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();
 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()->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;
 808         c->as_Call()->extract_projections(&projs, true, false);
 809         if (projs.fallthrough_memproj != nullptr) {
 810           if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) {
 811             if (projs.catchall_memproj == nullptr) {
 812               mem = projs.fallthrough_memproj;
 813             } else {
 814               if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) {
 815                 mem = projs.fallthrough_memproj;
 816               } else {
 817                 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier");
 818                 mem = projs.catchall_memproj;
 819               }
 820             }
 821           }
 822         } else {
 823           Node* proj = c->as_Call()->proj_out(TypeFunc::Memory);
 824           if (proj != nullptr &&
 825               proj->adr_type() == TypePtr::BOTTOM) {
 826             mem = proj;
 827           }
 828         }
 829       } else {
 830         for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
 831           Node* u = c->fast_out(i);
 832           if (u->is_Proj() &&
 833               u->bottom_type() == Type::MEMORY &&
 834               u->adr_type() == TypePtr::BOTTOM) {
 835               assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), "");
 836               assert(mem == nullptr, "only one proj");
 837               mem = u;
 838           }
 839         }
 840         assert(!c->is_Call() || c->as_Call()->adr_type() != nullptr || mem == nullptr, "no mem projection expected");
 841       }
 842     }
 843     c = phase->idom(c);
 844   } while (mem == nullptr);
 845   return mem;
 846 }
 847 
 848 void ShenandoahBarrierC2Support::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) {
 849   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 850     Node* u = n->fast_out(i);
 851     if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) {
 852       uses.push(u);
 853     }
 854   }
 855 }
 856 
 857 static void hide_strip_mined_loop(OuterStripMinedLoopNode* outer, CountedLoopNode* inner, PhaseIdealLoop* phase) {
 858   OuterStripMinedLoopEndNode* le = inner->outer_loop_end();
 859   Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl));
 860   phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl));
 861   Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt);
 862   phase->register_control(new_le, phase->get_loop(le), le->in(0));
 863   phase->lazy_replace(outer, new_outer);
 864   phase->lazy_replace(le, new_le);
 865   inner->clear_strip_mined();
 866 }
 867 
 868 void ShenandoahBarrierC2Support::test_gc_state(Node*& ctrl, Node* raw_mem, Node*& test_fail_ctrl,
 869                                                PhaseIdealLoop* phase, int flags) {
 870   PhaseIterGVN& igvn = phase->igvn();
 871   Node* old_ctrl = ctrl;
 872 
 873   Node* thread          = new ThreadLocalNode();
 874   Node* gc_state_offset = igvn.MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 875   Node* gc_state_addr   = new AddPNode(phase->C->top(), thread, gc_state_offset);
 876   Node* gc_state        = new LoadBNode(old_ctrl, raw_mem, gc_state_addr,
 877                                         DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(nullptr),
 878                                         TypeInt::BYTE, MemNode::unordered);
 879   Node* gc_state_and    = new AndINode(gc_state, igvn.intcon(flags));
 880   Node* gc_state_cmp    = new CmpINode(gc_state_and, igvn.zerocon(T_INT));
 881   Node* gc_state_bool   = new BoolNode(gc_state_cmp, BoolTest::ne);
 882 
 883   IfNode* gc_state_iff  = new IfNode(old_ctrl, gc_state_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
 884   ctrl                  = new IfTrueNode(gc_state_iff);
 885   test_fail_ctrl        = new IfFalseNode(gc_state_iff);
 886 
 887   IdealLoopTree* loop = phase->get_loop(old_ctrl);
 888   phase->register_control(gc_state_iff,   loop, old_ctrl);
 889   phase->register_control(ctrl,           loop, gc_state_iff);
 890   phase->register_control(test_fail_ctrl, loop, gc_state_iff);
 891 
 892   phase->register_new_node(thread,        old_ctrl);
 893   phase->register_new_node(gc_state_addr, old_ctrl);
 894   phase->register_new_node(gc_state,      old_ctrl);
 895   phase->register_new_node(gc_state_and,  old_ctrl);
 896   phase->register_new_node(gc_state_cmp,  old_ctrl);
 897   phase->register_new_node(gc_state_bool, old_ctrl);
 898 
 899   phase->set_ctrl(gc_state_offset, phase->C->root());
 900 
 901   assert(is_gc_state_test(gc_state_iff, flags), "Should match the shape");
 902 }
 903 
 904 void ShenandoahBarrierC2Support::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) {
 905   Node* old_ctrl = ctrl;
 906   PhaseIterGVN& igvn = phase->igvn();
 907 
 908   const Type* val_t = igvn.type(val);
 909   if (val_t->meet(TypePtr::NULL_PTR) == val_t) {
 910     Node* null_cmp   = new CmpPNode(val, igvn.zerocon(T_OBJECT));
 911     Node* null_test  = new BoolNode(null_cmp, BoolTest::ne);
 912 
 913     IfNode* null_iff = new IfNode(old_ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
 914     ctrl             = new IfTrueNode(null_iff);
 915     null_ctrl        = new IfFalseNode(null_iff);
 916 
 917     IdealLoopTree* loop = phase->get_loop(old_ctrl);
 918     phase->register_control(null_iff,  loop, old_ctrl);
 919     phase->register_control(ctrl,      loop, null_iff);
 920     phase->register_control(null_ctrl, loop, null_iff);
 921 
 922     phase->register_new_node(null_cmp,  old_ctrl);
 923     phase->register_new_node(null_test, old_ctrl);
 924   }
 925 }
 926 
 927 void ShenandoahBarrierC2Support::test_in_cset(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) {
 928   Node* old_ctrl = ctrl;
 929   PhaseIterGVN& igvn = phase->igvn();
 930 
 931   Node* raw_val        = new CastP2XNode(old_ctrl, val);
 932   Node* cset_idx       = new URShiftXNode(raw_val, igvn.intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint()));
 933 
 934   // Figure out the target cset address with raw pointer math.
 935   // This avoids matching AddP+LoadB that would emit inefficient code.
 936   // See JDK-8245465.
 937   Node* cset_addr_ptr  = igvn.makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr()));
 938   Node* cset_addr      = new CastP2XNode(old_ctrl, cset_addr_ptr);
 939   Node* cset_load_addr = new AddXNode(cset_addr, cset_idx);
 940   Node* cset_load_ptr  = new CastX2PNode(cset_load_addr);
 941 
 942   Node* cset_load      = new LoadBNode(old_ctrl, raw_mem, cset_load_ptr,
 943                                        DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(nullptr),
 944                                        TypeInt::BYTE, MemNode::unordered);
 945   Node* cset_cmp       = new CmpINode(cset_load, igvn.zerocon(T_INT));
 946   Node* cset_bool      = new BoolNode(cset_cmp, BoolTest::ne);
 947 
 948   IfNode* cset_iff     = new IfNode(old_ctrl, cset_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
 949   ctrl                 = new IfTrueNode(cset_iff);
 950   not_cset_ctrl        = new IfFalseNode(cset_iff);
 951 
 952   IdealLoopTree *loop = phase->get_loop(old_ctrl);
 953   phase->register_control(cset_iff,      loop, old_ctrl);
 954   phase->register_control(ctrl,          loop, cset_iff);
 955   phase->register_control(not_cset_ctrl, loop, cset_iff);
 956 
 957   phase->set_ctrl(cset_addr_ptr, phase->C->root());
 958 
 959   phase->register_new_node(raw_val,        old_ctrl);
 960   phase->register_new_node(cset_idx,       old_ctrl);
 961   phase->register_new_node(cset_addr,      old_ctrl);
 962   phase->register_new_node(cset_load_addr, old_ctrl);
 963   phase->register_new_node(cset_load_ptr,  old_ctrl);
 964   phase->register_new_node(cset_load,      old_ctrl);
 965   phase->register_new_node(cset_cmp,       old_ctrl);
 966   phase->register_new_node(cset_bool,      old_ctrl);
 967 }
 968 
 969 void ShenandoahBarrierC2Support::call_lrb_stub(Node*& ctrl, Node*& val, Node* load_addr,
 970                                                DecoratorSet decorators, PhaseIdealLoop* phase) {
 971   IdealLoopTree*loop = phase->get_loop(ctrl);
 972   const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr();
 973 
 974   address calladdr = nullptr;
 975   const char* name = nullptr;
 976   bool is_strong  = ShenandoahBarrierSet::is_strong_access(decorators);
 977   bool is_weak    = ShenandoahBarrierSet::is_weak_access(decorators);
 978   bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
 979   bool is_native  = ShenandoahBarrierSet::is_native_access(decorators);
 980   bool is_narrow  = UseCompressedOops && !is_native;
 981   if (is_strong) {
 982     if (is_narrow) {
 983       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow);
 984       name = "load_reference_barrier_strong_narrow";
 985     } else {
 986       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong);
 987       name = "load_reference_barrier_strong";
 988     }
 989   } else if (is_weak) {
 990     if (is_narrow) {
 991       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow);
 992       name = "load_reference_barrier_weak_narrow";
 993     } else {
 994       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak);
 995       name = "load_reference_barrier_weak";
 996     }
 997   } else {
 998     assert(is_phantom, "only remaining strength");
 999     if (is_narrow) {
1000       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom_narrow);
1001       name = "load_reference_barrier_phantom_narrow";
1002     } else {
1003       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom);
1004       name = "load_reference_barrier_phantom";
1005     }
1006   }
1007   Node* call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type(), calladdr, name, TypeRawPtr::BOTTOM);
1008 
1009   call->init_req(TypeFunc::Control, ctrl);
1010   call->init_req(TypeFunc::I_O, phase->C->top());
1011   call->init_req(TypeFunc::Memory, phase->C->top());
1012   call->init_req(TypeFunc::FramePtr, phase->C->top());
1013   call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1014   call->init_req(TypeFunc::Parms, val);
1015   call->init_req(TypeFunc::Parms+1, load_addr);
1016   phase->register_control(call, loop, ctrl);
1017   ctrl = new ProjNode(call, TypeFunc::Control);
1018   phase->register_control(ctrl, loop, call);
1019   val = new ProjNode(call, TypeFunc::Parms);
1020   phase->register_new_node(val, call);
1021   val = new CheckCastPPNode(ctrl, val, obj_type);
1022   phase->register_new_node(val, ctrl);
1023 }
1024 
1025 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) {
1026   Node* ctrl = phase->get_ctrl(barrier);
1027   Node* init_raw_mem = fixer.find_mem(ctrl, barrier);
1028 
1029   // Update the control of all nodes that should be after the
1030   // barrier control flow
1031   uses.clear();
1032   // Every node that is control dependent on the barrier's input
1033   // control will be after the expanded barrier. The raw memory (if
1034   // its memory is control dependent on the barrier's input control)
1035   // must stay above the barrier.
1036   uses_to_ignore.clear();
1037   if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) {
1038     uses_to_ignore.push(init_raw_mem);
1039   }
1040   for (uint next = 0; next < uses_to_ignore.size(); next++) {
1041     Node *n = uses_to_ignore.at(next);
1042     for (uint i = 0; i < n->req(); i++) {
1043       Node* in = n->in(i);
1044       if (in != nullptr && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) {
1045         uses_to_ignore.push(in);
1046       }
1047     }
1048   }
1049   for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) {
1050     Node* u = ctrl->fast_out(i);
1051     if (u->_idx < last &&
1052         u != barrier &&
1053         !uses_to_ignore.member(u) &&
1054         (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) &&
1055         (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) {
1056       Node* old_c = phase->ctrl_or_self(u);
1057       Node* c = old_c;
1058       if (c != ctrl ||
1059           is_dominator_same_ctrl(old_c, barrier, u, phase) ||
1060           ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) {
1061         phase->igvn().rehash_node_delayed(u);
1062         int nb = u->replace_edge(ctrl, region, &phase->igvn());
1063         if (u->is_CFG()) {
1064           if (phase->idom(u) == ctrl) {
1065             phase->set_idom(u, region, phase->dom_depth(region));
1066           }
1067         } else if (phase->get_ctrl(u) == ctrl) {
1068           assert(u != init_raw_mem, "should leave input raw mem above the barrier");
1069           uses.push(u);
1070         }
1071         assert(nb == 1, "more than 1 ctrl input?");
1072         --i, imax -= nb;
1073       }
1074     }
1075   }
1076 }
1077 
1078 static Node* create_phis_on_call_return(Node* ctrl, Node* c, Node* n, Node* n_clone, const CallProjections& projs, PhaseIdealLoop* phase) {
1079   Node* region = nullptr;
1080   while (c != ctrl) {
1081     if (c->is_Region()) {
1082       region = c;
1083     }
1084     c = phase->idom(c);
1085   }
1086   assert(region != nullptr, "");
1087   Node* phi = new PhiNode(region, n->bottom_type());
1088   for (uint j = 1; j < region->req(); j++) {
1089     Node* in = region->in(j);
1090     if (phase->is_dominator(projs.fallthrough_catchproj, in)) {
1091       phi->init_req(j, n);
1092     } else if (phase->is_dominator(projs.catchall_catchproj, in)) {
1093       phi->init_req(j, n_clone);
1094     } else {
1095       phi->init_req(j, create_phis_on_call_return(ctrl, in, n, n_clone, projs, phase));
1096     }
1097   }
1098   phase->register_new_node(phi, region);
1099   return phi;
1100 }
1101 
1102 void ShenandoahBarrierC2Support::pin_and_expand(PhaseIdealLoop* phase) {
1103   ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
1104 
1105   Unique_Node_List uses;
1106   for (int i = 0; i < state->iu_barriers_count(); i++) {
1107     Node* barrier = state->iu_barrier(i);
1108     Node* ctrl = phase->get_ctrl(barrier);
1109     IdealLoopTree* loop = phase->get_loop(ctrl);
1110     Node* head = loop->head();
1111     if (head->is_OuterStripMinedLoop()) {
1112       // Expanding a barrier here will break loop strip mining
1113       // verification. Transform the loop so the loop nest doesn't
1114       // appear as strip mined.
1115       OuterStripMinedLoopNode* outer = head->as_OuterStripMinedLoop();
1116       hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1117     }
1118   }
1119 
1120   Node_Stack stack(0);
1121   Node_List clones;
1122   for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1123     ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1124 
1125     Node* ctrl = phase->get_ctrl(lrb);
1126     Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1127 
1128     CallStaticJavaNode* unc = nullptr;
1129     Node* unc_ctrl = nullptr;
1130     Node* uncasted_val = val;
1131 
1132     for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) {
1133       Node* u = lrb->fast_out(i);
1134       if (u->Opcode() == Op_CastPP &&
1135           u->in(0) != nullptr &&
1136           phase->is_dominator(u->in(0), ctrl)) {
1137         const Type* u_t = phase->igvn().type(u);
1138 
1139         if (u_t->meet(TypePtr::NULL_PTR) != u_t &&
1140             u->in(0)->Opcode() == Op_IfTrue &&
1141             u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
1142             u->in(0)->in(0)->is_If() &&
1143             u->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
1144             u->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
1145             u->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
1146             u->in(0)->in(0)->in(1)->in(1)->in(1) == val &&
1147             u->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
1148           IdealLoopTree* loop = phase->get_loop(ctrl);
1149           IdealLoopTree* unc_loop = phase->get_loop(u->in(0));
1150 
1151           if (!unc_loop->is_member(loop)) {
1152             continue;
1153           }
1154 
1155           Node* branch = no_branches(ctrl, u->in(0), false, phase);
1156           assert(branch == nullptr || branch == NodeSentinel, "was not looking for a branch");
1157           if (branch == NodeSentinel) {
1158             continue;
1159           }
1160 
1161           Node* iff = u->in(0)->in(0);
1162           Node* bol = iff->in(1)->clone();
1163           Node* cmp = bol->in(1)->clone();
1164           cmp->set_req(1, lrb);
1165           bol->set_req(1, cmp);
1166           phase->igvn().replace_input_of(iff, 1, bol);
1167           phase->set_ctrl(lrb, iff->in(0));
1168           phase->register_new_node(cmp, iff->in(0));
1169           phase->register_new_node(bol, iff->in(0));
1170           break;
1171         }
1172       }
1173     }
1174     if ((ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) || ctrl->is_CallJava()) {
1175       CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_CallJava() : ctrl->as_CallJava();
1176       if (call->entry_point() == OptoRuntime::rethrow_stub()) {
1177         // The rethrow call may have too many projections to be
1178         // properly handled here. Given there's no reason for a
1179         // barrier to depend on the call, move it above the call
1180         stack.push(lrb, 0);
1181         do {
1182           Node* n = stack.node();
1183           uint idx = stack.index();
1184           if (idx < n->req()) {
1185             Node* in = n->in(idx);
1186             stack.set_index(idx+1);
1187             if (in != nullptr) {
1188               if (phase->has_ctrl(in)) {
1189                 if (phase->is_dominator(call, phase->get_ctrl(in))) {
1190 #ifdef ASSERT
1191                   for (uint i = 0; i < stack.size(); i++) {
1192                     assert(stack.node_at(i) != in, "node shouldn't have been seen yet");
1193                   }
1194 #endif
1195                   stack.push(in, 0);
1196                 }
1197               } else {
1198                 assert(phase->is_dominator(in, call->in(0)), "no dependency on the call");
1199               }
1200             }
1201           } else {
1202             phase->set_ctrl(n, call->in(0));
1203             stack.pop();
1204           }
1205         } while(stack.size() > 0);
1206         continue;
1207       }
1208       CallProjections projs;
1209       call->extract_projections(&projs, false, false);
1210 
1211 #ifdef ASSERT
1212       VectorSet cloned;
1213 #endif
1214       Node* lrb_clone = lrb->clone();
1215       phase->register_new_node(lrb_clone, projs.catchall_catchproj);
1216       phase->set_ctrl(lrb, projs.fallthrough_catchproj);
1217 
1218       stack.push(lrb, 0);
1219       clones.push(lrb_clone);
1220 
1221       do {
1222         assert(stack.size() == clones.size(), "");
1223         Node* n = stack.node();
1224 #ifdef ASSERT
1225         if (n->is_Load()) {
1226           Node* mem = n->in(MemNode::Memory);
1227           for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
1228             Node* u = mem->fast_out(j);
1229             assert(!u->is_Store() || !u->is_LoadStore() || phase->get_ctrl(u) != ctrl, "anti dependent store?");
1230           }
1231         }
1232 #endif
1233         uint idx = stack.index();
1234         Node* n_clone = clones.at(clones.size()-1);
1235         if (idx < n->outcnt()) {
1236           Node* u = n->raw_out(idx);
1237           Node* c = phase->ctrl_or_self(u);
1238           if (phase->is_dominator(call, c) && phase->is_dominator(c, projs.fallthrough_proj)) {
1239             stack.set_index(idx+1);
1240             assert(!u->is_CFG(), "");
1241             stack.push(u, 0);
1242             assert(!cloned.test_set(u->_idx), "only one clone");
1243             Node* u_clone = u->clone();
1244             int nb = u_clone->replace_edge(n, n_clone, &phase->igvn());
1245             assert(nb > 0, "should have replaced some uses");
1246             phase->register_new_node(u_clone, projs.catchall_catchproj);
1247             clones.push(u_clone);
1248             phase->set_ctrl(u, projs.fallthrough_catchproj);
1249           } else {
1250             bool replaced = false;
1251             if (u->is_Phi()) {
1252               for (uint k = 1; k < u->req(); k++) {
1253                 if (u->in(k) == n) {
1254                   if (phase->is_dominator(projs.catchall_catchproj, u->in(0)->in(k))) {
1255                     phase->igvn().replace_input_of(u, k, n_clone);
1256                     replaced = true;
1257                   } else if (!phase->is_dominator(projs.fallthrough_catchproj, u->in(0)->in(k))) {
1258                     phase->igvn().replace_input_of(u, k, create_phis_on_call_return(ctrl, u->in(0)->in(k), n, n_clone, projs, phase));
1259                     replaced = true;
1260                   }
1261                 }
1262               }
1263             } else {
1264               if (phase->is_dominator(projs.catchall_catchproj, c)) {
1265                 phase->igvn().rehash_node_delayed(u);
1266                 int nb = u->replace_edge(n, n_clone, &phase->igvn());
1267                 assert(nb > 0, "should have replaced some uses");
1268                 replaced = true;
1269               } else if (!phase->is_dominator(projs.fallthrough_catchproj, c)) {
1270                 if (u->is_If()) {
1271                   // Can't break If/Bool/Cmp chain
1272                   assert(n->is_Bool(), "unexpected If shape");
1273                   assert(stack.node_at(stack.size()-2)->is_Cmp(), "unexpected If shape");
1274                   assert(n_clone->is_Bool(), "unexpected clone");
1275                   assert(clones.at(clones.size()-2)->is_Cmp(), "unexpected clone");
1276                   Node* bol_clone = n->clone();
1277                   Node* cmp_clone = stack.node_at(stack.size()-2)->clone();
1278                   bol_clone->set_req(1, cmp_clone);
1279 
1280                   Node* nn = stack.node_at(stack.size()-3);
1281                   Node* nn_clone = clones.at(clones.size()-3);
1282                   assert(nn->Opcode() == nn_clone->Opcode(), "mismatch");
1283 
1284                   int nb = cmp_clone->replace_edge(nn, create_phis_on_call_return(ctrl, c, nn, nn_clone, projs, phase),
1285                                                    &phase->igvn());
1286                   assert(nb > 0, "should have replaced some uses");
1287 
1288                   phase->register_new_node(bol_clone, u->in(0));
1289                   phase->register_new_node(cmp_clone, u->in(0));
1290 
1291                   phase->igvn().replace_input_of(u, 1, bol_clone);
1292 
1293                 } else {
1294                   phase->igvn().rehash_node_delayed(u);
1295                   int nb = u->replace_edge(n, create_phis_on_call_return(ctrl, c, n, n_clone, projs, phase), &phase->igvn());
1296                   assert(nb > 0, "should have replaced some uses");
1297                 }
1298                 replaced = true;
1299               }
1300             }
1301             if (!replaced) {
1302               stack.set_index(idx+1);
1303             }
1304           }
1305         } else {
1306           stack.pop();
1307           clones.pop();
1308         }
1309       } while (stack.size() > 0);
1310       assert(stack.size() == 0 && clones.size() == 0, "");
1311     }
1312   }
1313 
1314   for (int i = 0; i < state->load_reference_barriers_count(); i++) {
1315     ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1316     Node* ctrl = phase->get_ctrl(lrb);
1317     IdealLoopTree* loop = phase->get_loop(ctrl);
1318     Node* head = loop->head();
1319     if (head->is_OuterStripMinedLoop()) {
1320       // Expanding a barrier here will break loop strip mining
1321       // verification. Transform the loop so the loop nest doesn't
1322       // appear as strip mined.
1323       OuterStripMinedLoopNode* outer = head->as_OuterStripMinedLoop();
1324       hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1325     }
1326   }
1327 
1328   // Expand load-reference-barriers
1329   MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
1330   Unique_Node_List uses_to_ignore;
1331   for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1332     ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1333     uint last = phase->C->unique();
1334     Node* ctrl = phase->get_ctrl(lrb);
1335     Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1336 
1337     Node* orig_ctrl = ctrl;
1338 
1339     Node* raw_mem = fixer.find_mem(ctrl, lrb);
1340     Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, nullptr);
1341 
1342     IdealLoopTree *loop = phase->get_loop(ctrl);
1343 
1344     Node* heap_stable_ctrl = nullptr;
1345     Node* null_ctrl = nullptr;
1346 
1347     assert(val->bottom_type()->make_oopptr(), "need oop");
1348     assert(val->bottom_type()->make_oopptr()->const_oop() == nullptr, "expect non-constant");
1349 
1350     enum { _heap_stable = 1, _evac_path, _not_cset, PATH_LIMIT };
1351     Node* region = new RegionNode(PATH_LIMIT);
1352     Node* val_phi = new PhiNode(region, val->bottom_type()->is_oopptr());
1353 
1354     // Stable path.
1355     int flags = ShenandoahHeap::HAS_FORWARDED;
1356     if (!ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1357       flags |= ShenandoahHeap::WEAK_ROOTS;
1358     }
1359     test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, flags);
1360     IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
1361 
1362     // Heap stable case
1363     region->init_req(_heap_stable, heap_stable_ctrl);
1364     val_phi->init_req(_heap_stable, val);
1365 
1366     // Test for in-cset, unless it's a native-LRB. Native LRBs need to return null
1367     // even for non-cset objects to prevent resurrection of such objects.
1368     // Wires !in_cset(obj) to slot 2 of region and phis
1369     Node* not_cset_ctrl = nullptr;
1370     if (ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1371       test_in_cset(ctrl, not_cset_ctrl, val, raw_mem, phase);
1372     }
1373     if (not_cset_ctrl != nullptr) {
1374       region->init_req(_not_cset, not_cset_ctrl);
1375       val_phi->init_req(_not_cset, val);
1376     } else {
1377       region->del_req(_not_cset);
1378       val_phi->del_req(_not_cset);
1379     }
1380 
1381     // Resolve object when orig-value is in cset.
1382     // Make the unconditional resolve for fwdptr.
1383 
1384     // Call lrb-stub and wire up that path in slots 4
1385     Node* result_mem = nullptr;
1386 
1387     Node* addr;
1388     if (ShenandoahSelfFixing) {
1389       VectorSet visited;
1390       addr = get_load_addr(phase, visited, lrb);
1391     } else {
1392       addr = phase->igvn().zerocon(T_OBJECT);
1393     }
1394     if (addr->Opcode() == Op_AddP) {
1395       Node* orig_base = addr->in(AddPNode::Base);
1396       Node* base = new CheckCastPPNode(ctrl, orig_base, orig_base->bottom_type(), ConstraintCastNode::StrongDependency);
1397       phase->register_new_node(base, ctrl);
1398       if (addr->in(AddPNode::Base) == addr->in((AddPNode::Address))) {
1399         // Field access
1400         addr = addr->clone();
1401         addr->set_req(AddPNode::Base, base);
1402         addr->set_req(AddPNode::Address, base);
1403         phase->register_new_node(addr, ctrl);
1404       } else {
1405         Node* addr2 = addr->in(AddPNode::Address);
1406         if (addr2->Opcode() == Op_AddP && addr2->in(AddPNode::Base) == addr2->in(AddPNode::Address) &&
1407               addr2->in(AddPNode::Base) == orig_base) {
1408           addr2 = addr2->clone();
1409           addr2->set_req(AddPNode::Base, base);
1410           addr2->set_req(AddPNode::Address, base);
1411           phase->register_new_node(addr2, ctrl);
1412           addr = addr->clone();
1413           addr->set_req(AddPNode::Base, base);
1414           addr->set_req(AddPNode::Address, addr2);
1415           phase->register_new_node(addr, ctrl);
1416         }
1417       }
1418     }
1419     call_lrb_stub(ctrl, val, addr, lrb->decorators(), phase);
1420     region->init_req(_evac_path, ctrl);
1421     val_phi->init_req(_evac_path, val);
1422 
1423     phase->register_control(region, loop, heap_stable_iff);
1424     Node* out_val = val_phi;
1425     phase->register_new_node(val_phi, region);
1426 
1427     fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase);
1428 
1429     ctrl = orig_ctrl;
1430 
1431     phase->igvn().replace_node(lrb, out_val);
1432 
1433     follow_barrier_uses(out_val, ctrl, uses, phase);
1434 
1435     for(uint next = 0; next < uses.size(); next++ ) {
1436       Node *n = uses.at(next);
1437       assert(phase->get_ctrl(n) == ctrl, "bad control");
1438       assert(n != raw_mem, "should leave input raw mem above the barrier");
1439       phase->set_ctrl(n, region);
1440       follow_barrier_uses(n, ctrl, uses, phase);
1441     }
1442     fixer.record_new_ctrl(ctrl, region, raw_mem, raw_mem_for_ctrl);
1443   }
1444   // Done expanding load-reference-barriers.
1445   assert(ShenandoahBarrierSetC2::bsc2()->state()->load_reference_barriers_count() == 0, "all load reference barrier nodes should have been replaced");
1446 
1447   for (int i = state->iu_barriers_count() - 1; i >= 0; i--) {
1448     Node* barrier = state->iu_barrier(i);
1449     Node* pre_val = barrier->in(1);
1450 
1451     if (phase->igvn().type(pre_val)->higher_equal(TypePtr::NULL_PTR)) {
1452       ShouldNotReachHere();
1453       continue;
1454     }
1455 
1456     Node* ctrl = phase->get_ctrl(barrier);
1457 
1458     if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) {
1459       assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0)->in(0), pre_val, ctrl->in(0), phase), "can't move");
1460       ctrl = ctrl->in(0)->in(0);
1461       phase->set_ctrl(barrier, ctrl);
1462     } else if (ctrl->is_CallRuntime()) {
1463       assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0), pre_val, ctrl, phase), "can't move");
1464       ctrl = ctrl->in(0);
1465       phase->set_ctrl(barrier, ctrl);
1466     }
1467 
1468     Node* init_ctrl = ctrl;
1469     IdealLoopTree* loop = phase->get_loop(ctrl);
1470     Node* raw_mem = fixer.find_mem(ctrl, barrier);
1471     Node* init_raw_mem = raw_mem;
1472     Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, nullptr);
1473     Node* heap_stable_ctrl = nullptr;
1474     Node* null_ctrl = nullptr;
1475     uint last = phase->C->unique();
1476 
1477     enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
1478     Node* region = new RegionNode(PATH_LIMIT);
1479     Node* phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1480 
1481     enum { _fast_path = 1, _slow_path, _null_path, PATH_LIMIT2 };
1482     Node* region2 = new RegionNode(PATH_LIMIT2);
1483     Node* phi2 = PhiNode::make(region2, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1484 
1485     // Stable path.
1486     test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, ShenandoahHeap::MARKING);
1487     region->init_req(_heap_stable, heap_stable_ctrl);
1488     phi->init_req(_heap_stable, raw_mem);
1489 
1490     // Null path
1491     Node* reg2_ctrl = nullptr;
1492     test_null(ctrl, pre_val, null_ctrl, phase);
1493     if (null_ctrl != nullptr) {
1494       reg2_ctrl = null_ctrl->in(0);
1495       region2->init_req(_null_path, null_ctrl);
1496       phi2->init_req(_null_path, raw_mem);
1497     } else {
1498       region2->del_req(_null_path);
1499       phi2->del_req(_null_path);
1500     }
1501 
1502     const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
1503     const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1504     Node* thread = new ThreadLocalNode();
1505     phase->register_new_node(thread, ctrl);
1506     Node* buffer_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(buffer_offset));
1507     phase->register_new_node(buffer_adr, ctrl);
1508     Node* index_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(index_offset));
1509     phase->register_new_node(index_adr, ctrl);
1510 
1511     BasicType index_bt = TypeX_X->basic_type();
1512     assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading Shenandoah SATBMarkQueue::_index with wrong size.");
1513     const TypePtr* adr_type = TypeRawPtr::BOTTOM;
1514     Node* index = new LoadXNode(ctrl, raw_mem, index_adr, adr_type, TypeX_X, MemNode::unordered);
1515     phase->register_new_node(index, ctrl);
1516     Node* index_cmp = new CmpXNode(index, phase->igvn().MakeConX(0));
1517     phase->register_new_node(index_cmp, ctrl);
1518     Node* index_test = new BoolNode(index_cmp, BoolTest::ne);
1519     phase->register_new_node(index_test, ctrl);
1520     IfNode* queue_full_iff = new IfNode(ctrl, index_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
1521     if (reg2_ctrl == nullptr) reg2_ctrl = queue_full_iff;
1522     phase->register_control(queue_full_iff, loop, ctrl);
1523     Node* not_full = new IfTrueNode(queue_full_iff);
1524     phase->register_control(not_full, loop, queue_full_iff);
1525     Node* full = new IfFalseNode(queue_full_iff);
1526     phase->register_control(full, loop, queue_full_iff);
1527 
1528     ctrl = not_full;
1529 
1530     Node* next_index = new SubXNode(index, phase->igvn().MakeConX(sizeof(intptr_t)));
1531     phase->register_new_node(next_index, ctrl);
1532 
1533     Node* buffer  = new LoadPNode(ctrl, raw_mem, buffer_adr, adr_type, TypeRawPtr::NOTNULL, MemNode::unordered);
1534     phase->register_new_node(buffer, ctrl);
1535     Node *log_addr = new AddPNode(phase->C->top(), buffer, next_index);
1536     phase->register_new_node(log_addr, ctrl);
1537     Node* log_store = new StorePNode(ctrl, raw_mem, log_addr, adr_type, pre_val, MemNode::unordered);
1538     phase->register_new_node(log_store, ctrl);
1539     // update the index
1540     Node* index_update = new StoreXNode(ctrl, log_store, index_adr, adr_type, next_index, MemNode::unordered);
1541     phase->register_new_node(index_update, ctrl);
1542 
1543     // Fast-path case
1544     region2->init_req(_fast_path, ctrl);
1545     phi2->init_req(_fast_path, index_update);
1546 
1547     ctrl = full;
1548 
1549     Node* base = find_bottom_mem(ctrl, phase);
1550 
1551     MergeMemNode* mm = MergeMemNode::make(base);
1552     mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
1553     phase->register_new_node(mm, ctrl);
1554 
1555     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);
1556     call->init_req(TypeFunc::Control, ctrl);
1557     call->init_req(TypeFunc::I_O, phase->C->top());
1558     call->init_req(TypeFunc::Memory, mm);
1559     call->init_req(TypeFunc::FramePtr, phase->C->top());
1560     call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1561     call->init_req(TypeFunc::Parms, pre_val);
1562     call->init_req(TypeFunc::Parms+1, thread);
1563     phase->register_control(call, loop, ctrl);
1564 
1565     Node* ctrl_proj = new ProjNode(call, TypeFunc::Control);
1566     phase->register_control(ctrl_proj, loop, call);
1567     Node* mem_proj = new ProjNode(call, TypeFunc::Memory);
1568     phase->register_new_node(mem_proj, call);
1569 
1570     // Slow-path case
1571     region2->init_req(_slow_path, ctrl_proj);
1572     phi2->init_req(_slow_path, mem_proj);
1573 
1574     phase->register_control(region2, loop, reg2_ctrl);
1575     phase->register_new_node(phi2, region2);
1576 
1577     region->init_req(_heap_unstable, region2);
1578     phi->init_req(_heap_unstable, phi2);
1579 
1580     phase->register_control(region, loop, heap_stable_ctrl->in(0));
1581     phase->register_new_node(phi, region);
1582 
1583     fix_ctrl(barrier, region, fixer, uses, uses_to_ignore, last, phase);
1584     for(uint next = 0; next < uses.size(); next++ ) {
1585       Node *n = uses.at(next);
1586       assert(phase->get_ctrl(n) == init_ctrl, "bad control");
1587       assert(n != init_raw_mem, "should leave input raw mem above the barrier");
1588       phase->set_ctrl(n, region);
1589       follow_barrier_uses(n, init_ctrl, uses, phase);
1590     }
1591     fixer.fix_mem(init_ctrl, region, init_raw_mem, raw_mem_for_ctrl, phi, uses);
1592 
1593     phase->igvn().replace_node(barrier, pre_val);
1594   }
1595   assert(state->iu_barriers_count() == 0, "all enqueue barrier nodes should have been replaced");
1596 
1597 }
1598 
1599 Node* ShenandoahBarrierC2Support::get_load_addr(PhaseIdealLoop* phase, VectorSet& visited, Node* in) {
1600   if (visited.test_set(in->_idx)) {
1601     return nullptr;
1602   }
1603   switch (in->Opcode()) {
1604     case Op_Proj:
1605       return get_load_addr(phase, visited, in->in(0));
1606     case Op_CastPP:
1607     case Op_CheckCastPP:
1608     case Op_DecodeN:
1609     case Op_EncodeP:
1610       return get_load_addr(phase, visited, in->in(1));
1611     case Op_LoadN:
1612     case Op_LoadP:
1613       return in->in(MemNode::Address);
1614     case Op_CompareAndExchangeN:
1615     case Op_CompareAndExchangeP:
1616     case Op_GetAndSetN:
1617     case Op_GetAndSetP:
1618     case Op_ShenandoahCompareAndExchangeP:
1619     case Op_ShenandoahCompareAndExchangeN:
1620       // Those instructions would just have stored a different
1621       // value into the field. No use to attempt to fix it at this point.
1622       return phase->igvn().zerocon(T_OBJECT);
1623     case Op_CMoveP:
1624     case Op_CMoveN: {
1625       Node* t = get_load_addr(phase, visited, in->in(CMoveNode::IfTrue));
1626       Node* f = get_load_addr(phase, visited, in->in(CMoveNode::IfFalse));
1627       // Handle unambiguous cases: single address reported on both branches.
1628       if (t != nullptr && f == nullptr) return t;
1629       if (t == nullptr && f != nullptr) return f;
1630       if (t != nullptr && t == f)    return t;
1631       // Ambiguity.
1632       return phase->igvn().zerocon(T_OBJECT);
1633     }
1634     case Op_Phi: {
1635       Node* addr = nullptr;
1636       for (uint i = 1; i < in->req(); i++) {
1637         Node* addr1 = get_load_addr(phase, visited, in->in(i));
1638         if (addr == nullptr) {
1639           addr = addr1;
1640         }
1641         if (addr != addr1) {
1642           return phase->igvn().zerocon(T_OBJECT);
1643         }
1644       }
1645       return addr;
1646     }
1647     case Op_ShenandoahLoadReferenceBarrier:
1648       return get_load_addr(phase, visited, in->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1649     case Op_ShenandoahIUBarrier:
1650       return get_load_addr(phase, visited, in->in(1));
1651     case Op_CallDynamicJava:
1652     case Op_CallLeaf:
1653     case Op_CallStaticJava:
1654     case Op_ConN:
1655     case Op_ConP:
1656     case Op_Parm:
1657     case Op_CreateEx:
1658       return phase->igvn().zerocon(T_OBJECT);
1659     default:
1660 #ifdef ASSERT
1661       fatal("Unknown node in get_load_addr: %s", NodeClassNames[in->Opcode()]);
1662 #endif
1663       return phase->igvn().zerocon(T_OBJECT);
1664   }
1665 
1666 }
1667 
1668 void ShenandoahBarrierC2Support::move_gc_state_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
1669   IdealLoopTree *loop = phase->get_loop(iff);
1670   Node* loop_head = loop->_head;
1671   Node* entry_c = loop_head->in(LoopNode::EntryControl);
1672 
1673   Node* bol = iff->in(1);
1674   Node* cmp = bol->in(1);
1675   Node* andi = cmp->in(1);
1676   Node* load = andi->in(1);
1677 
1678   assert(is_gc_state_load(load), "broken");
1679   if (!phase->is_dominator(load->in(0), entry_c)) {
1680     Node* mem_ctrl = nullptr;
1681     Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
1682     load = load->clone();
1683     load->set_req(MemNode::Memory, mem);
1684     load->set_req(0, entry_c);
1685     phase->register_new_node(load, entry_c);
1686     andi = andi->clone();
1687     andi->set_req(1, load);
1688     phase->register_new_node(andi, entry_c);
1689     cmp = cmp->clone();
1690     cmp->set_req(1, andi);
1691     phase->register_new_node(cmp, entry_c);
1692     bol = bol->clone();
1693     bol->set_req(1, cmp);
1694     phase->register_new_node(bol, entry_c);
1695 
1696     phase->igvn().replace_input_of(iff, 1, bol);
1697   }
1698 }
1699 
1700 bool ShenandoahBarrierC2Support::identical_backtoback_ifs(Node* n, PhaseIdealLoop* phase) {
1701   if (!n->is_If() || n->is_CountedLoopEnd()) {
1702     return false;
1703   }
1704   Node* region = n->in(0);
1705 
1706   if (!region->is_Region()) {
1707     return false;
1708   }
1709   Node* dom = phase->idom(region);
1710   if (!dom->is_If()) {
1711     return false;
1712   }
1713 
1714   if (!is_heap_stable_test(n) || !is_heap_stable_test(dom)) {
1715     return false;
1716   }
1717 
1718   IfNode* dom_if = dom->as_If();
1719   Node* proj_true = dom_if->proj_out(1);
1720   Node* proj_false = dom_if->proj_out(0);
1721 
1722   for (uint i = 1; i < region->req(); i++) {
1723     if (phase->is_dominator(proj_true, region->in(i))) {
1724       continue;
1725     }
1726     if (phase->is_dominator(proj_false, region->in(i))) {
1727       continue;
1728     }
1729     return false;
1730   }
1731 
1732   return true;
1733 }
1734 
1735 void ShenandoahBarrierC2Support::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) {
1736   assert(is_heap_stable_test(n), "no other tests");
1737   if (identical_backtoback_ifs(n, phase)) {
1738     Node* n_ctrl = n->in(0);
1739     if (phase->can_split_if(n_ctrl)) {
1740       IfNode* dom_if = phase->idom(n_ctrl)->as_If();
1741       if (is_heap_stable_test(n)) {
1742         Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1);
1743         assert(is_gc_state_load(gc_state_load), "broken");
1744         Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1);
1745         assert(is_gc_state_load(dom_gc_state_load), "broken");
1746         if (gc_state_load != dom_gc_state_load) {
1747           phase->igvn().replace_node(gc_state_load, dom_gc_state_load);
1748         }
1749       }
1750       PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1751       Node* proj_true = dom_if->proj_out(1);
1752       Node* proj_false = dom_if->proj_out(0);
1753       Node* con_true = phase->igvn().makecon(TypeInt::ONE);
1754       Node* con_false = phase->igvn().makecon(TypeInt::ZERO);
1755 
1756       for (uint i = 1; i < n_ctrl->req(); i++) {
1757         if (phase->is_dominator(proj_true, n_ctrl->in(i))) {
1758           bolphi->init_req(i, con_true);
1759         } else {
1760           assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1761           bolphi->init_req(i, con_false);
1762         }
1763       }
1764       phase->register_new_node(bolphi, n_ctrl);
1765       phase->igvn().replace_input_of(n, 1, bolphi);
1766       phase->do_split_if(n);
1767     }
1768   }
1769 }
1770 
1771 IfNode* ShenandoahBarrierC2Support::find_unswitching_candidate(const IdealLoopTree* loop, PhaseIdealLoop* phase) {
1772   // Find first invariant test that doesn't exit the loop
1773   LoopNode *head = loop->_head->as_Loop();
1774   IfNode* unswitch_iff = nullptr;
1775   Node* n = head->in(LoopNode::LoopBackControl);
1776   int loop_has_sfpts = -1;
1777   while (n != head) {
1778     Node* n_dom = phase->idom(n);
1779     if (n->is_Region()) {
1780       if (n_dom->is_If()) {
1781         IfNode* iff = n_dom->as_If();
1782         if (iff->in(1)->is_Bool()) {
1783           BoolNode* bol = iff->in(1)->as_Bool();
1784           if (bol->in(1)->is_Cmp()) {
1785             // If condition is invariant and not a loop exit,
1786             // then found reason to unswitch.
1787             if (is_heap_stable_test(iff) &&
1788                 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) {
1789               assert(!loop->is_loop_exit(iff), "both branches should be in the loop");
1790               if (loop_has_sfpts == -1) {
1791                 for(uint i = 0; i < loop->_body.size(); i++) {
1792                   Node *m = loop->_body[i];
1793                   if (m->is_SafePoint() && !m->is_CallLeaf()) {
1794                     loop_has_sfpts = 1;
1795                     break;
1796                   }
1797                 }
1798                 if (loop_has_sfpts == -1) {
1799                   loop_has_sfpts = 0;
1800                 }
1801               }
1802               if (!loop_has_sfpts) {
1803                 unswitch_iff = iff;
1804               }
1805             }
1806           }
1807         }
1808       }
1809     }
1810     n = n_dom;
1811   }
1812   return unswitch_iff;
1813 }
1814 
1815 
1816 void ShenandoahBarrierC2Support::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) {
1817   Node_List heap_stable_tests;
1818   stack.push(phase->C->start(), 0);
1819   do {
1820     Node* n = stack.node();
1821     uint i = stack.index();
1822 
1823     if (i < n->outcnt()) {
1824       Node* u = n->raw_out(i);
1825       stack.set_index(i+1);
1826       if (!visited.test_set(u->_idx)) {
1827         stack.push(u, 0);
1828       }
1829     } else {
1830       stack.pop();
1831       if (n->is_If() && is_heap_stable_test(n)) {
1832         heap_stable_tests.push(n);
1833       }
1834     }
1835   } while (stack.size() > 0);
1836 
1837   for (uint i = 0; i < heap_stable_tests.size(); i++) {
1838     Node* n = heap_stable_tests.at(i);
1839     assert(is_heap_stable_test(n), "only evacuation test");
1840     merge_back_to_back_tests(n, phase);
1841   }
1842 
1843   if (!phase->C->major_progress()) {
1844     VectorSet seen;
1845     for (uint i = 0; i < heap_stable_tests.size(); i++) {
1846       Node* n = heap_stable_tests.at(i);
1847       IdealLoopTree* loop = phase->get_loop(n);
1848       if (loop != phase->ltree_root() &&
1849           loop->_child == nullptr &&
1850           !loop->_irreducible) {
1851         Node* head = loop->_head;
1852         if (head->is_Loop() &&
1853             (!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) &&
1854             !seen.test_set(head->_idx)) {
1855           IfNode* iff = find_unswitching_candidate(loop, phase);
1856           if (iff != nullptr) {
1857             Node* bol = iff->in(1);
1858             if (head->as_Loop()->is_strip_mined()) {
1859               head->as_Loop()->verify_strip_mined(0);
1860             }
1861             move_gc_state_test_out_of_loop(iff, phase);
1862 
1863             AutoNodeBudget node_budget(phase);
1864 
1865             if (loop->policy_unswitching(phase)) {
1866               if (head->as_Loop()->is_strip_mined()) {
1867                 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop();
1868                 hide_strip_mined_loop(outer, head->as_CountedLoop(), phase);
1869               }
1870               phase->do_unswitching(loop, old_new);
1871             } else {
1872               // Not proceeding with unswitching. Move load back in
1873               // the loop.
1874               phase->igvn().replace_input_of(iff, 1, bol);
1875             }
1876           }
1877         }
1878       }
1879     }
1880   }
1881 }
1882 
1883 ShenandoahIUBarrierNode::ShenandoahIUBarrierNode(Node* val) : Node(nullptr, val) {
1884   ShenandoahBarrierSetC2::bsc2()->state()->add_iu_barrier(this);
1885 }
1886 
1887 const Type* ShenandoahIUBarrierNode::bottom_type() const {
1888   if (in(1) == nullptr || in(1)->is_top()) {
1889     return Type::TOP;
1890   }
1891   const Type* t = in(1)->bottom_type();
1892   if (t == TypePtr::NULL_PTR) {
1893     return t;
1894   }
1895   return t->is_oopptr();
1896 }
1897 
1898 const Type* ShenandoahIUBarrierNode::Value(PhaseGVN* phase) const {
1899   if (in(1) == nullptr) {
1900     return Type::TOP;
1901   }
1902   const Type* t = phase->type(in(1));
1903   if (t == Type::TOP) {
1904     return Type::TOP;
1905   }
1906   if (t == TypePtr::NULL_PTR) {
1907     return t;
1908   }
1909   return t->is_oopptr();
1910 }
1911 
1912 int ShenandoahIUBarrierNode::needed(Node* n) {
1913   if (n == nullptr ||
1914       n->is_Allocate() ||
1915       n->Opcode() == Op_ShenandoahIUBarrier ||
1916       n->bottom_type() == TypePtr::NULL_PTR ||
1917       (n->bottom_type()->make_oopptr() != nullptr && n->bottom_type()->make_oopptr()->const_oop() != nullptr)) {
1918     return NotNeeded;
1919   }
1920   if (n->is_Phi() ||
1921       n->is_CMove()) {
1922     return MaybeNeeded;
1923   }
1924   return Needed;
1925 }
1926 
1927 Node* ShenandoahIUBarrierNode::next(Node* n) {
1928   for (;;) {
1929     if (n == nullptr) {
1930       return n;
1931     } else if (n->bottom_type() == TypePtr::NULL_PTR) {
1932       return n;
1933     } else if (n->bottom_type()->make_oopptr() != nullptr && n->bottom_type()->make_oopptr()->const_oop() != nullptr) {
1934       return n;
1935     } else if (n->is_ConstraintCast() ||
1936                n->Opcode() == Op_DecodeN ||
1937                n->Opcode() == Op_EncodeP) {
1938       n = n->in(1);
1939     } else if (n->is_Proj()) {
1940       n = n->in(0);
1941     } else {
1942       return n;
1943     }
1944   }
1945   ShouldNotReachHere();
1946   return nullptr;
1947 }
1948 
1949 Node* ShenandoahIUBarrierNode::Identity(PhaseGVN* phase) {
1950   PhaseIterGVN* igvn = phase->is_IterGVN();
1951 
1952   Node* n = next(in(1));
1953 
1954   int cont = needed(n);
1955 
1956   if (cont == NotNeeded) {
1957     return in(1);
1958   } else if (cont == MaybeNeeded) {
1959     if (igvn == nullptr) {
1960       phase->record_for_igvn(this);
1961       return this;
1962     } else {
1963       ResourceMark rm;
1964       Unique_Node_List wq;
1965       uint wq_i = 0;
1966 
1967       for (;;) {
1968         if (n->is_Phi()) {
1969           for (uint i = 1; i < n->req(); i++) {
1970             Node* m = n->in(i);
1971             if (m != nullptr) {
1972               wq.push(m);
1973             }
1974           }
1975         } else {
1976           assert(n->is_CMove(), "nothing else here");
1977           Node* m = n->in(CMoveNode::IfFalse);
1978           wq.push(m);
1979           m = n->in(CMoveNode::IfTrue);
1980           wq.push(m);
1981         }
1982         Node* orig_n = nullptr;
1983         do {
1984           if (wq_i >= wq.size()) {
1985             return in(1);
1986           }
1987           n = wq.at(wq_i);
1988           wq_i++;
1989           orig_n = n;
1990           n = next(n);
1991           cont = needed(n);
1992           if (cont == Needed) {
1993             return this;
1994           }
1995         } while (cont != MaybeNeeded || (orig_n != n && wq.member(n)));
1996       }
1997     }
1998   }
1999 
2000   return this;
2001 }
2002 
2003 #ifdef ASSERT
2004 static bool has_never_branch(Node* root) {
2005   for (uint i = 1; i < root->req(); i++) {
2006     Node* in = root->in(i);
2007     if (in != nullptr && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->is_NeverBranch()) {
2008       return true;
2009     }
2010   }
2011   return false;
2012 }
2013 #endif
2014 
2015 void MemoryGraphFixer::collect_memory_nodes() {
2016   Node_Stack stack(0);
2017   VectorSet visited;
2018   Node_List regions;
2019 
2020   // Walk the raw memory graph and create a mapping from CFG node to
2021   // memory node. Exclude phis for now.
2022   stack.push(_phase->C->root(), 1);
2023   do {
2024     Node* n = stack.node();
2025     int opc = n->Opcode();
2026     uint i = stack.index();
2027     if (i < n->req()) {
2028       Node* mem = nullptr;
2029       if (opc == Op_Root) {
2030         Node* in = n->in(i);
2031         int in_opc = in->Opcode();
2032         if (in_opc == Op_Return || in_opc == Op_Rethrow) {
2033           mem = in->in(TypeFunc::Memory);
2034         } else if (in_opc == Op_Halt) {
2035           if (in->in(0)->is_Region()) {
2036             Node* r = in->in(0);
2037             for (uint j = 1; j < r->req(); j++) {
2038               assert(!r->in(j)->is_NeverBranch(), "");
2039             }
2040           } else {
2041             Node* proj = in->in(0);
2042             assert(proj->is_Proj(), "");
2043             Node* in = proj->in(0);
2044             assert(in->is_CallStaticJava() || in->is_NeverBranch() || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
2045             if (in->is_CallStaticJava()) {
2046               mem = in->in(TypeFunc::Memory);
2047             } else if (in->Opcode() == Op_Catch) {
2048               Node* call = in->in(0)->in(0);
2049               assert(call->is_Call(), "");
2050               mem = call->in(TypeFunc::Memory);
2051             } else if (in->is_NeverBranch()) {
2052               mem = collect_memory_for_infinite_loop(in);
2053             }
2054           }
2055         } else {
2056 #ifdef ASSERT
2057           n->dump();
2058           in->dump();
2059 #endif
2060           ShouldNotReachHere();
2061         }
2062       } else {
2063         assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
2064         assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
2065         mem = n->in(i);
2066       }
2067       i++;
2068       stack.set_index(i);
2069       if (mem == nullptr) {
2070         continue;
2071       }
2072       for (;;) {
2073         if (visited.test_set(mem->_idx) || mem->is_Start()) {
2074           break;
2075         }
2076         if (mem->is_Phi()) {
2077           stack.push(mem, 2);
2078           mem = mem->in(1);
2079         } else if (mem->is_Proj()) {
2080           stack.push(mem, mem->req());
2081           mem = mem->in(0);
2082         } else if (mem->is_SafePoint() || mem->is_MemBar()) {
2083           mem = mem->in(TypeFunc::Memory);
2084         } else if (mem->is_MergeMem()) {
2085           MergeMemNode* mm = mem->as_MergeMem();
2086           mem = mm->memory_at(_alias);
2087         } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
2088           assert(_alias == Compile::AliasIdxRaw, "");
2089           stack.push(mem, mem->req());
2090           mem = mem->in(MemNode::Memory);
2091         } else {
2092 #ifdef ASSERT
2093           mem->dump();
2094 #endif
2095           ShouldNotReachHere();
2096         }
2097       }
2098     } else {
2099       if (n->is_Phi()) {
2100         // Nothing
2101       } else if (!n->is_Root()) {
2102         Node* c = get_ctrl(n);
2103         _memory_nodes.map(c->_idx, n);
2104       }
2105       stack.pop();
2106     }
2107   } while(stack.is_nonempty());
2108 
2109   // Iterate over CFG nodes in rpo and propagate memory state to
2110   // compute memory state at regions, creating new phis if needed.
2111   Node_List rpo_list;
2112   visited.clear();
2113   _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
2114   Node* root = rpo_list.pop();
2115   assert(root == _phase->C->root(), "");
2116 
2117   const bool trace = false;
2118 #ifdef ASSERT
2119   if (trace) {
2120     for (int i = rpo_list.size() - 1; i >= 0; i--) {
2121       Node* c = rpo_list.at(i);
2122       if (_memory_nodes[c->_idx] != nullptr) {
2123         tty->print("X %d", c->_idx);  _memory_nodes[c->_idx]->dump();
2124       }
2125     }
2126   }
2127 #endif
2128   uint last = _phase->C->unique();
2129 
2130 #ifdef ASSERT
2131   uint16_t max_depth = 0;
2132   for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
2133     IdealLoopTree* lpt = iter.current();
2134     max_depth = MAX2(max_depth, lpt->_nest);
2135   }
2136 #endif
2137 
2138   bool progress = true;
2139   int iteration = 0;
2140   Node_List dead_phis;
2141   while (progress) {
2142     progress = false;
2143     iteration++;
2144     assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2145     if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
2146 
2147     for (int i = rpo_list.size() - 1; i >= 0; i--) {
2148       Node* c = rpo_list.at(i);
2149 
2150       Node* prev_mem = _memory_nodes[c->_idx];
2151       if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2152         Node* prev_region = regions[c->_idx];
2153         Node* unique = nullptr;
2154         for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
2155           Node* m = _memory_nodes[c->in(j)->_idx];
2156           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");
2157           if (m != nullptr) {
2158             if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
2159               assert(c->is_Loop() && j == LoopNode::LoopBackControl || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2160               // continue
2161             } else if (unique == nullptr) {
2162               unique = m;
2163             } else if (m == unique) {
2164               // continue
2165             } else {
2166               unique = NodeSentinel;
2167             }
2168           }
2169         }
2170         assert(unique != nullptr, "empty phi???");
2171         if (unique != NodeSentinel) {
2172           if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c) {
2173             dead_phis.push(prev_region);
2174           }
2175           regions.map(c->_idx, unique);
2176         } else {
2177           Node* phi = nullptr;
2178           if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
2179             phi = prev_region;
2180             for (uint k = 1; k < c->req(); k++) {
2181               Node* m = _memory_nodes[c->in(k)->_idx];
2182               assert(m != nullptr, "expect memory state");
2183               phi->set_req(k, m);
2184             }
2185           } else {
2186             for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == nullptr; j++) {
2187               Node* u = c->fast_out(j);
2188               if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2189                   (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
2190                 phi = u;
2191                 for (uint k = 1; k < c->req() && phi != nullptr; k++) {
2192                   Node* m = _memory_nodes[c->in(k)->_idx];
2193                   assert(m != nullptr, "expect memory state");
2194                   if (u->in(k) != m) {
2195                     phi = NodeSentinel;
2196                   }
2197                 }
2198               }
2199             }
2200             if (phi == NodeSentinel) {
2201               phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
2202               for (uint k = 1; k < c->req(); k++) {
2203                 Node* m = _memory_nodes[c->in(k)->_idx];
2204                 assert(m != nullptr, "expect memory state");
2205                 phi->init_req(k, m);
2206               }
2207             }
2208           }
2209           if (phi != nullptr) {
2210             regions.map(c->_idx, phi);
2211           } else {
2212             assert(c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
2213           }
2214         }
2215         Node* current_region = regions[c->_idx];
2216         if (current_region != prev_region) {
2217           progress = true;
2218           if (prev_region == prev_mem) {
2219             _memory_nodes.map(c->_idx, current_region);
2220           }
2221         }
2222       } else if (prev_mem == nullptr || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
2223         Node* m = _memory_nodes[_phase->idom(c)->_idx];
2224         assert(m != nullptr || c->Opcode() == Op_Halt, "expect memory state");
2225         if (m != prev_mem) {
2226           _memory_nodes.map(c->_idx, m);
2227           progress = true;
2228         }
2229       }
2230 #ifdef ASSERT
2231       if (trace) { tty->print("X %d", c->_idx);  _memory_nodes[c->_idx]->dump(); }
2232 #endif
2233     }
2234   }
2235 
2236   // Replace existing phi with computed memory state for that region
2237   // if different (could be a new phi or a dominating memory node if
2238   // that phi was found to be useless).
2239   while (dead_phis.size() > 0) {
2240     Node* n = dead_phis.pop();
2241     n->replace_by(_phase->C->top());
2242     n->destruct(&_phase->igvn());
2243   }
2244   for (int i = rpo_list.size() - 1; i >= 0; i--) {
2245     Node* c = rpo_list.at(i);
2246     if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2247       Node* n = regions[c->_idx];
2248       assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
2249       if (n != nullptr && n->is_Phi() && n->_idx >= last && n->in(0) == c) {
2250         _phase->register_new_node(n, c);
2251       }
2252     }
2253   }
2254   for (int i = rpo_list.size() - 1; i >= 0; i--) {
2255     Node* c = rpo_list.at(i);
2256     if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2257       Node* n = regions[c->_idx];
2258       assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
2259       for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
2260         Node* u = c->fast_out(i);
2261         if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2262             u != n) {
2263           assert(c->unique_ctrl_out()->Opcode() != Op_Halt, "expected memory state");
2264           if (u->adr_type() == TypePtr::BOTTOM) {
2265             fix_memory_uses(u, n, n, c);
2266           } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2267             _phase->lazy_replace(u, n);
2268             --i; --imax;
2269           }
2270         }
2271       }
2272     }
2273   }
2274 }
2275 
2276 Node* MemoryGraphFixer::collect_memory_for_infinite_loop(const Node* in) {
2277   Node* mem = nullptr;
2278   Node* head = in->in(0);
2279   assert(head->is_Region(), "unexpected infinite loop graph shape");
2280 
2281   Node* phi_mem = nullptr;
2282   for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2283     Node* u = head->fast_out(j);
2284     if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
2285       if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2286         assert(phi_mem == nullptr || phi_mem->adr_type() == TypePtr::BOTTOM, "");
2287         phi_mem = u;
2288       } else if (u->adr_type() == TypePtr::BOTTOM) {
2289         assert(phi_mem == nullptr || _phase->C->get_alias_index(phi_mem->adr_type()) == _alias, "");
2290         if (phi_mem == nullptr) {
2291           phi_mem = u;
2292         }
2293       }
2294     }
2295   }
2296   if (phi_mem == nullptr) {
2297     ResourceMark rm;
2298     Node_Stack stack(0);
2299     stack.push(head, 1);
2300     do {
2301       Node* n = stack.node();
2302       uint i = stack.index();
2303       if (i >= n->req()) {
2304         stack.pop();
2305       } else {
2306         stack.set_index(i + 1);
2307         Node* c = n->in(i);
2308         assert(c != head, "should have found a safepoint on the way");
2309         if (stack.size() != 1 || _phase->is_dominator(head, c)) {
2310           for (;;) {
2311             if (c->is_Region()) {
2312               stack.push(c, 1);
2313               break;
2314             } else if (c->is_SafePoint() && !c->is_CallLeaf()) {
2315               Node* m = c->in(TypeFunc::Memory);
2316               if (m->is_MergeMem()) {
2317                 m = m->as_MergeMem()->memory_at(_alias);
2318               }
2319               assert(mem == nullptr || mem == m, "several memory states");
2320               mem = m;
2321               break;
2322             } else {
2323               assert(c != c->in(0), "");
2324               c = c->in(0);
2325             }
2326           }
2327         }
2328       }
2329     } while (stack.size() > 0);
2330     assert(mem != nullptr, "should have found safepoint");
2331   } else {
2332     mem = phi_mem;
2333   }
2334   return mem;
2335 }
2336 
2337 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
2338   Node* c = _phase->get_ctrl(n);
2339   if (n->is_Proj() && n->in(0) != nullptr && n->in(0)->is_Call()) {
2340     assert(c == n->in(0), "");
2341     CallNode* call = c->as_Call();
2342     CallProjections projs;
2343     call->extract_projections(&projs, true, false);
2344     if (projs.catchall_memproj != nullptr) {
2345       if (projs.fallthrough_memproj == n) {
2346         c = projs.fallthrough_catchproj;
2347       } else {
2348         assert(projs.catchall_memproj == n, "");
2349         c = projs.catchall_catchproj;
2350       }
2351     }
2352   }
2353   return c;
2354 }
2355 
2356 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
2357   if (_phase->has_ctrl(n))
2358     return get_ctrl(n);
2359   else {
2360     assert (n->is_CFG(), "must be a CFG node");
2361     return n;
2362   }
2363 }
2364 
2365 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
2366   return m != nullptr && get_ctrl(m) == c;
2367 }
2368 
2369 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
2370   assert(n == nullptr || _phase->ctrl_or_self(n) == ctrl, "");
2371   assert(!ctrl->is_Call() || ctrl == n, "projection expected");
2372 #ifdef ASSERT
2373   if ((ctrl->is_Proj() && ctrl->in(0)->is_Call()) ||
2374       (ctrl->is_Catch() && ctrl->in(0)->in(0)->is_Call())) {
2375     CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_Call() : ctrl->in(0)->in(0)->as_Call();
2376     int mems = 0;
2377     for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
2378       Node* u = call->fast_out(i);
2379       if (u->bottom_type() == Type::MEMORY) {
2380         mems++;
2381       }
2382     }
2383     assert(mems <= 1, "No node right after call if multiple mem projections");
2384   }
2385 #endif
2386   Node* mem = _memory_nodes[ctrl->_idx];
2387   Node* c = ctrl;
2388   while (!mem_is_valid(mem, c) &&
2389          (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
2390     c = _phase->idom(c);
2391     mem = _memory_nodes[c->_idx];
2392   }
2393   if (n != nullptr && mem_is_valid(mem, c)) {
2394     while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
2395       mem = next_mem(mem, _alias);
2396     }
2397     if (mem->is_MergeMem()) {
2398       mem = mem->as_MergeMem()->memory_at(_alias);
2399     }
2400     if (!mem_is_valid(mem, c)) {
2401       do {
2402         c = _phase->idom(c);
2403         mem = _memory_nodes[c->_idx];
2404       } while (!mem_is_valid(mem, c) &&
2405                (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
2406     }
2407   }
2408   assert(mem->bottom_type() == Type::MEMORY, "");
2409   return mem;
2410 }
2411 
2412 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
2413   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2414     Node* use = region->fast_out(i);
2415     if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
2416         (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
2417       return true;
2418     }
2419   }
2420   return false;
2421 }
2422 
2423 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
2424   assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
2425   const bool trace = false;
2426   DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
2427   DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
2428   GrowableArray<Node*> phis;
2429   if (mem_for_ctrl != mem) {
2430     Node* old = mem_for_ctrl;
2431     Node* prev = nullptr;
2432     while (old != mem) {
2433       prev = old;
2434       if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
2435         assert(_alias == Compile::AliasIdxRaw, "");
2436         old = old->in(MemNode::Memory);
2437       } else if (old->Opcode() == Op_SCMemProj) {
2438         assert(_alias == Compile::AliasIdxRaw, "");
2439         old = old->in(0);
2440       } else {
2441         ShouldNotReachHere();
2442       }
2443     }
2444     assert(prev != nullptr, "");
2445     if (new_ctrl != ctrl) {
2446       _memory_nodes.map(ctrl->_idx, mem);
2447       _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2448     }
2449     uint input = (uint)MemNode::Memory;
2450     _phase->igvn().replace_input_of(prev, input, new_mem);
2451   } else {
2452     uses.clear();
2453     _memory_nodes.map(new_ctrl->_idx, new_mem);
2454     uses.push(new_ctrl);
2455     for(uint next = 0; next < uses.size(); next++ ) {
2456       Node *n = uses.at(next);
2457       assert(n->is_CFG(), "");
2458       DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
2459       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2460         Node* u = n->fast_out(i);
2461         if (!u->is_Root() && u->is_CFG() && u != n) {
2462           Node* m = _memory_nodes[u->_idx];
2463           if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) &&
2464               !has_mem_phi(u) &&
2465               u->unique_ctrl_out()->Opcode() != Op_Halt) {
2466             DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
2467             DEBUG_ONLY(if (trace && m != nullptr) { tty->print("ZZZ mem"); m->dump(); });
2468 
2469             if (!mem_is_valid(m, u) || !m->is_Phi()) {
2470               bool push = true;
2471               bool create_phi = true;
2472               if (_phase->is_dominator(new_ctrl, u)) {
2473                 create_phi = false;
2474               }
2475               if (create_phi) {
2476                 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
2477                 _phase->register_new_node(phi, u);
2478                 phis.push(phi);
2479                 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
2480                 if (!mem_is_valid(m, u)) {
2481                   DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
2482                   _memory_nodes.map(u->_idx, phi);
2483                 } else {
2484                   DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
2485                   for (;;) {
2486                     assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), "");
2487                     Node* next = nullptr;
2488                     if (m->is_Proj()) {
2489                       next = m->in(0);
2490                     } else {
2491                       assert(m->is_Mem() || m->is_LoadStore(), "");
2492                       assert(_alias == Compile::AliasIdxRaw, "");
2493                       next = m->in(MemNode::Memory);
2494                     }
2495                     if (_phase->get_ctrl(next) != u) {
2496                       break;
2497                     }
2498                     if (next->is_MergeMem()) {
2499                       assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
2500                       break;
2501                     }
2502                     if (next->is_Phi()) {
2503                       assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
2504                       break;
2505                     }
2506                     m = next;
2507                   }
2508 
2509                   DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
2510                   assert(m->is_Mem() || m->is_LoadStore(), "");
2511                   uint input = (uint)MemNode::Memory;
2512                   _phase->igvn().replace_input_of(m, input, phi);
2513                   push = false;
2514                 }
2515               } else {
2516                 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
2517               }
2518               if (push) {
2519                 uses.push(u);
2520               }
2521             }
2522           } else if (!mem_is_valid(m, u) &&
2523                      !(u->Opcode() == Op_CProj && u->in(0)->is_NeverBranch() && u->as_Proj()->_con == 1)) {
2524             uses.push(u);
2525           }
2526         }
2527       }
2528     }
2529     for (int i = 0; i < phis.length(); i++) {
2530       Node* n = phis.at(i);
2531       Node* r = n->in(0);
2532       DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
2533       for (uint j = 1; j < n->req(); j++) {
2534         Node* m = find_mem(r->in(j), nullptr);
2535         _phase->igvn().replace_input_of(n, j, m);
2536         DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
2537       }
2538     }
2539   }
2540   uint last = _phase->C->unique();
2541   MergeMemNode* mm = nullptr;
2542   int alias = _alias;
2543   DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
2544   // Process loads first to not miss an anti-dependency: if the memory
2545   // edge of a store is updated before a load is processed then an
2546   // anti-dependency may be missed.
2547   for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2548     Node* u = mem->out(i);
2549     if (u->_idx < last && u->is_Load() && _phase->C->get_alias_index(u->adr_type()) == alias) {
2550       Node* m = find_mem(_phase->get_ctrl(u), u);
2551       if (m != mem) {
2552         DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2553         _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2554         --i;
2555       }
2556     }
2557   }
2558   for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2559     Node* u = mem->out(i);
2560     if (u->_idx < last) {
2561       if (u->is_Mem()) {
2562         if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2563           Node* m = find_mem(_phase->get_ctrl(u), u);
2564           if (m != mem) {
2565             DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2566             _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2567             --i;
2568           }
2569         }
2570       } else if (u->is_MergeMem()) {
2571         MergeMemNode* u_mm = u->as_MergeMem();
2572         if (u_mm->memory_at(alias) == mem) {
2573           MergeMemNode* newmm = nullptr;
2574           for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2575             Node* uu = u->fast_out(j);
2576             assert(!uu->is_MergeMem(), "chain of MergeMems?");
2577             if (uu->is_Phi()) {
2578               assert(uu->adr_type() == TypePtr::BOTTOM, "");
2579               Node* region = uu->in(0);
2580               int nb = 0;
2581               for (uint k = 1; k < uu->req(); k++) {
2582                 if (uu->in(k) == u) {
2583                   Node* m = find_mem(region->in(k), nullptr);
2584                   if (m != mem) {
2585                     DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
2586                     newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2587                     if (newmm != u) {
2588                       _phase->igvn().replace_input_of(uu, k, newmm);
2589                       nb++;
2590                       --jmax;
2591                     }
2592                   }
2593                 }
2594               }
2595               if (nb > 0) {
2596                 --j;
2597               }
2598             } else {
2599               Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
2600               if (m != mem) {
2601                 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
2602                 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2603                 if (newmm != u) {
2604                   _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2605                   --j, --jmax;
2606                 }
2607               }
2608             }
2609           }
2610         }
2611       } else if (u->is_Phi()) {
2612         assert(u->bottom_type() == Type::MEMORY, "what else?");
2613         if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
2614           Node* region = u->in(0);
2615           bool replaced = false;
2616           for (uint j = 1; j < u->req(); j++) {
2617             if (u->in(j) == mem) {
2618               Node* m = find_mem(region->in(j), nullptr);
2619               Node* nnew = m;
2620               if (m != mem) {
2621                 if (u->adr_type() == TypePtr::BOTTOM) {
2622                   mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
2623                   nnew = mm;
2624                 }
2625                 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
2626                 _phase->igvn().replace_input_of(u, j, nnew);
2627                 replaced = true;
2628               }
2629             }
2630           }
2631           if (replaced) {
2632             --i;
2633           }
2634         }
2635       } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2636                  u->adr_type() == nullptr) {
2637         assert(u->adr_type() != nullptr ||
2638                u->Opcode() == Op_Rethrow ||
2639                u->Opcode() == Op_Return ||
2640                u->Opcode() == Op_SafePoint ||
2641                (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2642                (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2643                u->Opcode() == Op_CallLeaf, "");
2644         Node* m = find_mem(_phase->ctrl_or_self(u), u);
2645         if (m != mem) {
2646           mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
2647           _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2648           --i;
2649         }
2650       } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2651         Node* m = find_mem(_phase->ctrl_or_self(u), u);
2652         if (m != mem) {
2653           DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2654           _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2655           --i;
2656         }
2657       } else if (u->adr_type() != TypePtr::BOTTOM &&
2658                  _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
2659         Node* m = find_mem(_phase->ctrl_or_self(u), u);
2660         assert(m != mem, "");
2661         // u is on the wrong slice...
2662         assert(u->is_ClearArray(), "");
2663         DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2664         _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2665         --i;
2666       }
2667     }
2668   }
2669 #ifdef ASSERT
2670   assert(new_mem->outcnt() > 0, "");
2671   for (int i = 0; i < phis.length(); i++) {
2672     Node* n = phis.at(i);
2673     assert(n->outcnt() > 0, "new phi must have uses now");
2674   }
2675 #endif
2676 }
2677 
2678 void MemoryGraphFixer::record_new_ctrl(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl) {
2679   if (mem_for_ctrl != mem && new_ctrl != ctrl) {
2680     _memory_nodes.map(ctrl->_idx, mem);
2681     _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2682   }
2683 }
2684 
2685 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
2686   MergeMemNode* mm = MergeMemNode::make(mem);
2687   mm->set_memory_at(_alias, rep_proj);
2688   _phase->register_new_node(mm, rep_ctrl);
2689   return mm;
2690 }
2691 
2692 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
2693   MergeMemNode* newmm = nullptr;
2694   MergeMemNode* u_mm = u->as_MergeMem();
2695   Node* c = _phase->get_ctrl(u);
2696   if (_phase->is_dominator(c, rep_ctrl)) {
2697     c = rep_ctrl;
2698   } else {
2699     assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
2700   }
2701   if (u->outcnt() == 1) {
2702     if (u->req() > (uint)_alias && u->in(_alias) == mem) {
2703       _phase->igvn().replace_input_of(u, _alias, rep_proj);
2704       --i;
2705     } else {
2706       _phase->igvn().rehash_node_delayed(u);
2707       u_mm->set_memory_at(_alias, rep_proj);
2708     }
2709     newmm = u_mm;
2710     _phase->set_ctrl_and_loop(u, c);
2711   } else {
2712     // can't simply clone u and then change one of its input because
2713     // it adds and then removes an edge which messes with the
2714     // DUIterator
2715     newmm = MergeMemNode::make(u_mm->base_memory());
2716     for (uint j = 0; j < u->req(); j++) {
2717       if (j < newmm->req()) {
2718         if (j == (uint)_alias) {
2719           newmm->set_req(j, rep_proj);
2720         } else if (newmm->in(j) != u->in(j)) {
2721           newmm->set_req(j, u->in(j));
2722         }
2723       } else if (j == (uint)_alias) {
2724         newmm->add_req(rep_proj);
2725       } else {
2726         newmm->add_req(u->in(j));
2727       }
2728     }
2729     if ((uint)_alias >= u->req()) {
2730       newmm->set_memory_at(_alias, rep_proj);
2731     }
2732     _phase->register_new_node(newmm, c);
2733   }
2734   return newmm;
2735 }
2736 
2737 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
2738   if (phi->adr_type() == TypePtr::BOTTOM) {
2739     Node* region = phi->in(0);
2740     for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
2741       Node* uu = region->fast_out(j);
2742       if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
2743         return false;
2744       }
2745     }
2746     return true;
2747   }
2748   return _phase->C->get_alias_index(phi->adr_type()) == _alias;
2749 }
2750 
2751 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
2752   uint last = _phase-> C->unique();
2753   MergeMemNode* mm = nullptr;
2754   assert(mem->bottom_type() == Type::MEMORY, "");
2755   for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2756     Node* u = mem->out(i);
2757     if (u != replacement && u->_idx < last) {
2758       if (u->is_MergeMem()) {
2759         MergeMemNode* u_mm = u->as_MergeMem();
2760         if (u_mm->memory_at(_alias) == mem) {
2761           MergeMemNode* newmm = nullptr;
2762           for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2763             Node* uu = u->fast_out(j);
2764             assert(!uu->is_MergeMem(), "chain of MergeMems?");
2765             if (uu->is_Phi()) {
2766               if (should_process_phi(uu)) {
2767                 Node* region = uu->in(0);
2768                 int nb = 0;
2769                 for (uint k = 1; k < uu->req(); k++) {
2770                   if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
2771                     if (newmm == nullptr) {
2772                       newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2773                     }
2774                     if (newmm != u) {
2775                       _phase->igvn().replace_input_of(uu, k, newmm);
2776                       nb++;
2777                       --jmax;
2778                     }
2779                   }
2780                 }
2781                 if (nb > 0) {
2782                   --j;
2783                 }
2784               }
2785             } else {
2786               if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
2787                 if (newmm == nullptr) {
2788                   newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2789                 }
2790                 if (newmm != u) {
2791                   _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2792                   --j, --jmax;
2793                 }
2794               }
2795             }
2796           }
2797         }
2798       } else if (u->is_Phi()) {
2799         assert(u->bottom_type() == Type::MEMORY, "what else?");
2800         Node* region = u->in(0);
2801         if (should_process_phi(u)) {
2802           bool replaced = false;
2803           for (uint j = 1; j < u->req(); j++) {
2804             if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
2805               Node* nnew = rep_proj;
2806               if (u->adr_type() == TypePtr::BOTTOM) {
2807                 if (mm == nullptr) {
2808                   mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2809                 }
2810                 nnew = mm;
2811               }
2812               _phase->igvn().replace_input_of(u, j, nnew);
2813               replaced = true;
2814             }
2815           }
2816           if (replaced) {
2817             --i;
2818           }
2819 
2820         }
2821       } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2822                  u->adr_type() == nullptr) {
2823         assert(u->adr_type() != nullptr ||
2824                u->Opcode() == Op_Rethrow ||
2825                u->Opcode() == Op_Return ||
2826                u->Opcode() == Op_SafePoint ||
2827                (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2828                (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2829                u->Opcode() == Op_CallLeaf, "%s", u->Name());
2830         if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2831           if (mm == nullptr) {
2832             mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2833           }
2834           _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2835           --i;
2836         }
2837       } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2838         if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2839           _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
2840           --i;
2841         }
2842       }
2843     }
2844   }
2845 }
2846 
2847 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj, DecoratorSet decorators)
2848 : Node(ctrl, obj), _decorators(decorators) {
2849   ShenandoahBarrierSetC2::bsc2()->state()->add_load_reference_barrier(this);
2850 }
2851 
2852 DecoratorSet ShenandoahLoadReferenceBarrierNode::decorators() const {
2853   return _decorators;
2854 }
2855 
2856 uint ShenandoahLoadReferenceBarrierNode::size_of() const {
2857   return sizeof(*this);
2858 }
2859 
2860 static DecoratorSet mask_decorators(DecoratorSet decorators) {
2861   return decorators & (ON_STRONG_OOP_REF | ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF | ON_UNKNOWN_OOP_REF | IN_NATIVE);
2862 }
2863 
2864 uint ShenandoahLoadReferenceBarrierNode::hash() const {
2865   uint hash = Node::hash();
2866   hash += mask_decorators(_decorators);
2867   return hash;
2868 }
2869 
2870 bool ShenandoahLoadReferenceBarrierNode::cmp( const Node &n ) const {
2871   return Node::cmp(n) && n.Opcode() == Op_ShenandoahLoadReferenceBarrier &&
2872          mask_decorators(_decorators) == mask_decorators(((const ShenandoahLoadReferenceBarrierNode&)n)._decorators);
2873 }
2874 
2875 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const {
2876   if (in(ValueIn) == nullptr || in(ValueIn)->is_top()) {
2877     return Type::TOP;
2878   }
2879   const Type* t = in(ValueIn)->bottom_type();
2880   if (t == TypePtr::NULL_PTR) {
2881     return t;
2882   }
2883 
2884   if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2885     return t;
2886   }
2887 
2888   return t->meet(TypePtr::NULL_PTR);
2889 }
2890 
2891 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseGVN* phase) const {
2892   // Either input is TOP ==> the result is TOP
2893   const Type *t2 = phase->type(in(ValueIn));
2894   if( t2 == Type::TOP ) return Type::TOP;
2895 
2896   if (t2 == TypePtr::NULL_PTR) {
2897     return t2;
2898   }
2899 
2900   if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2901     return t2;
2902   }
2903 
2904   return t2->meet(TypePtr::NULL_PTR);
2905 }
2906 
2907 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseGVN* phase) {
2908   Node* value = in(ValueIn);
2909   if (!needs_barrier(phase, value)) {
2910     return value;
2911   }
2912   return this;
2913 }
2914 
2915 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseGVN* phase, Node* n) {
2916   Unique_Node_List visited;
2917   return needs_barrier_impl(phase, n, visited);
2918 }
2919 
2920 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseGVN* phase, Node* n, Unique_Node_List &visited) {
2921   if (n == nullptr) return false;
2922   if (visited.member(n)) {
2923     return false; // Been there.
2924   }
2925   visited.push(n);
2926 
2927   if (n->is_Allocate()) {
2928     // tty->print_cr("optimize barrier on alloc");
2929     return false;
2930   }
2931   if (n->is_Call()) {
2932     // tty->print_cr("optimize barrier on call");
2933     return false;
2934   }
2935 
2936   const Type* type = phase->type(n);
2937   if (type == Type::TOP) {
2938     return false;
2939   }
2940   if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) {
2941     // tty->print_cr("optimize barrier on null");
2942     return false;
2943   }
2944   if (type->make_oopptr() && type->make_oopptr()->const_oop() != nullptr) {
2945     // tty->print_cr("optimize barrier on constant");
2946     return false;
2947   }
2948 
2949   switch (n->Opcode()) {
2950     case Op_AddP:
2951       return true; // TODO: Can refine?
2952     case Op_LoadP:
2953     case Op_ShenandoahCompareAndExchangeN:
2954     case Op_ShenandoahCompareAndExchangeP:
2955     case Op_CompareAndExchangeN:
2956     case Op_CompareAndExchangeP:
2957     case Op_GetAndSetN:
2958     case Op_GetAndSetP:
2959       return true;
2960     case Op_Phi: {
2961       for (uint i = 1; i < n->req(); i++) {
2962         if (needs_barrier_impl(phase, n->in(i), visited)) return true;
2963       }
2964       return false;
2965     }
2966     case Op_CheckCastPP:
2967     case Op_CastPP:
2968       return needs_barrier_impl(phase, n->in(1), visited);
2969     case Op_Proj:
2970       return needs_barrier_impl(phase, n->in(0), visited);
2971     case Op_ShenandoahLoadReferenceBarrier:
2972       // tty->print_cr("optimize barrier on barrier");
2973       return false;
2974     case Op_Parm:
2975       // tty->print_cr("optimize barrier on input arg");
2976       return false;
2977     case Op_DecodeN:
2978     case Op_EncodeP:
2979       return needs_barrier_impl(phase, n->in(1), visited);
2980     case Op_LoadN:
2981       return true;
2982     case Op_CMoveN:
2983     case Op_CMoveP:
2984       return needs_barrier_impl(phase, n->in(2), visited) ||
2985              needs_barrier_impl(phase, n->in(3), visited);
2986     case Op_ShenandoahIUBarrier:
2987       return needs_barrier_impl(phase, n->in(1), visited);
2988     case Op_CreateEx:
2989       return false;
2990     default:
2991       break;
2992   }
2993 #ifdef ASSERT
2994   tty->print("need barrier on?: ");
2995   tty->print_cr("ins:");
2996   n->dump(2);
2997   tty->print_cr("outs:");
2998   n->dump(-2);
2999   ShouldNotReachHere();
3000 #endif
3001   return true;
3002 }