1 /*
   2  * Copyright (c) 1998, 2026, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "asm/macroAssembler.inline.hpp"
  26 #include "gc/shared/gc_globals.hpp"
  27 #include "memory/allocation.inline.hpp"
  28 #include "oops/compressedOops.hpp"
  29 #include "opto/ad.hpp"
  30 #include "opto/block.hpp"
  31 #include "opto/c2compiler.hpp"
  32 #include "opto/callnode.hpp"
  33 #include "opto/cfgnode.hpp"
  34 #include "opto/chaitin.hpp"
  35 #include "opto/machnode.hpp"
  36 #include "opto/runtime.hpp"
  37 #include "runtime/os.inline.hpp"
  38 #include "runtime/sharedRuntime.hpp"
  39 
  40 // Optimization - Graph Style
  41 
  42 // Check whether val is not-null-decoded compressed oop,
  43 // i.e. will grab into the base of the heap if it represents null.
  44 static bool accesses_heap_base_zone(Node *val) {
  45   if (CompressedOops::base() != nullptr) { // Implies UseCompressedOops.
  46     if (val && val->is_Mach()) {
  47       if (val->as_Mach()->ideal_Opcode() == Op_DecodeN) {
  48         // This assumes all Decodes with TypePtr::NotNull are matched to nodes that
  49         // decode null to point to the heap base (Decode_NN).
  50         if (val->bottom_type()->is_oopptr()->ptr() == TypePtr::NotNull) {
  51           return true;
  52         }
  53       }
  54       // Must recognize load operation with Decode matched in memory operand.
  55       // We should not reach here except for PPC/AIX, as os::zero_page_read_protected()
  56       // returns true everywhere else. On PPC, no such memory operands
  57       // exist, therefore we did not yet implement a check for such operands.
  58       NOT_AIX(Unimplemented());
  59     }
  60   }
  61   return false;
  62 }
  63 
  64 static bool needs_explicit_null_check_for_read(Node *val) {
  65   // On some OSes (AIX) the page at address 0 is only write protected.
  66   // If so, only Store operations will trap.
  67   if (os::zero_page_read_protected()) {
  68     return false;  // Implicit null check will work.
  69   }
  70   // Also a read accessing the base of a heap-based compressed heap will trap.
  71   if (accesses_heap_base_zone(val) &&         // Hits the base zone page.
  72       CompressedOops::use_implicit_null_checks()) { // Base zone page is protected.
  73     return false;
  74   }
  75 
  76   return true;
  77 }
  78 
  79 void PhaseCFG::move_node_and_its_projections_to_block(Node* n, Block* b) {
  80   assert(!is_CFG(n), "cannot move CFG node");
  81   Block* old = get_block_for_node(n);
  82   old->find_remove(n);
  83   b->add_inst(n);
  84   map_node_to_block(n, b);
  85   // Check for Mach projections that also need to be moved.
  86   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
  87     Node* out = n->fast_out(i);
  88     if (!out->is_MachProj()) {
  89       continue;
  90     }
  91     assert(!n->is_MachProj(), "nested projections are not allowed");
  92     move_node_and_its_projections_to_block(out, b);
  93   }
  94 }
  95 
  96 void PhaseCFG::ensure_node_is_at_block_or_above(Node* n, Block* b) {
  97   assert(!is_CFG(n), "cannot move CFG node");
  98   Block* current = get_block_for_node(n);
  99   if (current->dominates(b)) {
 100     return; // n is already placed above b, do nothing.
 101   }
 102   // We only expect nodes without further inputs, like MachTemp or load Base.
 103   assert(n->req() == 0 || (n->req() == 1 && n->in(0) == (Node*)C->root()),
 104          "need for recursive hoisting not expected");
 105   assert(b->dominates(current), "precondition: can only move n to b if b dominates n");
 106   move_node_and_its_projections_to_block(n, b);
 107 }
 108 
 109 //------------------------------implicit_null_check----------------------------
 110 // Detect implicit-null-check opportunities.  Basically, find null checks
 111 // with suitable memory ops nearby.  Use the memory op to do the null check.
 112 // I can generate a memory op if there is not one nearby.
 113 // The proj is the control projection for the not-null case.
 114 // The val is the pointer being checked for nullness or
 115 // decodeHeapOop_not_null node if it did not fold into address.
 116 void PhaseCFG::implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons) {
 117   // Assume if null check need for 0 offset then always needed
 118   // Intel solaris doesn't support any null checks yet and no
 119   // mechanism exists (yet) to set the switches at an os_cpu level
 120   if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;
 121 
 122   // Make sure the ptr-is-null path appears to be uncommon!
 123   float f = block->end()->as_MachIf()->_prob;
 124   if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;
 125   if( f > PROB_UNLIKELY_MAG(4) ) return;
 126 
 127   uint bidx = 0;                // Capture index of value into memop
 128   bool was_store;               // Memory op is a store op
 129 
 130   // Get the successor block for if the test ptr is non-null
 131   Block* not_null_block;  // this one goes with the proj
 132   Block* null_block;
 133   if (block->get_node(block->number_of_nodes()-1) == proj) {
 134     null_block     = block->_succs[0];
 135     not_null_block = block->_succs[1];
 136   } else {
 137     assert(block->get_node(block->number_of_nodes()-2) == proj, "proj is one or the other");
 138     not_null_block = block->_succs[0];
 139     null_block     = block->_succs[1];
 140   }
 141   while (null_block->is_Empty() == Block::empty_with_goto) {
 142     null_block     = null_block->_succs[0];
 143   }
 144 
 145   // Search the exception block for an uncommon trap.
 146   // (See Parse::do_if and Parse::do_ifnull for the reason
 147   // we need an uncommon trap.  Briefly, we need a way to
 148   // detect failure of this optimization, as in 6366351.)
 149   {
 150     bool found_trap = false;
 151     for (uint i1 = 0; i1 < null_block->number_of_nodes(); i1++) {
 152       Node* nn = null_block->get_node(i1);
 153       if (nn->is_MachCall() &&
 154           nn->as_MachCall()->entry_point() == OptoRuntime::uncommon_trap_blob()->entry_point()) {
 155         const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();
 156         if (trtype->isa_int() && trtype->is_int()->is_con()) {
 157           jint tr_con = trtype->is_int()->get_con();
 158           Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
 159           Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
 160           assert((int)reason < (int)BitsPerInt, "recode bit map");
 161           if (is_set_nth_bit(allowed_reasons, (int) reason)
 162               && action != Deoptimization::Action_none) {
 163             // This uncommon trap is sure to recompile, eventually.
 164             // When that happens, C->too_many_traps will prevent
 165             // this transformation from happening again.
 166             found_trap = true;
 167           }
 168         }
 169         break;
 170       }
 171     }
 172     if (!found_trap) {
 173       // We did not find an uncommon trap.
 174       return;
 175     }
 176   }
 177 
 178   // Check for decodeHeapOop_not_null node which did not fold into address
 179   bool is_decoden = ((intptr_t)val) & 1;
 180   val = (Node*)(((intptr_t)val) & ~1);
 181 
 182   assert(!is_decoden ||
 183          ((val->in(0) == nullptr) && val->is_Mach() &&
 184           (val->as_Mach()->ideal_Opcode() == Op_DecodeN)), "sanity");
 185 
 186   // Search the successor block for a load or store who's base value is also
 187   // the tested value.  There may be several.
 188   MachNode *best = nullptr;        // Best found so far
 189   for (DUIterator i = val->outs(); val->has_out(i); i++) {
 190     Node *m = val->out(i);
 191     if( !m->is_Mach() ) continue;
 192     MachNode *mach = m->as_Mach();
 193     if (mach->barrier_data() != 0 &&
 194         !mach->is_late_expanded_null_check_candidate()) {
 195       // Using memory accesses with barriers to perform implicit null checks is
 196       // only supported if these are explicit marked as emitting a candidate
 197       // memory access instruction at their initial address. If not marked as
 198       // such, barrier-tagged operations might expand into one or several memory
 199       // access instructions located at arbitrary offsets from the initial
 200       // address, which would invalidate the implicit null exception table.
 201       continue;
 202     }
 203     was_store = false;
 204     int iop = mach->ideal_Opcode();
 205     switch( iop ) {
 206     case Op_LoadB:
 207     case Op_LoadUB:
 208     case Op_LoadUS:
 209     case Op_LoadD:
 210     case Op_LoadF:
 211     case Op_LoadI:
 212     case Op_LoadL:
 213     case Op_LoadP:
 214     case Op_LoadN:
 215     case Op_LoadS:
 216     case Op_LoadKlass:
 217     case Op_LoadNKlass:
 218     case Op_LoadRange:
 219     case Op_LoadD_unaligned:
 220     case Op_LoadL_unaligned:
 221       assert(mach->in(2) == val, "should be address");
 222       break;
 223     case Op_StoreB:
 224     case Op_StoreC:
 225     case Op_StoreD:
 226     case Op_StoreF:
 227     case Op_StoreI:
 228     case Op_StoreL:
 229     case Op_StoreP:
 230     case Op_StoreN:
 231     case Op_StoreNKlass:
 232       was_store = true;         // Memory op is a store op
 233       // Stores will have their address in slot 2 (memory in slot 1).
 234       // If the value being nul-checked is in another slot, it means we
 235       // are storing the checked value, which does NOT check the value!
 236       if( mach->in(2) != val ) continue;
 237       break;                    // Found a memory op?
 238     case Op_StrComp:
 239     case Op_StrEquals:
 240     case Op_StrIndexOf:
 241     case Op_StrIndexOfChar:
 242     case Op_AryEq:
 243     case Op_VectorizedHashCode:
 244     case Op_StrInflatedCopy:
 245     case Op_StrCompressedCopy:
 246     case Op_EncodeISOArray:
 247     case Op_CountPositives:
 248       // Not a legit memory op for implicit null check regardless of
 249       // embedded loads
 250       continue;
 251     default:                    // Also check for embedded loads
 252       if( !mach->needs_anti_dependence_check() )
 253         continue;               // Not an memory op; skip it
 254       if( must_clone[iop] ) {
 255         // Do not move nodes which produce flags because
 256         // RA will try to clone it to place near branch and
 257         // it will cause recompilation, see clone_node().
 258         continue;
 259       }
 260       {
 261         // Check that value is used in memory address in
 262         // instructions with embedded load (CmpP val1,(val2+off)).
 263         Node* base;
 264         Node* index;
 265         const MachOper* oper = mach->memory_inputs(base, index);
 266         if (oper == nullptr || oper == (MachOper*)-1) {
 267           continue;             // Not an memory op; skip it
 268         }
 269         if (val == base ||
 270             (val == index && val->bottom_type()->isa_narrowoop())) {
 271           break;                // Found it
 272         } else {
 273           continue;             // Skip it
 274         }
 275       }
 276       break;
 277     }
 278 
 279     // On some OSes (AIX) the page at address 0 is only write protected.
 280     // If so, only Store operations will trap.
 281     // But a read accessing the base of a heap-based compressed heap will trap.
 282     if (!was_store && needs_explicit_null_check_for_read(val)) {
 283       continue;
 284     }
 285 
 286     // Check that node's control edge is not-null block's head or dominates it,
 287     // otherwise we can't hoist it because there are other control dependencies.
 288     Node* ctrl = mach->in(0);
 289     if (ctrl != nullptr && !(ctrl == not_null_block->head() ||
 290         get_block_for_node(ctrl)->dominates(not_null_block))) {
 291       continue;
 292     }
 293 
 294     // check if the offset is not too high for implicit exception
 295     {
 296       intptr_t offset = 0;
 297       const TypePtr *adr_type = nullptr;  // Do not need this return value here
 298       const Node* base = mach->get_base_and_disp(offset, adr_type);
 299       if (base == nullptr || base == NodeSentinel) {
 300         // Narrow oop address doesn't have base, only index.
 301         // Give up if offset is beyond page size or if heap base is not protected.
 302         if (val->bottom_type()->isa_narrowoop() &&
 303             (MacroAssembler::needs_explicit_null_check(offset) ||
 304              !CompressedOops::use_implicit_null_checks()))
 305           continue;
 306         // cannot reason about it; is probably not implicit null exception
 307       } else {
 308         const TypePtr* tptr;
 309         if ((UseCompressedOops && CompressedOops::shift() == 0) || CompressedKlassPointers::shift() == 0) {
 310           // 32-bits narrow oop can be the base of address expressions
 311           tptr = base->get_ptr_type();
 312         } else {
 313           // only regular oops are expected here
 314           tptr = base->bottom_type()->is_ptr();
 315         }
 316         // Give up if offset is not a compile-time constant.
 317         if (offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot)
 318           continue;
 319         offset += tptr->_offset; // correct if base is offsetted
 320         // Give up if reference is beyond page size.
 321         if (MacroAssembler::needs_explicit_null_check(offset))
 322           continue;
 323         // Give up if base is a decode node and the heap base is not protected.
 324         if (base->is_Mach() && base->as_Mach()->ideal_Opcode() == Op_DecodeN &&
 325             !CompressedOops::use_implicit_null_checks())
 326           continue;
 327       }
 328     }
 329 
 330     // Check ctrl input to see if the null-check dominates the memory op
 331     Block *cb = get_block_for_node(mach);
 332     cb = cb->_idom;             // Always hoist at least 1 block
 333     if( !was_store ) {          // Stores can be hoisted only one block
 334       while( cb->_dom_depth > (block->_dom_depth + 1))
 335         cb = cb->_idom;         // Hoist loads as far as we want
 336       // The non-null-block should dominate the memory op, too. Live
 337       // range spilling will insert a spill in the non-null-block if it is
 338       // needs to spill the memory op for an implicit null check.
 339       if (cb->_dom_depth == (block->_dom_depth + 1)) {
 340         if (cb != not_null_block) continue;
 341         cb = cb->_idom;
 342       }
 343     }
 344     if( cb != block ) continue;
 345 
 346     // Found a memory user; see if it can be hoisted to check-block
 347     uint vidx = 0;              // Capture index of value into memop
 348     uint j;
 349     for( j = mach->req()-1; j > 0; j-- ) {
 350       if( mach->in(j) == val ) {
 351         vidx = j;
 352         // Ignore DecodeN val which could be hoisted to where needed.
 353         if( is_decoden ) continue;
 354       }
 355       if (mach->in(j)->is_MachTemp()) {
 356         assert(mach->in(j)->outcnt() == 1, "MachTemp nodes should not be shared");
 357         // Ignore MachTemp inputs, they can be safely hoisted with the candidate.
 358         // MachTemp nodes have no inputs themselves and are only used to reserve
 359         // a scratch register for the implementation of the node (e.g. in
 360         // late-expanded GC barriers).
 361         continue;
 362       }
 363       // Block of memory-op input
 364       Block *inb = get_block_for_node(mach->in(j));
 365       Block *b = block;          // Start from nul check
 366       while( b != inb && b->_dom_depth > inb->_dom_depth )
 367         b = b->_idom;           // search upwards for input
 368       // See if input dominates null check
 369       if( b != inb )
 370         break;
 371     }
 372     if( j > 0 )
 373       continue;
 374     Block *mb = get_block_for_node(mach);
 375     // Hoisting stores requires more checks for the anti-dependence case.
 376     // Give up hoisting if we have to move the store past any load.
 377     if (was_store) {
 378        // Make sure control does not do a merge (would have to check allpaths)
 379        if (mb->num_preds() != 2) {
 380          continue;
 381        }
 382        // mach is a store, hence block is the immediate dominator of mb.
 383        // Due to the null-check shape of block (where its successors cannot re-join),
 384        // block must be the direct predecessor of mb.
 385        assert(get_block_for_node(mb->pred(1)) == block, "Unexpected predecessor block");
 386        uint k;
 387        uint num_nodes = mb->number_of_nodes();
 388        for (k = 1; k < num_nodes; k++) {
 389          Node *n = mb->get_node(k);
 390          if (n->needs_anti_dependence_check() &&
 391              n->in(LoadNode::Memory) == mach->in(StoreNode::Memory)) {
 392            break;              // Found anti-dependent load
 393          }
 394        }
 395        if (k < num_nodes) {
 396          continue;             // Found anti-dependent load
 397        }
 398     }
 399 
 400     // Make sure this memory op is not already being used for a NullCheck
 401     Node *e = mb->end();
 402     if( e->is_MachNullCheck() && e->in(1) == mach )
 403       continue;                 // Already being used as a null check
 404 
 405     // Found a candidate!  Pick one with least dom depth - the highest
 406     // in the dom tree should be closest to the null check.
 407     if (best == nullptr || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) {
 408       best = mach;
 409       bidx = vidx;
 410     }
 411   }
 412   // No candidate!
 413   if (best == nullptr) {
 414     return;
 415   }
 416 
 417   // ---- Found an implicit null check
 418 #ifndef PRODUCT
 419   extern uint implicit_null_checks;
 420   implicit_null_checks++;
 421 #endif
 422 
 423   if( is_decoden ) {
 424     // Check if we need to hoist decodeHeapOop_not_null first.
 425     Block *valb = get_block_for_node(val);
 426     if( block != valb && block->_dom_depth < valb->_dom_depth ) {
 427       // Hoist it up to the end of the test block together with its inputs if they exist.
 428       for (uint i = 2; i < val->req(); i++) {
 429         // DecodeN has 2 regular inputs + optional MachTemp or load Base inputs.
 430         // Inputs of val may already be early enough, but if not move them together with val.
 431         ensure_node_is_at_block_or_above(val->in(i), block);
 432       }
 433       move_node_and_its_projections_to_block(val, block);
 434     }
 435   }
 436 
 437   // Move any MachTemp inputs to the end of the test block.
 438   for (uint i = 0; i < best->req(); i++) {
 439     Node* n = best->in(i);
 440     if (n == nullptr || !n->is_MachTemp()) {
 441       continue;
 442     }
 443     ensure_node_is_at_block_or_above(n, block);
 444   }
 445 
 446   // Hoist the memory candidate up to the end of the test block.
 447   move_node_and_its_projections_to_block(best, block);
 448 
 449   // Move the control dependence if it is pinned to not-null block.
 450   // Don't change it in other cases: null or dominating control.
 451   Node* ctrl = best->in(0);
 452   if (ctrl != nullptr && get_block_for_node(ctrl) == not_null_block) {
 453     // Set it to control edge of null check.
 454     best->set_req(0, proj->in(0)->in(0));
 455   }
 456 
 457   // proj==Op_True --> ne test; proj==Op_False --> eq test.
 458   // One of two graph shapes got matched:
 459   //   (IfTrue  (If (Bool NE (CmpP ptr null))))
 460   //   (IfFalse (If (Bool EQ (CmpP ptr null))))
 461   // null checks are always branch-if-eq.  If we see a IfTrue projection
 462   // then we are replacing a 'ne' test with a 'eq' null check test.
 463   // We need to flip the projections to keep the same semantics.
 464   if( proj->Opcode() == Op_IfTrue ) {
 465     // Swap order of projections in basic block to swap branch targets
 466     Node *tmp1 = block->get_node(block->end_idx()+1);
 467     Node *tmp2 = block->get_node(block->end_idx()+2);
 468     block->map_node(tmp2, block->end_idx()+1);
 469     block->map_node(tmp1, block->end_idx()+2);
 470     Node *tmp = new Node(C->top()); // Use not null input
 471     tmp1->replace_by(tmp);
 472     tmp2->replace_by(tmp1);
 473     tmp->replace_by(tmp2);
 474     tmp->destruct(nullptr);
 475   }
 476 
 477   // Remove the existing null check; use a new implicit null check instead.
 478   // Since schedule-local needs precise def-use info, we need to correct
 479   // it as well.
 480   Node *old_tst = proj->in(0);
 481   MachNode *nul_chk = new MachNullCheckNode(old_tst->in(0),best,bidx);
 482   block->map_node(nul_chk, block->end_idx());
 483   map_node_to_block(nul_chk, block);
 484   // Redirect users of old_test to nul_chk
 485   for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
 486     old_tst->last_out(i2)->set_req(0, nul_chk);
 487   // Clean-up any dead code
 488   for (uint i3 = 0; i3 < old_tst->req(); i3++) {
 489     Node* in = old_tst->in(i3);
 490     old_tst->set_req(i3, nullptr);
 491     if (in->outcnt() == 0) {
 492       // Remove dead input node
 493       in->disconnect_inputs(C);
 494       block->find_remove(in);
 495     }
 496   }
 497 
 498   latency_from_uses(nul_chk);
 499   latency_from_uses(best);
 500 
 501   // insert anti-dependences to defs in this block
 502   if (! best->needs_anti_dependence_check()) {
 503     for (uint k = 1; k < block->number_of_nodes(); k++) {
 504       Node *n = block->get_node(k);
 505       if (n->needs_anti_dependence_check() &&
 506           n->in(LoadNode::Memory) == best->in(StoreNode::Memory)) {
 507         // Found anti-dependent load
 508         raise_above_anti_dependences(block, n);
 509         if (C->failing()) {
 510           return;
 511         }
 512       }
 513     }
 514   }
 515 }
 516 
 517 
 518 //------------------------------select-----------------------------------------
 519 // Select a nice fellow from the worklist to schedule next. If there is only one
 520 // choice, then use it. CreateEx nodes that are initially ready must start their
 521 // blocks and are given the highest priority, by being placed at the beginning
 522 // of the worklist. Next after initially-ready CreateEx nodes are projections,
 523 // which must follow their parents, and CreateEx nodes with local input
 524 // dependencies. Next are constants and CheckCastPP nodes. There are a number of
 525 // other special cases, for instructions that consume condition codes, et al.
 526 // These are chosen immediately. Some instructions are required to immediately
 527 // precede the last instruction in the block, and these are taken last. Of the
 528 // remaining cases (most), choose the instruction with the greatest latency
 529 // (that is, the most number of pseudo-cycles required to the end of the
 530 // routine). If there is a tie, choose the instruction with the most inputs.
 531 Node* PhaseCFG::select(
 532   Block* block,
 533   Node_List &worklist,
 534   GrowableArray<int> &ready_cnt,
 535   VectorSet &next_call,
 536   uint sched_slot,
 537   intptr_t* recalc_pressure_nodes) {
 538 
 539   // If only a single entry on the stack, use it
 540   uint cnt = worklist.size();
 541   if (cnt == 1) {
 542     Node *n = worklist[0];
 543     worklist.map(0,worklist.pop());
 544     return n;
 545   }
 546 
 547   uint choice  = 0; // Bigger is most important
 548   uint latency = 0; // Bigger is scheduled first
 549   uint score   = 0; // Bigger is better
 550   int idx = -1;     // Index in worklist
 551   int cand_cnt = 0; // Candidate count
 552   bool block_size_threshold_ok = (recalc_pressure_nodes != nullptr) && (block->number_of_nodes() > 10);
 553 
 554   for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
 555     // Order in worklist is used to break ties.
 556     // See caller for how this is used to delay scheduling
 557     // of induction variable increments to after the other
 558     // uses of the phi are scheduled.
 559     Node *n = worklist[i];      // Get Node on worklist
 560 
 561     int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
 562     if (iop == Op_CreateEx || n->is_Proj()) {
 563       // CreateEx nodes that are initially ready must start the block (after Phi
 564       // and Parm nodes which are pre-scheduled) and get top priority. This is
 565       // currently enforced by placing them at the beginning of the initial
 566       // worklist and selecting them eagerly here. After these, projections and
 567       // other CreateEx nodes are selected with equal priority.
 568       worklist.map(i,worklist.pop());
 569       return n;
 570     }
 571 
 572     if (n->Opcode() == Op_Con || iop == Op_CheckCastPP) {
 573       // Constants and CheckCastPP nodes have higher priority than the rest of
 574       // the nodes tested below. Record as current winner, but keep looking for
 575       // higher-priority nodes in the worklist.
 576       choice  = 4;
 577       // Latency and score are only used to break ties among low-priority nodes.
 578       latency = 0;
 579       score   = 0;
 580       idx     = i;
 581       continue;
 582     }
 583 
 584     // Final call in a block must be adjacent to 'catch'
 585     Node *e = block->end();
 586     if( e->is_Catch() && e->in(0)->in(0) == n )
 587       continue;
 588 
 589     // Memory op for an implicit null check has to be at the end of the block
 590     if( e->is_MachNullCheck() && e->in(1) == n )
 591       continue;
 592 
 593     // Schedule IV increment last.
 594     if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd) {
 595       // Cmp might be matched into CountedLoopEnd node.
 596       Node *cmp = (e->in(1)->ideal_reg() == Op_RegFlags) ? e->in(1) : e;
 597       if (cmp->req() > 1 && cmp->in(1) == n && n->is_iteratively_computed()) {
 598         continue;
 599       }
 600     }
 601 
 602     uint n_choice = 2;
 603 
 604     // See if this instruction is consumed by a branch. If so, then (as the
 605     // branch is the last instruction in the basic block) force it to the
 606     // end of the basic block
 607     if ( must_clone[iop] ) {
 608       // See if any use is a branch
 609       bool found_machif = false;
 610 
 611       for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
 612         Node* use = n->fast_out(j);
 613 
 614         // The use is a conditional branch, make them adjacent
 615         if (use->is_MachIf() && get_block_for_node(use) == block) {
 616           found_machif = true;
 617           break;
 618         }
 619 
 620         // More than this instruction pending for successor to be ready,
 621         // don't choose this if other opportunities are ready
 622         if (ready_cnt.at(use->_idx) > 1)
 623           n_choice = 1;
 624       }
 625 
 626       // loop terminated, prefer not to use this instruction
 627       if (found_machif)
 628         continue;
 629     }
 630 
 631     // See if this has a predecessor that is "must_clone", i.e. sets the
 632     // condition code. If so, choose this first
 633     for (uint j = 0; j < n->req() ; j++) {
 634       Node *inn = n->in(j);
 635       if (inn) {
 636         if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
 637           n_choice = 3;
 638           break;
 639         }
 640       }
 641     }
 642 
 643     // MachTemps should be scheduled last so they are near their uses
 644     if (n->is_MachTemp()) {
 645       n_choice = 1;
 646     }
 647 
 648     uint n_latency = get_latency_for_node(n);
 649     uint n_score = n->req();   // Many inputs get high score to break ties
 650 
 651     if (OptoRegScheduling && block_size_threshold_ok) {
 652       if (recalc_pressure_nodes[n->_idx] == 0x7fff7fff) {
 653         _regalloc->_scratch_int_pressure.init(_regalloc->_sched_int_pressure.high_pressure_limit());
 654         _regalloc->_scratch_float_pressure.init(_regalloc->_sched_float_pressure.high_pressure_limit());
 655         // simulate the notion that we just picked this node to schedule
 656         n->add_flag(Node::Flag_is_scheduled);
 657         // now calculate its effect upon the graph if we did
 658         adjust_register_pressure(n, block, recalc_pressure_nodes, false);
 659         // return its state for finalize in case somebody else wins
 660         n->remove_flag(Node::Flag_is_scheduled);
 661         // now save the two final pressure components of register pressure, limiting pressure calcs to short size
 662         short int_pressure = (short)_regalloc->_scratch_int_pressure.current_pressure();
 663         short float_pressure = (short)_regalloc->_scratch_float_pressure.current_pressure();
 664         recalc_pressure_nodes[n->_idx] = int_pressure;
 665         recalc_pressure_nodes[n->_idx] |= (float_pressure << 16);
 666       }
 667 
 668       if (_scheduling_for_pressure) {
 669         latency = n_latency;
 670         if (n_choice != 3) {
 671           // Now evaluate each register pressure component based on threshold in the score.
 672           // In general the defining register type will dominate the score, ergo we will not see register pressure grow on both banks
 673           // on a single instruction, but we might see it shrink on both banks.
 674           // For each use of register that has a register class that is over the high pressure limit, we build n_score up for
 675           // live ranges that terminate on this instruction.
 676           if (_regalloc->_sched_int_pressure.current_pressure() > _regalloc->_sched_int_pressure.high_pressure_limit()) {
 677             short int_pressure = (short)recalc_pressure_nodes[n->_idx];
 678             n_score = (int_pressure < 0) ? ((score + n_score) - int_pressure) : (int_pressure > 0) ? 1 : n_score;
 679           }
 680           if (_regalloc->_sched_float_pressure.current_pressure() > _regalloc->_sched_float_pressure.high_pressure_limit()) {
 681             short float_pressure = (short)(recalc_pressure_nodes[n->_idx] >> 16);
 682             n_score = (float_pressure < 0) ? ((score + n_score) - float_pressure) : (float_pressure > 0) ? 1 : n_score;
 683           }
 684         } else {
 685           // make sure we choose these candidates
 686           score = 0;
 687         }
 688       }
 689     }
 690 
 691     // Keep best latency found
 692     cand_cnt++;
 693     if (choice < n_choice ||
 694         (choice == n_choice &&
 695          ((StressLCM && C->randomized_select(cand_cnt)) ||
 696           (!StressLCM &&
 697            (latency < n_latency ||
 698             (latency == n_latency &&
 699              (score < n_score))))))) {
 700       choice  = n_choice;
 701       latency = n_latency;
 702       score   = n_score;
 703       idx     = i;               // Also keep index in worklist
 704     }
 705   } // End of for all ready nodes in worklist
 706 
 707   guarantee(idx >= 0, "index should be set");
 708   Node *n = worklist[(uint)idx];      // Get the winner
 709 
 710   worklist.map((uint)idx, worklist.pop());     // Compress worklist
 711   return n;
 712 }
 713 
 714 //-------------------------adjust_register_pressure----------------------------
 715 void PhaseCFG::adjust_register_pressure(Node* n, Block* block, intptr_t* recalc_pressure_nodes, bool finalize_mode) {
 716   PhaseLive* liveinfo = _regalloc->get_live();
 717   IndexSet* liveout = liveinfo->live(block);
 718   // first adjust the register pressure for the sources
 719   for (uint i = 1; i < n->req(); i++) {
 720     bool lrg_ends = false;
 721     Node *src_n = n->in(i);
 722     if (src_n == nullptr) continue;
 723     if (!src_n->is_Mach()) continue;
 724     uint src = _regalloc->_lrg_map.find(src_n);
 725     if (src == 0) continue;
 726     LRG& lrg_src = _regalloc->lrgs(src);
 727     // detect if the live range ends or not
 728     if (liveout->member(src) == false) {
 729       lrg_ends = true;
 730       for (DUIterator_Fast jmax, j = src_n->fast_outs(jmax); j < jmax; j++) {
 731         Node* m = src_n->fast_out(j); // Get user
 732         if (m == n) continue;
 733         if (!m->is_Mach()) continue;
 734         MachNode *mach = m->as_Mach();
 735         bool src_matches = false;
 736         int iop = mach->ideal_Opcode();
 737 
 738         switch (iop) {
 739         case Op_StoreB:
 740         case Op_StoreC:
 741         case Op_StoreD:
 742         case Op_StoreF:
 743         case Op_StoreI:
 744         case Op_StoreL:
 745         case Op_StoreP:
 746         case Op_StoreN:
 747         case Op_StoreVector:
 748         case Op_StoreVectorMasked:
 749         case Op_StoreVectorScatter:
 750         case Op_StoreVectorScatterMasked:
 751         case Op_StoreNKlass:
 752           for (uint k = 1; k < m->req(); k++) {
 753             Node *in = m->in(k);
 754             if (in == src_n) {
 755               src_matches = true;
 756               break;
 757             }
 758           }
 759           break;
 760 
 761         default:
 762           src_matches = true;
 763           break;
 764         }
 765 
 766         // If we have a store as our use, ignore the non source operands
 767         if (src_matches == false) continue;
 768 
 769         // Mark every unscheduled use which is not n with a recalculation
 770         if ((get_block_for_node(m) == block) && (!m->is_scheduled())) {
 771           if (finalize_mode && !m->is_Phi()) {
 772             recalc_pressure_nodes[m->_idx] = 0x7fff7fff;
 773           }
 774           lrg_ends = false;
 775         }
 776       }
 777     }
 778     // if none, this live range ends and we can adjust register pressure
 779     if (lrg_ends) {
 780       if (finalize_mode) {
 781         _regalloc->lower_pressure(block, 0, lrg_src, nullptr, _regalloc->_sched_int_pressure, _regalloc->_sched_float_pressure);
 782       } else {
 783         _regalloc->lower_pressure(block, 0, lrg_src, nullptr, _regalloc->_scratch_int_pressure, _regalloc->_scratch_float_pressure);
 784       }
 785     }
 786   }
 787 
 788   // now add the register pressure from the dest and evaluate which heuristic we should use:
 789   // 1.) The default, latency scheduling
 790   // 2.) Register pressure scheduling based on the high pressure limit threshold for int or float register stacks
 791   uint dst = _regalloc->_lrg_map.find(n);
 792   if (dst != 0) {
 793     LRG& lrg_dst = _regalloc->lrgs(dst);
 794     if (finalize_mode) {
 795       _regalloc->raise_pressure(block, lrg_dst, _regalloc->_sched_int_pressure, _regalloc->_sched_float_pressure);
 796       // check to see if we fall over the register pressure cliff here
 797       if (_regalloc->_sched_int_pressure.current_pressure() > _regalloc->_sched_int_pressure.high_pressure_limit()) {
 798         _scheduling_for_pressure = true;
 799       } else if (_regalloc->_sched_float_pressure.current_pressure() > _regalloc->_sched_float_pressure.high_pressure_limit()) {
 800         _scheduling_for_pressure = true;
 801       } else {
 802         // restore latency scheduling mode
 803         _scheduling_for_pressure = false;
 804       }
 805     } else {
 806       _regalloc->raise_pressure(block, lrg_dst, _regalloc->_scratch_int_pressure, _regalloc->_scratch_float_pressure);
 807     }
 808   }
 809 }
 810 
 811 //------------------------------set_next_call----------------------------------
 812 void PhaseCFG::set_next_call(const Block* block, Node* init, VectorSet& next_call) const {
 813   Node_List worklist;
 814   worklist.push(init);
 815 
 816   while (worklist.size() > 0) {
 817     Node* n = worklist.pop();
 818     if (next_call.test_set(n->_idx)) continue;
 819     for (uint i = 0; i < n->len(); i++) {
 820       Node* m = n->in(i);
 821       if (m == nullptr) continue;  // must see all nodes in block that precede call
 822       if (get_block_for_node(m) == block) {
 823         worklist.push(m);
 824       }
 825     }
 826   }
 827 }
 828 
 829 //------------------------------needed_for_next_call---------------------------
 830 // Set the flag 'next_call' for each Node that is needed for the next call to
 831 // be scheduled.  This flag lets me bias scheduling so Nodes needed for the
 832 // next subroutine call get priority - basically it moves things NOT needed
 833 // for the next call till after the call.  This prevents me from trying to
 834 // carry lots of stuff live across a call.
 835 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) {
 836   // Find the next control-defining Node in this block
 837   Node* call = nullptr;
 838   for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
 839     Node* m = this_call->fast_out(i);
 840     if (get_block_for_node(m) == block && // Local-block user
 841         m != this_call &&       // Not self-start node
 842         m->is_MachCall()) {
 843       call = m;
 844       break;
 845     }
 846   }
 847   if (call == nullptr)  return;    // No next call (e.g., block end is near)
 848   // Set next-call for all inputs to this call
 849   set_next_call(block, call, next_call);
 850 }
 851 
 852 //------------------------------add_call_kills-------------------------------------
 853 // helper function that adds caller save registers to MachProjNode
 854 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) {
 855   // Fill in the kill mask for the call
 856   for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
 857     if (!regs.member(r)) { // Not already defined by the call
 858       // Save-on-call register?
 859       if ((save_policy[r] == 'C') ||
 860           (save_policy[r] == 'A') ||
 861           ((save_policy[r] == 'E') && exclude_soe)) {
 862         proj->_rout.insert(r);
 863       }
 864     }
 865   }
 866 }
 867 
 868 
 869 //------------------------------sched_call-------------------------------------
 870 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) {
 871   ResourceMark rm(C->regmask_arena());
 872   RegMask regs(C->regmask_arena());
 873 
 874   // Schedule all the users of the call right now.  All the users are
 875   // projection Nodes, so they must be scheduled next to the call.
 876   // Collect all the defined registers.
 877   for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
 878     Node* n = mcall->fast_out(i);
 879     assert( n->is_MachProj(), "" );
 880     int n_cnt = ready_cnt.at(n->_idx)-1;
 881     ready_cnt.at_put(n->_idx, n_cnt);
 882     assert( n_cnt == 0, "" );
 883     // Schedule next to call
 884     block->map_node(n, node_cnt++);
 885     // Collect defined registers
 886     regs.or_with(n->out_RegMask());
 887     // Check for scheduling the next control-definer
 888     if( n->bottom_type() == Type::CONTROL )
 889       // Warm up next pile of heuristic bits
 890       needed_for_next_call(block, n, next_call);
 891 
 892     // Children of projections are now all ready
 893     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
 894       Node* m = n->fast_out(j); // Get user
 895       if(get_block_for_node(m) != block) {
 896         continue;
 897       }
 898       if( m->is_Phi() ) continue;
 899       int m_cnt = ready_cnt.at(m->_idx) - 1;
 900       ready_cnt.at_put(m->_idx, m_cnt);
 901       if( m_cnt == 0 )
 902         worklist.push(m);
 903     }
 904 
 905   }
 906 
 907   // Act as if the call defines the Frame Pointer.
 908   // Certainly the FP is alive and well after the call.
 909   regs.insert(_matcher.c_frame_pointer());
 910 
 911   // Set all registers killed and not already defined by the call.
 912   uint r_cnt = mcall->tf()->range()->cnt();
 913   int op = mcall->ideal_Opcode();
 914   MachProjNode* proj = new MachProjNode(mcall, r_cnt + 1, RegMask::EMPTY, MachProjNode::fat_proj);
 915   map_node_to_block(proj, block);
 916   block->insert_node(proj, node_cnt++);
 917 
 918   // Select the right register save policy.
 919   const char *save_policy = nullptr;
 920   switch (op) {
 921     case Op_CallRuntime:
 922     case Op_CallLeaf:
 923     case Op_CallLeafNoFP:
 924     case Op_CallLeafVector:
 925       // Calling C code so use C calling convention
 926       save_policy = _matcher._c_reg_save_policy;
 927       break;
 928 
 929     case Op_CallStaticJava:
 930     case Op_CallDynamicJava:
 931       // Calling Java code so use Java calling convention
 932       save_policy = _matcher._register_save_policy;
 933       break;
 934 
 935     default:
 936       ShouldNotReachHere();
 937   }
 938 
 939   // When using CallRuntime mark SOE registers as killed by the call
 940   // so values that could show up in the RegisterMap aren't live in a
 941   // callee saved register since the register wouldn't know where to
 942   // find them.  CallLeaf and CallLeafNoFP are ok because they can't
 943   // have debug info on them.  Strictly speaking this only needs to be
 944   // done for oops since idealreg2debugmask takes care of debug info
 945   // references but there no way to handle oops differently than other
 946   // pointers as far as the kill mask goes.
 947   bool exclude_soe = op == Op_CallRuntime;
 948   add_call_kills(proj, regs, save_policy, exclude_soe);
 949 
 950   return node_cnt;
 951 }
 952 
 953 
 954 //------------------------------schedule_local---------------------------------
 955 // Topological sort within a block.  Someday become a real scheduler.
 956 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call, intptr_t *recalc_pressure_nodes) {
 957   // Already "sorted" are the block start Node (as the first entry), and
 958   // the block-ending Node and any trailing control projections.  We leave
 959   // these alone.  PhiNodes and ParmNodes are made to follow the block start
 960   // Node.  Everything else gets topo-sorted.
 961 
 962 #ifndef PRODUCT
 963     if (trace_opto_pipelining()) {
 964       tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order);
 965       for (uint i = 0;i < block->number_of_nodes(); i++) {
 966         tty->print("# ");
 967         block->get_node(i)->dump();
 968       }
 969       tty->print_cr("#");
 970     }
 971 #endif
 972 
 973   // RootNode is already sorted
 974   if (block->number_of_nodes() == 1) {
 975     return true;
 976   }
 977 
 978   bool block_size_threshold_ok = (recalc_pressure_nodes != nullptr) && (block->number_of_nodes() > 10);
 979 
 980   // We track the uses of local definitions as input dependences so that
 981   // we know when a given instruction is available to be scheduled.
 982   uint i;
 983   if (OptoRegScheduling && block_size_threshold_ok) {
 984     for (i = 1; i < block->number_of_nodes(); i++) { // setup nodes for pressure calc
 985       Node *n = block->get_node(i);
 986       n->remove_flag(Node::Flag_is_scheduled);
 987       if (!n->is_Phi()) {
 988         recalc_pressure_nodes[n->_idx] = 0x7fff7fff;
 989       }
 990     }
 991   }
 992 
 993   // Move PhiNodes and ParmNodes from 1 to cnt up to the start
 994   uint node_cnt = block->end_idx();
 995   uint phi_cnt = 1;
 996   for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
 997     Node *n = block->get_node(i);
 998     if( n->is_Phi() ||          // Found a PhiNode or ParmNode
 999         (n->is_Proj()  && n->in(0) == block->head()) ) {
1000       // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
1001       block->map_node(block->get_node(phi_cnt), i);
1002       block->map_node(n, phi_cnt++);  // swap Phi/Parm up front
1003       if (OptoRegScheduling && block_size_threshold_ok) {
1004         // mark n as scheduled
1005         n->add_flag(Node::Flag_is_scheduled);
1006       }
1007     } else {                    // All others
1008       // Count block-local inputs to 'n'
1009       uint cnt = n->len();      // Input count
1010       uint local = 0;
1011       for( uint j=0; j<cnt; j++ ) {
1012         Node *m = n->in(j);
1013         if( m && get_block_for_node(m) == block && !m->is_top() )
1014           local++;              // One more block-local input
1015       }
1016       ready_cnt.at_put(n->_idx, local); // Count em up
1017       // A few node types require changing a required edge to a precedence edge
1018       // before allocation.
1019       if( n->is_Mach() && n->req() > TypeFunc::Parms &&
1020           (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
1021            n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
1022         // MemBarAcquire could be created without Precedent edge.
1023         // del_req() replaces the specified edge with the last input edge
1024         // and then removes the last edge. If the specified edge > number of
1025         // edges the last edge will be moved outside of the input edges array
1026         // and the edge will be lost. This is why this code should be
1027         // executed only when Precedent (== TypeFunc::Parms) edge is present.
1028         Node *x = n->in(TypeFunc::Parms);
1029         if (x != nullptr && get_block_for_node(x) == block && n->find_prec_edge(x) != -1) {
1030           // Old edge to node within same block will get removed, but no precedence
1031           // edge will get added because it already exists. Update ready count.
1032           int cnt = ready_cnt.at(n->_idx);
1033           assert(cnt > 1, "MemBar node %d must not get ready here", n->_idx);
1034           ready_cnt.at_put(n->_idx, cnt-1);
1035         }
1036         n->del_req(TypeFunc::Parms);
1037         n->add_prec(x);
1038       }
1039     }
1040   }
1041   for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count
1042     ready_cnt.at_put(block->get_node(i2)->_idx, 0);
1043 
1044   // All the prescheduled guys do not hold back internal nodes
1045   uint i3;
1046   for (i3 = 0; i3 < phi_cnt; i3++) {  // For all pre-scheduled
1047     Node *n = block->get_node(i3);       // Get pre-scheduled
1048     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1049       Node* m = n->fast_out(j);
1050       if (get_block_for_node(m) == block) { // Local-block user
1051         int m_cnt = ready_cnt.at(m->_idx)-1;
1052         if (OptoRegScheduling && block_size_threshold_ok) {
1053           // mark m as scheduled
1054           if (m_cnt < 0) {
1055             m->add_flag(Node::Flag_is_scheduled);
1056           }
1057         }
1058         ready_cnt.at_put(m->_idx, m_cnt);   // Fix ready count
1059       }
1060     }
1061   }
1062 
1063   Node_List delay;
1064   // Make a worklist
1065   Node_List worklist;
1066   for(uint i4=i3; i4<node_cnt; i4++ ) {    // Put ready guys on worklist
1067     Node *m = block->get_node(i4);
1068     if( !ready_cnt.at(m->_idx) ) {   // Zero ready count?
1069       if (m->is_iteratively_computed()) {
1070         // Push induction variable increments last to allow other uses
1071         // of the phi to be scheduled first. The select() method breaks
1072         // ties in scheduling by worklist order.
1073         delay.push(m);
1074       } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
1075         // Place CreateEx nodes that are initially ready at the beginning of the
1076         // worklist so they are selected first and scheduled at the block start.
1077         worklist.insert(0, m);
1078       } else {
1079         worklist.push(m);         // Then on to worklist!
1080       }
1081     }
1082   }
1083   while (delay.size()) {
1084     Node* d = delay.pop();
1085     worklist.push(d);
1086   }
1087 
1088   if (OptoRegScheduling && block_size_threshold_ok) {
1089     // To stage register pressure calculations we need to examine the live set variables
1090     // breaking them up by register class to compartmentalize the calculations.
1091     _regalloc->_sched_int_pressure.init(Matcher::int_pressure_limit());
1092     _regalloc->_sched_float_pressure.init(Matcher::float_pressure_limit());
1093     _regalloc->_scratch_int_pressure.init(Matcher::int_pressure_limit());
1094     _regalloc->_scratch_float_pressure.init(Matcher::float_pressure_limit());
1095 
1096     _regalloc->compute_entry_block_pressure(block);
1097   }
1098 
1099   // Warm up the 'next_call' heuristic bits
1100   needed_for_next_call(block, block->head(), next_call);
1101 
1102 #ifndef PRODUCT
1103     if (trace_opto_pipelining()) {
1104       for (uint j=0; j< block->number_of_nodes(); j++) {
1105         Node     *n = block->get_node(j);
1106         int     idx = n->_idx;
1107         tty->print("#   ready cnt:%3d  ", ready_cnt.at(idx));
1108         tty->print("latency:%3d  ", get_latency_for_node(n));
1109         tty->print("%4d: %s\n", idx, n->Name());
1110       }
1111     }
1112 #endif
1113 
1114   uint max_idx = (uint)ready_cnt.length();
1115   // Pull from worklist and schedule
1116   while( worklist.size() ) {    // Worklist is not ready
1117 
1118 #ifndef PRODUCT
1119     if (trace_opto_pipelining()) {
1120       tty->print("#   ready list:");
1121       for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
1122         Node *n = worklist[i];      // Get Node on worklist
1123         tty->print(" %d", n->_idx);
1124       }
1125       tty->cr();
1126     }
1127 #endif
1128 
1129     // Select and pop a ready guy from worklist
1130     Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt, recalc_pressure_nodes);
1131     block->map_node(n, phi_cnt++);    // Schedule him next
1132 
1133     if (OptoRegScheduling && block_size_threshold_ok) {
1134       n->add_flag(Node::Flag_is_scheduled);
1135 
1136       // Now adjust the resister pressure with the node we selected
1137       if (!n->is_Phi()) {
1138         adjust_register_pressure(n, block, recalc_pressure_nodes, true);
1139       }
1140     }
1141 
1142 #ifndef PRODUCT
1143     if (trace_opto_pipelining()) {
1144       tty->print("#    select %d: %s", n->_idx, n->Name());
1145       tty->print(", latency:%d", get_latency_for_node(n));
1146       n->dump();
1147       if (Verbose) {
1148         tty->print("#   ready list:");
1149         for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
1150           Node *n = worklist[i];      // Get Node on worklist
1151           tty->print(" %d", n->_idx);
1152         }
1153         tty->cr();
1154       }
1155     }
1156 
1157 #endif
1158     if( n->is_MachCall() ) {
1159       MachCallNode *mcall = n->as_MachCall();
1160       phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call);
1161       continue;
1162     }
1163 
1164     if (n->is_Mach() && n->as_Mach()->has_call()) {
1165       RegMask regs;
1166       regs.insert(_matcher.c_frame_pointer());
1167       regs.or_with(n->out_RegMask());
1168 
1169       MachProjNode* proj = new MachProjNode(n, 1, RegMask::EMPTY, MachProjNode::fat_proj);
1170       map_node_to_block(proj, block);
1171       block->insert_node(proj, phi_cnt++);
1172 
1173       add_call_kills(proj, regs, _matcher._c_reg_save_policy, false);
1174     }
1175 
1176     // Children are now all ready
1177     for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
1178       Node* m = n->fast_out(i5); // Get user
1179       if (get_block_for_node(m) != block) {
1180         continue;
1181       }
1182       if( m->is_Phi() ) continue;
1183       if (m->_idx >= max_idx) { // new node, skip it
1184         assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types");
1185         continue;
1186       }
1187       int m_cnt = ready_cnt.at(m->_idx) - 1;
1188       ready_cnt.at_put(m->_idx, m_cnt);
1189       if( m_cnt == 0 )
1190         worklist.push(m);
1191     }
1192   }
1193 
1194   if( phi_cnt != block->end_idx() ) {
1195     // did not schedule all.  Retry, Bailout, or Die
1196     if (C->subsume_loads() == true && !C->failing()) {
1197       // Retry with subsume_loads == false
1198       // If this is the first failure, the sentinel string will "stick"
1199       // to the Compile object, and the C2Compiler will see it and retry.
1200       C->record_failure(C2Compiler::retry_no_subsuming_loads());
1201     } else {
1202       assert(C->failure_is_artificial(), "graph should be schedulable");
1203     }
1204     // assert( phi_cnt == end_idx(), "did not schedule all" );
1205     return false;
1206   }
1207 
1208   if (OptoRegScheduling && block_size_threshold_ok) {
1209     _regalloc->compute_exit_block_pressure(block);
1210     block->_reg_pressure = _regalloc->_sched_int_pressure.final_pressure();
1211     block->_freg_pressure = _regalloc->_sched_float_pressure.final_pressure();
1212   }
1213 
1214 #ifndef PRODUCT
1215   if (trace_opto_pipelining()) {
1216     tty->print_cr("#");
1217     tty->print_cr("# after schedule_local");
1218     for (uint i = 0;i < block->number_of_nodes();i++) {
1219       tty->print("# ");
1220       block->get_node(i)->dump();
1221     }
1222     tty->print_cr("# ");
1223 
1224     if (OptoRegScheduling && block_size_threshold_ok) {
1225       tty->print_cr("# pressure info : %d", block->_pre_order);
1226       _regalloc->print_pressure_info(_regalloc->_sched_int_pressure, "int register info");
1227       _regalloc->print_pressure_info(_regalloc->_sched_float_pressure, "float register info");
1228     }
1229     tty->cr();
1230   }
1231 #endif
1232 
1233   return true;
1234 }
1235 
1236 //--------------------------catch_cleanup_fix_all_inputs-----------------------
1237 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
1238   for (uint l = 0; l < use->len(); l++) {
1239     if (use->in(l) == old_def) {
1240       if (l < use->req()) {
1241         use->set_req(l, new_def);
1242       } else {
1243         use->rm_prec(l);
1244         use->add_prec(new_def);
1245         l--;
1246       }
1247     }
1248   }
1249 }
1250 
1251 //------------------------------catch_cleanup_find_cloned_def------------------
1252 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
1253   assert( use_blk != def_blk, "Inter-block cleanup only");
1254 
1255   // The use is some block below the Catch.  Find and return the clone of the def
1256   // that dominates the use. If there is no clone in a dominating block, then
1257   // create a phi for the def in a dominating block.
1258 
1259   // Find which successor block dominates this use.  The successor
1260   // blocks must all be single-entry (from the Catch only; I will have
1261   // split blocks to make this so), hence they all dominate.
1262   while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
1263     use_blk = use_blk->_idom;
1264 
1265   // Find the successor
1266   Node *fixup = nullptr;
1267 
1268   uint j;
1269   for( j = 0; j < def_blk->_num_succs; j++ )
1270     if( use_blk == def_blk->_succs[j] )
1271       break;
1272 
1273   if( j == def_blk->_num_succs ) {
1274     // Block at same level in dom-tree is not a successor.  It needs a
1275     // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
1276     Node_Array inputs;
1277     for(uint k = 1; k < use_blk->num_preds(); k++) {
1278       Block* block = get_block_for_node(use_blk->pred(k));
1279       inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx));
1280     }
1281 
1282     // Check to see if the use_blk already has an identical phi inserted.
1283     // If it exists, it will be at the first position since all uses of a
1284     // def are processed together.
1285     Node *phi = use_blk->get_node(1);
1286     if( phi->is_Phi() ) {
1287       fixup = phi;
1288       for (uint k = 1; k < use_blk->num_preds(); k++) {
1289         if (phi->in(k) != inputs[k]) {
1290           // Not a match
1291           fixup = nullptr;
1292           break;
1293         }
1294       }
1295     }
1296 
1297     // If an existing PhiNode was not found, make a new one.
1298     if (fixup == nullptr) {
1299       Node *new_phi = PhiNode::make(use_blk->head(), def);
1300       use_blk->insert_node(new_phi, 1);
1301       map_node_to_block(new_phi, use_blk);
1302       for (uint k = 1; k < use_blk->num_preds(); k++) {
1303         new_phi->set_req(k, inputs[k]);
1304       }
1305       fixup = new_phi;
1306     }
1307 
1308   } else {
1309     // Found the use just below the Catch.  Make it use the clone.
1310     fixup = use_blk->get_node(n_clone_idx);
1311   }
1312 
1313   return fixup;
1314 }
1315 
1316 //--------------------------catch_cleanup_intra_block--------------------------
1317 // Fix all input edges in use that reference "def".  The use is in the same
1318 // block as the def and both have been cloned in each successor block.
1319 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
1320 
1321   // Both the use and def have been cloned. For each successor block,
1322   // get the clone of the use, and make its input the clone of the def
1323   // found in that block.
1324 
1325   uint use_idx = blk->find_node(use);
1326   uint offset_idx = use_idx - beg;
1327   for( uint k = 0; k < blk->_num_succs; k++ ) {
1328     // Get clone in each successor block
1329     Block *sb = blk->_succs[k];
1330     Node *clone = sb->get_node(offset_idx+1);
1331     assert( clone->Opcode() == use->Opcode(), "" );
1332 
1333     // Make use-clone reference the def-clone
1334     catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx));
1335   }
1336 }
1337 
1338 //------------------------------catch_cleanup_inter_block---------------------
1339 // Fix all input edges in use that reference "def".  The use is in a different
1340 // block than the def.
1341 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
1342   if( !use_blk ) return;        // Can happen if the use is a precedence edge
1343 
1344   Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx);
1345   catch_cleanup_fix_all_inputs(use, def, new_def);
1346 }
1347 
1348 //------------------------------call_catch_cleanup-----------------------------
1349 // If we inserted any instructions between a Call and his CatchNode,
1350 // clone the instructions on all paths below the Catch.
1351 void PhaseCFG::call_catch_cleanup(Block* block) {
1352 
1353   // End of region to clone
1354   uint end = block->end_idx();
1355   if( !block->get_node(end)->is_Catch() ) return;
1356   // Start of region to clone
1357   uint beg = end;
1358   while(!block->get_node(beg-1)->is_MachProj() ||
1359         !block->get_node(beg-1)->in(0)->is_MachCall() ) {
1360     beg--;
1361     assert(beg > 0,"Catch cleanup walking beyond block boundary");
1362   }
1363   // Range of inserted instructions is [beg, end)
1364   if( beg == end ) return;
1365 
1366   // Clone along all Catch output paths.  Clone area between the 'beg' and
1367   // 'end' indices.
1368   for( uint i = 0; i < block->_num_succs; i++ ) {
1369     Block *sb = block->_succs[i];
1370     // Clone the entire area; ignoring the edge fixup for now.
1371     for( uint j = end; j > beg; j-- ) {
1372       Node *clone = block->get_node(j-1)->clone();
1373       sb->insert_node(clone, 1);
1374       map_node_to_block(clone, sb);
1375       if (clone->needs_anti_dependence_check()) {
1376         raise_above_anti_dependences(sb, clone);
1377         if (C->failing()) {
1378           return;
1379         }
1380       }
1381     }
1382   }
1383 
1384 
1385   // Fixup edges.  Check the def-use info per cloned Node
1386   for(uint i2 = beg; i2 < end; i2++ ) {
1387     uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
1388     Node *n = block->get_node(i2);        // Node that got cloned
1389     // Need DU safe iterator because of edge manipulation in calls.
1390     Unique_Node_List* out = new Unique_Node_List();
1391     for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
1392       out->push(n->fast_out(j1));
1393     }
1394     uint max = out->size();
1395     for (uint j = 0; j < max; j++) {// For all users
1396       Node *use = out->pop();
1397       Block *buse = get_block_for_node(use);
1398       if( use->is_Phi() ) {
1399         for( uint k = 1; k < use->req(); k++ )
1400           if( use->in(k) == n ) {
1401             Block* b = get_block_for_node(buse->pred(k));
1402             Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx);
1403             use->set_req(k, fixup);
1404           }
1405       } else {
1406         if (block == buse) {
1407           catch_cleanup_intra_block(use, n, block, beg, n_clone_idx);
1408         } else {
1409           catch_cleanup_inter_block(use, buse, n, block, n_clone_idx);
1410         }
1411       }
1412     } // End for all users
1413 
1414   } // End of for all Nodes in cloned area
1415 
1416   // Remove the now-dead cloned ops
1417   for(uint i3 = beg; i3 < end; i3++ ) {
1418     block->get_node(beg)->disconnect_inputs(C);
1419     block->remove_node(beg);
1420   }
1421 
1422   // If the successor blocks have a CreateEx node, move it back to the top
1423   for (uint i4 = 0; i4 < block->_num_succs; i4++) {
1424     Block *sb = block->_succs[i4];
1425     uint new_cnt = end - beg;
1426     // Remove any newly created, but dead, nodes by traversing their schedule
1427     // backwards. Here, a dead node is a node whose only outputs (if any) are
1428     // unused projections.
1429     for (uint j = new_cnt; j > 0; j--) {
1430       Node *n = sb->get_node(j);
1431       // Individual projections are examined together with all siblings when
1432       // their parent is visited.
1433       if (n->is_Proj()) {
1434         continue;
1435       }
1436       bool dead = true;
1437       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1438         Node* out = n->fast_out(i);
1439         // n is live if it has a non-projection output or a used projection.
1440         if (!out->is_Proj() || out->outcnt() > 0) {
1441           dead = false;
1442           break;
1443         }
1444       }
1445       if (dead) {
1446         // n's only outputs (if any) are unused projections scheduled next to n
1447         // (see PhaseCFG::select()). Remove these projections backwards.
1448         for (uint k = j + n->outcnt(); k > j; k--) {
1449           Node* proj = sb->get_node(k);
1450           assert(proj->is_Proj() && proj->in(0) == n,
1451                  "projection should correspond to dead node");
1452           proj->disconnect_inputs(C);
1453           sb->remove_node(k);
1454           new_cnt--;
1455         }
1456         // Now remove the node itself.
1457         n->disconnect_inputs(C);
1458         sb->remove_node(j);
1459         new_cnt--;
1460       }
1461     }
1462     // If any newly created nodes remain, move the CreateEx node to the top
1463     if (new_cnt > 0) {
1464       Node *cex = sb->get_node(1+new_cnt);
1465       if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
1466         sb->remove_node(1+new_cnt);
1467         sb->insert_node(cex, 1);
1468       }
1469     }
1470   }
1471 }