 1338 // fields. 1339 // 1340 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 1341 // oop flavor. 1342 // 1343 //============================================================================= 1344 // FastAllocateSizeLimit value is in DOUBLEWORDS. 1345 // Allocations bigger than this always go the slow route. 1346 // This value must be small enough that allocation attempts that need to 1347 // trigger exceptions go the slow route. Also, it must be small enough so 1348 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 1349 //=============================================================================j// 1350 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 1351 // The allocator will coalesce int->oop copies away. See comment in 1352 // coalesce.cpp about how this works. It depends critically on the exact 1353 // code shape produced here, so if you are changing this code shape 1354 // make sure the GC info for the heap-top is correct in and around the 1355 // slow-path call. 1356 // 1357 1358 void PhaseMacroExpand::expand_allocate_common( 1359 AllocateNode* alloc, // allocation node to be expanded 1360 Node* length, // array length for an array allocation 1361 const TypeFunc* slow_call_type, // Type of slow call 1362 address slow_call_address // Address of slow call 1363 ) 1364 { 1365 1366 Node* ctrl = alloc->in(TypeFunc::Control); 1367 Node* mem = alloc->in(TypeFunc::Memory); 1368 Node* i_o = alloc->in(TypeFunc::I_O); 1369 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 1370 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1371 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 1372 1373 assert(ctrl != NULL, "must have control"); 1374 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 1375 // they will not be used if "always_slow" is set 1376 enum { slow_result_path = 1, fast_result_path = 2 }; 1377 Node *result_region = NULL; 1378 Node *result_phi_rawmem = NULL; 1379 Node *result_phi_rawoop = NULL; 1380 Node *result_phi_i_o = NULL; 1381 1382 // The initial slow comparison is a size check, the comparison 1383 // we want to do is a BoolTest::gt 1384 bool always_slow = false; 1385 int tv = _igvn.find_int_con(initial_slow_test, -1); 1386 if (tv >= 0) { 1387 always_slow = (tv == 1); 1388 initial_slow_test = NULL; 1389 } else { 1390 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 1391 } 1392 1393 if (C->env()->dtrace_alloc_probes() || 1394 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) { 1395 // Force slow-path allocation 1396 always_slow = true; 1397 initial_slow_test = NULL; 1398 } 1399 1400 1401 enum { too_big_or_final_path = 1, need_gc_path = 2 }; 1402 Node *slow_region = NULL; 1403 Node *toobig_false = ctrl; 1404 1405 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); 1406 // generate the initial test if necessary 1407 if (initial_slow_test != NULL ) { 1408 slow_region = new RegionNode(3); 1409 1410 // Now make the initial failure test. Usually a too-big test but 1411 // might be a TRUE for finalizers or a fancy class check for 1412 // newInstance0. 1413 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1414 transform_later(toobig_iff); 1415 // Plug the failing-too-big test into the slow-path region 1416 Node *toobig_true = new IfTrueNode( toobig_iff ); 1417 transform_later(toobig_true); 1418 slow_region ->init_req( too_big_or_final_path, toobig_true ); 1419 toobig_false = new IfFalseNode( toobig_iff ); 1420 transform_later(toobig_false); 1421 } else { // No initial test, just fall into next case 1422 toobig_false = ctrl; 1423 debug_only(slow_region = NodeSentinel); 1424 } 1425 1426 Node *slow_mem = mem; // save the current memory state for slow path 1427 // generate the fast allocation code unless we know that the initial test will always go slow 1428 if (!always_slow) { 1429 // Fast path modifies only raw memory. 1430 if (mem->is_MergeMem()) { 1431 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1432 } 1433 1434 Node* eden_top_adr; 1435 Node* eden_end_adr; 1436 1437 set_eden_pointers(eden_top_adr, eden_end_adr); 1438 1439 // Load Eden::end. Loop invariant and hoisted. 1440 // 1441 // Note: We set the control input on "eden_end" and "old_eden_top" when using 1442 // a TLAB to work around a bug where these values were being moved across 1443 // a safepoint. These are not oops, so they cannot be include in the oop 1444 // map, but they can be changed by a GC. The proper way to fix this would 1445 // be to set the raw memory state when generating a SafepointNode. However 1446 // this will require extensive changes to the loop optimization in order to 1447 // prevent a degradation of the optimization. 1448 // See comment in memnode.hpp, around line 227 in class LoadPNode. 1449 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 1450 1451 // allocate the Region and Phi nodes for the result 1452 result_region = new RegionNode(3); 1453 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1454 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM); 1455 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch 1456 1457 // We need a Region for the loop-back contended case. 1458 enum { fall_in_path = 1, contended_loopback_path = 2 }; 1459 Node *contended_region; 1460 Node *contended_phi_rawmem; 1461 if (UseTLAB) { 1462 contended_region = toobig_false; 1463 contended_phi_rawmem = mem; 1464 } else { 1465 contended_region = new RegionNode(3); 1466 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1467 // Now handle the passing-too-big test. We fall into the contended 1468 // loop-back merge point. 1469 contended_region ->init_req(fall_in_path, toobig_false); 1470 contended_phi_rawmem->init_req(fall_in_path, mem); 1471 transform_later(contended_region); 1472 transform_later(contended_phi_rawmem); 1473 } 1474 1475 // Load(-locked) the heap top. 1476 // See note above concerning the control input when using a TLAB 1477 Node *old_eden_top = UseTLAB 1478 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered) 1479 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire); 1480 1481 transform_later(old_eden_top); 1482 // Add to heap top to get a new heap top 1483 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes); 1484 transform_later(new_eden_top); 1485 // Check for needing a GC; compare against heap end 1486 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end); 1487 transform_later(needgc_cmp); 1488 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge); 1489 transform_later(needgc_bol); 1490 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); 1491 transform_later(needgc_iff); 1492 1493 // Plug the failing-heap-space-need-gc test into the slow-path region 1494 Node *needgc_true = new IfTrueNode(needgc_iff); 1495 transform_later(needgc_true); 1496 if (initial_slow_test) { 1497 slow_region->init_req(need_gc_path, needgc_true); 1498 // This completes all paths into the slow merge point 1499 transform_later(slow_region); 1500 } else { // No initial slow path needed! 1501 // Just fall from the need-GC path straight into the VM call. 1502 slow_region = needgc_true; 1503 } 1504 // No need for a GC. Setup for the Store-Conditional 1505 Node *needgc_false = new IfFalseNode(needgc_iff); 1506 transform_later(needgc_false); 1507 1508 // Grab regular I/O before optional prefetch may change it. 1509 // Slow-path does no I/O so just set it to the original I/O. 1510 result_phi_i_o->init_req(slow_result_path, i_o); 1511 1512 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem, 1513 old_eden_top, new_eden_top, length); 1514 1515 // Name successful fast-path variables 1516 Node* fast_oop = old_eden_top; 1517 Node* fast_oop_ctrl; 1518 Node* fast_oop_rawmem; 1519 1520 // Store (-conditional) the modified eden top back down. 1521 // StorePConditional produces flags for a test PLUS a modified raw 1522 // memory state. 1523 if (UseTLAB) { 1524 Node* store_eden_top = 1525 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1526 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered); 1527 transform_later(store_eden_top); 1528 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path 1529 fast_oop_rawmem = store_eden_top; 1530 } else { 1531 Node* store_eden_top = 1532 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1533 new_eden_top, fast_oop/*old_eden_top*/); 1534 transform_later(store_eden_top); 1535 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne); 1536 transform_later(contention_check); 1537 store_eden_top = new SCMemProjNode(store_eden_top); 1538 transform_later(store_eden_top); 1539 1540 // If not using TLABs, check to see if there was contention. 1541 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN); 1542 transform_later(contention_iff); 1543 Node *contention_true = new IfTrueNode(contention_iff); 1544 transform_later(contention_true); 1545 // If contention, loopback and try again. 1546 contended_region->init_req(contended_loopback_path, contention_true); 1547 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top); 1548 1549 // Fast-path succeeded with no contention! 1550 Node *contention_false = new IfFalseNode(contention_iff); 1551 transform_later(contention_false); 1552 fast_oop_ctrl = contention_false; 1553 1554 // Bump total allocated bytes for this thread 1555 Node* thread = new ThreadLocalNode(); 1556 transform_later(thread); 1557 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread, 1558 in_bytes(JavaThread::allocated_bytes_offset())); 1559 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1560 0, TypeLong::LONG, T_LONG); 1561 #ifdef _LP64 1562 Node* alloc_size = size_in_bytes; 1563 #else 1564 Node* alloc_size = new ConvI2LNode(size_in_bytes); 1565 transform_later(alloc_size); 1566 #endif 1567 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size); 1568 transform_later(new_alloc_bytes); 1569 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1570 0, new_alloc_bytes, T_LONG); 1571 } 1572 1573 InitializeNode* init = alloc->initialization(); 1574 fast_oop_rawmem = initialize_object(alloc, 1575 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1576 klass_node, length, size_in_bytes); 1577 1578 // If initialization is performed by an array copy, any required 1579 // MemBarStoreStore was already added. If the object does not 1580 // escape no need for a MemBarStoreStore. If the object does not 1581 // escape in its initializer and memory barrier (MemBarStoreStore or 1582 // stronger) is already added at exit of initializer, also no need 1583 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore 1584 // so that stores that initialize this object can't be reordered 1585 // with a subsequent store that makes this object accessible by 1586 // other threads. 1587 // Other threads include java threads and JVM internal threads 1588 // (for example concurrent GC threads). Current concurrent GC 1589 // implementation: CMS and G1 will not scan newly created object, 1590 // so it's safe to skip storestore barrier when allocation does 1591 // not escape. 1592 if (!alloc->does_not_escape_thread() && 1593 !alloc->is_allocation_MemBar_redundant() && 1594 (init == NULL || !init->is_complete_with_arraycopy())) { 1595 if (init == NULL || init->req() < InitializeNode::RawStores) { 1596 // No InitializeNode or no stores captured by zeroing 1597 // elimination. Simply add the MemBarStoreStore after object 1598 // initialization. 1599 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1600 transform_later(mb); 1601 1602 mb->init_req(TypeFunc::Memory, fast_oop_rawmem); 1603 mb->init_req(TypeFunc::Control, fast_oop_ctrl); 1604 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control); 1605 transform_later(fast_oop_ctrl); 1606 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory); 1607 transform_later(fast_oop_rawmem); 1608 } else { 1609 // Add the MemBarStoreStore after the InitializeNode so that 1610 // all stores performing the initialization that were moved 1611 // before the InitializeNode happen before the storestore 1612 // barrier. 1613 1614 Node* init_ctrl = init->proj_out(TypeFunc::Control); 1615 Node* init_mem = init->proj_out(TypeFunc::Memory); 1616 1617 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1618 transform_later(mb); 1619 1620 Node* ctrl = new ProjNode(init,TypeFunc::Control); 1621 transform_later(ctrl); 1622 Node* mem = new ProjNode(init,TypeFunc::Memory); 1623 transform_later(mem); 1624 1625 // The MemBarStoreStore depends on control and memory coming 1626 // from the InitializeNode 1627 mb->init_req(TypeFunc::Memory, mem); 1628 mb->init_req(TypeFunc::Control, ctrl); 1629 1630 ctrl = new ProjNode(mb,TypeFunc::Control); 1631 transform_later(ctrl); 1632 mem = new ProjNode(mb,TypeFunc::Memory); 1633 transform_later(mem); 1634 1635 // All nodes that depended on the InitializeNode for control 1636 // and memory must now depend on the MemBarNode that itself 1637 // depends on the InitializeNode 1638 if (init_ctrl != NULL) { 1639 _igvn.replace_node(init_ctrl, ctrl); 1640 } 1641 if (init_mem != NULL) { 1642 _igvn.replace_node(init_mem, mem); 1643 } 1644 } 1645 } 1646 1647 if (C->env()->dtrace_extended_probes()) { 1648 // Slow-path call 1649 int size = TypeFunc::Parms + 2; 1650 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1651 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), 1652 "dtrace_object_alloc", 1653 TypeRawPtr::BOTTOM); 1654 1655 // Get base of thread-local storage area 1656 Node* thread = new ThreadLocalNode(); 1657 transform_later(thread); 1658 1659 call->init_req(TypeFunc::Parms+0, thread); 1660 call->init_req(TypeFunc::Parms+1, fast_oop); 1661 call->init_req(TypeFunc::Control, fast_oop_ctrl); 1662 call->init_req(TypeFunc::I_O , top()); // does no i/o 1663 call->init_req(TypeFunc::Memory , fast_oop_rawmem); 1664 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1665 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1666 transform_later(call); 1667 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control); 1668 transform_later(fast_oop_ctrl); 1669 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory); 1670 transform_later(fast_oop_rawmem); 1671 } 1672 1673 // Plug in the successful fast-path into the result merge point 1674 result_region ->init_req(fast_result_path, fast_oop_ctrl); 1675 result_phi_rawoop->init_req(fast_result_path, fast_oop); 1676 result_phi_i_o ->init_req(fast_result_path, i_o); 1677 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); 1678 } else { 1679 slow_region = ctrl; 1680 result_phi_i_o = i_o; // Rename it to use in the following code. 1681 } 1682 1683 // Generate slow-path call 1684 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address, 1685 OptoRuntime::stub_name(slow_call_address), 1686 alloc->jvms()->bci(), 1687 TypePtr::BOTTOM); 1688 call->init_req( TypeFunc::Control, slow_region ); 1689 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1690 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs 1691 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1692 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1693 1694 call->init_req(TypeFunc::Parms+0, klass_node); 1695 if (length != NULL) { 1696 call->init_req(TypeFunc::Parms+1, length); 1697 } 1698 1699 // Copy debug information and adjust JVMState information, then replace 1700 // allocate node with the call 1701 copy_call_debug_info((CallNode *) alloc, call); 1702 if (!always_slow) { 1703 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1704 } else { 1705 // Hook i_o projection to avoid its elimination during allocation 1706 // replacement (when only a slow call is generated). 1707 call->set_req(TypeFunc::I_O, result_phi_i_o); 1708 } 1709 _igvn.replace_node(alloc, call); 1710 transform_later(call); 1711 1712 // Identify the output projections from the allocate node and 1713 // adjust any references to them. 1714 // The control and io projections look like: 1715 // 1716 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1717 // Allocate Catch 1718 // ^---Proj(io) <-------+ ^---CatchProj(io) 1719 // 1720 // We are interested in the CatchProj nodes. 1721 // 1722 extract_call_projections(call); 1723 1724 // An allocate node has separate memory projections for the uses on 1725 // the control and i_o paths. Replace the control memory projection with 1726 // result_phi_rawmem (unless we are only generating a slow call when 1727 // both memory projections are combined) 1728 if (!always_slow && _memproj_fallthrough != NULL) { 1729 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { 1730 Node *use = _memproj_fallthrough->fast_out(i); 1731 _igvn.rehash_node_delayed(use); 1732 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); 1733 // back up iterator 1734 --i; 1735 } 1736 } 1737 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete 1738 // _memproj_catchall so we end up with a call that has only 1 memory projection. 1739 if (_memproj_catchall != NULL ) { 1740 if (_memproj_fallthrough == NULL) { 1741 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory); 1742 transform_later(_memproj_fallthrough); 1743 } 1744 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { 1745 Node *use = _memproj_catchall->fast_out(i); 1746 _igvn.rehash_node_delayed(use); 1747 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); 1748 // back up iterator 1749 --i; 1750 } 1751 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted"); 1752 _igvn.remove_dead_node(_memproj_catchall); 1753 } 1754 1755 // An allocate node has separate i_o projections for the uses on the control 1756 // and i_o paths. Always replace the control i_o projection with result i_o 1757 // otherwise incoming i_o become dead when only a slow call is generated 1758 // (it is different from memory projections where both projections are 1759 // combined in such case). 1760 if (_ioproj_fallthrough != NULL) { 1761 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { 1762 Node *use = _ioproj_fallthrough->fast_out(i); 1763 _igvn.rehash_node_delayed(use); 1764 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); 1765 // back up iterator 1766 --i; 1767 } 1768 } 1769 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete 1770 // _ioproj_catchall so we end up with a call that has only 1 i_o projection. 1771 if (_ioproj_catchall != NULL ) { 1772 if (_ioproj_fallthrough == NULL) { 1773 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O); 1774 transform_later(_ioproj_fallthrough); 1775 } 1776 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { 1777 Node *use = _ioproj_catchall->fast_out(i); 1778 _igvn.rehash_node_delayed(use); 1779 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); 1780 // back up iterator 1781 --i; 1782 } 1783 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted"); 1784 _igvn.remove_dead_node(_ioproj_catchall); 1785 } 1786 1787 // if we generated only a slow call, we are done 1788 if (always_slow) { 1789 // Now we can unhook i_o. 1790 if (result_phi_i_o->outcnt() > 1) { 1791 call->set_req(TypeFunc::I_O, top()); 1792 } else { 1793 assert(result_phi_i_o->unique_ctrl_out() == call, ""); 1794 // Case of new array with negative size known during compilation. 1795 // AllocateArrayNode::Ideal() optimization disconnect unreachable 1796 // following code since call to runtime will throw exception. 1797 // As result there will be no users of i_o after the call. 1798 // Leave i_o attached to this call to avoid problems in preceding graph. 1799 } 1800 return; 1801 } 1802 1803 1804 if (_fallthroughcatchproj != NULL) { 1805 ctrl = _fallthroughcatchproj->clone(); 1806 transform_later(ctrl); 1807 _igvn.replace_node(_fallthroughcatchproj, result_region); 1808 } else { 1809 ctrl = top(); 1810 } 1811 Node *slow_result; 1812 if (_resproj == NULL) { 1813 // no uses of the allocation result 1814 slow_result = top(); 1815 } else { 1816 slow_result = _resproj->clone(); 1817 transform_later(slow_result); 1818 _igvn.replace_node(_resproj, result_phi_rawoop); 1819 } 1820 1821 // Plug slow-path into result merge point 1822 result_region ->init_req( slow_result_path, ctrl ); 1823 result_phi_rawoop->init_req( slow_result_path, slow_result); 1824 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); 1825 transform_later(result_region); 1826 transform_later(result_phi_rawoop); 1827 transform_later(result_phi_rawmem); 1828 transform_later(result_phi_i_o); 1829 // This completes all paths into the result merge point 1830 } 1831 1832 1833 // Helper for PhaseMacroExpand::expand_allocate_common. 1834 // Initializes the newly-allocated storage. 1835 Node* 1836 PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1837 Node* control, Node* rawmem, Node* object, 1838 Node* klass_node, Node* length, 1839 Node* size_in_bytes) { 1840 InitializeNode* init = alloc->initialization(); 1841 // Store the klass & mark bits 1842 Node* mark_node = NULL; 1843 // For now only enable fast locking for non-array types 1844 if (UseBiasedLocking && (length == NULL)) { 1845 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 1846 } else { 1847 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); 1848 } 1849 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 1850 1851 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 1852 int header_size = alloc->minimum_header_size(); // conservatively small 1853 1854 // Array length 1855 if (length != NULL) { // Arrays need length field 1856 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1857 // conservatively small header size: 1858 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1859 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1860 if (k->is_array_klass()) // we know the exact header size in most cases: 1861 header_size = Klass::layout_helper_header_size(k->layout_helper()); 1862 } 1863 1864 // Clear the object body, if necessary. 1865 if (init == NULL) { 1866 // The init has somehow disappeared; be cautious and clear everything. 1867 // 1868 // This can happen if a node is allocated but an uncommon trap occurs 1869 // immediately. In this case, the Initialize gets associated with the 1870 // trap, and may be placed in a different (outer) loop, if the Allocate 1871 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1872 // there can be two Allocates to one Initialize. The answer in all these 1873 // edge cases is safety first. It is always safe to clear immediately 1874 // within an Allocate, and then (maybe or maybe not) clear some more later. 1875 if (!(UseTLAB && ZeroTLAB)) { 1876 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1877 header_size, size_in_bytes, 1878 &_igvn); 1879 } 1880 } else { 1881 if (!init->is_complete()) { 1882 // Try to win by zeroing only what the init does not store. 1883 // We can also try to do some peephole optimizations, 1884 // such as combining some adjacent subword stores. 1885 rawmem = init->complete_stores(control, rawmem, object, 1886 header_size, size_in_bytes, &_igvn); 1887 } 1888 // We have no more use for this link, since the AllocateNode goes away: 1889 init->set_req(InitializeNode::RawAddress, top()); 1890 // (If we keep the link, it just confuses the register allocator, 1891 // who thinks he sees a real use of the address by the membar.) 1892 } 1893 1894 return rawmem; 1895 } 1896 1897 // Generate prefetch instructions for next allocations. 1898 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1899 Node*& contended_phi_rawmem, 1900 Node* old_eden_top, Node* new_eden_top, 1901 Node* length) { 1902 enum { fall_in_path = 1, pf_path = 2 }; 1903 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1904 // Generate prefetch allocation with watermark check. 1905 // As an allocation hits the watermark, we will prefetch starting 1906 // at a "distance" away from watermark. 1907 1908 Node *pf_region = new RegionNode(3); 1909 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 1910 TypeRawPtr::BOTTOM ); 1911 // I/O is used for Prefetch 1912 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO ); 1913 1914 Node *thread = new ThreadLocalNode(); 1915 transform_later(thread); 1916 1917 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread, 1918 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1919 transform_later(eden_pf_adr); 1920 1921 Node *old_pf_wm = new LoadPNode(needgc_false, 1922 contended_phi_rawmem, eden_pf_adr, 1923 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, 1924 MemNode::unordered); 1925 transform_later(old_pf_wm); 1926 1927 // check against new_eden_top 1928 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm ); 1929 transform_later(need_pf_cmp); 1930 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge ); 1931 transform_later(need_pf_bol); 1932 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol, 1933 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1934 transform_later(need_pf_iff); 1935 1936 // true node, add prefetchdistance 1937 Node *need_pf_true = new IfTrueNode( need_pf_iff ); 1938 transform_later(need_pf_true); 1939 1940 Node *need_pf_false = new IfFalseNode( need_pf_iff ); 1941 transform_later(need_pf_false); 1942 1943 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm, 1944 _igvn.MakeConX(AllocatePrefetchDistance) ); 1945 transform_later(new_pf_wmt ); 1946 new_pf_wmt->set_req(0, need_pf_true); 1947 1948 Node *store_new_wmt = new StorePNode(need_pf_true, 1949 contended_phi_rawmem, eden_pf_adr, 1950 TypeRawPtr::BOTTOM, new_pf_wmt, 1951 MemNode::unordered); 1952 transform_later(store_new_wmt); 1953 1954 // adding prefetches 1955 pf_phi_abio->init_req( fall_in_path, i_o ); 1956 1957 Node *prefetch_adr; 1958 Node *prefetch; 1959 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1960 uint step_size = AllocatePrefetchStepSize; 1961 uint distance = 0; 1962 1963 for ( uint i = 0; i < lines; i++ ) { 1964 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt, 1965 _igvn.MakeConX(distance) ); 1966 transform_later(prefetch_adr); 1967 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1968 transform_later(prefetch); 1969 distance += step_size; 1970 i_o = prefetch; 1971 } 1972 pf_phi_abio->set_req( pf_path, i_o ); 1973 1974 pf_region->init_req( fall_in_path, need_pf_false ); 1975 pf_region->init_req( pf_path, need_pf_true ); 1976 1977 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 1978 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 1979 1980 transform_later(pf_region); 1981 transform_later(pf_phi_rawmem); 1982 transform_later(pf_phi_abio); 1983 1984 needgc_false = pf_region; 1985 contended_phi_rawmem = pf_phi_rawmem; 1986 i_o = pf_phi_abio; 1987 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { 1988 // Insert a prefetch instruction for each allocation. 1989 // This code is used to generate 1 prefetch instruction per cache line. 1990 1991 // Generate several prefetch instructions. 1992 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1993 uint step_size = AllocatePrefetchStepSize; 1994 uint distance = AllocatePrefetchDistance; 1995 1996 // Next cache address. 1997 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top, 1998 _igvn.MakeConX(step_size + distance)); 1999 transform_later(cache_adr); 2000 cache_adr = new CastP2XNode(needgc_false, cache_adr); 2001 transform_later(cache_adr); 2002 // Address is aligned to execute prefetch to the beginning of cache line size 2003 // (it is important when BIS instruction is used on SPARC as prefetch). 2004 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); 2005 cache_adr = new AndXNode(cache_adr, mask); 2006 transform_later(cache_adr); 2007 cache_adr = new CastX2PNode(cache_adr); 2008 transform_later(cache_adr); 2009 2010 // Prefetch 2011 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); 2012 prefetch->set_req(0, needgc_false); 2013 transform_later(prefetch); 2014 contended_phi_rawmem = prefetch; 2015 Node *prefetch_adr; 2016 distance = step_size; 2017 for ( uint i = 1; i < lines; i++ ) { 2018 prefetch_adr = new AddPNode( cache_adr, cache_adr, 2019 _igvn.MakeConX(distance) ); 2020 transform_later(prefetch_adr); 2021 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); 2022 transform_later(prefetch); 2023 distance += step_size; 2024 contended_phi_rawmem = prefetch; 2025 } 2026 } else if( AllocatePrefetchStyle > 0 ) { 2027 // Insert a prefetch for each allocation only on the fast-path 2028 Node *prefetch_adr; 2029 Node *prefetch; 2030 // Generate several prefetch instructions. 2031 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 2032 uint step_size = AllocatePrefetchStepSize; 2033 uint distance = AllocatePrefetchDistance; 2034 for ( uint i = 0; i < lines; i++ ) { 2035 prefetch_adr = new AddPNode( old_eden_top, new_eden_top, 2036 _igvn.MakeConX(distance) ); 2037 transform_later(prefetch_adr); 2038 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 2039 // Do not let it float too high, since if eden_top == eden_end, 2040 // both might be null. 2041 if( i == 0 ) { // Set control for first prefetch, next follows it 2042 prefetch->init_req(0, needgc_false); 2043 } 2044 transform_later(prefetch); 2045 distance += step_size; 2046 i_o = prefetch; 2047 } 2048 } 2049 return i_o; 2050 } 2051 2052 2053 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 2054 expand_allocate_common(alloc, NULL, 2055 OptoRuntime::new_instance_Type(), 2056 OptoRuntime::new_instance_Java()); 2057 } 2058 2059 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 2060 Node* length = alloc->in(AllocateNode::ALength); 2061 InitializeNode* init = alloc->initialization(); 2062 Node* klass_node = alloc->in(AllocateNode::KlassNode); 2063 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 2064 address slow_call_address; // Address of slow call 2065 if (init != NULL && init->is_complete_with_arraycopy() && 2066 k->is_type_array_klass()) { 2067 // Don't zero type array during slow allocation in VM since 2068 // it will be initialized later by arraycopy in compiled code. 2069 slow_call_address = OptoRuntime::new_array_nozero_Java(); 2070 } else { 2071 slow_call_address = OptoRuntime::new_array_Java(); 2072 } 2073 expand_allocate_common(alloc, length, 2074 OptoRuntime::new_array_Type(), 2075 slow_call_address); 2076 } 2077 2078 //-------------------mark_eliminated_box---------------------------------- 2079 // 2080 // During EA obj may point to several objects but after few ideal graph 2081 // transformations (CCP) it may point to only one non escaping object 2082 // (but still using phi), corresponding locks and unlocks will be marked 2083 // for elimination. Later obj could be replaced with a new node (new phi) 2084 // and which does not have escape information. And later after some graph 2085 // reshape other locks and unlocks (which were not marked for elimination 2086 // before) are connected to this new obj (phi) but they still will not be 2087 // marked for elimination since new obj has no escape information. 2088 // Mark all associated (same box and obj) lock and unlock nodes for 2089 // elimination if some of them marked already. 2090 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) { 2091 if (oldbox->as_BoxLock()->is_eliminated()) 2092 return; // This BoxLock node was processed already. 2093 2094 // New implementation (EliminateNestedLocks) has separate BoxLock 2095 // node for each locked region so mark all associated locks/unlocks as 2096 // eliminated even if different objects are referenced in one locked region 2097 // (for example, OSR compilation of nested loop inside locked scope). 2098 if (EliminateNestedLocks || 2099 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) { 2100 // Box is used only in one lock region. Mark this box as eliminated. 2101 _igvn.hash_delete(oldbox); 2102 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value 2103 _igvn.hash_insert(oldbox); 2104 2105 for (uint i = 0; i < oldbox->outcnt(); i++) { 2106 Node* u = oldbox->raw_out(i); 2107 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) { 2108 AbstractLockNode* alock = u->as_AbstractLock(); 2109 // Check lock's box since box could be referenced by Lock's debug info. 2110 if (alock->box_node() == oldbox) { 2111 // Mark eliminated all related locks and unlocks. 2112 #ifdef ASSERT 2113 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4"); 2114 #endif 2115 alock->set_non_esc_obj(); 2116 } 2117 } 2118 } 2119 return; 2120 } 2121 2122 // Create new "eliminated" BoxLock node and use it in monitor debug info 2123 // instead of oldbox for the same object. 2124 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 2125 2126 // Note: BoxLock node is marked eliminated only here and it is used 2127 // to indicate that all associated lock and unlock nodes are marked 2128 // for elimination. 2129 newbox->set_eliminated(); 2130 transform_later(newbox); 2131 2132 // Replace old box node with new box for all users of the same object. 2133 for (uint i = 0; i < oldbox->outcnt();) { 2134 bool next_edge = true; 2135 2136 Node* u = oldbox->raw_out(i); 2137 if (u->is_AbstractLock()) { 2138 AbstractLockNode* alock = u->as_AbstractLock(); 2139 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) { 2140 // Replace Box and mark eliminated all related locks and unlocks. 2141 #ifdef ASSERT 2142 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5"); 2143 #endif 2144 alock->set_non_esc_obj(); 2145 _igvn.rehash_node_delayed(alock); 2146 alock->set_box_node(newbox); 2147 next_edge = false; 2148 } 2149 } 2150 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) { 2151 FastLockNode* flock = u->as_FastLock(); 2152 assert(flock->box_node() == oldbox, "sanity"); 2153 _igvn.rehash_node_delayed(flock); 2154 flock->set_box_node(newbox); 2155 next_edge = false; 2156 } 2157 2158 // Replace old box in monitor debug info. 2159 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 2160 SafePointNode* sfn = u->as_SafePoint(); 2161 JVMState* youngest_jvms = sfn->jvms(); 2162 int max_depth = youngest_jvms->depth(); 2163 for (int depth = 1; depth <= max_depth; depth++) { 2164 JVMState* jvms = youngest_jvms->of_depth(depth); 2165 int num_mon = jvms->nof_monitors(); 2166 // Loop over monitors 2167 for (int idx = 0; idx < num_mon; idx++) { 2168 Node* obj_node = sfn->monitor_obj(jvms, idx); 2169 Node* box_node = sfn->monitor_box(jvms, idx); 2170 if (box_node == oldbox && obj_node->eqv_uncast(obj)) { 2171 int j = jvms->monitor_box_offset(idx); 2172 _igvn.replace_input_of(u, j, newbox); 2173 next_edge = false; 2174 } 2175 } 2176 } 2177 } 2178 if (next_edge) i++; 2179 } 2180 } 2181 2182 //-----------------------mark_eliminated_locking_nodes----------------------- 2183 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { 2184 if (EliminateNestedLocks) { 2185 if (alock->is_nested()) { 2186 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity"); 2187 return; 2188 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened 2189 // Only Lock node has JVMState needed here. 2190 // Not that preceding claim is documented anywhere else. 2191 if (alock->jvms() != NULL) { 2192 if (alock->as_Lock()->is_nested_lock_region()) { 2193 // Mark eliminated related nested locks and unlocks. 2194 Node* obj = alock->obj_node(); 2195 BoxLockNode* box_node = alock->box_node()->as_BoxLock(); 2196 assert(!box_node->is_eliminated(), "should not be marked yet"); 2197 // Note: BoxLock node is marked eliminated only here 2198 // and it is used to indicate that all associated lock 2199 // and unlock nodes are marked for elimination. 2200 box_node->set_eliminated(); // Box's hash is always NO_HASH here 2201 for (uint i = 0; i < box_node->outcnt(); i++) { 2202 Node* u = box_node->raw_out(i); 2203 if (u->is_AbstractLock()) { 2204 alock = u->as_AbstractLock(); 2205 if (alock->box_node() == box_node) { 2206 // Verify that this Box is referenced only by related locks. 2207 assert(alock->obj_node()->eqv_uncast(obj), ""); 2208 // Mark all related locks and unlocks. 2209 #ifdef ASSERT 2210 alock->log_lock_optimization(C, "eliminate_lock_set_nested"); 2211 #endif 2212 alock->set_nested(); 2213 } 2214 } 2215 } 2216 } else { 2217 #ifdef ASSERT 2218 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region"); 2219 if (C->log() != NULL) 2220 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output 2221 #endif 2222 } 2223 } 2224 return; 2225 } 2226 // Process locks for non escaping object 2227 assert(alock->is_non_esc_obj(), ""); 2228 } // EliminateNestedLocks 2229 2230 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object 2231 // Look for all locks of this object and mark them and 2232 // corresponding BoxLock nodes as eliminated. 2233 Node* obj = alock->obj_node(); 2234 for (uint j = 0; j < obj->outcnt(); j++) { 2235 Node* o = obj->raw_out(j); 2236 if (o->is_AbstractLock() && 2237 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) { 2238 alock = o->as_AbstractLock(); 2239 Node* box = alock->box_node(); 2240 // Replace old box node with new eliminated box for all users 2241 // of the same object and mark related locks as eliminated. 2242 mark_eliminated_box(box, obj); 2243 } 2244 } 2245 } 2246 } 2247 2248 // we have determined that this lock/unlock can be eliminated, we simply 2249 // eliminate the node without expanding it. 2250 // 2251 // Note: The membar's associated with the lock/unlock are currently not 2252 // eliminated. This should be investigated as a future enhancement. 2253 // 2254 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 2255 2256 if (!alock->is_eliminated()) { 2257 return false; 2258 } 2259 #ifdef ASSERT 2260 if (!alock->is_coarsened()) { 2261 // Check that new "eliminated" BoxLock node is created. 2262 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 2263 assert(oldbox->is_eliminated(), "should be done already"); 2264 } 2265 #endif 2266 2267 alock->log_lock_optimization(C, "eliminate_lock"); 2268 2269 #ifndef PRODUCT 2270 if (PrintEliminateLocks) { 2271 if (alock->is_Lock()) { 2272 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx); 2273 } else { 2274 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx); 2275 } 2276 } 2277 #endif 2278 2279 Node* mem = alock->in(TypeFunc::Memory); 2280 Node* ctrl = alock->in(TypeFunc::Control); 2281 2282 extract_call_projections(alock); 2283 // There are 2 projections from the lock. The lock node will 2284 // be deleted when its last use is subsumed below. 2285 assert(alock->outcnt() == 2 && 2286 _fallthroughproj != NULL && 2287 _memproj_fallthrough != NULL, 2288 "Unexpected projections from Lock/Unlock"); 2289 2290 Node* fallthroughproj = _fallthroughproj; 2291 Node* memproj_fallthrough = _memproj_fallthrough; 2292 2293 // The memory projection from a lock/unlock is RawMem 2294 // The input to a Lock is merged memory, so extract its RawMem input 2295 // (unless the MergeMem has been optimized away.) 2296 if (alock->is_Lock()) { 2297 // Seach for MemBarAcquireLock node and delete it also. 2298 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 2299 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, ""); 2300 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 2301 Node* memproj = membar->proj_out(TypeFunc::Memory); 2302 _igvn.replace_node(ctrlproj, fallthroughproj); 2303 _igvn.replace_node(memproj, memproj_fallthrough); 2304 2305 // Delete FastLock node also if this Lock node is unique user 2306 // (a loop peeling may clone a Lock node). 2307 Node* flock = alock->as_Lock()->fastlock_node(); 2308 if (flock->outcnt() == 1) { 2309 assert(flock->unique_out() == alock, "sanity"); 2310 _igvn.replace_node(flock, top()); 2311 } 2312 } 2313 2314 // Seach for MemBarReleaseLock node and delete it also. 2315 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() && 2316 ctrl->in(0)->is_MemBar()) { 2317 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 2318 assert(membar->Opcode() == Op_MemBarReleaseLock && 2319 mem->is_Proj() && membar == mem->in(0), ""); 2320 _igvn.replace_node(fallthroughproj, ctrl); 2321 _igvn.replace_node(memproj_fallthrough, mem); 2322 fallthroughproj = ctrl; 2323 memproj_fallthrough = mem; 2324 ctrl = membar->in(TypeFunc::Control); 2325 mem = membar->in(TypeFunc::Memory); 2326 } 2327 2328 _igvn.replace_node(fallthroughproj, ctrl); 2329 _igvn.replace_node(memproj_fallthrough, mem); 2330 return true; 2331 } 2332 2333 2334 //------------------------------expand_lock_node---------------------- 2335 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 2336 2337 Node* ctrl = lock->in(TypeFunc::Control); 2338 Node* mem = lock->in(TypeFunc::Memory); 2339 Node* obj = lock->obj_node(); 2340 Node* box = lock->box_node(); 2341 Node* flock = lock->fastlock_node(); 2342 2343 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2344 2345 // Make the merge point 2346 Node *region; 2347 Node *mem_phi; 2348 Node *slow_path; 2349 2350 if (UseOptoBiasInlining) { 2351 /* 2352 * See the full description in MacroAssembler::biased_locking_enter(). 2353 * 2354 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { 2355 * // The object is biased. 2356 * proto_node = klass->prototype_header; 2357 * o_node = thread | proto_node; 2358 * x_node = o_node ^ mark_word; 2359 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? 2360 * // Done. 2361 * } else { 2362 * if( (x_node & biased_lock_mask) != 0 ) { 2363 * // The klass's prototype header is no longer biased. 2364 * cas(&mark_word, mark_word, proto_node) 2365 * goto cas_lock; 2366 * } else { 2367 * // The klass's prototype header is still biased. 2368 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? 2369 * old = mark_word; 2370 * new = o_node; 2371 * } else { 2372 * // Different thread or anonymous biased. 2373 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); 2374 * new = thread | old; 2375 * } 2376 * // Try to rebias. 2377 * if( cas(&mark_word, old, new) == 0 ) { 2378 * // Done. 2379 * } else { 2380 * goto slow_path; // Failed. 2381 * } 2382 * } 2383 * } 2384 * } else { 2385 * // The object is not biased. 2386 * cas_lock: 2387 * if( FastLock(obj) == 0 ) { 2388 * // Done. 2389 * } else { 2390 * slow_path: 2391 * OptoRuntime::complete_monitor_locking_Java(obj); 2392 * } 2393 * } 2394 */ 2395 2396 region = new RegionNode(5); 2397 // create a Phi for the memory state 2398 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2399 2400 Node* fast_lock_region = new RegionNode(3); 2401 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); 2402 2403 // First, check mark word for the biased lock pattern. 2404 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2405 2406 // Get fast path - mark word has the biased lock pattern. 2407 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, 2408 markOopDesc::biased_lock_mask_in_place, 2409 markOopDesc::biased_lock_pattern, true); 2410 // fast_lock_region->in(1) is set to slow path. 2411 fast_lock_mem_phi->init_req(1, mem); 2412 2413 // Now check that the lock is biased to the current thread and has 2414 // the same epoch and bias as Klass::_prototype_header. 2415 2416 // Special-case a fresh allocation to avoid building nodes: 2417 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); 2418 if (klass_node == NULL) { 2419 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 2420 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr())); 2421 #ifdef _LP64 2422 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) { 2423 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity"); 2424 klass_node->in(1)->init_req(0, ctrl); 2425 } else 2426 #endif 2427 klass_node->init_req(0, ctrl); 2428 } 2429 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type()); 2430 2431 Node* thread = transform_later(new ThreadLocalNode()); 2432 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2433 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node)); 2434 Node* x_node = transform_later(new XorXNode(o_node, mark_node)); 2435 2436 // Get slow path - mark word does NOT match the value. 2437 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, 2438 (~markOopDesc::age_mask_in_place), 0); 2439 // region->in(3) is set to fast path - the object is biased to the current thread. 2440 mem_phi->init_req(3, mem); 2441 2442 2443 // Mark word does NOT match the value (thread | Klass::_prototype_header). 2444 2445 2446 // First, check biased pattern. 2447 // Get fast path - _prototype_header has the same biased lock pattern. 2448 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, 2449 markOopDesc::biased_lock_mask_in_place, 0, true); 2450 2451 not_biased_ctrl = fast_lock_region->in(2); // Slow path 2452 // fast_lock_region->in(2) - the prototype header is no longer biased 2453 // and we have to revoke the bias on this object. 2454 // We are going to try to reset the mark of this object to the prototype 2455 // value and fall through to the CAS-based locking scheme. 2456 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 2457 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr, 2458 proto_node, mark_node); 2459 transform_later(cas); 2460 Node* proj = transform_later(new SCMemProjNode(cas)); 2461 fast_lock_mem_phi->init_req(2, proj); 2462 2463 2464 // Second, check epoch bits. 2465 Node* rebiased_region = new RegionNode(3); 2466 Node* old_phi = new PhiNode( rebiased_region, TypeX_X); 2467 Node* new_phi = new PhiNode( rebiased_region, TypeX_X); 2468 2469 // Get slow path - mark word does NOT match epoch bits. 2470 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, 2471 markOopDesc::epoch_mask_in_place, 0); 2472 // The epoch of the current bias is not valid, attempt to rebias the object 2473 // toward the current thread. 2474 rebiased_region->init_req(2, epoch_ctrl); 2475 old_phi->init_req(2, mark_node); 2476 new_phi->init_req(2, o_node); 2477 2478 // rebiased_region->in(1) is set to fast path. 2479 // The epoch of the current bias is still valid but we know 2480 // nothing about the owner; it might be set or it might be clear. 2481 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place | 2482 markOopDesc::age_mask_in_place | 2483 markOopDesc::epoch_mask_in_place); 2484 Node* old = transform_later(new AndXNode(mark_node, cmask)); 2485 cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2486 Node* new_mark = transform_later(new OrXNode(cast_thread, old)); 2487 old_phi->init_req(1, old); 2488 new_phi->init_req(1, new_mark); 2489 2490 transform_later(rebiased_region); 2491 transform_later(old_phi); 2492 transform_later(new_phi); 2493 2494 // Try to acquire the bias of the object using an atomic operation. 2495 // If this fails we will go in to the runtime to revoke the object's bias. 2496 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi); 2497 transform_later(cas); 2498 proj = transform_later(new SCMemProjNode(cas)); 2499 2500 // Get slow path - Failed to CAS. 2501 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); 2502 mem_phi->init_req(4, proj); 2503 // region->in(4) is set to fast path - the object is rebiased to the current thread. 2504 2505 // Failed to CAS. 2506 slow_path = new RegionNode(3); 2507 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); 2508 2509 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control 2510 slow_mem->init_req(1, proj); 2511 2512 // Call CAS-based locking scheme (FastLock node). 2513 2514 transform_later(fast_lock_region); 2515 transform_later(fast_lock_mem_phi); 2516 2517 // Get slow path - FastLock failed to lock the object. 2518 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); 2519 mem_phi->init_req(2, fast_lock_mem_phi); 2520 // region->in(2) is set to fast path - the object is locked to the current thread. 2521 2522 slow_path->init_req(2, ctrl); // Capture slow-control 2523 slow_mem->init_req(2, fast_lock_mem_phi); 2524 2525 transform_later(slow_path); 2526 transform_later(slow_mem); 2527 // Reset lock's memory edge. 2528 lock->set_req(TypeFunc::Memory, slow_mem); 2529 2530 } else { 2531 region = new RegionNode(3); 2532 // create a Phi for the memory state 2533 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2534 2535 // Optimize test; set region slot 2 2536 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); 2537 mem_phi->init_req(2, mem); 2538 } 2539 2540 // Make slow path call 2541 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), 2542 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, 2543 obj, box, NULL); 2544 2545 extract_call_projections(call); 2546 2547 // Slow path can only throw asynchronous exceptions, which are always 2548 // de-opted. So the compiler thinks the slow-call can never throw an 2549 // exception. If it DOES throw an exception we would need the debug 2550 // info removed first (since if it throws there is no monitor). 2551 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2552 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2553 2554 // Capture slow path 2555 // disconnect fall-through projection from call and create a new one 2556 // hook up users of fall-through projection to region 2557 Node *slow_ctrl = _fallthroughproj->clone(); 2558 transform_later(slow_ctrl); 2559 _igvn.hash_delete(_fallthroughproj); 2560 _fallthroughproj->disconnect_inputs(NULL, C); 2561 region->init_req(1, slow_ctrl); 2562 // region inputs are now complete 2563 transform_later(region); 2564 _igvn.replace_node(_fallthroughproj, region); 2565 2566 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory)); 2567 mem_phi->init_req(1, memproj ); 2568 transform_later(mem_phi); 2569 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2570 } 2571 2572 //------------------------------expand_unlock_node---------------------- 2573 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 2574 2575 Node* ctrl = unlock->in(TypeFunc::Control); 2576 Node* mem = unlock->in(TypeFunc::Memory); 2577 Node* obj = unlock->obj_node(); 2578 Node* box = unlock->box_node(); 2579 2580 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2581 2582 // No need for a null check on unlock 2583 2584 // Make the merge point 2585 Node *region; 2586 Node *mem_phi; 2587 2588 if (UseOptoBiasInlining) { 2589 // Check for biased locking unlock case, which is a no-op. 2590 // See the full description in MacroAssembler::biased_locking_exit(). 2591 region = new RegionNode(4); 2592 // create a Phi for the memory state 2593 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2594 mem_phi->init_req(3, mem); 2595 2596 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2597 ctrl = opt_bits_test(ctrl, region, 3, mark_node, 2598 markOopDesc::biased_lock_mask_in_place, 2599 markOopDesc::biased_lock_pattern); 2600 } else { 2601 region = new RegionNode(3); 2602 // create a Phi for the memory state 2603 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2604 } 2605 2606 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box ); 2607 funlock = transform_later( funlock )->as_FastUnlock(); 2608 // Optimize test; set region slot 2 2609 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); 2610 Node *thread = transform_later(new ThreadLocalNode()); 2611 2612 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), 2613 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), 2614 "complete_monitor_unlocking_C", slow_path, obj, box, thread); 2615 2616 extract_call_projections(call); 2617 2618 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2619 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2620 2621 // No exceptions for unlocking 2622 // Capture slow path 2623 // disconnect fall-through projection from call and create a new one 2624 // hook up users of fall-through projection to region 2625 Node *slow_ctrl = _fallthroughproj->clone(); 2626 transform_later(slow_ctrl); 2627 _igvn.hash_delete(_fallthroughproj); 2628 _fallthroughproj->disconnect_inputs(NULL, C); 2629 region->init_req(1, slow_ctrl); 2630 // region inputs are now complete 2631 transform_later(region); 2632 _igvn.replace_node(_fallthroughproj, region); 2633 2634 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) ); 2635 mem_phi->init_req(1, memproj ); 2636 mem_phi->init_req(2, mem); 2637 transform_later(mem_phi); 2638 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2639 } 2640 2641 // A value type is returned from the call but we don't know its 2642 // type. Either we get a buffered value (and nothing needs to be done) 2643 // or one of the values being returned is the klass of the value type 2644 // and we need to allocate a value type instance of that type and 2645 // initialize it with other values being returned. In that case, we 2646 // first try a fast path allocation and initialize the value with the 2647 // value klass's pack handler or we fall back to a runtime call. 2648 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) { 2649 Node* ret = call->proj_out(TypeFunc::Parms); 2650 if (ret == NULL) { 2651 return; 2652 } 2653 assert(ret->bottom_type()->is_valuetypeptr()->klass() == C->env()->___Value_klass(), "unexpected return type from MH intrinsic"); 2654 const TypeFunc* tf = call->_tf; 2655 const TypeTuple* domain = OptoRuntime::store_value_type_fields_Type()->domain_cc(); 2656 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain); 2657 call->_tf = new_tf; 2658 // Make sure the change of type is applied before projections are 2659 // processed by igvn 2660 _igvn.set_type(call, call->Value(&_igvn)); 2661 _igvn.set_type(ret, ret->Value(&_igvn)); 2662 2663 // Before any new projection is added: 2664 CallProjections projs; 2665 call->extract_projections(&projs, true, true); 2666 2667 Node* ctl = new Node(1); 2668 Node* mem = new Node(1); 2669 Node* io = new Node(1); 2670 Node* ex_ctl = new Node(1); 2671 Node* ex_mem = new Node(1); 2672 Node* ex_io = new Node(1); 2673 Node* res = new Node(1); 2674 2675 Node* cast = transform_later(new CastP2XNode(ctl, res)); 2676 Node* mask = MakeConX(0x1); 2677 Node* masked = transform_later(new AndXNode(cast, mask)); 2678 Node* cmp = transform_later(new CmpXNode(masked, mask)); 2679 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq)); 2680 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN); 2681 transform_later(allocation_iff); 2682 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff)); 2683 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff)); 2684 2685 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeValueTypePtr::NOTNULL)); 2686 2687 Node* mask2 = MakeConX(-2); 2688 Node* masked2 = transform_later(new AndXNode(cast, mask2)); 2689 Node* rawklassptr = transform_later(new CastX2PNode(masked2)); 2690 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeKlassPtr::VALUE)); 2691 2692 Node* top_adr; 2693 Node* end_adr; 2694 2695 Node* slowpath_bol = NULL; 2696 Node* old_top = NULL; 2697 Node* new_top = NULL; 2698 if (UseTLAB) { 2699 set_eden_pointers(top_adr, end_adr); 2700 Node* end = make_load(ctl, mem, end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 2701 old_top = new LoadPNode(ctl, mem, top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered); 2702 transform_later(old_top); 2703 Node* layout_val = make_load(NULL, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT); 2704 Node* size_in_bytes = ConvI2X(layout_val); 2705 new_top = new AddPNode(top(), old_top, size_in_bytes); 2706 transform_later(new_top); 2707 Node* slowpath_cmp = new CmpPNode(new_top, end); 2708 transform_later(slowpath_cmp); 2709 slowpath_bol = new BoolNode(slowpath_cmp, BoolTest::ge); 2710 transform_later(slowpath_bol); 2711 } else { 2712 slowpath_bol = intcon(1); 2713 old_top = top(); 2714 new_top = top(); 2715 } 2716 IfNode* slowpath_iff = new IfNode(allocation_ctl, slowpath_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); 2717 transform_later(slowpath_iff); 2718 2719 Node* slowpath_true = new IfTrueNode(slowpath_iff); 2720 transform_later(slowpath_true); 2721 2722 2723 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_value_type_fields_Type(), 2724 StubRoutines::store_value_type_fields_to_buf(), 2725 "store_value_type_fields", 2726 call->jvms()->bci(), 2727 TypePtr::BOTTOM); 2728 slow_call->init_req(TypeFunc::Control, slowpath_true); 2729 slow_call->init_req(TypeFunc::Memory, mem); 2730 slow_call->init_req(TypeFunc::I_O, io); 2731 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr)); 2732 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr)); 2733 slow_call->init_req(TypeFunc::Parms, res); 2734 2735 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control)); 2736 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory)); 2737 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O)); 2738 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms)); 2739 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2)); 2740 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci)); 2741 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci)); 2742 2743 Node* ex_r = new RegionNode(3); 2744 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM); 2745 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO); 2746 ex_r->init_req(1, slow_excp); 2747 ex_mem_phi->init_req(1, slow_mem); 2748 ex_io_phi->init_req(1, slow_io); 2749 ex_r->init_req(2, ex_ctl); 2750 ex_mem_phi->init_req(2, ex_mem); 2751 ex_io_phi->init_req(2, ex_io); 2752 2753 transform_later(ex_r); 2754 transform_later(ex_mem_phi); 2755 transform_later(ex_io_phi); 2756 2757 Node* slowpath_false = new IfFalseNode(slowpath_iff); 2758 transform_later(slowpath_false); 2759 Node* rawmem = new StorePNode(slowpath_false, mem, top_adr, TypeRawPtr::BOTTOM, new_top, MemNode::unordered); 2760 transform_later(rawmem); 2761 Node* mark_node = NULL; 2762 // For now only enable fast locking for non-array types 2763 if (UseBiasedLocking) { 2764 mark_node = make_load(slowpath_false, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 2765 } else { 2766 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); 2767 } 2768 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 2769 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 2770 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT); 2771 Node* pack_handler = make_load(slowpath_false, rawmem, klass_node, in_bytes(ValueKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 2772 2773 CallLeafNoFPNode* handler_call = new CallLeafNoFPNode(OptoRuntime::pack_value_type_Type(), 2774 NULL, 2775 "pack handler", 2776 TypeRawPtr::BOTTOM); 2777 handler_call->init_req(TypeFunc::Control, slowpath_false); 2778 handler_call->init_req(TypeFunc::Memory, rawmem); 2779 handler_call->init_req(TypeFunc::I_O, top()); 2780 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr)); 2781 handler_call->init_req(TypeFunc::ReturnAdr, top()); 2782 handler_call->init_req(TypeFunc::Parms, pack_handler); 2783 handler_call->init_req(TypeFunc::Parms+1, old_top); 2784 2785 // We don't know how many values are returned. This assumes the 2786 // worst case, that all available registers are used. 2787 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) { 2788 if (domain->field_at(i) == Type::HALF) { 2789 slow_call->init_req(i, top()); 2790 handler_call->init_req(i+1, top()); 2791 continue; 2792 } 2793 Node* proj = transform_later(new ProjNode(call, i)); 2794 slow_call->init_req(i, proj); 2795 handler_call->init_req(i+1, proj); 2796 } 2797 2798 // We can safepoint at that new call 2799 C->add_safepoint_edges(slow_call, call->jvms()); 2800 transform_later(slow_call); 2801 transform_later(handler_call); 2802 2803 Node* handler_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control)); 2804 rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory)); 2805 Node* slowpath_false_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms)); 2806 2807 MergeMemNode* slowpath_false_mem = MergeMemNode::make(mem); 2808 slowpath_false_mem->set_memory_at(Compile::AliasIdxRaw, rawmem); 2809 transform_later(slowpath_false_mem); 2810 2811 Node* r = new RegionNode(4); 2812 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM); 2813 Node* io_phi = new PhiNode(r, Type::ABIO); 2814 Node* res_phi = new PhiNode(r, ret->bottom_type()); 2815 2816 r->init_req(1, no_allocation_ctl); 2817 mem_phi->init_req(1, mem); 2818 io_phi->init_req(1, io); 2819 res_phi->init_req(1, no_allocation_res); 2820 r->init_req(2, slow_norm); 2821 mem_phi->init_req(2, slow_mem); 2822 io_phi->init_req(2, slow_io); 2823 res_phi->init_req(2, slow_res); 2824 r->init_req(3, handler_ctl); 2825 mem_phi->init_req(3, slowpath_false_mem); 2826 io_phi->init_req(3, io); 2827 res_phi->init_req(3, slowpath_false_res); 2828 2829 transform_later(r); 2830 transform_later(mem_phi); 2831 transform_later(io_phi); 2832 transform_later(res_phi); 2833 2834 _igvn.replace_in_uses(projs.fallthrough_catchproj, r); 2835 _igvn.replace_in_uses(projs.fallthrough_memproj, mem_phi); 2836 _igvn.replace_in_uses(projs.fallthrough_ioproj, io_phi); 2837 _igvn.replace_in_uses(projs.resproj, res_phi); 2838 _igvn.replace_in_uses(projs.catchall_catchproj, ex_r); 2839 _igvn.replace_in_uses(projs.catchall_memproj, ex_mem_phi); 2840 _igvn.replace_in_uses(projs.catchall_ioproj, ex_io_phi); 2841 2842 _igvn.replace_node(ctl, projs.fallthrough_catchproj); 2843 _igvn.replace_node(mem, projs.fallthrough_memproj); 2844 _igvn.replace_node(io, projs.fallthrough_ioproj); 2845 _igvn.replace_node(res, projs.resproj); 2846 _igvn.replace_node(ex_ctl, projs.catchall_catchproj); 2847 _igvn.replace_node(ex_mem, projs.catchall_memproj); 2848 _igvn.replace_node(ex_io, projs.catchall_ioproj); 2849 } 2850 2851 //---------------------------eliminate_macro_nodes---------------------- 2852 // Eliminate scalar replaced allocations and associated locks. 2853 void PhaseMacroExpand::eliminate_macro_nodes() { 2854 if (C->macro_count() == 0) 2855 return; 2856 2857 // First, attempt to eliminate locks 2858 int cnt = C->macro_count(); 2859 for (int i=0; i < cnt; i++) { 2860 Node *n = C->macro_node(i); 2861 if (n->is_AbstractLock()) { // Lock and Unlock nodes 2862 // Before elimination mark all associated (same box and obj) 2863 // lock and unlock nodes. 2864 mark_eliminated_locking_nodes(n->as_AbstractLock()); 2865 } 2866 } 2867 bool progress = true; 2868 while (progress) { 2869 progress = false; 2870 for (int i = C->macro_count(); i > 0; i--) { 2871 Node * n = C->macro_node(i-1); 2872 bool success = false; 2873 debug_only(int old_macro_count = C->macro_count();); 2874 if (n->is_AbstractLock()) { 2875 success = eliminate_locking_node(n->as_AbstractLock()); 2876 } 2877 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2878 progress = progress || success; 2879 } 2880 } 2881 // Next, attempt to eliminate allocations 2882 _has_locks = false; 2883 progress = true; 2884 while (progress) { 2885 progress = false; 2886 for (int i = C->macro_count(); i > 0; i--) { 2887 Node * n = C->macro_node(i-1); 2888 bool success = false; 2889 debug_only(int old_macro_count = C->macro_count();); 2890 switch (n->class_id()) { 2891 case Node::Class_Allocate: 2892 case Node::Class_AllocateArray: 2893 success = eliminate_allocate_node(n->as_Allocate()); 2894 break; 2895 case Node::Class_CallStaticJava: { 2896 CallStaticJavaNode* call = n->as_CallStaticJava(); 2897 if (!call->method()->is_method_handle_intrinsic()) { 2898 success = eliminate_boxing_node(n->as_CallStaticJava()); 2899 } 2900 break; 2901 } 2902 case Node::Class_Lock: 2903 case Node::Class_Unlock: 2904 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 2905 _has_locks = true; 2906 break; 2907 case Node::Class_ArrayCopy: 2908 break; 2909 default: 2910 assert(n->Opcode() == Op_LoopLimit || 2911 n->Opcode() == Op_Opaque1 || 2912 n->Opcode() == Op_Opaque2 || 2913 n->Opcode() == Op_Opaque3 || 2914 n->Opcode() == Op_Opaque4, "unknown node type in macro list"); 2915 } 2916 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2917 progress = progress || success; 2918 } 2919 } 2920 } 2921 2922 //------------------------------expand_macro_nodes---------------------- 2923 // Returns true if a failure occurred. 2924 bool PhaseMacroExpand::expand_macro_nodes() { 2925 // Last attempt to eliminate macro nodes. 2926 eliminate_macro_nodes(); 2927 2928 // Make sure expansion will not cause node limit to be exceeded. 2929 // Worst case is a macro node gets expanded into about 200 nodes. 2930 // Allow 50% more for optimization. 2931 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) ) 2932 return true; 2933 2934 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. 2935 bool progress = true; 2936 while (progress) { 2937 progress = false; 2938 for (int i = C->macro_count(); i > 0; i--) { 2939 Node * n = C->macro_node(i-1); 2940 bool success = false; 2941 debug_only(int old_macro_count = C->macro_count();); 2942 if (n->Opcode() == Op_LoopLimit) { 2943 // Remove it from macro list and put on IGVN worklist to optimize. 2944 C->remove_macro_node(n); 2945 _igvn._worklist.push(n); 2946 success = true; 2947 } else if (n->Opcode() == Op_CallStaticJava) { 2948 CallStaticJavaNode* call = n->as_CallStaticJava(); 2949 if (!call->method()->is_method_handle_intrinsic()) { 2950 // Remove it from macro list and put on IGVN worklist to optimize. 2951 C->remove_macro_node(n); 2952 _igvn._worklist.push(n); 2953 success = true; 2954 } 2955 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { 2956 _igvn.replace_node(n, n->in(1)); 2957 success = true; 2958 #if INCLUDE_RTM_OPT 2959 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) { 2960 assert(C->profile_rtm(), "should be used only in rtm deoptimization code"); 2961 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), ""); 2962 Node* cmp = n->unique_out(); 2963 #ifdef ASSERT 2964 // Validate graph. 2965 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), ""); 2966 BoolNode* bol = cmp->unique_out()->as_Bool(); 2967 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() && 2968 (bol->_test._test == BoolTest::ne), ""); 2969 IfNode* ifn = bol->unique_out()->as_If(); 2970 assert((ifn->outcnt() == 2) && 2971 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, ""); 2972 #endif 2973 Node* repl = n->in(1); 2974 if (!_has_locks) { 2975 // Remove RTM state check if there are no locks in the code. 2976 // Replace input to compare the same value. 2977 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1); 2978 } 2979 _igvn.replace_node(n, repl); 2980 success = true; 2981 #endif 2982 } else if (n->Opcode() == Op_Opaque4) { 2983 _igvn.replace_node(n, n->in(2)); 2984 success = true; 2985 } 2986 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2987 progress = progress || success; 2988 } 2989 } 2990 2991 // expand arraycopy "macro" nodes first 2992 // For ReduceBulkZeroing, we must first process all arraycopy nodes 2993 // before the allocate nodes are expanded. 2994 int macro_idx = C->macro_count() - 1; 2995 while (macro_idx >= 0) { 2996 Node * n = C->macro_node(macro_idx); 2997 assert(n->is_macro(), "only macro nodes expected here"); 2998 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2999 // node is unreachable, so don't try to expand it 3000 C->remove_macro_node(n); 3001 } else if (n->is_ArrayCopy()){ 3002 int macro_count = C->macro_count(); 3003 expand_arraycopy_node(n->as_ArrayCopy()); 3004 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 3005 } 3006 if (C->failing()) return true; 3007 macro_idx --; 3008 } 3009 3010 // expand "macro" nodes 3011 // nodes are removed from the macro list as they are processed 3012 while (C->macro_count() > 0) { 3013 int macro_count = C->macro_count(); 3014 Node * n = C->macro_node(macro_count-1); 3015 assert(n->is_macro(), "only macro nodes expected here"); 3016 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 3017 // node is unreachable, so don't try to expand it 3018 C->remove_macro_node(n); 3019 continue; 3020 } 3021 switch (n->class_id()) { 3022 case Node::Class_Allocate: 3023 expand_allocate(n->as_Allocate()); 3024 break; 3025 case Node::Class_AllocateArray: 3026 expand_allocate_array(n->as_AllocateArray()); 3027 break; 3028 case Node::Class_Lock: 3029 expand_lock_node(n->as_Lock()); 3030 break; 3031 case Node::Class_Unlock: 3032 expand_unlock_node(n->as_Unlock()); 3033 break; 3034 case Node::Class_CallStaticJava: 3035 expand_mh_intrinsic_return(n->as_CallStaticJava()); 3036 C->remove_macro_node(n); 3037 break; 3038 default: 3039 assert(false, "unknown node type in macro list"); 3040 } 3041 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 3042 if (C->failing()) return true; 3043 } 3044 3045 _igvn.set_delay_transform(false); 3046 _igvn.optimize(); 3047 if (C->failing()) return true; 3048 return false; 3049 }