1 /*
   2  * Copyright (c) 2015, 2021, Red Hat, Inc. All rights reserved.
   3  * Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This code is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License version 2 only, as
   8  * published by the Free Software Foundation.
   9  *
  10  * This code is distributed in the hope that it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  13  * version 2 for more details (a copy is included in the LICENSE file that
  14  * accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License version
  17  * 2 along with this work; if not, write to the Free Software Foundation,
  18  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  19  *
  20  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  21  * or visit www.oracle.com if you need additional information or have any
  22  * questions.
  23  *
  24  */
  25 
  26 #include "precompiled.hpp"
  27 
  28 #include "classfile/javaClasses.hpp"
  29 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
  30 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
  31 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
  32 #include "gc/shenandoah/shenandoahForwarding.hpp"
  33 #include "gc/shenandoah/shenandoahHeap.hpp"
  34 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
  35 #include "gc/shenandoah/shenandoahRuntime.hpp"
  36 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
  37 #include "opto/arraycopynode.hpp"
  38 #include "opto/block.hpp"
  39 #include "opto/callnode.hpp"
  40 #include "opto/castnode.hpp"
  41 #include "opto/movenode.hpp"
  42 #include "opto/phaseX.hpp"
  43 #include "opto/rootnode.hpp"
  44 #include "opto/runtime.hpp"
  45 #include "opto/subnode.hpp"
  46 
  47 bool ShenandoahBarrierC2Support::expand(Compile* C, PhaseIterGVN& igvn) {
  48   ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
  49   if ((state->iu_barriers_count() +
  50        state->load_reference_barriers_count()) > 0) {
  51     assert(C->post_loop_opts_phase(), "no loop opts allowed");
  52     C->reset_post_loop_opts_phase(); // ... but we know what we are doing
  53     bool attempt_more_loopopts = ShenandoahLoopOptsAfterExpansion;
  54     C->clear_major_progress();
  55     PhaseIdealLoop::optimize(igvn, LoopOptsShenandoahExpand);
  56     if (C->failing()) return false;
  57     PhaseIdealLoop::verify(igvn);
  58     if (attempt_more_loopopts) {
  59       C->set_major_progress();
  60       if (!C->optimize_loops(igvn, LoopOptsShenandoahPostExpand)) {
  61         return false;
  62       }
  63       C->clear_major_progress();
  64 
  65       C->process_for_post_loop_opts_igvn(igvn);
  66     }
  67     C->set_post_loop_opts_phase(); // now for real!
  68   }
  69   return true;
  70 }
  71 
  72 bool ShenandoahBarrierC2Support::is_gc_state_test(Node* iff, int mask) {
  73   if (!UseShenandoahGC) {
  74     return false;
  75   }
  76   assert(iff->is_If(), "bad input");
  77   if (iff->Opcode() != Op_If) {
  78     return false;
  79   }
  80   Node* bol = iff->in(1);
  81   if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) {
  82     return false;
  83   }
  84   Node* cmp = bol->in(1);
  85   if (cmp->Opcode() != Op_CmpI) {
  86     return false;
  87   }
  88   Node* in1 = cmp->in(1);
  89   Node* in2 = cmp->in(2);
  90   if (in2->find_int_con(-1) != 0) {
  91     return false;
  92   }
  93   if (in1->Opcode() != Op_AndI) {
  94     return false;
  95   }
  96   in2 = in1->in(2);
  97   if (in2->find_int_con(-1) != mask) {
  98     return false;
  99   }
 100   in1 = in1->in(1);
 101 
 102   return is_gc_state_load(in1);
 103 }
 104 
 105 bool ShenandoahBarrierC2Support::is_heap_stable_test(Node* iff) {
 106   return is_gc_state_test(iff, ShenandoahHeap::HAS_FORWARDED);
 107 }
 108 
 109 bool ShenandoahBarrierC2Support::is_gc_state_load(Node *n) {
 110   if (!UseShenandoahGC) {
 111     return false;
 112   }
 113   if (n->Opcode() != Op_LoadB && n->Opcode() != Op_LoadUB) {
 114     return false;
 115   }
 116   Node* addp = n->in(MemNode::Address);
 117   if (!addp->is_AddP()) {
 118     return false;
 119   }
 120   Node* base = addp->in(AddPNode::Address);
 121   Node* off = addp->in(AddPNode::Offset);
 122   if (base->Opcode() != Op_ThreadLocal) {
 123     return false;
 124   }
 125   if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) {
 126     return false;
 127   }
 128   return true;
 129 }
 130 
 131 bool ShenandoahBarrierC2Support::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) {
 132   assert(phase->is_dominator(stop, start), "bad inputs");
 133   ResourceMark rm;
 134   Unique_Node_List wq;
 135   wq.push(start);
 136   for (uint next = 0; next < wq.size(); next++) {
 137     Node *m = wq.at(next);
 138     if (m == stop) {
 139       continue;
 140     }
 141     if (m->is_SafePoint() && !m->is_CallLeaf()) {
 142       return true;
 143     }
 144     if (m->is_Region()) {
 145       for (uint i = 1; i < m->req(); i++) {
 146         wq.push(m->in(i));
 147       }
 148     } else {
 149       wq.push(m->in(0));
 150     }
 151   }
 152   return false;
 153 }
 154 
 155 #ifdef ASSERT
 156 bool ShenandoahBarrierC2Support::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) {
 157   assert(phis.size() == 0, "");
 158 
 159   while (true) {
 160     if (in->bottom_type() == TypePtr::NULL_PTR) {
 161       if (trace) {tty->print_cr("NULL");}
 162     } else if (!in->bottom_type()->make_ptr()->make_oopptr()) {
 163       if (trace) {tty->print_cr("Non oop");}
 164     } else {
 165       if (in->is_ConstraintCast()) {
 166         in = in->in(1);
 167         continue;
 168       } else if (in->is_AddP()) {
 169         assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access");
 170         in = in->in(AddPNode::Address);
 171         continue;
 172       } else if (in->is_Con()) {
 173         if (trace) {
 174           tty->print("Found constant");
 175           in->dump();
 176         }
 177       } else if (in->Opcode() == Op_Parm) {
 178         if (trace) {
 179           tty->print("Found argument");
 180         }
 181       } else if (in->Opcode() == Op_CreateEx) {
 182         if (trace) {
 183           tty->print("Found create-exception");
 184         }
 185       } else if (in->Opcode() == Op_LoadP && in->adr_type() == TypeRawPtr::BOTTOM) {
 186         if (trace) {
 187           tty->print("Found raw LoadP (OSR argument?)");
 188         }
 189       } else if (in->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 190         if (t == ShenandoahOopStore) {
 191           uint i = 0;
 192           for (; i < phis.size(); i++) {
 193             Node* n = phis.node_at(i);
 194             if (n->Opcode() == Op_ShenandoahIUBarrier) {
 195               break;
 196             }
 197           }
 198           if (i == phis.size()) {
 199             return false;
 200           }
 201         }
 202         barriers_used.push(in);
 203         if (trace) {tty->print("Found barrier"); in->dump();}
 204       } else if (in->Opcode() == Op_ShenandoahIUBarrier) {
 205         if (t != ShenandoahOopStore) {
 206           in = in->in(1);
 207           continue;
 208         }
 209         if (trace) {tty->print("Found enqueue barrier"); in->dump();}
 210         phis.push(in, in->req());
 211         in = in->in(1);
 212         continue;
 213       } else if (in->is_Proj() && in->in(0)->is_Allocate()) {
 214         if (trace) {
 215           tty->print("Found alloc");
 216           in->in(0)->dump();
 217         }
 218       } else if (in->is_Proj() && (in->in(0)->Opcode() == Op_CallStaticJava || in->in(0)->Opcode() == Op_CallDynamicJava)) {
 219         if (trace) {
 220           tty->print("Found Java call");
 221         }
 222       } else if (in->is_Phi()) {
 223         if (!visited.test_set(in->_idx)) {
 224           if (trace) {tty->print("Pushed phi:"); in->dump();}
 225           phis.push(in, 2);
 226           in = in->in(1);
 227           continue;
 228         }
 229         if (trace) {tty->print("Already seen phi:"); in->dump();}
 230       } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) {
 231         if (!visited.test_set(in->_idx)) {
 232           if (trace) {tty->print("Pushed cmovep:"); in->dump();}
 233           phis.push(in, CMoveNode::IfTrue);
 234           in = in->in(CMoveNode::IfFalse);
 235           continue;
 236         }
 237         if (trace) {tty->print("Already seen cmovep:"); in->dump();}
 238       } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) {
 239         in = in->in(1);
 240         continue;
 241       } else {
 242         return false;
 243       }
 244     }
 245     bool cont = false;
 246     while (phis.is_nonempty()) {
 247       uint idx = phis.index();
 248       Node* phi = phis.node();
 249       if (idx >= phi->req()) {
 250         if (trace) {tty->print("Popped phi:"); phi->dump();}
 251         phis.pop();
 252         continue;
 253       }
 254       if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();}
 255       in = phi->in(idx);
 256       phis.set_index(idx+1);
 257       cont = true;
 258       break;
 259     }
 260     if (!cont) {
 261       break;
 262     }
 263   }
 264   return true;
 265 }
 266 
 267 void ShenandoahBarrierC2Support::report_verify_failure(const char* msg, Node* n1, Node* n2) {
 268   if (n1 != NULL) {
 269     n1->dump(+10);
 270   }
 271   if (n2 != NULL) {
 272     n2->dump(+10);
 273   }
 274   fatal("%s", msg);
 275 }
 276 
 277 void ShenandoahBarrierC2Support::verify(RootNode* root) {
 278   ResourceMark rm;
 279   Unique_Node_List wq;
 280   GrowableArray<Node*> barriers;
 281   Unique_Node_List barriers_used;
 282   Node_Stack phis(0);
 283   VectorSet visited;
 284   const bool trace = false;
 285   const bool verify_no_useless_barrier = false;
 286 
 287   wq.push(root);
 288   for (uint next = 0; next < wq.size(); next++) {
 289     Node *n = wq.at(next);
 290     if (n->is_Load()) {
 291       const bool trace = false;
 292       if (trace) {tty->print("Verifying"); n->dump();}
 293       if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) {
 294         if (trace) {tty->print_cr("Load range/klass");}
 295       } else {
 296         const TypePtr* adr_type = n->as_Load()->adr_type();
 297 
 298         if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) {
 299           if (trace) {tty->print_cr("Mark load");}
 300         } else if (adr_type->isa_instptr() &&
 301                    adr_type->is_instptr()->instance_klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
 302                    adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset()) {
 303           if (trace) {tty->print_cr("Reference.get()");}
 304         } else if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) {
 305           report_verify_failure("Shenandoah verification: Load should have barriers", n);
 306         }
 307       }
 308     } else if (n->is_Store()) {
 309       const bool trace = false;
 310 
 311       if (trace) {tty->print("Verifying"); n->dump();}
 312       if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) {
 313         Node* adr = n->in(MemNode::Address);
 314         bool verify = true;
 315 
 316         if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) {
 317           adr = adr->in(AddPNode::Address);
 318           if (adr->is_AddP()) {
 319             assert(adr->in(AddPNode::Base)->is_top(), "");
 320             adr = adr->in(AddPNode::Address);
 321             if (adr->Opcode() == Op_LoadP &&
 322                 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() &&
 323                 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
 324                 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) {
 325               if (trace) {tty->print_cr("SATB prebarrier");}
 326               verify = false;
 327             }
 328           }
 329         }
 330 
 331         if (verify && !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahIUBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
 332           report_verify_failure("Shenandoah verification: Store should have barriers", n);
 333         }
 334       }
 335       if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
 336         report_verify_failure("Shenandoah verification: Store (address) should have barriers", n);
 337       }
 338     } else if (n->Opcode() == Op_CmpP) {
 339       const bool trace = false;
 340 
 341       Node* in1 = n->in(1);
 342       Node* in2 = n->in(2);
 343       if (in1->bottom_type()->isa_oopptr()) {
 344         if (trace) {tty->print("Verifying"); n->dump();}
 345 
 346         bool mark_inputs = false;
 347         if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR ||
 348             (in1->is_Con() || in2->is_Con())) {
 349           if (trace) {tty->print_cr("Comparison against a constant");}
 350           mark_inputs = true;
 351         } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) ||
 352                    (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) {
 353           if (trace) {tty->print_cr("Comparison with newly alloc'ed object");}
 354           mark_inputs = true;
 355         } else {
 356           assert(in2->bottom_type()->isa_oopptr(), "");
 357 
 358           if (!verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) ||
 359               !verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) {
 360             report_verify_failure("Shenandoah verification: Cmp should have barriers", n);
 361           }
 362         }
 363         if (verify_no_useless_barrier &&
 364             mark_inputs &&
 365             (!verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) ||
 366              !verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) {
 367           phis.clear();
 368           visited.reset();
 369         }
 370       }
 371     } else if (n->is_LoadStore()) {
 372       if (n->in(MemNode::ValueIn)->bottom_type()->make_ptr() &&
 373           !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahIUBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
 374         report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n);
 375       }
 376 
 377       if (n->in(MemNode::Address)->bottom_type()->make_oopptr() && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
 378         report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n);
 379       }
 380     } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) {
 381       CallNode* call = n->as_Call();
 382 
 383       static struct {
 384         const char* name;
 385         struct {
 386           int pos;
 387           verify_type t;
 388         } args[6];
 389       } calls[] = {
 390         "aescrypt_encryptBlock",
 391         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 392           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 393         "aescrypt_decryptBlock",
 394         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 395           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 396         "multiplyToLen",
 397         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },   { TypeFunc::Parms+4, ShenandoahStore },
 398           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 399         "squareToLen",
 400         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },   { -1,  ShenandoahNone},
 401           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 402         "montgomery_multiply",
 403         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },
 404           { TypeFunc::Parms+6, ShenandoahStore }, { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 405         "montgomery_square",
 406         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahLoad },   { TypeFunc::Parms+5, ShenandoahStore },
 407           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 408         "mulAdd",
 409         { { TypeFunc::Parms, ShenandoahStore },  { TypeFunc::Parms+1, ShenandoahLoad },   { -1,  ShenandoahNone},
 410           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 411         "vectorizedMismatch",
 412         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahLoad },   { -1,  ShenandoahNone},
 413           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 414         "updateBytesCRC32",
 415         { { TypeFunc::Parms+1, ShenandoahLoad }, { -1,  ShenandoahNone},                  { -1,  ShenandoahNone},
 416           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 417         "updateBytesAdler32",
 418         { { TypeFunc::Parms+1, ShenandoahLoad }, { -1,  ShenandoahNone},                  { -1,  ShenandoahNone},
 419           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 420         "updateBytesCRC32C",
 421         { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad},    { -1,  ShenandoahNone},
 422           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 423         "counterMode_AESCrypt",
 424         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 425           { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } },
 426         "cipherBlockChaining_encryptAESCrypt",
 427         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 428           { TypeFunc::Parms+3, ShenandoahLoad },  { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 429         "cipherBlockChaining_decryptAESCrypt",
 430         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 431           { TypeFunc::Parms+3, ShenandoahLoad },  { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 432         "shenandoah_clone_barrier",
 433         { { TypeFunc::Parms, ShenandoahLoad },   { -1,  ShenandoahNone},                  { -1,  ShenandoahNone},
 434           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 435         "ghash_processBlocks",
 436         { { TypeFunc::Parms, ShenandoahStore },  { TypeFunc::Parms+1, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },
 437           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 438         "sha1_implCompress",
 439         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 440           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 441         "sha256_implCompress",
 442         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 443           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 444         "sha512_implCompress",
 445         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 446           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 447         "sha1_implCompressMB",
 448         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 449           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 450         "sha256_implCompressMB",
 451         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 452           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 453         "sha512_implCompressMB",
 454         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 455           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 456         "encodeBlock",
 457         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+3, ShenandoahStore },   { -1, ShenandoahNone },
 458           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 459         "decodeBlock",
 460         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+3, ShenandoahStore },   { -1, ShenandoahNone },
 461           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 462       };
 463 
 464       if (call->is_call_to_arraycopystub()) {
 465         Node* dest = NULL;
 466         const TypeTuple* args = n->as_Call()->_tf->domain();
 467         for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
 468           if (args->field_at(i)->isa_ptr()) {
 469             j++;
 470             if (j == 2) {
 471               dest = n->in(i);
 472               break;
 473             }
 474           }
 475         }
 476         if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) ||
 477             !verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) {
 478           report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n);
 479         }
 480       } else if (strlen(call->_name) > 5 &&
 481                  !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) {
 482         if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) {
 483           report_verify_failure("Shenandoah verification: _fill should have barriers", n);
 484         }
 485       } else if (!strcmp(call->_name, "shenandoah_wb_pre")) {
 486         // skip
 487       } else {
 488         const int calls_len = sizeof(calls) / sizeof(calls[0]);
 489         int i = 0;
 490         for (; i < calls_len; i++) {
 491           if (!strcmp(calls[i].name, call->_name)) {
 492             break;
 493           }
 494         }
 495         if (i != calls_len) {
 496           const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]);
 497           for (uint j = 0; j < args_len; j++) {
 498             int pos = calls[i].args[j].pos;
 499             if (pos == -1) {
 500               break;
 501             }
 502             if (!verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) {
 503               report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
 504             }
 505           }
 506           for (uint j = TypeFunc::Parms; j < call->req(); j++) {
 507             if (call->in(j)->bottom_type()->make_ptr() &&
 508                 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
 509               uint k = 0;
 510               for (; k < args_len && calls[i].args[k].pos != (int)j; k++);
 511               if (k == args_len) {
 512                 fatal("arg %d for call %s not covered", j, call->_name);
 513               }
 514             }
 515           }
 516         } else {
 517           for (uint j = TypeFunc::Parms; j < call->req(); j++) {
 518             if (call->in(j)->bottom_type()->make_ptr() &&
 519                 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
 520               fatal("%s not covered", call->_name);
 521             }
 522           }
 523         }
 524       }
 525     } else if (n->Opcode() == Op_ShenandoahIUBarrier || n->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 526       // skip
 527     } else if (n->is_AddP()
 528                || n->is_Phi()
 529                || n->is_ConstraintCast()
 530                || n->Opcode() == Op_Return
 531                || n->Opcode() == Op_CMoveP
 532                || n->Opcode() == Op_CMoveN
 533                || n->Opcode() == Op_Rethrow
 534                || n->is_MemBar()
 535                || n->Opcode() == Op_Conv2B
 536                || n->Opcode() == Op_SafePoint
 537                || n->is_CallJava()
 538                || n->Opcode() == Op_Unlock
 539                || n->Opcode() == Op_EncodeP
 540                || n->Opcode() == Op_DecodeN) {
 541       // nothing to do
 542     } else {
 543       static struct {
 544         int opcode;
 545         struct {
 546           int pos;
 547           verify_type t;
 548         } inputs[2];
 549       } others[] = {
 550         Op_FastLock,
 551         { { 1, ShenandoahLoad },                  { -1, ShenandoahNone} },
 552         Op_Lock,
 553         { { TypeFunc::Parms, ShenandoahLoad },    { -1, ShenandoahNone} },
 554         Op_ArrayCopy,
 555         { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } },
 556         Op_StrCompressedCopy,
 557         { { 2, ShenandoahLoad },                  { 3, ShenandoahStore } },
 558         Op_StrInflatedCopy,
 559         { { 2, ShenandoahLoad },                  { 3, ShenandoahStore } },
 560         Op_AryEq,
 561         { { 2, ShenandoahLoad },                  { 3, ShenandoahLoad } },
 562         Op_StrIndexOf,
 563         { { 2, ShenandoahLoad },                  { 4, ShenandoahLoad } },
 564         Op_StrComp,
 565         { { 2, ShenandoahLoad },                  { 4, ShenandoahLoad } },
 566         Op_StrEquals,
 567         { { 2, ShenandoahLoad },                  { 3, ShenandoahLoad } },
 568         Op_EncodeISOArray,
 569         { { 2, ShenandoahLoad },                  { 3, ShenandoahStore } },
 570         Op_CountPositives,
 571         { { 2, ShenandoahLoad },                  { -1, ShenandoahNone} },
 572         Op_CastP2X,
 573         { { 1, ShenandoahLoad },                  { -1, ShenandoahNone} },
 574         Op_StrIndexOfChar,
 575         { { 2, ShenandoahLoad },                  { -1, ShenandoahNone } },
 576       };
 577 
 578       const int others_len = sizeof(others) / sizeof(others[0]);
 579       int i = 0;
 580       for (; i < others_len; i++) {
 581         if (others[i].opcode == n->Opcode()) {
 582           break;
 583         }
 584       }
 585       uint stop = n->is_Call() ? n->as_Call()->tf()->domain()->cnt() : n->req();
 586       if (i != others_len) {
 587         const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]);
 588         for (uint j = 0; j < inputs_len; j++) {
 589           int pos = others[i].inputs[j].pos;
 590           if (pos == -1) {
 591             break;
 592           }
 593           if (!verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) {
 594             report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
 595           }
 596         }
 597         for (uint j = 1; j < stop; j++) {
 598           if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() &&
 599               n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
 600             uint k = 0;
 601             for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++);
 602             if (k == inputs_len) {
 603               fatal("arg %d for node %s not covered", j, n->Name());
 604             }
 605           }
 606         }
 607       } else {
 608         for (uint j = 1; j < stop; j++) {
 609           if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() &&
 610               n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
 611             fatal("%s not covered", n->Name());
 612           }
 613         }
 614       }
 615     }
 616 
 617     if (n->is_SafePoint()) {
 618       SafePointNode* sfpt = n->as_SafePoint();
 619       if (verify_no_useless_barrier && sfpt->jvms() != NULL) {
 620         for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) {
 621           if (!verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) {
 622             phis.clear();
 623             visited.reset();
 624           }
 625         }
 626       }
 627     }
 628   }
 629 
 630   if (verify_no_useless_barrier) {
 631     for (int i = 0; i < barriers.length(); i++) {
 632       Node* n = barriers.at(i);
 633       if (!barriers_used.member(n)) {
 634         tty->print("XXX useless barrier"); n->dump(-2);
 635         ShouldNotReachHere();
 636       }
 637     }
 638   }
 639 }
 640 #endif
 641 
 642 bool ShenandoahBarrierC2Support::is_dominator_same_ctrl(Node* c, Node* d, Node* n, PhaseIdealLoop* phase) {
 643   // That both nodes have the same control is not sufficient to prove
 644   // domination, verify that there's no path from d to n
 645   ResourceMark rm;
 646   Unique_Node_List wq;
 647   wq.push(d);
 648   for (uint next = 0; next < wq.size(); next++) {
 649     Node *m = wq.at(next);
 650     if (m == n) {
 651       return false;
 652     }
 653     if (m->is_Phi() && m->in(0)->is_Loop()) {
 654       assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control");
 655     } else {
 656       if (m->is_Store() || m->is_LoadStore()) {
 657         // Take anti-dependencies into account
 658         Node* mem = m->in(MemNode::Memory);
 659         for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
 660           Node* u = mem->fast_out(i);
 661           if (u->is_Load() && phase->C->can_alias(m->adr_type(), phase->C->get_alias_index(u->adr_type())) &&
 662               phase->ctrl_or_self(u) == c) {
 663             wq.push(u);
 664           }
 665         }
 666       }
 667       for (uint i = 0; i < m->req(); i++) {
 668         if (m->in(i) != NULL && phase->ctrl_or_self(m->in(i)) == c) {
 669           wq.push(m->in(i));
 670         }
 671       }
 672     }
 673   }
 674   return true;
 675 }
 676 
 677 bool ShenandoahBarrierC2Support::is_dominator(Node* d_c, Node* n_c, Node* d, Node* n, PhaseIdealLoop* phase) {
 678   if (d_c != n_c) {
 679     return phase->is_dominator(d_c, n_c);
 680   }
 681   return is_dominator_same_ctrl(d_c, d, n, phase);
 682 }
 683 
 684 Node* next_mem(Node* mem, int alias) {
 685   Node* res = NULL;
 686   if (mem->is_Proj()) {
 687     res = mem->in(0);
 688   } else if (mem->is_SafePoint() || mem->is_MemBar()) {
 689     res = mem->in(TypeFunc::Memory);
 690   } else if (mem->is_Phi()) {
 691     res = mem->in(1);
 692   } else if (mem->is_MergeMem()) {
 693     res = mem->as_MergeMem()->memory_at(alias);
 694   } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
 695     assert(alias == Compile::AliasIdxRaw, "following raw memory can't lead to a barrier");
 696     res = mem->in(MemNode::Memory);
 697   } else {
 698 #ifdef ASSERT
 699     mem->dump();
 700 #endif
 701     ShouldNotReachHere();
 702   }
 703   return res;
 704 }
 705 
 706 Node* ShenandoahBarrierC2Support::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) {
 707   Node* iffproj = NULL;
 708   while (c != dom) {
 709     Node* next = phase->idom(c);
 710     assert(next->unique_ctrl_out_or_null() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?");
 711     if (c->is_Region()) {
 712       ResourceMark rm;
 713       Unique_Node_List wq;
 714       wq.push(c);
 715       for (uint i = 0; i < wq.size(); i++) {
 716         Node *n = wq.at(i);
 717         if (n == next) {
 718           continue;
 719         }
 720         if (n->is_Region()) {
 721           for (uint j = 1; j < n->req(); j++) {
 722             wq.push(n->in(j));
 723           }
 724         } else {
 725           wq.push(n->in(0));
 726         }
 727       }
 728       for (uint i = 0; i < wq.size(); i++) {
 729         Node *n = wq.at(i);
 730         assert(n->is_CFG(), "");
 731         if (n->is_Multi()) {
 732           for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
 733             Node* u = n->fast_out(j);
 734             if (u->is_CFG()) {
 735               if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
 736                 return NodeSentinel;
 737               }
 738             }
 739           }
 740         }
 741       }
 742     } else  if (c->is_Proj()) {
 743       if (c->is_IfProj()) {
 744         if (c->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) != NULL) {
 745           // continue;
 746         } else {
 747           if (!allow_one_proj) {
 748             return NodeSentinel;
 749           }
 750           if (iffproj == NULL) {
 751             iffproj = c;
 752           } else {
 753             return NodeSentinel;
 754           }
 755         }
 756       } else if (c->Opcode() == Op_JumpProj) {
 757         return NodeSentinel; // unsupported
 758       } else if (c->Opcode() == Op_CatchProj) {
 759         return NodeSentinel; // unsupported
 760       } else if (c->Opcode() == Op_CProj && next->Opcode() == Op_NeverBranch) {
 761         return NodeSentinel; // unsupported
 762       } else {
 763         assert(next->unique_ctrl_out() == c, "unsupported branch pattern");
 764       }
 765     }
 766     c = next;
 767   }
 768   return iffproj;
 769 }
 770 
 771 Node* ShenandoahBarrierC2Support::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) {
 772   ResourceMark rm;
 773   VectorSet wq;
 774   wq.set(mem->_idx);
 775   mem_ctrl = phase->ctrl_or_self(mem);
 776   while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) {
 777     mem = next_mem(mem, alias);
 778     if (wq.test_set(mem->_idx)) {
 779       return NULL;
 780     }
 781     mem_ctrl = phase->ctrl_or_self(mem);
 782   }
 783   if (mem->is_MergeMem()) {
 784     mem = mem->as_MergeMem()->memory_at(alias);
 785     mem_ctrl = phase->ctrl_or_self(mem);
 786   }
 787   return mem;
 788 }
 789 
 790 Node* ShenandoahBarrierC2Support::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) {
 791   Node* mem = NULL;
 792   Node* c = ctrl;
 793   do {
 794     if (c->is_Region()) {
 795       for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == NULL; i++) {
 796         Node* u = c->fast_out(i);
 797         if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
 798           if (u->adr_type() == TypePtr::BOTTOM) {
 799             mem = u;
 800           }
 801         }
 802       }
 803     } else {
 804       if (c->is_Call() && c->as_Call()->adr_type() != NULL) {
 805         CallProjections projs;
 806         c->as_Call()->extract_projections(&projs, true, false);
 807         if (projs.fallthrough_memproj != NULL) {
 808           if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) {
 809             if (projs.catchall_memproj == NULL) {
 810               mem = projs.fallthrough_memproj;
 811             } else {
 812               if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) {
 813                 mem = projs.fallthrough_memproj;
 814               } else {
 815                 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier");
 816                 mem = projs.catchall_memproj;
 817               }
 818             }
 819           }
 820         } else {
 821           Node* proj = c->as_Call()->proj_out(TypeFunc::Memory);
 822           if (proj != NULL &&
 823               proj->adr_type() == TypePtr::BOTTOM) {
 824             mem = proj;
 825           }
 826         }
 827       } else {
 828         for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
 829           Node* u = c->fast_out(i);
 830           if (u->is_Proj() &&
 831               u->bottom_type() == Type::MEMORY &&
 832               u->adr_type() == TypePtr::BOTTOM) {
 833               assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), "");
 834               assert(mem == NULL, "only one proj");
 835               mem = u;
 836           }
 837         }
 838         assert(!c->is_Call() || c->as_Call()->adr_type() != NULL || mem == NULL, "no mem projection expected");
 839       }
 840     }
 841     c = phase->idom(c);
 842   } while (mem == NULL);
 843   return mem;
 844 }
 845 
 846 void ShenandoahBarrierC2Support::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) {
 847   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 848     Node* u = n->fast_out(i);
 849     if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) {
 850       uses.push(u);
 851     }
 852   }
 853 }
 854 
 855 static void hide_strip_mined_loop(OuterStripMinedLoopNode* outer, CountedLoopNode* inner, PhaseIdealLoop* phase) {
 856   OuterStripMinedLoopEndNode* le = inner->outer_loop_end();
 857   Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl));
 858   phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl));
 859   Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt);
 860   phase->register_control(new_le, phase->get_loop(le), le->in(0));
 861   phase->lazy_replace(outer, new_outer);
 862   phase->lazy_replace(le, new_le);
 863   inner->clear_strip_mined();
 864 }
 865 
 866 void ShenandoahBarrierC2Support::test_gc_state(Node*& ctrl, Node* raw_mem, Node*& test_fail_ctrl,
 867                                                PhaseIdealLoop* phase, int flags) {
 868   PhaseIterGVN& igvn = phase->igvn();
 869   Node* old_ctrl = ctrl;
 870 
 871   Node* thread          = new ThreadLocalNode();
 872   Node* gc_state_offset = igvn.MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 873   Node* gc_state_addr   = new AddPNode(phase->C->top(), thread, gc_state_offset);
 874   Node* gc_state        = new LoadBNode(old_ctrl, raw_mem, gc_state_addr,
 875                                         DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(NULL),
 876                                         TypeInt::BYTE, MemNode::unordered);
 877   Node* gc_state_and    = new AndINode(gc_state, igvn.intcon(flags));
 878   Node* gc_state_cmp    = new CmpINode(gc_state_and, igvn.zerocon(T_INT));
 879   Node* gc_state_bool   = new BoolNode(gc_state_cmp, BoolTest::ne);
 880 
 881   IfNode* gc_state_iff  = new IfNode(old_ctrl, gc_state_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
 882   ctrl                  = new IfTrueNode(gc_state_iff);
 883   test_fail_ctrl        = new IfFalseNode(gc_state_iff);
 884 
 885   IdealLoopTree* loop = phase->get_loop(old_ctrl);
 886   phase->register_control(gc_state_iff,   loop, old_ctrl);
 887   phase->register_control(ctrl,           loop, gc_state_iff);
 888   phase->register_control(test_fail_ctrl, loop, gc_state_iff);
 889 
 890   phase->register_new_node(thread,        old_ctrl);
 891   phase->register_new_node(gc_state_addr, old_ctrl);
 892   phase->register_new_node(gc_state,      old_ctrl);
 893   phase->register_new_node(gc_state_and,  old_ctrl);
 894   phase->register_new_node(gc_state_cmp,  old_ctrl);
 895   phase->register_new_node(gc_state_bool, old_ctrl);
 896 
 897   phase->set_ctrl(gc_state_offset, phase->C->root());
 898 
 899   assert(is_gc_state_test(gc_state_iff, flags), "Should match the shape");
 900 }
 901 
 902 void ShenandoahBarrierC2Support::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) {
 903   Node* old_ctrl = ctrl;
 904   PhaseIterGVN& igvn = phase->igvn();
 905 
 906   const Type* val_t = igvn.type(val);
 907   if (val_t->meet(TypePtr::NULL_PTR) == val_t) {
 908     Node* null_cmp   = new CmpPNode(val, igvn.zerocon(T_OBJECT));
 909     Node* null_test  = new BoolNode(null_cmp, BoolTest::ne);
 910 
 911     IfNode* null_iff = new IfNode(old_ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
 912     ctrl             = new IfTrueNode(null_iff);
 913     null_ctrl        = new IfFalseNode(null_iff);
 914 
 915     IdealLoopTree* loop = phase->get_loop(old_ctrl);
 916     phase->register_control(null_iff,  loop, old_ctrl);
 917     phase->register_control(ctrl,      loop, null_iff);
 918     phase->register_control(null_ctrl, loop, null_iff);
 919 
 920     phase->register_new_node(null_cmp,  old_ctrl);
 921     phase->register_new_node(null_test, old_ctrl);
 922   }
 923 }
 924 
 925 void ShenandoahBarrierC2Support::test_in_cset(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) {
 926   Node* old_ctrl = ctrl;
 927   PhaseIterGVN& igvn = phase->igvn();
 928 
 929   Node* raw_val        = new CastP2XNode(old_ctrl, val);
 930   Node* cset_idx       = new URShiftXNode(raw_val, igvn.intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint()));
 931 
 932   // Figure out the target cset address with raw pointer math.
 933   // This avoids matching AddP+LoadB that would emit inefficient code.
 934   // See JDK-8245465.
 935   Node* cset_addr_ptr  = igvn.makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr()));
 936   Node* cset_addr      = new CastP2XNode(old_ctrl, cset_addr_ptr);
 937   Node* cset_load_addr = new AddXNode(cset_addr, cset_idx);
 938   Node* cset_load_ptr  = new CastX2PNode(cset_load_addr);
 939 
 940   Node* cset_load      = new LoadBNode(old_ctrl, raw_mem, cset_load_ptr,
 941                                        DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(NULL),
 942                                        TypeInt::BYTE, MemNode::unordered);
 943   Node* cset_cmp       = new CmpINode(cset_load, igvn.zerocon(T_INT));
 944   Node* cset_bool      = new BoolNode(cset_cmp, BoolTest::ne);
 945 
 946   IfNode* cset_iff     = new IfNode(old_ctrl, cset_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
 947   ctrl                 = new IfTrueNode(cset_iff);
 948   not_cset_ctrl        = new IfFalseNode(cset_iff);
 949 
 950   IdealLoopTree *loop = phase->get_loop(old_ctrl);
 951   phase->register_control(cset_iff,      loop, old_ctrl);
 952   phase->register_control(ctrl,          loop, cset_iff);
 953   phase->register_control(not_cset_ctrl, loop, cset_iff);
 954 
 955   phase->set_ctrl(cset_addr_ptr, phase->C->root());
 956 
 957   phase->register_new_node(raw_val,        old_ctrl);
 958   phase->register_new_node(cset_idx,       old_ctrl);
 959   phase->register_new_node(cset_addr,      old_ctrl);
 960   phase->register_new_node(cset_load_addr, old_ctrl);
 961   phase->register_new_node(cset_load_ptr,  old_ctrl);
 962   phase->register_new_node(cset_load,      old_ctrl);
 963   phase->register_new_node(cset_cmp,       old_ctrl);
 964   phase->register_new_node(cset_bool,      old_ctrl);
 965 }
 966 
 967 void ShenandoahBarrierC2Support::call_lrb_stub(Node*& ctrl, Node*& val, Node* load_addr,
 968                                                DecoratorSet decorators, PhaseIdealLoop* phase) {
 969   IdealLoopTree*loop = phase->get_loop(ctrl);
 970   const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr();
 971 
 972   address calladdr = NULL;
 973   const char* name = NULL;
 974   bool is_strong  = ShenandoahBarrierSet::is_strong_access(decorators);
 975   bool is_weak    = ShenandoahBarrierSet::is_weak_access(decorators);
 976   bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
 977   bool is_native  = ShenandoahBarrierSet::is_native_access(decorators);
 978   bool is_narrow  = UseCompressedOops && !is_native;
 979   if (is_strong) {
 980     if (is_narrow) {
 981       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow);
 982       name = "load_reference_barrier_strong_narrow";
 983     } else {
 984       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong);
 985       name = "load_reference_barrier_strong";
 986     }
 987   } else if (is_weak) {
 988     if (is_narrow) {
 989       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow);
 990       name = "load_reference_barrier_weak_narrow";
 991     } else {
 992       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak);
 993       name = "load_reference_barrier_weak";
 994     }
 995   } else {
 996     assert(is_phantom, "only remaining strength");
 997     if (is_narrow) {
 998       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom_narrow);
 999       name = "load_reference_barrier_phantom_narrow";
1000     } else {
1001       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom);
1002       name = "load_reference_barrier_phantom";
1003     }
1004   }
1005   Node* call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type(), calladdr, name, TypeRawPtr::BOTTOM);
1006 
1007   call->init_req(TypeFunc::Control, ctrl);
1008   call->init_req(TypeFunc::I_O, phase->C->top());
1009   call->init_req(TypeFunc::Memory, phase->C->top());
1010   call->init_req(TypeFunc::FramePtr, phase->C->top());
1011   call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1012   call->init_req(TypeFunc::Parms, val);
1013   call->init_req(TypeFunc::Parms+1, load_addr);
1014   phase->register_control(call, loop, ctrl);
1015   ctrl = new ProjNode(call, TypeFunc::Control);
1016   phase->register_control(ctrl, loop, call);
1017   val = new ProjNode(call, TypeFunc::Parms);
1018   phase->register_new_node(val, call);
1019   val = new CheckCastPPNode(ctrl, val, obj_type);
1020   phase->register_new_node(val, ctrl);
1021 }
1022 
1023 void ShenandoahBarrierC2Support::fix_ctrl(Node* barrier, Node* region, const MemoryGraphFixer& fixer, Unique_Node_List& uses, Unique_Node_List& uses_to_ignore, uint last, PhaseIdealLoop* phase) {
1024   Node* ctrl = phase->get_ctrl(barrier);
1025   Node* init_raw_mem = fixer.find_mem(ctrl, barrier);
1026 
1027   // Update the control of all nodes that should be after the
1028   // barrier control flow
1029   uses.clear();
1030   // Every node that is control dependent on the barrier's input
1031   // control will be after the expanded barrier. The raw memory (if
1032   // its memory is control dependent on the barrier's input control)
1033   // must stay above the barrier.
1034   uses_to_ignore.clear();
1035   if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) {
1036     uses_to_ignore.push(init_raw_mem);
1037   }
1038   for (uint next = 0; next < uses_to_ignore.size(); next++) {
1039     Node *n = uses_to_ignore.at(next);
1040     for (uint i = 0; i < n->req(); i++) {
1041       Node* in = n->in(i);
1042       if (in != NULL && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) {
1043         uses_to_ignore.push(in);
1044       }
1045     }
1046   }
1047   for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) {
1048     Node* u = ctrl->fast_out(i);
1049     if (u->_idx < last &&
1050         u != barrier &&
1051         !uses_to_ignore.member(u) &&
1052         (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) &&
1053         (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) {
1054       Node* old_c = phase->ctrl_or_self(u);
1055       Node* c = old_c;
1056       if (c != ctrl ||
1057           is_dominator_same_ctrl(old_c, barrier, u, phase) ||
1058           ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) {
1059         phase->igvn().rehash_node_delayed(u);
1060         int nb = u->replace_edge(ctrl, region, &phase->igvn());
1061         if (u->is_CFG()) {
1062           if (phase->idom(u) == ctrl) {
1063             phase->set_idom(u, region, phase->dom_depth(region));
1064           }
1065         } else if (phase->get_ctrl(u) == ctrl) {
1066           assert(u != init_raw_mem, "should leave input raw mem above the barrier");
1067           uses.push(u);
1068         }
1069         assert(nb == 1, "more than 1 ctrl input?");
1070         --i, imax -= nb;
1071       }
1072     }
1073   }
1074 }
1075 
1076 static Node* create_phis_on_call_return(Node* ctrl, Node* c, Node* n, Node* n_clone, const CallProjections& projs, PhaseIdealLoop* phase) {
1077   Node* region = NULL;
1078   while (c != ctrl) {
1079     if (c->is_Region()) {
1080       region = c;
1081     }
1082     c = phase->idom(c);
1083   }
1084   assert(region != NULL, "");
1085   Node* phi = new PhiNode(region, n->bottom_type());
1086   for (uint j = 1; j < region->req(); j++) {
1087     Node* in = region->in(j);
1088     if (phase->is_dominator(projs.fallthrough_catchproj, in)) {
1089       phi->init_req(j, n);
1090     } else if (phase->is_dominator(projs.catchall_catchproj, in)) {
1091       phi->init_req(j, n_clone);
1092     } else {
1093       phi->init_req(j, create_phis_on_call_return(ctrl, in, n, n_clone, projs, phase));
1094     }
1095   }
1096   phase->register_new_node(phi, region);
1097   return phi;
1098 }
1099 
1100 void ShenandoahBarrierC2Support::pin_and_expand(PhaseIdealLoop* phase) {
1101   ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
1102 
1103   Unique_Node_List uses;
1104   for (int i = 0; i < state->iu_barriers_count(); i++) {
1105     Node* barrier = state->iu_barrier(i);
1106     Node* ctrl = phase->get_ctrl(barrier);
1107     IdealLoopTree* loop = phase->get_loop(ctrl);
1108     Node* head = loop->head();
1109     if (head->is_OuterStripMinedLoop()) {
1110       // Expanding a barrier here will break loop strip mining
1111       // verification. Transform the loop so the loop nest doesn't
1112       // appear as strip mined.
1113       OuterStripMinedLoopNode* outer = head->as_OuterStripMinedLoop();
1114       hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1115     }
1116   }
1117 
1118   Node_Stack stack(0);
1119   Node_List clones;
1120   for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1121     ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1122 
1123     Node* ctrl = phase->get_ctrl(lrb);
1124     Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1125 
1126     CallStaticJavaNode* unc = NULL;
1127     Node* unc_ctrl = NULL;
1128     Node* uncasted_val = val;
1129 
1130     for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) {
1131       Node* u = lrb->fast_out(i);
1132       if (u->Opcode() == Op_CastPP &&
1133           u->in(0) != NULL &&
1134           phase->is_dominator(u->in(0), ctrl)) {
1135         const Type* u_t = phase->igvn().type(u);
1136 
1137         if (u_t->meet(TypePtr::NULL_PTR) != u_t &&
1138             u->in(0)->Opcode() == Op_IfTrue &&
1139             u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
1140             u->in(0)->in(0)->is_If() &&
1141             u->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
1142             u->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
1143             u->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
1144             u->in(0)->in(0)->in(1)->in(1)->in(1) == val &&
1145             u->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
1146           IdealLoopTree* loop = phase->get_loop(ctrl);
1147           IdealLoopTree* unc_loop = phase->get_loop(u->in(0));
1148 
1149           if (!unc_loop->is_member(loop)) {
1150             continue;
1151           }
1152 
1153           Node* branch = no_branches(ctrl, u->in(0), false, phase);
1154           assert(branch == NULL || branch == NodeSentinel, "was not looking for a branch");
1155           if (branch == NodeSentinel) {
1156             continue;
1157           }
1158 
1159           Node* iff = u->in(0)->in(0);
1160           Node* bol = iff->in(1)->clone();
1161           Node* cmp = bol->in(1)->clone();
1162           cmp->set_req(1, lrb);
1163           bol->set_req(1, cmp);
1164           phase->igvn().replace_input_of(iff, 1, bol);
1165           phase->set_ctrl(lrb, iff->in(0));
1166           phase->register_new_node(cmp, iff->in(0));
1167           phase->register_new_node(bol, iff->in(0));
1168           break;
1169         }
1170       }
1171     }
1172     if ((ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) || ctrl->is_CallJava()) {
1173       CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_CallJava() : ctrl->as_CallJava();
1174       if (call->entry_point() == OptoRuntime::rethrow_stub()) {
1175         // The rethrow call may have too many projections to be
1176         // properly handled here. Given there's no reason for a
1177         // barrier to depend on the call, move it above the call
1178         stack.push(lrb, 0);
1179         do {
1180           Node* n = stack.node();
1181           uint idx = stack.index();
1182           if (idx < n->req()) {
1183             Node* in = n->in(idx);
1184             stack.set_index(idx+1);
1185             if (in != NULL) {
1186               if (phase->has_ctrl(in)) {
1187                 if (phase->is_dominator(call, phase->get_ctrl(in))) {
1188 #ifdef ASSERT
1189                   for (uint i = 0; i < stack.size(); i++) {
1190                     assert(stack.node_at(i) != in, "node shouldn't have been seen yet");
1191                   }
1192 #endif
1193                   stack.push(in, 0);
1194                 }
1195               } else {
1196                 assert(phase->is_dominator(in, call->in(0)), "no dependency on the call");
1197               }
1198             }
1199           } else {
1200             phase->set_ctrl(n, call->in(0));
1201             stack.pop();
1202           }
1203         } while(stack.size() > 0);
1204         continue;
1205       }
1206       CallProjections projs;
1207       call->extract_projections(&projs, false, false);
1208 
1209 #ifdef ASSERT
1210       VectorSet cloned;
1211 #endif
1212       Node* lrb_clone = lrb->clone();
1213       phase->register_new_node(lrb_clone, projs.catchall_catchproj);
1214       phase->set_ctrl(lrb, projs.fallthrough_catchproj);
1215 
1216       stack.push(lrb, 0);
1217       clones.push(lrb_clone);
1218 
1219       do {
1220         assert(stack.size() == clones.size(), "");
1221         Node* n = stack.node();
1222 #ifdef ASSERT
1223         if (n->is_Load()) {
1224           Node* mem = n->in(MemNode::Memory);
1225           for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
1226             Node* u = mem->fast_out(j);
1227             assert(!u->is_Store() || !u->is_LoadStore() || phase->get_ctrl(u) != ctrl, "anti dependent store?");
1228           }
1229         }
1230 #endif
1231         uint idx = stack.index();
1232         Node* n_clone = clones.at(clones.size()-1);
1233         if (idx < n->outcnt()) {
1234           Node* u = n->raw_out(idx);
1235           Node* c = phase->ctrl_or_self(u);
1236           if (phase->is_dominator(call, c) && phase->is_dominator(c, projs.fallthrough_proj)) {
1237             stack.set_index(idx+1);
1238             assert(!u->is_CFG(), "");
1239             stack.push(u, 0);
1240             assert(!cloned.test_set(u->_idx), "only one clone");
1241             Node* u_clone = u->clone();
1242             int nb = u_clone->replace_edge(n, n_clone, &phase->igvn());
1243             assert(nb > 0, "should have replaced some uses");
1244             phase->register_new_node(u_clone, projs.catchall_catchproj);
1245             clones.push(u_clone);
1246             phase->set_ctrl(u, projs.fallthrough_catchproj);
1247           } else {
1248             bool replaced = false;
1249             if (u->is_Phi()) {
1250               for (uint k = 1; k < u->req(); k++) {
1251                 if (u->in(k) == n) {
1252                   if (phase->is_dominator(projs.catchall_catchproj, u->in(0)->in(k))) {
1253                     phase->igvn().replace_input_of(u, k, n_clone);
1254                     replaced = true;
1255                   } else if (!phase->is_dominator(projs.fallthrough_catchproj, u->in(0)->in(k))) {
1256                     phase->igvn().replace_input_of(u, k, create_phis_on_call_return(ctrl, u->in(0)->in(k), n, n_clone, projs, phase));
1257                     replaced = true;
1258                   }
1259                 }
1260               }
1261             } else {
1262               if (phase->is_dominator(projs.catchall_catchproj, c)) {
1263                 phase->igvn().rehash_node_delayed(u);
1264                 int nb = u->replace_edge(n, n_clone, &phase->igvn());
1265                 assert(nb > 0, "should have replaced some uses");
1266                 replaced = true;
1267               } else if (!phase->is_dominator(projs.fallthrough_catchproj, c)) {
1268                 if (u->is_If()) {
1269                   // Can't break If/Bool/Cmp chain
1270                   assert(n->is_Bool(), "unexpected If shape");
1271                   assert(stack.node_at(stack.size()-2)->is_Cmp(), "unexpected If shape");
1272                   assert(n_clone->is_Bool(), "unexpected clone");
1273                   assert(clones.at(clones.size()-2)->is_Cmp(), "unexpected clone");
1274                   Node* bol_clone = n->clone();
1275                   Node* cmp_clone = stack.node_at(stack.size()-2)->clone();
1276                   bol_clone->set_req(1, cmp_clone);
1277 
1278                   Node* nn = stack.node_at(stack.size()-3);
1279                   Node* nn_clone = clones.at(clones.size()-3);
1280                   assert(nn->Opcode() == nn_clone->Opcode(), "mismatch");
1281 
1282                   int nb = cmp_clone->replace_edge(nn, create_phis_on_call_return(ctrl, c, nn, nn_clone, projs, phase),
1283                                                    &phase->igvn());
1284                   assert(nb > 0, "should have replaced some uses");
1285 
1286                   phase->register_new_node(bol_clone, u->in(0));
1287                   phase->register_new_node(cmp_clone, u->in(0));
1288 
1289                   phase->igvn().replace_input_of(u, 1, bol_clone);
1290 
1291                 } else {
1292                   phase->igvn().rehash_node_delayed(u);
1293                   int nb = u->replace_edge(n, create_phis_on_call_return(ctrl, c, n, n_clone, projs, phase), &phase->igvn());
1294                   assert(nb > 0, "should have replaced some uses");
1295                 }
1296                 replaced = true;
1297               }
1298             }
1299             if (!replaced) {
1300               stack.set_index(idx+1);
1301             }
1302           }
1303         } else {
1304           stack.pop();
1305           clones.pop();
1306         }
1307       } while (stack.size() > 0);
1308       assert(stack.size() == 0 && clones.size() == 0, "");
1309     }
1310   }
1311 
1312   for (int i = 0; i < state->load_reference_barriers_count(); i++) {
1313     ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1314     Node* ctrl = phase->get_ctrl(lrb);
1315     IdealLoopTree* loop = phase->get_loop(ctrl);
1316     Node* head = loop->head();
1317     if (head->is_OuterStripMinedLoop()) {
1318       // Expanding a barrier here will break loop strip mining
1319       // verification. Transform the loop so the loop nest doesn't
1320       // appear as strip mined.
1321       OuterStripMinedLoopNode* outer = head->as_OuterStripMinedLoop();
1322       hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1323     }
1324   }
1325 
1326   // Expand load-reference-barriers
1327   MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
1328   Unique_Node_List uses_to_ignore;
1329   for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1330     ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1331     uint last = phase->C->unique();
1332     Node* ctrl = phase->get_ctrl(lrb);
1333     Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1334 
1335     Node* orig_ctrl = ctrl;
1336 
1337     Node* raw_mem = fixer.find_mem(ctrl, lrb);
1338     Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL);
1339 
1340     IdealLoopTree *loop = phase->get_loop(ctrl);
1341 
1342     Node* heap_stable_ctrl = NULL;
1343     Node* null_ctrl = NULL;
1344 
1345     assert(val->bottom_type()->make_oopptr(), "need oop");
1346     assert(val->bottom_type()->make_oopptr()->const_oop() == NULL, "expect non-constant");
1347 
1348     enum { _heap_stable = 1, _evac_path, _not_cset, PATH_LIMIT };
1349     Node* region = new RegionNode(PATH_LIMIT);
1350     Node* val_phi = new PhiNode(region, val->bottom_type()->is_oopptr());
1351 
1352     // Stable path.
1353     int flags = ShenandoahHeap::HAS_FORWARDED;
1354     if (!ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1355       flags |= ShenandoahHeap::WEAK_ROOTS;
1356     }
1357     test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, flags);
1358     IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
1359 
1360     // Heap stable case
1361     region->init_req(_heap_stable, heap_stable_ctrl);
1362     val_phi->init_req(_heap_stable, val);
1363 
1364     // Test for in-cset, unless it's a native-LRB. Native LRBs need to return NULL
1365     // even for non-cset objects to prevent resurrection of such objects.
1366     // Wires !in_cset(obj) to slot 2 of region and phis
1367     Node* not_cset_ctrl = NULL;
1368     if (ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1369       test_in_cset(ctrl, not_cset_ctrl, val, raw_mem, phase);
1370     }
1371     if (not_cset_ctrl != NULL) {
1372       region->init_req(_not_cset, not_cset_ctrl);
1373       val_phi->init_req(_not_cset, val);
1374     } else {
1375       region->del_req(_not_cset);
1376       val_phi->del_req(_not_cset);
1377     }
1378 
1379     // Resolve object when orig-value is in cset.
1380     // Make the unconditional resolve for fwdptr.
1381 
1382     // Call lrb-stub and wire up that path in slots 4
1383     Node* result_mem = NULL;
1384 
1385     Node* addr;
1386     if (ShenandoahSelfFixing) {
1387       VectorSet visited;
1388       addr = get_load_addr(phase, visited, lrb);
1389     } else {
1390       addr = phase->igvn().zerocon(T_OBJECT);
1391     }
1392     if (addr->Opcode() == Op_AddP) {
1393       Node* orig_base = addr->in(AddPNode::Base);
1394       Node* base = new CheckCastPPNode(ctrl, orig_base, orig_base->bottom_type(), ConstraintCastNode::StrongDependency);
1395       phase->register_new_node(base, ctrl);
1396       if (addr->in(AddPNode::Base) == addr->in((AddPNode::Address))) {
1397         // Field access
1398         addr = addr->clone();
1399         addr->set_req(AddPNode::Base, base);
1400         addr->set_req(AddPNode::Address, base);
1401         phase->register_new_node(addr, ctrl);
1402       } else {
1403         Node* addr2 = addr->in(AddPNode::Address);
1404         if (addr2->Opcode() == Op_AddP && addr2->in(AddPNode::Base) == addr2->in(AddPNode::Address) &&
1405               addr2->in(AddPNode::Base) == orig_base) {
1406           addr2 = addr2->clone();
1407           addr2->set_req(AddPNode::Base, base);
1408           addr2->set_req(AddPNode::Address, base);
1409           phase->register_new_node(addr2, ctrl);
1410           addr = addr->clone();
1411           addr->set_req(AddPNode::Base, base);
1412           addr->set_req(AddPNode::Address, addr2);
1413           phase->register_new_node(addr, ctrl);
1414         }
1415       }
1416     }
1417     call_lrb_stub(ctrl, val, addr, lrb->decorators(), phase);
1418     region->init_req(_evac_path, ctrl);
1419     val_phi->init_req(_evac_path, val);
1420 
1421     phase->register_control(region, loop, heap_stable_iff);
1422     Node* out_val = val_phi;
1423     phase->register_new_node(val_phi, region);
1424 
1425     fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase);
1426 
1427     ctrl = orig_ctrl;
1428 
1429     phase->igvn().replace_node(lrb, out_val);
1430 
1431     follow_barrier_uses(out_val, ctrl, uses, phase);
1432 
1433     for(uint next = 0; next < uses.size(); next++ ) {
1434       Node *n = uses.at(next);
1435       assert(phase->get_ctrl(n) == ctrl, "bad control");
1436       assert(n != raw_mem, "should leave input raw mem above the barrier");
1437       phase->set_ctrl(n, region);
1438       follow_barrier_uses(n, ctrl, uses, phase);
1439     }
1440     fixer.record_new_ctrl(ctrl, region, raw_mem, raw_mem_for_ctrl);
1441   }
1442   // Done expanding load-reference-barriers.
1443   assert(ShenandoahBarrierSetC2::bsc2()->state()->load_reference_barriers_count() == 0, "all load reference barrier nodes should have been replaced");
1444 
1445   for (int i = state->iu_barriers_count() - 1; i >= 0; i--) {
1446     Node* barrier = state->iu_barrier(i);
1447     Node* pre_val = barrier->in(1);
1448 
1449     if (phase->igvn().type(pre_val)->higher_equal(TypePtr::NULL_PTR)) {
1450       ShouldNotReachHere();
1451       continue;
1452     }
1453 
1454     Node* ctrl = phase->get_ctrl(barrier);
1455 
1456     if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) {
1457       assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0)->in(0), pre_val, ctrl->in(0), phase), "can't move");
1458       ctrl = ctrl->in(0)->in(0);
1459       phase->set_ctrl(barrier, ctrl);
1460     } else if (ctrl->is_CallRuntime()) {
1461       assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0), pre_val, ctrl, phase), "can't move");
1462       ctrl = ctrl->in(0);
1463       phase->set_ctrl(barrier, ctrl);
1464     }
1465 
1466     Node* init_ctrl = ctrl;
1467     IdealLoopTree* loop = phase->get_loop(ctrl);
1468     Node* raw_mem = fixer.find_mem(ctrl, barrier);
1469     Node* init_raw_mem = raw_mem;
1470     Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL);
1471     Node* heap_stable_ctrl = NULL;
1472     Node* null_ctrl = NULL;
1473     uint last = phase->C->unique();
1474 
1475     enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
1476     Node* region = new RegionNode(PATH_LIMIT);
1477     Node* phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1478 
1479     enum { _fast_path = 1, _slow_path, _null_path, PATH_LIMIT2 };
1480     Node* region2 = new RegionNode(PATH_LIMIT2);
1481     Node* phi2 = PhiNode::make(region2, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1482 
1483     // Stable path.
1484     test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, ShenandoahHeap::MARKING);
1485     region->init_req(_heap_stable, heap_stable_ctrl);
1486     phi->init_req(_heap_stable, raw_mem);
1487 
1488     // Null path
1489     Node* reg2_ctrl = NULL;
1490     test_null(ctrl, pre_val, null_ctrl, phase);
1491     if (null_ctrl != NULL) {
1492       reg2_ctrl = null_ctrl->in(0);
1493       region2->init_req(_null_path, null_ctrl);
1494       phi2->init_req(_null_path, raw_mem);
1495     } else {
1496       region2->del_req(_null_path);
1497       phi2->del_req(_null_path);
1498     }
1499 
1500     const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
1501     const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1502     Node* thread = new ThreadLocalNode();
1503     phase->register_new_node(thread, ctrl);
1504     Node* buffer_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(buffer_offset));
1505     phase->register_new_node(buffer_adr, ctrl);
1506     Node* index_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(index_offset));
1507     phase->register_new_node(index_adr, ctrl);
1508 
1509     BasicType index_bt = TypeX_X->basic_type();
1510     assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading Shenandoah SATBMarkQueue::_index with wrong size.");
1511     const TypePtr* adr_type = TypeRawPtr::BOTTOM;
1512     Node* index = new LoadXNode(ctrl, raw_mem, index_adr, adr_type, TypeX_X, MemNode::unordered);
1513     phase->register_new_node(index, ctrl);
1514     Node* index_cmp = new CmpXNode(index, phase->igvn().MakeConX(0));
1515     phase->register_new_node(index_cmp, ctrl);
1516     Node* index_test = new BoolNode(index_cmp, BoolTest::ne);
1517     phase->register_new_node(index_test, ctrl);
1518     IfNode* queue_full_iff = new IfNode(ctrl, index_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
1519     if (reg2_ctrl == NULL) reg2_ctrl = queue_full_iff;
1520     phase->register_control(queue_full_iff, loop, ctrl);
1521     Node* not_full = new IfTrueNode(queue_full_iff);
1522     phase->register_control(not_full, loop, queue_full_iff);
1523     Node* full = new IfFalseNode(queue_full_iff);
1524     phase->register_control(full, loop, queue_full_iff);
1525 
1526     ctrl = not_full;
1527 
1528     Node* next_index = new SubXNode(index, phase->igvn().MakeConX(sizeof(intptr_t)));
1529     phase->register_new_node(next_index, ctrl);
1530 
1531     Node* buffer  = new LoadPNode(ctrl, raw_mem, buffer_adr, adr_type, TypeRawPtr::NOTNULL, MemNode::unordered);
1532     phase->register_new_node(buffer, ctrl);
1533     Node *log_addr = new AddPNode(phase->C->top(), buffer, next_index);
1534     phase->register_new_node(log_addr, ctrl);
1535     Node* log_store = new StorePNode(ctrl, raw_mem, log_addr, adr_type, pre_val, MemNode::unordered);
1536     phase->register_new_node(log_store, ctrl);
1537     // update the index
1538     Node* index_update = new StoreXNode(ctrl, log_store, index_adr, adr_type, next_index, MemNode::unordered);
1539     phase->register_new_node(index_update, ctrl);
1540 
1541     // Fast-path case
1542     region2->init_req(_fast_path, ctrl);
1543     phi2->init_req(_fast_path, index_update);
1544 
1545     ctrl = full;
1546 
1547     Node* base = find_bottom_mem(ctrl, phase);
1548 
1549     MergeMemNode* mm = MergeMemNode::make(base);
1550     mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
1551     phase->register_new_node(mm, ctrl);
1552 
1553     Node* call = new CallLeafNode(ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type(), CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", TypeRawPtr::BOTTOM);
1554     call->init_req(TypeFunc::Control, ctrl);
1555     call->init_req(TypeFunc::I_O, phase->C->top());
1556     call->init_req(TypeFunc::Memory, mm);
1557     call->init_req(TypeFunc::FramePtr, phase->C->top());
1558     call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1559     call->init_req(TypeFunc::Parms, pre_val);
1560     call->init_req(TypeFunc::Parms+1, thread);
1561     phase->register_control(call, loop, ctrl);
1562 
1563     Node* ctrl_proj = new ProjNode(call, TypeFunc::Control);
1564     phase->register_control(ctrl_proj, loop, call);
1565     Node* mem_proj = new ProjNode(call, TypeFunc::Memory);
1566     phase->register_new_node(mem_proj, call);
1567 
1568     // Slow-path case
1569     region2->init_req(_slow_path, ctrl_proj);
1570     phi2->init_req(_slow_path, mem_proj);
1571 
1572     phase->register_control(region2, loop, reg2_ctrl);
1573     phase->register_new_node(phi2, region2);
1574 
1575     region->init_req(_heap_unstable, region2);
1576     phi->init_req(_heap_unstable, phi2);
1577 
1578     phase->register_control(region, loop, heap_stable_ctrl->in(0));
1579     phase->register_new_node(phi, region);
1580 
1581     fix_ctrl(barrier, region, fixer, uses, uses_to_ignore, last, phase);
1582     for(uint next = 0; next < uses.size(); next++ ) {
1583       Node *n = uses.at(next);
1584       assert(phase->get_ctrl(n) == init_ctrl, "bad control");
1585       assert(n != init_raw_mem, "should leave input raw mem above the barrier");
1586       phase->set_ctrl(n, region);
1587       follow_barrier_uses(n, init_ctrl, uses, phase);
1588     }
1589     fixer.fix_mem(init_ctrl, region, init_raw_mem, raw_mem_for_ctrl, phi, uses);
1590 
1591     phase->igvn().replace_node(barrier, pre_val);
1592   }
1593   assert(state->iu_barriers_count() == 0, "all enqueue barrier nodes should have been replaced");
1594 
1595 }
1596 
1597 Node* ShenandoahBarrierC2Support::get_load_addr(PhaseIdealLoop* phase, VectorSet& visited, Node* in) {
1598   if (visited.test_set(in->_idx)) {
1599     return NULL;
1600   }
1601   switch (in->Opcode()) {
1602     case Op_Proj:
1603       return get_load_addr(phase, visited, in->in(0));
1604     case Op_CastPP:
1605     case Op_CheckCastPP:
1606     case Op_DecodeN:
1607     case Op_EncodeP:
1608       return get_load_addr(phase, visited, in->in(1));
1609     case Op_LoadN:
1610     case Op_LoadP:
1611       return in->in(MemNode::Address);
1612     case Op_CompareAndExchangeN:
1613     case Op_CompareAndExchangeP:
1614     case Op_GetAndSetN:
1615     case Op_GetAndSetP:
1616     case Op_ShenandoahCompareAndExchangeP:
1617     case Op_ShenandoahCompareAndExchangeN:
1618       // Those instructions would just have stored a different
1619       // value into the field. No use to attempt to fix it at this point.
1620       return phase->igvn().zerocon(T_OBJECT);
1621     case Op_CMoveP:
1622     case Op_CMoveN: {
1623       Node* t = get_load_addr(phase, visited, in->in(CMoveNode::IfTrue));
1624       Node* f = get_load_addr(phase, visited, in->in(CMoveNode::IfFalse));
1625       // Handle unambiguous cases: single address reported on both branches.
1626       if (t != NULL && f == NULL) return t;
1627       if (t == NULL && f != NULL) return f;
1628       if (t != NULL && t == f)    return t;
1629       // Ambiguity.
1630       return phase->igvn().zerocon(T_OBJECT);
1631     }
1632     case Op_Phi: {
1633       Node* addr = NULL;
1634       for (uint i = 1; i < in->req(); i++) {
1635         Node* addr1 = get_load_addr(phase, visited, in->in(i));
1636         if (addr == NULL) {
1637           addr = addr1;
1638         }
1639         if (addr != addr1) {
1640           return phase->igvn().zerocon(T_OBJECT);
1641         }
1642       }
1643       return addr;
1644     }
1645     case Op_ShenandoahLoadReferenceBarrier:
1646       return get_load_addr(phase, visited, in->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1647     case Op_ShenandoahIUBarrier:
1648       return get_load_addr(phase, visited, in->in(1));
1649     case Op_CallDynamicJava:
1650     case Op_CallLeaf:
1651     case Op_CallStaticJava:
1652     case Op_ConN:
1653     case Op_ConP:
1654     case Op_Parm:
1655     case Op_CreateEx:
1656       return phase->igvn().zerocon(T_OBJECT);
1657     default:
1658 #ifdef ASSERT
1659       fatal("Unknown node in get_load_addr: %s", NodeClassNames[in->Opcode()]);
1660 #endif
1661       return phase->igvn().zerocon(T_OBJECT);
1662   }
1663 
1664 }
1665 
1666 void ShenandoahBarrierC2Support::move_gc_state_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
1667   IdealLoopTree *loop = phase->get_loop(iff);
1668   Node* loop_head = loop->_head;
1669   Node* entry_c = loop_head->in(LoopNode::EntryControl);
1670 
1671   Node* bol = iff->in(1);
1672   Node* cmp = bol->in(1);
1673   Node* andi = cmp->in(1);
1674   Node* load = andi->in(1);
1675 
1676   assert(is_gc_state_load(load), "broken");
1677   if (!phase->is_dominator(load->in(0), entry_c)) {
1678     Node* mem_ctrl = NULL;
1679     Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
1680     load = load->clone();
1681     load->set_req(MemNode::Memory, mem);
1682     load->set_req(0, entry_c);
1683     phase->register_new_node(load, entry_c);
1684     andi = andi->clone();
1685     andi->set_req(1, load);
1686     phase->register_new_node(andi, entry_c);
1687     cmp = cmp->clone();
1688     cmp->set_req(1, andi);
1689     phase->register_new_node(cmp, entry_c);
1690     bol = bol->clone();
1691     bol->set_req(1, cmp);
1692     phase->register_new_node(bol, entry_c);
1693 
1694     phase->igvn().replace_input_of(iff, 1, bol);
1695   }
1696 }
1697 
1698 bool ShenandoahBarrierC2Support::identical_backtoback_ifs(Node* n, PhaseIdealLoop* phase) {
1699   if (!n->is_If() || n->is_CountedLoopEnd()) {
1700     return false;
1701   }
1702   Node* region = n->in(0);
1703 
1704   if (!region->is_Region()) {
1705     return false;
1706   }
1707   Node* dom = phase->idom(region);
1708   if (!dom->is_If()) {
1709     return false;
1710   }
1711 
1712   if (!is_heap_stable_test(n) || !is_heap_stable_test(dom)) {
1713     return false;
1714   }
1715 
1716   IfNode* dom_if = dom->as_If();
1717   Node* proj_true = dom_if->proj_out(1);
1718   Node* proj_false = dom_if->proj_out(0);
1719 
1720   for (uint i = 1; i < region->req(); i++) {
1721     if (phase->is_dominator(proj_true, region->in(i))) {
1722       continue;
1723     }
1724     if (phase->is_dominator(proj_false, region->in(i))) {
1725       continue;
1726     }
1727     return false;
1728   }
1729 
1730   return true;
1731 }
1732 
1733 void ShenandoahBarrierC2Support::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) {
1734   assert(is_heap_stable_test(n), "no other tests");
1735   if (identical_backtoback_ifs(n, phase)) {
1736     Node* n_ctrl = n->in(0);
1737     if (phase->can_split_if(n_ctrl)) {
1738       IfNode* dom_if = phase->idom(n_ctrl)->as_If();
1739       if (is_heap_stable_test(n)) {
1740         Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1);
1741         assert(is_gc_state_load(gc_state_load), "broken");
1742         Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1);
1743         assert(is_gc_state_load(dom_gc_state_load), "broken");
1744         if (gc_state_load != dom_gc_state_load) {
1745           phase->igvn().replace_node(gc_state_load, dom_gc_state_load);
1746         }
1747       }
1748       PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1749       Node* proj_true = dom_if->proj_out(1);
1750       Node* proj_false = dom_if->proj_out(0);
1751       Node* con_true = phase->igvn().makecon(TypeInt::ONE);
1752       Node* con_false = phase->igvn().makecon(TypeInt::ZERO);
1753 
1754       for (uint i = 1; i < n_ctrl->req(); i++) {
1755         if (phase->is_dominator(proj_true, n_ctrl->in(i))) {
1756           bolphi->init_req(i, con_true);
1757         } else {
1758           assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1759           bolphi->init_req(i, con_false);
1760         }
1761       }
1762       phase->register_new_node(bolphi, n_ctrl);
1763       phase->igvn().replace_input_of(n, 1, bolphi);
1764       phase->do_split_if(n);
1765     }
1766   }
1767 }
1768 
1769 IfNode* ShenandoahBarrierC2Support::find_unswitching_candidate(const IdealLoopTree* loop, PhaseIdealLoop* phase) {
1770   // Find first invariant test that doesn't exit the loop
1771   LoopNode *head = loop->_head->as_Loop();
1772   IfNode* unswitch_iff = NULL;
1773   Node* n = head->in(LoopNode::LoopBackControl);
1774   int loop_has_sfpts = -1;
1775   while (n != head) {
1776     Node* n_dom = phase->idom(n);
1777     if (n->is_Region()) {
1778       if (n_dom->is_If()) {
1779         IfNode* iff = n_dom->as_If();
1780         if (iff->in(1)->is_Bool()) {
1781           BoolNode* bol = iff->in(1)->as_Bool();
1782           if (bol->in(1)->is_Cmp()) {
1783             // If condition is invariant and not a loop exit,
1784             // then found reason to unswitch.
1785             if (is_heap_stable_test(iff) &&
1786                 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) {
1787               assert(!loop->is_loop_exit(iff), "both branches should be in the loop");
1788               if (loop_has_sfpts == -1) {
1789                 for(uint i = 0; i < loop->_body.size(); i++) {
1790                   Node *m = loop->_body[i];
1791                   if (m->is_SafePoint() && !m->is_CallLeaf()) {
1792                     loop_has_sfpts = 1;
1793                     break;
1794                   }
1795                 }
1796                 if (loop_has_sfpts == -1) {
1797                   loop_has_sfpts = 0;
1798                 }
1799               }
1800               if (!loop_has_sfpts) {
1801                 unswitch_iff = iff;
1802               }
1803             }
1804           }
1805         }
1806       }
1807     }
1808     n = n_dom;
1809   }
1810   return unswitch_iff;
1811 }
1812 
1813 
1814 void ShenandoahBarrierC2Support::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) {
1815   Node_List heap_stable_tests;
1816   stack.push(phase->C->start(), 0);
1817   do {
1818     Node* n = stack.node();
1819     uint i = stack.index();
1820 
1821     if (i < n->outcnt()) {
1822       Node* u = n->raw_out(i);
1823       stack.set_index(i+1);
1824       if (!visited.test_set(u->_idx)) {
1825         stack.push(u, 0);
1826       }
1827     } else {
1828       stack.pop();
1829       if (n->is_If() && is_heap_stable_test(n)) {
1830         heap_stable_tests.push(n);
1831       }
1832     }
1833   } while (stack.size() > 0);
1834 
1835   for (uint i = 0; i < heap_stable_tests.size(); i++) {
1836     Node* n = heap_stable_tests.at(i);
1837     assert(is_heap_stable_test(n), "only evacuation test");
1838     merge_back_to_back_tests(n, phase);
1839   }
1840 
1841   if (!phase->C->major_progress()) {
1842     VectorSet seen;
1843     for (uint i = 0; i < heap_stable_tests.size(); i++) {
1844       Node* n = heap_stable_tests.at(i);
1845       IdealLoopTree* loop = phase->get_loop(n);
1846       if (loop != phase->ltree_root() &&
1847           loop->_child == NULL &&
1848           !loop->_irreducible) {
1849         Node* head = loop->_head;
1850         if (head->is_Loop() &&
1851             (!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) &&
1852             !seen.test_set(head->_idx)) {
1853           IfNode* iff = find_unswitching_candidate(loop, phase);
1854           if (iff != NULL) {
1855             Node* bol = iff->in(1);
1856             if (head->as_Loop()->is_strip_mined()) {
1857               head->as_Loop()->verify_strip_mined(0);
1858             }
1859             move_gc_state_test_out_of_loop(iff, phase);
1860 
1861             AutoNodeBudget node_budget(phase);
1862 
1863             if (loop->policy_unswitching(phase)) {
1864               if (head->as_Loop()->is_strip_mined()) {
1865                 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop();
1866                 hide_strip_mined_loop(outer, head->as_CountedLoop(), phase);
1867               }
1868               phase->do_unswitching(loop, old_new);
1869             } else {
1870               // Not proceeding with unswitching. Move load back in
1871               // the loop.
1872               phase->igvn().replace_input_of(iff, 1, bol);
1873             }
1874           }
1875         }
1876       }
1877     }
1878   }
1879 }
1880 
1881 ShenandoahIUBarrierNode::ShenandoahIUBarrierNode(Node* val) : Node(NULL, val) {
1882   ShenandoahBarrierSetC2::bsc2()->state()->add_iu_barrier(this);
1883 }
1884 
1885 const Type* ShenandoahIUBarrierNode::bottom_type() const {
1886   if (in(1) == NULL || in(1)->is_top()) {
1887     return Type::TOP;
1888   }
1889   const Type* t = in(1)->bottom_type();
1890   if (t == TypePtr::NULL_PTR) {
1891     return t;
1892   }
1893   return t->is_oopptr();
1894 }
1895 
1896 const Type* ShenandoahIUBarrierNode::Value(PhaseGVN* phase) const {
1897   if (in(1) == NULL) {
1898     return Type::TOP;
1899   }
1900   const Type* t = phase->type(in(1));
1901   if (t == Type::TOP) {
1902     return Type::TOP;
1903   }
1904   if (t == TypePtr::NULL_PTR) {
1905     return t;
1906   }
1907   return t->is_oopptr();
1908 }
1909 
1910 int ShenandoahIUBarrierNode::needed(Node* n) {
1911   if (n == NULL ||
1912       n->is_Allocate() ||
1913       n->Opcode() == Op_ShenandoahIUBarrier ||
1914       n->bottom_type() == TypePtr::NULL_PTR ||
1915       (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL)) {
1916     return NotNeeded;
1917   }
1918   if (n->is_Phi() ||
1919       n->is_CMove()) {
1920     return MaybeNeeded;
1921   }
1922   return Needed;
1923 }
1924 
1925 Node* ShenandoahIUBarrierNode::next(Node* n) {
1926   for (;;) {
1927     if (n == NULL) {
1928       return n;
1929     } else if (n->bottom_type() == TypePtr::NULL_PTR) {
1930       return n;
1931     } else if (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL) {
1932       return n;
1933     } else if (n->is_ConstraintCast() ||
1934                n->Opcode() == Op_DecodeN ||
1935                n->Opcode() == Op_EncodeP) {
1936       n = n->in(1);
1937     } else if (n->is_Proj()) {
1938       n = n->in(0);
1939     } else {
1940       return n;
1941     }
1942   }
1943   ShouldNotReachHere();
1944   return NULL;
1945 }
1946 
1947 Node* ShenandoahIUBarrierNode::Identity(PhaseGVN* phase) {
1948   PhaseIterGVN* igvn = phase->is_IterGVN();
1949 
1950   Node* n = next(in(1));
1951 
1952   int cont = needed(n);
1953 
1954   if (cont == NotNeeded) {
1955     return in(1);
1956   } else if (cont == MaybeNeeded) {
1957     if (igvn == NULL) {
1958       phase->record_for_igvn(this);
1959       return this;
1960     } else {
1961       ResourceMark rm;
1962       Unique_Node_List wq;
1963       uint wq_i = 0;
1964 
1965       for (;;) {
1966         if (n->is_Phi()) {
1967           for (uint i = 1; i < n->req(); i++) {
1968             Node* m = n->in(i);
1969             if (m != NULL) {
1970               wq.push(m);
1971             }
1972           }
1973         } else {
1974           assert(n->is_CMove(), "nothing else here");
1975           Node* m = n->in(CMoveNode::IfFalse);
1976           wq.push(m);
1977           m = n->in(CMoveNode::IfTrue);
1978           wq.push(m);
1979         }
1980         Node* orig_n = NULL;
1981         do {
1982           if (wq_i >= wq.size()) {
1983             return in(1);
1984           }
1985           n = wq.at(wq_i);
1986           wq_i++;
1987           orig_n = n;
1988           n = next(n);
1989           cont = needed(n);
1990           if (cont == Needed) {
1991             return this;
1992           }
1993         } while (cont != MaybeNeeded || (orig_n != n && wq.member(n)));
1994       }
1995     }
1996   }
1997 
1998   return this;
1999 }
2000 
2001 #ifdef ASSERT
2002 static bool has_never_branch(Node* root) {
2003   for (uint i = 1; i < root->req(); i++) {
2004     Node* in = root->in(i);
2005     if (in != NULL && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->Opcode() == Op_NeverBranch) {
2006       return true;
2007     }
2008   }
2009   return false;
2010 }
2011 #endif
2012 
2013 void MemoryGraphFixer::collect_memory_nodes() {
2014   Node_Stack stack(0);
2015   VectorSet visited;
2016   Node_List regions;
2017 
2018   // Walk the raw memory graph and create a mapping from CFG node to
2019   // memory node. Exclude phis for now.
2020   stack.push(_phase->C->root(), 1);
2021   do {
2022     Node* n = stack.node();
2023     int opc = n->Opcode();
2024     uint i = stack.index();
2025     if (i < n->req()) {
2026       Node* mem = NULL;
2027       if (opc == Op_Root) {
2028         Node* in = n->in(i);
2029         int in_opc = in->Opcode();
2030         if (in_opc == Op_Return || in_opc == Op_Rethrow) {
2031           mem = in->in(TypeFunc::Memory);
2032         } else if (in_opc == Op_Halt) {
2033           if (in->in(0)->is_Region()) {
2034             Node* r = in->in(0);
2035             for (uint j = 1; j < r->req(); j++) {
2036               assert(r->in(j)->Opcode() != Op_NeverBranch, "");
2037             }
2038           } else {
2039             Node* proj = in->in(0);
2040             assert(proj->is_Proj(), "");
2041             Node* in = proj->in(0);
2042             assert(in->is_CallStaticJava() || in->Opcode() == Op_NeverBranch || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
2043             if (in->is_CallStaticJava()) {
2044               mem = in->in(TypeFunc::Memory);
2045             } else if (in->Opcode() == Op_Catch) {
2046               Node* call = in->in(0)->in(0);
2047               assert(call->is_Call(), "");
2048               mem = call->in(TypeFunc::Memory);
2049             } else if (in->Opcode() == Op_NeverBranch) {
2050               mem = collect_memory_for_infinite_loop(in);
2051             }
2052           }
2053         } else {
2054 #ifdef ASSERT
2055           n->dump();
2056           in->dump();
2057 #endif
2058           ShouldNotReachHere();
2059         }
2060       } else {
2061         assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
2062         assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
2063         mem = n->in(i);
2064       }
2065       i++;
2066       stack.set_index(i);
2067       if (mem == NULL) {
2068         continue;
2069       }
2070       for (;;) {
2071         if (visited.test_set(mem->_idx) || mem->is_Start()) {
2072           break;
2073         }
2074         if (mem->is_Phi()) {
2075           stack.push(mem, 2);
2076           mem = mem->in(1);
2077         } else if (mem->is_Proj()) {
2078           stack.push(mem, mem->req());
2079           mem = mem->in(0);
2080         } else if (mem->is_SafePoint() || mem->is_MemBar()) {
2081           mem = mem->in(TypeFunc::Memory);
2082         } else if (mem->is_MergeMem()) {
2083           MergeMemNode* mm = mem->as_MergeMem();
2084           mem = mm->memory_at(_alias);
2085         } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
2086           assert(_alias == Compile::AliasIdxRaw, "");
2087           stack.push(mem, mem->req());
2088           mem = mem->in(MemNode::Memory);
2089         } else {
2090 #ifdef ASSERT
2091           mem->dump();
2092 #endif
2093           ShouldNotReachHere();
2094         }
2095       }
2096     } else {
2097       if (n->is_Phi()) {
2098         // Nothing
2099       } else if (!n->is_Root()) {
2100         Node* c = get_ctrl(n);
2101         _memory_nodes.map(c->_idx, n);
2102       }
2103       stack.pop();
2104     }
2105   } while(stack.is_nonempty());
2106 
2107   // Iterate over CFG nodes in rpo and propagate memory state to
2108   // compute memory state at regions, creating new phis if needed.
2109   Node_List rpo_list;
2110   visited.clear();
2111   _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
2112   Node* root = rpo_list.pop();
2113   assert(root == _phase->C->root(), "");
2114 
2115   const bool trace = false;
2116 #ifdef ASSERT
2117   if (trace) {
2118     for (int i = rpo_list.size() - 1; i >= 0; i--) {
2119       Node* c = rpo_list.at(i);
2120       if (_memory_nodes[c->_idx] != NULL) {
2121         tty->print("X %d", c->_idx);  _memory_nodes[c->_idx]->dump();
2122       }
2123     }
2124   }
2125 #endif
2126   uint last = _phase->C->unique();
2127 
2128 #ifdef ASSERT
2129   uint16_t max_depth = 0;
2130   for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
2131     IdealLoopTree* lpt = iter.current();
2132     max_depth = MAX2(max_depth, lpt->_nest);
2133   }
2134 #endif
2135 
2136   bool progress = true;
2137   int iteration = 0;
2138   Node_List dead_phis;
2139   while (progress) {
2140     progress = false;
2141     iteration++;
2142     assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2143     if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
2144 
2145     for (int i = rpo_list.size() - 1; i >= 0; i--) {
2146       Node* c = rpo_list.at(i);
2147 
2148       Node* prev_mem = _memory_nodes[c->_idx];
2149       if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2150         Node* prev_region = regions[c->_idx];
2151         Node* unique = NULL;
2152         for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
2153           Node* m = _memory_nodes[c->in(j)->_idx];
2154           assert(m != NULL || (c->is_Loop() && j == LoopNode::LoopBackControl && iteration == 1) || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "expect memory state");
2155           if (m != NULL) {
2156             if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
2157               assert(c->is_Loop() && j == LoopNode::LoopBackControl || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2158               // continue
2159             } else if (unique == NULL) {
2160               unique = m;
2161             } else if (m == unique) {
2162               // continue
2163             } else {
2164               unique = NodeSentinel;
2165             }
2166           }
2167         }
2168         assert(unique != NULL, "empty phi???");
2169         if (unique != NodeSentinel) {
2170           if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c) {
2171             dead_phis.push(prev_region);
2172           }
2173           regions.map(c->_idx, unique);
2174         } else {
2175           Node* phi = NULL;
2176           if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
2177             phi = prev_region;
2178             for (uint k = 1; k < c->req(); k++) {
2179               Node* m = _memory_nodes[c->in(k)->_idx];
2180               assert(m != NULL, "expect memory state");
2181               phi->set_req(k, m);
2182             }
2183           } else {
2184             for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == NULL; j++) {
2185               Node* u = c->fast_out(j);
2186               if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2187                   (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
2188                 phi = u;
2189                 for (uint k = 1; k < c->req() && phi != NULL; k++) {
2190                   Node* m = _memory_nodes[c->in(k)->_idx];
2191                   assert(m != NULL, "expect memory state");
2192                   if (u->in(k) != m) {
2193                     phi = NodeSentinel;
2194                   }
2195                 }
2196               }
2197             }
2198             if (phi == NodeSentinel) {
2199               phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
2200               for (uint k = 1; k < c->req(); k++) {
2201                 Node* m = _memory_nodes[c->in(k)->_idx];
2202                 assert(m != NULL, "expect memory state");
2203                 phi->init_req(k, m);
2204               }
2205             }
2206           }
2207           if (phi != NULL) {
2208             regions.map(c->_idx, phi);
2209           } else {
2210             assert(c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
2211           }
2212         }
2213         Node* current_region = regions[c->_idx];
2214         if (current_region != prev_region) {
2215           progress = true;
2216           if (prev_region == prev_mem) {
2217             _memory_nodes.map(c->_idx, current_region);
2218           }
2219         }
2220       } else if (prev_mem == NULL || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
2221         Node* m = _memory_nodes[_phase->idom(c)->_idx];
2222         assert(m != NULL || c->Opcode() == Op_Halt, "expect memory state");
2223         if (m != prev_mem) {
2224           _memory_nodes.map(c->_idx, m);
2225           progress = true;
2226         }
2227       }
2228 #ifdef ASSERT
2229       if (trace) { tty->print("X %d", c->_idx);  _memory_nodes[c->_idx]->dump(); }
2230 #endif
2231     }
2232   }
2233 
2234   // Replace existing phi with computed memory state for that region
2235   // if different (could be a new phi or a dominating memory node if
2236   // that phi was found to be useless).
2237   while (dead_phis.size() > 0) {
2238     Node* n = dead_phis.pop();
2239     n->replace_by(_phase->C->top());
2240     n->destruct(&_phase->igvn());
2241   }
2242   for (int i = rpo_list.size() - 1; i >= 0; i--) {
2243     Node* c = rpo_list.at(i);
2244     if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2245       Node* n = regions[c->_idx];
2246       assert(n != NULL || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
2247       if (n != NULL && n->is_Phi() && n->_idx >= last && n->in(0) == c) {
2248         _phase->register_new_node(n, c);
2249       }
2250     }
2251   }
2252   for (int i = rpo_list.size() - 1; i >= 0; i--) {
2253     Node* c = rpo_list.at(i);
2254     if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2255       Node* n = regions[c->_idx];
2256       assert(n != NULL || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
2257       for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
2258         Node* u = c->fast_out(i);
2259         if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2260             u != n) {
2261           assert(c->unique_ctrl_out()->Opcode() != Op_Halt, "expected memory state");
2262           if (u->adr_type() == TypePtr::BOTTOM) {
2263             fix_memory_uses(u, n, n, c);
2264           } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2265             _phase->lazy_replace(u, n);
2266             --i; --imax;
2267           }
2268         }
2269       }
2270     }
2271   }
2272 }
2273 
2274 Node* MemoryGraphFixer::collect_memory_for_infinite_loop(const Node* in) {
2275   Node* mem = NULL;
2276   Node* head = in->in(0);
2277   assert(head->is_Region(), "unexpected infinite loop graph shape");
2278 
2279   Node* phi_mem = NULL;
2280   for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2281     Node* u = head->fast_out(j);
2282     if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
2283       if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2284         assert(phi_mem == NULL || phi_mem->adr_type() == TypePtr::BOTTOM, "");
2285         phi_mem = u;
2286       } else if (u->adr_type() == TypePtr::BOTTOM) {
2287         assert(phi_mem == NULL || _phase->C->get_alias_index(phi_mem->adr_type()) == _alias, "");
2288         if (phi_mem == NULL) {
2289           phi_mem = u;
2290         }
2291       }
2292     }
2293   }
2294   if (phi_mem == NULL) {
2295     ResourceMark rm;
2296     Node_Stack stack(0);
2297     stack.push(head, 1);
2298     do {
2299       Node* n = stack.node();
2300       uint i = stack.index();
2301       if (i >= n->req()) {
2302         stack.pop();
2303       } else {
2304         stack.set_index(i + 1);
2305         Node* c = n->in(i);
2306         assert(c != head, "should have found a safepoint on the way");
2307         if (stack.size() != 1 || _phase->is_dominator(head, c)) {
2308           for (;;) {
2309             if (c->is_Region()) {
2310               stack.push(c, 1);
2311               break;
2312             } else if (c->is_SafePoint() && !c->is_CallLeaf()) {
2313               Node* m = c->in(TypeFunc::Memory);
2314               if (m->is_MergeMem()) {
2315                 m = m->as_MergeMem()->memory_at(_alias);
2316               }
2317               assert(mem == NULL || mem == m, "several memory states");
2318               mem = m;
2319               break;
2320             } else {
2321               assert(c != c->in(0), "");
2322               c = c->in(0);
2323             }
2324           }
2325         }
2326       }
2327     } while (stack.size() > 0);
2328     assert(mem != NULL, "should have found safepoint");
2329   } else {
2330     mem = phi_mem;
2331   }
2332   return mem;
2333 }
2334 
2335 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
2336   Node* c = _phase->get_ctrl(n);
2337   if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Call()) {
2338     assert(c == n->in(0), "");
2339     CallNode* call = c->as_Call();
2340     CallProjections projs;
2341     call->extract_projections(&projs, true, false);
2342     if (projs.catchall_memproj != NULL) {
2343       if (projs.fallthrough_memproj == n) {
2344         c = projs.fallthrough_catchproj;
2345       } else {
2346         assert(projs.catchall_memproj == n, "");
2347         c = projs.catchall_catchproj;
2348       }
2349     }
2350   }
2351   return c;
2352 }
2353 
2354 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
2355   if (_phase->has_ctrl(n))
2356     return get_ctrl(n);
2357   else {
2358     assert (n->is_CFG(), "must be a CFG node");
2359     return n;
2360   }
2361 }
2362 
2363 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
2364   return m != NULL && get_ctrl(m) == c;
2365 }
2366 
2367 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
2368   assert(n == NULL || _phase->ctrl_or_self(n) == ctrl, "");
2369   assert(!ctrl->is_Call() || ctrl == n, "projection expected");
2370 #ifdef ASSERT
2371   if ((ctrl->is_Proj() && ctrl->in(0)->is_Call()) ||
2372       (ctrl->is_Catch() && ctrl->in(0)->in(0)->is_Call())) {
2373     CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_Call() : ctrl->in(0)->in(0)->as_Call();
2374     int mems = 0;
2375     for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
2376       Node* u = call->fast_out(i);
2377       if (u->bottom_type() == Type::MEMORY) {
2378         mems++;
2379       }
2380     }
2381     assert(mems <= 1, "No node right after call if multiple mem projections");
2382   }
2383 #endif
2384   Node* mem = _memory_nodes[ctrl->_idx];
2385   Node* c = ctrl;
2386   while (!mem_is_valid(mem, c) &&
2387          (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
2388     c = _phase->idom(c);
2389     mem = _memory_nodes[c->_idx];
2390   }
2391   if (n != NULL && mem_is_valid(mem, c)) {
2392     while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
2393       mem = next_mem(mem, _alias);
2394     }
2395     if (mem->is_MergeMem()) {
2396       mem = mem->as_MergeMem()->memory_at(_alias);
2397     }
2398     if (!mem_is_valid(mem, c)) {
2399       do {
2400         c = _phase->idom(c);
2401         mem = _memory_nodes[c->_idx];
2402       } while (!mem_is_valid(mem, c) &&
2403                (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
2404     }
2405   }
2406   assert(mem->bottom_type() == Type::MEMORY, "");
2407   return mem;
2408 }
2409 
2410 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
2411   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2412     Node* use = region->fast_out(i);
2413     if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
2414         (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
2415       return true;
2416     }
2417   }
2418   return false;
2419 }
2420 
2421 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
2422   assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
2423   const bool trace = false;
2424   DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
2425   DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
2426   GrowableArray<Node*> phis;
2427   if (mem_for_ctrl != mem) {
2428     Node* old = mem_for_ctrl;
2429     Node* prev = NULL;
2430     while (old != mem) {
2431       prev = old;
2432       if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
2433         assert(_alias == Compile::AliasIdxRaw, "");
2434         old = old->in(MemNode::Memory);
2435       } else if (old->Opcode() == Op_SCMemProj) {
2436         assert(_alias == Compile::AliasIdxRaw, "");
2437         old = old->in(0);
2438       } else {
2439         ShouldNotReachHere();
2440       }
2441     }
2442     assert(prev != NULL, "");
2443     if (new_ctrl != ctrl) {
2444       _memory_nodes.map(ctrl->_idx, mem);
2445       _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2446     }
2447     uint input = (uint)MemNode::Memory;
2448     _phase->igvn().replace_input_of(prev, input, new_mem);
2449   } else {
2450     uses.clear();
2451     _memory_nodes.map(new_ctrl->_idx, new_mem);
2452     uses.push(new_ctrl);
2453     for(uint next = 0; next < uses.size(); next++ ) {
2454       Node *n = uses.at(next);
2455       assert(n->is_CFG(), "");
2456       DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
2457       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2458         Node* u = n->fast_out(i);
2459         if (!u->is_Root() && u->is_CFG() && u != n) {
2460           Node* m = _memory_nodes[u->_idx];
2461           if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) &&
2462               !has_mem_phi(u) &&
2463               u->unique_ctrl_out()->Opcode() != Op_Halt) {
2464             DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
2465             DEBUG_ONLY(if (trace && m != NULL) { tty->print("ZZZ mem"); m->dump(); });
2466 
2467             if (!mem_is_valid(m, u) || !m->is_Phi()) {
2468               bool push = true;
2469               bool create_phi = true;
2470               if (_phase->is_dominator(new_ctrl, u)) {
2471                 create_phi = false;
2472               }
2473               if (create_phi) {
2474                 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
2475                 _phase->register_new_node(phi, u);
2476                 phis.push(phi);
2477                 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
2478                 if (!mem_is_valid(m, u)) {
2479                   DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
2480                   _memory_nodes.map(u->_idx, phi);
2481                 } else {
2482                   DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
2483                   for (;;) {
2484                     assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), "");
2485                     Node* next = NULL;
2486                     if (m->is_Proj()) {
2487                       next = m->in(0);
2488                     } else {
2489                       assert(m->is_Mem() || m->is_LoadStore(), "");
2490                       assert(_alias == Compile::AliasIdxRaw, "");
2491                       next = m->in(MemNode::Memory);
2492                     }
2493                     if (_phase->get_ctrl(next) != u) {
2494                       break;
2495                     }
2496                     if (next->is_MergeMem()) {
2497                       assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
2498                       break;
2499                     }
2500                     if (next->is_Phi()) {
2501                       assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
2502                       break;
2503                     }
2504                     m = next;
2505                   }
2506 
2507                   DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
2508                   assert(m->is_Mem() || m->is_LoadStore(), "");
2509                   uint input = (uint)MemNode::Memory;
2510                   _phase->igvn().replace_input_of(m, input, phi);
2511                   push = false;
2512                 }
2513               } else {
2514                 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
2515               }
2516               if (push) {
2517                 uses.push(u);
2518               }
2519             }
2520           } else if (!mem_is_valid(m, u) &&
2521                      !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1)) {
2522             uses.push(u);
2523           }
2524         }
2525       }
2526     }
2527     for (int i = 0; i < phis.length(); i++) {
2528       Node* n = phis.at(i);
2529       Node* r = n->in(0);
2530       DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
2531       for (uint j = 1; j < n->req(); j++) {
2532         Node* m = find_mem(r->in(j), NULL);
2533         _phase->igvn().replace_input_of(n, j, m);
2534         DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
2535       }
2536     }
2537   }
2538   uint last = _phase->C->unique();
2539   MergeMemNode* mm = NULL;
2540   int alias = _alias;
2541   DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
2542   // Process loads first to not miss an anti-dependency: if the memory
2543   // edge of a store is updated before a load is processed then an
2544   // anti-dependency may be missed.
2545   for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2546     Node* u = mem->out(i);
2547     if (u->_idx < last && u->is_Load() && _phase->C->get_alias_index(u->adr_type()) == alias) {
2548       Node* m = find_mem(_phase->get_ctrl(u), u);
2549       if (m != mem) {
2550         DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2551         _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2552         --i;
2553       }
2554     }
2555   }
2556   for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2557     Node* u = mem->out(i);
2558     if (u->_idx < last) {
2559       if (u->is_Mem()) {
2560         if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2561           Node* m = find_mem(_phase->get_ctrl(u), u);
2562           if (m != mem) {
2563             DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2564             _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2565             --i;
2566           }
2567         }
2568       } else if (u->is_MergeMem()) {
2569         MergeMemNode* u_mm = u->as_MergeMem();
2570         if (u_mm->memory_at(alias) == mem) {
2571           MergeMemNode* newmm = NULL;
2572           for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2573             Node* uu = u->fast_out(j);
2574             assert(!uu->is_MergeMem(), "chain of MergeMems?");
2575             if (uu->is_Phi()) {
2576               assert(uu->adr_type() == TypePtr::BOTTOM, "");
2577               Node* region = uu->in(0);
2578               int nb = 0;
2579               for (uint k = 1; k < uu->req(); k++) {
2580                 if (uu->in(k) == u) {
2581                   Node* m = find_mem(region->in(k), NULL);
2582                   if (m != mem) {
2583                     DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
2584                     newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2585                     if (newmm != u) {
2586                       _phase->igvn().replace_input_of(uu, k, newmm);
2587                       nb++;
2588                       --jmax;
2589                     }
2590                   }
2591                 }
2592               }
2593               if (nb > 0) {
2594                 --j;
2595               }
2596             } else {
2597               Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
2598               if (m != mem) {
2599                 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
2600                 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2601                 if (newmm != u) {
2602                   _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2603                   --j, --jmax;
2604                 }
2605               }
2606             }
2607           }
2608         }
2609       } else if (u->is_Phi()) {
2610         assert(u->bottom_type() == Type::MEMORY, "what else?");
2611         if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
2612           Node* region = u->in(0);
2613           bool replaced = false;
2614           for (uint j = 1; j < u->req(); j++) {
2615             if (u->in(j) == mem) {
2616               Node* m = find_mem(region->in(j), NULL);
2617               Node* nnew = m;
2618               if (m != mem) {
2619                 if (u->adr_type() == TypePtr::BOTTOM) {
2620                   mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
2621                   nnew = mm;
2622                 }
2623                 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
2624                 _phase->igvn().replace_input_of(u, j, nnew);
2625                 replaced = true;
2626               }
2627             }
2628           }
2629           if (replaced) {
2630             --i;
2631           }
2632         }
2633       } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2634                  u->adr_type() == NULL) {
2635         assert(u->adr_type() != NULL ||
2636                u->Opcode() == Op_Rethrow ||
2637                u->Opcode() == Op_Return ||
2638                u->Opcode() == Op_SafePoint ||
2639                (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2640                (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2641                u->Opcode() == Op_CallLeaf, "");
2642         Node* m = find_mem(_phase->ctrl_or_self(u), u);
2643         if (m != mem) {
2644           mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
2645           _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2646           --i;
2647         }
2648       } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2649         Node* m = find_mem(_phase->ctrl_or_self(u), u);
2650         if (m != mem) {
2651           DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2652           _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2653           --i;
2654         }
2655       } else if (u->adr_type() != TypePtr::BOTTOM &&
2656                  _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
2657         Node* m = find_mem(_phase->ctrl_or_self(u), u);
2658         assert(m != mem, "");
2659         // u is on the wrong slice...
2660         assert(u->is_ClearArray(), "");
2661         DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2662         _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2663         --i;
2664       }
2665     }
2666   }
2667 #ifdef ASSERT
2668   assert(new_mem->outcnt() > 0, "");
2669   for (int i = 0; i < phis.length(); i++) {
2670     Node* n = phis.at(i);
2671     assert(n->outcnt() > 0, "new phi must have uses now");
2672   }
2673 #endif
2674 }
2675 
2676 void MemoryGraphFixer::record_new_ctrl(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl) {
2677   if (mem_for_ctrl != mem && new_ctrl != ctrl) {
2678     _memory_nodes.map(ctrl->_idx, mem);
2679     _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2680   }
2681 }
2682 
2683 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
2684   MergeMemNode* mm = MergeMemNode::make(mem);
2685   mm->set_memory_at(_alias, rep_proj);
2686   _phase->register_new_node(mm, rep_ctrl);
2687   return mm;
2688 }
2689 
2690 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
2691   MergeMemNode* newmm = NULL;
2692   MergeMemNode* u_mm = u->as_MergeMem();
2693   Node* c = _phase->get_ctrl(u);
2694   if (_phase->is_dominator(c, rep_ctrl)) {
2695     c = rep_ctrl;
2696   } else {
2697     assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
2698   }
2699   if (u->outcnt() == 1) {
2700     if (u->req() > (uint)_alias && u->in(_alias) == mem) {
2701       _phase->igvn().replace_input_of(u, _alias, rep_proj);
2702       --i;
2703     } else {
2704       _phase->igvn().rehash_node_delayed(u);
2705       u_mm->set_memory_at(_alias, rep_proj);
2706     }
2707     newmm = u_mm;
2708     _phase->set_ctrl_and_loop(u, c);
2709   } else {
2710     // can't simply clone u and then change one of its input because
2711     // it adds and then removes an edge which messes with the
2712     // DUIterator
2713     newmm = MergeMemNode::make(u_mm->base_memory());
2714     for (uint j = 0; j < u->req(); j++) {
2715       if (j < newmm->req()) {
2716         if (j == (uint)_alias) {
2717           newmm->set_req(j, rep_proj);
2718         } else if (newmm->in(j) != u->in(j)) {
2719           newmm->set_req(j, u->in(j));
2720         }
2721       } else if (j == (uint)_alias) {
2722         newmm->add_req(rep_proj);
2723       } else {
2724         newmm->add_req(u->in(j));
2725       }
2726     }
2727     if ((uint)_alias >= u->req()) {
2728       newmm->set_memory_at(_alias, rep_proj);
2729     }
2730     _phase->register_new_node(newmm, c);
2731   }
2732   return newmm;
2733 }
2734 
2735 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
2736   if (phi->adr_type() == TypePtr::BOTTOM) {
2737     Node* region = phi->in(0);
2738     for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
2739       Node* uu = region->fast_out(j);
2740       if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
2741         return false;
2742       }
2743     }
2744     return true;
2745   }
2746   return _phase->C->get_alias_index(phi->adr_type()) == _alias;
2747 }
2748 
2749 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
2750   uint last = _phase-> C->unique();
2751   MergeMemNode* mm = NULL;
2752   assert(mem->bottom_type() == Type::MEMORY, "");
2753   for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2754     Node* u = mem->out(i);
2755     if (u != replacement && u->_idx < last) {
2756       if (u->is_MergeMem()) {
2757         MergeMemNode* u_mm = u->as_MergeMem();
2758         if (u_mm->memory_at(_alias) == mem) {
2759           MergeMemNode* newmm = NULL;
2760           for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2761             Node* uu = u->fast_out(j);
2762             assert(!uu->is_MergeMem(), "chain of MergeMems?");
2763             if (uu->is_Phi()) {
2764               if (should_process_phi(uu)) {
2765                 Node* region = uu->in(0);
2766                 int nb = 0;
2767                 for (uint k = 1; k < uu->req(); k++) {
2768                   if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
2769                     if (newmm == NULL) {
2770                       newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2771                     }
2772                     if (newmm != u) {
2773                       _phase->igvn().replace_input_of(uu, k, newmm);
2774                       nb++;
2775                       --jmax;
2776                     }
2777                   }
2778                 }
2779                 if (nb > 0) {
2780                   --j;
2781                 }
2782               }
2783             } else {
2784               if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
2785                 if (newmm == NULL) {
2786                   newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2787                 }
2788                 if (newmm != u) {
2789                   _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2790                   --j, --jmax;
2791                 }
2792               }
2793             }
2794           }
2795         }
2796       } else if (u->is_Phi()) {
2797         assert(u->bottom_type() == Type::MEMORY, "what else?");
2798         Node* region = u->in(0);
2799         if (should_process_phi(u)) {
2800           bool replaced = false;
2801           for (uint j = 1; j < u->req(); j++) {
2802             if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
2803               Node* nnew = rep_proj;
2804               if (u->adr_type() == TypePtr::BOTTOM) {
2805                 if (mm == NULL) {
2806                   mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2807                 }
2808                 nnew = mm;
2809               }
2810               _phase->igvn().replace_input_of(u, j, nnew);
2811               replaced = true;
2812             }
2813           }
2814           if (replaced) {
2815             --i;
2816           }
2817 
2818         }
2819       } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2820                  u->adr_type() == NULL) {
2821         assert(u->adr_type() != NULL ||
2822                u->Opcode() == Op_Rethrow ||
2823                u->Opcode() == Op_Return ||
2824                u->Opcode() == Op_SafePoint ||
2825                (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2826                (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2827                u->Opcode() == Op_CallLeaf, "%s", u->Name());
2828         if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2829           if (mm == NULL) {
2830             mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2831           }
2832           _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2833           --i;
2834         }
2835       } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2836         if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2837           _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
2838           --i;
2839         }
2840       }
2841     }
2842   }
2843 }
2844 
2845 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj, DecoratorSet decorators)
2846 : Node(ctrl, obj), _decorators(decorators) {
2847   ShenandoahBarrierSetC2::bsc2()->state()->add_load_reference_barrier(this);
2848 }
2849 
2850 DecoratorSet ShenandoahLoadReferenceBarrierNode::decorators() const {
2851   return _decorators;
2852 }
2853 
2854 uint ShenandoahLoadReferenceBarrierNode::size_of() const {
2855   return sizeof(*this);
2856 }
2857 
2858 static DecoratorSet mask_decorators(DecoratorSet decorators) {
2859   return decorators & (ON_STRONG_OOP_REF | ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF | ON_UNKNOWN_OOP_REF | IN_NATIVE);
2860 }
2861 
2862 uint ShenandoahLoadReferenceBarrierNode::hash() const {
2863   uint hash = Node::hash();
2864   hash += mask_decorators(_decorators);
2865   return hash;
2866 }
2867 
2868 bool ShenandoahLoadReferenceBarrierNode::cmp( const Node &n ) const {
2869   return Node::cmp(n) && n.Opcode() == Op_ShenandoahLoadReferenceBarrier &&
2870          mask_decorators(_decorators) == mask_decorators(((const ShenandoahLoadReferenceBarrierNode&)n)._decorators);
2871 }
2872 
2873 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const {
2874   if (in(ValueIn) == NULL || in(ValueIn)->is_top()) {
2875     return Type::TOP;
2876   }
2877   const Type* t = in(ValueIn)->bottom_type();
2878   if (t == TypePtr::NULL_PTR) {
2879     return t;
2880   }
2881 
2882   if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2883     return t;
2884   }
2885 
2886   return t->meet(TypePtr::NULL_PTR);
2887 }
2888 
2889 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseGVN* phase) const {
2890   // Either input is TOP ==> the result is TOP
2891   const Type *t2 = phase->type(in(ValueIn));
2892   if( t2 == Type::TOP ) return Type::TOP;
2893 
2894   if (t2 == TypePtr::NULL_PTR) {
2895     return t2;
2896   }
2897 
2898   if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2899     return t2;
2900   }
2901 
2902   return t2->meet(TypePtr::NULL_PTR);
2903 }
2904 
2905 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseGVN* phase) {
2906   Node* value = in(ValueIn);
2907   if (!needs_barrier(phase, value)) {
2908     return value;
2909   }
2910   return this;
2911 }
2912 
2913 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseGVN* phase, Node* n) {
2914   Unique_Node_List visited;
2915   return needs_barrier_impl(phase, n, visited);
2916 }
2917 
2918 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseGVN* phase, Node* n, Unique_Node_List &visited) {
2919   if (n == NULL) return false;
2920   if (visited.member(n)) {
2921     return false; // Been there.
2922   }
2923   visited.push(n);
2924 
2925   if (n->is_Allocate()) {
2926     // tty->print_cr("optimize barrier on alloc");
2927     return false;
2928   }
2929   if (n->is_Call()) {
2930     // tty->print_cr("optimize barrier on call");
2931     return false;
2932   }
2933 
2934   const Type* type = phase->type(n);
2935   if (type == Type::TOP) {
2936     return false;
2937   }
2938   if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) {
2939     // tty->print_cr("optimize barrier on null");
2940     return false;
2941   }
2942   if (type->make_oopptr() && type->make_oopptr()->const_oop() != NULL) {
2943     // tty->print_cr("optimize barrier on constant");
2944     return false;
2945   }
2946 
2947   switch (n->Opcode()) {
2948     case Op_AddP:
2949       return true; // TODO: Can refine?
2950     case Op_LoadP:
2951     case Op_ShenandoahCompareAndExchangeN:
2952     case Op_ShenandoahCompareAndExchangeP:
2953     case Op_CompareAndExchangeN:
2954     case Op_CompareAndExchangeP:
2955     case Op_GetAndSetN:
2956     case Op_GetAndSetP:
2957       return true;
2958     case Op_Phi: {
2959       for (uint i = 1; i < n->req(); i++) {
2960         if (needs_barrier_impl(phase, n->in(i), visited)) return true;
2961       }
2962       return false;
2963     }
2964     case Op_CheckCastPP:
2965     case Op_CastPP:
2966       return needs_barrier_impl(phase, n->in(1), visited);
2967     case Op_Proj:
2968       return needs_barrier_impl(phase, n->in(0), visited);
2969     case Op_ShenandoahLoadReferenceBarrier:
2970       // tty->print_cr("optimize barrier on barrier");
2971       return false;
2972     case Op_Parm:
2973       // tty->print_cr("optimize barrier on input arg");
2974       return false;
2975     case Op_DecodeN:
2976     case Op_EncodeP:
2977       return needs_barrier_impl(phase, n->in(1), visited);
2978     case Op_LoadN:
2979       return true;
2980     case Op_CMoveN:
2981     case Op_CMoveP:
2982       return needs_barrier_impl(phase, n->in(2), visited) ||
2983              needs_barrier_impl(phase, n->in(3), visited);
2984     case Op_ShenandoahIUBarrier:
2985       return needs_barrier_impl(phase, n->in(1), visited);
2986     case Op_CreateEx:
2987       return false;
2988     default:
2989       break;
2990   }
2991 #ifdef ASSERT
2992   tty->print("need barrier on?: ");
2993   tty->print_cr("ins:");
2994   n->dump(2);
2995   tty->print_cr("outs:");
2996   n->dump(-2);
2997   ShouldNotReachHere();
2998 #endif
2999   return true;
3000 }