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