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