1 /* 2 * Copyright (c) 2005, 2011, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "libadt/vectset.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/callnode.hpp" 30 #include "opto/cfgnode.hpp" 31 #include "opto/compile.hpp" 32 #include "opto/connode.hpp" 33 #include "opto/locknode.hpp" 34 #include "opto/loopnode.hpp" 35 #include "opto/macro.hpp" 36 #include "opto/memnode.hpp" 37 #include "opto/node.hpp" 38 #include "opto/phaseX.hpp" 39 #include "opto/rootnode.hpp" 40 #include "opto/runtime.hpp" 41 #include "opto/subnode.hpp" 42 #include "opto/type.hpp" 43 #include "runtime/sharedRuntime.hpp" 44 45 46 // 47 // Replace any references to "oldref" in inputs to "use" with "newref". 48 // Returns the number of replacements made. 49 // 50 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { 51 int nreplacements = 0; 52 uint req = use->req(); 53 for (uint j = 0; j < use->len(); j++) { 54 Node *uin = use->in(j); 55 if (uin == oldref) { 56 if (j < req) 57 use->set_req(j, newref); 58 else 59 use->set_prec(j, newref); 60 nreplacements++; 61 } else if (j >= req && uin == NULL) { 62 break; 63 } 64 } 65 return nreplacements; 66 } 67 68 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) { 69 // Copy debug information and adjust JVMState information 70 uint old_dbg_start = oldcall->tf()->domain()->cnt(); 71 uint new_dbg_start = newcall->tf()->domain()->cnt(); 72 int jvms_adj = new_dbg_start - old_dbg_start; 73 assert (new_dbg_start == newcall->req(), "argument count mismatch"); 74 75 Dict* sosn_map = new Dict(cmpkey,hashkey); 76 for (uint i = old_dbg_start; i < oldcall->req(); i++) { 77 Node* old_in = oldcall->in(i); 78 // Clone old SafePointScalarObjectNodes, adjusting their field contents. 79 if (old_in != NULL && old_in->is_SafePointScalarObject()) { 80 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject(); 81 uint old_unique = C->unique(); 82 Node* new_in = old_sosn->clone(jvms_adj, sosn_map); 83 if (old_unique != C->unique()) { 84 new_in->set_req(0, C->root()); // reset control edge 85 new_in = transform_later(new_in); // Register new node. 86 } 87 old_in = new_in; 88 } 89 newcall->add_req(old_in); 90 } 91 92 newcall->set_jvms(oldcall->jvms()); 93 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) { 94 jvms->set_map(newcall); 95 jvms->set_locoff(jvms->locoff()+jvms_adj); 96 jvms->set_stkoff(jvms->stkoff()+jvms_adj); 97 jvms->set_monoff(jvms->monoff()+jvms_adj); 98 jvms->set_scloff(jvms->scloff()+jvms_adj); 99 jvms->set_endoff(jvms->endoff()+jvms_adj); 100 } 101 } 102 103 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) { 104 Node* cmp; 105 if (mask != 0) { 106 Node* and_node = transform_later(new (C, 3) AndXNode(word, MakeConX(mask))); 107 cmp = transform_later(new (C, 3) CmpXNode(and_node, MakeConX(bits))); 108 } else { 109 cmp = word; 110 } 111 Node* bol = transform_later(new (C, 2) BoolNode(cmp, BoolTest::ne)); 112 IfNode* iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); 113 transform_later(iff); 114 115 // Fast path taken. 116 Node *fast_taken = transform_later( new (C, 1) IfFalseNode(iff) ); 117 118 // Fast path not-taken, i.e. slow path 119 Node *slow_taken = transform_later( new (C, 1) IfTrueNode(iff) ); 120 121 if (return_fast_path) { 122 region->init_req(edge, slow_taken); // Capture slow-control 123 return fast_taken; 124 } else { 125 region->init_req(edge, fast_taken); // Capture fast-control 126 return slow_taken; 127 } 128 } 129 130 //--------------------copy_predefined_input_for_runtime_call-------------------- 131 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { 132 // Set fixed predefined input arguments 133 call->init_req( TypeFunc::Control, ctrl ); 134 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); 135 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? 136 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); 137 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); 138 } 139 140 //------------------------------make_slow_call--------------------------------- 141 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, address slow_call, const char* leaf_name, Node* slow_path, Node* parm0, Node* parm1) { 142 143 // Slow-path call 144 int size = slow_call_type->domain()->cnt(); 145 CallNode *call = leaf_name 146 ? (CallNode*)new (C, size) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) 147 : (CallNode*)new (C, size) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM ); 148 149 // Slow path call has no side-effects, uses few values 150 copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); 151 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0); 152 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1); 153 copy_call_debug_info(oldcall, call); 154 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 155 _igvn.replace_node(oldcall, call); 156 transform_later(call); 157 158 return call; 159 } 160 161 void PhaseMacroExpand::extract_call_projections(CallNode *call) { 162 _fallthroughproj = NULL; 163 _fallthroughcatchproj = NULL; 164 _ioproj_fallthrough = NULL; 165 _ioproj_catchall = NULL; 166 _catchallcatchproj = NULL; 167 _memproj_fallthrough = NULL; 168 _memproj_catchall = NULL; 169 _resproj = NULL; 170 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) { 171 ProjNode *pn = call->fast_out(i)->as_Proj(); 172 switch (pn->_con) { 173 case TypeFunc::Control: 174 { 175 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj 176 _fallthroughproj = pn; 177 DUIterator_Fast jmax, j = pn->fast_outs(jmax); 178 const Node *cn = pn->fast_out(j); 179 if (cn->is_Catch()) { 180 ProjNode *cpn = NULL; 181 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { 182 cpn = cn->fast_out(k)->as_Proj(); 183 assert(cpn->is_CatchProj(), "must be a CatchProjNode"); 184 if (cpn->_con == CatchProjNode::fall_through_index) 185 _fallthroughcatchproj = cpn; 186 else { 187 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); 188 _catchallcatchproj = cpn; 189 } 190 } 191 } 192 break; 193 } 194 case TypeFunc::I_O: 195 if (pn->_is_io_use) 196 _ioproj_catchall = pn; 197 else 198 _ioproj_fallthrough = pn; 199 break; 200 case TypeFunc::Memory: 201 if (pn->_is_io_use) 202 _memproj_catchall = pn; 203 else 204 _memproj_fallthrough = pn; 205 break; 206 case TypeFunc::Parms: 207 _resproj = pn; 208 break; 209 default: 210 assert(false, "unexpected projection from allocation node."); 211 } 212 } 213 214 } 215 216 // Eliminate a card mark sequence. p2x is a ConvP2XNode 217 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) { 218 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required"); 219 if (!UseG1GC) { 220 // vanilla/CMS post barrier 221 Node *shift = p2x->unique_out(); 222 Node *addp = shift->unique_out(); 223 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) { 224 Node *mem = addp->last_out(j); 225 if (UseCondCardMark && mem->is_Load()) { 226 assert(mem->Opcode() == Op_LoadB, "unexpected code shape"); 227 // The load is checking if the card has been written so 228 // replace it with zero to fold the test. 229 _igvn.replace_node(mem, intcon(0)); 230 continue; 231 } 232 assert(mem->is_Store(), "store required"); 233 _igvn.replace_node(mem, mem->in(MemNode::Memory)); 234 } 235 } else { 236 // G1 pre/post barriers 237 assert(p2x->outcnt() == 2, "expects 2 users: Xor and URShift nodes"); 238 // It could be only one user, URShift node, in Object.clone() instrinsic 239 // but the new allocation is passed to arraycopy stub and it could not 240 // be scalar replaced. So we don't check the case. 241 242 // Remove G1 post barrier. 243 244 // Search for CastP2X->Xor->URShift->Cmp path which 245 // checks if the store done to a different from the value's region. 246 // And replace Cmp with #0 (false) to collapse G1 post barrier. 247 Node* xorx = NULL; 248 for (DUIterator_Fast imax, i = p2x->fast_outs(imax); i < imax; i++) { 249 Node* u = p2x->fast_out(i); 250 if (u->Opcode() == Op_XorX) { 251 xorx = u; 252 break; 253 } 254 } 255 assert(xorx != NULL, "missing G1 post barrier"); 256 Node* shift = xorx->unique_out(); 257 Node* cmpx = shift->unique_out(); 258 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() && 259 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne, 260 "missing region check in G1 post barrier"); 261 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 262 263 // Remove G1 pre barrier. 264 265 // Search "if (marking != 0)" check and set it to "false". 266 Node* this_region = p2x->in(0); 267 assert(this_region != NULL, ""); 268 // There is no G1 pre barrier if previous stored value is NULL 269 // (for example, after initialization). 270 if (this_region->is_Region() && this_region->req() == 3) { 271 int ind = 1; 272 if (!this_region->in(ind)->is_IfFalse()) { 273 ind = 2; 274 } 275 if (this_region->in(ind)->is_IfFalse()) { 276 Node* bol = this_region->in(ind)->in(0)->in(1); 277 assert(bol->is_Bool(), ""); 278 cmpx = bol->in(1); 279 if (bol->as_Bool()->_test._test == BoolTest::ne && 280 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) && 281 cmpx->in(1)->is_Load()) { 282 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address); 283 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + 284 PtrQueue::byte_offset_of_active()); 285 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() && 286 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && 287 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) { 288 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 289 } 290 } 291 } 292 } 293 // Now CastP2X can be removed since it is used only on dead path 294 // which currently still alive until igvn optimize it. 295 assert(p2x->unique_out()->Opcode() == Op_URShiftX, ""); 296 _igvn.replace_node(p2x, top()); 297 } 298 } 299 300 // Search for a memory operation for the specified memory slice. 301 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) { 302 Node *orig_mem = mem; 303 Node *alloc_mem = alloc->in(TypeFunc::Memory); 304 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr(); 305 while (true) { 306 if (mem == alloc_mem || mem == start_mem ) { 307 return mem; // hit one of our sentinels 308 } else if (mem->is_MergeMem()) { 309 mem = mem->as_MergeMem()->memory_at(alias_idx); 310 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { 311 Node *in = mem->in(0); 312 // we can safely skip over safepoints, calls, locks and membars because we 313 // already know that the object is safe to eliminate. 314 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { 315 return in; 316 } else if (in->is_Call()) { 317 CallNode *call = in->as_Call(); 318 if (!call->may_modify(tinst, phase)) { 319 mem = call->in(TypeFunc::Memory); 320 } 321 mem = in->in(TypeFunc::Memory); 322 } else if (in->is_MemBar()) { 323 mem = in->in(TypeFunc::Memory); 324 } else { 325 assert(false, "unexpected projection"); 326 } 327 } else if (mem->is_Store()) { 328 const TypePtr* atype = mem->as_Store()->adr_type(); 329 int adr_idx = Compile::current()->get_alias_index(atype); 330 if (adr_idx == alias_idx) { 331 assert(atype->isa_oopptr(), "address type must be oopptr"); 332 int adr_offset = atype->offset(); 333 uint adr_iid = atype->is_oopptr()->instance_id(); 334 // Array elements references have the same alias_idx 335 // but different offset and different instance_id. 336 if (adr_offset == offset && adr_iid == alloc->_idx) 337 return mem; 338 } else { 339 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw"); 340 } 341 mem = mem->in(MemNode::Memory); 342 } else if (mem->is_ClearArray()) { 343 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) { 344 // Can not bypass initialization of the instance 345 // we are looking. 346 debug_only(intptr_t offset;) 347 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity"); 348 InitializeNode* init = alloc->as_Allocate()->initialization(); 349 // We are looking for stored value, return Initialize node 350 // or memory edge from Allocate node. 351 if (init != NULL) 352 return init; 353 else 354 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers). 355 } 356 // Otherwise skip it (the call updated 'mem' value). 357 } else if (mem->Opcode() == Op_SCMemProj) { 358 assert(mem->in(0)->is_LoadStore(), "sanity"); 359 const TypePtr* atype = mem->in(0)->in(MemNode::Address)->bottom_type()->is_ptr(); 360 int adr_idx = Compile::current()->get_alias_index(atype); 361 if (adr_idx == alias_idx) { 362 assert(false, "Object is not scalar replaceable if a LoadStore node access its field"); 363 return NULL; 364 } 365 mem = mem->in(0)->in(MemNode::Memory); 366 } else { 367 return mem; 368 } 369 assert(mem != orig_mem, "dead memory loop"); 370 } 371 } 372 373 // 374 // Given a Memory Phi, compute a value Phi containing the values from stores 375 // on the input paths. 376 // Note: this function is recursive, its depth is limied by the "level" argument 377 // Returns the computed Phi, or NULL if it cannot compute it. 378 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) { 379 assert(mem->is_Phi(), "sanity"); 380 int alias_idx = C->get_alias_index(adr_t); 381 int offset = adr_t->offset(); 382 int instance_id = adr_t->instance_id(); 383 384 // Check if an appropriate value phi already exists. 385 Node* region = mem->in(0); 386 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { 387 Node* phi = region->fast_out(k); 388 if (phi->is_Phi() && phi != mem && 389 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) { 390 return phi; 391 } 392 } 393 // Check if an appropriate new value phi already exists. 394 Node* new_phi = value_phis->find(mem->_idx); 395 if (new_phi != NULL) 396 return new_phi; 397 398 if (level <= 0) { 399 return NULL; // Give up: phi tree too deep 400 } 401 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 402 Node *alloc_mem = alloc->in(TypeFunc::Memory); 403 404 uint length = mem->req(); 405 GrowableArray <Node *> values(length, length, NULL); 406 407 // create a new Phi for the value 408 PhiNode *phi = new (C, length) PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset); 409 transform_later(phi); 410 value_phis->push(phi, mem->_idx); 411 412 for (uint j = 1; j < length; j++) { 413 Node *in = mem->in(j); 414 if (in == NULL || in->is_top()) { 415 values.at_put(j, in); 416 } else { 417 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); 418 if (val == start_mem || val == alloc_mem) { 419 // hit a sentinel, return appropriate 0 value 420 values.at_put(j, _igvn.zerocon(ft)); 421 continue; 422 } 423 if (val->is_Initialize()) { 424 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 425 } 426 if (val == NULL) { 427 return NULL; // can't find a value on this path 428 } 429 if (val == mem) { 430 values.at_put(j, mem); 431 } else if (val->is_Store()) { 432 values.at_put(j, val->in(MemNode::ValueIn)); 433 } else if(val->is_Proj() && val->in(0) == alloc) { 434 values.at_put(j, _igvn.zerocon(ft)); 435 } else if (val->is_Phi()) { 436 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); 437 if (val == NULL) { 438 return NULL; 439 } 440 values.at_put(j, val); 441 } else if (val->Opcode() == Op_SCMemProj) { 442 assert(val->in(0)->is_LoadStore(), "sanity"); 443 assert(false, "Object is not scalar replaceable if a LoadStore node access its field"); 444 return NULL; 445 } else { 446 #ifdef ASSERT 447 val->dump(); 448 assert(false, "unknown node on this path"); 449 #endif 450 return NULL; // unknown node on this path 451 } 452 } 453 } 454 // Set Phi's inputs 455 for (uint j = 1; j < length; j++) { 456 if (values.at(j) == mem) { 457 phi->init_req(j, phi); 458 } else { 459 phi->init_req(j, values.at(j)); 460 } 461 } 462 return phi; 463 } 464 465 // Search the last value stored into the object's field. 466 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) { 467 assert(adr_t->is_known_instance_field(), "instance required"); 468 int instance_id = adr_t->instance_id(); 469 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 470 471 int alias_idx = C->get_alias_index(adr_t); 472 int offset = adr_t->offset(); 473 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 474 Node *alloc_ctrl = alloc->in(TypeFunc::Control); 475 Node *alloc_mem = alloc->in(TypeFunc::Memory); 476 Arena *a = Thread::current()->resource_area(); 477 VectorSet visited(a); 478 479 480 bool done = sfpt_mem == alloc_mem; 481 Node *mem = sfpt_mem; 482 while (!done) { 483 if (visited.test_set(mem->_idx)) { 484 return NULL; // found a loop, give up 485 } 486 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 487 if (mem == start_mem || mem == alloc_mem) { 488 done = true; // hit a sentinel, return appropriate 0 value 489 } else if (mem->is_Initialize()) { 490 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 491 if (mem == NULL) { 492 done = true; // Something go wrong. 493 } else if (mem->is_Store()) { 494 const TypePtr* atype = mem->as_Store()->adr_type(); 495 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 496 done = true; 497 } 498 } else if (mem->is_Store()) { 499 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 500 assert(atype != NULL, "address type must be oopptr"); 501 assert(C->get_alias_index(atype) == alias_idx && 502 atype->is_known_instance_field() && atype->offset() == offset && 503 atype->instance_id() == instance_id, "store is correct memory slice"); 504 done = true; 505 } else if (mem->is_Phi()) { 506 // try to find a phi's unique input 507 Node *unique_input = NULL; 508 Node *top = C->top(); 509 for (uint i = 1; i < mem->req(); i++) { 510 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 511 if (n == NULL || n == top || n == mem) { 512 continue; 513 } else if (unique_input == NULL) { 514 unique_input = n; 515 } else if (unique_input != n) { 516 unique_input = top; 517 break; 518 } 519 } 520 if (unique_input != NULL && unique_input != top) { 521 mem = unique_input; 522 } else { 523 done = true; 524 } 525 } else { 526 assert(false, "unexpected node"); 527 } 528 } 529 if (mem != NULL) { 530 if (mem == start_mem || mem == alloc_mem) { 531 // hit a sentinel, return appropriate 0 value 532 return _igvn.zerocon(ft); 533 } else if (mem->is_Store()) { 534 return mem->in(MemNode::ValueIn); 535 } else if (mem->is_Phi()) { 536 // attempt to produce a Phi reflecting the values on the input paths of the Phi 537 Node_Stack value_phis(a, 8); 538 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 539 if (phi != NULL) { 540 return phi; 541 } else { 542 // Kill all new Phis 543 while(value_phis.is_nonempty()) { 544 Node* n = value_phis.node(); 545 _igvn.replace_node(n, C->top()); 546 value_phis.pop(); 547 } 548 } 549 } 550 } 551 // Something go wrong. 552 return NULL; 553 } 554 555 // Check the possibility of scalar replacement. 556 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 557 // Scan the uses of the allocation to check for anything that would 558 // prevent us from eliminating it. 559 NOT_PRODUCT( const char* fail_eliminate = NULL; ) 560 DEBUG_ONLY( Node* disq_node = NULL; ) 561 bool can_eliminate = true; 562 563 Node* res = alloc->result_cast(); 564 const TypeOopPtr* res_type = NULL; 565 if (res == NULL) { 566 // All users were eliminated. 567 } else if (!res->is_CheckCastPP()) { 568 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 569 can_eliminate = false; 570 } else { 571 res_type = _igvn.type(res)->isa_oopptr(); 572 if (res_type == NULL) { 573 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 574 can_eliminate = false; 575 } else if (res_type->isa_aryptr()) { 576 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 577 if (length < 0) { 578 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 579 can_eliminate = false; 580 } 581 } 582 } 583 584 if (can_eliminate && res != NULL) { 585 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 586 j < jmax && can_eliminate; j++) { 587 Node* use = res->fast_out(j); 588 589 if (use->is_AddP()) { 590 const TypePtr* addp_type = _igvn.type(use)->is_ptr(); 591 int offset = addp_type->offset(); 592 593 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 594 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) 595 can_eliminate = false; 596 break; 597 } 598 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 599 k < kmax && can_eliminate; k++) { 600 Node* n = use->fast_out(k); 601 if (!n->is_Store() && n->Opcode() != Op_CastP2X) { 602 DEBUG_ONLY(disq_node = n;) 603 if (n->is_Load() || n->is_LoadStore()) { 604 NOT_PRODUCT(fail_eliminate = "Field load";) 605 } else { 606 NOT_PRODUCT(fail_eliminate = "Not store field referrence";) 607 } 608 can_eliminate = false; 609 } 610 } 611 } else if (use->is_SafePoint()) { 612 SafePointNode* sfpt = use->as_SafePoint(); 613 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 614 // Object is passed as argument. 615 DEBUG_ONLY(disq_node = use;) 616 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 617 can_eliminate = false; 618 } 619 Node* sfptMem = sfpt->memory(); 620 if (sfptMem == NULL || sfptMem->is_top()) { 621 DEBUG_ONLY(disq_node = use;) 622 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) 623 can_eliminate = false; 624 } else { 625 safepoints.append_if_missing(sfpt); 626 } 627 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 628 if (use->is_Phi()) { 629 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 630 NOT_PRODUCT(fail_eliminate = "Object is return value";) 631 } else { 632 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 633 } 634 DEBUG_ONLY(disq_node = use;) 635 } else { 636 if (use->Opcode() == Op_Return) { 637 NOT_PRODUCT(fail_eliminate = "Object is return value";) 638 }else { 639 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 640 } 641 DEBUG_ONLY(disq_node = use;) 642 } 643 can_eliminate = false; 644 } 645 } 646 } 647 648 #ifndef PRODUCT 649 if (PrintEliminateAllocations) { 650 if (can_eliminate) { 651 tty->print("Scalar "); 652 if (res == NULL) 653 alloc->dump(); 654 else 655 res->dump(); 656 } else { 657 tty->print("NotScalar (%s)", fail_eliminate); 658 if (res == NULL) 659 alloc->dump(); 660 else 661 res->dump(); 662 #ifdef ASSERT 663 if (disq_node != NULL) { 664 tty->print(" >>>> "); 665 disq_node->dump(); 666 } 667 #endif /*ASSERT*/ 668 } 669 } 670 #endif 671 return can_eliminate; 672 } 673 674 // Do scalar replacement. 675 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 676 GrowableArray <SafePointNode *> safepoints_done; 677 678 ciKlass* klass = NULL; 679 ciInstanceKlass* iklass = NULL; 680 int nfields = 0; 681 int array_base; 682 int element_size; 683 BasicType basic_elem_type; 684 ciType* elem_type; 685 686 Node* res = alloc->result_cast(); 687 const TypeOopPtr* res_type = NULL; 688 if (res != NULL) { // Could be NULL when there are no users 689 res_type = _igvn.type(res)->isa_oopptr(); 690 } 691 692 if (res != NULL) { 693 klass = res_type->klass(); 694 if (res_type->isa_instptr()) { 695 // find the fields of the class which will be needed for safepoint debug information 696 assert(klass->is_instance_klass(), "must be an instance klass."); 697 iklass = klass->as_instance_klass(); 698 nfields = iklass->nof_nonstatic_fields(); 699 } else { 700 // find the array's elements which will be needed for safepoint debug information 701 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 702 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); 703 elem_type = klass->as_array_klass()->element_type(); 704 basic_elem_type = elem_type->basic_type(); 705 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 706 element_size = type2aelembytes(basic_elem_type); 707 } 708 } 709 // 710 // Process the safepoint uses 711 // 712 while (safepoints.length() > 0) { 713 SafePointNode* sfpt = safepoints.pop(); 714 Node* mem = sfpt->memory(); 715 uint first_ind = sfpt->req(); 716 SafePointScalarObjectNode* sobj = new (C, 1) SafePointScalarObjectNode(res_type, 717 #ifdef ASSERT 718 alloc, 719 #endif 720 first_ind, nfields); 721 sobj->init_req(0, C->root()); 722 transform_later(sobj); 723 724 // Scan object's fields adding an input to the safepoint for each field. 725 for (int j = 0; j < nfields; j++) { 726 intptr_t offset; 727 ciField* field = NULL; 728 if (iklass != NULL) { 729 field = iklass->nonstatic_field_at(j); 730 offset = field->offset(); 731 elem_type = field->type(); 732 basic_elem_type = field->layout_type(); 733 } else { 734 offset = array_base + j * (intptr_t)element_size; 735 } 736 737 const Type *field_type; 738 // The next code is taken from Parse::do_get_xxx(). 739 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { 740 if (!elem_type->is_loaded()) { 741 field_type = TypeInstPtr::BOTTOM; 742 } else if (field != NULL && field->is_constant() && field->is_static()) { 743 // This can happen if the constant oop is non-perm. 744 ciObject* con = field->constant_value().as_object(); 745 // Do not "join" in the previous type; it doesn't add value, 746 // and may yield a vacuous result if the field is of interface type. 747 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 748 assert(field_type != NULL, "field singleton type must be consistent"); 749 } else { 750 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 751 } 752 if (UseCompressedOops) { 753 field_type = field_type->make_narrowoop(); 754 basic_elem_type = T_NARROWOOP; 755 } 756 } else { 757 field_type = Type::get_const_basic_type(basic_elem_type); 758 } 759 760 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 761 762 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc); 763 if (field_val == NULL) { 764 // We weren't able to find a value for this field, 765 // give up on eliminating this allocation. 766 767 // Remove any extra entries we added to the safepoint. 768 uint last = sfpt->req() - 1; 769 for (int k = 0; k < j; k++) { 770 sfpt->del_req(last--); 771 } 772 // rollback processed safepoints 773 while (safepoints_done.length() > 0) { 774 SafePointNode* sfpt_done = safepoints_done.pop(); 775 // remove any extra entries we added to the safepoint 776 last = sfpt_done->req() - 1; 777 for (int k = 0; k < nfields; k++) { 778 sfpt_done->del_req(last--); 779 } 780 JVMState *jvms = sfpt_done->jvms(); 781 jvms->set_endoff(sfpt_done->req()); 782 // Now make a pass over the debug information replacing any references 783 // to SafePointScalarObjectNode with the allocated object. 784 int start = jvms->debug_start(); 785 int end = jvms->debug_end(); 786 for (int i = start; i < end; i++) { 787 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 788 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 789 if (scobj->first_index() == sfpt_done->req() && 790 scobj->n_fields() == (uint)nfields) { 791 assert(scobj->alloc() == alloc, "sanity"); 792 sfpt_done->set_req(i, res); 793 } 794 } 795 } 796 } 797 #ifndef PRODUCT 798 if (PrintEliminateAllocations) { 799 if (field != NULL) { 800 tty->print("=== At SafePoint node %d can't find value of Field: ", 801 sfpt->_idx); 802 field->print(); 803 int field_idx = C->get_alias_index(field_addr_type); 804 tty->print(" (alias_idx=%d)", field_idx); 805 } else { // Array's element 806 tty->print("=== At SafePoint node %d can't find value of array element [%d]", 807 sfpt->_idx, j); 808 } 809 tty->print(", which prevents elimination of: "); 810 if (res == NULL) 811 alloc->dump(); 812 else 813 res->dump(); 814 } 815 #endif 816 return false; 817 } 818 if (UseCompressedOops && field_type->isa_narrowoop()) { 819 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 820 // to be able scalar replace the allocation. 821 if (field_val->is_EncodeP()) { 822 field_val = field_val->in(1); 823 } else { 824 field_val = transform_later(new (C, 2) DecodeNNode(field_val, field_val->bottom_type()->make_ptr())); 825 } 826 } 827 sfpt->add_req(field_val); 828 } 829 JVMState *jvms = sfpt->jvms(); 830 jvms->set_endoff(sfpt->req()); 831 // Now make a pass over the debug information replacing any references 832 // to the allocated object with "sobj" 833 int start = jvms->debug_start(); 834 int end = jvms->debug_end(); 835 for (int i = start; i < end; i++) { 836 if (sfpt->in(i) == res) { 837 sfpt->set_req(i, sobj); 838 } 839 } 840 safepoints_done.append_if_missing(sfpt); // keep it for rollback 841 } 842 return true; 843 } 844 845 // Process users of eliminated allocation. 846 void PhaseMacroExpand::process_users_of_allocation(AllocateNode *alloc) { 847 Node* res = alloc->result_cast(); 848 if (res != NULL) { 849 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 850 Node *use = res->last_out(j); 851 uint oc1 = res->outcnt(); 852 853 if (use->is_AddP()) { 854 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 855 Node *n = use->last_out(k); 856 uint oc2 = use->outcnt(); 857 if (n->is_Store()) { 858 #ifdef ASSERT 859 // Verify that there is no dependent MemBarVolatile nodes, 860 // they should be removed during IGVN, see MemBarNode::Ideal(). 861 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); 862 p < pmax; p++) { 863 Node* mb = n->fast_out(p); 864 assert(mb->is_Initialize() || !mb->is_MemBar() || 865 mb->req() <= MemBarNode::Precedent || 866 mb->in(MemBarNode::Precedent) != n, 867 "MemBarVolatile should be eliminated for non-escaping object"); 868 } 869 #endif 870 _igvn.replace_node(n, n->in(MemNode::Memory)); 871 } else { 872 eliminate_card_mark(n); 873 } 874 k -= (oc2 - use->outcnt()); 875 } 876 } else { 877 eliminate_card_mark(use); 878 } 879 j -= (oc1 - res->outcnt()); 880 } 881 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 882 _igvn.remove_dead_node(res); 883 } 884 885 // 886 // Process other users of allocation's projections 887 // 888 if (_resproj != NULL && _resproj->outcnt() != 0) { 889 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { 890 Node *use = _resproj->last_out(j); 891 uint oc1 = _resproj->outcnt(); 892 if (use->is_Initialize()) { 893 // Eliminate Initialize node. 894 InitializeNode *init = use->as_Initialize(); 895 assert(init->outcnt() <= 2, "only a control and memory projection expected"); 896 Node *ctrl_proj = init->proj_out(TypeFunc::Control); 897 if (ctrl_proj != NULL) { 898 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection"); 899 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj); 900 } 901 Node *mem_proj = init->proj_out(TypeFunc::Memory); 902 if (mem_proj != NULL) { 903 Node *mem = init->in(TypeFunc::Memory); 904 #ifdef ASSERT 905 if (mem->is_MergeMem()) { 906 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection"); 907 } else { 908 assert(mem == _memproj_fallthrough, "allocation memory projection"); 909 } 910 #endif 911 _igvn.replace_node(mem_proj, mem); 912 } 913 } else if (use->is_AddP()) { 914 // raw memory addresses used only by the initialization 915 _igvn.replace_node(use, C->top()); 916 } else { 917 assert(false, "only Initialize or AddP expected"); 918 } 919 j -= (oc1 - _resproj->outcnt()); 920 } 921 } 922 if (_fallthroughcatchproj != NULL) { 923 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); 924 } 925 if (_memproj_fallthrough != NULL) { 926 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); 927 } 928 if (_memproj_catchall != NULL) { 929 _igvn.replace_node(_memproj_catchall, C->top()); 930 } 931 if (_ioproj_fallthrough != NULL) { 932 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); 933 } 934 if (_ioproj_catchall != NULL) { 935 _igvn.replace_node(_ioproj_catchall, C->top()); 936 } 937 if (_catchallcatchproj != NULL) { 938 _igvn.replace_node(_catchallcatchproj, C->top()); 939 } 940 } 941 942 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 943 944 if (!EliminateAllocations || !alloc->_is_scalar_replaceable) { 945 return false; 946 } 947 948 extract_call_projections(alloc); 949 950 GrowableArray <SafePointNode *> safepoints; 951 if (!can_eliminate_allocation(alloc, safepoints)) { 952 return false; 953 } 954 955 if (!scalar_replacement(alloc, safepoints)) { 956 return false; 957 } 958 959 CompileLog* log = C->log(); 960 if (log != NULL) { 961 Node* klass = alloc->in(AllocateNode::KlassNode); 962 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); 963 log->head("eliminate_allocation type='%d'", 964 log->identify(tklass->klass())); 965 JVMState* p = alloc->jvms(); 966 while (p != NULL) { 967 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 968 p = p->caller(); 969 } 970 log->tail("eliminate_allocation"); 971 } 972 973 process_users_of_allocation(alloc); 974 975 #ifndef PRODUCT 976 if (PrintEliminateAllocations) { 977 if (alloc->is_AllocateArray()) 978 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 979 else 980 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 981 } 982 #endif 983 984 return true; 985 } 986 987 988 //---------------------------set_eden_pointers------------------------- 989 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { 990 if (UseTLAB) { // Private allocation: load from TLS 991 Node* thread = transform_later(new (C, 1) ThreadLocalNode()); 992 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); 993 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); 994 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); 995 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); 996 } else { // Shared allocation: load from globals 997 CollectedHeap* ch = Universe::heap(); 998 address top_adr = (address)ch->top_addr(); 999 address end_adr = (address)ch->end_addr(); 1000 eden_top_adr = makecon(TypeRawPtr::make(top_adr)); 1001 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); 1002 } 1003 } 1004 1005 1006 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 1007 Node* adr = basic_plus_adr(base, offset); 1008 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 1009 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt); 1010 transform_later(value); 1011 return value; 1012 } 1013 1014 1015 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 1016 Node* adr = basic_plus_adr(base, offset); 1017 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt); 1018 transform_later(mem); 1019 return mem; 1020 } 1021 1022 //============================================================================= 1023 // 1024 // A L L O C A T I O N 1025 // 1026 // Allocation attempts to be fast in the case of frequent small objects. 1027 // It breaks down like this: 1028 // 1029 // 1) Size in doublewords is computed. This is a constant for objects and 1030 // variable for most arrays. Doubleword units are used to avoid size 1031 // overflow of huge doubleword arrays. We need doublewords in the end for 1032 // rounding. 1033 // 1034 // 2) Size is checked for being 'too large'. Too-large allocations will go 1035 // the slow path into the VM. The slow path can throw any required 1036 // exceptions, and does all the special checks for very large arrays. The 1037 // size test can constant-fold away for objects. For objects with 1038 // finalizers it constant-folds the otherway: you always go slow with 1039 // finalizers. 1040 // 1041 // 3) If NOT using TLABs, this is the contended loop-back point. 1042 // Load-Locked the heap top. If using TLABs normal-load the heap top. 1043 // 1044 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 1045 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 1046 // "size*8" we always enter the VM, where "largish" is a constant picked small 1047 // enough that there's always space between the eden max and 4Gig (old space is 1048 // there so it's quite large) and large enough that the cost of entering the VM 1049 // is dwarfed by the cost to initialize the space. 1050 // 1051 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 1052 // down. If contended, repeat at step 3. If using TLABs normal-store 1053 // adjusted heap top back down; there is no contention. 1054 // 1055 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 1056 // fields. 1057 // 1058 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 1059 // oop flavor. 1060 // 1061 //============================================================================= 1062 // FastAllocateSizeLimit value is in DOUBLEWORDS. 1063 // Allocations bigger than this always go the slow route. 1064 // This value must be small enough that allocation attempts that need to 1065 // trigger exceptions go the slow route. Also, it must be small enough so 1066 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 1067 //=============================================================================j// 1068 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 1069 // The allocator will coalesce int->oop copies away. See comment in 1070 // coalesce.cpp about how this works. It depends critically on the exact 1071 // code shape produced here, so if you are changing this code shape 1072 // make sure the GC info for the heap-top is correct in and around the 1073 // slow-path call. 1074 // 1075 1076 void PhaseMacroExpand::expand_allocate_common( 1077 AllocateNode* alloc, // allocation node to be expanded 1078 Node* length, // array length for an array allocation 1079 const TypeFunc* slow_call_type, // Type of slow call 1080 address slow_call_address // Address of slow call 1081 ) 1082 { 1083 1084 Node* ctrl = alloc->in(TypeFunc::Control); 1085 Node* mem = alloc->in(TypeFunc::Memory); 1086 Node* i_o = alloc->in(TypeFunc::I_O); 1087 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 1088 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1089 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 1090 1091 Node* storestore = alloc->storestore(); 1092 if (storestore != NULL) { 1093 // Break this link that is no longer useful and confuses register allocation 1094 storestore->set_req(MemBarNode::Precedent, top()); 1095 } 1096 1097 assert(ctrl != NULL, "must have control"); 1098 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 1099 // they will not be used if "always_slow" is set 1100 enum { slow_result_path = 1, fast_result_path = 2 }; 1101 Node *result_region; 1102 Node *result_phi_rawmem; 1103 Node *result_phi_rawoop; 1104 Node *result_phi_i_o; 1105 1106 // The initial slow comparison is a size check, the comparison 1107 // we want to do is a BoolTest::gt 1108 bool always_slow = false; 1109 int tv = _igvn.find_int_con(initial_slow_test, -1); 1110 if (tv >= 0) { 1111 always_slow = (tv == 1); 1112 initial_slow_test = NULL; 1113 } else { 1114 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 1115 } 1116 1117 if (C->env()->dtrace_alloc_probes() || 1118 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc() || 1119 (UseConcMarkSweepGC && CMSIncrementalMode))) { 1120 // Force slow-path allocation 1121 always_slow = true; 1122 initial_slow_test = NULL; 1123 } 1124 1125 1126 enum { too_big_or_final_path = 1, need_gc_path = 2 }; 1127 Node *slow_region = NULL; 1128 Node *toobig_false = ctrl; 1129 1130 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); 1131 // generate the initial test if necessary 1132 if (initial_slow_test != NULL ) { 1133 slow_region = new (C, 3) RegionNode(3); 1134 1135 // Now make the initial failure test. Usually a too-big test but 1136 // might be a TRUE for finalizers or a fancy class check for 1137 // newInstance0. 1138 IfNode *toobig_iff = new (C, 2) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1139 transform_later(toobig_iff); 1140 // Plug the failing-too-big test into the slow-path region 1141 Node *toobig_true = new (C, 1) IfTrueNode( toobig_iff ); 1142 transform_later(toobig_true); 1143 slow_region ->init_req( too_big_or_final_path, toobig_true ); 1144 toobig_false = new (C, 1) IfFalseNode( toobig_iff ); 1145 transform_later(toobig_false); 1146 } else { // No initial test, just fall into next case 1147 toobig_false = ctrl; 1148 debug_only(slow_region = NodeSentinel); 1149 } 1150 1151 Node *slow_mem = mem; // save the current memory state for slow path 1152 // generate the fast allocation code unless we know that the initial test will always go slow 1153 if (!always_slow) { 1154 // Fast path modifies only raw memory. 1155 if (mem->is_MergeMem()) { 1156 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1157 } 1158 1159 Node* eden_top_adr; 1160 Node* eden_end_adr; 1161 1162 set_eden_pointers(eden_top_adr, eden_end_adr); 1163 1164 // Load Eden::end. Loop invariant and hoisted. 1165 // 1166 // Note: We set the control input on "eden_end" and "old_eden_top" when using 1167 // a TLAB to work around a bug where these values were being moved across 1168 // a safepoint. These are not oops, so they cannot be include in the oop 1169 // map, but they can be changed by a GC. The proper way to fix this would 1170 // be to set the raw memory state when generating a SafepointNode. However 1171 // this will require extensive changes to the loop optimization in order to 1172 // prevent a degradation of the optimization. 1173 // See comment in memnode.hpp, around line 227 in class LoadPNode. 1174 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 1175 1176 // allocate the Region and Phi nodes for the result 1177 result_region = new (C, 3) RegionNode(3); 1178 result_phi_rawmem = new (C, 3) PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1179 result_phi_rawoop = new (C, 3) PhiNode(result_region, TypeRawPtr::BOTTOM); 1180 result_phi_i_o = new (C, 3) PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch 1181 1182 // We need a Region for the loop-back contended case. 1183 enum { fall_in_path = 1, contended_loopback_path = 2 }; 1184 Node *contended_region; 1185 Node *contended_phi_rawmem; 1186 if (UseTLAB) { 1187 contended_region = toobig_false; 1188 contended_phi_rawmem = mem; 1189 } else { 1190 contended_region = new (C, 3) RegionNode(3); 1191 contended_phi_rawmem = new (C, 3) PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1192 // Now handle the passing-too-big test. We fall into the contended 1193 // loop-back merge point. 1194 contended_region ->init_req(fall_in_path, toobig_false); 1195 contended_phi_rawmem->init_req(fall_in_path, mem); 1196 transform_later(contended_region); 1197 transform_later(contended_phi_rawmem); 1198 } 1199 1200 // Load(-locked) the heap top. 1201 // See note above concerning the control input when using a TLAB 1202 Node *old_eden_top = UseTLAB 1203 ? new (C, 3) LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM) 1204 : new (C, 3) LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr); 1205 1206 transform_later(old_eden_top); 1207 // Add to heap top to get a new heap top 1208 Node *new_eden_top = new (C, 4) AddPNode(top(), old_eden_top, size_in_bytes); 1209 transform_later(new_eden_top); 1210 // Check for needing a GC; compare against heap end 1211 Node *needgc_cmp = new (C, 3) CmpPNode(new_eden_top, eden_end); 1212 transform_later(needgc_cmp); 1213 Node *needgc_bol = new (C, 2) BoolNode(needgc_cmp, BoolTest::ge); 1214 transform_later(needgc_bol); 1215 IfNode *needgc_iff = new (C, 2) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); 1216 transform_later(needgc_iff); 1217 1218 // Plug the failing-heap-space-need-gc test into the slow-path region 1219 Node *needgc_true = new (C, 1) IfTrueNode(needgc_iff); 1220 transform_later(needgc_true); 1221 if (initial_slow_test) { 1222 slow_region->init_req(need_gc_path, needgc_true); 1223 // This completes all paths into the slow merge point 1224 transform_later(slow_region); 1225 } else { // No initial slow path needed! 1226 // Just fall from the need-GC path straight into the VM call. 1227 slow_region = needgc_true; 1228 } 1229 // No need for a GC. Setup for the Store-Conditional 1230 Node *needgc_false = new (C, 1) IfFalseNode(needgc_iff); 1231 transform_later(needgc_false); 1232 1233 // Grab regular I/O before optional prefetch may change it. 1234 // Slow-path does no I/O so just set it to the original I/O. 1235 result_phi_i_o->init_req(slow_result_path, i_o); 1236 1237 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem, 1238 old_eden_top, new_eden_top, length); 1239 1240 // Name successful fast-path variables 1241 Node* fast_oop = old_eden_top; 1242 Node* fast_oop_ctrl; 1243 Node* fast_oop_rawmem; 1244 1245 // Store (-conditional) the modified eden top back down. 1246 // StorePConditional produces flags for a test PLUS a modified raw 1247 // memory state. 1248 if (UseTLAB) { 1249 Node* store_eden_top = 1250 new (C, 4) StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1251 TypeRawPtr::BOTTOM, new_eden_top); 1252 transform_later(store_eden_top); 1253 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path 1254 fast_oop_rawmem = store_eden_top; 1255 } else { 1256 Node* store_eden_top = 1257 new (C, 5) StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1258 new_eden_top, fast_oop/*old_eden_top*/); 1259 transform_later(store_eden_top); 1260 Node *contention_check = new (C, 2) BoolNode(store_eden_top, BoolTest::ne); 1261 transform_later(contention_check); 1262 store_eden_top = new (C, 1) SCMemProjNode(store_eden_top); 1263 transform_later(store_eden_top); 1264 1265 // If not using TLABs, check to see if there was contention. 1266 IfNode *contention_iff = new (C, 2) IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN); 1267 transform_later(contention_iff); 1268 Node *contention_true = new (C, 1) IfTrueNode(contention_iff); 1269 transform_later(contention_true); 1270 // If contention, loopback and try again. 1271 contended_region->init_req(contended_loopback_path, contention_true); 1272 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top); 1273 1274 // Fast-path succeeded with no contention! 1275 Node *contention_false = new (C, 1) IfFalseNode(contention_iff); 1276 transform_later(contention_false); 1277 fast_oop_ctrl = contention_false; 1278 1279 // Bump total allocated bytes for this thread 1280 Node* thread = new (C, 1) ThreadLocalNode(); 1281 transform_later(thread); 1282 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread, 1283 in_bytes(JavaThread::allocated_bytes_offset())); 1284 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1285 0, TypeLong::LONG, T_LONG); 1286 #ifdef _LP64 1287 Node* alloc_size = size_in_bytes; 1288 #else 1289 Node* alloc_size = new (C, 2) ConvI2LNode(size_in_bytes); 1290 transform_later(alloc_size); 1291 #endif 1292 Node* new_alloc_bytes = new (C, 3) AddLNode(alloc_bytes, alloc_size); 1293 transform_later(new_alloc_bytes); 1294 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1295 0, new_alloc_bytes, T_LONG); 1296 } 1297 1298 InitializeNode* init = alloc->initialization(); 1299 fast_oop_rawmem = initialize_object(alloc, 1300 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1301 klass_node, length, size_in_bytes); 1302 1303 // If initialization is performed by an array copy, any required 1304 // MemBarStoreStore was already added. If the object does not 1305 // escape no need for a MemBarStoreStore. Otherwise we need a 1306 // MemBarStoreStore so that stores that initialize this object 1307 // can't be reordered with a subsequent store that makes this 1308 // object accessible by other threads. 1309 if (init == NULL || (!init->is_complete_with_arraycopy() && !init->does_not_escape())) { 1310 if (init == NULL || init->req() < InitializeNode::RawStores) { 1311 // No InitializeNode or no stores captured by zeroing 1312 // elimination. Simply add the MemBarStoreStore after object 1313 // initialization. 1314 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot, fast_oop_rawmem); 1315 transform_later(mb); 1316 1317 mb->init_req(TypeFunc::Memory, fast_oop_rawmem); 1318 mb->init_req(TypeFunc::Control, fast_oop_ctrl); 1319 fast_oop_ctrl = new (C, 1) ProjNode(mb,TypeFunc::Control); 1320 transform_later(fast_oop_ctrl); 1321 fast_oop_rawmem = new (C, 1) ProjNode(mb,TypeFunc::Memory); 1322 transform_later(fast_oop_rawmem); 1323 } else { 1324 // Add the MemBarStoreStore after the InitializeNode so that 1325 // all stores performing the initialization that were moved 1326 // before the InitializeNode happen before the storestore 1327 // barrier. 1328 1329 Node* init_ctrl = init->proj_out(TypeFunc::Control); 1330 Node* init_mem = init->proj_out(TypeFunc::Memory); 1331 1332 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1333 transform_later(mb); 1334 1335 Node* ctrl = new (C, 1) ProjNode(init,TypeFunc::Control); 1336 transform_later(ctrl); 1337 Node* mem = new (C, 1) ProjNode(init,TypeFunc::Memory); 1338 transform_later(mem); 1339 1340 // The MemBarStoreStore depends on control and memory coming 1341 // from the InitializeNode 1342 mb->init_req(TypeFunc::Memory, mem); 1343 mb->init_req(TypeFunc::Control, ctrl); 1344 1345 ctrl = new (C, 1) ProjNode(mb,TypeFunc::Control); 1346 transform_later(ctrl); 1347 mem = new (C, 1) ProjNode(mb,TypeFunc::Memory); 1348 transform_later(mem); 1349 1350 // All nodes that depended on the InitializeNode for control 1351 // and memory must now depend on the MemBarNode that itself 1352 // depends on the InitializeNode 1353 _igvn.replace_node(init_ctrl, ctrl); 1354 _igvn.replace_node(init_mem, mem); 1355 } 1356 } 1357 1358 if (C->env()->dtrace_extended_probes()) { 1359 // Slow-path call 1360 int size = TypeFunc::Parms + 2; 1361 CallLeafNode *call = new (C, size) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1362 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), 1363 "dtrace_object_alloc", 1364 TypeRawPtr::BOTTOM); 1365 1366 // Get base of thread-local storage area 1367 Node* thread = new (C, 1) ThreadLocalNode(); 1368 transform_later(thread); 1369 1370 call->init_req(TypeFunc::Parms+0, thread); 1371 call->init_req(TypeFunc::Parms+1, fast_oop); 1372 call->init_req(TypeFunc::Control, fast_oop_ctrl); 1373 call->init_req(TypeFunc::I_O , top()); // does no i/o 1374 call->init_req(TypeFunc::Memory , fast_oop_rawmem); 1375 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1376 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1377 transform_later(call); 1378 fast_oop_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control); 1379 transform_later(fast_oop_ctrl); 1380 fast_oop_rawmem = new (C, 1) ProjNode(call,TypeFunc::Memory); 1381 transform_later(fast_oop_rawmem); 1382 } 1383 1384 // Plug in the successful fast-path into the result merge point 1385 result_region ->init_req(fast_result_path, fast_oop_ctrl); 1386 result_phi_rawoop->init_req(fast_result_path, fast_oop); 1387 result_phi_i_o ->init_req(fast_result_path, i_o); 1388 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); 1389 } else { 1390 slow_region = ctrl; 1391 result_phi_i_o = i_o; // Rename it to use in the following code. 1392 } 1393 1394 // Generate slow-path call 1395 CallNode *call = new (C, slow_call_type->domain()->cnt()) 1396 CallStaticJavaNode(slow_call_type, slow_call_address, 1397 OptoRuntime::stub_name(slow_call_address), 1398 alloc->jvms()->bci(), 1399 TypePtr::BOTTOM); 1400 call->init_req( TypeFunc::Control, slow_region ); 1401 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1402 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs 1403 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1404 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1405 1406 call->init_req(TypeFunc::Parms+0, klass_node); 1407 if (length != NULL) { 1408 call->init_req(TypeFunc::Parms+1, length); 1409 } 1410 1411 // Copy debug information and adjust JVMState information, then replace 1412 // allocate node with the call 1413 copy_call_debug_info((CallNode *) alloc, call); 1414 if (!always_slow) { 1415 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1416 } else { 1417 // Hook i_o projection to avoid its elimination during allocation 1418 // replacement (when only a slow call is generated). 1419 call->set_req(TypeFunc::I_O, result_phi_i_o); 1420 } 1421 _igvn.replace_node(alloc, call); 1422 transform_later(call); 1423 1424 // Identify the output projections from the allocate node and 1425 // adjust any references to them. 1426 // The control and io projections look like: 1427 // 1428 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1429 // Allocate Catch 1430 // ^---Proj(io) <-------+ ^---CatchProj(io) 1431 // 1432 // We are interested in the CatchProj nodes. 1433 // 1434 extract_call_projections(call); 1435 1436 // An allocate node has separate memory projections for the uses on 1437 // the control and i_o paths. Replace the control memory projection with 1438 // result_phi_rawmem (unless we are only generating a slow call when 1439 // both memory projections are combined) 1440 if (!always_slow && _memproj_fallthrough != NULL) { 1441 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { 1442 Node *use = _memproj_fallthrough->fast_out(i); 1443 _igvn.hash_delete(use); 1444 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); 1445 _igvn._worklist.push(use); 1446 // back up iterator 1447 --i; 1448 } 1449 } 1450 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete 1451 // _memproj_catchall so we end up with a call that has only 1 memory projection. 1452 if (_memproj_catchall != NULL ) { 1453 if (_memproj_fallthrough == NULL) { 1454 _memproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::Memory); 1455 transform_later(_memproj_fallthrough); 1456 } 1457 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { 1458 Node *use = _memproj_catchall->fast_out(i); 1459 _igvn.hash_delete(use); 1460 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); 1461 _igvn._worklist.push(use); 1462 // back up iterator 1463 --i; 1464 } 1465 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted"); 1466 _igvn.remove_dead_node(_memproj_catchall); 1467 } 1468 1469 // An allocate node has separate i_o projections for the uses on the control 1470 // and i_o paths. Always replace the control i_o projection with result i_o 1471 // otherwise incoming i_o become dead when only a slow call is generated 1472 // (it is different from memory projections where both projections are 1473 // combined in such case). 1474 if (_ioproj_fallthrough != NULL) { 1475 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { 1476 Node *use = _ioproj_fallthrough->fast_out(i); 1477 _igvn.hash_delete(use); 1478 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); 1479 _igvn._worklist.push(use); 1480 // back up iterator 1481 --i; 1482 } 1483 } 1484 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete 1485 // _ioproj_catchall so we end up with a call that has only 1 i_o projection. 1486 if (_ioproj_catchall != NULL ) { 1487 if (_ioproj_fallthrough == NULL) { 1488 _ioproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::I_O); 1489 transform_later(_ioproj_fallthrough); 1490 } 1491 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { 1492 Node *use = _ioproj_catchall->fast_out(i); 1493 _igvn.hash_delete(use); 1494 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); 1495 _igvn._worklist.push(use); 1496 // back up iterator 1497 --i; 1498 } 1499 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted"); 1500 _igvn.remove_dead_node(_ioproj_catchall); 1501 } 1502 1503 // if we generated only a slow call, we are done 1504 if (always_slow) { 1505 // Now we can unhook i_o. 1506 if (result_phi_i_o->outcnt() > 1) { 1507 call->set_req(TypeFunc::I_O, top()); 1508 } else { 1509 assert(result_phi_i_o->unique_ctrl_out() == call, ""); 1510 // Case of new array with negative size known during compilation. 1511 // AllocateArrayNode::Ideal() optimization disconnect unreachable 1512 // following code since call to runtime will throw exception. 1513 // As result there will be no users of i_o after the call. 1514 // Leave i_o attached to this call to avoid problems in preceding graph. 1515 } 1516 return; 1517 } 1518 1519 1520 if (_fallthroughcatchproj != NULL) { 1521 ctrl = _fallthroughcatchproj->clone(); 1522 transform_later(ctrl); 1523 _igvn.replace_node(_fallthroughcatchproj, result_region); 1524 } else { 1525 ctrl = top(); 1526 } 1527 Node *slow_result; 1528 if (_resproj == NULL) { 1529 // no uses of the allocation result 1530 slow_result = top(); 1531 } else { 1532 slow_result = _resproj->clone(); 1533 transform_later(slow_result); 1534 _igvn.replace_node(_resproj, result_phi_rawoop); 1535 } 1536 1537 // Plug slow-path into result merge point 1538 result_region ->init_req( slow_result_path, ctrl ); 1539 result_phi_rawoop->init_req( slow_result_path, slow_result); 1540 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); 1541 transform_later(result_region); 1542 transform_later(result_phi_rawoop); 1543 transform_later(result_phi_rawmem); 1544 transform_later(result_phi_i_o); 1545 // This completes all paths into the result merge point 1546 } 1547 1548 1549 // Helper for PhaseMacroExpand::expand_allocate_common. 1550 // Initializes the newly-allocated storage. 1551 Node* 1552 PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1553 Node* control, Node* rawmem, Node* object, 1554 Node* klass_node, Node* length, 1555 Node* size_in_bytes) { 1556 InitializeNode* init = alloc->initialization(); 1557 // Store the klass & mark bits 1558 Node* mark_node = NULL; 1559 // For now only enable fast locking for non-array types 1560 if (UseBiasedLocking && (length == NULL)) { 1561 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 1562 } else { 1563 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); 1564 } 1565 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 1566 1567 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_OBJECT); 1568 int header_size = alloc->minimum_header_size(); // conservatively small 1569 1570 // Array length 1571 if (length != NULL) { // Arrays need length field 1572 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1573 // conservatively small header size: 1574 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1575 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1576 if (k->is_array_klass()) // we know the exact header size in most cases: 1577 header_size = Klass::layout_helper_header_size(k->layout_helper()); 1578 } 1579 1580 // Clear the object body, if necessary. 1581 if (init == NULL) { 1582 // The init has somehow disappeared; be cautious and clear everything. 1583 // 1584 // This can happen if a node is allocated but an uncommon trap occurs 1585 // immediately. In this case, the Initialize gets associated with the 1586 // trap, and may be placed in a different (outer) loop, if the Allocate 1587 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1588 // there can be two Allocates to one Initialize. The answer in all these 1589 // edge cases is safety first. It is always safe to clear immediately 1590 // within an Allocate, and then (maybe or maybe not) clear some more later. 1591 if (!ZeroTLAB) 1592 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1593 header_size, size_in_bytes, 1594 &_igvn); 1595 } else { 1596 if (!init->is_complete()) { 1597 // Try to win by zeroing only what the init does not store. 1598 // We can also try to do some peephole optimizations, 1599 // such as combining some adjacent subword stores. 1600 rawmem = init->complete_stores(control, rawmem, object, 1601 header_size, size_in_bytes, &_igvn); 1602 } 1603 // We have no more use for this link, since the AllocateNode goes away: 1604 init->set_req(InitializeNode::RawAddress, top()); 1605 // (If we keep the link, it just confuses the register allocator, 1606 // who thinks he sees a real use of the address by the membar.) 1607 } 1608 1609 return rawmem; 1610 } 1611 1612 // Generate prefetch instructions for next allocations. 1613 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1614 Node*& contended_phi_rawmem, 1615 Node* old_eden_top, Node* new_eden_top, 1616 Node* length) { 1617 enum { fall_in_path = 1, pf_path = 2 }; 1618 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1619 // Generate prefetch allocation with watermark check. 1620 // As an allocation hits the watermark, we will prefetch starting 1621 // at a "distance" away from watermark. 1622 1623 Node *pf_region = new (C, 3) RegionNode(3); 1624 Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY, 1625 TypeRawPtr::BOTTOM ); 1626 // I/O is used for Prefetch 1627 Node *pf_phi_abio = new (C, 3) PhiNode( pf_region, Type::ABIO ); 1628 1629 Node *thread = new (C, 1) ThreadLocalNode(); 1630 transform_later(thread); 1631 1632 Node *eden_pf_adr = new (C, 4) AddPNode( top()/*not oop*/, thread, 1633 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1634 transform_later(eden_pf_adr); 1635 1636 Node *old_pf_wm = new (C, 3) LoadPNode( needgc_false, 1637 contended_phi_rawmem, eden_pf_adr, 1638 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM ); 1639 transform_later(old_pf_wm); 1640 1641 // check against new_eden_top 1642 Node *need_pf_cmp = new (C, 3) CmpPNode( new_eden_top, old_pf_wm ); 1643 transform_later(need_pf_cmp); 1644 Node *need_pf_bol = new (C, 2) BoolNode( need_pf_cmp, BoolTest::ge ); 1645 transform_later(need_pf_bol); 1646 IfNode *need_pf_iff = new (C, 2) IfNode( needgc_false, need_pf_bol, 1647 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1648 transform_later(need_pf_iff); 1649 1650 // true node, add prefetchdistance 1651 Node *need_pf_true = new (C, 1) IfTrueNode( need_pf_iff ); 1652 transform_later(need_pf_true); 1653 1654 Node *need_pf_false = new (C, 1) IfFalseNode( need_pf_iff ); 1655 transform_later(need_pf_false); 1656 1657 Node *new_pf_wmt = new (C, 4) AddPNode( top(), old_pf_wm, 1658 _igvn.MakeConX(AllocatePrefetchDistance) ); 1659 transform_later(new_pf_wmt ); 1660 new_pf_wmt->set_req(0, need_pf_true); 1661 1662 Node *store_new_wmt = new (C, 4) StorePNode( need_pf_true, 1663 contended_phi_rawmem, eden_pf_adr, 1664 TypeRawPtr::BOTTOM, new_pf_wmt ); 1665 transform_later(store_new_wmt); 1666 1667 // adding prefetches 1668 pf_phi_abio->init_req( fall_in_path, i_o ); 1669 1670 Node *prefetch_adr; 1671 Node *prefetch; 1672 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize; 1673 uint step_size = AllocatePrefetchStepSize; 1674 uint distance = 0; 1675 1676 for ( uint i = 0; i < lines; i++ ) { 1677 prefetch_adr = new (C, 4) AddPNode( old_pf_wm, new_pf_wmt, 1678 _igvn.MakeConX(distance) ); 1679 transform_later(prefetch_adr); 1680 prefetch = new (C, 3) PrefetchAllocationNode( i_o, prefetch_adr ); 1681 transform_later(prefetch); 1682 distance += step_size; 1683 i_o = prefetch; 1684 } 1685 pf_phi_abio->set_req( pf_path, i_o ); 1686 1687 pf_region->init_req( fall_in_path, need_pf_false ); 1688 pf_region->init_req( pf_path, need_pf_true ); 1689 1690 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 1691 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 1692 1693 transform_later(pf_region); 1694 transform_later(pf_phi_rawmem); 1695 transform_later(pf_phi_abio); 1696 1697 needgc_false = pf_region; 1698 contended_phi_rawmem = pf_phi_rawmem; 1699 i_o = pf_phi_abio; 1700 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { 1701 // Insert a prefetch for each allocation. 1702 // This code is used for Sparc with BIS. 1703 Node *pf_region = new (C, 3) RegionNode(3); 1704 Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY, 1705 TypeRawPtr::BOTTOM ); 1706 1707 // Generate several prefetch instructions. 1708 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1709 uint step_size = AllocatePrefetchStepSize; 1710 uint distance = AllocatePrefetchDistance; 1711 1712 // Next cache address. 1713 Node *cache_adr = new (C, 4) AddPNode(old_eden_top, old_eden_top, 1714 _igvn.MakeConX(distance)); 1715 transform_later(cache_adr); 1716 cache_adr = new (C, 2) CastP2XNode(needgc_false, cache_adr); 1717 transform_later(cache_adr); 1718 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); 1719 cache_adr = new (C, 3) AndXNode(cache_adr, mask); 1720 transform_later(cache_adr); 1721 cache_adr = new (C, 2) CastX2PNode(cache_adr); 1722 transform_later(cache_adr); 1723 1724 // Prefetch 1725 Node *prefetch = new (C, 3) PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); 1726 prefetch->set_req(0, needgc_false); 1727 transform_later(prefetch); 1728 contended_phi_rawmem = prefetch; 1729 Node *prefetch_adr; 1730 distance = step_size; 1731 for ( uint i = 1; i < lines; i++ ) { 1732 prefetch_adr = new (C, 4) AddPNode( cache_adr, cache_adr, 1733 _igvn.MakeConX(distance) ); 1734 transform_later(prefetch_adr); 1735 prefetch = new (C, 3) PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); 1736 transform_later(prefetch); 1737 distance += step_size; 1738 contended_phi_rawmem = prefetch; 1739 } 1740 } else if( AllocatePrefetchStyle > 0 ) { 1741 // Insert a prefetch for each allocation only on the fast-path 1742 Node *prefetch_adr; 1743 Node *prefetch; 1744 // Generate several prefetch instructions. 1745 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1746 uint step_size = AllocatePrefetchStepSize; 1747 uint distance = AllocatePrefetchDistance; 1748 for ( uint i = 0; i < lines; i++ ) { 1749 prefetch_adr = new (C, 4) AddPNode( old_eden_top, new_eden_top, 1750 _igvn.MakeConX(distance) ); 1751 transform_later(prefetch_adr); 1752 prefetch = new (C, 3) PrefetchAllocationNode( i_o, prefetch_adr ); 1753 // Do not let it float too high, since if eden_top == eden_end, 1754 // both might be null. 1755 if( i == 0 ) { // Set control for first prefetch, next follows it 1756 prefetch->init_req(0, needgc_false); 1757 } 1758 transform_later(prefetch); 1759 distance += step_size; 1760 i_o = prefetch; 1761 } 1762 } 1763 return i_o; 1764 } 1765 1766 1767 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 1768 expand_allocate_common(alloc, NULL, 1769 OptoRuntime::new_instance_Type(), 1770 OptoRuntime::new_instance_Java()); 1771 } 1772 1773 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 1774 Node* length = alloc->in(AllocateNode::ALength); 1775 InitializeNode* init = alloc->initialization(); 1776 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1777 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1778 address slow_call_address; // Address of slow call 1779 if (init != NULL && init->is_complete_with_arraycopy() && 1780 k->is_type_array_klass()) { 1781 // Don't zero type array during slow allocation in VM since 1782 // it will be initialized later by arraycopy in compiled code. 1783 slow_call_address = OptoRuntime::new_array_nozero_Java(); 1784 } else { 1785 slow_call_address = OptoRuntime::new_array_Java(); 1786 } 1787 expand_allocate_common(alloc, length, 1788 OptoRuntime::new_array_Type(), 1789 slow_call_address); 1790 } 1791 1792 //-----------------------mark_eliminated_locking_nodes----------------------- 1793 // During EA obj may point to several objects but after few ideal graph 1794 // transformations (CCP) it may point to only one non escaping object 1795 // (but still using phi), corresponding locks and unlocks will be marked 1796 // for elimination. Later obj could be replaced with a new node (new phi) 1797 // and which does not have escape information. And later after some graph 1798 // reshape other locks and unlocks (which were not marked for elimination 1799 // before) are connected to this new obj (phi) but they still will not be 1800 // marked for elimination since new obj has no escape information. 1801 // Mark all associated (same box and obj) lock and unlock nodes for 1802 // elimination if some of them marked already. 1803 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { 1804 if (!alock->is_eliminated()) { 1805 return; 1806 } 1807 if (!alock->is_coarsened()) { // Eliminated by EA 1808 // Create new "eliminated" BoxLock node and use it 1809 // in monitor debug info for the same object. 1810 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 1811 Node* obj = alock->obj_node(); 1812 if (!oldbox->is_eliminated()) { 1813 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 1814 // Note: BoxLock node is marked eliminated only here 1815 // and it is used to indicate that all associated lock 1816 // and unlock nodes are marked for elimination. 1817 newbox->set_eliminated(); 1818 transform_later(newbox); 1819 // Replace old box node with new box for all users 1820 // of the same object. 1821 for (uint i = 0; i < oldbox->outcnt();) { 1822 1823 bool next_edge = true; 1824 Node* u = oldbox->raw_out(i); 1825 if (u->is_AbstractLock() && 1826 u->as_AbstractLock()->obj_node() == obj && 1827 u->as_AbstractLock()->box_node() == oldbox) { 1828 // Mark all associated locks and unlocks. 1829 u->as_AbstractLock()->set_eliminated(); 1830 _igvn.hash_delete(u); 1831 u->set_req(TypeFunc::Parms + 1, newbox); 1832 next_edge = false; 1833 } 1834 // Replace old box in monitor debug info. 1835 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 1836 SafePointNode* sfn = u->as_SafePoint(); 1837 JVMState* youngest_jvms = sfn->jvms(); 1838 int max_depth = youngest_jvms->depth(); 1839 for (int depth = 1; depth <= max_depth; depth++) { 1840 JVMState* jvms = youngest_jvms->of_depth(depth); 1841 int num_mon = jvms->nof_monitors(); 1842 // Loop over monitors 1843 for (int idx = 0; idx < num_mon; idx++) { 1844 Node* obj_node = sfn->monitor_obj(jvms, idx); 1845 Node* box_node = sfn->monitor_box(jvms, idx); 1846 if (box_node == oldbox && obj_node == obj) { 1847 int j = jvms->monitor_box_offset(idx); 1848 _igvn.hash_delete(u); 1849 u->set_req(j, newbox); 1850 next_edge = false; 1851 } 1852 } // for (int idx = 0; 1853 } // for (int depth = 1; 1854 } // if (u->is_SafePoint() 1855 if (next_edge) i++; 1856 } // for (uint i = 0; i < oldbox->outcnt();) 1857 } // if (!oldbox->is_eliminated()) 1858 } // if (!alock->is_coarsened()) 1859 } 1860 1861 // we have determined that this lock/unlock can be eliminated, we simply 1862 // eliminate the node without expanding it. 1863 // 1864 // Note: The membar's associated with the lock/unlock are currently not 1865 // eliminated. This should be investigated as a future enhancement. 1866 // 1867 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 1868 1869 if (!alock->is_eliminated()) { 1870 return false; 1871 } 1872 #ifdef ASSERT 1873 if (alock->is_Lock() && !alock->is_coarsened()) { 1874 // Check that new "eliminated" BoxLock node is created. 1875 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 1876 assert(oldbox->is_eliminated(), "should be done already"); 1877 } 1878 #endif 1879 CompileLog* log = C->log(); 1880 if (log != NULL) { 1881 log->head("eliminate_lock lock='%d'", 1882 alock->is_Lock()); 1883 JVMState* p = alock->jvms(); 1884 while (p != NULL) { 1885 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1886 p = p->caller(); 1887 } 1888 log->tail("eliminate_lock"); 1889 } 1890 1891 #ifndef PRODUCT 1892 if (PrintEliminateLocks) { 1893 if (alock->is_Lock()) { 1894 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx); 1895 } else { 1896 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx); 1897 } 1898 } 1899 #endif 1900 1901 Node* mem = alock->in(TypeFunc::Memory); 1902 Node* ctrl = alock->in(TypeFunc::Control); 1903 1904 extract_call_projections(alock); 1905 // There are 2 projections from the lock. The lock node will 1906 // be deleted when its last use is subsumed below. 1907 assert(alock->outcnt() == 2 && 1908 _fallthroughproj != NULL && 1909 _memproj_fallthrough != NULL, 1910 "Unexpected projections from Lock/Unlock"); 1911 1912 Node* fallthroughproj = _fallthroughproj; 1913 Node* memproj_fallthrough = _memproj_fallthrough; 1914 1915 // The memory projection from a lock/unlock is RawMem 1916 // The input to a Lock is merged memory, so extract its RawMem input 1917 // (unless the MergeMem has been optimized away.) 1918 if (alock->is_Lock()) { 1919 // Seach for MemBarAcquireLock node and delete it also. 1920 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 1921 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, ""); 1922 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 1923 Node* memproj = membar->proj_out(TypeFunc::Memory); 1924 _igvn.replace_node(ctrlproj, fallthroughproj); 1925 _igvn.replace_node(memproj, memproj_fallthrough); 1926 1927 // Delete FastLock node also if this Lock node is unique user 1928 // (a loop peeling may clone a Lock node). 1929 Node* flock = alock->as_Lock()->fastlock_node(); 1930 if (flock->outcnt() == 1) { 1931 assert(flock->unique_out() == alock, "sanity"); 1932 _igvn.replace_node(flock, top()); 1933 } 1934 } 1935 1936 // Seach for MemBarReleaseLock node and delete it also. 1937 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() && 1938 ctrl->in(0)->is_MemBar()) { 1939 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 1940 assert(membar->Opcode() == Op_MemBarReleaseLock && 1941 mem->is_Proj() && membar == mem->in(0), ""); 1942 _igvn.replace_node(fallthroughproj, ctrl); 1943 _igvn.replace_node(memproj_fallthrough, mem); 1944 fallthroughproj = ctrl; 1945 memproj_fallthrough = mem; 1946 ctrl = membar->in(TypeFunc::Control); 1947 mem = membar->in(TypeFunc::Memory); 1948 } 1949 1950 _igvn.replace_node(fallthroughproj, ctrl); 1951 _igvn.replace_node(memproj_fallthrough, mem); 1952 return true; 1953 } 1954 1955 1956 //------------------------------expand_lock_node---------------------- 1957 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 1958 1959 Node* ctrl = lock->in(TypeFunc::Control); 1960 Node* mem = lock->in(TypeFunc::Memory); 1961 Node* obj = lock->obj_node(); 1962 Node* box = lock->box_node(); 1963 Node* flock = lock->fastlock_node(); 1964 1965 // Make the merge point 1966 Node *region; 1967 Node *mem_phi; 1968 Node *slow_path; 1969 1970 if (UseOptoBiasInlining) { 1971 /* 1972 * See the full description in MacroAssembler::biased_locking_enter(). 1973 * 1974 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { 1975 * // The object is biased. 1976 * proto_node = klass->prototype_header; 1977 * o_node = thread | proto_node; 1978 * x_node = o_node ^ mark_word; 1979 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? 1980 * // Done. 1981 * } else { 1982 * if( (x_node & biased_lock_mask) != 0 ) { 1983 * // The klass's prototype header is no longer biased. 1984 * cas(&mark_word, mark_word, proto_node) 1985 * goto cas_lock; 1986 * } else { 1987 * // The klass's prototype header is still biased. 1988 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? 1989 * old = mark_word; 1990 * new = o_node; 1991 * } else { 1992 * // Different thread or anonymous biased. 1993 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); 1994 * new = thread | old; 1995 * } 1996 * // Try to rebias. 1997 * if( cas(&mark_word, old, new) == 0 ) { 1998 * // Done. 1999 * } else { 2000 * goto slow_path; // Failed. 2001 * } 2002 * } 2003 * } 2004 * } else { 2005 * // The object is not biased. 2006 * cas_lock: 2007 * if( FastLock(obj) == 0 ) { 2008 * // Done. 2009 * } else { 2010 * slow_path: 2011 * OptoRuntime::complete_monitor_locking_Java(obj); 2012 * } 2013 * } 2014 */ 2015 2016 region = new (C, 5) RegionNode(5); 2017 // create a Phi for the memory state 2018 mem_phi = new (C, 5) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2019 2020 Node* fast_lock_region = new (C, 3) RegionNode(3); 2021 Node* fast_lock_mem_phi = new (C, 3) PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); 2022 2023 // First, check mark word for the biased lock pattern. 2024 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2025 2026 // Get fast path - mark word has the biased lock pattern. 2027 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, 2028 markOopDesc::biased_lock_mask_in_place, 2029 markOopDesc::biased_lock_pattern, true); 2030 // fast_lock_region->in(1) is set to slow path. 2031 fast_lock_mem_phi->init_req(1, mem); 2032 2033 // Now check that the lock is biased to the current thread and has 2034 // the same epoch and bias as Klass::_prototype_header. 2035 2036 // Special-case a fresh allocation to avoid building nodes: 2037 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); 2038 if (klass_node == NULL) { 2039 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 2040 klass_node = transform_later( LoadKlassNode::make(_igvn, mem, k_adr, _igvn.type(k_adr)->is_ptr()) ); 2041 #ifdef _LP64 2042 if (UseCompressedOops && klass_node->is_DecodeN()) { 2043 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity"); 2044 klass_node->in(1)->init_req(0, ctrl); 2045 } else 2046 #endif 2047 klass_node->init_req(0, ctrl); 2048 } 2049 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type()); 2050 2051 Node* thread = transform_later(new (C, 1) ThreadLocalNode()); 2052 Node* cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread)); 2053 Node* o_node = transform_later(new (C, 3) OrXNode(cast_thread, proto_node)); 2054 Node* x_node = transform_later(new (C, 3) XorXNode(o_node, mark_node)); 2055 2056 // Get slow path - mark word does NOT match the value. 2057 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, 2058 (~markOopDesc::age_mask_in_place), 0); 2059 // region->in(3) is set to fast path - the object is biased to the current thread. 2060 mem_phi->init_req(3, mem); 2061 2062 2063 // Mark word does NOT match the value (thread | Klass::_prototype_header). 2064 2065 2066 // First, check biased pattern. 2067 // Get fast path - _prototype_header has the same biased lock pattern. 2068 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, 2069 markOopDesc::biased_lock_mask_in_place, 0, true); 2070 2071 not_biased_ctrl = fast_lock_region->in(2); // Slow path 2072 // fast_lock_region->in(2) - the prototype header is no longer biased 2073 // and we have to revoke the bias on this object. 2074 // We are going to try to reset the mark of this object to the prototype 2075 // value and fall through to the CAS-based locking scheme. 2076 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 2077 Node* cas = new (C, 5) StoreXConditionalNode(not_biased_ctrl, mem, adr, 2078 proto_node, mark_node); 2079 transform_later(cas); 2080 Node* proj = transform_later( new (C, 1) SCMemProjNode(cas)); 2081 fast_lock_mem_phi->init_req(2, proj); 2082 2083 2084 // Second, check epoch bits. 2085 Node* rebiased_region = new (C, 3) RegionNode(3); 2086 Node* old_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X); 2087 Node* new_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X); 2088 2089 // Get slow path - mark word does NOT match epoch bits. 2090 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, 2091 markOopDesc::epoch_mask_in_place, 0); 2092 // The epoch of the current bias is not valid, attempt to rebias the object 2093 // toward the current thread. 2094 rebiased_region->init_req(2, epoch_ctrl); 2095 old_phi->init_req(2, mark_node); 2096 new_phi->init_req(2, o_node); 2097 2098 // rebiased_region->in(1) is set to fast path. 2099 // The epoch of the current bias is still valid but we know 2100 // nothing about the owner; it might be set or it might be clear. 2101 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place | 2102 markOopDesc::age_mask_in_place | 2103 markOopDesc::epoch_mask_in_place); 2104 Node* old = transform_later(new (C, 3) AndXNode(mark_node, cmask)); 2105 cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread)); 2106 Node* new_mark = transform_later(new (C, 3) OrXNode(cast_thread, old)); 2107 old_phi->init_req(1, old); 2108 new_phi->init_req(1, new_mark); 2109 2110 transform_later(rebiased_region); 2111 transform_later(old_phi); 2112 transform_later(new_phi); 2113 2114 // Try to acquire the bias of the object using an atomic operation. 2115 // If this fails we will go in to the runtime to revoke the object's bias. 2116 cas = new (C, 5) StoreXConditionalNode(rebiased_region, mem, adr, 2117 new_phi, old_phi); 2118 transform_later(cas); 2119 proj = transform_later( new (C, 1) SCMemProjNode(cas)); 2120 2121 // Get slow path - Failed to CAS. 2122 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); 2123 mem_phi->init_req(4, proj); 2124 // region->in(4) is set to fast path - the object is rebiased to the current thread. 2125 2126 // Failed to CAS. 2127 slow_path = new (C, 3) RegionNode(3); 2128 Node *slow_mem = new (C, 3) PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); 2129 2130 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control 2131 slow_mem->init_req(1, proj); 2132 2133 // Call CAS-based locking scheme (FastLock node). 2134 2135 transform_later(fast_lock_region); 2136 transform_later(fast_lock_mem_phi); 2137 2138 // Get slow path - FastLock failed to lock the object. 2139 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); 2140 mem_phi->init_req(2, fast_lock_mem_phi); 2141 // region->in(2) is set to fast path - the object is locked to the current thread. 2142 2143 slow_path->init_req(2, ctrl); // Capture slow-control 2144 slow_mem->init_req(2, fast_lock_mem_phi); 2145 2146 transform_later(slow_path); 2147 transform_later(slow_mem); 2148 // Reset lock's memory edge. 2149 lock->set_req(TypeFunc::Memory, slow_mem); 2150 2151 } else { 2152 region = new (C, 3) RegionNode(3); 2153 // create a Phi for the memory state 2154 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2155 2156 // Optimize test; set region slot 2 2157 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); 2158 mem_phi->init_req(2, mem); 2159 } 2160 2161 // Make slow path call 2162 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box ); 2163 2164 extract_call_projections(call); 2165 2166 // Slow path can only throw asynchronous exceptions, which are always 2167 // de-opted. So the compiler thinks the slow-call can never throw an 2168 // exception. If it DOES throw an exception we would need the debug 2169 // info removed first (since if it throws there is no monitor). 2170 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2171 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2172 2173 // Capture slow path 2174 // disconnect fall-through projection from call and create a new one 2175 // hook up users of fall-through projection to region 2176 Node *slow_ctrl = _fallthroughproj->clone(); 2177 transform_later(slow_ctrl); 2178 _igvn.hash_delete(_fallthroughproj); 2179 _fallthroughproj->disconnect_inputs(NULL); 2180 region->init_req(1, slow_ctrl); 2181 // region inputs are now complete 2182 transform_later(region); 2183 _igvn.replace_node(_fallthroughproj, region); 2184 2185 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) ); 2186 mem_phi->init_req(1, memproj ); 2187 transform_later(mem_phi); 2188 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2189 } 2190 2191 //------------------------------expand_unlock_node---------------------- 2192 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 2193 2194 Node* ctrl = unlock->in(TypeFunc::Control); 2195 Node* mem = unlock->in(TypeFunc::Memory); 2196 Node* obj = unlock->obj_node(); 2197 Node* box = unlock->box_node(); 2198 2199 // No need for a null check on unlock 2200 2201 // Make the merge point 2202 Node *region; 2203 Node *mem_phi; 2204 2205 if (UseOptoBiasInlining) { 2206 // Check for biased locking unlock case, which is a no-op. 2207 // See the full description in MacroAssembler::biased_locking_exit(). 2208 region = new (C, 4) RegionNode(4); 2209 // create a Phi for the memory state 2210 mem_phi = new (C, 4) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2211 mem_phi->init_req(3, mem); 2212 2213 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2214 ctrl = opt_bits_test(ctrl, region, 3, mark_node, 2215 markOopDesc::biased_lock_mask_in_place, 2216 markOopDesc::biased_lock_pattern); 2217 } else { 2218 region = new (C, 3) RegionNode(3); 2219 // create a Phi for the memory state 2220 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2221 } 2222 2223 FastUnlockNode *funlock = new (C, 3) FastUnlockNode( ctrl, obj, box ); 2224 funlock = transform_later( funlock )->as_FastUnlock(); 2225 // Optimize test; set region slot 2 2226 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); 2227 2228 CallNode *call = make_slow_call( (CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), "complete_monitor_unlocking_C", slow_path, obj, box ); 2229 2230 extract_call_projections(call); 2231 2232 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2233 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2234 2235 // No exceptions for unlocking 2236 // Capture slow path 2237 // disconnect fall-through projection from call and create a new one 2238 // hook up users of fall-through projection to region 2239 Node *slow_ctrl = _fallthroughproj->clone(); 2240 transform_later(slow_ctrl); 2241 _igvn.hash_delete(_fallthroughproj); 2242 _fallthroughproj->disconnect_inputs(NULL); 2243 region->init_req(1, slow_ctrl); 2244 // region inputs are now complete 2245 transform_later(region); 2246 _igvn.replace_node(_fallthroughproj, region); 2247 2248 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) ); 2249 mem_phi->init_req(1, memproj ); 2250 mem_phi->init_req(2, mem); 2251 transform_later(mem_phi); 2252 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2253 } 2254 2255 //---------------------------eliminate_macro_nodes---------------------- 2256 // Eliminate scalar replaced allocations and associated locks. 2257 void PhaseMacroExpand::eliminate_macro_nodes() { 2258 if (C->macro_count() == 0) 2259 return; 2260 2261 // First, attempt to eliminate locks 2262 int cnt = C->macro_count(); 2263 for (int i=0; i < cnt; i++) { 2264 Node *n = C->macro_node(i); 2265 if (n->is_AbstractLock()) { // Lock and Unlock nodes 2266 // Before elimination mark all associated (same box and obj) 2267 // lock and unlock nodes. 2268 mark_eliminated_locking_nodes(n->as_AbstractLock()); 2269 } 2270 } 2271 bool progress = true; 2272 while (progress) { 2273 progress = false; 2274 for (int i = C->macro_count(); i > 0; i--) { 2275 Node * n = C->macro_node(i-1); 2276 bool success = false; 2277 debug_only(int old_macro_count = C->macro_count();); 2278 if (n->is_AbstractLock()) { 2279 success = eliminate_locking_node(n->as_AbstractLock()); 2280 } 2281 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2282 progress = progress || success; 2283 } 2284 } 2285 // Next, attempt to eliminate allocations 2286 progress = true; 2287 while (progress) { 2288 progress = false; 2289 for (int i = C->macro_count(); i > 0; i--) { 2290 Node * n = C->macro_node(i-1); 2291 bool success = false; 2292 debug_only(int old_macro_count = C->macro_count();); 2293 switch (n->class_id()) { 2294 case Node::Class_Allocate: 2295 case Node::Class_AllocateArray: 2296 success = eliminate_allocate_node(n->as_Allocate()); 2297 break; 2298 case Node::Class_Lock: 2299 case Node::Class_Unlock: 2300 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 2301 break; 2302 default: 2303 assert(n->Opcode() == Op_LoopLimit || 2304 n->Opcode() == Op_Opaque1 || 2305 n->Opcode() == Op_Opaque2, "unknown node type in macro list"); 2306 } 2307 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2308 progress = progress || success; 2309 } 2310 } 2311 } 2312 2313 //------------------------------expand_macro_nodes---------------------- 2314 // Returns true if a failure occurred. 2315 bool PhaseMacroExpand::expand_macro_nodes() { 2316 // Last attempt to eliminate macro nodes. 2317 eliminate_macro_nodes(); 2318 2319 // Make sure expansion will not cause node limit to be exceeded. 2320 // Worst case is a macro node gets expanded into about 50 nodes. 2321 // Allow 50% more for optimization. 2322 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) ) 2323 return true; 2324 2325 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. 2326 bool progress = true; 2327 while (progress) { 2328 progress = false; 2329 for (int i = C->macro_count(); i > 0; i--) { 2330 Node * n = C->macro_node(i-1); 2331 bool success = false; 2332 debug_only(int old_macro_count = C->macro_count();); 2333 if (n->Opcode() == Op_LoopLimit) { 2334 // Remove it from macro list and put on IGVN worklist to optimize. 2335 C->remove_macro_node(n); 2336 _igvn._worklist.push(n); 2337 success = true; 2338 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { 2339 _igvn.replace_node(n, n->in(1)); 2340 success = true; 2341 } 2342 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2343 progress = progress || success; 2344 } 2345 } 2346 2347 // expand "macro" nodes 2348 // nodes are removed from the macro list as they are processed 2349 while (C->macro_count() > 0) { 2350 int macro_count = C->macro_count(); 2351 Node * n = C->macro_node(macro_count-1); 2352 assert(n->is_macro(), "only macro nodes expected here"); 2353 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2354 // node is unreachable, so don't try to expand it 2355 C->remove_macro_node(n); 2356 continue; 2357 } 2358 switch (n->class_id()) { 2359 case Node::Class_Allocate: 2360 expand_allocate(n->as_Allocate()); 2361 break; 2362 case Node::Class_AllocateArray: 2363 expand_allocate_array(n->as_AllocateArray()); 2364 break; 2365 case Node::Class_Lock: 2366 expand_lock_node(n->as_Lock()); 2367 break; 2368 case Node::Class_Unlock: 2369 expand_unlock_node(n->as_Unlock()); 2370 break; 2371 default: 2372 assert(false, "unknown node type in macro list"); 2373 } 2374 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2375 if (C->failing()) return true; 2376 } 2377 2378 _igvn.set_delay_transform(false); 2379 _igvn.optimize(); 2380 if (C->failing()) return true; 2381 return false; 2382 }