1 /*
   2  * Copyright (c) 2015, 2019, Red Hat, Inc. All rights reserved.
   3  *
   4  * This code is free software; you can redistribute it and/or modify it
   5  * under the terms of the GNU General Public License version 2 only, as
   6  * published by the Free Software Foundation.
   7  *
   8  * This code is distributed in the hope that it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  11  * version 2 for more details (a copy is included in the LICENSE file that
  12  * accompanied this code).
  13  *
  14  * You should have received a copy of the GNU General Public License version
  15  * 2 along with this work; if not, write to the Free Software Foundation,
  16  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  17  *
  18  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  19  * or visit www.oracle.com if you need additional information or have any
  20  * questions.
  21  *
  22  */
  23 
  24 #include "precompiled.hpp"
  25 
  26 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
  27 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
  28 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
  29 #include "gc/shenandoah/shenandoahForwarding.hpp"
  30 #include "gc/shenandoah/shenandoahHeap.hpp"
  31 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
  32 #include "gc/shenandoah/shenandoahRuntime.hpp"
  33 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
  34 #include "opto/arraycopynode.hpp"
  35 #include "opto/block.hpp"
  36 #include "opto/callnode.hpp"
  37 #include "opto/castnode.hpp"
  38 #include "opto/movenode.hpp"
  39 #include "opto/phaseX.hpp"
  40 #include "opto/rootnode.hpp"
  41 #include "opto/runtime.hpp"
  42 #include "opto/subnode.hpp"
  43 
  44 bool ShenandoahBarrierC2Support::expand(Compile* C, PhaseIterGVN& igvn) {
  45   ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
  46   if ((state->enqueue_barriers_count() +
  47        state->load_reference_barriers_count()) > 0) {
  48     bool attempt_more_loopopts = ShenandoahLoopOptsAfterExpansion;
  49     C->clear_major_progress();
  50     PhaseIdealLoop ideal_loop(igvn, LoopOptsShenandoahExpand);
  51     if (C->failing()) return false;
  52     PhaseIdealLoop::verify(igvn);
  53     DEBUG_ONLY(verify_raw_mem(C->root());)
  54     if (attempt_more_loopopts) {
  55       C->set_major_progress();
  56       if (!C->optimize_loops(igvn, LoopOptsShenandoahPostExpand)) {
  57         return false;
  58       }
  59       C->clear_major_progress();
  60     }
  61   }
  62   return true;
  63 }
  64 
  65 bool ShenandoahBarrierC2Support::is_heap_state_test(Node* iff, int mask) {
  66   if (!UseShenandoahGC) {
  67     return false;
  68   }
  69   assert(iff->is_If(), "bad input");
  70   if (iff->Opcode() != Op_If) {
  71     return false;
  72   }
  73   Node* bol = iff->in(1);
  74   if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) {
  75     return false;
  76   }
  77   Node* cmp = bol->in(1);
  78   if (cmp->Opcode() != Op_CmpI) {
  79     return false;
  80   }
  81   Node* in1 = cmp->in(1);
  82   Node* in2 = cmp->in(2);
  83   if (in2->find_int_con(-1) != 0) {
  84     return false;
  85   }
  86   if (in1->Opcode() != Op_AndI) {
  87     return false;
  88   }
  89   in2 = in1->in(2);
  90   if (in2->find_int_con(-1) != mask) {
  91     return false;
  92   }
  93   in1 = in1->in(1);
  94 
  95   return is_gc_state_load(in1);
  96 }
  97 
  98 bool ShenandoahBarrierC2Support::is_heap_stable_test(Node* iff) {
  99   return is_heap_state_test(iff, ShenandoahHeap::HAS_FORWARDED);
 100 }
 101 
 102 bool ShenandoahBarrierC2Support::is_gc_state_load(Node *n) {
 103   if (!UseShenandoahGC) {
 104     return false;
 105   }
 106   if (n->Opcode() != Op_LoadB && n->Opcode() != Op_LoadUB) {
 107     return false;
 108   }
 109   Node* addp = n->in(MemNode::Address);
 110   if (!addp->is_AddP()) {
 111     return false;
 112   }
 113   Node* base = addp->in(AddPNode::Address);
 114   Node* off = addp->in(AddPNode::Offset);
 115   if (base->Opcode() != Op_ThreadLocal) {
 116     return false;
 117   }
 118   if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) {
 119     return false;
 120   }
 121   return true;
 122 }
 123 
 124 bool ShenandoahBarrierC2Support::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) {
 125   assert(phase->is_dominator(stop, start), "bad inputs");
 126   ResourceMark rm;
 127   Unique_Node_List wq;
 128   wq.push(start);
 129   for (uint next = 0; next < wq.size(); next++) {
 130     Node *m = wq.at(next);
 131     if (m == stop) {
 132       continue;
 133     }
 134     if (m->is_SafePoint() && !m->is_CallLeaf()) {
 135       return true;
 136     }
 137     if (m->is_Region()) {
 138       for (uint i = 1; i < m->req(); i++) {
 139         wq.push(m->in(i));
 140       }
 141     } else {
 142       wq.push(m->in(0));
 143     }
 144   }
 145   return false;
 146 }
 147 
 148 bool ShenandoahBarrierC2Support::try_common_gc_state_load(Node *n, PhaseIdealLoop *phase) {
 149   assert(is_gc_state_load(n), "inconsistent");
 150   Node* addp = n->in(MemNode::Address);
 151   Node* dominator = NULL;
 152   for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) {
 153     Node* u = addp->fast_out(i);
 154     assert(is_gc_state_load(u), "inconsistent");
 155     if (u != n && phase->is_dominator(u->in(0), n->in(0))) {
 156       if (dominator == NULL) {
 157         dominator = u;
 158       } else {
 159         if (phase->dom_depth(u->in(0)) < phase->dom_depth(dominator->in(0))) {
 160           dominator = u;
 161         }
 162       }
 163     }
 164   }
 165   if (dominator == NULL || has_safepoint_between(n->in(0), dominator->in(0), phase)) {
 166     return false;
 167   }
 168   phase->igvn().replace_node(n, dominator);
 169 
 170   return true;
 171 }
 172 
 173 #ifdef ASSERT
 174 bool ShenandoahBarrierC2Support::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) {
 175   assert(phis.size() == 0, "");
 176 
 177   while (true) {
 178     if (in->bottom_type() == TypePtr::NULL_PTR) {
 179       if (trace) {tty->print_cr("NULL");}
 180     } else if (!in->bottom_type()->make_ptr()->make_oopptr()) {
 181       if (trace) {tty->print_cr("Non oop");}
 182     } else if (t == ShenandoahLoad && ShenandoahOptimizeStableFinals &&
 183                in->bottom_type()->make_ptr()->isa_aryptr() &&
 184                in->bottom_type()->make_ptr()->is_aryptr()->is_stable()) {
 185       if (trace) {tty->print_cr("Stable array load");}
 186     } else {
 187       if (in->is_ConstraintCast()) {
 188         in = in->in(1);
 189         continue;
 190       } else if (in->is_AddP()) {
 191         assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access");
 192         in = in->in(AddPNode::Address);
 193         continue;
 194       } else if (in->is_Con()) {
 195         if (trace) {
 196           tty->print("Found constant");
 197           in->dump();
 198         }
 199       } else if (in->Opcode() == Op_Parm) {
 200         if (trace) {
 201           tty->print("Found argument");
 202         }
 203       } else if (in->Opcode() == Op_CreateEx) {
 204         if (trace) {
 205           tty->print("Found create-exception");
 206         }
 207       } else if (in->Opcode() == Op_LoadP && in->adr_type() == TypeRawPtr::BOTTOM) {
 208         if (trace) {
 209           tty->print("Found raw LoadP (OSR argument?)");
 210         }
 211       } else if (in->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 212         if (t == ShenandoahOopStore) {
 213           uint i = 0;
 214           for (; i < phis.size(); i++) {
 215             Node* n = phis.node_at(i);
 216             if (n->Opcode() == Op_ShenandoahEnqueueBarrier) {
 217               break;
 218             }
 219           }
 220           if (i == phis.size()) {
 221             return false;
 222           }
 223         }
 224         barriers_used.push(in);
 225         if (trace) {tty->print("Found barrier"); in->dump();}
 226       } else if (in->Opcode() == Op_ShenandoahEnqueueBarrier) {
 227         if (t != ShenandoahOopStore) {
 228           in = in->in(1);
 229           continue;
 230         }
 231         if (trace) {tty->print("Found enqueue barrier"); in->dump();}
 232         phis.push(in, in->req());
 233         in = in->in(1);
 234         continue;
 235       } else if (in->is_Proj() && in->in(0)->is_Allocate()) {
 236         if (trace) {
 237           tty->print("Found alloc");
 238           in->in(0)->dump();
 239         }
 240       } else if (in->is_Proj() && (in->in(0)->Opcode() == Op_CallStaticJava || in->in(0)->Opcode() == Op_CallDynamicJava)) {
 241         if (trace) {
 242           tty->print("Found Java call");
 243         }
 244       } else if (in->is_Phi()) {
 245         if (!visited.test_set(in->_idx)) {
 246           if (trace) {tty->print("Pushed phi:"); in->dump();}
 247           phis.push(in, 2);
 248           in = in->in(1);
 249           continue;
 250         }
 251         if (trace) {tty->print("Already seen phi:"); in->dump();}
 252       } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) {
 253         if (!visited.test_set(in->_idx)) {
 254           if (trace) {tty->print("Pushed cmovep:"); in->dump();}
 255           phis.push(in, CMoveNode::IfTrue);
 256           in = in->in(CMoveNode::IfFalse);
 257           continue;
 258         }
 259         if (trace) {tty->print("Already seen cmovep:"); in->dump();}
 260       } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) {
 261         in = in->in(1);
 262         continue;
 263       } else {
 264         return false;
 265       }
 266     }
 267     bool cont = false;
 268     while (phis.is_nonempty()) {
 269       uint idx = phis.index();
 270       Node* phi = phis.node();
 271       if (idx >= phi->req()) {
 272         if (trace) {tty->print("Popped phi:"); phi->dump();}
 273         phis.pop();
 274         continue;
 275       }
 276       if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();}
 277       in = phi->in(idx);
 278       phis.set_index(idx+1);
 279       cont = true;
 280       break;
 281     }
 282     if (!cont) {
 283       break;
 284     }
 285   }
 286   return true;
 287 }
 288 
 289 void ShenandoahBarrierC2Support::report_verify_failure(const char* msg, Node* n1, Node* n2) {
 290   if (n1 != NULL) {
 291     n1->dump(+10);
 292   }
 293   if (n2 != NULL) {
 294     n2->dump(+10);
 295   }
 296   fatal("%s", msg);
 297 }
 298 
 299 void ShenandoahBarrierC2Support::verify(RootNode* root) {
 300   ResourceMark rm;
 301   Unique_Node_List wq;
 302   GrowableArray<Node*> barriers;
 303   Unique_Node_List barriers_used;
 304   Node_Stack phis(0);
 305   VectorSet visited(Thread::current()->resource_area());
 306   const bool trace = false;
 307   const bool verify_no_useless_barrier = false;
 308 
 309   wq.push(root);
 310   for (uint next = 0; next < wq.size(); next++) {
 311     Node *n = wq.at(next);
 312     if (n->is_Load()) {
 313       const bool trace = false;
 314       if (trace) {tty->print("Verifying"); n->dump();}
 315       if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) {
 316         if (trace) {tty->print_cr("Load range/klass");}
 317       } else {
 318         const TypePtr* adr_type = n->as_Load()->adr_type();
 319 
 320         if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) {
 321           if (trace) {tty->print_cr("Mark load");}
 322         } else if (adr_type->isa_instptr() &&
 323                    adr_type->is_instptr()->klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
 324                    adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset) {
 325           if (trace) {tty->print_cr("Reference.get()");}
 326         } else {
 327           bool verify = true;
 328           if (adr_type->isa_instptr()) {
 329             const TypeInstPtr* tinst = adr_type->is_instptr();
 330             ciKlass* k = tinst->klass();
 331             assert(k->is_instance_klass(), "");
 332             ciInstanceKlass* ik = (ciInstanceKlass*)k;
 333             int offset = adr_type->offset();
 334 
 335             if ((ik->debug_final_field_at(offset) && ShenandoahOptimizeInstanceFinals) ||
 336                 (ik->debug_stable_field_at(offset) && ShenandoahOptimizeStableFinals)) {
 337               if (trace) {tty->print_cr("Final/stable");}
 338               verify = false;
 339             } else if (k == ciEnv::current()->Class_klass() &&
 340                        tinst->const_oop() != NULL &&
 341                        tinst->offset() >= (ik->size_helper() * wordSize)) {
 342               ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
 343               ciField* field = k->get_field_by_offset(tinst->offset(), true);
 344               if ((ShenandoahOptimizeStaticFinals && field->is_final()) ||
 345                   (ShenandoahOptimizeStableFinals && field->is_stable())) {
 346                 verify = false;
 347               }
 348             }
 349           }
 350 
 351           if (verify && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) {
 352             report_verify_failure("Shenandoah verification: Load should have barriers", n);
 353           }
 354         }
 355       }
 356     } else if (n->is_Store()) {
 357       const bool trace = false;
 358 
 359       if (trace) {tty->print("Verifying"); n->dump();}
 360       if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) {
 361         Node* adr = n->in(MemNode::Address);
 362         bool verify = true;
 363 
 364         if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) {
 365           adr = adr->in(AddPNode::Address);
 366           if (adr->is_AddP()) {
 367             assert(adr->in(AddPNode::Base)->is_top(), "");
 368             adr = adr->in(AddPNode::Address);
 369             if (adr->Opcode() == Op_LoadP &&
 370                 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() &&
 371                 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
 372                 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) {
 373               if (trace) {tty->print_cr("SATB prebarrier");}
 374               verify = false;
 375             }
 376           }
 377         }
 378 
 379         if (verify && !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahStoreValEnqueueBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
 380           report_verify_failure("Shenandoah verification: Store should have barriers", n);
 381         }
 382       }
 383       if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
 384         report_verify_failure("Shenandoah verification: Store (address) should have barriers", n);
 385       }
 386     } else if (n->Opcode() == Op_CmpP) {
 387       const bool trace = false;
 388 
 389       Node* in1 = n->in(1);
 390       Node* in2 = n->in(2);
 391       if (in1->bottom_type()->isa_oopptr()) {
 392         if (trace) {tty->print("Verifying"); n->dump();}
 393 
 394         bool mark_inputs = false;
 395         if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR ||
 396             (in1->is_Con() || in2->is_Con())) {
 397           if (trace) {tty->print_cr("Comparison against a constant");}
 398           mark_inputs = true;
 399         } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) ||
 400                    (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) {
 401           if (trace) {tty->print_cr("Comparison with newly alloc'ed object");}
 402           mark_inputs = true;
 403         } else {
 404           assert(in2->bottom_type()->isa_oopptr(), "");
 405 
 406           if (!verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) ||
 407               !verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) {
 408             report_verify_failure("Shenandoah verification: Cmp should have barriers", n);
 409           }
 410         }
 411         if (verify_no_useless_barrier &&
 412             mark_inputs &&
 413             (!verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) ||
 414              !verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) {
 415           phis.clear();
 416           visited.Reset();
 417         }
 418       }
 419     } else if (n->is_LoadStore()) {
 420       if (n->in(MemNode::ValueIn)->bottom_type()->make_ptr() &&
 421           !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahStoreValEnqueueBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
 422         report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n);
 423       }
 424 
 425       if (n->in(MemNode::Address)->bottom_type()->make_oopptr() && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
 426         report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n);
 427       }
 428     } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) {
 429       CallNode* call = n->as_Call();
 430 
 431       static struct {
 432         const char* name;
 433         struct {
 434           int pos;
 435           verify_type t;
 436         } args[6];
 437       } calls[] = {
 438         "aescrypt_encryptBlock",
 439         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 440           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 441         "aescrypt_decryptBlock",
 442         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 443           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 444         "multiplyToLen",
 445         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },   { TypeFunc::Parms+4, ShenandoahStore },
 446           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 447         "squareToLen",
 448         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },   { -1,  ShenandoahNone},
 449           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 450         "montgomery_multiply",
 451         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },
 452           { TypeFunc::Parms+6, ShenandoahStore }, { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 453         "montgomery_square",
 454         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahLoad },   { TypeFunc::Parms+5, ShenandoahStore },
 455           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 456         "mulAdd",
 457         { { TypeFunc::Parms, ShenandoahStore },  { TypeFunc::Parms+1, ShenandoahLoad },   { -1,  ShenandoahNone},
 458           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 459         "vectorizedMismatch",
 460         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahLoad },   { -1,  ShenandoahNone},
 461           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 462         "updateBytesCRC32",
 463         { { TypeFunc::Parms+1, ShenandoahLoad }, { -1,  ShenandoahNone},                  { -1,  ShenandoahNone},
 464           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 465         "updateBytesAdler32",
 466         { { TypeFunc::Parms+1, ShenandoahLoad }, { -1,  ShenandoahNone},                  { -1,  ShenandoahNone},
 467           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 468         "updateBytesCRC32C",
 469         { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad},    { -1,  ShenandoahNone},
 470           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 471         "counterMode_AESCrypt",
 472         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 473           { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } },
 474         "cipherBlockChaining_encryptAESCrypt",
 475         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 476           { TypeFunc::Parms+3, ShenandoahLoad },  { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 477         "cipherBlockChaining_decryptAESCrypt",
 478         { { TypeFunc::Parms, ShenandoahLoad },   { TypeFunc::Parms+1, ShenandoahStore },  { TypeFunc::Parms+2, ShenandoahLoad },
 479           { TypeFunc::Parms+3, ShenandoahLoad },  { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 480         "shenandoah_clone_barrier",
 481         { { TypeFunc::Parms, ShenandoahLoad },   { -1,  ShenandoahNone},                  { -1,  ShenandoahNone},
 482           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 483         "ghash_processBlocks",
 484         { { TypeFunc::Parms, ShenandoahStore },  { TypeFunc::Parms+1, ShenandoahLoad },   { TypeFunc::Parms+2, ShenandoahLoad },
 485           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 486         "sha1_implCompress",
 487         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 488           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 489         "sha256_implCompress",
 490         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 491           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 492         "sha512_implCompress",
 493         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 494           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 495         "sha1_implCompressMB",
 496         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 497           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 498         "sha256_implCompressMB",
 499         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 500           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 501         "sha512_implCompressMB",
 502         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+1, ShenandoahStore },   { -1, ShenandoahNone },
 503           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 504         "encodeBlock",
 505         { { TypeFunc::Parms, ShenandoahLoad },  { TypeFunc::Parms+3, ShenandoahStore },   { -1, ShenandoahNone },
 506           { -1,  ShenandoahNone},                 { -1,  ShenandoahNone},                 { -1,  ShenandoahNone} },
 507       };
 508 
 509       if (call->is_call_to_arraycopystub()) {
 510         Node* dest = NULL;
 511         const TypeTuple* args = n->as_Call()->_tf->domain();
 512         for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
 513           if (args->field_at(i)->isa_ptr()) {
 514             j++;
 515             if (j == 2) {
 516               dest = n->in(i);
 517               break;
 518             }
 519           }
 520         }
 521         if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) ||
 522             !verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) {
 523           report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n);
 524         }
 525       } else if (strlen(call->_name) > 5 &&
 526                  !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) {
 527         if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) {
 528           report_verify_failure("Shenandoah verification: _fill should have barriers", n);
 529         }
 530       } else if (!strcmp(call->_name, "shenandoah_wb_pre")) {
 531         // skip
 532       } else {
 533         const int calls_len = sizeof(calls) / sizeof(calls[0]);
 534         int i = 0;
 535         for (; i < calls_len; i++) {
 536           if (!strcmp(calls[i].name, call->_name)) {
 537             break;
 538           }
 539         }
 540         if (i != calls_len) {
 541           const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]);
 542           for (uint j = 0; j < args_len; j++) {
 543             int pos = calls[i].args[j].pos;
 544             if (pos == -1) {
 545               break;
 546             }
 547             if (!verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) {
 548               report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
 549             }
 550           }
 551           for (uint j = TypeFunc::Parms; j < call->req(); j++) {
 552             if (call->in(j)->bottom_type()->make_ptr() &&
 553                 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
 554               uint k = 0;
 555               for (; k < args_len && calls[i].args[k].pos != (int)j; k++);
 556               if (k == args_len) {
 557                 fatal("arg %d for call %s not covered", j, call->_name);
 558               }
 559             }
 560           }
 561         } else {
 562           for (uint j = TypeFunc::Parms; j < call->req(); j++) {
 563             if (call->in(j)->bottom_type()->make_ptr() &&
 564                 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
 565               fatal("%s not covered", call->_name);
 566             }
 567           }
 568         }
 569       }
 570     } else if (n->Opcode() == Op_ShenandoahEnqueueBarrier || n->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
 571       // skip
 572     } else if (n->is_AddP()
 573                || n->is_Phi()
 574                || n->is_ConstraintCast()
 575                || n->Opcode() == Op_Return
 576                || n->Opcode() == Op_CMoveP
 577                || n->Opcode() == Op_CMoveN
 578                || n->Opcode() == Op_Rethrow
 579                || n->is_MemBar()
 580                || n->Opcode() == Op_Conv2B
 581                || n->Opcode() == Op_SafePoint
 582                || n->is_CallJava()
 583                || n->Opcode() == Op_Unlock
 584                || n->Opcode() == Op_EncodeP
 585                || n->Opcode() == Op_DecodeN) {
 586       // nothing to do
 587     } else {
 588       static struct {
 589         int opcode;
 590         struct {
 591           int pos;
 592           verify_type t;
 593         } inputs[2];
 594       } others[] = {
 595         Op_FastLock,
 596         { { 1, ShenandoahLoad },                  { -1, ShenandoahNone} },
 597         Op_Lock,
 598         { { TypeFunc::Parms, ShenandoahLoad },    { -1, ShenandoahNone} },
 599         Op_ArrayCopy,
 600         { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } },
 601         Op_StrCompressedCopy,
 602         { { 2, ShenandoahLoad },                  { 3, ShenandoahStore } },
 603         Op_StrInflatedCopy,
 604         { { 2, ShenandoahLoad },                  { 3, ShenandoahStore } },
 605         Op_AryEq,
 606         { { 2, ShenandoahLoad },                  { 3, ShenandoahLoad } },
 607         Op_StrIndexOf,
 608         { { 2, ShenandoahLoad },                  { 4, ShenandoahLoad } },
 609         Op_StrComp,
 610         { { 2, ShenandoahLoad },                  { 4, ShenandoahLoad } },
 611         Op_StrEquals,
 612         { { 2, ShenandoahLoad },                  { 3, ShenandoahLoad } },
 613         Op_EncodeISOArray,
 614         { { 2, ShenandoahLoad },                  { 3, ShenandoahStore } },
 615         Op_HasNegatives,
 616         { { 2, ShenandoahLoad },                  { -1, ShenandoahNone} },
 617         Op_CastP2X,
 618         { { 1, ShenandoahLoad },                  { -1, ShenandoahNone} },
 619         Op_StrIndexOfChar,
 620         { { 2, ShenandoahLoad },                  { -1, ShenandoahNone } },
 621       };
 622 
 623       const int others_len = sizeof(others) / sizeof(others[0]);
 624       int i = 0;
 625       for (; i < others_len; i++) {
 626         if (others[i].opcode == n->Opcode()) {
 627           break;
 628         }
 629       }
 630       uint stop = n->is_Call() ? n->as_Call()->tf()->domain()->cnt() : n->req();
 631       if (i != others_len) {
 632         const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]);
 633         for (uint j = 0; j < inputs_len; j++) {
 634           int pos = others[i].inputs[j].pos;
 635           if (pos == -1) {
 636             break;
 637           }
 638           if (!verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) {
 639             report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
 640           }
 641         }
 642         for (uint j = 1; j < stop; j++) {
 643           if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() &&
 644               n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
 645             uint k = 0;
 646             for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++);
 647             if (k == inputs_len) {
 648               fatal("arg %d for node %s not covered", j, n->Name());
 649             }
 650           }
 651         }
 652       } else {
 653         for (uint j = 1; j < stop; j++) {
 654           if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() &&
 655               n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
 656             fatal("%s not covered", n->Name());
 657           }
 658         }
 659       }
 660     }
 661 
 662     if (n->is_SafePoint()) {
 663       SafePointNode* sfpt = n->as_SafePoint();
 664       if (verify_no_useless_barrier && sfpt->jvms() != NULL) {
 665         for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) {
 666           if (!verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) {
 667             phis.clear();
 668             visited.Reset();
 669           }
 670         }
 671       }
 672     }
 673     for( uint i = 0; i < n->len(); ++i ) {
 674       Node *m = n->in(i);
 675       if (m == NULL) continue;
 676 
 677       // In most cases, inputs should be known to be non null. If it's
 678       // not the case, it could be a missing cast_not_null() in an
 679       // intrinsic or support might be needed in AddPNode::Ideal() to
 680       // avoid a NULL+offset input.
 681       if (!(n->is_Phi() ||
 682             (n->is_SafePoint() && (!n->is_CallRuntime() || !strcmp(n->as_Call()->_name, "shenandoah_wb_pre") || !strcmp(n->as_Call()->_name, "unsafe_arraycopy"))) ||
 683             n->Opcode() == Op_CmpP ||
 684             n->Opcode() == Op_CmpN ||
 685             (n->Opcode() == Op_StoreP && i == StoreNode::ValueIn) ||
 686             (n->Opcode() == Op_StoreN && i == StoreNode::ValueIn) ||
 687             n->is_ConstraintCast() ||
 688             n->Opcode() == Op_Return ||
 689             n->Opcode() == Op_Conv2B ||
 690             n->is_AddP() ||
 691             n->Opcode() == Op_CMoveP ||
 692             n->Opcode() == Op_CMoveN ||
 693             n->Opcode() == Op_Rethrow ||
 694             n->is_MemBar() ||
 695             n->is_Mem() ||
 696             n->Opcode() == Op_AryEq ||
 697             n->Opcode() == Op_SCMemProj ||
 698             n->Opcode() == Op_EncodeP ||
 699             n->Opcode() == Op_DecodeN ||
 700             n->Opcode() == Op_ShenandoahEnqueueBarrier ||
 701             n->Opcode() == Op_ShenandoahLoadReferenceBarrier)) {
 702         if (m->bottom_type()->make_oopptr() && m->bottom_type()->make_oopptr()->meet(TypePtr::NULL_PTR) == m->bottom_type()) {
 703           report_verify_failure("Shenandoah verification: null input", n, m);
 704         }
 705       }
 706 
 707       wq.push(m);
 708     }
 709   }
 710 
 711   if (verify_no_useless_barrier) {
 712     for (int i = 0; i < barriers.length(); i++) {
 713       Node* n = barriers.at(i);
 714       if (!barriers_used.member(n)) {
 715         tty->print("XXX useless barrier"); n->dump(-2);
 716         ShouldNotReachHere();
 717       }
 718     }
 719   }
 720 }
 721 #endif
 722 
 723 bool ShenandoahBarrierC2Support::is_dominator_same_ctrl(Node* c, Node* d, Node* n, PhaseIdealLoop* phase) {
 724   // That both nodes have the same control is not sufficient to prove
 725   // domination, verify that there's no path from d to n
 726   ResourceMark rm;
 727   Unique_Node_List wq;
 728   wq.push(d);
 729   for (uint next = 0; next < wq.size(); next++) {
 730     Node *m = wq.at(next);
 731     if (m == n) {
 732       return false;
 733     }
 734     if (m->is_Phi() && m->in(0)->is_Loop()) {
 735       assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control");
 736     } else {
 737       for (uint i = 0; i < m->req(); i++) {
 738         if (m->in(i) != NULL && phase->ctrl_or_self(m->in(i)) == c) {
 739           wq.push(m->in(i));
 740         }
 741       }
 742     }
 743   }
 744   return true;
 745 }
 746 
 747 bool ShenandoahBarrierC2Support::is_dominator(Node* d_c, Node* n_c, Node* d, Node* n, PhaseIdealLoop* phase) {
 748   if (d_c != n_c) {
 749     return phase->is_dominator(d_c, n_c);
 750   }
 751   return is_dominator_same_ctrl(d_c, d, n, phase);
 752 }
 753 
 754 Node* next_mem(Node* mem, int alias) {
 755   Node* res = NULL;
 756   if (mem->is_Proj()) {
 757     res = mem->in(0);
 758   } else if (mem->is_SafePoint() || mem->is_MemBar()) {
 759     res = mem->in(TypeFunc::Memory);
 760   } else if (mem->is_Phi()) {
 761     res = mem->in(1);
 762   } else if (mem->is_MergeMem()) {
 763     res = mem->as_MergeMem()->memory_at(alias);
 764   } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
 765     assert(alias = Compile::AliasIdxRaw, "following raw memory can't lead to a barrier");
 766     res = mem->in(MemNode::Memory);
 767   } else {
 768 #ifdef ASSERT
 769     mem->dump();
 770 #endif
 771     ShouldNotReachHere();
 772   }
 773   return res;
 774 }
 775 
 776 Node* ShenandoahBarrierC2Support::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) {
 777   Node* iffproj = NULL;
 778   while (c != dom) {
 779     Node* next = phase->idom(c);
 780     assert(next->unique_ctrl_out() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?");
 781     if (c->is_Region()) {
 782       ResourceMark rm;
 783       Unique_Node_List wq;
 784       wq.push(c);
 785       for (uint i = 0; i < wq.size(); i++) {
 786         Node *n = wq.at(i);
 787         if (n == next) {
 788           continue;
 789         }
 790         if (n->is_Region()) {
 791           for (uint j = 1; j < n->req(); j++) {
 792             wq.push(n->in(j));
 793           }
 794         } else {
 795           wq.push(n->in(0));
 796         }
 797       }
 798       for (uint i = 0; i < wq.size(); i++) {
 799         Node *n = wq.at(i);
 800         assert(n->is_CFG(), "");
 801         if (n->is_Multi()) {
 802           for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
 803             Node* u = n->fast_out(j);
 804             if (u->is_CFG()) {
 805               if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
 806                 return NodeSentinel;
 807               }
 808             }
 809           }
 810         }
 811       }
 812     } else  if (c->is_Proj()) {
 813       if (c->is_IfProj()) {
 814         if (c->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) != NULL) {
 815           // continue;
 816         } else {
 817           if (!allow_one_proj) {
 818             return NodeSentinel;
 819           }
 820           if (iffproj == NULL) {
 821             iffproj = c;
 822           } else {
 823             return NodeSentinel;
 824           }
 825         }
 826       } else if (c->Opcode() == Op_JumpProj) {
 827         return NodeSentinel; // unsupported
 828       } else if (c->Opcode() == Op_CatchProj) {
 829         return NodeSentinel; // unsupported
 830       } else if (c->Opcode() == Op_CProj && next->Opcode() == Op_NeverBranch) {
 831         return NodeSentinel; // unsupported
 832       } else {
 833         assert(next->unique_ctrl_out() == c, "unsupported branch pattern");
 834       }
 835     }
 836     c = next;
 837   }
 838   return iffproj;
 839 }
 840 
 841 Node* ShenandoahBarrierC2Support::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) {
 842   ResourceMark rm;
 843   VectorSet wq(Thread::current()->resource_area());
 844   wq.set(mem->_idx);
 845   mem_ctrl = phase->ctrl_or_self(mem);
 846   while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) {
 847     mem = next_mem(mem, alias);
 848     if (wq.test_set(mem->_idx)) {
 849       return NULL;
 850     }
 851     mem_ctrl = phase->ctrl_or_self(mem);
 852   }
 853   if (mem->is_MergeMem()) {
 854     mem = mem->as_MergeMem()->memory_at(alias);
 855     mem_ctrl = phase->ctrl_or_self(mem);
 856   }
 857   return mem;
 858 }
 859 
 860 Node* ShenandoahBarrierC2Support::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) {
 861   Node* mem = NULL;
 862   Node* c = ctrl;
 863   do {
 864     if (c->is_Region()) {
 865       Node* phi_bottom = NULL;
 866       for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == NULL; i++) {
 867         Node* u = c->fast_out(i);
 868         if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
 869           if (u->adr_type() == TypePtr::BOTTOM) {
 870             mem = u;
 871           }
 872         }
 873       }
 874     } else {
 875       if (c->is_Call() && c->as_Call()->adr_type() != NULL) {
 876         CallProjections projs;
 877         c->as_Call()->extract_projections(&projs, true, false);
 878         if (projs.fallthrough_memproj != NULL) {
 879           if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) {
 880             if (projs.catchall_memproj == NULL) {
 881               mem = projs.fallthrough_memproj;
 882             } else {
 883               if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) {
 884                 mem = projs.fallthrough_memproj;
 885               } else {
 886                 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier");
 887                 mem = projs.catchall_memproj;
 888               }
 889             }
 890           }
 891         } else {
 892           Node* proj = c->as_Call()->proj_out(TypeFunc::Memory);
 893           if (proj != NULL &&
 894               proj->adr_type() == TypePtr::BOTTOM) {
 895             mem = proj;
 896           }
 897         }
 898       } else {
 899         for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
 900           Node* u = c->fast_out(i);
 901           if (u->is_Proj() &&
 902               u->bottom_type() == Type::MEMORY &&
 903               u->adr_type() == TypePtr::BOTTOM) {
 904               assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), "");
 905               assert(mem == NULL, "only one proj");
 906               mem = u;
 907           }
 908         }
 909         assert(!c->is_Call() || c->as_Call()->adr_type() != NULL || mem == NULL, "no mem projection expected");
 910       }
 911     }
 912     c = phase->idom(c);
 913   } while (mem == NULL);
 914   return mem;
 915 }
 916 
 917 void ShenandoahBarrierC2Support::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) {
 918   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 919     Node* u = n->fast_out(i);
 920     if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) {
 921       uses.push(u);
 922     }
 923   }
 924 }
 925 
 926 static void hide_strip_mined_loop(OuterStripMinedLoopNode* outer, CountedLoopNode* inner, PhaseIdealLoop* phase) {
 927   OuterStripMinedLoopEndNode* le = inner->outer_loop_end();
 928   Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl));
 929   phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl));
 930   Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt);
 931   phase->register_control(new_le, phase->get_loop(le), le->in(0));
 932   phase->lazy_replace(outer, new_outer);
 933   phase->lazy_replace(le, new_le);
 934   inner->clear_strip_mined();
 935 }
 936 
 937 void ShenandoahBarrierC2Support::test_heap_stable(Node*& ctrl, Node* raw_mem, Node*& heap_stable_ctrl,
 938                                                   PhaseIdealLoop* phase) {
 939   IdealLoopTree* loop = phase->get_loop(ctrl);
 940   Node* thread = new ThreadLocalNode();
 941   phase->register_new_node(thread, ctrl);
 942   Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
 943   phase->set_ctrl(offset, phase->C->root());
 944   Node* gc_state_addr = new AddPNode(phase->C->top(), thread, offset);
 945   phase->register_new_node(gc_state_addr, ctrl);
 946   uint gc_state_idx = Compile::AliasIdxRaw;
 947   const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument
 948   debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
 949 
 950   Node* gc_state = new LoadBNode(ctrl, raw_mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered);
 951   phase->register_new_node(gc_state, ctrl);
 952   Node* heap_stable_and = new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::HAS_FORWARDED));
 953   phase->register_new_node(heap_stable_and, ctrl);
 954   Node* heap_stable_cmp = new CmpINode(heap_stable_and, phase->igvn().zerocon(T_INT));
 955   phase->register_new_node(heap_stable_cmp, ctrl);
 956   Node* heap_stable_test = new BoolNode(heap_stable_cmp, BoolTest::ne);
 957   phase->register_new_node(heap_stable_test, ctrl);
 958   IfNode* heap_stable_iff = new IfNode(ctrl, heap_stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
 959   phase->register_control(heap_stable_iff, loop, ctrl);
 960 
 961   heap_stable_ctrl = new IfFalseNode(heap_stable_iff);
 962   phase->register_control(heap_stable_ctrl, loop, heap_stable_iff);
 963   ctrl = new IfTrueNode(heap_stable_iff);
 964   phase->register_control(ctrl, loop, heap_stable_iff);
 965 
 966   assert(is_heap_stable_test(heap_stable_iff), "Should match the shape");
 967 }
 968 
 969 void ShenandoahBarrierC2Support::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) {
 970   const Type* val_t = phase->igvn().type(val);
 971   if (val_t->meet(TypePtr::NULL_PTR) == val_t) {
 972     IdealLoopTree* loop = phase->get_loop(ctrl);
 973     Node* null_cmp = new CmpPNode(val, phase->igvn().zerocon(T_OBJECT));
 974     phase->register_new_node(null_cmp, ctrl);
 975     Node* null_test = new BoolNode(null_cmp, BoolTest::ne);
 976     phase->register_new_node(null_test, ctrl);
 977     IfNode* null_iff = new IfNode(ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
 978     phase->register_control(null_iff, loop, ctrl);
 979     ctrl = new IfTrueNode(null_iff);
 980     phase->register_control(ctrl, loop, null_iff);
 981     null_ctrl = new IfFalseNode(null_iff);
 982     phase->register_control(null_ctrl, loop, null_iff);
 983   }
 984 }
 985 
 986 Node* ShenandoahBarrierC2Support::clone_null_check(Node*& c, Node* val, Node* unc_ctrl, PhaseIdealLoop* phase) {
 987   IdealLoopTree *loop = phase->get_loop(c);
 988   Node* iff = unc_ctrl->in(0);
 989   assert(iff->is_If(), "broken");
 990   Node* new_iff = iff->clone();
 991   new_iff->set_req(0, c);
 992   phase->register_control(new_iff, loop, c);
 993   Node* iffalse = new IfFalseNode(new_iff->as_If());
 994   phase->register_control(iffalse, loop, new_iff);
 995   Node* iftrue = new IfTrueNode(new_iff->as_If());
 996   phase->register_control(iftrue, loop, new_iff);
 997   c = iftrue;
 998   const Type *t = phase->igvn().type(val);
 999   assert(val->Opcode() == Op_CastPP, "expect cast to non null here");
1000   Node* uncasted_val = val->in(1);
1001   val = new CastPPNode(uncasted_val, t);
1002   val->init_req(0, c);
1003   phase->register_new_node(val, c);
1004   return val;
1005 }
1006 
1007 void ShenandoahBarrierC2Support::fix_null_check(Node* unc, Node* unc_ctrl, Node* new_unc_ctrl,
1008                                                 Unique_Node_List& uses, PhaseIdealLoop* phase) {
1009   IfNode* iff = unc_ctrl->in(0)->as_If();
1010   Node* proj = iff->proj_out(0);
1011   assert(proj != unc_ctrl, "bad projection");
1012   Node* use = proj->unique_ctrl_out();
1013 
1014   assert(use == unc || use->is_Region(), "what else?");
1015 
1016   uses.clear();
1017   if (use == unc) {
1018     phase->set_idom(use, new_unc_ctrl, phase->dom_depth(use));
1019     for (uint i = 1; i < unc->req(); i++) {
1020       Node* n = unc->in(i);
1021       if (phase->has_ctrl(n) && phase->get_ctrl(n) == proj) {
1022         uses.push(n);
1023       }
1024     }
1025   } else {
1026     assert(use->is_Region(), "what else?");
1027     uint idx = 1;
1028     for (; use->in(idx) != proj; idx++);
1029     for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
1030       Node* u = use->fast_out(i);
1031       if (u->is_Phi() && phase->get_ctrl(u->in(idx)) == proj) {
1032         uses.push(u->in(idx));
1033       }
1034     }
1035   }
1036   for(uint next = 0; next < uses.size(); next++ ) {
1037     Node *n = uses.at(next);
1038     assert(phase->get_ctrl(n) == proj, "bad control");
1039     phase->set_ctrl_and_loop(n, new_unc_ctrl);
1040     if (n->in(0) == proj) {
1041       phase->igvn().replace_input_of(n, 0, new_unc_ctrl);
1042     }
1043     for (uint i = 0; i < n->req(); i++) {
1044       Node* m = n->in(i);
1045       if (m != NULL && phase->has_ctrl(m) && phase->get_ctrl(m) == proj) {
1046         uses.push(m);
1047       }
1048     }
1049   }
1050 
1051   phase->igvn().rehash_node_delayed(use);
1052   int nb = use->replace_edge(proj, new_unc_ctrl);
1053   assert(nb == 1, "only use expected");
1054 }
1055 
1056 void ShenandoahBarrierC2Support::in_cset_fast_test(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) {
1057   IdealLoopTree *loop = phase->get_loop(ctrl);
1058   Node* raw_rbtrue = new CastP2XNode(ctrl, val);
1059   phase->register_new_node(raw_rbtrue, ctrl);
1060   Node* cset_offset = new URShiftXNode(raw_rbtrue, phase->igvn().intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint()));
1061   phase->register_new_node(cset_offset, ctrl);
1062   Node* in_cset_fast_test_base_addr = phase->igvn().makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr()));
1063   phase->set_ctrl(in_cset_fast_test_base_addr, phase->C->root());
1064   Node* in_cset_fast_test_adr = new AddPNode(phase->C->top(), in_cset_fast_test_base_addr, cset_offset);
1065   phase->register_new_node(in_cset_fast_test_adr, ctrl);
1066   uint in_cset_fast_test_idx = Compile::AliasIdxRaw;
1067   const TypePtr* in_cset_fast_test_adr_type = NULL; // debug-mode-only argument
1068   debug_only(in_cset_fast_test_adr_type = phase->C->get_adr_type(in_cset_fast_test_idx));
1069   Node* in_cset_fast_test_load = new LoadBNode(ctrl, raw_mem, in_cset_fast_test_adr, in_cset_fast_test_adr_type, TypeInt::BYTE, MemNode::unordered);
1070   phase->register_new_node(in_cset_fast_test_load, ctrl);
1071   Node* in_cset_fast_test_cmp = new CmpINode(in_cset_fast_test_load, phase->igvn().zerocon(T_INT));
1072   phase->register_new_node(in_cset_fast_test_cmp, ctrl);
1073   Node* in_cset_fast_test_test = new BoolNode(in_cset_fast_test_cmp, BoolTest::eq);
1074   phase->register_new_node(in_cset_fast_test_test, ctrl);
1075   IfNode* in_cset_fast_test_iff = new IfNode(ctrl, in_cset_fast_test_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
1076   phase->register_control(in_cset_fast_test_iff, loop, ctrl);
1077 
1078   not_cset_ctrl = new IfTrueNode(in_cset_fast_test_iff);
1079   phase->register_control(not_cset_ctrl, loop, in_cset_fast_test_iff);
1080 
1081   ctrl = new IfFalseNode(in_cset_fast_test_iff);
1082   phase->register_control(ctrl, loop, in_cset_fast_test_iff);
1083 }
1084 
1085 void ShenandoahBarrierC2Support::call_lrb_stub(Node*& ctrl, Node*& val, Node*& result_mem, Node* raw_mem, PhaseIdealLoop* phase) {
1086   IdealLoopTree*loop = phase->get_loop(ctrl);
1087   const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr()->cast_to_nonconst();
1088 
1089   // The slow path stub consumes and produces raw memory in addition
1090   // to the existing memory edges
1091   Node* base = find_bottom_mem(ctrl, phase);
1092   MergeMemNode* mm = MergeMemNode::make(base);
1093   mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
1094   phase->register_new_node(mm, ctrl);
1095 
1096   Node* call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type(), CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_JRT), "shenandoah_load_reference_barrier", TypeRawPtr::BOTTOM);
1097   call->init_req(TypeFunc::Control, ctrl);
1098   call->init_req(TypeFunc::I_O, phase->C->top());
1099   call->init_req(TypeFunc::Memory, mm);
1100   call->init_req(TypeFunc::FramePtr, phase->C->top());
1101   call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1102   call->init_req(TypeFunc::Parms, val);
1103   phase->register_control(call, loop, ctrl);
1104   ctrl = new ProjNode(call, TypeFunc::Control);
1105   phase->register_control(ctrl, loop, call);
1106   result_mem = new ProjNode(call, TypeFunc::Memory);
1107   phase->register_new_node(result_mem, call);
1108   val = new ProjNode(call, TypeFunc::Parms);
1109   phase->register_new_node(val, call);
1110   val = new CheckCastPPNode(ctrl, val, obj_type);
1111   phase->register_new_node(val, ctrl);
1112 }
1113 
1114 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) {
1115   Node* ctrl = phase->get_ctrl(barrier);
1116   Node* init_raw_mem = fixer.find_mem(ctrl, barrier);
1117 
1118   // Update the control of all nodes that should be after the
1119   // barrier control flow
1120   uses.clear();
1121   // Every node that is control dependent on the barrier's input
1122   // control will be after the expanded barrier. The raw memory (if
1123   // its memory is control dependent on the barrier's input control)
1124   // must stay above the barrier.
1125   uses_to_ignore.clear();
1126   if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) {
1127     uses_to_ignore.push(init_raw_mem);
1128   }
1129   for (uint next = 0; next < uses_to_ignore.size(); next++) {
1130     Node *n = uses_to_ignore.at(next);
1131     for (uint i = 0; i < n->req(); i++) {
1132       Node* in = n->in(i);
1133       if (in != NULL && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) {
1134         uses_to_ignore.push(in);
1135       }
1136     }
1137   }
1138   for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) {
1139     Node* u = ctrl->fast_out(i);
1140     if (u->_idx < last &&
1141         u != barrier &&
1142         !uses_to_ignore.member(u) &&
1143         (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) &&
1144         (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) {
1145       Node* old_c = phase->ctrl_or_self(u);
1146       Node* c = old_c;
1147       if (c != ctrl ||
1148           is_dominator_same_ctrl(old_c, barrier, u, phase) ||
1149           ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) {
1150         phase->igvn().rehash_node_delayed(u);
1151         int nb = u->replace_edge(ctrl, region);
1152         if (u->is_CFG()) {
1153           if (phase->idom(u) == ctrl) {
1154             phase->set_idom(u, region, phase->dom_depth(region));
1155           }
1156         } else if (phase->get_ctrl(u) == ctrl) {
1157           assert(u != init_raw_mem, "should leave input raw mem above the barrier");
1158           uses.push(u);
1159         }
1160         assert(nb == 1, "more than 1 ctrl input?");
1161         --i, imax -= nb;
1162       }
1163     }
1164   }
1165 }
1166 
1167 static Node* create_phis_on_call_return(Node* ctrl, Node* c, Node* n, Node* n_clone, const CallProjections& projs, PhaseIdealLoop* phase) {
1168   Node* region = NULL;
1169   while (c != ctrl) {
1170     if (c->is_Region()) {
1171       region = c;
1172     }
1173     c = phase->idom(c);
1174   }
1175   assert(region != NULL, "");
1176   Node* phi = new PhiNode(region, n->bottom_type());
1177   for (uint j = 1; j < region->req(); j++) {
1178     Node* in = region->in(j);
1179     if (phase->is_dominator(projs.fallthrough_catchproj, in)) {
1180       phi->init_req(j, n);
1181     } else if (phase->is_dominator(projs.catchall_catchproj, in)) {
1182       phi->init_req(j, n_clone);
1183     } else {
1184       phi->init_req(j, create_phis_on_call_return(ctrl, in, n, n_clone, projs, phase));
1185     }
1186   }
1187   phase->register_new_node(phi, region);
1188   return phi;
1189 }
1190 
1191 void ShenandoahBarrierC2Support::pin_and_expand(PhaseIdealLoop* phase) {
1192   ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
1193 
1194   Unique_Node_List uses;
1195   for (int i = 0; i < state->enqueue_barriers_count(); i++) {
1196     Node* barrier = state->enqueue_barrier(i);
1197     Node* ctrl = phase->get_ctrl(barrier);
1198     IdealLoopTree* loop = phase->get_loop(ctrl);
1199     if (loop->_head->is_OuterStripMinedLoop()) {
1200       // Expanding a barrier here will break loop strip mining
1201       // verification. Transform the loop so the loop nest doesn't
1202       // appear as strip mined.
1203       OuterStripMinedLoopNode* outer = loop->_head->as_OuterStripMinedLoop();
1204       hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1205     }
1206   }
1207 
1208   Node_Stack stack(0);
1209   Node_List clones;
1210   for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1211     ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1212     if (lrb->get_barrier_strength() == ShenandoahLoadReferenceBarrierNode::NONE) {
1213       continue;
1214     }
1215 
1216     Node* ctrl = phase->get_ctrl(lrb);
1217     Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1218 
1219     CallStaticJavaNode* unc = NULL;
1220     Node* unc_ctrl = NULL;
1221     Node* uncasted_val = val;
1222 
1223     for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) {
1224       Node* u = lrb->fast_out(i);
1225       if (u->Opcode() == Op_CastPP &&
1226           u->in(0) != NULL &&
1227           phase->is_dominator(u->in(0), ctrl)) {
1228         const Type* u_t = phase->igvn().type(u);
1229 
1230         if (u_t->meet(TypePtr::NULL_PTR) != u_t &&
1231             u->in(0)->Opcode() == Op_IfTrue &&
1232             u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
1233             u->in(0)->in(0)->is_If() &&
1234             u->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
1235             u->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
1236             u->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
1237             u->in(0)->in(0)->in(1)->in(1)->in(1) == val &&
1238             u->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
1239           IdealLoopTree* loop = phase->get_loop(ctrl);
1240           IdealLoopTree* unc_loop = phase->get_loop(u->in(0));
1241 
1242           if (!unc_loop->is_member(loop)) {
1243             continue;
1244           }
1245 
1246           Node* branch = no_branches(ctrl, u->in(0), false, phase);
1247           assert(branch == NULL || branch == NodeSentinel, "was not looking for a branch");
1248           if (branch == NodeSentinel) {
1249             continue;
1250           }
1251 
1252           phase->igvn().replace_input_of(u, 1, val);
1253           phase->igvn().replace_input_of(lrb, ShenandoahLoadReferenceBarrierNode::ValueIn, u);
1254           phase->set_ctrl(u, u->in(0));
1255           phase->set_ctrl(lrb, u->in(0));
1256           unc = u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
1257           unc_ctrl = u->in(0);
1258           val = u;
1259 
1260           for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
1261             Node* u = val->fast_out(j);
1262             if (u == lrb) continue;
1263             phase->igvn().rehash_node_delayed(u);
1264             int nb = u->replace_edge(val, lrb);
1265             --j; jmax -= nb;
1266           }
1267 
1268           RegionNode* r = new RegionNode(3);
1269           IfNode* iff = unc_ctrl->in(0)->as_If();
1270 
1271           Node* ctrl_use = unc_ctrl->unique_ctrl_out();
1272           Node* unc_ctrl_clone = unc_ctrl->clone();
1273           phase->register_control(unc_ctrl_clone, loop, iff);
1274           Node* c = unc_ctrl_clone;
1275           Node* new_cast = clone_null_check(c, val, unc_ctrl_clone, phase);
1276           r->init_req(1, new_cast->in(0)->in(0)->as_If()->proj_out(0));
1277 
1278           phase->igvn().replace_input_of(unc_ctrl, 0, c->in(0));
1279           phase->set_idom(unc_ctrl, c->in(0), phase->dom_depth(unc_ctrl));
1280           phase->lazy_replace(c, unc_ctrl);
1281           c = NULL;;
1282           phase->igvn().replace_input_of(val, 0, unc_ctrl_clone);
1283           phase->set_ctrl(val, unc_ctrl_clone);
1284 
1285           IfNode* new_iff = new_cast->in(0)->in(0)->as_If();
1286           fix_null_check(unc, unc_ctrl_clone, r, uses, phase);
1287           Node* iff_proj = iff->proj_out(0);
1288           r->init_req(2, iff_proj);
1289           phase->register_control(r, phase->ltree_root(), iff);
1290 
1291           Node* new_bol = new_iff->in(1)->clone();
1292           Node* new_cmp = new_bol->in(1)->clone();
1293           assert(new_cmp->Opcode() == Op_CmpP, "broken");
1294           assert(new_cmp->in(1) == val->in(1), "broken");
1295           new_bol->set_req(1, new_cmp);
1296           new_cmp->set_req(1, lrb);
1297           phase->register_new_node(new_bol, new_iff->in(0));
1298           phase->register_new_node(new_cmp, new_iff->in(0));
1299           phase->igvn().replace_input_of(new_iff, 1, new_bol);
1300           phase->igvn().replace_input_of(new_cast, 1, lrb);
1301 
1302           for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) {
1303             Node* u = lrb->fast_out(i);
1304             if (u == new_cast || u == new_cmp) {
1305               continue;
1306             }
1307             phase->igvn().rehash_node_delayed(u);
1308             int nb = u->replace_edge(lrb, new_cast);
1309             assert(nb > 0, "no update?");
1310             --i; imax -= nb;
1311           }
1312 
1313           for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
1314             Node* u = val->fast_out(i);
1315             if (u == lrb) {
1316               continue;
1317             }
1318             phase->igvn().rehash_node_delayed(u);
1319             int nb = u->replace_edge(val, new_cast);
1320             assert(nb > 0, "no update?");
1321             --i; imax -= nb;
1322           }
1323 
1324           ctrl = unc_ctrl_clone;
1325           phase->set_ctrl_and_loop(lrb, ctrl);
1326           break;
1327         }
1328       }
1329     }
1330     if ((ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) || ctrl->is_CallJava()) {
1331       CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_CallJava() : ctrl->as_CallJava();
1332       CallProjections projs;
1333       call->extract_projections(&projs, false, false);
1334 
1335       Node* lrb_clone = lrb->clone();
1336       phase->register_new_node(lrb_clone, projs.catchall_catchproj);
1337       phase->set_ctrl(lrb, projs.fallthrough_catchproj);
1338 
1339       stack.push(lrb, 0);
1340       clones.push(lrb_clone);
1341 
1342       do {
1343         assert(stack.size() == clones.size(), "");
1344         Node* n = stack.node();
1345 #ifdef ASSERT
1346         if (n->is_Load()) {
1347           Node* mem = n->in(MemNode::Memory);
1348           for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
1349             Node* u = mem->fast_out(j);
1350             assert(!u->is_Store() || !u->is_LoadStore() || phase->get_ctrl(u) != ctrl, "anti dependent store?");
1351           }
1352         }
1353 #endif
1354         uint idx = stack.index();
1355         Node* n_clone = clones.at(clones.size()-1);
1356         if (idx < n->outcnt()) {
1357           Node* u = n->raw_out(idx);
1358           Node* c = phase->ctrl_or_self(u);
1359           if (phase->is_dominator(call, c) && phase->is_dominator(c, projs.fallthrough_proj)) {
1360             stack.set_index(idx+1);
1361             assert(!u->is_CFG(), "");
1362             stack.push(u, 0);
1363             Node* u_clone = u->clone();
1364             int nb = u_clone->replace_edge(n, n_clone);
1365             assert(nb > 0, "should have replaced some uses");
1366             phase->register_new_node(u_clone, projs.catchall_catchproj);
1367             clones.push(u_clone);
1368             phase->set_ctrl(u, projs.fallthrough_catchproj);
1369           } else {
1370             bool replaced = false;
1371             if (u->is_Phi()) {
1372               for (uint k = 1; k < u->req(); k++) {
1373                 if (u->in(k) == n) {
1374                   if (phase->is_dominator(projs.catchall_catchproj, u->in(0)->in(k))) {
1375                     phase->igvn().replace_input_of(u, k, n_clone);
1376                     replaced = true;
1377                   } else if (!phase->is_dominator(projs.fallthrough_catchproj, u->in(0)->in(k))) {
1378                     phase->igvn().replace_input_of(u, k, create_phis_on_call_return(ctrl, u->in(0)->in(k), n, n_clone, projs, phase));
1379                     replaced = true;
1380                   }
1381                 }
1382               }
1383             } else {
1384               if (phase->is_dominator(projs.catchall_catchproj, c)) {
1385                 phase->igvn().rehash_node_delayed(u);
1386                 int nb = u->replace_edge(n, n_clone);
1387                 assert(nb > 0, "should have replaced some uses");
1388                 replaced = true;
1389               } else if (!phase->is_dominator(projs.fallthrough_catchproj, c)) {
1390                 phase->igvn().rehash_node_delayed(u);
1391                 int nb = u->replace_edge(n, create_phis_on_call_return(ctrl, c, n, n_clone, projs, phase));
1392                 assert(nb > 0, "should have replaced some uses");
1393                 replaced = true;
1394               }
1395             }
1396             if (!replaced) {
1397               stack.set_index(idx+1);
1398             }
1399           }
1400         } else {
1401           stack.pop();
1402           clones.pop();
1403         }
1404       } while (stack.size() > 0);
1405       assert(stack.size() == 0 && clones.size() == 0, "");
1406     }
1407   }
1408 
1409   for (int i = 0; i < state->load_reference_barriers_count(); i++) {
1410     ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1411     if (lrb->get_barrier_strength() == ShenandoahLoadReferenceBarrierNode::NONE) {
1412       continue;
1413     }
1414     Node* ctrl = phase->get_ctrl(lrb);
1415     IdealLoopTree* loop = phase->get_loop(ctrl);
1416     if (loop->_head->is_OuterStripMinedLoop()) {
1417       // Expanding a barrier here will break loop strip mining
1418       // verification. Transform the loop so the loop nest doesn't
1419       // appear as strip mined.
1420       OuterStripMinedLoopNode* outer = loop->_head->as_OuterStripMinedLoop();
1421       hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1422     }
1423   }
1424 
1425   // Expand load-reference-barriers
1426   MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
1427   Unique_Node_List uses_to_ignore;
1428   for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1429     ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1430     if (lrb->get_barrier_strength() == ShenandoahLoadReferenceBarrierNode::NONE) {
1431       phase->igvn().replace_node(lrb, lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1432       continue;
1433     }
1434     uint last = phase->C->unique();
1435     Node* ctrl = phase->get_ctrl(lrb);
1436     Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1437 
1438 
1439     Node* orig_ctrl = ctrl;
1440 
1441     Node* raw_mem = fixer.find_mem(ctrl, lrb);
1442     Node* init_raw_mem = raw_mem;
1443     Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL);
1444 
1445     IdealLoopTree *loop = phase->get_loop(ctrl);
1446     CallStaticJavaNode* unc = lrb->pin_and_expand_null_check(phase->igvn());
1447     Node* unc_ctrl = NULL;
1448     if (unc != NULL) {
1449       if (val->in(ShenandoahLoadReferenceBarrierNode::Control) != ctrl) {
1450         unc = NULL;
1451       } else {
1452         unc_ctrl = val->in(ShenandoahLoadReferenceBarrierNode::Control);
1453       }
1454     }
1455 
1456     Node* uncasted_val = val;
1457     if (unc != NULL) {
1458       uncasted_val = val->in(1);
1459     }
1460 
1461     Node* heap_stable_ctrl = NULL;
1462     Node* null_ctrl = NULL;
1463 
1464     assert(val->bottom_type()->make_oopptr(), "need oop");
1465     assert(val->bottom_type()->make_oopptr()->const_oop() == NULL, "expect non-constant");
1466 
1467     enum { _heap_stable = 1, _not_cset, _fwded, _evac_path, _null_path, PATH_LIMIT };
1468     Node* region = new RegionNode(PATH_LIMIT);
1469     Node* val_phi = new PhiNode(region, uncasted_val->bottom_type()->is_oopptr());
1470     Node* raw_mem_phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1471 
1472     // Stable path.
1473     test_heap_stable(ctrl, raw_mem, heap_stable_ctrl, phase);
1474     IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
1475 
1476     // Heap stable case
1477     region->init_req(_heap_stable, heap_stable_ctrl);
1478     val_phi->init_req(_heap_stable, uncasted_val);
1479     raw_mem_phi->init_req(_heap_stable, raw_mem);
1480 
1481     Node* reg2_ctrl = NULL;
1482     // Null case
1483     test_null(ctrl, val, null_ctrl, phase);
1484     if (null_ctrl != NULL) {
1485       reg2_ctrl = null_ctrl->in(0);
1486       region->init_req(_null_path, null_ctrl);
1487       val_phi->init_req(_null_path, uncasted_val);
1488       raw_mem_phi->init_req(_null_path, raw_mem);
1489     } else {
1490       region->del_req(_null_path);
1491       val_phi->del_req(_null_path);
1492       raw_mem_phi->del_req(_null_path);
1493     }
1494 
1495     // Test for in-cset.
1496     // Wires !in_cset(obj) to slot 2 of region and phis
1497     Node* not_cset_ctrl = NULL;
1498     in_cset_fast_test(ctrl, not_cset_ctrl, uncasted_val, raw_mem, phase);
1499     if (not_cset_ctrl != NULL) {
1500       if (reg2_ctrl == NULL) reg2_ctrl = not_cset_ctrl->in(0);
1501       region->init_req(_not_cset, not_cset_ctrl);
1502       val_phi->init_req(_not_cset, uncasted_val);
1503       raw_mem_phi->init_req(_not_cset, raw_mem);
1504     }
1505 
1506     // Resolve object when orig-value is in cset.
1507     // Make the unconditional resolve for fwdptr.
1508     Node* new_val = uncasted_val;
1509     if (unc_ctrl != NULL) {
1510       // Clone the null check in this branch to allow implicit null check
1511       new_val = clone_null_check(ctrl, val, unc_ctrl, phase);
1512       fix_null_check(unc, unc_ctrl, ctrl->in(0)->as_If()->proj_out(0), uses, phase);
1513 
1514       IfNode* iff = unc_ctrl->in(0)->as_If();
1515       phase->igvn().replace_input_of(iff, 1, phase->igvn().intcon(1));
1516     }
1517     Node* addr = new AddPNode(new_val, uncasted_val, phase->igvn().MakeConX(oopDesc::mark_offset_in_bytes()));
1518     phase->register_new_node(addr, ctrl);
1519     assert(new_val->bottom_type()->isa_oopptr(), "what else?");
1520     Node* markword = new LoadXNode(ctrl, raw_mem, addr, TypeRawPtr::BOTTOM, TypeX_X, MemNode::unordered);
1521     phase->register_new_node(markword, ctrl);
1522 
1523     // Test if object is forwarded. This is the case if lowest two bits are set.
1524     Node* masked = new AndXNode(markword, phase->igvn().MakeConX(markOopDesc::lock_mask_in_place));
1525     phase->register_new_node(masked, ctrl);
1526     Node* cmp = new CmpXNode(masked, phase->igvn().MakeConX(markOopDesc::marked_value));
1527     phase->register_new_node(cmp, ctrl);
1528 
1529     // Only branch to LRB stub if object is not forwarded; otherwise reply with fwd ptr
1530     Node* bol = new BoolNode(cmp, BoolTest::eq); // Equals 3 means it's forwarded
1531     phase->register_new_node(bol, ctrl);
1532 
1533     IfNode* iff = new IfNode(ctrl, bol, PROB_LIKELY(0.999), COUNT_UNKNOWN);
1534     phase->register_control(iff, loop, ctrl);
1535     Node* if_fwd = new IfTrueNode(iff);
1536     phase->register_control(if_fwd, loop, iff);
1537     Node* if_not_fwd = new IfFalseNode(iff);
1538     phase->register_control(if_not_fwd, loop, iff);
1539 
1540     // Decode forward pointer: since we already have the lowest bits, we can just subtract them
1541     // from the mark word without the need for large immediate mask.
1542     Node* masked2 = new SubXNode(markword, masked);
1543     phase->register_new_node(masked2, if_fwd);
1544     Node* fwdraw = new CastX2PNode(masked2);
1545     fwdraw->init_req(0, if_fwd);
1546     phase->register_new_node(fwdraw, if_fwd);
1547     Node* fwd = new CheckCastPPNode(NULL, fwdraw, val->bottom_type());
1548     phase->register_new_node(fwd, if_fwd);
1549 
1550     // Wire up not-equal-path in slots 3.
1551     region->init_req(_fwded, if_fwd);
1552     val_phi->init_req(_fwded, fwd);
1553     raw_mem_phi->init_req(_fwded, raw_mem);
1554 
1555     // Call lrb-stub and wire up that path in slots 4
1556     Node* result_mem = NULL;
1557     ctrl = if_not_fwd;
1558     fwd = new_val;
1559     call_lrb_stub(ctrl, fwd, result_mem, raw_mem, phase);
1560     region->init_req(_evac_path, ctrl);
1561     val_phi->init_req(_evac_path, fwd);
1562     raw_mem_phi->init_req(_evac_path, result_mem);
1563 
1564     phase->register_control(region, loop, heap_stable_iff);
1565     Node* out_val = val_phi;
1566     phase->register_new_node(val_phi, region);
1567     phase->register_new_node(raw_mem_phi, region);
1568 
1569     fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase);
1570 
1571     ctrl = orig_ctrl;
1572 
1573     if (unc != NULL) {
1574       for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
1575         Node* u = val->fast_out(i);
1576         Node* c = phase->ctrl_or_self(u);
1577         if (u != lrb && (c != ctrl || is_dominator_same_ctrl(c, lrb, u, phase))) {
1578           phase->igvn().rehash_node_delayed(u);
1579           int nb = u->replace_edge(val, out_val);
1580           --i, imax -= nb;
1581         }
1582       }
1583       if (val->outcnt() == 0) {
1584         phase->igvn()._worklist.push(val);
1585       }
1586     }
1587     phase->igvn().replace_node(lrb, out_val);
1588 
1589     follow_barrier_uses(out_val, ctrl, uses, phase);
1590 
1591     for(uint next = 0; next < uses.size(); next++ ) {
1592       Node *n = uses.at(next);
1593       assert(phase->get_ctrl(n) == ctrl, "bad control");
1594       assert(n != init_raw_mem, "should leave input raw mem above the barrier");
1595       phase->set_ctrl(n, region);
1596       follow_barrier_uses(n, ctrl, uses, phase);
1597     }
1598 
1599     // The slow path call produces memory: hook the raw memory phi
1600     // from the expanded load reference barrier with the rest of the graph
1601     // which may require adding memory phis at every post dominated
1602     // region and at enclosing loop heads. Use the memory state
1603     // collected in memory_nodes to fix the memory graph. Update that
1604     // memory state as we go.
1605     fixer.fix_mem(ctrl, region, init_raw_mem, raw_mem_for_ctrl, raw_mem_phi, uses);
1606   }
1607   // Done expanding load-reference-barriers.
1608   assert(ShenandoahBarrierSetC2::bsc2()->state()->load_reference_barriers_count() == 0, "all load reference barrier nodes should have been replaced");
1609 
1610   for (int i = state->enqueue_barriers_count() - 1; i >= 0; i--) {
1611     Node* barrier = state->enqueue_barrier(i);
1612     Node* pre_val = barrier->in(1);
1613 
1614     if (phase->igvn().type(pre_val)->higher_equal(TypePtr::NULL_PTR)) {
1615       ShouldNotReachHere();
1616       continue;
1617     }
1618 
1619     Node* ctrl = phase->get_ctrl(barrier);
1620 
1621     if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) {
1622       assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0)->in(0), pre_val, ctrl->in(0), phase), "can't move");
1623       ctrl = ctrl->in(0)->in(0);
1624       phase->set_ctrl(barrier, ctrl);
1625     } else if (ctrl->is_CallRuntime()) {
1626       assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0), pre_val, ctrl, phase), "can't move");
1627       ctrl = ctrl->in(0);
1628       phase->set_ctrl(barrier, ctrl);
1629     }
1630 
1631     Node* init_ctrl = ctrl;
1632     IdealLoopTree* loop = phase->get_loop(ctrl);
1633     Node* raw_mem = fixer.find_mem(ctrl, barrier);
1634     Node* init_raw_mem = raw_mem;
1635     Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL);
1636     Node* heap_stable_ctrl = NULL;
1637     Node* null_ctrl = NULL;
1638     uint last = phase->C->unique();
1639 
1640     enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
1641     Node* region = new RegionNode(PATH_LIMIT);
1642     Node* phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1643 
1644     enum { _fast_path = 1, _slow_path, _null_path, PATH_LIMIT2 };
1645     Node* region2 = new RegionNode(PATH_LIMIT2);
1646     Node* phi2 = PhiNode::make(region2, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1647 
1648     // Stable path.
1649     test_heap_stable(ctrl, raw_mem, heap_stable_ctrl, phase);
1650     region->init_req(_heap_stable, heap_stable_ctrl);
1651     phi->init_req(_heap_stable, raw_mem);
1652 
1653     // Null path
1654     Node* reg2_ctrl = NULL;
1655     test_null(ctrl, pre_val, null_ctrl, phase);
1656     if (null_ctrl != NULL) {
1657       reg2_ctrl = null_ctrl->in(0);
1658       region2->init_req(_null_path, null_ctrl);
1659       phi2->init_req(_null_path, raw_mem);
1660     } else {
1661       region2->del_req(_null_path);
1662       phi2->del_req(_null_path);
1663     }
1664 
1665     const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
1666     const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1667     Node* thread = new ThreadLocalNode();
1668     phase->register_new_node(thread, ctrl);
1669     Node* buffer_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(buffer_offset));
1670     phase->register_new_node(buffer_adr, ctrl);
1671     Node* index_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(index_offset));
1672     phase->register_new_node(index_adr, ctrl);
1673 
1674     BasicType index_bt = TypeX_X->basic_type();
1675     assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
1676     const TypePtr* adr_type = TypeRawPtr::BOTTOM;
1677     Node* index = new LoadXNode(ctrl, raw_mem, index_adr, adr_type, TypeX_X, MemNode::unordered);
1678     phase->register_new_node(index, ctrl);
1679     Node* index_cmp = new CmpXNode(index, phase->igvn().MakeConX(0));
1680     phase->register_new_node(index_cmp, ctrl);
1681     Node* index_test = new BoolNode(index_cmp, BoolTest::ne);
1682     phase->register_new_node(index_test, ctrl);
1683     IfNode* queue_full_iff = new IfNode(ctrl, index_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
1684     if (reg2_ctrl == NULL) reg2_ctrl = queue_full_iff;
1685     phase->register_control(queue_full_iff, loop, ctrl);
1686     Node* not_full = new IfTrueNode(queue_full_iff);
1687     phase->register_control(not_full, loop, queue_full_iff);
1688     Node* full = new IfFalseNode(queue_full_iff);
1689     phase->register_control(full, loop, queue_full_iff);
1690 
1691     ctrl = not_full;
1692 
1693     Node* next_index = new SubXNode(index, phase->igvn().MakeConX(sizeof(intptr_t)));
1694     phase->register_new_node(next_index, ctrl);
1695 
1696     Node* buffer  = new LoadPNode(ctrl, raw_mem, buffer_adr, adr_type, TypeRawPtr::NOTNULL, MemNode::unordered);
1697     phase->register_new_node(buffer, ctrl);
1698     Node *log_addr = new AddPNode(phase->C->top(), buffer, next_index);
1699     phase->register_new_node(log_addr, ctrl);
1700     Node* log_store = new StorePNode(ctrl, raw_mem, log_addr, adr_type, pre_val, MemNode::unordered);
1701     phase->register_new_node(log_store, ctrl);
1702     // update the index
1703     Node* index_update = new StoreXNode(ctrl, log_store, index_adr, adr_type, next_index, MemNode::unordered);
1704     phase->register_new_node(index_update, ctrl);
1705 
1706     // Fast-path case
1707     region2->init_req(_fast_path, ctrl);
1708     phi2->init_req(_fast_path, index_update);
1709 
1710     ctrl = full;
1711 
1712     Node* base = find_bottom_mem(ctrl, phase);
1713 
1714     MergeMemNode* mm = MergeMemNode::make(base);
1715     mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
1716     phase->register_new_node(mm, ctrl);
1717 
1718     Node* call = new CallLeafNode(ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type(), CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", TypeRawPtr::BOTTOM);
1719     call->init_req(TypeFunc::Control, ctrl);
1720     call->init_req(TypeFunc::I_O, phase->C->top());
1721     call->init_req(TypeFunc::Memory, mm);
1722     call->init_req(TypeFunc::FramePtr, phase->C->top());
1723     call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1724     call->init_req(TypeFunc::Parms, pre_val);
1725     call->init_req(TypeFunc::Parms+1, thread);
1726     phase->register_control(call, loop, ctrl);
1727 
1728     Node* ctrl_proj = new ProjNode(call, TypeFunc::Control);
1729     phase->register_control(ctrl_proj, loop, call);
1730     Node* mem_proj = new ProjNode(call, TypeFunc::Memory);
1731     phase->register_new_node(mem_proj, call);
1732 
1733     // Slow-path case
1734     region2->init_req(_slow_path, ctrl_proj);
1735     phi2->init_req(_slow_path, mem_proj);
1736 
1737     phase->register_control(region2, loop, reg2_ctrl);
1738     phase->register_new_node(phi2, region2);
1739 
1740     region->init_req(_heap_unstable, region2);
1741     phi->init_req(_heap_unstable, phi2);
1742 
1743     phase->register_control(region, loop, heap_stable_ctrl->in(0));
1744     phase->register_new_node(phi, region);
1745 
1746     fix_ctrl(barrier, region, fixer, uses, uses_to_ignore, last, phase);
1747     for(uint next = 0; next < uses.size(); next++ ) {
1748       Node *n = uses.at(next);
1749       assert(phase->get_ctrl(n) == init_ctrl, "bad control");
1750       assert(n != init_raw_mem, "should leave input raw mem above the barrier");
1751       phase->set_ctrl(n, region);
1752       follow_barrier_uses(n, init_ctrl, uses, phase);
1753     }
1754     fixer.fix_mem(init_ctrl, region, init_raw_mem, raw_mem_for_ctrl, phi, uses);
1755 
1756     phase->igvn().replace_node(barrier, pre_val);
1757   }
1758   assert(state->enqueue_barriers_count() == 0, "all enqueue barrier nodes should have been replaced");
1759 
1760 }
1761 
1762 void ShenandoahBarrierC2Support::move_heap_stable_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
1763   IdealLoopTree *loop = phase->get_loop(iff);
1764   Node* loop_head = loop->_head;
1765   Node* entry_c = loop_head->in(LoopNode::EntryControl);
1766 
1767   Node* bol = iff->in(1);
1768   Node* cmp = bol->in(1);
1769   Node* andi = cmp->in(1);
1770   Node* load = andi->in(1);
1771 
1772   assert(is_gc_state_load(load), "broken");
1773   if (!phase->is_dominator(load->in(0), entry_c)) {
1774     Node* mem_ctrl = NULL;
1775     Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
1776     load = load->clone();
1777     load->set_req(MemNode::Memory, mem);
1778     load->set_req(0, entry_c);
1779     phase->register_new_node(load, entry_c);
1780     andi = andi->clone();
1781     andi->set_req(1, load);
1782     phase->register_new_node(andi, entry_c);
1783     cmp = cmp->clone();
1784     cmp->set_req(1, andi);
1785     phase->register_new_node(cmp, entry_c);
1786     bol = bol->clone();
1787     bol->set_req(1, cmp);
1788     phase->register_new_node(bol, entry_c);
1789 
1790     Node* old_bol =iff->in(1);
1791     phase->igvn().replace_input_of(iff, 1, bol);
1792   }
1793 }
1794 
1795 bool ShenandoahBarrierC2Support::identical_backtoback_ifs(Node* n, PhaseIdealLoop* phase) {
1796   if (!n->is_If() || n->is_CountedLoopEnd()) {
1797     return false;
1798   }
1799   Node* region = n->in(0);
1800 
1801   if (!region->is_Region()) {
1802     return false;
1803   }
1804   Node* dom = phase->idom(region);
1805   if (!dom->is_If()) {
1806     return false;
1807   }
1808 
1809   if (!is_heap_stable_test(n) || !is_heap_stable_test(dom)) {
1810     return false;
1811   }
1812 
1813   IfNode* dom_if = dom->as_If();
1814   Node* proj_true = dom_if->proj_out(1);
1815   Node* proj_false = dom_if->proj_out(0);
1816 
1817   for (uint i = 1; i < region->req(); i++) {
1818     if (phase->is_dominator(proj_true, region->in(i))) {
1819       continue;
1820     }
1821     if (phase->is_dominator(proj_false, region->in(i))) {
1822       continue;
1823     }
1824     return false;
1825   }
1826 
1827   return true;
1828 }
1829 
1830 void ShenandoahBarrierC2Support::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) {
1831   assert(is_heap_stable_test(n), "no other tests");
1832   if (identical_backtoback_ifs(n, phase)) {
1833     Node* n_ctrl = n->in(0);
1834     if (phase->can_split_if(n_ctrl)) {
1835       IfNode* dom_if = phase->idom(n_ctrl)->as_If();
1836       if (is_heap_stable_test(n)) {
1837         Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1);
1838         assert(is_gc_state_load(gc_state_load), "broken");
1839         Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1);
1840         assert(is_gc_state_load(dom_gc_state_load), "broken");
1841         if (gc_state_load != dom_gc_state_load) {
1842           phase->igvn().replace_node(gc_state_load, dom_gc_state_load);
1843         }
1844       }
1845       PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1846       Node* proj_true = dom_if->proj_out(1);
1847       Node* proj_false = dom_if->proj_out(0);
1848       Node* con_true = phase->igvn().makecon(TypeInt::ONE);
1849       Node* con_false = phase->igvn().makecon(TypeInt::ZERO);
1850 
1851       for (uint i = 1; i < n_ctrl->req(); i++) {
1852         if (phase->is_dominator(proj_true, n_ctrl->in(i))) {
1853           bolphi->init_req(i, con_true);
1854         } else {
1855           assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1856           bolphi->init_req(i, con_false);
1857         }
1858       }
1859       phase->register_new_node(bolphi, n_ctrl);
1860       phase->igvn().replace_input_of(n, 1, bolphi);
1861       phase->do_split_if(n);
1862     }
1863   }
1864 }
1865 
1866 IfNode* ShenandoahBarrierC2Support::find_unswitching_candidate(const IdealLoopTree* loop, PhaseIdealLoop* phase) {
1867   // Find first invariant test that doesn't exit the loop
1868   LoopNode *head = loop->_head->as_Loop();
1869   IfNode* unswitch_iff = NULL;
1870   Node* n = head->in(LoopNode::LoopBackControl);
1871   int loop_has_sfpts = -1;
1872   while (n != head) {
1873     Node* n_dom = phase->idom(n);
1874     if (n->is_Region()) {
1875       if (n_dom->is_If()) {
1876         IfNode* iff = n_dom->as_If();
1877         if (iff->in(1)->is_Bool()) {
1878           BoolNode* bol = iff->in(1)->as_Bool();
1879           if (bol->in(1)->is_Cmp()) {
1880             // If condition is invariant and not a loop exit,
1881             // then found reason to unswitch.
1882             if (is_heap_stable_test(iff) &&
1883                 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) {
1884               assert(!loop->is_loop_exit(iff), "both branches should be in the loop");
1885               if (loop_has_sfpts == -1) {
1886                 for(uint i = 0; i < loop->_body.size(); i++) {
1887                   Node *m = loop->_body[i];
1888                   if (m->is_SafePoint() && !m->is_CallLeaf()) {
1889                     loop_has_sfpts = 1;
1890                     break;
1891                   }
1892                 }
1893                 if (loop_has_sfpts == -1) {
1894                   loop_has_sfpts = 0;
1895                 }
1896               }
1897               if (!loop_has_sfpts) {
1898                 unswitch_iff = iff;
1899               }
1900             }
1901           }
1902         }
1903       }
1904     }
1905     n = n_dom;
1906   }
1907   return unswitch_iff;
1908 }
1909 
1910 
1911 void ShenandoahBarrierC2Support::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) {
1912   Node_List heap_stable_tests;
1913   Node_List gc_state_loads;
1914   stack.push(phase->C->start(), 0);
1915   do {
1916     Node* n = stack.node();
1917     uint i = stack.index();
1918 
1919     if (i < n->outcnt()) {
1920       Node* u = n->raw_out(i);
1921       stack.set_index(i+1);
1922       if (!visited.test_set(u->_idx)) {
1923         stack.push(u, 0);
1924       }
1925     } else {
1926       stack.pop();
1927       if (ShenandoahCommonGCStateLoads && is_gc_state_load(n)) {
1928         gc_state_loads.push(n);
1929       }
1930       if (n->is_If() && is_heap_stable_test(n)) {
1931         heap_stable_tests.push(n);
1932       }
1933     }
1934   } while (stack.size() > 0);
1935 
1936   bool progress;
1937   do {
1938     progress = false;
1939     for (uint i = 0; i < gc_state_loads.size(); i++) {
1940       Node* n = gc_state_loads.at(i);
1941       if (n->outcnt() != 0) {
1942         progress |= try_common_gc_state_load(n, phase);
1943       }
1944     }
1945   } while (progress);
1946 
1947   for (uint i = 0; i < heap_stable_tests.size(); i++) {
1948     Node* n = heap_stable_tests.at(i);
1949     assert(is_heap_stable_test(n), "only evacuation test");
1950     merge_back_to_back_tests(n, phase);
1951   }
1952 
1953   if (!phase->C->major_progress()) {
1954     VectorSet seen(Thread::current()->resource_area());
1955     for (uint i = 0; i < heap_stable_tests.size(); i++) {
1956       Node* n = heap_stable_tests.at(i);
1957       IdealLoopTree* loop = phase->get_loop(n);
1958       if (loop != phase->ltree_root() &&
1959           loop->_child == NULL &&
1960           !loop->_irreducible) {
1961         LoopNode* head = loop->_head->as_Loop();
1962         if ((!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) &&
1963             !seen.test_set(head->_idx)) {
1964           IfNode* iff = find_unswitching_candidate(loop, phase);
1965           if (iff != NULL) {
1966             Node* bol = iff->in(1);
1967             if (head->is_strip_mined()) {
1968               head->verify_strip_mined(0);
1969             }
1970             move_heap_stable_test_out_of_loop(iff, phase);
1971 
1972             AutoNodeBudget node_budget(phase);
1973 
1974             if (loop->policy_unswitching(phase)) {
1975               if (head->is_strip_mined()) {
1976                 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop();
1977                 hide_strip_mined_loop(outer, head->as_CountedLoop(), phase);
1978               }
1979               phase->do_unswitching(loop, old_new);
1980             } else {
1981               // Not proceeding with unswitching. Move load back in
1982               // the loop.
1983               phase->igvn().replace_input_of(iff, 1, bol);
1984             }
1985           }
1986         }
1987       }
1988     }
1989   }
1990 }
1991 
1992 #ifdef ASSERT
1993 void ShenandoahBarrierC2Support::verify_raw_mem(RootNode* root) {
1994   const bool trace = false;
1995   ResourceMark rm;
1996   Unique_Node_List nodes;
1997   Unique_Node_List controls;
1998   Unique_Node_List memories;
1999 
2000   nodes.push(root);
2001   for (uint next = 0; next < nodes.size(); next++) {
2002     Node *n  = nodes.at(next);
2003     if (ShenandoahBarrierSetC2::is_shenandoah_lrb_call(n)) {
2004       controls.push(n);
2005       if (trace) { tty->print("XXXXXX verifying"); n->dump(); }
2006       for (uint next2 = 0; next2 < controls.size(); next2++) {
2007         Node *m = controls.at(next2);
2008         for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
2009           Node* u = m->fast_out(i);
2010           if (u->is_CFG() && !u->is_Root() &&
2011               !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1) &&
2012               !(u->is_Region() && u->unique_ctrl_out()->Opcode() == Op_Halt)) {
2013             if (trace) { tty->print("XXXXXX pushing control"); u->dump(); }
2014             controls.push(u);
2015           }
2016         }
2017       }
2018       memories.push(n->as_Call()->proj_out(TypeFunc::Memory));
2019       for (uint next2 = 0; next2 < memories.size(); next2++) {
2020         Node *m = memories.at(next2);
2021         assert(m->bottom_type() == Type::MEMORY, "");
2022         for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
2023           Node* u = m->fast_out(i);
2024           if (u->bottom_type() == Type::MEMORY && (u->is_Mem() || u->is_ClearArray())) {
2025             if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
2026             memories.push(u);
2027           } else if (u->is_LoadStore()) {
2028             if (trace) { tty->print("XXXXXX pushing memory"); u->find_out_with(Op_SCMemProj)->dump(); }
2029             memories.push(u->find_out_with(Op_SCMemProj));
2030           } else if (u->is_MergeMem() && u->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == m) {
2031             if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
2032             memories.push(u);
2033           } else if (u->is_Phi()) {
2034             assert(u->bottom_type() == Type::MEMORY, "");
2035             if (u->adr_type() == TypeRawPtr::BOTTOM || u->adr_type() == TypePtr::BOTTOM) {
2036               assert(controls.member(u->in(0)), "");
2037               if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
2038               memories.push(u);
2039             }
2040           } else if (u->is_SafePoint() || u->is_MemBar()) {
2041             for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2042               Node* uu = u->fast_out(j);
2043               if (uu->bottom_type() == Type::MEMORY) {
2044                 if (trace) { tty->print("XXXXXX pushing memory"); uu->dump(); }
2045                 memories.push(uu);
2046               }
2047             }
2048           }
2049         }
2050       }
2051       for (uint next2 = 0; next2 < controls.size(); next2++) {
2052         Node *m = controls.at(next2);
2053         if (m->is_Region()) {
2054           bool all_in = true;
2055           for (uint i = 1; i < m->req(); i++) {
2056             if (!controls.member(m->in(i))) {
2057               all_in = false;
2058               break;
2059             }
2060           }
2061           if (trace) { tty->print("XXX verifying %s", all_in ? "all in" : ""); m->dump(); }
2062           bool found_phi = false;
2063           for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax && !found_phi; j++) {
2064             Node* u = m->fast_out(j);
2065             if (u->is_Phi() && memories.member(u)) {
2066               found_phi = true;
2067               for (uint i = 1; i < u->req() && found_phi; i++) {
2068                 Node* k = u->in(i);
2069                 if (memories.member(k) != controls.member(m->in(i))) {
2070                   found_phi = false;
2071                 }
2072               }
2073             }
2074           }
2075           assert(found_phi || all_in, "");
2076         }
2077       }
2078       controls.clear();
2079       memories.clear();
2080     }
2081     for( uint i = 0; i < n->len(); ++i ) {
2082       Node *m = n->in(i);
2083       if (m != NULL) {
2084         nodes.push(m);
2085       }
2086     }
2087   }
2088 }
2089 #endif
2090 
2091 ShenandoahEnqueueBarrierNode::ShenandoahEnqueueBarrierNode(Node* val) : Node(NULL, val) {
2092   ShenandoahBarrierSetC2::bsc2()->state()->add_enqueue_barrier(this);
2093 }
2094 
2095 const Type* ShenandoahEnqueueBarrierNode::bottom_type() const {
2096   if (in(1) == NULL || in(1)->is_top()) {
2097     return Type::TOP;
2098   }
2099   const Type* t = in(1)->bottom_type();
2100   if (t == TypePtr::NULL_PTR) {
2101     return t;
2102   }
2103   return t->is_oopptr()->cast_to_nonconst();
2104 }
2105 
2106 const Type* ShenandoahEnqueueBarrierNode::Value(PhaseGVN* phase) const {
2107   if (in(1) == NULL) {
2108     return Type::TOP;
2109   }
2110   const Type* t = phase->type(in(1));
2111   if (t == Type::TOP) {
2112     return Type::TOP;
2113   }
2114   if (t == TypePtr::NULL_PTR) {
2115     return t;
2116   }
2117   return t->is_oopptr()->cast_to_nonconst();
2118 }
2119 
2120 int ShenandoahEnqueueBarrierNode::needed(Node* n) {
2121   if (n == NULL ||
2122       n->is_Allocate() ||
2123       n->Opcode() == Op_ShenandoahEnqueueBarrier ||
2124       n->bottom_type() == TypePtr::NULL_PTR ||
2125       (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL)) {
2126     return NotNeeded;
2127   }
2128   if (n->is_Phi() ||
2129       n->is_CMove()) {
2130     return MaybeNeeded;
2131   }
2132   return Needed;
2133 }
2134 
2135 Node* ShenandoahEnqueueBarrierNode::next(Node* n) {
2136   for (;;) {
2137     if (n == NULL) {
2138       return n;
2139     } else if (n->bottom_type() == TypePtr::NULL_PTR) {
2140       return n;
2141     } else if (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL) {
2142       return n;
2143     } else if (n->is_ConstraintCast() ||
2144                n->Opcode() == Op_DecodeN ||
2145                n->Opcode() == Op_EncodeP) {
2146       n = n->in(1);
2147     } else if (n->is_Proj()) {
2148       n = n->in(0);
2149     } else {
2150       return n;
2151     }
2152   }
2153   ShouldNotReachHere();
2154   return NULL;
2155 }
2156 
2157 Node* ShenandoahEnqueueBarrierNode::Identity(PhaseGVN* phase) {
2158   PhaseIterGVN* igvn = phase->is_IterGVN();
2159 
2160   Node* n = next(in(1));
2161 
2162   int cont = needed(n);
2163 
2164   if (cont == NotNeeded) {
2165     return in(1);
2166   } else if (cont == MaybeNeeded) {
2167     if (igvn == NULL) {
2168       phase->record_for_igvn(this);
2169       return this;
2170     } else {
2171       ResourceMark rm;
2172       Unique_Node_List wq;
2173       uint wq_i = 0;
2174 
2175       for (;;) {
2176         if (n->is_Phi()) {
2177           for (uint i = 1; i < n->req(); i++) {
2178             Node* m = n->in(i);
2179             if (m != NULL) {
2180               wq.push(m);
2181             }
2182           }
2183         } else {
2184           assert(n->is_CMove(), "nothing else here");
2185           Node* m = n->in(CMoveNode::IfFalse);
2186           wq.push(m);
2187           m = n->in(CMoveNode::IfTrue);
2188           wq.push(m);
2189         }
2190         Node* orig_n = NULL;
2191         do {
2192           if (wq_i >= wq.size()) {
2193             return in(1);
2194           }
2195           n = wq.at(wq_i);
2196           wq_i++;
2197           orig_n = n;
2198           n = next(n);
2199           cont = needed(n);
2200           if (cont == Needed) {
2201             return this;
2202           }
2203         } while (cont != MaybeNeeded || (orig_n != n && wq.member(n)));
2204       }
2205     }
2206   }
2207 
2208   return this;
2209 }
2210 
2211 #ifdef ASSERT
2212 static bool has_never_branch(Node* root) {
2213   for (uint i = 1; i < root->req(); i++) {
2214     Node* in = root->in(i);
2215     if (in != NULL && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->Opcode() == Op_NeverBranch) {
2216       return true;
2217     }
2218   }
2219   return false;
2220 }
2221 #endif
2222 
2223 void MemoryGraphFixer::collect_memory_nodes() {
2224   Node_Stack stack(0);
2225   VectorSet visited(Thread::current()->resource_area());
2226   Node_List regions;
2227 
2228   // Walk the raw memory graph and create a mapping from CFG node to
2229   // memory node. Exclude phis for now.
2230   stack.push(_phase->C->root(), 1);
2231   do {
2232     Node* n = stack.node();
2233     int opc = n->Opcode();
2234     uint i = stack.index();
2235     if (i < n->req()) {
2236       Node* mem = NULL;
2237       if (opc == Op_Root) {
2238         Node* in = n->in(i);
2239         int in_opc = in->Opcode();
2240         if (in_opc == Op_Return || in_opc == Op_Rethrow) {
2241           mem = in->in(TypeFunc::Memory);
2242         } else if (in_opc == Op_Halt) {
2243           if (!in->in(0)->is_Region()) {
2244             Node* proj = in->in(0);
2245             assert(proj->is_Proj(), "");
2246             Node* in = proj->in(0);
2247             assert(in->is_CallStaticJava() || in->Opcode() == Op_NeverBranch || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
2248             if (in->is_CallStaticJava()) {
2249               mem = in->in(TypeFunc::Memory);
2250             } else if (in->Opcode() == Op_Catch) {
2251               Node* call = in->in(0)->in(0);
2252               assert(call->is_Call(), "");
2253               mem = call->in(TypeFunc::Memory);
2254             } else if (in->Opcode() == Op_NeverBranch) {
2255               ResourceMark rm;
2256               Unique_Node_List wq;
2257               wq.push(in);
2258               wq.push(in->as_Multi()->proj_out(0));
2259               for (uint j = 1; j < wq.size(); j++) {
2260                 Node* c = wq.at(j);
2261                 assert(!c->is_Root(), "shouldn't leave loop");
2262                 if (c->is_SafePoint()) {
2263                   assert(mem == NULL, "only one safepoint");
2264                   mem = c->in(TypeFunc::Memory);
2265                 }
2266                 for (DUIterator_Fast kmax, k = c->fast_outs(kmax); k < kmax; k++) {
2267                   Node* u = c->fast_out(k);
2268                   if (u->is_CFG()) {
2269                     wq.push(u);
2270                   }
2271                 }
2272               }
2273               assert(mem != NULL, "should have found safepoint");
2274             }
2275           }
2276         } else {
2277 #ifdef ASSERT
2278           n->dump();
2279           in->dump();
2280 #endif
2281           ShouldNotReachHere();
2282         }
2283       } else {
2284         assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
2285         assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
2286         mem = n->in(i);
2287       }
2288       i++;
2289       stack.set_index(i);
2290       if (mem == NULL) {
2291         continue;
2292       }
2293       for (;;) {
2294         if (visited.test_set(mem->_idx) || mem->is_Start()) {
2295           break;
2296         }
2297         if (mem->is_Phi()) {
2298           stack.push(mem, 2);
2299           mem = mem->in(1);
2300         } else if (mem->is_Proj()) {
2301           stack.push(mem, mem->req());
2302           mem = mem->in(0);
2303         } else if (mem->is_SafePoint() || mem->is_MemBar()) {
2304           mem = mem->in(TypeFunc::Memory);
2305         } else if (mem->is_MergeMem()) {
2306           MergeMemNode* mm = mem->as_MergeMem();
2307           mem = mm->memory_at(_alias);
2308         } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
2309           assert(_alias == Compile::AliasIdxRaw, "");
2310           stack.push(mem, mem->req());
2311           mem = mem->in(MemNode::Memory);
2312         } else {
2313 #ifdef ASSERT
2314           mem->dump();
2315 #endif
2316           ShouldNotReachHere();
2317         }
2318       }
2319     } else {
2320       if (n->is_Phi()) {
2321         // Nothing
2322       } else if (!n->is_Root()) {
2323         Node* c = get_ctrl(n);
2324         _memory_nodes.map(c->_idx, n);
2325       }
2326       stack.pop();
2327     }
2328   } while(stack.is_nonempty());
2329 
2330   // Iterate over CFG nodes in rpo and propagate memory state to
2331   // compute memory state at regions, creating new phis if needed.
2332   Node_List rpo_list;
2333   visited.Clear();
2334   _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
2335   Node* root = rpo_list.pop();
2336   assert(root == _phase->C->root(), "");
2337 
2338   const bool trace = false;
2339 #ifdef ASSERT
2340   if (trace) {
2341     for (int i = rpo_list.size() - 1; i >= 0; i--) {
2342       Node* c = rpo_list.at(i);
2343       if (_memory_nodes[c->_idx] != NULL) {
2344         tty->print("X %d", c->_idx);  _memory_nodes[c->_idx]->dump();
2345       }
2346     }
2347   }
2348 #endif
2349   uint last = _phase->C->unique();
2350 
2351 #ifdef ASSERT
2352   uint8_t max_depth = 0;
2353   for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
2354     IdealLoopTree* lpt = iter.current();
2355     max_depth = MAX2(max_depth, lpt->_nest);
2356   }
2357 #endif
2358 
2359   bool progress = true;
2360   int iteration = 0;
2361   Node_List dead_phis;
2362   while (progress) {
2363     progress = false;
2364     iteration++;
2365     assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2366     if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
2367     IdealLoopTree* last_updated_ilt = NULL;
2368     for (int i = rpo_list.size() - 1; i >= 0; i--) {
2369       Node* c = rpo_list.at(i);
2370 
2371       Node* prev_mem = _memory_nodes[c->_idx];
2372       if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2373         Node* prev_region = regions[c->_idx];
2374         Node* unique = NULL;
2375         for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
2376           Node* m = _memory_nodes[c->in(j)->_idx];
2377           assert(m != NULL || (c->is_Loop() && j == LoopNode::LoopBackControl && iteration == 1) || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "expect memory state");
2378           if (m != NULL) {
2379             if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
2380               assert(c->is_Loop() && j == LoopNode::LoopBackControl || _phase->C->has_irreducible_loop(), "");
2381               // continue
2382             } else if (unique == NULL) {
2383               unique = m;
2384             } else if (m == unique) {
2385               // continue
2386             } else {
2387               unique = NodeSentinel;
2388             }
2389           }
2390         }
2391         assert(unique != NULL, "empty phi???");
2392         if (unique != NodeSentinel) {
2393           if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c) {
2394             dead_phis.push(prev_region);
2395           }
2396           regions.map(c->_idx, unique);
2397         } else {
2398           Node* phi = NULL;
2399           if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
2400             phi = prev_region;
2401             for (uint k = 1; k < c->req(); k++) {
2402               Node* m = _memory_nodes[c->in(k)->_idx];
2403               assert(m != NULL, "expect memory state");
2404               phi->set_req(k, m);
2405             }
2406           } else {
2407             for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == NULL; j++) {
2408               Node* u = c->fast_out(j);
2409               if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2410                   (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
2411                 phi = u;
2412                 for (uint k = 1; k < c->req() && phi != NULL; k++) {
2413                   Node* m = _memory_nodes[c->in(k)->_idx];
2414                   assert(m != NULL, "expect memory state");
2415                   if (u->in(k) != m) {
2416                     phi = NULL;
2417                   }
2418                 }
2419               }
2420             }
2421             if (phi == NULL) {
2422               phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
2423               for (uint k = 1; k < c->req(); k++) {
2424                 Node* m = _memory_nodes[c->in(k)->_idx];
2425                 assert(m != NULL, "expect memory state");
2426                 phi->init_req(k, m);
2427               }
2428             }
2429           }
2430           assert(phi != NULL, "");
2431           regions.map(c->_idx, phi);
2432         }
2433         Node* current_region = regions[c->_idx];
2434         if (current_region != prev_region) {
2435           progress = true;
2436           if (prev_region == prev_mem) {
2437             _memory_nodes.map(c->_idx, current_region);
2438           }
2439         }
2440       } else if (prev_mem == NULL || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
2441         Node* m = _memory_nodes[_phase->idom(c)->_idx];
2442         assert(m != NULL, "expect memory state");
2443         if (m != prev_mem) {
2444           _memory_nodes.map(c->_idx, m);
2445           progress = true;
2446         }
2447       }
2448 #ifdef ASSERT
2449       if (trace) { tty->print("X %d", c->_idx);  _memory_nodes[c->_idx]->dump(); }
2450 #endif
2451     }
2452   }
2453 
2454   // Replace existing phi with computed memory state for that region
2455   // if different (could be a new phi or a dominating memory node if
2456   // that phi was found to be useless).
2457   while (dead_phis.size() > 0) {
2458     Node* n = dead_phis.pop();
2459     n->replace_by(_phase->C->top());
2460     n->destruct();
2461   }
2462   for (int i = rpo_list.size() - 1; i >= 0; i--) {
2463     Node* c = rpo_list.at(i);
2464     if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2465       Node* n = regions[c->_idx];
2466       if (n->is_Phi() && n->_idx >= last && n->in(0) == c) {
2467         _phase->register_new_node(n, c);
2468       }
2469     }
2470   }
2471   for (int i = rpo_list.size() - 1; i >= 0; i--) {
2472     Node* c = rpo_list.at(i);
2473     if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2474       Node* n = regions[c->_idx];
2475       for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
2476         Node* u = c->fast_out(i);
2477         if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2478             u != n) {
2479           if (u->adr_type() == TypePtr::BOTTOM) {
2480             fix_memory_uses(u, n, n, c);
2481           } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2482             _phase->lazy_replace(u, n);
2483             --i; --imax;
2484           }
2485         }
2486       }
2487     }
2488   }
2489 }
2490 
2491 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
2492   Node* c = _phase->get_ctrl(n);
2493   if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Call()) {
2494     assert(c == n->in(0), "");
2495     CallNode* call = c->as_Call();
2496     CallProjections projs;
2497     call->extract_projections(&projs, true, false);
2498     if (projs.catchall_memproj != NULL) {
2499       if (projs.fallthrough_memproj == n) {
2500         c = projs.fallthrough_catchproj;
2501       } else {
2502         assert(projs.catchall_memproj == n, "");
2503         c = projs.catchall_catchproj;
2504       }
2505     }
2506   }
2507   return c;
2508 }
2509 
2510 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
2511   if (_phase->has_ctrl(n))
2512     return get_ctrl(n);
2513   else {
2514     assert (n->is_CFG(), "must be a CFG node");
2515     return n;
2516   }
2517 }
2518 
2519 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
2520   return m != NULL && get_ctrl(m) == c;
2521 }
2522 
2523 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
2524   assert(n == NULL || _phase->ctrl_or_self(n) == ctrl, "");
2525   Node* mem = _memory_nodes[ctrl->_idx];
2526   Node* c = ctrl;
2527   while (!mem_is_valid(mem, c) &&
2528          (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
2529     c = _phase->idom(c);
2530     mem = _memory_nodes[c->_idx];
2531   }
2532   if (n != NULL && mem_is_valid(mem, c)) {
2533     while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
2534       mem = next_mem(mem, _alias);
2535     }
2536     if (mem->is_MergeMem()) {
2537       mem = mem->as_MergeMem()->memory_at(_alias);
2538     }
2539     if (!mem_is_valid(mem, c)) {
2540       do {
2541         c = _phase->idom(c);
2542         mem = _memory_nodes[c->_idx];
2543       } while (!mem_is_valid(mem, c) &&
2544                (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
2545     }
2546   }
2547   assert(mem->bottom_type() == Type::MEMORY, "");
2548   return mem;
2549 }
2550 
2551 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
2552   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2553     Node* use = region->fast_out(i);
2554     if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
2555         (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
2556       return true;
2557     }
2558   }
2559   return false;
2560 }
2561 
2562 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
2563   assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
2564   const bool trace = false;
2565   DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
2566   DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
2567   GrowableArray<Node*> phis;
2568   if (mem_for_ctrl != mem) {
2569     Node* old = mem_for_ctrl;
2570     Node* prev = NULL;
2571     while (old != mem) {
2572       prev = old;
2573       if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
2574         assert(_alias == Compile::AliasIdxRaw, "");
2575         old = old->in(MemNode::Memory);
2576       } else if (old->Opcode() == Op_SCMemProj) {
2577         assert(_alias == Compile::AliasIdxRaw, "");
2578         old = old->in(0);
2579       } else {
2580         ShouldNotReachHere();
2581       }
2582     }
2583     assert(prev != NULL, "");
2584     if (new_ctrl != ctrl) {
2585       _memory_nodes.map(ctrl->_idx, mem);
2586       _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2587     }
2588     uint input = (uint)MemNode::Memory;
2589     _phase->igvn().replace_input_of(prev, input, new_mem);
2590   } else {
2591     uses.clear();
2592     _memory_nodes.map(new_ctrl->_idx, new_mem);
2593     uses.push(new_ctrl);
2594     for(uint next = 0; next < uses.size(); next++ ) {
2595       Node *n = uses.at(next);
2596       assert(n->is_CFG(), "");
2597       DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
2598       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2599         Node* u = n->fast_out(i);
2600         if (!u->is_Root() && u->is_CFG() && u != n) {
2601           Node* m = _memory_nodes[u->_idx];
2602           if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) &&
2603               !has_mem_phi(u) &&
2604               u->unique_ctrl_out()->Opcode() != Op_Halt) {
2605             DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
2606             DEBUG_ONLY(if (trace && m != NULL) { tty->print("ZZZ mem"); m->dump(); });
2607 
2608             if (!mem_is_valid(m, u) || !m->is_Phi()) {
2609               bool push = true;
2610               bool create_phi = true;
2611               if (_phase->is_dominator(new_ctrl, u)) {
2612                 create_phi = false;
2613               } else if (!_phase->C->has_irreducible_loop()) {
2614                 IdealLoopTree* loop = _phase->get_loop(ctrl);
2615                 bool do_check = true;
2616                 IdealLoopTree* l = loop;
2617                 create_phi = false;
2618                 while (l != _phase->ltree_root()) {
2619                   Node* head = l->_head;
2620                   if (head->in(0) == NULL) {
2621                     head = _phase->get_ctrl(head);
2622                   }
2623                   if (_phase->is_dominator(head, u) && _phase->is_dominator(_phase->idom(u), head)) {
2624                     create_phi = true;
2625                     do_check = false;
2626                     break;
2627                   }
2628                   l = l->_parent;
2629                 }
2630 
2631                 if (do_check) {
2632                   assert(!create_phi, "");
2633                   IdealLoopTree* u_loop = _phase->get_loop(u);
2634                   if (u_loop != _phase->ltree_root() && u_loop->is_member(loop)) {
2635                     Node* c = ctrl;
2636                     while (!_phase->is_dominator(c, u_loop->tail())) {
2637                       c = _phase->idom(c);
2638                     }
2639                     if (!_phase->is_dominator(c, u)) {
2640                       do_check = false;
2641                     }
2642                   }
2643                 }
2644 
2645                 if (do_check && _phase->is_dominator(_phase->idom(u), new_ctrl)) {
2646                   create_phi = true;
2647                 }
2648               }
2649               if (create_phi) {
2650                 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
2651                 _phase->register_new_node(phi, u);
2652                 phis.push(phi);
2653                 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
2654                 if (!mem_is_valid(m, u)) {
2655                   DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
2656                   _memory_nodes.map(u->_idx, phi);
2657                 } else {
2658                   DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
2659                   for (;;) {
2660                     assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), "");
2661                     Node* next = NULL;
2662                     if (m->is_Proj()) {
2663                       next = m->in(0);
2664                     } else {
2665                       assert(m->is_Mem() || m->is_LoadStore(), "");
2666                       assert(_alias == Compile::AliasIdxRaw, "");
2667                       next = m->in(MemNode::Memory);
2668                     }
2669                     if (_phase->get_ctrl(next) != u) {
2670                       break;
2671                     }
2672                     if (next->is_MergeMem()) {
2673                       assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
2674                       break;
2675                     }
2676                     if (next->is_Phi()) {
2677                       assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
2678                       break;
2679                     }
2680                     m = next;
2681                   }
2682 
2683                   DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
2684                   assert(m->is_Mem() || m->is_LoadStore(), "");
2685                   uint input = (uint)MemNode::Memory;
2686                   _phase->igvn().replace_input_of(m, input, phi);
2687                   push = false;
2688                 }
2689               } else {
2690                 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
2691               }
2692               if (push) {
2693                 uses.push(u);
2694               }
2695             }
2696           } else if (!mem_is_valid(m, u) &&
2697                      !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1)) {
2698             uses.push(u);
2699           }
2700         }
2701       }
2702     }
2703     for (int i = 0; i < phis.length(); i++) {
2704       Node* n = phis.at(i);
2705       Node* r = n->in(0);
2706       DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
2707       for (uint j = 1; j < n->req(); j++) {
2708         Node* m = find_mem(r->in(j), NULL);
2709         _phase->igvn().replace_input_of(n, j, m);
2710         DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
2711       }
2712     }
2713   }
2714   uint last = _phase->C->unique();
2715   MergeMemNode* mm = NULL;
2716   int alias = _alias;
2717   DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
2718   for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2719     Node* u = mem->out(i);
2720     if (u->_idx < last) {
2721       if (u->is_Mem()) {
2722         if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2723           Node* m = find_mem(_phase->get_ctrl(u), u);
2724           if (m != mem) {
2725             DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2726             _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2727             --i;
2728           }
2729         }
2730       } else if (u->is_MergeMem()) {
2731         MergeMemNode* u_mm = u->as_MergeMem();
2732         if (u_mm->memory_at(alias) == mem) {
2733           MergeMemNode* newmm = NULL;
2734           for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2735             Node* uu = u->fast_out(j);
2736             assert(!uu->is_MergeMem(), "chain of MergeMems?");
2737             if (uu->is_Phi()) {
2738               assert(uu->adr_type() == TypePtr::BOTTOM, "");
2739               Node* region = uu->in(0);
2740               int nb = 0;
2741               for (uint k = 1; k < uu->req(); k++) {
2742                 if (uu->in(k) == u) {
2743                   Node* m = find_mem(region->in(k), NULL);
2744                   if (m != mem) {
2745                     DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
2746                     newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2747                     if (newmm != u) {
2748                       _phase->igvn().replace_input_of(uu, k, newmm);
2749                       nb++;
2750                       --jmax;
2751                     }
2752                   }
2753                 }
2754               }
2755               if (nb > 0) {
2756                 --j;
2757               }
2758             } else {
2759               Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
2760               if (m != mem) {
2761                 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
2762                 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2763                 if (newmm != u) {
2764                   _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2765                   --j, --jmax;
2766                 }
2767               }
2768             }
2769           }
2770         }
2771       } else if (u->is_Phi()) {
2772         assert(u->bottom_type() == Type::MEMORY, "what else?");
2773         if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
2774           Node* region = u->in(0);
2775           bool replaced = false;
2776           for (uint j = 1; j < u->req(); j++) {
2777             if (u->in(j) == mem) {
2778               Node* m = find_mem(region->in(j), NULL);
2779               Node* nnew = m;
2780               if (m != mem) {
2781                 if (u->adr_type() == TypePtr::BOTTOM) {
2782                   mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
2783                   nnew = mm;
2784                 }
2785                 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
2786                 _phase->igvn().replace_input_of(u, j, nnew);
2787                 replaced = true;
2788               }
2789             }
2790           }
2791           if (replaced) {
2792             --i;
2793           }
2794         }
2795       } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2796                  u->adr_type() == NULL) {
2797         assert(u->adr_type() != NULL ||
2798                u->Opcode() == Op_Rethrow ||
2799                u->Opcode() == Op_Return ||
2800                u->Opcode() == Op_SafePoint ||
2801                (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2802                (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2803                u->Opcode() == Op_CallLeaf, "");
2804         Node* m = find_mem(_phase->ctrl_or_self(u), u);
2805         if (m != mem) {
2806           mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
2807           _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2808           --i;
2809         }
2810       } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2811         Node* m = find_mem(_phase->ctrl_or_self(u), u);
2812         if (m != mem) {
2813           DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2814           _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2815           --i;
2816         }
2817       } else if (u->adr_type() != TypePtr::BOTTOM &&
2818                  _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
2819         Node* m = find_mem(_phase->ctrl_or_self(u), u);
2820         assert(m != mem, "");
2821         // u is on the wrong slice...
2822         assert(u->is_ClearArray(), "");
2823         DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2824         _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2825         --i;
2826       }
2827     }
2828   }
2829 #ifdef ASSERT
2830   assert(new_mem->outcnt() > 0, "");
2831   for (int i = 0; i < phis.length(); i++) {
2832     Node* n = phis.at(i);
2833     assert(n->outcnt() > 0, "new phi must have uses now");
2834   }
2835 #endif
2836 }
2837 
2838 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
2839   MergeMemNode* mm = MergeMemNode::make(mem);
2840   mm->set_memory_at(_alias, rep_proj);
2841   _phase->register_new_node(mm, rep_ctrl);
2842   return mm;
2843 }
2844 
2845 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
2846   MergeMemNode* newmm = NULL;
2847   MergeMemNode* u_mm = u->as_MergeMem();
2848   Node* c = _phase->get_ctrl(u);
2849   if (_phase->is_dominator(c, rep_ctrl)) {
2850     c = rep_ctrl;
2851   } else {
2852     assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
2853   }
2854   if (u->outcnt() == 1) {
2855     if (u->req() > (uint)_alias && u->in(_alias) == mem) {
2856       _phase->igvn().replace_input_of(u, _alias, rep_proj);
2857       --i;
2858     } else {
2859       _phase->igvn().rehash_node_delayed(u);
2860       u_mm->set_memory_at(_alias, rep_proj);
2861     }
2862     newmm = u_mm;
2863     _phase->set_ctrl_and_loop(u, c);
2864   } else {
2865     // can't simply clone u and then change one of its input because
2866     // it adds and then removes an edge which messes with the
2867     // DUIterator
2868     newmm = MergeMemNode::make(u_mm->base_memory());
2869     for (uint j = 0; j < u->req(); j++) {
2870       if (j < newmm->req()) {
2871         if (j == (uint)_alias) {
2872           newmm->set_req(j, rep_proj);
2873         } else if (newmm->in(j) != u->in(j)) {
2874           newmm->set_req(j, u->in(j));
2875         }
2876       } else if (j == (uint)_alias) {
2877         newmm->add_req(rep_proj);
2878       } else {
2879         newmm->add_req(u->in(j));
2880       }
2881     }
2882     if ((uint)_alias >= u->req()) {
2883       newmm->set_memory_at(_alias, rep_proj);
2884     }
2885     _phase->register_new_node(newmm, c);
2886   }
2887   return newmm;
2888 }
2889 
2890 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
2891   if (phi->adr_type() == TypePtr::BOTTOM) {
2892     Node* region = phi->in(0);
2893     for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
2894       Node* uu = region->fast_out(j);
2895       if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
2896         return false;
2897       }
2898     }
2899     return true;
2900   }
2901   return _phase->C->get_alias_index(phi->adr_type()) == _alias;
2902 }
2903 
2904 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
2905   uint last = _phase-> C->unique();
2906   MergeMemNode* mm = NULL;
2907   assert(mem->bottom_type() == Type::MEMORY, "");
2908   for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2909     Node* u = mem->out(i);
2910     if (u != replacement && u->_idx < last) {
2911       if (u->is_MergeMem()) {
2912         MergeMemNode* u_mm = u->as_MergeMem();
2913         if (u_mm->memory_at(_alias) == mem) {
2914           MergeMemNode* newmm = NULL;
2915           for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2916             Node* uu = u->fast_out(j);
2917             assert(!uu->is_MergeMem(), "chain of MergeMems?");
2918             if (uu->is_Phi()) {
2919               if (should_process_phi(uu)) {
2920                 Node* region = uu->in(0);
2921                 int nb = 0;
2922                 for (uint k = 1; k < uu->req(); k++) {
2923                   if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
2924                     if (newmm == NULL) {
2925                       newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2926                     }
2927                     if (newmm != u) {
2928                       _phase->igvn().replace_input_of(uu, k, newmm);
2929                       nb++;
2930                       --jmax;
2931                     }
2932                   }
2933                 }
2934                 if (nb > 0) {
2935                   --j;
2936                 }
2937               }
2938             } else {
2939               if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
2940                 if (newmm == NULL) {
2941                   newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2942                 }
2943                 if (newmm != u) {
2944                   _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2945                   --j, --jmax;
2946                 }
2947               }
2948             }
2949           }
2950         }
2951       } else if (u->is_Phi()) {
2952         assert(u->bottom_type() == Type::MEMORY, "what else?");
2953         Node* region = u->in(0);
2954         if (should_process_phi(u)) {
2955           bool replaced = false;
2956           for (uint j = 1; j < u->req(); j++) {
2957             if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
2958               Node* nnew = rep_proj;
2959               if (u->adr_type() == TypePtr::BOTTOM) {
2960                 if (mm == NULL) {
2961                   mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2962                 }
2963                 nnew = mm;
2964               }
2965               _phase->igvn().replace_input_of(u, j, nnew);
2966               replaced = true;
2967             }
2968           }
2969           if (replaced) {
2970             --i;
2971           }
2972 
2973         }
2974       } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2975                  u->adr_type() == NULL) {
2976         assert(u->adr_type() != NULL ||
2977                u->Opcode() == Op_Rethrow ||
2978                u->Opcode() == Op_Return ||
2979                u->Opcode() == Op_SafePoint ||
2980                u->Opcode() == Op_StoreLConditional ||
2981                (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2982                (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2983                u->Opcode() == Op_CallLeaf, "");
2984         if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2985           if (mm == NULL) {
2986             mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2987           }
2988           _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2989           --i;
2990         }
2991       } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2992         if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2993           _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
2994           --i;
2995         }
2996       }
2997     }
2998   }
2999 }
3000 
3001 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj)
3002 : Node(ctrl, obj) {
3003   ShenandoahBarrierSetC2::bsc2()->state()->add_load_reference_barrier(this);
3004 }
3005 
3006 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const {
3007   if (in(ValueIn) == NULL || in(ValueIn)->is_top()) {
3008     return Type::TOP;
3009   }
3010   const Type* t = in(ValueIn)->bottom_type();
3011   if (t == TypePtr::NULL_PTR) {
3012     return t;
3013   }
3014   return t->is_oopptr();
3015 }
3016 
3017 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseGVN* phase) const {
3018   // Either input is TOP ==> the result is TOP
3019   const Type *t2 = phase->type(in(ValueIn));
3020   if( t2 == Type::TOP ) return Type::TOP;
3021 
3022   if (t2 == TypePtr::NULL_PTR) {
3023     return t2;
3024   }
3025 
3026   const Type* type = t2->is_oopptr()/*->cast_to_nonconst()*/;
3027   return type;
3028 }
3029 
3030 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseGVN* phase) {
3031   Node* value = in(ValueIn);
3032   if (!needs_barrier(phase, value)) {
3033     return value;
3034   }
3035   return this;
3036 }
3037 
3038 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseGVN* phase, Node* n) {
3039   Unique_Node_List visited;
3040   return needs_barrier_impl(phase, n, visited);
3041 }
3042 
3043 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseGVN* phase, Node* n, Unique_Node_List &visited) {
3044   if (n == NULL) return false;
3045   if (visited.member(n)) {
3046     return false; // Been there.
3047   }
3048   visited.push(n);
3049 
3050   if (n->is_Allocate()) {
3051     // tty->print_cr("optimize barrier on alloc");
3052     return false;
3053   }
3054   if (n->is_Call()) {
3055     // tty->print_cr("optimize barrier on call");
3056     return false;
3057   }
3058 
3059   const Type* type = phase->type(n);
3060   if (type == Type::TOP) {
3061     return false;
3062   }
3063   if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) {
3064     // tty->print_cr("optimize barrier on null");
3065     return false;
3066   }
3067   if (type->make_oopptr() && type->make_oopptr()->const_oop() != NULL) {
3068     // tty->print_cr("optimize barrier on constant");
3069     return false;
3070   }
3071 
3072   switch (n->Opcode()) {
3073     case Op_AddP:
3074       return true; // TODO: Can refine?
3075     case Op_LoadP:
3076     case Op_ShenandoahCompareAndExchangeN:
3077     case Op_ShenandoahCompareAndExchangeP:
3078     case Op_CompareAndExchangeN:
3079     case Op_CompareAndExchangeP:
3080     case Op_GetAndSetN:
3081     case Op_GetAndSetP:
3082       return true;
3083     case Op_Phi: {
3084       for (uint i = 1; i < n->req(); i++) {
3085         if (needs_barrier_impl(phase, n->in(i), visited)) return true;
3086       }
3087       return false;
3088     }
3089     case Op_CheckCastPP:
3090     case Op_CastPP:
3091       return needs_barrier_impl(phase, n->in(1), visited);
3092     case Op_Proj:
3093       return needs_barrier_impl(phase, n->in(0), visited);
3094     case Op_ShenandoahLoadReferenceBarrier:
3095       // tty->print_cr("optimize barrier on barrier");
3096       return false;
3097     case Op_Parm:
3098       // tty->print_cr("optimize barrier on input arg");
3099       return false;
3100     case Op_DecodeN:
3101     case Op_EncodeP:
3102       return needs_barrier_impl(phase, n->in(1), visited);
3103     case Op_LoadN:
3104       return true;
3105     case Op_CMoveP:
3106       return needs_barrier_impl(phase, n->in(2), visited) ||
3107              needs_barrier_impl(phase, n->in(3), visited);
3108     case Op_ShenandoahEnqueueBarrier:
3109       return needs_barrier_impl(phase, n->in(1), visited);
3110     default:
3111       break;
3112   }
3113 #ifdef ASSERT
3114   tty->print("need barrier on?: ");
3115   tty->print_cr("ins:");
3116   n->dump(2);
3117   tty->print_cr("outs:");
3118   n->dump(-2);
3119   ShouldNotReachHere();
3120 #endif
3121   return true;
3122 }
3123 
3124 ShenandoahLoadReferenceBarrierNode::Strength ShenandoahLoadReferenceBarrierNode::get_barrier_strength() {
3125   Unique_Node_List visited;
3126   Node_Stack stack(0);
3127   stack.push(this, 0);
3128   Strength strength = NONE;
3129   while (strength != STRONG && stack.size() > 0) {
3130     Node* n = stack.node();
3131     if (visited.member(n)) {
3132       stack.pop();
3133       continue;
3134     }
3135     visited.push(n);
3136     bool visit_users = false;
3137     switch (n->Opcode()) {
3138       case Op_StoreN:
3139       case Op_StoreP: {
3140         strength = STRONG;
3141         break;
3142       }
3143       case Op_CmpP: {
3144         if (!n->in(1)->bottom_type()->higher_equal(TypePtr::NULL_PTR) &&
3145             !n->in(2)->bottom_type()->higher_equal(TypePtr::NULL_PTR)) {
3146           strength = STRONG;
3147         }
3148         break;
3149       }
3150       case Op_CallStaticJava: {
3151         strength = STRONG;
3152         break;
3153       }
3154       case Op_CallDynamicJava:
3155       case Op_CallLeaf:
3156       case Op_CallLeafNoFP:
3157       case Op_CompareAndSwapL:
3158       case Op_CompareAndSwapI:
3159       case Op_CompareAndSwapB:
3160       case Op_CompareAndSwapS:
3161       case Op_CompareAndSwapN:
3162       case Op_CompareAndSwapP:
3163       case Op_CompareAndExchangeL:
3164       case Op_CompareAndExchangeI:
3165       case Op_CompareAndExchangeB:
3166       case Op_CompareAndExchangeS:
3167       case Op_CompareAndExchangeN:
3168       case Op_CompareAndExchangeP:
3169       case Op_WeakCompareAndSwapL:
3170       case Op_WeakCompareAndSwapI:
3171       case Op_WeakCompareAndSwapB:
3172       case Op_WeakCompareAndSwapS:
3173       case Op_WeakCompareAndSwapN:
3174       case Op_WeakCompareAndSwapP:
3175       case Op_ShenandoahCompareAndSwapN:
3176       case Op_ShenandoahCompareAndSwapP:
3177       case Op_ShenandoahWeakCompareAndSwapN:
3178       case Op_ShenandoahWeakCompareAndSwapP:
3179       case Op_ShenandoahCompareAndExchangeN:
3180       case Op_ShenandoahCompareAndExchangeP:
3181       case Op_GetAndSetL:
3182       case Op_GetAndSetI:
3183       case Op_GetAndSetB:
3184       case Op_GetAndSetS:
3185       case Op_GetAndSetP:
3186       case Op_GetAndSetN:
3187       case Op_GetAndAddL:
3188       case Op_GetAndAddI:
3189       case Op_GetAndAddB:
3190       case Op_GetAndAddS:
3191       case Op_ShenandoahEnqueueBarrier:
3192       case Op_FastLock:
3193       case Op_FastUnlock:
3194       case Op_Rethrow:
3195       case Op_Return:
3196       case Op_StoreB:
3197       case Op_StoreC:
3198       case Op_StoreD:
3199       case Op_StoreF:
3200       case Op_StoreL:
3201       case Op_StoreLConditional:
3202       case Op_StoreI:
3203       case Op_StoreVector:
3204       case Op_StrInflatedCopy:
3205       case Op_StrCompressedCopy:
3206       case Op_EncodeP:
3207       case Op_CastP2X:
3208       case Op_SafePoint:
3209       case Op_EncodeISOArray:
3210         strength = STRONG;
3211         break;
3212       case Op_LoadB:
3213       case Op_LoadUB:
3214       case Op_LoadUS:
3215       case Op_LoadD:
3216       case Op_LoadF:
3217       case Op_LoadL:
3218       case Op_LoadI:
3219       case Op_LoadS:
3220       case Op_LoadN:
3221       case Op_LoadP:
3222       case Op_LoadVector: {
3223         const TypePtr* adr_type = n->adr_type();
3224         int alias_idx = Compile::current()->get_alias_index(adr_type);
3225         Compile::AliasType* alias_type = Compile::current()->alias_type(alias_idx);
3226         ciField* field = alias_type->field();
3227         bool is_static = field != NULL && field->is_static();
3228         bool is_final = field != NULL && field->is_final();
3229         bool is_stable = field != NULL && field->is_stable();
3230         if (ShenandoahOptimizeStaticFinals && is_static && is_final) {
3231           // Leave strength as is.
3232         } else if (ShenandoahOptimizeInstanceFinals && !is_static && is_final) {
3233           // Leave strength as is.
3234         } else if (ShenandoahOptimizeStableFinals && (is_stable || (adr_type->isa_aryptr() && adr_type->isa_aryptr()->is_stable()))) {
3235           // Leave strength as is.
3236         } else {
3237           strength = WEAK;
3238         }
3239         break;
3240       }
3241       case Op_AryEq: {
3242         Node* n1 = n->in(2);
3243         Node* n2 = n->in(3);
3244         if (!ShenandoahOptimizeStableFinals ||
3245             !n1->bottom_type()->isa_aryptr() || !n1->bottom_type()->isa_aryptr()->is_stable() ||
3246             !n2->bottom_type()->isa_aryptr() || !n2->bottom_type()->isa_aryptr()->is_stable()) {
3247           strength = WEAK;
3248         }
3249         break;
3250       }
3251       case Op_StrEquals:
3252       case Op_StrComp:
3253       case Op_StrIndexOf:
3254       case Op_StrIndexOfChar:
3255         if (!ShenandoahOptimizeStableFinals) {
3256            strength = WEAK;
3257         }
3258         break;
3259       case Op_Conv2B:
3260       case Op_LoadRange:
3261       case Op_LoadKlass:
3262       case Op_LoadNKlass:
3263         // NONE, i.e. leave current strength as is
3264         break;
3265       case Op_AddP:
3266       case Op_CheckCastPP:
3267       case Op_CastPP:
3268       case Op_CMoveP:
3269       case Op_Phi:
3270       case Op_ShenandoahLoadReferenceBarrier:
3271         visit_users = true;
3272         break;
3273       default: {
3274 #ifdef ASSERT
3275         tty->print_cr("Unknown node in get_barrier_strength:");
3276         n->dump(1);
3277         ShouldNotReachHere();
3278 #else
3279         strength = STRONG;
3280 #endif
3281       }
3282     }
3283 #ifdef ASSERT
3284 /*
3285     if (strength == STRONG) {
3286       tty->print("strengthening node: ");
3287       n->dump();
3288     }
3289     */
3290 #endif
3291     stack.pop();
3292     if (visit_users) {
3293       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3294         Node* user = n->fast_out(i);
3295         if (user != NULL) {
3296           stack.push(user, 0);
3297         }
3298       }
3299     }
3300   }
3301   return strength;
3302 }
3303 
3304 CallStaticJavaNode* ShenandoahLoadReferenceBarrierNode::pin_and_expand_null_check(PhaseIterGVN& igvn) {
3305   Node* val = in(ValueIn);
3306 
3307   const Type* val_t = igvn.type(val);
3308 
3309   if (val_t->meet(TypePtr::NULL_PTR) != val_t &&
3310       val->Opcode() == Op_CastPP &&
3311       val->in(0) != NULL &&
3312       val->in(0)->Opcode() == Op_IfTrue &&
3313       val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
3314       val->in(0)->in(0)->is_If() &&
3315       val->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
3316       val->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
3317       val->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
3318       val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1) &&
3319       val->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
3320     assert(val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1), "");
3321     CallStaticJavaNode* unc = val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
3322     return unc;
3323   }
3324   return NULL;
3325 }