< prev index next >

src/hotspot/share/opto/matcher.cpp

Print this page

 166   Unique_Node_List worklist;
 167   VectorSet visited;
 168   worklist.push(xroot);
 169   while (worklist.size() > 0) {
 170     Node* n = worklist.pop();
 171     visited.set(n->_idx);
 172     assert(C->node_arena()->contains(n), "dead node");
 173     for (uint j = 0; j < n->req(); j++) {
 174       Node* in = n->in(j);
 175       if (in != NULL) {
 176         assert(C->node_arena()->contains(in), "dead node");
 177         if (!visited.test(in->_idx)) {
 178           worklist.push(in);
 179         }
 180       }
 181     }
 182   }
 183 }
 184 #endif
 185 












































 186 
 187 //---------------------------match---------------------------------------------
 188 void Matcher::match( ) {
 189   if( MaxLabelRootDepth < 100 ) { // Too small?
 190     assert(false, "invalid MaxLabelRootDepth, increase it to 100 minimum");
 191     MaxLabelRootDepth = 100;
 192   }
 193   // One-time initialization of some register masks.
 194   init_spill_mask( C->root()->in(1) );
 195   _return_addr_mask = return_addr();
 196 #ifdef _LP64
 197   // Pointers take 2 slots in 64-bit land
 198   _return_addr_mask.Insert(OptoReg::add(return_addr(),1));
 199 #endif
 200 
 201   // Map a Java-signature return type into return register-value
 202   // machine registers for 0, 1 and 2 returned values.
 203   const TypeTuple *range = C->tf()->range();
 204   if( range->cnt() > TypeFunc::Parms ) { // If not a void function
 205     // Get ideal-register return type
 206     uint ireg = range->field_at(TypeFunc::Parms)->ideal_reg();
 207     // Get machine return register
 208     uint sop = C->start()->Opcode();
 209     OptoRegPair regs = return_value(ireg);
 210 
 211     // And mask for same
 212     _return_value_mask = RegMask(regs.first());
 213     if( OptoReg::is_valid(regs.second()) )
 214       _return_value_mask.Insert(regs.second());
 215   }
 216 
 217   // ---------------
 218   // Frame Layout
 219 
 220   // Need the method signature to determine the incoming argument types,
 221   // because the types determine which registers the incoming arguments are
 222   // in, and this affects the matched code.
 223   const TypeTuple *domain = C->tf()->domain();
 224   uint             argcnt = domain->cnt() - TypeFunc::Parms;
 225   BasicType *sig_bt        = NEW_RESOURCE_ARRAY( BasicType, argcnt );
 226   VMRegPair *vm_parm_regs  = NEW_RESOURCE_ARRAY( VMRegPair, argcnt );
 227   _parm_regs               = NEW_RESOURCE_ARRAY( OptoRegPair, argcnt );
 228   _calling_convention_mask = NEW_RESOURCE_ARRAY( RegMask, argcnt );
 229   uint i;
 230   for( i = 0; i<argcnt; i++ ) {
 231     sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type();
 232   }
 233 
 234   // Pass array of ideal registers and length to USER code (from the AD file)
 235   // that will convert this to an array of register numbers.
 236   const StartNode *start = C->start();
 237   start->calling_convention( sig_bt, vm_parm_regs, argcnt );
 238 #ifdef ASSERT
 239   // Sanity check users' calling convention.  Real handy while trying to
 240   // get the initial port correct.
 241   { for (uint i = 0; i<argcnt; i++) {
 242       if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) {
 243         assert(domain->field_at(i+TypeFunc::Parms)==Type::HALF, "only allowed on halve" );

 507   idealreg2mhdebugmask[Op_VecS] = &rms[31];
 508   idealreg2mhdebugmask[Op_VecD] = &rms[32];
 509   idealreg2mhdebugmask[Op_VecX] = &rms[33];
 510   idealreg2mhdebugmask[Op_VecY] = &rms[34];
 511   idealreg2mhdebugmask[Op_VecZ] = &rms[35];
 512 
 513   idealreg2spillmask  [Op_RegVectMask] = &rms[36];
 514   idealreg2debugmask  [Op_RegVectMask] = &rms[37];
 515   idealreg2mhdebugmask[Op_RegVectMask] = &rms[38];
 516 
 517   OptoReg::Name i;
 518 
 519   // At first, start with the empty mask
 520   C->FIRST_STACK_mask().Clear();
 521 
 522   // Add in the incoming argument area
 523   OptoReg::Name init_in = OptoReg::add(_old_SP, C->out_preserve_stack_slots());
 524   for (i = init_in; i < _in_arg_limit; i = OptoReg::add(i,1)) {
 525     C->FIRST_STACK_mask().Insert(i);
 526   }

 527   // Add in all bits past the outgoing argument area
 528   guarantee(RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1)),
 529             "must be able to represent all call arguments in reg mask");
 530   OptoReg::Name init = _out_arg_limit;
 531   for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) {
 532     C->FIRST_STACK_mask().Insert(i);
 533   }
 534   // Finally, set the "infinite stack" bit.
 535   C->FIRST_STACK_mask().set_AllStack();
 536 
 537   // Make spill masks.  Registers for their class, plus FIRST_STACK_mask.
 538   RegMask aligned_stack_mask = C->FIRST_STACK_mask();
 539   // Keep spill masks aligned.
 540   aligned_stack_mask.clear_to_pairs();
 541   assert(aligned_stack_mask.is_AllStack(), "should be infinite stack");
 542   RegMask scalable_stack_mask = aligned_stack_mask;
 543 
 544   *idealreg2spillmask[Op_RegP] = *idealreg2regmask[Op_RegP];
 545 #ifdef _LP64
 546   *idealreg2spillmask[Op_RegN] = *idealreg2regmask[Op_RegN];

 765     _register_save_policy[reg] == 'E' ||
 766     _register_save_policy[reg] == 'A'; // Save-on-entry register?
 767 }
 768 
 769 //---------------------------Fixup_Save_On_Entry-------------------------------
 770 void Matcher::Fixup_Save_On_Entry( ) {
 771   init_first_stack_mask();
 772 
 773   Node *root = C->root();       // Short name for root
 774   // Count number of save-on-entry registers.
 775   uint soe_cnt = number_of_saved_registers();
 776   uint i;
 777 
 778   // Find the procedure Start Node
 779   StartNode *start = C->start();
 780   assert( start, "Expect a start node" );
 781 
 782   // Input RegMask array shared by all Returns.
 783   // The type for doubles and longs has a count of 2, but
 784   // there is only 1 returned value
 785   uint ret_edge_cnt = TypeFunc::Parms + ((C->tf()->range()->cnt() == TypeFunc::Parms) ? 0 : 1);
 786   RegMask *ret_rms  = init_input_masks( ret_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
 787   // Returns have 0 or 1 returned values depending on call signature.
 788   // Return register is specified by return_value in the AD file.
 789   if (ret_edge_cnt > TypeFunc::Parms)
 790     ret_rms[TypeFunc::Parms+0] = _return_value_mask;
 791 
 792   // Input RegMask array shared by all Rethrows.
 793   uint reth_edge_cnt = TypeFunc::Parms+1;
 794   RegMask *reth_rms  = init_input_masks( reth_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
 795   // Rethrow takes exception oop only, but in the argument 0 slot.
 796   OptoReg::Name reg = find_receiver();
 797   if (reg >= 0) {
 798     reth_rms[TypeFunc::Parms] = mreg2regmask[reg];
 799 #ifdef _LP64
 800     // Need two slots for ptrs in 64-bit land
 801     reth_rms[TypeFunc::Parms].Insert(OptoReg::add(OptoReg::Name(reg), 1));
 802 #endif
 803   }
 804 
 805   // Input RegMask array shared by all TailCalls
 806   uint tail_call_edge_cnt = TypeFunc::Parms+2;
 807   RegMask *tail_call_rms = init_input_masks( tail_call_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
 808 
 809   // Input RegMask array shared by all TailJumps
 810   uint tail_jump_edge_cnt = TypeFunc::Parms+2;

 837   }
 838 
 839   // Input RegMask array shared by all Halts
 840   uint halt_edge_cnt = TypeFunc::Parms;
 841   RegMask *halt_rms = init_input_masks( halt_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
 842 
 843   // Capture the return input masks into each exit flavor
 844   for( i=1; i < root->req(); i++ ) {
 845     MachReturnNode *exit = root->in(i)->as_MachReturn();
 846     switch( exit->ideal_Opcode() ) {
 847       case Op_Return   : exit->_in_rms = ret_rms;  break;
 848       case Op_Rethrow  : exit->_in_rms = reth_rms; break;
 849       case Op_TailCall : exit->_in_rms = tail_call_rms; break;
 850       case Op_TailJump : exit->_in_rms = tail_jump_rms; break;
 851       case Op_Halt     : exit->_in_rms = halt_rms; break;
 852       default          : ShouldNotReachHere();
 853     }
 854   }
 855 
 856   // Next unused projection number from Start.
 857   int proj_cnt = C->tf()->domain()->cnt();
 858 
 859   // Do all the save-on-entry registers.  Make projections from Start for
 860   // them, and give them a use at the exit points.  To the allocator, they
 861   // look like incoming register arguments.
 862   for( i = 0; i < _last_Mach_Reg; i++ ) {
 863     if( is_save_on_entry(i) ) {
 864 
 865       // Add the save-on-entry to the mask array
 866       ret_rms      [      ret_edge_cnt] = mreg2regmask[i];
 867       reth_rms     [     reth_edge_cnt] = mreg2regmask[i];
 868       tail_call_rms[tail_call_edge_cnt] = mreg2regmask[i];
 869       tail_jump_rms[tail_jump_edge_cnt] = mreg2regmask[i];
 870       // Halts need the SOE registers, but only in the stack as debug info.
 871       // A just-prior uncommon-trap or deoptimization will use the SOE regs.
 872       halt_rms     [     halt_edge_cnt] = *idealreg2spillmask[_register_save_type[i]];
 873 
 874       Node *mproj;
 875 
 876       // Is this a RegF low half of a RegD?  Double up 2 adjacent RegF's
 877       // into a single RegD.

1109       Node *oldn = n;
1110       // Old-space or new-space check
1111       if (!C->node_arena()->contains(n)) {
1112         // Old space!
1113         Node* m;
1114         if (has_new_node(n)) {  // Not yet Label/Reduced
1115           m = new_node(n);
1116         } else {
1117           if (!is_dontcare(n)) { // Matcher can match this guy
1118             // Calls match special.  They match alone with no children.
1119             // Their children, the incoming arguments, match normally.
1120             m = n->is_SafePoint() ? match_sfpt(n->as_SafePoint()):match_tree(n);
1121             if (C->failing())  return NULL;
1122             if (m == NULL) { Matcher::soft_match_failure(); return NULL; }
1123             if (n->is_MemBar()) {
1124               m->as_MachMemBar()->set_adr_type(n->adr_type());
1125             }
1126           } else {                  // Nothing the matcher cares about
1127             if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Multi()) {       // Projections?
1128               // Convert to machine-dependent projection
1129               m = n->in(0)->as_Multi()->match( n->as_Proj(), this );




1130               NOT_PRODUCT(record_new2old(m, n);)
1131               if (m->in(0) != NULL) // m might be top
1132                 collect_null_checks(m, n);
1133             } else {                // Else just a regular 'ol guy
1134               m = n->clone();       // So just clone into new-space
1135               NOT_PRODUCT(record_new2old(m, n);)
1136               // Def-Use edges will be added incrementally as Uses
1137               // of this node are matched.
1138               assert(m->outcnt() == 0, "no Uses of this clone yet");
1139             }
1140           }
1141 
1142           set_new_node(n, m);       // Map old to new
1143           if (_old_node_note_array != NULL) {
1144             Node_Notes* nn = C->locate_node_notes(_old_node_note_array,
1145                                                   n->_idx);
1146             C->set_node_notes_at(m->_idx, nn);
1147           }
1148           debug_only(match_alias_type(C, n, m));
1149         }

1248   }
1249   return OptoReg::as_OptoReg(reg);
1250 }
1251 
1252 
1253 //------------------------------match_sfpt-------------------------------------
1254 // Helper function to match call instructions.  Calls match special.
1255 // They match alone with no children.  Their children, the incoming
1256 // arguments, match normally.
1257 MachNode *Matcher::match_sfpt( SafePointNode *sfpt ) {
1258   MachSafePointNode *msfpt = NULL;
1259   MachCallNode      *mcall = NULL;
1260   uint               cnt;
1261   // Split out case for SafePoint vs Call
1262   CallNode *call;
1263   const TypeTuple *domain;
1264   ciMethod*        method = NULL;
1265   bool             is_method_handle_invoke = false;  // for special kill effects
1266   if( sfpt->is_Call() ) {
1267     call = sfpt->as_Call();
1268     domain = call->tf()->domain();
1269     cnt = domain->cnt();
1270 
1271     // Match just the call, nothing else
1272     MachNode *m = match_tree(call);
1273     if (C->failing())  return NULL;
1274     if( m == NULL ) { Matcher::soft_match_failure(); return NULL; }
1275 
1276     // Copy data from the Ideal SafePoint to the machine version
1277     mcall = m->as_MachCall();
1278 
1279     mcall->set_tf(                  call->tf());
1280     mcall->set_entry_point(         call->entry_point());
1281     mcall->set_cnt(                 call->cnt());
1282     mcall->set_guaranteed_safepoint(call->guaranteed_safepoint());
1283 
1284     if( mcall->is_MachCallJava() ) {
1285       MachCallJavaNode *mcall_java  = mcall->as_MachCallJava();
1286       const CallJavaNode *call_java =  call->as_CallJava();
1287       assert(call_java->validate_symbolic_info(), "inconsistent info");
1288       method = call_java->method();

1327   msfpt->_in_rms = NEW_RESOURCE_ARRAY( RegMask, cnt );
1328   // Empty them all.
1329   for (uint i = 0; i < cnt; i++) ::new (&(msfpt->_in_rms[i])) RegMask();
1330 
1331   // Do all the pre-defined non-Empty register masks
1332   msfpt->_in_rms[TypeFunc::ReturnAdr] = _return_addr_mask;
1333   msfpt->_in_rms[TypeFunc::FramePtr ] = c_frame_ptr_mask;
1334 
1335   // Place first outgoing argument can possibly be put.
1336   OptoReg::Name begin_out_arg_area = OptoReg::add(_new_SP, C->out_preserve_stack_slots());
1337   assert( is_even(begin_out_arg_area), "" );
1338   // Compute max outgoing register number per call site.
1339   OptoReg::Name out_arg_limit_per_call = begin_out_arg_area;
1340   // Calls to C may hammer extra stack slots above and beyond any arguments.
1341   // These are usually backing store for register arguments for varargs.
1342   if( call != NULL && call->is_CallRuntime() )
1343     out_arg_limit_per_call = OptoReg::add(out_arg_limit_per_call,C->varargs_C_out_slots_killed());
1344 
1345 
1346   // Do the normal argument list (parameters) register masks
1347   int argcnt = cnt - TypeFunc::Parms;



1348   if( argcnt > 0 ) {          // Skip it all if we have no args
1349     BasicType *sig_bt  = NEW_RESOURCE_ARRAY( BasicType, argcnt );
1350     VMRegPair *parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt );
1351     int i;
1352     for( i = 0; i < argcnt; i++ ) {
1353       sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type();
1354     }
1355     // V-call to pick proper calling convention
1356     call->calling_convention( sig_bt, parm_regs, argcnt );
1357 
1358 #ifdef ASSERT
1359     // Sanity check users' calling convention.  Really handy during
1360     // the initial porting effort.  Fairly expensive otherwise.
1361     { for (int i = 0; i<argcnt; i++) {
1362       if( !parm_regs[i].first()->is_valid() &&
1363           !parm_regs[i].second()->is_valid() ) continue;
1364       VMReg reg1 = parm_regs[i].first();
1365       VMReg reg2 = parm_regs[i].second();
1366       for (int j = 0; j < i; j++) {
1367         if( !parm_regs[j].first()->is_valid() &&
1368             !parm_regs[j].second()->is_valid() ) continue;
1369         VMReg reg3 = parm_regs[j].first();
1370         VMReg reg4 = parm_regs[j].second();
1371         if( !reg1->is_valid() ) {
1372           assert( !reg2->is_valid(), "valid halvsies" );
1373         } else if( !reg3->is_valid() ) {
1374           assert( !reg4->is_valid(), "valid halvsies" );
1375         } else {
1376           assert( reg1 != reg2, "calling conv. must produce distinct regs");
1377           assert( reg1 != reg3, "calling conv. must produce distinct regs");
1378           assert( reg1 != reg4, "calling conv. must produce distinct regs");
1379           assert( reg2 != reg3, "calling conv. must produce distinct regs");
1380           assert( reg2 != reg4 || !reg2->is_valid(), "calling conv. must produce distinct regs");
1381           assert( reg3 != reg4, "calling conv. must produce distinct regs");
1382         }
1383       }
1384     }
1385     }
1386 #endif
1387 
1388     // Visit each argument.  Compute its outgoing register mask.
1389     // Return results now can have 2 bits returned.
1390     // Compute max over all outgoing arguments both per call-site
1391     // and over the entire method.
1392     for( i = 0; i < argcnt; i++ ) {
1393       // Address of incoming argument mask to fill in
1394       RegMask *rm = &mcall->_in_rms[i+TypeFunc::Parms];
1395       VMReg first = parm_regs[i].first();
1396       VMReg second = parm_regs[i].second();
1397       if(!first->is_valid() &&
1398          !second->is_valid()) {
1399         continue;               // Avoid Halves
1400       }
1401       // Handle case where arguments are in vector registers.
1402       if(call->in(TypeFunc::Parms + i)->bottom_type()->isa_vect()) {
1403         OptoReg::Name reg_fst = OptoReg::as_OptoReg(first);
1404         OptoReg::Name reg_snd = OptoReg::as_OptoReg(second);
1405         assert (reg_fst <= reg_snd, "fst=%d snd=%d", reg_fst, reg_snd);
1406         for (OptoReg::Name r = reg_fst; r <= reg_snd; r++) {
1407           rm->Insert(r);
1408         }
1409       }
1410       // Grab first register, adjust stack slots and insert in mask.
1411       OptoReg::Name reg1 = warp_outgoing_stk_arg(first, begin_out_arg_area, out_arg_limit_per_call );
1412       if (OptoReg::is_valid(reg1))
1413         rm->Insert( reg1 );

1414       // Grab second register (if any), adjust stack slots and insert in mask.
1415       OptoReg::Name reg2 = warp_outgoing_stk_arg(second, begin_out_arg_area, out_arg_limit_per_call );
1416       if (OptoReg::is_valid(reg2))
1417         rm->Insert( reg2 );

1418     } // End of for all arguments
1419   }
1420 
1421   // Compute the max stack slot killed by any call.  These will not be
1422   // available for debug info, and will be used to adjust FIRST_STACK_mask
1423   // after all call sites have been visited.
1424   if( _out_arg_limit < out_arg_limit_per_call)
1425     _out_arg_limit = out_arg_limit_per_call;
1426 
1427   if (mcall) {
1428     // Kill the outgoing argument area, including any non-argument holes and
1429     // any legacy C-killed slots.  Use Fat-Projections to do the killing.
1430     // Since the max-per-method covers the max-per-call-site and debug info
1431     // is excluded on the max-per-method basis, debug info cannot land in
1432     // this killed area.
1433     uint r_cnt = mcall->tf()->range()->cnt();
1434     MachProjNode *proj = new MachProjNode( mcall, r_cnt+10000, RegMask::Empty, MachProjNode::fat_proj );
1435     if (!RegMask::can_represent_arg(OptoReg::Name(out_arg_limit_per_call-1))) {
1436       C->record_method_not_compilable("unsupported outgoing calling sequence");
1437     } else {
1438       for (int i = begin_out_arg_area; i < out_arg_limit_per_call; i++)
1439         proj->_rout.Insert(OptoReg::Name(i));
1440     }
1441     if (proj->_rout.is_NotEmpty()) {
1442       push_projection(proj);
1443     }
1444   }
1445   // Transfer the safepoint information from the call to the mcall
1446   // Move the JVMState list
1447   msfpt->set_jvms(sfpt->jvms());
1448   for (JVMState* jvms = msfpt->jvms(); jvms; jvms = jvms->caller()) {
1449     jvms->set_map(sfpt);
1450   }
1451 
1452   // Debug inputs begin just after the last incoming parameter
1453   assert((mcall == NULL) || (mcall->jvms() == NULL) ||
1454          (mcall->jvms()->debug_start() + mcall->_jvmadj == mcall->tf()->domain()->cnt()), "");
1455 
1456   // Add additional edges.
1457   if (msfpt->mach_constant_base_node_input() != (uint)-1 && !msfpt->is_MachCallLeaf()) {
1458     // For these calls we can not add MachConstantBase in expand(), as the
1459     // ins are not complete then.
1460     msfpt->ins_req(msfpt->mach_constant_base_node_input(), C->mach_constant_base_node());
1461     if (msfpt->jvms() &&
1462         msfpt->mach_constant_base_node_input() <= msfpt->jvms()->debug_start() + msfpt->_jvmadj) {
1463       // We added an edge before jvms, so we must adapt the position of the ins.
1464       msfpt->jvms()->adapt_position(+1);
1465     }
1466   }
1467 
1468   // Registers killed by the call are set in the local scheduling pass
1469   // of Global Code Motion.
1470   return msfpt;
1471 }
1472 
1473 //---------------------------match_tree----------------------------------------
1474 // Match a Ideal Node DAG - turn it into a tree; Label & Reduce.  Used as part

2117         set_shared(n);       // Flag as shared and
2118         if (n->is_DecodeNarrowPtr()) {
2119           // Oop field/array element loads must be shared but since
2120           // they are shared through a DecodeN they may appear to have
2121           // a single use so force sharing here.
2122           set_shared(n->in(1));
2123         }
2124         mstack.pop();        // remove node from stack
2125         continue;
2126       }
2127       nstate = Visit; // Not already visited; so visit now
2128     }
2129     if (nstate == Visit) {
2130       mstack.set_state(Post_Visit);
2131       set_visited(n);   // Flag as visited now
2132       bool mem_op = false;
2133       int mem_addr_idx = MemNode::Address;
2134       if (find_shared_visit(mstack, n, nop, mem_op, mem_addr_idx)) {
2135         continue;
2136       }
2137       for (int i = n->req() - 1; i >= 0; --i) { // For my children
2138         Node* m = n->in(i); // Get ith input
2139         if (m == NULL) {
2140           continue;  // Ignore NULLs
2141         }
2142         if (clone_node(n, m, mstack)) {
2143           continue;
2144         }
2145 
2146         // Clone addressing expressions as they are "free" in memory access instructions
2147         if (mem_op && i == mem_addr_idx && m->is_AddP() &&
2148             // When there are other uses besides address expressions
2149             // put it on stack and mark as shared.
2150             !is_visited(m)) {
2151           // Some inputs for address expression are not put on stack
2152           // to avoid marking them as shared and forcing them into register
2153           // if they are used only in address expressions.
2154           // But they should be marked as shared if there are other uses
2155           // besides address expressions.
2156 
2157           if (pd_clone_address_expressions(m->as_AddP(), mstack, address_visited)) {

2441     case Op_FmaD:
2442     case Op_FmaF:
2443     case Op_FmaVD:
2444     case Op_FmaVF: {
2445       // Restructure into a binary tree for Matching.
2446       Node* pair = new BinaryNode(n->in(1), n->in(2));
2447       n->set_req(2, pair);
2448       n->set_req(1, n->in(3));
2449       n->del_req(3);
2450       break;
2451     }
2452     case Op_MulAddS2I: {
2453       Node* pair1 = new BinaryNode(n->in(1), n->in(2));
2454       Node* pair2 = new BinaryNode(n->in(3), n->in(4));
2455       n->set_req(1, pair1);
2456       n->set_req(2, pair2);
2457       n->del_req(4);
2458       n->del_req(3);
2459       break;
2460     }







2461     case Op_CopySignD:
2462     case Op_SignumVF:
2463     case Op_SignumVD:
2464     case Op_SignumF:
2465     case Op_SignumD: {
2466       Node* pair = new BinaryNode(n->in(2), n->in(3));
2467       n->set_req(2, pair);
2468       n->del_req(3);
2469       break;
2470     }
2471     case Op_VectorBlend:
2472     case Op_VectorInsert: {
2473       Node* pair = new BinaryNode(n->in(1), n->in(2));
2474       n->set_req(1, pair);
2475       n->set_req(2, n->in(3));
2476       n->del_req(3);
2477       break;
2478     }
2479     case Op_LoadVectorGatherMasked:
2480     case Op_StoreVectorScatter: {

 166   Unique_Node_List worklist;
 167   VectorSet visited;
 168   worklist.push(xroot);
 169   while (worklist.size() > 0) {
 170     Node* n = worklist.pop();
 171     visited.set(n->_idx);
 172     assert(C->node_arena()->contains(n), "dead node");
 173     for (uint j = 0; j < n->req(); j++) {
 174       Node* in = n->in(j);
 175       if (in != NULL) {
 176         assert(C->node_arena()->contains(in), "dead node");
 177         if (!visited.test(in->_idx)) {
 178           worklist.push(in);
 179         }
 180       }
 181     }
 182   }
 183 }
 184 #endif
 185 
 186 // Array of RegMask, one per returned values (inline type instances can
 187 // be returned as multiple return values, one per field)
 188 RegMask* Matcher::return_values_mask(const TypeFunc* tf) {
 189   const TypeTuple* range = tf->range_cc();
 190   uint cnt = range->cnt() - TypeFunc::Parms;
 191   if (cnt == 0) {
 192     return NULL;
 193   }
 194   RegMask* mask = NEW_RESOURCE_ARRAY(RegMask, cnt);
 195   BasicType* sig_bt = NEW_RESOURCE_ARRAY(BasicType, cnt);
 196   VMRegPair* vm_parm_regs = NEW_RESOURCE_ARRAY(VMRegPair, cnt);
 197   for (uint i = 0; i < cnt; i++) {
 198     sig_bt[i] = range->field_at(i+TypeFunc::Parms)->basic_type();
 199   }
 200 
 201   int regs = SharedRuntime::java_return_convention(sig_bt, vm_parm_regs, cnt);
 202   if (regs <= 0) {
 203     // We ran out of registers to store the IsInit information for a nullable inline type return.
 204     // Since it is only set in the 'call_epilog', we can simply put it on the stack.
 205     assert(tf->returns_inline_type_as_fields(), "should have been tested during graph construction");
 206     // TODO 8284443 Can we teach the register allocator to reserve a stack slot instead?
 207     // mask[--cnt] = STACK_ONLY_mask does not work (test with -XX:+StressGCM)
 208     int slot = C->fixed_slots() - 2;
 209     if (C->needs_stack_repair()) {
 210       slot -= 2; // Account for stack increment value
 211     }
 212     mask[--cnt].Clear();
 213     mask[cnt].Insert(OptoReg::stack2reg(slot));
 214   }
 215   for (uint i = 0; i < cnt; i++) {
 216     mask[i].Clear();
 217 
 218     OptoReg::Name reg1 = OptoReg::as_OptoReg(vm_parm_regs[i].first());
 219     if (OptoReg::is_valid(reg1)) {
 220       mask[i].Insert(reg1);
 221     }
 222     OptoReg::Name reg2 = OptoReg::as_OptoReg(vm_parm_regs[i].second());
 223     if (OptoReg::is_valid(reg2)) {
 224       mask[i].Insert(reg2);
 225     }
 226   }
 227 
 228   return mask;
 229 }
 230 
 231 //---------------------------match---------------------------------------------
 232 void Matcher::match( ) {
 233   if( MaxLabelRootDepth < 100 ) { // Too small?
 234     assert(false, "invalid MaxLabelRootDepth, increase it to 100 minimum");
 235     MaxLabelRootDepth = 100;
 236   }
 237   // One-time initialization of some register masks.
 238   init_spill_mask( C->root()->in(1) );
 239   _return_addr_mask = return_addr();
 240 #ifdef _LP64
 241   // Pointers take 2 slots in 64-bit land
 242   _return_addr_mask.Insert(OptoReg::add(return_addr(),1));
 243 #endif
 244 
 245   // Map Java-signature return types into return register-value
 246   // machine registers.
 247   _return_values_mask = return_values_mask(C->tf());












 248 
 249   // ---------------
 250   // Frame Layout
 251 
 252   // Need the method signature to determine the incoming argument types,
 253   // because the types determine which registers the incoming arguments are
 254   // in, and this affects the matched code.
 255   const TypeTuple *domain = C->tf()->domain_cc();
 256   uint             argcnt = domain->cnt() - TypeFunc::Parms;
 257   BasicType *sig_bt        = NEW_RESOURCE_ARRAY( BasicType, argcnt );
 258   VMRegPair *vm_parm_regs  = NEW_RESOURCE_ARRAY( VMRegPair, argcnt );
 259   _parm_regs               = NEW_RESOURCE_ARRAY( OptoRegPair, argcnt );
 260   _calling_convention_mask = NEW_RESOURCE_ARRAY( RegMask, argcnt );
 261   uint i;
 262   for( i = 0; i<argcnt; i++ ) {
 263     sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type();
 264   }
 265 
 266   // Pass array of ideal registers and length to USER code (from the AD file)
 267   // that will convert this to an array of register numbers.
 268   const StartNode *start = C->start();
 269   start->calling_convention( sig_bt, vm_parm_regs, argcnt );
 270 #ifdef ASSERT
 271   // Sanity check users' calling convention.  Real handy while trying to
 272   // get the initial port correct.
 273   { for (uint i = 0; i<argcnt; i++) {
 274       if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) {
 275         assert(domain->field_at(i+TypeFunc::Parms)==Type::HALF, "only allowed on halve" );

 539   idealreg2mhdebugmask[Op_VecS] = &rms[31];
 540   idealreg2mhdebugmask[Op_VecD] = &rms[32];
 541   idealreg2mhdebugmask[Op_VecX] = &rms[33];
 542   idealreg2mhdebugmask[Op_VecY] = &rms[34];
 543   idealreg2mhdebugmask[Op_VecZ] = &rms[35];
 544 
 545   idealreg2spillmask  [Op_RegVectMask] = &rms[36];
 546   idealreg2debugmask  [Op_RegVectMask] = &rms[37];
 547   idealreg2mhdebugmask[Op_RegVectMask] = &rms[38];
 548 
 549   OptoReg::Name i;
 550 
 551   // At first, start with the empty mask
 552   C->FIRST_STACK_mask().Clear();
 553 
 554   // Add in the incoming argument area
 555   OptoReg::Name init_in = OptoReg::add(_old_SP, C->out_preserve_stack_slots());
 556   for (i = init_in; i < _in_arg_limit; i = OptoReg::add(i,1)) {
 557     C->FIRST_STACK_mask().Insert(i);
 558   }
 559 
 560   // Add in all bits past the outgoing argument area
 561   guarantee(RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1)),
 562             "must be able to represent all call arguments in reg mask");
 563   OptoReg::Name init = _out_arg_limit;
 564   for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) {
 565     C->FIRST_STACK_mask().Insert(i);
 566   }
 567   // Finally, set the "infinite stack" bit.
 568   C->FIRST_STACK_mask().set_AllStack();
 569 
 570   // Make spill masks.  Registers for their class, plus FIRST_STACK_mask.
 571   RegMask aligned_stack_mask = C->FIRST_STACK_mask();
 572   // Keep spill masks aligned.
 573   aligned_stack_mask.clear_to_pairs();
 574   assert(aligned_stack_mask.is_AllStack(), "should be infinite stack");
 575   RegMask scalable_stack_mask = aligned_stack_mask;
 576 
 577   *idealreg2spillmask[Op_RegP] = *idealreg2regmask[Op_RegP];
 578 #ifdef _LP64
 579   *idealreg2spillmask[Op_RegN] = *idealreg2regmask[Op_RegN];

 798     _register_save_policy[reg] == 'E' ||
 799     _register_save_policy[reg] == 'A'; // Save-on-entry register?
 800 }
 801 
 802 //---------------------------Fixup_Save_On_Entry-------------------------------
 803 void Matcher::Fixup_Save_On_Entry( ) {
 804   init_first_stack_mask();
 805 
 806   Node *root = C->root();       // Short name for root
 807   // Count number of save-on-entry registers.
 808   uint soe_cnt = number_of_saved_registers();
 809   uint i;
 810 
 811   // Find the procedure Start Node
 812   StartNode *start = C->start();
 813   assert( start, "Expect a start node" );
 814 
 815   // Input RegMask array shared by all Returns.
 816   // The type for doubles and longs has a count of 2, but
 817   // there is only 1 returned value
 818   uint ret_edge_cnt = C->tf()->range_cc()->cnt();
 819   RegMask *ret_rms  = init_input_masks( ret_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
 820   for (i = TypeFunc::Parms; i < ret_edge_cnt; i++) {
 821     ret_rms[i] = _return_values_mask[i-TypeFunc::Parms];
 822   }

 823 
 824   // Input RegMask array shared by all Rethrows.
 825   uint reth_edge_cnt = TypeFunc::Parms+1;
 826   RegMask *reth_rms  = init_input_masks( reth_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
 827   // Rethrow takes exception oop only, but in the argument 0 slot.
 828   OptoReg::Name reg = find_receiver();
 829   if (reg >= 0) {
 830     reth_rms[TypeFunc::Parms] = mreg2regmask[reg];
 831 #ifdef _LP64
 832     // Need two slots for ptrs in 64-bit land
 833     reth_rms[TypeFunc::Parms].Insert(OptoReg::add(OptoReg::Name(reg), 1));
 834 #endif
 835   }
 836 
 837   // Input RegMask array shared by all TailCalls
 838   uint tail_call_edge_cnt = TypeFunc::Parms+2;
 839   RegMask *tail_call_rms = init_input_masks( tail_call_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
 840 
 841   // Input RegMask array shared by all TailJumps
 842   uint tail_jump_edge_cnt = TypeFunc::Parms+2;

 869   }
 870 
 871   // Input RegMask array shared by all Halts
 872   uint halt_edge_cnt = TypeFunc::Parms;
 873   RegMask *halt_rms = init_input_masks( halt_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask );
 874 
 875   // Capture the return input masks into each exit flavor
 876   for( i=1; i < root->req(); i++ ) {
 877     MachReturnNode *exit = root->in(i)->as_MachReturn();
 878     switch( exit->ideal_Opcode() ) {
 879       case Op_Return   : exit->_in_rms = ret_rms;  break;
 880       case Op_Rethrow  : exit->_in_rms = reth_rms; break;
 881       case Op_TailCall : exit->_in_rms = tail_call_rms; break;
 882       case Op_TailJump : exit->_in_rms = tail_jump_rms; break;
 883       case Op_Halt     : exit->_in_rms = halt_rms; break;
 884       default          : ShouldNotReachHere();
 885     }
 886   }
 887 
 888   // Next unused projection number from Start.
 889   int proj_cnt = C->tf()->domain_cc()->cnt();
 890 
 891   // Do all the save-on-entry registers.  Make projections from Start for
 892   // them, and give them a use at the exit points.  To the allocator, they
 893   // look like incoming register arguments.
 894   for( i = 0; i < _last_Mach_Reg; i++ ) {
 895     if( is_save_on_entry(i) ) {
 896 
 897       // Add the save-on-entry to the mask array
 898       ret_rms      [      ret_edge_cnt] = mreg2regmask[i];
 899       reth_rms     [     reth_edge_cnt] = mreg2regmask[i];
 900       tail_call_rms[tail_call_edge_cnt] = mreg2regmask[i];
 901       tail_jump_rms[tail_jump_edge_cnt] = mreg2regmask[i];
 902       // Halts need the SOE registers, but only in the stack as debug info.
 903       // A just-prior uncommon-trap or deoptimization will use the SOE regs.
 904       halt_rms     [     halt_edge_cnt] = *idealreg2spillmask[_register_save_type[i]];
 905 
 906       Node *mproj;
 907 
 908       // Is this a RegF low half of a RegD?  Double up 2 adjacent RegF's
 909       // into a single RegD.

1141       Node *oldn = n;
1142       // Old-space or new-space check
1143       if (!C->node_arena()->contains(n)) {
1144         // Old space!
1145         Node* m;
1146         if (has_new_node(n)) {  // Not yet Label/Reduced
1147           m = new_node(n);
1148         } else {
1149           if (!is_dontcare(n)) { // Matcher can match this guy
1150             // Calls match special.  They match alone with no children.
1151             // Their children, the incoming arguments, match normally.
1152             m = n->is_SafePoint() ? match_sfpt(n->as_SafePoint()):match_tree(n);
1153             if (C->failing())  return NULL;
1154             if (m == NULL) { Matcher::soft_match_failure(); return NULL; }
1155             if (n->is_MemBar()) {
1156               m->as_MachMemBar()->set_adr_type(n->adr_type());
1157             }
1158           } else {                  // Nothing the matcher cares about
1159             if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Multi()) {       // Projections?
1160               // Convert to machine-dependent projection
1161               RegMask* mask = NULL;
1162               if (n->in(0)->is_Call() && n->in(0)->as_Call()->tf()->returns_inline_type_as_fields()) {
1163                 mask = return_values_mask(n->in(0)->as_Call()->tf());
1164               }
1165               m = n->in(0)->as_Multi()->match(n->as_Proj(), this, mask);
1166               NOT_PRODUCT(record_new2old(m, n);)
1167               if (m->in(0) != NULL) // m might be top
1168                 collect_null_checks(m, n);
1169             } else {                // Else just a regular 'ol guy
1170               m = n->clone();       // So just clone into new-space
1171               NOT_PRODUCT(record_new2old(m, n);)
1172               // Def-Use edges will be added incrementally as Uses
1173               // of this node are matched.
1174               assert(m->outcnt() == 0, "no Uses of this clone yet");
1175             }
1176           }
1177 
1178           set_new_node(n, m);       // Map old to new
1179           if (_old_node_note_array != NULL) {
1180             Node_Notes* nn = C->locate_node_notes(_old_node_note_array,
1181                                                   n->_idx);
1182             C->set_node_notes_at(m->_idx, nn);
1183           }
1184           debug_only(match_alias_type(C, n, m));
1185         }

1284   }
1285   return OptoReg::as_OptoReg(reg);
1286 }
1287 
1288 
1289 //------------------------------match_sfpt-------------------------------------
1290 // Helper function to match call instructions.  Calls match special.
1291 // They match alone with no children.  Their children, the incoming
1292 // arguments, match normally.
1293 MachNode *Matcher::match_sfpt( SafePointNode *sfpt ) {
1294   MachSafePointNode *msfpt = NULL;
1295   MachCallNode      *mcall = NULL;
1296   uint               cnt;
1297   // Split out case for SafePoint vs Call
1298   CallNode *call;
1299   const TypeTuple *domain;
1300   ciMethod*        method = NULL;
1301   bool             is_method_handle_invoke = false;  // for special kill effects
1302   if( sfpt->is_Call() ) {
1303     call = sfpt->as_Call();
1304     domain = call->tf()->domain_cc();
1305     cnt = domain->cnt();
1306 
1307     // Match just the call, nothing else
1308     MachNode *m = match_tree(call);
1309     if (C->failing())  return NULL;
1310     if( m == NULL ) { Matcher::soft_match_failure(); return NULL; }
1311 
1312     // Copy data from the Ideal SafePoint to the machine version
1313     mcall = m->as_MachCall();
1314 
1315     mcall->set_tf(                  call->tf());
1316     mcall->set_entry_point(         call->entry_point());
1317     mcall->set_cnt(                 call->cnt());
1318     mcall->set_guaranteed_safepoint(call->guaranteed_safepoint());
1319 
1320     if( mcall->is_MachCallJava() ) {
1321       MachCallJavaNode *mcall_java  = mcall->as_MachCallJava();
1322       const CallJavaNode *call_java =  call->as_CallJava();
1323       assert(call_java->validate_symbolic_info(), "inconsistent info");
1324       method = call_java->method();

1363   msfpt->_in_rms = NEW_RESOURCE_ARRAY( RegMask, cnt );
1364   // Empty them all.
1365   for (uint i = 0; i < cnt; i++) ::new (&(msfpt->_in_rms[i])) RegMask();
1366 
1367   // Do all the pre-defined non-Empty register masks
1368   msfpt->_in_rms[TypeFunc::ReturnAdr] = _return_addr_mask;
1369   msfpt->_in_rms[TypeFunc::FramePtr ] = c_frame_ptr_mask;
1370 
1371   // Place first outgoing argument can possibly be put.
1372   OptoReg::Name begin_out_arg_area = OptoReg::add(_new_SP, C->out_preserve_stack_slots());
1373   assert( is_even(begin_out_arg_area), "" );
1374   // Compute max outgoing register number per call site.
1375   OptoReg::Name out_arg_limit_per_call = begin_out_arg_area;
1376   // Calls to C may hammer extra stack slots above and beyond any arguments.
1377   // These are usually backing store for register arguments for varargs.
1378   if( call != NULL && call->is_CallRuntime() )
1379     out_arg_limit_per_call = OptoReg::add(out_arg_limit_per_call,C->varargs_C_out_slots_killed());
1380 
1381 
1382   // Do the normal argument list (parameters) register masks
1383   // Null entry point is a special cast where the target of the call
1384   // is in a register.
1385   int adj = (call != NULL && call->entry_point() == NULL) ? 1 : 0;
1386   int argcnt = cnt - TypeFunc::Parms - adj;
1387   if( argcnt > 0 ) {          // Skip it all if we have no args
1388     BasicType *sig_bt  = NEW_RESOURCE_ARRAY( BasicType, argcnt );
1389     VMRegPair *parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt );
1390     int i;
1391     for( i = 0; i < argcnt; i++ ) {
1392       sig_bt[i] = domain->field_at(i+TypeFunc::Parms+adj)->basic_type();
1393     }
1394     // V-call to pick proper calling convention
1395     call->calling_convention( sig_bt, parm_regs, argcnt );
1396 
1397 #ifdef ASSERT
1398     // Sanity check users' calling convention.  Really handy during
1399     // the initial porting effort.  Fairly expensive otherwise.
1400     { for (int i = 0; i<argcnt; i++) {
1401       if( !parm_regs[i].first()->is_valid() &&
1402           !parm_regs[i].second()->is_valid() ) continue;
1403       VMReg reg1 = parm_regs[i].first();
1404       VMReg reg2 = parm_regs[i].second();
1405       for (int j = 0; j < i; j++) {
1406         if( !parm_regs[j].first()->is_valid() &&
1407             !parm_regs[j].second()->is_valid() ) continue;
1408         VMReg reg3 = parm_regs[j].first();
1409         VMReg reg4 = parm_regs[j].second();
1410         if( !reg1->is_valid() ) {
1411           assert( !reg2->is_valid(), "valid halvsies" );
1412         } else if( !reg3->is_valid() ) {
1413           assert( !reg4->is_valid(), "valid halvsies" );
1414         } else {
1415           assert( reg1 != reg2, "calling conv. must produce distinct regs");
1416           assert( reg1 != reg3, "calling conv. must produce distinct regs");
1417           assert( reg1 != reg4, "calling conv. must produce distinct regs");
1418           assert( reg2 != reg3, "calling conv. must produce distinct regs");
1419           assert( reg2 != reg4 || !reg2->is_valid(), "calling conv. must produce distinct regs");
1420           assert( reg3 != reg4, "calling conv. must produce distinct regs");
1421         }
1422       }
1423     }
1424     }
1425 #endif
1426 
1427     // Visit each argument.  Compute its outgoing register mask.
1428     // Return results now can have 2 bits returned.
1429     // Compute max over all outgoing arguments both per call-site
1430     // and over the entire method.
1431     for( i = 0; i < argcnt; i++ ) {
1432       // Address of incoming argument mask to fill in
1433       RegMask *rm = &mcall->_in_rms[i+TypeFunc::Parms+adj];
1434       VMReg first = parm_regs[i].first();
1435       VMReg second = parm_regs[i].second();
1436       if(!first->is_valid() &&
1437          !second->is_valid()) {
1438         continue;               // Avoid Halves
1439       }
1440       // Handle case where arguments are in vector registers.
1441       if(call->in(TypeFunc::Parms + i)->bottom_type()->isa_vect()) {
1442         OptoReg::Name reg_fst = OptoReg::as_OptoReg(first);
1443         OptoReg::Name reg_snd = OptoReg::as_OptoReg(second);
1444         assert (reg_fst <= reg_snd, "fst=%d snd=%d", reg_fst, reg_snd);
1445         for (OptoReg::Name r = reg_fst; r <= reg_snd; r++) {
1446           rm->Insert(r);
1447         }
1448       }
1449       // Grab first register, adjust stack slots and insert in mask.
1450       OptoReg::Name reg1 = warp_outgoing_stk_arg(first, begin_out_arg_area, out_arg_limit_per_call );
1451       if (OptoReg::is_valid(reg1)) {
1452         rm->Insert( reg1 );
1453       }
1454       // Grab second register (if any), adjust stack slots and insert in mask.
1455       OptoReg::Name reg2 = warp_outgoing_stk_arg(second, begin_out_arg_area, out_arg_limit_per_call );
1456       if (OptoReg::is_valid(reg2)) {
1457         rm->Insert( reg2 );
1458       }
1459     } // End of for all arguments
1460   }
1461 
1462   // Compute the max stack slot killed by any call.  These will not be
1463   // available for debug info, and will be used to adjust FIRST_STACK_mask
1464   // after all call sites have been visited.
1465   if( _out_arg_limit < out_arg_limit_per_call)
1466     _out_arg_limit = out_arg_limit_per_call;
1467 
1468   if (mcall) {
1469     // Kill the outgoing argument area, including any non-argument holes and
1470     // any legacy C-killed slots.  Use Fat-Projections to do the killing.
1471     // Since the max-per-method covers the max-per-call-site and debug info
1472     // is excluded on the max-per-method basis, debug info cannot land in
1473     // this killed area.
1474     uint r_cnt = mcall->tf()->range_sig()->cnt();
1475     MachProjNode *proj = new MachProjNode( mcall, r_cnt+10000, RegMask::Empty, MachProjNode::fat_proj );
1476     if (!RegMask::can_represent_arg(OptoReg::Name(out_arg_limit_per_call-1))) {
1477       C->record_method_not_compilable("unsupported outgoing calling sequence");
1478     } else {
1479       for (int i = begin_out_arg_area; i < out_arg_limit_per_call; i++)
1480         proj->_rout.Insert(OptoReg::Name(i));
1481     }
1482     if (proj->_rout.is_NotEmpty()) {
1483       push_projection(proj);
1484     }
1485   }
1486   // Transfer the safepoint information from the call to the mcall
1487   // Move the JVMState list
1488   msfpt->set_jvms(sfpt->jvms());
1489   for (JVMState* jvms = msfpt->jvms(); jvms; jvms = jvms->caller()) {
1490     jvms->set_map(sfpt);
1491   }
1492 
1493   // Debug inputs begin just after the last incoming parameter
1494   assert((mcall == NULL) || (mcall->jvms() == NULL) ||
1495          (mcall->jvms()->debug_start() + mcall->_jvmadj == mcall->tf()->domain_cc()->cnt()), "");
1496 
1497   // Add additional edges.
1498   if (msfpt->mach_constant_base_node_input() != (uint)-1 && !msfpt->is_MachCallLeaf()) {
1499     // For these calls we can not add MachConstantBase in expand(), as the
1500     // ins are not complete then.
1501     msfpt->ins_req(msfpt->mach_constant_base_node_input(), C->mach_constant_base_node());
1502     if (msfpt->jvms() &&
1503         msfpt->mach_constant_base_node_input() <= msfpt->jvms()->debug_start() + msfpt->_jvmadj) {
1504       // We added an edge before jvms, so we must adapt the position of the ins.
1505       msfpt->jvms()->adapt_position(+1);
1506     }
1507   }
1508 
1509   // Registers killed by the call are set in the local scheduling pass
1510   // of Global Code Motion.
1511   return msfpt;
1512 }
1513 
1514 //---------------------------match_tree----------------------------------------
1515 // Match a Ideal Node DAG - turn it into a tree; Label & Reduce.  Used as part

2158         set_shared(n);       // Flag as shared and
2159         if (n->is_DecodeNarrowPtr()) {
2160           // Oop field/array element loads must be shared but since
2161           // they are shared through a DecodeN they may appear to have
2162           // a single use so force sharing here.
2163           set_shared(n->in(1));
2164         }
2165         mstack.pop();        // remove node from stack
2166         continue;
2167       }
2168       nstate = Visit; // Not already visited; so visit now
2169     }
2170     if (nstate == Visit) {
2171       mstack.set_state(Post_Visit);
2172       set_visited(n);   // Flag as visited now
2173       bool mem_op = false;
2174       int mem_addr_idx = MemNode::Address;
2175       if (find_shared_visit(mstack, n, nop, mem_op, mem_addr_idx)) {
2176         continue;
2177       }
2178       for (int i = n->len() - 1; i >= 0; --i) { // For my children
2179         Node* m = n->in(i); // Get ith input
2180         if (m == NULL) {
2181           continue;  // Ignore NULLs
2182         }
2183         if (clone_node(n, m, mstack)) {
2184           continue;
2185         }
2186 
2187         // Clone addressing expressions as they are "free" in memory access instructions
2188         if (mem_op && i == mem_addr_idx && m->is_AddP() &&
2189             // When there are other uses besides address expressions
2190             // put it on stack and mark as shared.
2191             !is_visited(m)) {
2192           // Some inputs for address expression are not put on stack
2193           // to avoid marking them as shared and forcing them into register
2194           // if they are used only in address expressions.
2195           // But they should be marked as shared if there are other uses
2196           // besides address expressions.
2197 
2198           if (pd_clone_address_expressions(m->as_AddP(), mstack, address_visited)) {

2482     case Op_FmaD:
2483     case Op_FmaF:
2484     case Op_FmaVD:
2485     case Op_FmaVF: {
2486       // Restructure into a binary tree for Matching.
2487       Node* pair = new BinaryNode(n->in(1), n->in(2));
2488       n->set_req(2, pair);
2489       n->set_req(1, n->in(3));
2490       n->del_req(3);
2491       break;
2492     }
2493     case Op_MulAddS2I: {
2494       Node* pair1 = new BinaryNode(n->in(1), n->in(2));
2495       Node* pair2 = new BinaryNode(n->in(3), n->in(4));
2496       n->set_req(1, pair1);
2497       n->set_req(2, pair2);
2498       n->del_req(4);
2499       n->del_req(3);
2500       break;
2501     }
2502     case Op_ClearArray: {
2503       Node* pair = new BinaryNode(n->in(2), n->in(3));
2504       n->set_req(2, pair);
2505       n->set_req(3, n->in(4));
2506       n->del_req(4);
2507       break;
2508     }
2509     case Op_CopySignD:
2510     case Op_SignumVF:
2511     case Op_SignumVD:
2512     case Op_SignumF:
2513     case Op_SignumD: {
2514       Node* pair = new BinaryNode(n->in(2), n->in(3));
2515       n->set_req(2, pair);
2516       n->del_req(3);
2517       break;
2518     }
2519     case Op_VectorBlend:
2520     case Op_VectorInsert: {
2521       Node* pair = new BinaryNode(n->in(1), n->in(2));
2522       n->set_req(1, pair);
2523       n->set_req(2, n->in(3));
2524       n->del_req(3);
2525       break;
2526     }
2527     case Op_LoadVectorGatherMasked:
2528     case Op_StoreVectorScatter: {
< prev index next >