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