1 /*
   2  * Copyright (c) 1997, 2025, 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 "libadt/vectset.hpp"
  26 #include "memory/allocation.inline.hpp"
  27 #include "memory/resourceArea.hpp"
  28 #include "opto/block.hpp"
  29 #include "opto/c2compiler.hpp"
  30 #include "opto/callnode.hpp"
  31 #include "opto/cfgnode.hpp"
  32 #include "opto/chaitin.hpp"
  33 #include "opto/machnode.hpp"
  34 #include "opto/opcodes.hpp"
  35 #include "opto/phaseX.hpp"
  36 #include "opto/rootnode.hpp"
  37 #include "opto/runtime.hpp"
  38 #include "runtime/deoptimization.hpp"
  39 
  40 // Portions of code courtesy of Clifford Click
  41 
  42 // Optimization - Graph Style
  43 
  44 // To avoid float value underflow
  45 #define MIN_BLOCK_FREQUENCY 1.e-35f
  46 
  47 //----------------------------schedule_node_into_block-------------------------
  48 // Insert node n into block b. Look for projections of n and make sure they
  49 // are in b also.
  50 void PhaseCFG::schedule_node_into_block( Node *n, Block *b ) {
  51   // Set basic block of n, Add n to b,
  52   map_node_to_block(n, b);
  53   b->add_inst(n);
  54 
  55   // After Matching, nearly any old Node may have projections trailing it.
  56   // These are usually machine-dependent flags.  In any case, they might
  57   // float to another block below this one.  Move them up.
  58   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
  59     Node*  use  = n->fast_out(i);
  60     if (use->is_Proj()) {
  61       Block* buse = get_block_for_node(use);
  62       if (buse != b) {              // In wrong block?
  63         if (buse != nullptr) {
  64           buse->find_remove(use);   // Remove from wrong block
  65         }
  66         map_node_to_block(use, b);
  67         b->add_inst(use);
  68       }
  69     }
  70   }
  71 }
  72 
  73 //----------------------------replace_block_proj_ctrl-------------------------
  74 // Nodes that have is_block_proj() nodes as their control need to use
  75 // the appropriate Region for their actual block as their control since
  76 // the projection will be in a predecessor block.
  77 void PhaseCFG::replace_block_proj_ctrl( Node *n ) {
  78   const Node *in0 = n->in(0);
  79   assert(in0 != nullptr, "Only control-dependent");
  80   const Node *p = in0->is_block_proj();
  81   if (p != nullptr && p != n) {    // Control from a block projection?
  82     assert(!n->pinned() || n->is_MachConstantBase(), "only pinned MachConstantBase node is expected here");
  83     // Find trailing Region
  84     Block *pb = get_block_for_node(in0); // Block-projection already has basic block
  85     uint j = 0;
  86     if (pb->_num_succs != 1) {  // More then 1 successor?
  87       // Search for successor
  88       uint max = pb->number_of_nodes();
  89       assert( max > 1, "" );
  90       uint start = max - pb->_num_succs;
  91       // Find which output path belongs to projection
  92       for (j = start; j < max; j++) {
  93         if( pb->get_node(j) == in0 )
  94           break;
  95       }
  96       assert( j < max, "must find" );
  97       // Change control to match head of successor basic block
  98       j -= start;
  99     }
 100     n->set_req(0, pb->_succs[j]->head());
 101   }
 102 }
 103 
 104 bool PhaseCFG::is_dominator(Node* dom_node, Node* node) {
 105   assert(is_CFG(node) && is_CFG(dom_node), "node and dom_node must be CFG nodes");
 106   if (dom_node == node) {
 107     return true;
 108   }
 109   Block* d = find_block_for_node(dom_node);
 110   Block* n = find_block_for_node(node);
 111   assert(n != nullptr && d != nullptr, "blocks must exist");
 112 
 113   if (d == n) {
 114     if (dom_node->is_block_start()) {
 115       return true;
 116     }
 117     if (node->is_block_start()) {
 118       return false;
 119     }
 120     if (dom_node->is_block_proj()) {
 121       return false;
 122     }
 123     if (node->is_block_proj()) {
 124       return true;
 125     }
 126 
 127     assert(is_control_proj_or_safepoint(node), "node must be control projection or safepoint");
 128     assert(is_control_proj_or_safepoint(dom_node), "dom_node must be control projection or safepoint");
 129 
 130     // Neither 'node' nor 'dom_node' is a block start or block projection.
 131     // Check if 'dom_node' is above 'node' in the control graph.
 132     if (is_dominating_control(dom_node, node)) {
 133       return true;
 134     }
 135 
 136 #ifdef ASSERT
 137     // If 'dom_node' does not dominate 'node' then 'node' has to dominate 'dom_node'
 138     if (!is_dominating_control(node, dom_node)) {
 139       node->dump();
 140       dom_node->dump();
 141       assert(false, "neither dom_node nor node dominates the other");
 142     }
 143 #endif
 144 
 145     return false;
 146   }
 147   return d->dom_lca(n) == d;
 148 }
 149 
 150 bool PhaseCFG::is_CFG(Node* n) {
 151   return n->is_block_proj() || n->is_block_start() || is_control_proj_or_safepoint(n);
 152 }
 153 
 154 bool PhaseCFG::is_control_proj_or_safepoint(Node* n) const {
 155   bool result = n->is_ReachabilityFence() ||
 156                 (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_SafePoint) ||
 157                 (n->is_Proj() && n->as_Proj()->bottom_type() == Type::CONTROL);
 158   assert(!n->is_Proj() ||
 159          n->as_Proj()->bottom_type() != Type::CONTROL ||
 160          n->as_Proj()->_con == 0, "If control projection, it must be projection 0");
 161   return result;
 162 }
 163 
 164 Block* PhaseCFG::find_block_for_node(Node* n) const {
 165   if (n->is_block_start() || n->is_block_proj()) {
 166     return get_block_for_node(n);
 167   } else {
 168     // Walk the control graph up if 'n' is not a block start nor a block projection. In this case 'n' must be
 169     // an unmatched control projection or a not yet matched safepoint precedence edge in the middle of a block.
 170     assert(is_control_proj_or_safepoint(n), "must be control projection or safepoint");
 171     Node* ctrl = n->in(0);
 172     while (!ctrl->is_block_start()) {
 173       ctrl = ctrl->in(0);
 174     }
 175     return get_block_for_node(ctrl);
 176   }
 177 }
 178 
 179 // Walk up the control graph from 'n' and check if 'dom_ctrl' is found.
 180 bool PhaseCFG::is_dominating_control(Node* dom_ctrl, Node* n) {
 181   Node* ctrl = n->in(0);
 182   while (!ctrl->is_block_start()) {
 183     if (ctrl == dom_ctrl) {
 184       return true;
 185     }
 186     ctrl = ctrl->in(0);
 187   }
 188   return false;
 189 }
 190 
 191 
 192 //------------------------------schedule_pinned_nodes--------------------------
 193 // Set the basic block for Nodes pinned into blocks
 194 void PhaseCFG::schedule_pinned_nodes(VectorSet &visited) {
 195   // Allocate node stack of size C->live_nodes()+8 to avoid frequent realloc
 196   GrowableArray <Node*> spstack(C->live_nodes() + 8);
 197   spstack.push(_root);
 198   while (spstack.is_nonempty()) {
 199     Node* node = spstack.pop();
 200     if (!visited.test_set(node->_idx)) { // Test node and flag it as visited
 201       if (node->pinned() && !has_block(node)) {  // Pinned?  Nail it down!
 202         assert(node->in(0), "pinned Node must have Control");
 203         // Before setting block replace block_proj control edge
 204         replace_block_proj_ctrl(node);
 205         Node* input = node->in(0);
 206         while (!input->is_block_start()) {
 207           input = input->in(0);
 208         }
 209         Block* block = get_block_for_node(input); // Basic block of controlling input
 210         schedule_node_into_block(node, block);
 211       }
 212 
 213       // If the node has precedence edges (added when CastPP nodes are
 214       // removed in final_graph_reshaping), fix the control of the
 215       // node to cover the precedence edges and remove the
 216       // dependencies.
 217       Node* n = nullptr;
 218       for (uint i = node->len()-1; i >= node->req(); i--) {
 219         Node* m = node->in(i);
 220         if (m == nullptr) continue;
 221         assert(is_CFG(m), "must be a CFG node");
 222         node->rm_prec(i);
 223         if (n == nullptr) {
 224           n = m;
 225         } else {
 226           assert(is_dominator(n, m) || is_dominator(m, n), "one must dominate the other");
 227           n = is_dominator(n, m) ? m : n;
 228         }
 229       }
 230       if (n != nullptr) {
 231         assert(node->in(0), "control should have been set");
 232         assert(is_dominator(n, node->in(0)) || is_dominator(node->in(0), n), "one must dominate the other");
 233         if (!is_dominator(n, node->in(0))) {
 234           node->set_req(0, n);
 235         }
 236       }
 237 
 238       // process all inputs that are non null
 239       for (int i = node->len()-1; i >= 0; --i) {
 240         if (node->in(i) != nullptr) {
 241           spstack.push(node->in(i));
 242         }
 243       }
 244     }
 245   }
 246 }
 247 
 248 // Assert that new input b2 is dominated by all previous inputs.
 249 // Check this by by seeing that it is dominated by b1, the deepest
 250 // input observed until b2.
 251 static void assert_dom(Block* b1, Block* b2, Node* n, const PhaseCFG* cfg) {
 252   if (b1 == nullptr)  return;
 253   assert(b1->_dom_depth < b2->_dom_depth, "sanity");
 254   Block* tmp = b2;
 255   while (tmp != b1 && tmp != nullptr) {
 256     tmp = tmp->_idom;
 257   }
 258   if (tmp != b1) {
 259 #ifdef ASSERT
 260     // Detected an unschedulable graph.  Print some nice stuff and die.
 261     tty->print_cr("!!! Unschedulable graph !!!");
 262     for (uint j=0; j<n->len(); j++) { // For all inputs
 263       Node* inn = n->in(j); // Get input
 264       if (inn == nullptr)  continue;  // Ignore null, missing inputs
 265       Block* inb = cfg->get_block_for_node(inn);
 266       tty->print("B%d idom=B%d depth=%2d ",inb->_pre_order,
 267                  inb->_idom ? inb->_idom->_pre_order : 0, inb->_dom_depth);
 268       inn->dump();
 269     }
 270     tty->print("Failing node: ");
 271     n->dump();
 272     assert(false, "unschedulable graph");
 273 #endif
 274     cfg->C->record_failure("unschedulable graph");
 275   }
 276 }
 277 
 278 static Block* find_deepest_input(Node* n, const PhaseCFG* cfg) {
 279   // Find the last input dominated by all other inputs.
 280   Block* deepb           = nullptr;     // Deepest block so far
 281   int    deepb_dom_depth = 0;
 282   for (uint k = 0; k < n->len(); k++) { // For all inputs
 283     Node* inn = n->in(k);               // Get input
 284     if (inn == nullptr)  continue;      // Ignore null, missing inputs
 285     Block* inb = cfg->get_block_for_node(inn);
 286     assert(inb != nullptr, "must already have scheduled this input");
 287     if (deepb_dom_depth < (int) inb->_dom_depth) {
 288       // The new inb must be dominated by the previous deepb.
 289       // The various inputs must be linearly ordered in the dom
 290       // tree, or else there will not be a unique deepest block.
 291       assert_dom(deepb, inb, n, cfg);
 292       if (cfg->C->failing()) {
 293         return nullptr;
 294       }
 295       deepb = inb;                      // Save deepest block
 296       deepb_dom_depth = deepb->_dom_depth;
 297     }
 298   }
 299   assert(deepb != nullptr, "must be at least one input to n");
 300   return deepb;
 301 }
 302 
 303 
 304 //------------------------------schedule_early---------------------------------
 305 // Find the earliest Block any instruction can be placed in.  Some instructions
 306 // are pinned into Blocks.  Unpinned instructions can appear in last block in
 307 // which all their inputs occur.
 308 bool PhaseCFG::schedule_early(VectorSet &visited, Node_Stack &roots) {
 309   // Allocate stack with enough space to avoid frequent realloc
 310   Node_Stack nstack(roots.size() + 8);
 311   // _root will be processed among C->top() inputs
 312   roots.push(C->top(), 0);
 313   visited.set(C->top()->_idx);
 314 
 315   while (roots.size() != 0) {
 316     // Use local variables nstack_top_n & nstack_top_i to cache values
 317     // on stack's top.
 318     Node* parent_node = roots.node();
 319     uint  input_index = 0;
 320     roots.pop();
 321 
 322     while (true) {
 323       if (input_index == 0) {
 324         // Fixup some control.  Constants without control get attached
 325         // to root and nodes that use is_block_proj() nodes should be attached
 326         // to the region that starts their block.
 327         const Node* control_input = parent_node->in(0);
 328         if (control_input != nullptr) {
 329           replace_block_proj_ctrl(parent_node);
 330         } else {
 331           // Is a constant with NO inputs?
 332           if (parent_node->req() == 1) {
 333             parent_node->set_req(0, _root);
 334           }
 335         }
 336       }
 337 
 338       // First, visit all inputs and force them to get a block.  If an
 339       // input is already in a block we quit following inputs (to avoid
 340       // cycles). Instead we put that Node on a worklist to be handled
 341       // later (since IT'S inputs may not have a block yet).
 342 
 343       // Assume all n's inputs will be processed
 344       bool done = true;
 345 
 346       while (input_index < parent_node->len()) {
 347         Node* in = parent_node->in(input_index++);
 348         if (in == nullptr) {
 349           continue;
 350         }
 351 
 352         int is_visited = visited.test_set(in->_idx);
 353         if (!has_block(in)) {
 354           if (is_visited) {
 355             assert(false, "graph should be schedulable");
 356             return false;
 357           }
 358           // Save parent node and next input's index.
 359           nstack.push(parent_node, input_index);
 360           // Process current input now.
 361           parent_node = in;
 362           input_index = 0;
 363           // Not all n's inputs processed.
 364           done = false;
 365           break;
 366         } else if (!is_visited) {
 367           // Visit this guy later, using worklist
 368           roots.push(in, 0);
 369         }
 370       }
 371 
 372       if (done) {
 373         // All of n's inputs have been processed, complete post-processing.
 374 
 375         // Some instructions are pinned into a block.  These include Region,
 376         // Phi, Start, Return, and other control-dependent instructions and
 377         // any projections which depend on them.
 378         if (!parent_node->pinned()) {
 379           // Set earliest legal block.
 380           Block* earliest_block = find_deepest_input(parent_node, this);
 381           if (C->failing()) {
 382             return false;
 383           }
 384           map_node_to_block(parent_node, earliest_block);
 385         } else {
 386           assert(get_block_for_node(parent_node) == get_block_for_node(parent_node->in(0)), "Pinned Node should be at the same block as its control edge");
 387         }
 388 
 389         if (nstack.is_empty()) {
 390           // Finished all nodes on stack.
 391           // Process next node on the worklist 'roots'.
 392           break;
 393         }
 394         // Get saved parent node and next input's index.
 395         parent_node = nstack.node();
 396         input_index = nstack.index();
 397         nstack.pop();
 398       }
 399     }
 400   }
 401   return true;
 402 }
 403 
 404 //------------------------------dom_lca----------------------------------------
 405 // Find least common ancestor in dominator tree
 406 // LCA is a current notion of LCA, to be raised above 'this'.
 407 // As a convenient boundary condition, return 'this' if LCA is null.
 408 // Find the LCA of those two nodes.
 409 Block* Block::dom_lca(Block* LCA) {
 410   if (LCA == nullptr || LCA == this)  return this;
 411 
 412   Block* anc = this;
 413   while (anc->_dom_depth > LCA->_dom_depth)
 414     anc = anc->_idom;           // Walk up till anc is as high as LCA
 415 
 416   while (LCA->_dom_depth > anc->_dom_depth)
 417     LCA = LCA->_idom;           // Walk up till LCA is as high as anc
 418 
 419   while (LCA != anc) {          // Walk both up till they are the same
 420     LCA = LCA->_idom;
 421     anc = anc->_idom;
 422   }
 423 
 424   return LCA;
 425 }
 426 
 427 //--------------------------raise_LCA_above_use--------------------------------
 428 // We are placing a definition, and have been given a def->use edge.
 429 // The definition must dominate the use, so move the LCA upward in the
 430 // dominator tree to dominate the use.  If the use is a phi, adjust
 431 // the LCA only with the phi input paths which actually use this def.
 432 static Block* raise_LCA_above_use(Block* LCA, Node* use, Node* def, const PhaseCFG* cfg) {
 433   Block* buse = cfg->get_block_for_node(use);
 434   if (buse == nullptr) return LCA;   // Unused killing Projs have no use block
 435   if (!use->is_Phi())  return buse->dom_lca(LCA);
 436   uint pmax = use->req();       // Number of Phi inputs
 437   // Why does not this loop just break after finding the matching input to
 438   // the Phi?  Well...it's like this.  I do not have true def-use/use-def
 439   // chains.  Means I cannot distinguish, from the def-use direction, which
 440   // of many use-defs lead from the same use to the same def.  That is, this
 441   // Phi might have several uses of the same def.  Each use appears in a
 442   // different predecessor block.  But when I enter here, I cannot distinguish
 443   // which use-def edge I should find the predecessor block for.  So I find
 444   // them all.  Means I do a little extra work if a Phi uses the same value
 445   // more than once.
 446   for (uint j=1; j<pmax; j++) { // For all inputs
 447     if (use->in(j) == def) {    // Found matching input?
 448       Block* pred = cfg->get_block_for_node(buse->pred(j));
 449       LCA = pred->dom_lca(LCA);
 450     }
 451   }
 452   return LCA;
 453 }
 454 
 455 //----------------------------raise_LCA_above_marks----------------------------
 456 // Return a new LCA that dominates LCA and any of its marked predecessors.
 457 // Search all my parents up to 'early' (exclusive), looking for predecessors
 458 // which are marked with the given index.  Return the LCA (in the dom tree)
 459 // of all marked blocks.  If there are none marked, return the original
 460 // LCA.
 461 static Block* raise_LCA_above_marks(Block* LCA, node_idx_t mark, Block* early, const PhaseCFG* cfg) {
 462   assert(early->dominates(LCA), "precondition failed");
 463   Block_List worklist;
 464   worklist.push(LCA);
 465   while (worklist.size() > 0) {
 466     Block* mid = worklist.pop();
 467     if (mid == early)  continue;  // stop searching here
 468 
 469     // Test and set the visited bit.
 470     if (mid->raise_LCA_visited() == mark)  continue;  // already visited
 471 
 472     // Don't process the current LCA, otherwise the search may terminate early
 473     if (mid != LCA && mid->raise_LCA_mark() == mark) {
 474       // Raise the LCA.
 475       LCA = mid->dom_lca(LCA);
 476       if (LCA == early)  break;   // stop searching everywhere
 477       assert(early->dominates(LCA), "unsound LCA update");
 478       // Resume searching at that point, skipping intermediate levels.
 479       worklist.push(LCA);
 480       if (LCA == mid)
 481         continue; // Don't mark as visited to avoid early termination.
 482     } else {
 483       // Keep searching through this block's predecessors.
 484       for (uint j = 1, jmax = mid->num_preds(); j < jmax; j++) {
 485         Block* mid_parent = cfg->get_block_for_node(mid->pred(j));
 486         worklist.push(mid_parent);
 487       }
 488     }
 489     mid->set_raise_LCA_visited(mark);
 490   }
 491   return LCA;
 492 }
 493 
 494 //--------------------------memory_early_block--------------------------------
 495 // This is a variation of find_deepest_input, the heart of schedule_early.
 496 // Find the "early" block for a load, if we considered only memory and
 497 // address inputs, that is, if other data inputs were ignored.
 498 //
 499 // Because a subset of edges are considered, the resulting block will
 500 // be earlier (at a shallower dom_depth) than the true schedule_early
 501 // point of the node. We compute this earlier block as a more permissive
 502 // site for anti-dependency insertion, but only if subsume_loads is enabled.
 503 static Block* memory_early_block(Node* load, Block* early, const PhaseCFG* cfg) {
 504   Node* base;
 505   Node* index;
 506   Node* store = load->in(MemNode::Memory);
 507   load->as_Mach()->memory_inputs(base, index);
 508 
 509   assert(base != NodeSentinel && index != NodeSentinel,
 510          "unexpected base/index inputs");
 511 
 512   Node* mem_inputs[4];
 513   int mem_inputs_length = 0;
 514   if (base != nullptr)  mem_inputs[mem_inputs_length++] = base;
 515   if (index != nullptr) mem_inputs[mem_inputs_length++] = index;
 516   if (store != nullptr) mem_inputs[mem_inputs_length++] = store;
 517 
 518   // In the comparison below, add one to account for the control input,
 519   // which may be null, but always takes up a spot in the in array.
 520   if (mem_inputs_length + 1 < (int) load->req()) {
 521     // This "load" has more inputs than just the memory, base and index inputs.
 522     // For purposes of checking anti-dependences, we need to start
 523     // from the early block of only the address portion of the instruction,
 524     // and ignore other blocks that may have factored into the wider
 525     // schedule_early calculation.
 526     if (load->in(0) != nullptr) mem_inputs[mem_inputs_length++] = load->in(0);
 527 
 528     Block* deepb           = nullptr;        // Deepest block so far
 529     int    deepb_dom_depth = 0;
 530     for (int i = 0; i < mem_inputs_length; i++) {
 531       Block* inb = cfg->get_block_for_node(mem_inputs[i]);
 532       if (deepb_dom_depth < (int) inb->_dom_depth) {
 533         // The new inb must be dominated by the previous deepb.
 534         // The various inputs must be linearly ordered in the dom
 535         // tree, or else there will not be a unique deepest block.
 536         assert_dom(deepb, inb, load, cfg);
 537         if (cfg->C->failing()) {
 538           return nullptr;
 539         }
 540         deepb = inb;                      // Save deepest block
 541         deepb_dom_depth = deepb->_dom_depth;
 542       }
 543     }
 544     early = deepb;
 545   }
 546 
 547   return early;
 548 }
 549 
 550 // This function is used by raise_above_anti_dependences to find unrelated loads for stores in implicit null checks.
 551 bool PhaseCFG::unrelated_load_in_store_null_block(Node* store, Node* load) {
 552   // We expect an anti-dependence edge from 'load' to 'store', except when
 553   // implicit_null_check() has hoisted 'store' above its early block to
 554   // perform an implicit null check, and 'load' is placed in the null
 555   // block. In this case it is safe to ignore the anti-dependence, as the
 556   // null block is only reached if 'store' tries to write to null object and
 557   // 'load' read from non-null object (there is preceding check for that)
 558   // These objects can't be the same.
 559   Block* store_block = get_block_for_node(store);
 560   Block* load_block = get_block_for_node(load);
 561   Node* end = store_block->end();
 562   if (end->is_MachNullCheck() && (end->in(1) == store) && store_block->dominates(load_block)) {
 563     Node* if_true = end->find_out_with(Op_IfTrue);
 564     assert(if_true != nullptr, "null check without null projection");
 565     Node* null_block_region = if_true->find_out_with(Op_Region);
 566     assert(null_block_region != nullptr, "null check without null region");
 567     return get_block_for_node(null_block_region) == load_block;
 568   }
 569   return false;
 570 }
 571 
 572 class DefUseMemStatesQueue : public StackObj {
 573 private:
 574   class DefUsePair : public StackObj {
 575   private:
 576     Node* _def; // memory state
 577     Node* _use; // use of the memory state that also modifies the memory state
 578 
 579   public:
 580     DefUsePair(Node* def, Node* use) :
 581       _def(def), _use(use) {
 582     }
 583 
 584     DefUsePair() :
 585       _def(nullptr), _use(nullptr) {
 586     }
 587 
 588     Node* def() const {
 589       return _def;
 590     }
 591 
 592     Node* use() const {
 593       return _use;
 594     }
 595   };
 596 
 597   GrowableArray<DefUsePair> _queue;
 598   Unique_Node_List _worklist_visited; // visited mergemem nodes
 599 
 600   bool already_enqueued(Node* def_mem, PhiNode* use_phi) const {
 601     // def_mem is one of the inputs of use_phi and at least one input of use_phi is
 602     // not def_mem. It's however possible that use_phi has def_mem as input multiple
 603     // times. If that happens, use_phi is recorded as a use of def_mem multiple
 604     // times as well. When PhaseCFG::raise_above_anti_dependences() goes over
 605     // uses of def_mem and enqueues them for processing, use_phi would then be
 606     // enqueued for processing multiple times when it only needs to be
 607     // processed once. The code below checks if use_phi as a use of def_mem was
 608     // already enqueued to avoid redundant processing of use_phi.
 609     int j = _queue.length()-1;
 610     // If there are any use of def_mem already enqueued, they were enqueued
 611     // last (all use of def_mem are processed in one go).
 612     for (; j >= 0; j--) {
 613       const DefUsePair& def_use_pair = _queue.at(j);
 614       if (def_use_pair.def() != def_mem) {
 615         // We're done with the uses of def_mem
 616         break;
 617       }
 618       if (def_use_pair.use() == use_phi) {
 619         return true;
 620       }
 621     }
 622 #ifdef ASSERT
 623     for (; j >= 0; j--) {
 624       const DefUsePair& def_use_pair = _queue.at(j);
 625       assert(def_use_pair.def() != def_mem, "Should be done with the uses of def_mem");
 626     }
 627 #endif
 628     return false;
 629   }
 630 
 631 public:
 632   DefUseMemStatesQueue(ResourceArea* area) {
 633   }
 634 
 635   void push(Node* def_mem_state, Node* use_mem_state) {
 636     if (use_mem_state->is_MergeMem()) {
 637       // Be sure we don't get into combinatorial problems.
 638       if (_worklist_visited.member(use_mem_state)) {
 639         // already on work list; do not repeat
 640         return;
 641       }
 642       _worklist_visited.push(use_mem_state);
 643     } else if (use_mem_state->is_Phi()) {
 644       // A Phi could have the same mem as input multiple times. If that's the case, we don't need to enqueue it
 645       // more than once. We otherwise allow phis to be repeated; they can merge two relevant states.
 646       if (already_enqueued(def_mem_state, use_mem_state->as_Phi())) {
 647         return;
 648       }
 649     }
 650 
 651     _queue.push(DefUsePair(def_mem_state, use_mem_state));
 652   }
 653 
 654   bool is_nonempty() const {
 655     return _queue.is_nonempty();
 656   }
 657 
 658   Node* top_def() const {
 659     return _queue.top().def();
 660   }
 661 
 662   Node* top_use() const {
 663     return _queue.top().use();
 664   }
 665 
 666   void pop() {
 667     _queue.pop();
 668   }
 669 };
 670 
 671 // Enforce a scheduling of the given 'load' that ensures anti-dependent stores
 672 // do not overwrite the load's input memory state before the load executes.
 673 //
 674 // The given 'load' has a current scheduling range in the dominator tree that
 675 // starts at the load's early block (computed in schedule_early) and ends at
 676 // the given 'LCA' block for the load. However, there may still exist
 677 // anti-dependent stores between the early block and the LCA that overwrite
 678 // memory that the load must witness. For such stores, we must
 679 //
 680 //   1. raise the load's LCA to force the load to (eventually) be scheduled at
 681 //      latest in the store's block, and
 682 //   2. if the load may get scheduled in the store's block, additionally insert
 683 //      an anti-dependence edge (i.e., precedence edge) from the load to the
 684 //      store to ensure LCM schedules the load before the store within the
 685 //      block.
 686 //
 687 // For a given store, we say that the store is on a _distinct_ control-flow
 688 // path relative to the load if there are no paths from early to LCA that go
 689 // through the store's block. Such stores are not anti-dependent, and there is
 690 // no need to update the LCA nor to add anti-dependence edges.
 691 //
 692 // Due to the presence of loops, we must also raise the LCA above
 693 // anti-dependent memory Phis. We defer the details (see later comments in the
 694 // method) and for now look at an example without loops.
 695 //
 696 //          CFG               DOMINATOR TREE
 697 //
 698 //       B1 (early,L)              B1
 699 //       |\________                /\\___
 700 //       |         \              /  \   \
 701 //       B2 (L,S)   \            B2  B7  B6
 702 //      /  \         \           /\\___
 703 //     B3  B4 (S)    B7 (S)     /  \   \
 704 //      \  /         /         B3  B4  B5
 705 //       B5 (LCA,L) /
 706 //        \    ____/
 707 //         \  /
 708 //          B6
 709 //
 710 // Here, the load's scheduling range when calling raise_above_anti_dependences
 711 // is between early and LCA in the dominator tree, i.e., in block B1, B2, or B5
 712 // (indicated with "L"). However, there are a number of stores (indicated with
 713 // "S") that overwrite the memory which the load must witness. First, consider
 714 // the store in B4. We cannot legally schedule the load in B4, so an
 715 // anti-dependence edge is redundant. However, we must raise the LCA above
 716 // B4, which means that the updated LCA is B2. Now, consider the store in B2.
 717 // The LCA is already B2, so we do not need to raise it any further.
 718 // If we, eventually, decide to schedule the load in B2, it could happen that
 719 // LCM decides to place the load after the anti-dependent store in B2.
 720 // Therefore, we now need to add an anti-dependence edge between the load and
 721 // the B2 store, ensuring that the load is scheduled before the store. Finally,
 722 // the store in B7 is on a distinct control-flow path. Therefore, B7 requires
 723 // no action.
 724 //
 725 // The raise_above_anti_dependences method returns the updated LCA and ensures
 726 // there are no anti-dependent stores in any block between the load's early
 727 // block and the updated LCA. Any stores in the updated LCA will have new
 728 // anti-dependence edges back to the load. The caller may schedule the load in
 729 // the updated LCA, or it may hoist the load above the updated LCA, if the
 730 // updated LCA is not the early block.
 731 Block* PhaseCFG::raise_above_anti_dependences(Block* LCA, Node* load, const bool verify) {
 732   ResourceMark rm;
 733   assert(load->needs_anti_dependence_check(), "must be a load of some sort");
 734   assert(LCA != nullptr, "");
 735   DEBUG_ONLY(Block* LCA_orig = LCA);
 736 
 737   // Compute the alias index.  Loads and stores with different alias indices
 738   // do not need anti-dependence edges.
 739   int load_alias_idx = C->get_alias_index(load->adr_type());
 740 #ifdef ASSERT
 741   assert(Compile::AliasIdxTop <= load_alias_idx && load_alias_idx < C->num_alias_types(), "Invalid alias index");
 742   if (load_alias_idx == Compile::AliasIdxBot && C->do_aliasing() &&
 743       (PrintOpto || VerifyAliases ||
 744        (PrintMiscellaneous && (WizardMode || Verbose)))) {
 745     // Load nodes should not consume all of memory.
 746     // Reporting a bottom type indicates a bug in adlc.
 747     // If some particular type of node validly consumes all of memory,
 748     // sharpen the preceding "if" to exclude it, so we can catch bugs here.
 749     tty->print_cr("*** Possible Anti-Dependence Bug:  Load consumes all of memory.");
 750     load->dump(2);
 751     if (VerifyAliases)  assert(load_alias_idx != Compile::AliasIdxBot, "");
 752   }
 753 #endif
 754 
 755   if (!C->alias_type(load_alias_idx)->is_rewritable()) {
 756     // It is impossible to spoil this load by putting stores before it,
 757     // because we know that the stores will never update the value
 758     // which 'load' must witness.
 759     return LCA;
 760   }
 761 
 762   node_idx_t load_index = load->_idx;
 763 
 764   // Record the earliest legal placement of 'load', as determined by the unique
 765   // point in the dominator tree where all memory effects and other inputs are
 766   // first available (computed by schedule_early). For normal loads, 'early' is
 767   // the shallowest place (dominator-tree wise) to look for anti-dependences
 768   // between this load and any store.
 769   Block* early = get_block_for_node(load);
 770 
 771   // If we are subsuming loads, compute an "early" block that only considers
 772   // memory or address inputs. This block may be different from the
 773   // schedule_early block when it is at an even shallower depth in the
 774   // dominator tree, and allow for a broader discovery of anti-dependences.
 775   if (C->subsume_loads()) {
 776     early = memory_early_block(load, early, this);
 777     if (C->failing()) {
 778       return nullptr;
 779     }
 780   }
 781 
 782   assert(early->dominates(LCA_orig), "precondition failed");
 783 
 784   ResourceArea* area = Thread::current()->resource_area();
 785 
 786   // Bookkeeping of possibly anti-dependent stores that we find outside the
 787   // early block and that may need anti-dependence edges. Note that stores in
 788   // non_early_stores are not necessarily dominated by early. The search starts
 789   // from initial_mem, which can reside in a block that dominates early, and
 790   // therefore, stores we find may be in blocks that are on completely distinct
 791   // control-flow paths compared to early. However, in the end, only stores in
 792   // blocks dominated by early matter. The reason for bookkeeping not only
 793   // relevant stores is efficiency: we lazily record all possible
 794   // anti-dependent stores and add anti-dependence edges only to the relevant
 795   // ones at the very end of this method when we know the final updated LCA.
 796   Node_List non_early_stores(area);
 797 
 798   // Whether we must raise the LCA after the main worklist loop below.
 799   bool must_raise_LCA_above_marks = false;
 800 
 801   // The input load uses some memory state (initial_mem).
 802   Node* initial_mem = load->in(MemNode::Memory);
 803   // To find anti-dependences we must look for users of the same memory state.
 804   // To do this, we search the memory graph downwards from initial_mem. During
 805   // this search, we encounter different types of nodes that we handle
 806   // according to the following three categories:
 807   //
 808   // - MergeMems
 809   // - Memory-state-modifying nodes (informally referred to as "stores" above
 810   //   and below)
 811   // - Memory Phis
 812   //
 813   // MergeMems do not modify the memory state. Anti-dependent stores or memory
 814   // Phis may, however, exist downstream of MergeMems. Therefore, we must
 815   // permit the search to continue through MergeMems. Stores may raise the LCA
 816   // and may potentially also require an anti-dependence edge. Memory Phis may
 817   // raise the LCA but never require anti-dependence edges. See the comments
 818   // throughout the worklist loop below for further details.
 819   //
 820   // It may be useful to think of the anti-dependence search as traversing a
 821   // tree rooted at initial_mem, with internal nodes of type MergeMem and
 822   // memory Phis and stores as (potentially repeated) leaves.
 823 
 824   // We don't optimize the memory graph for pinned loads, so we may need to raise the
 825   // root of our search tree through the corresponding slices of MergeMem nodes to
 826   // get to the node that really creates the memory state for this slice.
 827   if (load_alias_idx >= Compile::AliasIdxRaw) {
 828     while (initial_mem->is_MergeMem()) {
 829       MergeMemNode* mm = initial_mem->as_MergeMem();
 830       Node* p = mm->memory_at(load_alias_idx);
 831       if (p != mm->base_memory()) {
 832         initial_mem = p;
 833       } else {
 834         break;
 835       }
 836     }
 837   }
 838   // To administer the search, we use a worklist consisting of (def,use)-pairs
 839   // of memory states, corresponding to edges in the search tree (and edges
 840   // in the memory graph). We need to keep track of search tree edges in the
 841   // worklist rather than individual nodes due to memory Phis (see details
 842   // below).
 843   DefUseMemStatesQueue worklist(area);
 844   // We start the search at initial_mem and indicate the search root with the
 845   // edge (nullptr, initial_mem).
 846   worklist.push(nullptr, initial_mem);
 847 
 848   // The worklist loop
 849   while (worklist.is_nonempty()) {
 850     // Pop the next edge from the worklist
 851     Node* def_mem_state = worklist.top_def();
 852     Node* use_mem_state = worklist.top_use();
 853     worklist.pop();
 854 
 855     // We are either
 856     // - at the root of the search with the edge (nullptr, initial_mem),
 857     // - just past initial_mem with the edge (initial_mem, use_mem_state), or
 858     // - just past a MergeMem with the edge (MergeMem, use_mem_state).
 859     assert(def_mem_state == nullptr || def_mem_state == initial_mem ||
 860            def_mem_state->is_MergeMem(),
 861            "unexpected memory state");
 862 
 863     const uint op = use_mem_state->Opcode();
 864 
 865 #ifdef ASSERT
 866     // CacheWB nodes are peculiar in a sense that they both are anti-dependent and produce memory.
 867     // Allow them to be treated as a store.
 868     bool is_cache_wb = false;
 869     if (use_mem_state->is_Mach()) {
 870       int ideal_op = use_mem_state->as_Mach()->ideal_Opcode();
 871       is_cache_wb = (ideal_op == Op_CacheWB);
 872     }
 873     assert(!use_mem_state->needs_anti_dependence_check() || is_cache_wb, "no loads");
 874 #endif
 875 
 876     // If we are either at the search root or have found a MergeMem, we step
 877     // past use_mem_state and populate the search worklist with edges
 878     // (use_mem_state, child) for use_mem_state's children.
 879     if (def_mem_state == nullptr // root (exclusive) of tree we are searching
 880         || op == Op_MergeMem     // internal node of tree we are searching
 881     ) {
 882       def_mem_state = use_mem_state;
 883 
 884       for (DUIterator_Fast imax, i = def_mem_state->fast_outs(imax); i < imax; i++) {
 885         use_mem_state = def_mem_state->fast_out(i);
 886         if (use_mem_state->needs_anti_dependence_check()) {
 887           // use_mem_state is also a kind of load (i.e.,
 888           // needs_anti_dependence_check), and it is not a store nor a memory
 889           // Phi. Hence, it is not anti-dependent on the load.
 890           continue;
 891         }
 892         worklist.push(def_mem_state, use_mem_state);
 893       }
 894       // Nothing more to do for the current (nullptr, initial_mem) or
 895       // (initial_mem/MergeMem, MergeMem) edge, move on.
 896       continue;
 897     }
 898 
 899     assert(!use_mem_state->is_MergeMem(),
 900            "use_mem_state should be either a store or a memory Phi");
 901 
 902     if (op == Op_MachProj || op == Op_Catch)   continue;
 903 
 904     // Compute the alias index. If the use_mem_state has an alias index
 905     // different from the load's, it is not anti-dependent. Wide MemBar's
 906     // are anti-dependent with everything (except immutable memories).
 907     const TypePtr* adr_type = use_mem_state->adr_type();
 908     if (!C->can_alias(adr_type, load_alias_idx))  continue;
 909 
 910     // Most slow-path runtime calls do NOT modify Java memory, but
 911     // they can block and so write Raw memory.
 912     if (use_mem_state->is_Mach()) {
 913       MachNode* muse = use_mem_state->as_Mach();
 914       if (load_alias_idx != Compile::AliasIdxRaw) {
 915         // Check for call into the runtime using the Java calling
 916         // convention (and from there into a wrapper); it has no
 917         // _method.  Can't do this optimization for Native calls because
 918         // they CAN write to Java memory.
 919         if (muse->ideal_Opcode() == Op_CallStaticJava) {
 920           assert(muse->is_MachSafePoint(), "");
 921           MachSafePointNode* ms = (MachSafePointNode*)muse;
 922           assert(ms->is_MachCallJava(), "");
 923           MachCallJavaNode* mcj = (MachCallJavaNode*) ms;
 924           if (mcj->_method == nullptr) {
 925             // These runtime calls do not write to Java visible memory
 926             // (other than Raw) and so are not anti-dependent.
 927             continue;
 928           }
 929         }
 930         // Same for SafePoints: they read/write Raw but only read otherwise.
 931         // This is basically a workaround for SafePoints only defining control
 932         // instead of control + memory.
 933         if (muse->ideal_Opcode() == Op_SafePoint) {
 934           continue;
 935         }
 936       } else {
 937         // Some raw memory, such as the load of "top" at an allocation,
 938         // can be control dependent on the previous safepoint. See
 939         // comments in GraphKit::allocate_heap() about control input.
 940         // Inserting an anti-dependence edge between such a safepoint and a use
 941         // creates a cycle, and will cause a subsequent failure in
 942         // local scheduling.  (BugId 4919904)
 943         // (%%% How can a control input be a safepoint and not a projection??)
 944         if (muse->ideal_Opcode() == Op_SafePoint && load->in(0) == muse) {
 945           continue;
 946         }
 947       }
 948     }
 949 
 950     // Determine the block of the use_mem_state.
 951     Block* use_mem_state_block = get_block_for_node(use_mem_state);
 952     assert(use_mem_state_block != nullptr,
 953            "unused killing projections skipped above");
 954 
 955     // For efficiency, we take a lazy approach to both raising the LCA and
 956     // adding anti-dependence edges. In this worklist loop, we only mark blocks
 957     // which we must raise the LCA above (set_raise_LCA_mark), and keep
 958     // track of nodes that potentially need anti-dependence edges
 959     // (non_early_stores). The only exceptions to this are if we
 960     // immediately see that we have to raise the LCA all the way to the early
 961     // block, and if we find stores in the early block (which always need
 962     // anti-dependence edges).
 963     //
 964     // After the worklist loop, we perform an efficient combined LCA-raising
 965     // operation over all marks and only then add anti-dependence edges where
 966     // strictly necessary according to the new raised LCA.
 967 
 968     if (use_mem_state->is_Phi()) {
 969       // We have reached a memory Phi node. On our search from initial_mem to
 970       // the Phi, we have found no anti-dependences (otherwise, we would have
 971       // already terminated the search along this branch). Consider the example
 972       // below, indicating a Phi node and its node inputs (we omit the control
 973       // input).
 974       //
 975       //    def_mem_state
 976       //          |
 977       //          | ? ?
 978       //          \ | /
 979       //           Phi
 980       //
 981       // We reached the Phi from def_mem_state and know that, on this
 982       // particular input, the memory that the load must witness is not
 983       // overwritten. However, for the Phi's other inputs (? in the
 984       // illustration), we have no information and must thus conservatively
 985       // assume that the load's memory is overwritten at and below the Phi.
 986       //
 987       // It is impossible to schedule the load before the Phi in
 988       // the same block as the Phi (use_mem_state_block), and anti-dependence
 989       // edges are, therefore, redundant. We must, however, find the
 990       // predecessor block of use_mem_state_block that corresponds to
 991       // def_mem_state, and raise the LCA above that block. Note that this block
 992       // is not necessarily def_mem_state's block! See the continuation of our
 993       // previous example below (now illustrating blocks instead of nodes)
 994       //
 995       //    def_mem_state's block
 996       //          |
 997       //          |
 998       //      pred_block
 999       //          |
1000       //          |   ?   ?
1001       //          |   |   |
1002       //      use_mem_state_block
1003       //
1004       // Here, we must raise the LCA above pred_block rather than
1005       // def_mem_state's block.
1006       //
1007       // Do not assert(use_mem_state_block != early, "Phi merging memory after access")
1008       // PhiNode may be at start of block 'early' with backedge to 'early'
1009       if (LCA == early) {
1010         // Don't bother if LCA is already raised all the way
1011         continue;
1012       }
1013       DEBUG_ONLY(bool found_match = false);
1014       for (uint j = PhiNode::Input, jmax = use_mem_state->req(); j < jmax; j++) {
1015         if (use_mem_state->in(j) == def_mem_state) {   // Found matching input?
1016           DEBUG_ONLY(found_match = true);
1017           Block* pred_block = get_block_for_node(use_mem_state_block->pred(j));
1018           if (pred_block != early) {
1019             // Lazily set the LCA mark
1020             pred_block->set_raise_LCA_mark(load_index);
1021             must_raise_LCA_above_marks = true;
1022           } else /* if (pred_block == early) */ {
1023             // We know already now that we must raise LCA all the way to early.
1024             LCA = early;
1025             // This turns off the process of gathering non_early_stores.
1026           }
1027         }
1028       }
1029       assert(found_match, "no worklist bug");
1030     } else if (use_mem_state_block != early) {
1031       // We found an anti-dependent store outside the load's 'early' block. The
1032       // store may be between the current LCA and the earliest possible block
1033       // (but it could very well also be on a distinct control-flow path).
1034       // Lazily set the LCA mark and push to non_early_stores.
1035       if (LCA == early) {
1036         // Don't bother if LCA is already raised all the way
1037         continue;
1038       }
1039       if (unrelated_load_in_store_null_block(use_mem_state, load)) {
1040         continue;
1041       }
1042       use_mem_state_block->set_raise_LCA_mark(load_index);
1043       must_raise_LCA_above_marks = true;
1044       non_early_stores.push(use_mem_state);
1045     } else /* if (use_mem_state_block == early) */ {
1046       // We found an anti-dependent store in the load's 'early' block.
1047       // Therefore, we know already now that we must raise LCA all the way to
1048       // early and that we need to add an anti-dependence edge to the store.
1049       assert(use_mem_state != load->find_exact_control(load->in(0)), "dependence cycle found");
1050       if (verify) {
1051         assert(use_mem_state->find_edge(load) != -1 || unrelated_load_in_store_null_block(use_mem_state, load),
1052                "missing precedence edge");
1053       } else {
1054         use_mem_state->add_prec(load);
1055       }
1056       LCA = early;
1057       // This turns off the process of gathering non_early_stores.
1058     }
1059   }
1060   // Worklist is now empty; we have visited all possible anti-dependences.
1061 
1062   // Finished if 'load' must be scheduled in its 'early' block.
1063   // If we found any stores there, they have already been given
1064   // anti-dependence edges.
1065   if (LCA == early) {
1066     return LCA;
1067   }
1068 
1069   // We get here only if there are no anti-dependent stores in the load's
1070   // 'early' block and if no memory Phi has forced LCA to the early block. Now
1071   // we must raise the LCA above the blocks for all the anti-dependent stores
1072   // and above the predecessor blocks of anti-dependent memory Phis we reached
1073   // during the search.
1074   if (must_raise_LCA_above_marks) {
1075     LCA = raise_LCA_above_marks(LCA, load->_idx, early, this);
1076   }
1077 
1078   // If LCA == early at this point, there were no stores that required
1079   // anti-dependence edges in the early block. Otherwise, we would have eagerly
1080   // raised the LCA to early already in the worklist loop.
1081   if (LCA == early) {
1082     return LCA;
1083   }
1084 
1085   // The raised LCA block can now be a home to anti-dependent stores for which
1086   // we still need to add anti-dependence edges, but no LCA predecessor block
1087   // contains any such stores (otherwise, we would have raised the LCA even
1088   // higher).
1089   //
1090   // The raised LCA will be a lower bound for placing the load, preventing the
1091   // load from sinking past any block containing a store that may overwrite
1092   // memory that the load must witness.
1093   //
1094   // Now we need to insert the necessary anti-dependence edges from 'load' to
1095   // each store in the non-early LCA block. We have recorded all such potential
1096   // stores in non_early_stores.
1097   //
1098   // If LCA->raise_LCA_mark() != load_index, it means that we raised the LCA to
1099   // a block in which we did not find any anti-dependent stores. So, no need to
1100   // search for any such stores.
1101   if (LCA->raise_LCA_mark() == load_index) {
1102     while (non_early_stores.size() > 0) {
1103       Node* store = non_early_stores.pop();
1104       Block* store_block = get_block_for_node(store);
1105       if (store_block == LCA) {
1106         // Add anti-dependence edge from the load to the store in the non-early
1107         // LCA.
1108         assert(store != load->find_exact_control(load->in(0)), "dependence cycle found");
1109         if (verify) {
1110           assert(store->find_edge(load) != -1, "missing precedence edge");
1111         } else {
1112           store->add_prec(load);
1113         }
1114       } else {
1115         assert(store_block->raise_LCA_mark() == load_index, "block was marked");
1116       }
1117     }
1118   }
1119 
1120   assert(LCA->dominates(LCA_orig), "unsound updated LCA");
1121 
1122   // Return the highest block containing stores; any stores
1123   // within that block have been given anti-dependence edges.
1124   return LCA;
1125 }
1126 
1127 // This class is used to iterate backwards over the nodes in the graph.
1128 
1129 class Node_Backward_Iterator {
1130 
1131 private:
1132   Node_Backward_Iterator();
1133 
1134 public:
1135   // Constructor for the iterator
1136   Node_Backward_Iterator(Node *root, VectorSet &visited, Node_Stack &stack, PhaseCFG &cfg);
1137 
1138   // Postincrement operator to iterate over the nodes
1139   Node *next();
1140 
1141 private:
1142   VectorSet   &_visited;
1143   Node_Stack  &_stack;
1144   PhaseCFG &_cfg;
1145 };
1146 
1147 // Constructor for the Node_Backward_Iterator
1148 Node_Backward_Iterator::Node_Backward_Iterator( Node *root, VectorSet &visited, Node_Stack &stack, PhaseCFG &cfg)
1149   : _visited(visited), _stack(stack), _cfg(cfg) {
1150   // The stack should contain exactly the root
1151   stack.clear();
1152   stack.push(root, root->outcnt());
1153 
1154   // Clear the visited bits
1155   visited.clear();
1156 }
1157 
1158 // Iterator for the Node_Backward_Iterator
1159 Node *Node_Backward_Iterator::next() {
1160 
1161   // If the _stack is empty, then just return null: finished.
1162   if ( !_stack.size() )
1163     return nullptr;
1164 
1165   // I visit unvisited not-anti-dependence users first, then anti-dependent
1166   // children next. I iterate backwards to support removal of nodes.
1167   // The stack holds states consisting of 3 values:
1168   // current Def node, flag which indicates 1st/2nd pass, index of current out edge
1169   Node *self = (Node*)(((uintptr_t)_stack.node()) & ~1);
1170   bool iterate_anti_dep = (((uintptr_t)_stack.node()) & 1);
1171   uint idx = MIN2(_stack.index(), self->outcnt()); // Support removal of nodes.
1172   _stack.pop();
1173 
1174   // I cycle here when I am entering a deeper level of recursion.
1175   // The key variable 'self' was set prior to jumping here.
1176   while( 1 ) {
1177 
1178     _visited.set(self->_idx);
1179 
1180     // Now schedule all uses as late as possible.
1181     const Node* src = self->is_Proj() ? self->in(0) : self;
1182     uint src_rpo = _cfg.get_block_for_node(src)->_rpo;
1183 
1184     // Schedule all nodes in a post-order visit
1185     Node *unvisited = nullptr;  // Unvisited anti-dependent Node, if any
1186 
1187     // Scan for unvisited nodes
1188     while (idx > 0) {
1189       // For all uses, schedule late
1190       Node* n = self->raw_out(--idx); // Use
1191 
1192       // Skip already visited children
1193       if ( _visited.test(n->_idx) )
1194         continue;
1195 
1196       // do not traverse backward control edges
1197       Node *use = n->is_Proj() ? n->in(0) : n;
1198       uint use_rpo = _cfg.get_block_for_node(use)->_rpo;
1199 
1200       if ( use_rpo < src_rpo )
1201         continue;
1202 
1203       // Phi nodes always precede uses in a basic block
1204       if ( use_rpo == src_rpo && use->is_Phi() )
1205         continue;
1206 
1207       unvisited = n;      // Found unvisited
1208 
1209       // Check for possible-anti-dependent
1210       // 1st pass: No such nodes, 2nd pass: Only such nodes.
1211       if (n->needs_anti_dependence_check() == iterate_anti_dep) {
1212         unvisited = n;      // Found unvisited
1213         break;
1214       }
1215     }
1216 
1217     // Did I find an unvisited not-anti-dependent Node?
1218     if (!unvisited) {
1219       if (!iterate_anti_dep) {
1220         // 2nd pass: Iterate over nodes which needs_anti_dependence_check.
1221         iterate_anti_dep = true;
1222         idx = self->outcnt();
1223         continue;
1224       }
1225       break;                  // All done with children; post-visit 'self'
1226     }
1227 
1228     // Visit the unvisited Node.  Contains the obvious push to
1229     // indicate I'm entering a deeper level of recursion.  I push the
1230     // old state onto the _stack and set a new state and loop (recurse).
1231     _stack.push((Node*)((uintptr_t)self | (uintptr_t)iterate_anti_dep), idx);
1232     self = unvisited;
1233     iterate_anti_dep = false;
1234     idx = self->outcnt();
1235   } // End recursion loop
1236 
1237   return self;
1238 }
1239 
1240 //------------------------------ComputeLatenciesBackwards----------------------
1241 // Compute the latency of all the instructions.
1242 void PhaseCFG::compute_latencies_backwards(VectorSet &visited, Node_Stack &stack) {
1243 #ifndef PRODUCT
1244   if (trace_opto_pipelining())
1245     tty->print("\n#---- ComputeLatenciesBackwards ----\n");
1246 #endif
1247 
1248   Node_Backward_Iterator iter((Node *)_root, visited, stack, *this);
1249   Node *n;
1250 
1251   // Walk over all the nodes from last to first
1252   while ((n = iter.next())) {
1253     // Set the latency for the definitions of this instruction
1254     partial_latency_of_defs(n);
1255   }
1256 } // end ComputeLatenciesBackwards
1257 
1258 //------------------------------partial_latency_of_defs------------------------
1259 // Compute the latency impact of this node on all defs.  This computes
1260 // a number that increases as we approach the beginning of the routine.
1261 void PhaseCFG::partial_latency_of_defs(Node *n) {
1262   // Set the latency for this instruction
1263 #ifndef PRODUCT
1264   if (trace_opto_pipelining()) {
1265     tty->print("# latency_to_inputs: node_latency[%d] = %d for node", n->_idx, get_latency_for_node(n));
1266     dump();
1267   }
1268 #endif
1269 
1270   if (n->is_Proj()) {
1271     n = n->in(0);
1272   }
1273 
1274   if (n->is_Root()) {
1275     return;
1276   }
1277 
1278   uint nlen = n->len();
1279   uint use_latency = get_latency_for_node(n);
1280   uint use_pre_order = get_block_for_node(n)->_pre_order;
1281 
1282   for (uint j = 0; j < nlen; j++) {
1283     Node *def = n->in(j);
1284 
1285     if (!def || def == n) {
1286       continue;
1287     }
1288 
1289     // Walk backwards thru projections
1290     if (def->is_Proj()) {
1291       def = def->in(0);
1292     }
1293 
1294 #ifndef PRODUCT
1295     if (trace_opto_pipelining()) {
1296       tty->print("#    in(%2d): ", j);
1297       def->dump();
1298     }
1299 #endif
1300 
1301     // If the defining block is not known, assume it is ok
1302     Block *def_block = get_block_for_node(def);
1303     uint def_pre_order = def_block ? def_block->_pre_order : 0;
1304 
1305     if ((use_pre_order <  def_pre_order) || (use_pre_order == def_pre_order && n->is_Phi())) {
1306       continue;
1307     }
1308 
1309     uint delta_latency = n->latency(j);
1310     uint current_latency = delta_latency + use_latency;
1311 
1312     if (get_latency_for_node(def) < current_latency) {
1313       set_latency_for_node(def, current_latency);
1314     }
1315 
1316 #ifndef PRODUCT
1317     if (trace_opto_pipelining()) {
1318       tty->print_cr("#      %d + edge_latency(%d) == %d -> %d, node_latency[%d] = %d", use_latency, j, delta_latency, current_latency, def->_idx, get_latency_for_node(def));
1319     }
1320 #endif
1321   }
1322 }
1323 
1324 //------------------------------latency_from_use-------------------------------
1325 // Compute the latency of a specific use
1326 int PhaseCFG::latency_from_use(Node *n, const Node *def, Node *use) {
1327   // If self-reference, return no latency
1328   if (use == n || use->is_Root()) {
1329     return 0;
1330   }
1331 
1332   uint def_pre_order = get_block_for_node(def)->_pre_order;
1333   uint latency = 0;
1334 
1335   // If the use is not a projection, then it is simple...
1336   if (!use->is_Proj()) {
1337 #ifndef PRODUCT
1338     if (trace_opto_pipelining()) {
1339       tty->print("#    out(): ");
1340       use->dump();
1341     }
1342 #endif
1343 
1344     uint use_pre_order = get_block_for_node(use)->_pre_order;
1345 
1346     if (use_pre_order < def_pre_order)
1347       return 0;
1348 
1349     if (use_pre_order == def_pre_order && use->is_Phi())
1350       return 0;
1351 
1352     uint nlen = use->len();
1353     uint nl = get_latency_for_node(use);
1354 
1355     for ( uint j=0; j<nlen; j++ ) {
1356       if (use->in(j) == n) {
1357         // Change this if we want local latencies
1358         uint ul = use->latency(j);
1359         uint  l = ul + nl;
1360         if (latency < l) latency = l;
1361 #ifndef PRODUCT
1362         if (trace_opto_pipelining()) {
1363           tty->print_cr("#      %d + edge_latency(%d) == %d -> %d, latency = %d",
1364                         nl, j, ul, l, latency);
1365         }
1366 #endif
1367       }
1368     }
1369   } else {
1370     // This is a projection, just grab the latency of the use(s)
1371     for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) {
1372       uint l = latency_from_use(use, def, use->fast_out(j));
1373       if (latency < l) latency = l;
1374     }
1375   }
1376 
1377   return latency;
1378 }
1379 
1380 //------------------------------latency_from_uses------------------------------
1381 // Compute the latency of this instruction relative to all of it's uses.
1382 // This computes a number that increases as we approach the beginning of the
1383 // routine.
1384 void PhaseCFG::latency_from_uses(Node *n) {
1385   // Set the latency for this instruction
1386 #ifndef PRODUCT
1387   if (trace_opto_pipelining()) {
1388     tty->print("# latency_from_outputs: node_latency[%d] = %d for node", n->_idx, get_latency_for_node(n));
1389     dump();
1390   }
1391 #endif
1392   uint latency=0;
1393   const Node *def = n->is_Proj() ? n->in(0): n;
1394 
1395   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1396     uint l = latency_from_use(n, def, n->fast_out(i));
1397 
1398     if (latency < l) latency = l;
1399   }
1400 
1401   set_latency_for_node(n, latency);
1402 }
1403 
1404 //------------------------------is_cheaper_block-------------------------
1405 // Check if a block between early and LCA block of uses is cheaper by
1406 // frequency-based policy, latency-based policy and random-based policy
1407 bool PhaseCFG::is_cheaper_block(Block* LCA, Node* self, uint target_latency,
1408                                 uint end_latency, double least_freq,
1409                                 int cand_cnt, bool in_latency) {
1410   if (StressGCM) {
1411     // Should be randomly accepted in stress mode
1412     return C->randomized_select(cand_cnt);
1413   }
1414 
1415   const double delta = 1 + PROB_UNLIKELY_MAG(4);
1416 
1417   // Better Frequency. Add a small delta to the comparison to not needlessly
1418   // hoist because of, e.g., small numerical inaccuracies.
1419   if (LCA->_freq * delta < least_freq) {
1420     return true;
1421   }
1422 
1423   // Otherwise, choose with latency
1424   if (!in_latency                     &&  // No block containing latency
1425       LCA->_freq < least_freq * delta &&  // No worse frequency
1426       target_latency >= end_latency   &&  // within latency range
1427       !self->is_iteratively_computed()    // But don't hoist IV increments
1428             // because they may end up above other uses of their phi forcing
1429             // their result register to be different from their input.
1430   ) {
1431     return true;
1432   }
1433 
1434   return false;
1435 }
1436 
1437 //------------------------------hoist_to_cheaper_block-------------------------
1438 // Pick a block for node self, between early and LCA block of uses, that is a
1439 // cheaper alternative to LCA.
1440 Block* PhaseCFG::hoist_to_cheaper_block(Block* LCA, Block* early, Node* self) {
1441   Block* least       = LCA;
1442   double least_freq  = least->_freq;
1443   uint target        = get_latency_for_node(self);
1444   uint start_latency = get_latency_for_node(LCA->head());
1445   uint end_latency   = get_latency_for_node(LCA->get_node(LCA->end_idx()));
1446   bool in_latency    = (target <= start_latency);
1447   const Block* root_block = get_block_for_node(_root);
1448 
1449   // Turn off latency scheduling if scheduling is just plain off
1450   if (!C->do_scheduling())
1451     in_latency = true;
1452 
1453   // Do not hoist (to cover latency) instructions which target a
1454   // single register.  Hoisting stretches the live range of the
1455   // single register and may force spilling.
1456   MachNode* mach = self->is_Mach() ? self->as_Mach() : nullptr;
1457   if (mach != nullptr && mach->out_RegMask().is_bound1() && !mach->out_RegMask().is_empty()) {
1458     in_latency = true;
1459   }
1460 
1461 #ifndef PRODUCT
1462   if (trace_opto_pipelining()) {
1463     tty->print("# Find cheaper block for latency %d: ", get_latency_for_node(self));
1464     self->dump();
1465     tty->print_cr("#   B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1466       LCA->_pre_order,
1467       LCA->head()->_idx,
1468       start_latency,
1469       LCA->get_node(LCA->end_idx())->_idx,
1470       end_latency,
1471       least_freq);
1472   }
1473 #endif
1474 
1475   int cand_cnt = 0;  // number of candidates tried
1476 
1477   // Walk up the dominator tree from LCA (Lowest common ancestor) to
1478   // the earliest legal location. Capture the least execution frequency,
1479   // or choose a random block if -XX:+StressGCM, or using latency-based policy
1480   while (LCA != early) {
1481     LCA = LCA->_idom;         // Follow up the dominator tree
1482 
1483     if (LCA == nullptr) {
1484       // Bailout without retry
1485       assert(false, "graph should be schedulable");
1486       C->record_method_not_compilable("late schedule failed: LCA is null");
1487       return least;
1488     }
1489 
1490     // Don't hoist machine instructions to the root basic block
1491     if (mach != nullptr && LCA == root_block)
1492       break;
1493 
1494     if (self->is_memory_writer() &&
1495         (LCA->_loop->depth() > early->_loop->depth())) {
1496       // LCA is an invalid placement for a memory writer: choosing it would
1497       // cause memory interference, as illustrated in schedule_late().
1498       continue;
1499     }
1500     verify_memory_writer_placement(LCA, self);
1501 
1502     uint start_lat = get_latency_for_node(LCA->head());
1503     uint end_idx   = LCA->end_idx();
1504     uint end_lat   = get_latency_for_node(LCA->get_node(end_idx));
1505     double LCA_freq = LCA->_freq;
1506 #ifndef PRODUCT
1507     if (trace_opto_pipelining()) {
1508       tty->print_cr("#   B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1509         LCA->_pre_order, LCA->head()->_idx, start_lat, end_idx, end_lat, LCA_freq);
1510     }
1511 #endif
1512     cand_cnt++;
1513     if (is_cheaper_block(LCA, self, target, end_lat, least_freq, cand_cnt, in_latency)) {
1514       least = LCA;            // Found cheaper block
1515       least_freq = LCA_freq;
1516       start_latency = start_lat;
1517       end_latency = end_lat;
1518       if (target <= start_lat)
1519         in_latency = true;
1520     }
1521   }
1522 
1523 #ifndef PRODUCT
1524   if (trace_opto_pipelining()) {
1525     tty->print_cr("#  Choose block B%d with start latency=%d and freq=%g",
1526       least->_pre_order, start_latency, least_freq);
1527   }
1528 #endif
1529 
1530   // See if the latency needs to be updated
1531   if (target < end_latency) {
1532 #ifndef PRODUCT
1533     if (trace_opto_pipelining()) {
1534       tty->print_cr("#  Change latency for [%4d] from %d to %d", self->_idx, target, end_latency);
1535     }
1536 #endif
1537     set_latency_for_node(self, end_latency);
1538     partial_latency_of_defs(self);
1539   }
1540 
1541   return least;
1542 }
1543 
1544 
1545 //------------------------------schedule_late-----------------------------------
1546 // Now schedule all codes as LATE as possible.  This is the LCA in the
1547 // dominator tree of all USES of a value.  Pick the block with the least
1548 // loop nesting depth that is lowest in the dominator tree.
1549 extern const char must_clone[];
1550 void PhaseCFG::schedule_late(VectorSet &visited, Node_Stack &stack) {
1551 #ifndef PRODUCT
1552   if (trace_opto_pipelining())
1553     tty->print("\n#---- schedule_late ----\n");
1554 #endif
1555 
1556   Node_Backward_Iterator iter((Node *)_root, visited, stack, *this);
1557   Node *self;
1558 
1559   // Walk over all the nodes from last to first
1560   while ((self = iter.next())) {
1561     Block* early = get_block_for_node(self); // Earliest legal placement
1562 
1563     if (self->is_top()) {
1564       // Top node goes in bb #2 with other constants.
1565       // It must be special-cased, because it has no out edges.
1566       early->add_inst(self);
1567       continue;
1568     }
1569 
1570     // No uses, just terminate
1571     if (self->outcnt() == 0) {
1572       assert(self->is_MachProj(), "sanity");
1573       continue;                   // Must be a dead machine projection
1574     }
1575 
1576     // If node is pinned in the block, then no scheduling can be done.
1577     if( self->pinned() )          // Pinned in block?
1578       continue;
1579 
1580 #ifdef ASSERT
1581     // Assert that memory writers (e.g. stores) have a "home" block (the block
1582     // given by their control input), and that this block corresponds to their
1583     // earliest possible placement. This guarantees that
1584     // hoist_to_cheaper_block() will always have at least one valid choice.
1585     if (self->is_memory_writer()) {
1586       assert(find_block_for_node(self->in(0)) == early,
1587              "The home of a memory writer must also be its earliest placement");
1588     }
1589 #endif
1590 
1591     MachNode* mach = self->is_Mach() ? self->as_Mach() : nullptr;
1592     if (mach) {
1593       switch (mach->ideal_Opcode()) {
1594       case Op_CreateEx:
1595         // Don't move exception creation
1596         early->add_inst(self);
1597         continue;
1598         break;
1599       case Op_CastI2N:
1600         early->add_inst(self);
1601         continue;
1602       case Op_CheckCastPP: {
1603         // Don't move CheckCastPP nodes away from their input, if the input
1604         // is a rawptr (5071820).
1605         Node *def = self->in(1);
1606         if (def != nullptr && def->bottom_type()->base() == Type::RawPtr) {
1607           early->add_inst(self);
1608 #ifdef ASSERT
1609           _raw_oops.push(def);
1610 #endif
1611           continue;
1612         }
1613         break;
1614       }
1615       default:
1616         break;
1617       }
1618       if (C->has_irreducible_loop() && self->is_memory_writer()) {
1619         // If the CFG is irreducible, place memory writers in their home block.
1620         // This prevents hoist_to_cheaper_block() from accidentally placing such
1621         // nodes into deeper loops, as in the following example:
1622         //
1623         // Home placement of store in B1 (loop L1):
1624         //
1625         // B1 (L1):
1626         //   m1 <- ..
1627         //   m2 <- store m1, ..
1628         // B2 (L2):
1629         //   jump B2
1630         // B3 (L1):
1631         //   .. <- .. m2, ..
1632         //
1633         // Wrong "hoisting" of store to B2 (in loop L2, child of L1):
1634         //
1635         // B1 (L1):
1636         //   m1 <- ..
1637         // B2 (L2):
1638         //   m2 <- store m1, ..
1639         //   # Wrong: m1 and m2 interfere at this point.
1640         //   jump B2
1641         // B3 (L1):
1642         //   .. <- .. m2, ..
1643         //
1644         // This "hoist inversion" can happen due to different factors such as
1645         // inaccurate estimation of frequencies for irreducible CFGs, and loops
1646         // with always-taken exits in reducible CFGs. In the reducible case,
1647         // hoist inversion is prevented by discarding invalid blocks (those in
1648         // deeper loops than the home block). In the irreducible case, the
1649         // invalid blocks cannot be identified due to incomplete loop nesting
1650         // information, hence a conservative solution is taken.
1651 #ifndef PRODUCT
1652         if (trace_opto_pipelining()) {
1653           tty->print_cr("# Irreducible loops: schedule in home block B%d:",
1654                         early->_pre_order);
1655           self->dump();
1656         }
1657 #endif
1658         schedule_node_into_block(self, early);
1659         continue;
1660       }
1661     }
1662 
1663     // Gather LCA of all uses
1664     Block *LCA = nullptr;
1665     {
1666       for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
1667         // For all uses, find LCA
1668         Node* use = self->fast_out(i);
1669         LCA = raise_LCA_above_use(LCA, use, self, this);
1670       }
1671       guarantee(LCA != nullptr, "There must be a LCA");
1672     }  // (Hide defs of imax, i from rest of block.)
1673 
1674     // Place temps in the block of their use.  This isn't a
1675     // requirement for correctness but it reduces useless
1676     // interference between temps and other nodes.
1677     if (mach != nullptr && mach->is_MachTemp()) {
1678       map_node_to_block(self, LCA);
1679       LCA->add_inst(self);
1680       continue;
1681     }
1682 
1683     // Check if 'self' could be anti-dependent on memory
1684     if (self->needs_anti_dependence_check()) {
1685       // Hoist LCA above possible-defs and insert anti-dependences to
1686       // defs in new LCA block.
1687       LCA = raise_above_anti_dependences(LCA, self);
1688       if (C->failing()) {
1689         return;
1690       }
1691     }
1692 
1693     if (early->_dom_depth > LCA->_dom_depth) {
1694       // Somehow the LCA has moved above the earliest legal point.
1695       // (One way this can happen is via memory_early_block.)
1696       if (C->subsume_loads() == true && !C->failing()) {
1697         // Retry with subsume_loads == false
1698         // If this is the first failure, the sentinel string will "stick"
1699         // to the Compile object, and the C2Compiler will see it and retry.
1700         C->record_failure(C2Compiler::retry_no_subsuming_loads());
1701       } else {
1702         // Bailout without retry when (early->_dom_depth > LCA->_dom_depth)
1703         assert(C->failure_is_artificial(), "graph should be schedulable");
1704         C->record_method_not_compilable("late schedule failed: incorrect graph" DEBUG_ONLY(COMMA true));
1705       }
1706       return;
1707     }
1708 
1709     if (self->is_memory_writer()) {
1710       // If the LCA of a memory writer is a descendant of its home loop, hoist
1711       // it into a valid placement.
1712       while (LCA->_loop->depth() > early->_loop->depth()) {
1713         LCA = LCA->_idom;
1714       }
1715       assert(LCA != nullptr, "a valid LCA must exist");
1716       verify_memory_writer_placement(LCA, self);
1717     }
1718 
1719     // If there is no opportunity to hoist, then we're done.
1720     // In stress mode, try to hoist even the single operations.
1721     bool try_to_hoist = StressGCM || (LCA != early);
1722 
1723     // Must clone guys stay next to use; no hoisting allowed.
1724     // Also cannot hoist guys that alter memory or are otherwise not
1725     // allocatable (hoisting can make a value live longer, leading to
1726     // anti and output dependency problems which are normally resolved
1727     // by the register allocator giving everyone a different register).
1728     if (mach != nullptr && must_clone[mach->ideal_Opcode()])
1729       try_to_hoist = false;
1730 
1731     Block* late = nullptr;
1732     if (try_to_hoist) {
1733       // Now find the block with the least execution frequency.
1734       // Start at the latest schedule and work up to the earliest schedule
1735       // in the dominator tree.  Thus the Node will dominate all its uses.
1736       late = hoist_to_cheaper_block(LCA, early, self);
1737     } else {
1738       // Just use the LCA of the uses.
1739       late = LCA;
1740     }
1741 
1742     // Put the node into target block
1743     schedule_node_into_block(self, late);
1744 
1745 #ifdef ASSERT
1746     if (self->needs_anti_dependence_check()) {
1747       // since precedence edges are only inserted when we're sure they
1748       // are needed make sure that after placement in a block we don't
1749       // need any new precedence edges.
1750       verify_anti_dependences(late, self);
1751       if (C->failing()) {
1752         return;
1753       }
1754     }
1755 #endif
1756   } // Loop until all nodes have been visited
1757 
1758 } // end ScheduleLate
1759 
1760 //------------------------------GlobalCodeMotion-------------------------------
1761 void PhaseCFG::global_code_motion() {
1762   ResourceMark rm;
1763 
1764 #ifndef PRODUCT
1765   if (trace_opto_pipelining()) {
1766     tty->print("\n---- Start GlobalCodeMotion ----\n");
1767   }
1768 #endif
1769 
1770   // Initialize the node to block mapping for things on the proj_list
1771   for (uint i = 0; i < _matcher.number_of_projections(); i++) {
1772     unmap_node_from_block(_matcher.get_projection(i));
1773   }
1774 
1775   // Set the basic block for Nodes pinned into blocks
1776   VectorSet visited;
1777   schedule_pinned_nodes(visited);
1778 
1779   // Find the earliest Block any instruction can be placed in.  Some
1780   // instructions are pinned into Blocks.  Unpinned instructions can
1781   // appear in last block in which all their inputs occur.
1782   visited.clear();
1783   Node_Stack stack((C->live_nodes() >> 2) + 16); // pre-grow
1784   if (!schedule_early(visited, stack)) {
1785     // Bailout without retry
1786     assert(C->failure_is_artificial(), "early schedule failed");
1787     C->record_method_not_compilable("early schedule failed" DEBUG_ONLY(COMMA true));
1788     return;
1789   }
1790 
1791   // Build Def-Use edges.
1792   // Compute the latency information (via backwards walk) for all the
1793   // instructions in the graph
1794   _node_latency = new GrowableArray<uint>(); // resource_area allocation
1795 
1796   if (C->do_scheduling()) {
1797     compute_latencies_backwards(visited, stack);
1798   }
1799 
1800   // Now schedule all codes as LATE as possible.  This is the LCA in the
1801   // dominator tree of all USES of a value.  Pick the block with the least
1802   // loop nesting depth that is lowest in the dominator tree.
1803   // ( visited.clear() called in schedule_late()->Node_Backward_Iterator() )
1804   schedule_late(visited, stack);
1805   if (C->failing()) {
1806     return;
1807   }
1808 
1809 #ifndef PRODUCT
1810   if (trace_opto_pipelining()) {
1811     tty->print("\n---- Detect implicit null checks ----\n");
1812   }
1813 #endif
1814 
1815   // Detect implicit-null-check opportunities.  Basically, find null checks
1816   // with suitable memory ops nearby.  Use the memory op to do the null check.
1817   // I can generate a memory op if there is not one nearby.
1818   if (C->is_method_compilation()) {
1819     // By reversing the loop direction we get a very minor gain on mpegaudio.
1820     // Feel free to revert to a forward loop for clarity.
1821     // for( int i=0; i < (int)matcher._null_check_tests.size(); i+=2 ) {
1822     for (int i = _matcher._null_check_tests.size() - 2; i >= 0; i -= 2) {
1823       Node* proj = _matcher._null_check_tests[i];
1824       Node* val  = _matcher._null_check_tests[i + 1];
1825       Block* block = get_block_for_node(proj);
1826       implicit_null_check(block, proj, val, C->allowed_deopt_reasons());
1827       // The implicit_null_check will only perform the transformation
1828       // if the null branch is truly uncommon, *and* it leads to an
1829       // uncommon trap.  Combined with the too_many_traps guards
1830       // above, this prevents SEGV storms reported in 6366351,
1831       // by recompiling offending methods without this optimization.
1832       if (C->failing()) {
1833         return;
1834       }
1835     }
1836   }
1837 
1838   bool block_size_threshold_ok = false;
1839   intptr_t *recalc_pressure_nodes = nullptr;
1840   if (OptoRegScheduling) {
1841     for (uint i = 0; i < number_of_blocks(); i++) {
1842       Block* block = get_block(i);
1843       if (block->number_of_nodes() > 10) {
1844         block_size_threshold_ok = true;
1845         break;
1846       }
1847     }
1848   }
1849 
1850   // Enabling the scheduler for register pressure plus finding blocks of size to schedule for it
1851   // is key to enabling this feature.
1852   PhaseChaitin regalloc(C->unique(), *this, _matcher, true);
1853   ResourceArea live_arena(mtCompiler, Arena::Tag::tag_reglive);      // Arena for liveness
1854   ResourceMark rm_live(&live_arena);
1855   PhaseLive live(*this, regalloc._lrg_map.names(), &live_arena, true);
1856   PhaseIFG ifg(&live_arena);
1857   if (OptoRegScheduling && block_size_threshold_ok) {
1858     regalloc.mark_ssa();
1859     Compile::TracePhase tp(_t_computeLive);
1860     rm_live.reset_to_mark();           // Reclaim working storage
1861     IndexSet::reset_memory(C, &live_arena);
1862     uint node_size = regalloc._lrg_map.max_lrg_id();
1863     ifg.init(node_size); // Empty IFG
1864     regalloc.set_ifg(ifg);
1865     regalloc.set_live(live);
1866     regalloc.gather_lrg_masks(false);    // Collect LRG masks
1867     live.compute(node_size); // Compute liveness
1868 
1869     recalc_pressure_nodes = NEW_RESOURCE_ARRAY(intptr_t, node_size);
1870     for (uint i = 0; i < node_size; i++) {
1871       recalc_pressure_nodes[i] = 0;
1872     }
1873   }
1874   _regalloc = &regalloc;
1875 
1876 #ifndef PRODUCT
1877   if (trace_opto_pipelining()) {
1878     tty->print("\n---- Start Local Scheduling ----\n");
1879   }
1880 #endif
1881 
1882   // Schedule locally.  Right now a simple topological sort.
1883   // Later, do a real latency aware scheduler.
1884   GrowableArray<int> ready_cnt(C->unique(), C->unique(), -1);
1885   visited.reset();
1886   for (uint i = 0; i < number_of_blocks(); i++) {
1887     Block* block = get_block(i);
1888     if (!schedule_local(block, ready_cnt, visited, recalc_pressure_nodes)) {
1889       if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
1890         assert(C->failure_is_artificial(), "local schedule failed");
1891         C->record_method_not_compilable("local schedule failed" DEBUG_ONLY(COMMA true));
1892       }
1893       _regalloc = nullptr;
1894       return;
1895     }
1896   }
1897   _regalloc = nullptr;
1898 
1899   // If we inserted any instructions between a Call and his CatchNode,
1900   // clone the instructions on all paths below the Catch.
1901   for (uint i = 0; i < number_of_blocks(); i++) {
1902     Block* block = get_block(i);
1903     call_catch_cleanup(block);
1904     if (C->failing()) {
1905       return;
1906     }
1907   }
1908 
1909 #ifndef PRODUCT
1910   if (trace_opto_pipelining()) {
1911     tty->print("\n---- After GlobalCodeMotion ----\n");
1912     for (uint i = 0; i < number_of_blocks(); i++) {
1913       Block* block = get_block(i);
1914       block->dump();
1915     }
1916   }
1917 #endif
1918   // Dead.
1919   _node_latency = (GrowableArray<uint> *)((intptr_t)0xdeadbeef);
1920 }
1921 
1922 bool PhaseCFG::do_global_code_motion() {
1923 
1924   build_dominator_tree();
1925   if (C->failing()) {
1926     return false;
1927   }
1928 
1929   NOT_PRODUCT( C->verify_graph_edges(); )
1930 
1931   estimate_block_frequency();
1932 
1933   global_code_motion();
1934 
1935   if (C->failing()) {
1936     return false;
1937   }
1938 
1939   return true;
1940 }
1941 
1942 //------------------------------Estimate_Block_Frequency-----------------------
1943 // Estimate block frequencies based on IfNode probabilities.
1944 void PhaseCFG::estimate_block_frequency() {
1945 
1946   // Force conditional branches leading to uncommon traps to be unlikely,
1947   // not because we get to the uncommon_trap with less relative frequency,
1948   // but because an uncommon_trap typically causes a deopt, so we only get
1949   // there once.
1950   if (C->do_freq_based_layout()) {
1951     Block_List worklist;
1952     Block* root_blk = get_block(0);
1953     for (uint i = 1; i < root_blk->num_preds(); i++) {
1954       Block *pb = get_block_for_node(root_blk->pred(i));
1955       if (pb->has_uncommon_code()) {
1956         worklist.push(pb);
1957       }
1958     }
1959     while (worklist.size() > 0) {
1960       Block* uct = worklist.pop();
1961       if (uct == get_root_block()) {
1962         continue;
1963       }
1964       for (uint i = 1; i < uct->num_preds(); i++) {
1965         Block *pb = get_block_for_node(uct->pred(i));
1966         if (pb->_num_succs == 1) {
1967           worklist.push(pb);
1968         } else if (pb->num_fall_throughs() == 2) {
1969           pb->update_uncommon_branch(uct);
1970         }
1971       }
1972     }
1973   }
1974 
1975   // Create the loop tree and calculate loop depth.
1976   _root_loop = create_loop_tree();
1977   _root_loop->compute_loop_depth(0);
1978 
1979   // Compute block frequency of each block, relative to a single loop entry.
1980   _root_loop->compute_freq();
1981 
1982   // Adjust all frequencies to be relative to a single method entry
1983   _root_loop->_freq = 1.0;
1984   _root_loop->scale_freq();
1985 
1986   // Save outmost loop frequency for LRG frequency threshold
1987   _outer_loop_frequency = _root_loop->outer_loop_freq();
1988 
1989   // force paths ending at uncommon traps to be infrequent
1990   if (!C->do_freq_based_layout()) {
1991     Block_List worklist;
1992     Block* root_blk = get_block(0);
1993     for (uint i = 1; i < root_blk->num_preds(); i++) {
1994       Block *pb = get_block_for_node(root_blk->pred(i));
1995       if (pb->has_uncommon_code()) {
1996         worklist.push(pb);
1997       }
1998     }
1999     while (worklist.size() > 0) {
2000       Block* uct = worklist.pop();
2001       uct->_freq = PROB_MIN;
2002       for (uint i = 1; i < uct->num_preds(); i++) {
2003         Block *pb = get_block_for_node(uct->pred(i));
2004         if (pb->_num_succs == 1 && pb->_freq > PROB_MIN) {
2005           worklist.push(pb);
2006         }
2007       }
2008     }
2009   }
2010 
2011 #ifdef ASSERT
2012   for (uint i = 0; i < number_of_blocks(); i++) {
2013     Block* b = get_block(i);
2014     assert(b->_freq >= MIN_BLOCK_FREQUENCY, "Register Allocator requires meaningful block frequency");
2015   }
2016 #endif
2017 
2018 #ifndef PRODUCT
2019   if (PrintCFGBlockFreq) {
2020     tty->print_cr("CFG Block Frequencies");
2021     _root_loop->dump_tree();
2022     if (Verbose) {
2023       tty->print_cr("PhaseCFG dump");
2024       dump();
2025       tty->print_cr("Node dump");
2026       _root->dump(99999);
2027     }
2028   }
2029 #endif
2030 }
2031 
2032 //----------------------------create_loop_tree--------------------------------
2033 // Create a loop tree from the CFG
2034 CFGLoop* PhaseCFG::create_loop_tree() {
2035 
2036 #ifdef ASSERT
2037   assert(get_block(0) == get_root_block(), "first block should be root block");
2038   for (uint i = 0; i < number_of_blocks(); i++) {
2039     Block* block = get_block(i);
2040     // Check that _loop field are clear...we could clear them if not.
2041     assert(block->_loop == nullptr, "clear _loop expected");
2042     // Sanity check that the RPO numbering is reflected in the _blocks array.
2043     // It doesn't have to be for the loop tree to be built, but if it is not,
2044     // then the blocks have been reordered since dom graph building...which
2045     // may question the RPO numbering
2046     assert(block->_rpo == i, "unexpected reverse post order number");
2047   }
2048 #endif
2049 
2050   int idct = 0;
2051   CFGLoop* root_loop = new CFGLoop(idct++);
2052 
2053   Block_List worklist;
2054 
2055   // Assign blocks to loops
2056   for(uint i = number_of_blocks() - 1; i > 0; i-- ) { // skip Root block
2057     Block* block = get_block(i);
2058 
2059     if (block->head()->is_Loop()) {
2060       Block* loop_head = block;
2061       assert(loop_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
2062       Node* tail_n = loop_head->pred(LoopNode::LoopBackControl);
2063       Block* tail = get_block_for_node(tail_n);
2064 
2065       // Defensively filter out Loop nodes for non-single-entry loops.
2066       // For all reasonable loops, the head occurs before the tail in RPO.
2067       if (i <= tail->_rpo) {
2068 
2069         // The tail and (recursive) predecessors of the tail
2070         // are made members of a new loop.
2071 
2072         assert(worklist.size() == 0, "nonempty worklist");
2073         CFGLoop* nloop = new CFGLoop(idct++);
2074         assert(loop_head->_loop == nullptr, "just checking");
2075         loop_head->_loop = nloop;
2076         // Add to nloop so push_pred() will skip over inner loops
2077         nloop->add_member(loop_head);
2078         nloop->push_pred(loop_head, LoopNode::LoopBackControl, worklist, this);
2079 
2080         while (worklist.size() > 0) {
2081           Block* member = worklist.pop();
2082           if (member != loop_head) {
2083             for (uint j = 1; j < member->num_preds(); j++) {
2084               nloop->push_pred(member, j, worklist, this);
2085             }
2086           }
2087         }
2088       }
2089     }
2090   }
2091 
2092   // Create a member list for each loop consisting
2093   // of both blocks and (immediate child) loops.
2094   for (uint i = 0; i < number_of_blocks(); i++) {
2095     Block* block = get_block(i);
2096     CFGLoop* lp = block->_loop;
2097     if (lp == nullptr) {
2098       // Not assigned to a loop. Add it to the method's pseudo loop.
2099       block->_loop = root_loop;
2100       lp = root_loop;
2101     }
2102     if (lp == root_loop || block != lp->head()) { // loop heads are already members
2103       lp->add_member(block);
2104     }
2105     if (lp != root_loop) {
2106       if (lp->parent() == nullptr) {
2107         // Not a nested loop. Make it a child of the method's pseudo loop.
2108         root_loop->add_nested_loop(lp);
2109       }
2110       if (block == lp->head()) {
2111         // Add nested loop to member list of parent loop.
2112         lp->parent()->add_member(lp);
2113       }
2114     }
2115   }
2116 
2117   return root_loop;
2118 }
2119 
2120 //------------------------------push_pred--------------------------------------
2121 void CFGLoop::push_pred(Block* blk, int i, Block_List& worklist, PhaseCFG* cfg) {
2122   Node* pred_n = blk->pred(i);
2123   Block* pred = cfg->get_block_for_node(pred_n);
2124   CFGLoop *pred_loop = pred->_loop;
2125   if (pred_loop == nullptr) {
2126     // Filter out blocks for non-single-entry loops.
2127     // For all reasonable loops, the head occurs before the tail in RPO.
2128     if (pred->_rpo > head()->_rpo) {
2129       pred->_loop = this;
2130       worklist.push(pred);
2131     }
2132   } else if (pred_loop != this) {
2133     // Nested loop.
2134     while (pred_loop->_parent != nullptr && pred_loop->_parent != this) {
2135       pred_loop = pred_loop->_parent;
2136     }
2137     // Make pred's loop be a child
2138     if (pred_loop->_parent == nullptr) {
2139       add_nested_loop(pred_loop);
2140       // Continue with loop entry predecessor.
2141       Block* pred_head = pred_loop->head();
2142       assert(pred_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
2143       assert(pred_head != head(), "loop head in only one loop");
2144       push_pred(pred_head, LoopNode::EntryControl, worklist, cfg);
2145     } else {
2146       assert(pred_loop->_parent == this && _parent == nullptr, "just checking");
2147     }
2148   }
2149 }
2150 
2151 //------------------------------add_nested_loop--------------------------------
2152 // Make cl a child of the current loop in the loop tree.
2153 void CFGLoop::add_nested_loop(CFGLoop* cl) {
2154   assert(_parent == nullptr, "no parent yet");
2155   assert(cl != this, "not my own parent");
2156   cl->_parent = this;
2157   CFGLoop* ch = _child;
2158   if (ch == nullptr) {
2159     _child = cl;
2160   } else {
2161     while (ch->_sibling != nullptr) { ch = ch->_sibling; }
2162     ch->_sibling = cl;
2163   }
2164 }
2165 
2166 //------------------------------compute_loop_depth-----------------------------
2167 // Store the loop depth in each CFGLoop object.
2168 // Recursively walk the children to do the same for them.
2169 void CFGLoop::compute_loop_depth(int depth) {
2170   _depth = depth;
2171   CFGLoop* ch = _child;
2172   while (ch != nullptr) {
2173     ch->compute_loop_depth(depth + 1);
2174     ch = ch->_sibling;
2175   }
2176 }
2177 
2178 //------------------------------compute_freq-----------------------------------
2179 // Compute the frequency of each block and loop, relative to a single entry
2180 // into the dominating loop head.
2181 void CFGLoop::compute_freq() {
2182   // Bottom up traversal of loop tree (visit inner loops first.)
2183   // Set loop head frequency to 1.0, then transitively
2184   // compute frequency for all successors in the loop,
2185   // as well as for each exit edge.  Inner loops are
2186   // treated as single blocks with loop exit targets
2187   // as the successor blocks.
2188 
2189   // Nested loops first
2190   CFGLoop* ch = _child;
2191   while (ch != nullptr) {
2192     ch->compute_freq();
2193     ch = ch->_sibling;
2194   }
2195   assert (_members.length() > 0, "no empty loops");
2196   Block* hd = head();
2197   hd->_freq = 1.0;
2198   for (int i = 0; i < _members.length(); i++) {
2199     CFGElement* s = _members.at(i);
2200     double freq = s->_freq;
2201     if (s->is_block()) {
2202       Block* b = s->as_Block();
2203       for (uint j = 0; j < b->_num_succs; j++) {
2204         Block* sb = b->_succs[j];
2205         update_succ_freq(sb, freq * b->succ_prob(j));
2206       }
2207     } else {
2208       CFGLoop* lp = s->as_CFGLoop();
2209       assert(lp->_parent == this, "immediate child");
2210       for (int k = 0; k < lp->_exits.length(); k++) {
2211         Block* eb = lp->_exits.at(k).get_target();
2212         double prob = lp->_exits.at(k).get_prob();
2213         update_succ_freq(eb, freq * prob);
2214       }
2215     }
2216   }
2217 
2218   // For all loops other than the outer, "method" loop,
2219   // sum and normalize the exit probability. The "method" loop
2220   // should keep the initial exit probability of 1, so that
2221   // inner blocks do not get erroneously scaled.
2222   if (_depth != 0) {
2223     // Total the exit probabilities for this loop.
2224     double exits_sum = 0.0f;
2225     for (int i = 0; i < _exits.length(); i++) {
2226       exits_sum += _exits.at(i).get_prob();
2227     }
2228 
2229     // Normalize the exit probabilities. Until now, the
2230     // probabilities estimate the possibility of exit per
2231     // a single loop iteration; afterward, they estimate
2232     // the probability of exit per loop entry.
2233     for (int i = 0; i < _exits.length(); i++) {
2234       Block* et = _exits.at(i).get_target();
2235       float new_prob = 0.0f;
2236       if (_exits.at(i).get_prob() > 0.0f) {
2237         new_prob = _exits.at(i).get_prob() / exits_sum;
2238       }
2239       BlockProbPair bpp(et, new_prob);
2240       _exits.at_put(i, bpp);
2241     }
2242 
2243     // Save the total, but guard against unreasonable probability,
2244     // as the value is used to estimate the loop trip count.
2245     // An infinite trip count would blur relative block
2246     // frequencies.
2247     if (exits_sum > 1.0f) exits_sum = 1.0;
2248     if (exits_sum < PROB_MIN) exits_sum = PROB_MIN;
2249     _exit_prob = exits_sum;
2250   }
2251 }
2252 
2253 //------------------------------succ_prob-------------------------------------
2254 // Determine the probability of reaching successor 'i' from the receiver block.
2255 float Block::succ_prob(uint i) {
2256   int eidx = end_idx();
2257   Node *n = get_node(eidx);  // Get ending Node
2258 
2259   int op = n->Opcode();
2260   if (n->is_Mach()) {
2261     if (n->is_MachNullCheck()) {
2262       // Can only reach here if called after lcm. The original Op_If is gone,
2263       // so we attempt to infer the probability from one or both of the
2264       // successor blocks.
2265       assert(_num_succs == 2, "expecting 2 successors of a null check");
2266       // If either successor has only one predecessor, then the
2267       // probability estimate can be derived using the
2268       // relative frequency of the successor and this block.
2269       if (_succs[i]->num_preds() == 2) {
2270         return _succs[i]->_freq / _freq;
2271       } else if (_succs[1-i]->num_preds() == 2) {
2272         return 1 - (_succs[1-i]->_freq / _freq);
2273       } else {
2274         // Estimate using both successor frequencies
2275         float freq = _succs[i]->_freq;
2276         return freq / (freq + _succs[1-i]->_freq);
2277       }
2278     }
2279     op = n->as_Mach()->ideal_Opcode();
2280   }
2281 
2282 
2283   // Switch on branch type
2284   switch( op ) {
2285   case Op_CountedLoopEnd:
2286   case Op_If: {
2287     assert (i < 2, "just checking");
2288     // Conditionals pass on only part of their frequency
2289     float prob  = n->as_MachIf()->_prob;
2290     assert(prob >= 0.0 && prob <= 1.0, "out of range probability");
2291     // If succ[i] is the FALSE branch, invert path info
2292     if( get_node(i + eidx + 1)->Opcode() == Op_IfFalse ) {
2293       return 1.0f - prob; // not taken
2294     } else {
2295       return prob; // taken
2296     }
2297   }
2298 
2299   case Op_Jump:
2300     return n->as_MachJump()->_probs[get_node(i + eidx + 1)->as_JumpProj()->_con];
2301 
2302   case Op_Catch: {
2303     const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj();
2304     if (ci->_con == CatchProjNode::fall_through_index) {
2305       // Fall-thru path gets the lion's share.
2306       return 1.0f - PROB_UNLIKELY_MAG(5)*_num_succs;
2307     } else {
2308       // Presume exceptional paths are equally unlikely
2309       return PROB_UNLIKELY_MAG(5);
2310     }
2311   }
2312 
2313   case Op_Root:
2314   case Op_Goto:
2315     // Pass frequency straight thru to target
2316     return 1.0f;
2317 
2318   case Op_NeverBranch: {
2319     Node* succ = n->as_NeverBranch()->proj_out(0)->unique_ctrl_out();
2320     if (_succs[i]->head() == succ) {
2321       return 1.0f;
2322     }
2323     return 0.0f;
2324   }
2325 
2326   case Op_TailCall:
2327   case Op_TailJump:
2328   case Op_ForwardException:
2329   case Op_Return:
2330   case Op_Halt:
2331   case Op_Rethrow:
2332     // Do not push out freq to root block
2333     return 0.0f;
2334 
2335   default:
2336     ShouldNotReachHere();
2337   }
2338 
2339   return 0.0f;
2340 }
2341 
2342 //------------------------------num_fall_throughs-----------------------------
2343 // Return the number of fall-through candidates for a block
2344 int Block::num_fall_throughs() {
2345   int eidx = end_idx();
2346   Node *n = get_node(eidx);  // Get ending Node
2347 
2348   int op = n->Opcode();
2349   if (n->is_Mach()) {
2350     if (n->is_MachNullCheck()) {
2351       // In theory, either side can fall-thru, for simplicity sake,
2352       // let's say only the false branch can now.
2353       return 1;
2354     }
2355     op = n->as_Mach()->ideal_Opcode();
2356   }
2357 
2358   // Switch on branch type
2359   switch( op ) {
2360   case Op_CountedLoopEnd:
2361   case Op_If:
2362     return 2;
2363 
2364   case Op_Root:
2365   case Op_Goto:
2366     return 1;
2367 
2368   case Op_Catch: {
2369     for (uint i = 0; i < _num_succs; i++) {
2370       const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj();
2371       if (ci->_con == CatchProjNode::fall_through_index) {
2372         return 1;
2373       }
2374     }
2375     return 0;
2376   }
2377 
2378   case Op_Jump:
2379   case Op_NeverBranch:
2380   case Op_TailCall:
2381   case Op_TailJump:
2382   case Op_ForwardException:
2383   case Op_Return:
2384   case Op_Halt:
2385   case Op_Rethrow:
2386     return 0;
2387 
2388   default:
2389     ShouldNotReachHere();
2390   }
2391 
2392   return 0;
2393 }
2394 
2395 //------------------------------succ_fall_through-----------------------------
2396 // Return true if a specific successor could be fall-through target.
2397 bool Block::succ_fall_through(uint i) {
2398   int eidx = end_idx();
2399   Node *n = get_node(eidx);  // Get ending Node
2400 
2401   int op = n->Opcode();
2402   if (n->is_Mach()) {
2403     if (n->is_MachNullCheck()) {
2404       // In theory, either side can fall-thru, for simplicity sake,
2405       // let's say only the false branch can now.
2406       return get_node(i + eidx + 1)->Opcode() == Op_IfFalse;
2407     }
2408     op = n->as_Mach()->ideal_Opcode();
2409   }
2410 
2411   // Switch on branch type
2412   switch( op ) {
2413   case Op_CountedLoopEnd:
2414   case Op_If:
2415   case Op_Root:
2416   case Op_Goto:
2417     return true;
2418 
2419   case Op_Catch: {
2420     const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj();
2421     return ci->_con == CatchProjNode::fall_through_index;
2422   }
2423 
2424   case Op_Jump:
2425   case Op_NeverBranch:
2426   case Op_TailCall:
2427   case Op_TailJump:
2428   case Op_ForwardException:
2429   case Op_Return:
2430   case Op_Halt:
2431   case Op_Rethrow:
2432     return false;
2433 
2434   default:
2435     ShouldNotReachHere();
2436   }
2437 
2438   return false;
2439 }
2440 
2441 //------------------------------update_uncommon_branch------------------------
2442 // Update the probability of a two-branch to be uncommon
2443 void Block::update_uncommon_branch(Block* ub) {
2444   int eidx = end_idx();
2445   Node *n = get_node(eidx);  // Get ending Node
2446 
2447   int op = n->as_Mach()->ideal_Opcode();
2448 
2449   assert(op == Op_CountedLoopEnd || op == Op_If, "must be a If");
2450   assert(num_fall_throughs() == 2, "must be a two way branch block");
2451 
2452   // Which successor is ub?
2453   uint s;
2454   for (s = 0; s <_num_succs; s++) {
2455     if (_succs[s] == ub) break;
2456   }
2457   assert(s < 2, "uncommon successor must be found");
2458 
2459   // If ub is the true path, make the proability small, else
2460   // ub is the false path, and make the probability large
2461   bool invert = (get_node(s + eidx + 1)->Opcode() == Op_IfFalse);
2462 
2463   // Get existing probability
2464   float p = n->as_MachIf()->_prob;
2465 
2466   if (invert) p = 1.0 - p;
2467   if (p > PROB_MIN) {
2468     p = PROB_MIN;
2469   }
2470   if (invert) p = 1.0 - p;
2471 
2472   n->as_MachIf()->_prob = p;
2473 }
2474 
2475 //------------------------------update_succ_freq-------------------------------
2476 // Update the appropriate frequency associated with block 'b', a successor of
2477 // a block in this loop.
2478 void CFGLoop::update_succ_freq(Block* b, double freq) {
2479   if (b->_loop == this) {
2480     if (b == head()) {
2481       // back branch within the loop
2482       // Do nothing now, the loop carried frequency will be
2483       // adjust later in scale_freq().
2484     } else {
2485       // simple branch within the loop
2486       b->_freq += freq;
2487     }
2488   } else if (!in_loop_nest(b)) {
2489     // branch is exit from this loop
2490     BlockProbPair bpp(b, freq);
2491     _exits.append(bpp);
2492   } else {
2493     // branch into nested loop
2494     CFGLoop* ch = b->_loop;
2495     ch->_freq += freq;
2496   }
2497 }
2498 
2499 //------------------------------in_loop_nest-----------------------------------
2500 // Determine if block b is in the receiver's loop nest.
2501 bool CFGLoop::in_loop_nest(Block* b) {
2502   int depth = _depth;
2503   CFGLoop* b_loop = b->_loop;
2504   int b_depth = b_loop->_depth;
2505   if (depth == b_depth) {
2506     return true;
2507   }
2508   while (b_depth > depth) {
2509     b_loop = b_loop->_parent;
2510     b_depth = b_loop->_depth;
2511   }
2512   return b_loop == this;
2513 }
2514 
2515 //------------------------------scale_freq-------------------------------------
2516 // Scale frequency of loops and blocks by trip counts from outer loops
2517 // Do a top down traversal of loop tree (visit outer loops first.)
2518 void CFGLoop::scale_freq() {
2519   double loop_freq = _freq * trip_count();
2520   _freq = loop_freq;
2521   for (int i = 0; i < _members.length(); i++) {
2522     CFGElement* s = _members.at(i);
2523     double block_freq = s->_freq * loop_freq;
2524     if (g_isnan(block_freq) || block_freq < MIN_BLOCK_FREQUENCY)
2525       block_freq = MIN_BLOCK_FREQUENCY;
2526     s->_freq = block_freq;
2527   }
2528   CFGLoop* ch = _child;
2529   while (ch != nullptr) {
2530     ch->scale_freq();
2531     ch = ch->_sibling;
2532   }
2533 }
2534 
2535 // Frequency of outer loop
2536 double CFGLoop::outer_loop_freq() const {
2537   if (_child != nullptr) {
2538     return _child->_freq;
2539   }
2540   return _freq;
2541 }
2542 
2543 #ifndef PRODUCT
2544 //------------------------------dump_tree--------------------------------------
2545 void CFGLoop::dump_tree() const {
2546   dump();
2547   if (_child != nullptr)   _child->dump_tree();
2548   if (_sibling != nullptr) _sibling->dump_tree();
2549 }
2550 
2551 //------------------------------dump-------------------------------------------
2552 void CFGLoop::dump() const {
2553   for (int i = 0; i < _depth; i++) tty->print("   ");
2554   tty->print("%s: %d  trip_count: %6.0f freq: %6.0f\n",
2555              _depth == 0 ? "Method" : "Loop", _id, trip_count(), _freq);
2556   for (int i = 0; i < _depth; i++) tty->print("   ");
2557   tty->print("         members:");
2558   int k = 0;
2559   for (int i = 0; i < _members.length(); i++) {
2560     if (k++ >= 6) {
2561       tty->print("\n              ");
2562       for (int j = 0; j < _depth+1; j++) tty->print("   ");
2563       k = 0;
2564     }
2565     CFGElement *s = _members.at(i);
2566     if (s->is_block()) {
2567       Block *b = s->as_Block();
2568       tty->print(" B%d(%6.3f)", b->_pre_order, b->_freq);
2569     } else {
2570       CFGLoop* lp = s->as_CFGLoop();
2571       tty->print(" L%d(%6.3f)", lp->_id, lp->_freq);
2572     }
2573   }
2574   tty->print("\n");
2575   for (int i = 0; i < _depth; i++) tty->print("   ");
2576   tty->print("         exits:  ");
2577   k = 0;
2578   for (int i = 0; i < _exits.length(); i++) {
2579     if (k++ >= 7) {
2580       tty->print("\n              ");
2581       for (int j = 0; j < _depth+1; j++) tty->print("   ");
2582       k = 0;
2583     }
2584     Block *blk = _exits.at(i).get_target();
2585     double prob = _exits.at(i).get_prob();
2586     tty->print(" ->%d@%d%%", blk->_pre_order, (int)(prob*100));
2587   }
2588   tty->print("\n");
2589 }
2590 #endif