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