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