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