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     C->clear_major_progress();
  54     PhaseIdealLoop::optimize(igvn, LoopOptsShenandoahExpand);
  55     if (C->failing()) return false;
  56 
  57     C->set_major_progress();
  58     if (!C->optimize_loops(igvn, LoopOptsShenandoahPostExpand)) {
  59       return false;
  60     }
  61     C->clear_major_progress();
  62     C->process_for_post_loop_opts_igvn(igvn);
  63     if (C->failing()) return false;
  64 
  65     C->set_post_loop_opts_phase(); // now for real!
  66   }
  67   return true;
  68 }
  69 
  70 bool ShenandoahBarrierC2Support::is_gc_state_test(Node* iff, int mask) {
  71   if (!UseShenandoahGC) {
  72     return false;
  73   }
  74   assert(iff->is_If(), "bad input");
  75   if (iff->Opcode() != Op_If) {
  76     return false;
  77   }
  78   Node* bol = iff->in(1);
  79   if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) {
  80     return false;
  81   }
  82   Node* cmp = bol->in(1);
  83   if (cmp->Opcode() != Op_CmpI) {
  84     return false;
  85   }
  86   Node* in1 = cmp->in(1);
  87   Node* in2 = cmp->in(2);
  88   if (in2->find_int_con(-1) != 0) {
  89     return false;
  90   }
  91   if (in1->Opcode() != Op_AndI) {
  92     return false;
  93   }
  94   in2 = in1->in(2);
  95   if (in2->find_int_con(-1) != mask) {
  96     return false;
  97   }
  98   in1 = in1->in(1);
  99 
 100   return is_gc_state_load(in1);
 101 }
 102 
 103 bool ShenandoahBarrierC2Support::is_heap_stable_test(Node* iff) {
 104   return is_gc_state_test(iff, ShenandoahHeap::HAS_FORWARDED);
 105 }
 106 
 107 bool ShenandoahBarrierC2Support::is_gc_state_load(Node *n) {
 108   if (!UseShenandoahGC) {
 109     return false;
 110   }
 111   if (n->Opcode() != Op_LoadB && n->Opcode() != Op_LoadUB) {
 112     return false;
 113   }
 114   Node* addp = n->in(MemNode::Address);
 115   if (!addp->is_AddP()) {
 116     return false;
 117   }
 118   Node* base = addp->in(AddPNode::Address);
 119   Node* off = addp->in(AddPNode::Offset);
 120   if (base->Opcode() != Op_ThreadLocal) {
 121     return false;
 122   }
 123   if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) {
 124     return false;
 125   }
 126   return true;
 127 }
 128 
 129 bool ShenandoahBarrierC2Support::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) {
 130   assert(phase->is_dominator(stop, start), "bad inputs");
 131   ResourceMark rm;
 132   Unique_Node_List wq;
 133   wq.push(start);
 134   for (uint next = 0; next < wq.size(); next++) {
 135     Node *m = wq.at(next);
 136     if (m == stop) {
 137       continue;
 138     }
 139     if (m->is_SafePoint() && !m->is_CallLeaf()) {
 140       return true;
 141     }
 142     if (m->is_Region()) {
 143       for (uint i = 1; i < m->req(); i++) {
 144         wq.push(m->in(i));
 145       }
 146     } else {
 147       wq.push(m->in(0));
 148     }
 149   }
 150   return false;
 151 }
 152 
 153 #ifdef ASSERT
 154 bool ShenandoahBarrierC2Support::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) {
 155   assert(phis.size() == 0, "");
 156 
 157   while (true) {
 158     if (in->bottom_type() == TypePtr::NULL_PTR) {
 159       if (trace) {tty->print_cr("null");}
 160     } else if (!in->bottom_type()->make_ptr()->make_oopptr()) {
 161       if (trace) {tty->print_cr("Non oop");}
 162     } else {
 163       if (in->is_ConstraintCast()) {
 164         in = in->in(1);
 165         continue;
 166       } else if (in->is_AddP()) {
 167         assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access");
 168         in = in->in(AddPNode::Address);
 169         continue;
 170       } else if (in->is_Con()) {
 171         if (trace) {
 172           tty->print("Found constant");
 173           in->dump();
 174         }
 175       } else if (in->Opcode() == Op_Parm) {
 176         if (trace) {
 177           tty->print("Found argument");
 178         }
 179       } else if (in->Opcode() == Op_CreateEx) {
 180         if (trace) {
 181           tty->print("Found create-exception");
 182         }
 183       } else if (in->Opcode() == Op_LoadP && in->adr_type() == TypeRawPtr::BOTTOM) {
 184         if (trace) {
 185           tty->print("Found raw LoadP (OSR argument?)");
 186         }
 187       } else if (in->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 188         if (t == ShenandoahOopStore) {
 189           uint i = 0;
 190           for (; i < phis.size(); i++) {
 191             Node* n = phis.node_at(i);
 192             if (n->Opcode() == Op_ShenandoahIUBarrier) {
 193               break;
 194             }
 195           }
 196           if (i == phis.size()) {
 197             return false;
 198           }
 199         }
 200         barriers_used.push(in);
 201         if (trace) {tty->print("Found barrier"); in->dump();}
 202       } else if (in->Opcode() == Op_ShenandoahIUBarrier) {
 203         if (t != ShenandoahOopStore) {
 204           in = in->in(1);
 205           continue;
 206         }
 207         if (trace) {tty->print("Found enqueue barrier"); in->dump();}
 208         phis.push(in, in->req());
 209         in = in->in(1);
 210         continue;
 211       } else if (in->is_Proj() && in->in(0)->is_Allocate()) {
 212         if (trace) {
 213           tty->print("Found alloc");
 214           in->in(0)->dump();
 215         }
 216       } else if (in->is_Proj() && (in->in(0)->Opcode() == Op_CallStaticJava || in->in(0)->Opcode() == Op_CallDynamicJava)) {
 217         if (trace) {
 218           tty->print("Found Java call");
 219         }
 220       } else if (in->is_Phi()) {
 221         if (!visited.test_set(in->_idx)) {
 222           if (trace) {tty->print("Pushed phi:"); in->dump();}
 223           phis.push(in, 2);
 224           in = in->in(1);
 225           continue;
 226         }
 227         if (trace) {tty->print("Already seen phi:"); in->dump();}
 228       } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) {
 229         if (!visited.test_set(in->_idx)) {
 230           if (trace) {tty->print("Pushed cmovep:"); in->dump();}
 231           phis.push(in, CMoveNode::IfTrue);
 232           in = in->in(CMoveNode::IfFalse);
 233           continue;
 234         }
 235         if (trace) {tty->print("Already seen cmovep:"); in->dump();}
 236       } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) {
 237         in = in->in(1);
 238         continue;
 239       } else {
 240         return false;
 241       }
 242     }
 243     bool cont = false;
 244     while (phis.is_nonempty()) {
 245       uint idx = phis.index();
 246       Node* phi = phis.node();
 247       if (idx >= phi->req()) {
 248         if (trace) {tty->print("Popped phi:"); phi->dump();}
 249         phis.pop();
 250         continue;
 251       }
 252       if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();}
 253       in = phi->in(idx);
 254       phis.set_index(idx+1);
 255       cont = true;
 256       break;
 257     }
 258     if (!cont) {
 259       break;
 260     }
 261   }
 262   return true;
 263 }
 264 
 265 void ShenandoahBarrierC2Support::report_verify_failure(const char* msg, Node* n1, Node* n2) {
 266   if (n1 != nullptr) {
 267     n1->dump(+10);
 268   }
 269   if (n2 != nullptr) {
 270     n2->dump(+10);
 271   }
 272   fatal("%s", msg);
 273 }
 274 
 275 void ShenandoahBarrierC2Support::verify(RootNode* root) {
 276   ResourceMark rm;
 277   Unique_Node_List wq;
 278   GrowableArray<Node*> barriers;
 279   Unique_Node_List barriers_used;
 280   Node_Stack phis(0);
 281   VectorSet visited;
 282   const bool trace = false;
 283   const bool verify_no_useless_barrier = false;
 284 
 285   wq.push(root);
 286   for (uint next = 0; next < wq.size(); next++) {
 287     Node *n = wq.at(next);
 288     if (n->is_Load()) {
 289       const bool trace = false;
 290       if (trace) {tty->print("Verifying"); n->dump();}
 291       if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) {
 292         if (trace) {tty->print_cr("Load range/klass");}
 293       } else {
 294         const TypePtr* adr_type = n->as_Load()->adr_type();
 295 
 296         if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) {
 297           if (trace) {tty->print_cr("Mark load");}
 298         } else if (adr_type->isa_instptr() &&
 299                    adr_type->is_instptr()->instance_klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
 300                    adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset()) {
 301           if (trace) {tty->print_cr("Reference.get()");}
 302         } else if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) {
 303           report_verify_failure("Shenandoah verification: Load should have barriers", n);
 304         }
 305       }
 306     } else if (n->is_Store()) {
 307       const bool trace = false;
 308 
 309       if (trace) {tty->print("Verifying"); n->dump();}
 310       if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) {
 311         Node* adr = n->in(MemNode::Address);
 312         bool verify = true;
 313 
 314         if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) {
 315           adr = adr->in(AddPNode::Address);
 316           if (adr->is_AddP()) {
 317             assert(adr->in(AddPNode::Base)->is_top(), "");
 318             adr = adr->in(AddPNode::Address);
 319             if (adr->Opcode() == Op_LoadP &&
 320                 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() &&
 321                 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
 322                 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) {
 323               if (trace) {tty->print_cr("SATB prebarrier");}
 324               verify = false;
 325             }
 326           }
 327         }
 328 
 329         if (verify && !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahIUBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
 330           report_verify_failure("Shenandoah verification: Store should have barriers", n);
 331         }
 332       }
 333       if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
 334         report_verify_failure("Shenandoah verification: Store (address) should have barriers", n);
 335       }
 336     } else if (n->Opcode() == Op_CmpP) {
 337       const bool trace = false;
 338 
 339       Node* in1 = n->in(1);
 340       Node* in2 = n->in(2);
 341       if (in1->bottom_type()->isa_oopptr()) {
 342         if (trace) {tty->print("Verifying"); n->dump();}
 343 
 344         bool mark_inputs = false;
 345         if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR ||
 346             (in1->is_Con() || in2->is_Con())) {
 347           if (trace) {tty->print_cr("Comparison against a constant");}
 348           mark_inputs = true;
 349         } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) ||
 350                    (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) {
 351           if (trace) {tty->print_cr("Comparison with newly alloc'ed object");}
 352           mark_inputs = true;
 353         } else {
 354           assert(in2->bottom_type()->isa_oopptr(), "");
 355 
 356           if (!verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) ||
 357               !verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) {
 358             report_verify_failure("Shenandoah verification: Cmp should have barriers", n);
 359           }
 360         }
 361         if (verify_no_useless_barrier &&
 362             mark_inputs &&
 363             (!verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) ||
 364              !verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) {
 365           phis.clear();
 366           visited.reset();
 367         }
 368       }
 369     } else if (n->is_LoadStore()) {
 370       if (n->in(MemNode::ValueIn)->bottom_type()->make_ptr() &&
 371           !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahIUBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
 372         report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n);
 373       }
 374 
 375       if (n->in(MemNode::Address)->bottom_type()->make_oopptr() && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
 376         report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n);
 377       }
 378     } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) {
 379       CallNode* call = n->as_Call();
 380 
 381       static struct {
 382         const char* name;
 383         struct {
 384           int pos;
 385           verify_type t;
 386         } args[6];
 387       } calls[] = {
 388         "array_partition_stub",
 389         { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+4, ShenandoahStore },   { -1, ShenandoahNone },
 390           { -1, ShenandoahNone },                { -1, ShenandoahNone },                  { -1, ShenandoahNone } },
 391         "arraysort_stub",
 392         { { TypeFunc::Parms, ShenandoahStore },  { -1, ShenandoahNone },                  { -1, ShenandoahNone },
 393           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 394         "aescrypt_encryptBlock",
 395         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 396           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 397         "aescrypt_decryptBlock",
 398         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 399           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 400         "multiplyToLen",
 401         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },   { TypeFunc::Parms+4, ShenandoahStore },
 402           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 403         "squareToLen",
 404         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },   { -1,  ShenandoahNone},
 405           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 406         "montgomery_multiply",
 407         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },
 408           { TypeFunc::Parms+6, ShenandoahStore }, { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 409         "montgomery_square",
 410         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahLoad },   { TypeFunc::Parms+5, ShenandoahStore },
 411           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 412         "mulAdd",
 413         { { TypeFunc::Parms, ShenandoahStore },  { TypeFunc::Parms+1, ShenandoahLoad },   { -1,  ShenandoahNone},
 414           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 415         "vectorizedMismatch",
 416         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahLoad },   { -1,  ShenandoahNone},
 417           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 418         "updateBytesCRC32",
 419         { { TypeFunc::Parms+1, ShenandoahLoad }, { -1,  ShenandoahNone},                  { -1,  ShenandoahNone},
 420           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 421         "updateBytesAdler32",
 422         { { TypeFunc::Parms+1, ShenandoahLoad }, { -1,  ShenandoahNone},                  { -1,  ShenandoahNone},
 423           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 424         "updateBytesCRC32C",
 425         { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad},    { -1,  ShenandoahNone},
 426           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 427         "counterMode_AESCrypt",
 428         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 429           { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } },
 430         "cipherBlockChaining_encryptAESCrypt",
 431         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 432           { TypeFunc::Parms+3, ShenandoahLoad },  { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 433         "cipherBlockChaining_decryptAESCrypt",
 434         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 435           { TypeFunc::Parms+3, ShenandoahLoad },  { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 436         "shenandoah_clone_barrier",
 437         { { TypeFunc::Parms, ShenandoahLoad },   { -1,  ShenandoahNone},                  { -1,  ShenandoahNone},
 438           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 439         "ghash_processBlocks",
 440         { { TypeFunc::Parms, ShenandoahStore },  { TypeFunc::Parms+1, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },
 441           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 442         "sha1_implCompress",
 443         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 444           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 445         "sha256_implCompress",
 446         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 447           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 448         "sha512_implCompress",
 449         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 450           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 451         "sha1_implCompressMB",
 452         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 453           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 454         "sha256_implCompressMB",
 455         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 456           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 457         "sha512_implCompressMB",
 458         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 459           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 460         "encodeBlock",
 461         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+3, ShenandoahStore },   { -1, ShenandoahNone },
 462           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 463         "decodeBlock",
 464         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+3, ShenandoahStore },   { -1, ShenandoahNone },
 465           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 466         "intpoly_montgomeryMult_P256",
 467         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahLoad  },   { TypeFunc::Parms+2, ShenandoahStore },
 468           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 469         "intpoly_assign",
 470         { { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },  { -1, ShenandoahNone },
 471           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 472       };
 473 
 474       if (call->is_call_to_arraycopystub()) {
 475         Node* dest = nullptr;
 476         const TypeTuple* args = n->as_Call()->_tf->domain();
 477         for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
 478           if (args->field_at(i)->isa_ptr()) {
 479             j++;
 480             if (j == 2) {
 481               dest = n->in(i);
 482               break;
 483             }
 484           }
 485         }
 486         if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) ||
 487             !verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) {
 488           report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n);
 489         }
 490       } else if (strlen(call->_name) > 5 &&
 491                  !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) {
 492         if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) {
 493           report_verify_failure("Shenandoah verification: _fill should have barriers", n);
 494         }
 495       } else if (!strcmp(call->_name, "shenandoah_wb_pre")) {
 496         // skip
 497       } else {
 498         const int calls_len = sizeof(calls) / sizeof(calls[0]);
 499         int i = 0;
 500         for (; i < calls_len; i++) {
 501           if (!strcmp(calls[i].name, call->_name)) {
 502             break;
 503           }
 504         }
 505         if (i != calls_len) {
 506           const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]);
 507           for (uint j = 0; j < args_len; j++) {
 508             int pos = calls[i].args[j].pos;
 509             if (pos == -1) {
 510               break;
 511             }
 512             if (!verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) {
 513               report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
 514             }
 515           }
 516           for (uint j = TypeFunc::Parms; j < call->req(); j++) {
 517             if (call->in(j)->bottom_type()->make_ptr() &&
 518                 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
 519               uint k = 0;
 520               for (; k < args_len && calls[i].args[k].pos != (int)j; k++);
 521               if (k == args_len) {
 522                 fatal("arg %d for call %s not covered", j, call->_name);
 523               }
 524             }
 525           }
 526         } else {
 527           for (uint j = TypeFunc::Parms; j < call->req(); j++) {
 528             if (call->in(j)->bottom_type()->make_ptr() &&
 529                 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
 530               fatal("%s not covered", call->_name);
 531             }
 532           }
 533         }
 534       }
 535     } else if (n->Opcode() == Op_ShenandoahIUBarrier || n->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 536       // skip
 537     } else if (n->is_AddP()
 538                || n->is_Phi()
 539                || n->is_ConstraintCast()
 540                || n->Opcode() == Op_Return
 541                || n->Opcode() == Op_CMoveP
 542                || n->Opcode() == Op_CMoveN
 543                || n->Opcode() == Op_Rethrow
 544                || n->is_MemBar()
 545                || n->Opcode() == Op_Conv2B
 546                || n->Opcode() == Op_SafePoint
 547                || n->is_CallJava()
 548                || n->Opcode() == Op_Unlock
 549                || n->Opcode() == Op_EncodeP
 550                || n->Opcode() == Op_DecodeN) {
 551       // nothing to do
 552     } else {
 553       static struct {
 554         int opcode;
 555         struct {
 556           int pos;
 557           verify_type t;
 558         } inputs[2];
 559       } others[] = {
 560         Op_FastLock,
 561         { { 1, ShenandoahLoad },                  { -1, ShenandoahNone} },
 562         Op_Lock,
 563         { { TypeFunc::Parms, ShenandoahLoad },    { -1, ShenandoahNone} },
 564         Op_ArrayCopy,
 565         { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } },
 566         Op_StrCompressedCopy,
 567         { { 2, ShenandoahLoad },                  { 3, ShenandoahStore } },
 568         Op_StrInflatedCopy,
 569         { { 2, ShenandoahLoad },                  { 3, ShenandoahStore } },
 570         Op_AryEq,
 571         { { 2, ShenandoahLoad },                  { 3, ShenandoahLoad } },
 572         Op_StrIndexOf,
 573         { { 2, ShenandoahLoad },                  { 4, ShenandoahLoad } },
 574         Op_StrComp,
 575         { { 2, ShenandoahLoad },                  { 4, ShenandoahLoad } },
 576         Op_StrEquals,
 577         { { 2, ShenandoahLoad },                  { 3, ShenandoahLoad } },
 578         Op_VectorizedHashCode,
 579         { { 2, ShenandoahLoad },                  { -1, ShenandoahNone } },
 580         Op_EncodeISOArray,
 581         { { 2, ShenandoahLoad },                  { 3, ShenandoahStore } },
 582         Op_CountPositives,
 583         { { 2, ShenandoahLoad },                  { -1, ShenandoahNone} },
 584         Op_CastP2X,
 585         { { 1, ShenandoahLoad },                  { -1, ShenandoahNone} },
 586         Op_StrIndexOfChar,
 587         { { 2, ShenandoahLoad },                  { -1, ShenandoahNone } },
 588       };
 589 
 590       const int others_len = sizeof(others) / sizeof(others[0]);
 591       int i = 0;
 592       for (; i < others_len; i++) {
 593         if (others[i].opcode == n->Opcode()) {
 594           break;
 595         }
 596       }
 597       uint stop = n->is_Call() ? n->as_Call()->tf()->domain()->cnt() : n->req();
 598       if (i != others_len) {
 599         const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]);
 600         for (uint j = 0; j < inputs_len; j++) {
 601           int pos = others[i].inputs[j].pos;
 602           if (pos == -1) {
 603             break;
 604           }
 605           if (!verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) {
 606             report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
 607           }
 608         }
 609         for (uint j = 1; j < stop; j++) {
 610           if (n->in(j) != nullptr && n->in(j)->bottom_type()->make_ptr() &&
 611               n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
 612             uint k = 0;
 613             for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++);
 614             if (k == inputs_len) {
 615               fatal("arg %d for node %s not covered", j, n->Name());
 616             }
 617           }
 618         }
 619       } else {
 620         for (uint j = 1; j < stop; j++) {
 621           if (n->in(j) != nullptr && n->in(j)->bottom_type()->make_ptr() &&
 622               n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
 623             fatal("%s not covered", n->Name());
 624           }
 625         }
 626       }
 627     }
 628 
 629     if (n->is_SafePoint()) {
 630       SafePointNode* sfpt = n->as_SafePoint();
 631       if (verify_no_useless_barrier && sfpt->jvms() != nullptr) {
 632         for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) {
 633           if (!verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) {
 634             phis.clear();
 635             visited.reset();
 636           }
 637         }
 638       }
 639     }
 640   }
 641 
 642   if (verify_no_useless_barrier) {
 643     for (int i = 0; i < barriers.length(); i++) {
 644       Node* n = barriers.at(i);
 645       if (!barriers_used.member(n)) {
 646         tty->print("XXX useless barrier"); n->dump(-2);
 647         ShouldNotReachHere();
 648       }
 649     }
 650   }
 651 }
 652 #endif
 653 
 654 bool ShenandoahBarrierC2Support::is_dominator_same_ctrl(Node* c, Node* d, Node* n, PhaseIdealLoop* phase) {
 655   // That both nodes have the same control is not sufficient to prove
 656   // domination, verify that there's no path from d to n
 657   ResourceMark rm;
 658   Unique_Node_List wq;
 659   wq.push(d);
 660   for (uint next = 0; next < wq.size(); next++) {
 661     Node *m = wq.at(next);
 662     if (m == n) {
 663       return false;
 664     }
 665     if (m->is_Phi() && m->in(0)->is_Loop()) {
 666       assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control");
 667     } else {
 668       if (m->is_Store() || m->is_LoadStore()) {
 669         // Take anti-dependencies into account
 670         Node* mem = m->in(MemNode::Memory);
 671         for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
 672           Node* u = mem->fast_out(i);
 673           if (u->is_Load() && phase->C->can_alias(m->adr_type(), phase->C->get_alias_index(u->adr_type())) &&
 674               phase->ctrl_or_self(u) == c) {
 675             wq.push(u);
 676           }
 677         }
 678       }
 679       for (uint i = 0; i < m->req(); i++) {
 680         if (m->in(i) != nullptr && phase->ctrl_or_self(m->in(i)) == c) {
 681           wq.push(m->in(i));
 682         }
 683       }
 684     }
 685   }
 686   return true;
 687 }
 688 
 689 bool ShenandoahBarrierC2Support::is_dominator(Node* d_c, Node* n_c, Node* d, Node* n, PhaseIdealLoop* phase) {
 690   if (d_c != n_c) {
 691     return phase->is_dominator(d_c, n_c);
 692   }
 693   return is_dominator_same_ctrl(d_c, d, n, phase);
 694 }
 695 
 696 Node* next_mem(Node* mem, int alias) {
 697   Node* res = nullptr;
 698   if (mem->is_Proj()) {
 699     res = mem->in(0);
 700   } else if (mem->is_SafePoint() || mem->is_MemBar()) {
 701     res = mem->in(TypeFunc::Memory);
 702   } else if (mem->is_Phi()) {
 703     res = mem->in(1);
 704   } else if (mem->is_MergeMem()) {
 705     res = mem->as_MergeMem()->memory_at(alias);
 706   } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
 707     assert(alias == Compile::AliasIdxRaw, "following raw memory can't lead to a barrier");
 708     res = mem->in(MemNode::Memory);
 709   } else {
 710 #ifdef ASSERT
 711     mem->dump();
 712 #endif
 713     ShouldNotReachHere();
 714   }
 715   return res;
 716 }
 717 
 718 Node* ShenandoahBarrierC2Support::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) {
 719   Node* iffproj = nullptr;
 720   while (c != dom) {
 721     Node* next = phase->idom(c);
 722     assert(next->unique_ctrl_out_or_null() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?");
 723     if (c->is_Region()) {
 724       ResourceMark rm;
 725       Unique_Node_List wq;
 726       wq.push(c);
 727       for (uint i = 0; i < wq.size(); i++) {
 728         Node *n = wq.at(i);
 729         if (n == next) {
 730           continue;
 731         }
 732         if (n->is_Region()) {
 733           for (uint j = 1; j < n->req(); j++) {
 734             wq.push(n->in(j));
 735           }
 736         } else {
 737           wq.push(n->in(0));
 738         }
 739       }
 740       for (uint i = 0; i < wq.size(); i++) {
 741         Node *n = wq.at(i);
 742         assert(n->is_CFG(), "");
 743         if (n->is_Multi()) {
 744           for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
 745             Node* u = n->fast_out(j);
 746             if (u->is_CFG()) {
 747               if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj()) {
 748                 return NodeSentinel;
 749               }
 750             }
 751           }
 752         }
 753       }
 754     } else  if (c->is_Proj()) {
 755       if (c->is_IfProj()) {
 756         if (c->as_Proj()->is_uncommon_trap_if_pattern() != nullptr) {
 757           // continue;
 758         } else {
 759           if (!allow_one_proj) {
 760             return NodeSentinel;
 761           }
 762           if (iffproj == nullptr) {
 763             iffproj = c;
 764           } else {
 765             return NodeSentinel;
 766           }
 767         }
 768       } else if (c->Opcode() == Op_JumpProj) {
 769         return NodeSentinel; // unsupported
 770       } else if (c->Opcode() == Op_CatchProj) {
 771         return NodeSentinel; // unsupported
 772       } else if (c->Opcode() == Op_CProj && next->is_NeverBranch()) {
 773         return NodeSentinel; // unsupported
 774       } else {
 775         assert(next->unique_ctrl_out() == c, "unsupported branch pattern");
 776       }
 777     }
 778     c = next;
 779   }
 780   return iffproj;
 781 }
 782 
 783 Node* ShenandoahBarrierC2Support::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) {
 784   ResourceMark rm;
 785   VectorSet wq;
 786   wq.set(mem->_idx);
 787   mem_ctrl = phase->ctrl_or_self(mem);
 788   while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) {
 789     mem = next_mem(mem, alias);
 790     if (wq.test_set(mem->_idx)) {
 791       return nullptr;
 792     }
 793     mem_ctrl = phase->ctrl_or_self(mem);
 794   }
 795   if (mem->is_MergeMem()) {
 796     mem = mem->as_MergeMem()->memory_at(alias);
 797     mem_ctrl = phase->ctrl_or_self(mem);
 798   }
 799   return mem;
 800 }
 801 
 802 Node* ShenandoahBarrierC2Support::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) {
 803   Node* mem = nullptr;
 804   Node* c = ctrl;
 805   do {
 806     if (c->is_Region()) {
 807       for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == nullptr; i++) {
 808         Node* u = c->fast_out(i);
 809         if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
 810           if (u->adr_type() == TypePtr::BOTTOM) {
 811             mem = u;
 812           }
 813         }
 814       }
 815     } else {
 816       if (c->is_Call() && c->as_Call()->adr_type() != nullptr) {
 817         CallProjections projs;
 818         c->as_Call()->extract_projections(&projs, true, false);
 819         if (projs.fallthrough_memproj != nullptr) {
 820           if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) {
 821             if (projs.catchall_memproj == nullptr) {
 822               mem = projs.fallthrough_memproj;
 823             } else {
 824               if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) {
 825                 mem = projs.fallthrough_memproj;
 826               } else {
 827                 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier");
 828                 mem = projs.catchall_memproj;
 829               }
 830             }
 831           }
 832         } else {
 833           Node* proj = c->as_Call()->proj_out(TypeFunc::Memory);
 834           if (proj != nullptr &&
 835               proj->adr_type() == TypePtr::BOTTOM) {
 836             mem = proj;
 837           }
 838         }
 839       } else {
 840         for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
 841           Node* u = c->fast_out(i);
 842           if (u->is_Proj() &&
 843               u->bottom_type() == Type::MEMORY &&
 844               u->adr_type() == TypePtr::BOTTOM) {
 845               assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), "");
 846               assert(mem == nullptr, "only one proj");
 847               mem = u;
 848           }
 849         }
 850         assert(!c->is_Call() || c->as_Call()->adr_type() != nullptr || mem == nullptr, "no mem projection expected");
 851       }
 852     }
 853     c = phase->idom(c);
 854   } while (mem == nullptr);
 855   return mem;
 856 }
 857 
 858 void ShenandoahBarrierC2Support::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) {
 859   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 860     Node* u = n->fast_out(i);
 861     if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) {
 862       uses.push(u);
 863     }
 864   }
 865 }
 866 
 867 static void hide_strip_mined_loop(OuterStripMinedLoopNode* outer, CountedLoopNode* inner, PhaseIdealLoop* phase) {
 868   OuterStripMinedLoopEndNode* le = inner->outer_loop_end();
 869   Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl));
 870   phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl));
 871   Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt);
 872   phase->register_control(new_le, phase->get_loop(le), le->in(0));
 873   phase->lazy_replace(outer, new_outer);
 874   phase->lazy_replace(le, new_le);
 875   inner->clear_strip_mined();
 876 }
 877 
 878 void ShenandoahBarrierC2Support::test_gc_state(Node*& ctrl, Node* raw_mem, Node*& test_fail_ctrl,
 879                                                PhaseIdealLoop* phase, int flags) {
 880   PhaseIterGVN& igvn = phase->igvn();
 881   Node* old_ctrl = ctrl;
 882 
 883   Node* thread          = new ThreadLocalNode();
 884   Node* gc_state_offset = igvn.MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 885   Node* gc_state_addr   = new AddPNode(phase->C->top(), thread, gc_state_offset);
 886   Node* gc_state        = new LoadBNode(old_ctrl, raw_mem, gc_state_addr,
 887                                         DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(nullptr),
 888                                         TypeInt::BYTE, MemNode::unordered);
 889   Node* gc_state_and    = new AndINode(gc_state, igvn.intcon(flags));
 890   Node* gc_state_cmp    = new CmpINode(gc_state_and, igvn.zerocon(T_INT));
 891   Node* gc_state_bool   = new BoolNode(gc_state_cmp, BoolTest::ne);
 892 
 893   IfNode* gc_state_iff  = new IfNode(old_ctrl, gc_state_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
 894   ctrl                  = new IfTrueNode(gc_state_iff);
 895   test_fail_ctrl        = new IfFalseNode(gc_state_iff);
 896 
 897   IdealLoopTree* loop = phase->get_loop(old_ctrl);
 898   phase->register_control(gc_state_iff,   loop, old_ctrl);
 899   phase->register_control(ctrl,           loop, gc_state_iff);
 900   phase->register_control(test_fail_ctrl, loop, gc_state_iff);
 901 
 902   phase->register_new_node(thread,        old_ctrl);
 903   phase->register_new_node(gc_state_addr, old_ctrl);
 904   phase->register_new_node(gc_state,      old_ctrl);
 905   phase->register_new_node(gc_state_and,  old_ctrl);
 906   phase->register_new_node(gc_state_cmp,  old_ctrl);
 907   phase->register_new_node(gc_state_bool, old_ctrl);
 908 
 909   phase->set_ctrl(gc_state_offset, phase->C->root());
 910 
 911   assert(is_gc_state_test(gc_state_iff, flags), "Should match the shape");
 912 }
 913 
 914 void ShenandoahBarrierC2Support::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) {
 915   Node* old_ctrl = ctrl;
 916   PhaseIterGVN& igvn = phase->igvn();
 917 
 918   const Type* val_t = igvn.type(val);
 919   if (val_t->meet(TypePtr::NULL_PTR) == val_t) {
 920     Node* null_cmp   = new CmpPNode(val, igvn.zerocon(T_OBJECT));
 921     Node* null_test  = new BoolNode(null_cmp, BoolTest::ne);
 922 
 923     IfNode* null_iff = new IfNode(old_ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
 924     ctrl             = new IfTrueNode(null_iff);
 925     null_ctrl        = new IfFalseNode(null_iff);
 926 
 927     IdealLoopTree* loop = phase->get_loop(old_ctrl);
 928     phase->register_control(null_iff,  loop, old_ctrl);
 929     phase->register_control(ctrl,      loop, null_iff);
 930     phase->register_control(null_ctrl, loop, null_iff);
 931 
 932     phase->register_new_node(null_cmp,  old_ctrl);
 933     phase->register_new_node(null_test, old_ctrl);
 934   }
 935 }
 936 
 937 void ShenandoahBarrierC2Support::test_in_cset(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) {
 938   Node* old_ctrl = ctrl;
 939   PhaseIterGVN& igvn = phase->igvn();
 940 
 941   Node* raw_val        = new CastP2XNode(old_ctrl, val);
 942   Node* cset_idx       = new URShiftXNode(raw_val, igvn.intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint()));
 943 
 944   // Figure out the target cset address with raw pointer math.
 945   // This avoids matching AddP+LoadB that would emit inefficient code.
 946   // See JDK-8245465.
 947   Node* cset_addr_ptr  = igvn.makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr()));
 948   Node* cset_addr      = new CastP2XNode(old_ctrl, cset_addr_ptr);
 949   Node* cset_load_addr = new AddXNode(cset_addr, cset_idx);
 950   Node* cset_load_ptr  = new CastX2PNode(cset_load_addr);
 951 
 952   Node* cset_load      = new LoadBNode(old_ctrl, raw_mem, cset_load_ptr,
 953                                        DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(nullptr),
 954                                        TypeInt::BYTE, MemNode::unordered);
 955   Node* cset_cmp       = new CmpINode(cset_load, igvn.zerocon(T_INT));
 956   Node* cset_bool      = new BoolNode(cset_cmp, BoolTest::ne);
 957 
 958   IfNode* cset_iff     = new IfNode(old_ctrl, cset_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
 959   ctrl                 = new IfTrueNode(cset_iff);
 960   not_cset_ctrl        = new IfFalseNode(cset_iff);
 961 
 962   IdealLoopTree *loop = phase->get_loop(old_ctrl);
 963   phase->register_control(cset_iff,      loop, old_ctrl);
 964   phase->register_control(ctrl,          loop, cset_iff);
 965   phase->register_control(not_cset_ctrl, loop, cset_iff);
 966 
 967   phase->set_ctrl(cset_addr_ptr, phase->C->root());
 968 
 969   phase->register_new_node(raw_val,        old_ctrl);
 970   phase->register_new_node(cset_idx,       old_ctrl);
 971   phase->register_new_node(cset_addr,      old_ctrl);
 972   phase->register_new_node(cset_load_addr, old_ctrl);
 973   phase->register_new_node(cset_load_ptr,  old_ctrl);
 974   phase->register_new_node(cset_load,      old_ctrl);
 975   phase->register_new_node(cset_cmp,       old_ctrl);
 976   phase->register_new_node(cset_bool,      old_ctrl);
 977 }
 978 
 979 void ShenandoahBarrierC2Support::call_lrb_stub(Node*& ctrl, Node*& val, Node* load_addr,
 980                                                DecoratorSet decorators, PhaseIdealLoop* phase) {
 981   IdealLoopTree*loop = phase->get_loop(ctrl);
 982   const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr();
 983 
 984   address calladdr = nullptr;
 985   const char* name = nullptr;
 986   bool is_strong  = ShenandoahBarrierSet::is_strong_access(decorators);
 987   bool is_weak    = ShenandoahBarrierSet::is_weak_access(decorators);
 988   bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
 989   bool is_native  = ShenandoahBarrierSet::is_native_access(decorators);
 990   bool is_narrow  = UseCompressedOops && !is_native;
 991   if (is_strong) {
 992     if (is_narrow) {
 993       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow);
 994       name = "load_reference_barrier_strong_narrow";
 995     } else {
 996       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong);
 997       name = "load_reference_barrier_strong";
 998     }
 999   } else if (is_weak) {
1000     if (is_narrow) {
1001       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow);
1002       name = "load_reference_barrier_weak_narrow";
1003     } else {
1004       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak);
1005       name = "load_reference_barrier_weak";
1006     }
1007   } else {
1008     assert(is_phantom, "only remaining strength");
1009     if (is_narrow) {
1010       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom_narrow);
1011       name = "load_reference_barrier_phantom_narrow";
1012     } else {
1013       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom);
1014       name = "load_reference_barrier_phantom";
1015     }
1016   }
1017   Node* call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type(), calladdr, name, TypeRawPtr::BOTTOM);
1018 
1019   call->init_req(TypeFunc::Control, ctrl);
1020   call->init_req(TypeFunc::I_O, phase->C->top());
1021   call->init_req(TypeFunc::Memory, phase->C->top());
1022   call->init_req(TypeFunc::FramePtr, phase->C->top());
1023   call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1024   call->init_req(TypeFunc::Parms, val);
1025   call->init_req(TypeFunc::Parms+1, load_addr);
1026   phase->register_control(call, loop, ctrl);
1027   ctrl = new ProjNode(call, TypeFunc::Control);
1028   phase->register_control(ctrl, loop, call);
1029   val = new ProjNode(call, TypeFunc::Parms);
1030   phase->register_new_node(val, call);
1031   val = new CheckCastPPNode(ctrl, val, obj_type);
1032   phase->register_new_node(val, ctrl);
1033 }
1034 
1035 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) {
1036   Node* ctrl = phase->get_ctrl(barrier);
1037   Node* init_raw_mem = fixer.find_mem(ctrl, barrier);
1038 
1039   // Update the control of all nodes that should be after the
1040   // barrier control flow
1041   uses.clear();
1042   // Every node that is control dependent on the barrier's input
1043   // control will be after the expanded barrier. The raw memory (if
1044   // its memory is control dependent on the barrier's input control)
1045   // must stay above the barrier.
1046   uses_to_ignore.clear();
1047   if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) {
1048     uses_to_ignore.push(init_raw_mem);
1049   }
1050   for (uint next = 0; next < uses_to_ignore.size(); next++) {
1051     Node *n = uses_to_ignore.at(next);
1052     for (uint i = 0; i < n->req(); i++) {
1053       Node* in = n->in(i);
1054       if (in != nullptr && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) {
1055         uses_to_ignore.push(in);
1056       }
1057     }
1058   }
1059   for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) {
1060     Node* u = ctrl->fast_out(i);
1061     if (u->_idx < last &&
1062         u != barrier &&
1063         !uses_to_ignore.member(u) &&
1064         (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) &&
1065         (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) {
1066       Node* old_c = phase->ctrl_or_self(u);
1067       Node* c = old_c;
1068       if (c != ctrl ||
1069           is_dominator_same_ctrl(old_c, barrier, u, phase) ||
1070           ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) {
1071         phase->igvn().rehash_node_delayed(u);
1072         int nb = u->replace_edge(ctrl, region, &phase->igvn());
1073         if (u->is_CFG()) {
1074           if (phase->idom(u) == ctrl) {
1075             phase->set_idom(u, region, phase->dom_depth(region));
1076           }
1077         } else if (phase->get_ctrl(u) == ctrl) {
1078           assert(u != init_raw_mem, "should leave input raw mem above the barrier");
1079           uses.push(u);
1080         }
1081         assert(nb == 1, "more than 1 ctrl input?");
1082         --i, imax -= nb;
1083       }
1084     }
1085   }
1086 }
1087 
1088 static Node* create_phis_on_call_return(Node* ctrl, Node* c, Node* n, Node* n_clone, const CallProjections& projs, PhaseIdealLoop* phase) {
1089   Node* region = nullptr;
1090   while (c != ctrl) {
1091     if (c->is_Region()) {
1092       region = c;
1093     }
1094     c = phase->idom(c);
1095   }
1096   assert(region != nullptr, "");
1097   Node* phi = new PhiNode(region, n->bottom_type());
1098   for (uint j = 1; j < region->req(); j++) {
1099     Node* in = region->in(j);
1100     if (phase->is_dominator(projs.fallthrough_catchproj, in)) {
1101       phi->init_req(j, n);
1102     } else if (phase->is_dominator(projs.catchall_catchproj, in)) {
1103       phi->init_req(j, n_clone);
1104     } else {
1105       phi->init_req(j, create_phis_on_call_return(ctrl, in, n, n_clone, projs, phase));
1106     }
1107   }
1108   phase->register_new_node(phi, region);
1109   return phi;
1110 }
1111 
1112 void ShenandoahBarrierC2Support::pin_and_expand(PhaseIdealLoop* phase) {
1113   ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
1114 
1115   Unique_Node_List uses;
1116   for (int i = 0; i < state->iu_barriers_count(); i++) {
1117     Node* barrier = state->iu_barrier(i);
1118     Node* ctrl = phase->get_ctrl(barrier);
1119     IdealLoopTree* loop = phase->get_loop(ctrl);
1120     Node* head = loop->head();
1121     if (head->is_OuterStripMinedLoop()) {
1122       // Expanding a barrier here will break loop strip mining
1123       // verification. Transform the loop so the loop nest doesn't
1124       // appear as strip mined.
1125       OuterStripMinedLoopNode* outer = head->as_OuterStripMinedLoop();
1126       hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1127     }
1128   }
1129 
1130   Node_Stack stack(0);
1131   Node_List clones;
1132   for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1133     ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1134 
1135     Node* ctrl = phase->get_ctrl(lrb);
1136     Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1137 
1138     CallStaticJavaNode* unc = nullptr;
1139     Node* unc_ctrl = nullptr;
1140     Node* uncasted_val = val;
1141 
1142     for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) {
1143       Node* u = lrb->fast_out(i);
1144       if (u->Opcode() == Op_CastPP &&
1145           u->in(0) != nullptr &&
1146           phase->is_dominator(u->in(0), ctrl)) {
1147         const Type* u_t = phase->igvn().type(u);
1148 
1149         if (u_t->meet(TypePtr::NULL_PTR) != u_t &&
1150             u->in(0)->Opcode() == Op_IfTrue &&
1151             u->in(0)->as_Proj()->is_uncommon_trap_if_pattern() &&
1152             u->in(0)->in(0)->is_If() &&
1153             u->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
1154             u->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
1155             u->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
1156             u->in(0)->in(0)->in(1)->in(1)->in(1) == val &&
1157             u->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
1158           IdealLoopTree* loop = phase->get_loop(ctrl);
1159           IdealLoopTree* unc_loop = phase->get_loop(u->in(0));
1160 
1161           if (!unc_loop->is_member(loop)) {
1162             continue;
1163           }
1164 
1165           Node* branch = no_branches(ctrl, u->in(0), false, phase);
1166           assert(branch == nullptr || branch == NodeSentinel, "was not looking for a branch");
1167           if (branch == NodeSentinel) {
1168             continue;
1169           }
1170 
1171           Node* iff = u->in(0)->in(0);
1172           Node* bol = iff->in(1)->clone();
1173           Node* cmp = bol->in(1)->clone();
1174           cmp->set_req(1, lrb);
1175           bol->set_req(1, cmp);
1176           phase->igvn().replace_input_of(iff, 1, bol);
1177           phase->set_ctrl(lrb, iff->in(0));
1178           phase->register_new_node(cmp, iff->in(0));
1179           phase->register_new_node(bol, iff->in(0));
1180           break;
1181         }
1182       }
1183     }
1184     if ((ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) || ctrl->is_CallJava()) {
1185       CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_CallJava() : ctrl->as_CallJava();
1186       if (call->entry_point() == OptoRuntime::rethrow_stub()) {
1187         // The rethrow call may have too many projections to be
1188         // properly handled here. Given there's no reason for a
1189         // barrier to depend on the call, move it above the call
1190         stack.push(lrb, 0);
1191         do {
1192           Node* n = stack.node();
1193           uint idx = stack.index();
1194           if (idx < n->req()) {
1195             Node* in = n->in(idx);
1196             stack.set_index(idx+1);
1197             if (in != nullptr) {
1198               if (phase->has_ctrl(in)) {
1199                 if (phase->is_dominator(call, phase->get_ctrl(in))) {
1200 #ifdef ASSERT
1201                   for (uint i = 0; i < stack.size(); i++) {
1202                     assert(stack.node_at(i) != in, "node shouldn't have been seen yet");
1203                   }
1204 #endif
1205                   stack.push(in, 0);
1206                 }
1207               } else {
1208                 assert(phase->is_dominator(in, call->in(0)), "no dependency on the call");
1209               }
1210             }
1211           } else {
1212             phase->set_ctrl(n, call->in(0));
1213             stack.pop();
1214           }
1215         } while(stack.size() > 0);
1216         continue;
1217       }
1218       CallProjections projs;
1219       call->extract_projections(&projs, false, false);
1220 
1221 #ifdef ASSERT
1222       VectorSet cloned;
1223 #endif
1224       Node* lrb_clone = lrb->clone();
1225       phase->register_new_node(lrb_clone, projs.catchall_catchproj);
1226       phase->set_ctrl(lrb, projs.fallthrough_catchproj);
1227 
1228       stack.push(lrb, 0);
1229       clones.push(lrb_clone);
1230 
1231       do {
1232         assert(stack.size() == clones.size(), "");
1233         Node* n = stack.node();
1234 #ifdef ASSERT
1235         if (n->is_Load()) {
1236           Node* mem = n->in(MemNode::Memory);
1237           for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
1238             Node* u = mem->fast_out(j);
1239             assert(!u->is_Store() || !u->is_LoadStore() || phase->get_ctrl(u) != ctrl, "anti dependent store?");
1240           }
1241         }
1242 #endif
1243         uint idx = stack.index();
1244         Node* n_clone = clones.at(clones.size()-1);
1245         if (idx < n->outcnt()) {
1246           Node* u = n->raw_out(idx);
1247           Node* c = phase->ctrl_or_self(u);
1248           if (phase->is_dominator(call, c) && phase->is_dominator(c, projs.fallthrough_proj)) {
1249             stack.set_index(idx+1);
1250             assert(!u->is_CFG(), "");
1251             stack.push(u, 0);
1252             assert(!cloned.test_set(u->_idx), "only one clone");
1253             Node* u_clone = u->clone();
1254             int nb = u_clone->replace_edge(n, n_clone, &phase->igvn());
1255             assert(nb > 0, "should have replaced some uses");
1256             phase->register_new_node(u_clone, projs.catchall_catchproj);
1257             clones.push(u_clone);
1258             phase->set_ctrl(u, projs.fallthrough_catchproj);
1259           } else {
1260             bool replaced = false;
1261             if (u->is_Phi()) {
1262               for (uint k = 1; k < u->req(); k++) {
1263                 if (u->in(k) == n) {
1264                   if (phase->is_dominator(projs.catchall_catchproj, u->in(0)->in(k))) {
1265                     phase->igvn().replace_input_of(u, k, n_clone);
1266                     replaced = true;
1267                   } else if (!phase->is_dominator(projs.fallthrough_catchproj, u->in(0)->in(k))) {
1268                     phase->igvn().replace_input_of(u, k, create_phis_on_call_return(ctrl, u->in(0)->in(k), n, n_clone, projs, phase));
1269                     replaced = true;
1270                   }
1271                 }
1272               }
1273             } else {
1274               if (phase->is_dominator(projs.catchall_catchproj, c)) {
1275                 phase->igvn().rehash_node_delayed(u);
1276                 int nb = u->replace_edge(n, n_clone, &phase->igvn());
1277                 assert(nb > 0, "should have replaced some uses");
1278                 replaced = true;
1279               } else if (!phase->is_dominator(projs.fallthrough_catchproj, c)) {
1280                 if (u->is_If()) {
1281                   // Can't break If/Bool/Cmp chain
1282                   assert(n->is_Bool(), "unexpected If shape");
1283                   assert(stack.node_at(stack.size()-2)->is_Cmp(), "unexpected If shape");
1284                   assert(n_clone->is_Bool(), "unexpected clone");
1285                   assert(clones.at(clones.size()-2)->is_Cmp(), "unexpected clone");
1286                   Node* bol_clone = n->clone();
1287                   Node* cmp_clone = stack.node_at(stack.size()-2)->clone();
1288                   bol_clone->set_req(1, cmp_clone);
1289 
1290                   Node* nn = stack.node_at(stack.size()-3);
1291                   Node* nn_clone = clones.at(clones.size()-3);
1292                   assert(nn->Opcode() == nn_clone->Opcode(), "mismatch");
1293 
1294                   int nb = cmp_clone->replace_edge(nn, create_phis_on_call_return(ctrl, c, nn, nn_clone, projs, phase),
1295                                                    &phase->igvn());
1296                   assert(nb > 0, "should have replaced some uses");
1297 
1298                   phase->register_new_node(bol_clone, u->in(0));
1299                   phase->register_new_node(cmp_clone, u->in(0));
1300 
1301                   phase->igvn().replace_input_of(u, 1, bol_clone);
1302 
1303                 } else {
1304                   phase->igvn().rehash_node_delayed(u);
1305                   int nb = u->replace_edge(n, create_phis_on_call_return(ctrl, c, n, n_clone, projs, phase), &phase->igvn());
1306                   assert(nb > 0, "should have replaced some uses");
1307                 }
1308                 replaced = true;
1309               }
1310             }
1311             if (!replaced) {
1312               stack.set_index(idx+1);
1313             }
1314           }
1315         } else {
1316           stack.pop();
1317           clones.pop();
1318         }
1319       } while (stack.size() > 0);
1320       assert(stack.size() == 0 && clones.size() == 0, "");
1321     }
1322   }
1323 
1324   for (int i = 0; i < state->load_reference_barriers_count(); i++) {
1325     ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1326     Node* ctrl = phase->get_ctrl(lrb);
1327     IdealLoopTree* loop = phase->get_loop(ctrl);
1328     Node* head = loop->head();
1329     if (head->is_OuterStripMinedLoop()) {
1330       // Expanding a barrier here will break loop strip mining
1331       // verification. Transform the loop so the loop nest doesn't
1332       // appear as strip mined.
1333       OuterStripMinedLoopNode* outer = head->as_OuterStripMinedLoop();
1334       hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1335     }
1336     if (head->is_BaseCountedLoop() && ctrl->is_IfProj() && ctrl->in(0)->is_BaseCountedLoopEnd() &&
1337         head->as_BaseCountedLoop()->loopexit() == ctrl->in(0)) {
1338       Node* entry = head->in(LoopNode::EntryControl);
1339       Node* backedge = head->in(LoopNode::LoopBackControl);
1340       Node* new_head = new LoopNode(entry, backedge);
1341       phase->register_control(new_head, phase->get_loop(entry), entry);
1342       phase->lazy_replace(head, new_head);
1343     }
1344   }
1345 
1346   // Expand load-reference-barriers
1347   MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
1348   Unique_Node_List uses_to_ignore;
1349   for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1350     ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1351     uint last = phase->C->unique();
1352     Node* ctrl = phase->get_ctrl(lrb);
1353     Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1354 
1355     Node* orig_ctrl = ctrl;
1356 
1357     Node* raw_mem = fixer.find_mem(ctrl, lrb);
1358     Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, nullptr);
1359 
1360     IdealLoopTree *loop = phase->get_loop(ctrl);
1361 
1362     Node* heap_stable_ctrl = nullptr;
1363     Node* null_ctrl = nullptr;
1364 
1365     assert(val->bottom_type()->make_oopptr(), "need oop");
1366     assert(val->bottom_type()->make_oopptr()->const_oop() == nullptr, "expect non-constant");
1367 
1368     enum { _heap_stable = 1, _evac_path, _not_cset, PATH_LIMIT };
1369     Node* region = new RegionNode(PATH_LIMIT);
1370     Node* val_phi = new PhiNode(region, val->bottom_type()->is_oopptr());
1371 
1372     // Stable path.
1373     int flags = ShenandoahHeap::HAS_FORWARDED;
1374     if (!ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1375       flags |= ShenandoahHeap::WEAK_ROOTS;
1376     }
1377     test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, flags);
1378     IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
1379 
1380     // Heap stable case
1381     region->init_req(_heap_stable, heap_stable_ctrl);
1382     val_phi->init_req(_heap_stable, val);
1383 
1384     // Test for in-cset, unless it's a native-LRB. Native LRBs need to return null
1385     // even for non-cset objects to prevent resurrection of such objects.
1386     // Wires !in_cset(obj) to slot 2 of region and phis
1387     Node* not_cset_ctrl = nullptr;
1388     if (ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1389       test_in_cset(ctrl, not_cset_ctrl, val, raw_mem, phase);
1390     }
1391     if (not_cset_ctrl != nullptr) {
1392       region->init_req(_not_cset, not_cset_ctrl);
1393       val_phi->init_req(_not_cset, val);
1394     } else {
1395       region->del_req(_not_cset);
1396       val_phi->del_req(_not_cset);
1397     }
1398 
1399     // Resolve object when orig-value is in cset.
1400     // Make the unconditional resolve for fwdptr.
1401 
1402     // Call lrb-stub and wire up that path in slots 4
1403     Node* result_mem = nullptr;
1404 
1405     Node* addr;
1406     {
1407       VectorSet visited;
1408       addr = get_load_addr(phase, visited, lrb);
1409     }
1410     if (addr->Opcode() == Op_AddP) {
1411       Node* orig_base = addr->in(AddPNode::Base);
1412       Node* base = new CheckCastPPNode(ctrl, orig_base, orig_base->bottom_type(), ConstraintCastNode::StrongDependency);
1413       phase->register_new_node(base, ctrl);
1414       if (addr->in(AddPNode::Base) == addr->in((AddPNode::Address))) {
1415         // Field access
1416         addr = addr->clone();
1417         addr->set_req(AddPNode::Base, base);
1418         addr->set_req(AddPNode::Address, base);
1419         phase->register_new_node(addr, ctrl);
1420       } else {
1421         Node* addr2 = addr->in(AddPNode::Address);
1422         if (addr2->Opcode() == Op_AddP && addr2->in(AddPNode::Base) == addr2->in(AddPNode::Address) &&
1423               addr2->in(AddPNode::Base) == orig_base) {
1424           addr2 = addr2->clone();
1425           addr2->set_req(AddPNode::Base, base);
1426           addr2->set_req(AddPNode::Address, base);
1427           phase->register_new_node(addr2, ctrl);
1428           addr = addr->clone();
1429           addr->set_req(AddPNode::Base, base);
1430           addr->set_req(AddPNode::Address, addr2);
1431           phase->register_new_node(addr, ctrl);
1432         }
1433       }
1434     }
1435     call_lrb_stub(ctrl, val, addr, lrb->decorators(), phase);
1436     region->init_req(_evac_path, ctrl);
1437     val_phi->init_req(_evac_path, val);
1438 
1439     phase->register_control(region, loop, heap_stable_iff);
1440     Node* out_val = val_phi;
1441     phase->register_new_node(val_phi, region);
1442 
1443     fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase);
1444 
1445     ctrl = orig_ctrl;
1446 
1447     phase->igvn().replace_node(lrb, out_val);
1448 
1449     follow_barrier_uses(out_val, ctrl, uses, phase);
1450 
1451     for(uint next = 0; next < uses.size(); next++ ) {
1452       Node *n = uses.at(next);
1453       assert(phase->get_ctrl(n) == ctrl, "bad control");
1454       assert(n != raw_mem, "should leave input raw mem above the barrier");
1455       phase->set_ctrl(n, region);
1456       follow_barrier_uses(n, ctrl, uses, phase);
1457     }
1458     fixer.record_new_ctrl(ctrl, region, raw_mem, raw_mem_for_ctrl);
1459   }
1460   // Done expanding load-reference-barriers.
1461   assert(ShenandoahBarrierSetC2::bsc2()->state()->load_reference_barriers_count() == 0, "all load reference barrier nodes should have been replaced");
1462 
1463   for (int i = state->iu_barriers_count() - 1; i >= 0; i--) {
1464     Node* barrier = state->iu_barrier(i);
1465     Node* pre_val = barrier->in(1);
1466 
1467     if (phase->igvn().type(pre_val)->higher_equal(TypePtr::NULL_PTR)) {
1468       ShouldNotReachHere();
1469       continue;
1470     }
1471 
1472     Node* ctrl = phase->get_ctrl(barrier);
1473 
1474     if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) {
1475       assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0)->in(0), pre_val, ctrl->in(0), phase), "can't move");
1476       ctrl = ctrl->in(0)->in(0);
1477       phase->set_ctrl(barrier, ctrl);
1478     } else if (ctrl->is_CallRuntime()) {
1479       assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0), pre_val, ctrl, phase), "can't move");
1480       ctrl = ctrl->in(0);
1481       phase->set_ctrl(barrier, ctrl);
1482     }
1483 
1484     Node* init_ctrl = ctrl;
1485     IdealLoopTree* loop = phase->get_loop(ctrl);
1486     Node* raw_mem = fixer.find_mem(ctrl, barrier);
1487     Node* init_raw_mem = raw_mem;
1488     Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, nullptr);
1489     Node* heap_stable_ctrl = nullptr;
1490     Node* null_ctrl = nullptr;
1491     uint last = phase->C->unique();
1492 
1493     enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
1494     Node* region = new RegionNode(PATH_LIMIT);
1495     Node* phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1496 
1497     enum { _fast_path = 1, _slow_path, _null_path, PATH_LIMIT2 };
1498     Node* region2 = new RegionNode(PATH_LIMIT2);
1499     Node* phi2 = PhiNode::make(region2, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1500 
1501     // Stable path.
1502     test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, ShenandoahHeap::MARKING);
1503     region->init_req(_heap_stable, heap_stable_ctrl);
1504     phi->init_req(_heap_stable, raw_mem);
1505 
1506     // Null path
1507     Node* reg2_ctrl = nullptr;
1508     test_null(ctrl, pre_val, null_ctrl, phase);
1509     if (null_ctrl != nullptr) {
1510       reg2_ctrl = null_ctrl->in(0);
1511       region2->init_req(_null_path, null_ctrl);
1512       phi2->init_req(_null_path, raw_mem);
1513     } else {
1514       region2->del_req(_null_path);
1515       phi2->del_req(_null_path);
1516     }
1517 
1518     const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
1519     const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1520     Node* thread = new ThreadLocalNode();
1521     phase->register_new_node(thread, ctrl);
1522     Node* buffer_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(buffer_offset));
1523     phase->register_new_node(buffer_adr, ctrl);
1524     Node* index_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(index_offset));
1525     phase->register_new_node(index_adr, ctrl);
1526 
1527     BasicType index_bt = TypeX_X->basic_type();
1528     assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading Shenandoah SATBMarkQueue::_index with wrong size.");
1529     const TypePtr* adr_type = TypeRawPtr::BOTTOM;
1530     Node* index = new LoadXNode(ctrl, raw_mem, index_adr, adr_type, TypeX_X, MemNode::unordered);
1531     phase->register_new_node(index, ctrl);
1532     Node* index_cmp = new CmpXNode(index, phase->igvn().MakeConX(0));
1533     phase->register_new_node(index_cmp, ctrl);
1534     Node* index_test = new BoolNode(index_cmp, BoolTest::ne);
1535     phase->register_new_node(index_test, ctrl);
1536     IfNode* queue_full_iff = new IfNode(ctrl, index_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
1537     if (reg2_ctrl == nullptr) reg2_ctrl = queue_full_iff;
1538     phase->register_control(queue_full_iff, loop, ctrl);
1539     Node* not_full = new IfTrueNode(queue_full_iff);
1540     phase->register_control(not_full, loop, queue_full_iff);
1541     Node* full = new IfFalseNode(queue_full_iff);
1542     phase->register_control(full, loop, queue_full_iff);
1543 
1544     ctrl = not_full;
1545 
1546     Node* next_index = new SubXNode(index, phase->igvn().MakeConX(sizeof(intptr_t)));
1547     phase->register_new_node(next_index, ctrl);
1548 
1549     Node* buffer  = new LoadPNode(ctrl, raw_mem, buffer_adr, adr_type, TypeRawPtr::NOTNULL, MemNode::unordered);
1550     phase->register_new_node(buffer, ctrl);
1551     Node *log_addr = new AddPNode(phase->C->top(), buffer, next_index);
1552     phase->register_new_node(log_addr, ctrl);
1553     Node* log_store = new StorePNode(ctrl, raw_mem, log_addr, adr_type, pre_val, MemNode::unordered);
1554     phase->register_new_node(log_store, ctrl);
1555     // update the index
1556     Node* index_update = new StoreXNode(ctrl, log_store, index_adr, adr_type, next_index, MemNode::unordered);
1557     phase->register_new_node(index_update, ctrl);
1558 
1559     // Fast-path case
1560     region2->init_req(_fast_path, ctrl);
1561     phi2->init_req(_fast_path, index_update);
1562 
1563     ctrl = full;
1564 
1565     Node* base = find_bottom_mem(ctrl, phase);
1566 
1567     MergeMemNode* mm = MergeMemNode::make(base);
1568     mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
1569     phase->register_new_node(mm, ctrl);
1570 
1571     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);
1572     call->init_req(TypeFunc::Control, ctrl);
1573     call->init_req(TypeFunc::I_O, phase->C->top());
1574     call->init_req(TypeFunc::Memory, mm);
1575     call->init_req(TypeFunc::FramePtr, phase->C->top());
1576     call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1577     call->init_req(TypeFunc::Parms, pre_val);
1578     call->init_req(TypeFunc::Parms+1, thread);
1579     phase->register_control(call, loop, ctrl);
1580 
1581     Node* ctrl_proj = new ProjNode(call, TypeFunc::Control);
1582     phase->register_control(ctrl_proj, loop, call);
1583     Node* mem_proj = new ProjNode(call, TypeFunc::Memory);
1584     phase->register_new_node(mem_proj, call);
1585 
1586     // Slow-path case
1587     region2->init_req(_slow_path, ctrl_proj);
1588     phi2->init_req(_slow_path, mem_proj);
1589 
1590     phase->register_control(region2, loop, reg2_ctrl);
1591     phase->register_new_node(phi2, region2);
1592 
1593     region->init_req(_heap_unstable, region2);
1594     phi->init_req(_heap_unstable, phi2);
1595 
1596     phase->register_control(region, loop, heap_stable_ctrl->in(0));
1597     phase->register_new_node(phi, region);
1598 
1599     fix_ctrl(barrier, region, fixer, uses, uses_to_ignore, last, phase);
1600     for(uint next = 0; next < uses.size(); next++ ) {
1601       Node *n = uses.at(next);
1602       assert(phase->get_ctrl(n) == init_ctrl, "bad control");
1603       assert(n != init_raw_mem, "should leave input raw mem above the barrier");
1604       phase->set_ctrl(n, region);
1605       follow_barrier_uses(n, init_ctrl, uses, phase);
1606     }
1607     fixer.fix_mem(init_ctrl, region, init_raw_mem, raw_mem_for_ctrl, phi, uses);
1608 
1609     phase->igvn().replace_node(barrier, pre_val);
1610   }
1611   assert(state->iu_barriers_count() == 0, "all enqueue barrier nodes should have been replaced");
1612 
1613 }
1614 
1615 Node* ShenandoahBarrierC2Support::get_load_addr(PhaseIdealLoop* phase, VectorSet& visited, Node* in) {
1616   if (visited.test_set(in->_idx)) {
1617     return nullptr;
1618   }
1619   switch (in->Opcode()) {
1620     case Op_Proj:
1621       return get_load_addr(phase, visited, in->in(0));
1622     case Op_CastPP:
1623     case Op_CheckCastPP:
1624     case Op_DecodeN:
1625     case Op_EncodeP:
1626       return get_load_addr(phase, visited, in->in(1));
1627     case Op_LoadN:
1628     case Op_LoadP:
1629       return in->in(MemNode::Address);
1630     case Op_CompareAndExchangeN:
1631     case Op_CompareAndExchangeP:
1632     case Op_GetAndSetN:
1633     case Op_GetAndSetP:
1634     case Op_ShenandoahCompareAndExchangeP:
1635     case Op_ShenandoahCompareAndExchangeN:
1636       // Those instructions would just have stored a different
1637       // value into the field. No use to attempt to fix it at this point.
1638       return phase->igvn().zerocon(T_OBJECT);
1639     case Op_CMoveP:
1640     case Op_CMoveN: {
1641       Node* t = get_load_addr(phase, visited, in->in(CMoveNode::IfTrue));
1642       Node* f = get_load_addr(phase, visited, in->in(CMoveNode::IfFalse));
1643       // Handle unambiguous cases: single address reported on both branches.
1644       if (t != nullptr && f == nullptr) return t;
1645       if (t == nullptr && f != nullptr) return f;
1646       if (t != nullptr && t == f)    return t;
1647       // Ambiguity.
1648       return phase->igvn().zerocon(T_OBJECT);
1649     }
1650     case Op_Phi: {
1651       Node* addr = nullptr;
1652       for (uint i = 1; i < in->req(); i++) {
1653         Node* addr1 = get_load_addr(phase, visited, in->in(i));
1654         if (addr == nullptr) {
1655           addr = addr1;
1656         }
1657         if (addr != addr1) {
1658           return phase->igvn().zerocon(T_OBJECT);
1659         }
1660       }
1661       return addr;
1662     }
1663     case Op_ShenandoahLoadReferenceBarrier:
1664       return get_load_addr(phase, visited, in->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1665     case Op_ShenandoahIUBarrier:
1666       return get_load_addr(phase, visited, in->in(1));
1667     case Op_CallDynamicJava:
1668     case Op_CallLeaf:
1669     case Op_CallStaticJava:
1670     case Op_ConN:
1671     case Op_ConP:
1672     case Op_Parm:
1673     case Op_CreateEx:
1674       return phase->igvn().zerocon(T_OBJECT);
1675     default:
1676 #ifdef ASSERT
1677       fatal("Unknown node in get_load_addr: %s", NodeClassNames[in->Opcode()]);
1678 #endif
1679       return phase->igvn().zerocon(T_OBJECT);
1680   }
1681 
1682 }
1683 
1684 void ShenandoahBarrierC2Support::move_gc_state_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
1685   IdealLoopTree *loop = phase->get_loop(iff);
1686   Node* loop_head = loop->_head;
1687   Node* entry_c = loop_head->in(LoopNode::EntryControl);
1688 
1689   Node* bol = iff->in(1);
1690   Node* cmp = bol->in(1);
1691   Node* andi = cmp->in(1);
1692   Node* load = andi->in(1);
1693 
1694   assert(is_gc_state_load(load), "broken");
1695   if (!phase->is_dominator(load->in(0), entry_c)) {
1696     Node* mem_ctrl = nullptr;
1697     Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
1698     load = load->clone();
1699     load->set_req(MemNode::Memory, mem);
1700     load->set_req(0, entry_c);
1701     phase->register_new_node(load, entry_c);
1702     andi = andi->clone();
1703     andi->set_req(1, load);
1704     phase->register_new_node(andi, entry_c);
1705     cmp = cmp->clone();
1706     cmp->set_req(1, andi);
1707     phase->register_new_node(cmp, entry_c);
1708     bol = bol->clone();
1709     bol->set_req(1, cmp);
1710     phase->register_new_node(bol, entry_c);
1711 
1712     phase->igvn().replace_input_of(iff, 1, bol);
1713   }
1714 }
1715 
1716 bool ShenandoahBarrierC2Support::identical_backtoback_ifs(Node* n, PhaseIdealLoop* phase) {
1717   if (!n->is_If() || n->is_CountedLoopEnd()) {
1718     return false;
1719   }
1720   Node* region = n->in(0);
1721 
1722   if (!region->is_Region()) {
1723     return false;
1724   }
1725   Node* dom = phase->idom(region);
1726   if (!dom->is_If()) {
1727     return false;
1728   }
1729 
1730   if (!is_heap_stable_test(n) || !is_heap_stable_test(dom)) {
1731     return false;
1732   }
1733 
1734   IfNode* dom_if = dom->as_If();
1735   Node* proj_true = dom_if->proj_out(1);
1736   Node* proj_false = dom_if->proj_out(0);
1737 
1738   for (uint i = 1; i < region->req(); i++) {
1739     if (phase->is_dominator(proj_true, region->in(i))) {
1740       continue;
1741     }
1742     if (phase->is_dominator(proj_false, region->in(i))) {
1743       continue;
1744     }
1745     return false;
1746   }
1747 
1748   return true;
1749 }
1750 
1751 bool ShenandoahBarrierC2Support::merge_point_safe(Node* region) {
1752   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1753     Node* n = region->fast_out(i);
1754     if (n->is_LoadStore()) {
1755       // Splitting a LoadStore node through phi, causes it to lose its SCMemProj: the split if code doesn't have support
1756       // for a LoadStore at the region the if is split through because that's not expected to happen (LoadStore nodes
1757       // should be between barrier nodes). It does however happen with Shenandoah though because barriers can get
1758       // expanded around a LoadStore node.
1759       return false;
1760     }
1761   }
1762   return true;
1763 }
1764 
1765 
1766 void ShenandoahBarrierC2Support::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) {
1767   assert(is_heap_stable_test(n), "no other tests");
1768   if (identical_backtoback_ifs(n, phase)) {
1769     Node* n_ctrl = n->in(0);
1770     if (phase->can_split_if(n_ctrl) && merge_point_safe(n_ctrl)) {
1771       IfNode* dom_if = phase->idom(n_ctrl)->as_If();
1772       if (is_heap_stable_test(n)) {
1773         Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1);
1774         assert(is_gc_state_load(gc_state_load), "broken");
1775         Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1);
1776         assert(is_gc_state_load(dom_gc_state_load), "broken");
1777         if (gc_state_load != dom_gc_state_load) {
1778           phase->igvn().replace_node(gc_state_load, dom_gc_state_load);
1779         }
1780       }
1781       PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1782       Node* proj_true = dom_if->proj_out(1);
1783       Node* proj_false = dom_if->proj_out(0);
1784       Node* con_true = phase->igvn().makecon(TypeInt::ONE);
1785       Node* con_false = phase->igvn().makecon(TypeInt::ZERO);
1786 
1787       for (uint i = 1; i < n_ctrl->req(); i++) {
1788         if (phase->is_dominator(proj_true, n_ctrl->in(i))) {
1789           bolphi->init_req(i, con_true);
1790         } else {
1791           assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1792           bolphi->init_req(i, con_false);
1793         }
1794       }
1795       phase->register_new_node(bolphi, n_ctrl);
1796       phase->igvn().replace_input_of(n, 1, bolphi);
1797       phase->do_split_if(n);
1798     }
1799   }
1800 }
1801 
1802 IfNode* ShenandoahBarrierC2Support::find_unswitching_candidate(const IdealLoopTree* loop, PhaseIdealLoop* phase) {
1803   // Find first invariant test that doesn't exit the loop
1804   LoopNode *head = loop->_head->as_Loop();
1805   IfNode* unswitch_iff = nullptr;
1806   Node* n = head->in(LoopNode::LoopBackControl);
1807   int loop_has_sfpts = -1;
1808   while (n != head) {
1809     Node* n_dom = phase->idom(n);
1810     if (n->is_Region()) {
1811       if (n_dom->is_If()) {
1812         IfNode* iff = n_dom->as_If();
1813         if (iff->in(1)->is_Bool()) {
1814           BoolNode* bol = iff->in(1)->as_Bool();
1815           if (bol->in(1)->is_Cmp()) {
1816             // If condition is invariant and not a loop exit,
1817             // then found reason to unswitch.
1818             if (is_heap_stable_test(iff) &&
1819                 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) {
1820               assert(!loop->is_loop_exit(iff), "both branches should be in the loop");
1821               if (loop_has_sfpts == -1) {
1822                 for(uint i = 0; i < loop->_body.size(); i++) {
1823                   Node *m = loop->_body[i];
1824                   if (m->is_SafePoint() && !m->is_CallLeaf()) {
1825                     loop_has_sfpts = 1;
1826                     break;
1827                   }
1828                 }
1829                 if (loop_has_sfpts == -1) {
1830                   loop_has_sfpts = 0;
1831                 }
1832               }
1833               if (!loop_has_sfpts) {
1834                 unswitch_iff = iff;
1835               }
1836             }
1837           }
1838         }
1839       }
1840     }
1841     n = n_dom;
1842   }
1843   return unswitch_iff;
1844 }
1845 
1846 
1847 void ShenandoahBarrierC2Support::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) {
1848   Node_List heap_stable_tests;
1849   stack.push(phase->C->start(), 0);
1850   do {
1851     Node* n = stack.node();
1852     uint i = stack.index();
1853 
1854     if (i < n->outcnt()) {
1855       Node* u = n->raw_out(i);
1856       stack.set_index(i+1);
1857       if (!visited.test_set(u->_idx)) {
1858         stack.push(u, 0);
1859       }
1860     } else {
1861       stack.pop();
1862       if (n->is_If() && is_heap_stable_test(n)) {
1863         heap_stable_tests.push(n);
1864       }
1865     }
1866   } while (stack.size() > 0);
1867 
1868   for (uint i = 0; i < heap_stable_tests.size(); i++) {
1869     Node* n = heap_stable_tests.at(i);
1870     assert(is_heap_stable_test(n), "only evacuation test");
1871     merge_back_to_back_tests(n, phase);
1872   }
1873 
1874   if (!phase->C->major_progress()) {
1875     VectorSet seen;
1876     for (uint i = 0; i < heap_stable_tests.size(); i++) {
1877       Node* n = heap_stable_tests.at(i);
1878       IdealLoopTree* loop = phase->get_loop(n);
1879       if (loop != phase->ltree_root() &&
1880           loop->_child == nullptr &&
1881           !loop->_irreducible) {
1882         Node* head = loop->_head;
1883         if (head->is_Loop() &&
1884             (!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) &&
1885             !seen.test_set(head->_idx)) {
1886           IfNode* iff = find_unswitching_candidate(loop, phase);
1887           if (iff != nullptr) {
1888             Node* bol = iff->in(1);
1889             if (head->as_Loop()->is_strip_mined()) {
1890               head->as_Loop()->verify_strip_mined(0);
1891             }
1892             move_gc_state_test_out_of_loop(iff, phase);
1893 
1894             AutoNodeBudget node_budget(phase);
1895 
1896             if (loop->policy_unswitching(phase)) {
1897               if (head->as_Loop()->is_strip_mined()) {
1898                 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop();
1899                 hide_strip_mined_loop(outer, head->as_CountedLoop(), phase);
1900               }
1901               phase->do_unswitching(loop, old_new);
1902             } else {
1903               // Not proceeding with unswitching. Move load back in
1904               // the loop.
1905               phase->igvn().replace_input_of(iff, 1, bol);
1906             }
1907           }
1908         }
1909       }
1910     }
1911   }
1912 }
1913 
1914 ShenandoahIUBarrierNode::ShenandoahIUBarrierNode(Node* val) : Node(nullptr, val) {
1915   ShenandoahBarrierSetC2::bsc2()->state()->add_iu_barrier(this);
1916 }
1917 
1918 const Type* ShenandoahIUBarrierNode::bottom_type() const {
1919   if (in(1) == nullptr || in(1)->is_top()) {
1920     return Type::TOP;
1921   }
1922   const Type* t = in(1)->bottom_type();
1923   if (t == TypePtr::NULL_PTR) {
1924     return t;
1925   }
1926   return t->is_oopptr();
1927 }
1928 
1929 const Type* ShenandoahIUBarrierNode::Value(PhaseGVN* phase) const {
1930   if (in(1) == nullptr) {
1931     return Type::TOP;
1932   }
1933   const Type* t = phase->type(in(1));
1934   if (t == Type::TOP) {
1935     return Type::TOP;
1936   }
1937   if (t == TypePtr::NULL_PTR) {
1938     return t;
1939   }
1940   return t->is_oopptr();
1941 }
1942 
1943 int ShenandoahIUBarrierNode::needed(Node* n) {
1944   if (n == nullptr ||
1945       n->is_Allocate() ||
1946       n->Opcode() == Op_ShenandoahIUBarrier ||
1947       n->bottom_type() == TypePtr::NULL_PTR ||
1948       (n->bottom_type()->make_oopptr() != nullptr && n->bottom_type()->make_oopptr()->const_oop() != nullptr)) {
1949     return NotNeeded;
1950   }
1951   if (n->is_Phi() ||
1952       n->is_CMove()) {
1953     return MaybeNeeded;
1954   }
1955   return Needed;
1956 }
1957 
1958 Node* ShenandoahIUBarrierNode::next(Node* n) {
1959   for (;;) {
1960     if (n == nullptr) {
1961       return n;
1962     } else if (n->bottom_type() == TypePtr::NULL_PTR) {
1963       return n;
1964     } else if (n->bottom_type()->make_oopptr() != nullptr && n->bottom_type()->make_oopptr()->const_oop() != nullptr) {
1965       return n;
1966     } else if (n->is_ConstraintCast() ||
1967                n->Opcode() == Op_DecodeN ||
1968                n->Opcode() == Op_EncodeP) {
1969       n = n->in(1);
1970     } else if (n->is_Proj()) {
1971       n = n->in(0);
1972     } else {
1973       return n;
1974     }
1975   }
1976   ShouldNotReachHere();
1977   return nullptr;
1978 }
1979 
1980 Node* ShenandoahIUBarrierNode::Identity(PhaseGVN* phase) {
1981   PhaseIterGVN* igvn = phase->is_IterGVN();
1982 
1983   Node* n = next(in(1));
1984 
1985   int cont = needed(n);
1986 
1987   if (cont == NotNeeded) {
1988     return in(1);
1989   } else if (cont == MaybeNeeded) {
1990     if (igvn == nullptr) {
1991       phase->record_for_igvn(this);
1992       return this;
1993     } else {
1994       ResourceMark rm;
1995       Unique_Node_List wq;
1996       uint wq_i = 0;
1997 
1998       for (;;) {
1999         if (n->is_Phi()) {
2000           for (uint i = 1; i < n->req(); i++) {
2001             Node* m = n->in(i);
2002             if (m != nullptr) {
2003               wq.push(m);
2004             }
2005           }
2006         } else {
2007           assert(n->is_CMove(), "nothing else here");
2008           Node* m = n->in(CMoveNode::IfFalse);
2009           wq.push(m);
2010           m = n->in(CMoveNode::IfTrue);
2011           wq.push(m);
2012         }
2013         Node* orig_n = nullptr;
2014         do {
2015           if (wq_i >= wq.size()) {
2016             return in(1);
2017           }
2018           n = wq.at(wq_i);
2019           wq_i++;
2020           orig_n = n;
2021           n = next(n);
2022           cont = needed(n);
2023           if (cont == Needed) {
2024             return this;
2025           }
2026         } while (cont != MaybeNeeded || (orig_n != n && wq.member(n)));
2027       }
2028     }
2029   }
2030 
2031   return this;
2032 }
2033 
2034 #ifdef ASSERT
2035 static bool has_never_branch(Node* root) {
2036   for (uint i = 1; i < root->req(); i++) {
2037     Node* in = root->in(i);
2038     if (in != nullptr && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->is_NeverBranch()) {
2039       return true;
2040     }
2041   }
2042   return false;
2043 }
2044 #endif
2045 
2046 void MemoryGraphFixer::collect_memory_nodes() {
2047   Node_Stack stack(0);
2048   VectorSet visited;
2049   Node_List regions;
2050 
2051   // Walk the raw memory graph and create a mapping from CFG node to
2052   // memory node. Exclude phis for now.
2053   stack.push(_phase->C->root(), 1);
2054   do {
2055     Node* n = stack.node();
2056     int opc = n->Opcode();
2057     uint i = stack.index();
2058     if (i < n->req()) {
2059       Node* mem = nullptr;
2060       if (opc == Op_Root) {
2061         Node* in = n->in(i);
2062         int in_opc = in->Opcode();
2063         if (in_opc == Op_Return || in_opc == Op_Rethrow) {
2064           mem = in->in(TypeFunc::Memory);
2065         } else if (in_opc == Op_Halt) {
2066           if (in->in(0)->is_Region()) {
2067             Node* r = in->in(0);
2068             for (uint j = 1; j < r->req(); j++) {
2069               assert(!r->in(j)->is_NeverBranch(), "");
2070             }
2071           } else {
2072             Node* proj = in->in(0);
2073             assert(proj->is_Proj(), "");
2074             Node* in = proj->in(0);
2075             assert(in->is_CallStaticJava() || in->is_NeverBranch() || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
2076             if (in->is_CallStaticJava()) {
2077               mem = in->in(TypeFunc::Memory);
2078             } else if (in->Opcode() == Op_Catch) {
2079               Node* call = in->in(0)->in(0);
2080               assert(call->is_Call(), "");
2081               mem = call->in(TypeFunc::Memory);
2082             } else if (in->is_NeverBranch()) {
2083               mem = collect_memory_for_infinite_loop(in);
2084             }
2085           }
2086         } else {
2087 #ifdef ASSERT
2088           n->dump();
2089           in->dump();
2090 #endif
2091           ShouldNotReachHere();
2092         }
2093       } else {
2094         assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
2095         assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
2096         mem = n->in(i);
2097       }
2098       i++;
2099       stack.set_index(i);
2100       if (mem == nullptr) {
2101         continue;
2102       }
2103       for (;;) {
2104         if (visited.test_set(mem->_idx) || mem->is_Start()) {
2105           break;
2106         }
2107         if (mem->is_Phi()) {
2108           stack.push(mem, 2);
2109           mem = mem->in(1);
2110         } else if (mem->is_Proj()) {
2111           stack.push(mem, mem->req());
2112           mem = mem->in(0);
2113         } else if (mem->is_SafePoint() || mem->is_MemBar()) {
2114           mem = mem->in(TypeFunc::Memory);
2115         } else if (mem->is_MergeMem()) {
2116           MergeMemNode* mm = mem->as_MergeMem();
2117           mem = mm->memory_at(_alias);
2118         } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
2119           assert(_alias == Compile::AliasIdxRaw, "");
2120           stack.push(mem, mem->req());
2121           mem = mem->in(MemNode::Memory);
2122         } else {
2123 #ifdef ASSERT
2124           mem->dump();
2125 #endif
2126           ShouldNotReachHere();
2127         }
2128       }
2129     } else {
2130       if (n->is_Phi()) {
2131         // Nothing
2132       } else if (!n->is_Root()) {
2133         Node* c = get_ctrl(n);
2134         _memory_nodes.map(c->_idx, n);
2135       }
2136       stack.pop();
2137     }
2138   } while(stack.is_nonempty());
2139 
2140   // Iterate over CFG nodes in rpo and propagate memory state to
2141   // compute memory state at regions, creating new phis if needed.
2142   Node_List rpo_list;
2143   visited.clear();
2144   _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
2145   Node* root = rpo_list.pop();
2146   assert(root == _phase->C->root(), "");
2147 
2148   const bool trace = false;
2149 #ifdef ASSERT
2150   if (trace) {
2151     for (int i = rpo_list.size() - 1; i >= 0; i--) {
2152       Node* c = rpo_list.at(i);
2153       if (_memory_nodes[c->_idx] != nullptr) {
2154         tty->print("X %d", c->_idx);  _memory_nodes[c->_idx]->dump();
2155       }
2156     }
2157   }
2158 #endif
2159   uint last = _phase->C->unique();
2160 
2161 #ifdef ASSERT
2162   uint16_t max_depth = 0;
2163   for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
2164     IdealLoopTree* lpt = iter.current();
2165     max_depth = MAX2(max_depth, lpt->_nest);
2166   }
2167 #endif
2168 
2169   bool progress = true;
2170   int iteration = 0;
2171   Node_List dead_phis;
2172   while (progress) {
2173     progress = false;
2174     iteration++;
2175     assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2176     if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
2177 
2178     for (int i = rpo_list.size() - 1; i >= 0; i--) {
2179       Node* c = rpo_list.at(i);
2180 
2181       Node* prev_mem = _memory_nodes[c->_idx];
2182       if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2183         Node* prev_region = regions[c->_idx];
2184         Node* unique = nullptr;
2185         for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
2186           Node* m = _memory_nodes[c->in(j)->_idx];
2187           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");
2188           if (m != nullptr) {
2189             if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
2190               assert((c->is_Loop() && j == LoopNode::LoopBackControl) || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2191               // continue
2192             } else if (unique == nullptr) {
2193               unique = m;
2194             } else if (m == unique) {
2195               // continue
2196             } else {
2197               unique = NodeSentinel;
2198             }
2199           }
2200         }
2201         assert(unique != nullptr, "empty phi???");
2202         if (unique != NodeSentinel) {
2203           if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c) {
2204             dead_phis.push(prev_region);
2205           }
2206           regions.map(c->_idx, unique);
2207         } else {
2208           Node* phi = nullptr;
2209           if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
2210             phi = prev_region;
2211             for (uint k = 1; k < c->req(); k++) {
2212               Node* m = _memory_nodes[c->in(k)->_idx];
2213               assert(m != nullptr, "expect memory state");
2214               phi->set_req(k, m);
2215             }
2216           } else {
2217             for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == nullptr; j++) {
2218               Node* u = c->fast_out(j);
2219               if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2220                   (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
2221                 phi = u;
2222                 for (uint k = 1; k < c->req() && phi != nullptr; k++) {
2223                   Node* m = _memory_nodes[c->in(k)->_idx];
2224                   assert(m != nullptr, "expect memory state");
2225                   if (u->in(k) != m) {
2226                     phi = NodeSentinel;
2227                   }
2228                 }
2229               }
2230             }
2231             if (phi == NodeSentinel) {
2232               phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
2233               for (uint k = 1; k < c->req(); k++) {
2234                 Node* m = _memory_nodes[c->in(k)->_idx];
2235                 assert(m != nullptr, "expect memory state");
2236                 phi->init_req(k, m);
2237               }
2238             }
2239           }
2240           if (phi != nullptr) {
2241             regions.map(c->_idx, phi);
2242           } else {
2243             assert(c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
2244           }
2245         }
2246         Node* current_region = regions[c->_idx];
2247         if (current_region != prev_region) {
2248           progress = true;
2249           if (prev_region == prev_mem) {
2250             _memory_nodes.map(c->_idx, current_region);
2251           }
2252         }
2253       } else if (prev_mem == nullptr || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
2254         Node* m = _memory_nodes[_phase->idom(c)->_idx];
2255         assert(m != nullptr || c->Opcode() == Op_Halt, "expect memory state");
2256         if (m != prev_mem) {
2257           _memory_nodes.map(c->_idx, m);
2258           progress = true;
2259         }
2260       }
2261 #ifdef ASSERT
2262       if (trace) { tty->print("X %d", c->_idx);  _memory_nodes[c->_idx]->dump(); }
2263 #endif
2264     }
2265   }
2266 
2267   // Replace existing phi with computed memory state for that region
2268   // if different (could be a new phi or a dominating memory node if
2269   // that phi was found to be useless).
2270   while (dead_phis.size() > 0) {
2271     Node* n = dead_phis.pop();
2272     n->replace_by(_phase->C->top());
2273     n->destruct(&_phase->igvn());
2274   }
2275   for (int i = rpo_list.size() - 1; i >= 0; i--) {
2276     Node* c = rpo_list.at(i);
2277     if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2278       Node* n = regions[c->_idx];
2279       assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
2280       if (n != nullptr && n->is_Phi() && n->_idx >= last && n->in(0) == c) {
2281         _phase->register_new_node(n, c);
2282       }
2283     }
2284   }
2285   for (int i = rpo_list.size() - 1; i >= 0; i--) {
2286     Node* c = rpo_list.at(i);
2287     if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2288       Node* n = regions[c->_idx];
2289       assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
2290       for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
2291         Node* u = c->fast_out(i);
2292         if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2293             u != n) {
2294           assert(c->unique_ctrl_out()->Opcode() != Op_Halt, "expected memory state");
2295           if (u->adr_type() == TypePtr::BOTTOM) {
2296             fix_memory_uses(u, n, n, c);
2297           } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2298             _phase->lazy_replace(u, n);
2299             --i; --imax;
2300           }
2301         }
2302       }
2303     }
2304   }
2305 }
2306 
2307 Node* MemoryGraphFixer::collect_memory_for_infinite_loop(const Node* in) {
2308   Node* mem = nullptr;
2309   Node* head = in->in(0);
2310   assert(head->is_Region(), "unexpected infinite loop graph shape");
2311 
2312   Node* phi_mem = nullptr;
2313   for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2314     Node* u = head->fast_out(j);
2315     if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
2316       if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2317         assert(phi_mem == nullptr || phi_mem->adr_type() == TypePtr::BOTTOM, "");
2318         phi_mem = u;
2319       } else if (u->adr_type() == TypePtr::BOTTOM) {
2320         assert(phi_mem == nullptr || _phase->C->get_alias_index(phi_mem->adr_type()) == _alias, "");
2321         if (phi_mem == nullptr) {
2322           phi_mem = u;
2323         }
2324       }
2325     }
2326   }
2327   if (phi_mem == nullptr) {
2328     ResourceMark rm;
2329     Node_Stack stack(0);
2330     stack.push(head, 1);
2331     do {
2332       Node* n = stack.node();
2333       uint i = stack.index();
2334       if (i >= n->req()) {
2335         stack.pop();
2336       } else {
2337         stack.set_index(i + 1);
2338         Node* c = n->in(i);
2339         assert(c != head, "should have found a safepoint on the way");
2340         if (stack.size() != 1 || _phase->is_dominator(head, c)) {
2341           for (;;) {
2342             if (c->is_Region()) {
2343               stack.push(c, 1);
2344               break;
2345             } else if (c->is_SafePoint() && !c->is_CallLeaf()) {
2346               Node* m = c->in(TypeFunc::Memory);
2347               if (m->is_MergeMem()) {
2348                 m = m->as_MergeMem()->memory_at(_alias);
2349               }
2350               assert(mem == nullptr || mem == m, "several memory states");
2351               mem = m;
2352               break;
2353             } else {
2354               assert(c != c->in(0), "");
2355               c = c->in(0);
2356             }
2357           }
2358         }
2359       }
2360     } while (stack.size() > 0);
2361     assert(mem != nullptr, "should have found safepoint");
2362   } else {
2363     mem = phi_mem;
2364   }
2365   return mem;
2366 }
2367 
2368 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
2369   Node* c = _phase->get_ctrl(n);
2370   if (n->is_Proj() && n->in(0) != nullptr && n->in(0)->is_Call()) {
2371     assert(c == n->in(0), "");
2372     CallNode* call = c->as_Call();
2373     CallProjections projs;
2374     call->extract_projections(&projs, true, false);
2375     if (projs.catchall_memproj != nullptr) {
2376       if (projs.fallthrough_memproj == n) {
2377         c = projs.fallthrough_catchproj;
2378       } else {
2379         assert(projs.catchall_memproj == n, "");
2380         c = projs.catchall_catchproj;
2381       }
2382     }
2383   }
2384   return c;
2385 }
2386 
2387 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
2388   if (_phase->has_ctrl(n))
2389     return get_ctrl(n);
2390   else {
2391     assert (n->is_CFG(), "must be a CFG node");
2392     return n;
2393   }
2394 }
2395 
2396 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
2397   return m != nullptr && get_ctrl(m) == c;
2398 }
2399 
2400 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
2401   assert(n == nullptr || _phase->ctrl_or_self(n) == ctrl, "");
2402   assert(!ctrl->is_Call() || ctrl == n, "projection expected");
2403 #ifdef ASSERT
2404   if ((ctrl->is_Proj() && ctrl->in(0)->is_Call()) ||
2405       (ctrl->is_Catch() && ctrl->in(0)->in(0)->is_Call())) {
2406     CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_Call() : ctrl->in(0)->in(0)->as_Call();
2407     int mems = 0;
2408     for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
2409       Node* u = call->fast_out(i);
2410       if (u->bottom_type() == Type::MEMORY) {
2411         mems++;
2412       }
2413     }
2414     assert(mems <= 1, "No node right after call if multiple mem projections");
2415   }
2416 #endif
2417   Node* mem = _memory_nodes[ctrl->_idx];
2418   Node* c = ctrl;
2419   while (!mem_is_valid(mem, c) &&
2420          (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
2421     c = _phase->idom(c);
2422     mem = _memory_nodes[c->_idx];
2423   }
2424   if (n != nullptr && mem_is_valid(mem, c)) {
2425     while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
2426       mem = next_mem(mem, _alias);
2427     }
2428     if (mem->is_MergeMem()) {
2429       mem = mem->as_MergeMem()->memory_at(_alias);
2430     }
2431     if (!mem_is_valid(mem, c)) {
2432       do {
2433         c = _phase->idom(c);
2434         mem = _memory_nodes[c->_idx];
2435       } while (!mem_is_valid(mem, c) &&
2436                (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
2437     }
2438   }
2439   assert(mem->bottom_type() == Type::MEMORY, "");
2440   return mem;
2441 }
2442 
2443 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
2444   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2445     Node* use = region->fast_out(i);
2446     if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
2447         (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
2448       return true;
2449     }
2450   }
2451   return false;
2452 }
2453 
2454 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
2455   assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
2456   const bool trace = false;
2457   DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
2458   DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
2459   GrowableArray<Node*> phis;
2460   if (mem_for_ctrl != mem) {
2461     Node* old = mem_for_ctrl;
2462     Node* prev = nullptr;
2463     while (old != mem) {
2464       prev = old;
2465       if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
2466         assert(_alias == Compile::AliasIdxRaw, "");
2467         old = old->in(MemNode::Memory);
2468       } else if (old->Opcode() == Op_SCMemProj) {
2469         assert(_alias == Compile::AliasIdxRaw, "");
2470         old = old->in(0);
2471       } else {
2472         ShouldNotReachHere();
2473       }
2474     }
2475     assert(prev != nullptr, "");
2476     if (new_ctrl != ctrl) {
2477       _memory_nodes.map(ctrl->_idx, mem);
2478       _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2479     }
2480     uint input = (uint)MemNode::Memory;
2481     _phase->igvn().replace_input_of(prev, input, new_mem);
2482   } else {
2483     uses.clear();
2484     _memory_nodes.map(new_ctrl->_idx, new_mem);
2485     uses.push(new_ctrl);
2486     for(uint next = 0; next < uses.size(); next++ ) {
2487       Node *n = uses.at(next);
2488       assert(n->is_CFG(), "");
2489       DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
2490       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2491         Node* u = n->fast_out(i);
2492         if (!u->is_Root() && u->is_CFG() && u != n) {
2493           Node* m = _memory_nodes[u->_idx];
2494           if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) &&
2495               !has_mem_phi(u) &&
2496               u->unique_ctrl_out()->Opcode() != Op_Halt) {
2497             DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
2498             DEBUG_ONLY(if (trace && m != nullptr) { tty->print("ZZZ mem"); m->dump(); });
2499 
2500             if (!mem_is_valid(m, u) || !m->is_Phi()) {
2501               bool push = true;
2502               bool create_phi = true;
2503               if (_phase->is_dominator(new_ctrl, u)) {
2504                 create_phi = false;
2505               }
2506               if (create_phi) {
2507                 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
2508                 _phase->register_new_node(phi, u);
2509                 phis.push(phi);
2510                 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
2511                 if (!mem_is_valid(m, u)) {
2512                   DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
2513                   _memory_nodes.map(u->_idx, phi);
2514                 } else {
2515                   DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
2516                   for (;;) {
2517                     assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), "");
2518                     Node* next = nullptr;
2519                     if (m->is_Proj()) {
2520                       next = m->in(0);
2521                     } else {
2522                       assert(m->is_Mem() || m->is_LoadStore(), "");
2523                       assert(_alias == Compile::AliasIdxRaw, "");
2524                       next = m->in(MemNode::Memory);
2525                     }
2526                     if (_phase->get_ctrl(next) != u) {
2527                       break;
2528                     }
2529                     if (next->is_MergeMem()) {
2530                       assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
2531                       break;
2532                     }
2533                     if (next->is_Phi()) {
2534                       assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
2535                       break;
2536                     }
2537                     m = next;
2538                   }
2539 
2540                   DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
2541                   assert(m->is_Mem() || m->is_LoadStore(), "");
2542                   uint input = (uint)MemNode::Memory;
2543                   _phase->igvn().replace_input_of(m, input, phi);
2544                   push = false;
2545                 }
2546               } else {
2547                 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
2548               }
2549               if (push) {
2550                 uses.push(u);
2551               }
2552             }
2553           } else if (!mem_is_valid(m, u) &&
2554                      !(u->Opcode() == Op_CProj && u->in(0)->is_NeverBranch() && u->as_Proj()->_con == 1)) {
2555             uses.push(u);
2556           }
2557         }
2558       }
2559     }
2560     for (int i = 0; i < phis.length(); i++) {
2561       Node* n = phis.at(i);
2562       Node* r = n->in(0);
2563       DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
2564       for (uint j = 1; j < n->req(); j++) {
2565         Node* m = find_mem(r->in(j), nullptr);
2566         _phase->igvn().replace_input_of(n, j, m);
2567         DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
2568       }
2569     }
2570   }
2571   uint last = _phase->C->unique();
2572   MergeMemNode* mm = nullptr;
2573   int alias = _alias;
2574   DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
2575   // Process loads first to not miss an anti-dependency: if the memory
2576   // edge of a store is updated before a load is processed then an
2577   // anti-dependency may be missed.
2578   for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2579     Node* u = mem->out(i);
2580     if (u->_idx < last && u->is_Load() && _phase->C->get_alias_index(u->adr_type()) == alias) {
2581       Node* m = find_mem(_phase->get_ctrl(u), u);
2582       if (m != mem) {
2583         DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2584         _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2585         --i;
2586       }
2587     }
2588   }
2589   for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2590     Node* u = mem->out(i);
2591     if (u->_idx < last) {
2592       if (u->is_Mem()) {
2593         if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2594           Node* m = find_mem(_phase->get_ctrl(u), u);
2595           if (m != mem) {
2596             DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2597             _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2598             --i;
2599           }
2600         }
2601       } else if (u->is_MergeMem()) {
2602         MergeMemNode* u_mm = u->as_MergeMem();
2603         if (u_mm->memory_at(alias) == mem) {
2604           MergeMemNode* newmm = nullptr;
2605           for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2606             Node* uu = u->fast_out(j);
2607             assert(!uu->is_MergeMem(), "chain of MergeMems?");
2608             if (uu->is_Phi()) {
2609               assert(uu->adr_type() == TypePtr::BOTTOM, "");
2610               Node* region = uu->in(0);
2611               int nb = 0;
2612               for (uint k = 1; k < uu->req(); k++) {
2613                 if (uu->in(k) == u) {
2614                   Node* m = find_mem(region->in(k), nullptr);
2615                   if (m != mem) {
2616                     DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
2617                     newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2618                     if (newmm != u) {
2619                       _phase->igvn().replace_input_of(uu, k, newmm);
2620                       nb++;
2621                       --jmax;
2622                     }
2623                   }
2624                 }
2625               }
2626               if (nb > 0) {
2627                 --j;
2628               }
2629             } else {
2630               Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
2631               if (m != mem) {
2632                 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
2633                 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2634                 if (newmm != u) {
2635                   _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2636                   --j, --jmax;
2637                 }
2638               }
2639             }
2640           }
2641         }
2642       } else if (u->is_Phi()) {
2643         assert(u->bottom_type() == Type::MEMORY, "what else?");
2644         if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
2645           Node* region = u->in(0);
2646           bool replaced = false;
2647           for (uint j = 1; j < u->req(); j++) {
2648             if (u->in(j) == mem) {
2649               Node* m = find_mem(region->in(j), nullptr);
2650               Node* nnew = m;
2651               if (m != mem) {
2652                 if (u->adr_type() == TypePtr::BOTTOM) {
2653                   mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
2654                   nnew = mm;
2655                 }
2656                 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
2657                 _phase->igvn().replace_input_of(u, j, nnew);
2658                 replaced = true;
2659               }
2660             }
2661           }
2662           if (replaced) {
2663             --i;
2664           }
2665         }
2666       } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2667                  u->adr_type() == nullptr) {
2668         assert(u->adr_type() != nullptr ||
2669                u->Opcode() == Op_Rethrow ||
2670                u->Opcode() == Op_Return ||
2671                u->Opcode() == Op_SafePoint ||
2672                (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2673                (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2674                u->Opcode() == Op_CallLeaf, "");
2675         Node* m = find_mem(_phase->ctrl_or_self(u), u);
2676         if (m != mem) {
2677           mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
2678           _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2679           --i;
2680         }
2681       } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2682         Node* m = find_mem(_phase->ctrl_or_self(u), u);
2683         if (m != mem) {
2684           DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2685           _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2686           --i;
2687         }
2688       } else if (u->adr_type() != TypePtr::BOTTOM &&
2689                  _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
2690         Node* m = find_mem(_phase->ctrl_or_self(u), u);
2691         assert(m != mem, "");
2692         // u is on the wrong slice...
2693         assert(u->is_ClearArray(), "");
2694         DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2695         _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2696         --i;
2697       }
2698     }
2699   }
2700 #ifdef ASSERT
2701   assert(new_mem->outcnt() > 0, "");
2702   for (int i = 0; i < phis.length(); i++) {
2703     Node* n = phis.at(i);
2704     assert(n->outcnt() > 0, "new phi must have uses now");
2705   }
2706 #endif
2707 }
2708 
2709 void MemoryGraphFixer::record_new_ctrl(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl) {
2710   if (mem_for_ctrl != mem && new_ctrl != ctrl) {
2711     _memory_nodes.map(ctrl->_idx, mem);
2712     _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2713   }
2714 }
2715 
2716 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
2717   MergeMemNode* mm = MergeMemNode::make(mem);
2718   mm->set_memory_at(_alias, rep_proj);
2719   _phase->register_new_node(mm, rep_ctrl);
2720   return mm;
2721 }
2722 
2723 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
2724   MergeMemNode* newmm = nullptr;
2725   MergeMemNode* u_mm = u->as_MergeMem();
2726   Node* c = _phase->get_ctrl(u);
2727   if (_phase->is_dominator(c, rep_ctrl)) {
2728     c = rep_ctrl;
2729   } else {
2730     assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
2731   }
2732   if (u->outcnt() == 1) {
2733     if (u->req() > (uint)_alias && u->in(_alias) == mem) {
2734       _phase->igvn().replace_input_of(u, _alias, rep_proj);
2735       --i;
2736     } else {
2737       _phase->igvn().rehash_node_delayed(u);
2738       u_mm->set_memory_at(_alias, rep_proj);
2739     }
2740     newmm = u_mm;
2741     _phase->set_ctrl_and_loop(u, c);
2742   } else {
2743     // can't simply clone u and then change one of its input because
2744     // it adds and then removes an edge which messes with the
2745     // DUIterator
2746     newmm = MergeMemNode::make(u_mm->base_memory());
2747     for (uint j = 0; j < u->req(); j++) {
2748       if (j < newmm->req()) {
2749         if (j == (uint)_alias) {
2750           newmm->set_req(j, rep_proj);
2751         } else if (newmm->in(j) != u->in(j)) {
2752           newmm->set_req(j, u->in(j));
2753         }
2754       } else if (j == (uint)_alias) {
2755         newmm->add_req(rep_proj);
2756       } else {
2757         newmm->add_req(u->in(j));
2758       }
2759     }
2760     if ((uint)_alias >= u->req()) {
2761       newmm->set_memory_at(_alias, rep_proj);
2762     }
2763     _phase->register_new_node(newmm, c);
2764   }
2765   return newmm;
2766 }
2767 
2768 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
2769   if (phi->adr_type() == TypePtr::BOTTOM) {
2770     Node* region = phi->in(0);
2771     for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
2772       Node* uu = region->fast_out(j);
2773       if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
2774         return false;
2775       }
2776     }
2777     return true;
2778   }
2779   return _phase->C->get_alias_index(phi->adr_type()) == _alias;
2780 }
2781 
2782 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
2783   uint last = _phase-> C->unique();
2784   MergeMemNode* mm = nullptr;
2785   assert(mem->bottom_type() == Type::MEMORY, "");
2786   for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2787     Node* u = mem->out(i);
2788     if (u != replacement && u->_idx < last) {
2789       if (u->is_MergeMem()) {
2790         MergeMemNode* u_mm = u->as_MergeMem();
2791         if (u_mm->memory_at(_alias) == mem) {
2792           MergeMemNode* newmm = nullptr;
2793           for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2794             Node* uu = u->fast_out(j);
2795             assert(!uu->is_MergeMem(), "chain of MergeMems?");
2796             if (uu->is_Phi()) {
2797               if (should_process_phi(uu)) {
2798                 Node* region = uu->in(0);
2799                 int nb = 0;
2800                 for (uint k = 1; k < uu->req(); k++) {
2801                   if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
2802                     if (newmm == nullptr) {
2803                       newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2804                     }
2805                     if (newmm != u) {
2806                       _phase->igvn().replace_input_of(uu, k, newmm);
2807                       nb++;
2808                       --jmax;
2809                     }
2810                   }
2811                 }
2812                 if (nb > 0) {
2813                   --j;
2814                 }
2815               }
2816             } else {
2817               if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
2818                 if (newmm == nullptr) {
2819                   newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2820                 }
2821                 if (newmm != u) {
2822                   _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2823                   --j, --jmax;
2824                 }
2825               }
2826             }
2827           }
2828         }
2829       } else if (u->is_Phi()) {
2830         assert(u->bottom_type() == Type::MEMORY, "what else?");
2831         Node* region = u->in(0);
2832         if (should_process_phi(u)) {
2833           bool replaced = false;
2834           for (uint j = 1; j < u->req(); j++) {
2835             if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
2836               Node* nnew = rep_proj;
2837               if (u->adr_type() == TypePtr::BOTTOM) {
2838                 if (mm == nullptr) {
2839                   mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2840                 }
2841                 nnew = mm;
2842               }
2843               _phase->igvn().replace_input_of(u, j, nnew);
2844               replaced = true;
2845             }
2846           }
2847           if (replaced) {
2848             --i;
2849           }
2850 
2851         }
2852       } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2853                  u->adr_type() == nullptr) {
2854         assert(u->adr_type() != nullptr ||
2855                u->Opcode() == Op_Rethrow ||
2856                u->Opcode() == Op_Return ||
2857                u->Opcode() == Op_SafePoint ||
2858                (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2859                (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2860                u->Opcode() == Op_CallLeaf, "%s", u->Name());
2861         if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2862           if (mm == nullptr) {
2863             mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2864           }
2865           _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2866           --i;
2867         }
2868       } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2869         if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2870           _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
2871           --i;
2872         }
2873       }
2874     }
2875   }
2876 }
2877 
2878 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj, DecoratorSet decorators)
2879 : Node(ctrl, obj), _decorators(decorators) {
2880   ShenandoahBarrierSetC2::bsc2()->state()->add_load_reference_barrier(this);
2881 }
2882 
2883 DecoratorSet ShenandoahLoadReferenceBarrierNode::decorators() const {
2884   return _decorators;
2885 }
2886 
2887 uint ShenandoahLoadReferenceBarrierNode::size_of() const {
2888   return sizeof(*this);
2889 }
2890 
2891 static DecoratorSet mask_decorators(DecoratorSet decorators) {
2892   return decorators & (ON_STRONG_OOP_REF | ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF | ON_UNKNOWN_OOP_REF | IN_NATIVE);
2893 }
2894 
2895 uint ShenandoahLoadReferenceBarrierNode::hash() const {
2896   uint hash = Node::hash();
2897   hash += mask_decorators(_decorators);
2898   return hash;
2899 }
2900 
2901 bool ShenandoahLoadReferenceBarrierNode::cmp( const Node &n ) const {
2902   return Node::cmp(n) && n.Opcode() == Op_ShenandoahLoadReferenceBarrier &&
2903          mask_decorators(_decorators) == mask_decorators(((const ShenandoahLoadReferenceBarrierNode&)n)._decorators);
2904 }
2905 
2906 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const {
2907   if (in(ValueIn) == nullptr || in(ValueIn)->is_top()) {
2908     return Type::TOP;
2909   }
2910   const Type* t = in(ValueIn)->bottom_type();
2911   if (t == TypePtr::NULL_PTR) {
2912     return t;
2913   }
2914 
2915   if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2916     return t;
2917   }
2918 
2919   return t->meet(TypePtr::NULL_PTR);
2920 }
2921 
2922 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseGVN* phase) const {
2923   // Either input is TOP ==> the result is TOP
2924   const Type *t2 = phase->type(in(ValueIn));
2925   if( t2 == Type::TOP ) return Type::TOP;
2926 
2927   if (t2 == TypePtr::NULL_PTR) {
2928     return t2;
2929   }
2930 
2931   if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2932     return t2;
2933   }
2934 
2935   return t2->meet(TypePtr::NULL_PTR);
2936 }
2937 
2938 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseGVN* phase) {
2939   Node* value = in(ValueIn);
2940   if (!needs_barrier(phase, value)) {
2941     return value;
2942   }
2943   return this;
2944 }
2945 
2946 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseGVN* phase, Node* n) {
2947   Unique_Node_List visited;
2948   return needs_barrier_impl(phase, n, visited);
2949 }
2950 
2951 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseGVN* phase, Node* n, Unique_Node_List &visited) {
2952   if (n == nullptr) return false;
2953   if (visited.member(n)) {
2954     return false; // Been there.
2955   }
2956   visited.push(n);
2957 
2958   if (n->is_Allocate()) {
2959     // tty->print_cr("optimize barrier on alloc");
2960     return false;
2961   }
2962   if (n->is_Call()) {
2963     // tty->print_cr("optimize barrier on call");
2964     return false;
2965   }
2966 
2967   const Type* type = phase->type(n);
2968   if (type == Type::TOP) {
2969     return false;
2970   }
2971   if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) {
2972     // tty->print_cr("optimize barrier on null");
2973     return false;
2974   }
2975   if (type->make_oopptr() && type->make_oopptr()->const_oop() != nullptr) {
2976     // tty->print_cr("optimize barrier on constant");
2977     return false;
2978   }
2979 
2980   switch (n->Opcode()) {
2981     case Op_AddP:
2982       return true; // TODO: Can refine?
2983     case Op_LoadP:
2984     case Op_ShenandoahCompareAndExchangeN:
2985     case Op_ShenandoahCompareAndExchangeP:
2986     case Op_CompareAndExchangeN:
2987     case Op_CompareAndExchangeP:
2988     case Op_GetAndSetN:
2989     case Op_GetAndSetP:
2990       return true;
2991     case Op_Phi: {
2992       for (uint i = 1; i < n->req(); i++) {
2993         if (needs_barrier_impl(phase, n->in(i), visited)) return true;
2994       }
2995       return false;
2996     }
2997     case Op_CheckCastPP:
2998     case Op_CastPP:
2999       return needs_barrier_impl(phase, n->in(1), visited);
3000     case Op_Proj:
3001       return needs_barrier_impl(phase, n->in(0), visited);
3002     case Op_ShenandoahLoadReferenceBarrier:
3003       // tty->print_cr("optimize barrier on barrier");
3004       return false;
3005     case Op_Parm:
3006       // tty->print_cr("optimize barrier on input arg");
3007       return false;
3008     case Op_DecodeN:
3009     case Op_EncodeP:
3010       return needs_barrier_impl(phase, n->in(1), visited);
3011     case Op_LoadN:
3012       return true;
3013     case Op_CMoveN:
3014     case Op_CMoveP:
3015       return needs_barrier_impl(phase, n->in(2), visited) ||
3016              needs_barrier_impl(phase, n->in(3), visited);
3017     case Op_ShenandoahIUBarrier:
3018       return needs_barrier_impl(phase, n->in(1), visited);
3019     case Op_CreateEx:
3020       return false;
3021     default:
3022       break;
3023   }
3024 #ifdef ASSERT
3025   tty->print("need barrier on?: ");
3026   tty->print_cr("ins:");
3027   n->dump(2);
3028   tty->print_cr("outs:");
3029   n->dump(-2);
3030   ShouldNotReachHere();
3031 #endif
3032   return true;
3033 }