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, newcall->in(0)); // 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 = NULL; 395 uint size = value_phis->size(); 396 for (uint i=0; i < size; i++) { 397 if ( mem->_idx == value_phis->index_at(i) ) { 398 return value_phis->node_at(i); 399 } 400 } 401 402 if (level <= 0) { 403 return NULL; // Give up: phi tree too deep 404 } 405 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 406 Node *alloc_mem = alloc->in(TypeFunc::Memory); 407 408 uint length = mem->req(); 409 GrowableArray <Node *> values(length, length, NULL); 410 411 // create a new Phi for the value 412 PhiNode *phi = new (C, length) PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset); 413 transform_later(phi); 414 value_phis->push(phi, mem->_idx); 415 416 for (uint j = 1; j < length; j++) { 417 Node *in = mem->in(j); 418 if (in == NULL || in->is_top()) { 419 values.at_put(j, in); 420 } else { 421 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); 422 if (val == start_mem || val == alloc_mem) { 423 // hit a sentinel, return appropriate 0 value 424 values.at_put(j, _igvn.zerocon(ft)); 425 continue; 426 } 427 if (val->is_Initialize()) { 428 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 429 } 430 if (val == NULL) { 431 return NULL; // can't find a value on this path 432 } 433 if (val == mem) { 434 values.at_put(j, mem); 435 } else if (val->is_Store()) { 436 values.at_put(j, val->in(MemNode::ValueIn)); 437 } else if(val->is_Proj() && val->in(0) == alloc) { 438 values.at_put(j, _igvn.zerocon(ft)); 439 } else if (val->is_Phi()) { 440 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); 441 if (val == NULL) { 442 return NULL; 443 } 444 values.at_put(j, val); 445 } else if (val->Opcode() == Op_SCMemProj) { 446 assert(val->in(0)->is_LoadStore(), "sanity"); 447 assert(false, "Object is not scalar replaceable if a LoadStore node access its field"); 448 return NULL; 449 } else { 450 #ifdef ASSERT 451 val->dump(); 452 assert(false, "unknown node on this path"); 453 #endif 454 return NULL; // unknown node on this path 455 } 456 } 457 } 458 // Set Phi's inputs 459 for (uint j = 1; j < length; j++) { 460 if (values.at(j) == mem) { 461 phi->init_req(j, phi); 462 } else { 463 phi->init_req(j, values.at(j)); 464 } 465 } 466 return phi; 467 } 468 469 // Search the last value stored into the object's field. 470 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) { 471 assert(adr_t->is_known_instance_field(), "instance required"); 472 int instance_id = adr_t->instance_id(); 473 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 474 475 int alias_idx = C->get_alias_index(adr_t); 476 int offset = adr_t->offset(); 477 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 478 Node *alloc_ctrl = alloc->in(TypeFunc::Control); 479 Node *alloc_mem = alloc->in(TypeFunc::Memory); 480 Arena *a = Thread::current()->resource_area(); 481 VectorSet visited(a); 482 483 484 bool done = sfpt_mem == alloc_mem; 485 Node *mem = sfpt_mem; 486 while (!done) { 487 if (visited.test_set(mem->_idx)) { 488 return NULL; // found a loop, give up 489 } 490 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 491 if (mem == start_mem || mem == alloc_mem) { 492 done = true; // hit a sentinel, return appropriate 0 value 493 } else if (mem->is_Initialize()) { 494 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 495 if (mem == NULL) { 496 done = true; // Something go wrong. 497 } else if (mem->is_Store()) { 498 const TypePtr* atype = mem->as_Store()->adr_type(); 499 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 500 done = true; 501 } 502 } else if (mem->is_Store()) { 503 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 504 assert(atype != NULL, "address type must be oopptr"); 505 assert(C->get_alias_index(atype) == alias_idx && 506 atype->is_known_instance_field() && atype->offset() == offset && 507 atype->instance_id() == instance_id, "store is correct memory slice"); 508 done = true; 509 } else if (mem->is_Phi()) { 510 // try to find a phi's unique input 511 Node *unique_input = NULL; 512 Node *top = C->top(); 513 for (uint i = 1; i < mem->req(); i++) { 514 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 515 if (n == NULL || n == top || n == mem) { 516 continue; 517 } else if (unique_input == NULL) { 518 unique_input = n; 519 } else if (unique_input != n) { 520 unique_input = top; 521 break; 522 } 523 } 524 if (unique_input != NULL && unique_input != top) { 525 mem = unique_input; 526 } else { 527 done = true; 528 } 529 } else { 530 assert(false, "unexpected node"); 531 } 532 } 533 if (mem != NULL) { 534 if (mem == start_mem || mem == alloc_mem) { 535 // hit a sentinel, return appropriate 0 value 536 return _igvn.zerocon(ft); 537 } else if (mem->is_Store()) { 538 return mem->in(MemNode::ValueIn); 539 } else if (mem->is_Phi()) { 540 // attempt to produce a Phi reflecting the values on the input paths of the Phi 541 Node_Stack value_phis(a, 8); 542 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 543 if (phi != NULL) { 544 return phi; 545 } else { 546 // Kill all new Phis 547 while(value_phis.is_nonempty()) { 548 Node* n = value_phis.node(); 549 _igvn.replace_node(n, C->top()); 550 value_phis.pop(); 551 } 552 } 553 } 554 } 555 // Something go wrong. 556 return NULL; 557 } 558 559 // Check the possibility of scalar replacement. 560 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 561 // Scan the uses of the allocation to check for anything that would 562 // prevent us from eliminating it. 563 NOT_PRODUCT( const char* fail_eliminate = NULL; ) 564 DEBUG_ONLY( Node* disq_node = NULL; ) 565 bool can_eliminate = true; 566 567 Node* res = alloc->result_cast(); 568 const TypeOopPtr* res_type = NULL; 569 if (res == NULL) { 570 // All users were eliminated. 571 } else if (!res->is_CheckCastPP()) { 572 alloc->_is_scalar_replaceable = false; // don't try again 573 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 574 can_eliminate = false; 575 } else { 576 res_type = _igvn.type(res)->isa_oopptr(); 577 if (res_type == NULL) { 578 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 579 can_eliminate = false; 580 } else if (res_type->isa_aryptr()) { 581 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 582 if (length < 0) { 583 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 584 can_eliminate = false; 585 } 586 } 587 } 588 589 if (can_eliminate && res != NULL) { 590 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 591 j < jmax && can_eliminate; j++) { 592 Node* use = res->fast_out(j); 593 594 if (use->is_AddP()) { 595 const TypePtr* addp_type = _igvn.type(use)->is_ptr(); 596 int offset = addp_type->offset(); 597 598 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 599 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) 600 can_eliminate = false; 601 break; 602 } 603 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 604 k < kmax && can_eliminate; k++) { 605 Node* n = use->fast_out(k); 606 if (!n->is_Store() && n->Opcode() != Op_CastP2X) { 607 DEBUG_ONLY(disq_node = n;) 608 if (n->is_Load() || n->is_LoadStore()) { 609 NOT_PRODUCT(fail_eliminate = "Field load";) 610 } else { 611 NOT_PRODUCT(fail_eliminate = "Not store field referrence";) 612 } 613 can_eliminate = false; 614 } 615 } 616 } else if (use->is_SafePoint()) { 617 SafePointNode* sfpt = use->as_SafePoint(); 618 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 619 // Object is passed as argument. 620 DEBUG_ONLY(disq_node = use;) 621 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 622 can_eliminate = false; 623 } 624 Node* sfptMem = sfpt->memory(); 625 if (sfptMem == NULL || sfptMem->is_top()) { 626 DEBUG_ONLY(disq_node = use;) 627 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) 628 can_eliminate = false; 629 } else { 630 safepoints.append_if_missing(sfpt); 631 } 632 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 633 if (use->is_Phi()) { 634 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 635 NOT_PRODUCT(fail_eliminate = "Object is return value";) 636 } else { 637 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 638 } 639 DEBUG_ONLY(disq_node = use;) 640 } else { 641 if (use->Opcode() == Op_Return) { 642 NOT_PRODUCT(fail_eliminate = "Object is return value";) 643 }else { 644 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 645 } 646 DEBUG_ONLY(disq_node = use;) 647 } 648 can_eliminate = false; 649 } 650 } 651 } 652 653 #ifndef PRODUCT 654 if (PrintEliminateAllocations) { 655 if (can_eliminate) { 656 tty->print("Scalar "); 657 if (res == NULL) 658 alloc->dump(); 659 else 660 res->dump(); 661 } else { 662 tty->print("NotScalar (%s)", fail_eliminate); 663 if (res == NULL) 664 alloc->dump(); 665 else 666 res->dump(); 667 #ifdef ASSERT 668 if (disq_node != NULL) { 669 tty->print(" >>>> "); 670 disq_node->dump(); 671 } 672 #endif /*ASSERT*/ 673 } 674 } 675 #endif 676 return can_eliminate; 677 } 678 679 // Do scalar replacement. 680 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 681 GrowableArray <SafePointNode *> safepoints_done; 682 683 ciKlass* klass = NULL; 684 ciInstanceKlass* iklass = NULL; 685 int nfields = 0; 686 int array_base; 687 int element_size; 688 BasicType basic_elem_type; 689 ciType* elem_type; 690 691 Node* res = alloc->result_cast(); 692 const TypeOopPtr* res_type = NULL; 693 if (res != NULL) { // Could be NULL when there are no users 694 res_type = _igvn.type(res)->isa_oopptr(); 695 } 696 697 if (res != NULL) { 698 klass = res_type->klass(); 699 if (res_type->isa_instptr()) { 700 // find the fields of the class which will be needed for safepoint debug information 701 assert(klass->is_instance_klass(), "must be an instance klass."); 702 iklass = klass->as_instance_klass(); 703 nfields = iklass->nof_nonstatic_fields(); 704 } else { 705 // find the array's elements which will be needed for safepoint debug information 706 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 707 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); 708 elem_type = klass->as_array_klass()->element_type(); 709 basic_elem_type = elem_type->basic_type(); 710 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 711 element_size = type2aelembytes(basic_elem_type); 712 } 713 } 714 // 715 // Process the safepoint uses 716 // 717 while (safepoints.length() > 0) { 718 SafePointNode* sfpt = safepoints.pop(); 719 Node* mem = sfpt->memory(); 720 uint first_ind = sfpt->req(); 721 SafePointScalarObjectNode* sobj = new (C, 1) SafePointScalarObjectNode(res_type, 722 #ifdef ASSERT 723 alloc, 724 #endif 725 first_ind, nfields); 726 sobj->init_req(0, sfpt->in(TypeFunc::Control)); 727 transform_later(sobj); 728 729 // Scan object's fields adding an input to the safepoint for each field. 730 for (int j = 0; j < nfields; j++) { 731 intptr_t offset; 732 ciField* field = NULL; 733 if (iklass != NULL) { 734 field = iklass->nonstatic_field_at(j); 735 offset = field->offset(); 736 elem_type = field->type(); 737 basic_elem_type = field->layout_type(); 738 } else { 739 offset = array_base + j * (intptr_t)element_size; 740 } 741 742 const Type *field_type; 743 // The next code is taken from Parse::do_get_xxx(). 744 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { 745 if (!elem_type->is_loaded()) { 746 field_type = TypeInstPtr::BOTTOM; 747 } else if (field != NULL && field->is_constant() && field->is_static()) { 748 // This can happen if the constant oop is non-perm. 749 ciObject* con = field->constant_value().as_object(); 750 // Do not "join" in the previous type; it doesn't add value, 751 // and may yield a vacuous result if the field is of interface type. 752 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 753 assert(field_type != NULL, "field singleton type must be consistent"); 754 } else { 755 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 756 } 757 if (UseCompressedOops) { 758 field_type = field_type->make_narrowoop(); 759 basic_elem_type = T_NARROWOOP; 760 } 761 } else { 762 field_type = Type::get_const_basic_type(basic_elem_type); 763 } 764 765 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 766 767 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc); 768 if (field_val == NULL) { 769 // we weren't able to find a value for this field, 770 // give up on eliminating this allocation 771 alloc->_is_scalar_replaceable = false; // don't try again 772 // remove any extra entries we added to the safepoint 773 uint last = sfpt->req() - 1; 774 for (int k = 0; k < j; k++) { 775 sfpt->del_req(last--); 776 } 777 // rollback processed safepoints 778 while (safepoints_done.length() > 0) { 779 SafePointNode* sfpt_done = safepoints_done.pop(); 780 // remove any extra entries we added to the safepoint 781 last = sfpt_done->req() - 1; 782 for (int k = 0; k < nfields; k++) { 783 sfpt_done->del_req(last--); 784 } 785 JVMState *jvms = sfpt_done->jvms(); 786 jvms->set_endoff(sfpt_done->req()); 787 // Now make a pass over the debug information replacing any references 788 // to SafePointScalarObjectNode with the allocated object. 789 int start = jvms->debug_start(); 790 int end = jvms->debug_end(); 791 for (int i = start; i < end; i++) { 792 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 793 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 794 if (scobj->first_index() == sfpt_done->req() && 795 scobj->n_fields() == (uint)nfields) { 796 assert(scobj->alloc() == alloc, "sanity"); 797 sfpt_done->set_req(i, res); 798 } 799 } 800 } 801 } 802 #ifndef PRODUCT 803 if (PrintEliminateAllocations) { 804 if (field != NULL) { 805 tty->print("=== At SafePoint node %d can't find value of Field: ", 806 sfpt->_idx); 807 field->print(); 808 int field_idx = C->get_alias_index(field_addr_type); 809 tty->print(" (alias_idx=%d)", field_idx); 810 } else { // Array's element 811 tty->print("=== At SafePoint node %d can't find value of array element [%d]", 812 sfpt->_idx, j); 813 } 814 tty->print(", which prevents elimination of: "); 815 if (res == NULL) 816 alloc->dump(); 817 else 818 res->dump(); 819 } 820 #endif 821 return false; 822 } 823 if (UseCompressedOops && field_type->isa_narrowoop()) { 824 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 825 // to be able scalar replace the allocation. 826 if (field_val->is_EncodeP()) { 827 field_val = field_val->in(1); 828 } else { 829 field_val = transform_later(new (C, 2) DecodeNNode(field_val, field_val->bottom_type()->make_ptr())); 830 } 831 } 832 sfpt->add_req(field_val); 833 } 834 JVMState *jvms = sfpt->jvms(); 835 jvms->set_endoff(sfpt->req()); 836 // Now make a pass over the debug information replacing any references 837 // to the allocated object with "sobj" 838 int start = jvms->debug_start(); 839 int end = jvms->debug_end(); 840 for (int i = start; i < end; i++) { 841 if (sfpt->in(i) == res) { 842 sfpt->set_req(i, sobj); 843 } 844 } 845 safepoints_done.append_if_missing(sfpt); // keep it for rollback 846 } 847 return true; 848 } 849 850 // Process users of eliminated allocation. 851 void PhaseMacroExpand::process_users_of_allocation(AllocateNode *alloc) { 852 Node* res = alloc->result_cast(); 853 if (res != NULL) { 854 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 855 Node *use = res->last_out(j); 856 uint oc1 = res->outcnt(); 857 858 if (use->is_AddP()) { 859 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 860 Node *n = use->last_out(k); 861 uint oc2 = use->outcnt(); 862 if (n->is_Store()) { 863 #ifdef ASSERT 864 // Verify that there is no dependent MemBarVolatile nodes, 865 // they should be removed during IGVN, see MemBarNode::Ideal(). 866 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); 867 p < pmax; p++) { 868 Node* mb = n->fast_out(p); 869 assert(mb->is_Initialize() || !mb->is_MemBar() || 870 mb->req() <= MemBarNode::Precedent || 871 mb->in(MemBarNode::Precedent) != n, 872 "MemBarVolatile should be eliminated for non-escaping object"); 873 } 874 #endif 875 _igvn.replace_node(n, n->in(MemNode::Memory)); 876 } else { 877 eliminate_card_mark(n); 878 } 879 k -= (oc2 - use->outcnt()); 880 } 881 } else { 882 eliminate_card_mark(use); 883 } 884 j -= (oc1 - res->outcnt()); 885 } 886 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 887 _igvn.remove_dead_node(res); 888 } 889 890 // 891 // Process other users of allocation's projections 892 // 893 if (_resproj != NULL && _resproj->outcnt() != 0) { 894 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { 895 Node *use = _resproj->last_out(j); 896 uint oc1 = _resproj->outcnt(); 897 if (use->is_Initialize()) { 898 // Eliminate Initialize node. 899 InitializeNode *init = use->as_Initialize(); 900 assert(init->outcnt() <= 2, "only a control and memory projection expected"); 901 Node *ctrl_proj = init->proj_out(TypeFunc::Control); 902 if (ctrl_proj != NULL) { 903 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection"); 904 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj); 905 } 906 Node *mem_proj = init->proj_out(TypeFunc::Memory); 907 if (mem_proj != NULL) { 908 Node *mem = init->in(TypeFunc::Memory); 909 #ifdef ASSERT 910 if (mem->is_MergeMem()) { 911 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection"); 912 } else { 913 assert(mem == _memproj_fallthrough, "allocation memory projection"); 914 } 915 #endif 916 _igvn.replace_node(mem_proj, mem); 917 } 918 } else if (use->is_AddP()) { 919 // raw memory addresses used only by the initialization 920 _igvn.replace_node(use, C->top()); 921 } else { 922 assert(false, "only Initialize or AddP expected"); 923 } 924 j -= (oc1 - _resproj->outcnt()); 925 } 926 } 927 if (_fallthroughcatchproj != NULL) { 928 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); 929 } 930 if (_memproj_fallthrough != NULL) { 931 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); 932 } 933 if (_memproj_catchall != NULL) { 934 _igvn.replace_node(_memproj_catchall, C->top()); 935 } 936 if (_ioproj_fallthrough != NULL) { 937 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); 938 } 939 if (_ioproj_catchall != NULL) { 940 _igvn.replace_node(_ioproj_catchall, C->top()); 941 } 942 if (_catchallcatchproj != NULL) { 943 _igvn.replace_node(_catchallcatchproj, C->top()); 944 } 945 } 946 947 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 948 949 if (!EliminateAllocations || !alloc->_is_scalar_replaceable) { 950 return false; 951 } 952 953 extract_call_projections(alloc); 954 955 GrowableArray <SafePointNode *> safepoints; 956 if (!can_eliminate_allocation(alloc, safepoints)) { 957 return false; 958 } 959 960 if (!scalar_replacement(alloc, safepoints)) { 961 return false; 962 } 963 964 CompileLog* log = C->log(); 965 if (log != NULL) { 966 Node* klass = alloc->in(AllocateNode::KlassNode); 967 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); 968 log->head("eliminate_allocation type='%d'", 969 log->identify(tklass->klass())); 970 JVMState* p = alloc->jvms(); 971 while (p != NULL) { 972 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 973 p = p->caller(); 974 } 975 log->tail("eliminate_allocation"); 976 } 977 978 process_users_of_allocation(alloc); 979 980 #ifndef PRODUCT 981 if (PrintEliminateAllocations) { 982 if (alloc->is_AllocateArray()) 983 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 984 else 985 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 986 } 987 #endif 988 989 return true; 990 } 991 992 993 //---------------------------set_eden_pointers------------------------- 994 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { 995 if (UseTLAB) { // Private allocation: load from TLS 996 Node* thread = transform_later(new (C, 1) ThreadLocalNode()); 997 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); 998 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); 999 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); 1000 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); 1001 } else { // Shared allocation: load from globals 1002 CollectedHeap* ch = Universe::heap(); 1003 address top_adr = (address)ch->top_addr(); 1004 address end_adr = (address)ch->end_addr(); 1005 eden_top_adr = makecon(TypeRawPtr::make(top_adr)); 1006 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); 1007 } 1008 } 1009 1010 1011 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 1012 Node* adr = basic_plus_adr(base, offset); 1013 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 1014 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt); 1015 transform_later(value); 1016 return value; 1017 } 1018 1019 1020 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 1021 Node* adr = basic_plus_adr(base, offset); 1022 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt); 1023 transform_later(mem); 1024 return mem; 1025 } 1026 1027 //============================================================================= 1028 // 1029 // A L L O C A T I O N 1030 // 1031 // Allocation attempts to be fast in the case of frequent small objects. 1032 // It breaks down like this: 1033 // 1034 // 1) Size in doublewords is computed. This is a constant for objects and 1035 // variable for most arrays. Doubleword units are used to avoid size 1036 // overflow of huge doubleword arrays. We need doublewords in the end for 1037 // rounding. 1038 // 1039 // 2) Size is checked for being 'too large'. Too-large allocations will go 1040 // the slow path into the VM. The slow path can throw any required 1041 // exceptions, and does all the special checks for very large arrays. The 1042 // size test can constant-fold away for objects. For objects with 1043 // finalizers it constant-folds the otherway: you always go slow with 1044 // finalizers. 1045 // 1046 // 3) If NOT using TLABs, this is the contended loop-back point. 1047 // Load-Locked the heap top. If using TLABs normal-load the heap top. 1048 // 1049 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 1050 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 1051 // "size*8" we always enter the VM, where "largish" is a constant picked small 1052 // enough that there's always space between the eden max and 4Gig (old space is 1053 // there so it's quite large) and large enough that the cost of entering the VM 1054 // is dwarfed by the cost to initialize the space. 1055 // 1056 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 1057 // down. If contended, repeat at step 3. If using TLABs normal-store 1058 // adjusted heap top back down; there is no contention. 1059 // 1060 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 1061 // fields. 1062 // 1063 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 1064 // oop flavor. 1065 // 1066 //============================================================================= 1067 // FastAllocateSizeLimit value is in DOUBLEWORDS. 1068 // Allocations bigger than this always go the slow route. 1069 // This value must be small enough that allocation attempts that need to 1070 // trigger exceptions go the slow route. Also, it must be small enough so 1071 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 1072 //=============================================================================j// 1073 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 1074 // The allocator will coalesce int->oop copies away. See comment in 1075 // coalesce.cpp about how this works. It depends critically on the exact 1076 // code shape produced here, so if you are changing this code shape 1077 // make sure the GC info for the heap-top is correct in and around the 1078 // slow-path call. 1079 // 1080 1081 void PhaseMacroExpand::expand_allocate_common( 1082 AllocateNode* alloc, // allocation node to be expanded 1083 Node* length, // array length for an array allocation 1084 const TypeFunc* slow_call_type, // Type of slow call 1085 address slow_call_address // Address of slow call 1086 ) 1087 { 1088 1089 Node* ctrl = alloc->in(TypeFunc::Control); 1090 Node* mem = alloc->in(TypeFunc::Memory); 1091 Node* i_o = alloc->in(TypeFunc::I_O); 1092 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 1093 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1094 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 1095 1096 assert(ctrl != NULL, "must have control"); 1097 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 1098 // they will not be used if "always_slow" is set 1099 enum { slow_result_path = 1, fast_result_path = 2 }; 1100 Node *result_region; 1101 Node *result_phi_rawmem; 1102 Node *result_phi_rawoop; 1103 Node *result_phi_i_o; 1104 1105 // The initial slow comparison is a size check, the comparison 1106 // we want to do is a BoolTest::gt 1107 bool always_slow = false; 1108 int tv = _igvn.find_int_con(initial_slow_test, -1); 1109 if (tv >= 0) { 1110 always_slow = (tv == 1); 1111 initial_slow_test = NULL; 1112 } else { 1113 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 1114 } 1115 1116 if (C->env()->dtrace_alloc_probes() || 1117 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc() || 1118 (UseConcMarkSweepGC && CMSIncrementalMode))) { 1119 // Force slow-path allocation 1120 always_slow = true; 1121 initial_slow_test = NULL; 1122 } 1123 1124 1125 enum { too_big_or_final_path = 1, need_gc_path = 2 }; 1126 Node *slow_region = NULL; 1127 Node *toobig_false = ctrl; 1128 1129 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); 1130 // generate the initial test if necessary 1131 if (initial_slow_test != NULL ) { 1132 slow_region = new (C, 3) RegionNode(3); 1133 1134 // Now make the initial failure test. Usually a too-big test but 1135 // might be a TRUE for finalizers or a fancy class check for 1136 // newInstance0. 1137 IfNode *toobig_iff = new (C, 2) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1138 transform_later(toobig_iff); 1139 // Plug the failing-too-big test into the slow-path region 1140 Node *toobig_true = new (C, 1) IfTrueNode( toobig_iff ); 1141 transform_later(toobig_true); 1142 slow_region ->init_req( too_big_or_final_path, toobig_true ); 1143 toobig_false = new (C, 1) IfFalseNode( toobig_iff ); 1144 transform_later(toobig_false); 1145 } else { // No initial test, just fall into next case 1146 toobig_false = ctrl; 1147 debug_only(slow_region = NodeSentinel); 1148 } 1149 1150 Node *slow_mem = mem; // save the current memory state for slow path 1151 // generate the fast allocation code unless we know that the initial test will always go slow 1152 if (!always_slow) { 1153 // Fast path modifies only raw memory. 1154 if (mem->is_MergeMem()) { 1155 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1156 } 1157 1158 Node* eden_top_adr; 1159 Node* eden_end_adr; 1160 1161 set_eden_pointers(eden_top_adr, eden_end_adr); 1162 1163 // Load Eden::end. Loop invariant and hoisted. 1164 // 1165 // Note: We set the control input on "eden_end" and "old_eden_top" when using 1166 // a TLAB to work around a bug where these values were being moved across 1167 // a safepoint. These are not oops, so they cannot be include in the oop 1168 // map, but they can be changed by a GC. The proper way to fix this would 1169 // be to set the raw memory state when generating a SafepointNode. However 1170 // this will require extensive changes to the loop optimization in order to 1171 // prevent a degradation of the optimization. 1172 // See comment in memnode.hpp, around line 227 in class LoadPNode. 1173 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 1174 1175 // allocate the Region and Phi nodes for the result 1176 result_region = new (C, 3) RegionNode(3); 1177 result_phi_rawmem = new (C, 3) PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1178 result_phi_rawoop = new (C, 3) PhiNode(result_region, TypeRawPtr::BOTTOM); 1179 result_phi_i_o = new (C, 3) PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch 1180 1181 // We need a Region for the loop-back contended case. 1182 enum { fall_in_path = 1, contended_loopback_path = 2 }; 1183 Node *contended_region; 1184 Node *contended_phi_rawmem; 1185 if (UseTLAB) { 1186 contended_region = toobig_false; 1187 contended_phi_rawmem = mem; 1188 } else { 1189 contended_region = new (C, 3) RegionNode(3); 1190 contended_phi_rawmem = new (C, 3) PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1191 // Now handle the passing-too-big test. We fall into the contended 1192 // loop-back merge point. 1193 contended_region ->init_req(fall_in_path, toobig_false); 1194 contended_phi_rawmem->init_req(fall_in_path, mem); 1195 transform_later(contended_region); 1196 transform_later(contended_phi_rawmem); 1197 } 1198 1199 // Load(-locked) the heap top. 1200 // See note above concerning the control input when using a TLAB 1201 Node *old_eden_top = UseTLAB 1202 ? new (C, 3) LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM) 1203 : new (C, 3) LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr); 1204 1205 transform_later(old_eden_top); 1206 // Add to heap top to get a new heap top 1207 Node *new_eden_top = new (C, 4) AddPNode(top(), old_eden_top, size_in_bytes); 1208 transform_later(new_eden_top); 1209 // Check for needing a GC; compare against heap end 1210 Node *needgc_cmp = new (C, 3) CmpPNode(new_eden_top, eden_end); 1211 transform_later(needgc_cmp); 1212 Node *needgc_bol = new (C, 2) BoolNode(needgc_cmp, BoolTest::ge); 1213 transform_later(needgc_bol); 1214 IfNode *needgc_iff = new (C, 2) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); 1215 transform_later(needgc_iff); 1216 1217 // Plug the failing-heap-space-need-gc test into the slow-path region 1218 Node *needgc_true = new (C, 1) IfTrueNode(needgc_iff); 1219 transform_later(needgc_true); 1220 if (initial_slow_test) { 1221 slow_region->init_req(need_gc_path, needgc_true); 1222 // This completes all paths into the slow merge point 1223 transform_later(slow_region); 1224 } else { // No initial slow path needed! 1225 // Just fall from the need-GC path straight into the VM call. 1226 slow_region = needgc_true; 1227 } 1228 // No need for a GC. Setup for the Store-Conditional 1229 Node *needgc_false = new (C, 1) IfFalseNode(needgc_iff); 1230 transform_later(needgc_false); 1231 1232 // Grab regular I/O before optional prefetch may change it. 1233 // Slow-path does no I/O so just set it to the original I/O. 1234 result_phi_i_o->init_req(slow_result_path, i_o); 1235 1236 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem, 1237 old_eden_top, new_eden_top, length); 1238 1239 // Name successful fast-path variables 1240 Node* fast_oop = old_eden_top; 1241 Node* fast_oop_ctrl; 1242 Node* fast_oop_rawmem; 1243 1244 // Store (-conditional) the modified eden top back down. 1245 // StorePConditional produces flags for a test PLUS a modified raw 1246 // memory state. 1247 if (UseTLAB) { 1248 Node* store_eden_top = 1249 new (C, 4) StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1250 TypeRawPtr::BOTTOM, new_eden_top); 1251 transform_later(store_eden_top); 1252 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path 1253 fast_oop_rawmem = store_eden_top; 1254 } else { 1255 Node* store_eden_top = 1256 new (C, 5) StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1257 new_eden_top, fast_oop/*old_eden_top*/); 1258 transform_later(store_eden_top); 1259 Node *contention_check = new (C, 2) BoolNode(store_eden_top, BoolTest::ne); 1260 transform_later(contention_check); 1261 store_eden_top = new (C, 1) SCMemProjNode(store_eden_top); 1262 transform_later(store_eden_top); 1263 1264 // If not using TLABs, check to see if there was contention. 1265 IfNode *contention_iff = new (C, 2) IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN); 1266 transform_later(contention_iff); 1267 Node *contention_true = new (C, 1) IfTrueNode(contention_iff); 1268 transform_later(contention_true); 1269 // If contention, loopback and try again. 1270 contended_region->init_req(contended_loopback_path, contention_true); 1271 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top); 1272 1273 // Fast-path succeeded with no contention! 1274 Node *contention_false = new (C, 1) IfFalseNode(contention_iff); 1275 transform_later(contention_false); 1276 fast_oop_ctrl = contention_false; 1277 1278 // Bump total allocated bytes for this thread 1279 Node* thread = new (C, 1) ThreadLocalNode(); 1280 transform_later(thread); 1281 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread, 1282 in_bytes(JavaThread::allocated_bytes_offset())); 1283 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1284 0, TypeLong::LONG, T_LONG); 1285 #ifdef _LP64 1286 Node* alloc_size = size_in_bytes; 1287 #else 1288 Node* alloc_size = new (C, 2) ConvI2LNode(size_in_bytes); 1289 transform_later(alloc_size); 1290 #endif 1291 Node* new_alloc_bytes = new (C, 3) AddLNode(alloc_bytes, alloc_size); 1292 transform_later(new_alloc_bytes); 1293 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1294 0, new_alloc_bytes, T_LONG); 1295 } 1296 1297 fast_oop_rawmem = initialize_object(alloc, 1298 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1299 klass_node, length, size_in_bytes); 1300 1301 if (C->env()->dtrace_extended_probes()) { 1302 // Slow-path call 1303 int size = TypeFunc::Parms + 2; 1304 CallLeafNode *call = new (C, size) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1305 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), 1306 "dtrace_object_alloc", 1307 TypeRawPtr::BOTTOM); 1308 1309 // Get base of thread-local storage area 1310 Node* thread = new (C, 1) ThreadLocalNode(); 1311 transform_later(thread); 1312 1313 call->init_req(TypeFunc::Parms+0, thread); 1314 call->init_req(TypeFunc::Parms+1, fast_oop); 1315 call->init_req(TypeFunc::Control, fast_oop_ctrl); 1316 call->init_req(TypeFunc::I_O , top()); // does no i/o 1317 call->init_req(TypeFunc::Memory , fast_oop_rawmem); 1318 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1319 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1320 transform_later(call); 1321 fast_oop_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control); 1322 transform_later(fast_oop_ctrl); 1323 fast_oop_rawmem = new (C, 1) ProjNode(call,TypeFunc::Memory); 1324 transform_later(fast_oop_rawmem); 1325 } 1326 1327 // Plug in the successful fast-path into the result merge point 1328 result_region ->init_req(fast_result_path, fast_oop_ctrl); 1329 result_phi_rawoop->init_req(fast_result_path, fast_oop); 1330 result_phi_i_o ->init_req(fast_result_path, i_o); 1331 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); 1332 } else { 1333 slow_region = ctrl; 1334 } 1335 1336 // Generate slow-path call 1337 CallNode *call = new (C, slow_call_type->domain()->cnt()) 1338 CallStaticJavaNode(slow_call_type, slow_call_address, 1339 OptoRuntime::stub_name(slow_call_address), 1340 alloc->jvms()->bci(), 1341 TypePtr::BOTTOM); 1342 call->init_req( TypeFunc::Control, slow_region ); 1343 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1344 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs 1345 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1346 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1347 1348 call->init_req(TypeFunc::Parms+0, klass_node); 1349 if (length != NULL) { 1350 call->init_req(TypeFunc::Parms+1, length); 1351 } 1352 1353 // Copy debug information and adjust JVMState information, then replace 1354 // allocate node with the call 1355 copy_call_debug_info((CallNode *) alloc, call); 1356 if (!always_slow) { 1357 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1358 } 1359 _igvn.replace_node(alloc, call); 1360 transform_later(call); 1361 1362 // Identify the output projections from the allocate node and 1363 // adjust any references to them. 1364 // The control and io projections look like: 1365 // 1366 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1367 // Allocate Catch 1368 // ^---Proj(io) <-------+ ^---CatchProj(io) 1369 // 1370 // We are interested in the CatchProj nodes. 1371 // 1372 extract_call_projections(call); 1373 1374 // An allocate node has separate memory projections for the uses on the control and i_o paths 1375 // Replace uses of the control memory projection with result_phi_rawmem (unless we are only generating a slow call) 1376 if (!always_slow && _memproj_fallthrough != NULL) { 1377 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { 1378 Node *use = _memproj_fallthrough->fast_out(i); 1379 _igvn.hash_delete(use); 1380 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); 1381 _igvn._worklist.push(use); 1382 // back up iterator 1383 --i; 1384 } 1385 } 1386 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete _memproj_catchall so 1387 // we end up with a call that has only 1 memory projection 1388 if (_memproj_catchall != NULL ) { 1389 if (_memproj_fallthrough == NULL) { 1390 _memproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::Memory); 1391 transform_later(_memproj_fallthrough); 1392 } 1393 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { 1394 Node *use = _memproj_catchall->fast_out(i); 1395 _igvn.hash_delete(use); 1396 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); 1397 _igvn._worklist.push(use); 1398 // back up iterator 1399 --i; 1400 } 1401 } 1402 1403 // An allocate node has separate i_o projections for the uses on the control and i_o paths 1404 // Replace uses of the control i_o projection with result_phi_i_o (unless we are only generating a slow call) 1405 if (_ioproj_fallthrough == NULL) { 1406 _ioproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::I_O); 1407 transform_later(_ioproj_fallthrough); 1408 } else if (!always_slow) { 1409 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { 1410 Node *use = _ioproj_fallthrough->fast_out(i); 1411 1412 _igvn.hash_delete(use); 1413 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); 1414 _igvn._worklist.push(use); 1415 // back up iterator 1416 --i; 1417 } 1418 } 1419 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete _ioproj_catchall so 1420 // we end up with a call that has only 1 control projection 1421 if (_ioproj_catchall != NULL ) { 1422 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { 1423 Node *use = _ioproj_catchall->fast_out(i); 1424 _igvn.hash_delete(use); 1425 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); 1426 _igvn._worklist.push(use); 1427 // back up iterator 1428 --i; 1429 } 1430 } 1431 1432 // if we generated only a slow call, we are done 1433 if (always_slow) 1434 return; 1435 1436 1437 if (_fallthroughcatchproj != NULL) { 1438 ctrl = _fallthroughcatchproj->clone(); 1439 transform_later(ctrl); 1440 _igvn.replace_node(_fallthroughcatchproj, result_region); 1441 } else { 1442 ctrl = top(); 1443 } 1444 Node *slow_result; 1445 if (_resproj == NULL) { 1446 // no uses of the allocation result 1447 slow_result = top(); 1448 } else { 1449 slow_result = _resproj->clone(); 1450 transform_later(slow_result); 1451 _igvn.replace_node(_resproj, result_phi_rawoop); 1452 } 1453 1454 // Plug slow-path into result merge point 1455 result_region ->init_req( slow_result_path, ctrl ); 1456 result_phi_rawoop->init_req( slow_result_path, slow_result); 1457 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); 1458 transform_later(result_region); 1459 transform_later(result_phi_rawoop); 1460 transform_later(result_phi_rawmem); 1461 transform_later(result_phi_i_o); 1462 // This completes all paths into the result merge point 1463 } 1464 1465 1466 // Helper for PhaseMacroExpand::expand_allocate_common. 1467 // Initializes the newly-allocated storage. 1468 Node* 1469 PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1470 Node* control, Node* rawmem, Node* object, 1471 Node* klass_node, Node* length, 1472 Node* size_in_bytes) { 1473 InitializeNode* init = alloc->initialization(); 1474 // Store the klass & mark bits 1475 Node* mark_node = NULL; 1476 // For now only enable fast locking for non-array types 1477 if (UseBiasedLocking && (length == NULL)) { 1478 mark_node = make_load(control, rawmem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeRawPtr::BOTTOM, T_ADDRESS); 1479 } else { 1480 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); 1481 } 1482 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 1483 1484 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_OBJECT); 1485 int header_size = alloc->minimum_header_size(); // conservatively small 1486 1487 // Array length 1488 if (length != NULL) { // Arrays need length field 1489 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1490 // conservatively small header size: 1491 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1492 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1493 if (k->is_array_klass()) // we know the exact header size in most cases: 1494 header_size = Klass::layout_helper_header_size(k->layout_helper()); 1495 } 1496 1497 // Clear the object body, if necessary. 1498 if (init == NULL) { 1499 // The init has somehow disappeared; be cautious and clear everything. 1500 // 1501 // This can happen if a node is allocated but an uncommon trap occurs 1502 // immediately. In this case, the Initialize gets associated with the 1503 // trap, and may be placed in a different (outer) loop, if the Allocate 1504 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1505 // there can be two Allocates to one Initialize. The answer in all these 1506 // edge cases is safety first. It is always safe to clear immediately 1507 // within an Allocate, and then (maybe or maybe not) clear some more later. 1508 if (!ZeroTLAB) 1509 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1510 header_size, size_in_bytes, 1511 &_igvn); 1512 } else { 1513 if (!init->is_complete()) { 1514 // Try to win by zeroing only what the init does not store. 1515 // We can also try to do some peephole optimizations, 1516 // such as combining some adjacent subword stores. 1517 rawmem = init->complete_stores(control, rawmem, object, 1518 header_size, size_in_bytes, &_igvn); 1519 } 1520 // We have no more use for this link, since the AllocateNode goes away: 1521 init->set_req(InitializeNode::RawAddress, top()); 1522 // (If we keep the link, it just confuses the register allocator, 1523 // who thinks he sees a real use of the address by the membar.) 1524 } 1525 1526 return rawmem; 1527 } 1528 1529 // Generate prefetch instructions for next allocations. 1530 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1531 Node*& contended_phi_rawmem, 1532 Node* old_eden_top, Node* new_eden_top, 1533 Node* length) { 1534 enum { fall_in_path = 1, pf_path = 2 }; 1535 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1536 // Generate prefetch allocation with watermark check. 1537 // As an allocation hits the watermark, we will prefetch starting 1538 // at a "distance" away from watermark. 1539 1540 Node *pf_region = new (C, 3) RegionNode(3); 1541 Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY, 1542 TypeRawPtr::BOTTOM ); 1543 // I/O is used for Prefetch 1544 Node *pf_phi_abio = new (C, 3) PhiNode( pf_region, Type::ABIO ); 1545 1546 Node *thread = new (C, 1) ThreadLocalNode(); 1547 transform_later(thread); 1548 1549 Node *eden_pf_adr = new (C, 4) AddPNode( top()/*not oop*/, thread, 1550 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1551 transform_later(eden_pf_adr); 1552 1553 Node *old_pf_wm = new (C, 3) LoadPNode( needgc_false, 1554 contended_phi_rawmem, eden_pf_adr, 1555 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM ); 1556 transform_later(old_pf_wm); 1557 1558 // check against new_eden_top 1559 Node *need_pf_cmp = new (C, 3) CmpPNode( new_eden_top, old_pf_wm ); 1560 transform_later(need_pf_cmp); 1561 Node *need_pf_bol = new (C, 2) BoolNode( need_pf_cmp, BoolTest::ge ); 1562 transform_later(need_pf_bol); 1563 IfNode *need_pf_iff = new (C, 2) IfNode( needgc_false, need_pf_bol, 1564 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1565 transform_later(need_pf_iff); 1566 1567 // true node, add prefetchdistance 1568 Node *need_pf_true = new (C, 1) IfTrueNode( need_pf_iff ); 1569 transform_later(need_pf_true); 1570 1571 Node *need_pf_false = new (C, 1) IfFalseNode( need_pf_iff ); 1572 transform_later(need_pf_false); 1573 1574 Node *new_pf_wmt = new (C, 4) AddPNode( top(), old_pf_wm, 1575 _igvn.MakeConX(AllocatePrefetchDistance) ); 1576 transform_later(new_pf_wmt ); 1577 new_pf_wmt->set_req(0, need_pf_true); 1578 1579 Node *store_new_wmt = new (C, 4) StorePNode( need_pf_true, 1580 contended_phi_rawmem, eden_pf_adr, 1581 TypeRawPtr::BOTTOM, new_pf_wmt ); 1582 transform_later(store_new_wmt); 1583 1584 // adding prefetches 1585 pf_phi_abio->init_req( fall_in_path, i_o ); 1586 1587 Node *prefetch_adr; 1588 Node *prefetch; 1589 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize; 1590 uint step_size = AllocatePrefetchStepSize; 1591 uint distance = 0; 1592 1593 for ( uint i = 0; i < lines; i++ ) { 1594 prefetch_adr = new (C, 4) AddPNode( old_pf_wm, new_pf_wmt, 1595 _igvn.MakeConX(distance) ); 1596 transform_later(prefetch_adr); 1597 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr ); 1598 transform_later(prefetch); 1599 distance += step_size; 1600 i_o = prefetch; 1601 } 1602 pf_phi_abio->set_req( pf_path, i_o ); 1603 1604 pf_region->init_req( fall_in_path, need_pf_false ); 1605 pf_region->init_req( pf_path, need_pf_true ); 1606 1607 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 1608 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 1609 1610 transform_later(pf_region); 1611 transform_later(pf_phi_rawmem); 1612 transform_later(pf_phi_abio); 1613 1614 needgc_false = pf_region; 1615 contended_phi_rawmem = pf_phi_rawmem; 1616 i_o = pf_phi_abio; 1617 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { 1618 // Insert a prefetch for each allocation only on the fast-path 1619 Node *pf_region = new (C, 3) RegionNode(3); 1620 Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY, 1621 TypeRawPtr::BOTTOM ); 1622 1623 // Generate several prefetch instructions only for arrays. 1624 uint lines = (length != NULL) ? AllocatePrefetchLines : 1; 1625 uint step_size = AllocatePrefetchStepSize; 1626 uint distance = AllocatePrefetchDistance; 1627 1628 // Next cache address. 1629 Node *cache_adr = new (C, 4) AddPNode(old_eden_top, old_eden_top, 1630 _igvn.MakeConX(distance)); 1631 transform_later(cache_adr); 1632 cache_adr = new (C, 2) CastP2XNode(needgc_false, cache_adr); 1633 transform_later(cache_adr); 1634 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); 1635 cache_adr = new (C, 3) AndXNode(cache_adr, mask); 1636 transform_later(cache_adr); 1637 cache_adr = new (C, 2) CastX2PNode(cache_adr); 1638 transform_later(cache_adr); 1639 1640 // Prefetch 1641 Node *prefetch = new (C, 3) PrefetchWriteNode( contended_phi_rawmem, cache_adr ); 1642 prefetch->set_req(0, needgc_false); 1643 transform_later(prefetch); 1644 contended_phi_rawmem = prefetch; 1645 Node *prefetch_adr; 1646 distance = step_size; 1647 for ( uint i = 1; i < lines; i++ ) { 1648 prefetch_adr = new (C, 4) AddPNode( cache_adr, cache_adr, 1649 _igvn.MakeConX(distance) ); 1650 transform_later(prefetch_adr); 1651 prefetch = new (C, 3) PrefetchWriteNode( contended_phi_rawmem, prefetch_adr ); 1652 transform_later(prefetch); 1653 distance += step_size; 1654 contended_phi_rawmem = prefetch; 1655 } 1656 } else if( AllocatePrefetchStyle > 0 ) { 1657 // Insert a prefetch for each allocation only on the fast-path 1658 Node *prefetch_adr; 1659 Node *prefetch; 1660 // Generate several prefetch instructions only for arrays. 1661 uint lines = (length != NULL) ? AllocatePrefetchLines : 1; 1662 uint step_size = AllocatePrefetchStepSize; 1663 uint distance = AllocatePrefetchDistance; 1664 for ( uint i = 0; i < lines; i++ ) { 1665 prefetch_adr = new (C, 4) AddPNode( old_eden_top, new_eden_top, 1666 _igvn.MakeConX(distance) ); 1667 transform_later(prefetch_adr); 1668 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr ); 1669 // Do not let it float too high, since if eden_top == eden_end, 1670 // both might be null. 1671 if( i == 0 ) { // Set control for first prefetch, next follows it 1672 prefetch->init_req(0, needgc_false); 1673 } 1674 transform_later(prefetch); 1675 distance += step_size; 1676 i_o = prefetch; 1677 } 1678 } 1679 return i_o; 1680 } 1681 1682 1683 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 1684 expand_allocate_common(alloc, NULL, 1685 OptoRuntime::new_instance_Type(), 1686 OptoRuntime::new_instance_Java()); 1687 } 1688 1689 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 1690 Node* length = alloc->in(AllocateNode::ALength); 1691 expand_allocate_common(alloc, length, 1692 OptoRuntime::new_array_Type(), 1693 OptoRuntime::new_array_Java()); 1694 } 1695 1696 1697 // we have determined that this lock/unlock can be eliminated, we simply 1698 // eliminate the node without expanding it. 1699 // 1700 // Note: The membar's associated with the lock/unlock are currently not 1701 // eliminated. This should be investigated as a future enhancement. 1702 // 1703 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 1704 1705 if (!alock->is_eliminated()) { 1706 return false; 1707 } 1708 if (alock->is_Lock() && !alock->is_coarsened()) { 1709 // Create new "eliminated" BoxLock node and use it 1710 // in monitor debug info for the same object. 1711 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 1712 Node* obj = alock->obj_node(); 1713 if (!oldbox->is_eliminated()) { 1714 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 1715 newbox->set_eliminated(); 1716 transform_later(newbox); 1717 // Replace old box node with new box for all users 1718 // of the same object. 1719 for (uint i = 0; i < oldbox->outcnt();) { 1720 1721 bool next_edge = true; 1722 Node* u = oldbox->raw_out(i); 1723 if (u == alock) { 1724 i++; 1725 continue; // It will be removed below 1726 } 1727 if (u->is_Lock() && 1728 u->as_Lock()->obj_node() == obj && 1729 // oldbox could be referenced in debug info also 1730 u->as_Lock()->box_node() == oldbox) { 1731 assert(u->as_Lock()->is_eliminated(), "sanity"); 1732 _igvn.hash_delete(u); 1733 u->set_req(TypeFunc::Parms + 1, newbox); 1734 next_edge = false; 1735 #ifdef ASSERT 1736 } else if (u->is_Unlock() && u->as_Unlock()->obj_node() == obj) { 1737 assert(u->as_Unlock()->is_eliminated(), "sanity"); 1738 #endif 1739 } 1740 // Replace old box in monitor debug info. 1741 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 1742 SafePointNode* sfn = u->as_SafePoint(); 1743 JVMState* youngest_jvms = sfn->jvms(); 1744 int max_depth = youngest_jvms->depth(); 1745 for (int depth = 1; depth <= max_depth; depth++) { 1746 JVMState* jvms = youngest_jvms->of_depth(depth); 1747 int num_mon = jvms->nof_monitors(); 1748 // Loop over monitors 1749 for (int idx = 0; idx < num_mon; idx++) { 1750 Node* obj_node = sfn->monitor_obj(jvms, idx); 1751 Node* box_node = sfn->monitor_box(jvms, idx); 1752 if (box_node == oldbox && obj_node == obj) { 1753 int j = jvms->monitor_box_offset(idx); 1754 _igvn.hash_delete(u); 1755 u->set_req(j, newbox); 1756 next_edge = false; 1757 } 1758 } // for (int idx = 0; 1759 } // for (int depth = 1; 1760 } // if (u->is_SafePoint() 1761 if (next_edge) i++; 1762 } // for (uint i = 0; i < oldbox->outcnt();) 1763 } // if (!oldbox->is_eliminated()) 1764 } // if (alock->is_Lock() && !lock->is_coarsened()) 1765 1766 CompileLog* log = C->log(); 1767 if (log != NULL) { 1768 log->head("eliminate_lock lock='%d'", 1769 alock->is_Lock()); 1770 JVMState* p = alock->jvms(); 1771 while (p != NULL) { 1772 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1773 p = p->caller(); 1774 } 1775 log->tail("eliminate_lock"); 1776 } 1777 1778 #ifndef PRODUCT 1779 if (PrintEliminateLocks) { 1780 if (alock->is_Lock()) { 1781 tty->print_cr("++++ Eliminating: %d Lock", alock->_idx); 1782 } else { 1783 tty->print_cr("++++ Eliminating: %d Unlock", alock->_idx); 1784 } 1785 } 1786 #endif 1787 1788 Node* mem = alock->in(TypeFunc::Memory); 1789 Node* ctrl = alock->in(TypeFunc::Control); 1790 1791 extract_call_projections(alock); 1792 // There are 2 projections from the lock. The lock node will 1793 // be deleted when its last use is subsumed below. 1794 assert(alock->outcnt() == 2 && 1795 _fallthroughproj != NULL && 1796 _memproj_fallthrough != NULL, 1797 "Unexpected projections from Lock/Unlock"); 1798 1799 Node* fallthroughproj = _fallthroughproj; 1800 Node* memproj_fallthrough = _memproj_fallthrough; 1801 1802 // The memory projection from a lock/unlock is RawMem 1803 // The input to a Lock is merged memory, so extract its RawMem input 1804 // (unless the MergeMem has been optimized away.) 1805 if (alock->is_Lock()) { 1806 // Seach for MemBarAcquire node and delete it also. 1807 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 1808 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquire, ""); 1809 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 1810 Node* memproj = membar->proj_out(TypeFunc::Memory); 1811 _igvn.replace_node(ctrlproj, fallthroughproj); 1812 _igvn.replace_node(memproj, memproj_fallthrough); 1813 1814 // Delete FastLock node also if this Lock node is unique user 1815 // (a loop peeling may clone a Lock node). 1816 Node* flock = alock->as_Lock()->fastlock_node(); 1817 if (flock->outcnt() == 1) { 1818 assert(flock->unique_out() == alock, "sanity"); 1819 _igvn.replace_node(flock, top()); 1820 } 1821 } 1822 1823 // Seach for MemBarRelease node and delete it also. 1824 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() && 1825 ctrl->in(0)->is_MemBar()) { 1826 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 1827 assert(membar->Opcode() == Op_MemBarRelease && 1828 mem->is_Proj() && membar == mem->in(0), ""); 1829 _igvn.replace_node(fallthroughproj, ctrl); 1830 _igvn.replace_node(memproj_fallthrough, mem); 1831 fallthroughproj = ctrl; 1832 memproj_fallthrough = mem; 1833 ctrl = membar->in(TypeFunc::Control); 1834 mem = membar->in(TypeFunc::Memory); 1835 } 1836 1837 _igvn.replace_node(fallthroughproj, ctrl); 1838 _igvn.replace_node(memproj_fallthrough, mem); 1839 return true; 1840 } 1841 1842 1843 //------------------------------expand_lock_node---------------------- 1844 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 1845 1846 Node* ctrl = lock->in(TypeFunc::Control); 1847 Node* mem = lock->in(TypeFunc::Memory); 1848 Node* obj = lock->obj_node(); 1849 Node* box = lock->box_node(); 1850 Node* flock = lock->fastlock_node(); 1851 1852 // Make the merge point 1853 Node *region; 1854 Node *mem_phi; 1855 Node *slow_path; 1856 1857 if (UseOptoBiasInlining) { 1858 /* 1859 * See the full description in MacroAssembler::biased_locking_enter(). 1860 * 1861 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { 1862 * // The object is biased. 1863 * proto_node = klass->prototype_header; 1864 * o_node = thread | proto_node; 1865 * x_node = o_node ^ mark_word; 1866 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? 1867 * // Done. 1868 * } else { 1869 * if( (x_node & biased_lock_mask) != 0 ) { 1870 * // The klass's prototype header is no longer biased. 1871 * cas(&mark_word, mark_word, proto_node) 1872 * goto cas_lock; 1873 * } else { 1874 * // The klass's prototype header is still biased. 1875 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? 1876 * old = mark_word; 1877 * new = o_node; 1878 * } else { 1879 * // Different thread or anonymous biased. 1880 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); 1881 * new = thread | old; 1882 * } 1883 * // Try to rebias. 1884 * if( cas(&mark_word, old, new) == 0 ) { 1885 * // Done. 1886 * } else { 1887 * goto slow_path; // Failed. 1888 * } 1889 * } 1890 * } 1891 * } else { 1892 * // The object is not biased. 1893 * cas_lock: 1894 * if( FastLock(obj) == 0 ) { 1895 * // Done. 1896 * } else { 1897 * slow_path: 1898 * OptoRuntime::complete_monitor_locking_Java(obj); 1899 * } 1900 * } 1901 */ 1902 1903 region = new (C, 5) RegionNode(5); 1904 // create a Phi for the memory state 1905 mem_phi = new (C, 5) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 1906 1907 Node* fast_lock_region = new (C, 3) RegionNode(3); 1908 Node* fast_lock_mem_phi = new (C, 3) PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1909 1910 // First, check mark word for the biased lock pattern. 1911 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 1912 1913 // Get fast path - mark word has the biased lock pattern. 1914 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, 1915 markOopDesc::biased_lock_mask_in_place, 1916 markOopDesc::biased_lock_pattern, true); 1917 // fast_lock_region->in(1) is set to slow path. 1918 fast_lock_mem_phi->init_req(1, mem); 1919 1920 // Now check that the lock is biased to the current thread and has 1921 // the same epoch and bias as Klass::_prototype_header. 1922 1923 // Special-case a fresh allocation to avoid building nodes: 1924 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); 1925 if (klass_node == NULL) { 1926 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1927 klass_node = transform_later( LoadKlassNode::make(_igvn, mem, k_adr, _igvn.type(k_adr)->is_ptr()) ); 1928 #ifdef _LP64 1929 if (UseCompressedOops && klass_node->is_DecodeN()) { 1930 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity"); 1931 klass_node->in(1)->init_req(0, ctrl); 1932 } else 1933 #endif 1934 klass_node->init_req(0, ctrl); 1935 } 1936 Node *proto_node = make_load(ctrl, mem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeX_X, TypeX_X->basic_type()); 1937 1938 Node* thread = transform_later(new (C, 1) ThreadLocalNode()); 1939 Node* cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread)); 1940 Node* o_node = transform_later(new (C, 3) OrXNode(cast_thread, proto_node)); 1941 Node* x_node = transform_later(new (C, 3) XorXNode(o_node, mark_node)); 1942 1943 // Get slow path - mark word does NOT match the value. 1944 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, 1945 (~markOopDesc::age_mask_in_place), 0); 1946 // region->in(3) is set to fast path - the object is biased to the current thread. 1947 mem_phi->init_req(3, mem); 1948 1949 1950 // Mark word does NOT match the value (thread | Klass::_prototype_header). 1951 1952 1953 // First, check biased pattern. 1954 // Get fast path - _prototype_header has the same biased lock pattern. 1955 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, 1956 markOopDesc::biased_lock_mask_in_place, 0, true); 1957 1958 not_biased_ctrl = fast_lock_region->in(2); // Slow path 1959 // fast_lock_region->in(2) - the prototype header is no longer biased 1960 // and we have to revoke the bias on this object. 1961 // We are going to try to reset the mark of this object to the prototype 1962 // value and fall through to the CAS-based locking scheme. 1963 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 1964 Node* cas = new (C, 5) StoreXConditionalNode(not_biased_ctrl, mem, adr, 1965 proto_node, mark_node); 1966 transform_later(cas); 1967 Node* proj = transform_later( new (C, 1) SCMemProjNode(cas)); 1968 fast_lock_mem_phi->init_req(2, proj); 1969 1970 1971 // Second, check epoch bits. 1972 Node* rebiased_region = new (C, 3) RegionNode(3); 1973 Node* old_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X); 1974 Node* new_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X); 1975 1976 // Get slow path - mark word does NOT match epoch bits. 1977 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, 1978 markOopDesc::epoch_mask_in_place, 0); 1979 // The epoch of the current bias is not valid, attempt to rebias the object 1980 // toward the current thread. 1981 rebiased_region->init_req(2, epoch_ctrl); 1982 old_phi->init_req(2, mark_node); 1983 new_phi->init_req(2, o_node); 1984 1985 // rebiased_region->in(1) is set to fast path. 1986 // The epoch of the current bias is still valid but we know 1987 // nothing about the owner; it might be set or it might be clear. 1988 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place | 1989 markOopDesc::age_mask_in_place | 1990 markOopDesc::epoch_mask_in_place); 1991 Node* old = transform_later(new (C, 3) AndXNode(mark_node, cmask)); 1992 cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread)); 1993 Node* new_mark = transform_later(new (C, 3) OrXNode(cast_thread, old)); 1994 old_phi->init_req(1, old); 1995 new_phi->init_req(1, new_mark); 1996 1997 transform_later(rebiased_region); 1998 transform_later(old_phi); 1999 transform_later(new_phi); 2000 2001 // Try to acquire the bias of the object using an atomic operation. 2002 // If this fails we will go in to the runtime to revoke the object's bias. 2003 cas = new (C, 5) StoreXConditionalNode(rebiased_region, mem, adr, 2004 new_phi, old_phi); 2005 transform_later(cas); 2006 proj = transform_later( new (C, 1) SCMemProjNode(cas)); 2007 2008 // Get slow path - Failed to CAS. 2009 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); 2010 mem_phi->init_req(4, proj); 2011 // region->in(4) is set to fast path - the object is rebiased to the current thread. 2012 2013 // Failed to CAS. 2014 slow_path = new (C, 3) RegionNode(3); 2015 Node *slow_mem = new (C, 3) PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); 2016 2017 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control 2018 slow_mem->init_req(1, proj); 2019 2020 // Call CAS-based locking scheme (FastLock node). 2021 2022 transform_later(fast_lock_region); 2023 transform_later(fast_lock_mem_phi); 2024 2025 // Get slow path - FastLock failed to lock the object. 2026 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); 2027 mem_phi->init_req(2, fast_lock_mem_phi); 2028 // region->in(2) is set to fast path - the object is locked to the current thread. 2029 2030 slow_path->init_req(2, ctrl); // Capture slow-control 2031 slow_mem->init_req(2, fast_lock_mem_phi); 2032 2033 transform_later(slow_path); 2034 transform_later(slow_mem); 2035 // Reset lock's memory edge. 2036 lock->set_req(TypeFunc::Memory, slow_mem); 2037 2038 } else { 2039 region = new (C, 3) RegionNode(3); 2040 // create a Phi for the memory state 2041 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2042 2043 // Optimize test; set region slot 2 2044 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); 2045 mem_phi->init_req(2, mem); 2046 } 2047 2048 // Make slow path call 2049 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box ); 2050 2051 extract_call_projections(call); 2052 2053 // Slow path can only throw asynchronous exceptions, which are always 2054 // de-opted. So the compiler thinks the slow-call can never throw an 2055 // exception. If it DOES throw an exception we would need the debug 2056 // info removed first (since if it throws there is no monitor). 2057 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2058 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2059 2060 // Capture slow path 2061 // disconnect fall-through projection from call and create a new one 2062 // hook up users of fall-through projection to region 2063 Node *slow_ctrl = _fallthroughproj->clone(); 2064 transform_later(slow_ctrl); 2065 _igvn.hash_delete(_fallthroughproj); 2066 _fallthroughproj->disconnect_inputs(NULL); 2067 region->init_req(1, slow_ctrl); 2068 // region inputs are now complete 2069 transform_later(region); 2070 _igvn.replace_node(_fallthroughproj, region); 2071 2072 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) ); 2073 mem_phi->init_req(1, memproj ); 2074 transform_later(mem_phi); 2075 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2076 } 2077 2078 //------------------------------expand_unlock_node---------------------- 2079 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 2080 2081 Node* ctrl = unlock->in(TypeFunc::Control); 2082 Node* mem = unlock->in(TypeFunc::Memory); 2083 Node* obj = unlock->obj_node(); 2084 Node* box = unlock->box_node(); 2085 2086 // No need for a null check on unlock 2087 2088 // Make the merge point 2089 Node *region; 2090 Node *mem_phi; 2091 2092 if (UseOptoBiasInlining) { 2093 // Check for biased locking unlock case, which is a no-op. 2094 // See the full description in MacroAssembler::biased_locking_exit(). 2095 region = new (C, 4) RegionNode(4); 2096 // create a Phi for the memory state 2097 mem_phi = new (C, 4) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2098 mem_phi->init_req(3, mem); 2099 2100 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2101 ctrl = opt_bits_test(ctrl, region, 3, mark_node, 2102 markOopDesc::biased_lock_mask_in_place, 2103 markOopDesc::biased_lock_pattern); 2104 } else { 2105 region = new (C, 3) RegionNode(3); 2106 // create a Phi for the memory state 2107 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2108 } 2109 2110 FastUnlockNode *funlock = new (C, 3) FastUnlockNode( ctrl, obj, box ); 2111 funlock = transform_later( funlock )->as_FastUnlock(); 2112 // Optimize test; set region slot 2 2113 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); 2114 2115 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 ); 2116 2117 extract_call_projections(call); 2118 2119 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2120 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2121 2122 // No exceptions for unlocking 2123 // Capture slow path 2124 // disconnect fall-through projection from call and create a new one 2125 // hook up users of fall-through projection to region 2126 Node *slow_ctrl = _fallthroughproj->clone(); 2127 transform_later(slow_ctrl); 2128 _igvn.hash_delete(_fallthroughproj); 2129 _fallthroughproj->disconnect_inputs(NULL); 2130 region->init_req(1, slow_ctrl); 2131 // region inputs are now complete 2132 transform_later(region); 2133 _igvn.replace_node(_fallthroughproj, region); 2134 2135 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) ); 2136 mem_phi->init_req(1, memproj ); 2137 mem_phi->init_req(2, mem); 2138 transform_later(mem_phi); 2139 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2140 } 2141 2142 //------------------------------expand_macro_nodes---------------------- 2143 // Returns true if a failure occurred. 2144 bool PhaseMacroExpand::expand_macro_nodes() { 2145 if (C->macro_count() == 0) 2146 return false; 2147 // First, attempt to eliminate locks 2148 bool progress = true; 2149 while (progress) { 2150 progress = false; 2151 for (int i = C->macro_count(); i > 0; i--) { 2152 Node * n = C->macro_node(i-1); 2153 bool success = false; 2154 debug_only(int old_macro_count = C->macro_count();); 2155 if (n->is_AbstractLock()) { 2156 success = eliminate_locking_node(n->as_AbstractLock()); 2157 } else if (n->Opcode() == Op_LoopLimit) { 2158 // Remove it from macro list and put on IGVN worklist to optimize. 2159 C->remove_macro_node(n); 2160 _igvn._worklist.push(n); 2161 success = true; 2162 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { 2163 _igvn.replace_node(n, n->in(1)); 2164 success = true; 2165 } 2166 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2167 progress = progress || success; 2168 } 2169 } 2170 // Next, attempt to eliminate allocations 2171 progress = true; 2172 while (progress) { 2173 progress = false; 2174 for (int i = C->macro_count(); i > 0; i--) { 2175 Node * n = C->macro_node(i-1); 2176 bool success = false; 2177 debug_only(int old_macro_count = C->macro_count();); 2178 switch (n->class_id()) { 2179 case Node::Class_Allocate: 2180 case Node::Class_AllocateArray: 2181 success = eliminate_allocate_node(n->as_Allocate()); 2182 break; 2183 case Node::Class_Lock: 2184 case Node::Class_Unlock: 2185 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 2186 break; 2187 default: 2188 assert(false, "unknown node type in macro list"); 2189 } 2190 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2191 progress = progress || success; 2192 } 2193 } 2194 // Make sure expansion will not cause node limit to be exceeded. 2195 // Worst case is a macro node gets expanded into about 50 nodes. 2196 // Allow 50% more for optimization. 2197 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) ) 2198 return true; 2199 2200 // expand "macro" nodes 2201 // nodes are removed from the macro list as they are processed 2202 while (C->macro_count() > 0) { 2203 int macro_count = C->macro_count(); 2204 Node * n = C->macro_node(macro_count-1); 2205 assert(n->is_macro(), "only macro nodes expected here"); 2206 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2207 // node is unreachable, so don't try to expand it 2208 C->remove_macro_node(n); 2209 continue; 2210 } 2211 switch (n->class_id()) { 2212 case Node::Class_Allocate: 2213 expand_allocate(n->as_Allocate()); 2214 break; 2215 case Node::Class_AllocateArray: 2216 expand_allocate_array(n->as_AllocateArray()); 2217 break; 2218 case Node::Class_Lock: 2219 expand_lock_node(n->as_Lock()); 2220 break; 2221 case Node::Class_Unlock: 2222 expand_unlock_node(n->as_Unlock()); 2223 break; 2224 default: 2225 assert(false, "unknown node type in macro list"); 2226 } 2227 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2228 if (C->failing()) return true; 2229 } 2230 2231 _igvn.set_delay_transform(false); 2232 _igvn.optimize(); 2233 return false; 2234 }