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 
  27 #include "classfile/javaClasses.hpp"
  28 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
  29 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
  30 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
  31 #include "gc/shenandoah/shenandoahForwarding.hpp"
  32 #include "gc/shenandoah/shenandoahHeap.hpp"
  33 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
  34 #include "gc/shenandoah/shenandoahRuntime.hpp"
  35 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
  36 #include "opto/arraycopynode.hpp"
  37 #include "opto/block.hpp"
  38 #include "opto/callnode.hpp"
  39 #include "opto/castnode.hpp"
  40 #include "opto/movenode.hpp"
  41 #include "opto/phaseX.hpp"
  42 #include "opto/rootnode.hpp"
  43 #include "opto/runtime.hpp"
  44 #include "opto/subnode.hpp"
  45 
  46 bool ShenandoahBarrierC2Support::expand(Compile* C, PhaseIterGVN& igvn) {
  47   ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
  48   if (state->load_reference_barriers_count() > 0) {
  49     assert(C->post_loop_opts_phase(), "no loop opts allowed");
  50     C->reset_post_loop_opts_phase(); // ... but we know what we are doing
  51     C->clear_major_progress();
  52     PhaseIdealLoop::optimize(igvn, LoopOptsShenandoahExpand);
  53     if (C->failing()) return false;
  54     C->process_for_post_loop_opts_igvn(igvn);
  55     if (C->failing()) return false;
  56 
  57     C->set_post_loop_opts_phase(); // now for real!
  58   }
  59   return true;
  60 }
  61 
  62 bool ShenandoahBarrierC2Support::is_gc_state_test(Node* iff, int mask) {
  63   if (!UseShenandoahGC) {
  64     return false;
  65   }
  66   assert(iff->is_If(), "bad input");
  67   if (iff->Opcode() != Op_If) {
  68     return false;
  69   }
  70   Node* bol = iff->in(1);
  71   if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) {
  72     return false;
  73   }
  74   Node* cmp = bol->in(1);
  75   if (cmp->Opcode() != Op_CmpI) {
  76     return false;
  77   }
  78   Node* in1 = cmp->in(1);
  79   Node* in2 = cmp->in(2);
  80   if (in2->find_int_con(-1) != 0) {
  81     return false;
  82   }
  83   if (in1->Opcode() != Op_AndI) {
  84     return false;
  85   }
  86   in2 = in1->in(2);
  87   if (in2->find_int_con(-1) != mask) {
  88     return false;
  89   }
  90   in1 = in1->in(1);
  91 
  92   return is_gc_state_load(in1);
  93 }
  94 
  95 bool ShenandoahBarrierC2Support::is_heap_stable_test(Node* iff) {
  96   return is_gc_state_test(iff, ShenandoahHeap::HAS_FORWARDED);
  97 }
  98 
  99 bool ShenandoahBarrierC2Support::is_gc_state_load(Node *n) {
 100   if (!UseShenandoahGC) {
 101     return false;
 102   }
 103   if (n->Opcode() != Op_LoadB && n->Opcode() != Op_LoadUB) {
 104     return false;
 105   }
 106   Node* addp = n->in(MemNode::Address);
 107   if (!addp->is_AddP()) {
 108     return false;
 109   }
 110   Node* base = addp->in(AddPNode::Address);
 111   Node* off = addp->in(AddPNode::Offset);
 112   if (base->Opcode() != Op_ThreadLocal) {
 113     return false;
 114   }
 115   if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) {
 116     return false;
 117   }
 118   return true;
 119 }
 120 
 121 bool ShenandoahBarrierC2Support::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) {
 122   assert(phase->is_dominator(stop, start), "bad inputs");
 123   ResourceMark rm;
 124   Unique_Node_List wq;
 125   wq.push(start);
 126   for (uint next = 0; next < wq.size(); next++) {
 127     Node *m = wq.at(next);
 128     if (m == stop) {
 129       continue;
 130     }
 131     if (m->is_SafePoint() && !m->is_CallLeaf()) {
 132       return true;
 133     }
 134     if (m->is_Region()) {
 135       for (uint i = 1; i < m->req(); i++) {
 136         wq.push(m->in(i));
 137       }
 138     } else {
 139       wq.push(m->in(0));
 140     }
 141   }
 142   return false;
 143 }
 144 
 145 #ifdef ASSERT
 146 bool ShenandoahBarrierC2Support::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) {
 147   assert(phis.size() == 0, "");
 148 
 149   while (true) {
 150     if (in->bottom_type() == TypePtr::NULL_PTR) {
 151       if (trace) {tty->print_cr("null");}
 152     } else if (!in->bottom_type()->make_ptr()->make_oopptr()) {
 153       if (trace) {tty->print_cr("Non oop");}
 154     } else {
 155       if (in->is_ConstraintCast()) {
 156         in = in->in(1);
 157         continue;
 158       } else if (in->is_AddP()) {
 159         assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access");
 160         in = in->in(AddPNode::Address);
 161         continue;
 162       } else if (in->is_Con()) {
 163         if (trace) {
 164           tty->print("Found constant");
 165           in->dump();
 166         }
 167       } else if (in->Opcode() == Op_Parm) {
 168         if (trace) {
 169           tty->print("Found argument");
 170         }
 171       } else if (in->Opcode() == Op_CreateEx) {
 172         if (trace) {
 173           tty->print("Found create-exception");
 174         }
 175       } else if (in->Opcode() == Op_LoadP && in->adr_type() == TypeRawPtr::BOTTOM) {
 176         if (trace) {
 177           tty->print("Found raw LoadP (OSR argument?)");
 178         }
 179       } else if (in->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 180         if (t == ShenandoahOopStore) {
 181           return false;
 182         }
 183         barriers_used.push(in);
 184         if (trace) {tty->print("Found barrier"); in->dump();}
 185       } else if (in->is_Proj() && in->in(0)->is_Allocate()) {
 186         if (trace) {
 187           tty->print("Found alloc");
 188           in->in(0)->dump();
 189         }
 190       } else if (in->is_Proj() && (in->in(0)->Opcode() == Op_CallStaticJava || in->in(0)->Opcode() == Op_CallDynamicJava)) {
 191         if (trace) {
 192           tty->print("Found Java call");
 193         }
 194       } else if (in->is_Phi()) {
 195         if (!visited.test_set(in->_idx)) {
 196           if (trace) {tty->print("Pushed phi:"); in->dump();}
 197           phis.push(in, 2);
 198           in = in->in(1);
 199           continue;
 200         }
 201         if (trace) {tty->print("Already seen phi:"); in->dump();}
 202       } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) {
 203         if (!visited.test_set(in->_idx)) {
 204           if (trace) {tty->print("Pushed cmovep:"); in->dump();}
 205           phis.push(in, CMoveNode::IfTrue);
 206           in = in->in(CMoveNode::IfFalse);
 207           continue;
 208         }
 209         if (trace) {tty->print("Already seen cmovep:"); in->dump();}
 210       } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) {
 211         in = in->in(1);
 212         continue;
 213       } else {
 214         return false;
 215       }
 216     }
 217     bool cont = false;
 218     while (phis.is_nonempty()) {
 219       uint idx = phis.index();
 220       Node* phi = phis.node();
 221       if (idx >= phi->req()) {
 222         if (trace) {tty->print("Popped phi:"); phi->dump();}
 223         phis.pop();
 224         continue;
 225       }
 226       if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();}
 227       in = phi->in(idx);
 228       phis.set_index(idx+1);
 229       cont = true;
 230       break;
 231     }
 232     if (!cont) {
 233       break;
 234     }
 235   }
 236   return true;
 237 }
 238 
 239 void ShenandoahBarrierC2Support::report_verify_failure(const char* msg, Node* n1, Node* n2) {
 240   if (n1 != nullptr) {
 241     n1->dump(+10);
 242   }
 243   if (n2 != nullptr) {
 244     n2->dump(+10);
 245   }
 246   fatal("%s", msg);
 247 }
 248 
 249 void ShenandoahBarrierC2Support::verify(RootNode* root) {
 250   ResourceMark rm;
 251   Unique_Node_List wq;
 252   GrowableArray<Node*> barriers;
 253   Unique_Node_List barriers_used;
 254   Node_Stack phis(0);
 255   VectorSet visited;
 256   const bool trace = false;
 257   const bool verify_no_useless_barrier = false;
 258 
 259   wq.push(root);
 260   for (uint next = 0; next < wq.size(); next++) {
 261     Node *n = wq.at(next);
 262     if (n->is_Load()) {
 263       const bool trace = false;
 264       if (trace) {tty->print("Verifying"); n->dump();}
 265       if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) {
 266         if (trace) {tty->print_cr("Load range/klass");}
 267       } else {
 268         const TypePtr* adr_type = n->as_Load()->adr_type();
 269 
 270         if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) {
 271           if (trace) {tty->print_cr("Mark load");}
 272         } else if (adr_type->isa_instptr() &&
 273                    adr_type->is_instptr()->instance_klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
 274                    adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset()) {
 275           if (trace) {tty->print_cr("Reference.get()");}
 276         } else if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) {
 277           report_verify_failure("Shenandoah verification: Load should have barriers", n);
 278         }
 279       }
 280     } else if (n->is_Store()) {
 281       const bool trace = false;
 282 
 283       if (trace) {tty->print("Verifying"); n->dump();}
 284       if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) {
 285         Node* adr = n->in(MemNode::Address);
 286         bool verify = true;
 287 
 288         if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) {
 289           adr = adr->in(AddPNode::Address);
 290           if (adr->is_AddP()) {
 291             assert(adr->in(AddPNode::Base)->is_top(), "");
 292             adr = adr->in(AddPNode::Address);
 293             if (adr->Opcode() == Op_LoadP &&
 294                 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() &&
 295                 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
 296                 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) {
 297               if (trace) {tty->print_cr("SATB prebarrier");}
 298               verify = false;
 299             }
 300           }
 301         }
 302 
 303         if (verify && !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahValue, trace, barriers_used)) {
 304           report_verify_failure("Shenandoah verification: Store should have barriers", n);
 305         }
 306       }
 307       if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
 308         report_verify_failure("Shenandoah verification: Store (address) should have barriers", n);
 309       }
 310     } else if (n->Opcode() == Op_CmpP) {
 311       const bool trace = false;
 312 
 313       Node* in1 = n->in(1);
 314       Node* in2 = n->in(2);
 315       if (in1->bottom_type()->isa_oopptr()) {
 316         if (trace) {tty->print("Verifying"); n->dump();}
 317 
 318         bool mark_inputs = false;
 319         if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR ||
 320             (in1->is_Con() || in2->is_Con())) {
 321           if (trace) {tty->print_cr("Comparison against a constant");}
 322           mark_inputs = true;
 323         } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) ||
 324                    (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) {
 325           if (trace) {tty->print_cr("Comparison with newly alloc'ed object");}
 326           mark_inputs = true;
 327         } else {
 328           assert(in2->bottom_type()->isa_oopptr(), "");
 329 
 330           if (!verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) ||
 331               !verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) {
 332             report_verify_failure("Shenandoah verification: Cmp should have barriers", n);
 333           }
 334         }
 335         if (verify_no_useless_barrier &&
 336             mark_inputs &&
 337             (!verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) ||
 338              !verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) {
 339           phis.clear();
 340           visited.reset();
 341         }
 342       }
 343     } else if (n->is_LoadStore()) {
 344       if (n->in(MemNode::ValueIn)->bottom_type()->make_ptr() &&
 345           !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahValue, trace, barriers_used)) {
 346         report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n);
 347       }
 348 
 349       if (n->in(MemNode::Address)->bottom_type()->make_oopptr() && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
 350         report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n);
 351       }
 352     } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) {
 353       CallNode* call = n->as_Call();
 354 
 355       static struct {
 356         const char* name;
 357         struct {
 358           int pos;
 359           verify_type t;
 360         } args[6];
 361       } calls[] = {
 362         "array_partition_stub",
 363         { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+4, ShenandoahStore },   { -1, ShenandoahNone },
 364           { -1, ShenandoahNone },                { -1, ShenandoahNone },                  { -1, ShenandoahNone } },
 365         "arraysort_stub",
 366         { { TypeFunc::Parms, ShenandoahStore },  { -1, ShenandoahNone },                  { -1, ShenandoahNone },
 367           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 368         "aescrypt_encryptBlock",
 369         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 370           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 371         "aescrypt_decryptBlock",
 372         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 373           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 374         "multiplyToLen",
 375         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },   { TypeFunc::Parms+4, ShenandoahStore },
 376           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 377         "squareToLen",
 378         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },   { -1,  ShenandoahNone},
 379           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 380         "montgomery_multiply",
 381         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },
 382           { TypeFunc::Parms+6, ShenandoahStore }, { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 383         "montgomery_square",
 384         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahLoad },   { TypeFunc::Parms+5, ShenandoahStore },
 385           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 386         "mulAdd",
 387         { { TypeFunc::Parms, ShenandoahStore },  { TypeFunc::Parms+1, ShenandoahLoad },   { -1,  ShenandoahNone},
 388           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 389         "vectorizedMismatch",
 390         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahLoad },   { -1,  ShenandoahNone},
 391           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 392         "updateBytesCRC32",
 393         { { TypeFunc::Parms+1, ShenandoahLoad }, { -1,  ShenandoahNone},                  { -1,  ShenandoahNone},
 394           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 395         "updateBytesAdler32",
 396         { { TypeFunc::Parms+1, ShenandoahLoad }, { -1,  ShenandoahNone},                  { -1,  ShenandoahNone},
 397           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 398         "updateBytesCRC32C",
 399         { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad},    { -1,  ShenandoahNone},
 400           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 401         "counterMode_AESCrypt",
 402         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 403           { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } },
 404         "cipherBlockChaining_encryptAESCrypt",
 405         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 406           { TypeFunc::Parms+3, ShenandoahLoad },  { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 407         "cipherBlockChaining_decryptAESCrypt",
 408         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 409           { TypeFunc::Parms+3, ShenandoahLoad },  { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 410         "shenandoah_clone",
 411         { { TypeFunc::Parms, ShenandoahLoad },   { -1,  ShenandoahNone},                  { -1,  ShenandoahNone},
 412           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 413         "ghash_processBlocks",
 414         { { TypeFunc::Parms, ShenandoahStore },  { TypeFunc::Parms+1, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },
 415           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 416         "sha1_implCompress",
 417         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 418           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 419         "sha256_implCompress",
 420         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 421           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 422         "sha512_implCompress",
 423         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 424           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 425         "sha1_implCompressMB",
 426         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 427           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 428         "sha256_implCompressMB",
 429         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 430           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 431         "sha512_implCompressMB",
 432         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 433           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 434         "encodeBlock",
 435         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+3, ShenandoahStore },   { -1, ShenandoahNone },
 436           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 437         "decodeBlock",
 438         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+3, ShenandoahStore },   { -1, ShenandoahNone },
 439           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 440         "intpoly_montgomeryMult_P256",
 441         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahLoad  },   { TypeFunc::Parms+2, ShenandoahStore },
 442           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 443         "intpoly_assign",
 444         { { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },  { -1, ShenandoahNone },
 445           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 446       };
 447 
 448       if (call->is_call_to_arraycopystub()) {
 449         Node* dest = nullptr;
 450         const TypeTuple* args = n->as_Call()->_tf->domain_sig();
 451         for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
 452           if (args->field_at(i)->isa_ptr()) {
 453             j++;
 454             if (j == 2) {
 455               dest = n->in(i);
 456               break;
 457             }
 458           }
 459         }
 460         if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) ||
 461             !verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) {
 462           report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n);
 463         }
 464       } else if (strlen(call->_name) > 5 &&
 465                  !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) {
 466         if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) {
 467           report_verify_failure("Shenandoah verification: _fill should have barriers", n);
 468         }
 469       } else if (!strcmp(call->_name, "shenandoah_wb_pre")) {
 470         // skip
 471       } else {
 472         const int calls_len = sizeof(calls) / sizeof(calls[0]);
 473         int i = 0;
 474         for (; i < calls_len; i++) {
 475           if (!strcmp(calls[i].name, call->_name)) {
 476             break;
 477           }
 478         }
 479         if (i != calls_len) {
 480           const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]);
 481           for (uint j = 0; j < args_len; j++) {
 482             int pos = calls[i].args[j].pos;
 483             if (pos == -1) {
 484               break;
 485             }
 486             if (!verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) {
 487               report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
 488             }
 489           }
 490           for (uint j = TypeFunc::Parms; j < call->req(); j++) {
 491             if (call->in(j)->bottom_type()->make_ptr() &&
 492                 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
 493               uint k = 0;
 494               for (; k < args_len && calls[i].args[k].pos != (int)j; k++);
 495               if (k == args_len) {
 496                 fatal("arg %d for call %s not covered", j, call->_name);
 497               }
 498             }
 499           }
 500         } else {
 501           for (uint j = TypeFunc::Parms; j < call->req(); j++) {
 502             if (call->in(j)->bottom_type()->make_ptr() &&
 503                 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
 504               fatal("%s not covered", call->_name);
 505             }
 506           }
 507         }
 508       }
 509     } else if (n->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 510       // skip
 511     } else if (n->is_AddP()
 512                || n->is_Phi()
 513                || n->is_ConstraintCast()
 514                || n->Opcode() == Op_Return
 515                || n->Opcode() == Op_CMoveP
 516                || n->Opcode() == Op_CMoveN
 517                || n->Opcode() == Op_Rethrow
 518                || n->is_MemBar()
 519                || n->Opcode() == Op_Conv2B
 520                || n->Opcode() == Op_SafePoint
 521                || n->is_CallJava()
 522                || n->Opcode() == Op_Unlock
 523                || n->Opcode() == Op_EncodeP
 524                || n->Opcode() == Op_DecodeN) {
 525       // nothing to do
 526     } else {
 527       static struct {
 528         int opcode;
 529         struct {
 530           int pos;
 531           verify_type t;
 532         } inputs[2];
 533       } others[] = {
 534         Op_FastLock,
 535         { { 1, ShenandoahLoad },                  { -1, ShenandoahNone} },
 536         Op_Lock,
 537         { { TypeFunc::Parms, ShenandoahLoad },    { -1, ShenandoahNone} },
 538         Op_ArrayCopy,
 539         { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } },
 540         Op_StrCompressedCopy,
 541         { { 2, ShenandoahLoad },                  { 3, ShenandoahStore } },
 542         Op_StrInflatedCopy,
 543         { { 2, ShenandoahLoad },                  { 3, ShenandoahStore } },
 544         Op_AryEq,
 545         { { 2, ShenandoahLoad },                  { 3, ShenandoahLoad } },
 546         Op_StrIndexOf,
 547         { { 2, ShenandoahLoad },                  { 4, ShenandoahLoad } },
 548         Op_StrComp,
 549         { { 2, ShenandoahLoad },                  { 4, ShenandoahLoad } },
 550         Op_StrEquals,
 551         { { 2, ShenandoahLoad },                  { 3, ShenandoahLoad } },
 552         Op_VectorizedHashCode,
 553         { { 2, ShenandoahLoad },                  { -1, ShenandoahNone } },
 554         Op_EncodeISOArray,
 555         { { 2, ShenandoahLoad },                  { 3, ShenandoahStore } },
 556         Op_CountPositives,
 557         { { 2, ShenandoahLoad },                  { -1, ShenandoahNone} },
 558         Op_CastP2X,
 559         { { 1, ShenandoahLoad },                  { -1, ShenandoahNone} },
 560         Op_StrIndexOfChar,
 561         { { 2, ShenandoahLoad },                  { -1, ShenandoahNone } },
 562       };
 563 
 564       const int others_len = sizeof(others) / sizeof(others[0]);
 565       int i = 0;
 566       for (; i < others_len; i++) {
 567         if (others[i].opcode == n->Opcode()) {
 568           break;
 569         }
 570       }
 571       uint stop = n->is_Call() ? n->as_Call()->tf()->domain_sig()->cnt() : n->req();
 572       if (i != others_len) {
 573         const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]);
 574         for (uint j = 0; j < inputs_len; j++) {
 575           int pos = others[i].inputs[j].pos;
 576           if (pos == -1) {
 577             break;
 578           }
 579           if (!verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) {
 580             report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
 581           }
 582         }
 583         for (uint j = 1; j < stop; j++) {
 584           if (n->in(j) != nullptr && n->in(j)->bottom_type()->make_ptr() &&
 585               n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
 586             uint k = 0;
 587             for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++);
 588             if (k == inputs_len) {
 589               fatal("arg %d for node %s not covered", j, n->Name());
 590             }
 591           }
 592         }
 593       } else {
 594         for (uint j = 1; j < stop; j++) {
 595           if (n->in(j) != nullptr && n->in(j)->bottom_type()->make_ptr() &&
 596               n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
 597             fatal("%s not covered", n->Name());
 598           }
 599         }
 600       }
 601     }
 602 
 603     if (n->is_SafePoint()) {
 604       SafePointNode* sfpt = n->as_SafePoint();
 605       if (verify_no_useless_barrier && sfpt->jvms() != nullptr) {
 606         for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) {
 607           if (!verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) {
 608             phis.clear();
 609             visited.reset();
 610           }
 611         }
 612       }
 613     }
 614   }
 615 
 616   if (verify_no_useless_barrier) {
 617     for (int i = 0; i < barriers.length(); i++) {
 618       Node* n = barriers.at(i);
 619       if (!barriers_used.member(n)) {
 620         tty->print("XXX useless barrier"); n->dump(-2);
 621         ShouldNotReachHere();
 622       }
 623     }
 624   }
 625 }
 626 #endif
 627 
 628 bool ShenandoahBarrierC2Support::is_dominator_same_ctrl(Node* c, Node* d, Node* n, PhaseIdealLoop* phase) {
 629   // That both nodes have the same control is not sufficient to prove
 630   // domination, verify that there's no path from d to n
 631   ResourceMark rm;
 632   Unique_Node_List wq;
 633   wq.push(d);
 634   for (uint next = 0; next < wq.size(); next++) {
 635     Node *m = wq.at(next);
 636     if (m == n) {
 637       return false;
 638     }
 639     if (m->is_Phi() && m->in(0)->is_Loop()) {
 640       assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control");
 641     } else {
 642       if (m->is_Store() || m->is_LoadStore()) {
 643         // Take anti-dependencies into account
 644         Node* mem = m->in(MemNode::Memory);
 645         for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
 646           Node* u = mem->fast_out(i);
 647           if (u->is_Load() && phase->C->can_alias(m->adr_type(), phase->C->get_alias_index(u->adr_type())) &&
 648               phase->ctrl_or_self(u) == c) {
 649             wq.push(u);
 650           }
 651         }
 652       }
 653       for (uint i = 0; i < m->req(); i++) {
 654         if (m->in(i) != nullptr && phase->ctrl_or_self(m->in(i)) == c) {
 655           wq.push(m->in(i));
 656         }
 657       }
 658     }
 659   }
 660   return true;
 661 }
 662 
 663 bool ShenandoahBarrierC2Support::is_dominator(Node* d_c, Node* n_c, Node* d, Node* n, PhaseIdealLoop* phase) {
 664   if (d_c != n_c) {
 665     return phase->is_dominator(d_c, n_c);
 666   }
 667   return is_dominator_same_ctrl(d_c, d, n, phase);
 668 }
 669 
 670 Node* next_mem(Node* mem, int alias) {
 671   Node* res = nullptr;
 672   if (mem->is_Proj()) {
 673     res = mem->in(0);
 674   } else if (mem->is_SafePoint() || mem->is_MemBar()) {
 675     res = mem->in(TypeFunc::Memory);
 676   } else if (mem->is_Phi()) {
 677     res = mem->in(1);
 678   } else if (mem->is_MergeMem()) {
 679     res = mem->as_MergeMem()->memory_at(alias);
 680   } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
 681     assert(alias == Compile::AliasIdxRaw, "following raw memory can't lead to a barrier");
 682     res = mem->in(MemNode::Memory);
 683   } else {
 684 #ifdef ASSERT
 685     mem->dump();
 686 #endif
 687     ShouldNotReachHere();
 688   }
 689   return res;
 690 }
 691 
 692 Node* ShenandoahBarrierC2Support::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) {
 693   Node* iffproj = nullptr;
 694   while (c != dom) {
 695     Node* next = phase->idom(c);
 696     assert(next->unique_ctrl_out_or_null() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?");
 697     if (c->is_Region()) {
 698       ResourceMark rm;
 699       Unique_Node_List wq;
 700       wq.push(c);
 701       for (uint i = 0; i < wq.size(); i++) {
 702         Node *n = wq.at(i);
 703         if (n == next) {
 704           continue;
 705         }
 706         if (n->is_Region()) {
 707           for (uint j = 1; j < n->req(); j++) {
 708             wq.push(n->in(j));
 709           }
 710         } else {
 711           wq.push(n->in(0));
 712         }
 713       }
 714       for (uint i = 0; i < wq.size(); i++) {
 715         Node *n = wq.at(i);
 716         assert(n->is_CFG(), "");
 717         if (n->is_Multi()) {
 718           for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
 719             Node* u = n->fast_out(j);
 720             if (u->is_CFG()) {
 721               if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj()) {
 722                 return NodeSentinel;
 723               }
 724             }
 725           }
 726         }
 727       }
 728     } else  if (c->is_Proj()) {
 729       if (c->is_IfProj()) {
 730         if (c->as_Proj()->is_uncommon_trap_if_pattern() != nullptr) {
 731           // continue;
 732         } else {
 733           if (!allow_one_proj) {
 734             return NodeSentinel;
 735           }
 736           if (iffproj == nullptr) {
 737             iffproj = c;
 738           } else {
 739             return NodeSentinel;
 740           }
 741         }
 742       } else if (c->Opcode() == Op_JumpProj) {
 743         return NodeSentinel; // unsupported
 744       } else if (c->Opcode() == Op_CatchProj) {
 745         return NodeSentinel; // unsupported
 746       } else if (c->Opcode() == Op_CProj && next->is_NeverBranch()) {
 747         return NodeSentinel; // unsupported
 748       } else {
 749         assert(next->unique_ctrl_out() == c, "unsupported branch pattern");
 750       }
 751     }
 752     c = next;
 753   }
 754   return iffproj;
 755 }
 756 
 757 Node* ShenandoahBarrierC2Support::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) {
 758   ResourceMark rm;
 759   VectorSet wq;
 760   wq.set(mem->_idx);
 761   mem_ctrl = phase->ctrl_or_self(mem);
 762   while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) {
 763     mem = next_mem(mem, alias);
 764     if (wq.test_set(mem->_idx)) {
 765       return nullptr;
 766     }
 767     mem_ctrl = phase->ctrl_or_self(mem);
 768   }
 769   if (mem->is_MergeMem()) {
 770     mem = mem->as_MergeMem()->memory_at(alias);
 771     mem_ctrl = phase->ctrl_or_self(mem);
 772   }
 773   return mem;
 774 }
 775 
 776 Node* ShenandoahBarrierC2Support::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) {
 777   Node* mem = nullptr;
 778   Node* c = ctrl;
 779   do {
 780     if (c->is_Region()) {
 781       for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == nullptr; i++) {
 782         Node* u = c->fast_out(i);
 783         if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
 784           if (u->adr_type() == TypePtr::BOTTOM) {
 785             mem = u;
 786           }
 787         }
 788       }
 789     } else {
 790       if (c->is_Call() && c->as_Call()->adr_type() != nullptr) {
 791         CallProjections* projs = c->as_Call()->extract_projections(true, false);
 792         if (projs->fallthrough_memproj != nullptr) {
 793           if (projs->fallthrough_memproj->adr_type() == TypePtr::BOTTOM) {
 794             if (projs->catchall_memproj == nullptr) {
 795               mem = projs->fallthrough_memproj;
 796             } else {
 797               if (phase->is_dominator(projs->fallthrough_catchproj, ctrl)) {
 798                 mem = projs->fallthrough_memproj;
 799               } else {
 800                 assert(phase->is_dominator(projs->catchall_catchproj, ctrl), "one proj must dominate barrier");
 801                 mem = projs->catchall_memproj;
 802               }
 803             }
 804           }
 805         } else {
 806           Node* proj = c->as_Call()->proj_out(TypeFunc::Memory);
 807           if (proj != nullptr &&
 808               proj->adr_type() == TypePtr::BOTTOM) {
 809             mem = proj;
 810           }
 811         }
 812       } else {
 813         for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
 814           Node* u = c->fast_out(i);
 815           if (u->is_Proj() &&
 816               u->bottom_type() == Type::MEMORY &&
 817               u->adr_type() == TypePtr::BOTTOM) {
 818               assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), "");
 819               assert(mem == nullptr, "only one proj");
 820               mem = u;
 821           }
 822         }
 823         assert(!c->is_Call() || c->as_Call()->adr_type() != nullptr || mem == nullptr, "no mem projection expected");
 824       }
 825     }
 826     c = phase->idom(c);
 827   } while (mem == nullptr);
 828   return mem;
 829 }
 830 
 831 void ShenandoahBarrierC2Support::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) {
 832   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 833     Node* u = n->fast_out(i);
 834     if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) {
 835       uses.push(u);
 836     }
 837   }
 838 }
 839 
 840 static void hide_strip_mined_loop(OuterStripMinedLoopNode* outer, CountedLoopNode* inner, PhaseIdealLoop* phase) {
 841   OuterStripMinedLoopEndNode* le = inner->outer_loop_end();
 842   Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl));
 843   phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl));
 844   Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt);
 845   phase->register_control(new_le, phase->get_loop(le), le->in(0));
 846   phase->lazy_replace(outer, new_outer);
 847   phase->lazy_replace(le, new_le);
 848   inner->clear_strip_mined();
 849 }
 850 
 851 void ShenandoahBarrierC2Support::test_gc_state(Node*& ctrl, Node* raw_mem, Node*& test_fail_ctrl,
 852                                                PhaseIdealLoop* phase, int flags) {
 853   PhaseIterGVN& igvn = phase->igvn();
 854   Node* old_ctrl = ctrl;
 855 
 856   Node* thread          = new ThreadLocalNode();
 857   Node* gc_state_offset = igvn.MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 858   Node* gc_state_addr   = new AddPNode(phase->C->top(), thread, gc_state_offset);
 859   Node* gc_state        = new LoadBNode(old_ctrl, raw_mem, gc_state_addr,
 860                                         DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(nullptr),
 861                                         TypeInt::BYTE, MemNode::unordered);
 862   Node* gc_state_and    = new AndINode(gc_state, igvn.intcon(flags));
 863   Node* gc_state_cmp    = new CmpINode(gc_state_and, igvn.zerocon(T_INT));
 864   Node* gc_state_bool   = new BoolNode(gc_state_cmp, BoolTest::ne);
 865 
 866   IfNode* gc_state_iff  = new IfNode(old_ctrl, gc_state_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
 867   ctrl                  = new IfTrueNode(gc_state_iff);
 868   test_fail_ctrl        = new IfFalseNode(gc_state_iff);
 869 
 870   IdealLoopTree* loop = phase->get_loop(old_ctrl);
 871   phase->register_control(gc_state_iff,   loop, old_ctrl);
 872   phase->register_control(ctrl,           loop, gc_state_iff);
 873   phase->register_control(test_fail_ctrl, loop, gc_state_iff);
 874 
 875   phase->register_new_node(thread,        old_ctrl);
 876   phase->register_new_node(gc_state_addr, old_ctrl);
 877   phase->register_new_node(gc_state,      old_ctrl);
 878   phase->register_new_node(gc_state_and,  old_ctrl);
 879   phase->register_new_node(gc_state_cmp,  old_ctrl);
 880   phase->register_new_node(gc_state_bool, old_ctrl);
 881 
 882   phase->set_root_as_ctrl(gc_state_offset);
 883 
 884   assert(is_gc_state_test(gc_state_iff, flags), "Should match the shape");
 885 }
 886 
 887 void ShenandoahBarrierC2Support::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) {
 888   Node* old_ctrl = ctrl;
 889   PhaseIterGVN& igvn = phase->igvn();
 890 
 891   const Type* val_t = igvn.type(val);
 892   if (val_t->meet(TypePtr::NULL_PTR) == val_t) {
 893     Node* null_cmp   = new CmpPNode(val, igvn.zerocon(T_OBJECT));
 894     Node* null_test  = new BoolNode(null_cmp, BoolTest::ne);
 895 
 896     IfNode* null_iff = new IfNode(old_ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
 897     ctrl             = new IfTrueNode(null_iff);
 898     null_ctrl        = new IfFalseNode(null_iff);
 899 
 900     IdealLoopTree* loop = phase->get_loop(old_ctrl);
 901     phase->register_control(null_iff,  loop, old_ctrl);
 902     phase->register_control(ctrl,      loop, null_iff);
 903     phase->register_control(null_ctrl, loop, null_iff);
 904 
 905     phase->register_new_node(null_cmp,  old_ctrl);
 906     phase->register_new_node(null_test, old_ctrl);
 907   }
 908 }
 909 
 910 void ShenandoahBarrierC2Support::test_in_cset(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) {
 911   Node* old_ctrl = ctrl;
 912   PhaseIterGVN& igvn = phase->igvn();
 913 
 914   Node* raw_val        = new CastP2XNode(old_ctrl, val);
 915   Node* cset_idx       = new URShiftXNode(raw_val, igvn.intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint()));
 916 
 917   // Figure out the target cset address with raw pointer math.
 918   // This avoids matching AddP+LoadB that would emit inefficient code.
 919   // See JDK-8245465.
 920   Node* cset_addr_ptr  = igvn.makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr()));
 921   Node* cset_addr      = new CastP2XNode(old_ctrl, cset_addr_ptr);
 922   Node* cset_load_addr = new AddXNode(cset_addr, cset_idx);
 923   Node* cset_load_ptr  = new CastX2PNode(cset_load_addr);
 924 
 925   Node* cset_load      = new LoadBNode(old_ctrl, raw_mem, cset_load_ptr,
 926                                        DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(nullptr),
 927                                        TypeInt::BYTE, MemNode::unordered);
 928   Node* cset_cmp       = new CmpINode(cset_load, igvn.zerocon(T_INT));
 929   Node* cset_bool      = new BoolNode(cset_cmp, BoolTest::ne);
 930 
 931   IfNode* cset_iff     = new IfNode(old_ctrl, cset_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
 932   ctrl                 = new IfTrueNode(cset_iff);
 933   not_cset_ctrl        = new IfFalseNode(cset_iff);
 934 
 935   IdealLoopTree *loop = phase->get_loop(old_ctrl);
 936   phase->register_control(cset_iff,      loop, old_ctrl);
 937   phase->register_control(ctrl,          loop, cset_iff);
 938   phase->register_control(not_cset_ctrl, loop, cset_iff);
 939 
 940   phase->set_root_as_ctrl(cset_addr_ptr);
 941 
 942   phase->register_new_node(raw_val,        old_ctrl);
 943   phase->register_new_node(cset_idx,       old_ctrl);
 944   phase->register_new_node(cset_addr,      old_ctrl);
 945   phase->register_new_node(cset_load_addr, old_ctrl);
 946   phase->register_new_node(cset_load_ptr,  old_ctrl);
 947   phase->register_new_node(cset_load,      old_ctrl);
 948   phase->register_new_node(cset_cmp,       old_ctrl);
 949   phase->register_new_node(cset_bool,      old_ctrl);
 950 }
 951 
 952 void ShenandoahBarrierC2Support::call_lrb_stub(Node*& ctrl, Node*& val, Node* load_addr,
 953                                                DecoratorSet decorators, PhaseIdealLoop* phase) {
 954   IdealLoopTree*loop = phase->get_loop(ctrl);
 955   const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr();
 956 
 957   address calladdr = nullptr;
 958   const char* name = nullptr;
 959   bool is_strong  = ShenandoahBarrierSet::is_strong_access(decorators);
 960   bool is_weak    = ShenandoahBarrierSet::is_weak_access(decorators);
 961   bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
 962   bool is_native  = ShenandoahBarrierSet::is_native_access(decorators);
 963   bool is_narrow  = UseCompressedOops && !is_native;
 964   if (is_strong) {
 965     if (is_narrow) {
 966       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow);
 967       name = "load_reference_barrier_strong_narrow";
 968     } else {
 969       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong);
 970       name = "load_reference_barrier_strong";
 971     }
 972   } else if (is_weak) {
 973     if (is_narrow) {
 974       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow);
 975       name = "load_reference_barrier_weak_narrow";
 976     } else {
 977       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak);
 978       name = "load_reference_barrier_weak";
 979     }
 980   } else {
 981     assert(is_phantom, "only remaining strength");
 982     if (is_narrow) {
 983       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom_narrow);
 984       name = "load_reference_barrier_phantom_narrow";
 985     } else {
 986       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom);
 987       name = "load_reference_barrier_phantom";
 988     }
 989   }
 990   Node* call = new CallLeafNode(ShenandoahBarrierSetC2::load_reference_barrier_Type(), calladdr, name, TypeRawPtr::BOTTOM);
 991 
 992   call->init_req(TypeFunc::Control, ctrl);
 993   call->init_req(TypeFunc::I_O, phase->C->top());
 994   call->init_req(TypeFunc::Memory, phase->C->top());
 995   call->init_req(TypeFunc::FramePtr, phase->C->top());
 996   call->init_req(TypeFunc::ReturnAdr, phase->C->top());
 997   call->init_req(TypeFunc::Parms, val);
 998   call->init_req(TypeFunc::Parms+1, load_addr);
 999   phase->register_control(call, loop, ctrl);
1000   ctrl = new ProjNode(call, TypeFunc::Control);
1001   phase->register_control(ctrl, loop, call);
1002   val = new ProjNode(call, TypeFunc::Parms);
1003   phase->register_new_node(val, call);
1004   val = new CheckCastPPNode(ctrl, val, obj_type);
1005   phase->register_new_node(val, ctrl);
1006 }
1007 
1008 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) {
1009   Node* ctrl = phase->get_ctrl(barrier);
1010   Node* init_raw_mem = fixer.find_mem(ctrl, barrier);
1011 
1012   // Update the control of all nodes that should be after the
1013   // barrier control flow
1014   uses.clear();
1015   // Every node that is control dependent on the barrier's input
1016   // control will be after the expanded barrier. The raw memory (if
1017   // its memory is control dependent on the barrier's input control)
1018   // must stay above the barrier.
1019   uses_to_ignore.clear();
1020   if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) {
1021     uses_to_ignore.push(init_raw_mem);
1022   }
1023   for (uint next = 0; next < uses_to_ignore.size(); next++) {
1024     Node *n = uses_to_ignore.at(next);
1025     for (uint i = 0; i < n->req(); i++) {
1026       Node* in = n->in(i);
1027       if (in != nullptr && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) {
1028         uses_to_ignore.push(in);
1029       }
1030     }
1031   }
1032   for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) {
1033     Node* u = ctrl->fast_out(i);
1034     if (u->_idx < last &&
1035         u != barrier &&
1036         !u->depends_only_on_test() && // preserve dependency on test
1037         !uses_to_ignore.member(u) &&
1038         (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) &&
1039         (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) {
1040       Node* old_c = phase->ctrl_or_self(u);
1041       Node* c = old_c;
1042       if (c != ctrl ||
1043           is_dominator_same_ctrl(old_c, barrier, u, phase) ||
1044           ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) {
1045         phase->igvn().rehash_node_delayed(u);
1046         int nb = u->replace_edge(ctrl, region, &phase->igvn());
1047         if (u->is_CFG()) {
1048           if (phase->idom(u) == ctrl) {
1049             phase->set_idom(u, region, phase->dom_depth(region));
1050           }
1051         } else if (phase->get_ctrl(u) == ctrl) {
1052           assert(u != init_raw_mem, "should leave input raw mem above the barrier");
1053           uses.push(u);
1054         }
1055         assert(nb == 1, "more than 1 ctrl input?");
1056         --i, imax -= nb;
1057       }
1058     }
1059   }
1060 }
1061 
1062 static Node* create_phis_on_call_return(Node* ctrl, Node* c, Node* n, Node* n_clone, const CallProjections* projs, PhaseIdealLoop* phase) {
1063   Node* region = nullptr;
1064   while (c != ctrl) {
1065     if (c->is_Region()) {
1066       region = c;
1067     }
1068     c = phase->idom(c);
1069   }
1070   assert(region != nullptr, "");
1071   Node* phi = new PhiNode(region, n->bottom_type());
1072   for (uint j = 1; j < region->req(); j++) {
1073     Node* in = region->in(j);
1074     if (phase->is_dominator(projs->fallthrough_catchproj, in)) {
1075       phi->init_req(j, n);
1076     } else if (phase->is_dominator(projs->catchall_catchproj, in)) {
1077       phi->init_req(j, n_clone);
1078     } else {
1079       phi->init_req(j, create_phis_on_call_return(ctrl, in, n, n_clone, projs, phase));
1080     }
1081   }
1082   phase->register_new_node(phi, region);
1083   return phi;
1084 }
1085 
1086 void ShenandoahBarrierC2Support::pin_and_expand(PhaseIdealLoop* phase) {
1087   ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
1088 
1089   Unique_Node_List uses;
1090   Node_Stack stack(0);
1091   Node_List clones;
1092   for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1093     ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1094 
1095     Node* ctrl = phase->get_ctrl(lrb);
1096     Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1097 
1098     CallStaticJavaNode* unc = nullptr;
1099     Node* unc_ctrl = nullptr;
1100     Node* uncasted_val = val;
1101 
1102     for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) {
1103       Node* u = lrb->fast_out(i);
1104       if (u->Opcode() == Op_CastPP &&
1105           u->in(0) != nullptr &&
1106           phase->is_dominator(u->in(0), ctrl)) {
1107         const Type* u_t = phase->igvn().type(u);
1108 
1109         if (u_t->meet(TypePtr::NULL_PTR) != u_t &&
1110             u->in(0)->Opcode() == Op_IfTrue &&
1111             u->in(0)->as_Proj()->is_uncommon_trap_if_pattern() &&
1112             u->in(0)->in(0)->is_If() &&
1113             u->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
1114             u->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
1115             u->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
1116             u->in(0)->in(0)->in(1)->in(1)->in(1) == val &&
1117             u->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
1118           IdealLoopTree* loop = phase->get_loop(ctrl);
1119           IdealLoopTree* unc_loop = phase->get_loop(u->in(0));
1120 
1121           if (!unc_loop->is_member(loop)) {
1122             continue;
1123           }
1124 
1125           Node* branch = no_branches(ctrl, u->in(0), false, phase);
1126           assert(branch == nullptr || branch == NodeSentinel, "was not looking for a branch");
1127           if (branch == NodeSentinel) {
1128             continue;
1129           }
1130 
1131           Node* iff = u->in(0)->in(0);
1132           Node* bol = iff->in(1)->clone();
1133           Node* cmp = bol->in(1)->clone();
1134           cmp->set_req(1, lrb);
1135           bol->set_req(1, cmp);
1136           phase->igvn().replace_input_of(iff, 1, bol);
1137           phase->set_ctrl(lrb, iff->in(0));
1138           phase->register_new_node(cmp, iff->in(0));
1139           phase->register_new_node(bol, iff->in(0));
1140           break;
1141         }
1142       }
1143     }
1144     // Load barrier on the control output of a call
1145     if ((ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) || ctrl->is_CallJava()) {
1146       CallJavaNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_CallJava() : ctrl->as_CallJava();
1147       if (call->entry_point() == OptoRuntime::rethrow_stub()) {
1148         // The rethrow call may have too many projections to be
1149         // properly handled here. Given there's no reason for a
1150         // barrier to depend on the call, move it above the call
1151         stack.push(lrb, 0);
1152         do {
1153           Node* n = stack.node();
1154           uint idx = stack.index();
1155           if (idx < n->req()) {
1156             Node* in = n->in(idx);
1157             stack.set_index(idx+1);
1158             if (in != nullptr) {
1159               if (phase->has_ctrl(in)) {
1160                 if (phase->is_dominator(call, phase->get_ctrl(in))) {
1161 #ifdef ASSERT
1162                   for (uint i = 0; i < stack.size(); i++) {
1163                     assert(stack.node_at(i) != in, "node shouldn't have been seen yet");
1164                   }
1165 #endif
1166                   stack.push(in, 0);
1167                 }
1168               } else {
1169                 assert(phase->is_dominator(in, call->in(0)), "no dependency on the call");
1170               }
1171             }
1172           } else {
1173             phase->set_ctrl(n, call->in(0));
1174             stack.pop();
1175           }
1176         } while(stack.size() > 0);
1177         continue;
1178       }
1179       CallProjections* projs = call->extract_projections(false, false);
1180 
1181       // If this is a runtime call, it doesn't have an exception handling path
1182       if (projs->fallthrough_catchproj == nullptr) {
1183         assert(call->method() == nullptr, "should be runtime call");
1184         assert(projs->catchall_catchproj == nullptr, "runtime call should not have catch all projection");
1185         continue;
1186       }
1187 
1188       // Otherwise, clone the barrier so there's one for the fallthrough and one for the exception handling path
1189 #ifdef ASSERT
1190       VectorSet cloned;
1191 #endif
1192       Node* lrb_clone = lrb->clone();
1193       phase->register_new_node(lrb_clone, projs->catchall_catchproj);
1194       phase->set_ctrl(lrb, projs->fallthrough_catchproj);
1195 
1196       stack.push(lrb, 0);
1197       clones.push(lrb_clone);
1198 
1199       do {
1200         assert(stack.size() == clones.size(), "");
1201         Node* n = stack.node();
1202 #ifdef ASSERT
1203         if (n->is_Load()) {
1204           Node* mem = n->in(MemNode::Memory);
1205           for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
1206             Node* u = mem->fast_out(j);
1207             assert(!u->is_Store() || !u->is_LoadStore() || phase->get_ctrl(u) != ctrl, "anti dependent store?");
1208           }
1209         }
1210 #endif
1211         uint idx = stack.index();
1212         Node* n_clone = clones.at(clones.size()-1);
1213         if (idx < n->outcnt()) {
1214           Node* u = n->raw_out(idx);
1215           Node* c = phase->ctrl_or_self(u);
1216           if (phase->is_dominator(call, c) && phase->is_dominator(c, projs->fallthrough_proj)) {
1217             stack.set_index(idx+1);
1218             assert(!u->is_CFG(), "");
1219             stack.push(u, 0);
1220             assert(!cloned.test_set(u->_idx), "only one clone");
1221             Node* u_clone = u->clone();
1222             int nb = u_clone->replace_edge(n, n_clone, &phase->igvn());
1223             assert(nb > 0, "should have replaced some uses");
1224             phase->register_new_node(u_clone, projs->catchall_catchproj);
1225             clones.push(u_clone);
1226             phase->set_ctrl(u, projs->fallthrough_catchproj);
1227           } else {
1228             bool replaced = false;
1229             if (u->is_Phi()) {
1230               for (uint k = 1; k < u->req(); k++) {
1231                 if (u->in(k) == n) {
1232                   if (phase->is_dominator(projs->catchall_catchproj, u->in(0)->in(k))) {
1233                     phase->igvn().replace_input_of(u, k, n_clone);
1234                     replaced = true;
1235                   } else if (!phase->is_dominator(projs->fallthrough_catchproj, u->in(0)->in(k))) {
1236                     phase->igvn().replace_input_of(u, k, create_phis_on_call_return(ctrl, u->in(0)->in(k), n, n_clone, projs, phase));
1237                     replaced = true;
1238                   }
1239                 }
1240               }
1241             } else {
1242               if (phase->is_dominator(projs->catchall_catchproj, c)) {
1243                 phase->igvn().rehash_node_delayed(u);
1244                 int nb = u->replace_edge(n, n_clone, &phase->igvn());
1245                 assert(nb > 0, "should have replaced some uses");
1246                 replaced = true;
1247               } else if (!phase->is_dominator(projs->fallthrough_catchproj, c)) {
1248                 if (u->is_If()) {
1249                   // Can't break If/Bool/Cmp chain
1250                   assert(n->is_Bool(), "unexpected If shape");
1251                   assert(stack.node_at(stack.size()-2)->is_Cmp(), "unexpected If shape");
1252                   assert(n_clone->is_Bool(), "unexpected clone");
1253                   assert(clones.at(clones.size()-2)->is_Cmp(), "unexpected clone");
1254                   Node* bol_clone = n->clone();
1255                   Node* cmp_clone = stack.node_at(stack.size()-2)->clone();
1256                   bol_clone->set_req(1, cmp_clone);
1257 
1258                   Node* nn = stack.node_at(stack.size()-3);
1259                   Node* nn_clone = clones.at(clones.size()-3);
1260                   assert(nn->Opcode() == nn_clone->Opcode(), "mismatch");
1261 
1262                   int nb = cmp_clone->replace_edge(nn, create_phis_on_call_return(ctrl, c, nn, nn_clone, projs, phase),
1263                                                    &phase->igvn());
1264                   assert(nb > 0, "should have replaced some uses");
1265 
1266                   phase->register_new_node(bol_clone, u->in(0));
1267                   phase->register_new_node(cmp_clone, u->in(0));
1268 
1269                   phase->igvn().replace_input_of(u, 1, bol_clone);
1270 
1271                 } else {
1272                   phase->igvn().rehash_node_delayed(u);
1273                   int nb = u->replace_edge(n, create_phis_on_call_return(ctrl, c, n, n_clone, projs, phase), &phase->igvn());
1274                   assert(nb > 0, "should have replaced some uses");
1275                 }
1276                 replaced = true;
1277               }
1278             }
1279             if (!replaced) {
1280               stack.set_index(idx+1);
1281             }
1282           }
1283         } else {
1284           stack.pop();
1285           clones.pop();
1286         }
1287       } while (stack.size() > 0);
1288       assert(stack.size() == 0 && clones.size() == 0, "");
1289     }
1290   }
1291 
1292   for (int i = 0; i < state->load_reference_barriers_count(); i++) {
1293     ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1294     Node* ctrl = phase->get_ctrl(lrb);
1295     IdealLoopTree* loop = phase->get_loop(ctrl);
1296     Node* head = loop->head();
1297     if (head->is_OuterStripMinedLoop()) {
1298       // Expanding a barrier here will break loop strip mining
1299       // verification. Transform the loop so the loop nest doesn't
1300       // appear as strip mined.
1301       OuterStripMinedLoopNode* outer = head->as_OuterStripMinedLoop();
1302       hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1303     }
1304     if (head->is_BaseCountedLoop() && ctrl->is_IfProj() && ctrl->in(0)->is_BaseCountedLoopEnd() &&
1305         head->as_BaseCountedLoop()->loopexit() == ctrl->in(0)) {
1306       Node* entry = head->in(LoopNode::EntryControl);
1307       Node* backedge = head->in(LoopNode::LoopBackControl);
1308       Node* new_head = new LoopNode(entry, backedge);
1309       phase->register_control(new_head, phase->get_loop(entry), entry);
1310       phase->lazy_replace(head, new_head);
1311     }
1312   }
1313 
1314   // Expand load-reference-barriers
1315   MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
1316   Unique_Node_List uses_to_ignore;
1317   for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1318     ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1319     uint last = phase->C->unique();
1320     Node* ctrl = phase->get_ctrl(lrb);
1321     Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1322 
1323     Node* orig_ctrl = ctrl;
1324 
1325     Node* raw_mem = fixer.find_mem(ctrl, lrb);
1326     Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, nullptr);
1327 
1328     IdealLoopTree *loop = phase->get_loop(ctrl);
1329 
1330     Node* heap_stable_ctrl = nullptr;
1331     Node* null_ctrl = nullptr;
1332 
1333     assert(val->bottom_type()->make_oopptr(), "need oop");
1334     assert(val->bottom_type()->make_oopptr()->const_oop() == nullptr, "expect non-constant");
1335 
1336     enum { _heap_stable = 1, _evac_path, _not_cset, PATH_LIMIT };
1337     Node* region = new RegionNode(PATH_LIMIT);
1338     Node* val_phi = new PhiNode(region, val->bottom_type()->is_oopptr());
1339 
1340     // Stable path.
1341     int flags = ShenandoahHeap::HAS_FORWARDED;
1342     if (!ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1343       flags |= ShenandoahHeap::WEAK_ROOTS;
1344     }
1345     test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, flags);
1346     IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
1347 
1348     // Heap stable case
1349     region->init_req(_heap_stable, heap_stable_ctrl);
1350     val_phi->init_req(_heap_stable, val);
1351 
1352     // Test for in-cset, unless it's a native-LRB. Native LRBs need to return null
1353     // even for non-cset objects to prevent resurrection of such objects.
1354     // Wires !in_cset(obj) to slot 2 of region and phis
1355     Node* not_cset_ctrl = nullptr;
1356     if (ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1357       test_in_cset(ctrl, not_cset_ctrl, val, raw_mem, phase);
1358     }
1359     if (not_cset_ctrl != nullptr) {
1360       region->init_req(_not_cset, not_cset_ctrl);
1361       val_phi->init_req(_not_cset, val);
1362     } else {
1363       region->del_req(_not_cset);
1364       val_phi->del_req(_not_cset);
1365     }
1366 
1367     // Resolve object when orig-value is in cset.
1368     // Make the unconditional resolve for fwdptr.
1369 
1370     // Call lrb-stub and wire up that path in slots 4
1371     Node* result_mem = nullptr;
1372 
1373     Node* addr;
1374     {
1375       VectorSet visited;
1376       addr = get_load_addr(phase, visited, lrb);
1377     }
1378     if (addr->Opcode() == Op_AddP) {
1379       Node* orig_base = addr->in(AddPNode::Base);
1380       Node* base = new CheckCastPPNode(ctrl, orig_base, orig_base->bottom_type(), ConstraintCastNode::StrongDependency);
1381       phase->register_new_node(base, ctrl);
1382       if (addr->in(AddPNode::Base) == addr->in((AddPNode::Address))) {
1383         // Field access
1384         addr = addr->clone();
1385         addr->set_req(AddPNode::Base, base);
1386         addr->set_req(AddPNode::Address, base);
1387         phase->register_new_node(addr, ctrl);
1388       } else {
1389         Node* addr2 = addr->in(AddPNode::Address);
1390         if (addr2->Opcode() == Op_AddP && addr2->in(AddPNode::Base) == addr2->in(AddPNode::Address) &&
1391               addr2->in(AddPNode::Base) == orig_base) {
1392           addr2 = addr2->clone();
1393           addr2->set_req(AddPNode::Base, base);
1394           addr2->set_req(AddPNode::Address, base);
1395           phase->register_new_node(addr2, ctrl);
1396           addr = addr->clone();
1397           addr->set_req(AddPNode::Base, base);
1398           addr->set_req(AddPNode::Address, addr2);
1399           phase->register_new_node(addr, ctrl);
1400         }
1401       }
1402     }
1403     call_lrb_stub(ctrl, val, addr, lrb->decorators(), phase);
1404     region->init_req(_evac_path, ctrl);
1405     val_phi->init_req(_evac_path, val);
1406 
1407     phase->register_control(region, loop, heap_stable_iff);
1408     Node* out_val = val_phi;
1409     phase->register_new_node(val_phi, region);
1410 
1411     fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase);
1412 
1413     ctrl = orig_ctrl;
1414 
1415     phase->igvn().replace_node(lrb, out_val);
1416 
1417     follow_barrier_uses(out_val, ctrl, uses, phase);
1418 
1419     for(uint next = 0; next < uses.size(); next++ ) {
1420       Node *n = uses.at(next);
1421       assert(phase->get_ctrl(n) == ctrl, "bad control");
1422       assert(n != raw_mem, "should leave input raw mem above the barrier");
1423       phase->set_ctrl(n, region);
1424       follow_barrier_uses(n, ctrl, uses, phase);
1425     }
1426     fixer.record_new_ctrl(ctrl, region, raw_mem, raw_mem_for_ctrl);
1427   }
1428   // Done expanding load-reference-barriers.
1429   assert(ShenandoahBarrierSetC2::bsc2()->state()->load_reference_barriers_count() == 0, "all load reference barrier nodes should have been replaced");
1430 }
1431 
1432 Node* ShenandoahBarrierC2Support::get_load_addr(PhaseIdealLoop* phase, VectorSet& visited, Node* in) {
1433   if (visited.test_set(in->_idx)) {
1434     return nullptr;
1435   }
1436   switch (in->Opcode()) {
1437     case Op_Proj:
1438       return get_load_addr(phase, visited, in->in(0));
1439     case Op_CastPP:
1440     case Op_CheckCastPP:
1441     case Op_DecodeN:
1442     case Op_EncodeP:
1443       return get_load_addr(phase, visited, in->in(1));
1444     case Op_LoadN:
1445     case Op_LoadP:
1446       return in->in(MemNode::Address);
1447     case Op_CompareAndExchangeN:
1448     case Op_CompareAndExchangeP:
1449     case Op_GetAndSetN:
1450     case Op_GetAndSetP:
1451     case Op_ShenandoahCompareAndExchangeP:
1452     case Op_ShenandoahCompareAndExchangeN:
1453       // Those instructions would just have stored a different
1454       // value into the field. No use to attempt to fix it at this point.
1455       return phase->igvn().zerocon(T_OBJECT);
1456     case Op_CMoveP:
1457     case Op_CMoveN: {
1458       Node* t = get_load_addr(phase, visited, in->in(CMoveNode::IfTrue));
1459       Node* f = get_load_addr(phase, visited, in->in(CMoveNode::IfFalse));
1460       // Handle unambiguous cases: single address reported on both branches.
1461       if (t != nullptr && f == nullptr) return t;
1462       if (t == nullptr && f != nullptr) return f;
1463       if (t != nullptr && t == f)    return t;
1464       // Ambiguity.
1465       return phase->igvn().zerocon(T_OBJECT);
1466     }
1467     case Op_Phi: {
1468       Node* addr = nullptr;
1469       for (uint i = 1; i < in->req(); i++) {
1470         Node* addr1 = get_load_addr(phase, visited, in->in(i));
1471         if (addr == nullptr) {
1472           addr = addr1;
1473         }
1474         if (addr != addr1) {
1475           return phase->igvn().zerocon(T_OBJECT);
1476         }
1477       }
1478       return addr;
1479     }
1480     case Op_ShenandoahLoadReferenceBarrier:
1481       return get_load_addr(phase, visited, in->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1482     case Op_CallDynamicJava:
1483     case Op_CallLeaf:
1484     case Op_CallStaticJava:
1485     case Op_ConN:
1486     case Op_ConP:
1487     case Op_Parm:
1488     case Op_CreateEx:
1489       return phase->igvn().zerocon(T_OBJECT);
1490     default:
1491 #ifdef ASSERT
1492       fatal("Unknown node in get_load_addr: %s", NodeClassNames[in->Opcode()]);
1493 #endif
1494       return phase->igvn().zerocon(T_OBJECT);
1495   }
1496 
1497 }
1498 
1499 #ifdef ASSERT
1500 static bool has_never_branch(Node* root) {
1501   for (uint i = 1; i < root->req(); i++) {
1502     Node* in = root->in(i);
1503     if (in != nullptr && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->is_NeverBranch()) {
1504       return true;
1505     }
1506   }
1507   return false;
1508 }
1509 #endif
1510 
1511 void MemoryGraphFixer::collect_memory_nodes() {
1512   Node_Stack stack(0);
1513   VectorSet visited;
1514   Node_List regions;
1515 
1516   // Walk the raw memory graph and create a mapping from CFG node to
1517   // memory node. Exclude phis for now.
1518   stack.push(_phase->C->root(), 1);
1519   do {
1520     Node* n = stack.node();
1521     int opc = n->Opcode();
1522     uint i = stack.index();
1523     if (i < n->req()) {
1524       Node* mem = nullptr;
1525       if (opc == Op_Root) {
1526         Node* in = n->in(i);
1527         int in_opc = in->Opcode();
1528         if (in_opc == Op_Return || in_opc == Op_Rethrow) {
1529           mem = in->in(TypeFunc::Memory);
1530         } else if (in_opc == Op_Halt) {
1531           if (in->in(0)->is_Region()) {
1532             Node* r = in->in(0);
1533             for (uint j = 1; j < r->req(); j++) {
1534               assert(!r->in(j)->is_NeverBranch(), "");
1535             }
1536           } else {
1537             Node* proj = in->in(0);
1538             assert(proj->is_Proj(), "");
1539             Node* in = proj->in(0);
1540             assert(in->is_CallStaticJava() || in->is_NeverBranch() || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
1541             if (in->is_CallStaticJava()) {
1542               mem = in->in(TypeFunc::Memory);
1543             } else if (in->Opcode() == Op_Catch) {
1544               Node* call = in->in(0)->in(0);
1545               assert(call->is_Call(), "");
1546               mem = call->in(TypeFunc::Memory);
1547             } else if (in->is_NeverBranch()) {
1548               mem = collect_memory_for_infinite_loop(in);
1549             }
1550           }
1551         } else {
1552 #ifdef ASSERT
1553           n->dump();
1554           in->dump();
1555 #endif
1556           ShouldNotReachHere();
1557         }
1558       } else {
1559         assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
1560         assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
1561         mem = n->in(i);
1562       }
1563       i++;
1564       stack.set_index(i);
1565       if (mem == nullptr) {
1566         continue;
1567       }
1568       for (;;) {
1569         if (visited.test_set(mem->_idx) || mem->is_Start()) {
1570           break;
1571         }
1572         if (mem->is_Phi()) {
1573           stack.push(mem, 2);
1574           mem = mem->in(1);
1575         } else if (mem->is_Proj()) {
1576           stack.push(mem, mem->req());
1577           mem = mem->in(0);
1578         } else if (mem->is_SafePoint() || mem->is_MemBar()) {
1579           mem = mem->in(TypeFunc::Memory);
1580         } else if (mem->is_MergeMem()) {
1581           MergeMemNode* mm = mem->as_MergeMem();
1582           mem = mm->memory_at(_alias);
1583         } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
1584           assert(_alias == Compile::AliasIdxRaw, "");
1585           stack.push(mem, mem->req());
1586           mem = mem->in(MemNode::Memory);
1587         } else {
1588 #ifdef ASSERT
1589           mem->dump();
1590 #endif
1591           ShouldNotReachHere();
1592         }
1593       }
1594     } else {
1595       if (n->is_Phi()) {
1596         // Nothing
1597       } else if (!n->is_Root()) {
1598         Node* c = get_ctrl(n);
1599         _memory_nodes.map(c->_idx, n);
1600       }
1601       stack.pop();
1602     }
1603   } while(stack.is_nonempty());
1604 
1605   // Iterate over CFG nodes in rpo and propagate memory state to
1606   // compute memory state at regions, creating new phis if needed.
1607   Node_List rpo_list;
1608   visited.clear();
1609   _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
1610   Node* root = rpo_list.pop();
1611   assert(root == _phase->C->root(), "");
1612 
1613   const bool trace = false;
1614 #ifdef ASSERT
1615   if (trace) {
1616     for (int i = rpo_list.size() - 1; i >= 0; i--) {
1617       Node* c = rpo_list.at(i);
1618       if (_memory_nodes[c->_idx] != nullptr) {
1619         tty->print("X %d", c->_idx);  _memory_nodes[c->_idx]->dump();
1620       }
1621     }
1622   }
1623 #endif
1624   uint last = _phase->C->unique();
1625 
1626 #ifdef ASSERT
1627   uint16_t max_depth = 0;
1628   for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
1629     IdealLoopTree* lpt = iter.current();
1630     max_depth = MAX2(max_depth, lpt->_nest);
1631   }
1632 #endif
1633 
1634   bool progress = true;
1635   int iteration = 0;
1636   Node_List dead_phis;
1637   while (progress) {
1638     progress = false;
1639     iteration++;
1640     assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
1641     if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
1642 
1643     for (int i = rpo_list.size() - 1; i >= 0; i--) {
1644       Node* c = rpo_list.at(i);
1645 
1646       Node* prev_mem = _memory_nodes[c->_idx];
1647       if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
1648         Node* prev_region = regions[c->_idx];
1649         Node* unique = nullptr;
1650         for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
1651           Node* m = _memory_nodes[c->in(j)->_idx];
1652           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");
1653           if (m != nullptr) {
1654             if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
1655               assert((c->is_Loop() && j == LoopNode::LoopBackControl) || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
1656               // continue
1657             } else if (unique == nullptr) {
1658               unique = m;
1659             } else if (m == unique) {
1660               // continue
1661             } else {
1662               unique = NodeSentinel;
1663             }
1664           }
1665         }
1666         assert(unique != nullptr, "empty phi???");
1667         if (unique != NodeSentinel) {
1668           if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c) {
1669             dead_phis.push(prev_region);
1670           }
1671           regions.map(c->_idx, unique);
1672         } else {
1673           Node* phi = nullptr;
1674           if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
1675             phi = prev_region;
1676             for (uint k = 1; k < c->req(); k++) {
1677               Node* m = _memory_nodes[c->in(k)->_idx];
1678               assert(m != nullptr, "expect memory state");
1679               phi->set_req(k, m);
1680             }
1681           } else {
1682             for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == nullptr; j++) {
1683               Node* u = c->fast_out(j);
1684               if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
1685                   (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
1686                 phi = u;
1687                 for (uint k = 1; k < c->req() && phi != nullptr; k++) {
1688                   Node* m = _memory_nodes[c->in(k)->_idx];
1689                   assert(m != nullptr, "expect memory state");
1690                   if (u->in(k) != m) {
1691                     phi = NodeSentinel;
1692                   }
1693                 }
1694               }
1695             }
1696             if (phi == NodeSentinel) {
1697               phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
1698               for (uint k = 1; k < c->req(); k++) {
1699                 Node* m = _memory_nodes[c->in(k)->_idx];
1700                 assert(m != nullptr, "expect memory state");
1701                 phi->init_req(k, m);
1702               }
1703             }
1704           }
1705           if (phi != nullptr) {
1706             regions.map(c->_idx, phi);
1707           } else {
1708             assert(c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
1709           }
1710         }
1711         Node* current_region = regions[c->_idx];
1712         if (current_region != prev_region) {
1713           progress = true;
1714           if (prev_region == prev_mem) {
1715             _memory_nodes.map(c->_idx, current_region);
1716           }
1717         }
1718       } else if (prev_mem == nullptr || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
1719         Node* m = _memory_nodes[_phase->idom(c)->_idx];
1720         assert(m != nullptr || c->Opcode() == Op_Halt, "expect memory state");
1721         if (m != prev_mem) {
1722           _memory_nodes.map(c->_idx, m);
1723           progress = true;
1724         }
1725       }
1726 #ifdef ASSERT
1727       if (trace) { tty->print("X %d", c->_idx);  _memory_nodes[c->_idx]->dump(); }
1728 #endif
1729     }
1730   }
1731 
1732   // Replace existing phi with computed memory state for that region
1733   // if different (could be a new phi or a dominating memory node if
1734   // that phi was found to be useless).
1735   while (dead_phis.size() > 0) {
1736     Node* n = dead_phis.pop();
1737     n->replace_by(_phase->C->top());
1738     n->destruct(&_phase->igvn());
1739   }
1740   for (int i = rpo_list.size() - 1; i >= 0; i--) {
1741     Node* c = rpo_list.at(i);
1742     if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
1743       Node* n = regions[c->_idx];
1744       assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
1745       if (n != nullptr && n->is_Phi() && n->_idx >= last && n->in(0) == c) {
1746         _phase->register_new_node(n, c);
1747       }
1748     }
1749   }
1750   for (int i = rpo_list.size() - 1; i >= 0; i--) {
1751     Node* c = rpo_list.at(i);
1752     if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
1753       Node* n = regions[c->_idx];
1754       assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
1755       for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
1756         Node* u = c->fast_out(i);
1757         if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
1758             u != n) {
1759           assert(c->unique_ctrl_out()->Opcode() != Op_Halt, "expected memory state");
1760           if (u->adr_type() == TypePtr::BOTTOM) {
1761             fix_memory_uses(u, n, n, c);
1762           } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
1763             _phase->lazy_replace(u, n);
1764             --i; --imax;
1765           }
1766         }
1767       }
1768     }
1769   }
1770 }
1771 
1772 Node* MemoryGraphFixer::collect_memory_for_infinite_loop(const Node* in) {
1773   Node* mem = nullptr;
1774   Node* head = in->in(0);
1775   assert(head->is_Region(), "unexpected infinite loop graph shape");
1776 
1777   Node* phi_mem = nullptr;
1778   for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
1779     Node* u = head->fast_out(j);
1780     if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
1781       if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
1782         assert(phi_mem == nullptr || phi_mem->adr_type() == TypePtr::BOTTOM, "");
1783         phi_mem = u;
1784       } else if (u->adr_type() == TypePtr::BOTTOM) {
1785         assert(phi_mem == nullptr || _phase->C->get_alias_index(phi_mem->adr_type()) == _alias, "");
1786         if (phi_mem == nullptr) {
1787           phi_mem = u;
1788         }
1789       }
1790     }
1791   }
1792   if (phi_mem == nullptr) {
1793     ResourceMark rm;
1794     Node_Stack stack(0);
1795     stack.push(head, 1);
1796     do {
1797       Node* n = stack.node();
1798       uint i = stack.index();
1799       if (i >= n->req()) {
1800         stack.pop();
1801       } else {
1802         stack.set_index(i + 1);
1803         Node* c = n->in(i);
1804         assert(c != head, "should have found a safepoint on the way");
1805         if (stack.size() != 1 || _phase->is_dominator(head, c)) {
1806           for (;;) {
1807             if (c->is_Region()) {
1808               stack.push(c, 1);
1809               break;
1810             } else if (c->is_SafePoint() && !c->is_CallLeaf()) {
1811               Node* m = c->in(TypeFunc::Memory);
1812               if (m->is_MergeMem()) {
1813                 m = m->as_MergeMem()->memory_at(_alias);
1814               }
1815               assert(mem == nullptr || mem == m, "several memory states");
1816               mem = m;
1817               break;
1818             } else {
1819               assert(c != c->in(0), "");
1820               c = c->in(0);
1821             }
1822           }
1823         }
1824       }
1825     } while (stack.size() > 0);
1826     assert(mem != nullptr, "should have found safepoint");
1827   } else {
1828     mem = phi_mem;
1829   }
1830   return mem;
1831 }
1832 
1833 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
1834   Node* c = _phase->get_ctrl(n);
1835   if (n->is_Proj() && n->in(0) != nullptr && n->in(0)->is_Call()) {
1836     assert(c == n->in(0), "");
1837     CallNode* call = c->as_Call();
1838     CallProjections* projs = call->extract_projections(true, false);
1839     if (projs->catchall_memproj != nullptr) {
1840       if (projs->fallthrough_memproj == n) {
1841         c = projs->fallthrough_catchproj;
1842       } else {
1843         assert(projs->catchall_memproj == n, "");
1844         c = projs->catchall_catchproj;
1845       }
1846     }
1847   }
1848   return c;
1849 }
1850 
1851 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
1852   if (_phase->has_ctrl(n))
1853     return get_ctrl(n);
1854   else {
1855     assert (n->is_CFG(), "must be a CFG node");
1856     return n;
1857   }
1858 }
1859 
1860 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
1861   return m != nullptr && get_ctrl(m) == c;
1862 }
1863 
1864 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
1865   assert(n == nullptr || _phase->ctrl_or_self(n) == ctrl, "");
1866   assert(!ctrl->is_Call() || ctrl == n, "projection expected");
1867 #ifdef ASSERT
1868   if ((ctrl->is_Proj() && ctrl->in(0)->is_Call()) ||
1869       (ctrl->is_Catch() && ctrl->in(0)->in(0)->is_Call())) {
1870     CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_Call() : ctrl->in(0)->in(0)->as_Call();
1871     int mems = 0;
1872     for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
1873       Node* u = call->fast_out(i);
1874       if (u->bottom_type() == Type::MEMORY) {
1875         mems++;
1876       }
1877     }
1878     assert(mems <= 1, "No node right after call if multiple mem projections");
1879   }
1880 #endif
1881   Node* mem = _memory_nodes[ctrl->_idx];
1882   Node* c = ctrl;
1883   while (!mem_is_valid(mem, c) &&
1884          (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
1885     c = _phase->idom(c);
1886     mem = _memory_nodes[c->_idx];
1887   }
1888   if (n != nullptr && mem_is_valid(mem, c)) {
1889     while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
1890       mem = next_mem(mem, _alias);
1891     }
1892     if (mem->is_MergeMem()) {
1893       mem = mem->as_MergeMem()->memory_at(_alias);
1894     }
1895     if (!mem_is_valid(mem, c)) {
1896       do {
1897         c = _phase->idom(c);
1898         mem = _memory_nodes[c->_idx];
1899       } while (!mem_is_valid(mem, c) &&
1900                (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
1901     }
1902   }
1903   assert(mem->bottom_type() == Type::MEMORY, "");
1904   return mem;
1905 }
1906 
1907 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
1908   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1909     Node* use = region->fast_out(i);
1910     if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
1911         (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
1912       return true;
1913     }
1914   }
1915   return false;
1916 }
1917 
1918 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
1919   assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
1920   const bool trace = false;
1921   DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
1922   DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
1923   GrowableArray<Node*> phis;
1924   if (mem_for_ctrl != mem) {
1925     Node* old = mem_for_ctrl;
1926     Node* prev = nullptr;
1927     while (old != mem) {
1928       prev = old;
1929       if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
1930         assert(_alias == Compile::AliasIdxRaw, "");
1931         old = old->in(MemNode::Memory);
1932       } else if (old->Opcode() == Op_SCMemProj) {
1933         assert(_alias == Compile::AliasIdxRaw, "");
1934         old = old->in(0);
1935       } else {
1936         ShouldNotReachHere();
1937       }
1938     }
1939     assert(prev != nullptr, "");
1940     if (new_ctrl != ctrl) {
1941       _memory_nodes.map(ctrl->_idx, mem);
1942       _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
1943     }
1944     uint input = (uint)MemNode::Memory;
1945     _phase->igvn().replace_input_of(prev, input, new_mem);
1946   } else {
1947     uses.clear();
1948     _memory_nodes.map(new_ctrl->_idx, new_mem);
1949     uses.push(new_ctrl);
1950     for(uint next = 0; next < uses.size(); next++ ) {
1951       Node *n = uses.at(next);
1952       assert(n->is_CFG(), "");
1953       DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
1954       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1955         Node* u = n->fast_out(i);
1956         if (!u->is_Root() && u->is_CFG() && u != n) {
1957           Node* m = _memory_nodes[u->_idx];
1958           if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) &&
1959               !has_mem_phi(u) &&
1960               u->unique_ctrl_out()->Opcode() != Op_Halt) {
1961             DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
1962             DEBUG_ONLY(if (trace && m != nullptr) { tty->print("ZZZ mem"); m->dump(); });
1963 
1964             if (!mem_is_valid(m, u) || !m->is_Phi()) {
1965               bool push = true;
1966               bool create_phi = true;
1967               if (_phase->is_dominator(new_ctrl, u)) {
1968                 create_phi = false;
1969               }
1970               if (create_phi) {
1971                 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
1972                 _phase->register_new_node(phi, u);
1973                 phis.push(phi);
1974                 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
1975                 if (!mem_is_valid(m, u)) {
1976                   DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
1977                   _memory_nodes.map(u->_idx, phi);
1978                 } else {
1979                   DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
1980                   for (;;) {
1981                     assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), "");
1982                     Node* next = nullptr;
1983                     if (m->is_Proj()) {
1984                       next = m->in(0);
1985                     } else {
1986                       assert(m->is_Mem() || m->is_LoadStore(), "");
1987                       assert(_alias == Compile::AliasIdxRaw, "");
1988                       next = m->in(MemNode::Memory);
1989                     }
1990                     if (_phase->get_ctrl(next) != u) {
1991                       break;
1992                     }
1993                     if (next->is_MergeMem()) {
1994                       assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
1995                       break;
1996                     }
1997                     if (next->is_Phi()) {
1998                       assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
1999                       break;
2000                     }
2001                     m = next;
2002                   }
2003 
2004                   DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
2005                   assert(m->is_Mem() || m->is_LoadStore(), "");
2006                   uint input = (uint)MemNode::Memory;
2007                   _phase->igvn().replace_input_of(m, input, phi);
2008                   push = false;
2009                 }
2010               } else {
2011                 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
2012               }
2013               if (push) {
2014                 uses.push(u);
2015               }
2016             }
2017           } else if (!mem_is_valid(m, u) &&
2018                      !(u->Opcode() == Op_CProj && u->in(0)->is_NeverBranch() && u->as_Proj()->_con == 1)) {
2019             uses.push(u);
2020           }
2021         }
2022       }
2023     }
2024     for (int i = 0; i < phis.length(); i++) {
2025       Node* n = phis.at(i);
2026       Node* r = n->in(0);
2027       DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
2028       for (uint j = 1; j < n->req(); j++) {
2029         Node* m = find_mem(r->in(j), nullptr);
2030         _phase->igvn().replace_input_of(n, j, m);
2031         DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
2032       }
2033     }
2034   }
2035   uint last = _phase->C->unique();
2036   MergeMemNode* mm = nullptr;
2037   int alias = _alias;
2038   DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
2039   // Process loads first to not miss an anti-dependency: if the memory
2040   // edge of a store is updated before a load is processed then an
2041   // anti-dependency may be missed.
2042   for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2043     Node* u = mem->out(i);
2044     if (u->_idx < last && u->is_Load() && _phase->C->get_alias_index(u->adr_type()) == alias) {
2045       Node* m = find_mem(_phase->get_ctrl(u), u);
2046       if (m != mem) {
2047         DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2048         _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2049         --i;
2050       }
2051     }
2052   }
2053   for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2054     Node* u = mem->out(i);
2055     if (u->_idx < last) {
2056       if (u->is_Mem()) {
2057         if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2058           Node* m = find_mem(_phase->get_ctrl(u), u);
2059           if (m != mem) {
2060             DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2061             _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2062             --i;
2063           }
2064         }
2065       } else if (u->is_MergeMem()) {
2066         MergeMemNode* u_mm = u->as_MergeMem();
2067         if (u_mm->memory_at(alias) == mem) {
2068           MergeMemNode* newmm = nullptr;
2069           for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2070             Node* uu = u->fast_out(j);
2071             assert(!uu->is_MergeMem(), "chain of MergeMems?");
2072             if (uu->is_Phi()) {
2073               assert(uu->adr_type() == TypePtr::BOTTOM, "");
2074               Node* region = uu->in(0);
2075               int nb = 0;
2076               for (uint k = 1; k < uu->req(); k++) {
2077                 if (uu->in(k) == u) {
2078                   Node* m = find_mem(region->in(k), nullptr);
2079                   if (m != mem) {
2080                     DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
2081                     newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2082                     if (newmm != u) {
2083                       _phase->igvn().replace_input_of(uu, k, newmm);
2084                       nb++;
2085                       --jmax;
2086                     }
2087                   }
2088                 }
2089               }
2090               if (nb > 0) {
2091                 --j;
2092               }
2093             } else {
2094               Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
2095               if (m != mem) {
2096                 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
2097                 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2098                 if (newmm != u) {
2099                   _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2100                   --j, --jmax;
2101                 }
2102               }
2103             }
2104           }
2105         }
2106       } else if (u->is_Phi()) {
2107         assert(u->bottom_type() == Type::MEMORY, "what else?");
2108         if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
2109           Node* region = u->in(0);
2110           bool replaced = false;
2111           for (uint j = 1; j < u->req(); j++) {
2112             if (u->in(j) == mem) {
2113               Node* m = find_mem(region->in(j), nullptr);
2114               Node* nnew = m;
2115               if (m != mem) {
2116                 if (u->adr_type() == TypePtr::BOTTOM) {
2117                   mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
2118                   nnew = mm;
2119                 }
2120                 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
2121                 _phase->igvn().replace_input_of(u, j, nnew);
2122                 replaced = true;
2123               }
2124             }
2125           }
2126           if (replaced) {
2127             --i;
2128           }
2129         }
2130       } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2131                  u->adr_type() == nullptr) {
2132         assert(u->adr_type() != nullptr ||
2133                u->Opcode() == Op_Rethrow ||
2134                u->Opcode() == Op_Return ||
2135                u->Opcode() == Op_SafePoint ||
2136                (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2137                (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2138                u->Opcode() == Op_CallLeaf, "");
2139         Node* m = find_mem(_phase->ctrl_or_self(u), u);
2140         if (m != mem) {
2141           mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
2142           _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2143           --i;
2144         }
2145       } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2146         Node* m = find_mem(_phase->ctrl_or_self(u), u);
2147         if (m != mem) {
2148           DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2149           _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2150           --i;
2151         }
2152       } else if (u->adr_type() != TypePtr::BOTTOM &&
2153                  _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
2154         Node* m = find_mem(_phase->ctrl_or_self(u), u);
2155         assert(m != mem, "");
2156         // u is on the wrong slice...
2157         assert(u->is_ClearArray(), "");
2158         DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2159         _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2160         --i;
2161       }
2162     }
2163   }
2164 #ifdef ASSERT
2165   assert(new_mem->outcnt() > 0, "");
2166   for (int i = 0; i < phis.length(); i++) {
2167     Node* n = phis.at(i);
2168     assert(n->outcnt() > 0, "new phi must have uses now");
2169   }
2170 #endif
2171 }
2172 
2173 void MemoryGraphFixer::record_new_ctrl(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl) {
2174   if (mem_for_ctrl != mem && new_ctrl != ctrl) {
2175     _memory_nodes.map(ctrl->_idx, mem);
2176     _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2177   }
2178 }
2179 
2180 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
2181   MergeMemNode* mm = MergeMemNode::make(mem);
2182   mm->set_memory_at(_alias, rep_proj);
2183   _phase->register_new_node(mm, rep_ctrl);
2184   return mm;
2185 }
2186 
2187 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
2188   MergeMemNode* newmm = nullptr;
2189   MergeMemNode* u_mm = u->as_MergeMem();
2190   Node* c = _phase->get_ctrl(u);
2191   if (_phase->is_dominator(c, rep_ctrl)) {
2192     c = rep_ctrl;
2193   } else {
2194     assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
2195   }
2196   if (u->outcnt() == 1) {
2197     if (u->req() > (uint)_alias && u->in(_alias) == mem) {
2198       _phase->igvn().replace_input_of(u, _alias, rep_proj);
2199       --i;
2200     } else {
2201       _phase->igvn().rehash_node_delayed(u);
2202       u_mm->set_memory_at(_alias, rep_proj);
2203     }
2204     newmm = u_mm;
2205     _phase->set_ctrl_and_loop(u, c);
2206   } else {
2207     // can't simply clone u and then change one of its input because
2208     // it adds and then removes an edge which messes with the
2209     // DUIterator
2210     newmm = MergeMemNode::make(u_mm->base_memory());
2211     for (uint j = 0; j < u->req(); j++) {
2212       if (j < newmm->req()) {
2213         if (j == (uint)_alias) {
2214           newmm->set_req(j, rep_proj);
2215         } else if (newmm->in(j) != u->in(j)) {
2216           newmm->set_req(j, u->in(j));
2217         }
2218       } else if (j == (uint)_alias) {
2219         newmm->add_req(rep_proj);
2220       } else {
2221         newmm->add_req(u->in(j));
2222       }
2223     }
2224     if ((uint)_alias >= u->req()) {
2225       newmm->set_memory_at(_alias, rep_proj);
2226     }
2227     _phase->register_new_node(newmm, c);
2228   }
2229   return newmm;
2230 }
2231 
2232 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
2233   if (phi->adr_type() == TypePtr::BOTTOM) {
2234     Node* region = phi->in(0);
2235     for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
2236       Node* uu = region->fast_out(j);
2237       if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
2238         return false;
2239       }
2240     }
2241     return true;
2242   }
2243   return _phase->C->get_alias_index(phi->adr_type()) == _alias;
2244 }
2245 
2246 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
2247   uint last = _phase-> C->unique();
2248   MergeMemNode* mm = nullptr;
2249   assert(mem->bottom_type() == Type::MEMORY, "");
2250   for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2251     Node* u = mem->out(i);
2252     if (u != replacement && u->_idx < last) {
2253       if (u->is_MergeMem()) {
2254         MergeMemNode* u_mm = u->as_MergeMem();
2255         if (u_mm->memory_at(_alias) == mem) {
2256           MergeMemNode* newmm = nullptr;
2257           for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2258             Node* uu = u->fast_out(j);
2259             assert(!uu->is_MergeMem(), "chain of MergeMems?");
2260             if (uu->is_Phi()) {
2261               if (should_process_phi(uu)) {
2262                 Node* region = uu->in(0);
2263                 int nb = 0;
2264                 for (uint k = 1; k < uu->req(); k++) {
2265                   if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
2266                     if (newmm == nullptr) {
2267                       newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2268                     }
2269                     if (newmm != u) {
2270                       _phase->igvn().replace_input_of(uu, k, newmm);
2271                       nb++;
2272                       --jmax;
2273                     }
2274                   }
2275                 }
2276                 if (nb > 0) {
2277                   --j;
2278                 }
2279               }
2280             } else {
2281               if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
2282                 if (newmm == nullptr) {
2283                   newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2284                 }
2285                 if (newmm != u) {
2286                   _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2287                   --j, --jmax;
2288                 }
2289               }
2290             }
2291           }
2292         }
2293       } else if (u->is_Phi()) {
2294         assert(u->bottom_type() == Type::MEMORY, "what else?");
2295         Node* region = u->in(0);
2296         if (should_process_phi(u)) {
2297           bool replaced = false;
2298           for (uint j = 1; j < u->req(); j++) {
2299             if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
2300               Node* nnew = rep_proj;
2301               if (u->adr_type() == TypePtr::BOTTOM) {
2302                 if (mm == nullptr) {
2303                   mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2304                 }
2305                 nnew = mm;
2306               }
2307               _phase->igvn().replace_input_of(u, j, nnew);
2308               replaced = true;
2309             }
2310           }
2311           if (replaced) {
2312             --i;
2313           }
2314 
2315         }
2316       } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2317                  u->adr_type() == nullptr) {
2318         assert(u->adr_type() != nullptr ||
2319                u->Opcode() == Op_Rethrow ||
2320                u->Opcode() == Op_Return ||
2321                u->Opcode() == Op_SafePoint ||
2322                (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2323                (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2324                u->Opcode() == Op_CallLeaf, "%s", u->Name());
2325         if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2326           if (mm == nullptr) {
2327             mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2328           }
2329           _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2330           --i;
2331         }
2332       } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2333         if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2334           _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
2335           --i;
2336         }
2337       }
2338     }
2339   }
2340 }
2341 
2342 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj, DecoratorSet decorators)
2343 : Node(ctrl, obj), _decorators(decorators) {
2344   ShenandoahBarrierSetC2::bsc2()->state()->add_load_reference_barrier(this);
2345 }
2346 
2347 DecoratorSet ShenandoahLoadReferenceBarrierNode::decorators() const {
2348   return _decorators;
2349 }
2350 
2351 uint ShenandoahLoadReferenceBarrierNode::size_of() const {
2352   return sizeof(*this);
2353 }
2354 
2355 static DecoratorSet mask_decorators(DecoratorSet decorators) {
2356   return decorators & (ON_STRONG_OOP_REF | ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF | ON_UNKNOWN_OOP_REF | IN_NATIVE);
2357 }
2358 
2359 uint ShenandoahLoadReferenceBarrierNode::hash() const {
2360   uint hash = Node::hash();
2361   hash += mask_decorators(_decorators);
2362   return hash;
2363 }
2364 
2365 bool ShenandoahLoadReferenceBarrierNode::cmp( const Node &n ) const {
2366   return Node::cmp(n) && n.Opcode() == Op_ShenandoahLoadReferenceBarrier &&
2367          mask_decorators(_decorators) == mask_decorators(((const ShenandoahLoadReferenceBarrierNode&)n)._decorators);
2368 }
2369 
2370 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const {
2371   if (in(ValueIn) == nullptr || in(ValueIn)->is_top()) {
2372     return Type::TOP;
2373   }
2374   const Type* t = in(ValueIn)->bottom_type();
2375   if (t == TypePtr::NULL_PTR) {
2376     return t;
2377   }
2378 
2379   if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2380     return t;
2381   }
2382 
2383   return t->meet(TypePtr::NULL_PTR);
2384 }
2385 
2386 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseGVN* phase) const {
2387   // Either input is TOP ==> the result is TOP
2388   const Type *t2 = phase->type(in(ValueIn));
2389   if( t2 == Type::TOP ) return Type::TOP;
2390 
2391   if (t2 == TypePtr::NULL_PTR) {
2392     return t2;
2393   }
2394 
2395   if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2396     return t2;
2397   }
2398 
2399   return t2->meet(TypePtr::NULL_PTR);
2400 }
2401 
2402 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseGVN* phase) {
2403   Node* value = in(ValueIn);
2404   if (!needs_barrier(phase, value)) {
2405     return value;
2406   }
2407   return this;
2408 }
2409 
2410 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseGVN* phase, Node* n) {
2411   Unique_Node_List visited;
2412   return needs_barrier_impl(phase, n, visited);
2413 }
2414 
2415 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseGVN* phase, Node* n, Unique_Node_List &visited) {
2416   if (n == nullptr) return false;
2417   if (visited.member(n)) {
2418     return false; // Been there.
2419   }
2420   visited.push(n);
2421 
2422   if (n->is_Allocate()) {
2423     // tty->print_cr("optimize barrier on alloc");
2424     return false;
2425   }
2426   if (n->is_Call()) {
2427     // tty->print_cr("optimize barrier on call");
2428     return false;
2429   }
2430 
2431   const Type* type = phase->type(n);
2432   if (type == Type::TOP) {
2433     return false;
2434   }
2435   if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) {
2436     // tty->print_cr("optimize barrier on null");
2437     return false;
2438   }
2439   if (type->make_oopptr() && type->make_oopptr()->const_oop() != nullptr) {
2440     // tty->print_cr("optimize barrier on constant");
2441     return false;
2442   }
2443 
2444   switch (n->Opcode()) {
2445     case Op_AddP:
2446       return true; // TODO: Can refine?
2447     case Op_LoadP:
2448     case Op_ShenandoahCompareAndExchangeN:
2449     case Op_ShenandoahCompareAndExchangeP:
2450     case Op_CompareAndExchangeN:
2451     case Op_CompareAndExchangeP:
2452     case Op_GetAndSetN:
2453     case Op_GetAndSetP:
2454       return true;
2455     case Op_Phi: {
2456       for (uint i = 1; i < n->req(); i++) {
2457         if (needs_barrier_impl(phase, n->in(i), visited)) return true;
2458       }
2459       return false;
2460     }
2461     case Op_CheckCastPP:
2462     case Op_CastPP:
2463       return needs_barrier_impl(phase, n->in(1), visited);
2464     case Op_Proj:
2465       return needs_barrier_impl(phase, n->in(0), visited);
2466     case Op_ShenandoahLoadReferenceBarrier:
2467       // tty->print_cr("optimize barrier on barrier");
2468       return false;
2469     case Op_Parm:
2470       // tty->print_cr("optimize barrier on input arg");
2471       return false;
2472     case Op_DecodeN:
2473     case Op_EncodeP:
2474       return needs_barrier_impl(phase, n->in(1), visited);
2475     case Op_LoadN:
2476       return true;
2477     case Op_CMoveN:
2478     case Op_CMoveP:
2479       return needs_barrier_impl(phase, n->in(2), visited) ||
2480              needs_barrier_impl(phase, n->in(3), visited);
2481     case Op_CreateEx:
2482       return false;
2483     default:
2484       break;
2485   }
2486 #ifdef ASSERT
2487   tty->print("need barrier on?: ");
2488   tty->print_cr("ins:");
2489   n->dump(2);
2490   tty->print_cr("outs:");
2491   n->dump(-2);
2492   ShouldNotReachHere();
2493 #endif
2494   return true;
2495 }