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_StoreLSpecial:
 230     case Op_StoreP:
 231     case Op_StoreN:
 232     case Op_StoreNKlass:
 233       was_store = true;         // Memory op is a store op
 234       // Stores will have their address in slot 2 (memory in slot 1).
 235       // If the value being nul-checked is in another slot, it means we
 236       // are storing the checked value, which does NOT check the value!
 237       if( mach->in(2) != val ) continue;
 238       break;                    // Found a memory op?
 239     case Op_StrComp:
 240     case Op_StrEquals:
 241     case Op_StrIndexOf:
 242     case Op_StrIndexOfChar:
 243     case Op_AryEq:
 244     case Op_VectorizedHashCode:
 245     case Op_StrInflatedCopy:
 246     case Op_StrCompressedCopy:
 247     case Op_EncodeISOArray:
 248     case Op_CountPositives:
 249       // Not a legit memory op for implicit null check regardless of
 250       // embedded loads
 251       continue;
 252     default:                    // Also check for embedded loads
 253       if( !mach->needs_anti_dependence_check() )
 254         continue;               // Not an memory op; skip it
 255       if( must_clone[iop] ) {
 256         // Do not move nodes which produce flags because
 257         // RA will try to clone it to place near branch and
 258         // it will cause recompilation, see clone_node().
 259         continue;
 260       }
 261       {
 262         // Check that value is used in memory address in
 263         // instructions with embedded load (CmpP val1,(val2+off)).
 264         Node* base;
 265         Node* index;
 266         const MachOper* oper = mach->memory_inputs(base, index);
 267         if (oper == nullptr || oper == (MachOper*)-1) {
 268           continue;             // Not an memory op; skip it
 269         }
 270         if (val == base ||
 271             (val == index && val->bottom_type()->isa_narrowoop())) {
 272           break;                // Found it
 273         } else {
 274           continue;             // Skip it
 275         }
 276       }
 277       break;
 278     }
 279 
 280     // On some OSes (AIX) the page at address 0 is only write protected.
 281     // If so, only Store operations will trap.
 282     // But a read accessing the base of a heap-based compressed heap will trap.
 283     if (!was_store && needs_explicit_null_check_for_read(val)) {
 284       continue;
 285     }
 286 
 287     // Check that node's control edge is not-null block's head or dominates it,
 288     // otherwise we can't hoist it because there are other control dependencies.
 289     Node* ctrl = mach->in(0);
 290     if (ctrl != nullptr && !(ctrl == not_null_block->head() ||
 291         get_block_for_node(ctrl)->dominates(not_null_block))) {
 292       continue;
 293     }
 294 
 295     // check if the offset is not too high for implicit exception
 296     {
 297       intptr_t offset = 0;
 298       const TypePtr *adr_type = nullptr;  // Do not need this return value here
 299       const Node* base = mach->get_base_and_disp(offset, adr_type);
 300       if (base == nullptr || base == NodeSentinel) {
 301         // Narrow oop address doesn't have base, only index.
 302         // Give up if offset is beyond page size or if heap base is not protected.
 303         if (val->bottom_type()->isa_narrowoop() &&
 304             (MacroAssembler::needs_explicit_null_check(offset) ||
 305              !CompressedOops::use_implicit_null_checks()))
 306           continue;
 307         // cannot reason about it; is probably not implicit null exception
 308       } else {
 309         const TypePtr* tptr;
 310         if ((UseCompressedOops && CompressedOops::shift() == 0) || CompressedKlassPointers::shift() == 0) {
 311           // 32-bits narrow oop can be the base of address expressions
 312           tptr = base->get_ptr_type();
 313         } else {
 314           // only regular oops are expected here
 315           tptr = base->bottom_type()->is_ptr();
 316         }
 317         // Give up if offset is not a compile-time constant.
 318         if (offset == Type::OffsetBot || tptr->offset() == Type::OffsetBot)
 319           continue;
 320         offset += tptr->offset(); // correct if base is offsetted
 321         // Give up if reference is beyond page size.
 322         if (MacroAssembler::needs_explicit_null_check(offset))
 323           continue;
 324         // Give up if base is a decode node and the heap base is not protected.
 325         if (base->is_Mach() && base->as_Mach()->ideal_Opcode() == Op_DecodeN &&
 326             !CompressedOops::use_implicit_null_checks())
 327           continue;
 328       }
 329     }
 330 
 331     // Check ctrl input to see if the null-check dominates the memory op
 332     Block *cb = get_block_for_node(mach);
 333     cb = cb->_idom;             // Always hoist at least 1 block
 334     if( !was_store ) {          // Stores can be hoisted only one block
 335       while( cb->_dom_depth > (block->_dom_depth + 1))
 336         cb = cb->_idom;         // Hoist loads as far as we want
 337       // The non-null-block should dominate the memory op, too. Live
 338       // range spilling will insert a spill in the non-null-block if it is
 339       // needs to spill the memory op for an implicit null check.
 340       if (cb->_dom_depth == (block->_dom_depth + 1)) {
 341         if (cb != not_null_block) continue;
 342         cb = cb->_idom;
 343       }
 344     }
 345     if( cb != block ) continue;
 346 
 347     // Found a memory user; see if it can be hoisted to check-block
 348     uint vidx = 0;              // Capture index of value into memop
 349     uint j;
 350     for( j = mach->req()-1; j > 0; j-- ) {
 351       if( mach->in(j) == val ) {
 352         vidx = j;
 353         // Ignore DecodeN val which could be hoisted to where needed.
 354         if( is_decoden ) continue;
 355       }
 356       if (mach->in(j)->is_MachTemp()) {
 357         assert(mach->in(j)->outcnt() == 1, "MachTemp nodes should not be shared");
 358         // Ignore MachTemp inputs, they can be safely hoisted with the candidate.
 359         // MachTemp nodes have no inputs themselves and are only used to reserve
 360         // a scratch register for the implementation of the node (e.g. in
 361         // late-expanded GC barriers).
 362         continue;
 363       }
 364       // Block of memory-op input
 365       Block* inb = get_block_for_node(mach->in(j));
 366       if (mach->in(j)->is_Con() && mach->in(j)->req() == 1 && inb == get_block_for_node(mach)) {
 367         // Ignore constant loads scheduled in the same block (we can simply hoist them as well)
 368         continue;
 369       }
 370       Block *b = block;          // Start from nul check
 371       while( b != inb && b->_dom_depth > inb->_dom_depth )
 372         b = b->_idom;           // search upwards for input
 373       // See if input dominates null check
 374       if( b != inb )
 375         break;
 376     }
 377     if( j > 0 )
 378       continue;
 379     Block *mb = get_block_for_node(mach);
 380     // Hoisting stores requires more checks for the anti-dependence case.
 381     // Give up hoisting if we have to move the store past any load.
 382     if (was_store) {
 383        // Make sure control does not do a merge (would have to check allpaths)
 384        if (mb->num_preds() != 2) {
 385          continue;
 386        }
 387        // mach is a store, hence block is the immediate dominator of mb.
 388        // Due to the null-check shape of block (where its successors cannot re-join),
 389        // block must be the direct predecessor of mb.
 390        assert(get_block_for_node(mb->pred(1)) == block, "Unexpected predecessor block");
 391        uint k;
 392        uint num_nodes = mb->number_of_nodes();
 393        for (k = 1; k < num_nodes; k++) {
 394          Node *n = mb->get_node(k);
 395          if (n->needs_anti_dependence_check() &&
 396              n->in(LoadNode::Memory) == mach->in(StoreNode::Memory)) {
 397            break;              // Found anti-dependent load
 398          }
 399        }
 400        if (k < num_nodes) {
 401          continue;             // Found anti-dependent load
 402        }
 403     }
 404 
 405     // Make sure this memory op is not already being used for a NullCheck
 406     Node *e = mb->end();
 407     if( e->is_MachNullCheck() && e->in(1) == mach )
 408       continue;                 // Already being used as a null check
 409 
 410     // Found a candidate!  Pick one with least dom depth - the highest
 411     // in the dom tree should be closest to the null check.
 412     if (best == nullptr || get_block_for_node(mach)->_dom_depth < get_block_for_node(best)->_dom_depth) {
 413       best = mach;
 414       bidx = vidx;
 415     }
 416   }
 417   // No candidate!
 418   if (best == nullptr) {
 419     return;
 420   }
 421 
 422   // ---- Found an implicit null check
 423 #ifndef PRODUCT
 424   extern uint implicit_null_checks;
 425   implicit_null_checks++;
 426 #endif
 427 
 428   if( is_decoden ) {
 429     // Check if we need to hoist decodeHeapOop_not_null first.
 430     Block *valb = get_block_for_node(val);
 431     if( block != valb && block->_dom_depth < valb->_dom_depth ) {
 432       // Hoist it up to the end of the test block together with its inputs if they exist.
 433       for (uint i = 2; i < val->req(); i++) {
 434         // DecodeN has 2 regular inputs + optional MachTemp or load Base inputs.
 435         // Inputs of val may already be early enough, but if not move them together with val.
 436         ensure_node_is_at_block_or_above(val->in(i), block);
 437       }
 438       move_node_and_its_projections_to_block(val, block);
 439     }
 440   }
 441 
 442   // Hoist constant load inputs as well.
 443   for (uint i = 1; i < best->req(); ++i) {
 444     Node* n = best->in(i);
 445     if (n->is_Con() && get_block_for_node(n) == get_block_for_node(best)) {
 446       get_block_for_node(n)->find_remove(n);
 447       block->add_inst(n);
 448       map_node_to_block(n, block);
 449       // Constant loads may kill flags (for example, when XORing a register).
 450       // Check for flag-killing projections that also need to be hoisted.
 451       for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
 452         Node* proj = n->fast_out(j);
 453         if (proj->is_MachProj()) {
 454           get_block_for_node(proj)->find_remove(proj);
 455           block->add_inst(proj);
 456           map_node_to_block(proj, block);
 457         }
 458       }
 459     }
 460   }
 461 
 462 
 463   // Move any MachTemp inputs to the end of the test block.
 464   for (uint i = 0; i < best->req(); i++) {
 465     Node* n = best->in(i);
 466     if (n == nullptr || !n->is_MachTemp()) {
 467       continue;
 468     }
 469     ensure_node_is_at_block_or_above(n, block);
 470   }
 471 
 472   // Hoist the memory candidate up to the end of the test block.
 473   move_node_and_its_projections_to_block(best, block);
 474 
 475   // Move the control dependence if it is pinned to not-null block.
 476   // Don't change it in other cases: null or dominating control.
 477   Node* ctrl = best->in(0);
 478   if (ctrl != nullptr && get_block_for_node(ctrl) == not_null_block) {
 479     // Set it to control edge of null check.
 480     best->set_req(0, proj->in(0)->in(0));
 481   }
 482 
 483   // proj==Op_True --> ne test; proj==Op_False --> eq test.
 484   // One of two graph shapes got matched:
 485   //   (IfTrue  (If (Bool NE (CmpP ptr null))))
 486   //   (IfFalse (If (Bool EQ (CmpP ptr null))))
 487   // null checks are always branch-if-eq.  If we see a IfTrue projection
 488   // then we are replacing a 'ne' test with a 'eq' null check test.
 489   // We need to flip the projections to keep the same semantics.
 490   if( proj->Opcode() == Op_IfTrue ) {
 491     // Swap order of projections in basic block to swap branch targets
 492     Node *tmp1 = block->get_node(block->end_idx()+1);
 493     Node *tmp2 = block->get_node(block->end_idx()+2);
 494     block->map_node(tmp2, block->end_idx()+1);
 495     block->map_node(tmp1, block->end_idx()+2);
 496     Node *tmp = new Node(C->top()); // Use not null input
 497     tmp1->replace_by(tmp);
 498     tmp2->replace_by(tmp1);
 499     tmp->replace_by(tmp2);
 500     tmp->destruct(nullptr);
 501   }
 502 
 503   // Remove the existing null check; use a new implicit null check instead.
 504   // Since schedule-local needs precise def-use info, we need to correct
 505   // it as well.
 506   Node *old_tst = proj->in(0);
 507   MachNode *nul_chk = new MachNullCheckNode(old_tst->in(0),best,bidx);
 508   block->map_node(nul_chk, block->end_idx());
 509   map_node_to_block(nul_chk, block);
 510   // Redirect users of old_test to nul_chk
 511   for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
 512     old_tst->last_out(i2)->set_req(0, nul_chk);
 513   // Clean-up any dead code
 514   for (uint i3 = 0; i3 < old_tst->req(); i3++) {
 515     Node* in = old_tst->in(i3);
 516     old_tst->set_req(i3, nullptr);
 517     if (in->outcnt() == 0) {
 518       // Remove dead input node
 519       in->disconnect_inputs(C);
 520       block->find_remove(in);
 521     }
 522   }
 523 
 524   latency_from_uses(nul_chk);
 525   latency_from_uses(best);
 526 
 527   // insert anti-dependences to defs in this block
 528   if (! best->needs_anti_dependence_check()) {
 529     for (uint k = 1; k < block->number_of_nodes(); k++) {
 530       Node *n = block->get_node(k);
 531       if (n->needs_anti_dependence_check() &&
 532           n->in(LoadNode::Memory) == best->in(StoreNode::Memory)) {
 533         // Found anti-dependent load
 534         raise_above_anti_dependences(block, n);
 535         if (C->failing()) {
 536           return;
 537         }
 538       }
 539     }
 540   }
 541 }
 542 
 543 
 544 //------------------------------select-----------------------------------------
 545 // Select a nice fellow from the worklist to schedule next. If there is only one
 546 // choice, then use it. CreateEx nodes that are initially ready must start their
 547 // blocks and are given the highest priority, by being placed at the beginning
 548 // of the worklist. Next after initially-ready CreateEx nodes are projections,
 549 // which must follow their parents, and CreateEx nodes with local input
 550 // dependencies. Next are constants and CheckCastPP nodes. There are a number of
 551 // other special cases, for instructions that consume condition codes, et al.
 552 // These are chosen immediately. Some instructions are required to immediately
 553 // precede the last instruction in the block, and these are taken last. Of the
 554 // remaining cases (most), choose the instruction with the greatest latency
 555 // (that is, the most number of pseudo-cycles required to the end of the
 556 // routine). If there is a tie, choose the instruction with the most inputs.
 557 Node* PhaseCFG::select(
 558   Block* block,
 559   Node_List &worklist,
 560   GrowableArray<int> &ready_cnt,
 561   VectorSet &next_call,
 562   uint sched_slot,
 563   intptr_t* recalc_pressure_nodes) {
 564 
 565   // If only a single entry on the stack, use it
 566   uint cnt = worklist.size();
 567   if (cnt == 1) {
 568     Node *n = worklist[0];
 569     worklist.map(0,worklist.pop());
 570     return n;
 571   }
 572 
 573   uint choice  = 0; // Bigger is most important
 574   uint latency = 0; // Bigger is scheduled first
 575   uint score   = 0; // Bigger is better
 576   int idx = -1;     // Index in worklist
 577   int cand_cnt = 0; // Candidate count
 578   bool block_size_threshold_ok = (recalc_pressure_nodes != nullptr) && (block->number_of_nodes() > 10);
 579 
 580   for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
 581     // Order in worklist is used to break ties.
 582     // See caller for how this is used to delay scheduling
 583     // of induction variable increments to after the other
 584     // uses of the phi are scheduled.
 585     Node *n = worklist[i];      // Get Node on worklist
 586 
 587     int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
 588     if (iop == Op_CreateEx || n->is_Proj()) {
 589       // CreateEx nodes that are initially ready must start the block (after Phi
 590       // and Parm nodes which are pre-scheduled) and get top priority. This is
 591       // currently enforced by placing them at the beginning of the initial
 592       // worklist and selecting them eagerly here. After these, projections and
 593       // other CreateEx nodes are selected with equal priority.
 594       worklist.map(i,worklist.pop());
 595       return n;
 596     }
 597 
 598     if (n->Opcode() == Op_Con || iop == Op_CheckCastPP) {
 599       // Constants and CheckCastPP nodes have higher priority than the rest of
 600       // the nodes tested below. Record as current winner, but keep looking for
 601       // higher-priority nodes in the worklist.
 602       choice  = 4;
 603       // Latency and score are only used to break ties among low-priority nodes.
 604       latency = 0;
 605       score   = 0;
 606       idx     = i;
 607       continue;
 608     }
 609 
 610     // Final call in a block must be adjacent to 'catch'
 611     Node *e = block->end();
 612     if( e->is_Catch() && e->in(0)->in(0) == n )
 613       continue;
 614 
 615     // Memory op for an implicit null check has to be at the end of the block
 616     if( e->is_MachNullCheck() && e->in(1) == n )
 617       continue;
 618 
 619     // Schedule IV increment last.
 620     if (e->is_Mach() && e->as_Mach()->ideal_Opcode() == Op_CountedLoopEnd) {
 621       // Cmp might be matched into CountedLoopEnd node.
 622       Node *cmp = (e->in(1)->ideal_reg() == Op_RegFlags) ? e->in(1) : e;
 623       if (cmp->req() > 1 && cmp->in(1) == n && n->is_iteratively_computed()) {
 624         continue;
 625       }
 626     }
 627 
 628     uint n_choice = 2;
 629 
 630     // See if this instruction is consumed by a branch. If so, then (as the
 631     // branch is the last instruction in the basic block) force it to the
 632     // end of the basic block
 633     if ( must_clone[iop] ) {
 634       // See if any use is a branch
 635       bool found_machif = false;
 636 
 637       for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
 638         Node* use = n->fast_out(j);
 639 
 640         // The use is a conditional branch, make them adjacent
 641         if (use->is_MachIf() && get_block_for_node(use) == block) {
 642           found_machif = true;
 643           break;
 644         }
 645 
 646         // More than this instruction pending for successor to be ready,
 647         // don't choose this if other opportunities are ready
 648         if (ready_cnt.at(use->_idx) > 1)
 649           n_choice = 1;
 650       }
 651 
 652       // loop terminated, prefer not to use this instruction
 653       if (found_machif)
 654         continue;
 655     }
 656 
 657     // See if this has a predecessor that is "must_clone", i.e. sets the
 658     // condition code. If so, choose this first
 659     for (uint j = 0; j < n->req() ; j++) {
 660       Node *inn = n->in(j);
 661       if (inn) {
 662         if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
 663           n_choice = 3;
 664           break;
 665         }
 666       }
 667     }
 668 
 669     // MachTemps should be scheduled last so they are near their uses
 670     if (n->is_MachTemp()) {
 671       n_choice = 1;
 672     }
 673 
 674     uint n_latency = get_latency_for_node(n);
 675     uint n_score = n->req();   // Many inputs get high score to break ties
 676 
 677     if (OptoRegScheduling && block_size_threshold_ok) {
 678       if (recalc_pressure_nodes[n->_idx] == 0x7fff7fff) {
 679         _regalloc->_scratch_int_pressure.init(_regalloc->_sched_int_pressure.high_pressure_limit());
 680         _regalloc->_scratch_float_pressure.init(_regalloc->_sched_float_pressure.high_pressure_limit());
 681         // simulate the notion that we just picked this node to schedule
 682         n->add_flag(Node::Flag_is_scheduled);
 683         // now calculate its effect upon the graph if we did
 684         adjust_register_pressure(n, block, recalc_pressure_nodes, false);
 685         // return its state for finalize in case somebody else wins
 686         n->remove_flag(Node::Flag_is_scheduled);
 687         // now save the two final pressure components of register pressure, limiting pressure calcs to short size
 688         short int_pressure = (short)_regalloc->_scratch_int_pressure.current_pressure();
 689         short float_pressure = (short)_regalloc->_scratch_float_pressure.current_pressure();
 690         recalc_pressure_nodes[n->_idx] = int_pressure;
 691         recalc_pressure_nodes[n->_idx] |= (float_pressure << 16);
 692       }
 693 
 694       if (_scheduling_for_pressure) {
 695         latency = n_latency;
 696         if (n_choice != 3) {
 697           // Now evaluate each register pressure component based on threshold in the score.
 698           // In general the defining register type will dominate the score, ergo we will not see register pressure grow on both banks
 699           // on a single instruction, but we might see it shrink on both banks.
 700           // For each use of register that has a register class that is over the high pressure limit, we build n_score up for
 701           // live ranges that terminate on this instruction.
 702           if (_regalloc->_sched_int_pressure.current_pressure() > _regalloc->_sched_int_pressure.high_pressure_limit()) {
 703             short int_pressure = (short)recalc_pressure_nodes[n->_idx];
 704             n_score = (int_pressure < 0) ? ((score + n_score) - int_pressure) : (int_pressure > 0) ? 1 : n_score;
 705           }
 706           if (_regalloc->_sched_float_pressure.current_pressure() > _regalloc->_sched_float_pressure.high_pressure_limit()) {
 707             short float_pressure = (short)(recalc_pressure_nodes[n->_idx] >> 16);
 708             n_score = (float_pressure < 0) ? ((score + n_score) - float_pressure) : (float_pressure > 0) ? 1 : n_score;
 709           }
 710         } else {
 711           // make sure we choose these candidates
 712           score = 0;
 713         }
 714       }
 715     }
 716 
 717     // Keep best latency found
 718     cand_cnt++;
 719     if (choice < n_choice ||
 720         (choice == n_choice &&
 721          ((StressLCM && C->randomized_select(cand_cnt)) ||
 722           (!StressLCM &&
 723            (latency < n_latency ||
 724             (latency == n_latency &&
 725              (score < n_score))))))) {
 726       choice  = n_choice;
 727       latency = n_latency;
 728       score   = n_score;
 729       idx     = i;               // Also keep index in worklist
 730     }
 731   } // End of for all ready nodes in worklist
 732 
 733   guarantee(idx >= 0, "index should be set");
 734   Node *n = worklist[(uint)idx];      // Get the winner
 735 
 736   worklist.map((uint)idx, worklist.pop());     // Compress worklist
 737   return n;
 738 }
 739 
 740 //-------------------------adjust_register_pressure----------------------------
 741 void PhaseCFG::adjust_register_pressure(Node* n, Block* block, intptr_t* recalc_pressure_nodes, bool finalize_mode) {
 742   PhaseLive* liveinfo = _regalloc->get_live();
 743   IndexSet* liveout = liveinfo->live(block);
 744   // first adjust the register pressure for the sources
 745   for (uint i = 1; i < n->req(); i++) {
 746     bool lrg_ends = false;
 747     Node *src_n = n->in(i);
 748     if (src_n == nullptr) continue;
 749     if (!src_n->is_Mach()) continue;
 750     uint src = _regalloc->_lrg_map.find(src_n);
 751     if (src == 0) continue;
 752     LRG& lrg_src = _regalloc->lrgs(src);
 753     // detect if the live range ends or not
 754     if (liveout->member(src) == false) {
 755       lrg_ends = true;
 756       for (DUIterator_Fast jmax, j = src_n->fast_outs(jmax); j < jmax; j++) {
 757         Node* m = src_n->fast_out(j); // Get user
 758         if (m == n) continue;
 759         if (!m->is_Mach()) continue;
 760         MachNode *mach = m->as_Mach();
 761         bool src_matches = false;
 762         int iop = mach->ideal_Opcode();
 763 
 764         switch (iop) {
 765         case Op_StoreB:
 766         case Op_StoreC:
 767         case Op_StoreD:
 768         case Op_StoreF:
 769         case Op_StoreI:
 770         case Op_StoreL:
 771         case Op_StoreLSpecial:
 772         case Op_StoreP:
 773         case Op_StoreN:
 774         case Op_StoreVector:
 775         case Op_StoreVectorMasked:
 776         case Op_StoreVectorScatter:
 777         case Op_StoreVectorScatterMasked:
 778         case Op_StoreNKlass:
 779           for (uint k = 1; k < m->req(); k++) {
 780             Node *in = m->in(k);
 781             if (in == src_n) {
 782               src_matches = true;
 783               break;
 784             }
 785           }
 786           break;
 787 
 788         default:
 789           src_matches = true;
 790           break;
 791         }
 792 
 793         // If we have a store as our use, ignore the non source operands
 794         if (src_matches == false) continue;
 795 
 796         // Mark every unscheduled use which is not n with a recalculation
 797         if ((get_block_for_node(m) == block) && (!m->is_scheduled())) {
 798           if (finalize_mode && !m->is_Phi()) {
 799             recalc_pressure_nodes[m->_idx] = 0x7fff7fff;
 800           }
 801           lrg_ends = false;
 802         }
 803       }
 804     }
 805     // if none, this live range ends and we can adjust register pressure
 806     if (lrg_ends) {
 807       if (finalize_mode) {
 808         _regalloc->lower_pressure(block, 0, lrg_src, nullptr, _regalloc->_sched_int_pressure, _regalloc->_sched_float_pressure);
 809       } else {
 810         _regalloc->lower_pressure(block, 0, lrg_src, nullptr, _regalloc->_scratch_int_pressure, _regalloc->_scratch_float_pressure);
 811       }
 812     }
 813   }
 814 
 815   // now add the register pressure from the dest and evaluate which heuristic we should use:
 816   // 1.) The default, latency scheduling
 817   // 2.) Register pressure scheduling based on the high pressure limit threshold for int or float register stacks
 818   uint dst = _regalloc->_lrg_map.find(n);
 819   if (dst != 0) {
 820     LRG& lrg_dst = _regalloc->lrgs(dst);
 821     if (finalize_mode) {
 822       _regalloc->raise_pressure(block, lrg_dst, _regalloc->_sched_int_pressure, _regalloc->_sched_float_pressure);
 823       // check to see if we fall over the register pressure cliff here
 824       if (_regalloc->_sched_int_pressure.current_pressure() > _regalloc->_sched_int_pressure.high_pressure_limit()) {
 825         _scheduling_for_pressure = true;
 826       } else if (_regalloc->_sched_float_pressure.current_pressure() > _regalloc->_sched_float_pressure.high_pressure_limit()) {
 827         _scheduling_for_pressure = true;
 828       } else {
 829         // restore latency scheduling mode
 830         _scheduling_for_pressure = false;
 831       }
 832     } else {
 833       _regalloc->raise_pressure(block, lrg_dst, _regalloc->_scratch_int_pressure, _regalloc->_scratch_float_pressure);
 834     }
 835   }
 836 }
 837 
 838 //------------------------------set_next_call----------------------------------
 839 void PhaseCFG::set_next_call(const Block* block, Node* init, VectorSet& next_call) const {
 840   Node_List worklist;
 841   worklist.push(init);
 842 
 843   while (worklist.size() > 0) {
 844     Node* n = worklist.pop();
 845     if (next_call.test_set(n->_idx)) continue;
 846     for (uint i = 0; i < n->len(); i++) {
 847       Node* m = n->in(i);
 848       if (m == nullptr) continue;  // must see all nodes in block that precede call
 849       if (get_block_for_node(m) == block) {
 850         worklist.push(m);
 851       }
 852     }
 853   }
 854 }
 855 
 856 //------------------------------needed_for_next_call---------------------------
 857 // Set the flag 'next_call' for each Node that is needed for the next call to
 858 // be scheduled.  This flag lets me bias scheduling so Nodes needed for the
 859 // next subroutine call get priority - basically it moves things NOT needed
 860 // for the next call till after the call.  This prevents me from trying to
 861 // carry lots of stuff live across a call.
 862 void PhaseCFG::needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call) {
 863   // Find the next control-defining Node in this block
 864   Node* call = nullptr;
 865   for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
 866     Node* m = this_call->fast_out(i);
 867     if (get_block_for_node(m) == block && // Local-block user
 868         m != this_call &&       // Not self-start node
 869         m->is_MachCall()) {
 870       call = m;
 871       break;
 872     }
 873   }
 874   if (call == nullptr)  return;    // No next call (e.g., block end is near)
 875   // Set next-call for all inputs to this call
 876   set_next_call(block, call, next_call);
 877 }
 878 
 879 //------------------------------add_call_kills-------------------------------------
 880 // helper function that adds caller save registers to MachProjNode
 881 static void add_call_kills(MachProjNode *proj, RegMask& regs, const char* save_policy, bool exclude_soe) {
 882   // Fill in the kill mask for the call
 883   for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
 884     if (!regs.member(r)) { // Not already defined by the call
 885       // Save-on-call register?
 886       if ((save_policy[r] == 'C') ||
 887           (save_policy[r] == 'A') ||
 888           ((save_policy[r] == 'E') && exclude_soe)) {
 889         proj->_rout.insert(r);
 890       }
 891     }
 892   }
 893 }
 894 
 895 
 896 //------------------------------sched_call-------------------------------------
 897 uint PhaseCFG::sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call) {
 898   ResourceMark rm(C->regmask_arena());
 899   RegMask regs(C->regmask_arena());
 900 
 901   // Schedule all the users of the call right now.  All the users are
 902   // projection Nodes, so they must be scheduled next to the call.
 903   // Collect all the defined registers.
 904   for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
 905     Node* n = mcall->fast_out(i);
 906     assert( n->is_MachProj(), "" );
 907     int n_cnt = ready_cnt.at(n->_idx)-1;
 908     ready_cnt.at_put(n->_idx, n_cnt);
 909     assert( n_cnt == 0, "" );
 910     // Schedule next to call
 911     block->map_node(n, node_cnt++);
 912     // Collect defined registers
 913     regs.or_with(n->out_RegMask());
 914     // Check for scheduling the next control-definer
 915     if( n->bottom_type() == Type::CONTROL )
 916       // Warm up next pile of heuristic bits
 917       needed_for_next_call(block, n, next_call);
 918 
 919     // Children of projections are now all ready
 920     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
 921       Node* m = n->fast_out(j); // Get user
 922       if(get_block_for_node(m) != block) {
 923         continue;
 924       }
 925       if( m->is_Phi() ) continue;
 926       int m_cnt = ready_cnt.at(m->_idx) - 1;
 927       ready_cnt.at_put(m->_idx, m_cnt);
 928       if( m_cnt == 0 )
 929         worklist.push(m);
 930     }
 931 
 932   }
 933 
 934   // Act as if the call defines the Frame Pointer.
 935   // Certainly the FP is alive and well after the call.
 936   regs.insert(_matcher.c_frame_pointer());
 937 
 938   // Set all registers killed and not already defined by the call.
 939   uint r_cnt = mcall->tf()->range_cc()->cnt();
 940   int op = mcall->ideal_Opcode();
 941   MachProjNode* proj = new MachProjNode(mcall, r_cnt + 1, RegMask::EMPTY, MachProjNode::fat_proj);
 942   map_node_to_block(proj, block);
 943   block->insert_node(proj, node_cnt++);
 944 
 945   // Select the right register save policy.
 946   const char *save_policy = nullptr;
 947   switch (op) {
 948     case Op_CallRuntime:
 949     case Op_CallLeaf:
 950     case Op_CallLeafNoFP:
 951     case Op_CallLeafVector:
 952       // Calling C code so use C calling convention
 953       save_policy = _matcher._c_reg_save_policy;
 954       break;
 955 
 956     case Op_CallStaticJava:
 957     case Op_CallDynamicJava:
 958       // Calling Java code so use Java calling convention
 959       save_policy = _matcher._register_save_policy;
 960       break;
 961 
 962     default:
 963       ShouldNotReachHere();
 964   }
 965 
 966   // When using CallRuntime mark SOE registers as killed by the call
 967   // so values that could show up in the RegisterMap aren't live in a
 968   // callee saved register since the register wouldn't know where to
 969   // find them.  CallLeaf and CallLeafNoFP are ok because they can't
 970   // have debug info on them.  Strictly speaking this only needs to be
 971   // done for oops since idealreg2debugmask takes care of debug info
 972   // references but there no way to handle oops differently than other
 973   // pointers as far as the kill mask goes.
 974   bool exclude_soe = op == Op_CallRuntime;
 975   add_call_kills(proj, regs, save_policy, exclude_soe);
 976 
 977   return node_cnt;
 978 }
 979 
 980 
 981 //------------------------------schedule_local---------------------------------
 982 // Topological sort within a block.  Someday become a real scheduler.
 983 bool PhaseCFG::schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call, intptr_t *recalc_pressure_nodes) {
 984   // Already "sorted" are the block start Node (as the first entry), and
 985   // the block-ending Node and any trailing control projections.  We leave
 986   // these alone.  PhiNodes and ParmNodes are made to follow the block start
 987   // Node.  Everything else gets topo-sorted.
 988 
 989 #ifndef PRODUCT
 990     if (trace_opto_pipelining()) {
 991       tty->print_cr("# --- schedule_local B%d, before: ---", block->_pre_order);
 992       for (uint i = 0;i < block->number_of_nodes(); i++) {
 993         tty->print("# ");
 994         block->get_node(i)->dump();
 995       }
 996       tty->print_cr("#");
 997     }
 998 #endif
 999 
1000   // RootNode is already sorted
1001   if (block->number_of_nodes() == 1) {
1002     return true;
1003   }
1004 
1005   bool block_size_threshold_ok = (recalc_pressure_nodes != nullptr) && (block->number_of_nodes() > 10);
1006 
1007   // We track the uses of local definitions as input dependences so that
1008   // we know when a given instruction is available to be scheduled.
1009   uint i;
1010   if (OptoRegScheduling && block_size_threshold_ok) {
1011     for (i = 1; i < block->number_of_nodes(); i++) { // setup nodes for pressure calc
1012       Node *n = block->get_node(i);
1013       n->remove_flag(Node::Flag_is_scheduled);
1014       if (!n->is_Phi()) {
1015         recalc_pressure_nodes[n->_idx] = 0x7fff7fff;
1016       }
1017     }
1018   }
1019 
1020   // Move PhiNodes and ParmNodes from 1 to cnt up to the start
1021   uint node_cnt = block->end_idx();
1022   uint phi_cnt = 1;
1023   for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
1024     Node *n = block->get_node(i);
1025     if( n->is_Phi() ||          // Found a PhiNode or ParmNode
1026         (n->is_Proj()  && n->in(0) == block->head()) ) {
1027       // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
1028       block->map_node(block->get_node(phi_cnt), i);
1029       block->map_node(n, phi_cnt++);  // swap Phi/Parm up front
1030       if (OptoRegScheduling && block_size_threshold_ok) {
1031         // mark n as scheduled
1032         n->add_flag(Node::Flag_is_scheduled);
1033       }
1034     } else {                    // All others
1035       // Count block-local inputs to 'n'
1036       uint cnt = n->len();      // Input count
1037       uint local = 0;
1038       for( uint j=0; j<cnt; j++ ) {
1039         Node *m = n->in(j);
1040         if( m && get_block_for_node(m) == block && !m->is_top() )
1041           local++;              // One more block-local input
1042       }
1043       ready_cnt.at_put(n->_idx, local); // Count em up
1044       // A few node types require changing a required edge to a precedence edge
1045       // before allocation.
1046       if( n->is_Mach() && n->req() > TypeFunc::Parms &&
1047           (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ||
1048            n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) {
1049         // MemBarAcquire could be created without Precedent edge.
1050         // del_req() replaces the specified edge with the last input edge
1051         // and then removes the last edge. If the specified edge > number of
1052         // edges the last edge will be moved outside of the input edges array
1053         // and the edge will be lost. This is why this code should be
1054         // executed only when Precedent (== TypeFunc::Parms) edge is present.
1055         Node *x = n->in(TypeFunc::Parms);
1056         if (x != nullptr && get_block_for_node(x) == block && n->find_prec_edge(x) != -1) {
1057           // Old edge to node within same block will get removed, but no precedence
1058           // edge will get added because it already exists. Update ready count.
1059           int cnt = ready_cnt.at(n->_idx);
1060           assert(cnt > 1, "MemBar node %d must not get ready here", n->_idx);
1061           ready_cnt.at_put(n->_idx, cnt-1);
1062         }
1063         n->del_req(TypeFunc::Parms);
1064         n->add_prec(x);
1065       }
1066     }
1067   }
1068   for(uint i2=i; i2< block->number_of_nodes(); i2++ ) // Trailing guys get zapped count
1069     ready_cnt.at_put(block->get_node(i2)->_idx, 0);
1070 
1071   // All the prescheduled guys do not hold back internal nodes
1072   uint i3;
1073   for (i3 = 0; i3 < phi_cnt; i3++) {  // For all pre-scheduled
1074     Node *n = block->get_node(i3);       // Get pre-scheduled
1075     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1076       Node* m = n->fast_out(j);
1077       if (get_block_for_node(m) == block) { // Local-block user
1078         int m_cnt = ready_cnt.at(m->_idx)-1;
1079         if (OptoRegScheduling && block_size_threshold_ok) {
1080           // mark m as scheduled
1081           if (m_cnt < 0) {
1082             m->add_flag(Node::Flag_is_scheduled);
1083           }
1084         }
1085         ready_cnt.at_put(m->_idx, m_cnt);   // Fix ready count
1086       }
1087     }
1088   }
1089 
1090   Node_List delay;
1091   // Make a worklist
1092   Node_List worklist;
1093   for(uint i4=i3; i4<node_cnt; i4++ ) {    // Put ready guys on worklist
1094     Node *m = block->get_node(i4);
1095     if( !ready_cnt.at(m->_idx) ) {   // Zero ready count?
1096       if (m->is_iteratively_computed()) {
1097         // Push induction variable increments last to allow other uses
1098         // of the phi to be scheduled first. The select() method breaks
1099         // ties in scheduling by worklist order.
1100         delay.push(m);
1101       } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
1102         // Place CreateEx nodes that are initially ready at the beginning of the
1103         // worklist so they are selected first and scheduled at the block start.
1104         worklist.insert(0, m);
1105       } else {
1106         worklist.push(m);         // Then on to worklist!
1107       }
1108     }
1109   }
1110   while (delay.size()) {
1111     Node* d = delay.pop();
1112     worklist.push(d);
1113   }
1114 
1115   if (OptoRegScheduling && block_size_threshold_ok) {
1116     // To stage register pressure calculations we need to examine the live set variables
1117     // breaking them up by register class to compartmentalize the calculations.
1118     _regalloc->_sched_int_pressure.init(Matcher::int_pressure_limit());
1119     _regalloc->_sched_float_pressure.init(Matcher::float_pressure_limit());
1120     _regalloc->_scratch_int_pressure.init(Matcher::int_pressure_limit());
1121     _regalloc->_scratch_float_pressure.init(Matcher::float_pressure_limit());
1122 
1123     _regalloc->compute_entry_block_pressure(block);
1124   }
1125 
1126   // Warm up the 'next_call' heuristic bits
1127   needed_for_next_call(block, block->head(), next_call);
1128 
1129 #ifndef PRODUCT
1130     if (trace_opto_pipelining()) {
1131       for (uint j=0; j< block->number_of_nodes(); j++) {
1132         Node     *n = block->get_node(j);
1133         int     idx = n->_idx;
1134         tty->print("#   ready cnt:%3d  ", ready_cnt.at(idx));
1135         tty->print("latency:%3d  ", get_latency_for_node(n));
1136         tty->print("%4d: %s\n", idx, n->Name());
1137       }
1138     }
1139 #endif
1140 
1141   uint max_idx = (uint)ready_cnt.length();
1142   // Pull from worklist and schedule
1143   while( worklist.size() ) {    // Worklist is not ready
1144 
1145 #ifndef PRODUCT
1146     if (trace_opto_pipelining()) {
1147       tty->print("#   ready list:");
1148       for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
1149         Node *n = worklist[i];      // Get Node on worklist
1150         tty->print(" %d", n->_idx);
1151       }
1152       tty->cr();
1153     }
1154 #endif
1155 
1156     // Select and pop a ready guy from worklist
1157     Node* n = select(block, worklist, ready_cnt, next_call, phi_cnt, recalc_pressure_nodes);
1158     block->map_node(n, phi_cnt++);    // Schedule him next
1159 
1160     if (OptoRegScheduling && block_size_threshold_ok) {
1161       n->add_flag(Node::Flag_is_scheduled);
1162 
1163       // Now adjust the resister pressure with the node we selected
1164       if (!n->is_Phi()) {
1165         adjust_register_pressure(n, block, recalc_pressure_nodes, true);
1166       }
1167     }
1168 
1169 #ifndef PRODUCT
1170     if (trace_opto_pipelining()) {
1171       tty->print("#    select %d: %s", n->_idx, n->Name());
1172       tty->print(", latency:%d", get_latency_for_node(n));
1173       n->dump();
1174       if (Verbose) {
1175         tty->print("#   ready list:");
1176         for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
1177           Node *n = worklist[i];      // Get Node on worklist
1178           tty->print(" %d", n->_idx);
1179         }
1180         tty->cr();
1181       }
1182     }
1183 
1184 #endif
1185     if( n->is_MachCall() ) {
1186       MachCallNode *mcall = n->as_MachCall();
1187       phi_cnt = sched_call(block, phi_cnt, worklist, ready_cnt, mcall, next_call);
1188       continue;
1189     }
1190 
1191     if (n->is_Mach() && n->as_Mach()->has_call()) {
1192       RegMask regs;
1193       regs.insert(_matcher.c_frame_pointer());
1194       regs.or_with(n->out_RegMask());
1195 
1196       MachProjNode* proj = new MachProjNode(n, 1, RegMask::EMPTY, MachProjNode::fat_proj);
1197       map_node_to_block(proj, block);
1198       block->insert_node(proj, phi_cnt++);
1199 
1200       add_call_kills(proj, regs, _matcher._c_reg_save_policy, false);
1201     }
1202 
1203     // Children are now all ready
1204     for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
1205       Node* m = n->fast_out(i5); // Get user
1206       if (get_block_for_node(m) != block) {
1207         continue;
1208       }
1209       if( m->is_Phi() ) continue;
1210       if (m->_idx >= max_idx) { // new node, skip it
1211         assert(m->is_MachProj() && n->is_Mach() && n->as_Mach()->has_call(), "unexpected node types");
1212         continue;
1213       }
1214       int m_cnt = ready_cnt.at(m->_idx) - 1;
1215       ready_cnt.at_put(m->_idx, m_cnt);
1216       if( m_cnt == 0 )
1217         worklist.push(m);
1218     }
1219   }
1220 
1221   if( phi_cnt != block->end_idx() ) {
1222     // did not schedule all.  Retry, Bailout, or Die
1223     if (C->subsume_loads() == true && !C->failing()) {
1224       // Retry with subsume_loads == false
1225       // If this is the first failure, the sentinel string will "stick"
1226       // to the Compile object, and the C2Compiler will see it and retry.
1227       C->record_failure(C2Compiler::retry_no_subsuming_loads());
1228     } else {
1229       assert(C->failure_is_artificial(), "graph should be schedulable");
1230     }
1231     // assert( phi_cnt == end_idx(), "did not schedule all" );
1232     return false;
1233   }
1234 
1235   if (OptoRegScheduling && block_size_threshold_ok) {
1236     _regalloc->compute_exit_block_pressure(block);
1237     block->_reg_pressure = _regalloc->_sched_int_pressure.final_pressure();
1238     block->_freg_pressure = _regalloc->_sched_float_pressure.final_pressure();
1239   }
1240 
1241 #ifndef PRODUCT
1242   if (trace_opto_pipelining()) {
1243     tty->print_cr("#");
1244     tty->print_cr("# after schedule_local");
1245     for (uint i = 0;i < block->number_of_nodes();i++) {
1246       tty->print("# ");
1247       block->get_node(i)->dump();
1248     }
1249     tty->print_cr("# ");
1250 
1251     if (OptoRegScheduling && block_size_threshold_ok) {
1252       tty->print_cr("# pressure info : %d", block->_pre_order);
1253       _regalloc->print_pressure_info(_regalloc->_sched_int_pressure, "int register info");
1254       _regalloc->print_pressure_info(_regalloc->_sched_float_pressure, "float register info");
1255     }
1256     tty->cr();
1257   }
1258 #endif
1259 
1260   return true;
1261 }
1262 
1263 //--------------------------catch_cleanup_fix_all_inputs-----------------------
1264 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
1265   for (uint l = 0; l < use->len(); l++) {
1266     if (use->in(l) == old_def) {
1267       if (l < use->req()) {
1268         use->set_req(l, new_def);
1269       } else {
1270         use->rm_prec(l);
1271         use->add_prec(new_def);
1272         l--;
1273       }
1274     }
1275   }
1276 }
1277 
1278 //------------------------------catch_cleanup_find_cloned_def------------------
1279 Node* PhaseCFG::catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
1280   assert( use_blk != def_blk, "Inter-block cleanup only");
1281 
1282   // The use is some block below the Catch.  Find and return the clone of the def
1283   // that dominates the use. If there is no clone in a dominating block, then
1284   // create a phi for the def in a dominating block.
1285 
1286   // Find which successor block dominates this use.  The successor
1287   // blocks must all be single-entry (from the Catch only; I will have
1288   // split blocks to make this so), hence they all dominate.
1289   while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
1290     use_blk = use_blk->_idom;
1291 
1292   // Find the successor
1293   Node *fixup = nullptr;
1294 
1295   uint j;
1296   for( j = 0; j < def_blk->_num_succs; j++ )
1297     if( use_blk == def_blk->_succs[j] )
1298       break;
1299 
1300   if( j == def_blk->_num_succs ) {
1301     // Block at same level in dom-tree is not a successor.  It needs a
1302     // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
1303     Node_Array inputs;
1304     for(uint k = 1; k < use_blk->num_preds(); k++) {
1305       Block* block = get_block_for_node(use_blk->pred(k));
1306       inputs.map(k, catch_cleanup_find_cloned_def(block, def, def_blk, n_clone_idx));
1307     }
1308 
1309     // Check to see if the use_blk already has an identical phi inserted.
1310     // If it exists, it will be at the first position since all uses of a
1311     // def are processed together.
1312     Node *phi = use_blk->get_node(1);
1313     if( phi->is_Phi() ) {
1314       fixup = phi;
1315       for (uint k = 1; k < use_blk->num_preds(); k++) {
1316         if (phi->in(k) != inputs[k]) {
1317           // Not a match
1318           fixup = nullptr;
1319           break;
1320         }
1321       }
1322     }
1323 
1324     // If an existing PhiNode was not found, make a new one.
1325     if (fixup == nullptr) {
1326       Node *new_phi = PhiNode::make(use_blk->head(), def);
1327       use_blk->insert_node(new_phi, 1);
1328       map_node_to_block(new_phi, use_blk);
1329       for (uint k = 1; k < use_blk->num_preds(); k++) {
1330         new_phi->set_req(k, inputs[k]);
1331       }
1332       fixup = new_phi;
1333     }
1334 
1335   } else {
1336     // Found the use just below the Catch.  Make it use the clone.
1337     fixup = use_blk->get_node(n_clone_idx);
1338   }
1339 
1340   return fixup;
1341 }
1342 
1343 //--------------------------catch_cleanup_intra_block--------------------------
1344 // Fix all input edges in use that reference "def".  The use is in the same
1345 // block as the def and both have been cloned in each successor block.
1346 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
1347 
1348   // Both the use and def have been cloned. For each successor block,
1349   // get the clone of the use, and make its input the clone of the def
1350   // found in that block.
1351 
1352   uint use_idx = blk->find_node(use);
1353   uint offset_idx = use_idx - beg;
1354   for( uint k = 0; k < blk->_num_succs; k++ ) {
1355     // Get clone in each successor block
1356     Block *sb = blk->_succs[k];
1357     Node *clone = sb->get_node(offset_idx+1);
1358     assert( clone->Opcode() == use->Opcode(), "" );
1359 
1360     // Make use-clone reference the def-clone
1361     catch_cleanup_fix_all_inputs(clone, def, sb->get_node(n_clone_idx));
1362   }
1363 }
1364 
1365 //------------------------------catch_cleanup_inter_block---------------------
1366 // Fix all input edges in use that reference "def".  The use is in a different
1367 // block than the def.
1368 void PhaseCFG::catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx) {
1369   if( !use_blk ) return;        // Can happen if the use is a precedence edge
1370 
1371   Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, n_clone_idx);
1372   catch_cleanup_fix_all_inputs(use, def, new_def);
1373 }
1374 
1375 //------------------------------call_catch_cleanup-----------------------------
1376 // If we inserted any instructions between a Call and his CatchNode,
1377 // clone the instructions on all paths below the Catch.
1378 void PhaseCFG::call_catch_cleanup(Block* block) {
1379 
1380   // End of region to clone
1381   uint end = block->end_idx();
1382   if( !block->get_node(end)->is_Catch() ) return;
1383   // Start of region to clone
1384   uint beg = end;
1385   while(!block->get_node(beg-1)->is_MachProj() ||
1386         !block->get_node(beg-1)->in(0)->is_MachCall() ) {
1387     beg--;
1388     assert(beg > 0,"Catch cleanup walking beyond block boundary");
1389   }
1390   // Range of inserted instructions is [beg, end)
1391   if( beg == end ) return;
1392 
1393   // Clone along all Catch output paths.  Clone area between the 'beg' and
1394   // 'end' indices.
1395   for( uint i = 0; i < block->_num_succs; i++ ) {
1396     Block *sb = block->_succs[i];
1397     // Clone the entire area; ignoring the edge fixup for now.
1398     for( uint j = end; j > beg; j-- ) {
1399       Node *clone = block->get_node(j-1)->clone();
1400       sb->insert_node(clone, 1);
1401       map_node_to_block(clone, sb);
1402       if (clone->needs_anti_dependence_check()) {
1403         raise_above_anti_dependences(sb, clone);
1404         if (C->failing()) {
1405           return;
1406         }
1407       }
1408     }
1409   }
1410 
1411 
1412   // Fixup edges.  Check the def-use info per cloned Node
1413   for(uint i2 = beg; i2 < end; i2++ ) {
1414     uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
1415     Node *n = block->get_node(i2);        // Node that got cloned
1416     // Need DU safe iterator because of edge manipulation in calls.
1417     Unique_Node_List* out = new Unique_Node_List();
1418     for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
1419       out->push(n->fast_out(j1));
1420     }
1421     uint max = out->size();
1422     for (uint j = 0; j < max; j++) {// For all users
1423       Node *use = out->pop();
1424       Block *buse = get_block_for_node(use);
1425       if( use->is_Phi() ) {
1426         for( uint k = 1; k < use->req(); k++ )
1427           if( use->in(k) == n ) {
1428             Block* b = get_block_for_node(buse->pred(k));
1429             Node *fixup = catch_cleanup_find_cloned_def(b, n, block, n_clone_idx);
1430             use->set_req(k, fixup);
1431           }
1432       } else {
1433         if (block == buse) {
1434           catch_cleanup_intra_block(use, n, block, beg, n_clone_idx);
1435         } else {
1436           catch_cleanup_inter_block(use, buse, n, block, n_clone_idx);
1437         }
1438       }
1439     } // End for all users
1440 
1441   } // End of for all Nodes in cloned area
1442 
1443   // Remove the now-dead cloned ops
1444   for(uint i3 = beg; i3 < end; i3++ ) {
1445     block->get_node(beg)->disconnect_inputs(C);
1446     block->remove_node(beg);
1447   }
1448 
1449   // If the successor blocks have a CreateEx node, move it back to the top
1450   for (uint i4 = 0; i4 < block->_num_succs; i4++) {
1451     Block *sb = block->_succs[i4];
1452     uint new_cnt = end - beg;
1453     // Remove any newly created, but dead, nodes by traversing their schedule
1454     // backwards. Here, a dead node is a node whose only outputs (if any) are
1455     // unused projections.
1456     for (uint j = new_cnt; j > 0; j--) {
1457       Node *n = sb->get_node(j);
1458       // Individual projections are examined together with all siblings when
1459       // their parent is visited.
1460       if (n->is_Proj()) {
1461         continue;
1462       }
1463       bool dead = true;
1464       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1465         Node* out = n->fast_out(i);
1466         // n is live if it has a non-projection output or a used projection.
1467         if (!out->is_Proj() || out->outcnt() > 0) {
1468           dead = false;
1469           break;
1470         }
1471       }
1472       if (dead) {
1473         // n's only outputs (if any) are unused projections scheduled next to n
1474         // (see PhaseCFG::select()). Remove these projections backwards.
1475         for (uint k = j + n->outcnt(); k > j; k--) {
1476           Node* proj = sb->get_node(k);
1477           assert(proj->is_Proj() && proj->in(0) == n,
1478                  "projection should correspond to dead node");
1479           proj->disconnect_inputs(C);
1480           sb->remove_node(k);
1481           new_cnt--;
1482         }
1483         // Now remove the node itself.
1484         n->disconnect_inputs(C);
1485         sb->remove_node(j);
1486         new_cnt--;
1487       }
1488     }
1489     // If any newly created nodes remain, move the CreateEx node to the top
1490     if (new_cnt > 0) {
1491       Node *cex = sb->get_node(1+new_cnt);
1492       if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
1493         sb->remove_node(1+new_cnt);
1494         sb->insert_node(cex, 1);
1495       }
1496     }
1497   }
1498 }