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 int instance_id = adr_t->instance_id(); 575 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 576 577 int alias_idx = C->get_alias_index(adr_t); 578 int offset = adr_t->offset(); 579 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 580 Node *alloc_ctrl = alloc->in(TypeFunc::Control); 581 Node *alloc_mem = alloc->in(TypeFunc::Memory); 582 Arena *a = Thread::current()->resource_area(); 583 VectorSet visited(a); 584 585 bool done = sfpt_mem == alloc_mem; 586 Node *mem = sfpt_mem; 587 while (!done) { 588 if (visited.test_set(mem->_idx)) { 589 return NULL; // found a loop, give up 590 } 591 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 592 if (mem == start_mem || mem == alloc_mem) { 593 done = true; // hit a sentinel, return appropriate 0 value 594 } else if (mem->is_Initialize()) { 595 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 596 if (mem == NULL) { 597 done = true; // Something went wrong. 598 } else if (mem->is_Store()) { 599 const TypePtr* atype = mem->as_Store()->adr_type(); 600 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 601 done = true; 602 } 603 } else if (mem->is_Store()) { 604 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 605 assert(atype != NULL, "address type must be oopptr"); 606 assert(C->get_alias_index(atype) == alias_idx && 607 atype->is_known_instance_field() && atype->offset() == offset && 608 atype->instance_id() == instance_id, "store is correct memory slice"); 609 done = true; 610 } else if (mem->is_Phi()) { 611 // try to find a phi's unique input 612 Node *unique_input = NULL; 613 Node *top = C->top(); 614 for (uint i = 1; i < mem->req(); i++) { 615 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 616 if (n == NULL || n == top || n == mem) { 617 continue; 618 } else if (unique_input == NULL) { 619 unique_input = n; 620 } else if (unique_input != n) { 621 unique_input = top; 622 break; 623 } 624 } 625 if (unique_input != NULL && unique_input != top) { 626 mem = unique_input; 627 } else { 628 done = true; 629 } 630 } else if (mem->is_ArrayCopy()) { 631 done = true; 632 } else { 633 assert(false, "unexpected node"); 634 } 635 } 636 if (mem != NULL) { 637 if (mem == start_mem || mem == alloc_mem) { 638 // hit a sentinel, return appropriate 0 value 639 return _igvn.zerocon(ft); 640 } else if (mem->is_Store()) { 641 return mem->in(MemNode::ValueIn); 642 } else if (mem->is_Phi()) { 643 // attempt to produce a Phi reflecting the values on the input paths of the Phi 644 Node_Stack value_phis(a, 8); 645 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 646 if (phi != NULL) { 647 return phi; 648 } else { 649 // Kill all new Phis 650 while(value_phis.is_nonempty()) { 651 Node* n = value_phis.node(); 652 _igvn.replace_node(n, C->top()); 653 value_phis.pop(); 654 } 655 } 656 } else if (mem->is_ArrayCopy()) { 657 Node* ctl = mem->in(0); 658 Node* m = mem->in(TypeFunc::Memory); 659 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) { 660 // pin the loads in the uncommon trap path 661 ctl = sfpt_ctl; 662 m = sfpt_mem; 663 } 664 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc); 665 } 666 } 667 // Something went wrong. 668 return NULL; 669 } 670 671 // Search the last value stored into the value type's fields. 672 Node* PhaseMacroExpand::value_type_from_mem(Node* mem, Node* ctl, ciValueKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) { 673 // Subtract the offset of the first field to account for the missing oop header 674 offset -= vk->first_field_offset(); 675 // Create a new ValueTypeNode and retrieve the field values from memory 676 ValueTypeNode* vt = ValueTypeNode::make(_igvn, vk)->as_ValueType(); 677 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) { 678 ciType* field_type = vt->field_type(i); 679 int field_offset = offset + vt->field_offset(i); 680 // Each value type field has its own memory slice 681 adr_type = adr_type->with_field_offset(field_offset); 682 Node* value = NULL; 683 if (field_type->is_valuetype() && vt->field_is_flattened(i)) { 684 value = value_type_from_mem(mem, ctl, field_type->as_value_klass(), adr_type, field_offset, alloc); 685 } else { 686 const Type* ft = Type::get_const_type(field_type); 687 BasicType bt = field_type->basic_type(); 688 if (UseCompressedOops && !is_java_primitive(bt)) { 689 ft = ft->make_narrowoop(); 690 bt = T_NARROWOOP; 691 } 692 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc); 693 if (ft->isa_narrowoop()) { 694 assert(UseCompressedOops, "unexpected narrow oop"); 695 value = transform_later(new DecodeNNode(value, value->get_ptr_type())); 696 } 697 } 698 vt->set_field_value(i, value); 699 } 700 return vt; 701 } 702 703 // Check the possibility of scalar replacement. 704 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 705 // Scan the uses of the allocation to check for anything that would 706 // prevent us from eliminating it. 707 NOT_PRODUCT( const char* fail_eliminate = NULL; ) 708 DEBUG_ONLY( Node* disq_node = NULL; ) 709 bool can_eliminate = true; 710 711 Node* res = alloc->result_cast(); 712 const TypeOopPtr* res_type = NULL; 713 if (res == NULL) { 714 // All users were eliminated. 715 } else if (!res->is_CheckCastPP()) { 716 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 717 can_eliminate = false; 718 } else { 719 res_type = _igvn.type(res)->isa_oopptr(); 720 if (res_type == NULL) { 721 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 722 can_eliminate = false; 723 } else if (res_type->isa_aryptr()) { 724 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 725 if (length < 0) { 726 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 727 can_eliminate = false; 728 } 729 } 730 } 731 732 if (can_eliminate && res != NULL) { 733 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 734 j < jmax && can_eliminate; j++) { 735 Node* use = res->fast_out(j); 736 737 if (use->is_AddP()) { 738 const TypePtr* addp_type = _igvn.type(use)->is_ptr(); 739 int offset = addp_type->offset(); 740 741 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 742 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) 743 can_eliminate = false; 744 break; 745 } 746 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 747 k < kmax && can_eliminate; k++) { 748 Node* n = use->fast_out(k); 749 if (!n->is_Store() && n->Opcode() != Op_CastP2X && 750 !(n->is_ArrayCopy() && 751 n->as_ArrayCopy()->is_clonebasic() && 752 n->in(ArrayCopyNode::Dest) == use)) { 753 DEBUG_ONLY(disq_node = n;) 754 if (n->is_Load() || n->is_LoadStore()) { 755 NOT_PRODUCT(fail_eliminate = "Field load";) 756 } else { 757 NOT_PRODUCT(fail_eliminate = "Not store field reference";) 758 } 759 can_eliminate = false; 760 } 761 } 762 } else if (use->is_ArrayCopy() && 763 (use->as_ArrayCopy()->is_arraycopy_validated() || 764 use->as_ArrayCopy()->is_copyof_validated() || 765 use->as_ArrayCopy()->is_copyofrange_validated()) && 766 use->in(ArrayCopyNode::Dest) == res) { 767 // ok to eliminate 768 } else if (use->is_SafePoint()) { 769 SafePointNode* sfpt = use->as_SafePoint(); 770 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 771 // Object is passed as argument. 772 DEBUG_ONLY(disq_node = use;) 773 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 774 can_eliminate = false; 775 } 776 Node* sfptMem = sfpt->memory(); 777 if (sfptMem == NULL || sfptMem->is_top()) { 778 DEBUG_ONLY(disq_node = use;) 779 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) 780 can_eliminate = false; 781 } else { 782 safepoints.append_if_missing(sfpt); 783 } 784 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 785 if (use->is_Phi()) { 786 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 787 NOT_PRODUCT(fail_eliminate = "Object is return value";) 788 } else { 789 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 790 } 791 DEBUG_ONLY(disq_node = use;) 792 } else { 793 if (use->Opcode() == Op_Return) { 794 NOT_PRODUCT(fail_eliminate = "Object is return value";) 795 } else { 796 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 797 } 798 DEBUG_ONLY(disq_node = use;) 799 } 800 can_eliminate = false; 801 } 802 } 803 } 804 805 #ifndef PRODUCT 806 if (PrintEliminateAllocations) { 807 if (can_eliminate) { 808 tty->print("Scalar "); 809 if (res == NULL) 810 alloc->dump(); 811 else 812 res->dump(); 813 } else if (alloc->_is_scalar_replaceable) { 814 tty->print("NotScalar (%s)", fail_eliminate); 815 if (res == NULL) 816 alloc->dump(); 817 else 818 res->dump(); 819 #ifdef ASSERT 820 if (disq_node != NULL) { 821 tty->print(" >>>> "); 822 disq_node->dump(); 823 } 824 #endif /*ASSERT*/ 825 } 826 } 827 #endif 828 return can_eliminate; 829 } 830 831 // Do scalar replacement. 832 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 833 GrowableArray <SafePointNode *> safepoints_done; 834 835 ciKlass* klass = NULL; 836 ciInstanceKlass* iklass = NULL; 837 int nfields = 0; 838 int array_base = 0; 839 int element_size = 0; 840 BasicType basic_elem_type = T_ILLEGAL; 841 ciType* elem_type = NULL; 842 843 Node* res = alloc->result_cast(); 844 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result"); 845 const TypeOopPtr* res_type = NULL; 846 if (res != NULL) { // Could be NULL when there are no users 847 res_type = _igvn.type(res)->isa_oopptr(); 848 } 849 850 if (res != NULL) { 851 klass = res_type->klass(); 852 // Value types are only allocated on demand 853 if (res_type->isa_instptr() || res_type->isa_valuetypeptr()) { 854 // find the fields of the class which will be needed for safepoint debug information 855 assert(klass->is_instance_klass(), "must be an instance klass."); 856 iklass = klass->as_instance_klass(); 857 nfields = iklass->nof_nonstatic_fields(); 858 } else { 859 // find the array's elements which will be needed for safepoint debug information 860 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 861 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); 862 elem_type = klass->as_array_klass()->element_type(); 863 basic_elem_type = elem_type->basic_type(); 864 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 865 element_size = type2aelembytes(basic_elem_type); 866 } 867 } 868 // 869 // Process the safepoint uses 870 // 871 Unique_Node_List value_worklist; 872 while (safepoints.length() > 0) { 873 SafePointNode* sfpt = safepoints.pop(); 874 Node* mem = sfpt->memory(); 875 Node* ctl = sfpt->control(); 876 assert(sfpt->jvms() != NULL, "missed JVMS"); 877 // Fields of scalar objs are referenced only at the end 878 // of regular debuginfo at the last (youngest) JVMS. 879 // Record relative start index. 880 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff()); 881 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, 882 #ifdef ASSERT 883 alloc, 884 #endif 885 first_ind, nfields); 886 sobj->init_req(0, C->root()); 887 transform_later(sobj); 888 889 // Scan object's fields adding an input to the safepoint for each field. 890 for (int j = 0; j < nfields; j++) { 891 intptr_t offset; 892 ciField* field = NULL; 893 if (iklass != NULL) { 894 field = iklass->nonstatic_field_at(j); 895 offset = field->offset(); 896 elem_type = field->type(); 897 basic_elem_type = field->layout_type(); 898 assert(!field->is_flattened(), "flattened value type fields should not have safepoint uses"); 899 } else { 900 offset = array_base + j * (intptr_t)element_size; 901 } 902 903 const Type *field_type; 904 // The next code is taken from Parse::do_get_xxx(). 905 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { 906 if (!elem_type->is_loaded()) { 907 field_type = TypeInstPtr::BOTTOM; 908 } else if (field != NULL && field->is_static_constant()) { 909 // This can happen if the constant oop is non-perm. 910 ciObject* con = field->constant_value().as_object(); 911 // Do not "join" in the previous type; it doesn't add value, 912 // and may yield a vacuous result if the field is of interface type. 913 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 914 assert(field_type != NULL, "field singleton type must be consistent"); 915 } else { 916 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 917 } 918 if (UseCompressedOops) { 919 field_type = field_type->make_narrowoop(); 920 basic_elem_type = T_NARROWOOP; 921 } 922 } else { 923 field_type = Type::get_const_basic_type(basic_elem_type); 924 } 925 926 Node* field_val = NULL; 927 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 928 if (klass->is_value_array_klass()) { 929 ciValueKlass* vk = elem_type->as_value_klass(); 930 assert(vk->flatten_array(), "must be flattened"); 931 field_val = value_type_from_mem(mem, ctl, vk, field_addr_type->isa_aryptr(), 0, alloc); 932 } else { 933 field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc); 934 } 935 if (field_val == NULL) { 936 // We weren't able to find a value for this field, 937 // give up on eliminating this allocation. 938 939 // Remove any extra entries we added to the safepoint. 940 uint last = sfpt->req() - 1; 941 for (int k = 0; k < j; k++) { 942 sfpt->del_req(last--); 943 } 944 _igvn._worklist.push(sfpt); 945 // rollback processed safepoints 946 while (safepoints_done.length() > 0) { 947 SafePointNode* sfpt_done = safepoints_done.pop(); 948 // remove any extra entries we added to the safepoint 949 last = sfpt_done->req() - 1; 950 for (int k = 0; k < nfields; k++) { 951 sfpt_done->del_req(last--); 952 } 953 JVMState *jvms = sfpt_done->jvms(); 954 jvms->set_endoff(sfpt_done->req()); 955 // Now make a pass over the debug information replacing any references 956 // to SafePointScalarObjectNode with the allocated object. 957 int start = jvms->debug_start(); 958 int end = jvms->debug_end(); 959 for (int i = start; i < end; i++) { 960 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 961 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 962 if (scobj->first_index(jvms) == sfpt_done->req() && 963 scobj->n_fields() == (uint)nfields) { 964 assert(scobj->alloc() == alloc, "sanity"); 965 sfpt_done->set_req(i, res); 966 } 967 } 968 } 969 _igvn._worklist.push(sfpt_done); 970 } 971 #ifndef PRODUCT 972 if (PrintEliminateAllocations) { 973 if (field != NULL) { 974 tty->print("=== At SafePoint node %d can't find value of Field: ", 975 sfpt->_idx); 976 field->print(); 977 int field_idx = C->get_alias_index(field_addr_type); 978 tty->print(" (alias_idx=%d)", field_idx); 979 } else { // Array's element 980 tty->print("=== At SafePoint node %d can't find value of array element [%d]", 981 sfpt->_idx, j); 982 } 983 tty->print(", which prevents elimination of: "); 984 if (res == NULL) 985 alloc->dump(); 986 else 987 res->dump(); 988 } 989 #endif 990 return false; 991 } 992 if (UseCompressedOops && field_type->isa_narrowoop()) { 993 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 994 // to be able scalar replace the allocation. 995 if (field_val->is_EncodeP()) { 996 field_val = field_val->in(1); 997 } else { 998 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type())); 999 } 1000 } else if (field_val->is_ValueType()) { 1001 // Keep track of value types to scalarize them later 1002 value_worklist.push(field_val); 1003 } 1004 sfpt->add_req(field_val); 1005 } 1006 JVMState *jvms = sfpt->jvms(); 1007 jvms->set_endoff(sfpt->req()); 1008 // Now make a pass over the debug information replacing any references 1009 // to the allocated object with "sobj" 1010 int start = jvms->debug_start(); 1011 int end = jvms->debug_end(); 1012 sfpt->replace_edges_in_range(res, sobj, start, end); 1013 _igvn._worklist.push(sfpt); 1014 safepoints_done.append_if_missing(sfpt); // keep it for rollback 1015 } 1016 // Scalarize value types that were added to the safepoint 1017 for (uint i = 0; i < value_worklist.size(); ++i) { 1018 Node* vt = value_worklist.at(i); 1019 vt->as_ValueType()->make_scalar_in_safepoints(C->root(), &_igvn); 1020 } 1021 return true; 1022 } 1023 1024 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) { 1025 Node* ctl_proj = n->proj_out(TypeFunc::Control); 1026 Node* mem_proj = n->proj_out(TypeFunc::Memory); 1027 if (ctl_proj != NULL) { 1028 igvn.replace_node(ctl_proj, n->in(0)); 1029 } 1030 if (mem_proj != NULL) { 1031 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory)); 1032 } 1033 } 1034 1035 // Process users of eliminated allocation. 1036 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) { 1037 Node* res = alloc->result_cast(); 1038 if (res != NULL) { 1039 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 1040 Node *use = res->last_out(j); 1041 uint oc1 = res->outcnt(); 1042 1043 if (use->is_AddP()) { 1044 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 1045 Node *n = use->last_out(k); 1046 uint oc2 = use->outcnt(); 1047 if (n->is_Store()) { 1048 #ifdef ASSERT 1049 // Verify that there is no dependent MemBarVolatile nodes, 1050 // they should be removed during IGVN, see MemBarNode::Ideal(). 1051 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); 1052 p < pmax; p++) { 1053 Node* mb = n->fast_out(p); 1054 assert(mb->is_Initialize() || !mb->is_MemBar() || 1055 mb->req() <= MemBarNode::Precedent || 1056 mb->in(MemBarNode::Precedent) != n, 1057 "MemBarVolatile should be eliminated for non-escaping object"); 1058 } 1059 #endif 1060 _igvn.replace_node(n, n->in(MemNode::Memory)); 1061 } else if (n->is_ArrayCopy()) { 1062 // Disconnect ArrayCopy node 1063 ArrayCopyNode* ac = n->as_ArrayCopy(); 1064 assert(ac->is_clonebasic(), "unexpected array copy kind"); 1065 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out(); 1066 disconnect_projections(ac, _igvn); 1067 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation"); 1068 Node* membar_before = alloc->in(0)->in(0); 1069 disconnect_projections(membar_before->as_MemBar(), _igvn); 1070 if (membar_after->is_MemBar()) { 1071 disconnect_projections(membar_after->as_MemBar(), _igvn); 1072 } 1073 } else { 1074 eliminate_card_mark(n); 1075 } 1076 k -= (oc2 - use->outcnt()); 1077 } 1078 } else if (use->is_ArrayCopy()) { 1079 // Disconnect ArrayCopy node 1080 ArrayCopyNode* ac = use->as_ArrayCopy(); 1081 assert(ac->is_arraycopy_validated() || 1082 ac->is_copyof_validated() || 1083 ac->is_copyofrange_validated(), "unsupported"); 1084 CallProjections callprojs; 1085 ac->extract_projections(&callprojs, true); 1086 1087 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O)); 1088 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory)); 1089 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control)); 1090 1091 // Set control to top. IGVN will remove the remaining projections 1092 ac->set_req(0, top()); 1093 ac->replace_edge(res, top()); 1094 1095 // Disconnect src right away: it can help find new 1096 // opportunities for allocation elimination 1097 Node* src = ac->in(ArrayCopyNode::Src); 1098 ac->replace_edge(src, top()); 1099 // src can be top at this point if src and dest of the 1100 // arraycopy were the same 1101 if (src->outcnt() == 0 && !src->is_top()) { 1102 _igvn.remove_dead_node(src); 1103 } 1104 1105 _igvn._worklist.push(ac); 1106 } else { 1107 eliminate_card_mark(use); 1108 } 1109 j -= (oc1 - res->outcnt()); 1110 } 1111 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 1112 _igvn.remove_dead_node(res); 1113 } 1114 1115 // 1116 // Process other users of allocation's projections 1117 // 1118 if (_resproj != NULL && _resproj->outcnt() != 0) { 1119 // First disconnect stores captured by Initialize node. 1120 // If Initialize node is eliminated first in the following code, 1121 // it will kill such stores and DUIterator_Last will assert. 1122 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) { 1123 Node *use = _resproj->fast_out(j); 1124 if (use->is_AddP()) { 1125 // raw memory addresses used only by the initialization 1126 _igvn.replace_node(use, C->top()); 1127 --j; --jmax; 1128 } 1129 } 1130 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { 1131 Node *use = _resproj->last_out(j); 1132 uint oc1 = _resproj->outcnt(); 1133 if (use->is_Initialize()) { 1134 // Eliminate Initialize node. 1135 InitializeNode *init = use->as_Initialize(); 1136 assert(init->outcnt() <= 2, "only a control and memory projection expected"); 1137 Node *ctrl_proj = init->proj_out(TypeFunc::Control); 1138 if (ctrl_proj != NULL) { 1139 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection"); 1140 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj); 1141 } 1142 Node *mem_proj = init->proj_out(TypeFunc::Memory); 1143 if (mem_proj != NULL) { 1144 Node *mem = init->in(TypeFunc::Memory); 1145 #ifdef ASSERT 1146 if (mem->is_MergeMem()) { 1147 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection"); 1148 } else { 1149 assert(mem == _memproj_fallthrough, "allocation memory projection"); 1150 } 1151 #endif 1152 _igvn.replace_node(mem_proj, mem); 1153 } 1154 } else { 1155 assert(false, "only Initialize or AddP expected"); 1156 } 1157 j -= (oc1 - _resproj->outcnt()); 1158 } 1159 } 1160 if (_fallthroughcatchproj != NULL) { 1161 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); 1162 } 1163 if (_memproj_fallthrough != NULL) { 1164 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); 1165 } 1166 if (_memproj_catchall != NULL) { 1167 _igvn.replace_node(_memproj_catchall, C->top()); 1168 } 1169 if (_ioproj_fallthrough != NULL) { 1170 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); 1171 } 1172 if (_ioproj_catchall != NULL) { 1173 _igvn.replace_node(_ioproj_catchall, C->top()); 1174 } 1175 if (_catchallcatchproj != NULL) { 1176 _igvn.replace_node(_catchallcatchproj, C->top()); 1177 } 1178 } 1179 1180 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 1181 // Don't do scalar replacement if the frame can be popped by JVMTI: 1182 // if reallocation fails during deoptimization we'll pop all 1183 // interpreter frames for this compiled frame and that won't play 1184 // nice with JVMTI popframe. 1185 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) { 1186 return false; 1187 } 1188 Node* klass = alloc->in(AllocateNode::KlassNode); 1189 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); 1190 Node* res = alloc->result_cast(); 1191 // Eliminate boxing allocations which are not used 1192 // regardless scalar replacable status. 1193 bool boxing_alloc = C->eliminate_boxing() && 1194 tklass->klass()->is_instance_klass() && 1195 tklass->klass()->as_instance_klass()->is_box_klass(); 1196 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) { 1197 return false; 1198 } 1199 1200 extract_call_projections(alloc); 1201 1202 GrowableArray <SafePointNode *> safepoints; 1203 if (!can_eliminate_allocation(alloc, safepoints)) { 1204 return false; 1205 } 1206 1207 if (!alloc->_is_scalar_replaceable) { 1208 assert(res == NULL, "sanity"); 1209 // We can only eliminate allocation if all debug info references 1210 // are already replaced with SafePointScalarObject because 1211 // we can't search for a fields value without instance_id. 1212 if (safepoints.length() > 0) { 1213 return false; 1214 } 1215 } 1216 1217 if (!scalar_replacement(alloc, safepoints)) { 1218 return false; 1219 } 1220 1221 CompileLog* log = C->log(); 1222 if (log != NULL) { 1223 log->head("eliminate_allocation type='%d'", 1224 log->identify(tklass->klass())); 1225 JVMState* p = alloc->jvms(); 1226 while (p != NULL) { 1227 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1228 p = p->caller(); 1229 } 1230 log->tail("eliminate_allocation"); 1231 } 1232 1233 process_users_of_allocation(alloc); 1234 1235 #ifndef PRODUCT 1236 if (PrintEliminateAllocations) { 1237 if (alloc->is_AllocateArray()) 1238 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1239 else 1240 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1241 } 1242 #endif 1243 1244 return true; 1245 } 1246 1247 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) { 1248 // EA should remove all uses of non-escaping boxing node. 1249 if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) { 1250 return false; 1251 } 1252 1253 assert(boxing->result_cast() == NULL, "unexpected boxing node result"); 1254 1255 extract_call_projections(boxing); 1256 1257 const TypeTuple* r = boxing->tf()->range_sig(); 1258 assert(r->cnt() > TypeFunc::Parms, "sanity"); 1259 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr(); 1260 assert(t != NULL, "sanity"); 1261 1262 CompileLog* log = C->log(); 1263 if (log != NULL) { 1264 log->head("eliminate_boxing type='%d'", 1265 log->identify(t->klass())); 1266 JVMState* p = boxing->jvms(); 1267 while (p != NULL) { 1268 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1269 p = p->caller(); 1270 } 1271 log->tail("eliminate_boxing"); 1272 } 1273 1274 process_users_of_allocation(boxing); 1275 1276 #ifndef PRODUCT 1277 if (PrintEliminateAllocations) { 1278 tty->print("++++ Eliminated: %d ", boxing->_idx); 1279 boxing->method()->print_short_name(tty); 1280 tty->cr(); 1281 } 1282 #endif 1283 1284 return true; 1285 } 1286 1287 //---------------------------set_eden_pointers------------------------- 1288 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { 1289 if (UseTLAB) { // Private allocation: load from TLS 1290 Node* thread = transform_later(new ThreadLocalNode()); 1291 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); 1292 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); 1293 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); 1294 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); 1295 } else { // Shared allocation: load from globals 1296 CollectedHeap* ch = Universe::heap(); 1297 address top_adr = (address)ch->top_addr(); 1298 address end_adr = (address)ch->end_addr(); 1299 eden_top_adr = makecon(TypeRawPtr::make(top_adr)); 1300 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); 1301 } 1302 } 1303 1304 1305 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 1306 Node* adr = basic_plus_adr(base, offset); 1307 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 1308 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered); 1309 transform_later(value); 1310 return value; 1311 } 1312 1313 1314 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 1315 Node* adr = basic_plus_adr(base, offset); 1316 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered); 1317 transform_later(mem); 1318 return mem; 1319 } 1320 1321 //============================================================================= 1322 // 1323 // A L L O C A T I O N 1324 // 1325 // Allocation attempts to be fast in the case of frequent small objects. 1326 // It breaks down like this: 1327 // 1328 // 1) Size in doublewords is computed. This is a constant for objects and 1329 // variable for most arrays. Doubleword units are used to avoid size 1330 // overflow of huge doubleword arrays. We need doublewords in the end for 1331 // rounding. 1332 // 1333 // 2) Size is checked for being 'too large'. Too-large allocations will go 1334 // the slow path into the VM. The slow path can throw any required 1335 // exceptions, and does all the special checks for very large arrays. The 1336 // size test can constant-fold away for objects. For objects with 1337 // finalizers it constant-folds the otherway: you always go slow with 1338 // finalizers. 1339 // 1340 // 3) If NOT using TLABs, this is the contended loop-back point. 1341 // Load-Locked the heap top. If using TLABs normal-load the heap top. 1342 // 1343 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 1344 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 1345 // "size*8" we always enter the VM, where "largish" is a constant picked small 1346 // enough that there's always space between the eden max and 4Gig (old space is 1347 // there so it's quite large) and large enough that the cost of entering the VM 1348 // is dwarfed by the cost to initialize the space. 1349 // 1350 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 1351 // down. If contended, repeat at step 3. If using TLABs normal-store 1352 // adjusted heap top back down; there is no contention. 1353 // 1354 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 1355 // fields. 1356 // 1357 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 1358 // oop flavor. 1359 // 1360 //============================================================================= 1361 // FastAllocateSizeLimit value is in DOUBLEWORDS. 1362 // Allocations bigger than this always go the slow route. 1363 // This value must be small enough that allocation attempts that need to 1364 // trigger exceptions go the slow route. Also, it must be small enough so 1365 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 1366 //=============================================================================j// 1367 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 1368 // The allocator will coalesce int->oop copies away. See comment in 1369 // coalesce.cpp about how this works. It depends critically on the exact 1370 // code shape produced here, so if you are changing this code shape 1371 // make sure the GC info for the heap-top is correct in and around the 1372 // slow-path call. 1373 // 1374 1375 void PhaseMacroExpand::expand_allocate_common( 1376 AllocateNode* alloc, // allocation node to be expanded 1377 Node* length, // array length for an array allocation 1378 const TypeFunc* slow_call_type, // Type of slow call 1379 address slow_call_address // Address of slow call 1380 ) 1381 { 1382 1383 Node* ctrl = alloc->in(TypeFunc::Control); 1384 Node* mem = alloc->in(TypeFunc::Memory); 1385 Node* i_o = alloc->in(TypeFunc::I_O); 1386 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 1387 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1388 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 1389 1390 assert(ctrl != NULL, "must have control"); 1391 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 1392 // they will not be used if "always_slow" is set 1393 enum { slow_result_path = 1, fast_result_path = 2 }; 1394 Node *result_region = NULL; 1395 Node *result_phi_rawmem = NULL; 1396 Node *result_phi_rawoop = NULL; 1397 Node *result_phi_i_o = NULL; 1398 1399 // The initial slow comparison is a size check, the comparison 1400 // we want to do is a BoolTest::gt 1401 bool always_slow = false; 1402 int tv = _igvn.find_int_con(initial_slow_test, -1); 1403 if (tv >= 0) { 1404 always_slow = (tv == 1); 1405 initial_slow_test = NULL; 1406 } else { 1407 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 1408 } 1409 1410 if (C->env()->dtrace_alloc_probes() || 1411 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) { 1412 // Force slow-path allocation 1413 always_slow = true; 1414 initial_slow_test = NULL; 1415 } 1416 1417 1418 enum { too_big_or_final_path = 1, need_gc_path = 2 }; 1419 Node *slow_region = NULL; 1420 Node *toobig_false = ctrl; 1421 1422 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); 1423 // generate the initial test if necessary 1424 if (initial_slow_test != NULL ) { 1425 slow_region = new RegionNode(3); 1426 1427 // Now make the initial failure test. Usually a too-big test but 1428 // might be a TRUE for finalizers or a fancy class check for 1429 // newInstance0. 1430 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1431 transform_later(toobig_iff); 1432 // Plug the failing-too-big test into the slow-path region 1433 Node *toobig_true = new IfTrueNode( toobig_iff ); 1434 transform_later(toobig_true); 1435 slow_region ->init_req( too_big_or_final_path, toobig_true ); 1436 toobig_false = new IfFalseNode( toobig_iff ); 1437 transform_later(toobig_false); 1438 } else { // No initial test, just fall into next case 1439 toobig_false = ctrl; 1440 debug_only(slow_region = NodeSentinel); 1441 } 1442 1443 Node *slow_mem = mem; // save the current memory state for slow path 1444 // generate the fast allocation code unless we know that the initial test will always go slow 1445 if (!always_slow) { 1446 // Fast path modifies only raw memory. 1447 if (mem->is_MergeMem()) { 1448 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1449 } 1450 1451 Node* eden_top_adr; 1452 Node* eden_end_adr; 1453 1454 set_eden_pointers(eden_top_adr, eden_end_adr); 1455 1456 // Load Eden::end. Loop invariant and hoisted. 1457 // 1458 // Note: We set the control input on "eden_end" and "old_eden_top" when using 1459 // a TLAB to work around a bug where these values were being moved across 1460 // a safepoint. These are not oops, so they cannot be include in the oop 1461 // map, but they can be changed by a GC. The proper way to fix this would 1462 // be to set the raw memory state when generating a SafepointNode. However 1463 // this will require extensive changes to the loop optimization in order to 1464 // prevent a degradation of the optimization. 1465 // See comment in memnode.hpp, around line 227 in class LoadPNode. 1466 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 1467 1468 // allocate the Region and Phi nodes for the result 1469 result_region = new RegionNode(3); 1470 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1471 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM); 1472 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch 1473 1474 // We need a Region for the loop-back contended case. 1475 enum { fall_in_path = 1, contended_loopback_path = 2 }; 1476 Node *contended_region; 1477 Node *contended_phi_rawmem; 1478 if (UseTLAB) { 1479 contended_region = toobig_false; 1480 contended_phi_rawmem = mem; 1481 } else { 1482 contended_region = new RegionNode(3); 1483 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1484 // Now handle the passing-too-big test. We fall into the contended 1485 // loop-back merge point. 1486 contended_region ->init_req(fall_in_path, toobig_false); 1487 contended_phi_rawmem->init_req(fall_in_path, mem); 1488 transform_later(contended_region); 1489 transform_later(contended_phi_rawmem); 1490 } 1491 1492 // Load(-locked) the heap top. 1493 // See note above concerning the control input when using a TLAB 1494 Node *old_eden_top = UseTLAB 1495 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered) 1496 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire); 1497 1498 transform_later(old_eden_top); 1499 // Add to heap top to get a new heap top 1500 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes); 1501 transform_later(new_eden_top); 1502 // Check for needing a GC; compare against heap end 1503 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end); 1504 transform_later(needgc_cmp); 1505 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge); 1506 transform_later(needgc_bol); 1507 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); 1508 transform_later(needgc_iff); 1509 1510 // Plug the failing-heap-space-need-gc test into the slow-path region 1511 Node *needgc_true = new IfTrueNode(needgc_iff); 1512 transform_later(needgc_true); 1513 if (initial_slow_test) { 1514 slow_region->init_req(need_gc_path, needgc_true); 1515 // This completes all paths into the slow merge point 1516 transform_later(slow_region); 1517 } else { // No initial slow path needed! 1518 // Just fall from the need-GC path straight into the VM call. 1519 slow_region = needgc_true; 1520 } 1521 // No need for a GC. Setup for the Store-Conditional 1522 Node *needgc_false = new IfFalseNode(needgc_iff); 1523 transform_later(needgc_false); 1524 1525 // Grab regular I/O before optional prefetch may change it. 1526 // Slow-path does no I/O so just set it to the original I/O. 1527 result_phi_i_o->init_req(slow_result_path, i_o); 1528 1529 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem, 1530 old_eden_top, new_eden_top, length); 1531 1532 // Name successful fast-path variables 1533 Node* fast_oop = old_eden_top; 1534 Node* fast_oop_ctrl; 1535 Node* fast_oop_rawmem; 1536 1537 // Store (-conditional) the modified eden top back down. 1538 // StorePConditional produces flags for a test PLUS a modified raw 1539 // memory state. 1540 if (UseTLAB) { 1541 Node* store_eden_top = 1542 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1543 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered); 1544 transform_later(store_eden_top); 1545 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path 1546 fast_oop_rawmem = store_eden_top; 1547 } else { 1548 Node* store_eden_top = 1549 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1550 new_eden_top, fast_oop/*old_eden_top*/); 1551 transform_later(store_eden_top); 1552 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne); 1553 transform_later(contention_check); 1554 store_eden_top = new SCMemProjNode(store_eden_top); 1555 transform_later(store_eden_top); 1556 1557 // If not using TLABs, check to see if there was contention. 1558 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN); 1559 transform_later(contention_iff); 1560 Node *contention_true = new IfTrueNode(contention_iff); 1561 transform_later(contention_true); 1562 // If contention, loopback and try again. 1563 contended_region->init_req(contended_loopback_path, contention_true); 1564 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top); 1565 1566 // Fast-path succeeded with no contention! 1567 Node *contention_false = new IfFalseNode(contention_iff); 1568 transform_later(contention_false); 1569 fast_oop_ctrl = contention_false; 1570 1571 // Bump total allocated bytes for this thread 1572 Node* thread = new ThreadLocalNode(); 1573 transform_later(thread); 1574 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread, 1575 in_bytes(JavaThread::allocated_bytes_offset())); 1576 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1577 0, TypeLong::LONG, T_LONG); 1578 #ifdef _LP64 1579 Node* alloc_size = size_in_bytes; 1580 #else 1581 Node* alloc_size = new ConvI2LNode(size_in_bytes); 1582 transform_later(alloc_size); 1583 #endif 1584 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size); 1585 transform_later(new_alloc_bytes); 1586 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1587 0, new_alloc_bytes, T_LONG); 1588 } 1589 1590 InitializeNode* init = alloc->initialization(); 1591 fast_oop_rawmem = initialize_object(alloc, 1592 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1593 klass_node, length, size_in_bytes); 1594 1595 // If initialization is performed by an array copy, any required 1596 // MemBarStoreStore was already added. If the object does not 1597 // escape no need for a MemBarStoreStore. If the object does not 1598 // escape in its initializer and memory barrier (MemBarStoreStore or 1599 // stronger) is already added at exit of initializer, also no need 1600 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore 1601 // so that stores that initialize this object can't be reordered 1602 // with a subsequent store that makes this object accessible by 1603 // other threads. 1604 // Other threads include java threads and JVM internal threads 1605 // (for example concurrent GC threads). Current concurrent GC 1606 // implementation: CMS and G1 will not scan newly created object, 1607 // so it's safe to skip storestore barrier when allocation does 1608 // not escape. 1609 if (!alloc->does_not_escape_thread() && 1610 !alloc->is_allocation_MemBar_redundant() && 1611 (init == NULL || !init->is_complete_with_arraycopy())) { 1612 if (init == NULL || init->req() < InitializeNode::RawStores) { 1613 // No InitializeNode or no stores captured by zeroing 1614 // elimination. Simply add the MemBarStoreStore after object 1615 // initialization. 1616 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1617 transform_later(mb); 1618 1619 mb->init_req(TypeFunc::Memory, fast_oop_rawmem); 1620 mb->init_req(TypeFunc::Control, fast_oop_ctrl); 1621 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control); 1622 transform_later(fast_oop_ctrl); 1623 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory); 1624 transform_later(fast_oop_rawmem); 1625 } else { 1626 // Add the MemBarStoreStore after the InitializeNode so that 1627 // all stores performing the initialization that were moved 1628 // before the InitializeNode happen before the storestore 1629 // barrier. 1630 1631 Node* init_ctrl = init->proj_out(TypeFunc::Control); 1632 Node* init_mem = init->proj_out(TypeFunc::Memory); 1633 1634 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1635 transform_later(mb); 1636 1637 Node* ctrl = new ProjNode(init,TypeFunc::Control); 1638 transform_later(ctrl); 1639 Node* mem = new ProjNode(init,TypeFunc::Memory); 1640 transform_later(mem); 1641 1642 // The MemBarStoreStore depends on control and memory coming 1643 // from the InitializeNode 1644 mb->init_req(TypeFunc::Memory, mem); 1645 mb->init_req(TypeFunc::Control, ctrl); 1646 1647 ctrl = new ProjNode(mb,TypeFunc::Control); 1648 transform_later(ctrl); 1649 mem = new ProjNode(mb,TypeFunc::Memory); 1650 transform_later(mem); 1651 1652 // All nodes that depended on the InitializeNode for control 1653 // and memory must now depend on the MemBarNode that itself 1654 // depends on the InitializeNode 1655 if (init_ctrl != NULL) { 1656 _igvn.replace_node(init_ctrl, ctrl); 1657 } 1658 if (init_mem != NULL) { 1659 _igvn.replace_node(init_mem, mem); 1660 } 1661 } 1662 } 1663 1664 if (C->env()->dtrace_extended_probes()) { 1665 // Slow-path call 1666 int size = TypeFunc::Parms + 2; 1667 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1668 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), 1669 "dtrace_object_alloc", 1670 TypeRawPtr::BOTTOM); 1671 1672 // Get base of thread-local storage area 1673 Node* thread = new ThreadLocalNode(); 1674 transform_later(thread); 1675 1676 call->init_req(TypeFunc::Parms+0, thread); 1677 call->init_req(TypeFunc::Parms+1, fast_oop); 1678 call->init_req(TypeFunc::Control, fast_oop_ctrl); 1679 call->init_req(TypeFunc::I_O , top()); // does no i/o 1680 call->init_req(TypeFunc::Memory , fast_oop_rawmem); 1681 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1682 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1683 transform_later(call); 1684 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control); 1685 transform_later(fast_oop_ctrl); 1686 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory); 1687 transform_later(fast_oop_rawmem); 1688 } 1689 1690 // Plug in the successful fast-path into the result merge point 1691 result_region ->init_req(fast_result_path, fast_oop_ctrl); 1692 result_phi_rawoop->init_req(fast_result_path, fast_oop); 1693 result_phi_i_o ->init_req(fast_result_path, i_o); 1694 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); 1695 } else { 1696 slow_region = ctrl; 1697 result_phi_i_o = i_o; // Rename it to use in the following code. 1698 } 1699 1700 // Generate slow-path call 1701 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address, 1702 OptoRuntime::stub_name(slow_call_address), 1703 alloc->jvms()->bci(), 1704 TypePtr::BOTTOM); 1705 call->init_req( TypeFunc::Control, slow_region ); 1706 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1707 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs 1708 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1709 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1710 1711 call->init_req(TypeFunc::Parms+0, klass_node); 1712 if (length != NULL) { 1713 call->init_req(TypeFunc::Parms+1, length); 1714 } 1715 1716 // Copy debug information and adjust JVMState information, then replace 1717 // allocate node with the call 1718 copy_call_debug_info((CallNode *) alloc, call); 1719 if (!always_slow) { 1720 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1721 } else { 1722 // Hook i_o projection to avoid its elimination during allocation 1723 // replacement (when only a slow call is generated). 1724 call->set_req(TypeFunc::I_O, result_phi_i_o); 1725 } 1726 _igvn.replace_node(alloc, call); 1727 transform_later(call); 1728 1729 // Identify the output projections from the allocate node and 1730 // adjust any references to them. 1731 // The control and io projections look like: 1732 // 1733 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1734 // Allocate Catch 1735 // ^---Proj(io) <-------+ ^---CatchProj(io) 1736 // 1737 // We are interested in the CatchProj nodes. 1738 // 1739 extract_call_projections(call); 1740 1741 // An allocate node has separate memory projections for the uses on 1742 // the control and i_o paths. Replace the control memory projection with 1743 // result_phi_rawmem (unless we are only generating a slow call when 1744 // both memory projections are combined) 1745 if (!always_slow && _memproj_fallthrough != NULL) { 1746 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { 1747 Node *use = _memproj_fallthrough->fast_out(i); 1748 _igvn.rehash_node_delayed(use); 1749 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); 1750 // back up iterator 1751 --i; 1752 } 1753 } 1754 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete 1755 // _memproj_catchall so we end up with a call that has only 1 memory projection. 1756 if (_memproj_catchall != NULL ) { 1757 if (_memproj_fallthrough == NULL) { 1758 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory); 1759 transform_later(_memproj_fallthrough); 1760 } 1761 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { 1762 Node *use = _memproj_catchall->fast_out(i); 1763 _igvn.rehash_node_delayed(use); 1764 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); 1765 // back up iterator 1766 --i; 1767 } 1768 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted"); 1769 _igvn.remove_dead_node(_memproj_catchall); 1770 } 1771 1772 // An allocate node has separate i_o projections for the uses on the control 1773 // and i_o paths. Always replace the control i_o projection with result i_o 1774 // otherwise incoming i_o become dead when only a slow call is generated 1775 // (it is different from memory projections where both projections are 1776 // combined in such case). 1777 if (_ioproj_fallthrough != NULL) { 1778 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { 1779 Node *use = _ioproj_fallthrough->fast_out(i); 1780 _igvn.rehash_node_delayed(use); 1781 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); 1782 // back up iterator 1783 --i; 1784 } 1785 } 1786 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete 1787 // _ioproj_catchall so we end up with a call that has only 1 i_o projection. 1788 if (_ioproj_catchall != NULL ) { 1789 if (_ioproj_fallthrough == NULL) { 1790 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O); 1791 transform_later(_ioproj_fallthrough); 1792 } 1793 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { 1794 Node *use = _ioproj_catchall->fast_out(i); 1795 _igvn.rehash_node_delayed(use); 1796 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); 1797 // back up iterator 1798 --i; 1799 } 1800 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted"); 1801 _igvn.remove_dead_node(_ioproj_catchall); 1802 } 1803 1804 // if we generated only a slow call, we are done 1805 if (always_slow) { 1806 // Now we can unhook i_o. 1807 if (result_phi_i_o->outcnt() > 1) { 1808 call->set_req(TypeFunc::I_O, top()); 1809 } else { 1810 assert(result_phi_i_o->unique_ctrl_out() == call, ""); 1811 // Case of new array with negative size known during compilation. 1812 // AllocateArrayNode::Ideal() optimization disconnect unreachable 1813 // following code since call to runtime will throw exception. 1814 // As result there will be no users of i_o after the call. 1815 // Leave i_o attached to this call to avoid problems in preceding graph. 1816 } 1817 return; 1818 } 1819 1820 1821 if (_fallthroughcatchproj != NULL) { 1822 ctrl = _fallthroughcatchproj->clone(); 1823 transform_later(ctrl); 1824 _igvn.replace_node(_fallthroughcatchproj, result_region); 1825 } else { 1826 ctrl = top(); 1827 } 1828 Node *slow_result; 1829 if (_resproj == NULL) { 1830 // no uses of the allocation result 1831 slow_result = top(); 1832 } else { 1833 slow_result = _resproj->clone(); 1834 transform_later(slow_result); 1835 _igvn.replace_node(_resproj, result_phi_rawoop); 1836 } 1837 1838 // Plug slow-path into result merge point 1839 result_region ->init_req( slow_result_path, ctrl ); 1840 result_phi_rawoop->init_req( slow_result_path, slow_result); 1841 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); 1842 transform_later(result_region); 1843 transform_later(result_phi_rawoop); 1844 transform_later(result_phi_rawmem); 1845 transform_later(result_phi_i_o); 1846 // This completes all paths into the result merge point 1847 } 1848 1849 1850 // Helper for PhaseMacroExpand::expand_allocate_common. 1851 // Initializes the newly-allocated storage. 1852 Node* 1853 PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1854 Node* control, Node* rawmem, Node* object, 1855 Node* klass_node, Node* length, 1856 Node* size_in_bytes) { 1857 InitializeNode* init = alloc->initialization(); 1858 // Store the klass & mark bits 1859 Node* mark_node = NULL; 1860 // For now only enable fast locking for non-array types 1861 if (UseBiasedLocking && (length == NULL)) { 1862 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 1863 } else { 1864 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); 1865 } 1866 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 1867 1868 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 1869 int header_size = alloc->minimum_header_size(); // conservatively small 1870 1871 // Array length 1872 if (length != NULL) { // Arrays need length field 1873 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1874 // conservatively small header size: 1875 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1876 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1877 if (k->is_array_klass()) // we know the exact header size in most cases: 1878 header_size = Klass::layout_helper_header_size(k->layout_helper()); 1879 } 1880 1881 // Clear the object body, if necessary. 1882 if (init == NULL) { 1883 // The init has somehow disappeared; be cautious and clear everything. 1884 // 1885 // This can happen if a node is allocated but an uncommon trap occurs 1886 // immediately. In this case, the Initialize gets associated with the 1887 // trap, and may be placed in a different (outer) loop, if the Allocate 1888 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1889 // there can be two Allocates to one Initialize. The answer in all these 1890 // edge cases is safety first. It is always safe to clear immediately 1891 // within an Allocate, and then (maybe or maybe not) clear some more later. 1892 if (!(UseTLAB && ZeroTLAB)) { 1893 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1894 header_size, size_in_bytes, 1895 &_igvn); 1896 } 1897 } else { 1898 if (!init->is_complete()) { 1899 // Try to win by zeroing only what the init does not store. 1900 // We can also try to do some peephole optimizations, 1901 // such as combining some adjacent subword stores. 1902 rawmem = init->complete_stores(control, rawmem, object, 1903 header_size, size_in_bytes, &_igvn); 1904 } 1905 // We have no more use for this link, since the AllocateNode goes away: 1906 init->set_req(InitializeNode::RawAddress, top()); 1907 // (If we keep the link, it just confuses the register allocator, 1908 // who thinks he sees a real use of the address by the membar.) 1909 } 1910 1911 return rawmem; 1912 } 1913 1914 // Generate prefetch instructions for next allocations. 1915 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1916 Node*& contended_phi_rawmem, 1917 Node* old_eden_top, Node* new_eden_top, 1918 Node* length) { 1919 enum { fall_in_path = 1, pf_path = 2 }; 1920 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1921 // Generate prefetch allocation with watermark check. 1922 // As an allocation hits the watermark, we will prefetch starting 1923 // at a "distance" away from watermark. 1924 1925 Node *pf_region = new RegionNode(3); 1926 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 1927 TypeRawPtr::BOTTOM ); 1928 // I/O is used for Prefetch 1929 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO ); 1930 1931 Node *thread = new ThreadLocalNode(); 1932 transform_later(thread); 1933 1934 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread, 1935 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1936 transform_later(eden_pf_adr); 1937 1938 Node *old_pf_wm = new LoadPNode(needgc_false, 1939 contended_phi_rawmem, eden_pf_adr, 1940 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, 1941 MemNode::unordered); 1942 transform_later(old_pf_wm); 1943 1944 // check against new_eden_top 1945 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm ); 1946 transform_later(need_pf_cmp); 1947 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge ); 1948 transform_later(need_pf_bol); 1949 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol, 1950 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1951 transform_later(need_pf_iff); 1952 1953 // true node, add prefetchdistance 1954 Node *need_pf_true = new IfTrueNode( need_pf_iff ); 1955 transform_later(need_pf_true); 1956 1957 Node *need_pf_false = new IfFalseNode( need_pf_iff ); 1958 transform_later(need_pf_false); 1959 1960 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm, 1961 _igvn.MakeConX(AllocatePrefetchDistance) ); 1962 transform_later(new_pf_wmt ); 1963 new_pf_wmt->set_req(0, need_pf_true); 1964 1965 Node *store_new_wmt = new StorePNode(need_pf_true, 1966 contended_phi_rawmem, eden_pf_adr, 1967 TypeRawPtr::BOTTOM, new_pf_wmt, 1968 MemNode::unordered); 1969 transform_later(store_new_wmt); 1970 1971 // adding prefetches 1972 pf_phi_abio->init_req( fall_in_path, i_o ); 1973 1974 Node *prefetch_adr; 1975 Node *prefetch; 1976 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1977 uint step_size = AllocatePrefetchStepSize; 1978 uint distance = 0; 1979 1980 for ( uint i = 0; i < lines; i++ ) { 1981 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt, 1982 _igvn.MakeConX(distance) ); 1983 transform_later(prefetch_adr); 1984 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1985 transform_later(prefetch); 1986 distance += step_size; 1987 i_o = prefetch; 1988 } 1989 pf_phi_abio->set_req( pf_path, i_o ); 1990 1991 pf_region->init_req( fall_in_path, need_pf_false ); 1992 pf_region->init_req( pf_path, need_pf_true ); 1993 1994 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 1995 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 1996 1997 transform_later(pf_region); 1998 transform_later(pf_phi_rawmem); 1999 transform_later(pf_phi_abio); 2000 2001 needgc_false = pf_region; 2002 contended_phi_rawmem = pf_phi_rawmem; 2003 i_o = pf_phi_abio; 2004 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { 2005 // Insert a prefetch instruction for each allocation. 2006 // This code is used to generate 1 prefetch instruction per cache line. 2007 2008 // Generate several prefetch instructions. 2009 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 2010 uint step_size = AllocatePrefetchStepSize; 2011 uint distance = AllocatePrefetchDistance; 2012 2013 // Next cache address. 2014 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top, 2015 _igvn.MakeConX(step_size + distance)); 2016 transform_later(cache_adr); 2017 cache_adr = new CastP2XNode(needgc_false, cache_adr); 2018 transform_later(cache_adr); 2019 // Address is aligned to execute prefetch to the beginning of cache line size 2020 // (it is important when BIS instruction is used on SPARC as prefetch). 2021 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); 2022 cache_adr = new AndXNode(cache_adr, mask); 2023 transform_later(cache_adr); 2024 cache_adr = new CastX2PNode(cache_adr); 2025 transform_later(cache_adr); 2026 2027 // Prefetch 2028 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); 2029 prefetch->set_req(0, needgc_false); 2030 transform_later(prefetch); 2031 contended_phi_rawmem = prefetch; 2032 Node *prefetch_adr; 2033 distance = step_size; 2034 for ( uint i = 1; i < lines; i++ ) { 2035 prefetch_adr = new AddPNode( cache_adr, cache_adr, 2036 _igvn.MakeConX(distance) ); 2037 transform_later(prefetch_adr); 2038 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); 2039 transform_later(prefetch); 2040 distance += step_size; 2041 contended_phi_rawmem = prefetch; 2042 } 2043 } else if( AllocatePrefetchStyle > 0 ) { 2044 // Insert a prefetch for each allocation only on the fast-path 2045 Node *prefetch_adr; 2046 Node *prefetch; 2047 // Generate several prefetch instructions. 2048 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 2049 uint step_size = AllocatePrefetchStepSize; 2050 uint distance = AllocatePrefetchDistance; 2051 for ( uint i = 0; i < lines; i++ ) { 2052 prefetch_adr = new AddPNode( old_eden_top, new_eden_top, 2053 _igvn.MakeConX(distance) ); 2054 transform_later(prefetch_adr); 2055 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 2056 // Do not let it float too high, since if eden_top == eden_end, 2057 // both might be null. 2058 if( i == 0 ) { // Set control for first prefetch, next follows it 2059 prefetch->init_req(0, needgc_false); 2060 } 2061 transform_later(prefetch); 2062 distance += step_size; 2063 i_o = prefetch; 2064 } 2065 } 2066 return i_o; 2067 } 2068 2069 2070 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 2071 expand_allocate_common(alloc, NULL, 2072 OptoRuntime::new_instance_Type(), 2073 OptoRuntime::new_instance_Java()); 2074 } 2075 2076 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 2077 Node* length = alloc->in(AllocateNode::ALength); 2078 InitializeNode* init = alloc->initialization(); 2079 Node* klass_node = alloc->in(AllocateNode::KlassNode); 2080 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 2081 address slow_call_address; // Address of slow call 2082 if (init != NULL && init->is_complete_with_arraycopy() && 2083 k->is_type_array_klass()) { 2084 // Don't zero type array during slow allocation in VM since 2085 // it will be initialized later by arraycopy in compiled code. 2086 slow_call_address = OptoRuntime::new_array_nozero_Java(); 2087 } else { 2088 slow_call_address = OptoRuntime::new_array_Java(); 2089 } 2090 expand_allocate_common(alloc, length, 2091 OptoRuntime::new_array_Type(), 2092 slow_call_address); 2093 } 2094 2095 //-------------------mark_eliminated_box---------------------------------- 2096 // 2097 // During EA obj may point to several objects but after few ideal graph 2098 // transformations (CCP) it may point to only one non escaping object 2099 // (but still using phi), corresponding locks and unlocks will be marked 2100 // for elimination. Later obj could be replaced with a new node (new phi) 2101 // and which does not have escape information. And later after some graph 2102 // reshape other locks and unlocks (which were not marked for elimination 2103 // before) are connected to this new obj (phi) but they still will not be 2104 // marked for elimination since new obj has no escape information. 2105 // Mark all associated (same box and obj) lock and unlock nodes for 2106 // elimination if some of them marked already. 2107 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) { 2108 if (oldbox->as_BoxLock()->is_eliminated()) 2109 return; // This BoxLock node was processed already. 2110 2111 // New implementation (EliminateNestedLocks) has separate BoxLock 2112 // node for each locked region so mark all associated locks/unlocks as 2113 // eliminated even if different objects are referenced in one locked region 2114 // (for example, OSR compilation of nested loop inside locked scope). 2115 if (EliminateNestedLocks || 2116 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) { 2117 // Box is used only in one lock region. Mark this box as eliminated. 2118 _igvn.hash_delete(oldbox); 2119 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value 2120 _igvn.hash_insert(oldbox); 2121 2122 for (uint i = 0; i < oldbox->outcnt(); i++) { 2123 Node* u = oldbox->raw_out(i); 2124 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) { 2125 AbstractLockNode* alock = u->as_AbstractLock(); 2126 // Check lock's box since box could be referenced by Lock's debug info. 2127 if (alock->box_node() == oldbox) { 2128 // Mark eliminated all related locks and unlocks. 2129 #ifdef ASSERT 2130 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4"); 2131 #endif 2132 alock->set_non_esc_obj(); 2133 } 2134 } 2135 } 2136 return; 2137 } 2138 2139 // Create new "eliminated" BoxLock node and use it in monitor debug info 2140 // instead of oldbox for the same object. 2141 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 2142 2143 // Note: BoxLock node is marked eliminated only here and it is used 2144 // to indicate that all associated lock and unlock nodes are marked 2145 // for elimination. 2146 newbox->set_eliminated(); 2147 transform_later(newbox); 2148 2149 // Replace old box node with new box for all users of the same object. 2150 for (uint i = 0; i < oldbox->outcnt();) { 2151 bool next_edge = true; 2152 2153 Node* u = oldbox->raw_out(i); 2154 if (u->is_AbstractLock()) { 2155 AbstractLockNode* alock = u->as_AbstractLock(); 2156 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) { 2157 // Replace Box and mark eliminated all related locks and unlocks. 2158 #ifdef ASSERT 2159 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5"); 2160 #endif 2161 alock->set_non_esc_obj(); 2162 _igvn.rehash_node_delayed(alock); 2163 alock->set_box_node(newbox); 2164 next_edge = false; 2165 } 2166 } 2167 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) { 2168 FastLockNode* flock = u->as_FastLock(); 2169 assert(flock->box_node() == oldbox, "sanity"); 2170 _igvn.rehash_node_delayed(flock); 2171 flock->set_box_node(newbox); 2172 next_edge = false; 2173 } 2174 2175 // Replace old box in monitor debug info. 2176 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 2177 SafePointNode* sfn = u->as_SafePoint(); 2178 JVMState* youngest_jvms = sfn->jvms(); 2179 int max_depth = youngest_jvms->depth(); 2180 for (int depth = 1; depth <= max_depth; depth++) { 2181 JVMState* jvms = youngest_jvms->of_depth(depth); 2182 int num_mon = jvms->nof_monitors(); 2183 // Loop over monitors 2184 for (int idx = 0; idx < num_mon; idx++) { 2185 Node* obj_node = sfn->monitor_obj(jvms, idx); 2186 Node* box_node = sfn->monitor_box(jvms, idx); 2187 if (box_node == oldbox && obj_node->eqv_uncast(obj)) { 2188 int j = jvms->monitor_box_offset(idx); 2189 _igvn.replace_input_of(u, j, newbox); 2190 next_edge = false; 2191 } 2192 } 2193 } 2194 } 2195 if (next_edge) i++; 2196 } 2197 } 2198 2199 //-----------------------mark_eliminated_locking_nodes----------------------- 2200 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { 2201 if (EliminateNestedLocks) { 2202 if (alock->is_nested()) { 2203 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity"); 2204 return; 2205 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened 2206 // Only Lock node has JVMState needed here. 2207 // Not that preceding claim is documented anywhere else. 2208 if (alock->jvms() != NULL) { 2209 if (alock->as_Lock()->is_nested_lock_region()) { 2210 // Mark eliminated related nested locks and unlocks. 2211 Node* obj = alock->obj_node(); 2212 BoxLockNode* box_node = alock->box_node()->as_BoxLock(); 2213 assert(!box_node->is_eliminated(), "should not be marked yet"); 2214 // Note: BoxLock node is marked eliminated only here 2215 // and it is used to indicate that all associated lock 2216 // and unlock nodes are marked for elimination. 2217 box_node->set_eliminated(); // Box's hash is always NO_HASH here 2218 for (uint i = 0; i < box_node->outcnt(); i++) { 2219 Node* u = box_node->raw_out(i); 2220 if (u->is_AbstractLock()) { 2221 alock = u->as_AbstractLock(); 2222 if (alock->box_node() == box_node) { 2223 // Verify that this Box is referenced only by related locks. 2224 assert(alock->obj_node()->eqv_uncast(obj), ""); 2225 // Mark all related locks and unlocks. 2226 #ifdef ASSERT 2227 alock->log_lock_optimization(C, "eliminate_lock_set_nested"); 2228 #endif 2229 alock->set_nested(); 2230 } 2231 } 2232 } 2233 } else { 2234 #ifdef ASSERT 2235 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region"); 2236 if (C->log() != NULL) 2237 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output 2238 #endif 2239 } 2240 } 2241 return; 2242 } 2243 // Process locks for non escaping object 2244 assert(alock->is_non_esc_obj(), ""); 2245 } // EliminateNestedLocks 2246 2247 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object 2248 // Look for all locks of this object and mark them and 2249 // corresponding BoxLock nodes as eliminated. 2250 Node* obj = alock->obj_node(); 2251 for (uint j = 0; j < obj->outcnt(); j++) { 2252 Node* o = obj->raw_out(j); 2253 if (o->is_AbstractLock() && 2254 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) { 2255 alock = o->as_AbstractLock(); 2256 Node* box = alock->box_node(); 2257 // Replace old box node with new eliminated box for all users 2258 // of the same object and mark related locks as eliminated. 2259 mark_eliminated_box(box, obj); 2260 } 2261 } 2262 } 2263 } 2264 2265 // we have determined that this lock/unlock can be eliminated, we simply 2266 // eliminate the node without expanding it. 2267 // 2268 // Note: The membar's associated with the lock/unlock are currently not 2269 // eliminated. This should be investigated as a future enhancement. 2270 // 2271 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 2272 2273 if (!alock->is_eliminated()) { 2274 return false; 2275 } 2276 #ifdef ASSERT 2277 if (!alock->is_coarsened()) { 2278 // Check that new "eliminated" BoxLock node is created. 2279 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 2280 assert(oldbox->is_eliminated(), "should be done already"); 2281 } 2282 #endif 2283 2284 alock->log_lock_optimization(C, "eliminate_lock"); 2285 2286 #ifndef PRODUCT 2287 if (PrintEliminateLocks) { 2288 if (alock->is_Lock()) { 2289 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx); 2290 } else { 2291 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx); 2292 } 2293 } 2294 #endif 2295 2296 Node* mem = alock->in(TypeFunc::Memory); 2297 Node* ctrl = alock->in(TypeFunc::Control); 2298 2299 extract_call_projections(alock); 2300 // There are 2 projections from the lock. The lock node will 2301 // be deleted when its last use is subsumed below. 2302 assert(alock->outcnt() == 2 && 2303 _fallthroughproj != NULL && 2304 _memproj_fallthrough != NULL, 2305 "Unexpected projections from Lock/Unlock"); 2306 2307 Node* fallthroughproj = _fallthroughproj; 2308 Node* memproj_fallthrough = _memproj_fallthrough; 2309 2310 // The memory projection from a lock/unlock is RawMem 2311 // The input to a Lock is merged memory, so extract its RawMem input 2312 // (unless the MergeMem has been optimized away.) 2313 if (alock->is_Lock()) { 2314 // Seach for MemBarAcquireLock node and delete it also. 2315 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 2316 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, ""); 2317 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 2318 Node* memproj = membar->proj_out(TypeFunc::Memory); 2319 _igvn.replace_node(ctrlproj, fallthroughproj); 2320 _igvn.replace_node(memproj, memproj_fallthrough); 2321 2322 // Delete FastLock node also if this Lock node is unique user 2323 // (a loop peeling may clone a Lock node). 2324 Node* flock = alock->as_Lock()->fastlock_node(); 2325 if (flock->outcnt() == 1) { 2326 assert(flock->unique_out() == alock, "sanity"); 2327 _igvn.replace_node(flock, top()); 2328 } 2329 } 2330 2331 // Seach for MemBarReleaseLock node and delete it also. 2332 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() && 2333 ctrl->in(0)->is_MemBar()) { 2334 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 2335 assert(membar->Opcode() == Op_MemBarReleaseLock && 2336 mem->is_Proj() && membar == mem->in(0), ""); 2337 _igvn.replace_node(fallthroughproj, ctrl); 2338 _igvn.replace_node(memproj_fallthrough, mem); 2339 fallthroughproj = ctrl; 2340 memproj_fallthrough = mem; 2341 ctrl = membar->in(TypeFunc::Control); 2342 mem = membar->in(TypeFunc::Memory); 2343 } 2344 2345 _igvn.replace_node(fallthroughproj, ctrl); 2346 _igvn.replace_node(memproj_fallthrough, mem); 2347 return true; 2348 } 2349 2350 2351 //------------------------------expand_lock_node---------------------- 2352 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 2353 2354 Node* ctrl = lock->in(TypeFunc::Control); 2355 Node* mem = lock->in(TypeFunc::Memory); 2356 Node* obj = lock->obj_node(); 2357 Node* box = lock->box_node(); 2358 Node* flock = lock->fastlock_node(); 2359 2360 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2361 2362 // Make the merge point 2363 Node *region; 2364 Node *mem_phi; 2365 Node *slow_path; 2366 2367 if (UseOptoBiasInlining) { 2368 /* 2369 * See the full description in MacroAssembler::biased_locking_enter(). 2370 * 2371 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { 2372 * // The object is biased. 2373 * proto_node = klass->prototype_header; 2374 * o_node = thread | proto_node; 2375 * x_node = o_node ^ mark_word; 2376 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? 2377 * // Done. 2378 * } else { 2379 * if( (x_node & biased_lock_mask) != 0 ) { 2380 * // The klass's prototype header is no longer biased. 2381 * cas(&mark_word, mark_word, proto_node) 2382 * goto cas_lock; 2383 * } else { 2384 * // The klass's prototype header is still biased. 2385 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? 2386 * old = mark_word; 2387 * new = o_node; 2388 * } else { 2389 * // Different thread or anonymous biased. 2390 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); 2391 * new = thread | old; 2392 * } 2393 * // Try to rebias. 2394 * if( cas(&mark_word, old, new) == 0 ) { 2395 * // Done. 2396 * } else { 2397 * goto slow_path; // Failed. 2398 * } 2399 * } 2400 * } 2401 * } else { 2402 * // The object is not biased. 2403 * cas_lock: 2404 * if( FastLock(obj) == 0 ) { 2405 * // Done. 2406 * } else { 2407 * slow_path: 2408 * OptoRuntime::complete_monitor_locking_Java(obj); 2409 * } 2410 * } 2411 */ 2412 2413 region = new RegionNode(5); 2414 // create a Phi for the memory state 2415 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2416 2417 Node* fast_lock_region = new RegionNode(3); 2418 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); 2419 2420 // First, check mark word for the biased lock pattern. 2421 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2422 2423 // Get fast path - mark word has the biased lock pattern. 2424 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, 2425 markOopDesc::biased_lock_mask_in_place, 2426 markOopDesc::biased_lock_pattern, true); 2427 // fast_lock_region->in(1) is set to slow path. 2428 fast_lock_mem_phi->init_req(1, mem); 2429 2430 // Now check that the lock is biased to the current thread and has 2431 // the same epoch and bias as Klass::_prototype_header. 2432 2433 // Special-case a fresh allocation to avoid building nodes: 2434 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); 2435 if (klass_node == NULL) { 2436 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 2437 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr())); 2438 #ifdef _LP64 2439 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) { 2440 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity"); 2441 klass_node->in(1)->init_req(0, ctrl); 2442 } else 2443 #endif 2444 klass_node->init_req(0, ctrl); 2445 } 2446 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type()); 2447 2448 Node* thread = transform_later(new ThreadLocalNode()); 2449 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2450 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node)); 2451 Node* x_node = transform_later(new XorXNode(o_node, mark_node)); 2452 2453 // Get slow path - mark word does NOT match the value. 2454 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, 2455 (~markOopDesc::age_mask_in_place), 0); 2456 // region->in(3) is set to fast path - the object is biased to the current thread. 2457 mem_phi->init_req(3, mem); 2458 2459 2460 // Mark word does NOT match the value (thread | Klass::_prototype_header). 2461 2462 2463 // First, check biased pattern. 2464 // Get fast path - _prototype_header has the same biased lock pattern. 2465 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, 2466 markOopDesc::biased_lock_mask_in_place, 0, true); 2467 2468 not_biased_ctrl = fast_lock_region->in(2); // Slow path 2469 // fast_lock_region->in(2) - the prototype header is no longer biased 2470 // and we have to revoke the bias on this object. 2471 // We are going to try to reset the mark of this object to the prototype 2472 // value and fall through to the CAS-based locking scheme. 2473 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 2474 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr, 2475 proto_node, mark_node); 2476 transform_later(cas); 2477 Node* proj = transform_later(new SCMemProjNode(cas)); 2478 fast_lock_mem_phi->init_req(2, proj); 2479 2480 2481 // Second, check epoch bits. 2482 Node* rebiased_region = new RegionNode(3); 2483 Node* old_phi = new PhiNode( rebiased_region, TypeX_X); 2484 Node* new_phi = new PhiNode( rebiased_region, TypeX_X); 2485 2486 // Get slow path - mark word does NOT match epoch bits. 2487 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, 2488 markOopDesc::epoch_mask_in_place, 0); 2489 // The epoch of the current bias is not valid, attempt to rebias the object 2490 // toward the current thread. 2491 rebiased_region->init_req(2, epoch_ctrl); 2492 old_phi->init_req(2, mark_node); 2493 new_phi->init_req(2, o_node); 2494 2495 // rebiased_region->in(1) is set to fast path. 2496 // The epoch of the current bias is still valid but we know 2497 // nothing about the owner; it might be set or it might be clear. 2498 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place | 2499 markOopDesc::age_mask_in_place | 2500 markOopDesc::epoch_mask_in_place); 2501 Node* old = transform_later(new AndXNode(mark_node, cmask)); 2502 cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2503 Node* new_mark = transform_later(new OrXNode(cast_thread, old)); 2504 old_phi->init_req(1, old); 2505 new_phi->init_req(1, new_mark); 2506 2507 transform_later(rebiased_region); 2508 transform_later(old_phi); 2509 transform_later(new_phi); 2510 2511 // Try to acquire the bias of the object using an atomic operation. 2512 // If this fails we will go in to the runtime to revoke the object's bias. 2513 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi); 2514 transform_later(cas); 2515 proj = transform_later(new SCMemProjNode(cas)); 2516 2517 // Get slow path - Failed to CAS. 2518 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); 2519 mem_phi->init_req(4, proj); 2520 // region->in(4) is set to fast path - the object is rebiased to the current thread. 2521 2522 // Failed to CAS. 2523 slow_path = new RegionNode(3); 2524 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); 2525 2526 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control 2527 slow_mem->init_req(1, proj); 2528 2529 // Call CAS-based locking scheme (FastLock node). 2530 2531 transform_later(fast_lock_region); 2532 transform_later(fast_lock_mem_phi); 2533 2534 // Get slow path - FastLock failed to lock the object. 2535 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); 2536 mem_phi->init_req(2, fast_lock_mem_phi); 2537 // region->in(2) is set to fast path - the object is locked to the current thread. 2538 2539 slow_path->init_req(2, ctrl); // Capture slow-control 2540 slow_mem->init_req(2, fast_lock_mem_phi); 2541 2542 transform_later(slow_path); 2543 transform_later(slow_mem); 2544 // Reset lock's memory edge. 2545 lock->set_req(TypeFunc::Memory, slow_mem); 2546 2547 } else { 2548 region = new RegionNode(3); 2549 // create a Phi for the memory state 2550 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2551 2552 // Optimize test; set region slot 2 2553 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); 2554 mem_phi->init_req(2, mem); 2555 } 2556 2557 // Make slow path call 2558 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), 2559 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, 2560 obj, box, NULL); 2561 2562 extract_call_projections(call); 2563 2564 // Slow path can only throw asynchronous exceptions, which are always 2565 // de-opted. So the compiler thinks the slow-call can never throw an 2566 // exception. If it DOES throw an exception we would need the debug 2567 // info removed first (since if it throws there is no monitor). 2568 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2569 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2570 2571 // Capture slow path 2572 // disconnect fall-through projection from call and create a new one 2573 // hook up users of fall-through projection to region 2574 Node *slow_ctrl = _fallthroughproj->clone(); 2575 transform_later(slow_ctrl); 2576 _igvn.hash_delete(_fallthroughproj); 2577 _fallthroughproj->disconnect_inputs(NULL, C); 2578 region->init_req(1, slow_ctrl); 2579 // region inputs are now complete 2580 transform_later(region); 2581 _igvn.replace_node(_fallthroughproj, region); 2582 2583 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory)); 2584 mem_phi->init_req(1, memproj ); 2585 transform_later(mem_phi); 2586 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2587 } 2588 2589 //------------------------------expand_unlock_node---------------------- 2590 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 2591 2592 Node* ctrl = unlock->in(TypeFunc::Control); 2593 Node* mem = unlock->in(TypeFunc::Memory); 2594 Node* obj = unlock->obj_node(); 2595 Node* box = unlock->box_node(); 2596 2597 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2598 2599 // No need for a null check on unlock 2600 2601 // Make the merge point 2602 Node *region; 2603 Node *mem_phi; 2604 2605 if (UseOptoBiasInlining) { 2606 // Check for biased locking unlock case, which is a no-op. 2607 // See the full description in MacroAssembler::biased_locking_exit(). 2608 region = new RegionNode(4); 2609 // create a Phi for the memory state 2610 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2611 mem_phi->init_req(3, mem); 2612 2613 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2614 ctrl = opt_bits_test(ctrl, region, 3, mark_node, 2615 markOopDesc::biased_lock_mask_in_place, 2616 markOopDesc::biased_lock_pattern); 2617 } else { 2618 region = new RegionNode(3); 2619 // create a Phi for the memory state 2620 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2621 } 2622 2623 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box ); 2624 funlock = transform_later( funlock )->as_FastUnlock(); 2625 // Optimize test; set region slot 2 2626 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); 2627 Node *thread = transform_later(new ThreadLocalNode()); 2628 2629 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), 2630 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), 2631 "complete_monitor_unlocking_C", slow_path, obj, box, thread); 2632 2633 extract_call_projections(call); 2634 2635 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2636 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2637 2638 // No exceptions for unlocking 2639 // Capture slow path 2640 // disconnect fall-through projection from call and create a new one 2641 // hook up users of fall-through projection to region 2642 Node *slow_ctrl = _fallthroughproj->clone(); 2643 transform_later(slow_ctrl); 2644 _igvn.hash_delete(_fallthroughproj); 2645 _fallthroughproj->disconnect_inputs(NULL, C); 2646 region->init_req(1, slow_ctrl); 2647 // region inputs are now complete 2648 transform_later(region); 2649 _igvn.replace_node(_fallthroughproj, region); 2650 2651 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) ); 2652 mem_phi->init_req(1, memproj ); 2653 mem_phi->init_req(2, mem); 2654 transform_later(mem_phi); 2655 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2656 } 2657 2658 // A value type is returned from the call but we don't know its 2659 // type. Either we get a buffered value (and nothing needs to be done) 2660 // or one of the values being returned is the klass of the value type 2661 // and we need to allocate a value type instance of that type and 2662 // initialize it with other values being returned. In that case, we 2663 // first try a fast path allocation and initialize the value with the 2664 // value klass's pack handler or we fall back to a runtime call. 2665 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) { 2666 Node* ret = call->proj_out(TypeFunc::Parms); 2667 if (ret == NULL) { 2668 return; 2669 } 2670 assert(ret->bottom_type()->is_valuetypeptr()->is__Value(), "unexpected return type from MH intrinsic"); 2671 const TypeFunc* tf = call->_tf; 2672 const TypeTuple* domain = OptoRuntime::store_value_type_fields_Type()->domain_cc(); 2673 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain); 2674 call->_tf = new_tf; 2675 // Make sure the change of type is applied before projections are 2676 // processed by igvn 2677 _igvn.set_type(call, call->Value(&_igvn)); 2678 _igvn.set_type(ret, ret->Value(&_igvn)); 2679 2680 // Before any new projection is added: 2681 CallProjections projs; 2682 call->extract_projections(&projs, true, true); 2683 2684 Node* ctl = new Node(1); 2685 Node* mem = new Node(1); 2686 Node* io = new Node(1); 2687 Node* ex_ctl = new Node(1); 2688 Node* ex_mem = new Node(1); 2689 Node* ex_io = new Node(1); 2690 Node* res = new Node(1); 2691 2692 Node* cast = transform_later(new CastP2XNode(ctl, res)); 2693 Node* mask = MakeConX(0x1); 2694 Node* masked = transform_later(new AndXNode(cast, mask)); 2695 Node* cmp = transform_later(new CmpXNode(masked, mask)); 2696 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq)); 2697 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN); 2698 transform_later(allocation_iff); 2699 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff)); 2700 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff)); 2701 2702 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeValueTypePtr::NOTNULL)); 2703 2704 Node* mask2 = MakeConX(-2); 2705 Node* masked2 = transform_later(new AndXNode(cast, mask2)); 2706 Node* rawklassptr = transform_later(new CastX2PNode(masked2)); 2707 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeKlassPtr::VALUE)); 2708 2709 Node* top_adr; 2710 Node* end_adr; 2711 2712 Node* slowpath_bol = NULL; 2713 Node* old_top = NULL; 2714 Node* new_top = NULL; 2715 if (UseTLAB) { 2716 set_eden_pointers(top_adr, end_adr); 2717 Node* end = make_load(ctl, mem, end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 2718 old_top = new LoadPNode(ctl, mem, top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered); 2719 transform_later(old_top); 2720 Node* layout_val = make_load(NULL, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT); 2721 Node* size_in_bytes = ConvI2X(layout_val); 2722 new_top = new AddPNode(top(), old_top, size_in_bytes); 2723 transform_later(new_top); 2724 Node* slowpath_cmp = new CmpPNode(new_top, end); 2725 transform_later(slowpath_cmp); 2726 slowpath_bol = new BoolNode(slowpath_cmp, BoolTest::ge); 2727 transform_later(slowpath_bol); 2728 } else { 2729 slowpath_bol = intcon(1); 2730 old_top = top(); 2731 new_top = top(); 2732 } 2733 IfNode* slowpath_iff = new IfNode(allocation_ctl, slowpath_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); 2734 transform_later(slowpath_iff); 2735 2736 Node* slowpath_true = new IfTrueNode(slowpath_iff); 2737 transform_later(slowpath_true); 2738 2739 2740 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_value_type_fields_Type(), 2741 StubRoutines::store_value_type_fields_to_buf(), 2742 "store_value_type_fields", 2743 call->jvms()->bci(), 2744 TypePtr::BOTTOM); 2745 slow_call->init_req(TypeFunc::Control, slowpath_true); 2746 slow_call->init_req(TypeFunc::Memory, mem); 2747 slow_call->init_req(TypeFunc::I_O, io); 2748 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr)); 2749 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr)); 2750 slow_call->init_req(TypeFunc::Parms, res); 2751 2752 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control)); 2753 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory)); 2754 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O)); 2755 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms)); 2756 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2)); 2757 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci)); 2758 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci)); 2759 2760 Node* ex_r = new RegionNode(3); 2761 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM); 2762 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO); 2763 ex_r->init_req(1, slow_excp); 2764 ex_mem_phi->init_req(1, slow_mem); 2765 ex_io_phi->init_req(1, slow_io); 2766 ex_r->init_req(2, ex_ctl); 2767 ex_mem_phi->init_req(2, ex_mem); 2768 ex_io_phi->init_req(2, ex_io); 2769 2770 transform_later(ex_r); 2771 transform_later(ex_mem_phi); 2772 transform_later(ex_io_phi); 2773 2774 Node* slowpath_false = new IfFalseNode(slowpath_iff); 2775 transform_later(slowpath_false); 2776 Node* rawmem = new StorePNode(slowpath_false, mem, top_adr, TypeRawPtr::BOTTOM, new_top, MemNode::unordered); 2777 transform_later(rawmem); 2778 Node* mark_node = NULL; 2779 // For now only enable fast locking for non-array types 2780 if (UseBiasedLocking) { 2781 mark_node = make_load(slowpath_false, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 2782 } else { 2783 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); 2784 } 2785 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 2786 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 2787 if (UseCompressedClassPointers) { 2788 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT); 2789 } 2790 Node* pack_handler = make_load(slowpath_false, rawmem, klass_node, in_bytes(ValueKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 2791 2792 CallLeafNoFPNode* handler_call = new CallLeafNoFPNode(OptoRuntime::pack_value_type_Type(), 2793 NULL, 2794 "pack handler", 2795 TypeRawPtr::BOTTOM); 2796 handler_call->init_req(TypeFunc::Control, slowpath_false); 2797 handler_call->init_req(TypeFunc::Memory, rawmem); 2798 handler_call->init_req(TypeFunc::I_O, top()); 2799 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr)); 2800 handler_call->init_req(TypeFunc::ReturnAdr, top()); 2801 handler_call->init_req(TypeFunc::Parms, pack_handler); 2802 handler_call->init_req(TypeFunc::Parms+1, old_top); 2803 2804 // We don't know how many values are returned. This assumes the 2805 // worst case, that all available registers are used. 2806 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) { 2807 if (domain->field_at(i) == Type::HALF) { 2808 slow_call->init_req(i, top()); 2809 handler_call->init_req(i+1, top()); 2810 continue; 2811 } 2812 Node* proj = transform_later(new ProjNode(call, i)); 2813 slow_call->init_req(i, proj); 2814 handler_call->init_req(i+1, proj); 2815 } 2816 2817 // We can safepoint at that new call 2818 C->add_safepoint_edges(slow_call, call->jvms()); 2819 transform_later(slow_call); 2820 transform_later(handler_call); 2821 2822 Node* handler_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control)); 2823 rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory)); 2824 Node* slowpath_false_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms)); 2825 2826 MergeMemNode* slowpath_false_mem = MergeMemNode::make(mem); 2827 slowpath_false_mem->set_memory_at(Compile::AliasIdxRaw, rawmem); 2828 transform_later(slowpath_false_mem); 2829 2830 Node* r = new RegionNode(4); 2831 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM); 2832 Node* io_phi = new PhiNode(r, Type::ABIO); 2833 Node* res_phi = new PhiNode(r, ret->bottom_type()); 2834 2835 r->init_req(1, no_allocation_ctl); 2836 mem_phi->init_req(1, mem); 2837 io_phi->init_req(1, io); 2838 res_phi->init_req(1, no_allocation_res); 2839 r->init_req(2, slow_norm); 2840 mem_phi->init_req(2, slow_mem); 2841 io_phi->init_req(2, slow_io); 2842 res_phi->init_req(2, slow_res); 2843 r->init_req(3, handler_ctl); 2844 mem_phi->init_req(3, slowpath_false_mem); 2845 io_phi->init_req(3, io); 2846 res_phi->init_req(3, slowpath_false_res); 2847 2848 transform_later(r); 2849 transform_later(mem_phi); 2850 transform_later(io_phi); 2851 transform_later(res_phi); 2852 2853 _igvn.replace_in_uses(projs.fallthrough_catchproj, r); 2854 _igvn.replace_in_uses(projs.fallthrough_memproj, mem_phi); 2855 _igvn.replace_in_uses(projs.fallthrough_ioproj, io_phi); 2856 _igvn.replace_in_uses(projs.resproj, res_phi); 2857 _igvn.replace_in_uses(projs.catchall_catchproj, ex_r); 2858 _igvn.replace_in_uses(projs.catchall_memproj, ex_mem_phi); 2859 _igvn.replace_in_uses(projs.catchall_ioproj, ex_io_phi); 2860 2861 _igvn.replace_node(ctl, projs.fallthrough_catchproj); 2862 _igvn.replace_node(mem, projs.fallthrough_memproj); 2863 _igvn.replace_node(io, projs.fallthrough_ioproj); 2864 _igvn.replace_node(res, projs.resproj); 2865 _igvn.replace_node(ex_ctl, projs.catchall_catchproj); 2866 _igvn.replace_node(ex_mem, projs.catchall_memproj); 2867 _igvn.replace_node(ex_io, projs.catchall_ioproj); 2868 } 2869 2870 //---------------------------eliminate_macro_nodes---------------------- 2871 // Eliminate scalar replaced allocations and associated locks. 2872 void PhaseMacroExpand::eliminate_macro_nodes() { 2873 if (C->macro_count() == 0) 2874 return; 2875 2876 // First, attempt to eliminate locks 2877 int cnt = C->macro_count(); 2878 for (int i=0; i < cnt; i++) { 2879 Node *n = C->macro_node(i); 2880 if (n->is_AbstractLock()) { // Lock and Unlock nodes 2881 // Before elimination mark all associated (same box and obj) 2882 // lock and unlock nodes. 2883 mark_eliminated_locking_nodes(n->as_AbstractLock()); 2884 } 2885 } 2886 bool progress = true; 2887 while (progress) { 2888 progress = false; 2889 for (int i = C->macro_count(); i > 0; i--) { 2890 Node * n = C->macro_node(i-1); 2891 bool success = false; 2892 debug_only(int old_macro_count = C->macro_count();); 2893 if (n->is_AbstractLock()) { 2894 success = eliminate_locking_node(n->as_AbstractLock()); 2895 } 2896 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2897 progress = progress || success; 2898 } 2899 } 2900 // Next, attempt to eliminate allocations 2901 _has_locks = false; 2902 progress = true; 2903 while (progress) { 2904 progress = false; 2905 for (int i = C->macro_count(); i > 0; i--) { 2906 Node * n = C->macro_node(i-1); 2907 bool success = false; 2908 debug_only(int old_macro_count = C->macro_count();); 2909 switch (n->class_id()) { 2910 case Node::Class_Allocate: 2911 case Node::Class_AllocateArray: 2912 success = eliminate_allocate_node(n->as_Allocate()); 2913 break; 2914 case Node::Class_CallStaticJava: { 2915 CallStaticJavaNode* call = n->as_CallStaticJava(); 2916 if (!call->method()->is_method_handle_intrinsic()) { 2917 success = eliminate_boxing_node(n->as_CallStaticJava()); 2918 } 2919 break; 2920 } 2921 case Node::Class_Lock: 2922 case Node::Class_Unlock: 2923 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 2924 _has_locks = true; 2925 break; 2926 case Node::Class_ArrayCopy: 2927 break; 2928 default: 2929 assert(n->Opcode() == Op_LoopLimit || 2930 n->Opcode() == Op_Opaque1 || 2931 n->Opcode() == Op_Opaque2 || 2932 n->Opcode() == Op_Opaque3 || 2933 n->Opcode() == Op_Opaque4, "unknown node type in macro list"); 2934 } 2935 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2936 progress = progress || success; 2937 } 2938 } 2939 } 2940 2941 //------------------------------expand_macro_nodes---------------------- 2942 // Returns true if a failure occurred. 2943 bool PhaseMacroExpand::expand_macro_nodes() { 2944 // Last attempt to eliminate macro nodes. 2945 eliminate_macro_nodes(); 2946 2947 // Make sure expansion will not cause node limit to be exceeded. 2948 // Worst case is a macro node gets expanded into about 200 nodes. 2949 // Allow 50% more for optimization. 2950 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) ) 2951 return true; 2952 2953 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. 2954 bool progress = true; 2955 while (progress) { 2956 progress = false; 2957 for (int i = C->macro_count(); i > 0; i--) { 2958 Node * n = C->macro_node(i-1); 2959 bool success = false; 2960 debug_only(int old_macro_count = C->macro_count();); 2961 if (n->Opcode() == Op_LoopLimit) { 2962 // Remove it from macro list and put on IGVN worklist to optimize. 2963 C->remove_macro_node(n); 2964 _igvn._worklist.push(n); 2965 success = true; 2966 } else if (n->Opcode() == Op_CallStaticJava) { 2967 CallStaticJavaNode* call = n->as_CallStaticJava(); 2968 if (!call->method()->is_method_handle_intrinsic()) { 2969 // Remove it from macro list and put on IGVN worklist to optimize. 2970 C->remove_macro_node(n); 2971 _igvn._worklist.push(n); 2972 success = true; 2973 } 2974 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { 2975 _igvn.replace_node(n, n->in(1)); 2976 success = true; 2977 #if INCLUDE_RTM_OPT 2978 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) { 2979 assert(C->profile_rtm(), "should be used only in rtm deoptimization code"); 2980 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), ""); 2981 Node* cmp = n->unique_out(); 2982 #ifdef ASSERT 2983 // Validate graph. 2984 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), ""); 2985 BoolNode* bol = cmp->unique_out()->as_Bool(); 2986 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() && 2987 (bol->_test._test == BoolTest::ne), ""); 2988 IfNode* ifn = bol->unique_out()->as_If(); 2989 assert((ifn->outcnt() == 2) && 2990 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, ""); 2991 #endif 2992 Node* repl = n->in(1); 2993 if (!_has_locks) { 2994 // Remove RTM state check if there are no locks in the code. 2995 // Replace input to compare the same value. 2996 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1); 2997 } 2998 _igvn.replace_node(n, repl); 2999 success = true; 3000 #endif 3001 } else if (n->Opcode() == Op_Opaque4) { 3002 _igvn.replace_node(n, n->in(2)); 3003 success = true; 3004 } 3005 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 3006 progress = progress || success; 3007 } 3008 } 3009 3010 // expand arraycopy "macro" nodes first 3011 // For ReduceBulkZeroing, we must first process all arraycopy nodes 3012 // before the allocate nodes are expanded. 3013 int macro_idx = C->macro_count() - 1; 3014 while (macro_idx >= 0) { 3015 Node * n = C->macro_node(macro_idx); 3016 assert(n->is_macro(), "only macro nodes expected here"); 3017 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 3018 // node is unreachable, so don't try to expand it 3019 C->remove_macro_node(n); 3020 } else if (n->is_ArrayCopy()){ 3021 int macro_count = C->macro_count(); 3022 expand_arraycopy_node(n->as_ArrayCopy()); 3023 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 3024 } 3025 if (C->failing()) return true; 3026 macro_idx --; 3027 } 3028 3029 // expand "macro" nodes 3030 // nodes are removed from the macro list as they are processed 3031 while (C->macro_count() > 0) { 3032 int macro_count = C->macro_count(); 3033 Node * n = C->macro_node(macro_count-1); 3034 assert(n->is_macro(), "only macro nodes expected here"); 3035 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 3036 // node is unreachable, so don't try to expand it 3037 C->remove_macro_node(n); 3038 continue; 3039 } 3040 switch (n->class_id()) { 3041 case Node::Class_Allocate: 3042 expand_allocate(n->as_Allocate()); 3043 break; 3044 case Node::Class_AllocateArray: 3045 expand_allocate_array(n->as_AllocateArray()); 3046 break; 3047 case Node::Class_Lock: 3048 expand_lock_node(n->as_Lock()); 3049 break; 3050 case Node::Class_Unlock: 3051 expand_unlock_node(n->as_Unlock()); 3052 break; 3053 case Node::Class_CallStaticJava: 3054 expand_mh_intrinsic_return(n->as_CallStaticJava()); 3055 C->remove_macro_node(n); 3056 break; 3057 default: 3058 assert(false, "unknown node type in macro list"); 3059 } 3060 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 3061 if (C->failing()) return true; 3062 } 3063 3064 _igvn.set_delay_transform(false); 3065 _igvn.optimize(); 3066 if (C->failing()) return true; 3067 return false; 3068 }