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