1 /* 2 * Copyright (c) 2005, 2016, 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 "gc/shenandoah/brooksPointer.hpp" 28 #include "libadt/vectset.hpp" 29 #include "opto/addnode.hpp" 30 #include "opto/arraycopynode.hpp" 31 #include "opto/callnode.hpp" 32 #include "opto/castnode.hpp" 33 #include "opto/cfgnode.hpp" 34 #include "opto/compile.hpp" 35 #include "opto/convertnode.hpp" 36 #include "opto/graphKit.hpp" 37 #include "opto/locknode.hpp" 38 #include "opto/loopnode.hpp" 39 #include "opto/macro.hpp" 40 #include "opto/memnode.hpp" 41 #include "opto/narrowptrnode.hpp" 42 #include "opto/node.hpp" 43 #include "opto/opaquenode.hpp" 44 #include "opto/phaseX.hpp" 45 #include "opto/rootnode.hpp" 46 #include "opto/runtime.hpp" 47 #include "opto/shenandoahSupport.hpp" 48 #include "opto/subnode.hpp" 49 #include "opto/type.hpp" 50 #include "runtime/sharedRuntime.hpp" 51 52 53 // 54 // Replace any references to "oldref" in inputs to "use" with "newref". 55 // Returns the number of replacements made. 56 // 57 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { 58 int nreplacements = 0; 59 uint req = use->req(); 60 for (uint j = 0; j < use->len(); j++) { 61 Node *uin = use->in(j); 62 if (uin == oldref) { 63 if (j < req) 64 use->set_req(j, newref); 65 else 66 use->set_prec(j, newref); 67 nreplacements++; 68 } else if (j >= req && uin == NULL) { 69 break; 70 } 71 } 72 return nreplacements; 73 } 74 75 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) { 76 // Copy debug information and adjust JVMState information 77 uint old_dbg_start = oldcall->tf()->domain()->cnt(); 78 uint new_dbg_start = newcall->tf()->domain()->cnt(); 79 int jvms_adj = new_dbg_start - old_dbg_start; 80 assert (new_dbg_start == newcall->req(), "argument count mismatch"); 81 82 // SafePointScalarObject node could be referenced several times in debug info. 83 // Use Dict to record cloned nodes. 84 Dict* sosn_map = new Dict(cmpkey,hashkey); 85 for (uint i = old_dbg_start; i < oldcall->req(); i++) { 86 Node* old_in = oldcall->in(i); 87 // Clone old SafePointScalarObjectNodes, adjusting their field contents. 88 if (old_in != NULL && old_in->is_SafePointScalarObject()) { 89 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject(); 90 uint old_unique = C->unique(); 91 Node* new_in = old_sosn->clone(sosn_map); 92 if (old_unique != C->unique()) { // New node? 93 new_in->set_req(0, C->root()); // reset control edge 94 new_in = transform_later(new_in); // Register new node. 95 } 96 old_in = new_in; 97 } 98 newcall->add_req(old_in); 99 } 100 101 // JVMS may be shared so clone it before we modify it 102 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL); 103 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) { 104 jvms->set_map(newcall); 105 jvms->set_locoff(jvms->locoff()+jvms_adj); 106 jvms->set_stkoff(jvms->stkoff()+jvms_adj); 107 jvms->set_monoff(jvms->monoff()+jvms_adj); 108 jvms->set_scloff(jvms->scloff()+jvms_adj); 109 jvms->set_endoff(jvms->endoff()+jvms_adj); 110 } 111 } 112 113 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) { 114 Node* cmp; 115 if (mask != 0) { 116 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask))); 117 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits))); 118 } else { 119 cmp = word; 120 } 121 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne)); 122 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); 123 transform_later(iff); 124 125 // Fast path taken. 126 Node *fast_taken = transform_later(new IfFalseNode(iff)); 127 128 // Fast path not-taken, i.e. slow path 129 Node *slow_taken = transform_later(new IfTrueNode(iff)); 130 131 if (return_fast_path) { 132 region->init_req(edge, slow_taken); // Capture slow-control 133 return fast_taken; 134 } else { 135 region->init_req(edge, fast_taken); // Capture fast-control 136 return slow_taken; 137 } 138 } 139 140 //--------------------copy_predefined_input_for_runtime_call-------------------- 141 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { 142 // Set fixed predefined input arguments 143 call->init_req( TypeFunc::Control, ctrl ); 144 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); 145 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? 146 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); 147 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); 148 } 149 150 //------------------------------make_slow_call--------------------------------- 151 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, 152 address slow_call, const char* leaf_name, Node* slow_path, 153 Node* parm0, Node* parm1, Node* parm2) { 154 155 // Slow-path call 156 CallNode *call = leaf_name 157 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) 158 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM ); 159 160 // Slow path call has no side-effects, uses few values 161 copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); 162 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0); 163 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1); 164 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2); 165 copy_call_debug_info(oldcall, call); 166 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 167 _igvn.replace_node(oldcall, call); 168 transform_later(call); 169 170 return call; 171 } 172 173 void PhaseMacroExpand::extract_call_projections(CallNode *call) { 174 _fallthroughproj = NULL; 175 _fallthroughcatchproj = NULL; 176 _ioproj_fallthrough = NULL; 177 _ioproj_catchall = NULL; 178 _catchallcatchproj = NULL; 179 _memproj_fallthrough = NULL; 180 _memproj_catchall = NULL; 181 _resproj = NULL; 182 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) { 183 ProjNode *pn = call->fast_out(i)->as_Proj(); 184 switch (pn->_con) { 185 case TypeFunc::Control: 186 { 187 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj 188 _fallthroughproj = pn; 189 DUIterator_Fast jmax, j = pn->fast_outs(jmax); 190 const Node *cn = pn->fast_out(j); 191 if (cn->is_Catch()) { 192 ProjNode *cpn = NULL; 193 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { 194 cpn = cn->fast_out(k)->as_Proj(); 195 assert(cpn->is_CatchProj(), "must be a CatchProjNode"); 196 if (cpn->_con == CatchProjNode::fall_through_index) 197 _fallthroughcatchproj = cpn; 198 else { 199 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); 200 _catchallcatchproj = cpn; 201 } 202 } 203 } 204 break; 205 } 206 case TypeFunc::I_O: 207 if (pn->_is_io_use) 208 _ioproj_catchall = pn; 209 else 210 _ioproj_fallthrough = pn; 211 break; 212 case TypeFunc::Memory: 213 if (pn->_is_io_use) 214 _memproj_catchall = pn; 215 else 216 _memproj_fallthrough = pn; 217 break; 218 case TypeFunc::Parms: 219 _resproj = pn; 220 break; 221 default: 222 assert(false, "unexpected projection from allocation node."); 223 } 224 } 225 226 } 227 228 // Eliminate a card mark sequence. p2x is a ConvP2XNode 229 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) { 230 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required"); 231 if (!UseG1GC) { 232 // vanilla/CMS post barrier 233 Node *shift = p2x->unique_out(); 234 Node *addp = shift->unique_out(); 235 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) { 236 Node *mem = addp->last_out(j); 237 if (UseCondCardMark && mem->is_Load()) { 238 assert(mem->Opcode() == Op_LoadB, "unexpected code shape"); 239 // The load is checking if the card has been written so 240 // replace it with zero to fold the test. 241 _igvn.replace_node(mem, intcon(0)); 242 continue; 243 } 244 assert(mem->is_Store(), "store required"); 245 _igvn.replace_node(mem, mem->in(MemNode::Memory)); 246 } 247 } else { 248 // G1 pre/post barriers 249 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes"); 250 // It could be only one user, URShift node, in Object.clone() intrinsic 251 // but the new allocation is passed to arraycopy stub and it could not 252 // be scalar replaced. So we don't check the case. 253 254 // An other case of only one user (Xor) is when the value check for NULL 255 // in G1 post barrier is folded after CCP so the code which used URShift 256 // is removed. 257 258 // Take Region node before eliminating post barrier since it also 259 // eliminates CastP2X node when it has only one user. 260 Node* this_region = p2x->in(0); 261 assert(this_region != NULL, ""); 262 263 // Remove G1 post barrier. 264 265 // Search for CastP2X->Xor->URShift->Cmp path which 266 // checks if the store done to a different from the value's region. 267 // And replace Cmp with #0 (false) to collapse G1 post barrier. 268 Node* xorx = p2x->find_out_with(Op_XorX); 269 if (xorx != NULL) { 270 Node* shift = xorx->unique_out(); 271 Node* cmpx = shift->unique_out(); 272 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() && 273 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne, 274 "missing region check in G1 post barrier"); 275 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 276 277 // Remove G1 pre barrier. 278 279 // Search "if (marking != 0)" check and set it to "false". 280 // There is no G1 pre barrier if previous stored value is NULL 281 // (for example, after initialization). 282 if (this_region->is_Region() && this_region->req() == 3) { 283 int ind = 1; 284 if (!this_region->in(ind)->is_IfFalse()) { 285 ind = 2; 286 } 287 if (this_region->in(ind)->is_IfFalse() && 288 this_region->in(ind)->in(0)->is_g1_marking_if(&_igvn)) { 289 Node* cmpx = this_region->in(ind)->in(0)->in(1)->in(1); 290 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 291 } 292 } 293 } else { 294 assert(!GraphKit::use_ReduceInitialCardMarks(), "can only happen with card marking"); 295 // This is a G1 post barrier emitted by the Object.clone() intrinsic. 296 // Search for the CastP2X->URShiftX->AddP->LoadB->Cmp path which checks if the card 297 // is marked as young_gen and replace the Cmp with 0 (false) to collapse the barrier. 298 Node* shift = p2x->find_out_with(Op_URShiftX); 299 assert(shift != NULL, "missing G1 post barrier"); 300 Node* addp = shift->unique_out(); 301 Node* load = addp->find_out_with(Op_LoadB); 302 assert(load != NULL, "missing G1 post barrier"); 303 Node* cmpx = load->unique_out(); 304 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() && 305 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne, 306 "missing card value check in G1 post barrier"); 307 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 308 // There is no G1 pre barrier in this case 309 } 310 // Now CastP2X can be removed since it is used only on dead path 311 // which currently still alive until igvn optimize it. 312 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, ""); 313 _igvn.replace_node(p2x, top()); 314 } 315 } 316 317 void PhaseMacroExpand::eliminate_g1_wb_pre(Node* n) { 318 Node* c = n->as_Call()->proj_out(TypeFunc::Control); 319 c = c->unique_ctrl_out(); 320 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 321 c = c->unique_ctrl_out(); 322 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 323 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0); 324 assert(iff->is_If(), "expect test"); 325 if (!iff->is_g1_marking_if(&_igvn)) { 326 c = c->unique_ctrl_out(); 327 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 328 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0); 329 assert(iff->is_g1_marking_if(&_igvn), "expect marking test"); 330 } 331 Node* cmpx = iff->in(1)->in(1); 332 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); 333 _igvn.rehash_node_delayed(n); 334 n->del_req(n->req()-1); 335 } 336 337 // Search for a memory operation for the specified memory slice. 338 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) { 339 Node *orig_mem = mem; 340 Node *alloc_mem = alloc->in(TypeFunc::Memory); 341 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr(); 342 while (true) { 343 if (mem == alloc_mem || mem == start_mem ) { 344 return mem; // hit one of our sentinels 345 } else if (mem->is_MergeMem()) { 346 mem = mem->as_MergeMem()->memory_at(alias_idx); 347 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { 348 Node *in = mem->in(0); 349 // we can safely skip over safepoints, calls, locks and membars because we 350 // already know that the object is safe to eliminate. 351 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { 352 return in; 353 } else if (in->is_Call()) { 354 CallNode *call = in->as_Call(); 355 if (call->may_modify(tinst, phase)) { 356 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape"); 357 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) { 358 return in; 359 } 360 } 361 mem = in->in(TypeFunc::Memory); 362 } else if (in->is_MemBar()) { 363 ArrayCopyNode* ac = NULL; 364 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) { 365 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone"); 366 return ac; 367 } 368 mem = in->in(TypeFunc::Memory); 369 } else { 370 assert(false, "unexpected projection"); 371 } 372 } else if (mem->is_Store()) { 373 const TypePtr* atype = mem->as_Store()->adr_type(); 374 int adr_idx = phase->C->get_alias_index(atype); 375 if (adr_idx == alias_idx) { 376 assert(atype->isa_oopptr(), "address type must be oopptr"); 377 int adr_offset = atype->offset(); 378 uint adr_iid = atype->is_oopptr()->instance_id(); 379 // Array elements references have the same alias_idx 380 // but different offset and different instance_id. 381 if (adr_offset == offset && adr_iid == alloc->_idx) 382 return mem; 383 } else { 384 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw"); 385 } 386 mem = mem->in(MemNode::Memory); 387 } else if (mem->is_ClearArray()) { 388 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) { 389 // Can not bypass initialization of the instance 390 // we are looking. 391 debug_only(intptr_t offset;) 392 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity"); 393 InitializeNode* init = alloc->as_Allocate()->initialization(); 394 // We are looking for stored value, return Initialize node 395 // or memory edge from Allocate node. 396 if (init != NULL) 397 return init; 398 else 399 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers). 400 } 401 // Otherwise skip it (the call updated 'mem' value). 402 } else if (mem->Opcode() == Op_SCMemProj) { 403 mem = mem->in(0); 404 Node* adr = NULL; 405 if (mem->is_LoadStore()) { 406 adr = mem->in(MemNode::Address); 407 } else { 408 assert(mem->Opcode() == Op_EncodeISOArray || 409 mem->Opcode() == Op_StrCompressedCopy, "sanity"); 410 adr = mem->in(3); // Destination array 411 } 412 const TypePtr* atype = adr->bottom_type()->is_ptr(); 413 int adr_idx = phase->C->get_alias_index(atype); 414 if (adr_idx == alias_idx) { 415 DEBUG_ONLY(mem->dump();) 416 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 417 return NULL; 418 } 419 mem = mem->in(MemNode::Memory); 420 } else if (mem->Opcode() == Op_StrInflatedCopy) { 421 Node* adr = mem->in(3); // Destination array 422 const TypePtr* atype = adr->bottom_type()->is_ptr(); 423 int adr_idx = phase->C->get_alias_index(atype); 424 if (adr_idx == alias_idx) { 425 DEBUG_ONLY(mem->dump();) 426 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field"); 427 return NULL; 428 } 429 mem = mem->in(MemNode::Memory); 430 } else { 431 return mem; 432 } 433 assert(mem != orig_mem, "dead memory loop"); 434 } 435 } 436 437 // Generate loads from source of the arraycopy for fields of 438 // destination needed at a deoptimization point 439 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) { 440 BasicType bt = ft; 441 const Type *type = ftype; 442 if (ft == T_NARROWOOP) { 443 bt = T_OBJECT; 444 type = ftype->make_oopptr(); 445 } 446 Node* res = NULL; 447 if (ac->is_clonebasic()) { 448 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base); 449 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset))); 450 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset); 451 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned); 452 } else { 453 if (ac->modifies(offset, offset, &_igvn, true)) { 454 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result"); 455 uint shift = exact_log2(type2aelembytes(bt)); 456 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); 457 #ifdef _LP64 458 diff = _igvn.transform(new ConvI2LNode(diff)); 459 #endif 460 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift))); 461 462 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff)); 463 Node* base = ac->in(ArrayCopyNode::Src); 464 Node* adr = _igvn.transform(new AddPNode(base, base, off)); 465 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset); 466 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned); 467 } 468 } 469 if (res != NULL) { 470 res = _igvn.transform(res); 471 if (ftype->isa_narrowoop()) { 472 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes 473 res = _igvn.transform(new EncodePNode(res, ftype)); 474 } 475 return res; 476 } 477 return NULL; 478 } 479 480 // 481 // Given a Memory Phi, compute a value Phi containing the values from stores 482 // on the input paths. 483 // Note: this function is recursive, its depth is limited by the "level" argument 484 // Returns the computed Phi, or NULL if it cannot compute it. 485 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, AllocateNode *alloc, Node_Stack *value_phis, int level) { 486 assert(mem->is_Phi(), "sanity"); 487 int alias_idx = C->get_alias_index(adr_t); 488 int offset = adr_t->offset(); 489 int instance_id = adr_t->instance_id(); 490 491 // Check if an appropriate value phi already exists. 492 Node* region = mem->in(0); 493 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { 494 Node* phi = region->fast_out(k); 495 if (phi->is_Phi() && phi != mem && 496 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) { 497 return phi; 498 } 499 } 500 // Check if an appropriate new value phi already exists. 501 Node* new_phi = value_phis->find(mem->_idx); 502 if (new_phi != NULL) 503 return new_phi; 504 505 if (level <= 0) { 506 return NULL; // Give up: phi tree too deep 507 } 508 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 509 Node *alloc_mem = alloc->in(TypeFunc::Memory); 510 511 uint length = mem->req(); 512 GrowableArray <Node *> values(length, length, NULL, false); 513 514 // create a new Phi for the value 515 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset); 516 transform_later(phi); 517 value_phis->push(phi, mem->_idx); 518 519 for (uint j = 1; j < length; j++) { 520 Node *in = mem->in(j); 521 if (in == NULL || in->is_top()) { 522 values.at_put(j, in); 523 } else { 524 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); 525 if (val == start_mem || val == alloc_mem) { 526 // hit a sentinel, return appropriate 0 value 527 values.at_put(j, _igvn.zerocon(ft)); 528 continue; 529 } 530 if (val->is_Initialize()) { 531 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 532 } 533 if (val == NULL) { 534 return NULL; // can't find a value on this path 535 } 536 if (val == mem) { 537 values.at_put(j, mem); 538 } else if (val->is_Store()) { 539 values.at_put(j, ShenandoahBarrierNode::skip_through_barrier(val->in(MemNode::ValueIn))); 540 } else if(val->is_Proj() && val->in(0) == alloc) { 541 values.at_put(j, _igvn.zerocon(ft)); 542 } else if (val->is_Phi()) { 543 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); 544 if (val == NULL) { 545 return NULL; 546 } 547 values.at_put(j, val); 548 } else if (val->Opcode() == Op_SCMemProj) { 549 assert(val->in(0)->is_LoadStore() || 550 val->in(0)->Opcode() == Op_EncodeISOArray || 551 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity"); 552 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field"); 553 return NULL; 554 } else if (val->is_ArrayCopy()) { 555 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc); 556 if (res == NULL) { 557 return NULL; 558 } 559 values.at_put(j, res); 560 } else { 561 #ifdef ASSERT 562 val->dump(); 563 assert(false, "unknown node on this path"); 564 #endif 565 return NULL; // unknown node on this path 566 } 567 } 568 } 569 // Set Phi's inputs 570 for (uint j = 1; j < length; j++) { 571 if (values.at(j) == mem) { 572 phi->init_req(j, phi); 573 } else { 574 phi->init_req(j, values.at(j)); 575 } 576 } 577 return phi; 578 } 579 580 // Search the last value stored into the object's field. 581 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) { 582 assert(adr_t->is_known_instance_field(), "instance required"); 583 int instance_id = adr_t->instance_id(); 584 assert((uint)instance_id == alloc->_idx, "wrong allocation"); 585 586 int alias_idx = C->get_alias_index(adr_t); 587 int offset = adr_t->offset(); 588 Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 589 Node *alloc_ctrl = alloc->in(TypeFunc::Control); 590 Node *alloc_mem = alloc->in(TypeFunc::Memory); 591 Arena *a = Thread::current()->resource_area(); 592 VectorSet visited(a); 593 594 595 bool done = sfpt_mem == alloc_mem; 596 Node *mem = sfpt_mem; 597 while (!done) { 598 if (visited.test_set(mem->_idx)) { 599 return NULL; // found a loop, give up 600 } 601 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); 602 if (mem == start_mem || mem == alloc_mem) { 603 done = true; // hit a sentinel, return appropriate 0 value 604 } else if (mem->is_Initialize()) { 605 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 606 if (mem == NULL) { 607 done = true; // Something go wrong. 608 } else if (mem->is_Store()) { 609 const TypePtr* atype = mem->as_Store()->adr_type(); 610 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 611 done = true; 612 } 613 } else if (mem->is_Store()) { 614 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 615 assert(atype != NULL, "address type must be oopptr"); 616 assert(C->get_alias_index(atype) == alias_idx && 617 atype->is_known_instance_field() && atype->offset() == offset && 618 atype->instance_id() == instance_id, "store is correct memory slice"); 619 done = true; 620 } else if (mem->is_Phi()) { 621 // try to find a phi's unique input 622 Node *unique_input = NULL; 623 Node *top = C->top(); 624 for (uint i = 1; i < mem->req(); i++) { 625 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); 626 if (n == NULL || n == top || n == mem) { 627 continue; 628 } else if (unique_input == NULL) { 629 unique_input = n; 630 } else if (unique_input != n) { 631 unique_input = top; 632 break; 633 } 634 } 635 if (unique_input != NULL && unique_input != top) { 636 mem = unique_input; 637 } else { 638 done = true; 639 } 640 } else if (mem->is_ArrayCopy()) { 641 done = true; 642 } else { 643 assert(false, "unexpected node"); 644 } 645 } 646 if (mem != NULL) { 647 if (mem == start_mem || mem == alloc_mem) { 648 // hit a sentinel, return appropriate 0 value 649 return _igvn.zerocon(ft); 650 } else if (mem->is_Store()) { 651 return ShenandoahBarrierNode::skip_through_barrier(mem->in(MemNode::ValueIn)); 652 } else if (mem->is_Phi()) { 653 // attempt to produce a Phi reflecting the values on the input paths of the Phi 654 Node_Stack value_phis(a, 8); 655 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); 656 if (phi != NULL) { 657 return phi; 658 } else { 659 // Kill all new Phis 660 while(value_phis.is_nonempty()) { 661 Node* n = value_phis.node(); 662 _igvn.replace_node(n, C->top()); 663 value_phis.pop(); 664 } 665 } 666 } else if (mem->is_ArrayCopy()) { 667 Node* ctl = mem->in(0); 668 Node* m = mem->in(TypeFunc::Memory); 669 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) { 670 // pin the loads in the uncommon trap path 671 ctl = sfpt_ctl; 672 m = sfpt_mem; 673 } 674 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc); 675 } 676 } 677 // Something go wrong. 678 return NULL; 679 } 680 681 // Check the possibility of scalar replacement. 682 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 683 // Scan the uses of the allocation to check for anything that would 684 // prevent us from eliminating it. 685 NOT_PRODUCT( const char* fail_eliminate = NULL; ) 686 DEBUG_ONLY( Node* disq_node = NULL; ) 687 bool can_eliminate = true; 688 689 Node* res = alloc->result_cast(); 690 const TypeOopPtr* res_type = NULL; 691 if (res == NULL) { 692 // All users were eliminated. 693 } else if (!res->is_CheckCastPP()) { 694 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 695 can_eliminate = false; 696 } else { 697 res_type = _igvn.type(res)->isa_oopptr(); 698 if (res_type == NULL) { 699 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 700 can_eliminate = false; 701 } else if (res_type->isa_aryptr()) { 702 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 703 if (length < 0) { 704 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 705 can_eliminate = false; 706 } 707 } 708 } 709 710 if (can_eliminate && res != NULL) { 711 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 712 j < jmax && can_eliminate; j++) { 713 Node* use = res->fast_out(j); 714 715 if (use->is_AddP()) { 716 const TypePtr* addp_type = _igvn.type(use)->is_ptr(); 717 int offset = addp_type->offset(); 718 719 if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 720 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) 721 can_eliminate = false; 722 break; 723 } 724 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 725 k < kmax && can_eliminate; k++) { 726 Node* n = use->fast_out(k); 727 if (!n->is_Store() && n->Opcode() != Op_CastP2X && 728 !n->is_g1_wb_pre_call() && 729 !(n->is_ArrayCopy() && 730 n->as_ArrayCopy()->is_clonebasic() && 731 n->in(ArrayCopyNode::Dest) == use)) { 732 DEBUG_ONLY(disq_node = n;) 733 if (n->is_Load() || n->is_LoadStore()) { 734 NOT_PRODUCT(fail_eliminate = "Field load";) 735 } else { 736 NOT_PRODUCT(fail_eliminate = "Not store field referrence";) 737 } 738 can_eliminate = false; 739 } 740 } 741 } else if (use->is_ArrayCopy() && 742 (use->as_ArrayCopy()->is_arraycopy_validated() || 743 use->as_ArrayCopy()->is_copyof_validated() || 744 use->as_ArrayCopy()->is_copyofrange_validated()) && 745 use->in(ArrayCopyNode::Dest) == res) { 746 // ok to eliminate 747 } else if (use->is_SafePoint()) { 748 SafePointNode* sfpt = use->as_SafePoint(); 749 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 750 // Object is passed as argument. 751 DEBUG_ONLY(disq_node = use;) 752 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 753 can_eliminate = false; 754 } 755 Node* sfptMem = sfpt->memory(); 756 if (sfptMem == NULL || sfptMem->is_top()) { 757 DEBUG_ONLY(disq_node = use;) 758 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) 759 can_eliminate = false; 760 } else { 761 safepoints.append_if_missing(sfpt); 762 } 763 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 764 if (use->is_Phi()) { 765 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 766 NOT_PRODUCT(fail_eliminate = "Object is return value";) 767 } else { 768 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 769 } 770 DEBUG_ONLY(disq_node = use;) 771 } else { 772 if (use->Opcode() == Op_Return) { 773 NOT_PRODUCT(fail_eliminate = "Object is return value";) 774 }else { 775 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 776 } 777 DEBUG_ONLY(disq_node = use;) 778 } 779 can_eliminate = false; 780 } 781 } 782 } 783 784 #ifndef PRODUCT 785 if (PrintEliminateAllocations) { 786 if (can_eliminate) { 787 tty->print("Scalar "); 788 if (res == NULL) 789 alloc->dump(); 790 else 791 res->dump(); 792 } else if (alloc->_is_scalar_replaceable) { 793 tty->print("NotScalar (%s)", fail_eliminate); 794 if (res == NULL) 795 alloc->dump(); 796 else 797 res->dump(); 798 #ifdef ASSERT 799 if (disq_node != NULL) { 800 tty->print(" >>>> "); 801 disq_node->dump(); 802 } 803 #endif /*ASSERT*/ 804 } 805 } 806 #endif 807 return can_eliminate; 808 } 809 810 // Do scalar replacement. 811 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 812 GrowableArray <SafePointNode *> safepoints_done; 813 814 ciKlass* klass = NULL; 815 ciInstanceKlass* iklass = NULL; 816 int nfields = 0; 817 int array_base = 0; 818 int element_size = 0; 819 BasicType basic_elem_type = T_ILLEGAL; 820 ciType* elem_type = NULL; 821 822 Node* res = alloc->result_cast(); 823 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result"); 824 const TypeOopPtr* res_type = NULL; 825 if (res != NULL) { // Could be NULL when there are no users 826 res_type = _igvn.type(res)->isa_oopptr(); 827 } 828 829 if (res != NULL) { 830 klass = res_type->klass(); 831 if (res_type->isa_instptr()) { 832 // find the fields of the class which will be needed for safepoint debug information 833 assert(klass->is_instance_klass(), "must be an instance klass."); 834 iklass = klass->as_instance_klass(); 835 nfields = iklass->nof_nonstatic_fields(); 836 } else { 837 // find the array's elements which will be needed for safepoint debug information 838 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 839 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); 840 elem_type = klass->as_array_klass()->element_type(); 841 basic_elem_type = elem_type->basic_type(); 842 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 843 element_size = type2aelembytes(basic_elem_type); 844 } 845 } 846 // 847 // Process the safepoint uses 848 // 849 while (safepoints.length() > 0) { 850 SafePointNode* sfpt = safepoints.pop(); 851 Node* mem = sfpt->memory(); 852 Node* ctl = sfpt->control(); 853 assert(sfpt->jvms() != NULL, "missed JVMS"); 854 // Fields of scalar objs are referenced only at the end 855 // of regular debuginfo at the last (youngest) JVMS. 856 // Record relative start index. 857 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff()); 858 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, 859 #ifdef ASSERT 860 alloc, 861 #endif 862 first_ind, nfields); 863 sobj->init_req(0, C->root()); 864 transform_later(sobj); 865 866 // Scan object's fields adding an input to the safepoint for each field. 867 for (int j = 0; j < nfields; j++) { 868 intptr_t offset; 869 ciField* field = NULL; 870 if (iklass != NULL) { 871 field = iklass->nonstatic_field_at(j); 872 offset = field->offset(); 873 elem_type = field->type(); 874 basic_elem_type = field->layout_type(); 875 } else { 876 offset = array_base + j * (intptr_t)element_size; 877 } 878 879 const Type *field_type; 880 // The next code is taken from Parse::do_get_xxx(). 881 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { 882 if (!elem_type->is_loaded()) { 883 field_type = TypeInstPtr::BOTTOM; 884 } else if (field != NULL && field->is_static_constant()) { 885 // This can happen if the constant oop is non-perm. 886 ciObject* con = field->constant_value().as_object(); 887 // Do not "join" in the previous type; it doesn't add value, 888 // and may yield a vacuous result if the field is of interface type. 889 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 890 assert(field_type != NULL, "field singleton type must be consistent"); 891 } else { 892 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 893 } 894 if (UseCompressedOops) { 895 field_type = field_type->make_narrowoop(); 896 basic_elem_type = T_NARROWOOP; 897 } 898 } else { 899 field_type = Type::get_const_basic_type(basic_elem_type); 900 } 901 902 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 903 904 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc); 905 if (field_val == NULL) { 906 // We weren't able to find a value for this field, 907 // give up on eliminating this allocation. 908 909 // Remove any extra entries we added to the safepoint. 910 uint last = sfpt->req() - 1; 911 for (int k = 0; k < j; k++) { 912 sfpt->del_req(last--); 913 } 914 _igvn._worklist.push(sfpt); 915 // rollback processed safepoints 916 while (safepoints_done.length() > 0) { 917 SafePointNode* sfpt_done = safepoints_done.pop(); 918 // remove any extra entries we added to the safepoint 919 last = sfpt_done->req() - 1; 920 for (int k = 0; k < nfields; k++) { 921 sfpt_done->del_req(last--); 922 } 923 JVMState *jvms = sfpt_done->jvms(); 924 jvms->set_endoff(sfpt_done->req()); 925 // Now make a pass over the debug information replacing any references 926 // to SafePointScalarObjectNode with the allocated object. 927 int start = jvms->debug_start(); 928 int end = jvms->debug_end(); 929 for (int i = start; i < end; i++) { 930 if (sfpt_done->in(i)->is_SafePointScalarObject()) { 931 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 932 if (scobj->first_index(jvms) == sfpt_done->req() && 933 scobj->n_fields() == (uint)nfields) { 934 assert(scobj->alloc() == alloc, "sanity"); 935 sfpt_done->set_req(i, res); 936 } 937 } 938 } 939 _igvn._worklist.push(sfpt_done); 940 } 941 #ifndef PRODUCT 942 if (PrintEliminateAllocations) { 943 if (field != NULL) { 944 tty->print("=== At SafePoint node %d can't find value of Field: ", 945 sfpt->_idx); 946 field->print(); 947 int field_idx = C->get_alias_index(field_addr_type); 948 tty->print(" (alias_idx=%d)", field_idx); 949 } else { // Array's element 950 tty->print("=== At SafePoint node %d can't find value of array element [%d]", 951 sfpt->_idx, j); 952 } 953 tty->print(", which prevents elimination of: "); 954 if (res == NULL) 955 alloc->dump(); 956 else 957 res->dump(); 958 } 959 #endif 960 return false; 961 } 962 if (UseCompressedOops && field_type->isa_narrowoop()) { 963 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 964 // to be able scalar replace the allocation. 965 if (field_val->is_EncodeP()) { 966 field_val = field_val->in(1); 967 } else { 968 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type())); 969 } 970 } 971 sfpt->add_req(field_val); 972 } 973 JVMState *jvms = sfpt->jvms(); 974 jvms->set_endoff(sfpt->req()); 975 // Now make a pass over the debug information replacing any references 976 // to the allocated object with "sobj" 977 int start = jvms->debug_start(); 978 int end = jvms->debug_end(); 979 sfpt->replace_edges_in_range(res, sobj, start, end); 980 _igvn._worklist.push(sfpt); 981 safepoints_done.append_if_missing(sfpt); // keep it for rollback 982 } 983 return true; 984 } 985 986 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) { 987 Node* ctl_proj = n->proj_out(TypeFunc::Control); 988 Node* mem_proj = n->proj_out(TypeFunc::Memory); 989 if (ctl_proj != NULL) { 990 igvn.replace_node(ctl_proj, n->in(0)); 991 } 992 if (mem_proj != NULL) { 993 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory)); 994 } 995 } 996 997 // Process users of eliminated allocation. 998 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) { 999 Node* res = alloc->result_cast(); 1000 if (res != NULL) { 1001 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 1002 Node *use = res->last_out(j); 1003 uint oc1 = res->outcnt(); 1004 1005 if (use->is_AddP()) { 1006 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 1007 Node *n = use->last_out(k); 1008 uint oc2 = use->outcnt(); 1009 if (n->is_Store()) { 1010 #ifdef ASSERT 1011 // Verify that there is no dependent MemBarVolatile nodes, 1012 // they should be removed during IGVN, see MemBarNode::Ideal(). 1013 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); 1014 p < pmax; p++) { 1015 Node* mb = n->fast_out(p); 1016 assert(mb->is_Initialize() || !mb->is_MemBar() || 1017 mb->req() <= MemBarNode::Precedent || 1018 mb->in(MemBarNode::Precedent) != n, 1019 "MemBarVolatile should be eliminated for non-escaping object"); 1020 } 1021 #endif 1022 _igvn.replace_node(n, n->in(MemNode::Memory)); 1023 } else if (n->is_ArrayCopy()) { 1024 // Disconnect ArrayCopy node 1025 ArrayCopyNode* ac = n->as_ArrayCopy(); 1026 assert(ac->is_clonebasic(), "unexpected array copy kind"); 1027 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out(); 1028 disconnect_projections(ac, _igvn); 1029 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation"); 1030 Node* membar_before = alloc->in(0)->in(0); 1031 disconnect_projections(membar_before->as_MemBar(), _igvn); 1032 if (membar_after->is_MemBar()) { 1033 disconnect_projections(membar_after->as_MemBar(), _igvn); 1034 } 1035 } else if (n->is_g1_wb_pre_call()) { 1036 eliminate_g1_wb_pre(n); 1037 } else { 1038 eliminate_card_mark(n); 1039 } 1040 k -= (oc2 - use->outcnt()); 1041 } 1042 _igvn.remove_dead_node(use); 1043 } else if (use->is_ArrayCopy()) { 1044 // Disconnect ArrayCopy node 1045 ArrayCopyNode* ac = use->as_ArrayCopy(); 1046 assert(ac->is_arraycopy_validated() || 1047 ac->is_copyof_validated() || 1048 ac->is_copyofrange_validated(), "unsupported"); 1049 CallProjections callprojs; 1050 ac->extract_projections(&callprojs, true); 1051 1052 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O)); 1053 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory)); 1054 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control)); 1055 1056 // Set control to top. IGVN will remove the remaining projections 1057 ac->set_req(0, top()); 1058 ac->replace_edge(res, top()); 1059 1060 // Disconnect src right away: it can help find new 1061 // opportunities for allocation elimination 1062 Node* src = ac->in(ArrayCopyNode::Src); 1063 ac->replace_edge(src, top()); 1064 if (src->outcnt() == 0) { 1065 _igvn.remove_dead_node(src); 1066 } 1067 1068 _igvn._worklist.push(ac); 1069 } else { 1070 eliminate_card_mark(use); 1071 } 1072 j -= (oc1 - res->outcnt()); 1073 } 1074 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 1075 _igvn.remove_dead_node(res); 1076 } 1077 1078 // 1079 // Process other users of allocation's projections 1080 // 1081 if (_resproj != NULL && _resproj->outcnt() != 0) { 1082 // First disconnect stores captured by Initialize node. 1083 // If Initialize node is eliminated first in the following code, 1084 // it will kill such stores and DUIterator_Last will assert. 1085 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) { 1086 Node *use = _resproj->fast_out(j); 1087 if (use->is_AddP()) { 1088 // raw memory addresses used only by the initialization 1089 _igvn.replace_node(use, C->top()); 1090 --j; --jmax; 1091 } 1092 } 1093 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { 1094 Node *use = _resproj->last_out(j); 1095 uint oc1 = _resproj->outcnt(); 1096 if (use->is_Initialize()) { 1097 // Eliminate Initialize node. 1098 InitializeNode *init = use->as_Initialize(); 1099 assert(init->outcnt() <= 2, "only a control and memory projection expected"); 1100 Node *ctrl_proj = init->proj_out(TypeFunc::Control); 1101 if (ctrl_proj != NULL) { 1102 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection"); 1103 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj); 1104 } 1105 Node *mem_proj = init->proj_out(TypeFunc::Memory); 1106 if (mem_proj != NULL) { 1107 Node *mem = init->in(TypeFunc::Memory); 1108 #ifdef ASSERT 1109 if (mem->is_MergeMem()) { 1110 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection"); 1111 } else { 1112 assert(mem == _memproj_fallthrough, "allocation memory projection"); 1113 } 1114 #endif 1115 _igvn.replace_node(mem_proj, mem); 1116 } 1117 } else { 1118 assert(false, "only Initialize or AddP expected"); 1119 } 1120 j -= (oc1 - _resproj->outcnt()); 1121 } 1122 } 1123 if (_fallthroughcatchproj != NULL) { 1124 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); 1125 } 1126 if (_memproj_fallthrough != NULL) { 1127 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); 1128 } 1129 if (_memproj_catchall != NULL) { 1130 _igvn.replace_node(_memproj_catchall, C->top()); 1131 } 1132 if (_ioproj_fallthrough != NULL) { 1133 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); 1134 } 1135 if (_ioproj_catchall != NULL) { 1136 _igvn.replace_node(_ioproj_catchall, C->top()); 1137 } 1138 if (_catchallcatchproj != NULL) { 1139 _igvn.replace_node(_catchallcatchproj, C->top()); 1140 } 1141 } 1142 1143 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 1144 // Don't do scalar replacement if the frame can be popped by JVMTI: 1145 // if reallocation fails during deoptimization we'll pop all 1146 // interpreter frames for this compiled frame and that won't play 1147 // nice with JVMTI popframe. 1148 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) { 1149 return false; 1150 } 1151 Node* klass = alloc->in(AllocateNode::KlassNode); 1152 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); 1153 Node* res = alloc->result_cast(); 1154 // Eliminate boxing allocations which are not used 1155 // regardless scalar replacable status. 1156 bool boxing_alloc = C->eliminate_boxing() && 1157 tklass->klass()->is_instance_klass() && 1158 tklass->klass()->as_instance_klass()->is_box_klass(); 1159 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) { 1160 return false; 1161 } 1162 1163 extract_call_projections(alloc); 1164 1165 GrowableArray <SafePointNode *> safepoints; 1166 if (!can_eliminate_allocation(alloc, safepoints)) { 1167 return false; 1168 } 1169 1170 if (!alloc->_is_scalar_replaceable) { 1171 assert(res == NULL, "sanity"); 1172 // We can only eliminate allocation if all debug info references 1173 // are already replaced with SafePointScalarObject because 1174 // we can't search for a fields value without instance_id. 1175 if (safepoints.length() > 0) { 1176 return false; 1177 } 1178 } 1179 1180 if (!scalar_replacement(alloc, safepoints)) { 1181 return false; 1182 } 1183 1184 CompileLog* log = C->log(); 1185 if (log != NULL) { 1186 log->head("eliminate_allocation type='%d'", 1187 log->identify(tklass->klass())); 1188 JVMState* p = alloc->jvms(); 1189 while (p != NULL) { 1190 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1191 p = p->caller(); 1192 } 1193 log->tail("eliminate_allocation"); 1194 } 1195 1196 process_users_of_allocation(alloc); 1197 1198 #ifndef PRODUCT 1199 if (PrintEliminateAllocations) { 1200 if (alloc->is_AllocateArray()) 1201 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 1202 else 1203 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 1204 } 1205 #endif 1206 1207 return true; 1208 } 1209 1210 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) { 1211 // EA should remove all uses of non-escaping boxing node. 1212 if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) { 1213 return false; 1214 } 1215 1216 assert(boxing->result_cast() == NULL, "unexpected boxing node result"); 1217 1218 extract_call_projections(boxing); 1219 1220 const TypeTuple* r = boxing->tf()->range(); 1221 assert(r->cnt() > TypeFunc::Parms, "sanity"); 1222 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr(); 1223 assert(t != NULL, "sanity"); 1224 1225 CompileLog* log = C->log(); 1226 if (log != NULL) { 1227 log->head("eliminate_boxing type='%d'", 1228 log->identify(t->klass())); 1229 JVMState* p = boxing->jvms(); 1230 while (p != NULL) { 1231 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 1232 p = p->caller(); 1233 } 1234 log->tail("eliminate_boxing"); 1235 } 1236 1237 process_users_of_allocation(boxing); 1238 1239 #ifndef PRODUCT 1240 if (PrintEliminateAllocations) { 1241 tty->print("++++ Eliminated: %d ", boxing->_idx); 1242 boxing->method()->print_short_name(tty); 1243 tty->cr(); 1244 } 1245 #endif 1246 1247 return true; 1248 } 1249 1250 //---------------------------set_eden_pointers------------------------- 1251 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { 1252 if (UseTLAB) { // Private allocation: load from TLS 1253 Node* thread = transform_later(new ThreadLocalNode()); 1254 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); 1255 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); 1256 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); 1257 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); 1258 } else { // Shared allocation: load from globals 1259 CollectedHeap* ch = Universe::heap(); 1260 address top_adr = (address)ch->top_addr(); 1261 address end_adr = (address)ch->end_addr(); 1262 eden_top_adr = makecon(TypeRawPtr::make(top_adr)); 1263 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); 1264 } 1265 } 1266 1267 1268 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 1269 Node* adr = basic_plus_adr(base, offset); 1270 const TypePtr* adr_type = adr->bottom_type()->is_ptr(); 1271 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered); 1272 transform_later(value); 1273 return value; 1274 } 1275 1276 1277 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 1278 Node* adr = basic_plus_adr(base, offset); 1279 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered); 1280 transform_later(mem); 1281 return mem; 1282 } 1283 1284 //============================================================================= 1285 // 1286 // A L L O C A T I O N 1287 // 1288 // Allocation attempts to be fast in the case of frequent small objects. 1289 // It breaks down like this: 1290 // 1291 // 1) Size in doublewords is computed. This is a constant for objects and 1292 // variable for most arrays. Doubleword units are used to avoid size 1293 // overflow of huge doubleword arrays. We need doublewords in the end for 1294 // rounding. 1295 // 1296 // 2) Size is checked for being 'too large'. Too-large allocations will go 1297 // the slow path into the VM. The slow path can throw any required 1298 // exceptions, and does all the special checks for very large arrays. The 1299 // size test can constant-fold away for objects. For objects with 1300 // finalizers it constant-folds the otherway: you always go slow with 1301 // finalizers. 1302 // 1303 // 3) If NOT using TLABs, this is the contended loop-back point. 1304 // Load-Locked the heap top. If using TLABs normal-load the heap top. 1305 // 1306 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 1307 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 1308 // "size*8" we always enter the VM, where "largish" is a constant picked small 1309 // enough that there's always space between the eden max and 4Gig (old space is 1310 // there so it's quite large) and large enough that the cost of entering the VM 1311 // is dwarfed by the cost to initialize the space. 1312 // 1313 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 1314 // down. If contended, repeat at step 3. If using TLABs normal-store 1315 // adjusted heap top back down; there is no contention. 1316 // 1317 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 1318 // fields. 1319 // 1320 // 7) Merge with the slow-path; cast the raw memory pointer to the correct 1321 // oop flavor. 1322 // 1323 //============================================================================= 1324 // FastAllocateSizeLimit value is in DOUBLEWORDS. 1325 // Allocations bigger than this always go the slow route. 1326 // This value must be small enough that allocation attempts that need to 1327 // trigger exceptions go the slow route. Also, it must be small enough so 1328 // that heap_top + size_in_bytes does not wrap around the 4Gig limit. 1329 //=============================================================================j// 1330 // %%% Here is an old comment from parseHelper.cpp; is it outdated? 1331 // The allocator will coalesce int->oop copies away. See comment in 1332 // coalesce.cpp about how this works. It depends critically on the exact 1333 // code shape produced here, so if you are changing this code shape 1334 // make sure the GC info for the heap-top is correct in and around the 1335 // slow-path call. 1336 // 1337 1338 void PhaseMacroExpand::expand_allocate_common( 1339 AllocateNode* alloc, // allocation node to be expanded 1340 Node* length, // array length for an array allocation 1341 const TypeFunc* slow_call_type, // Type of slow call 1342 address slow_call_address // Address of slow call 1343 ) 1344 { 1345 1346 Node* ctrl = alloc->in(TypeFunc::Control); 1347 Node* mem = alloc->in(TypeFunc::Memory); 1348 Node* i_o = alloc->in(TypeFunc::I_O); 1349 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 1350 Node* klass_node = alloc->in(AllocateNode::KlassNode); 1351 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 1352 1353 assert(ctrl != NULL, "must have control"); 1354 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 1355 // they will not be used if "always_slow" is set 1356 enum { slow_result_path = 1, fast_result_path = 2 }; 1357 Node *result_region = NULL; 1358 Node *result_phi_rawmem = NULL; 1359 Node *result_phi_rawoop = NULL; 1360 Node *result_phi_i_o = NULL; 1361 1362 // The initial slow comparison is a size check, the comparison 1363 // we want to do is a BoolTest::gt 1364 bool always_slow = false; 1365 int tv = _igvn.find_int_con(initial_slow_test, -1); 1366 if (tv >= 0) { 1367 always_slow = (tv == 1); 1368 initial_slow_test = NULL; 1369 } else { 1370 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 1371 } 1372 1373 if (C->env()->dtrace_alloc_probes() || 1374 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc())) { 1375 // Force slow-path allocation 1376 always_slow = true; 1377 initial_slow_test = NULL; 1378 } 1379 1380 1381 enum { too_big_or_final_path = 1, need_gc_path = 2 }; 1382 Node *slow_region = NULL; 1383 Node *toobig_false = ctrl; 1384 1385 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); 1386 // generate the initial test if necessary 1387 if (initial_slow_test != NULL ) { 1388 slow_region = new RegionNode(3); 1389 1390 // Now make the initial failure test. Usually a too-big test but 1391 // might be a TRUE for finalizers or a fancy class check for 1392 // newInstance0. 1393 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 1394 transform_later(toobig_iff); 1395 // Plug the failing-too-big test into the slow-path region 1396 Node *toobig_true = new IfTrueNode( toobig_iff ); 1397 transform_later(toobig_true); 1398 slow_region ->init_req( too_big_or_final_path, toobig_true ); 1399 toobig_false = new IfFalseNode( toobig_iff ); 1400 transform_later(toobig_false); 1401 } else { // No initial test, just fall into next case 1402 toobig_false = ctrl; 1403 debug_only(slow_region = NodeSentinel); 1404 } 1405 1406 Node *slow_mem = mem; // save the current memory state for slow path 1407 // generate the fast allocation code unless we know that the initial test will always go slow 1408 if (!always_slow) { 1409 // Fast path modifies only raw memory. 1410 if (mem->is_MergeMem()) { 1411 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1412 } 1413 1414 Node* eden_top_adr; 1415 Node* eden_end_adr; 1416 1417 set_eden_pointers(eden_top_adr, eden_end_adr); 1418 1419 // Load Eden::end. Loop invariant and hoisted. 1420 // 1421 // Note: We set the control input on "eden_end" and "old_eden_top" when using 1422 // a TLAB to work around a bug where these values were being moved across 1423 // a safepoint. These are not oops, so they cannot be include in the oop 1424 // map, but they can be changed by a GC. The proper way to fix this would 1425 // be to set the raw memory state when generating a SafepointNode. However 1426 // this will require extensive changes to the loop optimization in order to 1427 // prevent a degradation of the optimization. 1428 // See comment in memnode.hpp, around line 227 in class LoadPNode. 1429 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 1430 1431 // allocate the Region and Phi nodes for the result 1432 result_region = new RegionNode(3); 1433 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1434 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM); 1435 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch 1436 1437 // We need a Region for the loop-back contended case. 1438 enum { fall_in_path = 1, contended_loopback_path = 2 }; 1439 Node *contended_region; 1440 Node *contended_phi_rawmem; 1441 if (UseTLAB) { 1442 contended_region = toobig_false; 1443 contended_phi_rawmem = mem; 1444 } else { 1445 contended_region = new RegionNode(3); 1446 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1447 // Now handle the passing-too-big test. We fall into the contended 1448 // loop-back merge point. 1449 contended_region ->init_req(fall_in_path, toobig_false); 1450 contended_phi_rawmem->init_req(fall_in_path, mem); 1451 transform_later(contended_region); 1452 transform_later(contended_phi_rawmem); 1453 } 1454 1455 // Load(-locked) the heap top. 1456 // See note above concerning the control input when using a TLAB 1457 Node *old_eden_top = UseTLAB 1458 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered) 1459 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire); 1460 1461 transform_later(old_eden_top); 1462 // Add to heap top to get a new heap top 1463 1464 Node* init_size_in_bytes = size_in_bytes; 1465 if (UseShenandoahGC) { 1466 // Allocate several words more for the Shenandoah brooks pointer. 1467 size_in_bytes = new AddLNode(size_in_bytes, _igvn.MakeConX(BrooksPointer::byte_size())); 1468 transform_later(size_in_bytes); 1469 } 1470 1471 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes); 1472 transform_later(new_eden_top); 1473 // Check for needing a GC; compare against heap end 1474 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end); 1475 transform_later(needgc_cmp); 1476 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge); 1477 transform_later(needgc_bol); 1478 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); 1479 transform_later(needgc_iff); 1480 1481 // Plug the failing-heap-space-need-gc test into the slow-path region 1482 Node *needgc_true = new IfTrueNode(needgc_iff); 1483 transform_later(needgc_true); 1484 if (initial_slow_test) { 1485 slow_region->init_req(need_gc_path, needgc_true); 1486 // This completes all paths into the slow merge point 1487 transform_later(slow_region); 1488 } else { // No initial slow path needed! 1489 // Just fall from the need-GC path straight into the VM call. 1490 slow_region = needgc_true; 1491 } 1492 // No need for a GC. Setup for the Store-Conditional 1493 Node *needgc_false = new IfFalseNode(needgc_iff); 1494 transform_later(needgc_false); 1495 1496 // Grab regular I/O before optional prefetch may change it. 1497 // Slow-path does no I/O so just set it to the original I/O. 1498 result_phi_i_o->init_req(slow_result_path, i_o); 1499 1500 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem, 1501 old_eden_top, new_eden_top, length); 1502 1503 // Name successful fast-path variables 1504 Node* fast_oop = old_eden_top; 1505 Node* fast_oop_ctrl; 1506 Node* fast_oop_rawmem; 1507 1508 // Store (-conditional) the modified eden top back down. 1509 // StorePConditional produces flags for a test PLUS a modified raw 1510 // memory state. 1511 if (UseTLAB) { 1512 Node* store_eden_top = 1513 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1514 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered); 1515 transform_later(store_eden_top); 1516 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path 1517 fast_oop_rawmem = store_eden_top; 1518 } else { 1519 Node* store_eden_top = 1520 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr, 1521 new_eden_top, fast_oop/*old_eden_top*/); 1522 transform_later(store_eden_top); 1523 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne); 1524 transform_later(contention_check); 1525 store_eden_top = new SCMemProjNode(store_eden_top); 1526 transform_later(store_eden_top); 1527 1528 // If not using TLABs, check to see if there was contention. 1529 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN); 1530 transform_later(contention_iff); 1531 Node *contention_true = new IfTrueNode(contention_iff); 1532 transform_later(contention_true); 1533 // If contention, loopback and try again. 1534 contended_region->init_req(contended_loopback_path, contention_true); 1535 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top); 1536 1537 // Fast-path succeeded with no contention! 1538 Node *contention_false = new IfFalseNode(contention_iff); 1539 transform_later(contention_false); 1540 fast_oop_ctrl = contention_false; 1541 1542 // Bump total allocated bytes for this thread 1543 Node* thread = new ThreadLocalNode(); 1544 transform_later(thread); 1545 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread, 1546 in_bytes(JavaThread::allocated_bytes_offset())); 1547 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1548 0, TypeLong::LONG, T_LONG); 1549 #ifdef _LP64 1550 Node* alloc_size = size_in_bytes; 1551 #else 1552 Node* alloc_size = new ConvI2LNode(size_in_bytes); 1553 transform_later(alloc_size); 1554 #endif 1555 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size); 1556 transform_later(new_alloc_bytes); 1557 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, 1558 0, new_alloc_bytes, T_LONG); 1559 } 1560 1561 if (UseShenandoahGC) { 1562 // Bump up object for Shenandoah brooks pointer. 1563 fast_oop = new AddPNode(top(), fast_oop, _igvn.MakeConX(BrooksPointer::byte_size())); 1564 transform_later(fast_oop); 1565 } 1566 1567 InitializeNode* init = alloc->initialization(); 1568 fast_oop_rawmem = initialize_object(alloc, 1569 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1570 klass_node, length, init_size_in_bytes); 1571 1572 // If initialization is performed by an array copy, any required 1573 // MemBarStoreStore was already added. If the object does not 1574 // escape no need for a MemBarStoreStore. If the object does not 1575 // escape in its initializer and memory barrier (MemBarStoreStore or 1576 // stronger) is already added at exit of initializer, also no need 1577 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore 1578 // so that stores that initialize this object can't be reordered 1579 // with a subsequent store that makes this object accessible by 1580 // other threads. 1581 // Other threads include java threads and JVM internal threads 1582 // (for example concurrent GC threads). Current concurrent GC 1583 // implementation: CMS and G1 will not scan newly created object, 1584 // so it's safe to skip storestore barrier when allocation does 1585 // not escape. 1586 if (!alloc->does_not_escape_thread() && 1587 !alloc->is_allocation_MemBar_redundant() && 1588 (init == NULL || !init->is_complete_with_arraycopy())) { 1589 if (init == NULL || init->req() < InitializeNode::RawStores) { 1590 // No InitializeNode or no stores captured by zeroing 1591 // elimination. Simply add the MemBarStoreStore after object 1592 // initialization. 1593 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1594 transform_later(mb); 1595 1596 mb->init_req(TypeFunc::Memory, fast_oop_rawmem); 1597 mb->init_req(TypeFunc::Control, fast_oop_ctrl); 1598 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control); 1599 transform_later(fast_oop_ctrl); 1600 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory); 1601 transform_later(fast_oop_rawmem); 1602 } else { 1603 // Add the MemBarStoreStore after the InitializeNode so that 1604 // all stores performing the initialization that were moved 1605 // before the InitializeNode happen before the storestore 1606 // barrier. 1607 1608 Node* init_ctrl = init->proj_out(TypeFunc::Control); 1609 Node* init_mem = init->proj_out(TypeFunc::Memory); 1610 1611 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); 1612 transform_later(mb); 1613 1614 Node* ctrl = new ProjNode(init,TypeFunc::Control); 1615 transform_later(ctrl); 1616 Node* mem = new ProjNode(init,TypeFunc::Memory); 1617 transform_later(mem); 1618 1619 // The MemBarStoreStore depends on control and memory coming 1620 // from the InitializeNode 1621 mb->init_req(TypeFunc::Memory, mem); 1622 mb->init_req(TypeFunc::Control, ctrl); 1623 1624 ctrl = new ProjNode(mb,TypeFunc::Control); 1625 transform_later(ctrl); 1626 mem = new ProjNode(mb,TypeFunc::Memory); 1627 transform_later(mem); 1628 1629 // All nodes that depended on the InitializeNode for control 1630 // and memory must now depend on the MemBarNode that itself 1631 // depends on the InitializeNode 1632 if (init_ctrl != NULL) { 1633 _igvn.replace_node(init_ctrl, ctrl); 1634 } 1635 if (init_mem != NULL) { 1636 _igvn.replace_node(init_mem, mem); 1637 } 1638 } 1639 } 1640 1641 if (C->env()->dtrace_extended_probes()) { 1642 // Slow-path call 1643 int size = TypeFunc::Parms + 2; 1644 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1645 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), 1646 "dtrace_object_alloc", 1647 TypeRawPtr::BOTTOM); 1648 1649 // Get base of thread-local storage area 1650 Node* thread = new ThreadLocalNode(); 1651 transform_later(thread); 1652 1653 call->init_req(TypeFunc::Parms+0, thread); 1654 call->init_req(TypeFunc::Parms+1, fast_oop); 1655 call->init_req(TypeFunc::Control, fast_oop_ctrl); 1656 call->init_req(TypeFunc::I_O , top()); // does no i/o 1657 call->init_req(TypeFunc::Memory , fast_oop_rawmem); 1658 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); 1659 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); 1660 transform_later(call); 1661 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control); 1662 transform_later(fast_oop_ctrl); 1663 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory); 1664 transform_later(fast_oop_rawmem); 1665 } 1666 1667 // Plug in the successful fast-path into the result merge point 1668 result_region ->init_req(fast_result_path, fast_oop_ctrl); 1669 result_phi_rawoop->init_req(fast_result_path, fast_oop); 1670 result_phi_i_o ->init_req(fast_result_path, i_o); 1671 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); 1672 } else { 1673 slow_region = ctrl; 1674 result_phi_i_o = i_o; // Rename it to use in the following code. 1675 } 1676 1677 // Generate slow-path call 1678 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address, 1679 OptoRuntime::stub_name(slow_call_address), 1680 alloc->jvms()->bci(), 1681 TypePtr::BOTTOM); 1682 call->init_req( TypeFunc::Control, slow_region ); 1683 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1684 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs 1685 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1686 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1687 1688 call->init_req(TypeFunc::Parms+0, klass_node); 1689 if (length != NULL) { 1690 call->init_req(TypeFunc::Parms+1, length); 1691 } 1692 1693 // Copy debug information and adjust JVMState information, then replace 1694 // allocate node with the call 1695 copy_call_debug_info((CallNode *) alloc, call); 1696 if (!always_slow) { 1697 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1698 } else { 1699 // Hook i_o projection to avoid its elimination during allocation 1700 // replacement (when only a slow call is generated). 1701 call->set_req(TypeFunc::I_O, result_phi_i_o); 1702 } 1703 _igvn.replace_node(alloc, call); 1704 transform_later(call); 1705 1706 // Identify the output projections from the allocate node and 1707 // adjust any references to them. 1708 // The control and io projections look like: 1709 // 1710 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1711 // Allocate Catch 1712 // ^---Proj(io) <-------+ ^---CatchProj(io) 1713 // 1714 // We are interested in the CatchProj nodes. 1715 // 1716 extract_call_projections(call); 1717 1718 // An allocate node has separate memory projections for the uses on 1719 // the control and i_o paths. Replace the control memory projection with 1720 // result_phi_rawmem (unless we are only generating a slow call when 1721 // both memory projections are combined) 1722 if (!always_slow && _memproj_fallthrough != NULL) { 1723 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { 1724 Node *use = _memproj_fallthrough->fast_out(i); 1725 _igvn.rehash_node_delayed(use); 1726 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); 1727 // back up iterator 1728 --i; 1729 } 1730 } 1731 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete 1732 // _memproj_catchall so we end up with a call that has only 1 memory projection. 1733 if (_memproj_catchall != NULL ) { 1734 if (_memproj_fallthrough == NULL) { 1735 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory); 1736 transform_later(_memproj_fallthrough); 1737 } 1738 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { 1739 Node *use = _memproj_catchall->fast_out(i); 1740 _igvn.rehash_node_delayed(use); 1741 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); 1742 // back up iterator 1743 --i; 1744 } 1745 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted"); 1746 _igvn.remove_dead_node(_memproj_catchall); 1747 } 1748 1749 // An allocate node has separate i_o projections for the uses on the control 1750 // and i_o paths. Always replace the control i_o projection with result i_o 1751 // otherwise incoming i_o become dead when only a slow call is generated 1752 // (it is different from memory projections where both projections are 1753 // combined in such case). 1754 if (_ioproj_fallthrough != NULL) { 1755 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { 1756 Node *use = _ioproj_fallthrough->fast_out(i); 1757 _igvn.rehash_node_delayed(use); 1758 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); 1759 // back up iterator 1760 --i; 1761 } 1762 } 1763 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete 1764 // _ioproj_catchall so we end up with a call that has only 1 i_o projection. 1765 if (_ioproj_catchall != NULL ) { 1766 if (_ioproj_fallthrough == NULL) { 1767 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O); 1768 transform_later(_ioproj_fallthrough); 1769 } 1770 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { 1771 Node *use = _ioproj_catchall->fast_out(i); 1772 _igvn.rehash_node_delayed(use); 1773 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); 1774 // back up iterator 1775 --i; 1776 } 1777 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted"); 1778 _igvn.remove_dead_node(_ioproj_catchall); 1779 } 1780 1781 // if we generated only a slow call, we are done 1782 if (always_slow) { 1783 // Now we can unhook i_o. 1784 if (result_phi_i_o->outcnt() > 1) { 1785 call->set_req(TypeFunc::I_O, top()); 1786 } else { 1787 assert(result_phi_i_o->unique_ctrl_out() == call, ""); 1788 // Case of new array with negative size known during compilation. 1789 // AllocateArrayNode::Ideal() optimization disconnect unreachable 1790 // following code since call to runtime will throw exception. 1791 // As result there will be no users of i_o after the call. 1792 // Leave i_o attached to this call to avoid problems in preceding graph. 1793 } 1794 return; 1795 } 1796 1797 1798 if (_fallthroughcatchproj != NULL) { 1799 ctrl = _fallthroughcatchproj->clone(); 1800 transform_later(ctrl); 1801 _igvn.replace_node(_fallthroughcatchproj, result_region); 1802 } else { 1803 ctrl = top(); 1804 } 1805 Node *slow_result; 1806 if (_resproj == NULL) { 1807 // no uses of the allocation result 1808 slow_result = top(); 1809 } else { 1810 slow_result = _resproj->clone(); 1811 transform_later(slow_result); 1812 _igvn.replace_node(_resproj, result_phi_rawoop); 1813 } 1814 1815 // Plug slow-path into result merge point 1816 result_region ->init_req( slow_result_path, ctrl ); 1817 result_phi_rawoop->init_req( slow_result_path, slow_result); 1818 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); 1819 transform_later(result_region); 1820 transform_later(result_phi_rawoop); 1821 transform_later(result_phi_rawmem); 1822 transform_later(result_phi_i_o); 1823 // This completes all paths into the result merge point 1824 } 1825 1826 1827 // Helper for PhaseMacroExpand::expand_allocate_common. 1828 // Initializes the newly-allocated storage. 1829 Node* 1830 PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1831 Node* control, Node* rawmem, Node* object, 1832 Node* klass_node, Node* length, 1833 Node* size_in_bytes) { 1834 InitializeNode* init = alloc->initialization(); 1835 // Store the klass & mark bits 1836 Node* mark_node = NULL; 1837 // For now only enable fast locking for non-array types 1838 if (UseBiasedLocking && (length == NULL)) { 1839 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); 1840 } else { 1841 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); 1842 } 1843 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 1844 1845 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); 1846 int header_size = alloc->minimum_header_size(); // conservatively small 1847 1848 // Array length 1849 if (length != NULL) { // Arrays need length field 1850 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1851 // conservatively small header size: 1852 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1853 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1854 if (k->is_array_klass()) // we know the exact header size in most cases: 1855 header_size = Klass::layout_helper_header_size(k->layout_helper()); 1856 } 1857 1858 if (UseShenandoahGC) { 1859 // Initialize Shenandoah brooks pointer to point to the object itself. 1860 rawmem = make_store(control, rawmem, object, BrooksPointer::byte_offset(), object, T_OBJECT); 1861 } 1862 1863 // Clear the object body, if necessary. 1864 if (init == NULL) { 1865 // The init has somehow disappeared; be cautious and clear everything. 1866 // 1867 // This can happen if a node is allocated but an uncommon trap occurs 1868 // immediately. In this case, the Initialize gets associated with the 1869 // trap, and may be placed in a different (outer) loop, if the Allocate 1870 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1871 // there can be two Allocates to one Initialize. The answer in all these 1872 // edge cases is safety first. It is always safe to clear immediately 1873 // within an Allocate, and then (maybe or maybe not) clear some more later. 1874 if (!(UseTLAB && ZeroTLAB)) { 1875 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1876 header_size, size_in_bytes, 1877 &_igvn); 1878 } 1879 } else { 1880 if (!init->is_complete()) { 1881 // Try to win by zeroing only what the init does not store. 1882 // We can also try to do some peephole optimizations, 1883 // such as combining some adjacent subword stores. 1884 rawmem = init->complete_stores(control, rawmem, object, 1885 header_size, size_in_bytes, &_igvn); 1886 } 1887 // We have no more use for this link, since the AllocateNode goes away: 1888 init->set_req(InitializeNode::RawAddress, top()); 1889 // (If we keep the link, it just confuses the register allocator, 1890 // who thinks he sees a real use of the address by the membar.) 1891 } 1892 1893 return rawmem; 1894 } 1895 1896 // Generate prefetch instructions for next allocations. 1897 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1898 Node*& contended_phi_rawmem, 1899 Node* old_eden_top, Node* new_eden_top, 1900 Node* length) { 1901 enum { fall_in_path = 1, pf_path = 2 }; 1902 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1903 // Generate prefetch allocation with watermark check. 1904 // As an allocation hits the watermark, we will prefetch starting 1905 // at a "distance" away from watermark. 1906 1907 Node *pf_region = new RegionNode(3); 1908 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, 1909 TypeRawPtr::BOTTOM ); 1910 // I/O is used for Prefetch 1911 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO ); 1912 1913 Node *thread = new ThreadLocalNode(); 1914 transform_later(thread); 1915 1916 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread, 1917 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1918 transform_later(eden_pf_adr); 1919 1920 Node *old_pf_wm = new LoadPNode(needgc_false, 1921 contended_phi_rawmem, eden_pf_adr, 1922 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, 1923 MemNode::unordered); 1924 transform_later(old_pf_wm); 1925 1926 // check against new_eden_top 1927 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm ); 1928 transform_later(need_pf_cmp); 1929 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge ); 1930 transform_later(need_pf_bol); 1931 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol, 1932 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1933 transform_later(need_pf_iff); 1934 1935 // true node, add prefetchdistance 1936 Node *need_pf_true = new IfTrueNode( need_pf_iff ); 1937 transform_later(need_pf_true); 1938 1939 Node *need_pf_false = new IfFalseNode( need_pf_iff ); 1940 transform_later(need_pf_false); 1941 1942 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm, 1943 _igvn.MakeConX(AllocatePrefetchDistance) ); 1944 transform_later(new_pf_wmt ); 1945 new_pf_wmt->set_req(0, need_pf_true); 1946 1947 Node *store_new_wmt = new StorePNode(need_pf_true, 1948 contended_phi_rawmem, eden_pf_adr, 1949 TypeRawPtr::BOTTOM, new_pf_wmt, 1950 MemNode::unordered); 1951 transform_later(store_new_wmt); 1952 1953 // adding prefetches 1954 pf_phi_abio->init_req( fall_in_path, i_o ); 1955 1956 Node *prefetch_adr; 1957 Node *prefetch; 1958 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1959 uint step_size = AllocatePrefetchStepSize; 1960 uint distance = 0; 1961 1962 for ( uint i = 0; i < lines; i++ ) { 1963 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt, 1964 _igvn.MakeConX(distance) ); 1965 transform_later(prefetch_adr); 1966 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 1967 transform_later(prefetch); 1968 distance += step_size; 1969 i_o = prefetch; 1970 } 1971 pf_phi_abio->set_req( pf_path, i_o ); 1972 1973 pf_region->init_req( fall_in_path, need_pf_false ); 1974 pf_region->init_req( pf_path, need_pf_true ); 1975 1976 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 1977 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 1978 1979 transform_later(pf_region); 1980 transform_later(pf_phi_rawmem); 1981 transform_later(pf_phi_abio); 1982 1983 needgc_false = pf_region; 1984 contended_phi_rawmem = pf_phi_rawmem; 1985 i_o = pf_phi_abio; 1986 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { 1987 // Insert a prefetch instruction for each allocation. 1988 // This code is used to generate 1 prefetch instruction per cache line. 1989 1990 // Generate several prefetch instructions. 1991 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 1992 uint step_size = AllocatePrefetchStepSize; 1993 uint distance = AllocatePrefetchDistance; 1994 1995 // Next cache address. 1996 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top, 1997 _igvn.MakeConX(step_size + distance)); 1998 transform_later(cache_adr); 1999 cache_adr = new CastP2XNode(needgc_false, cache_adr); 2000 transform_later(cache_adr); 2001 // Address is aligned to execute prefetch to the beginning of cache line size 2002 // (it is important when BIS instruction is used on SPARC as prefetch). 2003 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); 2004 cache_adr = new AndXNode(cache_adr, mask); 2005 transform_later(cache_adr); 2006 cache_adr = new CastX2PNode(cache_adr); 2007 transform_later(cache_adr); 2008 2009 // Prefetch 2010 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); 2011 prefetch->set_req(0, needgc_false); 2012 transform_later(prefetch); 2013 contended_phi_rawmem = prefetch; 2014 Node *prefetch_adr; 2015 distance = step_size; 2016 for ( uint i = 1; i < lines; i++ ) { 2017 prefetch_adr = new AddPNode( cache_adr, cache_adr, 2018 _igvn.MakeConX(distance) ); 2019 transform_later(prefetch_adr); 2020 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); 2021 transform_later(prefetch); 2022 distance += step_size; 2023 contended_phi_rawmem = prefetch; 2024 } 2025 } else if( AllocatePrefetchStyle > 0 ) { 2026 // Insert a prefetch for each allocation only on the fast-path 2027 Node *prefetch_adr; 2028 Node *prefetch; 2029 // Generate several prefetch instructions. 2030 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; 2031 uint step_size = AllocatePrefetchStepSize; 2032 uint distance = AllocatePrefetchDistance; 2033 for ( uint i = 0; i < lines; i++ ) { 2034 prefetch_adr = new AddPNode( old_eden_top, new_eden_top, 2035 _igvn.MakeConX(distance) ); 2036 transform_later(prefetch_adr); 2037 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); 2038 // Do not let it float too high, since if eden_top == eden_end, 2039 // both might be null. 2040 if( i == 0 ) { // Set control for first prefetch, next follows it 2041 prefetch->init_req(0, needgc_false); 2042 } 2043 transform_later(prefetch); 2044 distance += step_size; 2045 i_o = prefetch; 2046 } 2047 } 2048 return i_o; 2049 } 2050 2051 2052 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 2053 expand_allocate_common(alloc, NULL, 2054 OptoRuntime::new_instance_Type(), 2055 OptoRuntime::new_instance_Java()); 2056 } 2057 2058 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 2059 Node* length = alloc->in(AllocateNode::ALength); 2060 InitializeNode* init = alloc->initialization(); 2061 Node* klass_node = alloc->in(AllocateNode::KlassNode); 2062 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 2063 address slow_call_address; // Address of slow call 2064 if (init != NULL && init->is_complete_with_arraycopy() && 2065 k->is_type_array_klass()) { 2066 // Don't zero type array during slow allocation in VM since 2067 // it will be initialized later by arraycopy in compiled code. 2068 slow_call_address = OptoRuntime::new_array_nozero_Java(); 2069 } else { 2070 slow_call_address = OptoRuntime::new_array_Java(); 2071 } 2072 expand_allocate_common(alloc, length, 2073 OptoRuntime::new_array_Type(), 2074 slow_call_address); 2075 } 2076 2077 //-------------------mark_eliminated_box---------------------------------- 2078 // 2079 // During EA obj may point to several objects but after few ideal graph 2080 // transformations (CCP) it may point to only one non escaping object 2081 // (but still using phi), corresponding locks and unlocks will be marked 2082 // for elimination. Later obj could be replaced with a new node (new phi) 2083 // and which does not have escape information. And later after some graph 2084 // reshape other locks and unlocks (which were not marked for elimination 2085 // before) are connected to this new obj (phi) but they still will not be 2086 // marked for elimination since new obj has no escape information. 2087 // Mark all associated (same box and obj) lock and unlock nodes for 2088 // elimination if some of them marked already. 2089 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) { 2090 if (oldbox->as_BoxLock()->is_eliminated()) 2091 return; // This BoxLock node was processed already. 2092 2093 // New implementation (EliminateNestedLocks) has separate BoxLock 2094 // node for each locked region so mark all associated locks/unlocks as 2095 // eliminated even if different objects are referenced in one locked region 2096 // (for example, OSR compilation of nested loop inside locked scope). 2097 if (EliminateNestedLocks || 2098 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) { 2099 // Box is used only in one lock region. Mark this box as eliminated. 2100 _igvn.hash_delete(oldbox); 2101 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value 2102 _igvn.hash_insert(oldbox); 2103 2104 for (uint i = 0; i < oldbox->outcnt(); i++) { 2105 Node* u = oldbox->raw_out(i); 2106 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) { 2107 AbstractLockNode* alock = u->as_AbstractLock(); 2108 // Check lock's box since box could be referenced by Lock's debug info. 2109 if (alock->box_node() == oldbox) { 2110 // Mark eliminated all related locks and unlocks. 2111 #ifdef ASSERT 2112 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4"); 2113 #endif 2114 alock->set_non_esc_obj(); 2115 } 2116 } 2117 } 2118 return; 2119 } 2120 2121 // Create new "eliminated" BoxLock node and use it in monitor debug info 2122 // instead of oldbox for the same object. 2123 BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); 2124 2125 // Note: BoxLock node is marked eliminated only here and it is used 2126 // to indicate that all associated lock and unlock nodes are marked 2127 // for elimination. 2128 newbox->set_eliminated(); 2129 transform_later(newbox); 2130 2131 // Replace old box node with new box for all users of the same object. 2132 for (uint i = 0; i < oldbox->outcnt();) { 2133 bool next_edge = true; 2134 2135 Node* u = oldbox->raw_out(i); 2136 if (u->is_AbstractLock()) { 2137 AbstractLockNode* alock = u->as_AbstractLock(); 2138 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) { 2139 // Replace Box and mark eliminated all related locks and unlocks. 2140 #ifdef ASSERT 2141 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5"); 2142 #endif 2143 alock->set_non_esc_obj(); 2144 _igvn.rehash_node_delayed(alock); 2145 alock->set_box_node(newbox); 2146 next_edge = false; 2147 } 2148 } 2149 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) { 2150 FastLockNode* flock = u->as_FastLock(); 2151 assert(flock->box_node() == oldbox, "sanity"); 2152 _igvn.rehash_node_delayed(flock); 2153 flock->set_box_node(newbox); 2154 next_edge = false; 2155 } 2156 2157 // Replace old box in monitor debug info. 2158 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { 2159 SafePointNode* sfn = u->as_SafePoint(); 2160 JVMState* youngest_jvms = sfn->jvms(); 2161 int max_depth = youngest_jvms->depth(); 2162 for (int depth = 1; depth <= max_depth; depth++) { 2163 JVMState* jvms = youngest_jvms->of_depth(depth); 2164 int num_mon = jvms->nof_monitors(); 2165 // Loop over monitors 2166 for (int idx = 0; idx < num_mon; idx++) { 2167 Node* obj_node = sfn->monitor_obj(jvms, idx); 2168 Node* box_node = sfn->monitor_box(jvms, idx); 2169 if (box_node == oldbox && obj_node->eqv_uncast(obj)) { 2170 int j = jvms->monitor_box_offset(idx); 2171 _igvn.replace_input_of(u, j, newbox); 2172 next_edge = false; 2173 } 2174 } 2175 } 2176 } 2177 if (next_edge) i++; 2178 } 2179 } 2180 2181 //-----------------------mark_eliminated_locking_nodes----------------------- 2182 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { 2183 if (EliminateNestedLocks) { 2184 if (alock->is_nested()) { 2185 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity"); 2186 return; 2187 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened 2188 // Only Lock node has JVMState needed here. 2189 // Not that preceding claim is documented anywhere else. 2190 if (alock->jvms() != NULL) { 2191 if (alock->as_Lock()->is_nested_lock_region()) { 2192 // Mark eliminated related nested locks and unlocks. 2193 Node* obj = alock->obj_node(); 2194 BoxLockNode* box_node = alock->box_node()->as_BoxLock(); 2195 assert(!box_node->is_eliminated(), "should not be marked yet"); 2196 // Note: BoxLock node is marked eliminated only here 2197 // and it is used to indicate that all associated lock 2198 // and unlock nodes are marked for elimination. 2199 box_node->set_eliminated(); // Box's hash is always NO_HASH here 2200 for (uint i = 0; i < box_node->outcnt(); i++) { 2201 Node* u = box_node->raw_out(i); 2202 if (u->is_AbstractLock()) { 2203 alock = u->as_AbstractLock(); 2204 if (alock->box_node() == box_node) { 2205 // Verify that this Box is referenced only by related locks. 2206 assert(alock->obj_node()->eqv_uncast(obj), ""); 2207 // Mark all related locks and unlocks. 2208 #ifdef ASSERT 2209 alock->log_lock_optimization(C, "eliminate_lock_set_nested"); 2210 #endif 2211 alock->set_nested(); 2212 } 2213 } 2214 } 2215 } else { 2216 #ifdef ASSERT 2217 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region"); 2218 if (C->log() != NULL) 2219 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output 2220 #endif 2221 } 2222 } 2223 return; 2224 } 2225 // Process locks for non escaping object 2226 assert(alock->is_non_esc_obj(), ""); 2227 } // EliminateNestedLocks 2228 2229 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object 2230 // Look for all locks of this object and mark them and 2231 // corresponding BoxLock nodes as eliminated. 2232 Node* obj = alock->obj_node(); 2233 for (uint j = 0; j < obj->outcnt(); j++) { 2234 Node* o = obj->raw_out(j); 2235 if (o->is_AbstractLock() && 2236 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) { 2237 alock = o->as_AbstractLock(); 2238 Node* box = alock->box_node(); 2239 // Replace old box node with new eliminated box for all users 2240 // of the same object and mark related locks as eliminated. 2241 mark_eliminated_box(box, obj); 2242 } 2243 } 2244 } 2245 } 2246 2247 // we have determined that this lock/unlock can be eliminated, we simply 2248 // eliminate the node without expanding it. 2249 // 2250 // Note: The membar's associated with the lock/unlock are currently not 2251 // eliminated. This should be investigated as a future enhancement. 2252 // 2253 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 2254 2255 if (!alock->is_eliminated()) { 2256 return false; 2257 } 2258 #ifdef ASSERT 2259 if (!alock->is_coarsened()) { 2260 // Check that new "eliminated" BoxLock node is created. 2261 BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); 2262 assert(oldbox->is_eliminated(), "should be done already"); 2263 } 2264 #endif 2265 2266 alock->log_lock_optimization(C, "eliminate_lock"); 2267 2268 #ifndef PRODUCT 2269 if (PrintEliminateLocks) { 2270 if (alock->is_Lock()) { 2271 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx); 2272 } else { 2273 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx); 2274 } 2275 } 2276 #endif 2277 2278 Node* mem = alock->in(TypeFunc::Memory); 2279 Node* ctrl = alock->in(TypeFunc::Control); 2280 2281 extract_call_projections(alock); 2282 // There are 2 projections from the lock. The lock node will 2283 // be deleted when its last use is subsumed below. 2284 assert(alock->outcnt() == 2 && 2285 _fallthroughproj != NULL && 2286 _memproj_fallthrough != NULL, 2287 "Unexpected projections from Lock/Unlock"); 2288 2289 Node* fallthroughproj = _fallthroughproj; 2290 Node* memproj_fallthrough = _memproj_fallthrough; 2291 2292 // The memory projection from a lock/unlock is RawMem 2293 // The input to a Lock is merged memory, so extract its RawMem input 2294 // (unless the MergeMem has been optimized away.) 2295 if (alock->is_Lock()) { 2296 // Seach for MemBarAcquireLock node and delete it also. 2297 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 2298 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, ""); 2299 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 2300 Node* memproj = membar->proj_out(TypeFunc::Memory); 2301 _igvn.replace_node(ctrlproj, fallthroughproj); 2302 _igvn.replace_node(memproj, memproj_fallthrough); 2303 2304 // Delete FastLock node also if this Lock node is unique user 2305 // (a loop peeling may clone a Lock node). 2306 Node* flock = alock->as_Lock()->fastlock_node(); 2307 if (flock->outcnt() == 1) { 2308 assert(flock->unique_out() == alock, "sanity"); 2309 _igvn.replace_node(flock, top()); 2310 } 2311 } 2312 2313 // Seach for MemBarReleaseLock node and delete it also. 2314 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() && 2315 ctrl->in(0)->is_MemBar()) { 2316 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 2317 assert(membar->Opcode() == Op_MemBarReleaseLock && 2318 mem->is_Proj() && membar == mem->in(0), ""); 2319 _igvn.replace_node(fallthroughproj, ctrl); 2320 _igvn.replace_node(memproj_fallthrough, mem); 2321 fallthroughproj = ctrl; 2322 memproj_fallthrough = mem; 2323 ctrl = membar->in(TypeFunc::Control); 2324 mem = membar->in(TypeFunc::Memory); 2325 } 2326 2327 _igvn.replace_node(fallthroughproj, ctrl); 2328 _igvn.replace_node(memproj_fallthrough, mem); 2329 return true; 2330 } 2331 2332 2333 //------------------------------expand_lock_node---------------------- 2334 void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 2335 2336 Node* ctrl = lock->in(TypeFunc::Control); 2337 Node* mem = lock->in(TypeFunc::Memory); 2338 Node* obj = lock->obj_node(); 2339 Node* box = lock->box_node(); 2340 Node* flock = lock->fastlock_node(); 2341 2342 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2343 2344 // Make the merge point 2345 Node *region; 2346 Node *mem_phi; 2347 Node *slow_path; 2348 2349 if (UseOptoBiasInlining) { 2350 /* 2351 * See the full description in MacroAssembler::biased_locking_enter(). 2352 * 2353 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { 2354 * // The object is biased. 2355 * proto_node = klass->prototype_header; 2356 * o_node = thread | proto_node; 2357 * x_node = o_node ^ mark_word; 2358 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? 2359 * // Done. 2360 * } else { 2361 * if( (x_node & biased_lock_mask) != 0 ) { 2362 * // The klass's prototype header is no longer biased. 2363 * cas(&mark_word, mark_word, proto_node) 2364 * goto cas_lock; 2365 * } else { 2366 * // The klass's prototype header is still biased. 2367 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? 2368 * old = mark_word; 2369 * new = o_node; 2370 * } else { 2371 * // Different thread or anonymous biased. 2372 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); 2373 * new = thread | old; 2374 * } 2375 * // Try to rebias. 2376 * if( cas(&mark_word, old, new) == 0 ) { 2377 * // Done. 2378 * } else { 2379 * goto slow_path; // Failed. 2380 * } 2381 * } 2382 * } 2383 * } else { 2384 * // The object is not biased. 2385 * cas_lock: 2386 * if( FastLock(obj) == 0 ) { 2387 * // Done. 2388 * } else { 2389 * slow_path: 2390 * OptoRuntime::complete_monitor_locking_Java(obj); 2391 * } 2392 * } 2393 */ 2394 2395 region = new RegionNode(5); 2396 // create a Phi for the memory state 2397 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2398 2399 Node* fast_lock_region = new RegionNode(3); 2400 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); 2401 2402 // First, check mark word for the biased lock pattern. 2403 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2404 2405 // Get fast path - mark word has the biased lock pattern. 2406 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, 2407 markOopDesc::biased_lock_mask_in_place, 2408 markOopDesc::biased_lock_pattern, true); 2409 // fast_lock_region->in(1) is set to slow path. 2410 fast_lock_mem_phi->init_req(1, mem); 2411 2412 // Now check that the lock is biased to the current thread and has 2413 // the same epoch and bias as Klass::_prototype_header. 2414 2415 // Special-case a fresh allocation to avoid building nodes: 2416 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); 2417 if (klass_node == NULL) { 2418 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 2419 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr())); 2420 #ifdef _LP64 2421 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) { 2422 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity"); 2423 klass_node->in(1)->init_req(0, ctrl); 2424 } else 2425 #endif 2426 klass_node->init_req(0, ctrl); 2427 } 2428 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type()); 2429 2430 Node* thread = transform_later(new ThreadLocalNode()); 2431 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2432 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node)); 2433 Node* x_node = transform_later(new XorXNode(o_node, mark_node)); 2434 2435 // Get slow path - mark word does NOT match the value. 2436 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, 2437 (~markOopDesc::age_mask_in_place), 0); 2438 // region->in(3) is set to fast path - the object is biased to the current thread. 2439 mem_phi->init_req(3, mem); 2440 2441 2442 // Mark word does NOT match the value (thread | Klass::_prototype_header). 2443 2444 2445 // First, check biased pattern. 2446 // Get fast path - _prototype_header has the same biased lock pattern. 2447 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, 2448 markOopDesc::biased_lock_mask_in_place, 0, true); 2449 2450 not_biased_ctrl = fast_lock_region->in(2); // Slow path 2451 // fast_lock_region->in(2) - the prototype header is no longer biased 2452 // and we have to revoke the bias on this object. 2453 // We are going to try to reset the mark of this object to the prototype 2454 // value and fall through to the CAS-based locking scheme. 2455 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 2456 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr, 2457 proto_node, mark_node); 2458 transform_later(cas); 2459 Node* proj = transform_later(new SCMemProjNode(cas)); 2460 fast_lock_mem_phi->init_req(2, proj); 2461 2462 2463 // Second, check epoch bits. 2464 Node* rebiased_region = new RegionNode(3); 2465 Node* old_phi = new PhiNode( rebiased_region, TypeX_X); 2466 Node* new_phi = new PhiNode( rebiased_region, TypeX_X); 2467 2468 // Get slow path - mark word does NOT match epoch bits. 2469 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, 2470 markOopDesc::epoch_mask_in_place, 0); 2471 // The epoch of the current bias is not valid, attempt to rebias the object 2472 // toward the current thread. 2473 rebiased_region->init_req(2, epoch_ctrl); 2474 old_phi->init_req(2, mark_node); 2475 new_phi->init_req(2, o_node); 2476 2477 // rebiased_region->in(1) is set to fast path. 2478 // The epoch of the current bias is still valid but we know 2479 // nothing about the owner; it might be set or it might be clear. 2480 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place | 2481 markOopDesc::age_mask_in_place | 2482 markOopDesc::epoch_mask_in_place); 2483 Node* old = transform_later(new AndXNode(mark_node, cmask)); 2484 cast_thread = transform_later(new CastP2XNode(ctrl, thread)); 2485 Node* new_mark = transform_later(new OrXNode(cast_thread, old)); 2486 old_phi->init_req(1, old); 2487 new_phi->init_req(1, new_mark); 2488 2489 transform_later(rebiased_region); 2490 transform_later(old_phi); 2491 transform_later(new_phi); 2492 2493 // Try to acquire the bias of the object using an atomic operation. 2494 // If this fails we will go in to the runtime to revoke the object's bias. 2495 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi); 2496 transform_later(cas); 2497 proj = transform_later(new SCMemProjNode(cas)); 2498 2499 // Get slow path - Failed to CAS. 2500 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); 2501 mem_phi->init_req(4, proj); 2502 // region->in(4) is set to fast path - the object is rebiased to the current thread. 2503 2504 // Failed to CAS. 2505 slow_path = new RegionNode(3); 2506 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); 2507 2508 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control 2509 slow_mem->init_req(1, proj); 2510 2511 // Call CAS-based locking scheme (FastLock node). 2512 2513 transform_later(fast_lock_region); 2514 transform_later(fast_lock_mem_phi); 2515 2516 // Get slow path - FastLock failed to lock the object. 2517 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); 2518 mem_phi->init_req(2, fast_lock_mem_phi); 2519 // region->in(2) is set to fast path - the object is locked to the current thread. 2520 2521 slow_path->init_req(2, ctrl); // Capture slow-control 2522 slow_mem->init_req(2, fast_lock_mem_phi); 2523 2524 transform_later(slow_path); 2525 transform_later(slow_mem); 2526 // Reset lock's memory edge. 2527 lock->set_req(TypeFunc::Memory, slow_mem); 2528 2529 } else { 2530 region = new RegionNode(3); 2531 // create a Phi for the memory state 2532 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2533 2534 // Optimize test; set region slot 2 2535 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); 2536 mem_phi->init_req(2, mem); 2537 } 2538 2539 // Make slow path call 2540 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), 2541 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, 2542 obj, box, NULL); 2543 2544 extract_call_projections(call); 2545 2546 // Slow path can only throw asynchronous exceptions, which are always 2547 // de-opted. So the compiler thinks the slow-call can never throw an 2548 // exception. If it DOES throw an exception we would need the debug 2549 // info removed first (since if it throws there is no monitor). 2550 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2551 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2552 2553 // Capture slow path 2554 // disconnect fall-through projection from call and create a new one 2555 // hook up users of fall-through projection to region 2556 Node *slow_ctrl = _fallthroughproj->clone(); 2557 transform_later(slow_ctrl); 2558 _igvn.hash_delete(_fallthroughproj); 2559 _fallthroughproj->disconnect_inputs(NULL, C); 2560 region->init_req(1, slow_ctrl); 2561 // region inputs are now complete 2562 transform_later(region); 2563 _igvn.replace_node(_fallthroughproj, region); 2564 2565 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory)); 2566 mem_phi->init_req(1, memproj ); 2567 transform_later(mem_phi); 2568 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2569 } 2570 2571 //------------------------------expand_unlock_node---------------------- 2572 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 2573 2574 Node* ctrl = unlock->in(TypeFunc::Control); 2575 Node* mem = unlock->in(TypeFunc::Memory); 2576 Node* obj = unlock->obj_node(); 2577 Node* box = unlock->box_node(); 2578 2579 assert(!box->as_BoxLock()->is_eliminated(), "sanity"); 2580 2581 // No need for a null check on unlock 2582 2583 // Make the merge point 2584 Node *region; 2585 Node *mem_phi; 2586 2587 if (UseOptoBiasInlining) { 2588 // Check for biased locking unlock case, which is a no-op. 2589 // See the full description in MacroAssembler::biased_locking_exit(). 2590 region = new RegionNode(4); 2591 // create a Phi for the memory state 2592 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2593 mem_phi->init_req(3, mem); 2594 2595 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); 2596 ctrl = opt_bits_test(ctrl, region, 3, mark_node, 2597 markOopDesc::biased_lock_mask_in_place, 2598 markOopDesc::biased_lock_pattern); 2599 } else { 2600 region = new RegionNode(3); 2601 // create a Phi for the memory state 2602 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 2603 } 2604 2605 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box ); 2606 funlock = transform_later( funlock )->as_FastUnlock(); 2607 // Optimize test; set region slot 2 2608 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); 2609 Node *thread = transform_later(new ThreadLocalNode()); 2610 2611 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), 2612 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), 2613 "complete_monitor_unlocking_C", slow_path, obj, box, thread); 2614 2615 extract_call_projections(call); 2616 2617 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 2618 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 2619 2620 // No exceptions for unlocking 2621 // Capture slow path 2622 // disconnect fall-through projection from call and create a new one 2623 // hook up users of fall-through projection to region 2624 Node *slow_ctrl = _fallthroughproj->clone(); 2625 transform_later(slow_ctrl); 2626 _igvn.hash_delete(_fallthroughproj); 2627 _fallthroughproj->disconnect_inputs(NULL, C); 2628 region->init_req(1, slow_ctrl); 2629 // region inputs are now complete 2630 transform_later(region); 2631 _igvn.replace_node(_fallthroughproj, region); 2632 2633 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) ); 2634 mem_phi->init_req(1, memproj ); 2635 mem_phi->init_req(2, mem); 2636 transform_later(mem_phi); 2637 _igvn.replace_node(_memproj_fallthrough, mem_phi); 2638 } 2639 2640 //---------------------------eliminate_macro_nodes---------------------- 2641 // Eliminate scalar replaced allocations and associated locks. 2642 void PhaseMacroExpand::eliminate_macro_nodes() { 2643 if (C->macro_count() == 0) 2644 return; 2645 2646 // First, attempt to eliminate locks 2647 int cnt = C->macro_count(); 2648 for (int i=0; i < cnt; i++) { 2649 Node *n = C->macro_node(i); 2650 if (n->is_AbstractLock()) { // Lock and Unlock nodes 2651 // Before elimination mark all associated (same box and obj) 2652 // lock and unlock nodes. 2653 mark_eliminated_locking_nodes(n->as_AbstractLock()); 2654 } 2655 } 2656 bool progress = true; 2657 while (progress) { 2658 progress = false; 2659 for (int i = C->macro_count(); i > 0; i--) { 2660 Node * n = C->macro_node(i-1); 2661 bool success = false; 2662 debug_only(int old_macro_count = C->macro_count();); 2663 if (n->is_AbstractLock()) { 2664 success = eliminate_locking_node(n->as_AbstractLock()); 2665 } 2666 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2667 progress = progress || success; 2668 } 2669 } 2670 // Next, attempt to eliminate allocations 2671 _has_locks = false; 2672 progress = true; 2673 while (progress) { 2674 progress = false; 2675 for (int i = C->macro_count(); i > 0; i--) { 2676 Node * n = C->macro_node(i-1); 2677 bool success = false; 2678 debug_only(int old_macro_count = C->macro_count();); 2679 switch (n->class_id()) { 2680 case Node::Class_Allocate: 2681 case Node::Class_AllocateArray: 2682 success = eliminate_allocate_node(n->as_Allocate()); 2683 break; 2684 case Node::Class_CallStaticJava: 2685 success = eliminate_boxing_node(n->as_CallStaticJava()); 2686 break; 2687 case Node::Class_Lock: 2688 case Node::Class_Unlock: 2689 assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); 2690 _has_locks = true; 2691 break; 2692 case Node::Class_ArrayCopy: 2693 break; 2694 default: 2695 assert(n->Opcode() == Op_LoopLimit || 2696 n->Opcode() == Op_Opaque1 || 2697 n->Opcode() == Op_Opaque2 || 2698 n->Opcode() == Op_Opaque3 || 2699 n->Opcode() == Op_Opaque4, "unknown node type in macro list"); 2700 } 2701 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2702 progress = progress || success; 2703 } 2704 } 2705 } 2706 2707 //------------------------------expand_macro_nodes---------------------- 2708 // Returns true if a failure occurred. 2709 bool PhaseMacroExpand::expand_macro_nodes() { 2710 // Last attempt to eliminate macro nodes. 2711 eliminate_macro_nodes(); 2712 2713 // Make sure expansion will not cause node limit to be exceeded. 2714 // Worst case is a macro node gets expanded into about 200 nodes. 2715 // Allow 50% more for optimization. 2716 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) ) 2717 return true; 2718 2719 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. 2720 bool progress = true; 2721 while (progress) { 2722 progress = false; 2723 for (int i = C->macro_count(); i > 0; i--) { 2724 Node * n = C->macro_node(i-1); 2725 bool success = false; 2726 debug_only(int old_macro_count = C->macro_count();); 2727 if (n->Opcode() == Op_LoopLimit) { 2728 // Remove it from macro list and put on IGVN worklist to optimize. 2729 C->remove_macro_node(n); 2730 _igvn._worklist.push(n); 2731 success = true; 2732 } else if (n->Opcode() == Op_CallStaticJava) { 2733 // Remove it from macro list and put on IGVN worklist to optimize. 2734 C->remove_macro_node(n); 2735 _igvn._worklist.push(n); 2736 success = true; 2737 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { 2738 _igvn.replace_node(n, n->in(1)); 2739 success = true; 2740 #if INCLUDE_RTM_OPT 2741 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) { 2742 assert(C->profile_rtm(), "should be used only in rtm deoptimization code"); 2743 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), ""); 2744 Node* cmp = n->unique_out(); 2745 #ifdef ASSERT 2746 // Validate graph. 2747 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), ""); 2748 BoolNode* bol = cmp->unique_out()->as_Bool(); 2749 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() && 2750 (bol->_test._test == BoolTest::ne), ""); 2751 IfNode* ifn = bol->unique_out()->as_If(); 2752 assert((ifn->outcnt() == 2) && 2753 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, ""); 2754 #endif 2755 Node* repl = n->in(1); 2756 if (!_has_locks) { 2757 // Remove RTM state check if there are no locks in the code. 2758 // Replace input to compare the same value. 2759 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1); 2760 } 2761 _igvn.replace_node(n, repl); 2762 success = true; 2763 #endif 2764 } else if (n->Opcode() == Op_Opaque4) { 2765 Node* res = ((Opaque4Node*)n)->adjust_strip_mined_loop(&_igvn); 2766 guarantee(res != NULL, "strip mined adjustment failed"); 2767 _igvn.replace_node(n, res); 2768 success = true; 2769 } 2770 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 2771 progress = progress || success; 2772 } 2773 } 2774 2775 // expand arraycopy "macro" nodes first 2776 // For ReduceBulkZeroing, we must first process all arraycopy nodes 2777 // before the allocate nodes are expanded. 2778 int macro_idx = C->macro_count() - 1; 2779 while (macro_idx >= 0) { 2780 Node * n = C->macro_node(macro_idx); 2781 assert(n->is_macro(), "only macro nodes expected here"); 2782 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2783 // node is unreachable, so don't try to expand it 2784 C->remove_macro_node(n); 2785 } else if (n->is_ArrayCopy()){ 2786 int macro_count = C->macro_count(); 2787 expand_arraycopy_node(n->as_ArrayCopy()); 2788 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2789 } 2790 if (C->failing()) return true; 2791 macro_idx --; 2792 } 2793 2794 // expand "macro" nodes 2795 // nodes are removed from the macro list as they are processed 2796 while (C->macro_count() > 0) { 2797 int macro_count = C->macro_count(); 2798 Node * n = C->macro_node(macro_count-1); 2799 assert(n->is_macro(), "only macro nodes expected here"); 2800 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 2801 // node is unreachable, so don't try to expand it 2802 C->remove_macro_node(n); 2803 continue; 2804 } 2805 switch (n->class_id()) { 2806 case Node::Class_Allocate: 2807 expand_allocate(n->as_Allocate()); 2808 break; 2809 case Node::Class_AllocateArray: 2810 expand_allocate_array(n->as_AllocateArray()); 2811 break; 2812 case Node::Class_Lock: 2813 expand_lock_node(n->as_Lock()); 2814 break; 2815 case Node::Class_Unlock: 2816 expand_unlock_node(n->as_Unlock()); 2817 break; 2818 default: 2819 assert(false, "unknown node type in macro list"); 2820 } 2821 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 2822 if (C->failing()) return true; 2823 } 2824 2825 _igvn.set_delay_transform(false); 2826 _igvn.optimize(); 2827 if (C->failing()) return true; 2828 return false; 2829 }