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