macro.cpp revision 196:d1605aabd0a1
1185089Sraj/* 2209131Sraj * Copyright 2005-2008 Sun Microsystems, Inc. All Rights Reserved. 3209131Sraj * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4240489Sgber * 5185089Sraj * This code is free software; you can redistribute it and/or modify it 6185089Sraj * under the terms of the GNU General Public License version 2 only, as 7185089Sraj * published by the Free Software Foundation. 8185089Sraj * 9209131Sraj * This code is distributed in the hope that it will be useful, but WITHOUT 10209131Sraj * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11209131Sraj * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12185089Sraj * version 2 for more details (a copy is included in the LICENSE file that 13185089Sraj * accompanied this code). 14185089Sraj * 15185089Sraj * You should have received a copy of the GNU General Public License version 16185089Sraj * 2 along with this work; if not, write to the Free Software Foundation, 17185089Sraj * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18185089Sraj * 19185089Sraj * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20185089Sraj * CA 95054 USA or visit www.sun.com if you need additional information or 21185089Sraj * have any questions. 22185089Sraj * 23185089Sraj */ 24185089Sraj 25185089Sraj#include "incls/_precompiled.incl" 26185089Sraj#include "incls/_macro.cpp.incl" 27185089Sraj 28185089Sraj 29185089Sraj// 30185089Sraj// Replace any references to "oldref" in inputs to "use" with "newref". 31185089Sraj// Returns the number of replacements made. 32185089Sraj// 33185089Srajint PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { 34185089Sraj int nreplacements = 0; 35185089Sraj uint req = use->req(); 36185089Sraj for (uint j = 0; j < use->len(); j++) { 37185089Sraj Node *uin = use->in(j); 38185089Sraj if (uin == oldref) { 39185089Sraj if (j < req) 40185089Sraj use->set_req(j, newref); 41185089Sraj else 42185089Sraj use->set_prec(j, newref); 43185089Sraj nreplacements++; 44185089Sraj } else if (j >= req && uin == NULL) { 45185089Sraj break; 46185089Sraj } 47185089Sraj } 48185089Sraj return nreplacements; 49185089Sraj} 50185089Sraj 51185089Srajvoid PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) { 52185089Sraj // Copy debug information and adjust JVMState information 53185089Sraj uint old_dbg_start = oldcall->tf()->domain()->cnt(); 54185089Sraj uint new_dbg_start = newcall->tf()->domain()->cnt(); 55185089Sraj int jvms_adj = new_dbg_start - old_dbg_start; 56240493Sgber assert (new_dbg_start == newcall->req(), "argument count mismatch"); 57240493Sgber 58185089Sraj Dict* sosn_map = new Dict(cmpkey,hashkey); 59185089Sraj for (uint i = old_dbg_start; i < oldcall->req(); i++) { 60185089Sraj Node* old_in = oldcall->in(i); 61209131Sraj // Clone old SafePointScalarObjectNodes, adjusting their field contents. 62209131Sraj if (old_in->is_SafePointScalarObject()) { 63209131Sraj SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject(); 64185089Sraj uint old_unique = C->unique(); 65185089Sraj Node* new_in = old_sosn->clone(jvms_adj, sosn_map); 66185089Sraj if (old_unique != C->unique()) { 67185089Sraj new_in = transform_later(new_in); // Register new node. 68209131Sraj } 69185089Sraj old_in = new_in; 70185089Sraj } 71185089Sraj newcall->add_req(old_in); 72185089Sraj } 73185089Sraj 74185089Sraj newcall->set_jvms(oldcall->jvms()); 75185089Sraj for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) { 76209131Sraj jvms->set_map(newcall); 77185089Sraj jvms->set_locoff(jvms->locoff()+jvms_adj); 78240493Sgber jvms->set_stkoff(jvms->stkoff()+jvms_adj); 79240493Sgber jvms->set_monoff(jvms->monoff()+jvms_adj); 80240493Sgber jvms->set_scloff(jvms->scloff()+jvms_adj); 81240493Sgber jvms->set_endoff(jvms->endoff()+jvms_adj); 82240493Sgber } 83240493Sgber} 84185089Sraj 85185089SrajNode* PhaseMacroExpand::opt_iff(Node* region, Node* iff) { 86185089Sraj IfNode *opt_iff = transform_later(iff)->as_If(); 87185089Sraj 88185089Sraj // Fast path taken; set region slot 2 89185089Sraj Node *fast_taken = transform_later( new (C, 1) IfFalseNode(opt_iff) ); 90185089Sraj region->init_req(2,fast_taken); // Capture fast-control 91185089Sraj 92185089Sraj // Fast path not-taken, i.e. slow path 93185089Sraj Node *slow_taken = transform_later( new (C, 1) IfTrueNode(opt_iff) ); 94185089Sraj return slow_taken; 95185089Sraj} 96185089Sraj 97185089Sraj//--------------------copy_predefined_input_for_runtime_call-------------------- 98185089Srajvoid PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { 99185089Sraj // Set fixed predefined input arguments 100240489Sgber call->init_req( TypeFunc::Control, ctrl ); 101240489Sgber call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); 102185089Sraj call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? 103240489Sgber call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); 104185089Sraj call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); 105240489Sgber} 106240489Sgber 107185089Sraj//------------------------------make_slow_call--------------------------------- 108240489SgberCallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, address slow_call, const char* leaf_name, Node* slow_path, Node* parm0, Node* parm1) { 109240489Sgber 110240489Sgber // Slow-path call 111240489Sgber int size = slow_call_type->domain()->cnt(); 112240489Sgber CallNode *call = leaf_name 113240489Sgber ? (CallNode*)new (C, size) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) 114209131Sraj : (CallNode*)new (C, size) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM ); 115185089Sraj 116185089Sraj // Slow path call has no side-effects, uses few values 117209131Sraj copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); 118209131Sraj if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0); 119209131Sraj if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1); 120240489Sgber copy_call_debug_info(oldcall, call); 121240489Sgber call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 122240489Sgber _igvn.hash_delete(oldcall); 123209131Sraj _igvn.subsume_node(oldcall, call); 124185089Sraj transform_later(call); 125209131Sraj 126209131Sraj return call; 127209131Sraj} 128240489Sgber 129240489Sgbervoid PhaseMacroExpand::extract_call_projections(CallNode *call) { 130209131Sraj _fallthroughproj = NULL; 131185089Sraj _fallthroughcatchproj = NULL; 132185089Sraj _ioproj_fallthrough = NULL; 133185089Sraj _ioproj_catchall = NULL; 134185089Sraj _catchallcatchproj = NULL; 135185089Sraj _memproj_fallthrough = NULL; 136185089Sraj _memproj_catchall = NULL; 137240493Sgber _resproj = NULL; 138240493Sgber for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) { 139240493Sgber ProjNode *pn = call->fast_out(i)->as_Proj(); 140185089Sraj switch (pn->_con) { 141185089Sraj case TypeFunc::Control: 142209131Sraj { 143240489Sgber // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj 144185089Sraj _fallthroughproj = pn; 145209131Sraj DUIterator_Fast jmax, j = pn->fast_outs(jmax); 146185089Sraj const Node *cn = pn->fast_out(j); 147185089Sraj if (cn->is_Catch()) { 148209131Sraj ProjNode *cpn = NULL; 149209131Sraj for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { 150209131Sraj cpn = cn->fast_out(k)->as_Proj(); 151209131Sraj assert(cpn->is_CatchProj(), "must be a CatchProjNode"); 152209131Sraj if (cpn->_con == CatchProjNode::fall_through_index) 153209131Sraj _fallthroughcatchproj = cpn; 154209131Sraj else { 155209131Sraj assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); 156209131Sraj _catchallcatchproj = cpn; 157209131Sraj } 158209131Sraj } 159209131Sraj } 160240489Sgber break; 161240489Sgber } 162185089Sraj case TypeFunc::I_O: 163209131Sraj if (pn->_is_io_use) 164209131Sraj _ioproj_catchall = pn; 165209131Sraj else 166209131Sraj _ioproj_fallthrough = pn; 167209131Sraj break; 168185089Sraj case TypeFunc::Memory: 169209131Sraj if (pn->_is_io_use) 170185089Sraj _memproj_catchall = pn; 171209131Sraj else 172209131Sraj _memproj_fallthrough = pn; 173185089Sraj break; 174209131Sraj case TypeFunc::Parms: 175209131Sraj _resproj = pn; 176209131Sraj break; 177185089Sraj default: 178209131Sraj assert(false, "unexpected projection from allocation node."); 179240493Sgber } 180240493Sgber } 181240493Sgber 182240493Sgber} 183240493Sgber 184185089Sraj// Eliminate a card mark sequence. p2x is a ConvP2XNode 185185089Srajvoid PhaseMacroExpand::eliminate_card_mark(Node *p2x) { 186185089Sraj assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required"); 187185089Sraj Node *shift = p2x->unique_out(); 188209131Sraj Node *addp = shift->unique_out(); 189185089Sraj for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) { 190209131Sraj Node *st = addp->last_out(j); 191209131Sraj assert(st->is_Store(), "store required"); 192185089Sraj _igvn.replace_node(st, st->in(MemNode::Memory)); 193185089Sraj } 194209131Sraj} 195209131Sraj 196209131Sraj// Search for a memory operation for the specified memory slice. 197209131Srajstatic Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc) { 198185089Sraj Node *orig_mem = mem; 199185089Sraj Node *alloc_mem = alloc->in(TypeFunc::Memory); 200185089Sraj while (true) { 201185089Sraj if (mem == alloc_mem || mem == start_mem ) { 202185089Sraj return mem; // hit one of our sentinals 203185089Sraj } else if (mem->is_MergeMem()) { 204209131Sraj mem = mem->as_MergeMem()->memory_at(alias_idx); 205209131Sraj } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { 206209131Sraj Node *in = mem->in(0); 207209131Sraj // we can safely skip over safepoints, calls, locks and membars because we 208240493Sgber // already know that the object is safe to eliminate. 209240493Sgber if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { 210240493Sgber return in; 211240493Sgber } else if (in->is_Call() || in->is_MemBar()) { 212240493Sgber mem = in->in(TypeFunc::Memory); 213240493Sgber } else { 214240493Sgber assert(false, "unexpected projection"); 215209131Sraj } 216209131Sraj } else if (mem->is_Store()) { 217209131Sraj const TypePtr* atype = mem->as_Store()->adr_type(); 218209131Sraj int adr_idx = Compile::current()->get_alias_index(atype); 219209131Sraj if (adr_idx == alias_idx) { 220209131Sraj assert(atype->isa_oopptr(), "address type must be oopptr"); 221209131Sraj int adr_offset = atype->offset(); 222227843Smarius uint adr_iid = atype->is_oopptr()->instance_id(); 223185089Sraj // Array elements references have the same alias_idx 224185089Sraj // but different offset and different instance_id. 225209131Sraj if (adr_offset == offset && adr_iid == alloc->_idx) 226185089Sraj return mem; 227209131Sraj } else { 228209131Sraj assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw"); 229185089Sraj } 230185089Sraj mem = mem->in(MemNode::Memory); 231185089Sraj } else { 232185089Sraj return mem; 233209131Sraj } 234185089Sraj if (mem == orig_mem) 235185089Sraj return mem; 236185089Sraj } 237185089Sraj} 238209131Sraj 239185089Sraj// 240218228Smarcel// Given a Memory Phi, compute a value Phi containing the values from stores 241185089Sraj// on the input paths. 242218228Smarcel// Note: this function is recursive, its depth is limied by the "level" argument 243218228Smarcel// Returns the computed Phi, or NULL if it cannot compute it. 244209131SrajNode *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, Node *alloc, int level) { 245218228Smarcel 246218228Smarcel if (level <= 0) { 247218228Smarcel return NULL; 248209131Sraj } 249185089Sraj int alias_idx = C->get_alias_index(adr_t); 250209131Sraj int offset = adr_t->offset(); 251209131Sraj int instance_id = adr_t->instance_id(); 252185089Sraj 253185089Sraj Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 254185089Sraj Node *alloc_mem = alloc->in(TypeFunc::Memory); 255209131Sraj 256185089Sraj uint length = mem->req(); 257209131Sraj GrowableArray <Node *> values(length, length, NULL); 258209131Sraj 259240489Sgber for (uint j = 1; j < length; j++) { 260209131Sraj Node *in = mem->in(j); 261185089Sraj if (in == NULL || in->is_top()) { 262185089Sraj values.at_put(j, in); 263209131Sraj } else { 264240489Sgber Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc); 265185089Sraj if (val == start_mem || val == alloc_mem) { 266240489Sgber // hit a sentinel, return appropriate 0 value 267218228Smarcel values.at_put(j, _igvn.zerocon(ft)); 268218228Smarcel continue; 269218228Smarcel } 270209131Sraj if (val->is_Initialize()) { 271240489Sgber val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 272240489Sgber } 273240489Sgber if (val == NULL) { 274218228Smarcel return NULL; // can't find a value on this path 275209131Sraj } 276240489Sgber if (val == mem) { 277209131Sraj values.at_put(j, mem); 278209131Sraj } else if (val->is_Store()) { 279209131Sraj values.at_put(j, val->in(MemNode::ValueIn)); 280185089Sraj } else if(val->is_Proj() && val->in(0) == alloc) { 281185089Sraj values.at_put(j, _igvn.zerocon(ft)); 282209131Sraj } else if (val->is_Phi()) { 283209131Sraj // Check if an appropriate node already exists. 284209131Sraj Node* region = val->in(0); 285185089Sraj Node* old_phi = NULL; 286185089Sraj for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { 287185089Sraj Node* phi = region->fast_out(k); 288185089Sraj if (phi->is_Phi() && phi != val && 289209131Sraj phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) { 290185089Sraj old_phi = phi; 291185089Sraj break; 292185089Sraj } 293185089Sraj } 294185089Sraj if (old_phi == NULL) { 295240489Sgber val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, level-1); 296240489Sgber if (val == NULL) { 297240489Sgber return NULL; 298240489Sgber } 299209131Sraj values.at_put(j, val); 300240489Sgber } else { 301240489Sgber values.at_put(j, old_phi); 302240489Sgber } 303240489Sgber } else { 304240489Sgber return NULL; // unknown node on this path 305240489Sgber } 306240489Sgber } 307240489Sgber } 308240489Sgber // create a new Phi for the value 309209131Sraj PhiNode *phi = new (C, length) PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset); 310209131Sraj for (uint j = 1; j < length; j++) { 311209131Sraj if (values.at(j) == mem) { 312209131Sraj phi->init_req(j, phi); 313209131Sraj } else { 314209131Sraj phi->init_req(j, values.at(j)); 315209131Sraj } 316209131Sraj } 317240489Sgber transform_later(phi); 318209131Sraj return phi; 319240489Sgber} 320185089Sraj 321240489Sgber// Search the last value stored into the object's field. 322240489SgberNode *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) { 323240489Sgber assert(adr_t->is_instance_field(), "instance required"); 324240489Sgber uint instance_id = adr_t->instance_id(); 325240489Sgber assert(instance_id == alloc->_idx, "wrong allocation"); 326240489Sgber 327185089Sraj int alias_idx = C->get_alias_index(adr_t); 328240489Sgber int offset = adr_t->offset(); 329240489Sgber Node *start_mem = C->start()->proj_out(TypeFunc::Memory); 330240489Sgber Node *alloc_ctrl = alloc->in(TypeFunc::Control); 331240489Sgber Node *alloc_mem = alloc->in(TypeFunc::Memory); 332240489Sgber VectorSet visited(Thread::current()->resource_area()); 333240489Sgber 334240489Sgber 335240489Sgber bool done = sfpt_mem == alloc_mem; 336240489Sgber Node *mem = sfpt_mem; 337240489Sgber while (!done) { 338240489Sgber if (visited.test_set(mem->_idx)) { 339240489Sgber return NULL; // found a loop, give up 340240489Sgber } 341240489Sgber mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc); 342240493Sgber if (mem == start_mem || mem == alloc_mem) { 343240489Sgber done = true; // hit a sentinel, return appropriate 0 value 344240489Sgber } else if (mem->is_Initialize()) { 345240489Sgber mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); 346240489Sgber if (mem == NULL) { 347240489Sgber done = true; // Something go wrong. 348240489Sgber } else if (mem->is_Store()) { 349240489Sgber const TypePtr* atype = mem->as_Store()->adr_type(); 350240489Sgber assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); 351240489Sgber done = true; 352240489Sgber } 353240489Sgber } else if (mem->is_Store()) { 354240489Sgber const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); 355240489Sgber assert(atype != NULL, "address type must be oopptr"); 356240489Sgber assert(C->get_alias_index(atype) == alias_idx && 357240489Sgber atype->is_instance_field() && atype->offset() == offset && 358240489Sgber atype->instance_id() == instance_id, "store is correct memory slice"); 359240489Sgber done = true; 360240489Sgber } else if (mem->is_Phi()) { 361240489Sgber // try to find a phi's unique input 362240489Sgber Node *unique_input = NULL; 363240489Sgber Node *top = C->top(); 364240489Sgber for (uint i = 1; i < mem->req(); i++) { 365240489Sgber Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc); 366240489Sgber if (n == NULL || n == top || n == mem) { 367240489Sgber continue; 368240489Sgber } else if (unique_input == NULL) { 369240489Sgber unique_input = n; 370240489Sgber } else if (unique_input != n) { 371240489Sgber unique_input = top; 372240489Sgber break; 373240489Sgber } 374240489Sgber } 375240489Sgber if (unique_input != NULL && unique_input != top) { 376240489Sgber mem = unique_input; 377240489Sgber } else { 378240489Sgber done = true; 379240489Sgber } 380240489Sgber } else { 381240489Sgber assert(false, "unexpected node"); 382240489Sgber } 383240489Sgber } 384240489Sgber if (mem != NULL) { 385240489Sgber if (mem == start_mem || mem == alloc_mem) { 386240489Sgber // hit a sentinel, return appropriate 0 value 387240489Sgber return _igvn.zerocon(ft); 388240489Sgber } else if (mem->is_Store()) { 389240489Sgber return mem->in(MemNode::ValueIn); 390240489Sgber } else if (mem->is_Phi()) { 391240489Sgber // attempt to produce a Phi reflecting the values on the input paths of the Phi 392240489Sgber Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, 8); 393240489Sgber if (phi != NULL) { 394240489Sgber return phi; 395240489Sgber } 396240489Sgber } 397240489Sgber } 398240489Sgber // Something go wrong. 399209131Sraj return NULL; 400209131Sraj} 401186932Sraj 402186932Sraj// Check the possibility of scalar replacement. 403186932Srajbool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 404209131Sraj // Scan the uses of the allocation to check for anything that would 405209131Sraj // prevent us from eliminating it. 406186932Sraj NOT_PRODUCT( const char* fail_eliminate = NULL; ) 407209131Sraj DEBUG_ONLY( Node* disq_node = NULL; ) 408186932Sraj bool can_eliminate = true; 409186932Sraj 410186932Sraj Node* res = alloc->result_cast(); 411209131Sraj const TypeOopPtr* res_type = NULL; 412209131Sraj if (res == NULL) { 413186932Sraj // All users were eliminated. 414186932Sraj } else if (!res->is_CheckCastPP()) { 415186932Sraj alloc->_is_scalar_replaceable = false; // don't try again 416209131Sraj NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) 417209131Sraj can_eliminate = false; 418186932Sraj } else { 419186932Sraj res_type = _igvn.type(res)->isa_oopptr(); 420186932Sraj if (res_type == NULL) { 421240489Sgber NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) 422185089Sraj can_eliminate = false; 423186932Sraj } else if (res_type->isa_aryptr()) { 424209131Sraj int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); 425209131Sraj if (length < 0) { 426240489Sgber NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) 427186932Sraj can_eliminate = false; 428185089Sraj } 429185089Sraj } 430240489Sgber } 431240489Sgber 432240489Sgber if (can_eliminate && res != NULL) { 433240489Sgber for (DUIterator_Fast jmax, j = res->fast_outs(jmax); 434240489Sgber j < jmax && can_eliminate; j++) { 435240489Sgber Node* use = res->fast_out(j); 436240489Sgber 437240489Sgber if (use->is_AddP()) { 438240489Sgber const TypePtr* addp_type = _igvn.type(use)->is_ptr(); 439240489Sgber int offset = addp_type->offset(); 440240489Sgber 441240489Sgber if (offset == Type::OffsetTop || offset == Type::OffsetBot) { 442240489Sgber NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) 443240489Sgber can_eliminate = false; 444240489Sgber break; 445240489Sgber } 446240489Sgber for (DUIterator_Fast kmax, k = use->fast_outs(kmax); 447240489Sgber k < kmax && can_eliminate; k++) { 448240489Sgber Node* n = use->fast_out(k); 449240489Sgber if (!n->is_Store() && n->Opcode() != Op_CastP2X) { 450240489Sgber DEBUG_ONLY(disq_node = n;) 451240489Sgber if (n->is_Load()) { 452240489Sgber NOT_PRODUCT(fail_eliminate = "Field load";) 453240489Sgber } else { 454240489Sgber NOT_PRODUCT(fail_eliminate = "Not store field referrence";) 455240489Sgber } 456240489Sgber can_eliminate = false; 457240489Sgber } 458240489Sgber } 459240489Sgber } else if (use->is_SafePoint()) { 460240489Sgber SafePointNode* sfpt = use->as_SafePoint(); 461240489Sgber if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { 462240489Sgber // Object is passed as argument. 463240489Sgber DEBUG_ONLY(disq_node = use;) 464240489Sgber NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) 465240489Sgber can_eliminate = false; 466240489Sgber } 467240489Sgber Node* sfptMem = sfpt->memory(); 468240489Sgber if (sfptMem == NULL || sfptMem->is_top()) { 469240489Sgber DEBUG_ONLY(disq_node = use;) 470240489Sgber NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) 471240489Sgber can_eliminate = false; 472240489Sgber } else { 473240489Sgber safepoints.append_if_missing(sfpt); 474240489Sgber } 475240489Sgber } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark 476240489Sgber if (use->is_Phi()) { 477240489Sgber if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { 478240489Sgber NOT_PRODUCT(fail_eliminate = "Object is return value";) 479240489Sgber } else { 480240489Sgber NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) 481240489Sgber } 482240489Sgber DEBUG_ONLY(disq_node = use;) 483240489Sgber } else { 484240489Sgber if (use->Opcode() == Op_Return) { 485240489Sgber NOT_PRODUCT(fail_eliminate = "Object is return value";) 486240489Sgber }else { 487240489Sgber NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) 488240489Sgber } 489240489Sgber DEBUG_ONLY(disq_node = use;) 490240489Sgber } 491240489Sgber can_eliminate = false; 492240489Sgber } 493240489Sgber } 494240489Sgber } 495240489Sgber 496240489Sgber#ifndef PRODUCT 497240489Sgber if (PrintEliminateAllocations) { 498240489Sgber if (can_eliminate) { 499240489Sgber tty->print("Scalar "); 500240489Sgber if (res == NULL) 501240489Sgber alloc->dump(); 502240489Sgber else 503240489Sgber res->dump(); 504240489Sgber } else { 505240489Sgber tty->print("NotScalar (%s)", fail_eliminate); 506240489Sgber if (res == NULL) 507185089Sraj alloc->dump(); 508209131Sraj else 509185089Sraj res->dump(); 510185089Sraj#ifdef ASSERT 511240489Sgber if (disq_node != NULL) { 512185089Sraj tty->print(" >>>> "); 513185089Sraj disq_node->dump(); 514185089Sraj } 515240489Sgber#endif /*ASSERT*/ 516240489Sgber } 517240489Sgber } 518240489Sgber#endif 519240489Sgber return can_eliminate; 520240489Sgber} 521209131Sraj 522185089Sraj// Do scalar replacement. 523185089Srajbool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { 524185089Sraj GrowableArray <SafePointNode *> safepoints_done; 525185089Sraj 526185089Sraj ciKlass* klass = NULL; 527185089Sraj ciInstanceKlass* iklass = NULL; 528240489Sgber int nfields = 0; 529240489Sgber int array_base; 530185089Sraj int element_size; 531185089Sraj BasicType basic_elem_type; 532185089Sraj ciType* elem_type; 533240489Sgber 534240489Sgber Node* res = alloc->result_cast(); 535185089Sraj const TypeOopPtr* res_type = NULL; 536209131Sraj if (res != NULL) { // Could be NULL when there are no users 537185089Sraj res_type = _igvn.type(res)->isa_oopptr(); 538209131Sraj } 539185089Sraj 540185089Sraj if (res != NULL) { 541185089Sraj klass = res_type->klass(); 542185089Sraj if (res_type->isa_instptr()) { 543185089Sraj // find the fields of the class which will be needed for safepoint debug information 544185089Sraj assert(klass->is_instance_klass(), "must be an instance klass."); 545209131Sraj iklass = klass->as_instance_klass(); 546185089Sraj nfields = iklass->nof_nonstatic_fields(); 547185089Sraj } else { 548185089Sraj // find the array's elements which will be needed for safepoint debug information 549185089Sraj nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); 550185089Sraj assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); 551209131Sraj elem_type = klass->as_array_klass()->element_type(); 552209131Sraj basic_elem_type = elem_type->basic_type(); 553209131Sraj array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 554209131Sraj element_size = type2aelembytes(basic_elem_type); 555185089Sraj } 556185089Sraj } 557209131Sraj // 558185639Sraj // Process the safepoint uses 559209131Sraj // 560185639Sraj while (safepoints.length() > 0) { 561209131Sraj SafePointNode* sfpt = safepoints.pop(); 562185089Sraj Node* mem = sfpt->memory(); 563209131Sraj uint first_ind = sfpt->req(); 564185089Sraj SafePointScalarObjectNode* sobj = new (C, 1) SafePointScalarObjectNode(res_type, 565185089Sraj#ifdef ASSERT 566185089Sraj alloc, 567209131Sraj#endif 568185089Sraj first_ind, nfields); 569209131Sraj sobj->init_req(0, sfpt->in(TypeFunc::Control)); 570185089Sraj transform_later(sobj); 571185089Sraj 572185089Sraj // Scan object's fields adding an input to the safepoint for each field. 573209131Sraj for (int j = 0; j < nfields; j++) { 574185089Sraj int offset; 575209131Sraj ciField* field = NULL; 576185089Sraj if (iklass != NULL) { 577209131Sraj field = iklass->nonstatic_field_at(j); 578185089Sraj offset = field->offset(); 579209131Sraj elem_type = field->type(); 580185089Sraj basic_elem_type = field->layout_type(); 581185089Sraj } else { 582209131Sraj offset = array_base + j * element_size; 583185089Sraj } 584185089Sraj 585185089Sraj const Type *field_type; 586209131Sraj // The next code is taken from Parse::do_get_xxx(). 587185089Sraj if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { 588185089Sraj if (!elem_type->is_loaded()) { 589185089Sraj field_type = TypeInstPtr::BOTTOM; 590209131Sraj } else if (field != NULL && field->is_constant()) { 591185089Sraj // This can happen if the constant oop is non-perm. 592185089Sraj ciObject* con = field->constant_value().as_object(); 593185089Sraj // Do not "join" in the previous type; it doesn't add value, 594185089Sraj // and may yield a vacuous result if the field is of interface type. 595185089Sraj field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); 596185089Sraj assert(field_type != NULL, "field singleton type must be consistent"); 597185089Sraj } else { 598209131Sraj field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); 599185089Sraj } 600185089Sraj if (UseCompressedOops) { 601185089Sraj field_type = field_type->is_oopptr()->make_narrowoop(); 602185089Sraj basic_elem_type = T_NARROWOOP; 603185089Sraj } 604185089Sraj } else { 605185089Sraj field_type = Type::get_const_basic_type(basic_elem_type); 606185089Sraj } 607209131Sraj 608185089Sraj const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr(); 609185089Sraj 610209131Sraj Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc); 611185089Sraj if (field_val == NULL) { 612185089Sraj // we weren't able to find a value for this field, 613209131Sraj // give up on eliminating this allocation 614185089Sraj alloc->_is_scalar_replaceable = false; // don't try again 615185089Sraj // remove any extra entries we added to the safepoint 616185089Sraj uint last = sfpt->req() - 1; 617185089Sraj for (int k = 0; k < j; k++) { 618185089Sraj sfpt->del_req(last--); 619185089Sraj } 620185089Sraj // rollback processed safepoints 621209131Sraj while (safepoints_done.length() > 0) { 622185089Sraj SafePointNode* sfpt_done = safepoints_done.pop(); 623185089Sraj // remove any extra entries we added to the safepoint 624185089Sraj last = sfpt_done->req() - 1; 625209131Sraj for (int k = 0; k < nfields; k++) { 626185089Sraj sfpt_done->del_req(last--); 627209131Sraj } 628185089Sraj JVMState *jvms = sfpt_done->jvms(); 629185089Sraj jvms->set_endoff(sfpt_done->req()); 630185089Sraj // Now make a pass over the debug information replacing any references 631185089Sraj // to SafePointScalarObjectNode with the allocated object. 632185089Sraj int start = jvms->debug_start(); 633185089Sraj int end = jvms->debug_end(); 634209131Sraj for (int i = start; i < end; i++) { 635185089Sraj if (sfpt_done->in(i)->is_SafePointScalarObject()) { 636185089Sraj SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); 637185089Sraj if (scobj->first_index() == sfpt_done->req() && 638185089Sraj scobj->n_fields() == (uint)nfields) { 639185089Sraj assert(scobj->alloc() == alloc, "sanity"); 640185089Sraj sfpt_done->set_req(i, res); 641185089Sraj } 642185089Sraj } 643185089Sraj } 644209131Sraj } 645209131Sraj#ifndef PRODUCT 646209131Sraj if (PrintEliminateAllocations) { 647209131Sraj if (field != NULL) { 648209131Sraj tty->print("=== At SafePoint node %d can't find value of Field: ", 649209131Sraj sfpt->_idx); 650209131Sraj field->print(); 651209131Sraj int field_idx = C->get_alias_index(field_addr_type); 652209131Sraj tty->print(" (alias_idx=%d)", field_idx); 653209131Sraj } else { // Array's element 654209131Sraj tty->print("=== At SafePoint node %d can't find value of array element [%d]", 655209131Sraj sfpt->_idx, j); 656209131Sraj } 657209131Sraj tty->print(", which prevents elimination of: "); 658209131Sraj if (res == NULL) 659209131Sraj alloc->dump(); 660209131Sraj else 661209131Sraj res->dump(); 662209131Sraj } 663209131Sraj#endif 664209131Sraj return false; 665209131Sraj } 666209131Sraj if (UseCompressedOops && field_type->isa_narrowoop()) { 667185089Sraj // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation 668209131Sraj // to be able scalar replace the allocation. 669185089Sraj _igvn.set_delay_transform(false); 670185089Sraj field_val = DecodeNNode::decode(&_igvn, field_val); 671209131Sraj _igvn.set_delay_transform(true); 672186932Sraj } 673186932Sraj sfpt->add_req(field_val); 674185089Sraj } 675186932Sraj JVMState *jvms = sfpt->jvms(); 676186932Sraj jvms->set_endoff(sfpt->req()); 677209131Sraj // Now make a pass over the debug information replacing any references 678186932Sraj // to the allocated object with "sobj" 679186932Sraj int start = jvms->debug_start(); 680209131Sraj int end = jvms->debug_end(); 681186932Sraj for (int i = start; i < end; i++) { 682186932Sraj if (sfpt->in(i) == res) { 683240489Sgber sfpt->set_req(i, sobj); 684186932Sraj } 685186932Sraj } 686186932Sraj safepoints_done.append_if_missing(sfpt); // keep it for rollback 687240489Sgber } 688240489Sgber return true; 689240489Sgber} 690240489Sgber 691240489Sgber// Process users of eliminated allocation. 692240489Sgbervoid PhaseMacroExpand::process_users_of_allocation(AllocateNode *alloc) { 693240489Sgber Node* res = alloc->result_cast(); 694240489Sgber if (res != NULL) { 695240489Sgber for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { 696240489Sgber Node *use = res->last_out(j); 697186932Sraj uint oc1 = res->outcnt(); 698186932Sraj 699186932Sraj if (use->is_AddP()) { 700186932Sraj for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { 701186932Sraj Node *n = use->last_out(k); 702209131Sraj uint oc2 = use->outcnt(); 703186932Sraj if (n->is_Store()) { 704186932Sraj _igvn.replace_node(n, n->in(MemNode::Memory)); 705186932Sraj } else { 706186932Sraj assert( n->Opcode() == Op_CastP2X, "CastP2X required"); 707186932Sraj eliminate_card_mark(n); 708186932Sraj } 709186932Sraj k -= (oc2 - use->outcnt()); 710186932Sraj } 711186932Sraj } else { 712185089Sraj assert( !use->is_SafePoint(), "safepoint uses must have been already elimiated"); 713185089Sraj assert( use->Opcode() == Op_CastP2X, "CastP2X required"); 714185089Sraj eliminate_card_mark(use); 715209131Sraj } 716185089Sraj j -= (oc1 - res->outcnt()); 717185089Sraj } 718185089Sraj assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); 719186932Sraj _igvn.remove_dead_node(res); 720186932Sraj } 721186932Sraj 722186932Sraj // 723186932Sraj // Process other users of allocation's projections 724185089Sraj // 725185089Sraj if (_resproj != NULL && _resproj->outcnt() != 0) { 726185089Sraj for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { 727209131Sraj Node *use = _resproj->last_out(j); 728185089Sraj uint oc1 = _resproj->outcnt(); 729209131Sraj if (use->is_Initialize()) { 730185089Sraj // Eliminate Initialize node. 731185089Sraj InitializeNode *init = use->as_Initialize(); 732185089Sraj assert(init->outcnt() <= 2, "only a control and memory projection expected"); 733185089Sraj Node *ctrl_proj = init->proj_out(TypeFunc::Control); 734185089Sraj if (ctrl_proj != NULL) { 735185089Sraj assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection"); 736185089Sraj _igvn.replace_node(ctrl_proj, _fallthroughcatchproj); 737185089Sraj } 738185089Sraj Node *mem_proj = init->proj_out(TypeFunc::Memory); 739185089Sraj if (mem_proj != NULL) { 740185089Sraj Node *mem = init->in(TypeFunc::Memory); 741185089Sraj#ifdef ASSERT 742185089Sraj if (mem->is_MergeMem()) { 743185089Sraj assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection"); 744209131Sraj } else { 745185089Sraj assert(mem == _memproj_fallthrough, "allocation memory projection"); 746209131Sraj } 747185089Sraj#endif 748185089Sraj _igvn.replace_node(mem_proj, mem); 749185089Sraj } 750185089Sraj } else if (use->is_AddP()) { 751185089Sraj // raw memory addresses used only by the initialization 752185089Sraj _igvn.hash_delete(use); 753185089Sraj _igvn.subsume_node(use, C->top()); 754185089Sraj } else { 755185089Sraj assert(false, "only Initialize or AddP expected"); 756209131Sraj } 757209131Sraj j -= (oc1 - _resproj->outcnt()); 758209131Sraj } 759209131Sraj } 760209131Sraj if (_fallthroughcatchproj != NULL) { 761209131Sraj _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); 762209131Sraj } 763209131Sraj if (_memproj_fallthrough != NULL) { 764209131Sraj _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); 765209131Sraj } 766209131Sraj if (_memproj_catchall != NULL) { 767209131Sraj _igvn.replace_node(_memproj_catchall, C->top()); 768209131Sraj } 769209131Sraj if (_ioproj_fallthrough != NULL) { 770209131Sraj _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); 771209131Sraj } 772209131Sraj if (_ioproj_catchall != NULL) { 773209131Sraj _igvn.replace_node(_ioproj_catchall, C->top()); 774209131Sraj } 775209131Sraj if (_catchallcatchproj != NULL) { 776209131Sraj _igvn.replace_node(_catchallcatchproj, C->top()); 777209131Sraj } 778209131Sraj} 779209131Sraj 780209131Srajbool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { 781209131Sraj 782209131Sraj if (!EliminateAllocations || !alloc->_is_scalar_replaceable) { 783209131Sraj return false; 784209131Sraj } 785209131Sraj 786209131Sraj extract_call_projections(alloc); 787209131Sraj 788209131Sraj GrowableArray <SafePointNode *> safepoints; 789209131Sraj if (!can_eliminate_allocation(alloc, safepoints)) { 790209131Sraj return false; 791209131Sraj } 792209131Sraj 793209131Sraj if (!scalar_replacement(alloc, safepoints)) { 794209131Sraj return false; 795209131Sraj } 796209131Sraj 797209131Sraj process_users_of_allocation(alloc); 798209131Sraj 799209131Sraj#ifndef PRODUCT 800209131Srajif (PrintEliminateAllocations) { 801209131Sraj if (alloc->is_AllocateArray()) 802209131Sraj tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); 803209131Sraj else 804209131Sraj tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); 805209131Sraj} 806209131Sraj#endif 807209131Sraj 808209131Sraj return true; 809209131Sraj} 810209131Sraj 811209131Sraj 812209131Sraj//---------------------------set_eden_pointers------------------------- 813209131Srajvoid PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { 814209131Sraj if (UseTLAB) { // Private allocation: load from TLS 815209131Sraj Node* thread = transform_later(new (C, 1) ThreadLocalNode()); 816209131Sraj int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); 817209131Sraj int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); 818209131Sraj eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); 819209131Sraj eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); 820209131Sraj } else { // Shared allocation: load from globals 821209131Sraj CollectedHeap* ch = Universe::heap(); 822209131Sraj address top_adr = (address)ch->top_addr(); 823209131Sraj address end_adr = (address)ch->end_addr(); 824209131Sraj eden_top_adr = makecon(TypeRawPtr::make(top_adr)); 825209131Sraj eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); 826209131Sraj } 827209131Sraj} 828209131Sraj 829209131Sraj 830209131SrajNode* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { 831209131Sraj Node* adr = basic_plus_adr(base, offset); 832209131Sraj const TypePtr* adr_type = TypeRawPtr::BOTTOM; 833209131Sraj Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt); 834209131Sraj transform_later(value); 835209131Sraj return value; 836209131Sraj} 837209131Sraj 838209131Sraj 839209131SrajNode* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { 840209131Sraj Node* adr = basic_plus_adr(base, offset); 841209131Sraj mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt); 842209131Sraj transform_later(mem); 843209131Sraj return mem; 844209131Sraj} 845209131Sraj 846209131Sraj//============================================================================= 847209131Sraj// 848209131Sraj// A L L O C A T I O N 849209131Sraj// 850209131Sraj// Allocation attempts to be fast in the case of frequent small objects. 851209131Sraj// It breaks down like this: 852209131Sraj// 853209131Sraj// 1) Size in doublewords is computed. This is a constant for objects and 854209131Sraj// variable for most arrays. Doubleword units are used to avoid size 855209131Sraj// overflow of huge doubleword arrays. We need doublewords in the end for 856209131Sraj// rounding. 857209131Sraj// 858209131Sraj// 2) Size is checked for being 'too large'. Too-large allocations will go 859209131Sraj// the slow path into the VM. The slow path can throw any required 860209131Sraj// exceptions, and does all the special checks for very large arrays. The 861240489Sgber// size test can constant-fold away for objects. For objects with 862240489Sgber// finalizers it constant-folds the otherway: you always go slow with 863240489Sgber// finalizers. 864209131Sraj// 865209131Sraj// 3) If NOT using TLABs, this is the contended loop-back point. 866209131Sraj// Load-Locked the heap top. If using TLABs normal-load the heap top. 867209131Sraj// 868209131Sraj// 4) Check that heap top + size*8 < max. If we fail go the slow ` route. 869209131Sraj// NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish 870209131Sraj// "size*8" we always enter the VM, where "largish" is a constant picked small 871209131Sraj// enough that there's always space between the eden max and 4Gig (old space is 872209131Sraj// there so it's quite large) and large enough that the cost of entering the VM 873209131Sraj// is dwarfed by the cost to initialize the space. 874209131Sraj// 875240489Sgber// 5) If NOT using TLABs, Store-Conditional the adjusted heap top back 876240489Sgber// down. If contended, repeat at step 3. If using TLABs normal-store 877240489Sgber// adjusted heap top back down; there is no contention. 878209131Sraj// 879209131Sraj// 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark 880209131Sraj// fields. 881209131Sraj// 882209131Sraj// 7) Merge with the slow-path; cast the raw memory pointer to the correct 883209131Sraj// oop flavor. 884209131Sraj// 885209131Sraj//============================================================================= 886209131Sraj// FastAllocateSizeLimit value is in DOUBLEWORDS. 887209131Sraj// Allocations bigger than this always go the slow route. 888209131Sraj// This value must be small enough that allocation attempts that need to 889209131Sraj// trigger exceptions go the slow route. Also, it must be small enough so 890209131Sraj// that heap_top + size_in_bytes does not wrap around the 4Gig limit. 891209131Sraj//=============================================================================j// 892209131Sraj// %%% Here is an old comment from parseHelper.cpp; is it outdated? 893209131Sraj// The allocator will coalesce int->oop copies away. See comment in 894209131Sraj// coalesce.cpp about how this works. It depends critically on the exact 895209131Sraj// code shape produced here, so if you are changing this code shape 896209131Sraj// make sure the GC info for the heap-top is correct in and around the 897209131Sraj// slow-path call. 898209131Sraj// 899209131Sraj 900209131Srajvoid PhaseMacroExpand::expand_allocate_common( 901209131Sraj AllocateNode* alloc, // allocation node to be expanded 902209131Sraj Node* length, // array length for an array allocation 903209131Sraj const TypeFunc* slow_call_type, // Type of slow call 904209131Sraj address slow_call_address // Address of slow call 905209131Sraj ) 906209131Sraj{ 907209131Sraj 908209131Sraj Node* ctrl = alloc->in(TypeFunc::Control); 909209131Sraj Node* mem = alloc->in(TypeFunc::Memory); 910209131Sraj Node* i_o = alloc->in(TypeFunc::I_O); 911209131Sraj Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); 912209131Sraj Node* klass_node = alloc->in(AllocateNode::KlassNode); 913209131Sraj Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); 914209131Sraj 915209131Sraj // With escape analysis, the entire memory state was needed to be able to 916240489Sgber // eliminate the allocation. Since the allocations cannot be eliminated, 917240489Sgber // optimize it to the raw slice. 918209131Sraj if (mem->is_MergeMem()) { 919209131Sraj mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 920209131Sraj } 921209131Sraj 922209131Sraj Node* eden_top_adr; 923240489Sgber Node* eden_end_adr; 924240489Sgber set_eden_pointers(eden_top_adr, eden_end_adr); 925240489Sgber 926209131Sraj uint raw_idx = C->get_alias_index(TypeRawPtr::BOTTOM); 927209131Sraj assert(ctrl != NULL, "must have control"); 928209131Sraj 929209131Sraj // Load Eden::end. Loop invariant and hoisted. 930209131Sraj // 931209131Sraj // Note: We set the control input on "eden_end" and "old_eden_top" when using 932209131Sraj // a TLAB to work around a bug where these values were being moved across 933209131Sraj // a safepoint. These are not oops, so they cannot be include in the oop 934209131Sraj // map, but the can be changed by a GC. The proper way to fix this would 935209131Sraj // be to set the raw memory state when generating a SafepointNode. However 936209131Sraj // this will require extensive changes to the loop optimization in order to 937209131Sraj // prevent a degradation of the optimization. 938209131Sraj // See comment in memnode.hpp, around line 227 in class LoadPNode. 939209131Sraj Node* eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); 940209131Sraj 941209131Sraj // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. 942209131Sraj // they will not be used if "always_slow" is set 943209131Sraj enum { slow_result_path = 1, fast_result_path = 2 }; 944209131Sraj Node *result_region; 945209131Sraj Node *result_phi_rawmem; 946209131Sraj Node *result_phi_rawoop; 947209131Sraj Node *result_phi_i_o; 948209131Sraj 949209131Sraj // The initial slow comparison is a size check, the comparison 950209131Sraj // we want to do is a BoolTest::gt 951209131Sraj bool always_slow = false; 952240493Sgber int tv = _igvn.find_int_con(initial_slow_test, -1); 953240493Sgber if (tv >= 0) { 954240493Sgber always_slow = (tv == 1); 955240493Sgber initial_slow_test = NULL; 956240493Sgber } else { 957240493Sgber initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); 958240493Sgber } 959240493Sgber 960240493Sgber if (DTraceAllocProbes) { 961240493Sgber // Force slow-path allocation 962240493Sgber always_slow = true; 963240493Sgber initial_slow_test = NULL; 964240493Sgber } 965240493Sgber 966240493Sgber enum { too_big_or_final_path = 1, need_gc_path = 2 }; 967240493Sgber Node *slow_region = NULL; 968240493Sgber Node *toobig_false = ctrl; 969240493Sgber 970240493Sgber assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); 971240493Sgber // generate the initial test if necessary 972240493Sgber if (initial_slow_test != NULL ) { 973240493Sgber slow_region = new (C, 3) RegionNode(3); 974240493Sgber 975240493Sgber // Now make the initial failure test. Usually a too-big test but 976240493Sgber // might be a TRUE for finalizers or a fancy class check for 977240493Sgber // newInstance0. 978240493Sgber IfNode *toobig_iff = new (C, 2) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 979240493Sgber transform_later(toobig_iff); 980240493Sgber // Plug the failing-too-big test into the slow-path region 981240493Sgber Node *toobig_true = new (C, 1) IfTrueNode( toobig_iff ); 982240493Sgber transform_later(toobig_true); 983240493Sgber slow_region ->init_req( too_big_or_final_path, toobig_true ); 984240493Sgber toobig_false = new (C, 1) IfFalseNode( toobig_iff ); 985240493Sgber transform_later(toobig_false); 986240493Sgber } else { // No initial test, just fall into next case 987240493Sgber toobig_false = ctrl; 988240493Sgber debug_only(slow_region = NodeSentinel); 989240493Sgber } 990240493Sgber 991240493Sgber Node *slow_mem = mem; // save the current memory state for slow path 992240493Sgber // generate the fast allocation code unless we know that the initial test will always go slow 993240493Sgber if (!always_slow) { 994240493Sgber // allocate the Region and Phi nodes for the result 995240493Sgber result_region = new (C, 3) RegionNode(3); 996240493Sgber result_phi_rawmem = new (C, 3) PhiNode( result_region, Type::MEMORY, TypeRawPtr::BOTTOM ); 997240493Sgber result_phi_rawoop = new (C, 3) PhiNode( result_region, TypeRawPtr::BOTTOM ); 998240493Sgber result_phi_i_o = new (C, 3) PhiNode( result_region, Type::ABIO ); // I/O is used for Prefetch 999240493Sgber 1000240493Sgber // We need a Region for the loop-back contended case. 1001240493Sgber enum { fall_in_path = 1, contended_loopback_path = 2 }; 1002240493Sgber Node *contended_region; 1003240493Sgber Node *contended_phi_rawmem; 1004240493Sgber if( UseTLAB ) { 1005240493Sgber contended_region = toobig_false; 1006240493Sgber contended_phi_rawmem = mem; 1007240493Sgber } else { 1008240493Sgber contended_region = new (C, 3) RegionNode(3); 1009240493Sgber contended_phi_rawmem = new (C, 3) PhiNode( contended_region, Type::MEMORY, TypeRawPtr::BOTTOM); 1010240493Sgber // Now handle the passing-too-big test. We fall into the contended 1011240493Sgber // loop-back merge point. 1012240493Sgber contended_region ->init_req( fall_in_path, toobig_false ); 1013240493Sgber contended_phi_rawmem->init_req( fall_in_path, mem ); 1014240493Sgber transform_later(contended_region); 1015240493Sgber transform_later(contended_phi_rawmem); 1016240493Sgber } 1017240493Sgber 1018240493Sgber // Load(-locked) the heap top. 1019240493Sgber // See note above concerning the control input when using a TLAB 1020240493Sgber Node *old_eden_top = UseTLAB 1021240493Sgber ? new (C, 3) LoadPNode ( ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM ) 1022240493Sgber : new (C, 3) LoadPLockedNode( contended_region, contended_phi_rawmem, eden_top_adr ); 1023240493Sgber 1024240493Sgber transform_later(old_eden_top); 1025240493Sgber // Add to heap top to get a new heap top 1026240493Sgber Node *new_eden_top = new (C, 4) AddPNode( top(), old_eden_top, size_in_bytes ); 1027240493Sgber transform_later(new_eden_top); 1028240493Sgber // Check for needing a GC; compare against heap end 1029 Node *needgc_cmp = new (C, 3) CmpPNode( new_eden_top, eden_end ); 1030 transform_later(needgc_cmp); 1031 Node *needgc_bol = new (C, 2) BoolNode( needgc_cmp, BoolTest::ge ); 1032 transform_later(needgc_bol); 1033 IfNode *needgc_iff = new (C, 2) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1034 transform_later(needgc_iff); 1035 1036 // Plug the failing-heap-space-need-gc test into the slow-path region 1037 Node *needgc_true = new (C, 1) IfTrueNode( needgc_iff ); 1038 transform_later(needgc_true); 1039 if( initial_slow_test ) { 1040 slow_region ->init_req( need_gc_path, needgc_true ); 1041 // This completes all paths into the slow merge point 1042 transform_later(slow_region); 1043 } else { // No initial slow path needed! 1044 // Just fall from the need-GC path straight into the VM call. 1045 slow_region = needgc_true; 1046 } 1047 // No need for a GC. Setup for the Store-Conditional 1048 Node *needgc_false = new (C, 1) IfFalseNode( needgc_iff ); 1049 transform_later(needgc_false); 1050 1051 // Grab regular I/O before optional prefetch may change it. 1052 // Slow-path does no I/O so just set it to the original I/O. 1053 result_phi_i_o->init_req( slow_result_path, i_o ); 1054 1055 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem, 1056 old_eden_top, new_eden_top, length); 1057 1058 // Store (-conditional) the modified eden top back down. 1059 // StorePConditional produces flags for a test PLUS a modified raw 1060 // memory state. 1061 Node *store_eden_top; 1062 Node *fast_oop_ctrl; 1063 if( UseTLAB ) { 1064 store_eden_top = new (C, 4) StorePNode( needgc_false, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, new_eden_top ); 1065 transform_later(store_eden_top); 1066 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path 1067 } else { 1068 store_eden_top = new (C, 5) StorePConditionalNode( needgc_false, contended_phi_rawmem, eden_top_adr, new_eden_top, old_eden_top ); 1069 transform_later(store_eden_top); 1070 Node *contention_check = new (C, 2) BoolNode( store_eden_top, BoolTest::ne ); 1071 transform_later(contention_check); 1072 store_eden_top = new (C, 1) SCMemProjNode(store_eden_top); 1073 transform_later(store_eden_top); 1074 1075 // If not using TLABs, check to see if there was contention. 1076 IfNode *contention_iff = new (C, 2) IfNode ( needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN ); 1077 transform_later(contention_iff); 1078 Node *contention_true = new (C, 1) IfTrueNode( contention_iff ); 1079 transform_later(contention_true); 1080 // If contention, loopback and try again. 1081 contended_region->init_req( contended_loopback_path, contention_true ); 1082 contended_phi_rawmem->init_req( contended_loopback_path, store_eden_top ); 1083 1084 // Fast-path succeeded with no contention! 1085 Node *contention_false = new (C, 1) IfFalseNode( contention_iff ); 1086 transform_later(contention_false); 1087 fast_oop_ctrl = contention_false; 1088 } 1089 1090 // Rename successful fast-path variables to make meaning more obvious 1091 Node* fast_oop = old_eden_top; 1092 Node* fast_oop_rawmem = store_eden_top; 1093 fast_oop_rawmem = initialize_object(alloc, 1094 fast_oop_ctrl, fast_oop_rawmem, fast_oop, 1095 klass_node, length, size_in_bytes); 1096 1097 if (ExtendedDTraceProbes) { 1098 // Slow-path call 1099 int size = TypeFunc::Parms + 2; 1100 CallLeafNode *call = new (C, size) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), 1101 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), 1102 "dtrace_object_alloc", 1103 TypeRawPtr::BOTTOM); 1104 1105 // Get base of thread-local storage area 1106 Node* thread = new (C, 1) ThreadLocalNode(); 1107 transform_later(thread); 1108 1109 call->init_req(TypeFunc::Parms+0, thread); 1110 call->init_req(TypeFunc::Parms+1, fast_oop); 1111 call->init_req( TypeFunc::Control, fast_oop_ctrl ); 1112 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1113 call->init_req( TypeFunc::Memory , fast_oop_rawmem ); 1114 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1115 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1116 transform_later(call); 1117 fast_oop_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control); 1118 transform_later(fast_oop_ctrl); 1119 fast_oop_rawmem = new (C, 1) ProjNode(call,TypeFunc::Memory); 1120 transform_later(fast_oop_rawmem); 1121 } 1122 1123 // Plug in the successful fast-path into the result merge point 1124 result_region ->init_req( fast_result_path, fast_oop_ctrl ); 1125 result_phi_rawoop->init_req( fast_result_path, fast_oop ); 1126 result_phi_i_o ->init_req( fast_result_path, i_o ); 1127 result_phi_rawmem->init_req( fast_result_path, fast_oop_rawmem ); 1128 } else { 1129 slow_region = ctrl; 1130 } 1131 1132 // Generate slow-path call 1133 CallNode *call = new (C, slow_call_type->domain()->cnt()) 1134 CallStaticJavaNode(slow_call_type, slow_call_address, 1135 OptoRuntime::stub_name(slow_call_address), 1136 alloc->jvms()->bci(), 1137 TypePtr::BOTTOM); 1138 call->init_req( TypeFunc::Control, slow_region ); 1139 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o 1140 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs 1141 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); 1142 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); 1143 1144 call->init_req(TypeFunc::Parms+0, klass_node); 1145 if (length != NULL) { 1146 call->init_req(TypeFunc::Parms+1, length); 1147 } 1148 1149 // Copy debug information and adjust JVMState information, then replace 1150 // allocate node with the call 1151 copy_call_debug_info((CallNode *) alloc, call); 1152 if (!always_slow) { 1153 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. 1154 } 1155 _igvn.hash_delete(alloc); 1156 _igvn.subsume_node(alloc, call); 1157 transform_later(call); 1158 1159 // Identify the output projections from the allocate node and 1160 // adjust any references to them. 1161 // The control and io projections look like: 1162 // 1163 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) 1164 // Allocate Catch 1165 // ^---Proj(io) <-------+ ^---CatchProj(io) 1166 // 1167 // We are interested in the CatchProj nodes. 1168 // 1169 extract_call_projections(call); 1170 1171 // An allocate node has separate memory projections for the uses on the control and i_o paths 1172 // Replace uses of the control memory projection with result_phi_rawmem (unless we are only generating a slow call) 1173 if (!always_slow && _memproj_fallthrough != NULL) { 1174 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { 1175 Node *use = _memproj_fallthrough->fast_out(i); 1176 _igvn.hash_delete(use); 1177 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); 1178 _igvn._worklist.push(use); 1179 // back up iterator 1180 --i; 1181 } 1182 } 1183 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete _memproj_catchall so 1184 // we end up with a call that has only 1 memory projection 1185 if (_memproj_catchall != NULL ) { 1186 if (_memproj_fallthrough == NULL) { 1187 _memproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::Memory); 1188 transform_later(_memproj_fallthrough); 1189 } 1190 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { 1191 Node *use = _memproj_catchall->fast_out(i); 1192 _igvn.hash_delete(use); 1193 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); 1194 _igvn._worklist.push(use); 1195 // back up iterator 1196 --i; 1197 } 1198 } 1199 1200 mem = result_phi_rawmem; 1201 1202 // An allocate node has separate i_o projections for the uses on the control and i_o paths 1203 // Replace uses of the control i_o projection with result_phi_i_o (unless we are only generating a slow call) 1204 if (_ioproj_fallthrough == NULL) { 1205 _ioproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::I_O); 1206 transform_later(_ioproj_fallthrough); 1207 } else if (!always_slow) { 1208 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { 1209 Node *use = _ioproj_fallthrough->fast_out(i); 1210 1211 _igvn.hash_delete(use); 1212 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); 1213 _igvn._worklist.push(use); 1214 // back up iterator 1215 --i; 1216 } 1217 } 1218 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete _ioproj_catchall so 1219 // we end up with a call that has only 1 control projection 1220 if (_ioproj_catchall != NULL ) { 1221 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { 1222 Node *use = _ioproj_catchall->fast_out(i); 1223 _igvn.hash_delete(use); 1224 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); 1225 _igvn._worklist.push(use); 1226 // back up iterator 1227 --i; 1228 } 1229 } 1230 1231 // if we generated only a slow call, we are done 1232 if (always_slow) 1233 return; 1234 1235 1236 if (_fallthroughcatchproj != NULL) { 1237 ctrl = _fallthroughcatchproj->clone(); 1238 transform_later(ctrl); 1239 _igvn.hash_delete(_fallthroughcatchproj); 1240 _igvn.subsume_node(_fallthroughcatchproj, result_region); 1241 } else { 1242 ctrl = top(); 1243 } 1244 Node *slow_result; 1245 if (_resproj == NULL) { 1246 // no uses of the allocation result 1247 slow_result = top(); 1248 } else { 1249 slow_result = _resproj->clone(); 1250 transform_later(slow_result); 1251 _igvn.hash_delete(_resproj); 1252 _igvn.subsume_node(_resproj, result_phi_rawoop); 1253 } 1254 1255 // Plug slow-path into result merge point 1256 result_region ->init_req( slow_result_path, ctrl ); 1257 result_phi_rawoop->init_req( slow_result_path, slow_result); 1258 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); 1259 transform_later(result_region); 1260 transform_later(result_phi_rawoop); 1261 transform_later(result_phi_rawmem); 1262 transform_later(result_phi_i_o); 1263 // This completes all paths into the result merge point 1264} 1265 1266 1267// Helper for PhaseMacroExpand::expand_allocate_common. 1268// Initializes the newly-allocated storage. 1269Node* 1270PhaseMacroExpand::initialize_object(AllocateNode* alloc, 1271 Node* control, Node* rawmem, Node* object, 1272 Node* klass_node, Node* length, 1273 Node* size_in_bytes) { 1274 InitializeNode* init = alloc->initialization(); 1275 // Store the klass & mark bits 1276 Node* mark_node = NULL; 1277 // For now only enable fast locking for non-array types 1278 if (UseBiasedLocking && (length == NULL)) { 1279 mark_node = make_load(NULL, rawmem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeRawPtr::BOTTOM, T_ADDRESS); 1280 } else { 1281 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); 1282 } 1283 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); 1284 1285 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_OBJECT); 1286 int header_size = alloc->minimum_header_size(); // conservatively small 1287 1288 // Array length 1289 if (length != NULL) { // Arrays need length field 1290 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); 1291 // conservatively small header size: 1292 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1293 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); 1294 if (k->is_array_klass()) // we know the exact header size in most cases: 1295 header_size = Klass::layout_helper_header_size(k->layout_helper()); 1296 } 1297 1298 // Clear the object body, if necessary. 1299 if (init == NULL) { 1300 // The init has somehow disappeared; be cautious and clear everything. 1301 // 1302 // This can happen if a node is allocated but an uncommon trap occurs 1303 // immediately. In this case, the Initialize gets associated with the 1304 // trap, and may be placed in a different (outer) loop, if the Allocate 1305 // is in a loop. If (this is rare) the inner loop gets unrolled, then 1306 // there can be two Allocates to one Initialize. The answer in all these 1307 // edge cases is safety first. It is always safe to clear immediately 1308 // within an Allocate, and then (maybe or maybe not) clear some more later. 1309 if (!ZeroTLAB) 1310 rawmem = ClearArrayNode::clear_memory(control, rawmem, object, 1311 header_size, size_in_bytes, 1312 &_igvn); 1313 } else { 1314 if (!init->is_complete()) { 1315 // Try to win by zeroing only what the init does not store. 1316 // We can also try to do some peephole optimizations, 1317 // such as combining some adjacent subword stores. 1318 rawmem = init->complete_stores(control, rawmem, object, 1319 header_size, size_in_bytes, &_igvn); 1320 } 1321 // We have no more use for this link, since the AllocateNode goes away: 1322 init->set_req(InitializeNode::RawAddress, top()); 1323 // (If we keep the link, it just confuses the register allocator, 1324 // who thinks he sees a real use of the address by the membar.) 1325 } 1326 1327 return rawmem; 1328} 1329 1330// Generate prefetch instructions for next allocations. 1331Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, 1332 Node*& contended_phi_rawmem, 1333 Node* old_eden_top, Node* new_eden_top, 1334 Node* length) { 1335 if( UseTLAB && AllocatePrefetchStyle == 2 ) { 1336 // Generate prefetch allocation with watermark check. 1337 // As an allocation hits the watermark, we will prefetch starting 1338 // at a "distance" away from watermark. 1339 enum { fall_in_path = 1, pf_path = 2 }; 1340 1341 Node *pf_region = new (C, 3) RegionNode(3); 1342 Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY, 1343 TypeRawPtr::BOTTOM ); 1344 // I/O is used for Prefetch 1345 Node *pf_phi_abio = new (C, 3) PhiNode( pf_region, Type::ABIO ); 1346 1347 Node *thread = new (C, 1) ThreadLocalNode(); 1348 transform_later(thread); 1349 1350 Node *eden_pf_adr = new (C, 4) AddPNode( top()/*not oop*/, thread, 1351 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); 1352 transform_later(eden_pf_adr); 1353 1354 Node *old_pf_wm = new (C, 3) LoadPNode( needgc_false, 1355 contended_phi_rawmem, eden_pf_adr, 1356 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM ); 1357 transform_later(old_pf_wm); 1358 1359 // check against new_eden_top 1360 Node *need_pf_cmp = new (C, 3) CmpPNode( new_eden_top, old_pf_wm ); 1361 transform_later(need_pf_cmp); 1362 Node *need_pf_bol = new (C, 2) BoolNode( need_pf_cmp, BoolTest::ge ); 1363 transform_later(need_pf_bol); 1364 IfNode *need_pf_iff = new (C, 2) IfNode( needgc_false, need_pf_bol, 1365 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); 1366 transform_later(need_pf_iff); 1367 1368 // true node, add prefetchdistance 1369 Node *need_pf_true = new (C, 1) IfTrueNode( need_pf_iff ); 1370 transform_later(need_pf_true); 1371 1372 Node *need_pf_false = new (C, 1) IfFalseNode( need_pf_iff ); 1373 transform_later(need_pf_false); 1374 1375 Node *new_pf_wmt = new (C, 4) AddPNode( top(), old_pf_wm, 1376 _igvn.MakeConX(AllocatePrefetchDistance) ); 1377 transform_later(new_pf_wmt ); 1378 new_pf_wmt->set_req(0, need_pf_true); 1379 1380 Node *store_new_wmt = new (C, 4) StorePNode( need_pf_true, 1381 contended_phi_rawmem, eden_pf_adr, 1382 TypeRawPtr::BOTTOM, new_pf_wmt ); 1383 transform_later(store_new_wmt); 1384 1385 // adding prefetches 1386 pf_phi_abio->init_req( fall_in_path, i_o ); 1387 1388 Node *prefetch_adr; 1389 Node *prefetch; 1390 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize; 1391 uint step_size = AllocatePrefetchStepSize; 1392 uint distance = 0; 1393 1394 for ( uint i = 0; i < lines; i++ ) { 1395 prefetch_adr = new (C, 4) AddPNode( old_pf_wm, new_pf_wmt, 1396 _igvn.MakeConX(distance) ); 1397 transform_later(prefetch_adr); 1398 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr ); 1399 transform_later(prefetch); 1400 distance += step_size; 1401 i_o = prefetch; 1402 } 1403 pf_phi_abio->set_req( pf_path, i_o ); 1404 1405 pf_region->init_req( fall_in_path, need_pf_false ); 1406 pf_region->init_req( pf_path, need_pf_true ); 1407 1408 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); 1409 pf_phi_rawmem->init_req( pf_path, store_new_wmt ); 1410 1411 transform_later(pf_region); 1412 transform_later(pf_phi_rawmem); 1413 transform_later(pf_phi_abio); 1414 1415 needgc_false = pf_region; 1416 contended_phi_rawmem = pf_phi_rawmem; 1417 i_o = pf_phi_abio; 1418 } else if( AllocatePrefetchStyle > 0 ) { 1419 // Insert a prefetch for each allocation only on the fast-path 1420 Node *prefetch_adr; 1421 Node *prefetch; 1422 // Generate several prefetch instructions only for arrays. 1423 uint lines = (length != NULL) ? AllocatePrefetchLines : 1; 1424 uint step_size = AllocatePrefetchStepSize; 1425 uint distance = AllocatePrefetchDistance; 1426 for ( uint i = 0; i < lines; i++ ) { 1427 prefetch_adr = new (C, 4) AddPNode( old_eden_top, new_eden_top, 1428 _igvn.MakeConX(distance) ); 1429 transform_later(prefetch_adr); 1430 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr ); 1431 // Do not let it float too high, since if eden_top == eden_end, 1432 // both might be null. 1433 if( i == 0 ) { // Set control for first prefetch, next follows it 1434 prefetch->init_req(0, needgc_false); 1435 } 1436 transform_later(prefetch); 1437 distance += step_size; 1438 i_o = prefetch; 1439 } 1440 } 1441 return i_o; 1442} 1443 1444 1445void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { 1446 expand_allocate_common(alloc, NULL, 1447 OptoRuntime::new_instance_Type(), 1448 OptoRuntime::new_instance_Java()); 1449} 1450 1451void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { 1452 Node* length = alloc->in(AllocateNode::ALength); 1453 expand_allocate_common(alloc, length, 1454 OptoRuntime::new_array_Type(), 1455 OptoRuntime::new_array_Java()); 1456} 1457 1458 1459// we have determined that this lock/unlock can be eliminated, we simply 1460// eliminate the node without expanding it. 1461// 1462// Note: The membar's associated with the lock/unlock are currently not 1463// eliminated. This should be investigated as a future enhancement. 1464// 1465bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { 1466 1467 if (!alock->is_eliminated()) { 1468 return false; 1469 } 1470 // Mark the box lock as eliminated if all correspondent locks are eliminated 1471 // to construct correct debug info. 1472 BoxLockNode* box = alock->box_node()->as_BoxLock(); 1473 if (!box->is_eliminated()) { 1474 bool eliminate = true; 1475 for (DUIterator_Fast imax, i = box->fast_outs(imax); i < imax; i++) { 1476 Node *lck = box->fast_out(i); 1477 if (lck->is_Lock() && !lck->as_AbstractLock()->is_eliminated()) { 1478 eliminate = false; 1479 break; 1480 } 1481 } 1482 if (eliminate) 1483 box->set_eliminated(); 1484 } 1485 1486 #ifndef PRODUCT 1487 if (PrintEliminateLocks) { 1488 if (alock->is_Lock()) { 1489 tty->print_cr("++++ Eliminating: %d Lock", alock->_idx); 1490 } else { 1491 tty->print_cr("++++ Eliminating: %d Unlock", alock->_idx); 1492 } 1493 } 1494 #endif 1495 1496 Node* mem = alock->in(TypeFunc::Memory); 1497 Node* ctrl = alock->in(TypeFunc::Control); 1498 1499 extract_call_projections(alock); 1500 // There are 2 projections from the lock. The lock node will 1501 // be deleted when its last use is subsumed below. 1502 assert(alock->outcnt() == 2 && 1503 _fallthroughproj != NULL && 1504 _memproj_fallthrough != NULL, 1505 "Unexpected projections from Lock/Unlock"); 1506 1507 Node* fallthroughproj = _fallthroughproj; 1508 Node* memproj_fallthrough = _memproj_fallthrough; 1509 1510 // The memory projection from a lock/unlock is RawMem 1511 // The input to a Lock is merged memory, so extract its RawMem input 1512 // (unless the MergeMem has been optimized away.) 1513 if (alock->is_Lock()) { 1514 // Seach for MemBarAcquire node and delete it also. 1515 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); 1516 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquire, ""); 1517 Node* ctrlproj = membar->proj_out(TypeFunc::Control); 1518 Node* memproj = membar->proj_out(TypeFunc::Memory); 1519 _igvn.hash_delete(ctrlproj); 1520 _igvn.subsume_node(ctrlproj, fallthroughproj); 1521 _igvn.hash_delete(memproj); 1522 _igvn.subsume_node(memproj, memproj_fallthrough); 1523 } 1524 1525 // Seach for MemBarRelease node and delete it also. 1526 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() && 1527 ctrl->in(0)->is_MemBar()) { 1528 MemBarNode* membar = ctrl->in(0)->as_MemBar(); 1529 assert(membar->Opcode() == Op_MemBarRelease && 1530 mem->is_Proj() && membar == mem->in(0), ""); 1531 _igvn.hash_delete(fallthroughproj); 1532 _igvn.subsume_node(fallthroughproj, ctrl); 1533 _igvn.hash_delete(memproj_fallthrough); 1534 _igvn.subsume_node(memproj_fallthrough, mem); 1535 fallthroughproj = ctrl; 1536 memproj_fallthrough = mem; 1537 ctrl = membar->in(TypeFunc::Control); 1538 mem = membar->in(TypeFunc::Memory); 1539 } 1540 1541 _igvn.hash_delete(fallthroughproj); 1542 _igvn.subsume_node(fallthroughproj, ctrl); 1543 _igvn.hash_delete(memproj_fallthrough); 1544 _igvn.subsume_node(memproj_fallthrough, mem); 1545 return true; 1546} 1547 1548 1549//------------------------------expand_lock_node---------------------- 1550void PhaseMacroExpand::expand_lock_node(LockNode *lock) { 1551 1552 Node* ctrl = lock->in(TypeFunc::Control); 1553 Node* mem = lock->in(TypeFunc::Memory); 1554 Node* obj = lock->obj_node(); 1555 Node* box = lock->box_node(); 1556 Node* flock = lock->fastlock_node(); 1557 1558 // Make the merge point 1559 Node *region = new (C, 3) RegionNode(3); 1560 1561 Node *bol = transform_later(new (C, 2) BoolNode(flock,BoolTest::ne)); 1562 Node *iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); 1563 // Optimize test; set region slot 2 1564 Node *slow_path = opt_iff(region,iff); 1565 1566 // Make slow path call 1567 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box ); 1568 1569 extract_call_projections(call); 1570 1571 // Slow path can only throw asynchronous exceptions, which are always 1572 // de-opted. So the compiler thinks the slow-call can never throw an 1573 // exception. If it DOES throw an exception we would need the debug 1574 // info removed first (since if it throws there is no monitor). 1575 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 1576 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 1577 1578 // Capture slow path 1579 // disconnect fall-through projection from call and create a new one 1580 // hook up users of fall-through projection to region 1581 Node *slow_ctrl = _fallthroughproj->clone(); 1582 transform_later(slow_ctrl); 1583 _igvn.hash_delete(_fallthroughproj); 1584 _fallthroughproj->disconnect_inputs(NULL); 1585 region->init_req(1, slow_ctrl); 1586 // region inputs are now complete 1587 transform_later(region); 1588 _igvn.subsume_node(_fallthroughproj, region); 1589 1590 // create a Phi for the memory state 1591 Node *mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 1592 Node *memproj = transform_later( new (C, 1) ProjNode(call, TypeFunc::Memory) ); 1593 mem_phi->init_req(1, memproj ); 1594 mem_phi->init_req(2, mem); 1595 transform_later(mem_phi); 1596 _igvn.hash_delete(_memproj_fallthrough); 1597 _igvn.subsume_node(_memproj_fallthrough, mem_phi); 1598 1599 1600} 1601 1602//------------------------------expand_unlock_node---------------------- 1603void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { 1604 1605 Node* ctrl = unlock->in(TypeFunc::Control); 1606 Node* mem = unlock->in(TypeFunc::Memory); 1607 Node* obj = unlock->obj_node(); 1608 Node* box = unlock->box_node(); 1609 1610 // No need for a null check on unlock 1611 1612 // Make the merge point 1613 RegionNode *region = new (C, 3) RegionNode(3); 1614 1615 FastUnlockNode *funlock = new (C, 3) FastUnlockNode( ctrl, obj, box ); 1616 funlock = transform_later( funlock )->as_FastUnlock(); 1617 Node *bol = transform_later(new (C, 2) BoolNode(funlock,BoolTest::ne)); 1618 Node *iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); 1619 // Optimize test; set region slot 2 1620 Node *slow_path = opt_iff(region,iff); 1621 1622 CallNode *call = make_slow_call( (CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), "complete_monitor_unlocking_C", slow_path, obj, box ); 1623 1624 extract_call_projections(call); 1625 1626 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && 1627 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); 1628 1629 // No exceptions for unlocking 1630 // Capture slow path 1631 // disconnect fall-through projection from call and create a new one 1632 // hook up users of fall-through projection to region 1633 Node *slow_ctrl = _fallthroughproj->clone(); 1634 transform_later(slow_ctrl); 1635 _igvn.hash_delete(_fallthroughproj); 1636 _fallthroughproj->disconnect_inputs(NULL); 1637 region->init_req(1, slow_ctrl); 1638 // region inputs are now complete 1639 transform_later(region); 1640 _igvn.subsume_node(_fallthroughproj, region); 1641 1642 // create a Phi for the memory state 1643 Node *mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); 1644 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) ); 1645 mem_phi->init_req(1, memproj ); 1646 mem_phi->init_req(2, mem); 1647 transform_later(mem_phi); 1648 _igvn.hash_delete(_memproj_fallthrough); 1649 _igvn.subsume_node(_memproj_fallthrough, mem_phi); 1650 1651 1652} 1653 1654//------------------------------expand_macro_nodes---------------------- 1655// Returns true if a failure occurred. 1656bool PhaseMacroExpand::expand_macro_nodes() { 1657 if (C->macro_count() == 0) 1658 return false; 1659 // attempt to eliminate allocations 1660 bool progress = true; 1661 while (progress) { 1662 progress = false; 1663 for (int i = C->macro_count(); i > 0; i--) { 1664 Node * n = C->macro_node(i-1); 1665 bool success = false; 1666 debug_only(int old_macro_count = C->macro_count();); 1667 switch (n->class_id()) { 1668 case Node::Class_Allocate: 1669 case Node::Class_AllocateArray: 1670 success = eliminate_allocate_node(n->as_Allocate()); 1671 break; 1672 case Node::Class_Lock: 1673 case Node::Class_Unlock: 1674 success = eliminate_locking_node(n->as_AbstractLock()); 1675 break; 1676 default: 1677 assert(false, "unknown node type in macro list"); 1678 } 1679 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); 1680 progress = progress || success; 1681 } 1682 } 1683 // Make sure expansion will not cause node limit to be exceeded. 1684 // Worst case is a macro node gets expanded into about 50 nodes. 1685 // Allow 50% more for optimization. 1686 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) ) 1687 return true; 1688 1689 // expand "macro" nodes 1690 // nodes are removed from the macro list as they are processed 1691 while (C->macro_count() > 0) { 1692 int macro_count = C->macro_count(); 1693 Node * n = C->macro_node(macro_count-1); 1694 assert(n->is_macro(), "only macro nodes expected here"); 1695 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { 1696 // node is unreachable, so don't try to expand it 1697 C->remove_macro_node(n); 1698 continue; 1699 } 1700 switch (n->class_id()) { 1701 case Node::Class_Allocate: 1702 expand_allocate(n->as_Allocate()); 1703 break; 1704 case Node::Class_AllocateArray: 1705 expand_allocate_array(n->as_AllocateArray()); 1706 break; 1707 case Node::Class_Lock: 1708 expand_lock_node(n->as_Lock()); 1709 break; 1710 case Node::Class_Unlock: 1711 expand_unlock_node(n->as_Unlock()); 1712 break; 1713 default: 1714 assert(false, "unknown node type in macro list"); 1715 } 1716 assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); 1717 if (C->failing()) return true; 1718 } 1719 1720 _igvn.set_delay_transform(false); 1721 _igvn.optimize(); 1722 return false; 1723} 1724