1 /* 2 * Copyright (c) 2005, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "libadt/vectset.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/callnode.hpp" 30 #include "opto/castnode.hpp" 31 #include "opto/cfgnode.hpp" 32 #include "opto/compile.hpp" 33 #include "opto/convertnode.hpp" 34 #include "opto/locknode.hpp" 35 #include "opto/loopnode.hpp" 36 #include "opto/macro.hpp" 37 #include "opto/memnode.hpp" 38 #include "opto/narrowptrnode.hpp" 39 #include "opto/node.hpp" 40 #include "opto/opaquenode.hpp" 41 #include "opto/phaseX.hpp" 42 #include "opto/rootnode.hpp" 43 #include "opto/runtime.hpp" 44 #include "opto/subnode.hpp" 45 #include "opto/type.hpp" 46 #include "runtime/sharedRuntime.hpp" 47 48 49 // 50 // Replace any references to "oldref" in inputs to "use" with "newref". 51 // Returns the number of replacements made. 52 // 53 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { 54 int nreplacements = 0; 55 uint req = use->req(); 56 for (uint j = 0; j < use->len(); j++) { 57 Node *uin = use->in(j); 58 if (uin == oldref) { 59 if (j < req) 60 use->set_req(j, newref); 61 else 62 use->set_prec(j, newref); 63 nreplacements++; 64 } else if (j >= req && uin == NULL) { 65 break; 66 } 67 } 68 return nreplacements; 69 } 70 71 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) { 72 // Copy debug information and adjust JVMState information 73 uint old_dbg_start = oldcall->tf()->domain()->cnt(); 74 uint new_dbg_start = newcall->tf()->domain()->cnt(); 75 int jvms_adj = new_dbg_start - old_dbg_start; 76 assert (new_dbg_start == newcall->req(), "argument count mismatch"); 77 78 // SafePointScalarObject node could be referenced several times in debug info. 79 // Use Dict to record cloned nodes. 80 Dict* sosn_map = new Dict(cmpkey,hashkey); 81 for (uint i = old_dbg_start; i < oldcall->req(); i++) { 82 Node* old_in = oldcall->in(i); 83 // Clone old SafePointScalarObjectNodes, adjusting their field contents. 84 if (old_in != NULL && old_in->is_SafePointScalarObject()) { 85 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject(); 86 uint old_unique = C->unique(); 87 Node* new_in = old_sosn->clone(sosn_map); 88 if (old_unique != C->unique()) { // New node? 89 new_in->set_req(0, C->root()); // reset control edge 90 new_in = transform_later(new_in); // Register new node. 91 } 92 old_in = new_in; 93 } 94 newcall->add_req(old_in); 95 } 96 97 // JVMS may be shared so clone it before we modify it 98 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL); 99 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) { 100 jvms->set_map(newcall); 101 jvms->set_locoff(jvms->locoff()+jvms_adj); 102 jvms->set_stkoff(jvms->stkoff()+jvms_adj); 103 jvms->set_monoff(jvms->monoff()+jvms_adj); 104 jvms->set_scloff(jvms->scloff()+jvms_adj); 105 jvms->set_endoff(jvms->endoff()+jvms_adj); 106 } 107 } 108 109 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) { 110 Node* cmp; 111 if (mask != 0) { 112 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask))); 113 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits))); 114 } else { 115 cmp = word; 116 } 117 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne)); 118 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); 119 transform_later(iff); 120 121 // Fast path taken. 122 Node *fast_taken = transform_later(new IfFalseNode(iff)); 123 124 // Fast path not-taken, i.e. slow path 125 Node *slow_taken = transform_later(new IfTrueNode(iff)); 126 127 if (return_fast_path) { 128 region->init_req(edge, slow_taken); // Capture slow-control 129 return fast_taken; 130 } else { 131 region->init_req(edge, fast_taken); // Capture fast-control 132 return slow_taken; 133 } 134 } 135 136 //--------------------copy_predefined_input_for_runtime_call-------------------- 137 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { 138 // Set fixed predefined input arguments 139 call->init_req( TypeFunc::Control, ctrl ); 140 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); 141 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? 142 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); 143 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); 144 } 145 146 //------------------------------make_slow_call--------------------------------- 147 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, 148 address slow_call, const char* leaf_name, Node* slow_path, 149 Node* parm0, Node* parm1, Node* parm2) { 150 151 // Slow-path call 152 CallNode *call = leaf_name 153 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) 154 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM ); 155 156 // Slow path call has no side-effects, uses few values 157 copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); 158 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0); 159 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1); 160 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2); 161 copy_call_debug_info(oldcall, call); 162 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 163 _igvn.replace_node(oldcall, call); 164 transform_later(call); 165 166 return call; 167 } 168 169 void PhaseMacroExpand::extract_call_projections(CallNode *call) { 170 _fallthroughproj = NULL; 171 _fallthroughcatchproj = NULL; 172 _ioproj_fallthrough = NULL; 173 _ioproj_catchall = NULL; 174 _catchallcatchproj = NULL; 175 _memproj_fallthrough = NULL; 176 _memproj_catchall = NULL; 177 _resproj = NULL; 178 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) { 179 ProjNode *pn = call->fast_out(i)->as_Proj(); 180 switch (pn->_con) { 181 case TypeFunc::Control: 182 { 183 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj 184 _fallthroughproj = pn; 185 DUIterator_Fast jmax, j = pn->fast_outs(jmax); 186 const Node *cn = pn->fast_out(j); 187 if (cn->is_Catch()) { 188 ProjNode *cpn = NULL; 189 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { 190 cpn = cn->fast_out(k)->as_Proj(); 191 assert(cpn->is_CatchProj(), "must be a CatchProjNode"); 192 if (cpn->_con == CatchProjNode::fall_through_index) 193 _fallthroughcatchproj = cpn; 194 else { 195 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); 196 _catchallcatchproj = cpn; 197 } 198 } 199 } 200 break; 201 } 202 case TypeFunc::I_O: 203 if (pn->_is_io_use) 204 _ioproj_catchall = pn; 205 else 206 _ioproj_fallthrough = pn; 207 break; 208 case TypeFunc::Memory: 209 if (pn->_is_io_use) 210 _memproj_catchall = pn; 211 else 212 _memproj_fallthrough = pn; 213 break; 214 case TypeFunc::Parms: 215 _resproj = pn; 216 break; 217 default: 218 assert(false, "unexpected projection from allocation node."); 219 } 220 } 221 222 } 223 224 // Eliminate a card mark sequence. p2x is a ConvP2XNode 225 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) { 226 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required"); 227 if (!UseG1GC) { 228 // vanilla/CMS post barrier 229 Node *shift = p2x->unique_out(); 230 Node *addp = shift->unique_out(); 231 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) { 232 Node *mem = addp->last_out(j); 233 if (UseCondCardMark && mem->is_Load()) { 234 assert(mem->Opcode() == Op_LoadB, "unexpected code shape"); 235 // The load is checking if the card has been written so 236 // replace it with zero to fold the test. 237 _igvn.replace_node(mem, intcon(0)); 238 continue; 239 } 240 assert(mem->is_Store(), "store required"); 241 _igvn.replace_node(mem, mem->in(MemNode::Memory)); 242 } 243 } else { 244 // G1 pre/post barriers 245 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes"); 246 // It could be only one user, URShift node, in Object.clone() instrinsic 247 // but the new allocation is passed to arraycopy stub and it could not 248 // be scalar replaced. So we don't check the case. 249 250 // An other case of only one user (Xor) is when the value check for NULL 251 // in G1 post barrier is folded after CCP so the code which used URShift 252 // is removed. 253 254 // Take Region node before eliminating post barrier since it also 255 // eliminates CastP2X node when it has only one user. 256 Node* this_region = p2x->in(0); 257 assert(this_region != NULL, ""); 258 259 // Remove G1 post barrier. 260 261 // Search for CastP2X->Xor->URShift->Cmp path which 262 // checks if the store done to a different from the value's region. 263 // And replace Cmp with #0 (false) to collapse G1 post barrier. 264 Node* xorx = p2x->find_out_with(Op_XorX); 265 assert(xorx != NULL, "missing G1 post barrier"); 266 Node* shift = xorx->unique_out(); 267 Node* cmpx = shift->unique_out(); 268 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() && 269 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne, 270 "missing region check in G1 post barrier"); 271 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 272 273 // Remove G1 pre barrier. 274 275 // Search "if (marking != 0)" check and set it to "false". 276 // There is no G1 pre barrier if previous stored value is NULL 277 // (for example, after initialization). 278 if (this_region->is_Region() && this_region->req() == 3) { 279 int ind = 1; 280 if (!this_region->in(ind)->is_IfFalse()) { 281 ind = 2; 282 } 283 if (this_region->in(ind)->is_IfFalse()) { 284 Node* bol = this_region->in(ind)->in(0)->in(1); 285 assert(bol->is_Bool(), ""); 286 cmpx = bol->in(1); 287 if (bol->as_Bool()->_test._test == BoolTest::ne && 288 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) && 289 cmpx->in(1)->is_Load()) { 290 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address); 291 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + 292 PtrQueue::byte_offset_of_active()); 293 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() && 294 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 295 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) { 296 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 297 } 298 } 299 } 300 } 301 // Now CastP2X can be removed since it is used only on dead path 302 // which currently still alive until igvn optimize it. 303 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, ""); 304 _igvn.replace_node(p2x, top()); 305 } 306 } 307 308 // Search for a memory operation for the specified memory slice. 309 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) { 310 Node *orig_mem = mem; 311 Node *alloc_mem = alloc->in(TypeFunc::Memory); 312 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr(); 313 while (true) { 314 if (mem == alloc_mem || mem == start_mem ) { 315 return mem; // hit one of our sentinels 316 } else if (mem->is_MergeMem()) { 317 mem = mem->as_MergeMem()->memory_at(alias_idx); 318 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { 319 Node *in = mem->in(0); 320 // we can safely skip over safepoints, calls, locks and membars because we 321 // already know that the object is safe to eliminate. 322 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { 323 return in; 324 } else if (in->is_Call()) { 325 CallNode *call = in->as_Call(); 326 if (!call->may_modify(tinst, phase)) { 327 mem = call->in(TypeFunc::Memory); 328 } 329 mem = in->in(TypeFunc::Memory); 330 } else if (in->is_MemBar()) { 331 mem = in->in(TypeFunc::Memory); 332 } else { 333 assert(false, "unexpected projection"); 334 } 335 } else if (mem->is_Store()) { 336 const TypePtr* atype = mem->as_Store()->adr_type(); 337 int adr_idx = Compile::current()->get_alias_index(atype); 338 if (adr_idx == alias_idx) { 339 assert(atype->isa_oopptr(), "address type must be oopptr"); 340 int adr_offset = atype->offset(); 341 uint adr_iid = atype->is_oopptr()->instance_id(); 342 // Array elements references have the same alias_idx 343 // but different offset and different instance_id. 344 if (adr_offset == offset && adr_iid == alloc->_idx) 345 return mem; 346 } else { 347 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw"); 348 } 349 mem = mem->in(MemNode::Memory); 350 } else if (mem->is_ClearArray()) { 351 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) { 352 // Can not bypass initialization of the instance 353 // we are looking. 354 debug_only(intptr_t offset;) 355 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity"); 356 InitializeNode* init = alloc->as_Allocate()->initialization(); 357 // We are looking for stored value, return Initialize node 358 // or memory edge from Allocate node. 359 if (init != NULL) 360 return init; 361 else 362 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers). 363 } 364 // Otherwise skip it (the call updated 'mem' value). 365 } else if (mem->Opcode() == Op_SCMemProj) { 366 mem = mem->in(0); 367 Node* adr = NULL; 368 if (mem->is_LoadStore()) { 369 adr = mem->in(MemNode::Address); 370 } else { 371 assert(mem->Opcode() == Op_EncodeISOArray, "sanity"); 372 adr = mem->in(3); // Destination array 373 } 374 const TypePtr* atype = adr->bottom_type()->is_ptr(); 375 int adr_idx = Compile::current()->get_alias_index(atype); 376 if (adr_idx == alias_idx) { 377 assert(false, "Object is not scalar replaceable if a LoadStore node access its field"); 378 return NULL; 379 } 380 mem = mem->in(MemNode::Memory); 381 } else { 382 return mem; 383 } 384 assert(mem != orig_mem, "dead memory loop"); 385 } 386 } 387 388 // 389 // Given a Memory Phi, compute a value Phi containing the values from stores 390 // on the input paths. 391 // Note: this function is recursive, its depth is limied by the "level" argument 392 // Returns the computed Phi, or NULL if it cannot compute it. 393 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, Node *alloc, Node_Stack *value_phis, int level) { 394 assert(mem->is_Phi(), "sanity"); 395 int alias_idx = C->get_alias_index(adr_t); 396 int offset = adr_t->offset(); 397 int instance_id = adr_t->instance_id(); 398 399 // Check if an appropriate value phi already exists. 400 Node* region = mem->in(0); 401 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { 402 Node* phi = region->fast_out(k); 403 if (phi->is_Phi() && phi != mem && 404 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) { 405 return phi; 406 } 407 } 408 // Check if an appropriate new value phi already exists. 409 Node* new_phi = value_phis->find(mem->_idx); 410 if (new_phi != NULL) 411 return new_phi; 412 413 if (level <= 0) { 414 return NULL; // Give up: phi tree too deep 415 } 416 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 417 Node *alloc_mem = alloc->in(TypeFunc::Memory); 418 419 uint length = mem->req(); 420 GrowableArray <Node *> values(length, length, NULL, false); 421 422 // create a new Phi for the value 423 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset); 424 transform_later(phi); 425 value_phis->push(phi, mem->_idx); 426 427 for (uint j = 1; j < length; j++) { 428 Node *in = mem->in(j); 429 if (in == NULL || in->is_top()) { 430 values.at_put(j, in); 431 } else { 432 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); 433 if (val == start_mem || val == alloc_mem) { 434 // hit a sentinel, return appropriate 0 value 435 values.at_put(j, _igvn.zerocon(ft)); 436 continue; 437 } 438 if (val->is_Initialize()) { 439 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 440 } 441 if (val == NULL) { 442 return NULL; // can't find a value on this path 443 } 444 if (val == mem) { 445 values.at_put(j, mem); 446 } else if (val->is_Store()) { 447 values.at_put(j, val->in(MemNode::ValueIn)); 448 } else if(val->is_Proj() && val->in(0) == alloc) { 449 values.at_put(j, _igvn.zerocon(ft)); 450 } else if (val->is_Phi()) { 451 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); 452 if (val == NULL) { 453 return NULL; 454 } 455 values.at_put(j, val); 456 } else if (val->Opcode() == Op_SCMemProj) { 457 assert(val->in(0)->is_LoadStore() || val->in(0)->Opcode() == Op_EncodeISOArray, "sanity"); 458 assert(false, "Object is not scalar replaceable if a LoadStore node access its field"); 459 return NULL; 460 } else { 461 #ifdef ASSERT 462 val->dump(); 463 assert(false, "unknown node on this path"); 464 #endif 465 return NULL; // unknown node on this path 466 } 467 } 468 } 469 // Set Phi's inputs 470 for (uint j = 1; j < length; j++) { 471 if (values.at(j) == mem) { 472 phi->init_req(j, phi); 473 } else { 474 phi->init_req(j, values.at(j)); 475 } 476 } 477 return phi; 478 } 479 480 // Search the last value stored into the object's field. 481 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) { 482 assert(adr_t->is_known_instance_field(), "instance required"); 483 int instance_id = adr_t->instance_id(); 484 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 485 486 int alias_idx = C->get_alias_index(adr_t); 487 int offset = adr_t->offset(); 488 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 489 Node *alloc_ctrl = alloc->in(TypeFunc::Control); 490 Node *alloc_mem = alloc->in(TypeFunc::Memory); 491 Arena *a = Thread::current()->resource_area(); 492 VectorSet visited(a); 493 494 495 bool done = sfpt_mem == alloc_mem; 496 Node *mem = sfpt_mem; 497 while (!done) { 498 if (visited.test_set(mem->_idx)) { 499 return NULL; // found a loop, give up 500 } 501 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 502 if (mem == start_mem || mem == alloc_mem) { 503 done = true; // hit a sentinel, return appropriate 0 value 504 } else if (mem->is_Initialize()) { 505 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 506 if (mem == NULL) { 507 done = true; // Something go wrong. 508 } else if (mem->is_Store()) { 509 const TypePtr* atype = mem->as_Store()->adr_type(); 510 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 511 done = true; 512 } 513 } else if (mem->is_Store()) { 514 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 515 assert(atype != NULL, "address type must be oopptr"); 516 assert(C->get_alias_index(atype) == alias_idx && 517 atype->is_known_instance_field() && atype->offset() == offset && 518 atype->instance_id() == instance_id, "store is correct memory slice"); 519 done = true; 520 } else if (mem->is_Phi()) { 521 // try to find a phi's unique input 522 Node *unique_input = NULL; 523 Node *top = C->top(); 524 for (uint i = 1; i < mem->req(); i++) { 525 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 526 if (n == NULL || n == top || n == mem) { 527 continue; 528 } else if (unique_input == NULL) { 529 unique_input = n; 530 } else if (unique_input != n) { 531 unique_input = top; 532 break; 533 } 534 } 535 if (unique_input != NULL && unique_input != top) { 536 mem = unique_input; 537 } else { 538 done = true; 539 } 540 } else { 541 assert(false, "unexpected node"); 542 } 543 } 544 if (mem != NULL) { 545 if (mem == start_mem || mem == alloc_mem) { 546 // hit a sentinel, return appropriate 0 value 547 return _igvn.zerocon(ft); 548 } else if (mem->is_Store()) { 549 return mem->in(MemNode::ValueIn); 550 } else if (mem->is_Phi()) { 551 // attempt to produce a Phi reflecting the values on the input paths of the Phi 552 Node_Stack value_phis(a, 8); 553 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 554 if (phi != NULL) { 555 return phi; 556 } else { 557 // Kill all new Phis 558 while(value_phis.is_nonempty()) { 559 Node* n = value_phis.node(); 560 _igvn.replace_node(n, C->top()); 561 value_phis.pop(); 562 } 563 } 564 } 565 } 566 // Something go wrong. 567 return NULL; 568 } 569 570 // Check the possibility of scalar replacement. 571 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 572 // Scan the uses of the allocation to check for anything that would 573 // prevent us from eliminating it. 574 NOT_PRODUCT( const char* fail_eliminate = NULL; ) 575 DEBUG_ONLY( Node* disq_node = NULL; ) 576 bool can_eliminate = true; 577 578 Node* res = alloc->result_cast(); 579 const TypeOopPtr* res_type = NULL; 580 if (res == NULL) { 581 // All users were eliminated. 582 } else if (!res->is_CheckCastPP()) { 583 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 584 can_eliminate = false; 585 } else { 586 res_type = _igvn.type(res)->isa_oopptr(); 587 if (res_type == NULL) { 588 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 589 can_eliminate = false; 590 } else if (res_type->isa_aryptr()) { 591 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 592 if (length < 0) { 593 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 594 can_eliminate = false; 595 } 596 } 597 } 598 599 if (can_eliminate && res != NULL) { 600 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 601 j < jmax && can_eliminate; j++) { 602 Node* use = res->fast_out(j); 603 604 if (use->is_AddP()) { 605 const TypePtr* addp_type = _igvn.type(use)->is_ptr(); 606 int offset = addp_type->offset(); 607 608 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 609 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) 610 can_eliminate = false; 611 break; 612 } 613 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 614 k < kmax && can_eliminate; k++) { 615 Node* n = use->fast_out(k); 616 if (!n->is_Store() && n->Opcode() != Op_CastP2X) { 617 DEBUG_ONLY(disq_node = n;) 618 if (n->is_Load() || n->is_LoadStore()) { 619 NOT_PRODUCT(fail_eliminate = "Field load";) 620 } else { 621 NOT_PRODUCT(fail_eliminate = "Not store field referrence";) 622 } 623 can_eliminate = false; 624 } 625 } 626 } else if (use->is_SafePoint()) { 627 SafePointNode* sfpt = use->as_SafePoint(); 628 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 629 // Object is passed as argument. 630 DEBUG_ONLY(disq_node = use;) 631 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 632 can_eliminate = false; 633 } 634 Node* sfptMem = sfpt->memory(); 635 if (sfptMem == NULL || sfptMem->is_top()) { 636 DEBUG_ONLY(disq_node = use;) 637 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) 638 can_eliminate = false; 639 } else { 640 safepoints.append_if_missing(sfpt); 641 } 642 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 643 if (use->is_Phi()) { 644 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 645 NOT_PRODUCT(fail_eliminate = "Object is return value";) 646 } else { 647 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 648 } 649 DEBUG_ONLY(disq_node = use;) 650 } else { 651 if (use->Opcode() == Op_Return) { 652 NOT_PRODUCT(fail_eliminate = "Object is return value";) 653 }else { 654 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 655 } 656 DEBUG_ONLY(disq_node = use;) 657 } 658 can_eliminate = false; 659 } 660 } 661 } 662 663 #ifndef PRODUCT 664 if (PrintEliminateAllocations) { 665 if (can_eliminate) { 666 tty->print("Scalar "); 667 if (res == NULL) 668 alloc->dump(); 669 else 670 res->dump(); 671 } else if (alloc->_is_scalar_replaceable) { 672 tty->print("NotScalar (%s)", fail_eliminate); 673 if (res == NULL) 674 alloc->dump(); 675 else 676 res->dump(); 677 #ifdef ASSERT 678 if (disq_node != NULL) { 679 tty->print(" >>>> "); 680 disq_node->dump(); 681 } 682 #endif /*ASSERT*/ 683 } 684 } 685 #endif 686 return can_eliminate; 687 } 688 689 // Do scalar replacement. 690 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 691 GrowableArray <SafePointNode *> safepoints_done; 692 693 ciKlass* klass = NULL; 694 ciInstanceKlass* iklass = NULL; 695 int nfields = 0; 696 int array_base; 697 int element_size; 698 BasicType basic_elem_type; 699 ciType* elem_type; 700 701 Node* res = alloc->result_cast(); 702 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result"); 703 const TypeOopPtr* res_type = NULL; 704 if (res != NULL) { // Could be NULL when there are no users 705 res_type = _igvn.type(res)->isa_oopptr(); 706 } 707 708 if (res != NULL) { 709 klass = res_type->klass(); 710 if (res_type->isa_instptr()) { 711 // find the fields of the class which will be needed for safepoint debug information 712 assert(klass->is_instance_klass(), "must be an instance klass."); 713 iklass = klass->as_instance_klass(); 714 nfields = iklass->nof_nonstatic_fields(); 715 } else { 716 // find the array's elements which will be needed for safepoint debug information 717 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 718 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); 719 elem_type = klass->as_array_klass()->element_type(); 720 basic_elem_type = elem_type->basic_type(); 721 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 722 element_size = type2aelembytes(basic_elem_type); 723 } 724 } 725 // 726 // Process the safepoint uses 727 // 728 while (safepoints.length() > 0) { 729 SafePointNode* sfpt = safepoints.pop(); 730 Node* mem = sfpt->memory(); 731 assert(sfpt->jvms() != NULL, "missed JVMS"); 732 // Fields of scalar objs are referenced only at the end 733 // of regular debuginfo at the last (youngest) JVMS. 734 // Record relative start index. 735 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff()); 736 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, 737 #ifdef ASSERT 738 alloc, 739 #endif 740 first_ind, nfields); 741 sobj->init_req(0, C->root()); 742 transform_later(sobj); 743 744 // Scan object's fields adding an input to the safepoint for each field. 745 for (int j = 0; j < nfields; j++) { 746 intptr_t offset; 747 ciField* field = NULL; 748 if (iklass != NULL) { 749 field = iklass->nonstatic_field_at(j); 750 offset = field->offset(); 751 elem_type = field->type(); 752 basic_elem_type = field->layout_type(); 753 } else { 754 offset = array_base + j * (intptr_t)element_size; 755 } 756 757 const Type *field_type; 758 // The next code is taken from Parse::do_get_xxx(). 759 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { 760 if (!elem_type->is_loaded()) { 761 field_type = TypeInstPtr::BOTTOM; 762 } else if (field != NULL && field->is_constant() && field->is_static()) { 763 // This can happen if the constant oop is non-perm. 764 ciObject* con = field->constant_value().as_object(); 765 // Do not "join" in the previous type; it doesn't add value, 766 // and may yield a vacuous result if the field is of interface type. 767 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 768 assert(field_type != NULL, "field singleton type must be consistent"); 769 } else { 770 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 771 } 772 if (UseCompressedOops) { 773 field_type = field_type->make_narrowoop(); 774 basic_elem_type = T_NARROWOOP; 775 } 776 } else { 777 field_type = Type::get_const_basic_type(basic_elem_type); 778 } 779 780 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 781 782 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc); 783 if (field_val == NULL) { 784 // We weren't able to find a value for this field, 785 // give up on eliminating this allocation. 786 787 // Remove any extra entries we added to the safepoint. 788 uint last = sfpt->req() - 1; 789 for (int k = 0; k < j; k++) { 790 sfpt->del_req(last--); 791 } 792 _igvn._worklist.push(sfpt); 793 // rollback processed safepoints 794 while (safepoints_done.length() > 0) { 795 SafePointNode* sfpt_done = safepoints_done.pop(); 796 // remove any extra entries we added to the safepoint 797 last = sfpt_done->req() - 1; 798 for (int k = 0; k < nfields; k++) { 799 sfpt_done->del_req(last--); 800 } 801 JVMState *jvms = sfpt_done->jvms(); 802 jvms->set_endoff(sfpt_done->req()); 803 // Now make a pass over the debug information replacing any references 804 // to SafePointScalarObjectNode with the allocated object. 805 int start = jvms->debug_start(); 806 int end = jvms->debug_end(); 807 for (int i = start; i < end; i++) { 808 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 809 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 810 if (scobj->first_index(jvms) == sfpt_done->req() && 811 scobj->n_fields() == (uint)nfields) { 812 assert(scobj->alloc() == alloc, "sanity"); 813 sfpt_done->set_req(i, res); 814 } 815 } 816 } 817 _igvn._worklist.push(sfpt_done); 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 // Don't do scalar replacement if the frame can be popped by JVMTI: 971 // if reallocation fails during deoptimization we'll pop all 972 // interpreter frames for this compiled frame and that won't play 973 // nice with JVMTI popframe. 974 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) { 975 return false; 976 } 977 Node* klass = alloc->in(AllocateNode::KlassNode); 978 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); 979 Node* res = alloc->result_cast(); 980 // Eliminate boxing allocations which are not used 981 // regardless scalar replacable status. 982 bool boxing_alloc = C->eliminate_boxing() && 983 tklass->klass()->is_instance_klass() && 984 tklass->klass()->as_instance_klass()->is_box_klass(); 985 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) { 986 return false; 987 } 988 989 extract_call_projections(alloc); 990 991 GrowableArray <SafePointNode *> safepoints; 992 if (!can_eliminate_allocation(alloc, safepoints)) { 993 return false; 994 } 995 996 if (!alloc->_is_scalar_replaceable) { 997 assert(res == NULL, "sanity"); 998 // We can only eliminate allocation if all debug info references 999 // are already replaced with SafePointScalarObject because 1000 // we can't search for a fields value without instance_id. 1001 if (safepoints.length() > 0) { 1002 return false; 1003 } 1004 } 1005 1006 if (!scalar_replacement(alloc, safepoints)) { 1007 return false; 1008 } 1009 1010 CompileLog* log = C->log(); 1011 if (log != NULL) { 1012 log->head("eliminate_allocation type='%d'", 1013 log->identify(tklass->klass())); 1014 JVMState* p = alloc->jvms(); 1015 while (p != NULL) { 1016 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1017 p = p->caller(); 1018 } 1019 log->tail("eliminate_allocation"); 1020 } 1021 1022 process_users_of_allocation(alloc); 1023 1024 #ifndef PRODUCT 1025 if (PrintEliminateAllocations) { 1026 if (alloc->is_AllocateArray()) 1027 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1028 else 1029 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1030 } 1031 #endif 1032 1033 return true; 1034 } 1035 1036 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) { 1037 // EA should remove all uses of non-escaping boxing node. 1038 if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) { 1039 return false; 1040 } 1041 1042 assert(boxing->result_cast() == NULL, "unexpected boxing node result"); 1043 1044 extract_call_projections(boxing); 1045 1046 const TypeTuple* r = boxing->tf()->range(); 1047 assert(r->cnt() > TypeFunc::Parms, "sanity"); 1048 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr(); 1049 assert(t != NULL, "sanity"); 1050 1051 CompileLog* log = C->log(); 1052 if (log != NULL) { 1053 log->head("eliminate_boxing type='%d'", 1054 log->identify(t->klass())); 1055 JVMState* p = boxing->jvms(); 1056 while (p != NULL) { 1057 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1058 p = p->caller(); 1059 } 1060 log->tail("eliminate_boxing"); 1061 } 1062 1063 process_users_of_allocation(boxing); 1064 1065 #ifndef PRODUCT 1066 if (PrintEliminateAllocations) { 1067 tty->print("++++ Eliminated: %d ", boxing->_idx); 1068 boxing->method()->print_short_name(tty); 1069 tty->cr(); 1070 } 1071 #endif 1072 1073 return true; 1074 } 1075 1076 //---------------------------set_eden_pointers------------------------- 1077 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { 1078 if (UseTLAB) { // Private allocation: load from TLS 1079 Node* thread = transform_later(new ThreadLocalNode()); 1080 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); 1081 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); 1082 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); 1083 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); 1084 } else { // Shared allocation: load from globals 1085 CollectedHeap* ch = Universe::heap(); 1086 address top_adr = (address)ch->top_addr(); 1087 address end_adr = (address)ch->end_addr(); 1088 eden_top_adr = makecon(TypeRawPtr::make(top_adr)); 1089 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); 1090 } 1091 } 1092 1093 1094 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 1095 Node* adr = basic_plus_adr(base, offset); 1096 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 1097 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered); 1098 transform_later(value); 1099 return value; 1100 } 1101 1102 1103 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 1104 Node* adr = basic_plus_adr(base, offset); 1105 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered); 1106 transform_later(mem); 1107 return mem; 1108 } 1109 1110 //============================================================================= 1111 // 1112 // A L L O C A T I O N 1113 // 1114 // Allocation attempts to be fast in the case of frequent small objects. 1115 // It breaks down like this: 1116 // 1117 // 1) Size in doublewords is computed. This is a constant for objects and 1118 // variable for most arrays. Doubleword units are used to avoid size 1119 // overflow of huge doubleword arrays. We need doublewords in the end for 1120 // rounding. 1121 // 1122 // 2) Size is checked for being 'too large'. Too-large allocations will go 1123 // the slow path into the VM. The slow path can throw any required 1124 // exceptions, and does all the special checks for very large arrays. The 1125 // size test can constant-fold away for objects. For objects with 1126 // finalizers it constant-folds the otherway: you always go slow with 1127 // finalizers. 1128 // 1129 // 3) If NOT using TLABs, this is the contended loop-back point. 1130 // Load-Locked the heap top. If using TLABs normal-load the heap top. 1131 // 1132 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 1133 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 1134 // "size*8" we always enter the VM, where "largish" is a constant picked small 1135 // enough that there's always space between the eden max and 4Gig (old space is 1136 // there so it's quite large) and large enough that the cost of entering the VM 1137 // is dwarfed by the cost to initialize the space. 1138 // 1139 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 1140 // down. If contended, repeat at step 3. If using TLABs normal-store 1141 // adjusted heap top back down; there is no contention. 1142 // 1143 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 1144 // fields. 1145 // 1146 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 1147 // oop flavor. 1148 // 1149 //============================================================================= 1150 // FastAllocateSizeLimit value is in DOUBLEWORDS. 1151 // Allocations bigger than this always go the slow route. 1152 // This value must be small enough that allocation attempts that need to 1153 // trigger exceptions go the slow route. Also, it must be small enough so 1154 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 1155 //=============================================================================j// 1156 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 1157 // The allocator will coalesce int->oop copies away. See comment in 1158 // coalesce.cpp about how this works. It depends critically on the exact 1159 // code shape produced here, so if you are changing this code shape 1160 // make sure the GC info for the heap-top is correct in and around the 1161 // slow-path call. 1162 // 1163 1164 void PhaseMacroExpand::expand_allocate_common( 1165 AllocateNode* alloc, // allocation node to be expanded 1166 Node* length, // array length for an array allocation 1167 const TypeFunc* slow_call_type, // Type of slow call 1168 address slow_call_address // Address of slow call 1169 ) 1170 { 1171 1172 Node* ctrl = alloc->in(TypeFunc::Control); 1173 Node* mem = alloc->in(TypeFunc::Memory); 1174 Node* i_o = alloc->in(TypeFunc::I_O); 1175 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 1176 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1177 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 1178 1179 assert(ctrl != NULL, "must have control"); 1180 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 1181 // they will not be used if "always_slow" is set 1182 enum { slow_result_path = 1, fast_result_path = 2 }; 1183 Node *result_region; 1184 Node *result_phi_rawmem; 1185 Node *result_phi_rawoop; 1186 Node *result_phi_i_o; 1187 1188 // The initial slow comparison is a size check, the comparison 1189 // we want to do is a BoolTest::gt 1190 bool always_slow = false; 1191 int tv = _igvn.find_int_con(initial_slow_test, -1); 1192 if (tv >= 0) { 1193 always_slow = (tv == 1); 1194 initial_slow_test = NULL; 1195 } else { 1196 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 1197 } 1198 1199 if (C->env()->dtrace_alloc_probes() || 1200 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc())) { 1201 // Force slow-path allocation 1202 always_slow = true; 1203 initial_slow_test = NULL; 1204 } 1205 1206 1207 enum { too_big_or_final_path = 1, need_gc_path = 2 }; 1208 Node *slow_region = NULL; 1209 Node *toobig_false = ctrl; 1210 1211 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); 1212 // generate the initial test if necessary 1213 if (initial_slow_test != NULL ) { 1214 slow_region = new RegionNode(3); 1215 1216 // Now make the initial failure test. Usually a too-big test but 1217 // might be a TRUE for finalizers or a fancy class check for 1218 // newInstance0. 1219 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1220 transform_later(toobig_iff); 1221 // Plug the failing-too-big test into the slow-path region 1222 Node *toobig_true = new IfTrueNode( toobig_iff ); 1223 transform_later(toobig_true); 1224 slow_region ->init_req( too_big_or_final_path, toobig_true ); 1225 toobig_false = new IfFalseNode( toobig_iff ); 1226 transform_later(toobig_false); 1227 } else { // No initial test, just fall into next case 1228 toobig_false = ctrl; 1229 debug_only(slow_region = NodeSentinel); 1230 } 1231 1232 Node *slow_mem = mem; // save the current memory state for slow path 1233 // generate the fast allocation code unless we know that the initial test will always go slow 1234 if (!always_slow) { 1235 // Fast path modifies only raw memory. 1236 if (mem->is_MergeMem()) { 1237 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1238 } 1239 1240 Node* eden_top_adr; 1241 Node* eden_end_adr; 1242 1243 set_eden_pointers(eden_top_adr, eden_end_adr); 1244 1245 // Load Eden::end. Loop invariant and hoisted. 1246 // 1247 // Note: We set the control input on "eden_end" and "old_eden_top" when using 1248 // a TLAB to work around a bug where these values were being moved across 1249 // a safepoint. These are not oops, so they cannot be include in the oop 1250 // map, but they can be changed by a GC. The proper way to fix this would 1251 // be to set the raw memory state when generating a SafepointNode. However 1252 // this will require extensive changes to the loop optimization in order to 1253 // prevent a degradation of the optimization. 1254 // See comment in memnode.hpp, around line 227 in class LoadPNode. 1255 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 1256 1257 // allocate the Region and Phi nodes for the result 1258 result_region = new RegionNode(3); 1259 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1260 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM); 1261 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch 1262 1263 // We need a Region for the loop-back contended case. 1264 enum { fall_in_path = 1, contended_loopback_path = 2 }; 1265 Node *contended_region; 1266 Node *contended_phi_rawmem; 1267 if (UseTLAB) { 1268 contended_region = toobig_false; 1269 contended_phi_rawmem = mem; 1270 } else { 1271 contended_region = new RegionNode(3); 1272 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1273 // Now handle the passing-too-big test. We fall into the contended 1274 // loop-back merge point. 1275 contended_region ->init_req(fall_in_path, toobig_false); 1276 contended_phi_rawmem->init_req(fall_in_path, mem); 1277 transform_later(contended_region); 1278 transform_later(contended_phi_rawmem); 1279 } 1280 1281 // Load(-locked) the heap top. 1282 // See note above concerning the control input when using a TLAB 1283 Node *old_eden_top = UseTLAB 1284 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered) 1285 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire); 1286 1287 transform_later(old_eden_top); 1288 // Add to heap top to get a new heap top 1289 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes); 1290 transform_later(new_eden_top); 1291 // Check for needing a GC; compare against heap end 1292 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end); 1293 transform_later(needgc_cmp); 1294 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge); 1295 transform_later(needgc_bol); 1296 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); 1297 transform_later(needgc_iff); 1298 1299 // Plug the failing-heap-space-need-gc test into the slow-path region 1300 Node *needgc_true = new IfTrueNode(needgc_iff); 1301 transform_later(needgc_true); 1302 if (initial_slow_test) { 1303 slow_region->init_req(need_gc_path, needgc_true); 1304 // This completes all paths into the slow merge point 1305 transform_later(slow_region); 1306 } else { // No initial slow path needed! 1307 // Just fall from the need-GC path straight into the VM call. 1308 slow_region = needgc_true; 1309 } 1310 // No need for a GC. Setup for the Store-Conditional 1311 Node *needgc_false = new IfFalseNode(needgc_iff); 1312 transform_later(needgc_false); 1313 1314 // Grab regular I/O before optional prefetch may change it. 1315 // Slow-path does no I/O so just set it to the original I/O. 1316 result_phi_i_o->init_req(slow_result_path, i_o); 1317 1318 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem, 1319 old_eden_top, new_eden_top, length); 1320 1321 // Name successful fast-path variables 1322 Node* fast_oop = old_eden_top; 1323 Node* fast_oop_ctrl; 1324 Node* fast_oop_rawmem; 1325 1326 // Store (-conditional) the modified eden top back down. 1327 // StorePConditional produces flags for a test PLUS a modified raw 1328 // memory state. 1329 if (UseTLAB) { 1330 Node* store_eden_top = 1331 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1332 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered); 1333 transform_later(store_eden_top); 1334 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path 1335 fast_oop_rawmem = store_eden_top; 1336 } else { 1337 Node* store_eden_top = 1338 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1339 new_eden_top, fast_oop/*old_eden_top*/); 1340 transform_later(store_eden_top); 1341 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne); 1342 transform_later(contention_check); 1343 store_eden_top = new SCMemProjNode(store_eden_top); 1344 transform_later(store_eden_top); 1345 1346 // If not using TLABs, check to see if there was contention. 1347 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN); 1348 transform_later(contention_iff); 1349 Node *contention_true = new IfTrueNode(contention_iff); 1350 transform_later(contention_true); 1351 // If contention, loopback and try again. 1352 contended_region->init_req(contended_loopback_path, contention_true); 1353 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top); 1354 1355 // Fast-path succeeded with no contention! 1356 Node *contention_false = new IfFalseNode(contention_iff); 1357 transform_later(contention_false); 1358 fast_oop_ctrl = contention_false; 1359 1360 // Bump total allocated bytes for this thread 1361 Node* thread = new ThreadLocalNode(); 1362 transform_later(thread); 1363 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread, 1364 in_bytes(JavaThread::allocated_bytes_offset())); 1365 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1366 0, TypeLong::LONG, T_LONG); 1367 #ifdef _LP64 1368 Node* alloc_size = size_in_bytes; 1369 #else 1370 Node* alloc_size = new ConvI2LNode(size_in_bytes); 1371 transform_later(alloc_size); 1372 #endif 1373 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size); 1374 transform_later(new_alloc_bytes); 1375 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1376 0, new_alloc_bytes, T_LONG); 1377 } 1378 1379 InitializeNode* init = alloc->initialization(); 1380 fast_oop_rawmem = initialize_object(alloc, 1381 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1382 klass_node, length, size_in_bytes); 1383 1384 // If initialization is performed by an array copy, any required 1385 // MemBarStoreStore was already added. If the object does not 1386 // escape no need for a MemBarStoreStore. Otherwise we need a 1387 // MemBarStoreStore so that stores that initialize this object 1388 // can't be reordered with a subsequent store that makes this 1389 // object accessible by other threads. 1390 if (init == NULL || (!init->is_complete_with_arraycopy() && !init->does_not_escape())) { 1391 if (init == NULL || init->req() < InitializeNode::RawStores) { 1392 // No InitializeNode or no stores captured by zeroing 1393 // elimination. Simply add the MemBarStoreStore after object 1394 // initialization. 1395 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1396 transform_later(mb); 1397 1398 mb->init_req(TypeFunc::Memory, fast_oop_rawmem); 1399 mb->init_req(TypeFunc::Control, fast_oop_ctrl); 1400 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control); 1401 transform_later(fast_oop_ctrl); 1402 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory); 1403 transform_later(fast_oop_rawmem); 1404 } else { 1405 // Add the MemBarStoreStore after the InitializeNode so that 1406 // all stores performing the initialization that were moved 1407 // before the InitializeNode happen before the storestore 1408 // barrier. 1409 1410 Node* init_ctrl = init->proj_out(TypeFunc::Control); 1411 Node* init_mem = init->proj_out(TypeFunc::Memory); 1412 1413 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1414 transform_later(mb); 1415 1416 Node* ctrl = new ProjNode(init,TypeFunc::Control); 1417 transform_later(ctrl); 1418 Node* mem = new ProjNode(init,TypeFunc::Memory); 1419 transform_later(mem); 1420 1421 // The MemBarStoreStore depends on control and memory coming 1422 // from the InitializeNode 1423 mb->init_req(TypeFunc::Memory, mem); 1424 mb->init_req(TypeFunc::Control, ctrl); 1425 1426 ctrl = new ProjNode(mb,TypeFunc::Control); 1427 transform_later(ctrl); 1428 mem = new ProjNode(mb,TypeFunc::Memory); 1429 transform_later(mem); 1430 1431 // All nodes that depended on the InitializeNode for control 1432 // and memory must now depend on the MemBarNode that itself 1433 // depends on the InitializeNode 1434 _igvn.replace_node(init_ctrl, ctrl); 1435 _igvn.replace_node(init_mem, mem); 1436 } 1437 } 1438 1439 if (C->env()->dtrace_extended_probes()) { 1440 // Slow-path call 1441 int size = TypeFunc::Parms + 2; 1442 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1443 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), 1444 "dtrace_object_alloc", 1445 TypeRawPtr::BOTTOM); 1446 1447 // Get base of thread-local storage area 1448 Node* thread = new ThreadLocalNode(); 1449 transform_later(thread); 1450 1451 call->init_req(TypeFunc::Parms+0, thread); 1452 call->init_req(TypeFunc::Parms+1, fast_oop); 1453 call->init_req(TypeFunc::Control, fast_oop_ctrl); 1454 call->init_req(TypeFunc::I_O , top()); // does no i/o 1455 call->init_req(TypeFunc::Memory , fast_oop_rawmem); 1456 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1457 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1458 transform_later(call); 1459 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control); 1460 transform_later(fast_oop_ctrl); 1461 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory); 1462 transform_later(fast_oop_rawmem); 1463 } 1464 1465 // Plug in the successful fast-path into the result merge point 1466 result_region ->init_req(fast_result_path, fast_oop_ctrl); 1467 result_phi_rawoop->init_req(fast_result_path, fast_oop); 1468 result_phi_i_o ->init_req(fast_result_path, i_o); 1469 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); 1470 } else { 1471 slow_region = ctrl; 1472 result_phi_i_o = i_o; // Rename it to use in the following code. 1473 } 1474 1475 // Generate slow-path call 1476 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address, 1477 OptoRuntime::stub_name(slow_call_address), 1478 alloc->jvms()->bci(), 1479 TypePtr::BOTTOM); 1480 call->init_req( TypeFunc::Control, slow_region ); 1481 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1482 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs 1483 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1484 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1485 1486 call->init_req(TypeFunc::Parms+0, klass_node); 1487 if (length != NULL) { 1488 call->init_req(TypeFunc::Parms+1, length); 1489 } 1490 1491 // Copy debug information and adjust JVMState information, then replace 1492 // allocate node with the call 1493 copy_call_debug_info((CallNode *) alloc, call); 1494 if (!always_slow) { 1495 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1496 } else { 1497 // Hook i_o projection to avoid its elimination during allocation 1498 // replacement (when only a slow call is generated). 1499 call->set_req(TypeFunc::I_O, result_phi_i_o); 1500 } 1501 _igvn.replace_node(alloc, call); 1502 transform_later(call); 1503 1504 // Identify the output projections from the allocate node and 1505 // adjust any references to them. 1506 // The control and io projections look like: 1507 // 1508 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1509 // Allocate Catch 1510 // ^---Proj(io) <-------+ ^---CatchProj(io) 1511 // 1512 // We are interested in the CatchProj nodes. 1513 // 1514 extract_call_projections(call); 1515 1516 // An allocate node has separate memory projections for the uses on 1517 // the control and i_o paths. Replace the control memory projection with 1518 // result_phi_rawmem (unless we are only generating a slow call when 1519 // both memory projections are combined) 1520 if (!always_slow && _memproj_fallthrough != NULL) { 1521 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { 1522 Node *use = _memproj_fallthrough->fast_out(i); 1523 _igvn.rehash_node_delayed(use); 1524 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); 1525 // back up iterator 1526 --i; 1527 } 1528 } 1529 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete 1530 // _memproj_catchall so we end up with a call that has only 1 memory projection. 1531 if (_memproj_catchall != NULL ) { 1532 if (_memproj_fallthrough == NULL) { 1533 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory); 1534 transform_later(_memproj_fallthrough); 1535 } 1536 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { 1537 Node *use = _memproj_catchall->fast_out(i); 1538 _igvn.rehash_node_delayed(use); 1539 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); 1540 // back up iterator 1541 --i; 1542 } 1543 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted"); 1544 _igvn.remove_dead_node(_memproj_catchall); 1545 } 1546 1547 // An allocate node has separate i_o projections for the uses on the control 1548 // and i_o paths. Always replace the control i_o projection with result i_o 1549 // otherwise incoming i_o become dead when only a slow call is generated 1550 // (it is different from memory projections where both projections are 1551 // combined in such case). 1552 if (_ioproj_fallthrough != NULL) { 1553 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { 1554 Node *use = _ioproj_fallthrough->fast_out(i); 1555 _igvn.rehash_node_delayed(use); 1556 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); 1557 // back up iterator 1558 --i; 1559 } 1560 } 1561 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete 1562 // _ioproj_catchall so we end up with a call that has only 1 i_o projection. 1563 if (_ioproj_catchall != NULL ) { 1564 if (_ioproj_fallthrough == NULL) { 1565 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O); 1566 transform_later(_ioproj_fallthrough); 1567 } 1568 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { 1569 Node *use = _ioproj_catchall->fast_out(i); 1570 _igvn.rehash_node_delayed(use); 1571 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); 1572 // back up iterator 1573 --i; 1574 } 1575 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted"); 1576 _igvn.remove_dead_node(_ioproj_catchall); 1577 } 1578 1579 // if we generated only a slow call, we are done 1580 if (always_slow) { 1581 // Now we can unhook i_o. 1582 if (result_phi_i_o->outcnt() > 1) { 1583 call->set_req(TypeFunc::I_O, top()); 1584 } else { 1585 assert(result_phi_i_o->unique_ctrl_out() == call, ""); 1586 // Case of new array with negative size known during compilation. 1587 // AllocateArrayNode::Ideal() optimization disconnect unreachable 1588 // following code since call to runtime will throw exception. 1589 // As result there will be no users of i_o after the call. 1590 // Leave i_o attached to this call to avoid problems in preceding graph. 1591 } 1592 return; 1593 } 1594 1595 1596 if (_fallthroughcatchproj != NULL) { 1597 ctrl = _fallthroughcatchproj->clone(); 1598 transform_later(ctrl); 1599 _igvn.replace_node(_fallthroughcatchproj, result_region); 1600 } else { 1601 ctrl = top(); 1602 } 1603 Node *slow_result; 1604 if (_resproj == NULL) { 1605 // no uses of the allocation result 1606 slow_result = top(); 1607 } else { 1608 slow_result = _resproj->clone(); 1609 transform_later(slow_result); 1610 _igvn.replace_node(_resproj, result_phi_rawoop); 1611 } 1612 1613 // Plug slow-path into result merge point 1614 result_region ->init_req( slow_result_path, ctrl ); 1615 result_phi_rawoop->init_req( slow_result_path, slow_result); 1616 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); 1617 transform_later(result_region); 1618 transform_later(result_phi_rawoop); 1619 transform_later(result_phi_rawmem); 1620 transform_later(result_phi_i_o); 1621 // This completes all paths into the result merge point 1622 } 1623 1624 1625 // Helper for PhaseMacroExpand::expand_allocate_common. 1626 // Initializes the newly-allocated storage. 1627 Node* 1628 PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1629 Node* control, Node* rawmem, Node* object, 1630 Node* klass_node, Node* length, 1631 Node* size_in_bytes) { 1632 InitializeNode* init = alloc->initialization(); 1633 // Store the klass & mark bits 1634 Node* mark_node = NULL; 1635 // For now only enable fast locking for non-array types 1636 if (UseBiasedLocking && (length == NULL)) { 1637 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 1638 } else { 1639 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); 1640 } 1641 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 1642 1643 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 1644 int header_size = alloc->minimum_header_size(); // conservatively small 1645 1646 // Array length 1647 if (length != NULL) { // Arrays need length field 1648 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1649 // conservatively small header size: 1650 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1651 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1652 if (k->is_array_klass()) // we know the exact header size in most cases: 1653 header_size = Klass::layout_helper_header_size(k->layout_helper()); 1654 } 1655 1656 // Clear the object body, if necessary. 1657 if (init == NULL) { 1658 // The init has somehow disappeared; be cautious and clear everything. 1659 // 1660 // This can happen if a node is allocated but an uncommon trap occurs 1661 // immediately. In this case, the Initialize gets associated with the 1662 // trap, and may be placed in a different (outer) loop, if the Allocate 1663 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1664 // there can be two Allocates to one Initialize. The answer in all these 1665 // edge cases is safety first. It is always safe to clear immediately 1666 // within an Allocate, and then (maybe or maybe not) clear some more later. 1667 if (!ZeroTLAB) 1668 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1669 header_size, size_in_bytes, 1670 &_igvn); 1671 } else { 1672 if (!init->is_complete()) { 1673 // Try to win by zeroing only what the init does not store. 1674 // We can also try to do some peephole optimizations, 1675 // such as combining some adjacent subword stores. 1676 rawmem = init->complete_stores(control, rawmem, object, 1677 header_size, size_in_bytes, &_igvn); 1678 } 1679 // We have no more use for this link, since the AllocateNode goes away: 1680 init->set_req(InitializeNode::RawAddress, top()); 1681 // (If we keep the link, it just confuses the register allocator, 1682 // who thinks he sees a real use of the address by the membar.) 1683 } 1684 1685 return rawmem; 1686 } 1687 1688 // Generate prefetch instructions for next allocations. 1689 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1690 Node*& contended_phi_rawmem, 1691 Node* old_eden_top, Node* new_eden_top, 1692 Node* length) { 1693 enum { fall_in_path = 1, pf_path = 2 }; 1694 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1695 // Generate prefetch allocation with watermark check. 1696 // As an allocation hits the watermark, we will prefetch starting 1697 // at a "distance" away from watermark. 1698 1699 Node *pf_region = new RegionNode(3); 1700 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 1701 TypeRawPtr::BOTTOM ); 1702 // I/O is used for Prefetch 1703 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO ); 1704 1705 Node *thread = new ThreadLocalNode(); 1706 transform_later(thread); 1707 1708 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread, 1709 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1710 transform_later(eden_pf_adr); 1711 1712 Node *old_pf_wm = new LoadPNode(needgc_false, 1713 contended_phi_rawmem, eden_pf_adr, 1714 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, 1715 MemNode::unordered); 1716 transform_later(old_pf_wm); 1717 1718 // check against new_eden_top 1719 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm ); 1720 transform_later(need_pf_cmp); 1721 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge ); 1722 transform_later(need_pf_bol); 1723 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol, 1724 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1725 transform_later(need_pf_iff); 1726 1727 // true node, add prefetchdistance 1728 Node *need_pf_true = new IfTrueNode( need_pf_iff ); 1729 transform_later(need_pf_true); 1730 1731 Node *need_pf_false = new IfFalseNode( need_pf_iff ); 1732 transform_later(need_pf_false); 1733 1734 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm, 1735 _igvn.MakeConX(AllocatePrefetchDistance) ); 1736 transform_later(new_pf_wmt ); 1737 new_pf_wmt->set_req(0, need_pf_true); 1738 1739 Node *store_new_wmt = new StorePNode(need_pf_true, 1740 contended_phi_rawmem, eden_pf_adr, 1741 TypeRawPtr::BOTTOM, new_pf_wmt, 1742 MemNode::unordered); 1743 transform_later(store_new_wmt); 1744 1745 // adding prefetches 1746 pf_phi_abio->init_req( fall_in_path, i_o ); 1747 1748 Node *prefetch_adr; 1749 Node *prefetch; 1750 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize; 1751 uint step_size = AllocatePrefetchStepSize; 1752 uint distance = 0; 1753 1754 for ( uint i = 0; i < lines; i++ ) { 1755 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt, 1756 _igvn.MakeConX(distance) ); 1757 transform_later(prefetch_adr); 1758 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1759 transform_later(prefetch); 1760 distance += step_size; 1761 i_o = prefetch; 1762 } 1763 pf_phi_abio->set_req( pf_path, i_o ); 1764 1765 pf_region->init_req( fall_in_path, need_pf_false ); 1766 pf_region->init_req( pf_path, need_pf_true ); 1767 1768 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 1769 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 1770 1771 transform_later(pf_region); 1772 transform_later(pf_phi_rawmem); 1773 transform_later(pf_phi_abio); 1774 1775 needgc_false = pf_region; 1776 contended_phi_rawmem = pf_phi_rawmem; 1777 i_o = pf_phi_abio; 1778 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { 1779 // Insert a prefetch for each allocation. 1780 // This code is used for Sparc with BIS. 1781 Node *pf_region = new RegionNode(3); 1782 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 1783 TypeRawPtr::BOTTOM ); 1784 transform_later(pf_region); 1785 1786 // Generate several prefetch instructions. 1787 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1788 uint step_size = AllocatePrefetchStepSize; 1789 uint distance = AllocatePrefetchDistance; 1790 1791 // Next cache address. 1792 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top, 1793 _igvn.MakeConX(distance)); 1794 transform_later(cache_adr); 1795 cache_adr = new CastP2XNode(needgc_false, cache_adr); 1796 transform_later(cache_adr); 1797 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); 1798 cache_adr = new AndXNode(cache_adr, mask); 1799 transform_later(cache_adr); 1800 cache_adr = new CastX2PNode(cache_adr); 1801 transform_later(cache_adr); 1802 1803 // Prefetch 1804 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); 1805 prefetch->set_req(0, needgc_false); 1806 transform_later(prefetch); 1807 contended_phi_rawmem = prefetch; 1808 Node *prefetch_adr; 1809 distance = step_size; 1810 for ( uint i = 1; i < lines; i++ ) { 1811 prefetch_adr = new AddPNode( cache_adr, cache_adr, 1812 _igvn.MakeConX(distance) ); 1813 transform_later(prefetch_adr); 1814 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); 1815 transform_later(prefetch); 1816 distance += step_size; 1817 contended_phi_rawmem = prefetch; 1818 } 1819 } else if( AllocatePrefetchStyle > 0 ) { 1820 // Insert a prefetch for each allocation only on the fast-path 1821 Node *prefetch_adr; 1822 Node *prefetch; 1823 // Generate several prefetch instructions. 1824 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1825 uint step_size = AllocatePrefetchStepSize; 1826 uint distance = AllocatePrefetchDistance; 1827 for ( uint i = 0; i < lines; i++ ) { 1828 prefetch_adr = new AddPNode( old_eden_top, new_eden_top, 1829 _igvn.MakeConX(distance) ); 1830 transform_later(prefetch_adr); 1831 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1832 // Do not let it float too high, since if eden_top == eden_end, 1833 // both might be null. 1834 if( i == 0 ) { // Set control for first prefetch, next follows it 1835 prefetch->init_req(0, needgc_false); 1836 } 1837 transform_later(prefetch); 1838 distance += step_size; 1839 i_o = prefetch; 1840 } 1841 } 1842 return i_o; 1843 } 1844 1845 1846 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 1847 expand_allocate_common(alloc, NULL, 1848 OptoRuntime::new_instance_Type(), 1849 OptoRuntime::new_instance_Java()); 1850 } 1851 1852 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 1853 Node* length = alloc->in(AllocateNode::ALength); 1854 InitializeNode* init = alloc->initialization(); 1855 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1856 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1857 address slow_call_address; // Address of slow call 1858 if (init != NULL && init->is_complete_with_arraycopy() && 1859 k->is_type_array_klass()) { 1860 // Don't zero type array during slow allocation in VM since 1861 // it will be initialized later by arraycopy in compiled code. 1862 slow_call_address = OptoRuntime::new_array_nozero_Java(); 1863 } else { 1864 slow_call_address = OptoRuntime::new_array_Java(); 1865 } 1866 expand_allocate_common(alloc, length, 1867 OptoRuntime::new_array_Type(), 1868 slow_call_address); 1869 } 1870 1871 //-------------------mark_eliminated_box---------------------------------- 1872 // 1873 // During EA obj may point to several objects but after few ideal graph 1874 // transformations (CCP) it may point to only one non escaping object 1875 // (but still using phi), corresponding locks and unlocks will be marked 1876 // for elimination. Later obj could be replaced with a new node (new phi) 1877 // and which does not have escape information. And later after some graph 1878 // reshape other locks and unlocks (which were not marked for elimination 1879 // before) are connected to this new obj (phi) but they still will not be 1880 // marked for elimination since new obj has no escape information. 1881 // Mark all associated (same box and obj) lock and unlock nodes for 1882 // elimination if some of them marked already. 1883 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) { 1884 if (oldbox->as_BoxLock()->is_eliminated()) 1885 return; // This BoxLock node was processed already. 1886 1887 // New implementation (EliminateNestedLocks) has separate BoxLock 1888 // node for each locked region so mark all associated locks/unlocks as 1889 // eliminated even if different objects are referenced in one locked region 1890 // (for example, OSR compilation of nested loop inside locked scope). 1891 if (EliminateNestedLocks || 1892 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) { 1893 // Box is used only in one lock region. Mark this box as eliminated. 1894 _igvn.hash_delete(oldbox); 1895 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value 1896 _igvn.hash_insert(oldbox); 1897 1898 for (uint i = 0; i < oldbox->outcnt(); i++) { 1899 Node* u = oldbox->raw_out(i); 1900 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) { 1901 AbstractLockNode* alock = u->as_AbstractLock(); 1902 // Check lock's box since box could be referenced by Lock's debug info. 1903 if (alock->box_node() == oldbox) { 1904 // Mark eliminated all related locks and unlocks. 1905 #ifdef ASSERT 1906 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4"); 1907 #endif 1908 alock->set_non_esc_obj(); 1909 } 1910 } 1911 } 1912 return; 1913 } 1914 1915 // Create new "eliminated" BoxLock node and use it in monitor debug info 1916 // instead of oldbox for the same object. 1917 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 1918 1919 // Note: BoxLock node is marked eliminated only here and it is used 1920 // to indicate that all associated lock and unlock nodes are marked 1921 // for elimination. 1922 newbox->set_eliminated(); 1923 transform_later(newbox); 1924 1925 // Replace old box node with new box for all users of the same object. 1926 for (uint i = 0; i < oldbox->outcnt();) { 1927 bool next_edge = true; 1928 1929 Node* u = oldbox->raw_out(i); 1930 if (u->is_AbstractLock()) { 1931 AbstractLockNode* alock = u->as_AbstractLock(); 1932 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) { 1933 // Replace Box and mark eliminated all related locks and unlocks. 1934 #ifdef ASSERT 1935 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5"); 1936 #endif 1937 alock->set_non_esc_obj(); 1938 _igvn.rehash_node_delayed(alock); 1939 alock->set_box_node(newbox); 1940 next_edge = false; 1941 } 1942 } 1943 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) { 1944 FastLockNode* flock = u->as_FastLock(); 1945 assert(flock->box_node() == oldbox, "sanity"); 1946 _igvn.rehash_node_delayed(flock); 1947 flock->set_box_node(newbox); 1948 next_edge = false; 1949 } 1950 1951 // Replace old box in monitor debug info. 1952 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 1953 SafePointNode* sfn = u->as_SafePoint(); 1954 JVMState* youngest_jvms = sfn->jvms(); 1955 int max_depth = youngest_jvms->depth(); 1956 for (int depth = 1; depth <= max_depth; depth++) { 1957 JVMState* jvms = youngest_jvms->of_depth(depth); 1958 int num_mon = jvms->nof_monitors(); 1959 // Loop over monitors 1960 for (int idx = 0; idx < num_mon; idx++) { 1961 Node* obj_node = sfn->monitor_obj(jvms, idx); 1962 Node* box_node = sfn->monitor_box(jvms, idx); 1963 if (box_node == oldbox && obj_node->eqv_uncast(obj)) { 1964 int j = jvms->monitor_box_offset(idx); 1965 _igvn.replace_input_of(u, j, newbox); 1966 next_edge = false; 1967 } 1968 } 1969 } 1970 } 1971 if (next_edge) i++; 1972 } 1973 } 1974 1975 //-----------------------mark_eliminated_locking_nodes----------------------- 1976 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { 1977 if (EliminateNestedLocks) { 1978 if (alock->is_nested()) { 1979 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity"); 1980 return; 1981 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened 1982 // Only Lock node has JVMState needed here. 1983 // Not that preceding claim is documented anywhere else. 1984 if (alock->jvms() != NULL) { 1985 if (alock->as_Lock()->is_nested_lock_region()) { 1986 // Mark eliminated related nested locks and unlocks. 1987 Node* obj = alock->obj_node(); 1988 BoxLockNode* box_node = alock->box_node()->as_BoxLock(); 1989 assert(!box_node->is_eliminated(), "should not be marked yet"); 1990 // Note: BoxLock node is marked eliminated only here 1991 // and it is used to indicate that all associated lock 1992 // and unlock nodes are marked for elimination. 1993 box_node->set_eliminated(); // Box's hash is always NO_HASH here 1994 for (uint i = 0; i < box_node->outcnt(); i++) { 1995 Node* u = box_node->raw_out(i); 1996 if (u->is_AbstractLock()) { 1997 alock = u->as_AbstractLock(); 1998 if (alock->box_node() == box_node) { 1999 // Verify that this Box is referenced only by related locks. 2000 assert(alock->obj_node()->eqv_uncast(obj), ""); 2001 // Mark all related locks and unlocks. 2002 #ifdef ASSERT 2003 alock->log_lock_optimization(C, "eliminate_lock_set_nested"); 2004 #endif 2005 alock->set_nested(); 2006 } 2007 } 2008 } 2009 } else { 2010 #ifdef ASSERT 2011 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region"); 2012 if (C->log() != NULL) 2013 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output 2014 #endif 2015 } 2016 } 2017 return; 2018 } 2019 // Process locks for non escaping object 2020 assert(alock->is_non_esc_obj(), ""); 2021 } // EliminateNestedLocks 2022 2023 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object 2024 // Look for all locks of this object and mark them and 2025 // corresponding BoxLock nodes as eliminated. 2026 Node* obj = alock->obj_node(); 2027 for (uint j = 0; j < obj->outcnt(); j++) { 2028 Node* o = obj->raw_out(j); 2029 if (o->is_AbstractLock() && 2030 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) { 2031 alock = o->as_AbstractLock(); 2032 Node* box = alock->box_node(); 2033 // Replace old box node with new eliminated box for all users 2034 // of the same object and mark related locks as eliminated. 2035 mark_eliminated_box(box, obj); 2036 } 2037 } 2038 } 2039 } 2040 2041 // we have determined that this lock/unlock can be eliminated, we simply 2042 // eliminate the node without expanding it. 2043 // 2044 // Note: The membar's associated with the lock/unlock are currently not 2045 // eliminated. This should be investigated as a future enhancement. 2046 // 2047 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 2048 2049 if (!alock->is_eliminated()) { 2050 return false; 2051 } 2052 #ifdef ASSERT 2053 if (!alock->is_coarsened()) { 2054 // Check that new "eliminated" BoxLock node is created. 2055 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 2056 assert(oldbox->is_eliminated(), "should be done already"); 2057 } 2058 #endif 2059 2060 alock->log_lock_optimization(C, "eliminate_lock"); 2061 2062 #ifndef PRODUCT 2063 if (PrintEliminateLocks) { 2064 if (alock->is_Lock()) { 2065 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx); 2066 } else { 2067 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx); 2068 } 2069 } 2070 #endif 2071 2072 Node* mem = alock->in(TypeFunc::Memory); 2073 Node* ctrl = alock->in(TypeFunc::Control); 2074 2075 extract_call_projections(alock); 2076 // There are 2 projections from the lock. The lock node will 2077 // be deleted when its last use is subsumed below. 2078 assert(alock->outcnt() == 2 && 2079 _fallthroughproj != NULL && 2080 _memproj_fallthrough != NULL, 2081 "Unexpected projections from Lock/Unlock"); 2082 2083 Node* fallthroughproj = _fallthroughproj; 2084 Node* memproj_fallthrough = _memproj_fallthrough; 2085 2086 // The memory projection from a lock/unlock is RawMem 2087 // The input to a Lock is merged memory, so extract its RawMem input 2088 // (unless the MergeMem has been optimized away.) 2089 if (alock->is_Lock()) { 2090 // Seach for MemBarAcquireLock node and delete it also. 2091 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 2092 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, ""); 2093 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 2094 Node* memproj = membar->proj_out(TypeFunc::Memory); 2095 _igvn.replace_node(ctrlproj, fallthroughproj); 2096 _igvn.replace_node(memproj, memproj_fallthrough); 2097 2098 // Delete FastLock node also if this Lock node is unique user 2099 // (a loop peeling may clone a Lock node). 2100 Node* flock = alock->as_Lock()->fastlock_node(); 2101 if (flock->outcnt() == 1) { 2102 assert(flock->unique_out() == alock, "sanity"); 2103 _igvn.replace_node(flock, top()); 2104 } 2105 } 2106 2107 // Seach for MemBarReleaseLock node and delete it also. 2108 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() && 2109 ctrl->in(0)->is_MemBar()) { 2110 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 2111 assert(membar->Opcode() == Op_MemBarReleaseLock && 2112 mem->is_Proj() && membar == mem->in(0), ""); 2113 _igvn.replace_node(fallthroughproj, ctrl); 2114 _igvn.replace_node(memproj_fallthrough, mem); 2115 fallthroughproj = ctrl; 2116 memproj_fallthrough = mem; 2117 ctrl = membar->in(TypeFunc::Control); 2118 mem = membar->in(TypeFunc::Memory); 2119 } 2120 2121 _igvn.replace_node(fallthroughproj, ctrl); 2122 _igvn.replace_node(memproj_fallthrough, mem); 2123 return true; 2124 } 2125 2126 2127 //------------------------------expand_lock_node---------------------- 2128 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 2129 2130 Node* ctrl = lock->in(TypeFunc::Control); 2131 Node* mem = lock->in(TypeFunc::Memory); 2132 Node* obj = lock->obj_node(); 2133 Node* box = lock->box_node(); 2134 Node* flock = lock->fastlock_node(); 2135 2136 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2137 2138 // Make the merge point 2139 Node *region; 2140 Node *mem_phi; 2141 Node *slow_path; 2142 2143 if (UseOptoBiasInlining) { 2144 /* 2145 * See the full description in MacroAssembler::biased_locking_enter(). 2146 * 2147 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { 2148 * // The object is biased. 2149 * proto_node = klass->prototype_header; 2150 * o_node = thread | proto_node; 2151 * x_node = o_node ^ mark_word; 2152 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? 2153 * // Done. 2154 * } else { 2155 * if( (x_node & biased_lock_mask) != 0 ) { 2156 * // The klass's prototype header is no longer biased. 2157 * cas(&mark_word, mark_word, proto_node) 2158 * goto cas_lock; 2159 * } else { 2160 * // The klass's prototype header is still biased. 2161 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? 2162 * old = mark_word; 2163 * new = o_node; 2164 * } else { 2165 * // Different thread or anonymous biased. 2166 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); 2167 * new = thread | old; 2168 * } 2169 * // Try to rebias. 2170 * if( cas(&mark_word, old, new) == 0 ) { 2171 * // Done. 2172 * } else { 2173 * goto slow_path; // Failed. 2174 * } 2175 * } 2176 * } 2177 * } else { 2178 * // The object is not biased. 2179 * cas_lock: 2180 * if( FastLock(obj) == 0 ) { 2181 * // Done. 2182 * } else { 2183 * slow_path: 2184 * OptoRuntime::complete_monitor_locking_Java(obj); 2185 * } 2186 * } 2187 */ 2188 2189 region = new RegionNode(5); 2190 // create a Phi for the memory state 2191 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2192 2193 Node* fast_lock_region = new RegionNode(3); 2194 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); 2195 2196 // First, check mark word for the biased lock pattern. 2197 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2198 2199 // Get fast path - mark word has the biased lock pattern. 2200 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, 2201 markOopDesc::biased_lock_mask_in_place, 2202 markOopDesc::biased_lock_pattern, true); 2203 // fast_lock_region->in(1) is set to slow path. 2204 fast_lock_mem_phi->init_req(1, mem); 2205 2206 // Now check that the lock is biased to the current thread and has 2207 // the same epoch and bias as Klass::_prototype_header. 2208 2209 // Special-case a fresh allocation to avoid building nodes: 2210 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); 2211 if (klass_node == NULL) { 2212 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 2213 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr())); 2214 #ifdef _LP64 2215 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) { 2216 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity"); 2217 klass_node->in(1)->init_req(0, ctrl); 2218 } else 2219 #endif 2220 klass_node->init_req(0, ctrl); 2221 } 2222 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type()); 2223 2224 Node* thread = transform_later(new ThreadLocalNode()); 2225 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2226 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node)); 2227 Node* x_node = transform_later(new XorXNode(o_node, mark_node)); 2228 2229 // Get slow path - mark word does NOT match the value. 2230 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, 2231 (~markOopDesc::age_mask_in_place), 0); 2232 // region->in(3) is set to fast path - the object is biased to the current thread. 2233 mem_phi->init_req(3, mem); 2234 2235 2236 // Mark word does NOT match the value (thread | Klass::_prototype_header). 2237 2238 2239 // First, check biased pattern. 2240 // Get fast path - _prototype_header has the same biased lock pattern. 2241 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, 2242 markOopDesc::biased_lock_mask_in_place, 0, true); 2243 2244 not_biased_ctrl = fast_lock_region->in(2); // Slow path 2245 // fast_lock_region->in(2) - the prototype header is no longer biased 2246 // and we have to revoke the bias on this object. 2247 // We are going to try to reset the mark of this object to the prototype 2248 // value and fall through to the CAS-based locking scheme. 2249 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 2250 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr, 2251 proto_node, mark_node); 2252 transform_later(cas); 2253 Node* proj = transform_later(new SCMemProjNode(cas)); 2254 fast_lock_mem_phi->init_req(2, proj); 2255 2256 2257 // Second, check epoch bits. 2258 Node* rebiased_region = new RegionNode(3); 2259 Node* old_phi = new PhiNode( rebiased_region, TypeX_X); 2260 Node* new_phi = new PhiNode( rebiased_region, TypeX_X); 2261 2262 // Get slow path - mark word does NOT match epoch bits. 2263 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, 2264 markOopDesc::epoch_mask_in_place, 0); 2265 // The epoch of the current bias is not valid, attempt to rebias the object 2266 // toward the current thread. 2267 rebiased_region->init_req(2, epoch_ctrl); 2268 old_phi->init_req(2, mark_node); 2269 new_phi->init_req(2, o_node); 2270 2271 // rebiased_region->in(1) is set to fast path. 2272 // The epoch of the current bias is still valid but we know 2273 // nothing about the owner; it might be set or it might be clear. 2274 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place | 2275 markOopDesc::age_mask_in_place | 2276 markOopDesc::epoch_mask_in_place); 2277 Node* old = transform_later(new AndXNode(mark_node, cmask)); 2278 cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2279 Node* new_mark = transform_later(new OrXNode(cast_thread, old)); 2280 old_phi->init_req(1, old); 2281 new_phi->init_req(1, new_mark); 2282 2283 transform_later(rebiased_region); 2284 transform_later(old_phi); 2285 transform_later(new_phi); 2286 2287 // Try to acquire the bias of the object using an atomic operation. 2288 // If this fails we will go in to the runtime to revoke the object's bias. 2289 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi); 2290 transform_later(cas); 2291 proj = transform_later(new SCMemProjNode(cas)); 2292 2293 // Get slow path - Failed to CAS. 2294 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); 2295 mem_phi->init_req(4, proj); 2296 // region->in(4) is set to fast path - the object is rebiased to the current thread. 2297 2298 // Failed to CAS. 2299 slow_path = new RegionNode(3); 2300 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); 2301 2302 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control 2303 slow_mem->init_req(1, proj); 2304 2305 // Call CAS-based locking scheme (FastLock node). 2306 2307 transform_later(fast_lock_region); 2308 transform_later(fast_lock_mem_phi); 2309 2310 // Get slow path - FastLock failed to lock the object. 2311 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); 2312 mem_phi->init_req(2, fast_lock_mem_phi); 2313 // region->in(2) is set to fast path - the object is locked to the current thread. 2314 2315 slow_path->init_req(2, ctrl); // Capture slow-control 2316 slow_mem->init_req(2, fast_lock_mem_phi); 2317 2318 transform_later(slow_path); 2319 transform_later(slow_mem); 2320 // Reset lock's memory edge. 2321 lock->set_req(TypeFunc::Memory, slow_mem); 2322 2323 } else { 2324 region = new RegionNode(3); 2325 // create a Phi for the memory state 2326 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2327 2328 // Optimize test; set region slot 2 2329 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); 2330 mem_phi->init_req(2, mem); 2331 } 2332 2333 // Make slow path call 2334 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), 2335 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, 2336 obj, box, NULL); 2337 2338 extract_call_projections(call); 2339 2340 // Slow path can only throw asynchronous exceptions, which are always 2341 // de-opted. So the compiler thinks the slow-call can never throw an 2342 // exception. If it DOES throw an exception we would need the debug 2343 // info removed first (since if it throws there is no monitor). 2344 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2345 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2346 2347 // Capture slow path 2348 // disconnect fall-through projection from call and create a new one 2349 // hook up users of fall-through projection to region 2350 Node *slow_ctrl = _fallthroughproj->clone(); 2351 transform_later(slow_ctrl); 2352 _igvn.hash_delete(_fallthroughproj); 2353 _fallthroughproj->disconnect_inputs(NULL, C); 2354 region->init_req(1, slow_ctrl); 2355 // region inputs are now complete 2356 transform_later(region); 2357 _igvn.replace_node(_fallthroughproj, region); 2358 2359 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory)); 2360 mem_phi->init_req(1, memproj ); 2361 transform_later(mem_phi); 2362 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2363 } 2364 2365 //------------------------------expand_unlock_node---------------------- 2366 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 2367 2368 Node* ctrl = unlock->in(TypeFunc::Control); 2369 Node* mem = unlock->in(TypeFunc::Memory); 2370 Node* obj = unlock->obj_node(); 2371 Node* box = unlock->box_node(); 2372 2373 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2374 2375 // No need for a null check on unlock 2376 2377 // Make the merge point 2378 Node *region; 2379 Node *mem_phi; 2380 2381 if (UseOptoBiasInlining) { 2382 // Check for biased locking unlock case, which is a no-op. 2383 // See the full description in MacroAssembler::biased_locking_exit(). 2384 region = new RegionNode(4); 2385 // create a Phi for the memory state 2386 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2387 mem_phi->init_req(3, mem); 2388 2389 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2390 ctrl = opt_bits_test(ctrl, region, 3, mark_node, 2391 markOopDesc::biased_lock_mask_in_place, 2392 markOopDesc::biased_lock_pattern); 2393 } else { 2394 region = new RegionNode(3); 2395 // create a Phi for the memory state 2396 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2397 } 2398 2399 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box ); 2400 funlock = transform_later( funlock )->as_FastUnlock(); 2401 // Optimize test; set region slot 2 2402 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); 2403 Node *thread = transform_later(new ThreadLocalNode()); 2404 2405 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), 2406 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), 2407 "complete_monitor_unlocking_C", slow_path, obj, box, thread); 2408 2409 extract_call_projections(call); 2410 2411 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2412 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2413 2414 // No exceptions for unlocking 2415 // Capture slow path 2416 // disconnect fall-through projection from call and create a new one 2417 // hook up users of fall-through projection to region 2418 Node *slow_ctrl = _fallthroughproj->clone(); 2419 transform_later(slow_ctrl); 2420 _igvn.hash_delete(_fallthroughproj); 2421 _fallthroughproj->disconnect_inputs(NULL, C); 2422 region->init_req(1, slow_ctrl); 2423 // region inputs are now complete 2424 transform_later(region); 2425 _igvn.replace_node(_fallthroughproj, region); 2426 2427 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) ); 2428 mem_phi->init_req(1, memproj ); 2429 mem_phi->init_req(2, mem); 2430 transform_later(mem_phi); 2431 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2432 } 2433 2434 //---------------------------eliminate_macro_nodes---------------------- 2435 // Eliminate scalar replaced allocations and associated locks. 2436 void PhaseMacroExpand::eliminate_macro_nodes() { 2437 if (C->macro_count() == 0) 2438 return; 2439 2440 // First, attempt to eliminate locks 2441 int cnt = C->macro_count(); 2442 for (int i=0; i < cnt; i++) { 2443 Node *n = C->macro_node(i); 2444 if (n->is_AbstractLock()) { // Lock and Unlock nodes 2445 // Before elimination mark all associated (same box and obj) 2446 // lock and unlock nodes. 2447 mark_eliminated_locking_nodes(n->as_AbstractLock()); 2448 } 2449 } 2450 bool progress = true; 2451 while (progress) { 2452 progress = false; 2453 for (int i = C->macro_count(); i > 0; i--) { 2454 Node * n = C->macro_node(i-1); 2455 bool success = false; 2456 debug_only(int old_macro_count = C->macro_count();); 2457 if (n->is_AbstractLock()) { 2458 success = eliminate_locking_node(n->as_AbstractLock()); 2459 } 2460 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2461 progress = progress || success; 2462 } 2463 } 2464 // Next, attempt to eliminate allocations 2465 _has_locks = false; 2466 progress = true; 2467 while (progress) { 2468 progress = false; 2469 for (int i = C->macro_count(); i > 0; i--) { 2470 Node * n = C->macro_node(i-1); 2471 bool success = false; 2472 debug_only(int old_macro_count = C->macro_count();); 2473 switch (n->class_id()) { 2474 case Node::Class_Allocate: 2475 case Node::Class_AllocateArray: 2476 success = eliminate_allocate_node(n->as_Allocate()); 2477 break; 2478 case Node::Class_CallStaticJava: 2479 success = eliminate_boxing_node(n->as_CallStaticJava()); 2480 break; 2481 case Node::Class_Lock: 2482 case Node::Class_Unlock: 2483 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 2484 _has_locks = true; 2485 break; 2486 case Node::Class_ArrayCopy: 2487 break; 2488 default: 2489 assert(n->Opcode() == Op_LoopLimit || 2490 n->Opcode() == Op_Opaque1 || 2491 n->Opcode() == Op_Opaque2 || 2492 n->Opcode() == Op_Opaque3, "unknown node type in macro list"); 2493 } 2494 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2495 progress = progress || success; 2496 } 2497 } 2498 } 2499 2500 //------------------------------expand_macro_nodes---------------------- 2501 // Returns true if a failure occurred. 2502 bool PhaseMacroExpand::expand_macro_nodes() { 2503 // Last attempt to eliminate macro nodes. 2504 eliminate_macro_nodes(); 2505 2506 // Make sure expansion will not cause node limit to be exceeded. 2507 // Worst case is a macro node gets expanded into about 50 nodes. 2508 // Allow 50% more for optimization. 2509 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) ) 2510 return true; 2511 2512 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. 2513 bool progress = true; 2514 while (progress) { 2515 progress = false; 2516 for (int i = C->macro_count(); i > 0; i--) { 2517 Node * n = C->macro_node(i-1); 2518 bool success = false; 2519 debug_only(int old_macro_count = C->macro_count();); 2520 if (n->Opcode() == Op_LoopLimit) { 2521 // Remove it from macro list and put on IGVN worklist to optimize. 2522 C->remove_macro_node(n); 2523 _igvn._worklist.push(n); 2524 success = true; 2525 } else if (n->Opcode() == Op_CallStaticJava) { 2526 // Remove it from macro list and put on IGVN worklist to optimize. 2527 C->remove_macro_node(n); 2528 _igvn._worklist.push(n); 2529 success = true; 2530 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { 2531 _igvn.replace_node(n, n->in(1)); 2532 success = true; 2533 #if INCLUDE_RTM_OPT 2534 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) { 2535 assert(C->profile_rtm(), "should be used only in rtm deoptimization code"); 2536 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), ""); 2537 Node* cmp = n->unique_out(); 2538 #ifdef ASSERT 2539 // Validate graph. 2540 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), ""); 2541 BoolNode* bol = cmp->unique_out()->as_Bool(); 2542 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() && 2543 (bol->_test._test == BoolTest::ne), ""); 2544 IfNode* ifn = bol->unique_out()->as_If(); 2545 assert((ifn->outcnt() == 2) && 2546 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change), ""); 2547 #endif 2548 Node* repl = n->in(1); 2549 if (!_has_locks) { 2550 // Remove RTM state check if there are no locks in the code. 2551 // Replace input to compare the same value. 2552 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1); 2553 } 2554 _igvn.replace_node(n, repl); 2555 success = true; 2556 #endif 2557 } 2558 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2559 progress = progress || success; 2560 } 2561 } 2562 2563 // expand arraycopy "macro" nodes first 2564 // For ReduceBulkZeroing, we must first process all arraycopy nodes 2565 // before the allocate nodes are expanded. 2566 int macro_idx = C->macro_count() - 1; 2567 while (macro_idx >= 0) { 2568 Node * n = C->macro_node(macro_idx); 2569 assert(n->is_macro(), "only macro nodes expected here"); 2570 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2571 // node is unreachable, so don't try to expand it 2572 C->remove_macro_node(n); 2573 } else if (n->is_ArrayCopy()){ 2574 int macro_count = C->macro_count(); 2575 expand_arraycopy_node(n->as_ArrayCopy()); 2576 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2577 } 2578 if (C->failing()) return true; 2579 macro_idx --; 2580 } 2581 2582 // expand "macro" nodes 2583 // nodes are removed from the macro list as they are processed 2584 while (C->macro_count() > 0) { 2585 int macro_count = C->macro_count(); 2586 Node * n = C->macro_node(macro_count-1); 2587 assert(n->is_macro(), "only macro nodes expected here"); 2588 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2589 // node is unreachable, so don't try to expand it 2590 C->remove_macro_node(n); 2591 continue; 2592 } 2593 switch (n->class_id()) { 2594 case Node::Class_Allocate: 2595 expand_allocate(n->as_Allocate()); 2596 break; 2597 case Node::Class_AllocateArray: 2598 expand_allocate_array(n->as_AllocateArray()); 2599 break; 2600 case Node::Class_Lock: 2601 expand_lock_node(n->as_Lock()); 2602 break; 2603 case Node::Class_Unlock: 2604 expand_unlock_node(n->as_Unlock()); 2605 break; 2606 default: 2607 assert(false, "unknown node type in macro list"); 2608 } 2609 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2610 if (C->failing()) return true; 2611 } 2612 2613 _igvn.set_delay_transform(false); 2614 _igvn.optimize(); 2615 if (C->failing()) return true; 2616 return false; 2617 }