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