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