1 /*
   2  * Copyright (c) 2015, 2021, Red Hat, Inc. All rights reserved.
   3  * Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This code is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License version 2 only, as
   8  * published by the Free Software Foundation.
   9  *
  10  * This code is distributed in the hope that it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  13  * version 2 for more details (a copy is included in the LICENSE file that
  14  * accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License version
  17  * 2 along with this work; if not, write to the Free Software Foundation,
  18  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  19  *
  20  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  21  * or visit www.oracle.com if you need additional information or have any
  22  * questions.
  23  *
  24  */
  25 
  26 #include "precompiled.hpp"
  27 
  28 #include "classfile/javaClasses.hpp"
  29 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
  30 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
  31 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
  32 #include "gc/shenandoah/shenandoahForwarding.hpp"
  33 #include "gc/shenandoah/shenandoahHeap.hpp"
  34 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
  35 #include "gc/shenandoah/shenandoahRuntime.hpp"
  36 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
  37 #include "opto/arraycopynode.hpp"
  38 #include "opto/block.hpp"
  39 #include "opto/callnode.hpp"
  40 #include "opto/castnode.hpp"
  41 #include "opto/movenode.hpp"
  42 #include "opto/phaseX.hpp"
  43 #include "opto/rootnode.hpp"
  44 #include "opto/runtime.hpp"
  45 #include "opto/subnode.hpp"
  46 
  47 bool ShenandoahBarrierC2Support::expand(Compile* C, PhaseIterGVN& igvn) {
  48   ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
  49   if (state->load_reference_barriers_count() > 0) {
  50     assert(C->post_loop_opts_phase(), "no loop opts allowed");
  51     C->reset_post_loop_opts_phase(); // ... but we know what we are doing
  52     C->clear_major_progress();
  53     PhaseIdealLoop::optimize(igvn, LoopOptsShenandoahExpand);
  54     if (C->failing()) return false;
  55     C->process_for_post_loop_opts_igvn(igvn);
  56     if (C->failing()) return false;
  57 
  58     C->set_post_loop_opts_phase(); // now for real!
  59   }
  60   return true;
  61 }
  62 
  63 bool ShenandoahBarrierC2Support::is_gc_state_test(Node* iff, int mask) {
  64   if (!UseShenandoahGC) {
  65     return false;
  66   }
  67   assert(iff->is_If(), "bad input");
  68   if (iff->Opcode() != Op_If) {
  69     return false;
  70   }
  71   Node* bol = iff->in(1);
  72   if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) {
  73     return false;
  74   }
  75   Node* cmp = bol->in(1);
  76   if (cmp->Opcode() != Op_CmpI) {
  77     return false;
  78   }
  79   Node* in1 = cmp->in(1);
  80   Node* in2 = cmp->in(2);
  81   if (in2->find_int_con(-1) != 0) {
  82     return false;
  83   }
  84   if (in1->Opcode() != Op_AndI) {
  85     return false;
  86   }
  87   in2 = in1->in(2);
  88   if (in2->find_int_con(-1) != mask) {
  89     return false;
  90   }
  91   in1 = in1->in(1);
  92 
  93   return is_gc_state_load(in1);
  94 }
  95 
  96 bool ShenandoahBarrierC2Support::is_heap_stable_test(Node* iff) {
  97   return is_gc_state_test(iff, ShenandoahHeap::HAS_FORWARDED);
  98 }
  99 
 100 bool ShenandoahBarrierC2Support::is_gc_state_load(Node *n) {
 101   if (!UseShenandoahGC) {
 102     return false;
 103   }
 104   if (n->Opcode() != Op_LoadB && n->Opcode() != Op_LoadUB) {
 105     return false;
 106   }
 107   Node* addp = n->in(MemNode::Address);
 108   if (!addp->is_AddP()) {
 109     return false;
 110   }
 111   Node* base = addp->in(AddPNode::Address);
 112   Node* off = addp->in(AddPNode::Offset);
 113   if (base->Opcode() != Op_ThreadLocal) {
 114     return false;
 115   }
 116   if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) {
 117     return false;
 118   }
 119   return true;
 120 }
 121 
 122 bool ShenandoahBarrierC2Support::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) {
 123   assert(phase->is_dominator(stop, start), "bad inputs");
 124   ResourceMark rm;
 125   Unique_Node_List wq;
 126   wq.push(start);
 127   for (uint next = 0; next < wq.size(); next++) {
 128     Node *m = wq.at(next);
 129     if (m == stop) {
 130       continue;
 131     }
 132     if (m->is_SafePoint() && !m->is_CallLeaf()) {
 133       return true;
 134     }
 135     if (m->is_Region()) {
 136       for (uint i = 1; i < m->req(); i++) {
 137         wq.push(m->in(i));
 138       }
 139     } else {
 140       wq.push(m->in(0));
 141     }
 142   }
 143   return false;
 144 }
 145 
 146 #ifdef ASSERT
 147 bool ShenandoahBarrierC2Support::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) {
 148   assert(phis.size() == 0, "");
 149 
 150   while (true) {
 151     if (in->bottom_type() == TypePtr::NULL_PTR) {
 152       if (trace) {tty->print_cr("null");}
 153     } else if (!in->bottom_type()->make_ptr()->make_oopptr()) {
 154       if (trace) {tty->print_cr("Non oop");}
 155     } else {
 156       if (in->is_ConstraintCast()) {
 157         in = in->in(1);
 158         continue;
 159       } else if (in->is_AddP()) {
 160         assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access");
 161         in = in->in(AddPNode::Address);
 162         continue;
 163       } else if (in->is_Con()) {
 164         if (trace) {
 165           tty->print("Found constant");
 166           in->dump();
 167         }
 168       } else if (in->Opcode() == Op_Parm) {
 169         if (trace) {
 170           tty->print("Found argument");
 171         }
 172       } else if (in->Opcode() == Op_CreateEx) {
 173         if (trace) {
 174           tty->print("Found create-exception");
 175         }
 176       } else if (in->Opcode() == Op_LoadP && in->adr_type() == TypeRawPtr::BOTTOM) {
 177         if (trace) {
 178           tty->print("Found raw LoadP (OSR argument?)");
 179         }
 180       } else if (in->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 181         if (t == ShenandoahOopStore) {
 182           return false;
 183         }
 184         barriers_used.push(in);
 185         if (trace) {tty->print("Found barrier"); in->dump();}
 186       } else if (in->is_Proj() && in->in(0)->is_Allocate()) {
 187         if (trace) {
 188           tty->print("Found alloc");
 189           in->in(0)->dump();
 190         }
 191       } else if (in->is_Proj() && (in->in(0)->Opcode() == Op_CallStaticJava || in->in(0)->Opcode() == Op_CallDynamicJava)) {
 192         if (trace) {
 193           tty->print("Found Java call");
 194         }
 195       } else if (in->is_Phi()) {
 196         if (!visited.test_set(in->_idx)) {
 197           if (trace) {tty->print("Pushed phi:"); in->dump();}
 198           phis.push(in, 2);
 199           in = in->in(1);
 200           continue;
 201         }
 202         if (trace) {tty->print("Already seen phi:"); in->dump();}
 203       } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) {
 204         if (!visited.test_set(in->_idx)) {
 205           if (trace) {tty->print("Pushed cmovep:"); in->dump();}
 206           phis.push(in, CMoveNode::IfTrue);
 207           in = in->in(CMoveNode::IfFalse);
 208           continue;
 209         }
 210         if (trace) {tty->print("Already seen cmovep:"); in->dump();}
 211       } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) {
 212         in = in->in(1);
 213         continue;
 214       } else {
 215         return false;
 216       }
 217     }
 218     bool cont = false;
 219     while (phis.is_nonempty()) {
 220       uint idx = phis.index();
 221       Node* phi = phis.node();
 222       if (idx >= phi->req()) {
 223         if (trace) {tty->print("Popped phi:"); phi->dump();}
 224         phis.pop();
 225         continue;
 226       }
 227       if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();}
 228       in = phi->in(idx);
 229       phis.set_index(idx+1);
 230       cont = true;
 231       break;
 232     }
 233     if (!cont) {
 234       break;
 235     }
 236   }
 237   return true;
 238 }
 239 
 240 void ShenandoahBarrierC2Support::report_verify_failure(const char* msg, Node* n1, Node* n2) {
 241   if (n1 != nullptr) {
 242     n1->dump(+10);
 243   }
 244   if (n2 != nullptr) {
 245     n2->dump(+10);
 246   }
 247   fatal("%s", msg);
 248 }
 249 
 250 void ShenandoahBarrierC2Support::verify(RootNode* root) {
 251   ResourceMark rm;
 252   Unique_Node_List wq;
 253   GrowableArray<Node*> barriers;
 254   Unique_Node_List barriers_used;
 255   Node_Stack phis(0);
 256   VectorSet visited;
 257   const bool trace = false;
 258   const bool verify_no_useless_barrier = false;
 259 
 260   wq.push(root);
 261   for (uint next = 0; next < wq.size(); next++) {
 262     Node *n = wq.at(next);
 263     if (n->is_Load()) {
 264       const bool trace = false;
 265       if (trace) {tty->print("Verifying"); n->dump();}
 266       if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) {
 267         if (trace) {tty->print_cr("Load range/klass");}
 268       } else {
 269         const TypePtr* adr_type = n->as_Load()->adr_type();
 270 
 271         if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) {
 272           if (trace) {tty->print_cr("Mark load");}
 273         } else if (adr_type->isa_instptr() &&
 274                    adr_type->is_instptr()->instance_klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
 275                    adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset()) {
 276           if (trace) {tty->print_cr("Reference.get()");}
 277         } else if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) {
 278           report_verify_failure("Shenandoah verification: Load should have barriers", n);
 279         }
 280       }
 281     } else if (n->is_Store()) {
 282       const bool trace = false;
 283 
 284       if (trace) {tty->print("Verifying"); n->dump();}
 285       if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) {
 286         Node* adr = n->in(MemNode::Address);
 287         bool verify = true;
 288 
 289         if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) {
 290           adr = adr->in(AddPNode::Address);
 291           if (adr->is_AddP()) {
 292             assert(adr->in(AddPNode::Base)->is_top(), "");
 293             adr = adr->in(AddPNode::Address);
 294             if (adr->Opcode() == Op_LoadP &&
 295                 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() &&
 296                 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
 297                 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) {
 298               if (trace) {tty->print_cr("SATB prebarrier");}
 299               verify = false;
 300             }
 301           }
 302         }
 303 
 304         if (verify && !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahValue, trace, barriers_used)) {
 305           report_verify_failure("Shenandoah verification: Store should have barriers", n);
 306         }
 307       }
 308       if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
 309         report_verify_failure("Shenandoah verification: Store (address) should have barriers", n);
 310       }
 311     } else if (n->Opcode() == Op_CmpP) {
 312       const bool trace = false;
 313 
 314       Node* in1 = n->in(1);
 315       Node* in2 = n->in(2);
 316       if (in1->bottom_type()->isa_oopptr()) {
 317         if (trace) {tty->print("Verifying"); n->dump();}
 318 
 319         bool mark_inputs = false;
 320         if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR ||
 321             (in1->is_Con() || in2->is_Con())) {
 322           if (trace) {tty->print_cr("Comparison against a constant");}
 323           mark_inputs = true;
 324         } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) ||
 325                    (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) {
 326           if (trace) {tty->print_cr("Comparison with newly alloc'ed object");}
 327           mark_inputs = true;
 328         } else {
 329           assert(in2->bottom_type()->isa_oopptr(), "");
 330 
 331           if (!verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) ||
 332               !verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) {
 333             report_verify_failure("Shenandoah verification: Cmp should have barriers", n);
 334           }
 335         }
 336         if (verify_no_useless_barrier &&
 337             mark_inputs &&
 338             (!verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) ||
 339              !verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) {
 340           phis.clear();
 341           visited.reset();
 342         }
 343       }
 344     } else if (n->is_LoadStore()) {
 345       if (n->in(MemNode::ValueIn)->bottom_type()->make_ptr() &&
 346           !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahValue, trace, barriers_used)) {
 347         report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n);
 348       }
 349 
 350       if (n->in(MemNode::Address)->bottom_type()->make_oopptr() && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
 351         report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n);
 352       }
 353     } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) {
 354       CallNode* call = n->as_Call();
 355 
 356       static struct {
 357         const char* name;
 358         struct {
 359           int pos;
 360           verify_type t;
 361         } args[6];
 362       } calls[] = {
 363         "array_partition_stub",
 364         { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+4, ShenandoahStore },   { -1, ShenandoahNone },
 365           { -1, ShenandoahNone },                { -1, ShenandoahNone },                  { -1, ShenandoahNone } },
 366         "arraysort_stub",
 367         { { TypeFunc::Parms, ShenandoahStore },  { -1, ShenandoahNone },                  { -1, ShenandoahNone },
 368           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 369         "aescrypt_encryptBlock",
 370         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 371           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 372         "aescrypt_decryptBlock",
 373         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 374           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 375         "multiplyToLen",
 376         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },   { TypeFunc::Parms+4, ShenandoahStore },
 377           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 378         "squareToLen",
 379         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },   { -1,  ShenandoahNone},
 380           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 381         "montgomery_multiply",
 382         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },
 383           { TypeFunc::Parms+6, ShenandoahStore }, { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 384         "montgomery_square",
 385         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahLoad },   { TypeFunc::Parms+5, ShenandoahStore },
 386           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 387         "mulAdd",
 388         { { TypeFunc::Parms, ShenandoahStore },  { TypeFunc::Parms+1, ShenandoahLoad },   { -1,  ShenandoahNone},
 389           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 390         "vectorizedMismatch",
 391         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahLoad },   { -1,  ShenandoahNone},
 392           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 393         "updateBytesCRC32",
 394         { { TypeFunc::Parms+1, ShenandoahLoad }, { -1,  ShenandoahNone},                  { -1,  ShenandoahNone},
 395           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 396         "updateBytesAdler32",
 397         { { TypeFunc::Parms+1, ShenandoahLoad }, { -1,  ShenandoahNone},                  { -1,  ShenandoahNone},
 398           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 399         "updateBytesCRC32C",
 400         { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad},    { -1,  ShenandoahNone},
 401           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 402         "counterMode_AESCrypt",
 403         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 404           { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } },
 405         "cipherBlockChaining_encryptAESCrypt",
 406         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 407           { TypeFunc::Parms+3, ShenandoahLoad },  { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 408         "cipherBlockChaining_decryptAESCrypt",
 409         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 410           { TypeFunc::Parms+3, ShenandoahLoad },  { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 411         "shenandoah_clone",
 412         { { TypeFunc::Parms, ShenandoahLoad },   { -1,  ShenandoahNone},                  { -1,  ShenandoahNone},
 413           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 414         "ghash_processBlocks",
 415         { { TypeFunc::Parms, ShenandoahStore },  { TypeFunc::Parms+1, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },
 416           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 417         "sha1_implCompress",
 418         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 419           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 420         "sha256_implCompress",
 421         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 422           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 423         "sha512_implCompress",
 424         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 425           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 426         "sha1_implCompressMB",
 427         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 428           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 429         "sha256_implCompressMB",
 430         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 431           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 432         "sha512_implCompressMB",
 433         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 434           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 435         "encodeBlock",
 436         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+3, ShenandoahStore },   { -1, ShenandoahNone },
 437           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 438         "decodeBlock",
 439         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+3, ShenandoahStore },   { -1, ShenandoahNone },
 440           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 441       };
 442 
 443       if (call->is_call_to_arraycopystub()) {
 444         Node* dest = nullptr;
 445         const TypeTuple* args = n->as_Call()->_tf->domain();
 446         for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
 447           if (args->field_at(i)->isa_ptr()) {
 448             j++;
 449             if (j == 2) {
 450               dest = n->in(i);
 451               break;
 452             }
 453           }
 454         }
 455         if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) ||
 456             !verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) {
 457           report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n);
 458         }
 459       } else if (strlen(call->_name) > 5 &&
 460                  !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) {
 461         if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) {
 462           report_verify_failure("Shenandoah verification: _fill should have barriers", n);
 463         }
 464       } else if (!strcmp(call->_name, "shenandoah_wb_pre")) {
 465         // skip
 466       } else {
 467         const int calls_len = sizeof(calls) / sizeof(calls[0]);
 468         int i = 0;
 469         for (; i < calls_len; i++) {
 470           if (!strcmp(calls[i].name, call->_name)) {
 471             break;
 472           }
 473         }
 474         if (i != calls_len) {
 475           const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]);
 476           for (uint j = 0; j < args_len; j++) {
 477             int pos = calls[i].args[j].pos;
 478             if (pos == -1) {
 479               break;
 480             }
 481             if (!verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) {
 482               report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
 483             }
 484           }
 485           for (uint j = TypeFunc::Parms; j < call->req(); j++) {
 486             if (call->in(j)->bottom_type()->make_ptr() &&
 487                 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
 488               uint k = 0;
 489               for (; k < args_len && calls[i].args[k].pos != (int)j; k++);
 490               if (k == args_len) {
 491                 fatal("arg %d for call %s not covered", j, call->_name);
 492               }
 493             }
 494           }
 495         } else {
 496           for (uint j = TypeFunc::Parms; j < call->req(); j++) {
 497             if (call->in(j)->bottom_type()->make_ptr() &&
 498                 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
 499               fatal("%s not covered", call->_name);
 500             }
 501           }
 502         }
 503       }
 504     } else if (n->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 505       // skip
 506     } else if (n->is_AddP()
 507                || n->is_Phi()
 508                || n->is_ConstraintCast()
 509                || n->Opcode() == Op_Return
 510                || n->Opcode() == Op_CMoveP
 511                || n->Opcode() == Op_CMoveN
 512                || n->Opcode() == Op_Rethrow
 513                || n->is_MemBar()
 514                || n->Opcode() == Op_Conv2B
 515                || n->Opcode() == Op_SafePoint
 516                || n->is_CallJava()
 517                || n->Opcode() == Op_Unlock
 518                || n->Opcode() == Op_EncodeP
 519                || n->Opcode() == Op_DecodeN) {
 520       // nothing to do
 521     } else {
 522       static struct {
 523         int opcode;
 524         struct {
 525           int pos;
 526           verify_type t;
 527         } inputs[2];
 528       } others[] = {
 529         Op_FastLock,
 530         { { 1, ShenandoahLoad },                  { -1, ShenandoahNone} },
 531         Op_Lock,
 532         { { TypeFunc::Parms, ShenandoahLoad },    { -1, ShenandoahNone} },
 533         Op_ArrayCopy,
 534         { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } },
 535         Op_StrCompressedCopy,
 536         { { 2, ShenandoahLoad },                  { 3, ShenandoahStore } },
 537         Op_StrInflatedCopy,
 538         { { 2, ShenandoahLoad },                  { 3, ShenandoahStore } },
 539         Op_AryEq,
 540         { { 2, ShenandoahLoad },                  { 3, ShenandoahLoad } },
 541         Op_StrIndexOf,
 542         { { 2, ShenandoahLoad },                  { 4, ShenandoahLoad } },
 543         Op_StrComp,
 544         { { 2, ShenandoahLoad },                  { 4, ShenandoahLoad } },
 545         Op_StrEquals,
 546         { { 2, ShenandoahLoad },                  { 3, ShenandoahLoad } },
 547         Op_VectorizedHashCode,
 548         { { 2, ShenandoahLoad },                  { -1, ShenandoahNone } },
 549         Op_EncodeISOArray,
 550         { { 2, ShenandoahLoad },                  { 3, ShenandoahStore } },
 551         Op_CountPositives,
 552         { { 2, ShenandoahLoad },                  { -1, ShenandoahNone} },
 553         Op_CastP2X,
 554         { { 1, ShenandoahLoad },                  { -1, ShenandoahNone} },
 555         Op_StrIndexOfChar,
 556         { { 2, ShenandoahLoad },                  { -1, ShenandoahNone } },
 557       };
 558 
 559       const int others_len = sizeof(others) / sizeof(others[0]);
 560       int i = 0;
 561       for (; i < others_len; i++) {
 562         if (others[i].opcode == n->Opcode()) {
 563           break;
 564         }
 565       }
 566       uint stop = n->is_Call() ? n->as_Call()->tf()->domain()->cnt() : n->req();
 567       if (i != others_len) {
 568         const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]);
 569         for (uint j = 0; j < inputs_len; j++) {
 570           int pos = others[i].inputs[j].pos;
 571           if (pos == -1) {
 572             break;
 573           }
 574           if (!verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) {
 575             report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
 576           }
 577         }
 578         for (uint j = 1; j < stop; j++) {
 579           if (n->in(j) != nullptr && n->in(j)->bottom_type()->make_ptr() &&
 580               n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
 581             uint k = 0;
 582             for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++);
 583             if (k == inputs_len) {
 584               fatal("arg %d for node %s not covered", j, n->Name());
 585             }
 586           }
 587         }
 588       } else {
 589         for (uint j = 1; j < stop; j++) {
 590           if (n->in(j) != nullptr && n->in(j)->bottom_type()->make_ptr() &&
 591               n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
 592             fatal("%s not covered", n->Name());
 593           }
 594         }
 595       }
 596     }
 597 
 598     if (n->is_SafePoint()) {
 599       SafePointNode* sfpt = n->as_SafePoint();
 600       if (verify_no_useless_barrier && sfpt->jvms() != nullptr) {
 601         for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) {
 602           if (!verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) {
 603             phis.clear();
 604             visited.reset();
 605           }
 606         }
 607       }
 608     }
 609   }
 610 
 611   if (verify_no_useless_barrier) {
 612     for (int i = 0; i < barriers.length(); i++) {
 613       Node* n = barriers.at(i);
 614       if (!barriers_used.member(n)) {
 615         tty->print("XXX useless barrier"); n->dump(-2);
 616         ShouldNotReachHere();
 617       }
 618     }
 619   }
 620 }
 621 #endif
 622 
 623 bool ShenandoahBarrierC2Support::is_dominator_same_ctrl(Node* c, Node* d, Node* n, PhaseIdealLoop* phase) {
 624   // That both nodes have the same control is not sufficient to prove
 625   // domination, verify that there's no path from d to n
 626   ResourceMark rm;
 627   Unique_Node_List wq;
 628   wq.push(d);
 629   for (uint next = 0; next < wq.size(); next++) {
 630     Node *m = wq.at(next);
 631     if (m == n) {
 632       return false;
 633     }
 634     if (m->is_Phi() && m->in(0)->is_Loop()) {
 635       assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control");
 636     } else {
 637       if (m->is_Store() || m->is_LoadStore()) {
 638         // Take anti-dependencies into account
 639         Node* mem = m->in(MemNode::Memory);
 640         for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
 641           Node* u = mem->fast_out(i);
 642           if (u->is_Load() && phase->C->can_alias(m->adr_type(), phase->C->get_alias_index(u->adr_type())) &&
 643               phase->ctrl_or_self(u) == c) {
 644             wq.push(u);
 645           }
 646         }
 647       }
 648       for (uint i = 0; i < m->req(); i++) {
 649         if (m->in(i) != nullptr && phase->ctrl_or_self(m->in(i)) == c) {
 650           wq.push(m->in(i));
 651         }
 652       }
 653     }
 654   }
 655   return true;
 656 }
 657 
 658 bool ShenandoahBarrierC2Support::is_dominator(Node* d_c, Node* n_c, Node* d, Node* n, PhaseIdealLoop* phase) {
 659   if (d_c != n_c) {
 660     return phase->is_dominator(d_c, n_c);
 661   }
 662   return is_dominator_same_ctrl(d_c, d, n, phase);
 663 }
 664 
 665 Node* next_mem(Node* mem, int alias) {
 666   Node* res = nullptr;
 667   if (mem->is_Proj()) {
 668     res = mem->in(0);
 669   } else if (mem->is_SafePoint() || mem->is_MemBar()) {
 670     res = mem->in(TypeFunc::Memory);
 671   } else if (mem->is_Phi()) {
 672     res = mem->in(1);
 673   } else if (mem->is_MergeMem()) {
 674     res = mem->as_MergeMem()->memory_at(alias);
 675   } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
 676     assert(alias == Compile::AliasIdxRaw, "following raw memory can't lead to a barrier");
 677     res = mem->in(MemNode::Memory);
 678   } else {
 679 #ifdef ASSERT
 680     mem->dump();
 681 #endif
 682     ShouldNotReachHere();
 683   }
 684   return res;
 685 }
 686 
 687 Node* ShenandoahBarrierC2Support::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) {
 688   Node* iffproj = nullptr;
 689   while (c != dom) {
 690     Node* next = phase->idom(c);
 691     assert(next->unique_ctrl_out_or_null() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?");
 692     if (c->is_Region()) {
 693       ResourceMark rm;
 694       Unique_Node_List wq;
 695       wq.push(c);
 696       for (uint i = 0; i < wq.size(); i++) {
 697         Node *n = wq.at(i);
 698         if (n == next) {
 699           continue;
 700         }
 701         if (n->is_Region()) {
 702           for (uint j = 1; j < n->req(); j++) {
 703             wq.push(n->in(j));
 704           }
 705         } else {
 706           wq.push(n->in(0));
 707         }
 708       }
 709       for (uint i = 0; i < wq.size(); i++) {
 710         Node *n = wq.at(i);
 711         assert(n->is_CFG(), "");
 712         if (n->is_Multi()) {
 713           for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
 714             Node* u = n->fast_out(j);
 715             if (u->is_CFG()) {
 716               if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
 717                 return NodeSentinel;
 718               }
 719             }
 720           }
 721         }
 722       }
 723     } else  if (c->is_Proj()) {
 724       if (c->is_IfProj()) {
 725         if (c->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) != nullptr) {
 726           // continue;
 727         } else {
 728           if (!allow_one_proj) {
 729             return NodeSentinel;
 730           }
 731           if (iffproj == nullptr) {
 732             iffproj = c;
 733           } else {
 734             return NodeSentinel;
 735           }
 736         }
 737       } else if (c->Opcode() == Op_JumpProj) {
 738         return NodeSentinel; // unsupported
 739       } else if (c->Opcode() == Op_CatchProj) {
 740         return NodeSentinel; // unsupported
 741       } else if (c->Opcode() == Op_CProj && next->is_NeverBranch()) {
 742         return NodeSentinel; // unsupported
 743       } else {
 744         assert(next->unique_ctrl_out() == c, "unsupported branch pattern");
 745       }
 746     }
 747     c = next;
 748   }
 749   return iffproj;
 750 }
 751 
 752 Node* ShenandoahBarrierC2Support::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) {
 753   ResourceMark rm;
 754   VectorSet wq;
 755   wq.set(mem->_idx);
 756   mem_ctrl = phase->ctrl_or_self(mem);
 757   while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) {
 758     mem = next_mem(mem, alias);
 759     if (wq.test_set(mem->_idx)) {
 760       return nullptr;
 761     }
 762     mem_ctrl = phase->ctrl_or_self(mem);
 763   }
 764   if (mem->is_MergeMem()) {
 765     mem = mem->as_MergeMem()->memory_at(alias);
 766     mem_ctrl = phase->ctrl_or_self(mem);
 767   }
 768   return mem;
 769 }
 770 
 771 Node* ShenandoahBarrierC2Support::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) {
 772   Node* mem = nullptr;
 773   Node* c = ctrl;
 774   do {
 775     if (c->is_Region()) {
 776       for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == nullptr; i++) {
 777         Node* u = c->fast_out(i);
 778         if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
 779           if (u->adr_type() == TypePtr::BOTTOM) {
 780             mem = u;
 781           }
 782         }
 783       }
 784     } else {
 785       if (c->is_Call() && c->as_Call()->adr_type() != nullptr) {
 786         CallProjections projs;
 787         c->as_Call()->extract_projections(&projs, true, false);
 788         if (projs.fallthrough_memproj != nullptr) {
 789           if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) {
 790             if (projs.catchall_memproj == nullptr) {
 791               mem = projs.fallthrough_memproj;
 792             } else {
 793               if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) {
 794                 mem = projs.fallthrough_memproj;
 795               } else {
 796                 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier");
 797                 mem = projs.catchall_memproj;
 798               }
 799             }
 800           }
 801         } else {
 802           Node* proj = c->as_Call()->proj_out(TypeFunc::Memory);
 803           if (proj != nullptr &&
 804               proj->adr_type() == TypePtr::BOTTOM) {
 805             mem = proj;
 806           }
 807         }
 808       } else {
 809         for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
 810           Node* u = c->fast_out(i);
 811           if (u->is_Proj() &&
 812               u->bottom_type() == Type::MEMORY &&
 813               u->adr_type() == TypePtr::BOTTOM) {
 814               assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), "");
 815               assert(mem == nullptr, "only one proj");
 816               mem = u;
 817           }
 818         }
 819         assert(!c->is_Call() || c->as_Call()->adr_type() != nullptr || mem == nullptr, "no mem projection expected");
 820       }
 821     }
 822     c = phase->idom(c);
 823   } while (mem == nullptr);
 824   return mem;
 825 }
 826 
 827 void ShenandoahBarrierC2Support::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) {
 828   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 829     Node* u = n->fast_out(i);
 830     if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) {
 831       uses.push(u);
 832     }
 833   }
 834 }
 835 
 836 static void hide_strip_mined_loop(OuterStripMinedLoopNode* outer, CountedLoopNode* inner, PhaseIdealLoop* phase) {
 837   OuterStripMinedLoopEndNode* le = inner->outer_loop_end();
 838   Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl));
 839   phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl));
 840   Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt);
 841   phase->register_control(new_le, phase->get_loop(le), le->in(0));
 842   phase->lazy_replace(outer, new_outer);
 843   phase->lazy_replace(le, new_le);
 844   inner->clear_strip_mined();
 845 }
 846 
 847 void ShenandoahBarrierC2Support::test_gc_state(Node*& ctrl, Node* raw_mem, Node*& test_fail_ctrl,
 848                                                PhaseIdealLoop* phase, int flags) {
 849   PhaseIterGVN& igvn = phase->igvn();
 850   Node* old_ctrl = ctrl;
 851 
 852   Node* thread          = new ThreadLocalNode();
 853   Node* gc_state_offset = igvn.MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 854   Node* gc_state_addr   = new AddPNode(phase->C->top(), thread, gc_state_offset);
 855   Node* gc_state        = new LoadBNode(old_ctrl, raw_mem, gc_state_addr,
 856                                         DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(nullptr),
 857                                         TypeInt::BYTE, MemNode::unordered);
 858   Node* gc_state_and    = new AndINode(gc_state, igvn.intcon(flags));
 859   Node* gc_state_cmp    = new CmpINode(gc_state_and, igvn.zerocon(T_INT));
 860   Node* gc_state_bool   = new BoolNode(gc_state_cmp, BoolTest::ne);
 861 
 862   IfNode* gc_state_iff  = new IfNode(old_ctrl, gc_state_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
 863   ctrl                  = new IfTrueNode(gc_state_iff);
 864   test_fail_ctrl        = new IfFalseNode(gc_state_iff);
 865 
 866   IdealLoopTree* loop = phase->get_loop(old_ctrl);
 867   phase->register_control(gc_state_iff,   loop, old_ctrl);
 868   phase->register_control(ctrl,           loop, gc_state_iff);
 869   phase->register_control(test_fail_ctrl, loop, gc_state_iff);
 870 
 871   phase->register_new_node(thread,        old_ctrl);
 872   phase->register_new_node(gc_state_addr, old_ctrl);
 873   phase->register_new_node(gc_state,      old_ctrl);
 874   phase->register_new_node(gc_state_and,  old_ctrl);
 875   phase->register_new_node(gc_state_cmp,  old_ctrl);
 876   phase->register_new_node(gc_state_bool, old_ctrl);
 877 
 878   phase->set_ctrl(gc_state_offset, phase->C->root());
 879 
 880   assert(is_gc_state_test(gc_state_iff, flags), "Should match the shape");
 881 }
 882 
 883 void ShenandoahBarrierC2Support::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) {
 884   Node* old_ctrl = ctrl;
 885   PhaseIterGVN& igvn = phase->igvn();
 886 
 887   const Type* val_t = igvn.type(val);
 888   if (val_t->meet(TypePtr::NULL_PTR) == val_t) {
 889     Node* null_cmp   = new CmpPNode(val, igvn.zerocon(T_OBJECT));
 890     Node* null_test  = new BoolNode(null_cmp, BoolTest::ne);
 891 
 892     IfNode* null_iff = new IfNode(old_ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
 893     ctrl             = new IfTrueNode(null_iff);
 894     null_ctrl        = new IfFalseNode(null_iff);
 895 
 896     IdealLoopTree* loop = phase->get_loop(old_ctrl);
 897     phase->register_control(null_iff,  loop, old_ctrl);
 898     phase->register_control(ctrl,      loop, null_iff);
 899     phase->register_control(null_ctrl, loop, null_iff);
 900 
 901     phase->register_new_node(null_cmp,  old_ctrl);
 902     phase->register_new_node(null_test, old_ctrl);
 903   }
 904 }
 905 
 906 void ShenandoahBarrierC2Support::test_in_cset(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) {
 907   Node* old_ctrl = ctrl;
 908   PhaseIterGVN& igvn = phase->igvn();
 909 
 910   Node* raw_val        = new CastP2XNode(old_ctrl, val);
 911   Node* cset_idx       = new URShiftXNode(raw_val, igvn.intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint()));
 912 
 913   // Figure out the target cset address with raw pointer math.
 914   // This avoids matching AddP+LoadB that would emit inefficient code.
 915   // See JDK-8245465.
 916   Node* cset_addr_ptr  = igvn.makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr()));
 917   Node* cset_addr      = new CastP2XNode(old_ctrl, cset_addr_ptr);
 918   Node* cset_load_addr = new AddXNode(cset_addr, cset_idx);
 919   Node* cset_load_ptr  = new CastX2PNode(cset_load_addr);
 920 
 921   Node* cset_load      = new LoadBNode(old_ctrl, raw_mem, cset_load_ptr,
 922                                        DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(nullptr),
 923                                        TypeInt::BYTE, MemNode::unordered);
 924   Node* cset_cmp       = new CmpINode(cset_load, igvn.zerocon(T_INT));
 925   Node* cset_bool      = new BoolNode(cset_cmp, BoolTest::ne);
 926 
 927   IfNode* cset_iff     = new IfNode(old_ctrl, cset_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
 928   ctrl                 = new IfTrueNode(cset_iff);
 929   not_cset_ctrl        = new IfFalseNode(cset_iff);
 930 
 931   IdealLoopTree *loop = phase->get_loop(old_ctrl);
 932   phase->register_control(cset_iff,      loop, old_ctrl);
 933   phase->register_control(ctrl,          loop, cset_iff);
 934   phase->register_control(not_cset_ctrl, loop, cset_iff);
 935 
 936   phase->set_ctrl(cset_addr_ptr, phase->C->root());
 937 
 938   phase->register_new_node(raw_val,        old_ctrl);
 939   phase->register_new_node(cset_idx,       old_ctrl);
 940   phase->register_new_node(cset_addr,      old_ctrl);
 941   phase->register_new_node(cset_load_addr, old_ctrl);
 942   phase->register_new_node(cset_load_ptr,  old_ctrl);
 943   phase->register_new_node(cset_load,      old_ctrl);
 944   phase->register_new_node(cset_cmp,       old_ctrl);
 945   phase->register_new_node(cset_bool,      old_ctrl);
 946 }
 947 
 948 void ShenandoahBarrierC2Support::call_lrb_stub(Node*& ctrl, Node*& val, Node* load_addr,
 949                                                DecoratorSet decorators, PhaseIdealLoop* phase) {
 950   IdealLoopTree*loop = phase->get_loop(ctrl);
 951   const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr();
 952 
 953   address calladdr = nullptr;
 954   const char* name = nullptr;
 955   bool is_strong  = ShenandoahBarrierSet::is_strong_access(decorators);
 956   bool is_weak    = ShenandoahBarrierSet::is_weak_access(decorators);
 957   bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
 958   bool is_native  = ShenandoahBarrierSet::is_native_access(decorators);
 959   bool is_narrow  = UseCompressedOops && !is_native;
 960   if (is_strong) {
 961     if (is_narrow) {
 962       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow);
 963       name = "load_reference_barrier_strong_narrow";
 964     } else {
 965       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong);
 966       name = "load_reference_barrier_strong";
 967     }
 968   } else if (is_weak) {
 969     if (is_narrow) {
 970       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow);
 971       name = "load_reference_barrier_weak_narrow";
 972     } else {
 973       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak);
 974       name = "load_reference_barrier_weak";
 975     }
 976   } else {
 977     assert(is_phantom, "only remaining strength");
 978     if (is_narrow) {
 979       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom_narrow);
 980       name = "load_reference_barrier_phantom_narrow";
 981     } else {
 982       calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom);
 983       name = "load_reference_barrier_phantom";
 984     }
 985   }
 986   Node* call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type(), calladdr, name, TypeRawPtr::BOTTOM);
 987 
 988   call->init_req(TypeFunc::Control, ctrl);
 989   call->init_req(TypeFunc::I_O, phase->C->top());
 990   call->init_req(TypeFunc::Memory, phase->C->top());
 991   call->init_req(TypeFunc::FramePtr, phase->C->top());
 992   call->init_req(TypeFunc::ReturnAdr, phase->C->top());
 993   call->init_req(TypeFunc::Parms, val);
 994   call->init_req(TypeFunc::Parms+1, load_addr);
 995   phase->register_control(call, loop, ctrl);
 996   ctrl = new ProjNode(call, TypeFunc::Control);
 997   phase->register_control(ctrl, loop, call);
 998   val = new ProjNode(call, TypeFunc::Parms);
 999   phase->register_new_node(val, call);
1000   val = new CheckCastPPNode(ctrl, val, obj_type);
1001   phase->register_new_node(val, ctrl);
1002 }
1003 
1004 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) {
1005   Node* ctrl = phase->get_ctrl(barrier);
1006   Node* init_raw_mem = fixer.find_mem(ctrl, barrier);
1007 
1008   // Update the control of all nodes that should be after the
1009   // barrier control flow
1010   uses.clear();
1011   // Every node that is control dependent on the barrier's input
1012   // control will be after the expanded barrier. The raw memory (if
1013   // its memory is control dependent on the barrier's input control)
1014   // must stay above the barrier.
1015   uses_to_ignore.clear();
1016   if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) {
1017     uses_to_ignore.push(init_raw_mem);
1018   }
1019   for (uint next = 0; next < uses_to_ignore.size(); next++) {
1020     Node *n = uses_to_ignore.at(next);
1021     for (uint i = 0; i < n->req(); i++) {
1022       Node* in = n->in(i);
1023       if (in != nullptr && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) {
1024         uses_to_ignore.push(in);
1025       }
1026     }
1027   }
1028   for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) {
1029     Node* u = ctrl->fast_out(i);
1030     if (u->_idx < last &&
1031         u != barrier &&
1032         !u->depends_only_on_test() && // preserve dependency on test
1033         !uses_to_ignore.member(u) &&
1034         (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) &&
1035         (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) {
1036       Node* old_c = phase->ctrl_or_self(u);
1037       Node* c = old_c;
1038       if (c != ctrl ||
1039           is_dominator_same_ctrl(old_c, barrier, u, phase) ||
1040           ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) {
1041         phase->igvn().rehash_node_delayed(u);
1042         int nb = u->replace_edge(ctrl, region, &phase->igvn());
1043         if (u->is_CFG()) {
1044           if (phase->idom(u) == ctrl) {
1045             phase->set_idom(u, region, phase->dom_depth(region));
1046           }
1047         } else if (phase->get_ctrl(u) == ctrl) {
1048           assert(u != init_raw_mem, "should leave input raw mem above the barrier");
1049           uses.push(u);
1050         }
1051         assert(nb == 1, "more than 1 ctrl input?");
1052         --i, imax -= nb;
1053       }
1054     }
1055   }
1056 }
1057 
1058 static Node* create_phis_on_call_return(Node* ctrl, Node* c, Node* n, Node* n_clone, const CallProjections& projs, PhaseIdealLoop* phase) {
1059   Node* region = nullptr;
1060   while (c != ctrl) {
1061     if (c->is_Region()) {
1062       region = c;
1063     }
1064     c = phase->idom(c);
1065   }
1066   assert(region != nullptr, "");
1067   Node* phi = new PhiNode(region, n->bottom_type());
1068   for (uint j = 1; j < region->req(); j++) {
1069     Node* in = region->in(j);
1070     if (phase->is_dominator(projs.fallthrough_catchproj, in)) {
1071       phi->init_req(j, n);
1072     } else if (phase->is_dominator(projs.catchall_catchproj, in)) {
1073       phi->init_req(j, n_clone);
1074     } else {
1075       phi->init_req(j, create_phis_on_call_return(ctrl, in, n, n_clone, projs, phase));
1076     }
1077   }
1078   phase->register_new_node(phi, region);
1079   return phi;
1080 }
1081 
1082 void ShenandoahBarrierC2Support::pin_and_expand(PhaseIdealLoop* phase) {
1083   ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
1084 
1085   Unique_Node_List uses;
1086   Node_Stack stack(0);
1087   Node_List clones;
1088   for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1089     ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1090 
1091     Node* ctrl = phase->get_ctrl(lrb);
1092     Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1093 
1094     CallStaticJavaNode* unc = nullptr;
1095     Node* unc_ctrl = nullptr;
1096     Node* uncasted_val = val;
1097 
1098     for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) {
1099       Node* u = lrb->fast_out(i);
1100       if (u->Opcode() == Op_CastPP &&
1101           u->in(0) != nullptr &&
1102           phase->is_dominator(u->in(0), ctrl)) {
1103         const Type* u_t = phase->igvn().type(u);
1104 
1105         if (u_t->meet(TypePtr::NULL_PTR) != u_t &&
1106             u->in(0)->Opcode() == Op_IfTrue &&
1107             u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
1108             u->in(0)->in(0)->is_If() &&
1109             u->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
1110             u->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
1111             u->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
1112             u->in(0)->in(0)->in(1)->in(1)->in(1) == val &&
1113             u->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
1114           IdealLoopTree* loop = phase->get_loop(ctrl);
1115           IdealLoopTree* unc_loop = phase->get_loop(u->in(0));
1116 
1117           if (!unc_loop->is_member(loop)) {
1118             continue;
1119           }
1120 
1121           Node* branch = no_branches(ctrl, u->in(0), false, phase);
1122           assert(branch == nullptr || branch == NodeSentinel, "was not looking for a branch");
1123           if (branch == NodeSentinel) {
1124             continue;
1125           }
1126 
1127           Node* iff = u->in(0)->in(0);
1128           Node* bol = iff->in(1)->clone();
1129           Node* cmp = bol->in(1)->clone();
1130           cmp->set_req(1, lrb);
1131           bol->set_req(1, cmp);
1132           phase->igvn().replace_input_of(iff, 1, bol);
1133           phase->set_ctrl(lrb, iff->in(0));
1134           phase->register_new_node(cmp, iff->in(0));
1135           phase->register_new_node(bol, iff->in(0));
1136           break;
1137         }
1138       }
1139     }
1140     if ((ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) || ctrl->is_CallJava()) {
1141       CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_CallJava() : ctrl->as_CallJava();
1142       if (call->entry_point() == OptoRuntime::rethrow_stub()) {
1143         // The rethrow call may have too many projections to be
1144         // properly handled here. Given there's no reason for a
1145         // barrier to depend on the call, move it above the call
1146         stack.push(lrb, 0);
1147         do {
1148           Node* n = stack.node();
1149           uint idx = stack.index();
1150           if (idx < n->req()) {
1151             Node* in = n->in(idx);
1152             stack.set_index(idx+1);
1153             if (in != nullptr) {
1154               if (phase->has_ctrl(in)) {
1155                 if (phase->is_dominator(call, phase->get_ctrl(in))) {
1156 #ifdef ASSERT
1157                   for (uint i = 0; i < stack.size(); i++) {
1158                     assert(stack.node_at(i) != in, "node shouldn't have been seen yet");
1159                   }
1160 #endif
1161                   stack.push(in, 0);
1162                 }
1163               } else {
1164                 assert(phase->is_dominator(in, call->in(0)), "no dependency on the call");
1165               }
1166             }
1167           } else {
1168             phase->set_ctrl(n, call->in(0));
1169             stack.pop();
1170           }
1171         } while(stack.size() > 0);
1172         continue;
1173       }
1174       CallProjections projs;
1175       call->extract_projections(&projs, false, false);
1176 
1177 #ifdef ASSERT
1178       VectorSet cloned;
1179 #endif
1180       Node* lrb_clone = lrb->clone();
1181       phase->register_new_node(lrb_clone, projs.catchall_catchproj);
1182       phase->set_ctrl(lrb, projs.fallthrough_catchproj);
1183 
1184       stack.push(lrb, 0);
1185       clones.push(lrb_clone);
1186 
1187       do {
1188         assert(stack.size() == clones.size(), "");
1189         Node* n = stack.node();
1190 #ifdef ASSERT
1191         if (n->is_Load()) {
1192           Node* mem = n->in(MemNode::Memory);
1193           for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
1194             Node* u = mem->fast_out(j);
1195             assert(!u->is_Store() || !u->is_LoadStore() || phase->get_ctrl(u) != ctrl, "anti dependent store?");
1196           }
1197         }
1198 #endif
1199         uint idx = stack.index();
1200         Node* n_clone = clones.at(clones.size()-1);
1201         if (idx < n->outcnt()) {
1202           Node* u = n->raw_out(idx);
1203           Node* c = phase->ctrl_or_self(u);
1204           if (phase->is_dominator(call, c) && phase->is_dominator(c, projs.fallthrough_proj)) {
1205             stack.set_index(idx+1);
1206             assert(!u->is_CFG(), "");
1207             stack.push(u, 0);
1208             assert(!cloned.test_set(u->_idx), "only one clone");
1209             Node* u_clone = u->clone();
1210             int nb = u_clone->replace_edge(n, n_clone, &phase->igvn());
1211             assert(nb > 0, "should have replaced some uses");
1212             phase->register_new_node(u_clone, projs.catchall_catchproj);
1213             clones.push(u_clone);
1214             phase->set_ctrl(u, projs.fallthrough_catchproj);
1215           } else {
1216             bool replaced = false;
1217             if (u->is_Phi()) {
1218               for (uint k = 1; k < u->req(); k++) {
1219                 if (u->in(k) == n) {
1220                   if (phase->is_dominator(projs.catchall_catchproj, u->in(0)->in(k))) {
1221                     phase->igvn().replace_input_of(u, k, n_clone);
1222                     replaced = true;
1223                   } else if (!phase->is_dominator(projs.fallthrough_catchproj, u->in(0)->in(k))) {
1224                     phase->igvn().replace_input_of(u, k, create_phis_on_call_return(ctrl, u->in(0)->in(k), n, n_clone, projs, phase));
1225                     replaced = true;
1226                   }
1227                 }
1228               }
1229             } else {
1230               if (phase->is_dominator(projs.catchall_catchproj, c)) {
1231                 phase->igvn().rehash_node_delayed(u);
1232                 int nb = u->replace_edge(n, n_clone, &phase->igvn());
1233                 assert(nb > 0, "should have replaced some uses");
1234                 replaced = true;
1235               } else if (!phase->is_dominator(projs.fallthrough_catchproj, c)) {
1236                 if (u->is_If()) {
1237                   // Can't break If/Bool/Cmp chain
1238                   assert(n->is_Bool(), "unexpected If shape");
1239                   assert(stack.node_at(stack.size()-2)->is_Cmp(), "unexpected If shape");
1240                   assert(n_clone->is_Bool(), "unexpected clone");
1241                   assert(clones.at(clones.size()-2)->is_Cmp(), "unexpected clone");
1242                   Node* bol_clone = n->clone();
1243                   Node* cmp_clone = stack.node_at(stack.size()-2)->clone();
1244                   bol_clone->set_req(1, cmp_clone);
1245 
1246                   Node* nn = stack.node_at(stack.size()-3);
1247                   Node* nn_clone = clones.at(clones.size()-3);
1248                   assert(nn->Opcode() == nn_clone->Opcode(), "mismatch");
1249 
1250                   int nb = cmp_clone->replace_edge(nn, create_phis_on_call_return(ctrl, c, nn, nn_clone, projs, phase),
1251                                                    &phase->igvn());
1252                   assert(nb > 0, "should have replaced some uses");
1253 
1254                   phase->register_new_node(bol_clone, u->in(0));
1255                   phase->register_new_node(cmp_clone, u->in(0));
1256 
1257                   phase->igvn().replace_input_of(u, 1, bol_clone);
1258 
1259                 } else {
1260                   phase->igvn().rehash_node_delayed(u);
1261                   int nb = u->replace_edge(n, create_phis_on_call_return(ctrl, c, n, n_clone, projs, phase), &phase->igvn());
1262                   assert(nb > 0, "should have replaced some uses");
1263                 }
1264                 replaced = true;
1265               }
1266             }
1267             if (!replaced) {
1268               stack.set_index(idx+1);
1269             }
1270           }
1271         } else {
1272           stack.pop();
1273           clones.pop();
1274         }
1275       } while (stack.size() > 0);
1276       assert(stack.size() == 0 && clones.size() == 0, "");
1277     }
1278   }
1279 
1280   for (int i = 0; i < state->load_reference_barriers_count(); i++) {
1281     ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1282     Node* ctrl = phase->get_ctrl(lrb);
1283     IdealLoopTree* loop = phase->get_loop(ctrl);
1284     Node* head = loop->head();
1285     if (head->is_OuterStripMinedLoop()) {
1286       // Expanding a barrier here will break loop strip mining
1287       // verification. Transform the loop so the loop nest doesn't
1288       // appear as strip mined.
1289       OuterStripMinedLoopNode* outer = head->as_OuterStripMinedLoop();
1290       hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1291     }
1292     if (head->is_BaseCountedLoop() && ctrl->is_IfProj() && ctrl->in(0)->is_BaseCountedLoopEnd() &&
1293         head->as_BaseCountedLoop()->loopexit() == ctrl->in(0)) {
1294       Node* entry = head->in(LoopNode::EntryControl);
1295       Node* backedge = head->in(LoopNode::LoopBackControl);
1296       Node* new_head = new LoopNode(entry, backedge);
1297       phase->register_control(new_head, phase->get_loop(entry), entry);
1298       phase->lazy_replace(head, new_head);
1299     }
1300   }
1301 
1302   // Expand load-reference-barriers
1303   MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
1304   Unique_Node_List uses_to_ignore;
1305   for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1306     ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1307     uint last = phase->C->unique();
1308     Node* ctrl = phase->get_ctrl(lrb);
1309     Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1310 
1311     Node* orig_ctrl = ctrl;
1312 
1313     Node* raw_mem = fixer.find_mem(ctrl, lrb);
1314     Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, nullptr);
1315 
1316     IdealLoopTree *loop = phase->get_loop(ctrl);
1317 
1318     Node* heap_stable_ctrl = nullptr;
1319     Node* null_ctrl = nullptr;
1320 
1321     assert(val->bottom_type()->make_oopptr(), "need oop");
1322     assert(val->bottom_type()->make_oopptr()->const_oop() == nullptr, "expect non-constant");
1323 
1324     enum { _heap_stable = 1, _evac_path, _not_cset, PATH_LIMIT };
1325     Node* region = new RegionNode(PATH_LIMIT);
1326     Node* val_phi = new PhiNode(region, val->bottom_type()->is_oopptr());
1327 
1328     // Stable path.
1329     int flags = ShenandoahHeap::HAS_FORWARDED;
1330     if (!ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1331       flags |= ShenandoahHeap::WEAK_ROOTS;
1332     }
1333     test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, flags);
1334     IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
1335 
1336     // Heap stable case
1337     region->init_req(_heap_stable, heap_stable_ctrl);
1338     val_phi->init_req(_heap_stable, val);
1339 
1340     // Test for in-cset, unless it's a native-LRB. Native LRBs need to return null
1341     // even for non-cset objects to prevent resurrection of such objects.
1342     // Wires !in_cset(obj) to slot 2 of region and phis
1343     Node* not_cset_ctrl = nullptr;
1344     if (ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1345       test_in_cset(ctrl, not_cset_ctrl, val, raw_mem, phase);
1346     }
1347     if (not_cset_ctrl != nullptr) {
1348       region->init_req(_not_cset, not_cset_ctrl);
1349       val_phi->init_req(_not_cset, val);
1350     } else {
1351       region->del_req(_not_cset);
1352       val_phi->del_req(_not_cset);
1353     }
1354 
1355     // Resolve object when orig-value is in cset.
1356     // Make the unconditional resolve for fwdptr.
1357 
1358     // Call lrb-stub and wire up that path in slots 4
1359     Node* result_mem = nullptr;
1360 
1361     Node* addr;
1362     {
1363       VectorSet visited;
1364       addr = get_load_addr(phase, visited, lrb);
1365     }
1366     if (addr->Opcode() == Op_AddP) {
1367       Node* orig_base = addr->in(AddPNode::Base);
1368       Node* base = new CheckCastPPNode(ctrl, orig_base, orig_base->bottom_type(), ConstraintCastNode::StrongDependency);
1369       phase->register_new_node(base, ctrl);
1370       if (addr->in(AddPNode::Base) == addr->in((AddPNode::Address))) {
1371         // Field access
1372         addr = addr->clone();
1373         addr->set_req(AddPNode::Base, base);
1374         addr->set_req(AddPNode::Address, base);
1375         phase->register_new_node(addr, ctrl);
1376       } else {
1377         Node* addr2 = addr->in(AddPNode::Address);
1378         if (addr2->Opcode() == Op_AddP && addr2->in(AddPNode::Base) == addr2->in(AddPNode::Address) &&
1379               addr2->in(AddPNode::Base) == orig_base) {
1380           addr2 = addr2->clone();
1381           addr2->set_req(AddPNode::Base, base);
1382           addr2->set_req(AddPNode::Address, base);
1383           phase->register_new_node(addr2, ctrl);
1384           addr = addr->clone();
1385           addr->set_req(AddPNode::Base, base);
1386           addr->set_req(AddPNode::Address, addr2);
1387           phase->register_new_node(addr, ctrl);
1388         }
1389       }
1390     }
1391     call_lrb_stub(ctrl, val, addr, lrb->decorators(), phase);
1392     region->init_req(_evac_path, ctrl);
1393     val_phi->init_req(_evac_path, val);
1394 
1395     phase->register_control(region, loop, heap_stable_iff);
1396     Node* out_val = val_phi;
1397     phase->register_new_node(val_phi, region);
1398 
1399     fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase);
1400 
1401     ctrl = orig_ctrl;
1402 
1403     phase->igvn().replace_node(lrb, out_val);
1404 
1405     follow_barrier_uses(out_val, ctrl, uses, phase);
1406 
1407     for(uint next = 0; next < uses.size(); next++ ) {
1408       Node *n = uses.at(next);
1409       assert(phase->get_ctrl(n) == ctrl, "bad control");
1410       assert(n != raw_mem, "should leave input raw mem above the barrier");
1411       phase->set_ctrl(n, region);
1412       follow_barrier_uses(n, ctrl, uses, phase);
1413     }
1414     fixer.record_new_ctrl(ctrl, region, raw_mem, raw_mem_for_ctrl);
1415   }
1416   // Done expanding load-reference-barriers.
1417   assert(ShenandoahBarrierSetC2::bsc2()->state()->load_reference_barriers_count() == 0, "all load reference barrier nodes should have been replaced");
1418 }
1419 
1420 Node* ShenandoahBarrierC2Support::get_load_addr(PhaseIdealLoop* phase, VectorSet& visited, Node* in) {
1421   if (visited.test_set(in->_idx)) {
1422     return nullptr;
1423   }
1424   switch (in->Opcode()) {
1425     case Op_Proj:
1426       return get_load_addr(phase, visited, in->in(0));
1427     case Op_CastPP:
1428     case Op_CheckCastPP:
1429     case Op_DecodeN:
1430     case Op_EncodeP:
1431       return get_load_addr(phase, visited, in->in(1));
1432     case Op_LoadN:
1433     case Op_LoadP:
1434       return in->in(MemNode::Address);
1435     case Op_CompareAndExchangeN:
1436     case Op_CompareAndExchangeP:
1437     case Op_GetAndSetN:
1438     case Op_GetAndSetP:
1439     case Op_ShenandoahCompareAndExchangeP:
1440     case Op_ShenandoahCompareAndExchangeN:
1441       // Those instructions would just have stored a different
1442       // value into the field. No use to attempt to fix it at this point.
1443       return phase->igvn().zerocon(T_OBJECT);
1444     case Op_CMoveP:
1445     case Op_CMoveN: {
1446       Node* t = get_load_addr(phase, visited, in->in(CMoveNode::IfTrue));
1447       Node* f = get_load_addr(phase, visited, in->in(CMoveNode::IfFalse));
1448       // Handle unambiguous cases: single address reported on both branches.
1449       if (t != nullptr && f == nullptr) return t;
1450       if (t == nullptr && f != nullptr) return f;
1451       if (t != nullptr && t == f)    return t;
1452       // Ambiguity.
1453       return phase->igvn().zerocon(T_OBJECT);
1454     }
1455     case Op_Phi: {
1456       Node* addr = nullptr;
1457       for (uint i = 1; i < in->req(); i++) {
1458         Node* addr1 = get_load_addr(phase, visited, in->in(i));
1459         if (addr == nullptr) {
1460           addr = addr1;
1461         }
1462         if (addr != addr1) {
1463           return phase->igvn().zerocon(T_OBJECT);
1464         }
1465       }
1466       return addr;
1467     }
1468     case Op_ShenandoahLoadReferenceBarrier:
1469       return get_load_addr(phase, visited, in->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1470     case Op_CallDynamicJava:
1471     case Op_CallLeaf:
1472     case Op_CallStaticJava:
1473     case Op_ConN:
1474     case Op_ConP:
1475     case Op_Parm:
1476     case Op_CreateEx:
1477       return phase->igvn().zerocon(T_OBJECT);
1478     default:
1479 #ifdef ASSERT
1480       fatal("Unknown node in get_load_addr: %s", NodeClassNames[in->Opcode()]);
1481 #endif
1482       return phase->igvn().zerocon(T_OBJECT);
1483   }
1484 
1485 }
1486 
1487 #ifdef ASSERT
1488 static bool has_never_branch(Node* root) {
1489   for (uint i = 1; i < root->req(); i++) {
1490     Node* in = root->in(i);
1491     if (in != nullptr && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->is_NeverBranch()) {
1492       return true;
1493     }
1494   }
1495   return false;
1496 }
1497 #endif
1498 
1499 void MemoryGraphFixer::collect_memory_nodes() {
1500   Node_Stack stack(0);
1501   VectorSet visited;
1502   Node_List regions;
1503 
1504   // Walk the raw memory graph and create a mapping from CFG node to
1505   // memory node. Exclude phis for now.
1506   stack.push(_phase->C->root(), 1);
1507   do {
1508     Node* n = stack.node();
1509     int opc = n->Opcode();
1510     uint i = stack.index();
1511     if (i < n->req()) {
1512       Node* mem = nullptr;
1513       if (opc == Op_Root) {
1514         Node* in = n->in(i);
1515         int in_opc = in->Opcode();
1516         if (in_opc == Op_Return || in_opc == Op_Rethrow) {
1517           mem = in->in(TypeFunc::Memory);
1518         } else if (in_opc == Op_Halt) {
1519           if (in->in(0)->is_Region()) {
1520             Node* r = in->in(0);
1521             for (uint j = 1; j < r->req(); j++) {
1522               assert(!r->in(j)->is_NeverBranch(), "");
1523             }
1524           } else {
1525             Node* proj = in->in(0);
1526             assert(proj->is_Proj(), "");
1527             Node* in = proj->in(0);
1528             assert(in->is_CallStaticJava() || in->is_NeverBranch() || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
1529             if (in->is_CallStaticJava()) {
1530               mem = in->in(TypeFunc::Memory);
1531             } else if (in->Opcode() == Op_Catch) {
1532               Node* call = in->in(0)->in(0);
1533               assert(call->is_Call(), "");
1534               mem = call->in(TypeFunc::Memory);
1535             } else if (in->is_NeverBranch()) {
1536               mem = collect_memory_for_infinite_loop(in);
1537             }
1538           }
1539         } else {
1540 #ifdef ASSERT
1541           n->dump();
1542           in->dump();
1543 #endif
1544           ShouldNotReachHere();
1545         }
1546       } else {
1547         assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
1548         assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
1549         mem = n->in(i);
1550       }
1551       i++;
1552       stack.set_index(i);
1553       if (mem == nullptr) {
1554         continue;
1555       }
1556       for (;;) {
1557         if (visited.test_set(mem->_idx) || mem->is_Start()) {
1558           break;
1559         }
1560         if (mem->is_Phi()) {
1561           stack.push(mem, 2);
1562           mem = mem->in(1);
1563         } else if (mem->is_Proj()) {
1564           stack.push(mem, mem->req());
1565           mem = mem->in(0);
1566         } else if (mem->is_SafePoint() || mem->is_MemBar()) {
1567           mem = mem->in(TypeFunc::Memory);
1568         } else if (mem->is_MergeMem()) {
1569           MergeMemNode* mm = mem->as_MergeMem();
1570           mem = mm->memory_at(_alias);
1571         } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
1572           assert(_alias == Compile::AliasIdxRaw, "");
1573           stack.push(mem, mem->req());
1574           mem = mem->in(MemNode::Memory);
1575         } else {
1576 #ifdef ASSERT
1577           mem->dump();
1578 #endif
1579           ShouldNotReachHere();
1580         }
1581       }
1582     } else {
1583       if (n->is_Phi()) {
1584         // Nothing
1585       } else if (!n->is_Root()) {
1586         Node* c = get_ctrl(n);
1587         _memory_nodes.map(c->_idx, n);
1588       }
1589       stack.pop();
1590     }
1591   } while(stack.is_nonempty());
1592 
1593   // Iterate over CFG nodes in rpo and propagate memory state to
1594   // compute memory state at regions, creating new phis if needed.
1595   Node_List rpo_list;
1596   visited.clear();
1597   _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
1598   Node* root = rpo_list.pop();
1599   assert(root == _phase->C->root(), "");
1600 
1601   const bool trace = false;
1602 #ifdef ASSERT
1603   if (trace) {
1604     for (int i = rpo_list.size() - 1; i >= 0; i--) {
1605       Node* c = rpo_list.at(i);
1606       if (_memory_nodes[c->_idx] != nullptr) {
1607         tty->print("X %d", c->_idx);  _memory_nodes[c->_idx]->dump();
1608       }
1609     }
1610   }
1611 #endif
1612   uint last = _phase->C->unique();
1613 
1614 #ifdef ASSERT
1615   uint16_t max_depth = 0;
1616   for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
1617     IdealLoopTree* lpt = iter.current();
1618     max_depth = MAX2(max_depth, lpt->_nest);
1619   }
1620 #endif
1621 
1622   bool progress = true;
1623   int iteration = 0;
1624   Node_List dead_phis;
1625   while (progress) {
1626     progress = false;
1627     iteration++;
1628     assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
1629     if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
1630 
1631     for (int i = rpo_list.size() - 1; i >= 0; i--) {
1632       Node* c = rpo_list.at(i);
1633 
1634       Node* prev_mem = _memory_nodes[c->_idx];
1635       if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
1636         Node* prev_region = regions[c->_idx];
1637         Node* unique = nullptr;
1638         for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
1639           Node* m = _memory_nodes[c->in(j)->_idx];
1640           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");
1641           if (m != nullptr) {
1642             if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
1643               assert((c->is_Loop() && j == LoopNode::LoopBackControl) || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
1644               // continue
1645             } else if (unique == nullptr) {
1646               unique = m;
1647             } else if (m == unique) {
1648               // continue
1649             } else {
1650               unique = NodeSentinel;
1651             }
1652           }
1653         }
1654         assert(unique != nullptr, "empty phi???");
1655         if (unique != NodeSentinel) {
1656           if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c) {
1657             dead_phis.push(prev_region);
1658           }
1659           regions.map(c->_idx, unique);
1660         } else {
1661           Node* phi = nullptr;
1662           if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
1663             phi = prev_region;
1664             for (uint k = 1; k < c->req(); k++) {
1665               Node* m = _memory_nodes[c->in(k)->_idx];
1666               assert(m != nullptr, "expect memory state");
1667               phi->set_req(k, m);
1668             }
1669           } else {
1670             for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == nullptr; j++) {
1671               Node* u = c->fast_out(j);
1672               if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
1673                   (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
1674                 phi = u;
1675                 for (uint k = 1; k < c->req() && phi != nullptr; k++) {
1676                   Node* m = _memory_nodes[c->in(k)->_idx];
1677                   assert(m != nullptr, "expect memory state");
1678                   if (u->in(k) != m) {
1679                     phi = NodeSentinel;
1680                   }
1681                 }
1682               }
1683             }
1684             if (phi == NodeSentinel) {
1685               phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
1686               for (uint k = 1; k < c->req(); k++) {
1687                 Node* m = _memory_nodes[c->in(k)->_idx];
1688                 assert(m != nullptr, "expect memory state");
1689                 phi->init_req(k, m);
1690               }
1691             }
1692           }
1693           if (phi != nullptr) {
1694             regions.map(c->_idx, phi);
1695           } else {
1696             assert(c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
1697           }
1698         }
1699         Node* current_region = regions[c->_idx];
1700         if (current_region != prev_region) {
1701           progress = true;
1702           if (prev_region == prev_mem) {
1703             _memory_nodes.map(c->_idx, current_region);
1704           }
1705         }
1706       } else if (prev_mem == nullptr || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
1707         Node* m = _memory_nodes[_phase->idom(c)->_idx];
1708         assert(m != nullptr || c->Opcode() == Op_Halt, "expect memory state");
1709         if (m != prev_mem) {
1710           _memory_nodes.map(c->_idx, m);
1711           progress = true;
1712         }
1713       }
1714 #ifdef ASSERT
1715       if (trace) { tty->print("X %d", c->_idx);  _memory_nodes[c->_idx]->dump(); }
1716 #endif
1717     }
1718   }
1719 
1720   // Replace existing phi with computed memory state for that region
1721   // if different (could be a new phi or a dominating memory node if
1722   // that phi was found to be useless).
1723   while (dead_phis.size() > 0) {
1724     Node* n = dead_phis.pop();
1725     n->replace_by(_phase->C->top());
1726     n->destruct(&_phase->igvn());
1727   }
1728   for (int i = rpo_list.size() - 1; i >= 0; i--) {
1729     Node* c = rpo_list.at(i);
1730     if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
1731       Node* n = regions[c->_idx];
1732       assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
1733       if (n != nullptr && n->is_Phi() && n->_idx >= last && n->in(0) == c) {
1734         _phase->register_new_node(n, c);
1735       }
1736     }
1737   }
1738   for (int i = rpo_list.size() - 1; i >= 0; i--) {
1739     Node* c = rpo_list.at(i);
1740     if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
1741       Node* n = regions[c->_idx];
1742       assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
1743       for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
1744         Node* u = c->fast_out(i);
1745         if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
1746             u != n) {
1747           assert(c->unique_ctrl_out()->Opcode() != Op_Halt, "expected memory state");
1748           if (u->adr_type() == TypePtr::BOTTOM) {
1749             fix_memory_uses(u, n, n, c);
1750           } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
1751             _phase->lazy_replace(u, n);
1752             --i; --imax;
1753           }
1754         }
1755       }
1756     }
1757   }
1758 }
1759 
1760 Node* MemoryGraphFixer::collect_memory_for_infinite_loop(const Node* in) {
1761   Node* mem = nullptr;
1762   Node* head = in->in(0);
1763   assert(head->is_Region(), "unexpected infinite loop graph shape");
1764 
1765   Node* phi_mem = nullptr;
1766   for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
1767     Node* u = head->fast_out(j);
1768     if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
1769       if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
1770         assert(phi_mem == nullptr || phi_mem->adr_type() == TypePtr::BOTTOM, "");
1771         phi_mem = u;
1772       } else if (u->adr_type() == TypePtr::BOTTOM) {
1773         assert(phi_mem == nullptr || _phase->C->get_alias_index(phi_mem->adr_type()) == _alias, "");
1774         if (phi_mem == nullptr) {
1775           phi_mem = u;
1776         }
1777       }
1778     }
1779   }
1780   if (phi_mem == nullptr) {
1781     ResourceMark rm;
1782     Node_Stack stack(0);
1783     stack.push(head, 1);
1784     do {
1785       Node* n = stack.node();
1786       uint i = stack.index();
1787       if (i >= n->req()) {
1788         stack.pop();
1789       } else {
1790         stack.set_index(i + 1);
1791         Node* c = n->in(i);
1792         assert(c != head, "should have found a safepoint on the way");
1793         if (stack.size() != 1 || _phase->is_dominator(head, c)) {
1794           for (;;) {
1795             if (c->is_Region()) {
1796               stack.push(c, 1);
1797               break;
1798             } else if (c->is_SafePoint() && !c->is_CallLeaf()) {
1799               Node* m = c->in(TypeFunc::Memory);
1800               if (m->is_MergeMem()) {
1801                 m = m->as_MergeMem()->memory_at(_alias);
1802               }
1803               assert(mem == nullptr || mem == m, "several memory states");
1804               mem = m;
1805               break;
1806             } else {
1807               assert(c != c->in(0), "");
1808               c = c->in(0);
1809             }
1810           }
1811         }
1812       }
1813     } while (stack.size() > 0);
1814     assert(mem != nullptr, "should have found safepoint");
1815   } else {
1816     mem = phi_mem;
1817   }
1818   return mem;
1819 }
1820 
1821 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
1822   Node* c = _phase->get_ctrl(n);
1823   if (n->is_Proj() && n->in(0) != nullptr && n->in(0)->is_Call()) {
1824     assert(c == n->in(0), "");
1825     CallNode* call = c->as_Call();
1826     CallProjections projs;
1827     call->extract_projections(&projs, true, false);
1828     if (projs.catchall_memproj != nullptr) {
1829       if (projs.fallthrough_memproj == n) {
1830         c = projs.fallthrough_catchproj;
1831       } else {
1832         assert(projs.catchall_memproj == n, "");
1833         c = projs.catchall_catchproj;
1834       }
1835     }
1836   }
1837   return c;
1838 }
1839 
1840 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
1841   if (_phase->has_ctrl(n))
1842     return get_ctrl(n);
1843   else {
1844     assert (n->is_CFG(), "must be a CFG node");
1845     return n;
1846   }
1847 }
1848 
1849 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
1850   return m != nullptr && get_ctrl(m) == c;
1851 }
1852 
1853 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
1854   assert(n == nullptr || _phase->ctrl_or_self(n) == ctrl, "");
1855   assert(!ctrl->is_Call() || ctrl == n, "projection expected");
1856 #ifdef ASSERT
1857   if ((ctrl->is_Proj() && ctrl->in(0)->is_Call()) ||
1858       (ctrl->is_Catch() && ctrl->in(0)->in(0)->is_Call())) {
1859     CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_Call() : ctrl->in(0)->in(0)->as_Call();
1860     int mems = 0;
1861     for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
1862       Node* u = call->fast_out(i);
1863       if (u->bottom_type() == Type::MEMORY) {
1864         mems++;
1865       }
1866     }
1867     assert(mems <= 1, "No node right after call if multiple mem projections");
1868   }
1869 #endif
1870   Node* mem = _memory_nodes[ctrl->_idx];
1871   Node* c = ctrl;
1872   while (!mem_is_valid(mem, c) &&
1873          (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
1874     c = _phase->idom(c);
1875     mem = _memory_nodes[c->_idx];
1876   }
1877   if (n != nullptr && mem_is_valid(mem, c)) {
1878     while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
1879       mem = next_mem(mem, _alias);
1880     }
1881     if (mem->is_MergeMem()) {
1882       mem = mem->as_MergeMem()->memory_at(_alias);
1883     }
1884     if (!mem_is_valid(mem, c)) {
1885       do {
1886         c = _phase->idom(c);
1887         mem = _memory_nodes[c->_idx];
1888       } while (!mem_is_valid(mem, c) &&
1889                (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
1890     }
1891   }
1892   assert(mem->bottom_type() == Type::MEMORY, "");
1893   return mem;
1894 }
1895 
1896 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
1897   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1898     Node* use = region->fast_out(i);
1899     if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
1900         (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
1901       return true;
1902     }
1903   }
1904   return false;
1905 }
1906 
1907 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
1908   assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
1909   const bool trace = false;
1910   DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
1911   DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
1912   GrowableArray<Node*> phis;
1913   if (mem_for_ctrl != mem) {
1914     Node* old = mem_for_ctrl;
1915     Node* prev = nullptr;
1916     while (old != mem) {
1917       prev = old;
1918       if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
1919         assert(_alias == Compile::AliasIdxRaw, "");
1920         old = old->in(MemNode::Memory);
1921       } else if (old->Opcode() == Op_SCMemProj) {
1922         assert(_alias == Compile::AliasIdxRaw, "");
1923         old = old->in(0);
1924       } else {
1925         ShouldNotReachHere();
1926       }
1927     }
1928     assert(prev != nullptr, "");
1929     if (new_ctrl != ctrl) {
1930       _memory_nodes.map(ctrl->_idx, mem);
1931       _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
1932     }
1933     uint input = (uint)MemNode::Memory;
1934     _phase->igvn().replace_input_of(prev, input, new_mem);
1935   } else {
1936     uses.clear();
1937     _memory_nodes.map(new_ctrl->_idx, new_mem);
1938     uses.push(new_ctrl);
1939     for(uint next = 0; next < uses.size(); next++ ) {
1940       Node *n = uses.at(next);
1941       assert(n->is_CFG(), "");
1942       DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
1943       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1944         Node* u = n->fast_out(i);
1945         if (!u->is_Root() && u->is_CFG() && u != n) {
1946           Node* m = _memory_nodes[u->_idx];
1947           if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) &&
1948               !has_mem_phi(u) &&
1949               u->unique_ctrl_out()->Opcode() != Op_Halt) {
1950             DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
1951             DEBUG_ONLY(if (trace && m != nullptr) { tty->print("ZZZ mem"); m->dump(); });
1952 
1953             if (!mem_is_valid(m, u) || !m->is_Phi()) {
1954               bool push = true;
1955               bool create_phi = true;
1956               if (_phase->is_dominator(new_ctrl, u)) {
1957                 create_phi = false;
1958               }
1959               if (create_phi) {
1960                 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
1961                 _phase->register_new_node(phi, u);
1962                 phis.push(phi);
1963                 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
1964                 if (!mem_is_valid(m, u)) {
1965                   DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
1966                   _memory_nodes.map(u->_idx, phi);
1967                 } else {
1968                   DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
1969                   for (;;) {
1970                     assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), "");
1971                     Node* next = nullptr;
1972                     if (m->is_Proj()) {
1973                       next = m->in(0);
1974                     } else {
1975                       assert(m->is_Mem() || m->is_LoadStore(), "");
1976                       assert(_alias == Compile::AliasIdxRaw, "");
1977                       next = m->in(MemNode::Memory);
1978                     }
1979                     if (_phase->get_ctrl(next) != u) {
1980                       break;
1981                     }
1982                     if (next->is_MergeMem()) {
1983                       assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
1984                       break;
1985                     }
1986                     if (next->is_Phi()) {
1987                       assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
1988                       break;
1989                     }
1990                     m = next;
1991                   }
1992 
1993                   DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
1994                   assert(m->is_Mem() || m->is_LoadStore(), "");
1995                   uint input = (uint)MemNode::Memory;
1996                   _phase->igvn().replace_input_of(m, input, phi);
1997                   push = false;
1998                 }
1999               } else {
2000                 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
2001               }
2002               if (push) {
2003                 uses.push(u);
2004               }
2005             }
2006           } else if (!mem_is_valid(m, u) &&
2007                      !(u->Opcode() == Op_CProj && u->in(0)->is_NeverBranch() && u->as_Proj()->_con == 1)) {
2008             uses.push(u);
2009           }
2010         }
2011       }
2012     }
2013     for (int i = 0; i < phis.length(); i++) {
2014       Node* n = phis.at(i);
2015       Node* r = n->in(0);
2016       DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
2017       for (uint j = 1; j < n->req(); j++) {
2018         Node* m = find_mem(r->in(j), nullptr);
2019         _phase->igvn().replace_input_of(n, j, m);
2020         DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
2021       }
2022     }
2023   }
2024   uint last = _phase->C->unique();
2025   MergeMemNode* mm = nullptr;
2026   int alias = _alias;
2027   DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
2028   // Process loads first to not miss an anti-dependency: if the memory
2029   // edge of a store is updated before a load is processed then an
2030   // anti-dependency may be missed.
2031   for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2032     Node* u = mem->out(i);
2033     if (u->_idx < last && u->is_Load() && _phase->C->get_alias_index(u->adr_type()) == alias) {
2034       Node* m = find_mem(_phase->get_ctrl(u), u);
2035       if (m != mem) {
2036         DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2037         _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2038         --i;
2039       }
2040     }
2041   }
2042   for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2043     Node* u = mem->out(i);
2044     if (u->_idx < last) {
2045       if (u->is_Mem()) {
2046         if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2047           Node* m = find_mem(_phase->get_ctrl(u), u);
2048           if (m != mem) {
2049             DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2050             _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2051             --i;
2052           }
2053         }
2054       } else if (u->is_MergeMem()) {
2055         MergeMemNode* u_mm = u->as_MergeMem();
2056         if (u_mm->memory_at(alias) == mem) {
2057           MergeMemNode* newmm = nullptr;
2058           for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2059             Node* uu = u->fast_out(j);
2060             assert(!uu->is_MergeMem(), "chain of MergeMems?");
2061             if (uu->is_Phi()) {
2062               assert(uu->adr_type() == TypePtr::BOTTOM, "");
2063               Node* region = uu->in(0);
2064               int nb = 0;
2065               for (uint k = 1; k < uu->req(); k++) {
2066                 if (uu->in(k) == u) {
2067                   Node* m = find_mem(region->in(k), nullptr);
2068                   if (m != mem) {
2069                     DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
2070                     newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2071                     if (newmm != u) {
2072                       _phase->igvn().replace_input_of(uu, k, newmm);
2073                       nb++;
2074                       --jmax;
2075                     }
2076                   }
2077                 }
2078               }
2079               if (nb > 0) {
2080                 --j;
2081               }
2082             } else {
2083               Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
2084               if (m != mem) {
2085                 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
2086                 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2087                 if (newmm != u) {
2088                   _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2089                   --j, --jmax;
2090                 }
2091               }
2092             }
2093           }
2094         }
2095       } else if (u->is_Phi()) {
2096         assert(u->bottom_type() == Type::MEMORY, "what else?");
2097         if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
2098           Node* region = u->in(0);
2099           bool replaced = false;
2100           for (uint j = 1; j < u->req(); j++) {
2101             if (u->in(j) == mem) {
2102               Node* m = find_mem(region->in(j), nullptr);
2103               Node* nnew = m;
2104               if (m != mem) {
2105                 if (u->adr_type() == TypePtr::BOTTOM) {
2106                   mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
2107                   nnew = mm;
2108                 }
2109                 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
2110                 _phase->igvn().replace_input_of(u, j, nnew);
2111                 replaced = true;
2112               }
2113             }
2114           }
2115           if (replaced) {
2116             --i;
2117           }
2118         }
2119       } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2120                  u->adr_type() == nullptr) {
2121         assert(u->adr_type() != nullptr ||
2122                u->Opcode() == Op_Rethrow ||
2123                u->Opcode() == Op_Return ||
2124                u->Opcode() == Op_SafePoint ||
2125                (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2126                (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2127                u->Opcode() == Op_CallLeaf, "");
2128         Node* m = find_mem(_phase->ctrl_or_self(u), u);
2129         if (m != mem) {
2130           mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
2131           _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2132           --i;
2133         }
2134       } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2135         Node* m = find_mem(_phase->ctrl_or_self(u), u);
2136         if (m != mem) {
2137           DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2138           _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2139           --i;
2140         }
2141       } else if (u->adr_type() != TypePtr::BOTTOM &&
2142                  _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
2143         Node* m = find_mem(_phase->ctrl_or_self(u), u);
2144         assert(m != mem, "");
2145         // u is on the wrong slice...
2146         assert(u->is_ClearArray(), "");
2147         DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2148         _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2149         --i;
2150       }
2151     }
2152   }
2153 #ifdef ASSERT
2154   assert(new_mem->outcnt() > 0, "");
2155   for (int i = 0; i < phis.length(); i++) {
2156     Node* n = phis.at(i);
2157     assert(n->outcnt() > 0, "new phi must have uses now");
2158   }
2159 #endif
2160 }
2161 
2162 void MemoryGraphFixer::record_new_ctrl(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl) {
2163   if (mem_for_ctrl != mem && new_ctrl != ctrl) {
2164     _memory_nodes.map(ctrl->_idx, mem);
2165     _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2166   }
2167 }
2168 
2169 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
2170   MergeMemNode* mm = MergeMemNode::make(mem);
2171   mm->set_memory_at(_alias, rep_proj);
2172   _phase->register_new_node(mm, rep_ctrl);
2173   return mm;
2174 }
2175 
2176 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
2177   MergeMemNode* newmm = nullptr;
2178   MergeMemNode* u_mm = u->as_MergeMem();
2179   Node* c = _phase->get_ctrl(u);
2180   if (_phase->is_dominator(c, rep_ctrl)) {
2181     c = rep_ctrl;
2182   } else {
2183     assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
2184   }
2185   if (u->outcnt() == 1) {
2186     if (u->req() > (uint)_alias && u->in(_alias) == mem) {
2187       _phase->igvn().replace_input_of(u, _alias, rep_proj);
2188       --i;
2189     } else {
2190       _phase->igvn().rehash_node_delayed(u);
2191       u_mm->set_memory_at(_alias, rep_proj);
2192     }
2193     newmm = u_mm;
2194     _phase->set_ctrl_and_loop(u, c);
2195   } else {
2196     // can't simply clone u and then change one of its input because
2197     // it adds and then removes an edge which messes with the
2198     // DUIterator
2199     newmm = MergeMemNode::make(u_mm->base_memory());
2200     for (uint j = 0; j < u->req(); j++) {
2201       if (j < newmm->req()) {
2202         if (j == (uint)_alias) {
2203           newmm->set_req(j, rep_proj);
2204         } else if (newmm->in(j) != u->in(j)) {
2205           newmm->set_req(j, u->in(j));
2206         }
2207       } else if (j == (uint)_alias) {
2208         newmm->add_req(rep_proj);
2209       } else {
2210         newmm->add_req(u->in(j));
2211       }
2212     }
2213     if ((uint)_alias >= u->req()) {
2214       newmm->set_memory_at(_alias, rep_proj);
2215     }
2216     _phase->register_new_node(newmm, c);
2217   }
2218   return newmm;
2219 }
2220 
2221 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
2222   if (phi->adr_type() == TypePtr::BOTTOM) {
2223     Node* region = phi->in(0);
2224     for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
2225       Node* uu = region->fast_out(j);
2226       if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
2227         return false;
2228       }
2229     }
2230     return true;
2231   }
2232   return _phase->C->get_alias_index(phi->adr_type()) == _alias;
2233 }
2234 
2235 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
2236   uint last = _phase-> C->unique();
2237   MergeMemNode* mm = nullptr;
2238   assert(mem->bottom_type() == Type::MEMORY, "");
2239   for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2240     Node* u = mem->out(i);
2241     if (u != replacement && u->_idx < last) {
2242       if (u->is_MergeMem()) {
2243         MergeMemNode* u_mm = u->as_MergeMem();
2244         if (u_mm->memory_at(_alias) == mem) {
2245           MergeMemNode* newmm = nullptr;
2246           for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2247             Node* uu = u->fast_out(j);
2248             assert(!uu->is_MergeMem(), "chain of MergeMems?");
2249             if (uu->is_Phi()) {
2250               if (should_process_phi(uu)) {
2251                 Node* region = uu->in(0);
2252                 int nb = 0;
2253                 for (uint k = 1; k < uu->req(); k++) {
2254                   if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
2255                     if (newmm == nullptr) {
2256                       newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2257                     }
2258                     if (newmm != u) {
2259                       _phase->igvn().replace_input_of(uu, k, newmm);
2260                       nb++;
2261                       --jmax;
2262                     }
2263                   }
2264                 }
2265                 if (nb > 0) {
2266                   --j;
2267                 }
2268               }
2269             } else {
2270               if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
2271                 if (newmm == nullptr) {
2272                   newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2273                 }
2274                 if (newmm != u) {
2275                   _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2276                   --j, --jmax;
2277                 }
2278               }
2279             }
2280           }
2281         }
2282       } else if (u->is_Phi()) {
2283         assert(u->bottom_type() == Type::MEMORY, "what else?");
2284         Node* region = u->in(0);
2285         if (should_process_phi(u)) {
2286           bool replaced = false;
2287           for (uint j = 1; j < u->req(); j++) {
2288             if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
2289               Node* nnew = rep_proj;
2290               if (u->adr_type() == TypePtr::BOTTOM) {
2291                 if (mm == nullptr) {
2292                   mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2293                 }
2294                 nnew = mm;
2295               }
2296               _phase->igvn().replace_input_of(u, j, nnew);
2297               replaced = true;
2298             }
2299           }
2300           if (replaced) {
2301             --i;
2302           }
2303 
2304         }
2305       } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2306                  u->adr_type() == nullptr) {
2307         assert(u->adr_type() != nullptr ||
2308                u->Opcode() == Op_Rethrow ||
2309                u->Opcode() == Op_Return ||
2310                u->Opcode() == Op_SafePoint ||
2311                (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2312                (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2313                u->Opcode() == Op_CallLeaf, "%s", u->Name());
2314         if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2315           if (mm == nullptr) {
2316             mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2317           }
2318           _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2319           --i;
2320         }
2321       } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2322         if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2323           _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
2324           --i;
2325         }
2326       }
2327     }
2328   }
2329 }
2330 
2331 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj, DecoratorSet decorators)
2332 : Node(ctrl, obj), _decorators(decorators) {
2333   ShenandoahBarrierSetC2::bsc2()->state()->add_load_reference_barrier(this);
2334 }
2335 
2336 DecoratorSet ShenandoahLoadReferenceBarrierNode::decorators() const {
2337   return _decorators;
2338 }
2339 
2340 uint ShenandoahLoadReferenceBarrierNode::size_of() const {
2341   return sizeof(*this);
2342 }
2343 
2344 static DecoratorSet mask_decorators(DecoratorSet decorators) {
2345   return decorators & (ON_STRONG_OOP_REF | ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF | ON_UNKNOWN_OOP_REF | IN_NATIVE);
2346 }
2347 
2348 uint ShenandoahLoadReferenceBarrierNode::hash() const {
2349   uint hash = Node::hash();
2350   hash += mask_decorators(_decorators);
2351   return hash;
2352 }
2353 
2354 bool ShenandoahLoadReferenceBarrierNode::cmp( const Node &n ) const {
2355   return Node::cmp(n) && n.Opcode() == Op_ShenandoahLoadReferenceBarrier &&
2356          mask_decorators(_decorators) == mask_decorators(((const ShenandoahLoadReferenceBarrierNode&)n)._decorators);
2357 }
2358 
2359 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const {
2360   if (in(ValueIn) == nullptr || in(ValueIn)->is_top()) {
2361     return Type::TOP;
2362   }
2363   const Type* t = in(ValueIn)->bottom_type();
2364   if (t == TypePtr::NULL_PTR) {
2365     return t;
2366   }
2367 
2368   if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2369     return t;
2370   }
2371 
2372   return t->meet(TypePtr::NULL_PTR);
2373 }
2374 
2375 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseGVN* phase) const {
2376   // Either input is TOP ==> the result is TOP
2377   const Type *t2 = phase->type(in(ValueIn));
2378   if( t2 == Type::TOP ) return Type::TOP;
2379 
2380   if (t2 == TypePtr::NULL_PTR) {
2381     return t2;
2382   }
2383 
2384   if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2385     return t2;
2386   }
2387 
2388   return t2->meet(TypePtr::NULL_PTR);
2389 }
2390 
2391 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseGVN* phase) {
2392   Node* value = in(ValueIn);
2393   if (!needs_barrier(phase, value)) {
2394     return value;
2395   }
2396   return this;
2397 }
2398 
2399 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseGVN* phase, Node* n) {
2400   Unique_Node_List visited;
2401   return needs_barrier_impl(phase, n, visited);
2402 }
2403 
2404 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseGVN* phase, Node* n, Unique_Node_List &visited) {
2405   if (n == nullptr) return false;
2406   if (visited.member(n)) {
2407     return false; // Been there.
2408   }
2409   visited.push(n);
2410 
2411   if (n->is_Allocate()) {
2412     // tty->print_cr("optimize barrier on alloc");
2413     return false;
2414   }
2415   if (n->is_Call()) {
2416     // tty->print_cr("optimize barrier on call");
2417     return false;
2418   }
2419 
2420   const Type* type = phase->type(n);
2421   if (type == Type::TOP) {
2422     return false;
2423   }
2424   if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) {
2425     // tty->print_cr("optimize barrier on null");
2426     return false;
2427   }
2428   if (type->make_oopptr() && type->make_oopptr()->const_oop() != nullptr) {
2429     // tty->print_cr("optimize barrier on constant");
2430     return false;
2431   }
2432 
2433   switch (n->Opcode()) {
2434     case Op_AddP:
2435       return true; // TODO: Can refine?
2436     case Op_LoadP:
2437     case Op_ShenandoahCompareAndExchangeN:
2438     case Op_ShenandoahCompareAndExchangeP:
2439     case Op_CompareAndExchangeN:
2440     case Op_CompareAndExchangeP:
2441     case Op_GetAndSetN:
2442     case Op_GetAndSetP:
2443       return true;
2444     case Op_Phi: {
2445       for (uint i = 1; i < n->req(); i++) {
2446         if (needs_barrier_impl(phase, n->in(i), visited)) return true;
2447       }
2448       return false;
2449     }
2450     case Op_CheckCastPP:
2451     case Op_CastPP:
2452       return needs_barrier_impl(phase, n->in(1), visited);
2453     case Op_Proj:
2454       return needs_barrier_impl(phase, n->in(0), visited);
2455     case Op_ShenandoahLoadReferenceBarrier:
2456       // tty->print_cr("optimize barrier on barrier");
2457       return false;
2458     case Op_Parm:
2459       // tty->print_cr("optimize barrier on input arg");
2460       return false;
2461     case Op_DecodeN:
2462     case Op_EncodeP:
2463       return needs_barrier_impl(phase, n->in(1), visited);
2464     case Op_LoadN:
2465       return true;
2466     case Op_CMoveN:
2467     case Op_CMoveP:
2468       return needs_barrier_impl(phase, n->in(2), visited) ||
2469              needs_barrier_impl(phase, n->in(3), visited);
2470     case Op_CreateEx:
2471       return false;
2472     default:
2473       break;
2474   }
2475 #ifdef ASSERT
2476   tty->print("need barrier on?: ");
2477   tty->print_cr("ins:");
2478   n->dump(2);
2479   tty->print_cr("outs:");
2480   n->dump(-2);
2481   ShouldNotReachHere();
2482 #endif
2483   return true;
2484 }