1 /* 2 * Copyright 2005-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 #include "incls/_precompiled.incl" 26 #include "incls/_macro.cpp.incl" 27 28 29 // 30 // Replace any references to "oldref" in inputs to "use" with "newref". 31 // Returns the number of replacements made. 32 // 33 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { 34 int nreplacements = 0; 35 uint req = use->req(); 36 for (uint j = 0; j < use->len(); j++) { 37 Node *uin = use->in(j); 38 if (uin == oldref) { 39 if (j < req) 40 use->set_req(j, newref); 41 else 42 use->set_prec(j, newref); 43 nreplacements++; 44 } else if (j >= req && uin == NULL) { 45 break; 46 } 47 } 48 return nreplacements; 49 } 50 51 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) { 52 // Copy debug information and adjust JVMState information 53 uint old_dbg_start = oldcall->tf()->domain()->cnt(); 54 uint new_dbg_start = newcall->tf()->domain()->cnt(); 55 int jvms_adj = new_dbg_start - old_dbg_start; 56 assert (new_dbg_start == newcall->req(), "argument count mismatch"); 57 58 Dict* sosn_map = new Dict(cmpkey,hashkey); 59 for (uint i = old_dbg_start; i < oldcall->req(); i++) { 60 Node* old_in = oldcall->in(i); 61 // Clone old SafePointScalarObjectNodes, adjusting their field contents. 62 if (old_in != NULL && old_in->is_SafePointScalarObject()) { 63 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject(); 64 uint old_unique = C->unique(); 65 Node* new_in = old_sosn->clone(jvms_adj, sosn_map); 66 if (old_unique != C->unique()) { 67 new_in->set_req(0, newcall->in(0)); // reset control edge 68 new_in = transform_later(new_in); // Register new node. 69 } 70 old_in = new_in; 71 } 72 newcall->add_req(old_in); 73 } 74 75 newcall->set_jvms(oldcall->jvms()); 76 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) { 77 jvms->set_map(newcall); 78 jvms->set_locoff(jvms->locoff()+jvms_adj); 79 jvms->set_stkoff(jvms->stkoff()+jvms_adj); 80 jvms->set_monoff(jvms->monoff()+jvms_adj); 81 jvms->set_scloff(jvms->scloff()+jvms_adj); 82 jvms->set_endoff(jvms->endoff()+jvms_adj); 83 } 84 } 85 86 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) { 87 Node* cmp; 88 if (mask != 0) { 89 Node* and_node = transform_later(new (C, 3) AndXNode(word, MakeConX(mask))); 90 cmp = transform_later(new (C, 3) CmpXNode(and_node, MakeConX(bits))); 91 } else { 92 cmp = word; 93 } 94 Node* bol = transform_later(new (C, 2) BoolNode(cmp, BoolTest::ne)); 95 IfNode* iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); 96 transform_later(iff); 97 98 // Fast path taken. 99 Node *fast_taken = transform_later( new (C, 1) IfFalseNode(iff) ); 100 101 // Fast path not-taken, i.e. slow path 102 Node *slow_taken = transform_later( new (C, 1) IfTrueNode(iff) ); 103 104 if (return_fast_path) { 105 region->init_req(edge, slow_taken); // Capture slow-control 106 return fast_taken; 107 } else { 108 region->init_req(edge, fast_taken); // Capture fast-control 109 return slow_taken; 110 } 111 } 112 113 //--------------------copy_predefined_input_for_runtime_call-------------------- 114 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { 115 // Set fixed predefined input arguments 116 call->init_req( TypeFunc::Control, ctrl ); 117 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); 118 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? 119 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); 120 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); 121 } 122 123 //------------------------------make_slow_call--------------------------------- 124 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) { 125 126 // Slow-path call 127 int size = slow_call_type->domain()->cnt(); 128 CallNode *call = leaf_name 129 ? (CallNode*)new (C, size) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) 130 : (CallNode*)new (C, size) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM ); 131 132 // Slow path call has no side-effects, uses few values 133 copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); 134 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0); 135 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1); 136 copy_call_debug_info(oldcall, call); 137 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 138 _igvn.hash_delete(oldcall); 139 _igvn.subsume_node(oldcall, call); 140 transform_later(call); 141 142 return call; 143 } 144 145 void PhaseMacroExpand::extract_call_projections(CallNode *call) { 146 _fallthroughproj = NULL; 147 _fallthroughcatchproj = NULL; 148 _ioproj_fallthrough = NULL; 149 _ioproj_catchall = NULL; 150 _catchallcatchproj = NULL; 151 _memproj_fallthrough = NULL; 152 _memproj_catchall = NULL; 153 _resproj = NULL; 154 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) { 155 ProjNode *pn = call->fast_out(i)->as_Proj(); 156 switch (pn->_con) { 157 case TypeFunc::Control: 158 { 159 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj 160 _fallthroughproj = pn; 161 DUIterator_Fast jmax, j = pn->fast_outs(jmax); 162 const Node *cn = pn->fast_out(j); 163 if (cn->is_Catch()) { 164 ProjNode *cpn = NULL; 165 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { 166 cpn = cn->fast_out(k)->as_Proj(); 167 assert(cpn->is_CatchProj(), "must be a CatchProjNode"); 168 if (cpn->_con == CatchProjNode::fall_through_index) 169 _fallthroughcatchproj = cpn; 170 else { 171 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); 172 _catchallcatchproj = cpn; 173 } 174 } 175 } 176 break; 177 } 178 case TypeFunc::I_O: 179 if (pn->_is_io_use) 180 _ioproj_catchall = pn; 181 else 182 _ioproj_fallthrough = pn; 183 break; 184 case TypeFunc::Memory: 185 if (pn->_is_io_use) 186 _memproj_catchall = pn; 187 else 188 _memproj_fallthrough = pn; 189 break; 190 case TypeFunc::Parms: 191 _resproj = pn; 192 break; 193 default: 194 assert(false, "unexpected projection from allocation node."); 195 } 196 } 197 198 } 199 200 // Eliminate a card mark sequence. p2x is a ConvP2XNode 201 void PhaseMacroExpand::eliminate_card_mark(Node *p2x) { 202 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required"); 203 Node *shift = p2x->unique_out(); 204 Node *addp = shift->unique_out(); 205 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) { 206 Node *st = addp->last_out(j); 207 assert(st->is_Store(), "store required"); 208 _igvn.replace_node(st, st->in(MemNode::Memory)); 209 } 210 } 211 212 // Search for a memory operation for the specified memory slice. 213 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) { 214 Node *orig_mem = mem; 215 Node *alloc_mem = alloc->in(TypeFunc::Memory); 216 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr(); 217 while (true) { 218 if (mem == alloc_mem || mem == start_mem ) { 219 return mem; // hit one of our sentinels 220 } else if (mem->is_MergeMem()) { 221 mem = mem->as_MergeMem()->memory_at(alias_idx); 222 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { 223 Node *in = mem->in(0); 224 // we can safely skip over safepoints, calls, locks and membars because we 225 // already know that the object is safe to eliminate. 226 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { 227 return in; 228 } else if (in->is_Call()) { 229 CallNode *call = in->as_Call(); 230 if (!call->may_modify(tinst, phase)) { 231 mem = call->in(TypeFunc::Memory); 232 } 233 mem = in->in(TypeFunc::Memory); 234 } else if (in->is_MemBar()) { 235 mem = in->in(TypeFunc::Memory); 236 } else { 237 assert(false, "unexpected projection"); 238 } 239 } else if (mem->is_Store()) { 240 const TypePtr* atype = mem->as_Store()->adr_type(); 241 int adr_idx = Compile::current()->get_alias_index(atype); 242 if (adr_idx == alias_idx) { 243 assert(atype->isa_oopptr(), "address type must be oopptr"); 244 int adr_offset = atype->offset(); 245 uint adr_iid = atype->is_oopptr()->instance_id(); 246 // Array elements references have the same alias_idx 247 // but different offset and different instance_id. 248 if (adr_offset == offset && adr_iid == alloc->_idx) 249 return mem; 250 } else { 251 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw"); 252 } 253 mem = mem->in(MemNode::Memory); 254 } else if (mem->Opcode() == Op_SCMemProj) { 255 assert(mem->in(0)->is_LoadStore(), "sanity"); 256 const TypePtr* atype = mem->in(0)->in(MemNode::Address)->bottom_type()->is_ptr(); 257 int adr_idx = Compile::current()->get_alias_index(atype); 258 if (adr_idx == alias_idx) { 259 assert(false, "Object is not scalar replaceable if a LoadStore node access its field"); 260 return NULL; 261 } 262 mem = mem->in(0)->in(MemNode::Memory); 263 } else { 264 return mem; 265 } 266 assert(mem != orig_mem, "dead memory loop"); 267 } 268 } 269 270 // 271 // Given a Memory Phi, compute a value Phi containing the values from stores 272 // on the input paths. 273 // Note: this function is recursive, its depth is limied by the "level" argument 274 // Returns the computed Phi, or NULL if it cannot compute it. 275 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) { 276 assert(mem->is_Phi(), "sanity"); 277 int alias_idx = C->get_alias_index(adr_t); 278 int offset = adr_t->offset(); 279 int instance_id = adr_t->instance_id(); 280 281 // Check if an appropriate value phi already exists. 282 Node* region = mem->in(0); 283 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { 284 Node* phi = region->fast_out(k); 285 if (phi->is_Phi() && phi != mem && 286 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) { 287 return phi; 288 } 289 } 290 // Check if an appropriate new value phi already exists. 291 Node* new_phi = NULL; 292 uint size = value_phis->size(); 293 for (uint i=0; i < size; i++) { 294 if ( mem->_idx == value_phis->index_at(i) ) { 295 return value_phis->node_at(i); 296 } 297 } 298 299 if (level <= 0) { 300 return NULL; // Give up: phi tree too deep 301 } 302 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 303 Node *alloc_mem = alloc->in(TypeFunc::Memory); 304 305 uint length = mem->req(); 306 GrowableArray <Node *> values(length, length, NULL); 307 308 // create a new Phi for the value 309 PhiNode *phi = new (C, length) PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset); 310 transform_later(phi); 311 value_phis->push(phi, mem->_idx); 312 313 for (uint j = 1; j < length; j++) { 314 Node *in = mem->in(j); 315 if (in == NULL || in->is_top()) { 316 values.at_put(j, in); 317 } else { 318 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); 319 if (val == start_mem || val == alloc_mem) { 320 // hit a sentinel, return appropriate 0 value 321 values.at_put(j, _igvn.zerocon(ft)); 322 continue; 323 } 324 if (val->is_Initialize()) { 325 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 326 } 327 if (val == NULL) { 328 return NULL; // can't find a value on this path 329 } 330 if (val == mem) { 331 values.at_put(j, mem); 332 } else if (val->is_Store()) { 333 values.at_put(j, val->in(MemNode::ValueIn)); 334 } else if(val->is_Proj() && val->in(0) == alloc) { 335 values.at_put(j, _igvn.zerocon(ft)); 336 } else if (val->is_Phi()) { 337 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); 338 if (val == NULL) { 339 return NULL; 340 } 341 values.at_put(j, val); 342 } else if (val->Opcode() == Op_SCMemProj) { 343 assert(val->in(0)->is_LoadStore(), "sanity"); 344 assert(false, "Object is not scalar replaceable if a LoadStore node access its field"); 345 return NULL; 346 } else { 347 #ifdef ASSERT 348 val->dump(); 349 assert(false, "unknown node on this path"); 350 #endif 351 return NULL; // unknown node on this path 352 } 353 } 354 } 355 // Set Phi's inputs 356 for (uint j = 1; j < length; j++) { 357 if (values.at(j) == mem) { 358 phi->init_req(j, phi); 359 } else { 360 phi->init_req(j, values.at(j)); 361 } 362 } 363 return phi; 364 } 365 366 // Search the last value stored into the object's field. 367 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) { 368 assert(adr_t->is_known_instance_field(), "instance required"); 369 int instance_id = adr_t->instance_id(); 370 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 371 372 int alias_idx = C->get_alias_index(adr_t); 373 int offset = adr_t->offset(); 374 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 375 Node *alloc_ctrl = alloc->in(TypeFunc::Control); 376 Node *alloc_mem = alloc->in(TypeFunc::Memory); 377 Arena *a = Thread::current()->resource_area(); 378 VectorSet visited(a); 379 380 381 bool done = sfpt_mem == alloc_mem; 382 Node *mem = sfpt_mem; 383 while (!done) { 384 if (visited.test_set(mem->_idx)) { 385 return NULL; // found a loop, give up 386 } 387 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 388 if (mem == start_mem || mem == alloc_mem) { 389 done = true; // hit a sentinel, return appropriate 0 value 390 } else if (mem->is_Initialize()) { 391 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 392 if (mem == NULL) { 393 done = true; // Something go wrong. 394 } else if (mem->is_Store()) { 395 const TypePtr* atype = mem->as_Store()->adr_type(); 396 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 397 done = true; 398 } 399 } else if (mem->is_Store()) { 400 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 401 assert(atype != NULL, "address type must be oopptr"); 402 assert(C->get_alias_index(atype) == alias_idx && 403 atype->is_known_instance_field() && atype->offset() == offset && 404 atype->instance_id() == instance_id, "store is correct memory slice"); 405 done = true; 406 } else if (mem->is_Phi()) { 407 // try to find a phi's unique input 408 Node *unique_input = NULL; 409 Node *top = C->top(); 410 for (uint i = 1; i < mem->req(); i++) { 411 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 412 if (n == NULL || n == top || n == mem) { 413 continue; 414 } else if (unique_input == NULL) { 415 unique_input = n; 416 } else if (unique_input != n) { 417 unique_input = top; 418 break; 419 } 420 } 421 if (unique_input != NULL && unique_input != top) { 422 mem = unique_input; 423 } else { 424 done = true; 425 } 426 } else { 427 assert(false, "unexpected node"); 428 } 429 } 430 if (mem != NULL) { 431 if (mem == start_mem || mem == alloc_mem) { 432 // hit a sentinel, return appropriate 0 value 433 return _igvn.zerocon(ft); 434 } else if (mem->is_Store()) { 435 return mem->in(MemNode::ValueIn); 436 } else if (mem->is_Phi()) { 437 // attempt to produce a Phi reflecting the values on the input paths of the Phi 438 Node_Stack value_phis(a, 8); 439 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 440 if (phi != NULL) { 441 return phi; 442 } else { 443 // Kill all new Phis 444 while(value_phis.is_nonempty()) { 445 Node* n = value_phis.node(); 446 _igvn.hash_delete(n); 447 _igvn.subsume_node(n, C->top()); 448 value_phis.pop(); 449 } 450 } 451 } 452 } 453 // Something go wrong. 454 return NULL; 455 } 456 457 // Check the possibility of scalar replacement. 458 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 459 // Scan the uses of the allocation to check for anything that would 460 // prevent us from eliminating it. 461 NOT_PRODUCT( const char* fail_eliminate = NULL; ) 462 DEBUG_ONLY( Node* disq_node = NULL; ) 463 bool can_eliminate = true; 464 465 Node* res = alloc->result_cast(); 466 const TypeOopPtr* res_type = NULL; 467 if (res == NULL) { 468 // All users were eliminated. 469 } else if (!res->is_CheckCastPP()) { 470 alloc->_is_scalar_replaceable = false; // don't try again 471 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 472 can_eliminate = false; 473 } else { 474 res_type = _igvn.type(res)->isa_oopptr(); 475 if (res_type == NULL) { 476 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 477 can_eliminate = false; 478 } else if (res_type->isa_aryptr()) { 479 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 480 if (length < 0) { 481 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 482 can_eliminate = false; 483 } 484 } 485 } 486 487 if (can_eliminate && res != NULL) { 488 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 489 j < jmax && can_eliminate; j++) { 490 Node* use = res->fast_out(j); 491 492 if (use->is_AddP()) { 493 const TypePtr* addp_type = _igvn.type(use)->is_ptr(); 494 int offset = addp_type->offset(); 495 496 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 497 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) 498 can_eliminate = false; 499 break; 500 } 501 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 502 k < kmax && can_eliminate; k++) { 503 Node* n = use->fast_out(k); 504 if (!n->is_Store() && n->Opcode() != Op_CastP2X) { 505 DEBUG_ONLY(disq_node = n;) 506 if (n->is_Load() || n->is_LoadStore()) { 507 NOT_PRODUCT(fail_eliminate = "Field load";) 508 } else { 509 NOT_PRODUCT(fail_eliminate = "Not store field referrence";) 510 } 511 can_eliminate = false; 512 } 513 } 514 } else if (use->is_SafePoint()) { 515 SafePointNode* sfpt = use->as_SafePoint(); 516 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 517 // Object is passed as argument. 518 DEBUG_ONLY(disq_node = use;) 519 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 520 can_eliminate = false; 521 } 522 Node* sfptMem = sfpt->memory(); 523 if (sfptMem == NULL || sfptMem->is_top()) { 524 DEBUG_ONLY(disq_node = use;) 525 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) 526 can_eliminate = false; 527 } else { 528 safepoints.append_if_missing(sfpt); 529 } 530 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 531 if (use->is_Phi()) { 532 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 533 NOT_PRODUCT(fail_eliminate = "Object is return value";) 534 } else { 535 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 536 } 537 DEBUG_ONLY(disq_node = use;) 538 } else { 539 if (use->Opcode() == Op_Return) { 540 NOT_PRODUCT(fail_eliminate = "Object is return value";) 541 }else { 542 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 543 } 544 DEBUG_ONLY(disq_node = use;) 545 } 546 can_eliminate = false; 547 } 548 } 549 } 550 551 #ifndef PRODUCT 552 if (PrintEliminateAllocations) { 553 if (can_eliminate) { 554 tty->print("Scalar "); 555 if (res == NULL) 556 alloc->dump(); 557 else 558 res->dump(); 559 } else { 560 tty->print("NotScalar (%s)", fail_eliminate); 561 if (res == NULL) 562 alloc->dump(); 563 else 564 res->dump(); 565 #ifdef ASSERT 566 if (disq_node != NULL) { 567 tty->print(" >>>> "); 568 disq_node->dump(); 569 } 570 #endif /*ASSERT*/ 571 } 572 } 573 #endif 574 return can_eliminate; 575 } 576 577 // Do scalar replacement. 578 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 579 GrowableArray <SafePointNode *> safepoints_done; 580 581 ciKlass* klass = NULL; 582 ciInstanceKlass* iklass = NULL; 583 int nfields = 0; 584 int array_base; 585 int element_size; 586 BasicType basic_elem_type; 587 ciType* elem_type; 588 589 Node* res = alloc->result_cast(); 590 const TypeOopPtr* res_type = NULL; 591 if (res != NULL) { // Could be NULL when there are no users 592 res_type = _igvn.type(res)->isa_oopptr(); 593 } 594 595 if (res != NULL) { 596 klass = res_type->klass(); 597 if (res_type->isa_instptr()) { 598 // find the fields of the class which will be needed for safepoint debug information 599 assert(klass->is_instance_klass(), "must be an instance klass."); 600 iklass = klass->as_instance_klass(); 601 nfields = iklass->nof_nonstatic_fields(); 602 } else { 603 // find the array's elements which will be needed for safepoint debug information 604 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 605 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); 606 elem_type = klass->as_array_klass()->element_type(); 607 basic_elem_type = elem_type->basic_type(); 608 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 609 element_size = type2aelembytes(basic_elem_type); 610 } 611 } 612 // 613 // Process the safepoint uses 614 // 615 while (safepoints.length() > 0) { 616 SafePointNode* sfpt = safepoints.pop(); 617 Node* mem = sfpt->memory(); 618 uint first_ind = sfpt->req(); 619 SafePointScalarObjectNode* sobj = new (C, 1) SafePointScalarObjectNode(res_type, 620 #ifdef ASSERT 621 alloc, 622 #endif 623 first_ind, nfields); 624 sobj->init_req(0, sfpt->in(TypeFunc::Control)); 625 transform_later(sobj); 626 627 // Scan object's fields adding an input to the safepoint for each field. 628 for (int j = 0; j < nfields; j++) { 629 intptr_t offset; 630 ciField* field = NULL; 631 if (iklass != NULL) { 632 field = iklass->nonstatic_field_at(j); 633 offset = field->offset(); 634 elem_type = field->type(); 635 basic_elem_type = field->layout_type(); 636 } else { 637 offset = array_base + j * (intptr_t)element_size; 638 } 639 640 const Type *field_type; 641 // The next code is taken from Parse::do_get_xxx(). 642 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { 643 if (!elem_type->is_loaded()) { 644 field_type = TypeInstPtr::BOTTOM; 645 } else if (field != NULL && field->is_constant()) { 646 // This can happen if the constant oop is non-perm. 647 ciObject* con = field->constant_value().as_object(); 648 // Do not "join" in the previous type; it doesn't add value, 649 // and may yield a vacuous result if the field is of interface type. 650 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 651 assert(field_type != NULL, "field singleton type must be consistent"); 652 } else { 653 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 654 } 655 if (UseCompressedOops) { 656 field_type = field_type->make_narrowoop(); 657 basic_elem_type = T_NARROWOOP; 658 } 659 } else { 660 field_type = Type::get_const_basic_type(basic_elem_type); 661 } 662 663 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 664 665 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc); 666 if (field_val == NULL) { 667 // we weren't able to find a value for this field, 668 // give up on eliminating this allocation 669 alloc->_is_scalar_replaceable = false; // don't try again 670 // remove any extra entries we added to the safepoint 671 uint last = sfpt->req() - 1; 672 for (int k = 0; k < j; k++) { 673 sfpt->del_req(last--); 674 } 675 // rollback processed safepoints 676 while (safepoints_done.length() > 0) { 677 SafePointNode* sfpt_done = safepoints_done.pop(); 678 // remove any extra entries we added to the safepoint 679 last = sfpt_done->req() - 1; 680 for (int k = 0; k < nfields; k++) { 681 sfpt_done->del_req(last--); 682 } 683 JVMState *jvms = sfpt_done->jvms(); 684 jvms->set_endoff(sfpt_done->req()); 685 // Now make a pass over the debug information replacing any references 686 // to SafePointScalarObjectNode with the allocated object. 687 int start = jvms->debug_start(); 688 int end = jvms->debug_end(); 689 for (int i = start; i < end; i++) { 690 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 691 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 692 if (scobj->first_index() == sfpt_done->req() && 693 scobj->n_fields() == (uint)nfields) { 694 assert(scobj->alloc() == alloc, "sanity"); 695 sfpt_done->set_req(i, res); 696 } 697 } 698 } 699 } 700 #ifndef PRODUCT 701 if (PrintEliminateAllocations) { 702 if (field != NULL) { 703 tty->print("=== At SafePoint node %d can't find value of Field: ", 704 sfpt->_idx); 705 field->print(); 706 int field_idx = C->get_alias_index(field_addr_type); 707 tty->print(" (alias_idx=%d)", field_idx); 708 } else { // Array's element 709 tty->print("=== At SafePoint node %d can't find value of array element [%d]", 710 sfpt->_idx, j); 711 } 712 tty->print(", which prevents elimination of: "); 713 if (res == NULL) 714 alloc->dump(); 715 else 716 res->dump(); 717 } 718 #endif 719 return false; 720 } 721 if (UseCompressedOops && field_type->isa_narrowoop()) { 722 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 723 // to be able scalar replace the allocation. 724 if (field_val->is_EncodeP()) { 725 field_val = field_val->in(1); 726 } else { 727 field_val = transform_later(new (C, 2) DecodeNNode(field_val, field_val->bottom_type()->make_ptr())); 728 } 729 } 730 sfpt->add_req(field_val); 731 } 732 JVMState *jvms = sfpt->jvms(); 733 jvms->set_endoff(sfpt->req()); 734 // Now make a pass over the debug information replacing any references 735 // to the allocated object with "sobj" 736 int start = jvms->debug_start(); 737 int end = jvms->debug_end(); 738 for (int i = start; i < end; i++) { 739 if (sfpt->in(i) == res) { 740 sfpt->set_req(i, sobj); 741 } 742 } 743 safepoints_done.append_if_missing(sfpt); // keep it for rollback 744 } 745 return true; 746 } 747 748 // Process users of eliminated allocation. 749 void PhaseMacroExpand::process_users_of_allocation(AllocateNode *alloc) { 750 Node* res = alloc->result_cast(); 751 if (res != NULL) { 752 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 753 Node *use = res->last_out(j); 754 uint oc1 = res->outcnt(); 755 756 if (use->is_AddP()) { 757 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 758 Node *n = use->last_out(k); 759 uint oc2 = use->outcnt(); 760 if (n->is_Store()) { 761 _igvn.replace_node(n, n->in(MemNode::Memory)); 762 } else { 763 assert( n->Opcode() == Op_CastP2X, "CastP2X required"); 764 eliminate_card_mark(n); 765 } 766 k -= (oc2 - use->outcnt()); 767 } 768 } else { 769 assert( !use->is_SafePoint(), "safepoint uses must have been already elimiated"); 770 assert( use->Opcode() == Op_CastP2X, "CastP2X required"); 771 eliminate_card_mark(use); 772 } 773 j -= (oc1 - res->outcnt()); 774 } 775 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 776 _igvn.remove_dead_node(res); 777 } 778 779 // 780 // Process other users of allocation's projections 781 // 782 if (_resproj != NULL && _resproj->outcnt() != 0) { 783 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { 784 Node *use = _resproj->last_out(j); 785 uint oc1 = _resproj->outcnt(); 786 if (use->is_Initialize()) { 787 // Eliminate Initialize node. 788 InitializeNode *init = use->as_Initialize(); 789 assert(init->outcnt() <= 2, "only a control and memory projection expected"); 790 Node *ctrl_proj = init->proj_out(TypeFunc::Control); 791 if (ctrl_proj != NULL) { 792 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection"); 793 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj); 794 } 795 Node *mem_proj = init->proj_out(TypeFunc::Memory); 796 if (mem_proj != NULL) { 797 Node *mem = init->in(TypeFunc::Memory); 798 #ifdef ASSERT 799 if (mem->is_MergeMem()) { 800 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection"); 801 } else { 802 assert(mem == _memproj_fallthrough, "allocation memory projection"); 803 } 804 #endif 805 _igvn.replace_node(mem_proj, mem); 806 } 807 } else if (use->is_AddP()) { 808 // raw memory addresses used only by the initialization 809 _igvn.hash_delete(use); 810 _igvn.subsume_node(use, C->top()); 811 } else { 812 assert(false, "only Initialize or AddP expected"); 813 } 814 j -= (oc1 - _resproj->outcnt()); 815 } 816 } 817 if (_fallthroughcatchproj != NULL) { 818 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); 819 } 820 if (_memproj_fallthrough != NULL) { 821 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); 822 } 823 if (_memproj_catchall != NULL) { 824 _igvn.replace_node(_memproj_catchall, C->top()); 825 } 826 if (_ioproj_fallthrough != NULL) { 827 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); 828 } 829 if (_ioproj_catchall != NULL) { 830 _igvn.replace_node(_ioproj_catchall, C->top()); 831 } 832 if (_catchallcatchproj != NULL) { 833 _igvn.replace_node(_catchallcatchproj, C->top()); 834 } 835 } 836 837 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 838 839 if (!EliminateAllocations || !alloc->_is_scalar_replaceable) { 840 return false; 841 } 842 843 extract_call_projections(alloc); 844 845 GrowableArray <SafePointNode *> safepoints; 846 if (!can_eliminate_allocation(alloc, safepoints)) { 847 return false; 848 } 849 850 if (!scalar_replacement(alloc, safepoints)) { 851 return false; 852 } 853 854 process_users_of_allocation(alloc); 855 856 #ifndef PRODUCT 857 if (PrintEliminateAllocations) { 858 if (alloc->is_AllocateArray()) 859 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 860 else 861 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 862 } 863 #endif 864 865 return true; 866 } 867 868 869 //---------------------------set_eden_pointers------------------------- 870 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { 871 if (UseTLAB) { // Private allocation: load from TLS 872 Node* thread = transform_later(new (C, 1) ThreadLocalNode()); 873 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); 874 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); 875 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); 876 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); 877 } else { // Shared allocation: load from globals 878 CollectedHeap* ch = Universe::heap(); 879 address top_adr = (address)ch->top_addr(); 880 address end_adr = (address)ch->end_addr(); 881 eden_top_adr = makecon(TypeRawPtr::make(top_adr)); 882 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); 883 } 884 } 885 886 887 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 888 Node* adr = basic_plus_adr(base, offset); 889 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 890 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt); 891 transform_later(value); 892 return value; 893 } 894 895 896 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 897 Node* adr = basic_plus_adr(base, offset); 898 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt); 899 transform_later(mem); 900 return mem; 901 } 902 903 //============================================================================= 904 // 905 // A L L O C A T I O N 906 // 907 // Allocation attempts to be fast in the case of frequent small objects. 908 // It breaks down like this: 909 // 910 // 1) Size in doublewords is computed. This is a constant for objects and 911 // variable for most arrays. Doubleword units are used to avoid size 912 // overflow of huge doubleword arrays. We need doublewords in the end for 913 // rounding. 914 // 915 // 2) Size is checked for being 'too large'. Too-large allocations will go 916 // the slow path into the VM. The slow path can throw any required 917 // exceptions, and does all the special checks for very large arrays. The 918 // size test can constant-fold away for objects. For objects with 919 // finalizers it constant-folds the otherway: you always go slow with 920 // finalizers. 921 // 922 // 3) If NOT using TLABs, this is the contended loop-back point. 923 // Load-Locked the heap top. If using TLABs normal-load the heap top. 924 // 925 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 926 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 927 // "size*8" we always enter the VM, where "largish" is a constant picked small 928 // enough that there's always space between the eden max and 4Gig (old space is 929 // there so it's quite large) and large enough that the cost of entering the VM 930 // is dwarfed by the cost to initialize the space. 931 // 932 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 933 // down. If contended, repeat at step 3. If using TLABs normal-store 934 // adjusted heap top back down; there is no contention. 935 // 936 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 937 // fields. 938 // 939 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 940 // oop flavor. 941 // 942 //============================================================================= 943 // FastAllocateSizeLimit value is in DOUBLEWORDS. 944 // Allocations bigger than this always go the slow route. 945 // This value must be small enough that allocation attempts that need to 946 // trigger exceptions go the slow route. Also, it must be small enough so 947 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 948 //=============================================================================j// 949 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 950 // The allocator will coalesce int->oop copies away. See comment in 951 // coalesce.cpp about how this works. It depends critically on the exact 952 // code shape produced here, so if you are changing this code shape 953 // make sure the GC info for the heap-top is correct in and around the 954 // slow-path call. 955 // 956 957 void PhaseMacroExpand::expand_allocate_common( 958 AllocateNode* alloc, // allocation node to be expanded 959 Node* length, // array length for an array allocation 960 const TypeFunc* slow_call_type, // Type of slow call 961 address slow_call_address // Address of slow call 962 ) 963 { 964 965 Node* ctrl = alloc->in(TypeFunc::Control); 966 Node* mem = alloc->in(TypeFunc::Memory); 967 Node* i_o = alloc->in(TypeFunc::I_O); 968 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 969 Node* klass_node = alloc->in(AllocateNode::KlassNode); 970 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 971 972 assert(ctrl != NULL, "must have control"); 973 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 974 // they will not be used if "always_slow" is set 975 enum { slow_result_path = 1, fast_result_path = 2 }; 976 Node *result_region; 977 Node *result_phi_rawmem; 978 Node *result_phi_rawoop; 979 Node *result_phi_i_o; 980 981 // The initial slow comparison is a size check, the comparison 982 // we want to do is a BoolTest::gt 983 bool always_slow = false; 984 int tv = _igvn.find_int_con(initial_slow_test, -1); 985 if (tv >= 0) { 986 always_slow = (tv == 1); 987 initial_slow_test = NULL; 988 } else { 989 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 990 } 991 992 if (DTraceAllocProbes || 993 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc() || 994 (UseConcMarkSweepGC && CMSIncrementalMode))) { 995 // Force slow-path allocation 996 always_slow = true; 997 initial_slow_test = NULL; 998 } 999 1000 1001 enum { too_big_or_final_path = 1, need_gc_path = 2 }; 1002 Node *slow_region = NULL; 1003 Node *toobig_false = ctrl; 1004 1005 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); 1006 // generate the initial test if necessary 1007 if (initial_slow_test != NULL ) { 1008 slow_region = new (C, 3) RegionNode(3); 1009 1010 // Now make the initial failure test. Usually a too-big test but 1011 // might be a TRUE for finalizers or a fancy class check for 1012 // newInstance0. 1013 IfNode *toobig_iff = new (C, 2) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1014 transform_later(toobig_iff); 1015 // Plug the failing-too-big test into the slow-path region 1016 Node *toobig_true = new (C, 1) IfTrueNode( toobig_iff ); 1017 transform_later(toobig_true); 1018 slow_region ->init_req( too_big_or_final_path, toobig_true ); 1019 toobig_false = new (C, 1) IfFalseNode( toobig_iff ); 1020 transform_later(toobig_false); 1021 } else { // No initial test, just fall into next case 1022 toobig_false = ctrl; 1023 debug_only(slow_region = NodeSentinel); 1024 } 1025 1026 Node *slow_mem = mem; // save the current memory state for slow path 1027 // generate the fast allocation code unless we know that the initial test will always go slow 1028 if (!always_slow) { 1029 // Fast path modifies only raw memory. 1030 if (mem->is_MergeMem()) { 1031 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1032 } 1033 1034 Node* eden_top_adr; 1035 Node* eden_end_adr; 1036 1037 set_eden_pointers(eden_top_adr, eden_end_adr); 1038 1039 // Load Eden::end. Loop invariant and hoisted. 1040 // 1041 // Note: We set the control input on "eden_end" and "old_eden_top" when using 1042 // a TLAB to work around a bug where these values were being moved across 1043 // a safepoint. These are not oops, so they cannot be include in the oop 1044 // map, but the can be changed by a GC. The proper way to fix this would 1045 // be to set the raw memory state when generating a SafepointNode. However 1046 // this will require extensive changes to the loop optimization in order to 1047 // prevent a degradation of the optimization. 1048 // See comment in memnode.hpp, around line 227 in class LoadPNode. 1049 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 1050 1051 // allocate the Region and Phi nodes for the result 1052 result_region = new (C, 3) RegionNode(3); 1053 result_phi_rawmem = new (C, 3) PhiNode( result_region, Type::MEMORY, TypeRawPtr::BOTTOM ); 1054 result_phi_rawoop = new (C, 3) PhiNode( result_region, TypeRawPtr::BOTTOM ); 1055 result_phi_i_o = new (C, 3) PhiNode( result_region, Type::ABIO ); // I/O is used for Prefetch 1056 1057 // We need a Region for the loop-back contended case. 1058 enum { fall_in_path = 1, contended_loopback_path = 2 }; 1059 Node *contended_region; 1060 Node *contended_phi_rawmem; 1061 if( UseTLAB ) { 1062 contended_region = toobig_false; 1063 contended_phi_rawmem = mem; 1064 } else { 1065 contended_region = new (C, 3) RegionNode(3); 1066 contended_phi_rawmem = new (C, 3) PhiNode( contended_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1067 // Now handle the passing-too-big test. We fall into the contended 1068 // loop-back merge point. 1069 contended_region ->init_req( fall_in_path, toobig_false ); 1070 contended_phi_rawmem->init_req( fall_in_path, mem ); 1071 transform_later(contended_region); 1072 transform_later(contended_phi_rawmem); 1073 } 1074 1075 // Load(-locked) the heap top. 1076 // See note above concerning the control input when using a TLAB 1077 Node *old_eden_top = UseTLAB 1078 ? new (C, 3) LoadPNode ( ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM ) 1079 : new (C, 3) LoadPLockedNode( contended_region, contended_phi_rawmem, eden_top_adr ); 1080 1081 transform_later(old_eden_top); 1082 // Add to heap top to get a new heap top 1083 Node *new_eden_top = new (C, 4) AddPNode( top(), old_eden_top, size_in_bytes ); 1084 transform_later(new_eden_top); 1085 // Check for needing a GC; compare against heap end 1086 Node *needgc_cmp = new (C, 3) CmpPNode( new_eden_top, eden_end ); 1087 transform_later(needgc_cmp); 1088 Node *needgc_bol = new (C, 2) BoolNode( needgc_cmp, BoolTest::ge ); 1089 transform_later(needgc_bol); 1090 IfNode *needgc_iff = new (C, 2) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1091 transform_later(needgc_iff); 1092 1093 // Plug the failing-heap-space-need-gc test into the slow-path region 1094 Node *needgc_true = new (C, 1) IfTrueNode( needgc_iff ); 1095 transform_later(needgc_true); 1096 if( initial_slow_test ) { 1097 slow_region ->init_req( need_gc_path, needgc_true ); 1098 // This completes all paths into the slow merge point 1099 transform_later(slow_region); 1100 } else { // No initial slow path needed! 1101 // Just fall from the need-GC path straight into the VM call. 1102 slow_region = needgc_true; 1103 } 1104 // No need for a GC. Setup for the Store-Conditional 1105 Node *needgc_false = new (C, 1) IfFalseNode( needgc_iff ); 1106 transform_later(needgc_false); 1107 1108 // Grab regular I/O before optional prefetch may change it. 1109 // Slow-path does no I/O so just set it to the original I/O. 1110 result_phi_i_o->init_req( slow_result_path, i_o ); 1111 1112 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem, 1113 old_eden_top, new_eden_top, length); 1114 1115 // Store (-conditional) the modified eden top back down. 1116 // StorePConditional produces flags for a test PLUS a modified raw 1117 // memory state. 1118 Node *store_eden_top; 1119 Node *fast_oop_ctrl; 1120 if( UseTLAB ) { 1121 store_eden_top = new (C, 4) StorePNode( needgc_false, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, new_eden_top ); 1122 transform_later(store_eden_top); 1123 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path 1124 } else { 1125 store_eden_top = new (C, 5) StorePConditionalNode( needgc_false, contended_phi_rawmem, eden_top_adr, new_eden_top, old_eden_top ); 1126 transform_later(store_eden_top); 1127 Node *contention_check = new (C, 2) BoolNode( store_eden_top, BoolTest::ne ); 1128 transform_later(contention_check); 1129 store_eden_top = new (C, 1) SCMemProjNode(store_eden_top); 1130 transform_later(store_eden_top); 1131 1132 // If not using TLABs, check to see if there was contention. 1133 IfNode *contention_iff = new (C, 2) IfNode ( needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN ); 1134 transform_later(contention_iff); 1135 Node *contention_true = new (C, 1) IfTrueNode( contention_iff ); 1136 transform_later(contention_true); 1137 // If contention, loopback and try again. 1138 contended_region->init_req( contended_loopback_path, contention_true ); 1139 contended_phi_rawmem->init_req( contended_loopback_path, store_eden_top ); 1140 1141 // Fast-path succeeded with no contention! 1142 Node *contention_false = new (C, 1) IfFalseNode( contention_iff ); 1143 transform_later(contention_false); 1144 fast_oop_ctrl = contention_false; 1145 } 1146 1147 // Rename successful fast-path variables to make meaning more obvious 1148 Node* fast_oop = old_eden_top; 1149 Node* fast_oop_rawmem = store_eden_top; 1150 fast_oop_rawmem = initialize_object(alloc, 1151 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1152 klass_node, length, size_in_bytes); 1153 1154 if (ExtendedDTraceProbes) { 1155 // Slow-path call 1156 int size = TypeFunc::Parms + 2; 1157 CallLeafNode *call = new (C, size) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1158 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), 1159 "dtrace_object_alloc", 1160 TypeRawPtr::BOTTOM); 1161 1162 // Get base of thread-local storage area 1163 Node* thread = new (C, 1) ThreadLocalNode(); 1164 transform_later(thread); 1165 1166 call->init_req(TypeFunc::Parms+0, thread); 1167 call->init_req(TypeFunc::Parms+1, fast_oop); 1168 call->init_req( TypeFunc::Control, fast_oop_ctrl ); 1169 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1170 call->init_req( TypeFunc::Memory , fast_oop_rawmem ); 1171 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1172 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1173 transform_later(call); 1174 fast_oop_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control); 1175 transform_later(fast_oop_ctrl); 1176 fast_oop_rawmem = new (C, 1) ProjNode(call,TypeFunc::Memory); 1177 transform_later(fast_oop_rawmem); 1178 } 1179 1180 // Plug in the successful fast-path into the result merge point 1181 result_region ->init_req( fast_result_path, fast_oop_ctrl ); 1182 result_phi_rawoop->init_req( fast_result_path, fast_oop ); 1183 result_phi_i_o ->init_req( fast_result_path, i_o ); 1184 result_phi_rawmem->init_req( fast_result_path, fast_oop_rawmem ); 1185 } else { 1186 slow_region = ctrl; 1187 } 1188 1189 // Generate slow-path call 1190 CallNode *call = new (C, slow_call_type->domain()->cnt()) 1191 CallStaticJavaNode(slow_call_type, slow_call_address, 1192 OptoRuntime::stub_name(slow_call_address), 1193 alloc->jvms()->bci(), 1194 TypePtr::BOTTOM); 1195 call->init_req( TypeFunc::Control, slow_region ); 1196 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1197 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs 1198 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1199 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1200 1201 call->init_req(TypeFunc::Parms+0, klass_node); 1202 if (length != NULL) { 1203 call->init_req(TypeFunc::Parms+1, length); 1204 } 1205 1206 // Copy debug information and adjust JVMState information, then replace 1207 // allocate node with the call 1208 copy_call_debug_info((CallNode *) alloc, call); 1209 if (!always_slow) { 1210 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1211 } 1212 _igvn.hash_delete(alloc); 1213 _igvn.subsume_node(alloc, call); 1214 transform_later(call); 1215 1216 // Identify the output projections from the allocate node and 1217 // adjust any references to them. 1218 // The control and io projections look like: 1219 // 1220 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1221 // Allocate Catch 1222 // ^---Proj(io) <-------+ ^---CatchProj(io) 1223 // 1224 // We are interested in the CatchProj nodes. 1225 // 1226 extract_call_projections(call); 1227 1228 // An allocate node has separate memory projections for the uses on the control and i_o paths 1229 // Replace uses of the control memory projection with result_phi_rawmem (unless we are only generating a slow call) 1230 if (!always_slow && _memproj_fallthrough != NULL) { 1231 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { 1232 Node *use = _memproj_fallthrough->fast_out(i); 1233 _igvn.hash_delete(use); 1234 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); 1235 _igvn._worklist.push(use); 1236 // back up iterator 1237 --i; 1238 } 1239 } 1240 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete _memproj_catchall so 1241 // we end up with a call that has only 1 memory projection 1242 if (_memproj_catchall != NULL ) { 1243 if (_memproj_fallthrough == NULL) { 1244 _memproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::Memory); 1245 transform_later(_memproj_fallthrough); 1246 } 1247 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { 1248 Node *use = _memproj_catchall->fast_out(i); 1249 _igvn.hash_delete(use); 1250 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); 1251 _igvn._worklist.push(use); 1252 // back up iterator 1253 --i; 1254 } 1255 } 1256 1257 // An allocate node has separate i_o projections for the uses on the control and i_o paths 1258 // Replace uses of the control i_o projection with result_phi_i_o (unless we are only generating a slow call) 1259 if (_ioproj_fallthrough == NULL) { 1260 _ioproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::I_O); 1261 transform_later(_ioproj_fallthrough); 1262 } else if (!always_slow) { 1263 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { 1264 Node *use = _ioproj_fallthrough->fast_out(i); 1265 1266 _igvn.hash_delete(use); 1267 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); 1268 _igvn._worklist.push(use); 1269 // back up iterator 1270 --i; 1271 } 1272 } 1273 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete _ioproj_catchall so 1274 // we end up with a call that has only 1 control projection 1275 if (_ioproj_catchall != NULL ) { 1276 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { 1277 Node *use = _ioproj_catchall->fast_out(i); 1278 _igvn.hash_delete(use); 1279 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); 1280 _igvn._worklist.push(use); 1281 // back up iterator 1282 --i; 1283 } 1284 } 1285 1286 // if we generated only a slow call, we are done 1287 if (always_slow) 1288 return; 1289 1290 1291 if (_fallthroughcatchproj != NULL) { 1292 ctrl = _fallthroughcatchproj->clone(); 1293 transform_later(ctrl); 1294 _igvn.hash_delete(_fallthroughcatchproj); 1295 _igvn.subsume_node(_fallthroughcatchproj, result_region); 1296 } else { 1297 ctrl = top(); 1298 } 1299 Node *slow_result; 1300 if (_resproj == NULL) { 1301 // no uses of the allocation result 1302 slow_result = top(); 1303 } else { 1304 slow_result = _resproj->clone(); 1305 transform_later(slow_result); 1306 _igvn.hash_delete(_resproj); 1307 _igvn.subsume_node(_resproj, result_phi_rawoop); 1308 } 1309 1310 // Plug slow-path into result merge point 1311 result_region ->init_req( slow_result_path, ctrl ); 1312 result_phi_rawoop->init_req( slow_result_path, slow_result); 1313 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); 1314 transform_later(result_region); 1315 transform_later(result_phi_rawoop); 1316 transform_later(result_phi_rawmem); 1317 transform_later(result_phi_i_o); 1318 // This completes all paths into the result merge point 1319 } 1320 1321 1322 // Helper for PhaseMacroExpand::expand_allocate_common. 1323 // Initializes the newly-allocated storage. 1324 Node* 1325 PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1326 Node* control, Node* rawmem, Node* object, 1327 Node* klass_node, Node* length, 1328 Node* size_in_bytes) { 1329 InitializeNode* init = alloc->initialization(); 1330 // Store the klass & mark bits 1331 Node* mark_node = NULL; 1332 // For now only enable fast locking for non-array types 1333 if (UseBiasedLocking && (length == NULL)) { 1334 mark_node = make_load(NULL, rawmem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeRawPtr::BOTTOM, T_ADDRESS); 1335 } else { 1336 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); 1337 } 1338 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 1339 1340 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_OBJECT); 1341 int header_size = alloc->minimum_header_size(); // conservatively small 1342 1343 // Array length 1344 if (length != NULL) { // Arrays need length field 1345 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1346 // conservatively small header size: 1347 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1348 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1349 if (k->is_array_klass()) // we know the exact header size in most cases: 1350 header_size = Klass::layout_helper_header_size(k->layout_helper()); 1351 } 1352 1353 // Clear the object body, if necessary. 1354 if (init == NULL) { 1355 // The init has somehow disappeared; be cautious and clear everything. 1356 // 1357 // This can happen if a node is allocated but an uncommon trap occurs 1358 // immediately. In this case, the Initialize gets associated with the 1359 // trap, and may be placed in a different (outer) loop, if the Allocate 1360 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1361 // there can be two Allocates to one Initialize. The answer in all these 1362 // edge cases is safety first. It is always safe to clear immediately 1363 // within an Allocate, and then (maybe or maybe not) clear some more later. 1364 if (!ZeroTLAB) 1365 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1366 header_size, size_in_bytes, 1367 &_igvn); 1368 } else { 1369 if (!init->is_complete()) { 1370 // Try to win by zeroing only what the init does not store. 1371 // We can also try to do some peephole optimizations, 1372 // such as combining some adjacent subword stores. 1373 rawmem = init->complete_stores(control, rawmem, object, 1374 header_size, size_in_bytes, &_igvn); 1375 } 1376 // We have no more use for this link, since the AllocateNode goes away: 1377 init->set_req(InitializeNode::RawAddress, top()); 1378 // (If we keep the link, it just confuses the register allocator, 1379 // who thinks he sees a real use of the address by the membar.) 1380 } 1381 1382 return rawmem; 1383 } 1384 1385 // Generate prefetch instructions for next allocations. 1386 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1387 Node*& contended_phi_rawmem, 1388 Node* old_eden_top, Node* new_eden_top, 1389 Node* length) { 1390 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1391 // Generate prefetch allocation with watermark check. 1392 // As an allocation hits the watermark, we will prefetch starting 1393 // at a "distance" away from watermark. 1394 enum { fall_in_path = 1, pf_path = 2 }; 1395 1396 Node *pf_region = new (C, 3) RegionNode(3); 1397 Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY, 1398 TypeRawPtr::BOTTOM ); 1399 // I/O is used for Prefetch 1400 Node *pf_phi_abio = new (C, 3) PhiNode( pf_region, Type::ABIO ); 1401 1402 Node *thread = new (C, 1) ThreadLocalNode(); 1403 transform_later(thread); 1404 1405 Node *eden_pf_adr = new (C, 4) AddPNode( top()/*not oop*/, thread, 1406 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1407 transform_later(eden_pf_adr); 1408 1409 Node *old_pf_wm = new (C, 3) LoadPNode( needgc_false, 1410 contended_phi_rawmem, eden_pf_adr, 1411 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM ); 1412 transform_later(old_pf_wm); 1413 1414 // check against new_eden_top 1415 Node *need_pf_cmp = new (C, 3) CmpPNode( new_eden_top, old_pf_wm ); 1416 transform_later(need_pf_cmp); 1417 Node *need_pf_bol = new (C, 2) BoolNode( need_pf_cmp, BoolTest::ge ); 1418 transform_later(need_pf_bol); 1419 IfNode *need_pf_iff = new (C, 2) IfNode( needgc_false, need_pf_bol, 1420 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1421 transform_later(need_pf_iff); 1422 1423 // true node, add prefetchdistance 1424 Node *need_pf_true = new (C, 1) IfTrueNode( need_pf_iff ); 1425 transform_later(need_pf_true); 1426 1427 Node *need_pf_false = new (C, 1) IfFalseNode( need_pf_iff ); 1428 transform_later(need_pf_false); 1429 1430 Node *new_pf_wmt = new (C, 4) AddPNode( top(), old_pf_wm, 1431 _igvn.MakeConX(AllocatePrefetchDistance) ); 1432 transform_later(new_pf_wmt ); 1433 new_pf_wmt->set_req(0, need_pf_true); 1434 1435 Node *store_new_wmt = new (C, 4) StorePNode( need_pf_true, 1436 contended_phi_rawmem, eden_pf_adr, 1437 TypeRawPtr::BOTTOM, new_pf_wmt ); 1438 transform_later(store_new_wmt); 1439 1440 // adding prefetches 1441 pf_phi_abio->init_req( fall_in_path, i_o ); 1442 1443 Node *prefetch_adr; 1444 Node *prefetch; 1445 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize; 1446 uint step_size = AllocatePrefetchStepSize; 1447 uint distance = 0; 1448 1449 for ( uint i = 0; i < lines; i++ ) { 1450 prefetch_adr = new (C, 4) AddPNode( old_pf_wm, new_pf_wmt, 1451 _igvn.MakeConX(distance) ); 1452 transform_later(prefetch_adr); 1453 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr ); 1454 transform_later(prefetch); 1455 distance += step_size; 1456 i_o = prefetch; 1457 } 1458 pf_phi_abio->set_req( pf_path, i_o ); 1459 1460 pf_region->init_req( fall_in_path, need_pf_false ); 1461 pf_region->init_req( pf_path, need_pf_true ); 1462 1463 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 1464 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 1465 1466 transform_later(pf_region); 1467 transform_later(pf_phi_rawmem); 1468 transform_later(pf_phi_abio); 1469 1470 needgc_false = pf_region; 1471 contended_phi_rawmem = pf_phi_rawmem; 1472 i_o = pf_phi_abio; 1473 } else if( AllocatePrefetchStyle > 0 ) { 1474 // Insert a prefetch for each allocation only on the fast-path 1475 Node *prefetch_adr; 1476 Node *prefetch; 1477 // Generate several prefetch instructions only for arrays. 1478 uint lines = (length != NULL) ? AllocatePrefetchLines : 1; 1479 uint step_size = AllocatePrefetchStepSize; 1480 uint distance = AllocatePrefetchDistance; 1481 for ( uint i = 0; i < lines; i++ ) { 1482 prefetch_adr = new (C, 4) AddPNode( old_eden_top, new_eden_top, 1483 _igvn.MakeConX(distance) ); 1484 transform_later(prefetch_adr); 1485 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr ); 1486 // Do not let it float too high, since if eden_top == eden_end, 1487 // both might be null. 1488 if( i == 0 ) { // Set control for first prefetch, next follows it 1489 prefetch->init_req(0, needgc_false); 1490 } 1491 transform_later(prefetch); 1492 distance += step_size; 1493 i_o = prefetch; 1494 } 1495 } 1496 return i_o; 1497 } 1498 1499 1500 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 1501 expand_allocate_common(alloc, NULL, 1502 OptoRuntime::new_instance_Type(), 1503 OptoRuntime::new_instance_Java()); 1504 } 1505 1506 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 1507 Node* length = alloc->in(AllocateNode::ALength); 1508 expand_allocate_common(alloc, length, 1509 OptoRuntime::new_array_Type(), 1510 OptoRuntime::new_array_Java()); 1511 } 1512 1513 1514 // we have determined that this lock/unlock can be eliminated, we simply 1515 // eliminate the node without expanding it. 1516 // 1517 // Note: The membar's associated with the lock/unlock are currently not 1518 // eliminated. This should be investigated as a future enhancement. 1519 // 1520 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 1521 1522 if (!alock->is_eliminated()) { 1523 return false; 1524 } 1525 if (alock->is_Lock() && !alock->is_coarsened()) { 1526 // Create new "eliminated" BoxLock node and use it 1527 // in monitor debug info for the same object. 1528 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 1529 Node* obj = alock->obj_node(); 1530 if (!oldbox->is_eliminated()) { 1531 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 1532 newbox->set_eliminated(); 1533 transform_later(newbox); 1534 // Replace old box node with new box for all users 1535 // of the same object. 1536 for (uint i = 0; i < oldbox->outcnt();) { 1537 1538 bool next_edge = true; 1539 Node* u = oldbox->raw_out(i); 1540 if (u == alock) { 1541 i++; 1542 continue; // It will be removed below 1543 } 1544 if (u->is_Lock() && 1545 u->as_Lock()->obj_node() == obj && 1546 // oldbox could be referenced in debug info also 1547 u->as_Lock()->box_node() == oldbox) { 1548 assert(u->as_Lock()->is_eliminated(), "sanity"); 1549 _igvn.hash_delete(u); 1550 u->set_req(TypeFunc::Parms + 1, newbox); 1551 next_edge = false; 1552 #ifdef ASSERT 1553 } else if (u->is_Unlock() && u->as_Unlock()->obj_node() == obj) { 1554 assert(u->as_Unlock()->is_eliminated(), "sanity"); 1555 #endif 1556 } 1557 // Replace old box in monitor debug info. 1558 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 1559 SafePointNode* sfn = u->as_SafePoint(); 1560 JVMState* youngest_jvms = sfn->jvms(); 1561 int max_depth = youngest_jvms->depth(); 1562 for (int depth = 1; depth <= max_depth; depth++) { 1563 JVMState* jvms = youngest_jvms->of_depth(depth); 1564 int num_mon = jvms->nof_monitors(); 1565 // Loop over monitors 1566 for (int idx = 0; idx < num_mon; idx++) { 1567 Node* obj_node = sfn->monitor_obj(jvms, idx); 1568 Node* box_node = sfn->monitor_box(jvms, idx); 1569 if (box_node == oldbox && obj_node == obj) { 1570 int j = jvms->monitor_box_offset(idx); 1571 _igvn.hash_delete(u); 1572 u->set_req(j, newbox); 1573 next_edge = false; 1574 } 1575 } // for (int idx = 0; 1576 } // for (int depth = 1; 1577 } // if (u->is_SafePoint() 1578 if (next_edge) i++; 1579 } // for (uint i = 0; i < oldbox->outcnt();) 1580 } // if (!oldbox->is_eliminated()) 1581 } // if (alock->is_Lock() && !lock->is_coarsened()) 1582 1583 #ifndef PRODUCT 1584 if (PrintEliminateLocks) { 1585 if (alock->is_Lock()) { 1586 tty->print_cr("++++ Eliminating: %d Lock", alock->_idx); 1587 } else { 1588 tty->print_cr("++++ Eliminating: %d Unlock", alock->_idx); 1589 } 1590 } 1591 #endif 1592 1593 Node* mem = alock->in(TypeFunc::Memory); 1594 Node* ctrl = alock->in(TypeFunc::Control); 1595 1596 extract_call_projections(alock); 1597 // There are 2 projections from the lock. The lock node will 1598 // be deleted when its last use is subsumed below. 1599 assert(alock->outcnt() == 2 && 1600 _fallthroughproj != NULL && 1601 _memproj_fallthrough != NULL, 1602 "Unexpected projections from Lock/Unlock"); 1603 1604 Node* fallthroughproj = _fallthroughproj; 1605 Node* memproj_fallthrough = _memproj_fallthrough; 1606 1607 // The memory projection from a lock/unlock is RawMem 1608 // The input to a Lock is merged memory, so extract its RawMem input 1609 // (unless the MergeMem has been optimized away.) 1610 if (alock->is_Lock()) { 1611 // Seach for MemBarAcquire node and delete it also. 1612 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 1613 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquire, ""); 1614 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 1615 Node* memproj = membar->proj_out(TypeFunc::Memory); 1616 _igvn.hash_delete(ctrlproj); 1617 _igvn.subsume_node(ctrlproj, fallthroughproj); 1618 _igvn.hash_delete(memproj); 1619 _igvn.subsume_node(memproj, memproj_fallthrough); 1620 1621 // Delete FastLock node also if this Lock node is unique user 1622 // (a loop peeling may clone a Lock node). 1623 Node* flock = alock->as_Lock()->fastlock_node(); 1624 if (flock->outcnt() == 1) { 1625 assert(flock->unique_out() == alock, "sanity"); 1626 _igvn.hash_delete(flock); 1627 _igvn.subsume_node(flock, top()); 1628 } 1629 } 1630 1631 // Seach for MemBarRelease node and delete it also. 1632 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() && 1633 ctrl->in(0)->is_MemBar()) { 1634 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 1635 assert(membar->Opcode() == Op_MemBarRelease && 1636 mem->is_Proj() && membar == mem->in(0), ""); 1637 _igvn.hash_delete(fallthroughproj); 1638 _igvn.subsume_node(fallthroughproj, ctrl); 1639 _igvn.hash_delete(memproj_fallthrough); 1640 _igvn.subsume_node(memproj_fallthrough, mem); 1641 fallthroughproj = ctrl; 1642 memproj_fallthrough = mem; 1643 ctrl = membar->in(TypeFunc::Control); 1644 mem = membar->in(TypeFunc::Memory); 1645 } 1646 1647 _igvn.hash_delete(fallthroughproj); 1648 _igvn.subsume_node(fallthroughproj, ctrl); 1649 _igvn.hash_delete(memproj_fallthrough); 1650 _igvn.subsume_node(memproj_fallthrough, mem); 1651 return true; 1652 } 1653 1654 1655 //------------------------------expand_lock_node---------------------- 1656 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 1657 1658 Node* ctrl = lock->in(TypeFunc::Control); 1659 Node* mem = lock->in(TypeFunc::Memory); 1660 Node* obj = lock->obj_node(); 1661 Node* box = lock->box_node(); 1662 Node* flock = lock->fastlock_node(); 1663 1664 // Make the merge point 1665 Node *region; 1666 Node *mem_phi; 1667 Node *slow_path; 1668 1669 if (UseOptoBiasInlining) { 1670 /* 1671 * See the full description in MacroAssembler::biased_locking_enter(). 1672 * 1673 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { 1674 * // The object is biased. 1675 * proto_node = klass->prototype_header; 1676 * o_node = thread | proto_node; 1677 * x_node = o_node ^ mark_word; 1678 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? 1679 * // Done. 1680 * } else { 1681 * if( (x_node & biased_lock_mask) != 0 ) { 1682 * // The klass's prototype header is no longer biased. 1683 * cas(&mark_word, mark_word, proto_node) 1684 * goto cas_lock; 1685 * } else { 1686 * // The klass's prototype header is still biased. 1687 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? 1688 * old = mark_word; 1689 * new = o_node; 1690 * } else { 1691 * // Different thread or anonymous biased. 1692 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); 1693 * new = thread | old; 1694 * } 1695 * // Try to rebias. 1696 * if( cas(&mark_word, old, new) == 0 ) { 1697 * // Done. 1698 * } else { 1699 * goto slow_path; // Failed. 1700 * } 1701 * } 1702 * } 1703 * } else { 1704 * // The object is not biased. 1705 * cas_lock: 1706 * if( FastLock(obj) == 0 ) { 1707 * // Done. 1708 * } else { 1709 * slow_path: 1710 * OptoRuntime::complete_monitor_locking_Java(obj); 1711 * } 1712 * } 1713 */ 1714 1715 region = new (C, 5) RegionNode(5); 1716 // create a Phi for the memory state 1717 mem_phi = new (C, 5) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 1718 1719 Node* fast_lock_region = new (C, 3) RegionNode(3); 1720 Node* fast_lock_mem_phi = new (C, 3) PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1721 1722 // First, check mark word for the biased lock pattern. 1723 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 1724 1725 // Get fast path - mark word has the biased lock pattern. 1726 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, 1727 markOopDesc::biased_lock_mask_in_place, 1728 markOopDesc::biased_lock_pattern, true); 1729 // fast_lock_region->in(1) is set to slow path. 1730 fast_lock_mem_phi->init_req(1, mem); 1731 1732 // Now check that the lock is biased to the current thread and has 1733 // the same epoch and bias as Klass::_prototype_header. 1734 1735 // Special-case a fresh allocation to avoid building nodes: 1736 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); 1737 if (klass_node == NULL) { 1738 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1739 klass_node = transform_later( LoadKlassNode::make(_igvn, mem, k_adr, _igvn.type(k_adr)->is_ptr()) ); 1740 #ifdef _LP64 1741 if (UseCompressedOops && klass_node->is_DecodeN()) { 1742 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity"); 1743 klass_node->in(1)->init_req(0, ctrl); 1744 } else 1745 #endif 1746 klass_node->init_req(0, ctrl); 1747 } 1748 Node *proto_node = make_load(ctrl, mem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeX_X, TypeX_X->basic_type()); 1749 1750 Node* thread = transform_later(new (C, 1) ThreadLocalNode()); 1751 Node* cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread)); 1752 Node* o_node = transform_later(new (C, 3) OrXNode(cast_thread, proto_node)); 1753 Node* x_node = transform_later(new (C, 3) XorXNode(o_node, mark_node)); 1754 1755 // Get slow path - mark word does NOT match the value. 1756 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, 1757 (~markOopDesc::age_mask_in_place), 0); 1758 // region->in(3) is set to fast path - the object is biased to the current thread. 1759 mem_phi->init_req(3, mem); 1760 1761 1762 // Mark word does NOT match the value (thread | Klass::_prototype_header). 1763 1764 1765 // First, check biased pattern. 1766 // Get fast path - _prototype_header has the same biased lock pattern. 1767 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, 1768 markOopDesc::biased_lock_mask_in_place, 0, true); 1769 1770 not_biased_ctrl = fast_lock_region->in(2); // Slow path 1771 // fast_lock_region->in(2) - the prototype header is no longer biased 1772 // and we have to revoke the bias on this object. 1773 // We are going to try to reset the mark of this object to the prototype 1774 // value and fall through to the CAS-based locking scheme. 1775 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 1776 Node* cas = new (C, 5) StoreXConditionalNode(not_biased_ctrl, mem, adr, 1777 proto_node, mark_node); 1778 transform_later(cas); 1779 Node* proj = transform_later( new (C, 1) SCMemProjNode(cas)); 1780 fast_lock_mem_phi->init_req(2, proj); 1781 1782 1783 // Second, check epoch bits. 1784 Node* rebiased_region = new (C, 3) RegionNode(3); 1785 Node* old_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X); 1786 Node* new_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X); 1787 1788 // Get slow path - mark word does NOT match epoch bits. 1789 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, 1790 markOopDesc::epoch_mask_in_place, 0); 1791 // The epoch of the current bias is not valid, attempt to rebias the object 1792 // toward the current thread. 1793 rebiased_region->init_req(2, epoch_ctrl); 1794 old_phi->init_req(2, mark_node); 1795 new_phi->init_req(2, o_node); 1796 1797 // rebiased_region->in(1) is set to fast path. 1798 // The epoch of the current bias is still valid but we know 1799 // nothing about the owner; it might be set or it might be clear. 1800 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place | 1801 markOopDesc::age_mask_in_place | 1802 markOopDesc::epoch_mask_in_place); 1803 Node* old = transform_later(new (C, 3) AndXNode(mark_node, cmask)); 1804 cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread)); 1805 Node* new_mark = transform_later(new (C, 3) OrXNode(cast_thread, old)); 1806 old_phi->init_req(1, old); 1807 new_phi->init_req(1, new_mark); 1808 1809 transform_later(rebiased_region); 1810 transform_later(old_phi); 1811 transform_later(new_phi); 1812 1813 // Try to acquire the bias of the object using an atomic operation. 1814 // If this fails we will go in to the runtime to revoke the object's bias. 1815 cas = new (C, 5) StoreXConditionalNode(rebiased_region, mem, adr, 1816 new_phi, old_phi); 1817 transform_later(cas); 1818 proj = transform_later( new (C, 1) SCMemProjNode(cas)); 1819 1820 // Get slow path - Failed to CAS. 1821 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); 1822 mem_phi->init_req(4, proj); 1823 // region->in(4) is set to fast path - the object is rebiased to the current thread. 1824 1825 // Failed to CAS. 1826 slow_path = new (C, 3) RegionNode(3); 1827 Node *slow_mem = new (C, 3) PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); 1828 1829 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control 1830 slow_mem->init_req(1, proj); 1831 1832 // Call CAS-based locking scheme (FastLock node). 1833 1834 transform_later(fast_lock_region); 1835 transform_later(fast_lock_mem_phi); 1836 1837 // Get slow path - FastLock failed to lock the object. 1838 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); 1839 mem_phi->init_req(2, fast_lock_mem_phi); 1840 // region->in(2) is set to fast path - the object is locked to the current thread. 1841 1842 slow_path->init_req(2, ctrl); // Capture slow-control 1843 slow_mem->init_req(2, fast_lock_mem_phi); 1844 1845 transform_later(slow_path); 1846 transform_later(slow_mem); 1847 // Reset lock's memory edge. 1848 lock->set_req(TypeFunc::Memory, slow_mem); 1849 1850 } else { 1851 region = new (C, 3) RegionNode(3); 1852 // create a Phi for the memory state 1853 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 1854 1855 // Optimize test; set region slot 2 1856 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); 1857 mem_phi->init_req(2, mem); 1858 } 1859 1860 // Make slow path call 1861 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box ); 1862 1863 extract_call_projections(call); 1864 1865 // Slow path can only throw asynchronous exceptions, which are always 1866 // de-opted. So the compiler thinks the slow-call can never throw an 1867 // exception. If it DOES throw an exception we would need the debug 1868 // info removed first (since if it throws there is no monitor). 1869 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 1870 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 1871 1872 // Capture slow path 1873 // disconnect fall-through projection from call and create a new one 1874 // hook up users of fall-through projection to region 1875 Node *slow_ctrl = _fallthroughproj->clone(); 1876 transform_later(slow_ctrl); 1877 _igvn.hash_delete(_fallthroughproj); 1878 _fallthroughproj->disconnect_inputs(NULL); 1879 region->init_req(1, slow_ctrl); 1880 // region inputs are now complete 1881 transform_later(region); 1882 _igvn.subsume_node(_fallthroughproj, region); 1883 1884 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) ); 1885 mem_phi->init_req(1, memproj ); 1886 transform_later(mem_phi); 1887 _igvn.hash_delete(_memproj_fallthrough); 1888 _igvn.subsume_node(_memproj_fallthrough, mem_phi); 1889 } 1890 1891 //------------------------------expand_unlock_node---------------------- 1892 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 1893 1894 Node* ctrl = unlock->in(TypeFunc::Control); 1895 Node* mem = unlock->in(TypeFunc::Memory); 1896 Node* obj = unlock->obj_node(); 1897 Node* box = unlock->box_node(); 1898 1899 // No need for a null check on unlock 1900 1901 // Make the merge point 1902 Node *region; 1903 Node *mem_phi; 1904 1905 if (UseOptoBiasInlining) { 1906 // Check for biased locking unlock case, which is a no-op. 1907 // See the full description in MacroAssembler::biased_locking_exit(). 1908 region = new (C, 4) RegionNode(4); 1909 // create a Phi for the memory state 1910 mem_phi = new (C, 4) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 1911 mem_phi->init_req(3, mem); 1912 1913 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 1914 ctrl = opt_bits_test(ctrl, region, 3, mark_node, 1915 markOopDesc::biased_lock_mask_in_place, 1916 markOopDesc::biased_lock_pattern); 1917 } else { 1918 region = new (C, 3) RegionNode(3); 1919 // create a Phi for the memory state 1920 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 1921 } 1922 1923 FastUnlockNode *funlock = new (C, 3) FastUnlockNode( ctrl, obj, box ); 1924 funlock = transform_later( funlock )->as_FastUnlock(); 1925 // Optimize test; set region slot 2 1926 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); 1927 1928 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 ); 1929 1930 extract_call_projections(call); 1931 1932 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 1933 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 1934 1935 // No exceptions for unlocking 1936 // Capture slow path 1937 // disconnect fall-through projection from call and create a new one 1938 // hook up users of fall-through projection to region 1939 Node *slow_ctrl = _fallthroughproj->clone(); 1940 transform_later(slow_ctrl); 1941 _igvn.hash_delete(_fallthroughproj); 1942 _fallthroughproj->disconnect_inputs(NULL); 1943 region->init_req(1, slow_ctrl); 1944 // region inputs are now complete 1945 transform_later(region); 1946 _igvn.subsume_node(_fallthroughproj, region); 1947 1948 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) ); 1949 mem_phi->init_req(1, memproj ); 1950 mem_phi->init_req(2, mem); 1951 transform_later(mem_phi); 1952 _igvn.hash_delete(_memproj_fallthrough); 1953 _igvn.subsume_node(_memproj_fallthrough, mem_phi); 1954 } 1955 1956 //------------------------------expand_macro_nodes---------------------- 1957 // Returns true if a failure occurred. 1958 bool PhaseMacroExpand::expand_macro_nodes() { 1959 if (C->macro_count() == 0) 1960 return false; 1961 // First, attempt to eliminate locks 1962 bool progress = true; 1963 while (progress) { 1964 progress = false; 1965 for (int i = C->macro_count(); i > 0; i--) { 1966 Node * n = C->macro_node(i-1); 1967 bool success = false; 1968 debug_only(int old_macro_count = C->macro_count();); 1969 if (n->is_AbstractLock()) { 1970 success = eliminate_locking_node(n->as_AbstractLock()); 1971 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { 1972 _igvn.add_users_to_worklist(n); 1973 _igvn.hash_delete(n); 1974 _igvn.subsume_node(n, n->in(1)); 1975 success = true; 1976 } 1977 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 1978 progress = progress || success; 1979 } 1980 } 1981 // Next, attempt to eliminate allocations 1982 progress = true; 1983 while (progress) { 1984 progress = false; 1985 for (int i = C->macro_count(); i > 0; i--) { 1986 Node * n = C->macro_node(i-1); 1987 bool success = false; 1988 debug_only(int old_macro_count = C->macro_count();); 1989 switch (n->class_id()) { 1990 case Node::Class_Allocate: 1991 case Node::Class_AllocateArray: 1992 success = eliminate_allocate_node(n->as_Allocate()); 1993 break; 1994 case Node::Class_Lock: 1995 case Node::Class_Unlock: 1996 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 1997 break; 1998 default: 1999 assert(false, "unknown node type in macro list"); 2000 } 2001 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2002 progress = progress || success; 2003 } 2004 } 2005 // Make sure expansion will not cause node limit to be exceeded. 2006 // Worst case is a macro node gets expanded into about 50 nodes. 2007 // Allow 50% more for optimization. 2008 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) ) 2009 return true; 2010 2011 // expand "macro" nodes 2012 // nodes are removed from the macro list as they are processed 2013 while (C->macro_count() > 0) { 2014 int macro_count = C->macro_count(); 2015 Node * n = C->macro_node(macro_count-1); 2016 assert(n->is_macro(), "only macro nodes expected here"); 2017 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2018 // node is unreachable, so don't try to expand it 2019 C->remove_macro_node(n); 2020 continue; 2021 } 2022 switch (n->class_id()) { 2023 case Node::Class_Allocate: 2024 expand_allocate(n->as_Allocate()); 2025 break; 2026 case Node::Class_AllocateArray: 2027 expand_allocate_array(n->as_AllocateArray()); 2028 break; 2029 case Node::Class_Lock: 2030 expand_lock_node(n->as_Lock()); 2031 break; 2032 case Node::Class_Unlock: 2033 expand_unlock_node(n->as_Unlock()); 2034 break; 2035 default: 2036 assert(false, "unknown node type in macro list"); 2037 } 2038 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2039 if (C->failing()) return true; 2040 } 2041 2042 _igvn.set_delay_transform(false); 2043 _igvn.optimize(); 2044 return false; 2045 }