graphKit.cpp revision 7890:f83851ae258e
1193323Sed/* 2193323Sed * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved. 3193323Sed * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4193323Sed * 5193323Sed * This code is free software; you can redistribute it and/or modify it 6193323Sed * under the terms of the GNU General Public License version 2 only, as 7193323Sed * published by the Free Software Foundation. 8193323Sed * 9193323Sed * This code is distributed in the hope that it will be useful, but WITHOUT 10193323Sed * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11193323Sed * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12193323Sed * version 2 for more details (a copy is included in the LICENSE file that 13193323Sed * accompanied this code). 14193323Sed * 15193323Sed * You should have received a copy of the GNU General Public License version 16193323Sed * 2 along with this work; if not, write to the Free Software Foundation, 17193323Sed * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18193323Sed * 19193323Sed * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20239462Sdim * or visit www.oracle.com if you need additional information or have any 21218893Sdim * questions. 22223017Sdim * 23223017Sdim */ 24239462Sdim 25239462Sdim#include "precompiled.hpp" 26239462Sdim#include "compiler/compileLog.hpp" 27193323Sed#include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp" 28193323Sed#include "gc_implementation/g1/heapRegion.hpp" 29193323Sed#include "gc_interface/collectedHeap.hpp" 30193323Sed#include "memory/barrierSet.hpp" 31193323Sed#include "memory/cardTableModRefBS.hpp" 32193323Sed#include "opto/addnode.hpp" 33239462Sdim#include "opto/castnode.hpp" 34193323Sed#include "opto/convertnode.hpp" 35193323Sed#include "opto/graphKit.hpp" 36193323Sed#include "opto/idealKit.hpp" 37208599Srdivacky#include "opto/intrinsicnode.hpp" 38193323Sed#include "opto/locknode.hpp" 39193323Sed#include "opto/machnode.hpp" 40239462Sdim#include "opto/opaquenode.hpp" 41239462Sdim#include "opto/parse.hpp" 42221345Sdim#include "opto/rootnode.hpp" 43221345Sdim#include "opto/runtime.hpp" 44239462Sdim#include "runtime/deoptimization.hpp" 45221345Sdim#include "runtime/sharedRuntime.hpp" 46221345Sdim 47221345Sdim//----------------------------GraphKit----------------------------------------- 48221345Sdim// Main utility constructor. 49263508SdimGraphKit::GraphKit(JVMState* jvms) 50221345Sdim : Phase(Phase::Parser), 51221345Sdim _env(C->env()), 52221345Sdim _gvn(*C->initial_gvn()) 53221345Sdim{ 54221345Sdim _exceptions = jvms->map()->next_exception(); 55221345Sdim if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 56221345Sdim set_jvms(jvms); 57221345Sdim} 58263508Sdim 59221345Sdim// Private constructor for parser. 60221345SdimGraphKit::GraphKit() 61221345Sdim : Phase(Phase::Parser), 62221345Sdim _env(C->env()), 63263508Sdim _gvn(*C->initial_gvn()) 64221345Sdim{ 65221345Sdim _exceptions = NULL; 66221345Sdim set_map(NULL); 67263508Sdim debug_only(_sp = -99); 68221345Sdim debug_only(set_bci(-99)); 69239462Sdim} 70193323Sed 71193323Sed 72193323Sed 73193323Sed//---------------------------clean_stack--------------------------------------- 74239462Sdim// Clear away rubbish from the stack area of the JVM state. 75193323Sed// This destroys any arguments that may be waiting on the stack. 76221345Sdimvoid GraphKit::clean_stack(int from_sp) { 77221345Sdim SafePointNode* map = this->map(); 78221345Sdim JVMState* jvms = this->jvms(); 79221345Sdim int stk_size = jvms->stk_size(); 80263508Sdim int stkoff = jvms->stkoff(); 81221345Sdim Node* top = this->top(); 82263508Sdim for (int i = from_sp; i < stk_size; i++) { 83221345Sdim if (map->in(stkoff + i) != top) { 84221345Sdim map->set_req(stkoff + i, top); 85193323Sed } 86221345Sdim } 87221345Sdim} 88221345Sdim 89221345Sdim 90221345Sdim//--------------------------------sync_jvms----------------------------------- 91221345Sdim// Make sure our current jvms agrees with our parse state. 92221345SdimJVMState* GraphKit::sync_jvms() const { 93263508Sdim JVMState* jvms = this->jvms(); 94221345Sdim jvms->set_bci(bci()); // Record the new bci in the JVMState 95221345Sdim jvms->set_sp(sp()); // Record the new sp in the JVMState 96221345Sdim assert(jvms_in_sync(), "jvms is now in sync"); 97263508Sdim return jvms; 98221345Sdim} 99263508Sdim 100221345Sdim//--------------------------------sync_jvms_for_reexecute--------------------- 101221345Sdim// Make sure our current jvms agrees with our parse state. This version 102221345Sdim// uses the reexecute_sp for reexecuting bytecodes. 103221345SdimJVMState* GraphKit::sync_jvms_for_reexecute() { 104221345Sdim JVMState* jvms = this->jvms(); 105221345Sdim jvms->set_bci(bci()); // Record the new bci in the JVMState 106221345Sdim jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 107221345Sdim return jvms; 108221345Sdim} 109221345Sdim 110221345Sdim#ifdef ASSERT 111221345Sdimbool GraphKit::jvms_in_sync() const { 112221345Sdim Parse* parse = is_Parse(); 113221345Sdim if (parse == NULL) { 114263508Sdim if (bci() != jvms()->bci()) return false; 115263508Sdim if (sp() != (int)jvms()->sp()) return false; 116221345Sdim return true; 117221345Sdim } 118193323Sed if (jvms()->method() != parse->method()) return false; 119193323Sed if (jvms()->bci() != parse->bci()) return false; 120193323Sed int jvms_sp = jvms()->sp(); 121221345Sdim if (jvms_sp != parse->sp()) return false; 122221345Sdim int jvms_depth = jvms()->depth(); 123221345Sdim if (jvms_depth != parse->depth()) return false; 124193323Sed return true; 125193323Sed} 126193323Sed 127221345Sdim// Local helper checks for special internal merge points 128221345Sdim// used to accumulate and merge exception states. 129221345Sdim// They are marked by the region's in(0) edge being the map itself. 130221345Sdim// Such merge points must never "escape" into the parser at large, 131221345Sdim// until they have been handed to gvn.transform. 132221345Sdimstatic bool is_hidden_merge(Node* reg) { 133263508Sdim if (reg == NULL) return false; 134221345Sdim if (reg->is_Phi()) { 135221345Sdim reg = reg->in(0); 136221345Sdim if (reg == NULL) return false; 137221345Sdim } 138221345Sdim return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 139221345Sdim} 140221345Sdim 141221345Sdimvoid GraphKit::verify_map() const { 142221345Sdim if (map() == NULL) return; // null map is OK 143221345Sdim assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 144221345Sdim assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 145263508Sdim assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 146221345Sdim} 147221345Sdim 148221345Sdimvoid GraphKit::verify_exception_state(SafePointNode* ex_map) { 149221345Sdim assert(ex_map->next_exception() == NULL, "not already part of a chain"); 150263508Sdim assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 151221345Sdim} 152221345Sdim#endif 153263508Sdim 154221345Sdim//---------------------------stop_and_kill_map--------------------------------- 155221345Sdim// Set _map to NULL, signalling a stop to further bytecode execution. 156263508Sdim// First smash the current map's control to a constant, to mark it dead. 157221345Sdimvoid GraphKit::stop_and_kill_map() { 158263508Sdim SafePointNode* dead_map = stop(); 159221345Sdim if (dead_map != NULL) { 160221345Sdim dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 161221345Sdim assert(dead_map->is_killed(), "must be so marked"); 162221345Sdim } 163263508Sdim} 164263508Sdim 165221345Sdim 166221345Sdim//--------------------------------stopped-------------------------------------- 167221345Sdim// Tell if _map is NULL, or control is top. 168221345Sdimbool GraphKit::stopped() { 169263508Sdim if (map() == NULL) return true; 170193323Sed else if (control() == top()) return true; 171193323Sed else return false; 172193323Sed} 173193323Sed 174221345Sdim 175221345Sdim//-----------------------------has_ex_handler---------------------------------- 176263508Sdim// Tell if this method or any caller method has exception handlers. 177263508Sdimbool GraphKit::has_ex_handler() { 178221345Sdim for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 179221345Sdim if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 180263508Sdim return true; 181221345Sdim } 182221345Sdim } 183221345Sdim return false; 184193323Sed} 185263508Sdim 186221345Sdim//------------------------------save_ex_oop------------------------------------ 187221345Sdim// Save an exception without blowing stack contents or other JVM state. 188221345Sdimvoid GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 189221345Sdim assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 190218893Sdim ex_map->add_req(ex_oop); 191193323Sed debug_only(verify_exception_state(ex_map)); 192218893Sdim} 193193323Sed 194193323Sedinline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 195218893Sdim assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 196221345Sdim Node* ex_oop = ex_map->in(ex_map->req()-1); 197221345Sdim if (clear_it) ex_map->del_req(ex_map->req()-1); 198221345Sdim return ex_oop; 199221345Sdim} 200221345Sdim 201221345Sdim//-----------------------------saved_ex_oop------------------------------------ 202221345Sdim// Recover a saved exception from its map. 203221345SdimNode* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 204221345Sdim return common_saved_ex_oop(ex_map, false); 205221345Sdim} 206221345Sdim 207221345Sdim//--------------------------clear_saved_ex_oop--------------------------------- 208221345Sdim// Erase a previously saved exception from its map. 209221345SdimNode* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 210221345Sdim return common_saved_ex_oop(ex_map, true); 211221345Sdim} 212221345Sdim 213221345Sdim#ifdef ASSERT 214263508Sdim//---------------------------has_saved_ex_oop---------------------------------- 215221345Sdim// Erase a previously saved exception from its map. 216221345Sdimbool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 217263508Sdim return ex_map->req() == ex_map->jvms()->endoff()+1; 218221345Sdim} 219221345Sdim#endif 220221345Sdim 221263508Sdim//-------------------------make_exception_state-------------------------------- 222221345Sdim// Turn the current JVM state into an exception state, appending the ex_oop. 223221345SdimSafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 224221345Sdim sync_jvms(); 225221345Sdim SafePointNode* ex_map = stop(); // do not manipulate this map any more 226263508Sdim set_saved_ex_oop(ex_map, ex_oop); 227193323Sed return ex_map; 228205407Srdivacky} 229221345Sdim 230193323Sed 231218893Sdim//--------------------------add_exception_state-------------------------------- 232193323Sed// Add an exception to my list of exceptions. 233221345Sdimvoid GraphKit::add_exception_state(SafePointNode* ex_map) { 234221345Sdim if (ex_map == NULL || ex_map->control() == top()) { 235193323Sed return; 236193323Sed } 237193323Sed#ifdef ASSERT 238193323Sed verify_exception_state(ex_map); 239193323Sed if (has_exceptions()) { 240263508Sdim assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 241221345Sdim } 242221345Sdim#endif 243221345Sdim 244221345Sdim // If there is already an exception of exactly this type, merge with it. 245221345Sdim // In particular, null-checks and other low-level exceptions common up here. 246221345Sdim Node* ex_oop = saved_ex_oop(ex_map); 247221345Sdim const Type* ex_type = _gvn.type(ex_oop); 248221345Sdim if (ex_oop == top()) { 249243830Sdim // No action needed. 250193323Sed return; 251193323Sed } 252193323Sed assert(ex_type->isa_instptr(), "exception must be an instance"); 253263508Sdim for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 254193323Sed const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 255193323Sed // We check sp also because call bytecodes can generate exceptions 256224145Sdim // both before and after arguments are popped! 257224145Sdim if (ex_type2 == ex_type 258193323Sed && e2->_jvms->sp() == ex_map->_jvms->sp()) { 259193323Sed combine_exception_states(ex_map, e2); 260193323Sed return; 261224145Sdim } 262263508Sdim } 263263508Sdim 264263508Sdim // No pre-existing exception of the same type. Chain it on the list. 265193323Sed push_exception_state(ex_map); 266218893Sdim} 267212904Sdim 268193323Sed//-----------------------add_exception_states_from----------------------------- 269193323Sedvoid GraphKit::add_exception_states_from(JVMState* jvms) { 270212904Sdim SafePointNode* ex_map = jvms->map()->next_exception(); 271212904Sdim if (ex_map != NULL) { 272212904Sdim jvms->map()->set_next_exception(NULL); 273193323Sed for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 274212904Sdim next_map = ex_map->next_exception(); 275193323Sed ex_map->set_next_exception(NULL); 276193323Sed add_exception_state(ex_map); 277193323Sed } 278221345Sdim } 279193323Sed} 280193323Sed 281193323Sed//-----------------------transfer_exceptions_into_jvms------------------------- 282193323SedJVMState* GraphKit::transfer_exceptions_into_jvms() { 283195340Sed if (map() == NULL) { 284193323Sed // We need a JVMS to carry the exceptions, but the map has gone away. 285221345Sdim // Create a scratch JVMS, cloned from any of the exception states... 286208599Srdivacky if (has_exceptions()) { 287221345Sdim _map = _exceptions; 288193323Sed _map = clone_map(); 289221345Sdim _map->set_next_exception(NULL); 290234353Sdim clear_saved_ex_oop(_map); 291193323Sed debug_only(verify_map()); 292223017Sdim } else { 293193323Sed // ...or created from scratch 294193323Sed JVMState* jvms = new (C) JVMState(_method, NULL); 295193323Sed jvms->set_bci(_bci); 296193323Sed jvms->set_sp(_sp); 297193323Sed jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 298193323Sed set_jvms(jvms); 299195340Sed for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 300221345Sdim set_all_memory(top()); 301193323Sed while (map()->req() < jvms->endoff()) map()->add_req(top()); 302193323Sed } 303193323Sed // (This is a kludge, in case you didn't notice.) 304193323Sed set_control(top()); 305193323Sed } 306193323Sed JVMState* jvms = sync_jvms(); 307193323Sed assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 308193323Sed jvms->map()->set_next_exception(_exceptions); 309193323Sed _exceptions = NULL; // done with this set of exceptions 310221345Sdim return jvms; 311208599Srdivacky} 312193323Sed 313221345Sdimstatic inline void add_n_reqs(Node* dstphi, Node* srcphi) { 314193323Sed assert(is_hidden_merge(dstphi), "must be a special merge node"); 315193323Sed assert(is_hidden_merge(srcphi), "must be a special merge node"); 316221345Sdim uint limit = srcphi->req(); 317193323Sed for (uint i = PhiNode::Input; i < limit; i++) { 318193323Sed dstphi->add_req(srcphi->in(i)); 319193323Sed } 320223017Sdim} 321193323Sedstatic inline void add_one_req(Node* dstphi, Node* src) { 322193323Sed assert(is_hidden_merge(dstphi), "must be a special merge node"); 323193323Sed assert(!is_hidden_merge(src), "must not be a special merge node"); 324193323Sed dstphi->add_req(src); 325195340Sed} 326193323Sed 327221345Sdim//-----------------------combine_exception_states------------------------------ 328208599Srdivacky// This helper function combines exception states by building phis on a 329221345Sdim// specially marked state-merging region. These regions and phis are 330193323Sed// untransformed, and can build up gradually. The region is marked by 331221345Sdim// having a control input of its exception map, rather than NULL. Such 332193323Sed// regions do not appear except in this function, and in use_exception_state. 333193323Sedvoid GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 334223017Sdim if (failing()) return; // dying anyway... 335193323Sed JVMState* ex_jvms = ex_map->_jvms; 336193323Sed assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 337193323Sed assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 338193323Sed assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 339193323Sed assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 340193323Sed assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 341193323Sed assert(ex_map->req() == phi_map->req(), "matching maps"); 342221345Sdim uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 343221345Sdim Node* hidden_merge_mark = root(); 344221345Sdim Node* region = phi_map->control(); 345193323Sed MergeMemNode* phi_mem = phi_map->merged_memory(); 346193323Sed MergeMemNode* ex_mem = ex_map->merged_memory(); 347193323Sed if (region->in(0) != hidden_merge_mark) { 348193323Sed // The control input is not (yet) a specially-marked region in phi_map. 349193323Sed // Make it so, and build some phis. 350193323Sed region = new RegionNode(2); 351193323Sed _gvn.set_type(region, Type::CONTROL); 352193323Sed region->set_req(0, hidden_merge_mark); // marks an internal ex-state 353221345Sdim region->init_req(1, phi_map->control()); 354193323Sed phi_map->set_control(region); 355193323Sed Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 356193323Sed record_for_igvn(io_phi); 357221345Sdim _gvn.set_type(io_phi, Type::ABIO); 358221345Sdim phi_map->set_i_o(io_phi); 359221345Sdim for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 360221345Sdim Node* m = mms.memory(); 361193323Sed Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 362193323Sed record_for_igvn(m_phi); 363239462Sdim _gvn.set_type(m_phi, Type::MEMORY); 364193323Sed mms.set_memory(m_phi); 365239462Sdim } 366193323Sed } 367193323Sed 368193323Sed // Either or both of phi_map and ex_map might already be converted into phis. 369193323Sed Node* ex_control = ex_map->control(); 370193323Sed // if there is special marking on ex_map also, we add multiple edges from src 371218893Sdim bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 372212904Sdim // how wide was the destination phi_map, originally? 373193323Sed uint orig_width = region->req(); 374193323Sed 375193323Sed if (add_multiple) { 376221345Sdim add_n_reqs(region, ex_control); 377221345Sdim add_n_reqs(phi_map->i_o(), ex_map->i_o()); 378263508Sdim } else { 379193323Sed // ex_map has no merges, so we just add single edges everywhere 380221345Sdim add_one_req(region, ex_control); 381193323Sed add_one_req(phi_map->i_o(), ex_map->i_o()); 382193323Sed } 383193323Sed for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 384221345Sdim if (mms.is_empty()) { 385221345Sdim // get a copy of the base memory, and patch some inputs into it 386221345Sdim const TypePtr* adr_type = mms.adr_type(C); 387193323Sed Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 388239462Sdim assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 389193323Sed mms.set_memory(phi); 390193323Sed // Prepare to append interesting stuff onto the newly sliced phi: 391193323Sed while (phi->req() > orig_width) phi->del_req(phi->req()-1); 392193323Sed } 393193323Sed // Append stuff from ex_map: 394193323Sed if (add_multiple) { 395193323Sed add_n_reqs(mms.memory(), mms.memory2()); 396193323Sed } else { 397193323Sed add_one_req(mms.memory(), mms.memory2()); 398193323Sed } 399193323Sed } 400193323Sed uint limit = ex_map->req(); 401193323Sed for (uint i = TypeFunc::Parms; i < limit; i++) { 402193323Sed // Skip everything in the JVMS after tos. (The ex_oop follows.) 403193323Sed if (i == tos) i = ex_jvms->monoff(); 404193323Sed Node* src = ex_map->in(i); 405221345Sdim Node* dst = phi_map->in(i); 406221345Sdim if (src != dst) { 407221345Sdim PhiNode* phi; 408221345Sdim if (dst->in(0) != region) { 409221345Sdim dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 410221345Sdim record_for_igvn(phi); 411221345Sdim _gvn.set_type(phi, phi->type()); 412243830Sdim phi_map->set_req(i, dst); 413221345Sdim // Prepare to append interesting stuff onto the new phi: 414221345Sdim while (dst->req() > orig_width) dst->del_req(dst->req()-1); 415221345Sdim } else { 416243830Sdim assert(dst->is_Phi(), "nobody else uses a hidden region"); 417243830Sdim phi = dst->as_Phi(); 418221345Sdim } 419224145Sdim if (add_multiple && src->in(0) == ex_control) { 420221345Sdim // Both are phis. 421221345Sdim add_n_reqs(dst, src); 422221345Sdim } else { 423221345Sdim while (dst->req() < region->req()) add_one_req(dst, src); 424193323Sed } 425193323Sed const Type* srctype = _gvn.type(src); 426193323Sed if (phi->type() != srctype) { 427193323Sed const Type* dsttype = phi->type()->meet_speculative(srctype); 428193323Sed if (phi->type() != dsttype) { 429193323Sed phi->set_type(dsttype); 430193323Sed _gvn.set_type(phi, dsttype); 431193323Sed } 432193323Sed } 433193323Sed } 434193323Sed } 435193323Sed phi_map->merge_replaced_nodes_with(ex_map); 436193323Sed} 437193323Sed 438193323Sed//--------------------------use_exception_state-------------------------------- 439193323SedNode* GraphKit::use_exception_state(SafePointNode* phi_map) { 440193323Sed if (failing()) { stop(); return top(); } 441193323Sed Node* region = phi_map->control(); 442205407Srdivacky Node* hidden_merge_mark = root(); 443221345Sdim assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 444193323Sed Node* ex_oop = clear_saved_ex_oop(phi_map); 445218893Sdim if (region->in(0) == hidden_merge_mark) { 446193323Sed // Special marking for internal ex-states. Process the phis now. 447193323Sed region->set_req(0, region); // now it's an ordinary region 448193323Sed set_jvms(phi_map->jvms()); // ...so now we can use it as a map 449193323Sed // Note: Setting the jvms also sets the bci and sp. 450193323Sed set_control(_gvn.transform(region)); 451193323Sed uint tos = jvms()->stkoff() + sp(); 452193323Sed for (uint i = 1; i < tos; i++) { 453193323Sed Node* x = phi_map->in(i); 454193323Sed if (x->in(0) == region) { 455193323Sed assert(x->is_Phi(), "expected a special phi"); 456193323Sed phi_map->set_req(i, _gvn.transform(x)); 457193323Sed } 458193323Sed } 459193323Sed for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 460193323Sed Node* x = mms.memory(); 461193323Sed if (x->in(0) == region) { 462193323Sed assert(x->is_Phi(), "nobody else uses a hidden region"); 463208599Srdivacky mms.set_memory(_gvn.transform(x)); 464193323Sed } 465218893Sdim } 466224145Sdim if (ex_oop->in(0) == region) { 467224145Sdim assert(ex_oop->is_Phi(), "expected a special phi"); 468193323Sed ex_oop = _gvn.transform(ex_oop); 469193323Sed } 470193323Sed } else { 471193323Sed set_jvms(phi_map->jvms()); 472193323Sed } 473193323Sed 474212904Sdim assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 475212904Sdim assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 476212904Sdim return ex_oop; 477218893Sdim} 478243830Sdim 479208599Srdivacky//---------------------------------java_bc------------------------------------- 480208599SrdivackyBytecodes::Code GraphKit::java_bc() const { 481208599Srdivacky ciMethod* method = this->method(); 482208599Srdivacky int bci = this->bci(); 483193323Sed if (method != NULL && bci != InvocationEntryBci) 484193323Sed return method->java_code_at_bci(bci); 485193323Sed else 486193323Sed return Bytecodes::_illegal; 487193323Sed} 488193323Sed 489193323Sedvoid GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 490206083Srdivacky bool must_throw) { 491206083Srdivacky // if the exception capability is set, then we will generate code 492218893Sdim // to check the JavaThread.should_post_on_exceptions flag to see 493193323Sed // if we actually need to report exception events (for this 494193323Sed // thread). If we don't need to report exception events, we will 495205407Srdivacky // take the normal fast path provided by add_exception_events. If 496205407Srdivacky // exception event reporting is enabled for this thread, we will 497193323Sed // take the uncommon_trap in the BuildCutout below. 498205407Srdivacky 499205407Srdivacky // first must access the should_post_on_exceptions_flag in this thread's JavaThread 500205407Srdivacky Node* jthread = _gvn.transform(new ThreadLocalNode()); 501205407Srdivacky Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 502193323Sed Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 503193323Sed 504193323Sed // Test the should_post_on_exceptions_flag vs. 0 505193323Sed Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 506193323Sed Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 507193323Sed 508193323Sed // Branch to slow_path if should_post_on_exceptions_flag was true 509263508Sdim { BuildCutout unless(this, tst, PROB_MAX); 510263508Sdim // Do not try anything fancy if we're notifying the VM on every throw. 511193323Sed // Cf. case Bytecodes::_athrow in parse2.cpp. 512218893Sdim uncommon_trap(reason, Deoptimization::Action_none, 513193323Sed (ciKlass*)NULL, (char*)NULL, must_throw); 514221345Sdim } 515226633Sdim 516193323Sed} 517221345Sdim 518221345Sdim//------------------------------builtin_throw---------------------------------- 519221345Sdimvoid GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 520221345Sdim bool must_throw = true; 521221345Sdim 522193323Sed if (env()->jvmti_can_post_on_exceptions()) { 523221345Sdim // check if we must post exception events, take uncommon trap if so 524221345Sdim uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 525221345Sdim // here if should_post_on_exceptions is false 526221345Sdim // continue on with the normal codegen 527221345Sdim } 528221345Sdim 529221345Sdim // If this particular condition has not yet happened at this 530193323Sed // bytecode, then use the uncommon trap mechanism, and allow for 531221345Sdim // a future recompilation if several traps occur here. 532193323Sed // If the throw is hot, try to use a more complicated inline mechanism 533218893Sdim // which keeps execution inside the compiled code. 534193323Sed bool treat_throw_as_hot = false; 535193323Sed ciMethodData* md = method()->method_data(); 536193323Sed 537221345Sdim if (ProfileTraps) { 538221345Sdim if (too_many_traps(reason)) { 539221345Sdim treat_throw_as_hot = true; 540221345Sdim } 541221345Sdim // (If there is no MDO at all, assume it is early in 542221345Sdim // execution, and that any deopts are part of the 543221345Sdim // startup transient, and don't need to be remembered.) 544193323Sed 545193323Sed // Also, if there is a local exception handler, treat all throws 546193323Sed // as hot if there has been at least one in this method. 547193323Sed if (C->trap_count(reason) != 0 548193323Sed && method()->method_data()->trap_count(reason) != 0 549193323Sed && has_ex_handler()) { 550193323Sed treat_throw_as_hot = true; 551193323Sed } 552193323Sed } 553193323Sed 554263508Sdim // If this throw happens frequently, an uncommon trap might cause 555221345Sdim // a performance pothole. If there is a local exception handler, 556243830Sdim // and if this particular bytecode appears to be deoptimizing often, 557221345Sdim // let us handle the throw inline, with a preconstructed instance. 558218893Sdim // Note: If the deopt count has blown up, the uncommon trap 559193323Sed // runtime is going to flush this nmethod, not matter what. 560263508Sdim if (treat_throw_as_hot 561221345Sdim && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 562221345Sdim // If the throw is local, we use a pre-existing instance and 563221345Sdim // punt on the backtrace. This would lead to a missing backtrace 564221345Sdim // (a repeat of 4292742) if the backtrace object is ever asked 565221345Sdim // for its backtrace. 566221345Sdim // Fixing this remaining case of 4292742 requires some flavor of 567221345Sdim // escape analysis. Leave that for the future. 568193323Sed ciInstance* ex_obj = NULL; 569193323Sed switch (reason) { 570193323Sed case Deoptimization::Reason_null_check: 571221345Sdim ex_obj = env()->NullPointerException_instance(); 572221345Sdim break; 573221345Sdim case Deoptimization::Reason_div0_check: 574263508Sdim ex_obj = env()->ArithmeticException_instance(); 575221345Sdim break; 576221345Sdim case Deoptimization::Reason_range_check: 577221345Sdim ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 578221345Sdim break; 579221345Sdim case Deoptimization::Reason_class_check: 580221345Sdim if (java_bc() == Bytecodes::_aastore) { 581263508Sdim ex_obj = env()->ArrayStoreException_instance(); 582221345Sdim } else { 583221345Sdim ex_obj = env()->ClassCastException_instance(); 584221345Sdim } 585221345Sdim break; 586221345Sdim } 587193323Sed if (failing()) { stop(); return; } // exception allocation might fail 588193323Sed if (ex_obj != NULL) { 589193323Sed // Cheat with a preallocated exception object. 590193323Sed if (C->log() != NULL) 591193323Sed C->log()->elem("hot_throw preallocated='1' reason='%s'", 592193323Sed Deoptimization::trap_reason_name(reason)); 593193323Sed const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 594193323Sed Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 595193323Sed 596193323Sed // Clear the detail message of the preallocated exception object. 597193323Sed // Weblogic sometimes mutates the detail message of exceptions 598193323Sed // using reflection. 599193323Sed int offset = java_lang_Throwable::get_detailMessage_offset(); 600193323Sed const TypePtr* adr_typ = ex_con->add_offset(offset); 601193323Sed 602193323Sed Node *adr = basic_plus_adr(ex_node, ex_node, offset); 603193323Sed const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 604193323Sed // Conservatively release stores of object references. 605193323Sed Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 606193323Sed 607193323Sed add_exception_state(make_exception_state(ex_node)); 608193323Sed return; 609193323Sed } 610193323Sed } 611193323Sed 612193323Sed // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 613193323Sed // It won't be much cheaper than bailing to the interp., since we'll 614193323Sed // have to pass up all the debug-info, and the runtime will have to 615193323Sed // create the stack trace. 616193323Sed 617221345Sdim // Usual case: Bail to interpreter. 618193323Sed // Reserve the right to recompile if we haven't seen anything yet. 619193323Sed 620193323Sed ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 621193323Sed Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 622193323Sed if (treat_throw_as_hot 623193323Sed && (method()->method_data()->trap_recompiled_at(bci(), m) 624193323Sed || C->too_many_traps(reason))) { 625193323Sed // We cannot afford to take more traps here. Suffer in the interpreter. 626193323Sed if (C->log() != NULL) 627193323Sed C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 628218893Sdim Deoptimization::trap_reason_name(reason), 629193323Sed C->trap_count(reason)); 630193323Sed action = Deoptimization::Action_none; 631193323Sed } 632193323Sed 633193323Sed // "must_throw" prunes the JVM state to include only the stack, if there 634193323Sed // are no local exception handlers. This should cut down on register 635193323Sed // allocation time and code size, by drastically reducing the number 636193323Sed // of in-edges on the call to the uncommon trap. 637193323Sed 638218893Sdim uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 639193323Sed} 640193323Sed 641210299Sed 642210299Sed//----------------------------PreserveJVMState--------------------------------- 643210299SedPreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 644193323Sed debug_only(kit->verify_map()); 645218893Sdim _kit = kit; 646193323Sed _map = kit->map(); // preserve the map 647208599Srdivacky _sp = kit->sp(); 648221345Sdim kit->set_map(clone_map ? kit->clone_map() : NULL); 649221345Sdim#ifdef ASSERT 650193323Sed _bci = kit->bci(); 651239462Sdim Parse* parser = kit->is_Parse(); 652193323Sed int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 653193323Sed _block = block; 654193323Sed#endif 655193323Sed} 656193323SedPreserveJVMState::~PreserveJVMState() { 657193323Sed GraphKit* kit = _kit; 658221345Sdim#ifdef ASSERT 659193323Sed assert(kit->bci() == _bci, "bci must not shift"); 660218893Sdim Parse* parser = kit->is_Parse(); 661193323Sed int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 662193323Sed assert(block == _block, "block must not shift"); 663218893Sdim#endif 664193323Sed kit->set_map(_map); 665193323Sed kit->set_sp(_sp); 666193323Sed} 667193323Sed 668193323Sed 669193323Sed//-----------------------------BuildCutout------------------------------------- 670193323SedBuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 671218893Sdim : PreserveJVMState(kit) 672193323Sed{ 673193323Sed assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 674193323Sed SafePointNode* outer_map = _map; // preserved map is caller's 675193323Sed SafePointNode* inner_map = kit->map(); 676221345Sdim IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 677198090Srdivacky outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 678193323Sed inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 679193323Sed} 680193323SedBuildCutout::~BuildCutout() { 681198090Srdivacky GraphKit* kit = _kit; 682193323Sed assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 683193323Sed} 684193323Sed 685193323Sed//---------------------------PreserveReexecuteState---------------------------- 686193323SedPreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 687193323Sed assert(!kit->stopped(), "must call stopped() before"); 688221345Sdim _kit = kit; 689193323Sed _sp = kit->sp(); 690193323Sed _reexecute = kit->jvms()->_reexecute; 691193323Sed} 692193323SedPreserveReexecuteState::~PreserveReexecuteState() { 693193323Sed if (_kit->stopped()) return; 694218893Sdim _kit->jvms()->_reexecute = _reexecute; 695193323Sed _kit->set_sp(_sp); 696193323Sed} 697193323Sed 698193323Sed//------------------------------clone_map-------------------------------------- 699193323Sed// Implementation of PreserveJVMState 700221345Sdim// 701198090Srdivacky// Only clone_map(...) here. If this function is only used in the 702193323Sed// PreserveJVMState class we may want to get rid of this extra 703193323Sed// function eventually and do it all there. 704193323Sed 705193323SedSafePointNode* GraphKit::clone_map() { 706218893Sdim if (map() == NULL) return NULL; 707198090Srdivacky 708193323Sed // Clone the memory edge first 709218893Sdim Node* mem = MergeMemNode::make(map()->memory()); 710193323Sed gvn().set_type_bottom(mem); 711193323Sed 712218893Sdim SafePointNode *clonemap = (SafePointNode*)map()->clone(); 713193323Sed JVMState* jvms = this->jvms(); 714193323Sed JVMState* clonejvms = jvms->clone_shallow(C); 715193323Sed clonemap->set_memory(mem); 716193323Sed clonemap->set_jvms(clonejvms); 717193323Sed clonejvms->set_map(clonemap); 718193323Sed record_for_igvn(clonemap); 719193323Sed gvn().set_type_bottom(clonemap); 720193323Sed return clonemap; 721193323Sed} 722193323Sed 723193323Sed 724193323Sed//-----------------------------set_map_clone----------------------------------- 725193323Sedvoid GraphKit::set_map_clone(SafePointNode* m) { 726193323Sed _map = m; 727193323Sed _map = clone_map(); 728193323Sed _map->set_next_exception(NULL); 729218893Sdim debug_only(verify_map()); 730210299Sed} 731210299Sed 732210299Sed 733210299Sed//----------------------------kill_dead_locals--------------------------------- 734210299Sed// Detect any locals which are known to be dead, and force them to top. 735210299Sedvoid GraphKit::kill_dead_locals() { 736218893Sdim // Consult the liveness information for the locals. If any 737193323Sed // of them are unused, then they can be replaced by top(). This 738218893Sdim // should help register allocation time and cut down on the size 739208599Srdivacky // of the deoptimization information. 740221345Sdim 741221345Sdim // This call is made from many of the bytecode handling 742193323Sed // subroutines called from the Big Switch in do_one_bytecode. 743239462Sdim // Every bytecode which might include a slow path is responsible 744193323Sed // for killing its dead locals. The more consistent we 745193323Sed // are about killing deads, the fewer useless phis will be 746193323Sed // constructed for them at various merge points. 747193323Sed 748193323Sed // bci can be -1 (InvocationEntryBci). We return the entry 749208599Srdivacky // liveness for the method. 750208599Srdivacky 751193323Sed if (method() == NULL || method()->code_size() == 0) { 752193323Sed // We are building a graph for a call to a native method. 753218893Sdim // All locals are live. 754193323Sed return; 755193323Sed } 756193323Sed 757218893Sdim ResourceMark rm; 758193323Sed 759193323Sed // Consult the liveness information for the locals. If any 760193323Sed // of them are unused, then they can be replaced by top(). This 761193323Sed // should help register allocation time and cut down on the size 762193323Sed // of the deoptimization information. 763193323Sed MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 764193323Sed 765193323Sed int len = (int)live_locals.size(); 766193323Sed assert(len <= jvms()->loc_size(), "too many live locals"); 767193323Sed for (int local = 0; local < len; local++) { 768193323Sed if (!live_locals.at(local)) { 769221345Sdim set_local(local, top()); 770198090Srdivacky } 771193323Sed } 772193323Sed} 773193323Sed 774193323Sed#ifdef ASSERT 775198090Srdivacky//-------------------------dead_locals_are_killed------------------------------ 776193323Sed// Return true if all dead locals are set to top in the map. 777193323Sed// Used to assert "clean" debug info at various points. 778193323Sedbool GraphKit::dead_locals_are_killed() { 779193323Sed if (method() == NULL || method()->code_size() == 0) { 780193323Sed // No locals need to be dead, so all is as it should be. 781193323Sed return true; 782221345Sdim } 783193323Sed 784193323Sed // Make sure somebody called kill_dead_locals upstream. 785193323Sed ResourceMark rm; 786193323Sed for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 787193323Sed if (jvms->loc_size() == 0) continue; // no locals to consult 788193323Sed SafePointNode* map = jvms->map(); 789193323Sed ciMethod* method = jvms->method(); 790193323Sed int bci = jvms->bci(); 791193323Sed if (jvms == this->jvms()) { 792193323Sed bci = this->bci(); // it might not yet be synched 793193323Sed } 794193323Sed MethodLivenessResult live_locals = method->liveness_at_bci(bci); 795193323Sed int len = (int)live_locals.size(); 796193323Sed if (!live_locals.is_valid() || len == 0) 797193323Sed // This method is trivial, or is poisoned by a breakpoint. 798193323Sed return true; 799193323Sed assert(len == jvms->loc_size(), "live map consistent with locals map"); 800193323Sed for (int local = 0; local < len; local++) { 801221345Sdim if (!live_locals.at(local) && map->local(jvms, local) != top()) { 802202375Srdivacky if (PrintMiscellaneous && (Verbose || WizardMode)) { 803193323Sed tty->print_cr("Zombie local %d: ", local); 804193323Sed jvms->dump(); 805193323Sed } 806193323Sed return false; 807263508Sdim } 808263508Sdim } 809263508Sdim } 810221345Sdim return true; 811263508Sdim} 812221345Sdim 813221345Sdim#endif //ASSERT 814221345Sdim 815221345Sdim// Helper function for enforcing certain bytecodes to reexecute if 816221345Sdim// deoptimization happens 817221345Sdimstatic bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 818221345Sdim ciMethod* cur_method = jvms->method(); 819263508Sdim int cur_bci = jvms->bci(); 820221345Sdim if (cur_method != NULL && cur_bci != InvocationEntryBci) { 821221345Sdim Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 822221345Sdim return Interpreter::bytecode_should_reexecute(code) || 823221345Sdim is_anewarray && code == Bytecodes::_multianewarray; 824221345Sdim // Reexecute _multianewarray bytecode which was replaced with 825221345Sdim // sequence of [a]newarray. See Parse::do_multianewarray(). 826221345Sdim // 827221345Sdim // Note: interpreter should not have it set since this optimization 828221345Sdim // is limited by dimensions and guarded by flag so in some cases 829221345Sdim // multianewarray() runtime calls will be generated and 830221345Sdim // the bytecode should not be reexecutes (stack will not be reset). 831221345Sdim } else 832221345Sdim return false; 833221345Sdim} 834263508Sdim 835221345Sdim// Helper function for adding JVMState and debug information to node 836221345Sdimvoid GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 837221345Sdim // Add the safepoint edges to the call (or other safepoint). 838263508Sdim 839193323Sed // Make sure dead locals are set to top. This 840193323Sed // should help register allocation time and cut down on the size 841193323Sed // of the deoptimization information. 842193323Sed assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 843193323Sed 844193323Sed // Walk the inline list to fill in the correct set of JVMState's 845193323Sed // Also fill in the associated edges for each JVMState. 846221345Sdim 847193323Sed // If the bytecode needs to be reexecuted we need to put 848193323Sed // the arguments back on the stack. 849193323Sed const bool should_reexecute = jvms()->should_reexecute(); 850193323Sed JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 851193323Sed 852193323Sed // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 853193323Sed // undefined if the bci is different. This is normal for Parse but it 854193323Sed // should not happen for LibraryCallKit because only one bci is processed. 855193323Sed assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 856193323Sed "in LibraryCallKit the reexecute bit should not change"); 857193323Sed 858263508Sdim // If we are guaranteed to throw, we can prune everything but the 859221345Sdim // input to the current bytecode. 860193323Sed bool can_prune_locals = false; 861193323Sed uint stack_slots_not_pruned = 0; 862239462Sdim int inputs = 0, depth = 0; 863239462Sdim if (must_throw) { 864239462Sdim assert(method() == youngest_jvms->method(), "sanity"); 865239462Sdim if (compute_stack_effects(inputs, depth)) { 866193323Sed can_prune_locals = true; 867239462Sdim stack_slots_not_pruned = inputs; 868239462Sdim } 869193323Sed } 870193323Sed 871193323Sed if (env()->should_retain_local_variables()) { 872193323Sed // At any safepoint, this method can get breakpointed, which would 873193323Sed // then require an immediate deoptimization. 874193323Sed can_prune_locals = false; // do not prune locals 875221345Sdim stack_slots_not_pruned = 0; 876221345Sdim } 877221345Sdim 878193323Sed // do not scribble on the input jvms 879193323Sed JVMState* out_jvms = youngest_jvms->clone_deep(C); 880239462Sdim call->set_jvms(out_jvms); // Start jvms list for call node 881 882 // For a known set of bytecodes, the interpreter should reexecute them if 883 // deoptimization happens. We set the reexecute state for them here 884 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 885 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 886 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 887 } 888 889 // Presize the call: 890 DEBUG_ONLY(uint non_debug_edges = call->req()); 891 call->add_req_batch(top(), youngest_jvms->debug_depth()); 892 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 893 894 // Set up edges so that the call looks like this: 895 // Call [state:] ctl io mem fptr retadr 896 // [parms:] parm0 ... parmN 897 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 898 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 899 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 900 // Note that caller debug info precedes callee debug info. 901 902 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 903 uint debug_ptr = call->req(); 904 905 // Loop over the map input edges associated with jvms, add them 906 // to the call node, & reset all offsets to match call node array. 907 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 908 uint debug_end = debug_ptr; 909 uint debug_start = debug_ptr - in_jvms->debug_size(); 910 debug_ptr = debug_start; // back up the ptr 911 912 uint p = debug_start; // walks forward in [debug_start, debug_end) 913 uint j, k, l; 914 SafePointNode* in_map = in_jvms->map(); 915 out_jvms->set_map(call); 916 917 if (can_prune_locals) { 918 assert(in_jvms->method() == out_jvms->method(), "sanity"); 919 // If the current throw can reach an exception handler in this JVMS, 920 // then we must keep everything live that can reach that handler. 921 // As a quick and dirty approximation, we look for any handlers at all. 922 if (in_jvms->method()->has_exception_handlers()) { 923 can_prune_locals = false; 924 } 925 } 926 927 // Add the Locals 928 k = in_jvms->locoff(); 929 l = in_jvms->loc_size(); 930 out_jvms->set_locoff(p); 931 if (!can_prune_locals) { 932 for (j = 0; j < l; j++) 933 call->set_req(p++, in_map->in(k+j)); 934 } else { 935 p += l; // already set to top above by add_req_batch 936 } 937 938 // Add the Expression Stack 939 k = in_jvms->stkoff(); 940 l = in_jvms->sp(); 941 out_jvms->set_stkoff(p); 942 if (!can_prune_locals) { 943 for (j = 0; j < l; j++) 944 call->set_req(p++, in_map->in(k+j)); 945 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 946 // Divide stack into {S0,...,S1}, where S0 is set to top. 947 uint s1 = stack_slots_not_pruned; 948 stack_slots_not_pruned = 0; // for next iteration 949 if (s1 > l) s1 = l; 950 uint s0 = l - s1; 951 p += s0; // skip the tops preinstalled by add_req_batch 952 for (j = s0; j < l; j++) 953 call->set_req(p++, in_map->in(k+j)); 954 } else { 955 p += l; // already set to top above by add_req_batch 956 } 957 958 // Add the Monitors 959 k = in_jvms->monoff(); 960 l = in_jvms->mon_size(); 961 out_jvms->set_monoff(p); 962 for (j = 0; j < l; j++) 963 call->set_req(p++, in_map->in(k+j)); 964 965 // Copy any scalar object fields. 966 k = in_jvms->scloff(); 967 l = in_jvms->scl_size(); 968 out_jvms->set_scloff(p); 969 for (j = 0; j < l; j++) 970 call->set_req(p++, in_map->in(k+j)); 971 972 // Finish the new jvms. 973 out_jvms->set_endoff(p); 974 975 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 976 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 977 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 978 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 979 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 980 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 981 982 // Update the two tail pointers in parallel. 983 out_jvms = out_jvms->caller(); 984 in_jvms = in_jvms->caller(); 985 } 986 987 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 988 989 // Test the correctness of JVMState::debug_xxx accessors: 990 assert(call->jvms()->debug_start() == non_debug_edges, ""); 991 assert(call->jvms()->debug_end() == call->req(), ""); 992 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 993} 994 995bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 996 Bytecodes::Code code = java_bc(); 997 if (code == Bytecodes::_wide) { 998 code = method()->java_code_at_bci(bci() + 1); 999 } 1000 1001 BasicType rtype = T_ILLEGAL; 1002 int rsize = 0; 1003 1004 if (code != Bytecodes::_illegal) { 1005 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1006 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1007 if (rtype < T_CONFLICT) 1008 rsize = type2size[rtype]; 1009 } 1010 1011 switch (code) { 1012 case Bytecodes::_illegal: 1013 return false; 1014 1015 case Bytecodes::_ldc: 1016 case Bytecodes::_ldc_w: 1017 case Bytecodes::_ldc2_w: 1018 inputs = 0; 1019 break; 1020 1021 case Bytecodes::_dup: inputs = 1; break; 1022 case Bytecodes::_dup_x1: inputs = 2; break; 1023 case Bytecodes::_dup_x2: inputs = 3; break; 1024 case Bytecodes::_dup2: inputs = 2; break; 1025 case Bytecodes::_dup2_x1: inputs = 3; break; 1026 case Bytecodes::_dup2_x2: inputs = 4; break; 1027 case Bytecodes::_swap: inputs = 2; break; 1028 case Bytecodes::_arraylength: inputs = 1; break; 1029 1030 case Bytecodes::_getstatic: 1031 case Bytecodes::_putstatic: 1032 case Bytecodes::_getfield: 1033 case Bytecodes::_putfield: 1034 { 1035 bool ignored_will_link; 1036 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1037 int size = field->type()->size(); 1038 bool is_get = (depth >= 0), is_static = (depth & 1); 1039 inputs = (is_static ? 0 : 1); 1040 if (is_get) { 1041 depth = size - inputs; 1042 } else { 1043 inputs += size; // putxxx pops the value from the stack 1044 depth = - inputs; 1045 } 1046 } 1047 break; 1048 1049 case Bytecodes::_invokevirtual: 1050 case Bytecodes::_invokespecial: 1051 case Bytecodes::_invokestatic: 1052 case Bytecodes::_invokedynamic: 1053 case Bytecodes::_invokeinterface: 1054 { 1055 bool ignored_will_link; 1056 ciSignature* declared_signature = NULL; 1057 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1058 assert(declared_signature != NULL, "cannot be null"); 1059 inputs = declared_signature->arg_size_for_bc(code); 1060 int size = declared_signature->return_type()->size(); 1061 depth = size - inputs; 1062 } 1063 break; 1064 1065 case Bytecodes::_multianewarray: 1066 { 1067 ciBytecodeStream iter(method()); 1068 iter.reset_to_bci(bci()); 1069 iter.next(); 1070 inputs = iter.get_dimensions(); 1071 assert(rsize == 1, ""); 1072 depth = rsize - inputs; 1073 } 1074 break; 1075 1076 case Bytecodes::_ireturn: 1077 case Bytecodes::_lreturn: 1078 case Bytecodes::_freturn: 1079 case Bytecodes::_dreturn: 1080 case Bytecodes::_areturn: 1081 assert(rsize = -depth, ""); 1082 inputs = rsize; 1083 break; 1084 1085 case Bytecodes::_jsr: 1086 case Bytecodes::_jsr_w: 1087 inputs = 0; 1088 depth = 1; // S.B. depth=1, not zero 1089 break; 1090 1091 default: 1092 // bytecode produces a typed result 1093 inputs = rsize - depth; 1094 assert(inputs >= 0, ""); 1095 break; 1096 } 1097 1098#ifdef ASSERT 1099 // spot check 1100 int outputs = depth + inputs; 1101 assert(outputs >= 0, "sanity"); 1102 switch (code) { 1103 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1104 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1105 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1106 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1107 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1108 } 1109#endif //ASSERT 1110 1111 return true; 1112} 1113 1114 1115 1116//------------------------------basic_plus_adr--------------------------------- 1117Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1118 // short-circuit a common case 1119 if (offset == intcon(0)) return ptr; 1120 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1121} 1122 1123Node* GraphKit::ConvI2L(Node* offset) { 1124 // short-circuit a common case 1125 jint offset_con = find_int_con(offset, Type::OffsetBot); 1126 if (offset_con != Type::OffsetBot) { 1127 return longcon((jlong) offset_con); 1128 } 1129 return _gvn.transform( new ConvI2LNode(offset)); 1130} 1131 1132Node* GraphKit::ConvI2UL(Node* offset) { 1133 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1134 if (offset_con != (juint) Type::OffsetBot) { 1135 return longcon((julong) offset_con); 1136 } 1137 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1138 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1139 return _gvn.transform( new AndLNode(conv, mask) ); 1140} 1141 1142Node* GraphKit::ConvL2I(Node* offset) { 1143 // short-circuit a common case 1144 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1145 if (offset_con != (jlong)Type::OffsetBot) { 1146 return intcon((int) offset_con); 1147 } 1148 return _gvn.transform( new ConvL2INode(offset)); 1149} 1150 1151//-------------------------load_object_klass----------------------------------- 1152Node* GraphKit::load_object_klass(Node* obj) { 1153 // Special-case a fresh allocation to avoid building nodes: 1154 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1155 if (akls != NULL) return akls; 1156 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1157 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1158} 1159 1160//-------------------------load_array_length----------------------------------- 1161Node* GraphKit::load_array_length(Node* array) { 1162 // Special-case a fresh allocation to avoid building nodes: 1163 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1164 Node *alen; 1165 if (alloc == NULL) { 1166 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1167 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1168 } else { 1169 alen = alloc->Ideal_length(); 1170 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1171 if (ccast != alen) { 1172 alen = _gvn.transform(ccast); 1173 } 1174 } 1175 return alen; 1176} 1177 1178//------------------------------do_null_check---------------------------------- 1179// Helper function to do a NULL pointer check. Returned value is 1180// the incoming address with NULL casted away. You are allowed to use the 1181// not-null value only if you are control dependent on the test. 1182extern int explicit_null_checks_inserted, 1183 explicit_null_checks_elided; 1184Node* GraphKit::null_check_common(Node* value, BasicType type, 1185 // optional arguments for variations: 1186 bool assert_null, 1187 Node* *null_control, 1188 bool speculative) { 1189 assert(!assert_null || null_control == NULL, "not both at once"); 1190 if (stopped()) return top(); 1191 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1192 // For some performance testing, we may wish to suppress null checking. 1193 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1194 return value; 1195 } 1196 explicit_null_checks_inserted++; 1197 1198 // Construct NULL check 1199 Node *chk = NULL; 1200 switch(type) { 1201 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1202 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1203 case T_ARRAY : // fall through 1204 type = T_OBJECT; // simplify further tests 1205 case T_OBJECT : { 1206 const Type *t = _gvn.type( value ); 1207 1208 const TypeOopPtr* tp = t->isa_oopptr(); 1209 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1210 // Only for do_null_check, not any of its siblings: 1211 && !assert_null && null_control == NULL) { 1212 // Usually, any field access or invocation on an unloaded oop type 1213 // will simply fail to link, since the statically linked class is 1214 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1215 // the static class is loaded but the sharper oop type is not. 1216 // Rather than checking for this obscure case in lots of places, 1217 // we simply observe that a null check on an unloaded class 1218 // will always be followed by a nonsense operation, so we 1219 // can just issue the uncommon trap here. 1220 // Our access to the unloaded class will only be correct 1221 // after it has been loaded and initialized, which requires 1222 // a trip through the interpreter. 1223#ifndef PRODUCT 1224 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1225#endif 1226 uncommon_trap(Deoptimization::Reason_unloaded, 1227 Deoptimization::Action_reinterpret, 1228 tp->klass(), "!loaded"); 1229 return top(); 1230 } 1231 1232 if (assert_null) { 1233 // See if the type is contained in NULL_PTR. 1234 // If so, then the value is already null. 1235 if (t->higher_equal(TypePtr::NULL_PTR)) { 1236 explicit_null_checks_elided++; 1237 return value; // Elided null assert quickly! 1238 } 1239 } else { 1240 // See if mixing in the NULL pointer changes type. 1241 // If so, then the NULL pointer was not allowed in the original 1242 // type. In other words, "value" was not-null. 1243 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1244 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1245 explicit_null_checks_elided++; 1246 return value; // Elided null check quickly! 1247 } 1248 } 1249 chk = new CmpPNode( value, null() ); 1250 break; 1251 } 1252 1253 default: 1254 fatal(err_msg_res("unexpected type: %s", type2name(type))); 1255 } 1256 assert(chk != NULL, "sanity check"); 1257 chk = _gvn.transform(chk); 1258 1259 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1260 BoolNode *btst = new BoolNode( chk, btest); 1261 Node *tst = _gvn.transform( btst ); 1262 1263 //----------- 1264 // if peephole optimizations occurred, a prior test existed. 1265 // If a prior test existed, maybe it dominates as we can avoid this test. 1266 if (tst != btst && type == T_OBJECT) { 1267 // At this point we want to scan up the CFG to see if we can 1268 // find an identical test (and so avoid this test altogether). 1269 Node *cfg = control(); 1270 int depth = 0; 1271 while( depth < 16 ) { // Limit search depth for speed 1272 if( cfg->Opcode() == Op_IfTrue && 1273 cfg->in(0)->in(1) == tst ) { 1274 // Found prior test. Use "cast_not_null" to construct an identical 1275 // CastPP (and hence hash to) as already exists for the prior test. 1276 // Return that casted value. 1277 if (assert_null) { 1278 replace_in_map(value, null()); 1279 return null(); // do not issue the redundant test 1280 } 1281 Node *oldcontrol = control(); 1282 set_control(cfg); 1283 Node *res = cast_not_null(value); 1284 set_control(oldcontrol); 1285 explicit_null_checks_elided++; 1286 return res; 1287 } 1288 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1289 if (cfg == NULL) break; // Quit at region nodes 1290 depth++; 1291 } 1292 } 1293 1294 //----------- 1295 // Branch to failure if null 1296 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1297 Deoptimization::DeoptReason reason; 1298 if (assert_null) { 1299 reason = Deoptimization::Reason_null_assert; 1300 } else if (type == T_OBJECT) { 1301 reason = Deoptimization::reason_null_check(speculative); 1302 } else { 1303 reason = Deoptimization::Reason_div0_check; 1304 } 1305 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1306 // ciMethodData::has_trap_at will return a conservative -1 if any 1307 // must-be-null assertion has failed. This could cause performance 1308 // problems for a method after its first do_null_assert failure. 1309 // Consider using 'Reason_class_check' instead? 1310 1311 // To cause an implicit null check, we set the not-null probability 1312 // to the maximum (PROB_MAX). For an explicit check the probability 1313 // is set to a smaller value. 1314 if (null_control != NULL || too_many_traps(reason)) { 1315 // probability is less likely 1316 ok_prob = PROB_LIKELY_MAG(3); 1317 } else if (!assert_null && 1318 (ImplicitNullCheckThreshold > 0) && 1319 method() != NULL && 1320 (method()->method_data()->trap_count(reason) 1321 >= (uint)ImplicitNullCheckThreshold)) { 1322 ok_prob = PROB_LIKELY_MAG(3); 1323 } 1324 1325 if (null_control != NULL) { 1326 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1327 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1328 set_control( _gvn.transform( new IfTrueNode(iff))); 1329 if (null_true == top()) 1330 explicit_null_checks_elided++; 1331 (*null_control) = null_true; 1332 } else { 1333 BuildCutout unless(this, tst, ok_prob); 1334 // Check for optimizer eliding test at parse time 1335 if (stopped()) { 1336 // Failure not possible; do not bother making uncommon trap. 1337 explicit_null_checks_elided++; 1338 } else if (assert_null) { 1339 uncommon_trap(reason, 1340 Deoptimization::Action_make_not_entrant, 1341 NULL, "assert_null"); 1342 } else { 1343 replace_in_map(value, zerocon(type)); 1344 builtin_throw(reason); 1345 } 1346 } 1347 1348 // Must throw exception, fall-thru not possible? 1349 if (stopped()) { 1350 return top(); // No result 1351 } 1352 1353 if (assert_null) { 1354 // Cast obj to null on this path. 1355 replace_in_map(value, zerocon(type)); 1356 return zerocon(type); 1357 } 1358 1359 // Cast obj to not-null on this path, if there is no null_control. 1360 // (If there is a null_control, a non-null value may come back to haunt us.) 1361 if (type == T_OBJECT) { 1362 Node* cast = cast_not_null(value, false); 1363 if (null_control == NULL || (*null_control) == top()) 1364 replace_in_map(value, cast); 1365 value = cast; 1366 } 1367 1368 return value; 1369} 1370 1371 1372//------------------------------cast_not_null---------------------------------- 1373// Cast obj to not-null on this path 1374Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1375 const Type *t = _gvn.type(obj); 1376 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1377 // Object is already not-null? 1378 if( t == t_not_null ) return obj; 1379 1380 Node *cast = new CastPPNode(obj,t_not_null); 1381 cast->init_req(0, control()); 1382 cast = _gvn.transform( cast ); 1383 1384 // Scan for instances of 'obj' in the current JVM mapping. 1385 // These instances are known to be not-null after the test. 1386 if (do_replace_in_map) 1387 replace_in_map(obj, cast); 1388 1389 return cast; // Return casted value 1390} 1391 1392 1393//--------------------------replace_in_map------------------------------------- 1394void GraphKit::replace_in_map(Node* old, Node* neww) { 1395 if (old == neww) { 1396 return; 1397 } 1398 1399 map()->replace_edge(old, neww); 1400 1401 // Note: This operation potentially replaces any edge 1402 // on the map. This includes locals, stack, and monitors 1403 // of the current (innermost) JVM state. 1404 1405 // don't let inconsistent types from profiling escape this 1406 // method 1407 1408 const Type* told = _gvn.type(old); 1409 const Type* tnew = _gvn.type(neww); 1410 1411 if (!tnew->higher_equal(told)) { 1412 return; 1413 } 1414 1415 map()->record_replaced_node(old, neww); 1416} 1417 1418 1419//============================================================================= 1420//--------------------------------memory--------------------------------------- 1421Node* GraphKit::memory(uint alias_idx) { 1422 MergeMemNode* mem = merged_memory(); 1423 Node* p = mem->memory_at(alias_idx); 1424 _gvn.set_type(p, Type::MEMORY); // must be mapped 1425 return p; 1426} 1427 1428//-----------------------------reset_memory------------------------------------ 1429Node* GraphKit::reset_memory() { 1430 Node* mem = map()->memory(); 1431 // do not use this node for any more parsing! 1432 debug_only( map()->set_memory((Node*)NULL) ); 1433 return _gvn.transform( mem ); 1434} 1435 1436//------------------------------set_all_memory--------------------------------- 1437void GraphKit::set_all_memory(Node* newmem) { 1438 Node* mergemem = MergeMemNode::make(newmem); 1439 gvn().set_type_bottom(mergemem); 1440 map()->set_memory(mergemem); 1441} 1442 1443//------------------------------set_all_memory_call---------------------------- 1444void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1445 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1446 set_all_memory(newmem); 1447} 1448 1449//============================================================================= 1450// 1451// parser factory methods for MemNodes 1452// 1453// These are layered on top of the factory methods in LoadNode and StoreNode, 1454// and integrate with the parser's memory state and _gvn engine. 1455// 1456 1457// factory methods in "int adr_idx" 1458Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1459 int adr_idx, 1460 MemNode::MemOrd mo, bool require_atomic_access) { 1461 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1462 const TypePtr* adr_type = NULL; // debug-mode-only argument 1463 debug_only(adr_type = C->get_adr_type(adr_idx)); 1464 Node* mem = memory(adr_idx); 1465 Node* ld; 1466 if (require_atomic_access && bt == T_LONG) { 1467 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo); 1468 } else if (require_atomic_access && bt == T_DOUBLE) { 1469 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo); 1470 } else { 1471 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo); 1472 } 1473 ld = _gvn.transform(ld); 1474 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1475 // Improve graph before escape analysis and boxing elimination. 1476 record_for_igvn(ld); 1477 } 1478 return ld; 1479} 1480 1481Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1482 int adr_idx, 1483 MemNode::MemOrd mo, 1484 bool require_atomic_access) { 1485 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1486 const TypePtr* adr_type = NULL; 1487 debug_only(adr_type = C->get_adr_type(adr_idx)); 1488 Node *mem = memory(adr_idx); 1489 Node* st; 1490 if (require_atomic_access && bt == T_LONG) { 1491 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1492 } else if (require_atomic_access && bt == T_DOUBLE) { 1493 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1494 } else { 1495 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1496 } 1497 st = _gvn.transform(st); 1498 set_memory(st, adr_idx); 1499 // Back-to-back stores can only remove intermediate store with DU info 1500 // so push on worklist for optimizer. 1501 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1502 record_for_igvn(st); 1503 1504 return st; 1505} 1506 1507 1508void GraphKit::pre_barrier(bool do_load, 1509 Node* ctl, 1510 Node* obj, 1511 Node* adr, 1512 uint adr_idx, 1513 Node* val, 1514 const TypeOopPtr* val_type, 1515 Node* pre_val, 1516 BasicType bt) { 1517 1518 BarrierSet* bs = Universe::heap()->barrier_set(); 1519 set_control(ctl); 1520 switch (bs->kind()) { 1521 case BarrierSet::G1SATBCT: 1522 case BarrierSet::G1SATBCTLogging: 1523 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1524 break; 1525 1526 case BarrierSet::CardTableModRef: 1527 case BarrierSet::CardTableExtension: 1528 case BarrierSet::ModRef: 1529 break; 1530 1531 default : 1532 ShouldNotReachHere(); 1533 1534 } 1535} 1536 1537bool GraphKit::can_move_pre_barrier() const { 1538 BarrierSet* bs = Universe::heap()->barrier_set(); 1539 switch (bs->kind()) { 1540 case BarrierSet::G1SATBCT: 1541 case BarrierSet::G1SATBCTLogging: 1542 return true; // Can move it if no safepoint 1543 1544 case BarrierSet::CardTableModRef: 1545 case BarrierSet::CardTableExtension: 1546 case BarrierSet::ModRef: 1547 return true; // There is no pre-barrier 1548 1549 default : 1550 ShouldNotReachHere(); 1551 } 1552 return false; 1553} 1554 1555void GraphKit::post_barrier(Node* ctl, 1556 Node* store, 1557 Node* obj, 1558 Node* adr, 1559 uint adr_idx, 1560 Node* val, 1561 BasicType bt, 1562 bool use_precise) { 1563 BarrierSet* bs = Universe::heap()->barrier_set(); 1564 set_control(ctl); 1565 switch (bs->kind()) { 1566 case BarrierSet::G1SATBCT: 1567 case BarrierSet::G1SATBCTLogging: 1568 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1569 break; 1570 1571 case BarrierSet::CardTableModRef: 1572 case BarrierSet::CardTableExtension: 1573 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1574 break; 1575 1576 case BarrierSet::ModRef: 1577 break; 1578 1579 default : 1580 ShouldNotReachHere(); 1581 1582 } 1583} 1584 1585Node* GraphKit::store_oop(Node* ctl, 1586 Node* obj, 1587 Node* adr, 1588 const TypePtr* adr_type, 1589 Node* val, 1590 const TypeOopPtr* val_type, 1591 BasicType bt, 1592 bool use_precise, 1593 MemNode::MemOrd mo) { 1594 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1595 // could be delayed during Parse (for example, in adjust_map_after_if()). 1596 // Execute transformation here to avoid barrier generation in such case. 1597 if (_gvn.type(val) == TypePtr::NULL_PTR) 1598 val = _gvn.makecon(TypePtr::NULL_PTR); 1599 1600 set_control(ctl); 1601 if (stopped()) return top(); // Dead path ? 1602 1603 assert(bt == T_OBJECT, "sanity"); 1604 assert(val != NULL, "not dead path"); 1605 uint adr_idx = C->get_alias_index(adr_type); 1606 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1607 1608 pre_barrier(true /* do_load */, 1609 control(), obj, adr, adr_idx, val, val_type, 1610 NULL /* pre_val */, 1611 bt); 1612 1613 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo); 1614 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1615 return store; 1616} 1617 1618// Could be an array or object we don't know at compile time (unsafe ref.) 1619Node* GraphKit::store_oop_to_unknown(Node* ctl, 1620 Node* obj, // containing obj 1621 Node* adr, // actual adress to store val at 1622 const TypePtr* adr_type, 1623 Node* val, 1624 BasicType bt, 1625 MemNode::MemOrd mo) { 1626 Compile::AliasType* at = C->alias_type(adr_type); 1627 const TypeOopPtr* val_type = NULL; 1628 if (adr_type->isa_instptr()) { 1629 if (at->field() != NULL) { 1630 // known field. This code is a copy of the do_put_xxx logic. 1631 ciField* field = at->field(); 1632 if (!field->type()->is_loaded()) { 1633 val_type = TypeInstPtr::BOTTOM; 1634 } else { 1635 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1636 } 1637 } 1638 } else if (adr_type->isa_aryptr()) { 1639 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1640 } 1641 if (val_type == NULL) { 1642 val_type = TypeInstPtr::BOTTOM; 1643 } 1644 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo); 1645} 1646 1647 1648//-------------------------array_element_address------------------------- 1649Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1650 const TypeInt* sizetype) { 1651 uint shift = exact_log2(type2aelembytes(elembt)); 1652 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1653 1654 // short-circuit a common case (saves lots of confusing waste motion) 1655 jint idx_con = find_int_con(idx, -1); 1656 if (idx_con >= 0) { 1657 intptr_t offset = header + ((intptr_t)idx_con << shift); 1658 return basic_plus_adr(ary, offset); 1659 } 1660 1661 // must be correct type for alignment purposes 1662 Node* base = basic_plus_adr(ary, header); 1663#ifdef _LP64 1664 // The scaled index operand to AddP must be a clean 64-bit value. 1665 // Java allows a 32-bit int to be incremented to a negative 1666 // value, which appears in a 64-bit register as a large 1667 // positive number. Using that large positive number as an 1668 // operand in pointer arithmetic has bad consequences. 1669 // On the other hand, 32-bit overflow is rare, and the possibility 1670 // can often be excluded, if we annotate the ConvI2L node with 1671 // a type assertion that its value is known to be a small positive 1672 // number. (The prior range check has ensured this.) 1673 // This assertion is used by ConvI2LNode::Ideal. 1674 int index_max = max_jint - 1; // array size is max_jint, index is one less 1675 if (sizetype != NULL) index_max = sizetype->_hi - 1; 1676 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 1677 idx = _gvn.transform( new ConvI2LNode(idx, lidxtype) ); 1678#endif 1679 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1680 return basic_plus_adr(ary, base, scale); 1681} 1682 1683//-------------------------load_array_element------------------------- 1684Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1685 const Type* elemtype = arytype->elem(); 1686 BasicType elembt = elemtype->array_element_basic_type(); 1687 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1688 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1689 return ld; 1690} 1691 1692//-------------------------set_arguments_for_java_call------------------------- 1693// Arguments (pre-popped from the stack) are taken from the JVMS. 1694void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1695 // Add the call arguments: 1696 uint nargs = call->method()->arg_size(); 1697 for (uint i = 0; i < nargs; i++) { 1698 Node* arg = argument(i); 1699 call->init_req(i + TypeFunc::Parms, arg); 1700 } 1701} 1702 1703//---------------------------set_edges_for_java_call--------------------------- 1704// Connect a newly created call into the current JVMS. 1705// A return value node (if any) is returned from set_edges_for_java_call. 1706void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1707 1708 // Add the predefined inputs: 1709 call->init_req( TypeFunc::Control, control() ); 1710 call->init_req( TypeFunc::I_O , i_o() ); 1711 call->init_req( TypeFunc::Memory , reset_memory() ); 1712 call->init_req( TypeFunc::FramePtr, frameptr() ); 1713 call->init_req( TypeFunc::ReturnAdr, top() ); 1714 1715 add_safepoint_edges(call, must_throw); 1716 1717 Node* xcall = _gvn.transform(call); 1718 1719 if (xcall == top()) { 1720 set_control(top()); 1721 return; 1722 } 1723 assert(xcall == call, "call identity is stable"); 1724 1725 // Re-use the current map to produce the result. 1726 1727 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1728 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1729 set_all_memory_call(xcall, separate_io_proj); 1730 1731 //return xcall; // no need, caller already has it 1732} 1733 1734Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1735 if (stopped()) return top(); // maybe the call folded up? 1736 1737 // Capture the return value, if any. 1738 Node* ret; 1739 if (call->method() == NULL || 1740 call->method()->return_type()->basic_type() == T_VOID) 1741 ret = top(); 1742 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1743 1744 // Note: Since any out-of-line call can produce an exception, 1745 // we always insert an I_O projection from the call into the result. 1746 1747 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1748 1749 if (separate_io_proj) { 1750 // The caller requested separate projections be used by the fall 1751 // through and exceptional paths, so replace the projections for 1752 // the fall through path. 1753 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1754 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1755 } 1756 return ret; 1757} 1758 1759//--------------------set_predefined_input_for_runtime_call-------------------- 1760// Reading and setting the memory state is way conservative here. 1761// The real problem is that I am not doing real Type analysis on memory, 1762// so I cannot distinguish card mark stores from other stores. Across a GC 1763// point the Store Barrier and the card mark memory has to agree. I cannot 1764// have a card mark store and its barrier split across the GC point from 1765// either above or below. Here I get that to happen by reading ALL of memory. 1766// A better answer would be to separate out card marks from other memory. 1767// For now, return the input memory state, so that it can be reused 1768// after the call, if this call has restricted memory effects. 1769Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1770 // Set fixed predefined input arguments 1771 Node* memory = reset_memory(); 1772 call->init_req( TypeFunc::Control, control() ); 1773 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1774 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1775 call->init_req( TypeFunc::FramePtr, frameptr() ); 1776 call->init_req( TypeFunc::ReturnAdr, top() ); 1777 return memory; 1778} 1779 1780//-------------------set_predefined_output_for_runtime_call-------------------- 1781// Set control and memory (not i_o) from the call. 1782// If keep_mem is not NULL, use it for the output state, 1783// except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1784// If hook_mem is NULL, this call produces no memory effects at all. 1785// If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1786// then only that memory slice is taken from the call. 1787// In the last case, we must put an appropriate memory barrier before 1788// the call, so as to create the correct anti-dependencies on loads 1789// preceding the call. 1790void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1791 Node* keep_mem, 1792 const TypePtr* hook_mem) { 1793 // no i/o 1794 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1795 if (keep_mem) { 1796 // First clone the existing memory state 1797 set_all_memory(keep_mem); 1798 if (hook_mem != NULL) { 1799 // Make memory for the call 1800 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1801 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1802 // We also use hook_mem to extract specific effects from arraycopy stubs. 1803 set_memory(mem, hook_mem); 1804 } 1805 // ...else the call has NO memory effects. 1806 1807 // Make sure the call advertises its memory effects precisely. 1808 // This lets us build accurate anti-dependences in gcm.cpp. 1809 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1810 "call node must be constructed correctly"); 1811 } else { 1812 assert(hook_mem == NULL, ""); 1813 // This is not a "slow path" call; all memory comes from the call. 1814 set_all_memory_call(call); 1815 } 1816} 1817 1818 1819// Replace the call with the current state of the kit. 1820void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1821 JVMState* ejvms = NULL; 1822 if (has_exceptions()) { 1823 ejvms = transfer_exceptions_into_jvms(); 1824 } 1825 1826 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1827 ReplacedNodes replaced_nodes_exception; 1828 Node* ex_ctl = top(); 1829 1830 SafePointNode* final_state = stop(); 1831 1832 // Find all the needed outputs of this call 1833 CallProjections callprojs; 1834 call->extract_projections(&callprojs, true); 1835 1836 Node* init_mem = call->in(TypeFunc::Memory); 1837 Node* final_mem = final_state->in(TypeFunc::Memory); 1838 Node* final_ctl = final_state->in(TypeFunc::Control); 1839 Node* final_io = final_state->in(TypeFunc::I_O); 1840 1841 // Replace all the old call edges with the edges from the inlining result 1842 if (callprojs.fallthrough_catchproj != NULL) { 1843 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1844 } 1845 if (callprojs.fallthrough_memproj != NULL) { 1846 if (final_mem->is_MergeMem()) { 1847 // Parser's exits MergeMem was not transformed but may be optimized 1848 final_mem = _gvn.transform(final_mem); 1849 } 1850 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1851 } 1852 if (callprojs.fallthrough_ioproj != NULL) { 1853 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1854 } 1855 1856 // Replace the result with the new result if it exists and is used 1857 if (callprojs.resproj != NULL && result != NULL) { 1858 C->gvn_replace_by(callprojs.resproj, result); 1859 } 1860 1861 if (ejvms == NULL) { 1862 // No exception edges to simply kill off those paths 1863 if (callprojs.catchall_catchproj != NULL) { 1864 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1865 } 1866 if (callprojs.catchall_memproj != NULL) { 1867 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1868 } 1869 if (callprojs.catchall_ioproj != NULL) { 1870 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1871 } 1872 // Replace the old exception object with top 1873 if (callprojs.exobj != NULL) { 1874 C->gvn_replace_by(callprojs.exobj, C->top()); 1875 } 1876 } else { 1877 GraphKit ekit(ejvms); 1878 1879 // Load my combined exception state into the kit, with all phis transformed: 1880 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1881 replaced_nodes_exception = ex_map->replaced_nodes(); 1882 1883 Node* ex_oop = ekit.use_exception_state(ex_map); 1884 1885 if (callprojs.catchall_catchproj != NULL) { 1886 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1887 ex_ctl = ekit.control(); 1888 } 1889 if (callprojs.catchall_memproj != NULL) { 1890 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1891 } 1892 if (callprojs.catchall_ioproj != NULL) { 1893 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1894 } 1895 1896 // Replace the old exception object with the newly created one 1897 if (callprojs.exobj != NULL) { 1898 C->gvn_replace_by(callprojs.exobj, ex_oop); 1899 } 1900 } 1901 1902 // Disconnect the call from the graph 1903 call->disconnect_inputs(NULL, C); 1904 C->gvn_replace_by(call, C->top()); 1905 1906 // Clean up any MergeMems that feed other MergeMems since the 1907 // optimizer doesn't like that. 1908 if (final_mem->is_MergeMem()) { 1909 Node_List wl; 1910 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1911 Node* m = i.get(); 1912 if (m->is_MergeMem() && !wl.contains(m)) { 1913 wl.push(m); 1914 } 1915 } 1916 while (wl.size() > 0) { 1917 _gvn.transform(wl.pop()); 1918 } 1919 } 1920 1921 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1922 replaced_nodes.apply(C, final_ctl); 1923 } 1924 if (!ex_ctl->is_top() && do_replaced_nodes) { 1925 replaced_nodes_exception.apply(C, ex_ctl); 1926 } 1927} 1928 1929 1930//------------------------------increment_counter------------------------------ 1931// for statistics: increment a VM counter by 1 1932 1933void GraphKit::increment_counter(address counter_addr) { 1934 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1935 increment_counter(adr1); 1936} 1937 1938void GraphKit::increment_counter(Node* counter_addr) { 1939 int adr_type = Compile::AliasIdxRaw; 1940 Node* ctrl = control(); 1941 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1942 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1943 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1944} 1945 1946 1947//------------------------------uncommon_trap---------------------------------- 1948// Bail out to the interpreter in mid-method. Implemented by calling the 1949// uncommon_trap blob. This helper function inserts a runtime call with the 1950// right debug info. 1951void GraphKit::uncommon_trap(int trap_request, 1952 ciKlass* klass, const char* comment, 1953 bool must_throw, 1954 bool keep_exact_action) { 1955 if (failing()) stop(); 1956 if (stopped()) return; // trap reachable? 1957 1958 // Note: If ProfileTraps is true, and if a deopt. actually 1959 // occurs here, the runtime will make sure an MDO exists. There is 1960 // no need to call method()->ensure_method_data() at this point. 1961 1962 // Set the stack pointer to the right value for reexecution: 1963 set_sp(reexecute_sp()); 1964 1965#ifdef ASSERT 1966 if (!must_throw) { 1967 // Make sure the stack has at least enough depth to execute 1968 // the current bytecode. 1969 int inputs, ignored_depth; 1970 if (compute_stack_effects(inputs, ignored_depth)) { 1971 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1972 Bytecodes::name(java_bc()), sp(), inputs)); 1973 } 1974 } 1975#endif 1976 1977 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1978 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1979 1980 switch (action) { 1981 case Deoptimization::Action_maybe_recompile: 1982 case Deoptimization::Action_reinterpret: 1983 // Temporary fix for 6529811 to allow virtual calls to be sure they 1984 // get the chance to go from mono->bi->mega 1985 if (!keep_exact_action && 1986 Deoptimization::trap_request_index(trap_request) < 0 && 1987 too_many_recompiles(reason)) { 1988 // This BCI is causing too many recompilations. 1989 if (C->log() != NULL) { 1990 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 1991 Deoptimization::trap_reason_name(reason), 1992 Deoptimization::trap_action_name(action)); 1993 } 1994 action = Deoptimization::Action_none; 1995 trap_request = Deoptimization::make_trap_request(reason, action); 1996 } else { 1997 C->set_trap_can_recompile(true); 1998 } 1999 break; 2000 case Deoptimization::Action_make_not_entrant: 2001 C->set_trap_can_recompile(true); 2002 break; 2003#ifdef ASSERT 2004 case Deoptimization::Action_none: 2005 case Deoptimization::Action_make_not_compilable: 2006 break; 2007 default: 2008 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action))); 2009 break; 2010#endif 2011 } 2012 2013 if (TraceOptoParse) { 2014 char buf[100]; 2015 tty->print_cr("Uncommon trap %s at bci:%d", 2016 Deoptimization::format_trap_request(buf, sizeof(buf), 2017 trap_request), bci()); 2018 } 2019 2020 CompileLog* log = C->log(); 2021 if (log != NULL) { 2022 int kid = (klass == NULL)? -1: log->identify(klass); 2023 log->begin_elem("uncommon_trap bci='%d'", bci()); 2024 char buf[100]; 2025 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2026 trap_request)); 2027 if (kid >= 0) log->print(" klass='%d'", kid); 2028 if (comment != NULL) log->print(" comment='%s'", comment); 2029 log->end_elem(); 2030 } 2031 2032 // Make sure any guarding test views this path as very unlikely 2033 Node *i0 = control()->in(0); 2034 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2035 IfNode *iff = i0->as_If(); 2036 float f = iff->_prob; // Get prob 2037 if (control()->Opcode() == Op_IfTrue) { 2038 if (f > PROB_UNLIKELY_MAG(4)) 2039 iff->_prob = PROB_MIN; 2040 } else { 2041 if (f < PROB_LIKELY_MAG(4)) 2042 iff->_prob = PROB_MAX; 2043 } 2044 } 2045 2046 // Clear out dead values from the debug info. 2047 kill_dead_locals(); 2048 2049 // Now insert the uncommon trap subroutine call 2050 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2051 const TypePtr* no_memory_effects = NULL; 2052 // Pass the index of the class to be loaded 2053 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2054 (must_throw ? RC_MUST_THROW : 0), 2055 OptoRuntime::uncommon_trap_Type(), 2056 call_addr, "uncommon_trap", no_memory_effects, 2057 intcon(trap_request)); 2058 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2059 "must extract request correctly from the graph"); 2060 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2061 2062 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2063 // The debug info is the only real input to this call. 2064 2065 // Halt-and-catch fire here. The above call should never return! 2066 HaltNode* halt = new HaltNode(control(), frameptr()); 2067 _gvn.set_type_bottom(halt); 2068 root()->add_req(halt); 2069 2070 stop_and_kill_map(); 2071} 2072 2073 2074//--------------------------just_allocated_object------------------------------ 2075// Report the object that was just allocated. 2076// It must be the case that there are no intervening safepoints. 2077// We use this to determine if an object is so "fresh" that 2078// it does not require card marks. 2079Node* GraphKit::just_allocated_object(Node* current_control) { 2080 if (C->recent_alloc_ctl() == current_control) 2081 return C->recent_alloc_obj(); 2082 return NULL; 2083} 2084 2085 2086void GraphKit::round_double_arguments(ciMethod* dest_method) { 2087 // (Note: TypeFunc::make has a cache that makes this fast.) 2088 const TypeFunc* tf = TypeFunc::make(dest_method); 2089 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2090 for (int j = 0; j < nargs; j++) { 2091 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2092 if( targ->basic_type() == T_DOUBLE ) { 2093 // If any parameters are doubles, they must be rounded before 2094 // the call, dstore_rounding does gvn.transform 2095 Node *arg = argument(j); 2096 arg = dstore_rounding(arg); 2097 set_argument(j, arg); 2098 } 2099 } 2100} 2101 2102/** 2103 * Record profiling data exact_kls for Node n with the type system so 2104 * that it can propagate it (speculation) 2105 * 2106 * @param n node that the type applies to 2107 * @param exact_kls type from profiling 2108 * @param maybe_null did profiling see null? 2109 * 2110 * @return node with improved type 2111 */ 2112Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) { 2113 const Type* current_type = _gvn.type(n); 2114 assert(UseTypeSpeculation, "type speculation must be on"); 2115 2116 const TypePtr* speculative = current_type->speculative(); 2117 2118 // Should the klass from the profile be recorded in the speculative type? 2119 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2120 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2121 const TypeOopPtr* xtype = tklass->as_instance_type(); 2122 assert(xtype->klass_is_exact(), "Should be exact"); 2123 // Any reason to believe n is not null (from this profiling or a previous one)? 2124 const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2125 // record the new speculative type's depth 2126 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2127 speculative = speculative->with_inline_depth(jvms()->depth()); 2128 } else if (current_type->would_improve_ptr(maybe_null)) { 2129 // Profiling report that null was never seen so we can change the 2130 // speculative type to non null ptr. 2131 assert(!maybe_null, "nothing to improve"); 2132 if (speculative == NULL) { 2133 speculative = TypePtr::NOTNULL; 2134 } else { 2135 const TypePtr* ptr = TypePtr::NOTNULL; 2136 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2137 } 2138 } 2139 2140 if (speculative != current_type->speculative()) { 2141 // Build a type with a speculative type (what we think we know 2142 // about the type but will need a guard when we use it) 2143 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2144 // We're changing the type, we need a new CheckCast node to carry 2145 // the new type. The new type depends on the control: what 2146 // profiling tells us is only valid from here as far as we can 2147 // tell. 2148 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2149 cast = _gvn.transform(cast); 2150 replace_in_map(n, cast); 2151 n = cast; 2152 } 2153 2154 return n; 2155} 2156 2157/** 2158 * Record profiling data from receiver profiling at an invoke with the 2159 * type system so that it can propagate it (speculation) 2160 * 2161 * @param n receiver node 2162 * 2163 * @return node with improved type 2164 */ 2165Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2166 if (!UseTypeSpeculation) { 2167 return n; 2168 } 2169 ciKlass* exact_kls = profile_has_unique_klass(); 2170 bool maybe_null = true; 2171 if (java_bc() == Bytecodes::_checkcast || 2172 java_bc() == Bytecodes::_instanceof || 2173 java_bc() == Bytecodes::_aastore) { 2174 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2175 bool maybe_null = data == NULL ? true : data->as_BitData()->null_seen(); 2176 } 2177 return record_profile_for_speculation(n, exact_kls, maybe_null); 2178 return n; 2179} 2180 2181/** 2182 * Record profiling data from argument profiling at an invoke with the 2183 * type system so that it can propagate it (speculation) 2184 * 2185 * @param dest_method target method for the call 2186 * @param bc what invoke bytecode is this? 2187 */ 2188void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2189 if (!UseTypeSpeculation) { 2190 return; 2191 } 2192 const TypeFunc* tf = TypeFunc::make(dest_method); 2193 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2194 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2195 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2196 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2197 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2198 bool maybe_null = true; 2199 ciKlass* better_type = NULL; 2200 if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) { 2201 record_profile_for_speculation(argument(j), better_type, maybe_null); 2202 } 2203 i++; 2204 } 2205 } 2206} 2207 2208/** 2209 * Record profiling data from parameter profiling at an invoke with 2210 * the type system so that it can propagate it (speculation) 2211 */ 2212void GraphKit::record_profiled_parameters_for_speculation() { 2213 if (!UseTypeSpeculation) { 2214 return; 2215 } 2216 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2217 if (_gvn.type(local(i))->isa_oopptr()) { 2218 bool maybe_null = true; 2219 ciKlass* better_type = NULL; 2220 if (method()->parameter_profiled_type(j, better_type, maybe_null)) { 2221 record_profile_for_speculation(local(i), better_type, maybe_null); 2222 } 2223 j++; 2224 } 2225 } 2226} 2227 2228/** 2229 * Record profiling data from return value profiling at an invoke with 2230 * the type system so that it can propagate it (speculation) 2231 */ 2232void GraphKit::record_profiled_return_for_speculation() { 2233 if (!UseTypeSpeculation) { 2234 return; 2235 } 2236 bool maybe_null = true; 2237 ciKlass* better_type = NULL; 2238 if (method()->return_profiled_type(bci(), better_type, maybe_null)) { 2239 // If profiling reports a single type for the return value, 2240 // feed it to the type system so it can propagate it as a 2241 // speculative type 2242 record_profile_for_speculation(stack(sp()-1), better_type, maybe_null); 2243 } 2244} 2245 2246void GraphKit::round_double_result(ciMethod* dest_method) { 2247 // A non-strict method may return a double value which has an extended 2248 // exponent, but this must not be visible in a caller which is 'strict' 2249 // If a strict caller invokes a non-strict callee, round a double result 2250 2251 BasicType result_type = dest_method->return_type()->basic_type(); 2252 assert( method() != NULL, "must have caller context"); 2253 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2254 // Destination method's return value is on top of stack 2255 // dstore_rounding() does gvn.transform 2256 Node *result = pop_pair(); 2257 result = dstore_rounding(result); 2258 push_pair(result); 2259 } 2260} 2261 2262// rounding for strict float precision conformance 2263Node* GraphKit::precision_rounding(Node* n) { 2264 return UseStrictFP && _method->flags().is_strict() 2265 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2266 ? _gvn.transform( new RoundFloatNode(0, n) ) 2267 : n; 2268} 2269 2270// rounding for strict double precision conformance 2271Node* GraphKit::dprecision_rounding(Node *n) { 2272 return UseStrictFP && _method->flags().is_strict() 2273 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2274 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2275 : n; 2276} 2277 2278// rounding for non-strict double stores 2279Node* GraphKit::dstore_rounding(Node* n) { 2280 return Matcher::strict_fp_requires_explicit_rounding 2281 && UseSSE <= 1 2282 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2283 : n; 2284} 2285 2286//============================================================================= 2287// Generate a fast path/slow path idiom. Graph looks like: 2288// [foo] indicates that 'foo' is a parameter 2289// 2290// [in] NULL 2291// \ / 2292// CmpP 2293// Bool ne 2294// If 2295// / \ 2296// True False-<2> 2297// / | 2298// / cast_not_null 2299// Load | | ^ 2300// [fast_test] | | 2301// gvn to opt_test | | 2302// / \ | <1> 2303// True False | 2304// | \\ | 2305// [slow_call] \[fast_result] 2306// Ctl Val \ \ 2307// | \ \ 2308// Catch <1> \ \ 2309// / \ ^ \ \ 2310// Ex No_Ex | \ \ 2311// | \ \ | \ <2> \ 2312// ... \ [slow_res] | | \ [null_result] 2313// \ \--+--+--- | | 2314// \ | / \ | / 2315// --------Region Phi 2316// 2317//============================================================================= 2318// Code is structured as a series of driver functions all called 'do_XXX' that 2319// call a set of helper functions. Helper functions first, then drivers. 2320 2321//------------------------------null_check_oop--------------------------------- 2322// Null check oop. Set null-path control into Region in slot 3. 2323// Make a cast-not-nullness use the other not-null control. Return cast. 2324Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2325 bool never_see_null, 2326 bool safe_for_replace, 2327 bool speculative) { 2328 // Initial NULL check taken path 2329 (*null_control) = top(); 2330 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2331 2332 // Generate uncommon_trap: 2333 if (never_see_null && (*null_control) != top()) { 2334 // If we see an unexpected null at a check-cast we record it and force a 2335 // recompile; the offending check-cast will be compiled to handle NULLs. 2336 // If we see more than one offending BCI, then all checkcasts in the 2337 // method will be compiled to handle NULLs. 2338 PreserveJVMState pjvms(this); 2339 set_control(*null_control); 2340 replace_in_map(value, null()); 2341 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2342 uncommon_trap(reason, 2343 Deoptimization::Action_make_not_entrant); 2344 (*null_control) = top(); // NULL path is dead 2345 } 2346 if ((*null_control) == top() && safe_for_replace) { 2347 replace_in_map(value, cast); 2348 } 2349 2350 // Cast away null-ness on the result 2351 return cast; 2352} 2353 2354//------------------------------opt_iff---------------------------------------- 2355// Optimize the fast-check IfNode. Set the fast-path region slot 2. 2356// Return slow-path control. 2357Node* GraphKit::opt_iff(Node* region, Node* iff) { 2358 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2359 2360 // Fast path taken; set region slot 2 2361 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2362 region->init_req(2,fast_taken); // Capture fast-control 2363 2364 // Fast path not-taken, i.e. slow path 2365 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2366 return slow_taken; 2367} 2368 2369//-----------------------------make_runtime_call------------------------------- 2370Node* GraphKit::make_runtime_call(int flags, 2371 const TypeFunc* call_type, address call_addr, 2372 const char* call_name, 2373 const TypePtr* adr_type, 2374 // The following parms are all optional. 2375 // The first NULL ends the list. 2376 Node* parm0, Node* parm1, 2377 Node* parm2, Node* parm3, 2378 Node* parm4, Node* parm5, 2379 Node* parm6, Node* parm7) { 2380 // Slow-path call 2381 bool is_leaf = !(flags & RC_NO_LEAF); 2382 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2383 if (call_name == NULL) { 2384 assert(!is_leaf, "must supply name for leaf"); 2385 call_name = OptoRuntime::stub_name(call_addr); 2386 } 2387 CallNode* call; 2388 if (!is_leaf) { 2389 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2390 bci(), adr_type); 2391 } else if (flags & RC_NO_FP) { 2392 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2393 } else { 2394 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2395 } 2396 2397 // The following is similar to set_edges_for_java_call, 2398 // except that the memory effects of the call are restricted to AliasIdxRaw. 2399 2400 // Slow path call has no side-effects, uses few values 2401 bool wide_in = !(flags & RC_NARROW_MEM); 2402 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2403 2404 Node* prev_mem = NULL; 2405 if (wide_in) { 2406 prev_mem = set_predefined_input_for_runtime_call(call); 2407 } else { 2408 assert(!wide_out, "narrow in => narrow out"); 2409 Node* narrow_mem = memory(adr_type); 2410 prev_mem = reset_memory(); 2411 map()->set_memory(narrow_mem); 2412 set_predefined_input_for_runtime_call(call); 2413 } 2414 2415 // Hook each parm in order. Stop looking at the first NULL. 2416 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2417 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2418 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2419 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2420 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2421 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2422 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2423 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2424 /* close each nested if ===> */ } } } } } } } } 2425 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2426 2427 if (!is_leaf) { 2428 // Non-leaves can block and take safepoints: 2429 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2430 } 2431 // Non-leaves can throw exceptions: 2432 if (has_io) { 2433 call->set_req(TypeFunc::I_O, i_o()); 2434 } 2435 2436 if (flags & RC_UNCOMMON) { 2437 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2438 // (An "if" probability corresponds roughly to an unconditional count. 2439 // Sort of.) 2440 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2441 } 2442 2443 Node* c = _gvn.transform(call); 2444 assert(c == call, "cannot disappear"); 2445 2446 if (wide_out) { 2447 // Slow path call has full side-effects. 2448 set_predefined_output_for_runtime_call(call); 2449 } else { 2450 // Slow path call has few side-effects, and/or sets few values. 2451 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2452 } 2453 2454 if (has_io) { 2455 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2456 } 2457 return call; 2458 2459} 2460 2461//------------------------------merge_memory----------------------------------- 2462// Merge memory from one path into the current memory state. 2463void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2464 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2465 Node* old_slice = mms.force_memory(); 2466 Node* new_slice = mms.memory2(); 2467 if (old_slice != new_slice) { 2468 PhiNode* phi; 2469 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2470 if (mms.is_empty()) { 2471 // clone base memory Phi's inputs for this memory slice 2472 assert(old_slice == mms.base_memory(), "sanity"); 2473 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2474 _gvn.set_type(phi, Type::MEMORY); 2475 for (uint i = 1; i < phi->req(); i++) { 2476 phi->init_req(i, old_slice->in(i)); 2477 } 2478 } else { 2479 phi = old_slice->as_Phi(); // Phi was generated already 2480 } 2481 } else { 2482 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2483 _gvn.set_type(phi, Type::MEMORY); 2484 } 2485 phi->set_req(new_path, new_slice); 2486 mms.set_memory(phi); 2487 } 2488 } 2489} 2490 2491//------------------------------make_slow_call_ex------------------------------ 2492// Make the exception handler hookups for the slow call 2493void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2494 if (stopped()) return; 2495 2496 // Make a catch node with just two handlers: fall-through and catch-all 2497 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2498 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2499 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2500 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2501 2502 { PreserveJVMState pjvms(this); 2503 set_control(excp); 2504 set_i_o(i_o); 2505 2506 if (excp != top()) { 2507 if (deoptimize) { 2508 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2509 uncommon_trap(Deoptimization::Reason_unhandled, 2510 Deoptimization::Action_none); 2511 } else { 2512 // Create an exception state also. 2513 // Use an exact type if the caller has specified a specific exception. 2514 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2515 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2516 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2517 } 2518 } 2519 } 2520 2521 // Get the no-exception control from the CatchNode. 2522 set_control(norm); 2523} 2524 2525static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2526 Node* cmp = NULL; 2527 switch(bt) { 2528 case T_INT: cmp = new CmpINode(in1, in2); break; 2529 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2530 default: fatal(err_msg("unexpected comparison type %s", type2name(bt))); 2531 } 2532 gvn->transform(cmp); 2533 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2534 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2535 gvn->transform(iff); 2536 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2537 return iff; 2538} 2539 2540 2541//-------------------------------gen_subtype_check----------------------------- 2542// Generate a subtyping check. Takes as input the subtype and supertype. 2543// Returns 2 values: sets the default control() to the true path and returns 2544// the false path. Only reads invariant memory; sets no (visible) memory. 2545// The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2546// but that's not exposed to the optimizer. This call also doesn't take in an 2547// Object; if you wish to check an Object you need to load the Object's class 2548// prior to coming here. 2549Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2550 Compile* C = gvn->C; 2551 // Fast check for identical types, perhaps identical constants. 2552 // The types can even be identical non-constants, in cases 2553 // involving Array.newInstance, Object.clone, etc. 2554 if (subklass == superklass) 2555 return C->top(); // false path is dead; no test needed. 2556 2557 if (gvn->type(superklass)->singleton()) { 2558 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2559 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2560 2561 // In the common case of an exact superklass, try to fold up the 2562 // test before generating code. You may ask, why not just generate 2563 // the code and then let it fold up? The answer is that the generated 2564 // code will necessarily include null checks, which do not always 2565 // completely fold away. If they are also needless, then they turn 2566 // into a performance loss. Example: 2567 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2568 // Here, the type of 'fa' is often exact, so the store check 2569 // of fa[1]=x will fold up, without testing the nullness of x. 2570 switch (C->static_subtype_check(superk, subk)) { 2571 case Compile::SSC_always_false: 2572 { 2573 Node* always_fail = *ctrl; 2574 *ctrl = gvn->C->top(); 2575 return always_fail; 2576 } 2577 case Compile::SSC_always_true: 2578 return C->top(); 2579 case Compile::SSC_easy_test: 2580 { 2581 // Just do a direct pointer compare and be done. 2582 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2583 *ctrl = gvn->transform(new IfTrueNode(iff)); 2584 return gvn->transform(new IfFalseNode(iff)); 2585 } 2586 case Compile::SSC_full_test: 2587 break; 2588 default: 2589 ShouldNotReachHere(); 2590 } 2591 } 2592 2593 // %%% Possible further optimization: Even if the superklass is not exact, 2594 // if the subklass is the unique subtype of the superklass, the check 2595 // will always succeed. We could leave a dependency behind to ensure this. 2596 2597 // First load the super-klass's check-offset 2598 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2599 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2600 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2601 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2602 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2603 2604 // Load from the sub-klass's super-class display list, or a 1-word cache of 2605 // the secondary superclass list, or a failing value with a sentinel offset 2606 // if the super-klass is an interface or exceptionally deep in the Java 2607 // hierarchy and we have to scan the secondary superclass list the hard way. 2608 // Worst-case type is a little odd: NULL is allowed as a result (usually 2609 // klass loads can never produce a NULL). 2610 Node *chk_off_X = chk_off; 2611#ifdef _LP64 2612 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2613#endif 2614 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2615 // For some types like interfaces the following loadKlass is from a 1-word 2616 // cache which is mutable so can't use immutable memory. Other 2617 // types load from the super-class display table which is immutable. 2618 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2619 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2620 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2621 2622 // Compile speed common case: ARE a subtype and we canNOT fail 2623 if( superklass == nkls ) 2624 return C->top(); // false path is dead; no test needed. 2625 2626 // See if we get an immediate positive hit. Happens roughly 83% of the 2627 // time. Test to see if the value loaded just previously from the subklass 2628 // is exactly the superklass. 2629 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2630 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2631 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2632 2633 // Compile speed common case: Check for being deterministic right now. If 2634 // chk_off is a constant and not equal to cacheoff then we are NOT a 2635 // subklass. In this case we need exactly the 1 test above and we can 2636 // return those results immediately. 2637 if (!might_be_cache) { 2638 Node* not_subtype_ctrl = *ctrl; 2639 *ctrl = iftrue1; // We need exactly the 1 test above 2640 return not_subtype_ctrl; 2641 } 2642 2643 // Gather the various success & failures here 2644 RegionNode *r_ok_subtype = new RegionNode(4); 2645 gvn->record_for_igvn(r_ok_subtype); 2646 RegionNode *r_not_subtype = new RegionNode(3); 2647 gvn->record_for_igvn(r_not_subtype); 2648 2649 r_ok_subtype->init_req(1, iftrue1); 2650 2651 // Check for immediate negative hit. Happens roughly 11% of the time (which 2652 // is roughly 63% of the remaining cases). Test to see if the loaded 2653 // check-offset points into the subklass display list or the 1-element 2654 // cache. If it points to the display (and NOT the cache) and the display 2655 // missed then it's not a subtype. 2656 Node *cacheoff = gvn->intcon(cacheoff_con); 2657 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2658 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2659 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2660 2661 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2662 // No performance impact (too rare) but allows sharing of secondary arrays 2663 // which has some footprint reduction. 2664 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2665 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2666 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2667 2668 // -- Roads not taken here: -- 2669 // We could also have chosen to perform the self-check at the beginning 2670 // of this code sequence, as the assembler does. This would not pay off 2671 // the same way, since the optimizer, unlike the assembler, can perform 2672 // static type analysis to fold away many successful self-checks. 2673 // Non-foldable self checks work better here in second position, because 2674 // the initial primary superclass check subsumes a self-check for most 2675 // types. An exception would be a secondary type like array-of-interface, 2676 // which does not appear in its own primary supertype display. 2677 // Finally, we could have chosen to move the self-check into the 2678 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2679 // dependent manner. But it is worthwhile to have the check here, 2680 // where it can be perhaps be optimized. The cost in code space is 2681 // small (register compare, branch). 2682 2683 // Now do a linear scan of the secondary super-klass array. Again, no real 2684 // performance impact (too rare) but it's gotta be done. 2685 // Since the code is rarely used, there is no penalty for moving it 2686 // out of line, and it can only improve I-cache density. 2687 // The decision to inline or out-of-line this final check is platform 2688 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2689 Node* psc = gvn->transform( 2690 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2691 2692 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2693 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2694 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2695 2696 // Return false path; set default control to true path. 2697 *ctrl = gvn->transform(r_ok_subtype); 2698 return gvn->transform(r_not_subtype); 2699} 2700 2701// Profile-driven exact type check: 2702Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2703 float prob, 2704 Node* *casted_receiver) { 2705 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2706 Node* recv_klass = load_object_klass(receiver); 2707 Node* want_klass = makecon(tklass); 2708 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2709 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2710 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2711 set_control( _gvn.transform( new IfTrueNode (iff) )); 2712 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2713 2714 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2715 assert(recv_xtype->klass_is_exact(), ""); 2716 2717 // Subsume downstream occurrences of receiver with a cast to 2718 // recv_xtype, since now we know what the type will be. 2719 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2720 (*casted_receiver) = _gvn.transform(cast); 2721 // (User must make the replace_in_map call.) 2722 2723 return fail; 2724} 2725 2726 2727//------------------------------seems_never_null------------------------------- 2728// Use null_seen information if it is available from the profile. 2729// If we see an unexpected null at a type check we record it and force a 2730// recompile; the offending check will be recompiled to handle NULLs. 2731// If we see several offending BCIs, then all checks in the 2732// method will be recompiled. 2733bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2734 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2735 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2736 if (UncommonNullCast // Cutout for this technique 2737 && obj != null() // And not the -Xcomp stupid case? 2738 && !too_many_traps(reason) 2739 ) { 2740 if (speculating) { 2741 return true; 2742 } 2743 if (data == NULL) 2744 // Edge case: no mature data. Be optimistic here. 2745 return true; 2746 // If the profile has not seen a null, assume it won't happen. 2747 assert(java_bc() == Bytecodes::_checkcast || 2748 java_bc() == Bytecodes::_instanceof || 2749 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2750 return !data->as_BitData()->null_seen(); 2751 } 2752 speculating = false; 2753 return false; 2754} 2755 2756//------------------------maybe_cast_profiled_receiver------------------------- 2757// If the profile has seen exactly one type, narrow to exactly that type. 2758// Subsequent type checks will always fold up. 2759Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2760 ciKlass* require_klass, 2761 ciKlass* spec_klass, 2762 bool safe_for_replace) { 2763 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2764 2765 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2766 2767 // Make sure we haven't already deoptimized from this tactic. 2768 if (too_many_traps(reason) || too_many_recompiles(reason)) 2769 return NULL; 2770 2771 // (No, this isn't a call, but it's enough like a virtual call 2772 // to use the same ciMethod accessor to get the profile info...) 2773 // If we have a speculative type use it instead of profiling (which 2774 // may not help us) 2775 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2776 if (exact_kls != NULL) {// no cast failures here 2777 if (require_klass == NULL || 2778 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2779 // If we narrow the type to match what the type profile sees or 2780 // the speculative type, we can then remove the rest of the 2781 // cast. 2782 // This is a win, even if the exact_kls is very specific, 2783 // because downstream operations, such as method calls, 2784 // will often benefit from the sharper type. 2785 Node* exact_obj = not_null_obj; // will get updated in place... 2786 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2787 &exact_obj); 2788 { PreserveJVMState pjvms(this); 2789 set_control(slow_ctl); 2790 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2791 } 2792 if (safe_for_replace) { 2793 replace_in_map(not_null_obj, exact_obj); 2794 } 2795 return exact_obj; 2796 } 2797 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2798 } 2799 2800 return NULL; 2801} 2802 2803/** 2804 * Cast obj to type and emit guard unless we had too many traps here 2805 * already 2806 * 2807 * @param obj node being casted 2808 * @param type type to cast the node to 2809 * @param not_null true if we know node cannot be null 2810 */ 2811Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2812 ciKlass* type, 2813 bool not_null, 2814 SafePointNode* sfpt) { 2815 // type == NULL if profiling tells us this object is always null 2816 if (type != NULL) { 2817 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2818 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2819 ciMethod* trap_method = (sfpt == NULL) ? method() : sfpt->jvms()->method(); 2820 int trap_bci = (sfpt == NULL) ? bci() : sfpt->jvms()->bci(); 2821 2822 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2823 !C->too_many_traps(trap_method, trap_bci, class_reason) && 2824 !C->too_many_recompiles(trap_method, trap_bci, class_reason)) { 2825 Node* not_null_obj = NULL; 2826 // not_null is true if we know the object is not null and 2827 // there's no need for a null check 2828 if (!not_null) { 2829 Node* null_ctl = top(); 2830 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2831 assert(null_ctl->is_top(), "no null control here"); 2832 } else { 2833 not_null_obj = obj; 2834 } 2835 2836 Node* exact_obj = not_null_obj; 2837 ciKlass* exact_kls = type; 2838 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2839 &exact_obj); 2840 if (sfpt != NULL) { 2841 GraphKit kit(sfpt->jvms()); 2842 PreserveJVMState pjvms(&kit); 2843 kit.set_control(slow_ctl); 2844 kit.uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2845 } else { 2846 PreserveJVMState pjvms(this); 2847 set_control(slow_ctl); 2848 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2849 } 2850 replace_in_map(not_null_obj, exact_obj); 2851 obj = exact_obj; 2852 } 2853 } else { 2854 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2855 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2856 Node* exact_obj = null_assert(obj); 2857 replace_in_map(obj, exact_obj); 2858 obj = exact_obj; 2859 } 2860 } 2861 return obj; 2862} 2863 2864//-------------------------------gen_instanceof-------------------------------- 2865// Generate an instance-of idiom. Used by both the instance-of bytecode 2866// and the reflective instance-of call. 2867Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2868 kill_dead_locals(); // Benefit all the uncommon traps 2869 assert( !stopped(), "dead parse path should be checked in callers" ); 2870 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2871 "must check for not-null not-dead klass in callers"); 2872 2873 // Make the merge point 2874 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2875 RegionNode* region = new RegionNode(PATH_LIMIT); 2876 Node* phi = new PhiNode(region, TypeInt::BOOL); 2877 C->set_has_split_ifs(true); // Has chance for split-if optimization 2878 2879 ciProfileData* data = NULL; 2880 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2881 data = method()->method_data()->bci_to_data(bci()); 2882 } 2883 bool speculative_not_null = false; 2884 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2885 && seems_never_null(obj, data, speculative_not_null)); 2886 2887 // Null check; get casted pointer; set region slot 3 2888 Node* null_ctl = top(); 2889 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2890 2891 // If not_null_obj is dead, only null-path is taken 2892 if (stopped()) { // Doing instance-of on a NULL? 2893 set_control(null_ctl); 2894 return intcon(0); 2895 } 2896 region->init_req(_null_path, null_ctl); 2897 phi ->init_req(_null_path, intcon(0)); // Set null path value 2898 if (null_ctl == top()) { 2899 // Do this eagerly, so that pattern matches like is_diamond_phi 2900 // will work even during parsing. 2901 assert(_null_path == PATH_LIMIT-1, "delete last"); 2902 region->del_req(_null_path); 2903 phi ->del_req(_null_path); 2904 } 2905 2906 // Do we know the type check always succeed? 2907 bool known_statically = false; 2908 if (_gvn.type(superklass)->singleton()) { 2909 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2910 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2911 if (subk != NULL && subk->is_loaded()) { 2912 int static_res = C->static_subtype_check(superk, subk); 2913 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2914 } 2915 } 2916 2917 if (known_statically && UseTypeSpeculation) { 2918 // If we know the type check always succeeds then we don't use the 2919 // profiling data at this bytecode. Don't lose it, feed it to the 2920 // type system as a speculative type. 2921 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2922 } else { 2923 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2924 // We may not have profiling here or it may not help us. If we 2925 // have a speculative type use it to perform an exact cast. 2926 ciKlass* spec_obj_type = obj_type->speculative_type(); 2927 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2928 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2929 if (stopped()) { // Profile disagrees with this path. 2930 set_control(null_ctl); // Null is the only remaining possibility. 2931 return intcon(0); 2932 } 2933 if (cast_obj != NULL) { 2934 not_null_obj = cast_obj; 2935 } 2936 } 2937 } 2938 2939 // Load the object's klass 2940 Node* obj_klass = load_object_klass(not_null_obj); 2941 2942 // Generate the subtype check 2943 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2944 2945 // Plug in the success path to the general merge in slot 1. 2946 region->init_req(_obj_path, control()); 2947 phi ->init_req(_obj_path, intcon(1)); 2948 2949 // Plug in the failing path to the general merge in slot 2. 2950 region->init_req(_fail_path, not_subtype_ctrl); 2951 phi ->init_req(_fail_path, intcon(0)); 2952 2953 // Return final merged results 2954 set_control( _gvn.transform(region) ); 2955 record_for_igvn(region); 2956 return _gvn.transform(phi); 2957} 2958 2959//-------------------------------gen_checkcast--------------------------------- 2960// Generate a checkcast idiom. Used by both the checkcast bytecode and the 2961// array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2962// uncommon-trap paths work. Adjust stack after this call. 2963// If failure_control is supplied and not null, it is filled in with 2964// the control edge for the cast failure. Otherwise, an appropriate 2965// uncommon trap or exception is thrown. 2966Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2967 Node* *failure_control) { 2968 kill_dead_locals(); // Benefit all the uncommon traps 2969 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2970 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2971 2972 // Fast cutout: Check the case that the cast is vacuously true. 2973 // This detects the common cases where the test will short-circuit 2974 // away completely. We do this before we perform the null check, 2975 // because if the test is going to turn into zero code, we don't 2976 // want a residual null check left around. (Causes a slowdown, 2977 // for example, in some objArray manipulations, such as a[i]=a[j].) 2978 if (tk->singleton()) { 2979 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2980 if (objtp != NULL && objtp->klass() != NULL) { 2981 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 2982 case Compile::SSC_always_true: 2983 // If we know the type check always succeed then we don't use 2984 // the profiling data at this bytecode. Don't lose it, feed it 2985 // to the type system as a speculative type. 2986 return record_profiled_receiver_for_speculation(obj); 2987 case Compile::SSC_always_false: 2988 // It needs a null check because a null will *pass* the cast check. 2989 // A non-null value will always produce an exception. 2990 return null_assert(obj); 2991 } 2992 } 2993 } 2994 2995 ciProfileData* data = NULL; 2996 bool safe_for_replace = false; 2997 if (failure_control == NULL) { // use MDO in regular case only 2998 assert(java_bc() == Bytecodes::_aastore || 2999 java_bc() == Bytecodes::_checkcast, 3000 "interpreter profiles type checks only for these BCs"); 3001 data = method()->method_data()->bci_to_data(bci()); 3002 safe_for_replace = true; 3003 } 3004 3005 // Make the merge point 3006 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3007 RegionNode* region = new RegionNode(PATH_LIMIT); 3008 Node* phi = new PhiNode(region, toop); 3009 C->set_has_split_ifs(true); // Has chance for split-if optimization 3010 3011 // Use null-cast information if it is available 3012 bool speculative_not_null = false; 3013 bool never_see_null = ((failure_control == NULL) // regular case only 3014 && seems_never_null(obj, data, speculative_not_null)); 3015 3016 // Null check; get casted pointer; set region slot 3 3017 Node* null_ctl = top(); 3018 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3019 3020 // If not_null_obj is dead, only null-path is taken 3021 if (stopped()) { // Doing instance-of on a NULL? 3022 set_control(null_ctl); 3023 return null(); 3024 } 3025 region->init_req(_null_path, null_ctl); 3026 phi ->init_req(_null_path, null()); // Set null path value 3027 if (null_ctl == top()) { 3028 // Do this eagerly, so that pattern matches like is_diamond_phi 3029 // will work even during parsing. 3030 assert(_null_path == PATH_LIMIT-1, "delete last"); 3031 region->del_req(_null_path); 3032 phi ->del_req(_null_path); 3033 } 3034 3035 Node* cast_obj = NULL; 3036 if (tk->klass_is_exact()) { 3037 // The following optimization tries to statically cast the speculative type of the object 3038 // (for example obtained during profiling) to the type of the superklass and then do a 3039 // dynamic check that the type of the object is what we expect. To work correctly 3040 // for checkcast and aastore the type of superklass should be exact. 3041 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3042 // We may not have profiling here or it may not help us. If we have 3043 // a speculative type use it to perform an exact cast. 3044 ciKlass* spec_obj_type = obj_type->speculative_type(); 3045 if (spec_obj_type != NULL || 3046 (data != NULL && 3047 // Counter has never been decremented (due to cast failure). 3048 // ...This is a reasonable thing to expect. It is true of 3049 // all casts inserted by javac to implement generic types. 3050 data->as_CounterData()->count() >= 0)) { 3051 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3052 if (cast_obj != NULL) { 3053 if (failure_control != NULL) // failure is now impossible 3054 (*failure_control) = top(); 3055 // adjust the type of the phi to the exact klass: 3056 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3057 } 3058 } 3059 } 3060 3061 if (cast_obj == NULL) { 3062 // Load the object's klass 3063 Node* obj_klass = load_object_klass(not_null_obj); 3064 3065 // Generate the subtype check 3066 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3067 3068 // Plug in success path into the merge 3069 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3070 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3071 if (failure_control == NULL) { 3072 if (not_subtype_ctrl != top()) { // If failure is possible 3073 PreserveJVMState pjvms(this); 3074 set_control(not_subtype_ctrl); 3075 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3076 } 3077 } else { 3078 (*failure_control) = not_subtype_ctrl; 3079 } 3080 } 3081 3082 region->init_req(_obj_path, control()); 3083 phi ->init_req(_obj_path, cast_obj); 3084 3085 // A merge of NULL or Casted-NotNull obj 3086 Node* res = _gvn.transform(phi); 3087 3088 // Note I do NOT always 'replace_in_map(obj,result)' here. 3089 // if( tk->klass()->can_be_primary_super() ) 3090 // This means that if I successfully store an Object into an array-of-String 3091 // I 'forget' that the Object is really now known to be a String. I have to 3092 // do this because we don't have true union types for interfaces - if I store 3093 // a Baz into an array-of-Interface and then tell the optimizer it's an 3094 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3095 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3096 // replace_in_map( obj, res ); 3097 3098 // Return final merged results 3099 set_control( _gvn.transform(region) ); 3100 record_for_igvn(region); 3101 return res; 3102} 3103 3104//------------------------------next_monitor----------------------------------- 3105// What number should be given to the next monitor? 3106int GraphKit::next_monitor() { 3107 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3108 int next = current + C->sync_stack_slots(); 3109 // Keep the toplevel high water mark current: 3110 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3111 return current; 3112} 3113 3114//------------------------------insert_mem_bar--------------------------------- 3115// Memory barrier to avoid floating things around 3116// The membar serves as a pinch point between both control and all memory slices. 3117Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3118 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3119 mb->init_req(TypeFunc::Control, control()); 3120 mb->init_req(TypeFunc::Memory, reset_memory()); 3121 Node* membar = _gvn.transform(mb); 3122 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3123 set_all_memory_call(membar); 3124 return membar; 3125} 3126 3127//-------------------------insert_mem_bar_volatile---------------------------- 3128// Memory barrier to avoid floating things around 3129// The membar serves as a pinch point between both control and memory(alias_idx). 3130// If you want to make a pinch point on all memory slices, do not use this 3131// function (even with AliasIdxBot); use insert_mem_bar() instead. 3132Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3133 // When Parse::do_put_xxx updates a volatile field, it appends a series 3134 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3135 // The first membar is on the same memory slice as the field store opcode. 3136 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3137 // All the other membars (for other volatile slices, including AliasIdxBot, 3138 // which stands for all unknown volatile slices) are control-dependent 3139 // on the first membar. This prevents later volatile loads or stores 3140 // from sliding up past the just-emitted store. 3141 3142 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3143 mb->set_req(TypeFunc::Control,control()); 3144 if (alias_idx == Compile::AliasIdxBot) { 3145 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3146 } else { 3147 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3148 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3149 } 3150 Node* membar = _gvn.transform(mb); 3151 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3152 if (alias_idx == Compile::AliasIdxBot) { 3153 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3154 } else { 3155 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3156 } 3157 return membar; 3158} 3159 3160//------------------------------shared_lock------------------------------------ 3161// Emit locking code. 3162FastLockNode* GraphKit::shared_lock(Node* obj) { 3163 // bci is either a monitorenter bc or InvocationEntryBci 3164 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3165 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3166 3167 if( !GenerateSynchronizationCode ) 3168 return NULL; // Not locking things? 3169 if (stopped()) // Dead monitor? 3170 return NULL; 3171 3172 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3173 3174 // Box the stack location 3175 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3176 Node* mem = reset_memory(); 3177 3178 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3179 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3180 // Create the counters for this fast lock. 3181 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3182 } 3183 3184 // Create the rtm counters for this fast lock if needed. 3185 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3186 3187 // Add monitor to debug info for the slow path. If we block inside the 3188 // slow path and de-opt, we need the monitor hanging around 3189 map()->push_monitor( flock ); 3190 3191 const TypeFunc *tf = LockNode::lock_type(); 3192 LockNode *lock = new LockNode(C, tf); 3193 3194 lock->init_req( TypeFunc::Control, control() ); 3195 lock->init_req( TypeFunc::Memory , mem ); 3196 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3197 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3198 lock->init_req( TypeFunc::ReturnAdr, top() ); 3199 3200 lock->init_req(TypeFunc::Parms + 0, obj); 3201 lock->init_req(TypeFunc::Parms + 1, box); 3202 lock->init_req(TypeFunc::Parms + 2, flock); 3203 add_safepoint_edges(lock); 3204 3205 lock = _gvn.transform( lock )->as_Lock(); 3206 3207 // lock has no side-effects, sets few values 3208 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3209 3210 insert_mem_bar(Op_MemBarAcquireLock); 3211 3212 // Add this to the worklist so that the lock can be eliminated 3213 record_for_igvn(lock); 3214 3215#ifndef PRODUCT 3216 if (PrintLockStatistics) { 3217 // Update the counter for this lock. Don't bother using an atomic 3218 // operation since we don't require absolute accuracy. 3219 lock->create_lock_counter(map()->jvms()); 3220 increment_counter(lock->counter()->addr()); 3221 } 3222#endif 3223 3224 return flock; 3225} 3226 3227 3228//------------------------------shared_unlock---------------------------------- 3229// Emit unlocking code. 3230void GraphKit::shared_unlock(Node* box, Node* obj) { 3231 // bci is either a monitorenter bc or InvocationEntryBci 3232 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3233 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3234 3235 if( !GenerateSynchronizationCode ) 3236 return; 3237 if (stopped()) { // Dead monitor? 3238 map()->pop_monitor(); // Kill monitor from debug info 3239 return; 3240 } 3241 3242 // Memory barrier to avoid floating things down past the locked region 3243 insert_mem_bar(Op_MemBarReleaseLock); 3244 3245 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3246 UnlockNode *unlock = new UnlockNode(C, tf); 3247#ifdef ASSERT 3248 unlock->set_dbg_jvms(sync_jvms()); 3249#endif 3250 uint raw_idx = Compile::AliasIdxRaw; 3251 unlock->init_req( TypeFunc::Control, control() ); 3252 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3253 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3254 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3255 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3256 3257 unlock->init_req(TypeFunc::Parms + 0, obj); 3258 unlock->init_req(TypeFunc::Parms + 1, box); 3259 unlock = _gvn.transform(unlock)->as_Unlock(); 3260 3261 Node* mem = reset_memory(); 3262 3263 // unlock has no side-effects, sets few values 3264 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3265 3266 // Kill monitor from debug info 3267 map()->pop_monitor( ); 3268} 3269 3270//-------------------------------get_layout_helper----------------------------- 3271// If the given klass is a constant or known to be an array, 3272// fetch the constant layout helper value into constant_value 3273// and return (Node*)NULL. Otherwise, load the non-constant 3274// layout helper value, and return the node which represents it. 3275// This two-faced routine is useful because allocation sites 3276// almost always feature constant types. 3277Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3278 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3279 if (!StressReflectiveCode && inst_klass != NULL) { 3280 ciKlass* klass = inst_klass->klass(); 3281 bool xklass = inst_klass->klass_is_exact(); 3282 if (xklass || klass->is_array_klass()) { 3283 jint lhelper = klass->layout_helper(); 3284 if (lhelper != Klass::_lh_neutral_value) { 3285 constant_value = lhelper; 3286 return (Node*) NULL; 3287 } 3288 } 3289 } 3290 constant_value = Klass::_lh_neutral_value; // put in a known value 3291 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3292 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3293} 3294 3295// We just put in an allocate/initialize with a big raw-memory effect. 3296// Hook selected additional alias categories on the initialization. 3297static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3298 MergeMemNode* init_in_merge, 3299 Node* init_out_raw) { 3300 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3301 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3302 3303 Node* prevmem = kit.memory(alias_idx); 3304 init_in_merge->set_memory_at(alias_idx, prevmem); 3305 kit.set_memory(init_out_raw, alias_idx); 3306} 3307 3308//---------------------------set_output_for_allocation------------------------- 3309Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3310 const TypeOopPtr* oop_type, 3311 bool deoptimize_on_exception) { 3312 int rawidx = Compile::AliasIdxRaw; 3313 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3314 add_safepoint_edges(alloc); 3315 Node* allocx = _gvn.transform(alloc); 3316 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3317 // create memory projection for i_o 3318 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3319 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3320 3321 // create a memory projection as for the normal control path 3322 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3323 set_memory(malloc, rawidx); 3324 3325 // a normal slow-call doesn't change i_o, but an allocation does 3326 // we create a separate i_o projection for the normal control path 3327 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3328 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3329 3330 // put in an initialization barrier 3331 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3332 rawoop)->as_Initialize(); 3333 assert(alloc->initialization() == init, "2-way macro link must work"); 3334 assert(init ->allocation() == alloc, "2-way macro link must work"); 3335 { 3336 // Extract memory strands which may participate in the new object's 3337 // initialization, and source them from the new InitializeNode. 3338 // This will allow us to observe initializations when they occur, 3339 // and link them properly (as a group) to the InitializeNode. 3340 assert(init->in(InitializeNode::Memory) == malloc, ""); 3341 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3342 init->set_req(InitializeNode::Memory, minit_in); 3343 record_for_igvn(minit_in); // fold it up later, if possible 3344 Node* minit_out = memory(rawidx); 3345 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3346 if (oop_type->isa_aryptr()) { 3347 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3348 int elemidx = C->get_alias_index(telemref); 3349 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3350 } else if (oop_type->isa_instptr()) { 3351 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3352 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3353 ciField* field = ik->nonstatic_field_at(i); 3354 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3355 continue; // do not bother to track really large numbers of fields 3356 // Find (or create) the alias category for this field: 3357 int fieldidx = C->alias_type(field)->index(); 3358 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3359 } 3360 } 3361 } 3362 3363 // Cast raw oop to the real thing... 3364 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3365 javaoop = _gvn.transform(javaoop); 3366 C->set_recent_alloc(control(), javaoop); 3367 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3368 3369#ifdef ASSERT 3370 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3371 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3372 "Ideal_allocation works"); 3373 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3374 "Ideal_allocation works"); 3375 if (alloc->is_AllocateArray()) { 3376 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3377 "Ideal_allocation works"); 3378 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3379 "Ideal_allocation works"); 3380 } else { 3381 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3382 } 3383 } 3384#endif //ASSERT 3385 3386 return javaoop; 3387} 3388 3389//---------------------------new_instance-------------------------------------- 3390// This routine takes a klass_node which may be constant (for a static type) 3391// or may be non-constant (for reflective code). It will work equally well 3392// for either, and the graph will fold nicely if the optimizer later reduces 3393// the type to a constant. 3394// The optional arguments are for specialized use by intrinsics: 3395// - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3396// - If 'return_size_val', report the the total object size to the caller. 3397// - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3398Node* GraphKit::new_instance(Node* klass_node, 3399 Node* extra_slow_test, 3400 Node* *return_size_val, 3401 bool deoptimize_on_exception) { 3402 // Compute size in doublewords 3403 // The size is always an integral number of doublewords, represented 3404 // as a positive bytewise size stored in the klass's layout_helper. 3405 // The layout_helper also encodes (in a low bit) the need for a slow path. 3406 jint layout_con = Klass::_lh_neutral_value; 3407 Node* layout_val = get_layout_helper(klass_node, layout_con); 3408 int layout_is_con = (layout_val == NULL); 3409 3410 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3411 // Generate the initial go-slow test. It's either ALWAYS (return a 3412 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3413 // case) a computed value derived from the layout_helper. 3414 Node* initial_slow_test = NULL; 3415 if (layout_is_con) { 3416 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3417 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3418 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; 3419 3420 } else { // reflective case 3421 // This reflective path is used by Unsafe.allocateInstance. 3422 // (It may be stress-tested by specifying StressReflectiveCode.) 3423 // Basically, we want to get into the VM is there's an illegal argument. 3424 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3425 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3426 if (extra_slow_test != intcon(0)) { 3427 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3428 } 3429 // (Macro-expander will further convert this to a Bool, if necessary.) 3430 } 3431 3432 // Find the size in bytes. This is easy; it's the layout_helper. 3433 // The size value must be valid even if the slow path is taken. 3434 Node* size = NULL; 3435 if (layout_is_con) { 3436 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3437 } else { // reflective case 3438 // This reflective path is used by clone and Unsafe.allocateInstance. 3439 size = ConvI2X(layout_val); 3440 3441 // Clear the low bits to extract layout_helper_size_in_bytes: 3442 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3443 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3444 size = _gvn.transform( new AndXNode(size, mask) ); 3445 } 3446 if (return_size_val != NULL) { 3447 (*return_size_val) = size; 3448 } 3449 3450 // This is a precise notnull oop of the klass. 3451 // (Actually, it need not be precise if this is a reflective allocation.) 3452 // It's what we cast the result to. 3453 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3454 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3455 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3456 3457 // Now generate allocation code 3458 3459 // The entire memory state is needed for slow path of the allocation 3460 // since GC and deoptimization can happened. 3461 Node *mem = reset_memory(); 3462 set_all_memory(mem); // Create new memory state 3463 3464 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3465 control(), mem, i_o(), 3466 size, klass_node, 3467 initial_slow_test); 3468 3469 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3470} 3471 3472//-------------------------------new_array------------------------------------- 3473// helper for both newarray and anewarray 3474// The 'length' parameter is (obviously) the length of the array. 3475// See comments on new_instance for the meaning of the other arguments. 3476Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3477 Node* length, // number of array elements 3478 int nargs, // number of arguments to push back for uncommon trap 3479 Node* *return_size_val, 3480 bool deoptimize_on_exception) { 3481 jint layout_con = Klass::_lh_neutral_value; 3482 Node* layout_val = get_layout_helper(klass_node, layout_con); 3483 int layout_is_con = (layout_val == NULL); 3484 3485 if (!layout_is_con && !StressReflectiveCode && 3486 !too_many_traps(Deoptimization::Reason_class_check)) { 3487 // This is a reflective array creation site. 3488 // Optimistically assume that it is a subtype of Object[], 3489 // so that we can fold up all the address arithmetic. 3490 layout_con = Klass::array_layout_helper(T_OBJECT); 3491 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3492 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3493 { BuildCutout unless(this, bol_lh, PROB_MAX); 3494 inc_sp(nargs); 3495 uncommon_trap(Deoptimization::Reason_class_check, 3496 Deoptimization::Action_maybe_recompile); 3497 } 3498 layout_val = NULL; 3499 layout_is_con = true; 3500 } 3501 3502 // Generate the initial go-slow test. Make sure we do not overflow 3503 // if length is huge (near 2Gig) or negative! We do not need 3504 // exact double-words here, just a close approximation of needed 3505 // double-words. We can't add any offset or rounding bits, lest we 3506 // take a size -1 of bytes and make it positive. Use an unsigned 3507 // compare, so negative sizes look hugely positive. 3508 int fast_size_limit = FastAllocateSizeLimit; 3509 if (layout_is_con) { 3510 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3511 // Increase the size limit if we have exact knowledge of array type. 3512 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3513 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3514 } 3515 3516 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3517 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3518 if (initial_slow_test->is_Bool()) { 3519 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3520 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3521 } 3522 3523 // --- Size Computation --- 3524 // array_size = round_to_heap(array_header + (length << elem_shift)); 3525 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3526 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3527 // The rounding mask is strength-reduced, if possible. 3528 int round_mask = MinObjAlignmentInBytes - 1; 3529 Node* header_size = NULL; 3530 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3531 // (T_BYTE has the weakest alignment and size restrictions...) 3532 if (layout_is_con) { 3533 int hsize = Klass::layout_helper_header_size(layout_con); 3534 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3535 BasicType etype = Klass::layout_helper_element_type(layout_con); 3536 if ((round_mask & ~right_n_bits(eshift)) == 0) 3537 round_mask = 0; // strength-reduce it if it goes away completely 3538 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3539 assert(header_size_min <= hsize, "generic minimum is smallest"); 3540 header_size_min = hsize; 3541 header_size = intcon(hsize + round_mask); 3542 } else { 3543 Node* hss = intcon(Klass::_lh_header_size_shift); 3544 Node* hsm = intcon(Klass::_lh_header_size_mask); 3545 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3546 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3547 Node* mask = intcon(round_mask); 3548 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3549 } 3550 3551 Node* elem_shift = NULL; 3552 if (layout_is_con) { 3553 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3554 if (eshift != 0) 3555 elem_shift = intcon(eshift); 3556 } else { 3557 // There is no need to mask or shift this value. 3558 // The semantics of LShiftINode include an implicit mask to 0x1F. 3559 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3560 elem_shift = layout_val; 3561 } 3562 3563 // Transition to native address size for all offset calculations: 3564 Node* lengthx = ConvI2X(length); 3565 Node* headerx = ConvI2X(header_size); 3566#ifdef _LP64 3567 { const TypeLong* tllen = _gvn.find_long_type(lengthx); 3568 if (tllen != NULL && tllen->_lo < 0) { 3569 // Add a manual constraint to a positive range. Cf. array_element_address. 3570 jlong size_max = arrayOopDesc::max_array_length(T_BYTE); 3571 if (size_max > tllen->_hi) size_max = tllen->_hi; 3572 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); 3573 lengthx = _gvn.transform( new ConvI2LNode(length, tlcon)); 3574 } 3575 } 3576#endif 3577 3578 // Combine header size (plus rounding) and body size. Then round down. 3579 // This computation cannot overflow, because it is used only in two 3580 // places, one where the length is sharply limited, and the other 3581 // after a successful allocation. 3582 Node* abody = lengthx; 3583 if (elem_shift != NULL) 3584 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3585 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3586 if (round_mask != 0) { 3587 Node* mask = MakeConX(~round_mask); 3588 size = _gvn.transform( new AndXNode(size, mask) ); 3589 } 3590 // else if round_mask == 0, the size computation is self-rounding 3591 3592 if (return_size_val != NULL) { 3593 // This is the size 3594 (*return_size_val) = size; 3595 } 3596 3597 // Now generate allocation code 3598 3599 // The entire memory state is needed for slow path of the allocation 3600 // since GC and deoptimization can happened. 3601 Node *mem = reset_memory(); 3602 set_all_memory(mem); // Create new memory state 3603 3604 // Create the AllocateArrayNode and its result projections 3605 AllocateArrayNode* alloc 3606 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3607 control(), mem, i_o(), 3608 size, klass_node, 3609 initial_slow_test, 3610 length); 3611 3612 // Cast to correct type. Note that the klass_node may be constant or not, 3613 // and in the latter case the actual array type will be inexact also. 3614 // (This happens via a non-constant argument to inline_native_newArray.) 3615 // In any case, the value of klass_node provides the desired array type. 3616 const TypeInt* length_type = _gvn.find_int_type(length); 3617 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3618 if (ary_type->isa_aryptr() && length_type != NULL) { 3619 // Try to get a better type than POS for the size 3620 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3621 } 3622 3623 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3624 3625 // Cast length on remaining path to be as narrow as possible 3626 if (map()->find_edge(length) >= 0) { 3627 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3628 if (ccast != length) { 3629 _gvn.set_type_bottom(ccast); 3630 record_for_igvn(ccast); 3631 replace_in_map(length, ccast); 3632 } 3633 } 3634 3635 return javaoop; 3636} 3637 3638// The following "Ideal_foo" functions are placed here because they recognize 3639// the graph shapes created by the functions immediately above. 3640 3641//---------------------------Ideal_allocation---------------------------------- 3642// Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3643AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3644 if (ptr == NULL) { // reduce dumb test in callers 3645 return NULL; 3646 } 3647 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3648 ptr = ptr->in(1); 3649 if (ptr == NULL) return NULL; 3650 } 3651 // Return NULL for allocations with several casts: 3652 // j.l.reflect.Array.newInstance(jobject, jint) 3653 // Object.clone() 3654 // to keep more precise type from last cast. 3655 if (ptr->is_Proj()) { 3656 Node* allo = ptr->in(0); 3657 if (allo != NULL && allo->is_Allocate()) { 3658 return allo->as_Allocate(); 3659 } 3660 } 3661 // Report failure to match. 3662 return NULL; 3663} 3664 3665// Fancy version which also strips off an offset (and reports it to caller). 3666AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3667 intptr_t& offset) { 3668 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3669 if (base == NULL) return NULL; 3670 return Ideal_allocation(base, phase); 3671} 3672 3673// Trace Initialize <- Proj[Parm] <- Allocate 3674AllocateNode* InitializeNode::allocation() { 3675 Node* rawoop = in(InitializeNode::RawAddress); 3676 if (rawoop->is_Proj()) { 3677 Node* alloc = rawoop->in(0); 3678 if (alloc->is_Allocate()) { 3679 return alloc->as_Allocate(); 3680 } 3681 } 3682 return NULL; 3683} 3684 3685// Trace Allocate -> Proj[Parm] -> Initialize 3686InitializeNode* AllocateNode::initialization() { 3687 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3688 if (rawoop == NULL) return NULL; 3689 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3690 Node* init = rawoop->fast_out(i); 3691 if (init->is_Initialize()) { 3692 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3693 return init->as_Initialize(); 3694 } 3695 } 3696 return NULL; 3697} 3698 3699//----------------------------- loop predicates --------------------------- 3700 3701//------------------------------add_predicate_impl---------------------------- 3702void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3703 // Too many traps seen? 3704 if (too_many_traps(reason)) { 3705#ifdef ASSERT 3706 if (TraceLoopPredicate) { 3707 int tc = C->trap_count(reason); 3708 tty->print("too many traps=%s tcount=%d in ", 3709 Deoptimization::trap_reason_name(reason), tc); 3710 method()->print(); // which method has too many predicate traps 3711 tty->cr(); 3712 } 3713#endif 3714 // We cannot afford to take more traps here, 3715 // do not generate predicate. 3716 return; 3717 } 3718 3719 Node *cont = _gvn.intcon(1); 3720 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3721 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3722 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3723 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3724 C->add_predicate_opaq(opq); 3725 { 3726 PreserveJVMState pjvms(this); 3727 set_control(iffalse); 3728 inc_sp(nargs); 3729 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3730 } 3731 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3732 set_control(iftrue); 3733} 3734 3735//------------------------------add_predicate--------------------------------- 3736void GraphKit::add_predicate(int nargs) { 3737 if (UseLoopPredicate) { 3738 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3739 } 3740 // loop's limit check predicate should be near the loop. 3741 if (LoopLimitCheck) { 3742 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3743 } 3744} 3745 3746//----------------------------- store barriers ---------------------------- 3747#define __ ideal. 3748 3749void GraphKit::sync_kit(IdealKit& ideal) { 3750 set_all_memory(__ merged_memory()); 3751 set_i_o(__ i_o()); 3752 set_control(__ ctrl()); 3753} 3754 3755void GraphKit::final_sync(IdealKit& ideal) { 3756 // Final sync IdealKit and graphKit. 3757 sync_kit(ideal); 3758} 3759 3760Node* GraphKit::byte_map_base_node() { 3761 // Get base of card map 3762 CardTableModRefBS* ct = (CardTableModRefBS*)(Universe::heap()->barrier_set()); 3763 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code"); 3764 if (ct->byte_map_base != NULL) { 3765 return makecon(TypeRawPtr::make((address)ct->byte_map_base)); 3766 } else { 3767 return null(); 3768 } 3769} 3770 3771// vanilla/CMS post barrier 3772// Insert a write-barrier store. This is to let generational GC work; we have 3773// to flag all oop-stores before the next GC point. 3774void GraphKit::write_barrier_post(Node* oop_store, 3775 Node* obj, 3776 Node* adr, 3777 uint adr_idx, 3778 Node* val, 3779 bool use_precise) { 3780 // No store check needed if we're storing a NULL or an old object 3781 // (latter case is probably a string constant). The concurrent 3782 // mark sweep garbage collector, however, needs to have all nonNull 3783 // oop updates flagged via card-marks. 3784 if (val != NULL && val->is_Con()) { 3785 // must be either an oop or NULL 3786 const Type* t = val->bottom_type(); 3787 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3788 // stores of null never (?) need barriers 3789 return; 3790 } 3791 3792 if (use_ReduceInitialCardMarks() 3793 && obj == just_allocated_object(control())) { 3794 // We can skip marks on a freshly-allocated object in Eden. 3795 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3796 // That routine informs GC to take appropriate compensating steps, 3797 // upon a slow-path allocation, so as to make this card-mark 3798 // elision safe. 3799 return; 3800 } 3801 3802 if (!use_precise) { 3803 // All card marks for a (non-array) instance are in one place: 3804 adr = obj; 3805 } 3806 // (Else it's an array (or unknown), and we want more precise card marks.) 3807 assert(adr != NULL, ""); 3808 3809 IdealKit ideal(this, true); 3810 3811 // Convert the pointer to an int prior to doing math on it 3812 Node* cast = __ CastPX(__ ctrl(), adr); 3813 3814 // Divide by card size 3815 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, 3816 "Only one we handle so far."); 3817 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3818 3819 // Combine card table base and card offset 3820 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3821 3822 // Get the alias_index for raw card-mark memory 3823 int adr_type = Compile::AliasIdxRaw; 3824 Node* zero = __ ConI(0); // Dirty card value 3825 BasicType bt = T_BYTE; 3826 3827 if (UseCondCardMark) { 3828 // The classic GC reference write barrier is typically implemented 3829 // as a store into the global card mark table. Unfortunately 3830 // unconditional stores can result in false sharing and excessive 3831 // coherence traffic as well as false transactional aborts. 3832 // UseCondCardMark enables MP "polite" conditional card mark 3833 // stores. In theory we could relax the load from ctrl() to 3834 // no_ctrl, but that doesn't buy much latitude. 3835 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3836 __ if_then(card_val, BoolTest::ne, zero); 3837 } 3838 3839 // Smash zero into card 3840 if( !UseConcMarkSweepGC ) { 3841 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 3842 } else { 3843 // Specialized path for CM store barrier 3844 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3845 } 3846 3847 if (UseCondCardMark) { 3848 __ end_if(); 3849 } 3850 3851 // Final sync IdealKit and GraphKit. 3852 final_sync(ideal); 3853} 3854/* 3855 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 3856 * required by SATB to make sure all objects live at the start of the 3857 * marking are kept alive, all reference updates need to any previous 3858 * reference stored before writing. 3859 * 3860 * If the previous value is NULL there is no need to save the old value. 3861 * References that are NULL are filtered during runtime by the barrier 3862 * code to avoid unnecessary queuing. 3863 * 3864 * However in the case of newly allocated objects it might be possible to 3865 * prove that the reference about to be overwritten is NULL during compile 3866 * time and avoid adding the barrier code completely. 3867 * 3868 * The compiler needs to determine that the object in which a field is about 3869 * to be written is newly allocated, and that no prior store to the same field 3870 * has happened since the allocation. 3871 * 3872 * Returns true if the pre-barrier can be removed 3873 */ 3874bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 3875 BasicType bt, uint adr_idx) { 3876 intptr_t offset = 0; 3877 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 3878 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 3879 3880 if (offset == Type::OffsetBot) { 3881 return false; // cannot unalias unless there are precise offsets 3882 } 3883 3884 if (alloc == NULL) { 3885 return false; // No allocation found 3886 } 3887 3888 intptr_t size_in_bytes = type2aelembytes(bt); 3889 3890 Node* mem = memory(adr_idx); // start searching here... 3891 3892 for (int cnt = 0; cnt < 50; cnt++) { 3893 3894 if (mem->is_Store()) { 3895 3896 Node* st_adr = mem->in(MemNode::Address); 3897 intptr_t st_offset = 0; 3898 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 3899 3900 if (st_base == NULL) { 3901 break; // inscrutable pointer 3902 } 3903 3904 // Break we have found a store with same base and offset as ours so break 3905 if (st_base == base && st_offset == offset) { 3906 break; 3907 } 3908 3909 if (st_offset != offset && st_offset != Type::OffsetBot) { 3910 const int MAX_STORE = BytesPerLong; 3911 if (st_offset >= offset + size_in_bytes || 3912 st_offset <= offset - MAX_STORE || 3913 st_offset <= offset - mem->as_Store()->memory_size()) { 3914 // Success: The offsets are provably independent. 3915 // (You may ask, why not just test st_offset != offset and be done? 3916 // The answer is that stores of different sizes can co-exist 3917 // in the same sequence of RawMem effects. We sometimes initialize 3918 // a whole 'tile' of array elements with a single jint or jlong.) 3919 mem = mem->in(MemNode::Memory); 3920 continue; // advance through independent store memory 3921 } 3922 } 3923 3924 if (st_base != base 3925 && MemNode::detect_ptr_independence(base, alloc, st_base, 3926 AllocateNode::Ideal_allocation(st_base, phase), 3927 phase)) { 3928 // Success: The bases are provably independent. 3929 mem = mem->in(MemNode::Memory); 3930 continue; // advance through independent store memory 3931 } 3932 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 3933 3934 InitializeNode* st_init = mem->in(0)->as_Initialize(); 3935 AllocateNode* st_alloc = st_init->allocation(); 3936 3937 // Make sure that we are looking at the same allocation site. 3938 // The alloc variable is guaranteed to not be null here from earlier check. 3939 if (alloc == st_alloc) { 3940 // Check that the initialization is storing NULL so that no previous store 3941 // has been moved up and directly write a reference 3942 Node* captured_store = st_init->find_captured_store(offset, 3943 type2aelembytes(T_OBJECT), 3944 phase); 3945 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 3946 return true; 3947 } 3948 } 3949 } 3950 3951 // Unless there is an explicit 'continue', we must bail out here, 3952 // because 'mem' is an inscrutable memory state (e.g., a call). 3953 break; 3954 } 3955 3956 return false; 3957} 3958 3959// G1 pre/post barriers 3960void GraphKit::g1_write_barrier_pre(bool do_load, 3961 Node* obj, 3962 Node* adr, 3963 uint alias_idx, 3964 Node* val, 3965 const TypeOopPtr* val_type, 3966 Node* pre_val, 3967 BasicType bt) { 3968 3969 // Some sanity checks 3970 // Note: val is unused in this routine. 3971 3972 if (do_load) { 3973 // We need to generate the load of the previous value 3974 assert(obj != NULL, "must have a base"); 3975 assert(adr != NULL, "where are loading from?"); 3976 assert(pre_val == NULL, "loaded already?"); 3977 assert(val_type != NULL, "need a type"); 3978 3979 if (use_ReduceInitialCardMarks() 3980 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 3981 return; 3982 } 3983 3984 } else { 3985 // In this case both val_type and alias_idx are unused. 3986 assert(pre_val != NULL, "must be loaded already"); 3987 // Nothing to be done if pre_val is null. 3988 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 3989 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 3990 } 3991 assert(bt == T_OBJECT, "or we shouldn't be here"); 3992 3993 IdealKit ideal(this, true); 3994 3995 Node* tls = __ thread(); // ThreadLocalStorage 3996 3997 Node* no_ctrl = NULL; 3998 Node* no_base = __ top(); 3999 Node* zero = __ ConI(0); 4000 Node* zeroX = __ ConX(0); 4001 4002 float likely = PROB_LIKELY(0.999); 4003 float unlikely = PROB_UNLIKELY(0.999); 4004 4005 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4006 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); 4007 4008 // Offsets into the thread 4009 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 4010 PtrQueue::byte_offset_of_active()); 4011 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4012 PtrQueue::byte_offset_of_index()); 4013 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4014 PtrQueue::byte_offset_of_buf()); 4015 4016 // Now the actual pointers into the thread 4017 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4018 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4019 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4020 4021 // Now some of the values 4022 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4023 4024 // if (!marking) 4025 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4026 BasicType index_bt = TypeX_X->basic_type(); 4027 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size."); 4028 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4029 4030 if (do_load) { 4031 // load original value 4032 // alias_idx correct?? 4033 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4034 } 4035 4036 // if (pre_val != NULL) 4037 __ if_then(pre_val, BoolTest::ne, null()); { 4038 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4039 4040 // is the queue for this thread full? 4041 __ if_then(index, BoolTest::ne, zeroX, likely); { 4042 4043 // decrement the index 4044 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4045 4046 // Now get the buffer location we will log the previous value into and store it 4047 Node *log_addr = __ AddP(no_base, buffer, next_index); 4048 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4049 // update the index 4050 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4051 4052 } __ else_(); { 4053 4054 // logging buffer is full, call the runtime 4055 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4056 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4057 } __ end_if(); // (!index) 4058 } __ end_if(); // (pre_val != NULL) 4059 } __ end_if(); // (!marking) 4060 4061 // Final sync IdealKit and GraphKit. 4062 final_sync(ideal); 4063} 4064 4065/* 4066 * G1 similar to any GC with a Young Generation requires a way to keep track of 4067 * references from Old Generation to Young Generation to make sure all live 4068 * objects are found. G1 also requires to keep track of object references 4069 * between different regions to enable evacuation of old regions, which is done 4070 * as part of mixed collections. References are tracked in remembered sets and 4071 * is continuously updated as reference are written to with the help of the 4072 * post-barrier. 4073 * 4074 * To reduce the number of updates to the remembered set the post-barrier 4075 * filters updates to fields in objects located in the Young Generation, 4076 * the same region as the reference, when the NULL is being written or 4077 * if the card is already marked as dirty by an earlier write. 4078 * 4079 * Under certain circumstances it is possible to avoid generating the 4080 * post-barrier completely if it is possible during compile time to prove 4081 * the object is newly allocated and that no safepoint exists between the 4082 * allocation and the store. 4083 * 4084 * In the case of slow allocation the allocation code must handle the barrier 4085 * as part of the allocation in the case the allocated object is not located 4086 * in the nursery, this would happen for humongous objects. This is similar to 4087 * how CMS is required to handle this case, see the comments for the method 4088 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier. 4089 * A deferred card mark is required for these objects and handled in the above 4090 * mentioned methods. 4091 * 4092 * Returns true if the post barrier can be removed 4093 */ 4094bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4095 Node* adr) { 4096 intptr_t offset = 0; 4097 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4098 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4099 4100 if (offset == Type::OffsetBot) { 4101 return false; // cannot unalias unless there are precise offsets 4102 } 4103 4104 if (alloc == NULL) { 4105 return false; // No allocation found 4106 } 4107 4108 // Start search from Store node 4109 Node* mem = store->in(MemNode::Control); 4110 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4111 4112 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4113 AllocateNode* st_alloc = st_init->allocation(); 4114 4115 // Make sure we are looking at the same allocation 4116 if (alloc == st_alloc) { 4117 return true; 4118 } 4119 } 4120 4121 return false; 4122} 4123 4124// 4125// Update the card table and add card address to the queue 4126// 4127void GraphKit::g1_mark_card(IdealKit& ideal, 4128 Node* card_adr, 4129 Node* oop_store, 4130 uint oop_alias_idx, 4131 Node* index, 4132 Node* index_adr, 4133 Node* buffer, 4134 const TypeFunc* tf) { 4135 4136 Node* zero = __ ConI(0); 4137 Node* zeroX = __ ConX(0); 4138 Node* no_base = __ top(); 4139 BasicType card_bt = T_BYTE; 4140 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4141 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4142 4143 // Now do the queue work 4144 __ if_then(index, BoolTest::ne, zeroX); { 4145 4146 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4147 Node* log_addr = __ AddP(no_base, buffer, next_index); 4148 4149 // Order, see storeCM. 4150 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4151 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4152 4153 } __ else_(); { 4154 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4155 } __ end_if(); 4156 4157} 4158 4159void GraphKit::g1_write_barrier_post(Node* oop_store, 4160 Node* obj, 4161 Node* adr, 4162 uint alias_idx, 4163 Node* val, 4164 BasicType bt, 4165 bool use_precise) { 4166 // If we are writing a NULL then we need no post barrier 4167 4168 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4169 // Must be NULL 4170 const Type* t = val->bottom_type(); 4171 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4172 // No post barrier if writing NULLx 4173 return; 4174 } 4175 4176 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4177 // We can skip marks on a freshly-allocated object in Eden. 4178 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 4179 // That routine informs GC to take appropriate compensating steps, 4180 // upon a slow-path allocation, so as to make this card-mark 4181 // elision safe. 4182 return; 4183 } 4184 4185 if (use_ReduceInitialCardMarks() 4186 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4187 return; 4188 } 4189 4190 if (!use_precise) { 4191 // All card marks for a (non-array) instance are in one place: 4192 adr = obj; 4193 } 4194 // (Else it's an array (or unknown), and we want more precise card marks.) 4195 assert(adr != NULL, ""); 4196 4197 IdealKit ideal(this, true); 4198 4199 Node* tls = __ thread(); // ThreadLocalStorage 4200 4201 Node* no_base = __ top(); 4202 float likely = PROB_LIKELY(0.999); 4203 float unlikely = PROB_UNLIKELY(0.999); 4204 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4205 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4206 Node* zeroX = __ ConX(0); 4207 4208 // Get the alias_index for raw card-mark memory 4209 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4210 4211 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4212 4213 // Offsets into the thread 4214 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4215 PtrQueue::byte_offset_of_index()); 4216 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4217 PtrQueue::byte_offset_of_buf()); 4218 4219 // Pointers into the thread 4220 4221 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4222 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4223 4224 // Now some values 4225 // Use ctrl to avoid hoisting these values past a safepoint, which could 4226 // potentially reset these fields in the JavaThread. 4227 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4228 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4229 4230 // Convert the store obj pointer to an int prior to doing math on it 4231 // Must use ctrl to prevent "integerized oop" existing across safepoint 4232 Node* cast = __ CastPX(__ ctrl(), adr); 4233 4234 // Divide pointer by card size 4235 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4236 4237 // Combine card table base and card offset 4238 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4239 4240 // If we know the value being stored does it cross regions? 4241 4242 if (val != NULL) { 4243 // Does the store cause us to cross regions? 4244 4245 // Should be able to do an unsigned compare of region_size instead of 4246 // and extra shift. Do we have an unsigned compare?? 4247 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4248 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4249 4250 // if (xor_res == 0) same region so skip 4251 __ if_then(xor_res, BoolTest::ne, zeroX); { 4252 4253 // No barrier if we are storing a NULL 4254 __ if_then(val, BoolTest::ne, null(), unlikely); { 4255 4256 // Ok must mark the card if not already dirty 4257 4258 // load the original value of the card 4259 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4260 4261 __ if_then(card_val, BoolTest::ne, young_card); { 4262 sync_kit(ideal); 4263 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4264 insert_mem_bar(Op_MemBarVolatile, oop_store); 4265 __ sync_kit(this); 4266 4267 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4268 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4269 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4270 } __ end_if(); 4271 } __ end_if(); 4272 } __ end_if(); 4273 } __ end_if(); 4274 } else { 4275 // Object.clone() instrinsic uses this path. 4276 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4277 } 4278 4279 // Final sync IdealKit and GraphKit. 4280 final_sync(ideal); 4281} 4282#undef __ 4283 4284 4285 4286Node* GraphKit::load_String_offset(Node* ctrl, Node* str) { 4287 if (java_lang_String::has_offset_field()) { 4288 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4289 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4290 false, NULL, 0); 4291 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4292 int offset_field_idx = C->get_alias_index(offset_field_type); 4293 return make_load(ctrl, 4294 basic_plus_adr(str, str, offset_offset), 4295 TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered); 4296 } else { 4297 return intcon(0); 4298 } 4299} 4300 4301Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4302 if (java_lang_String::has_count_field()) { 4303 int count_offset = java_lang_String::count_offset_in_bytes(); 4304 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4305 false, NULL, 0); 4306 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4307 int count_field_idx = C->get_alias_index(count_field_type); 4308 return make_load(ctrl, 4309 basic_plus_adr(str, str, count_offset), 4310 TypeInt::INT, T_INT, count_field_idx, MemNode::unordered); 4311 } else { 4312 return load_array_length(load_String_value(ctrl, str)); 4313 } 4314} 4315 4316Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4317 int value_offset = java_lang_String::value_offset_in_bytes(); 4318 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4319 false, NULL, 0); 4320 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4321 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4322 TypeAry::make(TypeInt::CHAR,TypeInt::POS), 4323 ciTypeArrayKlass::make(T_CHAR), true, 0); 4324 int value_field_idx = C->get_alias_index(value_field_type); 4325 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4326 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4327 // String.value field is known to be @Stable. 4328 if (UseImplicitStableValues) { 4329 load = cast_array_to_stable(load, value_type); 4330 } 4331 return load; 4332} 4333 4334void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) { 4335 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4336 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4337 false, NULL, 0); 4338 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4339 int offset_field_idx = C->get_alias_index(offset_field_type); 4340 store_to_memory(ctrl, basic_plus_adr(str, offset_offset), 4341 value, T_INT, offset_field_idx, MemNode::unordered); 4342} 4343 4344void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4345 int value_offset = java_lang_String::value_offset_in_bytes(); 4346 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4347 false, NULL, 0); 4348 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4349 4350 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4351 value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered); 4352} 4353 4354void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) { 4355 int count_offset = java_lang_String::count_offset_in_bytes(); 4356 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4357 false, NULL, 0); 4358 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4359 int count_field_idx = C->get_alias_index(count_field_type); 4360 store_to_memory(ctrl, basic_plus_adr(str, count_offset), 4361 value, T_INT, count_field_idx, MemNode::unordered); 4362} 4363 4364Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4365 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4366 // assumption of CCP analysis. 4367 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4368} 4369