1 /* 2 * Copyright (c) 2005, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "libadt/vectset.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/arraycopynode.hpp" 30 #include "opto/callnode.hpp" 31 #include "opto/castnode.hpp" 32 #include "opto/cfgnode.hpp" 33 #include "opto/compile.hpp" 34 #include "opto/convertnode.hpp" 35 #include "opto/graphKit.hpp" 36 #include "opto/locknode.hpp" 37 #include "opto/loopnode.hpp" 38 #include "opto/macro.hpp" 39 #include "opto/memnode.hpp" 40 #include "opto/narrowptrnode.hpp" 41 #include "opto/node.hpp" 42 #include "opto/opaquenode.hpp" 43 #include "opto/phaseX.hpp" 44 #include "opto/rootnode.hpp" 45 #include "opto/runtime.hpp" 46 #include "opto/subnode.hpp" 47 #include "opto/type.hpp" 48 #include "runtime/sharedRuntime.hpp" 49 50 51 // 52 // Replace any references to "oldref" in inputs to "use" with "newref". 53 // Returns the number of replacements made. 54 // 55 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { 56 int nreplacements = 0; 57 uint req = use->req(); 58 for (uint j = 0; j < use->len(); j++) { 59 Node *uin = use->in(j); 60 if (uin == oldref) { 61 if (j < req) 62 use->set_req(j, newref); 63 else 64 use->set_prec(j, newref); 65 nreplacements++; 66 } else if (j >= req && uin == NULL) { 67 break; 68 } 69 } 70 return nreplacements; 71 } 72 73 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) { 74 // Copy debug information and adjust JVMState information 75 uint old_dbg_start = oldcall->tf()->domain()->cnt(); 76 uint new_dbg_start = newcall->tf()->domain()->cnt(); 77 int jvms_adj = new_dbg_start - old_dbg_start; 78 assert (new_dbg_start == newcall->req(), "argument count mismatch"); 79 80 // SafePointScalarObject node could be referenced several times in debug info. 81 // Use Dict to record cloned nodes. 82 Dict* sosn_map = new Dict(cmpkey,hashkey); 83 for (uint i = old_dbg_start; i < oldcall->req(); i++) { 84 Node* old_in = oldcall->in(i); 85 // Clone old SafePointScalarObjectNodes, adjusting their field contents. 86 if (old_in != NULL && old_in->is_SafePointScalarObject()) { 87 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject(); 88 uint old_unique = C->unique(); 89 Node* new_in = old_sosn->clone(sosn_map); 90 if (old_unique != C->unique()) { // New node? 91 new_in->set_req(0, C->root()); // reset control edge 92 new_in = transform_later(new_in); // Register new node. 93 } 94 old_in = new_in; 95 } 96 newcall->add_req(old_in); 97 } 98 99 // JVMS may be shared so clone it before we modify it 100 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL); 101 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) { 102 jvms->set_map(newcall); 103 jvms->set_locoff(jvms->locoff()+jvms_adj); 104 jvms->set_stkoff(jvms->stkoff()+jvms_adj); 105 jvms->set_monoff(jvms->monoff()+jvms_adj); 106 jvms->set_scloff(jvms->scloff()+jvms_adj); 107 jvms->set_endoff(jvms->endoff()+jvms_adj); 108 } 109 } 110 111 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) { 112 Node* cmp; 113 if (mask != 0) { 114 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask))); 115 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits))); 116 } else { 117 cmp = word; 118 } 119 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne)); 120 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); 121 transform_later(iff); 122 123 // Fast path taken. 124 Node *fast_taken = transform_later(new IfFalseNode(iff)); 125 126 // Fast path not-taken, i.e. slow path 127 Node *slow_taken = transform_later(new IfTrueNode(iff)); 128 129 if (return_fast_path) { 130 region->init_req(edge, slow_taken); // Capture slow-control 131 return fast_taken; 132 } else { 133 region->init_req(edge, fast_taken); // Capture fast-control 134 return slow_taken; 135 } 136 } 137 138 //--------------------copy_predefined_input_for_runtime_call-------------------- 139 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { 140 // Set fixed predefined input arguments 141 call->init_req( TypeFunc::Control, ctrl ); 142 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); 143 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? 144 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); 145 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); 146 } 147 148 //------------------------------make_slow_call--------------------------------- 149 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, 150 address slow_call, const char* leaf_name, Node* slow_path, 151 Node* parm0, Node* parm1, Node* parm2) { 152 153 // Slow-path call 154 CallNode *call = leaf_name 155 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) 156 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM ); 157 158 // Slow path call has no side-effects, uses few values 159 copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); 160 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0); 161 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1); 162 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2); 163 copy_call_debug_info(oldcall, call); 164 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 165 _igvn.replace_node(oldcall, call); 166 transform_later(call); 167 168 return call; 169 } 170 171 void PhaseMacroExpand::extract_call_projections(CallNode *call) { 172 _fallthroughproj = NULL; 173 _fallthroughcatchproj = NULL; 174 _ioproj_fallthrough = NULL; 175 _ioproj_catchall = NULL; 176 _catchallcatchproj = NULL; 177 _memproj_fallthrough = NULL; 178 _memproj_catchall = NULL; 179 _resproj = NULL; 180 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) { 181 ProjNode *pn = call->fast_out(i)->as_Proj(); 182 switch (pn->_con) { 183 case TypeFunc::Control: 184 { 185 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj 186 _fallthroughproj = pn; 187 DUIterator_Fast jmax, j = pn->fast_outs(jmax); 188 const Node *cn = pn->fast_out(j); 189 if (cn->is_Catch()) { 190 ProjNode *cpn = NULL; 191 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { 192 cpn = cn->fast_out(k)->as_Proj(); 193 assert(cpn->is_CatchProj(), "must be a CatchProjNode"); 194 if (cpn->_con == CatchProjNode::fall_through_index) 195 _fallthroughcatchproj = cpn; 196 else { 197 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); 198 _catchallcatchproj = cpn; 199 } 200 } 201 } 202 break; 203 } 204 case TypeFunc::I_O: 205 if (pn->_is_io_use) 206 _ioproj_catchall = pn; 207 else 208 _ioproj_fallthrough = pn; 209 break; 210 case TypeFunc::Memory: 211 if (pn->_is_io_use) 212 _memproj_catchall = pn; 213 else 214 _memproj_fallthrough = pn; 215 break; 216 case TypeFunc::Parms: 217 _resproj = pn; 218 break; 219 default: 220 assert(false, "unexpected projection from allocation node."); 221 } 222 } 223 224 } 225 226 // Eliminate a card mark sequence. p2x is a ConvP2XNode 227 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) { 228 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required"); 229 if (!UseG1GC) { 230 // vanilla/CMS post barrier 231 Node *shift = p2x->unique_out(); 232 Node *addp = shift->unique_out(); 233 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) { 234 Node *mem = addp->last_out(j); 235 if (UseCondCardMark && mem->is_Load()) { 236 assert(mem->Opcode() == Op_LoadB, "unexpected code shape"); 237 // The load is checking if the card has been written so 238 // replace it with zero to fold the test. 239 _igvn.replace_node(mem, intcon(0)); 240 continue; 241 } 242 assert(mem->is_Store(), "store required"); 243 _igvn.replace_node(mem, mem->in(MemNode::Memory)); 244 } 245 } else { 246 // G1 pre/post barriers 247 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes"); 248 // It could be only one user, URShift node, in Object.clone() intrinsic 249 // but the new allocation is passed to arraycopy stub and it could not 250 // be scalar replaced. So we don't check the case. 251 252 // An other case of only one user (Xor) is when the value check for NULL 253 // in G1 post barrier is folded after CCP so the code which used URShift 254 // is removed. 255 256 // Take Region node before eliminating post barrier since it also 257 // eliminates CastP2X node when it has only one user. 258 Node* this_region = p2x->in(0); 259 assert(this_region != NULL, ""); 260 261 // Remove G1 post barrier. 262 263 // Search for CastP2X->Xor->URShift->Cmp path which 264 // checks if the store done to a different from the value's region. 265 // And replace Cmp with #0 (false) to collapse G1 post barrier. 266 Node* xorx = p2x->find_out_with(Op_XorX); 267 if (xorx != NULL) { 268 Node* shift = xorx->unique_out(); 269 Node* cmpx = shift->unique_out(); 270 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() && 271 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne, 272 "missing region check in G1 post barrier"); 273 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 274 275 // Remove G1 pre barrier. 276 277 // Search "if (marking != 0)" check and set it to "false". 278 // There is no G1 pre barrier if previous stored value is NULL 279 // (for example, after initialization). 280 if (this_region->is_Region() && this_region->req() == 3) { 281 int ind = 1; 282 if (!this_region->in(ind)->is_IfFalse()) { 283 ind = 2; 284 } 285 if (this_region->in(ind)->is_IfFalse()) { 286 Node* bol = this_region->in(ind)->in(0)->in(1); 287 assert(bol->is_Bool(), ""); 288 cmpx = bol->in(1); 289 if (bol->as_Bool()->_test._test == BoolTest::ne && 290 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) && 291 cmpx->in(1)->is_Load()) { 292 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address); 293 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + 294 SATBMarkQueue::byte_offset_of_active()); 295 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() && 296 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 297 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) { 298 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 299 } 300 } 301 } 302 } 303 } else { 304 assert(!GraphKit::use_ReduceInitialCardMarks(), "can only happen with card marking"); 305 // This is a G1 post barrier emitted by the Object.clone() intrinsic. 306 // Search for the CastP2X->URShiftX->AddP->LoadB->Cmp path which checks if the card 307 // is marked as young_gen and replace the Cmp with 0 (false) to collapse the barrier. 308 Node* shift = p2x->find_out_with(Op_URShiftX); 309 assert(shift != NULL, "missing G1 post barrier"); 310 Node* addp = shift->unique_out(); 311 Node* load = addp->find_out_with(Op_LoadB); 312 assert(load != NULL, "missing G1 post barrier"); 313 Node* cmpx = load->unique_out(); 314 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() && 315 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne, 316 "missing card value check in G1 post barrier"); 317 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 318 // There is no G1 pre barrier in this case 319 } 320 // Now CastP2X can be removed since it is used only on dead path 321 // which currently still alive until igvn optimize it. 322 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, ""); 323 _igvn.replace_node(p2x, top()); 324 } 325 } 326 327 // Search for a memory operation for the specified memory slice. 328 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) { 329 Node *orig_mem = mem; 330 Node *alloc_mem = alloc->in(TypeFunc::Memory); 331 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr(); 332 while (true) { 333 if (mem == alloc_mem || mem == start_mem ) { 334 return mem; // hit one of our sentinels 335 } else if (mem->is_MergeMem()) { 336 mem = mem->as_MergeMem()->memory_at(alias_idx); 337 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { 338 Node *in = mem->in(0); 339 // we can safely skip over safepoints, calls, locks and membars because we 340 // already know that the object is safe to eliminate. 341 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { 342 return in; 343 } else if (in->is_Call()) { 344 CallNode *call = in->as_Call(); 345 if (call->may_modify(tinst, phase)) { 346 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape"); 347 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) { 348 return in; 349 } 350 } 351 mem = in->in(TypeFunc::Memory); 352 } else if (in->is_MemBar()) { 353 ArrayCopyNode* ac = NULL; 354 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) { 355 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone"); 356 return ac; 357 } 358 mem = in->in(TypeFunc::Memory); 359 } else { 360 assert(false, "unexpected projection"); 361 } 362 } else if (mem->is_Store()) { 363 const TypePtr* atype = mem->as_Store()->adr_type(); 364 int adr_idx = phase->C->get_alias_index(atype); 365 if (adr_idx == alias_idx) { 366 assert(atype->isa_oopptr(), "address type must be oopptr"); 367 int adr_offset = atype->offset(); 368 uint adr_iid = atype->is_oopptr()->instance_id(); 369 // Array elements references have the same alias_idx 370 // but different offset and different instance_id. 371 if (adr_offset == offset && adr_iid == alloc->_idx) 372 return mem; 373 } else { 374 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw"); 375 } 376 mem = mem->in(MemNode::Memory); 377 } else if (mem->is_ClearArray()) { 378 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) { 379 // Can not bypass initialization of the instance 380 // we are looking. 381 debug_only(intptr_t offset;) 382 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity"); 383 InitializeNode* init = alloc->as_Allocate()->initialization(); 384 // We are looking for stored value, return Initialize node 385 // or memory edge from Allocate node. 386 if (init != NULL) 387 return init; 388 else 389 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers). 390 } 391 // Otherwise skip it (the call updated 'mem' value). 392 } else if (mem->Opcode() == Op_SCMemProj) { 393 mem = mem->in(0); 394 Node* adr = NULL; 395 if (mem->is_LoadStore()) { 396 adr = mem->in(MemNode::Address); 397 } else { 398 assert(mem->Opcode() == Op_EncodeISOArray || 399 mem->Opcode() == Op_StrCompressedCopy, "sanity"); 400 adr = mem->in(3); // Destination array 401 } 402 const TypePtr* atype = adr->bottom_type()->is_ptr(); 403 int adr_idx = phase->C->get_alias_index(atype); 404 if (adr_idx == alias_idx) { 405 DEBUG_ONLY(mem->dump();) 406 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 407 return NULL; 408 } 409 mem = mem->in(MemNode::Memory); 410 } else if (mem->Opcode() == Op_StrInflatedCopy) { 411 Node* adr = mem->in(3); // Destination array 412 const TypePtr* atype = adr->bottom_type()->is_ptr(); 413 int adr_idx = phase->C->get_alias_index(atype); 414 if (adr_idx == alias_idx) { 415 DEBUG_ONLY(mem->dump();) 416 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field"); 417 return NULL; 418 } 419 mem = mem->in(MemNode::Memory); 420 } else { 421 return mem; 422 } 423 assert(mem != orig_mem, "dead memory loop"); 424 } 425 } 426 427 // Generate loads from source of the arraycopy for fields of 428 // destination needed at a deoptimization point 429 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, BasicType ft, const Type *ftype, AllocateNode *alloc) { 430 BasicType bt = ft; 431 const Type *type = ftype; 432 if (ft == T_NARROWOOP) { 433 bt = T_OBJECT; 434 type = ftype->make_oopptr(); 435 } 436 Node* res = NULL; 437 if (ac->is_clonebasic()) { 438 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base); 439 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset))); 440 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset); 441 Node* m = ac->in(TypeFunc::Memory); 442 while (m->is_MergeMem()) { 443 m = m->as_MergeMem()->memory_at(C->get_alias_index(adr_type)); 444 if (m->is_Proj() && m->in(0)->is_MemBar()) { 445 m = m->in(0)->in(TypeFunc::Memory); 446 } 447 } 448 res = LoadNode::make(_igvn, ctl, m, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned); 449 } else { 450 if (ac->modifies(offset, offset, &_igvn, true)) { 451 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result"); 452 uint shift = exact_log2(type2aelembytes(bt)); 453 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 454 #ifdef _LP64 455 diff = _igvn.transform(new ConvI2LNode(diff)); 456 #endif 457 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift))); 458 459 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff)); 460 Node* base = ac->in(ArrayCopyNode::Src); 461 Node* adr = _igvn.transform(new AddPNode(base, base, off)); 462 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset); 463 Node* m = ac->in(TypeFunc::Memory); 464 res = LoadNode::make(_igvn, ctl, m, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned); 465 } 466 } 467 if (res != NULL) { 468 res = _igvn.transform(res); 469 if (ftype->isa_narrowoop()) { 470 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes 471 res = _igvn.transform(new EncodePNode(res, ftype)); 472 } 473 return res; 474 } 475 return NULL; 476 } 477 478 // 479 // Given a Memory Phi, compute a value Phi containing the values from stores 480 // on the input paths. 481 // Note: this function is recursive, its depth is limited by the "level" argument 482 // Returns the computed Phi, or NULL if it cannot compute it. 483 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) { 484 assert(mem->is_Phi(), "sanity"); 485 int alias_idx = C->get_alias_index(adr_t); 486 int offset = adr_t->offset(); 487 int instance_id = adr_t->instance_id(); 488 489 // Check if an appropriate value phi already exists. 490 Node* region = mem->in(0); 491 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { 492 Node* phi = region->fast_out(k); 493 if (phi->is_Phi() && phi != mem && 494 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) { 495 return phi; 496 } 497 } 498 // Check if an appropriate new value phi already exists. 499 Node* new_phi = value_phis->find(mem->_idx); 500 if (new_phi != NULL) 501 return new_phi; 502 503 if (level <= 0) { 504 return NULL; // Give up: phi tree too deep 505 } 506 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 507 Node *alloc_mem = alloc->in(TypeFunc::Memory); 508 509 uint length = mem->req(); 510 GrowableArray <Node *> values(length, length, NULL, false); 511 512 // create a new Phi for the value 513 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset); 514 transform_later(phi); 515 value_phis->push(phi, mem->_idx); 516 517 for (uint j = 1; j < length; j++) { 518 Node *in = mem->in(j); 519 if (in == NULL || in->is_top()) { 520 values.at_put(j, in); 521 } else { 522 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); 523 if (val == start_mem || val == alloc_mem) { 524 // hit a sentinel, return appropriate 0 value 525 values.at_put(j, _igvn.zerocon(ft)); 526 continue; 527 } 528 if (val->is_Initialize()) { 529 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 530 } 531 if (val == NULL) { 532 return NULL; // can't find a value on this path 533 } 534 if (val == mem) { 535 values.at_put(j, mem); 536 } else if (val->is_Store()) { 537 values.at_put(j, val->in(MemNode::ValueIn)); 538 } else if(val->is_Proj() && val->in(0) == alloc) { 539 values.at_put(j, _igvn.zerocon(ft)); 540 } else if (val->is_Phi()) { 541 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); 542 if (val == NULL) { 543 return NULL; 544 } 545 values.at_put(j, val); 546 } else if (val->Opcode() == Op_SCMemProj) { 547 assert(val->in(0)->is_LoadStore() || 548 val->in(0)->Opcode() == Op_EncodeISOArray || 549 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity"); 550 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 551 return NULL; 552 } else if (val->is_ArrayCopy()) { 553 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), ft, phi_type, alloc); 554 if (res == NULL) { 555 return NULL; 556 } 557 values.at_put(j, res); 558 } else { 559 #ifdef ASSERT 560 val->dump(); 561 assert(false, "unknown node on this path"); 562 #endif 563 return NULL; // unknown node on this path 564 } 565 } 566 } 567 // Set Phi's inputs 568 for (uint j = 1; j < length; j++) { 569 if (values.at(j) == mem) { 570 phi->init_req(j, phi); 571 } else { 572 phi->init_req(j, values.at(j)); 573 } 574 } 575 return phi; 576 } 577 578 // Search the last value stored into the object's field. 579 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) { 580 assert(adr_t->is_known_instance_field(), "instance required"); 581 int instance_id = adr_t->instance_id(); 582 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 583 584 int alias_idx = C->get_alias_index(adr_t); 585 int offset = adr_t->offset(); 586 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 587 Node *alloc_ctrl = alloc->in(TypeFunc::Control); 588 Node *alloc_mem = alloc->in(TypeFunc::Memory); 589 Arena *a = Thread::current()->resource_area(); 590 VectorSet visited(a); 591 592 593 bool done = sfpt_mem == alloc_mem; 594 Node *mem = sfpt_mem; 595 while (!done) { 596 if (visited.test_set(mem->_idx)) { 597 return NULL; // found a loop, give up 598 } 599 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 600 if (mem == start_mem || mem == alloc_mem) { 601 done = true; // hit a sentinel, return appropriate 0 value 602 } else if (mem->is_Initialize()) { 603 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 604 if (mem == NULL) { 605 done = true; // Something go wrong. 606 } else if (mem->is_Store()) { 607 const TypePtr* atype = mem->as_Store()->adr_type(); 608 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 609 done = true; 610 } 611 } else if (mem->is_Store()) { 612 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 613 assert(atype != NULL, "address type must be oopptr"); 614 assert(C->get_alias_index(atype) == alias_idx && 615 atype->is_known_instance_field() && atype->offset() == offset && 616 atype->instance_id() == instance_id, "store is correct memory slice"); 617 done = true; 618 } else if (mem->is_Phi()) { 619 // try to find a phi's unique input 620 Node *unique_input = NULL; 621 Node *top = C->top(); 622 for (uint i = 1; i < mem->req(); i++) { 623 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 624 if (n == NULL || n == top || n == mem) { 625 continue; 626 } else if (unique_input == NULL) { 627 unique_input = n; 628 } else if (unique_input != n) { 629 unique_input = top; 630 break; 631 } 632 } 633 if (unique_input != NULL && unique_input != top) { 634 mem = unique_input; 635 } else { 636 done = true; 637 } 638 } else if (mem->is_ArrayCopy()) { 639 done = true; 640 } else { 641 assert(false, "unexpected node"); 642 } 643 } 644 if (mem != NULL) { 645 if (mem == start_mem || mem == alloc_mem) { 646 // hit a sentinel, return appropriate 0 value 647 return _igvn.zerocon(ft); 648 } else if (mem->is_Store()) { 649 return mem->in(MemNode::ValueIn); 650 } else if (mem->is_Phi()) { 651 // attempt to produce a Phi reflecting the values on the input paths of the Phi 652 Node_Stack value_phis(a, 8); 653 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 654 if (phi != NULL) { 655 return phi; 656 } else { 657 // Kill all new Phis 658 while(value_phis.is_nonempty()) { 659 Node* n = value_phis.node(); 660 _igvn.replace_node(n, C->top()); 661 value_phis.pop(); 662 } 663 } 664 } else if (mem->is_ArrayCopy()) { 665 Node* ctl = mem->in(0); 666 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) { 667 // pin the loads in the uncommon trap path 668 ctl = sfpt_ctl; 669 } 670 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, ft, ftype, alloc); 671 } 672 } 673 // Something go wrong. 674 return NULL; 675 } 676 677 // Check the possibility of scalar replacement. 678 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 679 // Scan the uses of the allocation to check for anything that would 680 // prevent us from eliminating it. 681 NOT_PRODUCT( const char* fail_eliminate = NULL; ) 682 DEBUG_ONLY( Node* disq_node = NULL; ) 683 bool can_eliminate = true; 684 685 Node* res = alloc->result_cast(); 686 const TypeOopPtr* res_type = NULL; 687 if (res == NULL) { 688 // All users were eliminated. 689 } else if (!res->is_CheckCastPP()) { 690 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 691 can_eliminate = false; 692 } else { 693 res_type = _igvn.type(res)->isa_oopptr(); 694 if (res_type == NULL) { 695 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 696 can_eliminate = false; 697 } else if (res_type->isa_aryptr()) { 698 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 699 if (length < 0) { 700 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 701 can_eliminate = false; 702 } 703 } 704 } 705 706 if (can_eliminate && res != NULL) { 707 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 708 j < jmax && can_eliminate; j++) { 709 Node* use = res->fast_out(j); 710 711 if (use->is_AddP()) { 712 const TypePtr* addp_type = _igvn.type(use)->is_ptr(); 713 int offset = addp_type->offset(); 714 715 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 716 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) 717 can_eliminate = false; 718 break; 719 } 720 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 721 k < kmax && can_eliminate; k++) { 722 Node* n = use->fast_out(k); 723 if (!n->is_Store() && n->Opcode() != Op_CastP2X && 724 !(n->is_ArrayCopy() && 725 n->as_ArrayCopy()->is_clonebasic() && 726 n->in(ArrayCopyNode::Dest) == use)) { 727 DEBUG_ONLY(disq_node = n;) 728 if (n->is_Load() || n->is_LoadStore()) { 729 NOT_PRODUCT(fail_eliminate = "Field load";) 730 } else { 731 NOT_PRODUCT(fail_eliminate = "Not store field referrence";) 732 } 733 can_eliminate = false; 734 } 735 } 736 } else if (use->is_ArrayCopy() && 737 (use->as_ArrayCopy()->is_arraycopy_validated() || 738 use->as_ArrayCopy()->is_copyof_validated() || 739 use->as_ArrayCopy()->is_copyofrange_validated()) && 740 use->in(ArrayCopyNode::Dest) == res) { 741 // ok to eliminate 742 } else if (use->is_SafePoint()) { 743 SafePointNode* sfpt = use->as_SafePoint(); 744 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 745 // Object is passed as argument. 746 DEBUG_ONLY(disq_node = use;) 747 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 748 can_eliminate = false; 749 } 750 Node* sfptMem = sfpt->memory(); 751 if (sfptMem == NULL || sfptMem->is_top()) { 752 DEBUG_ONLY(disq_node = use;) 753 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) 754 can_eliminate = false; 755 } else { 756 safepoints.append_if_missing(sfpt); 757 } 758 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 759 if (use->is_Phi()) { 760 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 761 NOT_PRODUCT(fail_eliminate = "Object is return value";) 762 } else { 763 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 764 } 765 DEBUG_ONLY(disq_node = use;) 766 } else { 767 if (use->Opcode() == Op_Return) { 768 NOT_PRODUCT(fail_eliminate = "Object is return value";) 769 }else { 770 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 771 } 772 DEBUG_ONLY(disq_node = use;) 773 } 774 can_eliminate = false; 775 } 776 } 777 } 778 779 #ifndef PRODUCT 780 if (PrintEliminateAllocations) { 781 if (can_eliminate) { 782 tty->print("Scalar "); 783 if (res == NULL) 784 alloc->dump(); 785 else 786 res->dump(); 787 } else if (alloc->_is_scalar_replaceable) { 788 tty->print("NotScalar (%s)", fail_eliminate); 789 if (res == NULL) 790 alloc->dump(); 791 else 792 res->dump(); 793 #ifdef ASSERT 794 if (disq_node != NULL) { 795 tty->print(" >>>> "); 796 disq_node->dump(); 797 } 798 #endif /*ASSERT*/ 799 } 800 } 801 #endif 802 return can_eliminate; 803 } 804 805 // Do scalar replacement. 806 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 807 GrowableArray <SafePointNode *> safepoints_done; 808 809 ciKlass* klass = NULL; 810 ciInstanceKlass* iklass = NULL; 811 int nfields = 0; 812 int array_base = 0; 813 int element_size = 0; 814 BasicType basic_elem_type = T_ILLEGAL; 815 ciType* elem_type = NULL; 816 817 Node* res = alloc->result_cast(); 818 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result"); 819 const TypeOopPtr* res_type = NULL; 820 if (res != NULL) { // Could be NULL when there are no users 821 res_type = _igvn.type(res)->isa_oopptr(); 822 } 823 824 if (res != NULL) { 825 klass = res_type->klass(); 826 if (res_type->isa_instptr()) { 827 // find the fields of the class which will be needed for safepoint debug information 828 assert(klass->is_instance_klass(), "must be an instance klass."); 829 iklass = klass->as_instance_klass(); 830 nfields = iklass->nof_nonstatic_fields(); 831 } else { 832 // find the array's elements which will be needed for safepoint debug information 833 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 834 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); 835 elem_type = klass->as_array_klass()->element_type(); 836 basic_elem_type = elem_type->basic_type(); 837 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 838 element_size = type2aelembytes(basic_elem_type); 839 } 840 } 841 // 842 // Process the safepoint uses 843 // 844 while (safepoints.length() > 0) { 845 SafePointNode* sfpt = safepoints.pop(); 846 Node* mem = sfpt->memory(); 847 Node* ctl = sfpt->control(); 848 assert(sfpt->jvms() != NULL, "missed JVMS"); 849 // Fields of scalar objs are referenced only at the end 850 // of regular debuginfo at the last (youngest) JVMS. 851 // Record relative start index. 852 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff()); 853 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, 854 #ifdef ASSERT 855 alloc, 856 #endif 857 first_ind, nfields); 858 sobj->init_req(0, C->root()); 859 transform_later(sobj); 860 861 // Scan object's fields adding an input to the safepoint for each field. 862 for (int j = 0; j < nfields; j++) { 863 intptr_t offset; 864 ciField* field = NULL; 865 if (iklass != NULL) { 866 field = iklass->nonstatic_field_at(j); 867 offset = field->offset(); 868 elem_type = field->type(); 869 basic_elem_type = field->layout_type(); 870 } else { 871 offset = array_base + j * (intptr_t)element_size; 872 } 873 874 const Type *field_type; 875 // The next code is taken from Parse::do_get_xxx(). 876 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { 877 if (!elem_type->is_loaded()) { 878 field_type = TypeInstPtr::BOTTOM; 879 } else if (field != NULL && field->is_static_constant()) { 880 // This can happen if the constant oop is non-perm. 881 ciObject* con = field->constant_value().as_object(); 882 // Do not "join" in the previous type; it doesn't add value, 883 // and may yield a vacuous result if the field is of interface type. 884 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 885 assert(field_type != NULL, "field singleton type must be consistent"); 886 } else { 887 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 888 } 889 if (UseCompressedOops) { 890 field_type = field_type->make_narrowoop(); 891 basic_elem_type = T_NARROWOOP; 892 } 893 } else { 894 field_type = Type::get_const_basic_type(basic_elem_type); 895 } 896 897 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 898 899 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc); 900 if (field_val == NULL) { 901 // We weren't able to find a value for this field, 902 // give up on eliminating this allocation. 903 904 // Remove any extra entries we added to the safepoint. 905 uint last = sfpt->req() - 1; 906 for (int k = 0; k < j; k++) { 907 sfpt->del_req(last--); 908 } 909 _igvn._worklist.push(sfpt); 910 // rollback processed safepoints 911 while (safepoints_done.length() > 0) { 912 SafePointNode* sfpt_done = safepoints_done.pop(); 913 // remove any extra entries we added to the safepoint 914 last = sfpt_done->req() - 1; 915 for (int k = 0; k < nfields; k++) { 916 sfpt_done->del_req(last--); 917 } 918 JVMState *jvms = sfpt_done->jvms(); 919 jvms->set_endoff(sfpt_done->req()); 920 // Now make a pass over the debug information replacing any references 921 // to SafePointScalarObjectNode with the allocated object. 922 int start = jvms->debug_start(); 923 int end = jvms->debug_end(); 924 for (int i = start; i < end; i++) { 925 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 926 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 927 if (scobj->first_index(jvms) == sfpt_done->req() && 928 scobj->n_fields() == (uint)nfields) { 929 assert(scobj->alloc() == alloc, "sanity"); 930 sfpt_done->set_req(i, res); 931 } 932 } 933 } 934 _igvn._worklist.push(sfpt_done); 935 } 936 #ifndef PRODUCT 937 if (PrintEliminateAllocations) { 938 if (field != NULL) { 939 tty->print("=== At SafePoint node %d can't find value of Field: ", 940 sfpt->_idx); 941 field->print(); 942 int field_idx = C->get_alias_index(field_addr_type); 943 tty->print(" (alias_idx=%d)", field_idx); 944 } else { // Array's element 945 tty->print("=== At SafePoint node %d can't find value of array element [%d]", 946 sfpt->_idx, j); 947 } 948 tty->print(", which prevents elimination of: "); 949 if (res == NULL) 950 alloc->dump(); 951 else 952 res->dump(); 953 } 954 #endif 955 return false; 956 } 957 if (UseCompressedOops && field_type->isa_narrowoop()) { 958 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 959 // to be able scalar replace the allocation. 960 if (field_val->is_EncodeP()) { 961 field_val = field_val->in(1); 962 } else { 963 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type())); 964 } 965 } 966 sfpt->add_req(field_val); 967 } 968 JVMState *jvms = sfpt->jvms(); 969 jvms->set_endoff(sfpt->req()); 970 // Now make a pass over the debug information replacing any references 971 // to the allocated object with "sobj" 972 int start = jvms->debug_start(); 973 int end = jvms->debug_end(); 974 sfpt->replace_edges_in_range(res, sobj, start, end); 975 _igvn._worklist.push(sfpt); 976 safepoints_done.append_if_missing(sfpt); // keep it for rollback 977 } 978 return true; 979 } 980 981 // Process users of eliminated allocation. 982 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) { 983 Node* res = alloc->result_cast(); 984 if (res != NULL) { 985 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 986 Node *use = res->last_out(j); 987 uint oc1 = res->outcnt(); 988 989 if (use->is_AddP()) { 990 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 991 Node *n = use->last_out(k); 992 uint oc2 = use->outcnt(); 993 if (n->is_Store()) { 994 #ifdef ASSERT 995 // Verify that there is no dependent MemBarVolatile nodes, 996 // they should be removed during IGVN, see MemBarNode::Ideal(). 997 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); 998 p < pmax; p++) { 999 Node* mb = n->fast_out(p); 1000 assert(mb->is_Initialize() || !mb->is_MemBar() || 1001 mb->req() <= MemBarNode::Precedent || 1002 mb->in(MemBarNode::Precedent) != n, 1003 "MemBarVolatile should be eliminated for non-escaping object"); 1004 } 1005 #endif 1006 _igvn.replace_node(n, n->in(MemNode::Memory)); 1007 } else if (n->is_ArrayCopy()) { 1008 // Disconnect ArrayCopy node 1009 ArrayCopyNode* ac = n->as_ArrayCopy(); 1010 assert(ac->is_clonebasic(), "unexpected array copy kind"); 1011 Node* ctl_proj = ac->proj_out(TypeFunc::Control); 1012 Node* mem_proj = ac->proj_out(TypeFunc::Memory); 1013 if (ctl_proj != NULL) { 1014 _igvn.replace_node(ctl_proj, n->in(0)); 1015 } 1016 if (mem_proj != NULL) { 1017 _igvn.replace_node(mem_proj, n->in(TypeFunc::Memory)); 1018 } 1019 } else { 1020 eliminate_card_mark(n); 1021 } 1022 k -= (oc2 - use->outcnt()); 1023 } 1024 } else if (use->is_ArrayCopy()) { 1025 // Disconnect ArrayCopy node 1026 ArrayCopyNode* ac = use->as_ArrayCopy(); 1027 assert(ac->is_arraycopy_validated() || 1028 ac->is_copyof_validated() || 1029 ac->is_copyofrange_validated(), "unsupported"); 1030 CallProjections callprojs; 1031 ac->extract_projections(&callprojs, true); 1032 1033 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O)); 1034 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory)); 1035 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control)); 1036 1037 // Set control to top. IGVN will remove the remaining projections 1038 ac->set_req(0, top()); 1039 ac->replace_edge(res, top()); 1040 1041 // Disconnect src right away: it can help find new 1042 // opportunities for allocation elimination 1043 Node* src = ac->in(ArrayCopyNode::Src); 1044 ac->replace_edge(src, top()); 1045 if (src->outcnt() == 0) { 1046 _igvn.remove_dead_node(src); 1047 } 1048 1049 _igvn._worklist.push(ac); 1050 } else { 1051 eliminate_card_mark(use); 1052 } 1053 j -= (oc1 - res->outcnt()); 1054 } 1055 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 1056 _igvn.remove_dead_node(res); 1057 } 1058 1059 // 1060 // Process other users of allocation's projections 1061 // 1062 if (_resproj != NULL && _resproj->outcnt() != 0) { 1063 // First disconnect stores captured by Initialize node. 1064 // If Initialize node is eliminated first in the following code, 1065 // it will kill such stores and DUIterator_Last will assert. 1066 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) { 1067 Node *use = _resproj->fast_out(j); 1068 if (use->is_AddP()) { 1069 // raw memory addresses used only by the initialization 1070 _igvn.replace_node(use, C->top()); 1071 --j; --jmax; 1072 } 1073 } 1074 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { 1075 Node *use = _resproj->last_out(j); 1076 uint oc1 = _resproj->outcnt(); 1077 if (use->is_Initialize()) { 1078 // Eliminate Initialize node. 1079 InitializeNode *init = use->as_Initialize(); 1080 assert(init->outcnt() <= 2, "only a control and memory projection expected"); 1081 Node *ctrl_proj = init->proj_out(TypeFunc::Control); 1082 if (ctrl_proj != NULL) { 1083 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection"); 1084 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj); 1085 } 1086 Node *mem_proj = init->proj_out(TypeFunc::Memory); 1087 if (mem_proj != NULL) { 1088 Node *mem = init->in(TypeFunc::Memory); 1089 #ifdef ASSERT 1090 if (mem->is_MergeMem()) { 1091 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection"); 1092 } else { 1093 assert(mem == _memproj_fallthrough, "allocation memory projection"); 1094 } 1095 #endif 1096 _igvn.replace_node(mem_proj, mem); 1097 } 1098 } else { 1099 assert(false, "only Initialize or AddP expected"); 1100 } 1101 j -= (oc1 - _resproj->outcnt()); 1102 } 1103 } 1104 if (_fallthroughcatchproj != NULL) { 1105 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); 1106 } 1107 if (_memproj_fallthrough != NULL) { 1108 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); 1109 } 1110 if (_memproj_catchall != NULL) { 1111 _igvn.replace_node(_memproj_catchall, C->top()); 1112 } 1113 if (_ioproj_fallthrough != NULL) { 1114 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); 1115 } 1116 if (_ioproj_catchall != NULL) { 1117 _igvn.replace_node(_ioproj_catchall, C->top()); 1118 } 1119 if (_catchallcatchproj != NULL) { 1120 _igvn.replace_node(_catchallcatchproj, C->top()); 1121 } 1122 } 1123 1124 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 1125 // Don't do scalar replacement if the frame can be popped by JVMTI: 1126 // if reallocation fails during deoptimization we'll pop all 1127 // interpreter frames for this compiled frame and that won't play 1128 // nice with JVMTI popframe. 1129 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) { 1130 return false; 1131 } 1132 Node* klass = alloc->in(AllocateNode::KlassNode); 1133 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); 1134 Node* res = alloc->result_cast(); 1135 // Eliminate boxing allocations which are not used 1136 // regardless scalar replacable status. 1137 bool boxing_alloc = C->eliminate_boxing() && 1138 tklass->klass()->is_instance_klass() && 1139 tklass->klass()->as_instance_klass()->is_box_klass(); 1140 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) { 1141 return false; 1142 } 1143 1144 extract_call_projections(alloc); 1145 1146 GrowableArray <SafePointNode *> safepoints; 1147 if (!can_eliminate_allocation(alloc, safepoints)) { 1148 return false; 1149 } 1150 1151 if (!alloc->_is_scalar_replaceable) { 1152 assert(res == NULL, "sanity"); 1153 // We can only eliminate allocation if all debug info references 1154 // are already replaced with SafePointScalarObject because 1155 // we can't search for a fields value without instance_id. 1156 if (safepoints.length() > 0) { 1157 return false; 1158 } 1159 } 1160 1161 if (!scalar_replacement(alloc, safepoints)) { 1162 return false; 1163 } 1164 1165 CompileLog* log = C->log(); 1166 if (log != NULL) { 1167 log->head("eliminate_allocation type='%d'", 1168 log->identify(tklass->klass())); 1169 JVMState* p = alloc->jvms(); 1170 while (p != NULL) { 1171 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1172 p = p->caller(); 1173 } 1174 log->tail("eliminate_allocation"); 1175 } 1176 1177 process_users_of_allocation(alloc); 1178 1179 #ifndef PRODUCT 1180 if (PrintEliminateAllocations) { 1181 if (alloc->is_AllocateArray()) 1182 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1183 else 1184 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1185 } 1186 #endif 1187 1188 return true; 1189 } 1190 1191 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) { 1192 // EA should remove all uses of non-escaping boxing node. 1193 if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) { 1194 return false; 1195 } 1196 1197 assert(boxing->result_cast() == NULL, "unexpected boxing node result"); 1198 1199 extract_call_projections(boxing); 1200 1201 const TypeTuple* r = boxing->tf()->range(); 1202 assert(r->cnt() > TypeFunc::Parms, "sanity"); 1203 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr(); 1204 assert(t != NULL, "sanity"); 1205 1206 CompileLog* log = C->log(); 1207 if (log != NULL) { 1208 log->head("eliminate_boxing type='%d'", 1209 log->identify(t->klass())); 1210 JVMState* p = boxing->jvms(); 1211 while (p != NULL) { 1212 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1213 p = p->caller(); 1214 } 1215 log->tail("eliminate_boxing"); 1216 } 1217 1218 process_users_of_allocation(boxing); 1219 1220 #ifndef PRODUCT 1221 if (PrintEliminateAllocations) { 1222 tty->print("++++ Eliminated: %d ", boxing->_idx); 1223 boxing->method()->print_short_name(tty); 1224 tty->cr(); 1225 } 1226 #endif 1227 1228 return true; 1229 } 1230 1231 //---------------------------set_eden_pointers------------------------- 1232 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { 1233 if (UseTLAB) { // Private allocation: load from TLS 1234 Node* thread = transform_later(new ThreadLocalNode()); 1235 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); 1236 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); 1237 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); 1238 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); 1239 } else { // Shared allocation: load from globals 1240 CollectedHeap* ch = Universe::heap(); 1241 address top_adr = (address)ch->top_addr(); 1242 address end_adr = (address)ch->end_addr(); 1243 eden_top_adr = makecon(TypeRawPtr::make(top_adr)); 1244 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); 1245 } 1246 } 1247 1248 1249 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 1250 Node* adr = basic_plus_adr(base, offset); 1251 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 1252 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered); 1253 transform_later(value); 1254 return value; 1255 } 1256 1257 1258 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 1259 Node* adr = basic_plus_adr(base, offset); 1260 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered); 1261 transform_later(mem); 1262 return mem; 1263 } 1264 1265 //============================================================================= 1266 // 1267 // A L L O C A T I O N 1268 // 1269 // Allocation attempts to be fast in the case of frequent small objects. 1270 // It breaks down like this: 1271 // 1272 // 1) Size in doublewords is computed. This is a constant for objects and 1273 // variable for most arrays. Doubleword units are used to avoid size 1274 // overflow of huge doubleword arrays. We need doublewords in the end for 1275 // rounding. 1276 // 1277 // 2) Size is checked for being 'too large'. Too-large allocations will go 1278 // the slow path into the VM. The slow path can throw any required 1279 // exceptions, and does all the special checks for very large arrays. The 1280 // size test can constant-fold away for objects. For objects with 1281 // finalizers it constant-folds the otherway: you always go slow with 1282 // finalizers. 1283 // 1284 // 3) If NOT using TLABs, this is the contended loop-back point. 1285 // Load-Locked the heap top. If using TLABs normal-load the heap top. 1286 // 1287 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 1288 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 1289 // "size*8" we always enter the VM, where "largish" is a constant picked small 1290 // enough that there's always space between the eden max and 4Gig (old space is 1291 // there so it's quite large) and large enough that the cost of entering the VM 1292 // is dwarfed by the cost to initialize the space. 1293 // 1294 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 1295 // down. If contended, repeat at step 3. If using TLABs normal-store 1296 // adjusted heap top back down; there is no contention. 1297 // 1298 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 1299 // fields. 1300 // 1301 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 1302 // oop flavor. 1303 // 1304 //============================================================================= 1305 // FastAllocateSizeLimit value is in DOUBLEWORDS. 1306 // Allocations bigger than this always go the slow route. 1307 // This value must be small enough that allocation attempts that need to 1308 // trigger exceptions go the slow route. Also, it must be small enough so 1309 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 1310 //=============================================================================j// 1311 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 1312 // The allocator will coalesce int->oop copies away. See comment in 1313 // coalesce.cpp about how this works. It depends critically on the exact 1314 // code shape produced here, so if you are changing this code shape 1315 // make sure the GC info for the heap-top is correct in and around the 1316 // slow-path call. 1317 // 1318 1319 void PhaseMacroExpand::expand_allocate_common( 1320 AllocateNode* alloc, // allocation node to be expanded 1321 Node* length, // array length for an array allocation 1322 const TypeFunc* slow_call_type, // Type of slow call 1323 address slow_call_address // Address of slow call 1324 ) 1325 { 1326 1327 Node* ctrl = alloc->in(TypeFunc::Control); 1328 Node* mem = alloc->in(TypeFunc::Memory); 1329 Node* i_o = alloc->in(TypeFunc::I_O); 1330 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 1331 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1332 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 1333 1334 assert(ctrl != NULL, "must have control"); 1335 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 1336 // they will not be used if "always_slow" is set 1337 enum { slow_result_path = 1, fast_result_path = 2 }; 1338 Node *result_region = NULL; 1339 Node *result_phi_rawmem = NULL; 1340 Node *result_phi_rawoop = NULL; 1341 Node *result_phi_i_o = NULL; 1342 1343 // The initial slow comparison is a size check, the comparison 1344 // we want to do is a BoolTest::gt 1345 bool always_slow = false; 1346 int tv = _igvn.find_int_con(initial_slow_test, -1); 1347 if (tv >= 0) { 1348 always_slow = (tv == 1); 1349 initial_slow_test = NULL; 1350 } else { 1351 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 1352 } 1353 1354 if (C->env()->dtrace_alloc_probes() || 1355 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc())) { 1356 // Force slow-path allocation 1357 always_slow = true; 1358 initial_slow_test = NULL; 1359 } 1360 1361 1362 enum { too_big_or_final_path = 1, need_gc_path = 2 }; 1363 Node *slow_region = NULL; 1364 Node *toobig_false = ctrl; 1365 1366 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); 1367 // generate the initial test if necessary 1368 if (initial_slow_test != NULL ) { 1369 slow_region = new RegionNode(3); 1370 1371 // Now make the initial failure test. Usually a too-big test but 1372 // might be a TRUE for finalizers or a fancy class check for 1373 // newInstance0. 1374 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1375 transform_later(toobig_iff); 1376 // Plug the failing-too-big test into the slow-path region 1377 Node *toobig_true = new IfTrueNode( toobig_iff ); 1378 transform_later(toobig_true); 1379 slow_region ->init_req( too_big_or_final_path, toobig_true ); 1380 toobig_false = new IfFalseNode( toobig_iff ); 1381 transform_later(toobig_false); 1382 } else { // No initial test, just fall into next case 1383 toobig_false = ctrl; 1384 debug_only(slow_region = NodeSentinel); 1385 } 1386 1387 Node *slow_mem = mem; // save the current memory state for slow path 1388 // generate the fast allocation code unless we know that the initial test will always go slow 1389 if (!always_slow) { 1390 // Fast path modifies only raw memory. 1391 if (mem->is_MergeMem()) { 1392 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1393 } 1394 1395 Node* eden_top_adr; 1396 Node* eden_end_adr; 1397 1398 set_eden_pointers(eden_top_adr, eden_end_adr); 1399 1400 // Load Eden::end. Loop invariant and hoisted. 1401 // 1402 // Note: We set the control input on "eden_end" and "old_eden_top" when using 1403 // a TLAB to work around a bug where these values were being moved across 1404 // a safepoint. These are not oops, so they cannot be include in the oop 1405 // map, but they can be changed by a GC. The proper way to fix this would 1406 // be to set the raw memory state when generating a SafepointNode. However 1407 // this will require extensive changes to the loop optimization in order to 1408 // prevent a degradation of the optimization. 1409 // See comment in memnode.hpp, around line 227 in class LoadPNode. 1410 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 1411 1412 // allocate the Region and Phi nodes for the result 1413 result_region = new RegionNode(3); 1414 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1415 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM); 1416 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch 1417 1418 // We need a Region for the loop-back contended case. 1419 enum { fall_in_path = 1, contended_loopback_path = 2 }; 1420 Node *contended_region; 1421 Node *contended_phi_rawmem; 1422 if (UseTLAB) { 1423 contended_region = toobig_false; 1424 contended_phi_rawmem = mem; 1425 } else { 1426 contended_region = new RegionNode(3); 1427 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1428 // Now handle the passing-too-big test. We fall into the contended 1429 // loop-back merge point. 1430 contended_region ->init_req(fall_in_path, toobig_false); 1431 contended_phi_rawmem->init_req(fall_in_path, mem); 1432 transform_later(contended_region); 1433 transform_later(contended_phi_rawmem); 1434 } 1435 1436 // Load(-locked) the heap top. 1437 // See note above concerning the control input when using a TLAB 1438 Node *old_eden_top = UseTLAB 1439 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered) 1440 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire); 1441 1442 transform_later(old_eden_top); 1443 // Add to heap top to get a new heap top 1444 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes); 1445 transform_later(new_eden_top); 1446 // Check for needing a GC; compare against heap end 1447 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end); 1448 transform_later(needgc_cmp); 1449 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge); 1450 transform_later(needgc_bol); 1451 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); 1452 transform_later(needgc_iff); 1453 1454 // Plug the failing-heap-space-need-gc test into the slow-path region 1455 Node *needgc_true = new IfTrueNode(needgc_iff); 1456 transform_later(needgc_true); 1457 if (initial_slow_test) { 1458 slow_region->init_req(need_gc_path, needgc_true); 1459 // This completes all paths into the slow merge point 1460 transform_later(slow_region); 1461 } else { // No initial slow path needed! 1462 // Just fall from the need-GC path straight into the VM call. 1463 slow_region = needgc_true; 1464 } 1465 // No need for a GC. Setup for the Store-Conditional 1466 Node *needgc_false = new IfFalseNode(needgc_iff); 1467 transform_later(needgc_false); 1468 1469 // Grab regular I/O before optional prefetch may change it. 1470 // Slow-path does no I/O so just set it to the original I/O. 1471 result_phi_i_o->init_req(slow_result_path, i_o); 1472 1473 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem, 1474 old_eden_top, new_eden_top, length); 1475 1476 // Name successful fast-path variables 1477 Node* fast_oop = old_eden_top; 1478 Node* fast_oop_ctrl; 1479 Node* fast_oop_rawmem; 1480 1481 // Store (-conditional) the modified eden top back down. 1482 // StorePConditional produces flags for a test PLUS a modified raw 1483 // memory state. 1484 if (UseTLAB) { 1485 Node* store_eden_top = 1486 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1487 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered); 1488 transform_later(store_eden_top); 1489 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path 1490 fast_oop_rawmem = store_eden_top; 1491 } else { 1492 Node* store_eden_top = 1493 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1494 new_eden_top, fast_oop/*old_eden_top*/); 1495 transform_later(store_eden_top); 1496 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne); 1497 transform_later(contention_check); 1498 store_eden_top = new SCMemProjNode(store_eden_top); 1499 transform_later(store_eden_top); 1500 1501 // If not using TLABs, check to see if there was contention. 1502 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN); 1503 transform_later(contention_iff); 1504 Node *contention_true = new IfTrueNode(contention_iff); 1505 transform_later(contention_true); 1506 // If contention, loopback and try again. 1507 contended_region->init_req(contended_loopback_path, contention_true); 1508 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top); 1509 1510 // Fast-path succeeded with no contention! 1511 Node *contention_false = new IfFalseNode(contention_iff); 1512 transform_later(contention_false); 1513 fast_oop_ctrl = contention_false; 1514 1515 // Bump total allocated bytes for this thread 1516 Node* thread = new ThreadLocalNode(); 1517 transform_later(thread); 1518 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread, 1519 in_bytes(JavaThread::allocated_bytes_offset())); 1520 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1521 0, TypeLong::LONG, T_LONG); 1522 #ifdef _LP64 1523 Node* alloc_size = size_in_bytes; 1524 #else 1525 Node* alloc_size = new ConvI2LNode(size_in_bytes); 1526 transform_later(alloc_size); 1527 #endif 1528 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size); 1529 transform_later(new_alloc_bytes); 1530 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1531 0, new_alloc_bytes, T_LONG); 1532 } 1533 1534 InitializeNode* init = alloc->initialization(); 1535 fast_oop_rawmem = initialize_object(alloc, 1536 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1537 klass_node, length, size_in_bytes); 1538 1539 // If initialization is performed by an array copy, any required 1540 // MemBarStoreStore was already added. If the object does not 1541 // escape no need for a MemBarStoreStore. If the object does not 1542 // escape in its initializer and memory barrier (MemBarStoreStore or 1543 // stronger) is already added at exit of initializer, also no need 1544 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore 1545 // so that stores that initialize this object can't be reordered 1546 // with a subsequent store that makes this object accessible by 1547 // other threads. 1548 // Other threads include java threads and JVM internal threads 1549 // (for example concurrent GC threads). Current concurrent GC 1550 // implementation: CMS and G1 will not scan newly created object, 1551 // so it's safe to skip storestore barrier when allocation does 1552 // not escape. 1553 if (!alloc->does_not_escape_thread() && 1554 !alloc->is_allocation_MemBar_redundant() && 1555 (init == NULL || !init->is_complete_with_arraycopy())) { 1556 if (init == NULL || init->req() < InitializeNode::RawStores) { 1557 // No InitializeNode or no stores captured by zeroing 1558 // elimination. Simply add the MemBarStoreStore after object 1559 // initialization. 1560 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1561 transform_later(mb); 1562 1563 mb->init_req(TypeFunc::Memory, fast_oop_rawmem); 1564 mb->init_req(TypeFunc::Control, fast_oop_ctrl); 1565 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control); 1566 transform_later(fast_oop_ctrl); 1567 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory); 1568 transform_later(fast_oop_rawmem); 1569 } else { 1570 // Add the MemBarStoreStore after the InitializeNode so that 1571 // all stores performing the initialization that were moved 1572 // before the InitializeNode happen before the storestore 1573 // barrier. 1574 1575 Node* init_ctrl = init->proj_out(TypeFunc::Control); 1576 Node* init_mem = init->proj_out(TypeFunc::Memory); 1577 1578 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1579 transform_later(mb); 1580 1581 Node* ctrl = new ProjNode(init,TypeFunc::Control); 1582 transform_later(ctrl); 1583 Node* mem = new ProjNode(init,TypeFunc::Memory); 1584 transform_later(mem); 1585 1586 // The MemBarStoreStore depends on control and memory coming 1587 // from the InitializeNode 1588 mb->init_req(TypeFunc::Memory, mem); 1589 mb->init_req(TypeFunc::Control, ctrl); 1590 1591 ctrl = new ProjNode(mb,TypeFunc::Control); 1592 transform_later(ctrl); 1593 mem = new ProjNode(mb,TypeFunc::Memory); 1594 transform_later(mem); 1595 1596 // All nodes that depended on the InitializeNode for control 1597 // and memory must now depend on the MemBarNode that itself 1598 // depends on the InitializeNode 1599 if (init_ctrl != NULL) { 1600 _igvn.replace_node(init_ctrl, ctrl); 1601 } 1602 if (init_mem != NULL) { 1603 _igvn.replace_node(init_mem, mem); 1604 } 1605 } 1606 } 1607 1608 if (C->env()->dtrace_extended_probes()) { 1609 // Slow-path call 1610 int size = TypeFunc::Parms + 2; 1611 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1612 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), 1613 "dtrace_object_alloc", 1614 TypeRawPtr::BOTTOM); 1615 1616 // Get base of thread-local storage area 1617 Node* thread = new ThreadLocalNode(); 1618 transform_later(thread); 1619 1620 call->init_req(TypeFunc::Parms+0, thread); 1621 call->init_req(TypeFunc::Parms+1, fast_oop); 1622 call->init_req(TypeFunc::Control, fast_oop_ctrl); 1623 call->init_req(TypeFunc::I_O , top()); // does no i/o 1624 call->init_req(TypeFunc::Memory , fast_oop_rawmem); 1625 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1626 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1627 transform_later(call); 1628 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control); 1629 transform_later(fast_oop_ctrl); 1630 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory); 1631 transform_later(fast_oop_rawmem); 1632 } 1633 1634 // Plug in the successful fast-path into the result merge point 1635 result_region ->init_req(fast_result_path, fast_oop_ctrl); 1636 result_phi_rawoop->init_req(fast_result_path, fast_oop); 1637 result_phi_i_o ->init_req(fast_result_path, i_o); 1638 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); 1639 } else { 1640 slow_region = ctrl; 1641 result_phi_i_o = i_o; // Rename it to use in the following code. 1642 } 1643 1644 // Generate slow-path call 1645 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address, 1646 OptoRuntime::stub_name(slow_call_address), 1647 alloc->jvms()->bci(), 1648 TypePtr::BOTTOM); 1649 call->init_req( TypeFunc::Control, slow_region ); 1650 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1651 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs 1652 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1653 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1654 1655 call->init_req(TypeFunc::Parms+0, klass_node); 1656 if (length != NULL) { 1657 call->init_req(TypeFunc::Parms+1, length); 1658 } 1659 1660 // Copy debug information and adjust JVMState information, then replace 1661 // allocate node with the call 1662 copy_call_debug_info((CallNode *) alloc, call); 1663 if (!always_slow) { 1664 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1665 } else { 1666 // Hook i_o projection to avoid its elimination during allocation 1667 // replacement (when only a slow call is generated). 1668 call->set_req(TypeFunc::I_O, result_phi_i_o); 1669 } 1670 _igvn.replace_node(alloc, call); 1671 transform_later(call); 1672 1673 // Identify the output projections from the allocate node and 1674 // adjust any references to them. 1675 // The control and io projections look like: 1676 // 1677 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1678 // Allocate Catch 1679 // ^---Proj(io) <-------+ ^---CatchProj(io) 1680 // 1681 // We are interested in the CatchProj nodes. 1682 // 1683 extract_call_projections(call); 1684 1685 // An allocate node has separate memory projections for the uses on 1686 // the control and i_o paths. Replace the control memory projection with 1687 // result_phi_rawmem (unless we are only generating a slow call when 1688 // both memory projections are combined) 1689 if (!always_slow && _memproj_fallthrough != NULL) { 1690 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { 1691 Node *use = _memproj_fallthrough->fast_out(i); 1692 _igvn.rehash_node_delayed(use); 1693 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); 1694 // back up iterator 1695 --i; 1696 } 1697 } 1698 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete 1699 // _memproj_catchall so we end up with a call that has only 1 memory projection. 1700 if (_memproj_catchall != NULL ) { 1701 if (_memproj_fallthrough == NULL) { 1702 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory); 1703 transform_later(_memproj_fallthrough); 1704 } 1705 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { 1706 Node *use = _memproj_catchall->fast_out(i); 1707 _igvn.rehash_node_delayed(use); 1708 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); 1709 // back up iterator 1710 --i; 1711 } 1712 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted"); 1713 _igvn.remove_dead_node(_memproj_catchall); 1714 } 1715 1716 // An allocate node has separate i_o projections for the uses on the control 1717 // and i_o paths. Always replace the control i_o projection with result i_o 1718 // otherwise incoming i_o become dead when only a slow call is generated 1719 // (it is different from memory projections where both projections are 1720 // combined in such case). 1721 if (_ioproj_fallthrough != NULL) { 1722 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { 1723 Node *use = _ioproj_fallthrough->fast_out(i); 1724 _igvn.rehash_node_delayed(use); 1725 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); 1726 // back up iterator 1727 --i; 1728 } 1729 } 1730 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete 1731 // _ioproj_catchall so we end up with a call that has only 1 i_o projection. 1732 if (_ioproj_catchall != NULL ) { 1733 if (_ioproj_fallthrough == NULL) { 1734 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O); 1735 transform_later(_ioproj_fallthrough); 1736 } 1737 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { 1738 Node *use = _ioproj_catchall->fast_out(i); 1739 _igvn.rehash_node_delayed(use); 1740 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); 1741 // back up iterator 1742 --i; 1743 } 1744 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted"); 1745 _igvn.remove_dead_node(_ioproj_catchall); 1746 } 1747 1748 // if we generated only a slow call, we are done 1749 if (always_slow) { 1750 // Now we can unhook i_o. 1751 if (result_phi_i_o->outcnt() > 1) { 1752 call->set_req(TypeFunc::I_O, top()); 1753 } else { 1754 assert(result_phi_i_o->unique_ctrl_out() == call, ""); 1755 // Case of new array with negative size known during compilation. 1756 // AllocateArrayNode::Ideal() optimization disconnect unreachable 1757 // following code since call to runtime will throw exception. 1758 // As result there will be no users of i_o after the call. 1759 // Leave i_o attached to this call to avoid problems in preceding graph. 1760 } 1761 return; 1762 } 1763 1764 1765 if (_fallthroughcatchproj != NULL) { 1766 ctrl = _fallthroughcatchproj->clone(); 1767 transform_later(ctrl); 1768 _igvn.replace_node(_fallthroughcatchproj, result_region); 1769 } else { 1770 ctrl = top(); 1771 } 1772 Node *slow_result; 1773 if (_resproj == NULL) { 1774 // no uses of the allocation result 1775 slow_result = top(); 1776 } else { 1777 slow_result = _resproj->clone(); 1778 transform_later(slow_result); 1779 _igvn.replace_node(_resproj, result_phi_rawoop); 1780 } 1781 1782 // Plug slow-path into result merge point 1783 result_region ->init_req( slow_result_path, ctrl ); 1784 result_phi_rawoop->init_req( slow_result_path, slow_result); 1785 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); 1786 transform_later(result_region); 1787 transform_later(result_phi_rawoop); 1788 transform_later(result_phi_rawmem); 1789 transform_later(result_phi_i_o); 1790 // This completes all paths into the result merge point 1791 } 1792 1793 1794 // Helper for PhaseMacroExpand::expand_allocate_common. 1795 // Initializes the newly-allocated storage. 1796 Node* 1797 PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1798 Node* control, Node* rawmem, Node* object, 1799 Node* klass_node, Node* length, 1800 Node* size_in_bytes) { 1801 InitializeNode* init = alloc->initialization(); 1802 // Store the klass & mark bits 1803 Node* mark_node = NULL; 1804 // For now only enable fast locking for non-array types 1805 if (UseBiasedLocking && (length == NULL)) { 1806 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 1807 } else { 1808 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); 1809 } 1810 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 1811 1812 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 1813 int header_size = alloc->minimum_header_size(); // conservatively small 1814 1815 // Array length 1816 if (length != NULL) { // Arrays need length field 1817 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1818 // conservatively small header size: 1819 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1820 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1821 if (k->is_array_klass()) // we know the exact header size in most cases: 1822 header_size = Klass::layout_helper_header_size(k->layout_helper()); 1823 } 1824 1825 // Clear the object body, if necessary. 1826 if (init == NULL) { 1827 // The init has somehow disappeared; be cautious and clear everything. 1828 // 1829 // This can happen if a node is allocated but an uncommon trap occurs 1830 // immediately. In this case, the Initialize gets associated with the 1831 // trap, and may be placed in a different (outer) loop, if the Allocate 1832 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1833 // there can be two Allocates to one Initialize. The answer in all these 1834 // edge cases is safety first. It is always safe to clear immediately 1835 // within an Allocate, and then (maybe or maybe not) clear some more later. 1836 if (!(UseTLAB && ZeroTLAB)) { 1837 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1838 header_size, size_in_bytes, 1839 &_igvn); 1840 } 1841 } else { 1842 if (!init->is_complete()) { 1843 // Try to win by zeroing only what the init does not store. 1844 // We can also try to do some peephole optimizations, 1845 // such as combining some adjacent subword stores. 1846 rawmem = init->complete_stores(control, rawmem, object, 1847 header_size, size_in_bytes, &_igvn); 1848 } 1849 // We have no more use for this link, since the AllocateNode goes away: 1850 init->set_req(InitializeNode::RawAddress, top()); 1851 // (If we keep the link, it just confuses the register allocator, 1852 // who thinks he sees a real use of the address by the membar.) 1853 } 1854 1855 return rawmem; 1856 } 1857 1858 // Generate prefetch instructions for next allocations. 1859 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1860 Node*& contended_phi_rawmem, 1861 Node* old_eden_top, Node* new_eden_top, 1862 Node* length) { 1863 enum { fall_in_path = 1, pf_path = 2 }; 1864 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1865 // Generate prefetch allocation with watermark check. 1866 // As an allocation hits the watermark, we will prefetch starting 1867 // at a "distance" away from watermark. 1868 1869 Node *pf_region = new RegionNode(3); 1870 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 1871 TypeRawPtr::BOTTOM ); 1872 // I/O is used for Prefetch 1873 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO ); 1874 1875 Node *thread = new ThreadLocalNode(); 1876 transform_later(thread); 1877 1878 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread, 1879 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1880 transform_later(eden_pf_adr); 1881 1882 Node *old_pf_wm = new LoadPNode(needgc_false, 1883 contended_phi_rawmem, eden_pf_adr, 1884 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, 1885 MemNode::unordered); 1886 transform_later(old_pf_wm); 1887 1888 // check against new_eden_top 1889 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm ); 1890 transform_later(need_pf_cmp); 1891 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge ); 1892 transform_later(need_pf_bol); 1893 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol, 1894 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1895 transform_later(need_pf_iff); 1896 1897 // true node, add prefetchdistance 1898 Node *need_pf_true = new IfTrueNode( need_pf_iff ); 1899 transform_later(need_pf_true); 1900 1901 Node *need_pf_false = new IfFalseNode( need_pf_iff ); 1902 transform_later(need_pf_false); 1903 1904 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm, 1905 _igvn.MakeConX(AllocatePrefetchDistance) ); 1906 transform_later(new_pf_wmt ); 1907 new_pf_wmt->set_req(0, need_pf_true); 1908 1909 Node *store_new_wmt = new StorePNode(need_pf_true, 1910 contended_phi_rawmem, eden_pf_adr, 1911 TypeRawPtr::BOTTOM, new_pf_wmt, 1912 MemNode::unordered); 1913 transform_later(store_new_wmt); 1914 1915 // adding prefetches 1916 pf_phi_abio->init_req( fall_in_path, i_o ); 1917 1918 Node *prefetch_adr; 1919 Node *prefetch; 1920 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1921 uint step_size = AllocatePrefetchStepSize; 1922 uint distance = 0; 1923 1924 for ( uint i = 0; i < lines; i++ ) { 1925 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt, 1926 _igvn.MakeConX(distance) ); 1927 transform_later(prefetch_adr); 1928 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1929 transform_later(prefetch); 1930 distance += step_size; 1931 i_o = prefetch; 1932 } 1933 pf_phi_abio->set_req( pf_path, i_o ); 1934 1935 pf_region->init_req( fall_in_path, need_pf_false ); 1936 pf_region->init_req( pf_path, need_pf_true ); 1937 1938 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 1939 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 1940 1941 transform_later(pf_region); 1942 transform_later(pf_phi_rawmem); 1943 transform_later(pf_phi_abio); 1944 1945 needgc_false = pf_region; 1946 contended_phi_rawmem = pf_phi_rawmem; 1947 i_o = pf_phi_abio; 1948 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { 1949 // Insert a prefetch instruction for each allocation. 1950 // This code is used for SPARC with BIS. 1951 1952 // Generate several prefetch instructions. 1953 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1954 uint step_size = AllocatePrefetchStepSize; 1955 uint distance = AllocatePrefetchDistance; 1956 1957 // Next cache address. 1958 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top, 1959 _igvn.MakeConX(step_size + distance)); 1960 transform_later(cache_adr); 1961 cache_adr = new CastP2XNode(needgc_false, cache_adr); 1962 transform_later(cache_adr); 1963 // For BIS instructions to be emitted, the address must be aligned at cache line size. 1964 // (The VM sets AllocatePrefetchStepSize to the cache line size, unless a value is 1965 // specified at the command line.) If the address is not aligned at cache line size 1966 // boundary, a standard store instruction is triggered (instead of the BIS). For the 1967 // latter, 8-byte alignment is necessary. 1968 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); 1969 cache_adr = new AndXNode(cache_adr, mask); 1970 transform_later(cache_adr); 1971 cache_adr = new CastX2PNode(cache_adr); 1972 transform_later(cache_adr); 1973 1974 // Prefetch 1975 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); 1976 prefetch->set_req(0, needgc_false); 1977 transform_later(prefetch); 1978 contended_phi_rawmem = prefetch; 1979 Node *prefetch_adr; 1980 distance = step_size; 1981 for ( uint i = 1; i < lines; i++ ) { 1982 prefetch_adr = new AddPNode( cache_adr, cache_adr, 1983 _igvn.MakeConX(distance) ); 1984 transform_later(prefetch_adr); 1985 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); 1986 transform_later(prefetch); 1987 distance += step_size; 1988 contended_phi_rawmem = prefetch; 1989 } 1990 } else if( AllocatePrefetchStyle > 0 ) { 1991 // Insert a prefetch for each allocation only on the fast-path 1992 Node *prefetch_adr; 1993 Node *prefetch; 1994 // Generate several prefetch instructions. 1995 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1996 uint step_size = AllocatePrefetchStepSize; 1997 uint distance = AllocatePrefetchDistance; 1998 for ( uint i = 0; i < lines; i++ ) { 1999 prefetch_adr = new AddPNode( old_eden_top, new_eden_top, 2000 _igvn.MakeConX(distance) ); 2001 transform_later(prefetch_adr); 2002 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 2003 // Do not let it float too high, since if eden_top == eden_end, 2004 // both might be null. 2005 if( i == 0 ) { // Set control for first prefetch, next follows it 2006 prefetch->init_req(0, needgc_false); 2007 } 2008 transform_later(prefetch); 2009 distance += step_size; 2010 i_o = prefetch; 2011 } 2012 } 2013 return i_o; 2014 } 2015 2016 2017 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 2018 expand_allocate_common(alloc, NULL, 2019 OptoRuntime::new_instance_Type(), 2020 OptoRuntime::new_instance_Java()); 2021 } 2022 2023 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 2024 Node* length = alloc->in(AllocateNode::ALength); 2025 InitializeNode* init = alloc->initialization(); 2026 Node* klass_node = alloc->in(AllocateNode::KlassNode); 2027 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 2028 address slow_call_address; // Address of slow call 2029 if (init != NULL && init->is_complete_with_arraycopy() && 2030 k->is_type_array_klass()) { 2031 // Don't zero type array during slow allocation in VM since 2032 // it will be initialized later by arraycopy in compiled code. 2033 slow_call_address = OptoRuntime::new_array_nozero_Java(); 2034 } else { 2035 slow_call_address = OptoRuntime::new_array_Java(); 2036 } 2037 expand_allocate_common(alloc, length, 2038 OptoRuntime::new_array_Type(), 2039 slow_call_address); 2040 } 2041 2042 //-------------------mark_eliminated_box---------------------------------- 2043 // 2044 // During EA obj may point to several objects but after few ideal graph 2045 // transformations (CCP) it may point to only one non escaping object 2046 // (but still using phi), corresponding locks and unlocks will be marked 2047 // for elimination. Later obj could be replaced with a new node (new phi) 2048 // and which does not have escape information. And later after some graph 2049 // reshape other locks and unlocks (which were not marked for elimination 2050 // before) are connected to this new obj (phi) but they still will not be 2051 // marked for elimination since new obj has no escape information. 2052 // Mark all associated (same box and obj) lock and unlock nodes for 2053 // elimination if some of them marked already. 2054 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) { 2055 if (oldbox->as_BoxLock()->is_eliminated()) 2056 return; // This BoxLock node was processed already. 2057 2058 // New implementation (EliminateNestedLocks) has separate BoxLock 2059 // node for each locked region so mark all associated locks/unlocks as 2060 // eliminated even if different objects are referenced in one locked region 2061 // (for example, OSR compilation of nested loop inside locked scope). 2062 if (EliminateNestedLocks || 2063 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) { 2064 // Box is used only in one lock region. Mark this box as eliminated. 2065 _igvn.hash_delete(oldbox); 2066 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value 2067 _igvn.hash_insert(oldbox); 2068 2069 for (uint i = 0; i < oldbox->outcnt(); i++) { 2070 Node* u = oldbox->raw_out(i); 2071 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) { 2072 AbstractLockNode* alock = u->as_AbstractLock(); 2073 // Check lock's box since box could be referenced by Lock's debug info. 2074 if (alock->box_node() == oldbox) { 2075 // Mark eliminated all related locks and unlocks. 2076 #ifdef ASSERT 2077 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4"); 2078 #endif 2079 alock->set_non_esc_obj(); 2080 } 2081 } 2082 } 2083 return; 2084 } 2085 2086 // Create new "eliminated" BoxLock node and use it in monitor debug info 2087 // instead of oldbox for the same object. 2088 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 2089 2090 // Note: BoxLock node is marked eliminated only here and it is used 2091 // to indicate that all associated lock and unlock nodes are marked 2092 // for elimination. 2093 newbox->set_eliminated(); 2094 transform_later(newbox); 2095 2096 // Replace old box node with new box for all users of the same object. 2097 for (uint i = 0; i < oldbox->outcnt();) { 2098 bool next_edge = true; 2099 2100 Node* u = oldbox->raw_out(i); 2101 if (u->is_AbstractLock()) { 2102 AbstractLockNode* alock = u->as_AbstractLock(); 2103 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) { 2104 // Replace Box and mark eliminated all related locks and unlocks. 2105 #ifdef ASSERT 2106 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5"); 2107 #endif 2108 alock->set_non_esc_obj(); 2109 _igvn.rehash_node_delayed(alock); 2110 alock->set_box_node(newbox); 2111 next_edge = false; 2112 } 2113 } 2114 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) { 2115 FastLockNode* flock = u->as_FastLock(); 2116 assert(flock->box_node() == oldbox, "sanity"); 2117 _igvn.rehash_node_delayed(flock); 2118 flock->set_box_node(newbox); 2119 next_edge = false; 2120 } 2121 2122 // Replace old box in monitor debug info. 2123 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 2124 SafePointNode* sfn = u->as_SafePoint(); 2125 JVMState* youngest_jvms = sfn->jvms(); 2126 int max_depth = youngest_jvms->depth(); 2127 for (int depth = 1; depth <= max_depth; depth++) { 2128 JVMState* jvms = youngest_jvms->of_depth(depth); 2129 int num_mon = jvms->nof_monitors(); 2130 // Loop over monitors 2131 for (int idx = 0; idx < num_mon; idx++) { 2132 Node* obj_node = sfn->monitor_obj(jvms, idx); 2133 Node* box_node = sfn->monitor_box(jvms, idx); 2134 if (box_node == oldbox && obj_node->eqv_uncast(obj)) { 2135 int j = jvms->monitor_box_offset(idx); 2136 _igvn.replace_input_of(u, j, newbox); 2137 next_edge = false; 2138 } 2139 } 2140 } 2141 } 2142 if (next_edge) i++; 2143 } 2144 } 2145 2146 //-----------------------mark_eliminated_locking_nodes----------------------- 2147 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { 2148 if (EliminateNestedLocks) { 2149 if (alock->is_nested()) { 2150 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity"); 2151 return; 2152 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened 2153 // Only Lock node has JVMState needed here. 2154 // Not that preceding claim is documented anywhere else. 2155 if (alock->jvms() != NULL) { 2156 if (alock->as_Lock()->is_nested_lock_region()) { 2157 // Mark eliminated related nested locks and unlocks. 2158 Node* obj = alock->obj_node(); 2159 BoxLockNode* box_node = alock->box_node()->as_BoxLock(); 2160 assert(!box_node->is_eliminated(), "should not be marked yet"); 2161 // Note: BoxLock node is marked eliminated only here 2162 // and it is used to indicate that all associated lock 2163 // and unlock nodes are marked for elimination. 2164 box_node->set_eliminated(); // Box's hash is always NO_HASH here 2165 for (uint i = 0; i < box_node->outcnt(); i++) { 2166 Node* u = box_node->raw_out(i); 2167 if (u->is_AbstractLock()) { 2168 alock = u->as_AbstractLock(); 2169 if (alock->box_node() == box_node) { 2170 // Verify that this Box is referenced only by related locks. 2171 assert(alock->obj_node()->eqv_uncast(obj), ""); 2172 // Mark all related locks and unlocks. 2173 #ifdef ASSERT 2174 alock->log_lock_optimization(C, "eliminate_lock_set_nested"); 2175 #endif 2176 alock->set_nested(); 2177 } 2178 } 2179 } 2180 } else { 2181 #ifdef ASSERT 2182 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region"); 2183 if (C->log() != NULL) 2184 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output 2185 #endif 2186 } 2187 } 2188 return; 2189 } 2190 // Process locks for non escaping object 2191 assert(alock->is_non_esc_obj(), ""); 2192 } // EliminateNestedLocks 2193 2194 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object 2195 // Look for all locks of this object and mark them and 2196 // corresponding BoxLock nodes as eliminated. 2197 Node* obj = alock->obj_node(); 2198 for (uint j = 0; j < obj->outcnt(); j++) { 2199 Node* o = obj->raw_out(j); 2200 if (o->is_AbstractLock() && 2201 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) { 2202 alock = o->as_AbstractLock(); 2203 Node* box = alock->box_node(); 2204 // Replace old box node with new eliminated box for all users 2205 // of the same object and mark related locks as eliminated. 2206 mark_eliminated_box(box, obj); 2207 } 2208 } 2209 } 2210 } 2211 2212 // we have determined that this lock/unlock can be eliminated, we simply 2213 // eliminate the node without expanding it. 2214 // 2215 // Note: The membar's associated with the lock/unlock are currently not 2216 // eliminated. This should be investigated as a future enhancement. 2217 // 2218 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 2219 2220 if (!alock->is_eliminated()) { 2221 return false; 2222 } 2223 #ifdef ASSERT 2224 if (!alock->is_coarsened()) { 2225 // Check that new "eliminated" BoxLock node is created. 2226 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 2227 assert(oldbox->is_eliminated(), "should be done already"); 2228 } 2229 #endif 2230 2231 alock->log_lock_optimization(C, "eliminate_lock"); 2232 2233 #ifndef PRODUCT 2234 if (PrintEliminateLocks) { 2235 if (alock->is_Lock()) { 2236 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx); 2237 } else { 2238 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx); 2239 } 2240 } 2241 #endif 2242 2243 Node* mem = alock->in(TypeFunc::Memory); 2244 Node* ctrl = alock->in(TypeFunc::Control); 2245 2246 extract_call_projections(alock); 2247 // There are 2 projections from the lock. The lock node will 2248 // be deleted when its last use is subsumed below. 2249 assert(alock->outcnt() == 2 && 2250 _fallthroughproj != NULL && 2251 _memproj_fallthrough != NULL, 2252 "Unexpected projections from Lock/Unlock"); 2253 2254 Node* fallthroughproj = _fallthroughproj; 2255 Node* memproj_fallthrough = _memproj_fallthrough; 2256 2257 // The memory projection from a lock/unlock is RawMem 2258 // The input to a Lock is merged memory, so extract its RawMem input 2259 // (unless the MergeMem has been optimized away.) 2260 if (alock->is_Lock()) { 2261 // Seach for MemBarAcquireLock node and delete it also. 2262 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 2263 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, ""); 2264 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 2265 Node* memproj = membar->proj_out(TypeFunc::Memory); 2266 _igvn.replace_node(ctrlproj, fallthroughproj); 2267 _igvn.replace_node(memproj, memproj_fallthrough); 2268 2269 // Delete FastLock node also if this Lock node is unique user 2270 // (a loop peeling may clone a Lock node). 2271 Node* flock = alock->as_Lock()->fastlock_node(); 2272 if (flock->outcnt() == 1) { 2273 assert(flock->unique_out() == alock, "sanity"); 2274 _igvn.replace_node(flock, top()); 2275 } 2276 } 2277 2278 // Seach for MemBarReleaseLock node and delete it also. 2279 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() && 2280 ctrl->in(0)->is_MemBar()) { 2281 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 2282 assert(membar->Opcode() == Op_MemBarReleaseLock && 2283 mem->is_Proj() && membar == mem->in(0), ""); 2284 _igvn.replace_node(fallthroughproj, ctrl); 2285 _igvn.replace_node(memproj_fallthrough, mem); 2286 fallthroughproj = ctrl; 2287 memproj_fallthrough = mem; 2288 ctrl = membar->in(TypeFunc::Control); 2289 mem = membar->in(TypeFunc::Memory); 2290 } 2291 2292 _igvn.replace_node(fallthroughproj, ctrl); 2293 _igvn.replace_node(memproj_fallthrough, mem); 2294 return true; 2295 } 2296 2297 2298 //------------------------------expand_lock_node---------------------- 2299 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 2300 2301 Node* ctrl = lock->in(TypeFunc::Control); 2302 Node* mem = lock->in(TypeFunc::Memory); 2303 Node* obj = lock->obj_node(); 2304 Node* box = lock->box_node(); 2305 Node* flock = lock->fastlock_node(); 2306 2307 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2308 2309 // Make the merge point 2310 Node *region; 2311 Node *mem_phi; 2312 Node *slow_path; 2313 2314 if (UseOptoBiasInlining) { 2315 /* 2316 * See the full description in MacroAssembler::biased_locking_enter(). 2317 * 2318 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { 2319 * // The object is biased. 2320 * proto_node = klass->prototype_header; 2321 * o_node = thread | proto_node; 2322 * x_node = o_node ^ mark_word; 2323 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? 2324 * // Done. 2325 * } else { 2326 * if( (x_node & biased_lock_mask) != 0 ) { 2327 * // The klass's prototype header is no longer biased. 2328 * cas(&mark_word, mark_word, proto_node) 2329 * goto cas_lock; 2330 * } else { 2331 * // The klass's prototype header is still biased. 2332 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? 2333 * old = mark_word; 2334 * new = o_node; 2335 * } else { 2336 * // Different thread or anonymous biased. 2337 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); 2338 * new = thread | old; 2339 * } 2340 * // Try to rebias. 2341 * if( cas(&mark_word, old, new) == 0 ) { 2342 * // Done. 2343 * } else { 2344 * goto slow_path; // Failed. 2345 * } 2346 * } 2347 * } 2348 * } else { 2349 * // The object is not biased. 2350 * cas_lock: 2351 * if( FastLock(obj) == 0 ) { 2352 * // Done. 2353 * } else { 2354 * slow_path: 2355 * OptoRuntime::complete_monitor_locking_Java(obj); 2356 * } 2357 * } 2358 */ 2359 2360 region = new RegionNode(5); 2361 // create a Phi for the memory state 2362 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2363 2364 Node* fast_lock_region = new RegionNode(3); 2365 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); 2366 2367 // First, check mark word for the biased lock pattern. 2368 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2369 2370 // Get fast path - mark word has the biased lock pattern. 2371 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, 2372 markOopDesc::biased_lock_mask_in_place, 2373 markOopDesc::biased_lock_pattern, true); 2374 // fast_lock_region->in(1) is set to slow path. 2375 fast_lock_mem_phi->init_req(1, mem); 2376 2377 // Now check that the lock is biased to the current thread and has 2378 // the same epoch and bias as Klass::_prototype_header. 2379 2380 // Special-case a fresh allocation to avoid building nodes: 2381 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); 2382 if (klass_node == NULL) { 2383 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 2384 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr())); 2385 #ifdef _LP64 2386 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) { 2387 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity"); 2388 klass_node->in(1)->init_req(0, ctrl); 2389 } else 2390 #endif 2391 klass_node->init_req(0, ctrl); 2392 } 2393 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type()); 2394 2395 Node* thread = transform_later(new ThreadLocalNode()); 2396 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2397 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node)); 2398 Node* x_node = transform_later(new XorXNode(o_node, mark_node)); 2399 2400 // Get slow path - mark word does NOT match the value. 2401 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, 2402 (~markOopDesc::age_mask_in_place), 0); 2403 // region->in(3) is set to fast path - the object is biased to the current thread. 2404 mem_phi->init_req(3, mem); 2405 2406 2407 // Mark word does NOT match the value (thread | Klass::_prototype_header). 2408 2409 2410 // First, check biased pattern. 2411 // Get fast path - _prototype_header has the same biased lock pattern. 2412 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, 2413 markOopDesc::biased_lock_mask_in_place, 0, true); 2414 2415 not_biased_ctrl = fast_lock_region->in(2); // Slow path 2416 // fast_lock_region->in(2) - the prototype header is no longer biased 2417 // and we have to revoke the bias on this object. 2418 // We are going to try to reset the mark of this object to the prototype 2419 // value and fall through to the CAS-based locking scheme. 2420 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 2421 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr, 2422 proto_node, mark_node); 2423 transform_later(cas); 2424 Node* proj = transform_later(new SCMemProjNode(cas)); 2425 fast_lock_mem_phi->init_req(2, proj); 2426 2427 2428 // Second, check epoch bits. 2429 Node* rebiased_region = new RegionNode(3); 2430 Node* old_phi = new PhiNode( rebiased_region, TypeX_X); 2431 Node* new_phi = new PhiNode( rebiased_region, TypeX_X); 2432 2433 // Get slow path - mark word does NOT match epoch bits. 2434 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, 2435 markOopDesc::epoch_mask_in_place, 0); 2436 // The epoch of the current bias is not valid, attempt to rebias the object 2437 // toward the current thread. 2438 rebiased_region->init_req(2, epoch_ctrl); 2439 old_phi->init_req(2, mark_node); 2440 new_phi->init_req(2, o_node); 2441 2442 // rebiased_region->in(1) is set to fast path. 2443 // The epoch of the current bias is still valid but we know 2444 // nothing about the owner; it might be set or it might be clear. 2445 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place | 2446 markOopDesc::age_mask_in_place | 2447 markOopDesc::epoch_mask_in_place); 2448 Node* old = transform_later(new AndXNode(mark_node, cmask)); 2449 cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2450 Node* new_mark = transform_later(new OrXNode(cast_thread, old)); 2451 old_phi->init_req(1, old); 2452 new_phi->init_req(1, new_mark); 2453 2454 transform_later(rebiased_region); 2455 transform_later(old_phi); 2456 transform_later(new_phi); 2457 2458 // Try to acquire the bias of the object using an atomic operation. 2459 // If this fails we will go in to the runtime to revoke the object's bias. 2460 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi); 2461 transform_later(cas); 2462 proj = transform_later(new SCMemProjNode(cas)); 2463 2464 // Get slow path - Failed to CAS. 2465 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); 2466 mem_phi->init_req(4, proj); 2467 // region->in(4) is set to fast path - the object is rebiased to the current thread. 2468 2469 // Failed to CAS. 2470 slow_path = new RegionNode(3); 2471 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); 2472 2473 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control 2474 slow_mem->init_req(1, proj); 2475 2476 // Call CAS-based locking scheme (FastLock node). 2477 2478 transform_later(fast_lock_region); 2479 transform_later(fast_lock_mem_phi); 2480 2481 // Get slow path - FastLock failed to lock the object. 2482 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); 2483 mem_phi->init_req(2, fast_lock_mem_phi); 2484 // region->in(2) is set to fast path - the object is locked to the current thread. 2485 2486 slow_path->init_req(2, ctrl); // Capture slow-control 2487 slow_mem->init_req(2, fast_lock_mem_phi); 2488 2489 transform_later(slow_path); 2490 transform_later(slow_mem); 2491 // Reset lock's memory edge. 2492 lock->set_req(TypeFunc::Memory, slow_mem); 2493 2494 } else { 2495 region = new RegionNode(3); 2496 // create a Phi for the memory state 2497 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2498 2499 // Optimize test; set region slot 2 2500 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); 2501 mem_phi->init_req(2, mem); 2502 } 2503 2504 // Make slow path call 2505 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), 2506 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, 2507 obj, box, NULL); 2508 2509 extract_call_projections(call); 2510 2511 // Slow path can only throw asynchronous exceptions, which are always 2512 // de-opted. So the compiler thinks the slow-call can never throw an 2513 // exception. If it DOES throw an exception we would need the debug 2514 // info removed first (since if it throws there is no monitor). 2515 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2516 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2517 2518 // Capture slow path 2519 // disconnect fall-through projection from call and create a new one 2520 // hook up users of fall-through projection to region 2521 Node *slow_ctrl = _fallthroughproj->clone(); 2522 transform_later(slow_ctrl); 2523 _igvn.hash_delete(_fallthroughproj); 2524 _fallthroughproj->disconnect_inputs(NULL, C); 2525 region->init_req(1, slow_ctrl); 2526 // region inputs are now complete 2527 transform_later(region); 2528 _igvn.replace_node(_fallthroughproj, region); 2529 2530 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory)); 2531 mem_phi->init_req(1, memproj ); 2532 transform_later(mem_phi); 2533 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2534 } 2535 2536 //------------------------------expand_unlock_node---------------------- 2537 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 2538 2539 Node* ctrl = unlock->in(TypeFunc::Control); 2540 Node* mem = unlock->in(TypeFunc::Memory); 2541 Node* obj = unlock->obj_node(); 2542 Node* box = unlock->box_node(); 2543 2544 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2545 2546 // No need for a null check on unlock 2547 2548 // Make the merge point 2549 Node *region; 2550 Node *mem_phi; 2551 2552 if (UseOptoBiasInlining) { 2553 // Check for biased locking unlock case, which is a no-op. 2554 // See the full description in MacroAssembler::biased_locking_exit(). 2555 region = new RegionNode(4); 2556 // create a Phi for the memory state 2557 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2558 mem_phi->init_req(3, mem); 2559 2560 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2561 ctrl = opt_bits_test(ctrl, region, 3, mark_node, 2562 markOopDesc::biased_lock_mask_in_place, 2563 markOopDesc::biased_lock_pattern); 2564 } else { 2565 region = new RegionNode(3); 2566 // create a Phi for the memory state 2567 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2568 } 2569 2570 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box ); 2571 funlock = transform_later( funlock )->as_FastUnlock(); 2572 // Optimize test; set region slot 2 2573 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); 2574 Node *thread = transform_later(new ThreadLocalNode()); 2575 2576 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), 2577 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), 2578 "complete_monitor_unlocking_C", slow_path, obj, box, thread); 2579 2580 extract_call_projections(call); 2581 2582 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2583 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2584 2585 // No exceptions for unlocking 2586 // Capture slow path 2587 // disconnect fall-through projection from call and create a new one 2588 // hook up users of fall-through projection to region 2589 Node *slow_ctrl = _fallthroughproj->clone(); 2590 transform_later(slow_ctrl); 2591 _igvn.hash_delete(_fallthroughproj); 2592 _fallthroughproj->disconnect_inputs(NULL, C); 2593 region->init_req(1, slow_ctrl); 2594 // region inputs are now complete 2595 transform_later(region); 2596 _igvn.replace_node(_fallthroughproj, region); 2597 2598 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) ); 2599 mem_phi->init_req(1, memproj ); 2600 mem_phi->init_req(2, mem); 2601 transform_later(mem_phi); 2602 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2603 } 2604 2605 //---------------------------eliminate_macro_nodes---------------------- 2606 // Eliminate scalar replaced allocations and associated locks. 2607 void PhaseMacroExpand::eliminate_macro_nodes() { 2608 if (C->macro_count() == 0) 2609 return; 2610 2611 // First, attempt to eliminate locks 2612 int cnt = C->macro_count(); 2613 for (int i=0; i < cnt; i++) { 2614 Node *n = C->macro_node(i); 2615 if (n->is_AbstractLock()) { // Lock and Unlock nodes 2616 // Before elimination mark all associated (same box and obj) 2617 // lock and unlock nodes. 2618 mark_eliminated_locking_nodes(n->as_AbstractLock()); 2619 } 2620 } 2621 bool progress = true; 2622 while (progress) { 2623 progress = false; 2624 for (int i = C->macro_count(); i > 0; i--) { 2625 Node * n = C->macro_node(i-1); 2626 bool success = false; 2627 debug_only(int old_macro_count = C->macro_count();); 2628 if (n->is_AbstractLock()) { 2629 success = eliminate_locking_node(n->as_AbstractLock()); 2630 } 2631 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2632 progress = progress || success; 2633 } 2634 } 2635 // Next, attempt to eliminate allocations 2636 _has_locks = false; 2637 progress = true; 2638 while (progress) { 2639 progress = false; 2640 for (int i = C->macro_count(); i > 0; i--) { 2641 Node * n = C->macro_node(i-1); 2642 bool success = false; 2643 debug_only(int old_macro_count = C->macro_count();); 2644 switch (n->class_id()) { 2645 case Node::Class_Allocate: 2646 case Node::Class_AllocateArray: 2647 success = eliminate_allocate_node(n->as_Allocate()); 2648 break; 2649 case Node::Class_CallStaticJava: 2650 success = eliminate_boxing_node(n->as_CallStaticJava()); 2651 break; 2652 case Node::Class_Lock: 2653 case Node::Class_Unlock: 2654 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 2655 _has_locks = true; 2656 break; 2657 case Node::Class_ArrayCopy: 2658 break; 2659 default: 2660 assert(n->Opcode() == Op_LoopLimit || 2661 n->Opcode() == Op_Opaque1 || 2662 n->Opcode() == Op_Opaque2 || 2663 n->Opcode() == Op_Opaque3, "unknown node type in macro list"); 2664 } 2665 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2666 progress = progress || success; 2667 } 2668 } 2669 } 2670 2671 //------------------------------expand_macro_nodes---------------------- 2672 // Returns true if a failure occurred. 2673 bool PhaseMacroExpand::expand_macro_nodes() { 2674 // Last attempt to eliminate macro nodes. 2675 eliminate_macro_nodes(); 2676 2677 // Make sure expansion will not cause node limit to be exceeded. 2678 // Worst case is a macro node gets expanded into about 200 nodes. 2679 // Allow 50% more for optimization. 2680 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) ) 2681 return true; 2682 2683 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. 2684 bool progress = true; 2685 while (progress) { 2686 progress = false; 2687 for (int i = C->macro_count(); i > 0; i--) { 2688 Node * n = C->macro_node(i-1); 2689 bool success = false; 2690 debug_only(int old_macro_count = C->macro_count();); 2691 if (n->Opcode() == Op_LoopLimit) { 2692 // Remove it from macro list and put on IGVN worklist to optimize. 2693 C->remove_macro_node(n); 2694 _igvn._worklist.push(n); 2695 success = true; 2696 } else if (n->Opcode() == Op_CallStaticJava) { 2697 // Remove it from macro list and put on IGVN worklist to optimize. 2698 C->remove_macro_node(n); 2699 _igvn._worklist.push(n); 2700 success = true; 2701 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { 2702 _igvn.replace_node(n, n->in(1)); 2703 success = true; 2704 #if INCLUDE_RTM_OPT 2705 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) { 2706 assert(C->profile_rtm(), "should be used only in rtm deoptimization code"); 2707 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), ""); 2708 Node* cmp = n->unique_out(); 2709 #ifdef ASSERT 2710 // Validate graph. 2711 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), ""); 2712 BoolNode* bol = cmp->unique_out()->as_Bool(); 2713 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() && 2714 (bol->_test._test == BoolTest::ne), ""); 2715 IfNode* ifn = bol->unique_out()->as_If(); 2716 assert((ifn->outcnt() == 2) && 2717 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, ""); 2718 #endif 2719 Node* repl = n->in(1); 2720 if (!_has_locks) { 2721 // Remove RTM state check if there are no locks in the code. 2722 // Replace input to compare the same value. 2723 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1); 2724 } 2725 _igvn.replace_node(n, repl); 2726 success = true; 2727 #endif 2728 } 2729 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2730 progress = progress || success; 2731 } 2732 } 2733 2734 // expand arraycopy "macro" nodes first 2735 // For ReduceBulkZeroing, we must first process all arraycopy nodes 2736 // before the allocate nodes are expanded. 2737 int macro_idx = C->macro_count() - 1; 2738 while (macro_idx >= 0) { 2739 Node * n = C->macro_node(macro_idx); 2740 assert(n->is_macro(), "only macro nodes expected here"); 2741 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2742 // node is unreachable, so don't try to expand it 2743 C->remove_macro_node(n); 2744 } else if (n->is_ArrayCopy()){ 2745 int macro_count = C->macro_count(); 2746 expand_arraycopy_node(n->as_ArrayCopy()); 2747 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2748 } 2749 if (C->failing()) return true; 2750 macro_idx --; 2751 } 2752 2753 // expand "macro" nodes 2754 // nodes are removed from the macro list as they are processed 2755 while (C->macro_count() > 0) { 2756 int macro_count = C->macro_count(); 2757 Node * n = C->macro_node(macro_count-1); 2758 assert(n->is_macro(), "only macro nodes expected here"); 2759 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2760 // node is unreachable, so don't try to expand it 2761 C->remove_macro_node(n); 2762 continue; 2763 } 2764 switch (n->class_id()) { 2765 case Node::Class_Allocate: 2766 expand_allocate(n->as_Allocate()); 2767 break; 2768 case Node::Class_AllocateArray: 2769 expand_allocate_array(n->as_AllocateArray()); 2770 break; 2771 case Node::Class_Lock: 2772 expand_lock_node(n->as_Lock()); 2773 break; 2774 case Node::Class_Unlock: 2775 expand_unlock_node(n->as_Unlock()); 2776 break; 2777 default: 2778 assert(false, "unknown node type in macro list"); 2779 } 2780 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2781 if (C->failing()) return true; 2782 } 2783 2784 _igvn.set_delay_transform(false); 2785 _igvn.optimize(); 2786 if (C->failing()) return true; 2787 return false; 2788 }