1 /*
   2  * Copyright (c) 1997, 2012, 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 "precompiled.hpp"
  26 #include "libadt/vectset.hpp"
  27 #include "memory/allocation.inline.hpp"
  28 #include "opto/block.hpp"
  29 #include "opto/c2compiler.hpp"
  30 #include "opto/callnode.hpp"
  31 #include "opto/cfgnode.hpp"
  32 #include "opto/machnode.hpp"
  33 #include "opto/opcodes.hpp"
  34 #include "opto/phaseX.hpp"
  35 #include "opto/rootnode.hpp"
  36 #include "opto/runtime.hpp"
  37 #include "runtime/deoptimization.hpp"
  38 #ifdef TARGET_ARCH_MODEL_x86_32
  39 # include "adfiles/ad_x86_32.hpp"
  40 #endif
  41 #ifdef TARGET_ARCH_MODEL_x86_64
  42 # include "adfiles/ad_x86_64.hpp"
  43 #endif
  44 #ifdef TARGET_ARCH_MODEL_sparc
  45 # include "adfiles/ad_sparc.hpp"
  46 #endif
  47 #ifdef TARGET_ARCH_MODEL_zero
  48 # include "adfiles/ad_zero.hpp"
  49 #endif
  50 #ifdef TARGET_ARCH_MODEL_arm
  51 # include "adfiles/ad_arm.hpp"
  52 #endif
  53 #ifdef TARGET_ARCH_MODEL_ppc
  54 # include "adfiles/ad_ppc.hpp"
  55 #endif
  56 
  57 // Portions of code courtesy of Clifford Click
  58 
  59 // Optimization - Graph Style
  60 
  61 // To avoid float value underflow
  62 #define MIN_BLOCK_FREQUENCY 1.e-35f
  63 
  64 //----------------------------schedule_node_into_block-------------------------
  65 // Insert node n into block b. Look for projections of n and make sure they
  66 // are in b also.
  67 void PhaseCFG::schedule_node_into_block( Node *n, Block *b ) {
  68   // Set basic block of n, Add n to b,
  69   map_node_to_block(n, b);
  70   b->add_inst(n);
  71 
  72   // After Matching, nearly any old Node may have projections trailing it.
  73   // These are usually machine-dependent flags.  In any case, they might
  74   // float to another block below this one.  Move them up.
  75   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
  76     Node*  use  = n->fast_out(i);
  77     if (use->is_Proj()) {
  78       Block* buse = get_block_for_node(use);
  79       if (buse != b) {              // In wrong block?
  80         if (buse != NULL) {
  81           buse->find_remove(use);   // Remove from wrong block
  82         }
  83         map_node_to_block(use, b);
  84         b->add_inst(use);
  85       }
  86     }
  87   }
  88 }
  89 
  90 //----------------------------replace_block_proj_ctrl-------------------------
  91 // Nodes that have is_block_proj() nodes as their control need to use
  92 // the appropriate Region for their actual block as their control since
  93 // the projection will be in a predecessor block.
  94 void PhaseCFG::replace_block_proj_ctrl( Node *n ) {
  95   const Node *in0 = n->in(0);
  96   assert(in0 != NULL, "Only control-dependent");
  97   const Node *p = in0->is_block_proj();
  98   if (p != NULL && p != n) {    // Control from a block projection?
  99     assert(!n->pinned() || n->is_MachConstantBase(), "only pinned MachConstantBase node is expected here");
 100     // Find trailing Region
 101     Block *pb = get_block_for_node(in0); // Block-projection already has basic block
 102     uint j = 0;
 103     if (pb->_num_succs != 1) {  // More then 1 successor?
 104       // Search for successor
 105       uint max = pb->_nodes.size();
 106       assert( max > 1, "" );
 107       uint start = max - pb->_num_succs;
 108       // Find which output path belongs to projection
 109       for (j = start; j < max; j++) {
 110         if( pb->_nodes[j] == in0 )
 111           break;
 112       }
 113       assert( j < max, "must find" );
 114       // Change control to match head of successor basic block
 115       j -= start;
 116     }
 117     n->set_req(0, pb->_succs[j]->head());
 118   }
 119 }
 120 
 121 
 122 //------------------------------schedule_pinned_nodes--------------------------
 123 // Set the basic block for Nodes pinned into blocks
 124 void PhaseCFG::schedule_pinned_nodes( VectorSet &visited ) {
 125   // Allocate node stack of size C->unique()+8 to avoid frequent realloc
 126   GrowableArray <Node *> spstack(C->unique()+8);
 127   spstack.push(_root);
 128   while ( spstack.is_nonempty() ) {
 129     Node *n = spstack.pop();
 130     if( !visited.test_set(n->_idx) ) { // Test node and flag it as visited
 131       if( n->pinned() && !has_block(n)) {  // Pinned?  Nail it down!
 132         assert( n->in(0), "pinned Node must have Control" );
 133         // Before setting block replace block_proj control edge
 134         replace_block_proj_ctrl(n);
 135         Node *input = n->in(0);
 136         while (!input->is_block_start()) {
 137           input = input->in(0);
 138         }
 139         Block *b = get_block_for_node(input); // Basic block of controlling input
 140         schedule_node_into_block(n, b);
 141       }
 142       for( int i = n->req() - 1; i >= 0; --i ) {  // For all inputs
 143         if( n->in(i) != NULL )
 144           spstack.push(n->in(i));
 145       }
 146     }
 147   }
 148 }
 149 
 150 #ifdef ASSERT
 151 // Assert that new input b2 is dominated by all previous inputs.
 152 // Check this by by seeing that it is dominated by b1, the deepest
 153 // input observed until b2.
 154 static void assert_dom(Block* b1, Block* b2, Node* n, const PhaseCFG* cfg) {
 155   if (b1 == NULL)  return;
 156   assert(b1->_dom_depth < b2->_dom_depth, "sanity");
 157   Block* tmp = b2;
 158   while (tmp != b1 && tmp != NULL) {
 159     tmp = tmp->_idom;
 160   }
 161   if (tmp != b1) {
 162     // Detected an unschedulable graph.  Print some nice stuff and die.
 163     tty->print_cr("!!! Unschedulable graph !!!");
 164     for (uint j=0; j<n->len(); j++) { // For all inputs
 165       Node* inn = n->in(j); // Get input
 166       if (inn == NULL)  continue;  // Ignore NULL, missing inputs
 167       Block* inb = cfg->get_block_for_node(inn);
 168       tty->print("B%d idom=B%d depth=%2d ",inb->_pre_order,
 169                  inb->_idom ? inb->_idom->_pre_order : 0, inb->_dom_depth);
 170       inn->dump();
 171     }
 172     tty->print("Failing node: ");
 173     n->dump();
 174     assert(false, "unscheduable graph");
 175   }
 176 }
 177 #endif
 178 
 179 static Block* find_deepest_input(Node* n, const PhaseCFG* cfg) {
 180   // Find the last input dominated by all other inputs.
 181   Block* deepb           = NULL;        // Deepest block so far
 182   int    deepb_dom_depth = 0;
 183   for (uint k = 0; k < n->len(); k++) { // For all inputs
 184     Node* inn = n->in(k);               // Get input
 185     if (inn == NULL)  continue;         // Ignore NULL, missing inputs
 186     Block* inb = cfg->get_block_for_node(inn);
 187     assert(inb != NULL, "must already have scheduled this input");
 188     if (deepb_dom_depth < (int) inb->_dom_depth) {
 189       // The new inb must be dominated by the previous deepb.
 190       // The various inputs must be linearly ordered in the dom
 191       // tree, or else there will not be a unique deepest block.
 192       DEBUG_ONLY(assert_dom(deepb, inb, n, cfg));
 193       deepb = inb;                      // Save deepest block
 194       deepb_dom_depth = deepb->_dom_depth;
 195     }
 196   }
 197   assert(deepb != NULL, "must be at least one input to n");
 198   return deepb;
 199 }
 200 
 201 
 202 //------------------------------schedule_early---------------------------------
 203 // Find the earliest Block any instruction can be placed in.  Some instructions
 204 // are pinned into Blocks.  Unpinned instructions can appear in last block in
 205 // which all their inputs occur.
 206 bool PhaseCFG::schedule_early(VectorSet &visited, Node_List &roots) {
 207   // Allocate stack with enough space to avoid frequent realloc
 208   Node_Stack nstack(roots.Size() + 8); // (unique >> 1) + 24 from Java2D stats
 209   // roots.push(_root); _root will be processed among C->top() inputs
 210   roots.push(C->top());
 211   visited.set(C->top()->_idx);
 212 
 213   while (roots.size() != 0) {
 214     // Use local variables nstack_top_n & nstack_top_i to cache values
 215     // on stack's top.
 216     Node *nstack_top_n = roots.pop();
 217     uint  nstack_top_i = 0;
 218 //while_nstack_nonempty:
 219     while (true) {
 220       // Get parent node and next input's index from stack's top.
 221       Node *n = nstack_top_n;
 222       uint  i = nstack_top_i;
 223 
 224       if (i == 0) {
 225         // Fixup some control.  Constants without control get attached
 226         // to root and nodes that use is_block_proj() nodes should be attached
 227         // to the region that starts their block.
 228         const Node *in0 = n->in(0);
 229         if (in0 != NULL) {              // Control-dependent?
 230           replace_block_proj_ctrl(n);
 231         } else {               // n->in(0) == NULL
 232           if (n->req() == 1) { // This guy is a constant with NO inputs?
 233             n->set_req(0, _root);
 234           }
 235         }
 236       }
 237 
 238       // First, visit all inputs and force them to get a block.  If an
 239       // input is already in a block we quit following inputs (to avoid
 240       // cycles). Instead we put that Node on a worklist to be handled
 241       // later (since IT'S inputs may not have a block yet).
 242       bool done = true;              // Assume all n's inputs will be processed
 243       while (i < n->len()) {         // For all inputs
 244         Node *in = n->in(i);         // Get input
 245         ++i;
 246         if (in == NULL) continue;    // Ignore NULL, missing inputs
 247         int is_visited = visited.test_set(in->_idx);
 248         if (!has_block(in)) { // Missing block selection?
 249           if (is_visited) {
 250             // assert( !visited.test(in->_idx), "did not schedule early" );
 251             return false;
 252           }
 253           nstack.push(n, i);         // Save parent node and next input's index.
 254           nstack_top_n = in;         // Process current input now.
 255           nstack_top_i = 0;
 256           done = false;              // Not all n's inputs processed.
 257           break; // continue while_nstack_nonempty;
 258         } else if (!is_visited) {    // Input not yet visited?
 259           roots.push(in);            // Visit this guy later, using worklist
 260         }
 261       }
 262       if (done) {
 263         // All of n's inputs have been processed, complete post-processing.
 264 
 265         // Some instructions are pinned into a block.  These include Region,
 266         // Phi, Start, Return, and other control-dependent instructions and
 267         // any projections which depend on them.
 268         if (!n->pinned()) {
 269           // Set earliest legal block.
 270           map_node_to_block(n, find_deepest_input(n, this));
 271         } else {
 272           assert(get_block_for_node(n) == get_block_for_node(n->in(0)), "Pinned Node should be at the same block as its control edge");
 273         }
 274 
 275         if (nstack.is_empty()) {
 276           // Finished all nodes on stack.
 277           // Process next node on the worklist 'roots'.
 278           break;
 279         }
 280         // Get saved parent node and next input's index.
 281         nstack_top_n = nstack.node();
 282         nstack_top_i = nstack.index();
 283         nstack.pop();
 284       } //    if (done)
 285     }   // while (true)
 286   }     // while (roots.size() != 0)
 287   return true;
 288 }
 289 
 290 //------------------------------dom_lca----------------------------------------
 291 // Find least common ancestor in dominator tree
 292 // LCA is a current notion of LCA, to be raised above 'this'.
 293 // As a convenient boundary condition, return 'this' if LCA is NULL.
 294 // Find the LCA of those two nodes.
 295 Block* Block::dom_lca(Block* LCA) {
 296   if (LCA == NULL || LCA == this)  return this;
 297 
 298   Block* anc = this;
 299   while (anc->_dom_depth > LCA->_dom_depth)
 300     anc = anc->_idom;           // Walk up till anc is as high as LCA
 301 
 302   while (LCA->_dom_depth > anc->_dom_depth)
 303     LCA = LCA->_idom;           // Walk up till LCA is as high as anc
 304 
 305   while (LCA != anc) {          // Walk both up till they are the same
 306     LCA = LCA->_idom;
 307     anc = anc->_idom;
 308   }
 309 
 310   return LCA;
 311 }
 312 
 313 //--------------------------raise_LCA_above_use--------------------------------
 314 // We are placing a definition, and have been given a def->use edge.
 315 // The definition must dominate the use, so move the LCA upward in the
 316 // dominator tree to dominate the use.  If the use is a phi, adjust
 317 // the LCA only with the phi input paths which actually use this def.
 318 static Block* raise_LCA_above_use(Block* LCA, Node* use, Node* def, const PhaseCFG* cfg) {
 319   Block* buse = cfg->get_block_for_node(use);
 320   if (buse == NULL)    return LCA;   // Unused killing Projs have no use block
 321   if (!use->is_Phi())  return buse->dom_lca(LCA);
 322   uint pmax = use->req();       // Number of Phi inputs
 323   // Why does not this loop just break after finding the matching input to
 324   // the Phi?  Well...it's like this.  I do not have true def-use/use-def
 325   // chains.  Means I cannot distinguish, from the def-use direction, which
 326   // of many use-defs lead from the same use to the same def.  That is, this
 327   // Phi might have several uses of the same def.  Each use appears in a
 328   // different predecessor block.  But when I enter here, I cannot distinguish
 329   // which use-def edge I should find the predecessor block for.  So I find
 330   // them all.  Means I do a little extra work if a Phi uses the same value
 331   // more than once.
 332   for (uint j=1; j<pmax; j++) { // For all inputs
 333     if (use->in(j) == def) {    // Found matching input?
 334       Block* pred = cfg->get_block_for_node(buse->pred(j));
 335       LCA = pred->dom_lca(LCA);
 336     }
 337   }
 338   return LCA;
 339 }
 340 
 341 //----------------------------raise_LCA_above_marks----------------------------
 342 // Return a new LCA that dominates LCA and any of its marked predecessors.
 343 // Search all my parents up to 'early' (exclusive), looking for predecessors
 344 // which are marked with the given index.  Return the LCA (in the dom tree)
 345 // of all marked blocks.  If there are none marked, return the original
 346 // LCA.
 347 static Block* raise_LCA_above_marks(Block* LCA, node_idx_t mark, Block* early, const PhaseCFG* cfg) {
 348   Block_List worklist;
 349   worklist.push(LCA);
 350   while (worklist.size() > 0) {
 351     Block* mid = worklist.pop();
 352     if (mid == early)  continue;  // stop searching here
 353 
 354     // Test and set the visited bit.
 355     if (mid->raise_LCA_visited() == mark)  continue;  // already visited
 356 
 357     // Don't process the current LCA, otherwise the search may terminate early
 358     if (mid != LCA && mid->raise_LCA_mark() == mark) {
 359       // Raise the LCA.
 360       LCA = mid->dom_lca(LCA);
 361       if (LCA == early)  break;   // stop searching everywhere
 362       assert(early->dominates(LCA), "early is high enough");
 363       // Resume searching at that point, skipping intermediate levels.
 364       worklist.push(LCA);
 365       if (LCA == mid)
 366         continue; // Don't mark as visited to avoid early termination.
 367     } else {
 368       // Keep searching through this block's predecessors.
 369       for (uint j = 1, jmax = mid->num_preds(); j < jmax; j++) {
 370         Block* mid_parent = cfg->get_block_for_node(mid->pred(j));
 371         worklist.push(mid_parent);
 372       }
 373     }
 374     mid->set_raise_LCA_visited(mark);
 375   }
 376   return LCA;
 377 }
 378 
 379 //--------------------------memory_early_block--------------------------------
 380 // This is a variation of find_deepest_input, the heart of schedule_early.
 381 // Find the "early" block for a load, if we considered only memory and
 382 // address inputs, that is, if other data inputs were ignored.
 383 //
 384 // Because a subset of edges are considered, the resulting block will
 385 // be earlier (at a shallower dom_depth) than the true schedule_early
 386 // point of the node. We compute this earlier block as a more permissive
 387 // site for anti-dependency insertion, but only if subsume_loads is enabled.
 388 static Block* memory_early_block(Node* load, Block* early, const PhaseCFG* cfg) {
 389   Node* base;
 390   Node* index;
 391   Node* store = load->in(MemNode::Memory);
 392   load->as_Mach()->memory_inputs(base, index);
 393 
 394   assert(base != NodeSentinel && index != NodeSentinel,
 395          "unexpected base/index inputs");
 396 
 397   Node* mem_inputs[4];
 398   int mem_inputs_length = 0;
 399   if (base != NULL)  mem_inputs[mem_inputs_length++] = base;
 400   if (index != NULL) mem_inputs[mem_inputs_length++] = index;
 401   if (store != NULL) mem_inputs[mem_inputs_length++] = store;
 402 
 403   // In the comparision below, add one to account for the control input,
 404   // which may be null, but always takes up a spot in the in array.
 405   if (mem_inputs_length + 1 < (int) load->req()) {
 406     // This "load" has more inputs than just the memory, base and index inputs.
 407     // For purposes of checking anti-dependences, we need to start
 408     // from the early block of only the address portion of the instruction,
 409     // and ignore other blocks that may have factored into the wider
 410     // schedule_early calculation.
 411     if (load->in(0) != NULL) mem_inputs[mem_inputs_length++] = load->in(0);
 412 
 413     Block* deepb           = NULL;        // Deepest block so far
 414     int    deepb_dom_depth = 0;
 415     for (int i = 0; i < mem_inputs_length; i++) {
 416       Block* inb = cfg->get_block_for_node(mem_inputs[i]);
 417       if (deepb_dom_depth < (int) inb->_dom_depth) {
 418         // The new inb must be dominated by the previous deepb.
 419         // The various inputs must be linearly ordered in the dom
 420         // tree, or else there will not be a unique deepest block.
 421         DEBUG_ONLY(assert_dom(deepb, inb, load, cfg));
 422         deepb = inb;                      // Save deepest block
 423         deepb_dom_depth = deepb->_dom_depth;
 424       }
 425     }
 426     early = deepb;
 427   }
 428 
 429   return early;
 430 }
 431 
 432 //--------------------------insert_anti_dependences---------------------------
 433 // A load may need to witness memory that nearby stores can overwrite.
 434 // For each nearby store, either insert an "anti-dependence" edge
 435 // from the load to the store, or else move LCA upward to force the
 436 // load to (eventually) be scheduled in a block above the store.
 437 //
 438 // Do not add edges to stores on distinct control-flow paths;
 439 // only add edges to stores which might interfere.
 440 //
 441 // Return the (updated) LCA.  There will not be any possibly interfering
 442 // store between the load's "early block" and the updated LCA.
 443 // Any stores in the updated LCA will have new precedence edges
 444 // back to the load.  The caller is expected to schedule the load
 445 // in the LCA, in which case the precedence edges will make LCM
 446 // preserve anti-dependences.  The caller may also hoist the load
 447 // above the LCA, if it is not the early block.
 448 Block* PhaseCFG::insert_anti_dependences(Block* LCA, Node* load, bool verify) {
 449   assert(load->needs_anti_dependence_check(), "must be a load of some sort");
 450   assert(LCA != NULL, "");
 451   DEBUG_ONLY(Block* LCA_orig = LCA);
 452 
 453   // Compute the alias index.  Loads and stores with different alias indices
 454   // do not need anti-dependence edges.
 455   uint load_alias_idx = C->get_alias_index(load->adr_type());
 456 #ifdef ASSERT
 457   if (load_alias_idx == Compile::AliasIdxBot && C->AliasLevel() > 0 &&
 458       (PrintOpto || VerifyAliases ||
 459        PrintMiscellaneous && (WizardMode || Verbose))) {
 460     // Load nodes should not consume all of memory.
 461     // Reporting a bottom type indicates a bug in adlc.
 462     // If some particular type of node validly consumes all of memory,
 463     // sharpen the preceding "if" to exclude it, so we can catch bugs here.
 464     tty->print_cr("*** Possible Anti-Dependence Bug:  Load consumes all of memory.");
 465     load->dump(2);
 466     if (VerifyAliases)  assert(load_alias_idx != Compile::AliasIdxBot, "");
 467   }
 468 #endif
 469   assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrComp),
 470          "String compare is only known 'load' that does not conflict with any stores");
 471   assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrEquals),
 472          "String equals is a 'load' that does not conflict with any stores");
 473   assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrIndexOf),
 474          "String indexOf is a 'load' that does not conflict with any stores");
 475   assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_AryEq),
 476          "Arrays equals is a 'load' that do not conflict with any stores");
 477 
 478   if (!C->alias_type(load_alias_idx)->is_rewritable()) {
 479     // It is impossible to spoil this load by putting stores before it,
 480     // because we know that the stores will never update the value
 481     // which 'load' must witness.
 482     return LCA;
 483   }
 484 
 485   node_idx_t load_index = load->_idx;
 486 
 487   // Note the earliest legal placement of 'load', as determined by
 488   // by the unique point in the dom tree where all memory effects
 489   // and other inputs are first available.  (Computed by schedule_early.)
 490   // For normal loads, 'early' is the shallowest place (dom graph wise)
 491   // to look for anti-deps between this load and any store.
 492   Block* early = get_block_for_node(load);
 493 
 494   // If we are subsuming loads, compute an "early" block that only considers
 495   // memory or address inputs. This block may be different than the
 496   // schedule_early block in that it could be at an even shallower depth in the
 497   // dominator tree, and allow for a broader discovery of anti-dependences.
 498   if (C->subsume_loads()) {
 499     early = memory_early_block(load, early, this);
 500   }
 501 
 502   ResourceArea *area = Thread::current()->resource_area();
 503   Node_List worklist_mem(area);     // prior memory state to store
 504   Node_List worklist_store(area);   // possible-def to explore
 505   Node_List worklist_visited(area); // visited mergemem nodes
 506   Node_List non_early_stores(area); // all relevant stores outside of early
 507   bool must_raise_LCA = false;
 508 
 509 #ifdef TRACK_PHI_INPUTS
 510   // %%% This extra checking fails because MergeMem nodes are not GVNed.
 511   // Provide "phi_inputs" to check if every input to a PhiNode is from the
 512   // original memory state.  This indicates a PhiNode for which should not
 513   // prevent the load from sinking.  For such a block, set_raise_LCA_mark
 514   // may be overly conservative.
 515   // Mechanism: count inputs seen for each Phi encountered in worklist_store.
 516   DEBUG_ONLY(GrowableArray<uint> phi_inputs(area, C->unique(),0,0));
 517 #endif
 518 
 519   // 'load' uses some memory state; look for users of the same state.
 520   // Recurse through MergeMem nodes to the stores that use them.
 521 
 522   // Each of these stores is a possible definition of memory
 523   // that 'load' needs to use.  We need to force 'load'
 524   // to occur before each such store.  When the store is in
 525   // the same block as 'load', we insert an anti-dependence
 526   // edge load->store.
 527 
 528   // The relevant stores "nearby" the load consist of a tree rooted
 529   // at initial_mem, with internal nodes of type MergeMem.
 530   // Therefore, the branches visited by the worklist are of this form:
 531   //    initial_mem -> (MergeMem ->)* store
 532   // The anti-dependence constraints apply only to the fringe of this tree.
 533 
 534   Node* initial_mem = load->in(MemNode::Memory);
 535   worklist_store.push(initial_mem);
 536   worklist_visited.push(initial_mem);
 537   worklist_mem.push(NULL);
 538   while (worklist_store.size() > 0) {
 539     // Examine a nearby store to see if it might interfere with our load.
 540     Node* mem   = worklist_mem.pop();
 541     Node* store = worklist_store.pop();
 542     uint op = store->Opcode();
 543 
 544     // MergeMems do not directly have anti-deps.
 545     // Treat them as internal nodes in a forward tree of memory states,
 546     // the leaves of which are each a 'possible-def'.
 547     if (store == initial_mem    // root (exclusive) of tree we are searching
 548         || op == Op_MergeMem    // internal node of tree we are searching
 549         ) {
 550       mem = store;   // It's not a possibly interfering store.
 551       if (store == initial_mem)
 552         initial_mem = NULL;  // only process initial memory once
 553 
 554       for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
 555         store = mem->fast_out(i);
 556         if (store->is_MergeMem()) {
 557           // Be sure we don't get into combinatorial problems.
 558           // (Allow phis to be repeated; they can merge two relevant states.)
 559           uint j = worklist_visited.size();
 560           for (; j > 0; j--) {
 561             if (worklist_visited.at(j-1) == store)  break;
 562           }
 563           if (j > 0)  continue; // already on work list; do not repeat
 564           worklist_visited.push(store);
 565         }
 566         worklist_mem.push(mem);
 567         worklist_store.push(store);
 568       }
 569       continue;
 570     }
 571 
 572     if (op == Op_MachProj || op == Op_Catch)   continue;
 573     if (store->needs_anti_dependence_check())  continue;  // not really a store
 574 
 575     // Compute the alias index.  Loads and stores with different alias
 576     // indices do not need anti-dependence edges.  Wide MemBar's are
 577     // anti-dependent on everything (except immutable memories).
 578     const TypePtr* adr_type = store->adr_type();
 579     if (!C->can_alias(adr_type, load_alias_idx))  continue;
 580 
 581     // Most slow-path runtime calls do NOT modify Java memory, but
 582     // they can block and so write Raw memory.
 583     if (store->is_Mach()) {
 584       MachNode* mstore = store->as_Mach();
 585       if (load_alias_idx != Compile::AliasIdxRaw) {
 586         // Check for call into the runtime using the Java calling
 587         // convention (and from there into a wrapper); it has no
 588         // _method.  Can't do this optimization for Native calls because
 589         // they CAN write to Java memory.
 590         if (mstore->ideal_Opcode() == Op_CallStaticJava) {
 591           assert(mstore->is_MachSafePoint(), "");
 592           MachSafePointNode* ms = (MachSafePointNode*) mstore;
 593           assert(ms->is_MachCallJava(), "");
 594           MachCallJavaNode* mcj = (MachCallJavaNode*) ms;
 595           if (mcj->_method == NULL) {
 596             // These runtime calls do not write to Java visible memory
 597             // (other than Raw) and so do not require anti-dependence edges.
 598             continue;
 599           }
 600         }
 601         // Same for SafePoints: they read/write Raw but only read otherwise.
 602         // This is basically a workaround for SafePoints only defining control
 603         // instead of control + memory.
 604         if (mstore->ideal_Opcode() == Op_SafePoint)
 605           continue;
 606       } else {
 607         // Some raw memory, such as the load of "top" at an allocation,
 608         // can be control dependent on the previous safepoint. See
 609         // comments in GraphKit::allocate_heap() about control input.
 610         // Inserting an anti-dep between such a safepoint and a use
 611         // creates a cycle, and will cause a subsequent failure in
 612         // local scheduling.  (BugId 4919904)
 613         // (%%% How can a control input be a safepoint and not a projection??)
 614         if (mstore->ideal_Opcode() == Op_SafePoint && load->in(0) == mstore)
 615           continue;
 616       }
 617     }
 618 
 619     // Identify a block that the current load must be above,
 620     // or else observe that 'store' is all the way up in the
 621     // earliest legal block for 'load'.  In the latter case,
 622     // immediately insert an anti-dependence edge.
 623     Block* store_block = get_block_for_node(store);
 624     assert(store_block != NULL, "unused killing projections skipped above");
 625 
 626     if (store->is_Phi()) {
 627       // 'load' uses memory which is one (or more) of the Phi's inputs.
 628       // It must be scheduled not before the Phi, but rather before
 629       // each of the relevant Phi inputs.
 630       //
 631       // Instead of finding the LCA of all inputs to a Phi that match 'mem',
 632       // we mark each corresponding predecessor block and do a combined
 633       // hoisting operation later (raise_LCA_above_marks).
 634       //
 635       // Do not assert(store_block != early, "Phi merging memory after access")
 636       // PhiNode may be at start of block 'early' with backedge to 'early'
 637       DEBUG_ONLY(bool found_match = false);
 638       for (uint j = PhiNode::Input, jmax = store->req(); j < jmax; j++) {
 639         if (store->in(j) == mem) {   // Found matching input?
 640           DEBUG_ONLY(found_match = true);
 641           Block* pred_block = get_block_for_node(store_block->pred(j));
 642           if (pred_block != early) {
 643             // If any predecessor of the Phi matches the load's "early block",
 644             // we do not need a precedence edge between the Phi and 'load'
 645             // since the load will be forced into a block preceding the Phi.
 646             pred_block->set_raise_LCA_mark(load_index);
 647             assert(!LCA_orig->dominates(pred_block) ||
 648                    early->dominates(pred_block), "early is high enough");
 649             must_raise_LCA = true;
 650           } else {
 651             // anti-dependent upon PHI pinned below 'early', no edge needed
 652             LCA = early;             // but can not schedule below 'early'
 653           }
 654         }
 655       }
 656       assert(found_match, "no worklist bug");
 657 #ifdef TRACK_PHI_INPUTS
 658 #ifdef ASSERT
 659       // This assert asks about correct handling of PhiNodes, which may not
 660       // have all input edges directly from 'mem'. See BugId 4621264
 661       int num_mem_inputs = phi_inputs.at_grow(store->_idx,0) + 1;
 662       // Increment by exactly one even if there are multiple copies of 'mem'
 663       // coming into the phi, because we will run this block several times
 664       // if there are several copies of 'mem'.  (That's how DU iterators work.)
 665       phi_inputs.at_put(store->_idx, num_mem_inputs);
 666       assert(PhiNode::Input + num_mem_inputs < store->req(),
 667              "Expect at least one phi input will not be from original memory state");
 668 #endif //ASSERT
 669 #endif //TRACK_PHI_INPUTS
 670     } else if (store_block != early) {
 671       // 'store' is between the current LCA and earliest possible block.
 672       // Label its block, and decide later on how to raise the LCA
 673       // to include the effect on LCA of this store.
 674       // If this store's block gets chosen as the raised LCA, we
 675       // will find him on the non_early_stores list and stick him
 676       // with a precedence edge.
 677       // (But, don't bother if LCA is already raised all the way.)
 678       if (LCA != early) {
 679         store_block->set_raise_LCA_mark(load_index);
 680         must_raise_LCA = true;
 681         non_early_stores.push(store);
 682       }
 683     } else {
 684       // Found a possibly-interfering store in the load's 'early' block.
 685       // This means 'load' cannot sink at all in the dominator tree.
 686       // Add an anti-dep edge, and squeeze 'load' into the highest block.
 687       assert(store != load->in(0), "dependence cycle found");
 688       if (verify) {
 689         assert(store->find_edge(load) != -1, "missing precedence edge");
 690       } else {
 691         store->add_prec(load);
 692       }
 693       LCA = early;
 694       // This turns off the process of gathering non_early_stores.
 695     }
 696   }
 697   // (Worklist is now empty; all nearby stores have been visited.)
 698 
 699   // Finished if 'load' must be scheduled in its 'early' block.
 700   // If we found any stores there, they have already been given
 701   // precedence edges.
 702   if (LCA == early)  return LCA;
 703 
 704   // We get here only if there are no possibly-interfering stores
 705   // in the load's 'early' block.  Move LCA up above all predecessors
 706   // which contain stores we have noted.
 707   //
 708   // The raised LCA block can be a home to such interfering stores,
 709   // but its predecessors must not contain any such stores.
 710   //
 711   // The raised LCA will be a lower bound for placing the load,
 712   // preventing the load from sinking past any block containing
 713   // a store that may invalidate the memory state required by 'load'.
 714   if (must_raise_LCA)
 715     LCA = raise_LCA_above_marks(LCA, load->_idx, early, this);
 716   if (LCA == early)  return LCA;
 717 
 718   // Insert anti-dependence edges from 'load' to each store
 719   // in the non-early LCA block.
 720   // Mine the non_early_stores list for such stores.
 721   if (LCA->raise_LCA_mark() == load_index) {
 722     while (non_early_stores.size() > 0) {
 723       Node* store = non_early_stores.pop();
 724       Block* store_block = get_block_for_node(store);
 725       if (store_block == LCA) {
 726         // add anti_dependence from store to load in its own block
 727         assert(store != load->in(0), "dependence cycle found");
 728         if (verify) {
 729           assert(store->find_edge(load) != -1, "missing precedence edge");
 730         } else {
 731           store->add_prec(load);
 732         }
 733       } else {
 734         assert(store_block->raise_LCA_mark() == load_index, "block was marked");
 735         // Any other stores we found must be either inside the new LCA
 736         // or else outside the original LCA.  In the latter case, they
 737         // did not interfere with any use of 'load'.
 738         assert(LCA->dominates(store_block)
 739                || !LCA_orig->dominates(store_block), "no stray stores");
 740       }
 741     }
 742   }
 743 
 744   // Return the highest block containing stores; any stores
 745   // within that block have been given anti-dependence edges.
 746   return LCA;
 747 }
 748 
 749 // This class is used to iterate backwards over the nodes in the graph.
 750 
 751 class Node_Backward_Iterator {
 752 
 753 private:
 754   Node_Backward_Iterator();
 755 
 756 public:
 757   // Constructor for the iterator
 758   Node_Backward_Iterator(Node *root, VectorSet &visited, Node_List &stack, PhaseCFG &cfg);
 759 
 760   // Postincrement operator to iterate over the nodes
 761   Node *next();
 762 
 763 private:
 764   VectorSet   &_visited;
 765   Node_List   &_stack;
 766   PhaseCFG &_cfg;
 767 };
 768 
 769 // Constructor for the Node_Backward_Iterator
 770 Node_Backward_Iterator::Node_Backward_Iterator( Node *root, VectorSet &visited, Node_List &stack, PhaseCFG &cfg)
 771   : _visited(visited), _stack(stack), _cfg(cfg) {
 772   // The stack should contain exactly the root
 773   stack.clear();
 774   stack.push(root);
 775 
 776   // Clear the visited bits
 777   visited.Clear();
 778 }
 779 
 780 // Iterator for the Node_Backward_Iterator
 781 Node *Node_Backward_Iterator::next() {
 782 
 783   // If the _stack is empty, then just return NULL: finished.
 784   if ( !_stack.size() )
 785     return NULL;
 786 
 787   // '_stack' is emulating a real _stack.  The 'visit-all-users' loop has been
 788   // made stateless, so I do not need to record the index 'i' on my _stack.
 789   // Instead I visit all users each time, scanning for unvisited users.
 790   // I visit unvisited not-anti-dependence users first, then anti-dependent
 791   // children next.
 792   Node *self = _stack.pop();
 793 
 794   // I cycle here when I am entering a deeper level of recursion.
 795   // The key variable 'self' was set prior to jumping here.
 796   while( 1 ) {
 797 
 798     _visited.set(self->_idx);
 799 
 800     // Now schedule all uses as late as possible.
 801     const Node* src = self->is_Proj() ? self->in(0) : self;
 802     uint src_rpo = _cfg.get_block_for_node(src)->_rpo;
 803 
 804     // Schedule all nodes in a post-order visit
 805     Node *unvisited = NULL;  // Unvisited anti-dependent Node, if any
 806 
 807     // Scan for unvisited nodes
 808     for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
 809       // For all uses, schedule late
 810       Node* n = self->fast_out(i); // Use
 811 
 812       // Skip already visited children
 813       if ( _visited.test(n->_idx) )
 814         continue;
 815 
 816       // do not traverse backward control edges
 817       Node *use = n->is_Proj() ? n->in(0) : n;
 818       uint use_rpo = _cfg.get_block_for_node(use)->_rpo;
 819 
 820       if ( use_rpo < src_rpo )
 821         continue;
 822 
 823       // Phi nodes always precede uses in a basic block
 824       if ( use_rpo == src_rpo && use->is_Phi() )
 825         continue;
 826 
 827       unvisited = n;      // Found unvisited
 828 
 829       // Check for possible-anti-dependent
 830       if( !n->needs_anti_dependence_check() )
 831         break;            // Not visited, not anti-dep; schedule it NOW
 832     }
 833 
 834     // Did I find an unvisited not-anti-dependent Node?
 835     if ( !unvisited )
 836       break;                  // All done with children; post-visit 'self'
 837 
 838     // Visit the unvisited Node.  Contains the obvious push to
 839     // indicate I'm entering a deeper level of recursion.  I push the
 840     // old state onto the _stack and set a new state and loop (recurse).
 841     _stack.push(self);
 842     self = unvisited;
 843   } // End recursion loop
 844 
 845   return self;
 846 }
 847 
 848 //------------------------------ComputeLatenciesBackwards----------------------
 849 // Compute the latency of all the instructions.
 850 void PhaseCFG::ComputeLatenciesBackwards(VectorSet &visited, Node_List &stack) {
 851 #ifndef PRODUCT
 852   if (trace_opto_pipelining())
 853     tty->print("\n#---- ComputeLatenciesBackwards ----\n");
 854 #endif
 855 
 856   Node_Backward_Iterator iter((Node *)_root, visited, stack, *this);
 857   Node *n;
 858 
 859   // Walk over all the nodes from last to first
 860   while (n = iter.next()) {
 861     // Set the latency for the definitions of this instruction
 862     partial_latency_of_defs(n);
 863   }
 864 } // end ComputeLatenciesBackwards
 865 
 866 //------------------------------partial_latency_of_defs------------------------
 867 // Compute the latency impact of this node on all defs.  This computes
 868 // a number that increases as we approach the beginning of the routine.
 869 void PhaseCFG::partial_latency_of_defs(Node *n) {
 870   // Set the latency for this instruction
 871 #ifndef PRODUCT
 872   if (trace_opto_pipelining()) {
 873     tty->print("# latency_to_inputs: node_latency[%d] = %d for node",
 874                n->_idx, _node_latency->at_grow(n->_idx));
 875     dump();
 876   }
 877 #endif
 878 
 879   if (n->is_Proj())
 880     n = n->in(0);
 881 
 882   if (n->is_Root())
 883     return;
 884 
 885   uint nlen = n->len();
 886   uint use_latency = _node_latency->at_grow(n->_idx);
 887   uint use_pre_order = get_block_for_node(n)->_pre_order;
 888 
 889   for ( uint j=0; j<nlen; j++ ) {
 890     Node *def = n->in(j);
 891 
 892     if (!def || def == n)
 893       continue;
 894 
 895     // Walk backwards thru projections
 896     if (def->is_Proj())
 897       def = def->in(0);
 898 
 899 #ifndef PRODUCT
 900     if (trace_opto_pipelining()) {
 901       tty->print("#    in(%2d): ", j);
 902       def->dump();
 903     }
 904 #endif
 905 
 906     // If the defining block is not known, assume it is ok
 907     Block *def_block = get_block_for_node(def);
 908     uint def_pre_order = def_block ? def_block->_pre_order : 0;
 909 
 910     if ( (use_pre_order <  def_pre_order) ||
 911          (use_pre_order == def_pre_order && n->is_Phi()) )
 912       continue;
 913 
 914     uint delta_latency = n->latency(j);
 915     uint current_latency = delta_latency + use_latency;
 916 
 917     if (_node_latency->at_grow(def->_idx) < current_latency) {
 918       _node_latency->at_put_grow(def->_idx, current_latency);
 919     }
 920 
 921 #ifndef PRODUCT
 922     if (trace_opto_pipelining()) {
 923       tty->print_cr("#      %d + edge_latency(%d) == %d -> %d, node_latency[%d] = %d",
 924                     use_latency, j, delta_latency, current_latency, def->_idx,
 925                     _node_latency->at_grow(def->_idx));
 926     }
 927 #endif
 928   }
 929 }
 930 
 931 //------------------------------latency_from_use-------------------------------
 932 // Compute the latency of a specific use
 933 int PhaseCFG::latency_from_use(Node *n, const Node *def, Node *use) {
 934   // If self-reference, return no latency
 935   if (use == n || use->is_Root()) {
 936     return 0;
 937   }
 938 
 939   uint def_pre_order = get_block_for_node(def)->_pre_order;
 940   uint latency = 0;
 941 
 942   // If the use is not a projection, then it is simple...
 943   if (!use->is_Proj()) {
 944 #ifndef PRODUCT
 945     if (trace_opto_pipelining()) {
 946       tty->print("#    out(): ");
 947       use->dump();
 948     }
 949 #endif
 950 
 951     uint use_pre_order = get_block_for_node(use)->_pre_order;
 952 
 953     if (use_pre_order < def_pre_order)
 954       return 0;
 955 
 956     if (use_pre_order == def_pre_order && use->is_Phi())
 957       return 0;
 958 
 959     uint nlen = use->len();
 960     uint nl = _node_latency->at_grow(use->_idx);
 961 
 962     for ( uint j=0; j<nlen; j++ ) {
 963       if (use->in(j) == n) {
 964         // Change this if we want local latencies
 965         uint ul = use->latency(j);
 966         uint  l = ul + nl;
 967         if (latency < l) latency = l;
 968 #ifndef PRODUCT
 969         if (trace_opto_pipelining()) {
 970           tty->print_cr("#      %d + edge_latency(%d) == %d -> %d, latency = %d",
 971                         nl, j, ul, l, latency);
 972         }
 973 #endif
 974       }
 975     }
 976   } else {
 977     // This is a projection, just grab the latency of the use(s)
 978     for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) {
 979       uint l = latency_from_use(use, def, use->fast_out(j));
 980       if (latency < l) latency = l;
 981     }
 982   }
 983 
 984   return latency;
 985 }
 986 
 987 //------------------------------latency_from_uses------------------------------
 988 // Compute the latency of this instruction relative to all of it's uses.
 989 // This computes a number that increases as we approach the beginning of the
 990 // routine.
 991 void PhaseCFG::latency_from_uses(Node *n) {
 992   // Set the latency for this instruction
 993 #ifndef PRODUCT
 994   if (trace_opto_pipelining()) {
 995     tty->print("# latency_from_outputs: node_latency[%d] = %d for node",
 996                n->_idx, _node_latency->at_grow(n->_idx));
 997     dump();
 998   }
 999 #endif
1000   uint latency=0;
1001   const Node *def = n->is_Proj() ? n->in(0): n;
1002 
1003   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1004     uint l = latency_from_use(n, def, n->fast_out(i));
1005 
1006     if (latency < l) latency = l;
1007   }
1008 
1009   _node_latency->at_put_grow(n->_idx, latency);
1010 }
1011 
1012 //------------------------------hoist_to_cheaper_block-------------------------
1013 // Pick a block for node self, between early and LCA, that is a cheaper
1014 // alternative to LCA.
1015 Block* PhaseCFG::hoist_to_cheaper_block(Block* LCA, Block* early, Node* self) {
1016   const double delta = 1+PROB_UNLIKELY_MAG(4);
1017   Block* least       = LCA;
1018   double least_freq  = least->_freq;
1019   uint target        = _node_latency->at_grow(self->_idx);
1020   uint start_latency = _node_latency->at_grow(LCA->_nodes[0]->_idx);
1021   uint end_latency   = _node_latency->at_grow(LCA->_nodes[LCA->end_idx()]->_idx);
1022   bool in_latency    = (target <= start_latency);
1023   const Block* root_block = get_block_for_node(_root);
1024 
1025   // Turn off latency scheduling if scheduling is just plain off
1026   if (!C->do_scheduling())
1027     in_latency = true;
1028 
1029   // Do not hoist (to cover latency) instructions which target a
1030   // single register.  Hoisting stretches the live range of the
1031   // single register and may force spilling.
1032   MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL;
1033   if (mach && mach->out_RegMask().is_bound1() && mach->out_RegMask().is_NotEmpty())
1034     in_latency = true;
1035 
1036 #ifndef PRODUCT
1037   if (trace_opto_pipelining()) {
1038     tty->print("# Find cheaper block for latency %d: ",
1039       _node_latency->at_grow(self->_idx));
1040     self->dump();
1041     tty->print_cr("#   B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1042       LCA->_pre_order,
1043       LCA->_nodes[0]->_idx,
1044       start_latency,
1045       LCA->_nodes[LCA->end_idx()]->_idx,
1046       end_latency,
1047       least_freq);
1048   }
1049 #endif
1050 
1051   int cand_cnt = 0;  // number of candidates tried
1052 
1053   // Walk up the dominator tree from LCA (Lowest common ancestor) to
1054   // the earliest legal location.  Capture the least execution frequency.
1055   while (LCA != early) {
1056     LCA = LCA->_idom;         // Follow up the dominator tree
1057 
1058     if (LCA == NULL) {
1059       // Bailout without retry
1060       C->record_method_not_compilable("late schedule failed: LCA == NULL");
1061       return least;
1062     }
1063 
1064     // Don't hoist machine instructions to the root basic block
1065     if (mach && LCA == root_block)
1066       break;
1067 
1068     uint start_lat = _node_latency->at_grow(LCA->_nodes[0]->_idx);
1069     uint end_idx   = LCA->end_idx();
1070     uint end_lat   = _node_latency->at_grow(LCA->_nodes[end_idx]->_idx);
1071     double LCA_freq = LCA->_freq;
1072 #ifndef PRODUCT
1073     if (trace_opto_pipelining()) {
1074       tty->print_cr("#   B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
1075         LCA->_pre_order, LCA->_nodes[0]->_idx, start_lat, end_idx, end_lat, LCA_freq);
1076     }
1077 #endif
1078     cand_cnt++;
1079     if (LCA_freq < least_freq              || // Better Frequency
1080         (StressGCM && Compile::randomized_select(cand_cnt)) || // Should be randomly accepted in stress mode
1081          (!StressGCM                    &&    // Otherwise, choose with latency
1082           !in_latency                   &&    // No block containing latency
1083           LCA_freq < least_freq * delta &&    // No worse frequency
1084           target >= end_lat             &&    // within latency range
1085           !self->is_iteratively_computed() )  // But don't hoist IV increments
1086              // because they may end up above other uses of their phi forcing
1087              // their result register to be different from their input.
1088        ) {
1089       least = LCA;            // Found cheaper block
1090       least_freq = LCA_freq;
1091       start_latency = start_lat;
1092       end_latency = end_lat;
1093       if (target <= start_lat)
1094         in_latency = true;
1095     }
1096   }
1097 
1098 #ifndef PRODUCT
1099   if (trace_opto_pipelining()) {
1100     tty->print_cr("#  Choose block B%d with start latency=%d and freq=%g",
1101       least->_pre_order, start_latency, least_freq);
1102   }
1103 #endif
1104 
1105   // See if the latency needs to be updated
1106   if (target < end_latency) {
1107 #ifndef PRODUCT
1108     if (trace_opto_pipelining()) {
1109       tty->print_cr("#  Change latency for [%4d] from %d to %d", self->_idx, target, end_latency);
1110     }
1111 #endif
1112     _node_latency->at_put_grow(self->_idx, end_latency);
1113     partial_latency_of_defs(self);
1114   }
1115 
1116   return least;
1117 }
1118 
1119 
1120 //------------------------------schedule_late-----------------------------------
1121 // Now schedule all codes as LATE as possible.  This is the LCA in the
1122 // dominator tree of all USES of a value.  Pick the block with the least
1123 // loop nesting depth that is lowest in the dominator tree.
1124 extern const char must_clone[];
1125 void PhaseCFG::schedule_late(VectorSet &visited, Node_List &stack) {
1126 #ifndef PRODUCT
1127   if (trace_opto_pipelining())
1128     tty->print("\n#---- schedule_late ----\n");
1129 #endif
1130 
1131   Node_Backward_Iterator iter((Node *)_root, visited, stack, *this);
1132   Node *self;
1133 
1134   // Walk over all the nodes from last to first
1135   while (self = iter.next()) {
1136     Block* early = get_block_for_node(self); // Earliest legal placement
1137 
1138     if (self->is_top()) {
1139       // Top node goes in bb #2 with other constants.
1140       // It must be special-cased, because it has no out edges.
1141       early->add_inst(self);
1142       continue;
1143     }
1144 
1145     // No uses, just terminate
1146     if (self->outcnt() == 0) {
1147       assert(self->is_MachProj(), "sanity");
1148       continue;                   // Must be a dead machine projection
1149     }
1150 
1151     // If node is pinned in the block, then no scheduling can be done.
1152     if( self->pinned() )          // Pinned in block?
1153       continue;
1154 
1155     MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL;
1156     if (mach) {
1157       switch (mach->ideal_Opcode()) {
1158       case Op_CreateEx:
1159         // Don't move exception creation
1160         early->add_inst(self);
1161         continue;
1162         break;
1163       case Op_CheckCastPP:
1164         // Don't move CheckCastPP nodes away from their input, if the input
1165         // is a rawptr (5071820).
1166         Node *def = self->in(1);
1167         if (def != NULL && def->bottom_type()->base() == Type::RawPtr) {
1168           early->add_inst(self);
1169 #ifdef ASSERT
1170           _raw_oops.push(def);
1171 #endif
1172           continue;
1173         }
1174         break;
1175       }
1176     }
1177 
1178     // Gather LCA of all uses
1179     Block *LCA = NULL;
1180     {
1181       for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
1182         // For all uses, find LCA
1183         Node* use = self->fast_out(i);
1184         LCA = raise_LCA_above_use(LCA, use, self, this);
1185       }
1186     }  // (Hide defs of imax, i from rest of block.)
1187 
1188     // Place temps in the block of their use.  This isn't a
1189     // requirement for correctness but it reduces useless
1190     // interference between temps and other nodes.
1191     if (mach != NULL && mach->is_MachTemp()) {
1192       map_node_to_block(self, LCA);
1193       LCA->add_inst(self);
1194       continue;
1195     }
1196 
1197     // Check if 'self' could be anti-dependent on memory
1198     if (self->needs_anti_dependence_check()) {
1199       // Hoist LCA above possible-defs and insert anti-dependences to
1200       // defs in new LCA block.
1201       LCA = insert_anti_dependences(LCA, self);
1202     }
1203 
1204     if (early->_dom_depth > LCA->_dom_depth) {
1205       // Somehow the LCA has moved above the earliest legal point.
1206       // (One way this can happen is via memory_early_block.)
1207       if (C->subsume_loads() == true && !C->failing()) {
1208         // Retry with subsume_loads == false
1209         // If this is the first failure, the sentinel string will "stick"
1210         // to the Compile object, and the C2Compiler will see it and retry.
1211         C->record_failure(C2Compiler::retry_no_subsuming_loads());
1212       } else {
1213         // Bailout without retry when (early->_dom_depth > LCA->_dom_depth)
1214         C->record_method_not_compilable("late schedule failed: incorrect graph");
1215       }
1216       return;
1217     }
1218 
1219     // If there is no opportunity to hoist, then we're done.
1220     // In stress mode, try to hoist even the single operations.
1221     bool try_to_hoist = StressGCM || (LCA != early);
1222 
1223     // Must clone guys stay next to use; no hoisting allowed.
1224     // Also cannot hoist guys that alter memory or are otherwise not
1225     // allocatable (hoisting can make a value live longer, leading to
1226     // anti and output dependency problems which are normally resolved
1227     // by the register allocator giving everyone a different register).
1228     if (mach != NULL && must_clone[mach->ideal_Opcode()])
1229       try_to_hoist = false;
1230 
1231     Block* late = NULL;
1232     if (try_to_hoist) {
1233       // Now find the block with the least execution frequency.
1234       // Start at the latest schedule and work up to the earliest schedule
1235       // in the dominator tree.  Thus the Node will dominate all its uses.
1236       late = hoist_to_cheaper_block(LCA, early, self);
1237     } else {
1238       // Just use the LCA of the uses.
1239       late = LCA;
1240     }
1241 
1242     // Put the node into target block
1243     schedule_node_into_block(self, late);
1244 
1245 #ifdef ASSERT
1246     if (self->needs_anti_dependence_check()) {
1247       // since precedence edges are only inserted when we're sure they
1248       // are needed make sure that after placement in a block we don't
1249       // need any new precedence edges.
1250       verify_anti_dependences(late, self);
1251     }
1252 #endif
1253   } // Loop until all nodes have been visited
1254 
1255 } // end ScheduleLate
1256 
1257 //------------------------------GlobalCodeMotion-------------------------------
1258 void PhaseCFG::GlobalCodeMotion( Matcher &matcher, uint unique, Node_List &proj_list ) {
1259   ResourceMark rm;
1260 
1261 #ifndef PRODUCT
1262   if (trace_opto_pipelining()) {
1263     tty->print("\n---- Start GlobalCodeMotion ----\n");
1264   }
1265 #endif
1266 
1267   // Initialize the node to block mapping for things on the proj_list
1268   for (uint i = 0; i < proj_list.size(); i++) {
1269     unmap_node_from_block(proj_list[i]);
1270   }
1271 
1272   // Set the basic block for Nodes pinned into blocks
1273   Arena *a = Thread::current()->resource_area();
1274   VectorSet visited(a);
1275   schedule_pinned_nodes( visited );
1276 
1277   // Find the earliest Block any instruction can be placed in.  Some
1278   // instructions are pinned into Blocks.  Unpinned instructions can
1279   // appear in last block in which all their inputs occur.
1280   visited.Clear();
1281   Node_List stack(a);
1282   stack.map( (unique >> 1) + 16, NULL); // Pre-grow the list
1283   if (!schedule_early(visited, stack)) {
1284     // Bailout without retry
1285     C->record_method_not_compilable("early schedule failed");
1286     return;
1287   }
1288 
1289   // Build Def-Use edges.
1290   proj_list.push(_root);        // Add real root as another root
1291   proj_list.pop();
1292 
1293   // Compute the latency information (via backwards walk) for all the
1294   // instructions in the graph
1295   _node_latency = new GrowableArray<uint>(); // resource_area allocation
1296 
1297   if( C->do_scheduling() )
1298     ComputeLatenciesBackwards(visited, stack);
1299 
1300   // Now schedule all codes as LATE as possible.  This is the LCA in the
1301   // dominator tree of all USES of a value.  Pick the block with the least
1302   // loop nesting depth that is lowest in the dominator tree.
1303   // ( visited.Clear() called in schedule_late()->Node_Backward_Iterator() )
1304   schedule_late(visited, stack);
1305   if( C->failing() ) {
1306     // schedule_late fails only when graph is incorrect.
1307     assert(!VerifyGraphEdges, "verification should have failed");
1308     return;
1309   }
1310 
1311   unique = C->unique();
1312 
1313 #ifndef PRODUCT
1314   if (trace_opto_pipelining()) {
1315     tty->print("\n---- Detect implicit null checks ----\n");
1316   }
1317 #endif
1318 
1319   // Detect implicit-null-check opportunities.  Basically, find NULL checks
1320   // with suitable memory ops nearby.  Use the memory op to do the NULL check.
1321   // I can generate a memory op if there is not one nearby.
1322   if (C->is_method_compilation()) {
1323     // Don't do it for natives, adapters, or runtime stubs
1324     int allowed_reasons = 0;
1325     // ...and don't do it when there have been too many traps, globally.
1326     for (int reason = (int)Deoptimization::Reason_none+1;
1327          reason < Compile::trapHistLength; reason++) {
1328       assert(reason < BitsPerInt, "recode bit map");
1329       if (!C->too_many_traps((Deoptimization::DeoptReason) reason))
1330         allowed_reasons |= nth_bit(reason);
1331     }
1332     // By reversing the loop direction we get a very minor gain on mpegaudio.
1333     // Feel free to revert to a forward loop for clarity.
1334     // for( int i=0; i < (int)matcher._null_check_tests.size(); i+=2 ) {
1335     for( int i= matcher._null_check_tests.size()-2; i>=0; i-=2 ) {
1336       Node *proj = matcher._null_check_tests[i  ];
1337       Node *val  = matcher._null_check_tests[i+1];
1338       get_block_for_node(proj)->implicit_null_check(this, proj, val, allowed_reasons);
1339       // The implicit_null_check will only perform the transformation
1340       // if the null branch is truly uncommon, *and* it leads to an
1341       // uncommon trap.  Combined with the too_many_traps guards
1342       // above, this prevents SEGV storms reported in 6366351,
1343       // by recompiling offending methods without this optimization.
1344     }
1345   }
1346 
1347 #ifndef PRODUCT
1348   if (trace_opto_pipelining()) {
1349     tty->print("\n---- Start Local Scheduling ----\n");
1350   }
1351 #endif
1352 
1353   // Schedule locally.  Right now a simple topological sort.
1354   // Later, do a real latency aware scheduler.
1355   uint max_idx = C->unique();
1356   GrowableArray<int> ready_cnt(max_idx, max_idx, -1);
1357   visited.Clear();
1358   for (uint i = 0; i < _num_blocks; i++) {
1359     if (!_blocks[i]->schedule_local(this, matcher, ready_cnt, visited)) {
1360       if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
1361         C->record_method_not_compilable("local schedule failed");
1362       }
1363       return;
1364     }
1365   }
1366 
1367   // If we inserted any instructions between a Call and his CatchNode,
1368   // clone the instructions on all paths below the Catch.
1369   for (uint i = 0; i < _num_blocks; i++) {
1370     _blocks[i]->call_catch_cleanup(this, C);
1371   }
1372 
1373 #ifndef PRODUCT
1374   if (trace_opto_pipelining()) {
1375     tty->print("\n---- After GlobalCodeMotion ----\n");
1376     for (uint i = 0; i < _num_blocks; i++) {
1377       _blocks[i]->dump();
1378     }
1379   }
1380 #endif
1381   // Dead.
1382   _node_latency = (GrowableArray<uint> *)0xdeadbeef;
1383 }
1384 
1385 
1386 //------------------------------Estimate_Block_Frequency-----------------------
1387 // Estimate block frequencies based on IfNode probabilities.
1388 void PhaseCFG::Estimate_Block_Frequency() {
1389 
1390   // Force conditional branches leading to uncommon traps to be unlikely,
1391   // not because we get to the uncommon_trap with less relative frequency,
1392   // but because an uncommon_trap typically causes a deopt, so we only get
1393   // there once.
1394   if (C->do_freq_based_layout()) {
1395     Block_List worklist;
1396     Block* root_blk = _blocks[0];
1397     for (uint i = 1; i < root_blk->num_preds(); i++) {
1398       Block *pb = get_block_for_node(root_blk->pred(i));
1399       if (pb->has_uncommon_code()) {
1400         worklist.push(pb);
1401       }
1402     }
1403     while (worklist.size() > 0) {
1404       Block* uct = worklist.pop();
1405       if (uct == _broot) continue;
1406       for (uint i = 1; i < uct->num_preds(); i++) {
1407         Block *pb = get_block_for_node(uct->pred(i));
1408         if (pb->_num_succs == 1) {
1409           worklist.push(pb);
1410         } else if (pb->num_fall_throughs() == 2) {
1411           pb->update_uncommon_branch(uct);
1412         }
1413       }
1414     }
1415   }
1416 
1417   // Create the loop tree and calculate loop depth.
1418   _root_loop = create_loop_tree();
1419   _root_loop->compute_loop_depth(0);
1420 
1421   // Compute block frequency of each block, relative to a single loop entry.
1422   _root_loop->compute_freq();
1423 
1424   // Adjust all frequencies to be relative to a single method entry
1425   _root_loop->_freq = 1.0;
1426   _root_loop->scale_freq();
1427 
1428   // Save outmost loop frequency for LRG frequency threshold
1429   _outer_loop_freq = _root_loop->outer_loop_freq();
1430 
1431   // force paths ending at uncommon traps to be infrequent
1432   if (!C->do_freq_based_layout()) {
1433     Block_List worklist;
1434     Block* root_blk = _blocks[0];
1435     for (uint i = 1; i < root_blk->num_preds(); i++) {
1436       Block *pb = get_block_for_node(root_blk->pred(i));
1437       if (pb->has_uncommon_code()) {
1438         worklist.push(pb);
1439       }
1440     }
1441     while (worklist.size() > 0) {
1442       Block* uct = worklist.pop();
1443       uct->_freq = PROB_MIN;
1444       for (uint i = 1; i < uct->num_preds(); i++) {
1445         Block *pb = get_block_for_node(uct->pred(i));
1446         if (pb->_num_succs == 1 && pb->_freq > PROB_MIN) {
1447           worklist.push(pb);
1448         }
1449       }
1450     }
1451   }
1452 
1453 #ifdef ASSERT
1454   for (uint i = 0; i < _num_blocks; i++ ) {
1455     Block *b = _blocks[i];
1456     assert(b->_freq >= MIN_BLOCK_FREQUENCY, "Register Allocator requires meaningful block frequency");
1457   }
1458 #endif
1459 
1460 #ifndef PRODUCT
1461   if (PrintCFGBlockFreq) {
1462     tty->print_cr("CFG Block Frequencies");
1463     _root_loop->dump_tree();
1464     if (Verbose) {
1465       tty->print_cr("PhaseCFG dump");
1466       dump();
1467       tty->print_cr("Node dump");
1468       _root->dump(99999);
1469     }
1470   }
1471 #endif
1472 }
1473 
1474 //----------------------------create_loop_tree--------------------------------
1475 // Create a loop tree from the CFG
1476 CFGLoop* PhaseCFG::create_loop_tree() {
1477 
1478 #ifdef ASSERT
1479   assert( _blocks[0] == _broot, "" );
1480   for (uint i = 0; i < _num_blocks; i++ ) {
1481     Block *b = _blocks[i];
1482     // Check that _loop field are clear...we could clear them if not.
1483     assert(b->_loop == NULL, "clear _loop expected");
1484     // Sanity check that the RPO numbering is reflected in the _blocks array.
1485     // It doesn't have to be for the loop tree to be built, but if it is not,
1486     // then the blocks have been reordered since dom graph building...which
1487     // may question the RPO numbering
1488     assert(b->_rpo == i, "unexpected reverse post order number");
1489   }
1490 #endif
1491 
1492   int idct = 0;
1493   CFGLoop* root_loop = new CFGLoop(idct++);
1494 
1495   Block_List worklist;
1496 
1497   // Assign blocks to loops
1498   for(uint i = _num_blocks - 1; i > 0; i-- ) { // skip Root block
1499     Block *b = _blocks[i];
1500 
1501     if (b->head()->is_Loop()) {
1502       Block* loop_head = b;
1503       assert(loop_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1504       Node* tail_n = loop_head->pred(LoopNode::LoopBackControl);
1505       Block* tail = get_block_for_node(tail_n);
1506 
1507       // Defensively filter out Loop nodes for non-single-entry loops.
1508       // For all reasonable loops, the head occurs before the tail in RPO.
1509       if (i <= tail->_rpo) {
1510 
1511         // The tail and (recursive) predecessors of the tail
1512         // are made members of a new loop.
1513 
1514         assert(worklist.size() == 0, "nonempty worklist");
1515         CFGLoop* nloop = new CFGLoop(idct++);
1516         assert(loop_head->_loop == NULL, "just checking");
1517         loop_head->_loop = nloop;
1518         // Add to nloop so push_pred() will skip over inner loops
1519         nloop->add_member(loop_head);
1520         nloop->push_pred(loop_head, LoopNode::LoopBackControl, worklist, this);
1521 
1522         while (worklist.size() > 0) {
1523           Block* member = worklist.pop();
1524           if (member != loop_head) {
1525             for (uint j = 1; j < member->num_preds(); j++) {
1526               nloop->push_pred(member, j, worklist, this);
1527             }
1528           }
1529         }
1530       }
1531     }
1532   }
1533 
1534   // Create a member list for each loop consisting
1535   // of both blocks and (immediate child) loops.
1536   for (uint i = 0; i < _num_blocks; i++) {
1537     Block *b = _blocks[i];
1538     CFGLoop* lp = b->_loop;
1539     if (lp == NULL) {
1540       // Not assigned to a loop. Add it to the method's pseudo loop.
1541       b->_loop = root_loop;
1542       lp = root_loop;
1543     }
1544     if (lp == root_loop || b != lp->head()) { // loop heads are already members
1545       lp->add_member(b);
1546     }
1547     if (lp != root_loop) {
1548       if (lp->parent() == NULL) {
1549         // Not a nested loop. Make it a child of the method's pseudo loop.
1550         root_loop->add_nested_loop(lp);
1551       }
1552       if (b == lp->head()) {
1553         // Add nested loop to member list of parent loop.
1554         lp->parent()->add_member(lp);
1555       }
1556     }
1557   }
1558 
1559   return root_loop;
1560 }
1561 
1562 //------------------------------push_pred--------------------------------------
1563 void CFGLoop::push_pred(Block* blk, int i, Block_List& worklist, PhaseCFG* cfg) {
1564   Node* pred_n = blk->pred(i);
1565   Block* pred = cfg->get_block_for_node(pred_n);
1566   CFGLoop *pred_loop = pred->_loop;
1567   if (pred_loop == NULL) {
1568     // Filter out blocks for non-single-entry loops.
1569     // For all reasonable loops, the head occurs before the tail in RPO.
1570     if (pred->_rpo > head()->_rpo) {
1571       pred->_loop = this;
1572       worklist.push(pred);
1573     }
1574   } else if (pred_loop != this) {
1575     // Nested loop.
1576     while (pred_loop->_parent != NULL && pred_loop->_parent != this) {
1577       pred_loop = pred_loop->_parent;
1578     }
1579     // Make pred's loop be a child
1580     if (pred_loop->_parent == NULL) {
1581       add_nested_loop(pred_loop);
1582       // Continue with loop entry predecessor.
1583       Block* pred_head = pred_loop->head();
1584       assert(pred_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
1585       assert(pred_head != head(), "loop head in only one loop");
1586       push_pred(pred_head, LoopNode::EntryControl, worklist, cfg);
1587     } else {
1588       assert(pred_loop->_parent == this && _parent == NULL, "just checking");
1589     }
1590   }
1591 }
1592 
1593 //------------------------------add_nested_loop--------------------------------
1594 // Make cl a child of the current loop in the loop tree.
1595 void CFGLoop::add_nested_loop(CFGLoop* cl) {
1596   assert(_parent == NULL, "no parent yet");
1597   assert(cl != this, "not my own parent");
1598   cl->_parent = this;
1599   CFGLoop* ch = _child;
1600   if (ch == NULL) {
1601     _child = cl;
1602   } else {
1603     while (ch->_sibling != NULL) { ch = ch->_sibling; }
1604     ch->_sibling = cl;
1605   }
1606 }
1607 
1608 //------------------------------compute_loop_depth-----------------------------
1609 // Store the loop depth in each CFGLoop object.
1610 // Recursively walk the children to do the same for them.
1611 void CFGLoop::compute_loop_depth(int depth) {
1612   _depth = depth;
1613   CFGLoop* ch = _child;
1614   while (ch != NULL) {
1615     ch->compute_loop_depth(depth + 1);
1616     ch = ch->_sibling;
1617   }
1618 }
1619 
1620 //------------------------------compute_freq-----------------------------------
1621 // Compute the frequency of each block and loop, relative to a single entry
1622 // into the dominating loop head.
1623 void CFGLoop::compute_freq() {
1624   // Bottom up traversal of loop tree (visit inner loops first.)
1625   // Set loop head frequency to 1.0, then transitively
1626   // compute frequency for all successors in the loop,
1627   // as well as for each exit edge.  Inner loops are
1628   // treated as single blocks with loop exit targets
1629   // as the successor blocks.
1630 
1631   // Nested loops first
1632   CFGLoop* ch = _child;
1633   while (ch != NULL) {
1634     ch->compute_freq();
1635     ch = ch->_sibling;
1636   }
1637   assert (_members.length() > 0, "no empty loops");
1638   Block* hd = head();
1639   hd->_freq = 1.0f;
1640   for (int i = 0; i < _members.length(); i++) {
1641     CFGElement* s = _members.at(i);
1642     float freq = s->_freq;
1643     if (s->is_block()) {
1644       Block* b = s->as_Block();
1645       for (uint j = 0; j < b->_num_succs; j++) {
1646         Block* sb = b->_succs[j];
1647         update_succ_freq(sb, freq * b->succ_prob(j));
1648       }
1649     } else {
1650       CFGLoop* lp = s->as_CFGLoop();
1651       assert(lp->_parent == this, "immediate child");
1652       for (int k = 0; k < lp->_exits.length(); k++) {
1653         Block* eb = lp->_exits.at(k).get_target();
1654         float prob = lp->_exits.at(k).get_prob();
1655         update_succ_freq(eb, freq * prob);
1656       }
1657     }
1658   }
1659 
1660   // For all loops other than the outer, "method" loop,
1661   // sum and normalize the exit probability. The "method" loop
1662   // should keep the initial exit probability of 1, so that
1663   // inner blocks do not get erroneously scaled.
1664   if (_depth != 0) {
1665     // Total the exit probabilities for this loop.
1666     float exits_sum = 0.0f;
1667     for (int i = 0; i < _exits.length(); i++) {
1668       exits_sum += _exits.at(i).get_prob();
1669     }
1670 
1671     // Normalize the exit probabilities. Until now, the
1672     // probabilities estimate the possibility of exit per
1673     // a single loop iteration; afterward, they estimate
1674     // the probability of exit per loop entry.
1675     for (int i = 0; i < _exits.length(); i++) {
1676       Block* et = _exits.at(i).get_target();
1677       float new_prob = 0.0f;
1678       if (_exits.at(i).get_prob() > 0.0f) {
1679         new_prob = _exits.at(i).get_prob() / exits_sum;
1680       }
1681       BlockProbPair bpp(et, new_prob);
1682       _exits.at_put(i, bpp);
1683     }
1684 
1685     // Save the total, but guard against unreasonable probability,
1686     // as the value is used to estimate the loop trip count.
1687     // An infinite trip count would blur relative block
1688     // frequencies.
1689     if (exits_sum > 1.0f) exits_sum = 1.0;
1690     if (exits_sum < PROB_MIN) exits_sum = PROB_MIN;
1691     _exit_prob = exits_sum;
1692   }
1693 }
1694 
1695 //------------------------------succ_prob-------------------------------------
1696 // Determine the probability of reaching successor 'i' from the receiver block.
1697 float Block::succ_prob(uint i) {
1698   int eidx = end_idx();
1699   Node *n = _nodes[eidx];  // Get ending Node
1700 
1701   int op = n->Opcode();
1702   if (n->is_Mach()) {
1703     if (n->is_MachNullCheck()) {
1704       // Can only reach here if called after lcm. The original Op_If is gone,
1705       // so we attempt to infer the probability from one or both of the
1706       // successor blocks.
1707       assert(_num_succs == 2, "expecting 2 successors of a null check");
1708       // If either successor has only one predecessor, then the
1709       // probability estimate can be derived using the
1710       // relative frequency of the successor and this block.
1711       if (_succs[i]->num_preds() == 2) {
1712         return _succs[i]->_freq / _freq;
1713       } else if (_succs[1-i]->num_preds() == 2) {
1714         return 1 - (_succs[1-i]->_freq / _freq);
1715       } else {
1716         // Estimate using both successor frequencies
1717         float freq = _succs[i]->_freq;
1718         return freq / (freq + _succs[1-i]->_freq);
1719       }
1720     }
1721     op = n->as_Mach()->ideal_Opcode();
1722   }
1723 
1724 
1725   // Switch on branch type
1726   switch( op ) {
1727   case Op_CountedLoopEnd:
1728   case Op_If: {
1729     assert (i < 2, "just checking");
1730     // Conditionals pass on only part of their frequency
1731     float prob  = n->as_MachIf()->_prob;
1732     assert(prob >= 0.0 && prob <= 1.0, "out of range probability");
1733     // If succ[i] is the FALSE branch, invert path info
1734     if( _nodes[i + eidx + 1]->Opcode() == Op_IfFalse ) {
1735       return 1.0f - prob; // not taken
1736     } else {
1737       return prob; // taken
1738     }
1739   }
1740 
1741   case Op_Jump:
1742     // Divide the frequency between all successors evenly
1743     return 1.0f/_num_succs;
1744 
1745   case Op_Catch: {
1746     const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj();
1747     if (ci->_con == CatchProjNode::fall_through_index) {
1748       // Fall-thru path gets the lion's share.
1749       return 1.0f - PROB_UNLIKELY_MAG(5)*_num_succs;
1750     } else {
1751       // Presume exceptional paths are equally unlikely
1752       return PROB_UNLIKELY_MAG(5);
1753     }
1754   }
1755 
1756   case Op_Root:
1757   case Op_Goto:
1758     // Pass frequency straight thru to target
1759     return 1.0f;
1760 
1761   case Op_NeverBranch:
1762     return 0.0f;
1763 
1764   case Op_TailCall:
1765   case Op_TailJump:
1766   case Op_Return:
1767   case Op_Halt:
1768   case Op_Rethrow:
1769     // Do not push out freq to root block
1770     return 0.0f;
1771 
1772   default:
1773     ShouldNotReachHere();
1774   }
1775 
1776   return 0.0f;
1777 }
1778 
1779 //------------------------------num_fall_throughs-----------------------------
1780 // Return the number of fall-through candidates for a block
1781 int Block::num_fall_throughs() {
1782   int eidx = end_idx();
1783   Node *n = _nodes[eidx];  // Get ending Node
1784 
1785   int op = n->Opcode();
1786   if (n->is_Mach()) {
1787     if (n->is_MachNullCheck()) {
1788       // In theory, either side can fall-thru, for simplicity sake,
1789       // let's say only the false branch can now.
1790       return 1;
1791     }
1792     op = n->as_Mach()->ideal_Opcode();
1793   }
1794 
1795   // Switch on branch type
1796   switch( op ) {
1797   case Op_CountedLoopEnd:
1798   case Op_If:
1799     return 2;
1800 
1801   case Op_Root:
1802   case Op_Goto:
1803     return 1;
1804 
1805   case Op_Catch: {
1806     for (uint i = 0; i < _num_succs; i++) {
1807       const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj();
1808       if (ci->_con == CatchProjNode::fall_through_index) {
1809         return 1;
1810       }
1811     }
1812     return 0;
1813   }
1814 
1815   case Op_Jump:
1816   case Op_NeverBranch:
1817   case Op_TailCall:
1818   case Op_TailJump:
1819   case Op_Return:
1820   case Op_Halt:
1821   case Op_Rethrow:
1822     return 0;
1823 
1824   default:
1825     ShouldNotReachHere();
1826   }
1827 
1828   return 0;
1829 }
1830 
1831 //------------------------------succ_fall_through-----------------------------
1832 // Return true if a specific successor could be fall-through target.
1833 bool Block::succ_fall_through(uint i) {
1834   int eidx = end_idx();
1835   Node *n = _nodes[eidx];  // Get ending Node
1836 
1837   int op = n->Opcode();
1838   if (n->is_Mach()) {
1839     if (n->is_MachNullCheck()) {
1840       // In theory, either side can fall-thru, for simplicity sake,
1841       // let's say only the false branch can now.
1842       return _nodes[i + eidx + 1]->Opcode() == Op_IfFalse;
1843     }
1844     op = n->as_Mach()->ideal_Opcode();
1845   }
1846 
1847   // Switch on branch type
1848   switch( op ) {
1849   case Op_CountedLoopEnd:
1850   case Op_If:
1851   case Op_Root:
1852   case Op_Goto:
1853     return true;
1854 
1855   case Op_Catch: {
1856     const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj();
1857     return ci->_con == CatchProjNode::fall_through_index;
1858   }
1859 
1860   case Op_Jump:
1861   case Op_NeverBranch:
1862   case Op_TailCall:
1863   case Op_TailJump:
1864   case Op_Return:
1865   case Op_Halt:
1866   case Op_Rethrow:
1867     return false;
1868 
1869   default:
1870     ShouldNotReachHere();
1871   }
1872 
1873   return false;
1874 }
1875 
1876 //------------------------------update_uncommon_branch------------------------
1877 // Update the probability of a two-branch to be uncommon
1878 void Block::update_uncommon_branch(Block* ub) {
1879   int eidx = end_idx();
1880   Node *n = _nodes[eidx];  // Get ending Node
1881 
1882   int op = n->as_Mach()->ideal_Opcode();
1883 
1884   assert(op == Op_CountedLoopEnd || op == Op_If, "must be a If");
1885   assert(num_fall_throughs() == 2, "must be a two way branch block");
1886 
1887   // Which successor is ub?
1888   uint s;
1889   for (s = 0; s <_num_succs; s++) {
1890     if (_succs[s] == ub) break;
1891   }
1892   assert(s < 2, "uncommon successor must be found");
1893 
1894   // If ub is the true path, make the proability small, else
1895   // ub is the false path, and make the probability large
1896   bool invert = (_nodes[s + eidx + 1]->Opcode() == Op_IfFalse);
1897 
1898   // Get existing probability
1899   float p = n->as_MachIf()->_prob;
1900 
1901   if (invert) p = 1.0 - p;
1902   if (p > PROB_MIN) {
1903     p = PROB_MIN;
1904   }
1905   if (invert) p = 1.0 - p;
1906 
1907   n->as_MachIf()->_prob = p;
1908 }
1909 
1910 //------------------------------update_succ_freq-------------------------------
1911 // Update the appropriate frequency associated with block 'b', a successor of
1912 // a block in this loop.
1913 void CFGLoop::update_succ_freq(Block* b, float freq) {
1914   if (b->_loop == this) {
1915     if (b == head()) {
1916       // back branch within the loop
1917       // Do nothing now, the loop carried frequency will be
1918       // adjust later in scale_freq().
1919     } else {
1920       // simple branch within the loop
1921       b->_freq += freq;
1922     }
1923   } else if (!in_loop_nest(b)) {
1924     // branch is exit from this loop
1925     BlockProbPair bpp(b, freq);
1926     _exits.append(bpp);
1927   } else {
1928     // branch into nested loop
1929     CFGLoop* ch = b->_loop;
1930     ch->_freq += freq;
1931   }
1932 }
1933 
1934 //------------------------------in_loop_nest-----------------------------------
1935 // Determine if block b is in the receiver's loop nest.
1936 bool CFGLoop::in_loop_nest(Block* b) {
1937   int depth = _depth;
1938   CFGLoop* b_loop = b->_loop;
1939   int b_depth = b_loop->_depth;
1940   if (depth == b_depth) {
1941     return true;
1942   }
1943   while (b_depth > depth) {
1944     b_loop = b_loop->_parent;
1945     b_depth = b_loop->_depth;
1946   }
1947   return b_loop == this;
1948 }
1949 
1950 //------------------------------scale_freq-------------------------------------
1951 // Scale frequency of loops and blocks by trip counts from outer loops
1952 // Do a top down traversal of loop tree (visit outer loops first.)
1953 void CFGLoop::scale_freq() {
1954   float loop_freq = _freq * trip_count();
1955   _freq = loop_freq;
1956   for (int i = 0; i < _members.length(); i++) {
1957     CFGElement* s = _members.at(i);
1958     float block_freq = s->_freq * loop_freq;
1959     if (g_isnan(block_freq) || block_freq < MIN_BLOCK_FREQUENCY)
1960       block_freq = MIN_BLOCK_FREQUENCY;
1961     s->_freq = block_freq;
1962   }
1963   CFGLoop* ch = _child;
1964   while (ch != NULL) {
1965     ch->scale_freq();
1966     ch = ch->_sibling;
1967   }
1968 }
1969 
1970 // Frequency of outer loop
1971 float CFGLoop::outer_loop_freq() const {
1972   if (_child != NULL) {
1973     return _child->_freq;
1974   }
1975   return _freq;
1976 }
1977 
1978 #ifndef PRODUCT
1979 //------------------------------dump_tree--------------------------------------
1980 void CFGLoop::dump_tree() const {
1981   dump();
1982   if (_child != NULL)   _child->dump_tree();
1983   if (_sibling != NULL) _sibling->dump_tree();
1984 }
1985 
1986 //------------------------------dump-------------------------------------------
1987 void CFGLoop::dump() const {
1988   for (int i = 0; i < _depth; i++) tty->print("   ");
1989   tty->print("%s: %d  trip_count: %6.0f freq: %6.0f\n",
1990              _depth == 0 ? "Method" : "Loop", _id, trip_count(), _freq);
1991   for (int i = 0; i < _depth; i++) tty->print("   ");
1992   tty->print("         members:", _id);
1993   int k = 0;
1994   for (int i = 0; i < _members.length(); i++) {
1995     if (k++ >= 6) {
1996       tty->print("\n              ");
1997       for (int j = 0; j < _depth+1; j++) tty->print("   ");
1998       k = 0;
1999     }
2000     CFGElement *s = _members.at(i);
2001     if (s->is_block()) {
2002       Block *b = s->as_Block();
2003       tty->print(" B%d(%6.3f)", b->_pre_order, b->_freq);
2004     } else {
2005       CFGLoop* lp = s->as_CFGLoop();
2006       tty->print(" L%d(%6.3f)", lp->_id, lp->_freq);
2007     }
2008   }
2009   tty->print("\n");
2010   for (int i = 0; i < _depth; i++) tty->print("   ");
2011   tty->print("         exits:  ");
2012   k = 0;
2013   for (int i = 0; i < _exits.length(); i++) {
2014     if (k++ >= 7) {
2015       tty->print("\n              ");
2016       for (int j = 0; j < _depth+1; j++) tty->print("   ");
2017       k = 0;
2018     }
2019     Block *blk = _exits.at(i).get_target();
2020     float prob = _exits.at(i).get_prob();
2021     tty->print(" ->%d@%d%%", blk->_pre_order, (int)(prob*100));
2022   }
2023   tty->print("\n");
2024 }
2025 #endif