1 /*
   2  * Copyright (c) 2007, 2018, 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 #include "precompiled.hpp"
  25 #include "compiler/compileLog.hpp"
  26 #include "libadt/vectset.hpp"
  27 #include "memory/allocation.inline.hpp"
  28 #include "memory/resourceArea.hpp"
  29 #include "opto/addnode.hpp"
  30 #include "opto/callnode.hpp"
  31 #include "opto/castnode.hpp"
  32 #include "opto/convertnode.hpp"
  33 #include "opto/divnode.hpp"
  34 #include "opto/matcher.hpp"
  35 #include "opto/memnode.hpp"
  36 #include "opto/mulnode.hpp"
  37 #include "opto/opcodes.hpp"
  38 #include "opto/opaquenode.hpp"
  39 #include "opto/superword.hpp"
  40 #include "opto/vectornode.hpp"
  41 #include "opto/movenode.hpp"
  42 
  43 //
  44 //                  S U P E R W O R D   T R A N S F O R M
  45 //=============================================================================
  46 
  47 //------------------------------SuperWord---------------------------
  48 SuperWord::SuperWord(PhaseIdealLoop* phase) :
  49   _phase(phase),
  50   _igvn(phase->_igvn),
  51   _arena(phase->C->comp_arena()),
  52   _packset(arena(), 8,  0, NULL),         // packs for the current block
  53   _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
  54   _block(arena(), 8,  0, NULL),           // nodes in current block
  55   _post_block(arena(), 8, 0, NULL),       // nodes common to current block which are marked as post loop vectorizable
  56   _data_entry(arena(), 8,  0, NULL),      // nodes with all inputs from outside
  57   _mem_slice_head(arena(), 8,  0, NULL),  // memory slice heads
  58   _mem_slice_tail(arena(), 8,  0, NULL),  // memory slice tails
  59   _node_info(arena(), 8,  0, SWNodeInfo::initial), // info needed per node
  60   _clone_map(phase->C->clone_map()),      // map of nodes created in cloning
  61   _cmovev_kit(_arena, this),              // map to facilitate CMoveV creation
  62   _align_to_ref(NULL),                    // memory reference to align vectors to
  63   _disjoint_ptrs(arena(), 8,  0, OrderedPair::initial), // runtime disambiguated pointer pairs
  64   _dg(_arena),                            // dependence graph
  65   _visited(arena()),                      // visited node set
  66   _post_visited(arena()),                 // post visited node set
  67   _n_idx_list(arena(), 8),                // scratch list of (node,index) pairs
  68   _stk(arena(), 8, 0, NULL),              // scratch stack of nodes
  69   _nlist(arena(), 8, 0, NULL),            // scratch list of nodes
  70   _lpt(NULL),                             // loop tree node
  71   _lp(NULL),                              // LoopNode
  72   _bb(NULL),                              // basic block
  73   _iv(NULL),                              // induction var
  74   _race_possible(false),                  // cases where SDMU is true
  75   _early_return(true),                    // analysis evaluations routine
  76   _num_work_vecs(0),                      // amount of vector work we have
  77   _num_reductions(0),                     // amount of reduction work we have
  78   _do_vector_loop(phase->C->do_vector_loop()),  // whether to do vectorization/simd style
  79   _do_reserve_copy(DoReserveCopyInSuperWord),
  80   _ii_first(-1),                          // first loop generation index - only if do_vector_loop()
  81   _ii_last(-1),                           // last loop generation index - only if do_vector_loop()
  82   _ii_order(arena(), 8, 0, 0)
  83 {
  84 #ifndef PRODUCT
  85   _vector_loop_debug = 0;
  86   if (_phase->C->method() != NULL) {
  87     _vector_loop_debug = phase->C->directive()->VectorizeDebugOption;
  88   }
  89 
  90 #endif
  91 }
  92 
  93 //------------------------------transform_loop---------------------------
  94 void SuperWord::transform_loop(IdealLoopTree* lpt, bool do_optimization) {
  95   assert(UseSuperWord, "should be");
  96   // Do vectors exist on this architecture?
  97   if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return;
  98 
  99   assert(lpt->_head->is_CountedLoop(), "must be");
 100   CountedLoopNode *cl = lpt->_head->as_CountedLoop();
 101 
 102   if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
 103 
 104   bool post_loop_allowed = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
 105   if (post_loop_allowed) {
 106     if (cl->is_reduction_loop()) return; // no predication mapping
 107     Node *limit = cl->limit();
 108     if (limit->is_Con()) return; // non constant limits only
 109     // Now check the limit for expressions we do not handle
 110     if (limit->is_Add()) {
 111       Node *in2 = limit->in(2);
 112       if (in2->is_Con()) {
 113         int val = in2->get_int();
 114         // should not try to program these cases
 115         if (val < 0) return;
 116       }
 117     }
 118   }
 119 
 120   // skip any loop that has not been assigned max unroll by analysis
 121   if (do_optimization) {
 122     if (SuperWordLoopUnrollAnalysis && cl->slp_max_unroll() == 0) return;
 123   }
 124 
 125   // Check for no control flow in body (other than exit)
 126   Node *cl_exit = cl->loopexit();
 127   if (cl->is_main_loop() && (cl_exit->in(0) != lpt->_head)) {
 128     #ifndef PRODUCT
 129       if (TraceSuperWord) {
 130         tty->print_cr("SuperWord::transform_loop: loop too complicated, cl_exit->in(0) != lpt->_head");
 131         tty->print("cl_exit %d", cl_exit->_idx); cl_exit->dump();
 132         tty->print("cl_exit->in(0) %d", cl_exit->in(0)->_idx); cl_exit->in(0)->dump();
 133         tty->print("lpt->_head %d", lpt->_head->_idx); lpt->_head->dump();
 134         lpt->dump_head();
 135       }
 136     #endif
 137     return;
 138   }
 139 
 140   // Make sure the are no extra control users of the loop backedge
 141   if (cl->back_control()->outcnt() != 1) {
 142     return;
 143   }
 144 
 145   // Skip any loops already optimized by slp
 146   if (cl->is_vectorized_loop()) return;
 147 
 148   if (cl->do_unroll_only()) return;
 149 
 150   if (cl->is_main_loop()) {
 151     // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
 152     CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
 153     if (pre_end == NULL) return;
 154     Node *pre_opaq1 = pre_end->limit();
 155     if (pre_opaq1->Opcode() != Op_Opaque1) return;
 156   }
 157 
 158   init(); // initialize data structures
 159 
 160   set_lpt(lpt);
 161   set_lp(cl);
 162 
 163   // For now, define one block which is the entire loop body
 164   set_bb(cl);
 165 
 166   if (do_optimization) {
 167     assert(_packset.length() == 0, "packset must be empty");
 168     SLP_extract();
 169     if (PostLoopMultiversioning && Matcher::has_predicated_vectors()) {
 170       if (cl->is_vectorized_loop() && cl->is_main_loop() && !cl->is_reduction_loop()) {
 171         IdealLoopTree *lpt_next = lpt->_next;
 172         CountedLoopNode *cl_next = lpt_next->_head->as_CountedLoop();
 173         _phase->has_range_checks(lpt_next);
 174         if (cl_next->is_post_loop() && !cl_next->range_checks_present()) {
 175           if (!cl_next->is_vectorized_loop()) {
 176             int slp_max_unroll_factor = cl->slp_max_unroll();
 177             cl_next->set_slp_max_unroll(slp_max_unroll_factor);
 178           }
 179         }
 180       }
 181     }
 182   }
 183 }
 184 
 185 //------------------------------early unrolling analysis------------------------------
 186 void SuperWord::unrolling_analysis(int &local_loop_unroll_factor) {
 187   bool is_slp = true;
 188   ResourceMark rm;
 189   size_t ignored_size = lpt()->_body.size();
 190   int *ignored_loop_nodes = NEW_RESOURCE_ARRAY(int, ignored_size);
 191   Node_Stack nstack((int)ignored_size);
 192   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
 193   Node *cl_exit = cl->loopexit_or_null();
 194   int rpo_idx = _post_block.length();
 195 
 196   assert(rpo_idx == 0, "post loop block is empty");
 197 
 198   // First clear the entries
 199   for (uint i = 0; i < lpt()->_body.size(); i++) {
 200     ignored_loop_nodes[i] = -1;
 201   }
 202 
 203   int max_vector = Matcher::max_vector_size(T_BYTE);
 204   bool post_loop_allowed = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
 205 
 206   // Process the loop, some/all of the stack entries will not be in order, ergo
 207   // need to preprocess the ignored initial state before we process the loop
 208   for (uint i = 0; i < lpt()->_body.size(); i++) {
 209     Node* n = lpt()->_body.at(i);
 210     if (n == cl->incr() ||
 211       n->is_reduction() ||
 212       n->is_AddP() ||
 213       n->is_Cmp() ||
 214       n->is_IfTrue() ||
 215       n->is_CountedLoop() ||
 216       (n == cl_exit)) {
 217       ignored_loop_nodes[i] = n->_idx;
 218       continue;
 219     }
 220 
 221     if (n->is_If()) {
 222       IfNode *iff = n->as_If();
 223       if (iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN) {
 224         if (lpt()->is_loop_exit(iff)) {
 225           ignored_loop_nodes[i] = n->_idx;
 226           continue;
 227         }
 228       }
 229     }
 230 
 231     if (n->is_Phi() && (n->bottom_type() == Type::MEMORY)) {
 232       Node* n_tail = n->in(LoopNode::LoopBackControl);
 233       if (n_tail != n->in(LoopNode::EntryControl)) {
 234         if (!n_tail->is_Mem()) {
 235           is_slp = false;
 236           break;
 237         }
 238       }
 239     }
 240 
 241     // This must happen after check of phi/if
 242     if (n->is_Phi() || n->is_If()) {
 243       ignored_loop_nodes[i] = n->_idx;
 244       continue;
 245     }
 246 
 247     if (n->is_LoadStore() || n->is_MergeMem() ||
 248       (n->is_Proj() && !n->as_Proj()->is_CFG())) {
 249       is_slp = false;
 250       break;
 251     }
 252 
 253     // Ignore nodes with non-primitive type.
 254     BasicType bt;
 255     if (n->is_Mem()) {
 256       bt = n->as_Mem()->memory_type();
 257     } else {
 258       bt = n->bottom_type()->basic_type();
 259     }
 260     if (is_java_primitive(bt) == false) {
 261       ignored_loop_nodes[i] = n->_idx;
 262       continue;
 263     }
 264 
 265     if (n->is_Mem()) {
 266       MemNode* current = n->as_Mem();
 267       Node* adr = n->in(MemNode::Address);
 268       Node* n_ctrl = _phase->get_ctrl(adr);
 269 
 270       // save a queue of post process nodes
 271       if (n_ctrl != NULL && lpt()->is_member(_phase->get_loop(n_ctrl))) {
 272         // Process the memory expression
 273         int stack_idx = 0;
 274         bool have_side_effects = true;
 275         if (adr->is_AddP() == false) {
 276           nstack.push(adr, stack_idx++);
 277         } else {
 278           // Mark the components of the memory operation in nstack
 279           SWPointer p1(current, this, &nstack, true);
 280           have_side_effects = p1.node_stack()->is_nonempty();
 281         }
 282 
 283         // Process the pointer stack
 284         while (have_side_effects) {
 285           Node* pointer_node = nstack.node();
 286           for (uint j = 0; j < lpt()->_body.size(); j++) {
 287             Node* cur_node = lpt()->_body.at(j);
 288             if (cur_node == pointer_node) {
 289               ignored_loop_nodes[j] = cur_node->_idx;
 290               break;
 291             }
 292           }
 293           nstack.pop();
 294           have_side_effects = nstack.is_nonempty();
 295         }
 296       }
 297     }
 298   }
 299 
 300   if (is_slp) {
 301     // Now we try to find the maximum supported consistent vector which the machine
 302     // description can use
 303     bool small_basic_type = false;
 304     bool flag_small_bt = false;
 305     for (uint i = 0; i < lpt()->_body.size(); i++) {
 306       if (ignored_loop_nodes[i] != -1) continue;
 307 
 308       BasicType bt;
 309       Node* n = lpt()->_body.at(i);
 310       if (n->is_Mem()) {
 311         bt = n->as_Mem()->memory_type();
 312       } else {
 313         bt = n->bottom_type()->basic_type();
 314       }
 315 
 316       if (post_loop_allowed) {
 317         if (!small_basic_type) {
 318           switch (bt) {
 319           case T_CHAR:
 320           case T_BYTE:
 321           case T_SHORT:
 322             small_basic_type = true;
 323             break;
 324 
 325           case T_LONG:
 326             // TODO: Remove when support completed for mask context with LONG.
 327             //       Support needs to be augmented for logical qword operations, currently we map to dword
 328             //       buckets for vectors on logicals as these were legacy.
 329             small_basic_type = true;
 330             break;
 331 
 332           default:
 333             break;
 334           }
 335         }
 336       }
 337 
 338       if (is_java_primitive(bt) == false) continue;
 339 
 340          int cur_max_vector = Matcher::max_vector_size(bt);
 341 
 342       // If a max vector exists which is not larger than _local_loop_unroll_factor
 343       // stop looking, we already have the max vector to map to.
 344       if (cur_max_vector < local_loop_unroll_factor) {
 345         is_slp = false;
 346         if (TraceSuperWordLoopUnrollAnalysis) {
 347           tty->print_cr("slp analysis fails: unroll limit greater than max vector\n");
 348         }
 349         break;
 350       }
 351 
 352       // Map the maximal common vector
 353       if (VectorNode::implemented(n->Opcode(), cur_max_vector, bt)) {
 354         if (cur_max_vector < max_vector && !flag_small_bt) {
 355           max_vector = cur_max_vector;
 356         } else if (cur_max_vector > max_vector && UseSubwordForMaxVector) {
 357           // Analyse subword in the loop to set maximum vector size to take advantage of full vector width for subword types.
 358           // Here we analyze if narrowing is likely to happen and if it is we set vector size more aggressively.
 359           // We check for possibility of narrowing by looking through chain operations using subword types.
 360           if (is_subword_type(bt)) {
 361             uint start, end;
 362             VectorNode::vector_operands(n, &start, &end);
 363 
 364             for (uint j = start; j < end; j++) {
 365               Node* in = n->in(j);
 366               // Don't propagate through a memory
 367               if (!in->is_Mem() && in_bb(in) && in->bottom_type()->basic_type() == T_INT) {
 368                 bool same_type = true;
 369                 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
 370                   Node *use = in->fast_out(k);
 371                   if (!in_bb(use) && use->bottom_type()->basic_type() != bt) {
 372                     same_type = false;
 373                     break;
 374                   }
 375                 }
 376                 if (same_type) {
 377                   max_vector = cur_max_vector;
 378                   flag_small_bt = true;
 379                   cl->mark_subword_loop();
 380                 }
 381               }
 382             }
 383           }
 384         }
 385         // We only process post loops on predicated targets where we want to
 386         // mask map the loop to a single iteration
 387         if (post_loop_allowed) {
 388           _post_block.at_put_grow(rpo_idx++, n);
 389         }
 390       }
 391     }
 392     if (is_slp) {
 393       local_loop_unroll_factor = max_vector;
 394       cl->mark_passed_slp();
 395     }
 396     cl->mark_was_slp();
 397     if (cl->is_main_loop()) {
 398       cl->set_slp_max_unroll(local_loop_unroll_factor);
 399     } else if (post_loop_allowed) {
 400       if (!small_basic_type) {
 401         // avoid replication context for small basic types in programmable masked loops
 402         cl->set_slp_max_unroll(local_loop_unroll_factor);
 403       }
 404     }
 405   }
 406 }
 407 
 408 //------------------------------SLP_extract---------------------------
 409 // Extract the superword level parallelism
 410 //
 411 // 1) A reverse post-order of nodes in the block is constructed.  By scanning
 412 //    this list from first to last, all definitions are visited before their uses.
 413 //
 414 // 2) A point-to-point dependence graph is constructed between memory references.
 415 //    This simplies the upcoming "independence" checker.
 416 //
 417 // 3) The maximum depth in the node graph from the beginning of the block
 418 //    to each node is computed.  This is used to prune the graph search
 419 //    in the independence checker.
 420 //
 421 // 4) For integer types, the necessary bit width is propagated backwards
 422 //    from stores to allow packed operations on byte, char, and short
 423 //    integers.  This reverses the promotion to type "int" that javac
 424 //    did for operations like: char c1,c2,c3;  c1 = c2 + c3.
 425 //
 426 // 5) One of the memory references is picked to be an aligned vector reference.
 427 //    The pre-loop trip count is adjusted to align this reference in the
 428 //    unrolled body.
 429 //
 430 // 6) The initial set of pack pairs is seeded with memory references.
 431 //
 432 // 7) The set of pack pairs is extended by following use->def and def->use links.
 433 //
 434 // 8) The pairs are combined into vector sized packs.
 435 //
 436 // 9) Reorder the memory slices to co-locate members of the memory packs.
 437 //
 438 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
 439 //    inserting scalar promotion, vector creation from multiple scalars, and
 440 //    extraction of scalar values from vectors.
 441 //
 442 void SuperWord::SLP_extract() {
 443 
 444 #ifndef PRODUCT
 445   if (_do_vector_loop && TraceSuperWord) {
 446     tty->print("SuperWord::SLP_extract\n");
 447     tty->print("input loop\n");
 448     _lpt->dump_head();
 449     _lpt->dump();
 450     for (uint i = 0; i < _lpt->_body.size(); i++) {
 451       _lpt->_body.at(i)->dump();
 452     }
 453   }
 454 #endif
 455   // Ready the block
 456   if (!construct_bb()) {
 457     return; // Exit if no interesting nodes or complex graph.
 458   }
 459 
 460   // build    _dg, _disjoint_ptrs
 461   dependence_graph();
 462 
 463   // compute function depth(Node*)
 464   compute_max_depth();
 465 
 466   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
 467   bool post_loop_allowed = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
 468   if (cl->is_main_loop()) {
 469     if (_do_vector_loop) {
 470       if (mark_generations() != -1) {
 471         hoist_loads_in_graph(); // this only rebuild the graph; all basic structs need rebuild explicitly
 472 
 473         if (!construct_bb()) {
 474           return; // Exit if no interesting nodes or complex graph.
 475         }
 476         dependence_graph();
 477         compute_max_depth();
 478       }
 479 
 480 #ifndef PRODUCT
 481       if (TraceSuperWord) {
 482         tty->print_cr("\nSuperWord::_do_vector_loop: graph after hoist_loads_in_graph");
 483         _lpt->dump_head();
 484         for (int j = 0; j < _block.length(); j++) {
 485           Node* n = _block.at(j);
 486           int d = depth(n);
 487           for (int i = 0; i < d; i++) tty->print("%s", "  ");
 488           tty->print("%d :", d);
 489           n->dump();
 490         }
 491       }
 492 #endif
 493     }
 494 
 495     compute_vector_element_type();
 496 
 497     // Attempt vectorization
 498 
 499     find_adjacent_refs();
 500 
 501     extend_packlist();
 502 
 503     if (_do_vector_loop) {
 504       if (_packset.length() == 0) {
 505         if (TraceSuperWord) {
 506           tty->print_cr("\nSuperWord::_do_vector_loop DFA could not build packset, now trying to build anyway");
 507         }
 508         pack_parallel();
 509       }
 510     }
 511 
 512     combine_packs();
 513 
 514     construct_my_pack_map();
 515     if (UseVectorCmov) {
 516       merge_packs_to_cmovd();
 517     }
 518 
 519     filter_packs();
 520 
 521     schedule();
 522   } else if (post_loop_allowed) {
 523     int saved_mapped_unroll_factor = cl->slp_max_unroll();
 524     if (saved_mapped_unroll_factor) {
 525       int vector_mapped_unroll_factor = saved_mapped_unroll_factor;
 526 
 527       // now reset the slp_unroll_factor so that we can check the analysis mapped
 528       // what the vector loop was mapped to
 529       cl->set_slp_max_unroll(0);
 530 
 531       // do the analysis on the post loop
 532       unrolling_analysis(vector_mapped_unroll_factor);
 533 
 534       // if our analyzed loop is a canonical fit, start processing it
 535       if (vector_mapped_unroll_factor == saved_mapped_unroll_factor) {
 536         // now add the vector nodes to packsets
 537         for (int i = 0; i < _post_block.length(); i++) {
 538           Node* n = _post_block.at(i);
 539           Node_List* singleton = new Node_List();
 540           singleton->push(n);
 541           _packset.append(singleton);
 542           set_my_pack(n, singleton);
 543         }
 544 
 545         // map base types for vector usage
 546         compute_vector_element_type();
 547       } else {
 548         return;
 549       }
 550     } else {
 551       // for some reason we could not map the slp analysis state of the vectorized loop
 552       return;
 553     }
 554   }
 555 
 556   output();
 557 }
 558 
 559 //------------------------------find_adjacent_refs---------------------------
 560 // Find the adjacent memory references and create pack pairs for them.
 561 // This is the initial set of packs that will then be extended by
 562 // following use->def and def->use links.  The align positions are
 563 // assigned relative to the reference "align_to_ref"
 564 void SuperWord::find_adjacent_refs() {
 565   // Get list of memory operations
 566   Node_List memops;
 567   for (int i = 0; i < _block.length(); i++) {
 568     Node* n = _block.at(i);
 569     if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
 570         is_java_primitive(n->as_Mem()->memory_type())) {
 571       int align = memory_alignment(n->as_Mem(), 0);
 572       if (align != bottom_align) {
 573         memops.push(n);
 574       }
 575     }
 576   }
 577 
 578   Node_List align_to_refs;
 579   int best_iv_adjustment = 0;
 580   MemNode* best_align_to_mem_ref = NULL;
 581 
 582   while (memops.size() != 0) {
 583     // Find a memory reference to align to.
 584     MemNode* mem_ref = find_align_to_ref(memops);
 585     if (mem_ref == NULL) break;
 586     align_to_refs.push(mem_ref);
 587     int iv_adjustment = get_iv_adjustment(mem_ref);
 588 
 589     if (best_align_to_mem_ref == NULL) {
 590       // Set memory reference which is the best from all memory operations
 591       // to be used for alignment. The pre-loop trip count is modified to align
 592       // this reference to a vector-aligned address.
 593       best_align_to_mem_ref = mem_ref;
 594       best_iv_adjustment = iv_adjustment;
 595       NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
 596     }
 597 
 598     SWPointer align_to_ref_p(mem_ref, this, NULL, false);
 599     // Set alignment relative to "align_to_ref" for all related memory operations.
 600     for (int i = memops.size() - 1; i >= 0; i--) {
 601       MemNode* s = memops.at(i)->as_Mem();
 602       if (isomorphic(s, mem_ref) &&
 603            (!_do_vector_loop || same_origin_idx(s, mem_ref))) {
 604         SWPointer p2(s, this, NULL, false);
 605         if (p2.comparable(align_to_ref_p)) {
 606           int align = memory_alignment(s, iv_adjustment);
 607           set_alignment(s, align);
 608         }
 609       }
 610     }
 611 
 612     // Create initial pack pairs of memory operations for which
 613     // alignment is set and vectors will be aligned.
 614     bool create_pack = true;
 615     if (memory_alignment(mem_ref, best_iv_adjustment) == 0 || _do_vector_loop) {
 616       if (!Matcher::misaligned_vectors_ok() || AlignVector) {
 617         int vw = vector_width(mem_ref);
 618         int vw_best = vector_width(best_align_to_mem_ref);
 619         if (vw > vw_best) {
 620           // Do not vectorize a memory access with more elements per vector
 621           // if unaligned memory access is not allowed because number of
 622           // iterations in pre-loop will be not enough to align it.
 623           create_pack = false;
 624         } else {
 625           SWPointer p2(best_align_to_mem_ref, this, NULL, false);
 626           if (align_to_ref_p.invar() != p2.invar()) {
 627             // Do not vectorize memory accesses with different invariants
 628             // if unaligned memory accesses are not allowed.
 629             create_pack = false;
 630           }
 631         }
 632       }
 633     } else {
 634       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 635         // Can't allow vectorization of unaligned memory accesses with the
 636         // same type since it could be overlapped accesses to the same array.
 637         create_pack = false;
 638       } else {
 639         // Allow independent (different type) unaligned memory operations
 640         // if HW supports them.
 641         if (!Matcher::misaligned_vectors_ok() || AlignVector) {
 642           create_pack = false;
 643         } else {
 644           // Check if packs of the same memory type but
 645           // with a different alignment were created before.
 646           for (uint i = 0; i < align_to_refs.size(); i++) {
 647             MemNode* mr = align_to_refs.at(i)->as_Mem();
 648             if (same_velt_type(mr, mem_ref) &&
 649                 memory_alignment(mr, iv_adjustment) != 0)
 650               create_pack = false;
 651           }
 652         }
 653       }
 654     }
 655     if (create_pack) {
 656       for (uint i = 0; i < memops.size(); i++) {
 657         Node* s1 = memops.at(i);
 658         int align = alignment(s1);
 659         if (align == top_align) continue;
 660         for (uint j = 0; j < memops.size(); j++) {
 661           Node* s2 = memops.at(j);
 662           if (alignment(s2) == top_align) continue;
 663           if (s1 != s2 && are_adjacent_refs(s1, s2)) {
 664             if (stmts_can_pack(s1, s2, align)) {
 665               Node_List* pair = new Node_List();
 666               pair->push(s1);
 667               pair->push(s2);
 668               if (!_do_vector_loop || same_origin_idx(s1, s2)) {
 669                 _packset.append(pair);
 670               }
 671             }
 672           }
 673         }
 674       }
 675     } else { // Don't create unaligned pack
 676       // First, remove remaining memory ops of the same type from the list.
 677       for (int i = memops.size() - 1; i >= 0; i--) {
 678         MemNode* s = memops.at(i)->as_Mem();
 679         if (same_velt_type(s, mem_ref)) {
 680           memops.remove(i);
 681         }
 682       }
 683 
 684       // Second, remove already constructed packs of the same type.
 685       for (int i = _packset.length() - 1; i >= 0; i--) {
 686         Node_List* p = _packset.at(i);
 687         MemNode* s = p->at(0)->as_Mem();
 688         if (same_velt_type(s, mem_ref)) {
 689           remove_pack_at(i);
 690         }
 691       }
 692 
 693       // If needed find the best memory reference for loop alignment again.
 694       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 695         // Put memory ops from remaining packs back on memops list for
 696         // the best alignment search.
 697         uint orig_msize = memops.size();
 698         for (int i = 0; i < _packset.length(); i++) {
 699           Node_List* p = _packset.at(i);
 700           MemNode* s = p->at(0)->as_Mem();
 701           assert(!same_velt_type(s, mem_ref), "sanity");
 702           memops.push(s);
 703         }
 704         best_align_to_mem_ref = find_align_to_ref(memops);
 705         if (best_align_to_mem_ref == NULL) {
 706           if (TraceSuperWord) {
 707             tty->print_cr("SuperWord::find_adjacent_refs(): best_align_to_mem_ref == NULL");
 708           }
 709           break;
 710         }
 711         best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
 712         NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
 713         // Restore list.
 714         while (memops.size() > orig_msize)
 715           (void)memops.pop();
 716       }
 717     } // unaligned memory accesses
 718 
 719     // Remove used mem nodes.
 720     for (int i = memops.size() - 1; i >= 0; i--) {
 721       MemNode* m = memops.at(i)->as_Mem();
 722       if (alignment(m) != top_align) {
 723         memops.remove(i);
 724       }
 725     }
 726 
 727   } // while (memops.size() != 0
 728   set_align_to_ref(best_align_to_mem_ref);
 729 
 730   if (TraceSuperWord) {
 731     tty->print_cr("\nAfter find_adjacent_refs");
 732     print_packset();
 733   }
 734 }
 735 
 736 #ifndef PRODUCT
 737 void SuperWord::find_adjacent_refs_trace_1(Node* best_align_to_mem_ref, int best_iv_adjustment) {
 738   if (is_trace_adjacent()) {
 739     tty->print("SuperWord::find_adjacent_refs best_align_to_mem_ref = %d, best_iv_adjustment = %d",
 740        best_align_to_mem_ref->_idx, best_iv_adjustment);
 741        best_align_to_mem_ref->dump();
 742   }
 743 }
 744 #endif
 745 
 746 //------------------------------find_align_to_ref---------------------------
 747 // Find a memory reference to align the loop induction variable to.
 748 // Looks first at stores then at loads, looking for a memory reference
 749 // with the largest number of references similar to it.
 750 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
 751   GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
 752 
 753   // Count number of comparable memory ops
 754   for (uint i = 0; i < memops.size(); i++) {
 755     MemNode* s1 = memops.at(i)->as_Mem();
 756     SWPointer p1(s1, this, NULL, false);
 757     // Discard if pre loop can't align this reference
 758     if (!ref_is_alignable(p1)) {
 759       *cmp_ct.adr_at(i) = 0;
 760       continue;
 761     }
 762     for (uint j = i+1; j < memops.size(); j++) {
 763       MemNode* s2 = memops.at(j)->as_Mem();
 764       if (isomorphic(s1, s2)) {
 765         SWPointer p2(s2, this, NULL, false);
 766         if (p1.comparable(p2)) {
 767           (*cmp_ct.adr_at(i))++;
 768           (*cmp_ct.adr_at(j))++;
 769         }
 770       }
 771     }
 772   }
 773 
 774   // Find Store (or Load) with the greatest number of "comparable" references,
 775   // biggest vector size, smallest data size and smallest iv offset.
 776   int max_ct        = 0;
 777   int max_vw        = 0;
 778   int max_idx       = -1;
 779   int min_size      = max_jint;
 780   int min_iv_offset = max_jint;
 781   for (uint j = 0; j < memops.size(); j++) {
 782     MemNode* s = memops.at(j)->as_Mem();
 783     if (s->is_Store()) {
 784       int vw = vector_width_in_bytes(s);
 785       assert(vw > 1, "sanity");
 786       SWPointer p(s, this, NULL, false);
 787       if ( cmp_ct.at(j) >  max_ct ||
 788           (cmp_ct.at(j) == max_ct &&
 789             ( vw >  max_vw ||
 790              (vw == max_vw &&
 791               ( data_size(s) <  min_size ||
 792                (data_size(s) == min_size &&
 793                 p.offset_in_bytes() < min_iv_offset)))))) {
 794         max_ct = cmp_ct.at(j);
 795         max_vw = vw;
 796         max_idx = j;
 797         min_size = data_size(s);
 798         min_iv_offset = p.offset_in_bytes();
 799       }
 800     }
 801   }
 802   // If no stores, look at loads
 803   if (max_ct == 0) {
 804     for (uint j = 0; j < memops.size(); j++) {
 805       MemNode* s = memops.at(j)->as_Mem();
 806       if (s->is_Load()) {
 807         int vw = vector_width_in_bytes(s);
 808         assert(vw > 1, "sanity");
 809         SWPointer p(s, this, NULL, false);
 810         if ( cmp_ct.at(j) >  max_ct ||
 811             (cmp_ct.at(j) == max_ct &&
 812               ( vw >  max_vw ||
 813                (vw == max_vw &&
 814                 ( data_size(s) <  min_size ||
 815                  (data_size(s) == min_size &&
 816                   p.offset_in_bytes() < min_iv_offset)))))) {
 817           max_ct = cmp_ct.at(j);
 818           max_vw = vw;
 819           max_idx = j;
 820           min_size = data_size(s);
 821           min_iv_offset = p.offset_in_bytes();
 822         }
 823       }
 824     }
 825   }
 826 
 827 #ifdef ASSERT
 828   if (TraceSuperWord && Verbose) {
 829     tty->print_cr("\nVector memops after find_align_to_ref");
 830     for (uint i = 0; i < memops.size(); i++) {
 831       MemNode* s = memops.at(i)->as_Mem();
 832       s->dump();
 833     }
 834   }
 835 #endif
 836 
 837   if (max_ct > 0) {
 838 #ifdef ASSERT
 839     if (TraceSuperWord) {
 840       tty->print("\nVector align to node: ");
 841       memops.at(max_idx)->as_Mem()->dump();
 842     }
 843 #endif
 844     return memops.at(max_idx)->as_Mem();
 845   }
 846   return NULL;
 847 }
 848 
 849 //------------------------------ref_is_alignable---------------------------
 850 // Can the preloop align the reference to position zero in the vector?
 851 bool SuperWord::ref_is_alignable(SWPointer& p) {
 852   if (!p.has_iv()) {
 853     return true;   // no induction variable
 854   }
 855   CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
 856   assert(pre_end != NULL, "we must have a correct pre-loop");
 857   assert(pre_end->stride_is_con(), "pre loop stride is constant");
 858   int preloop_stride = pre_end->stride_con();
 859 
 860   int span = preloop_stride * p.scale_in_bytes();
 861   int mem_size = p.memory_size();
 862   int offset   = p.offset_in_bytes();
 863   // Stride one accesses are alignable if offset is aligned to memory operation size.
 864   // Offset can be unaligned when UseUnalignedAccesses is used.
 865   if (ABS(span) == mem_size && (ABS(offset) % mem_size) == 0) {
 866     return true;
 867   }
 868   // If the initial offset from start of the object is computable,
 869   // check if the pre-loop can align the final offset accordingly.
 870   //
 871   // In other words: Can we find an i such that the offset
 872   // after i pre-loop iterations is aligned to vw?
 873   //   (init_offset + pre_loop) % vw == 0              (1)
 874   // where
 875   //   pre_loop = i * span
 876   // is the number of bytes added to the offset by i pre-loop iterations.
 877   //
 878   // For this to hold we need pre_loop to increase init_offset by
 879   //   pre_loop = vw - (init_offset % vw)
 880   //
 881   // This is only possible if pre_loop is divisible by span because each
 882   // pre-loop iteration increases the initial offset by 'span' bytes:
 883   //   (vw - (init_offset % vw)) % span == 0
 884   //
 885   int vw = vector_width_in_bytes(p.mem());
 886   assert(vw > 1, "sanity");
 887   Node* init_nd = pre_end->init_trip();
 888   if (init_nd->is_Con() && p.invar() == NULL) {
 889     int init = init_nd->bottom_type()->is_int()->get_con();
 890     int init_offset = init * p.scale_in_bytes() + offset;
 891     if (init_offset < 0) { // negative offset from object start?
 892       return false;        // may happen in dead loop
 893     }
 894     if (vw % span == 0) {
 895       // If vm is a multiple of span, we use formula (1).
 896       if (span > 0) {
 897         return (vw - (init_offset % vw)) % span == 0;
 898       } else {
 899         assert(span < 0, "nonzero stride * scale");
 900         return (init_offset % vw) % -span == 0;
 901       }
 902     } else if (span % vw == 0) {
 903       // If span is a multiple of vw, we can simplify formula (1) to:
 904       //   (init_offset + i * span) % vw == 0
 905       //     =>
 906       //   (init_offset % vw) + ((i * span) % vw) == 0
 907       //     =>
 908       //   init_offset % vw == 0
 909       //
 910       // Because we add a multiple of vw to the initial offset, the final
 911       // offset is a multiple of vw if and only if init_offset is a multiple.
 912       //
 913       return (init_offset % vw) == 0;
 914     }
 915   }
 916   return false;
 917 }
 918 
 919 //---------------------------get_iv_adjustment---------------------------
 920 // Calculate loop's iv adjustment for this memory ops.
 921 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
 922   SWPointer align_to_ref_p(mem_ref, this, NULL, false);
 923   int offset = align_to_ref_p.offset_in_bytes();
 924   int scale  = align_to_ref_p.scale_in_bytes();
 925   int elt_size = align_to_ref_p.memory_size();
 926   int vw       = vector_width_in_bytes(mem_ref);
 927   assert(vw > 1, "sanity");
 928   int iv_adjustment;
 929   if (scale != 0) {
 930     int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
 931     // At least one iteration is executed in pre-loop by default. As result
 932     // several iterations are needed to align memory operations in main-loop even
 933     // if offset is 0.
 934     int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
 935     assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
 936            "(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size);
 937     iv_adjustment = iv_adjustment_in_bytes/elt_size;
 938   } else {
 939     // This memory op is not dependent on iv (scale == 0)
 940     iv_adjustment = 0;
 941   }
 942 
 943 #ifndef PRODUCT
 944   if (TraceSuperWord) {
 945     tty->print("SuperWord::get_iv_adjustment: n = %d, noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d: ",
 946       mem_ref->_idx, offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
 947     mem_ref->dump();
 948   }
 949 #endif
 950   return iv_adjustment;
 951 }
 952 
 953 //---------------------------dependence_graph---------------------------
 954 // Construct dependency graph.
 955 // Add dependence edges to load/store nodes for memory dependence
 956 //    A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
 957 void SuperWord::dependence_graph() {
 958   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
 959   // First, assign a dependence node to each memory node
 960   for (int i = 0; i < _block.length(); i++ ) {
 961     Node *n = _block.at(i);
 962     if (n->is_Mem() || (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
 963       _dg.make_node(n);
 964     }
 965   }
 966 
 967   // For each memory slice, create the dependences
 968   for (int i = 0; i < _mem_slice_head.length(); i++) {
 969     Node* n      = _mem_slice_head.at(i);
 970     Node* n_tail = _mem_slice_tail.at(i);
 971 
 972     // Get slice in predecessor order (last is first)
 973     if (cl->is_main_loop()) {
 974       mem_slice_preds(n_tail, n, _nlist);
 975     }
 976 
 977 #ifndef PRODUCT
 978     if(TraceSuperWord && Verbose) {
 979       tty->print_cr("SuperWord::dependence_graph: built a new mem slice");
 980       for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 981         _nlist.at(j)->dump();
 982       }
 983     }
 984 #endif
 985     // Make the slice dependent on the root
 986     DepMem* slice = _dg.dep(n);
 987     _dg.make_edge(_dg.root(), slice);
 988 
 989     // Create a sink for the slice
 990     DepMem* slice_sink = _dg.make_node(NULL);
 991     _dg.make_edge(slice_sink, _dg.tail());
 992 
 993     // Now visit each pair of memory ops, creating the edges
 994     for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 995       Node* s1 = _nlist.at(j);
 996 
 997       // If no dependency yet, use slice
 998       if (_dg.dep(s1)->in_cnt() == 0) {
 999         _dg.make_edge(slice, s1);
1000       }
1001       SWPointer p1(s1->as_Mem(), this, NULL, false);
1002       bool sink_dependent = true;
1003       for (int k = j - 1; k >= 0; k--) {
1004         Node* s2 = _nlist.at(k);
1005         if (s1->is_Load() && s2->is_Load())
1006           continue;
1007         SWPointer p2(s2->as_Mem(), this, NULL, false);
1008 
1009         int cmp = p1.cmp(p2);
1010         if (SuperWordRTDepCheck &&
1011             p1.base() != p2.base() && p1.valid() && p2.valid()) {
1012           // Create a runtime check to disambiguate
1013           OrderedPair pp(p1.base(), p2.base());
1014           _disjoint_ptrs.append_if_missing(pp);
1015         } else if (!SWPointer::not_equal(cmp)) {
1016           // Possibly same address
1017           _dg.make_edge(s1, s2);
1018           sink_dependent = false;
1019         }
1020       }
1021       if (sink_dependent) {
1022         _dg.make_edge(s1, slice_sink);
1023       }
1024     }
1025 
1026     if (TraceSuperWord) {
1027       tty->print_cr("\nDependence graph for slice: %d", n->_idx);
1028       for (int q = 0; q < _nlist.length(); q++) {
1029         _dg.print(_nlist.at(q));
1030       }
1031       tty->cr();
1032     }
1033 
1034     _nlist.clear();
1035   }
1036 
1037   if (TraceSuperWord) {
1038     tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
1039     for (int r = 0; r < _disjoint_ptrs.length(); r++) {
1040       _disjoint_ptrs.at(r).print();
1041       tty->cr();
1042     }
1043     tty->cr();
1044   }
1045 
1046 }
1047 
1048 //---------------------------mem_slice_preds---------------------------
1049 // Return a memory slice (node list) in predecessor order starting at "start"
1050 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
1051   assert(preds.length() == 0, "start empty");
1052   Node* n = start;
1053   Node* prev = NULL;
1054   while (true) {
1055     NOT_PRODUCT( if(is_trace_mem_slice()) tty->print_cr("SuperWord::mem_slice_preds: n %d", n->_idx);)
1056     assert(in_bb(n), "must be in block");
1057     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1058       Node* out = n->fast_out(i);
1059       if (out->is_Load()) {
1060         if (in_bb(out)) {
1061           preds.push(out);
1062           if (TraceSuperWord && Verbose) {
1063             tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", out->_idx);
1064           }
1065         }
1066       } else {
1067         // FIXME
1068         if (out->is_MergeMem() && !in_bb(out)) {
1069           // Either unrolling is causing a memory edge not to disappear,
1070           // or need to run igvn.optimize() again before SLP
1071         } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
1072           // Ditto.  Not sure what else to check further.
1073         } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
1074           // StoreCM has an input edge used as a precedence edge.
1075           // Maybe an issue when oop stores are vectorized.
1076         } else {
1077           assert(out == prev || prev == NULL, "no branches off of store slice");
1078         }
1079       }//else
1080     }//for
1081     if (n == stop) break;
1082     preds.push(n);
1083     if (TraceSuperWord && Verbose) {
1084       tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", n->_idx);
1085     }
1086     prev = n;
1087     assert(n->is_Mem(), "unexpected node %s", n->Name());
1088     n = n->in(MemNode::Memory);
1089   }
1090 }
1091 
1092 //------------------------------stmts_can_pack---------------------------
1093 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
1094 // s1 aligned at "align"
1095 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
1096 
1097   // Do not use superword for non-primitives
1098   BasicType bt1 = velt_basic_type(s1);
1099   BasicType bt2 = velt_basic_type(s2);
1100   if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
1101     return false;
1102   if (Matcher::max_vector_size(bt1) < 2) {
1103     return false; // No vectors for this type
1104   }
1105 
1106   if (isomorphic(s1, s2)) {
1107     if ((independent(s1, s2) && have_similar_inputs(s1, s2)) || reduction(s1, s2)) {
1108       if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
1109         if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
1110           int s1_align = alignment(s1);
1111           int s2_align = alignment(s2);
1112           if (s1_align == top_align || s1_align == align) {
1113             if (s2_align == top_align || s2_align == align + data_size(s1)) {
1114               return true;
1115             }
1116           }
1117         }
1118       }
1119     }
1120   }
1121   return false;
1122 }
1123 
1124 //------------------------------exists_at---------------------------
1125 // Does s exist in a pack at position pos?
1126 bool SuperWord::exists_at(Node* s, uint pos) {
1127   for (int i = 0; i < _packset.length(); i++) {
1128     Node_List* p = _packset.at(i);
1129     if (p->at(pos) == s) {
1130       return true;
1131     }
1132   }
1133   return false;
1134 }
1135 
1136 //------------------------------are_adjacent_refs---------------------------
1137 // Is s1 immediately before s2 in memory?
1138 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
1139   if (!s1->is_Mem() || !s2->is_Mem()) return false;
1140   if (!in_bb(s1)    || !in_bb(s2))    return false;
1141 
1142   // Do not use superword for non-primitives
1143   if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
1144       !is_java_primitive(s2->as_Mem()->memory_type())) {
1145     return false;
1146   }
1147 
1148   // FIXME - co_locate_pack fails on Stores in different mem-slices, so
1149   // only pack memops that are in the same alias set until that's fixed.
1150   if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
1151       _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
1152     return false;
1153   SWPointer p1(s1->as_Mem(), this, NULL, false);
1154   SWPointer p2(s2->as_Mem(), this, NULL, false);
1155   if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
1156   int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
1157   return diff == data_size(s1);
1158 }
1159 
1160 //------------------------------isomorphic---------------------------
1161 // Are s1 and s2 similar?
1162 bool SuperWord::isomorphic(Node* s1, Node* s2) {
1163   if (s1->Opcode() != s2->Opcode()) return false;
1164   if (s1->req() != s2->req()) return false;
1165   if (s1->in(0) != s2->in(0)) return false;
1166   if (!same_velt_type(s1, s2)) return false;
1167   return true;
1168 }
1169 
1170 //------------------------------independent---------------------------
1171 // Is there no data path from s1 to s2 or s2 to s1?
1172 bool SuperWord::independent(Node* s1, Node* s2) {
1173   //  assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
1174   int d1 = depth(s1);
1175   int d2 = depth(s2);
1176   if (d1 == d2) return s1 != s2;
1177   Node* deep    = d1 > d2 ? s1 : s2;
1178   Node* shallow = d1 > d2 ? s2 : s1;
1179 
1180   visited_clear();
1181 
1182   return independent_path(shallow, deep);
1183 }
1184 
1185 //--------------------------have_similar_inputs-----------------------
1186 // For a node pair (s1, s2) which is isomorphic and independent,
1187 // do s1 and s2 have similar input edges?
1188 bool SuperWord::have_similar_inputs(Node* s1, Node* s2) {
1189   // assert(isomorphic(s1, s2) == true, "check isomorphic");
1190   // assert(independent(s1, s2) == true, "check independent");
1191   if (s1->req() > 1 && !s1->is_Store() && !s1->is_Load()) {
1192     for (uint i = 1; i < s1->req(); i++) {
1193       if (s1->in(i)->Opcode() != s2->in(i)->Opcode()) return false;
1194     }
1195   }
1196   return true;
1197 }
1198 
1199 //------------------------------reduction---------------------------
1200 // Is there a data path between s1 and s2 and the nodes reductions?
1201 bool SuperWord::reduction(Node* s1, Node* s2) {
1202   bool retValue = false;
1203   int d1 = depth(s1);
1204   int d2 = depth(s2);
1205   if (d1 + 1 == d2) {
1206     if (s1->is_reduction() && s2->is_reduction()) {
1207       // This is an ordered set, so s1 should define s2
1208       for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1209         Node* t1 = s1->fast_out(i);
1210         if (t1 == s2) {
1211           // both nodes are reductions and connected
1212           retValue = true;
1213         }
1214       }
1215     }
1216   }
1217 
1218   return retValue;
1219 }
1220 
1221 //------------------------------independent_path------------------------------
1222 // Helper for independent
1223 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
1224   if (dp >= 1000) return false; // stop deep recursion
1225   visited_set(deep);
1226   int shal_depth = depth(shallow);
1227   assert(shal_depth <= depth(deep), "must be");
1228   for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
1229     Node* pred = preds.current();
1230     if (in_bb(pred) && !visited_test(pred)) {
1231       if (shallow == pred) {
1232         return false;
1233       }
1234       if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
1235         return false;
1236       }
1237     }
1238   }
1239   return true;
1240 }
1241 
1242 //------------------------------set_alignment---------------------------
1243 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
1244   set_alignment(s1, align);
1245   if (align == top_align || align == bottom_align) {
1246     set_alignment(s2, align);
1247   } else {
1248     set_alignment(s2, align + data_size(s1));
1249   }
1250 }
1251 
1252 //------------------------------data_size---------------------------
1253 int SuperWord::data_size(Node* s) {
1254   Node* use = NULL; //test if the node is a candidate for CMoveV optimization, then return the size of CMov
1255   if (UseVectorCmov) {
1256     use = _cmovev_kit.is_Bool_candidate(s);
1257     if (use != NULL) {
1258       return data_size(use);
1259     }
1260     use = _cmovev_kit.is_CmpD_candidate(s);
1261     if (use != NULL) {
1262       return data_size(use);
1263     }
1264   }
1265 
1266   int bsize = type2aelembytes(velt_basic_type(s));
1267   assert(bsize != 0, "valid size");
1268   return bsize;
1269 }
1270 
1271 //------------------------------extend_packlist---------------------------
1272 // Extend packset by following use->def and def->use links from pack members.
1273 void SuperWord::extend_packlist() {
1274   bool changed;
1275   do {
1276     packset_sort(_packset.length());
1277     changed = false;
1278     for (int i = 0; i < _packset.length(); i++) {
1279       Node_List* p = _packset.at(i);
1280       changed |= follow_use_defs(p);
1281       changed |= follow_def_uses(p);
1282     }
1283   } while (changed);
1284 
1285   if (_race_possible) {
1286     for (int i = 0; i < _packset.length(); i++) {
1287       Node_List* p = _packset.at(i);
1288       order_def_uses(p);
1289     }
1290   }
1291 
1292   if (TraceSuperWord) {
1293     tty->print_cr("\nAfter extend_packlist");
1294     print_packset();
1295   }
1296 }
1297 
1298 //------------------------------follow_use_defs---------------------------
1299 // Extend the packset by visiting operand definitions of nodes in pack p
1300 bool SuperWord::follow_use_defs(Node_List* p) {
1301   assert(p->size() == 2, "just checking");
1302   Node* s1 = p->at(0);
1303   Node* s2 = p->at(1);
1304   assert(s1->req() == s2->req(), "just checking");
1305   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1306 
1307   if (s1->is_Load()) return false;
1308 
1309   int align = alignment(s1);
1310   NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: s1 %d, align %d", s1->_idx, align);)
1311   bool changed = false;
1312   int start = s1->is_Store() ? MemNode::ValueIn   : 1;
1313   int end   = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
1314   for (int j = start; j < end; j++) {
1315     Node* t1 = s1->in(j);
1316     Node* t2 = s2->in(j);
1317     if (!in_bb(t1) || !in_bb(t2))
1318       continue;
1319     if (stmts_can_pack(t1, t2, align)) {
1320       if (est_savings(t1, t2) >= 0) {
1321         Node_List* pair = new Node_List();
1322         pair->push(t1);
1323         pair->push(t2);
1324         _packset.append(pair);
1325         NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: set_alignment(%d, %d, %d)", t1->_idx, t2->_idx, align);)
1326         set_alignment(t1, t2, align);
1327         changed = true;
1328       }
1329     }
1330   }
1331   return changed;
1332 }
1333 
1334 //------------------------------follow_def_uses---------------------------
1335 // Extend the packset by visiting uses of nodes in pack p
1336 bool SuperWord::follow_def_uses(Node_List* p) {
1337   bool changed = false;
1338   Node* s1 = p->at(0);
1339   Node* s2 = p->at(1);
1340   assert(p->size() == 2, "just checking");
1341   assert(s1->req() == s2->req(), "just checking");
1342   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1343 
1344   if (s1->is_Store()) return false;
1345 
1346   int align = alignment(s1);
1347   NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: s1 %d, align %d", s1->_idx, align);)
1348   int savings = -1;
1349   int num_s1_uses = 0;
1350   Node* u1 = NULL;
1351   Node* u2 = NULL;
1352   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1353     Node* t1 = s1->fast_out(i);
1354     num_s1_uses++;
1355     if (!in_bb(t1)) continue;
1356     for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1357       Node* t2 = s2->fast_out(j);
1358       if (!in_bb(t2)) continue;
1359       if (t2->Opcode() == Op_AddI && t2 == _lp->as_CountedLoop()->incr()) continue; // don't mess with the iv
1360       if (!opnd_positions_match(s1, t1, s2, t2))
1361         continue;
1362       if (stmts_can_pack(t1, t2, align)) {
1363         int my_savings = est_savings(t1, t2);
1364         if (my_savings > savings) {
1365           savings = my_savings;
1366           u1 = t1;
1367           u2 = t2;
1368         }
1369       }
1370     }
1371   }
1372   if (num_s1_uses > 1) {
1373     _race_possible = true;
1374   }
1375   if (savings >= 0) {
1376     Node_List* pair = new Node_List();
1377     pair->push(u1);
1378     pair->push(u2);
1379     _packset.append(pair);
1380     NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: set_alignment(%d, %d, %d)", u1->_idx, u2->_idx, align);)
1381     set_alignment(u1, u2, align);
1382     changed = true;
1383   }
1384   return changed;
1385 }
1386 
1387 //------------------------------order_def_uses---------------------------
1388 // For extended packsets, ordinally arrange uses packset by major component
1389 void SuperWord::order_def_uses(Node_List* p) {
1390   Node* s1 = p->at(0);
1391 
1392   if (s1->is_Store()) return;
1393 
1394   // reductions are always managed beforehand
1395   if (s1->is_reduction()) return;
1396 
1397   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1398     Node* t1 = s1->fast_out(i);
1399 
1400     // Only allow operand swap on commuting operations
1401     if (!t1->is_Add() && !t1->is_Mul()) {
1402       break;
1403     }
1404 
1405     // Now find t1's packset
1406     Node_List* p2 = NULL;
1407     for (int j = 0; j < _packset.length(); j++) {
1408       p2 = _packset.at(j);
1409       Node* first = p2->at(0);
1410       if (t1 == first) {
1411         break;
1412       }
1413       p2 = NULL;
1414     }
1415     // Arrange all sub components by the major component
1416     if (p2 != NULL) {
1417       for (uint j = 1; j < p->size(); j++) {
1418         Node* d1 = p->at(j);
1419         Node* u1 = p2->at(j);
1420         opnd_positions_match(s1, t1, d1, u1);
1421       }
1422     }
1423   }
1424 }
1425 
1426 //---------------------------opnd_positions_match-------------------------
1427 // Is the use of d1 in u1 at the same operand position as d2 in u2?
1428 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
1429   // check reductions to see if they are marshalled to represent the reduction
1430   // operator in a specified opnd
1431   if (u1->is_reduction() && u2->is_reduction()) {
1432     // ensure reductions have phis and reduction definitions feeding the 1st operand
1433     Node* first = u1->in(2);
1434     if (first->is_Phi() || first->is_reduction()) {
1435       u1->swap_edges(1, 2);
1436     }
1437     // ensure reductions have phis and reduction definitions feeding the 1st operand
1438     first = u2->in(2);
1439     if (first->is_Phi() || first->is_reduction()) {
1440       u2->swap_edges(1, 2);
1441     }
1442     return true;
1443   }
1444 
1445   uint ct = u1->req();
1446   if (ct != u2->req()) return false;
1447   uint i1 = 0;
1448   uint i2 = 0;
1449   do {
1450     for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
1451     for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
1452     if (i1 != i2) {
1453       if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
1454         // Further analysis relies on operands position matching.
1455         u2->swap_edges(i1, i2);
1456       } else {
1457         return false;
1458       }
1459     }
1460   } while (i1 < ct);
1461   return true;
1462 }
1463 
1464 //------------------------------est_savings---------------------------
1465 // Estimate the savings from executing s1 and s2 as a pack
1466 int SuperWord::est_savings(Node* s1, Node* s2) {
1467   int save_in = 2 - 1; // 2 operations per instruction in packed form
1468 
1469   // inputs
1470   for (uint i = 1; i < s1->req(); i++) {
1471     Node* x1 = s1->in(i);
1472     Node* x2 = s2->in(i);
1473     if (x1 != x2) {
1474       if (are_adjacent_refs(x1, x2)) {
1475         save_in += adjacent_profit(x1, x2);
1476       } else if (!in_packset(x1, x2)) {
1477         save_in -= pack_cost(2);
1478       } else {
1479         save_in += unpack_cost(2);
1480       }
1481     }
1482   }
1483 
1484   // uses of result
1485   uint ct = 0;
1486   int save_use = 0;
1487   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1488     Node* s1_use = s1->fast_out(i);
1489     for (int j = 0; j < _packset.length(); j++) {
1490       Node_List* p = _packset.at(j);
1491       if (p->at(0) == s1_use) {
1492         for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
1493           Node* s2_use = s2->fast_out(k);
1494           if (p->at(p->size()-1) == s2_use) {
1495             ct++;
1496             if (are_adjacent_refs(s1_use, s2_use)) {
1497               save_use += adjacent_profit(s1_use, s2_use);
1498             }
1499           }
1500         }
1501       }
1502     }
1503   }
1504 
1505   if (ct < s1->outcnt()) save_use += unpack_cost(1);
1506   if (ct < s2->outcnt()) save_use += unpack_cost(1);
1507 
1508   return MAX2(save_in, save_use);
1509 }
1510 
1511 //------------------------------costs---------------------------
1512 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
1513 int SuperWord::pack_cost(int ct)   { return ct; }
1514 int SuperWord::unpack_cost(int ct) { return ct; }
1515 
1516 //------------------------------combine_packs---------------------------
1517 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
1518 void SuperWord::combine_packs() {
1519   bool changed = true;
1520   // Combine packs regardless max vector size.
1521   while (changed) {
1522     changed = false;
1523     for (int i = 0; i < _packset.length(); i++) {
1524       Node_List* p1 = _packset.at(i);
1525       if (p1 == NULL) continue;
1526       // Because of sorting we can start at i + 1
1527       for (int j = i + 1; j < _packset.length(); j++) {
1528         Node_List* p2 = _packset.at(j);
1529         if (p2 == NULL) continue;
1530         if (i == j) continue;
1531         if (p1->at(p1->size()-1) == p2->at(0)) {
1532           for (uint k = 1; k < p2->size(); k++) {
1533             p1->push(p2->at(k));
1534           }
1535           _packset.at_put(j, NULL);
1536           changed = true;
1537         }
1538       }
1539     }
1540   }
1541 
1542   // Split packs which have size greater then max vector size.
1543   for (int i = 0; i < _packset.length(); i++) {
1544     Node_List* p1 = _packset.at(i);
1545     if (p1 != NULL) {
1546       BasicType bt = velt_basic_type(p1->at(0));
1547       uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1548       assert(is_power_of_2(max_vlen), "sanity");
1549       uint psize = p1->size();
1550       if (!is_power_of_2(psize)) {
1551         // Skip pack which can't be vector.
1552         // case1: for(...) { a[i] = i; }    elements values are different (i+x)
1553         // case2: for(...) { a[i] = b[i+1]; }  can't align both, load and store
1554         _packset.at_put(i, NULL);
1555         continue;
1556       }
1557       if (psize > max_vlen) {
1558         Node_List* pack = new Node_List();
1559         for (uint j = 0; j < psize; j++) {
1560           pack->push(p1->at(j));
1561           if (pack->size() >= max_vlen) {
1562             assert(is_power_of_2(pack->size()), "sanity");
1563             _packset.append(pack);
1564             pack = new Node_List();
1565           }
1566         }
1567         _packset.at_put(i, NULL);
1568       }
1569     }
1570   }
1571 
1572   // Compress list.
1573   for (int i = _packset.length() - 1; i >= 0; i--) {
1574     Node_List* p1 = _packset.at(i);
1575     if (p1 == NULL) {
1576       _packset.remove_at(i);
1577     }
1578   }
1579 
1580   if (TraceSuperWord) {
1581     tty->print_cr("\nAfter combine_packs");
1582     print_packset();
1583   }
1584 }
1585 
1586 //-----------------------------construct_my_pack_map--------------------------
1587 // Construct the map from nodes to packs.  Only valid after the
1588 // point where a node is only in one pack (after combine_packs).
1589 void SuperWord::construct_my_pack_map() {
1590   Node_List* rslt = NULL;
1591   for (int i = 0; i < _packset.length(); i++) {
1592     Node_List* p = _packset.at(i);
1593     for (uint j = 0; j < p->size(); j++) {
1594       Node* s = p->at(j);
1595       assert(my_pack(s) == NULL, "only in one pack");
1596       set_my_pack(s, p);
1597     }
1598   }
1599 }
1600 
1601 //------------------------------filter_packs---------------------------
1602 // Remove packs that are not implemented or not profitable.
1603 void SuperWord::filter_packs() {
1604   // Remove packs that are not implemented
1605   for (int i = _packset.length() - 1; i >= 0; i--) {
1606     Node_List* pk = _packset.at(i);
1607     bool impl = implemented(pk);
1608     if (!impl) {
1609 #ifndef PRODUCT
1610       if (TraceSuperWord && Verbose) {
1611         tty->print_cr("Unimplemented");
1612         pk->at(0)->dump();
1613       }
1614 #endif
1615       remove_pack_at(i);
1616     }
1617     Node *n = pk->at(0);
1618     if (n->is_reduction()) {
1619       _num_reductions++;
1620     } else {
1621       _num_work_vecs++;
1622     }
1623   }
1624 
1625   // Remove packs that are not profitable
1626   bool changed;
1627   do {
1628     changed = false;
1629     for (int i = _packset.length() - 1; i >= 0; i--) {
1630       Node_List* pk = _packset.at(i);
1631       bool prof = profitable(pk);
1632       if (!prof) {
1633 #ifndef PRODUCT
1634         if (TraceSuperWord && Verbose) {
1635           tty->print_cr("Unprofitable");
1636           pk->at(0)->dump();
1637         }
1638 #endif
1639         remove_pack_at(i);
1640         changed = true;
1641       }
1642     }
1643   } while (changed);
1644 
1645 #ifndef PRODUCT
1646   if (TraceSuperWord) {
1647     tty->print_cr("\nAfter filter_packs");
1648     print_packset();
1649     tty->cr();
1650   }
1651 #endif
1652 }
1653 
1654 //------------------------------merge_packs_to_cmovd---------------------------
1655 // Merge CMoveD into new vector-nodes
1656 // We want to catch this pattern and subsume CmpD and Bool into CMoveD
1657 //
1658 //                   SubD             ConD
1659 //                  /  |               /
1660 //                 /   |           /   /
1661 //                /    |       /      /
1662 //               /     |   /         /
1663 //              /      /            /
1664 //             /    /  |           /
1665 //            v /      |          /
1666 //         CmpD        |         /
1667 //          |          |        /
1668 //          v          |       /
1669 //         Bool        |      /
1670 //           \         |     /
1671 //             \       |    /
1672 //               \     |   /
1673 //                 \   |  /
1674 //                   \ v /
1675 //                   CMoveD
1676 //
1677 
1678 void SuperWord::merge_packs_to_cmovd() {
1679   for (int i = _packset.length() - 1; i >= 0; i--) {
1680     _cmovev_kit.make_cmovevd_pack(_packset.at(i));
1681   }
1682   #ifndef PRODUCT
1683     if (TraceSuperWord) {
1684       tty->print_cr("\nSuperWord::merge_packs_to_cmovd(): After merge");
1685       print_packset();
1686       tty->cr();
1687     }
1688   #endif
1689 }
1690 
1691 Node* CMoveKit::is_Bool_candidate(Node* def) const {
1692   Node* use = NULL;
1693   if (!def->is_Bool() || def->in(0) != NULL || def->outcnt() != 1) {
1694     return NULL;
1695   }
1696   for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1697     use = def->fast_out(j);
1698     if (!_sw->same_generation(def, use) || !use->is_CMove()) {
1699       return NULL;
1700     }
1701   }
1702   return use;
1703 }
1704 
1705 Node* CMoveKit::is_CmpD_candidate(Node* def) const {
1706   Node* use = NULL;
1707   if (!def->is_Cmp() || def->in(0) != NULL || def->outcnt() != 1) {
1708     return NULL;
1709   }
1710   for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1711     use = def->fast_out(j);
1712     if (!_sw->same_generation(def, use) || (use = is_Bool_candidate(use)) == NULL || !_sw->same_generation(def, use)) {
1713       return NULL;
1714     }
1715   }
1716   return use;
1717 }
1718 
1719 Node_List* CMoveKit::make_cmovevd_pack(Node_List* cmovd_pk) {
1720   Node *cmovd = cmovd_pk->at(0);
1721   if (!cmovd->is_CMove()) {
1722     return NULL;
1723   }
1724   if (cmovd->Opcode() != Op_CMoveF && cmovd->Opcode() != Op_CMoveD) {
1725     return NULL;
1726   }
1727   if (pack(cmovd) != NULL) { // already in the cmov pack
1728     return NULL;
1729   }
1730   if (cmovd->in(0) != NULL) {
1731     NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: CMoveD %d has control flow, escaping...", cmovd->_idx); cmovd->dump();})
1732     return NULL;
1733   }
1734 
1735   Node* bol = cmovd->as_CMove()->in(CMoveNode::Condition);
1736   if (!bol->is_Bool()
1737       || bol->outcnt() != 1
1738       || !_sw->same_generation(bol, cmovd)
1739       || bol->in(0) != NULL  // BoolNode has control flow!!
1740       || _sw->my_pack(bol) == NULL) {
1741       NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: Bool %d does not fit CMoveD %d for building vector, escaping...", bol->_idx, cmovd->_idx); bol->dump();})
1742       return NULL;
1743   }
1744   Node_List* bool_pk = _sw->my_pack(bol);
1745   if (bool_pk->size() != cmovd_pk->size() ) {
1746     return NULL;
1747   }
1748 
1749   Node* cmpd = bol->in(1);
1750   if (!cmpd->is_Cmp()
1751       || cmpd->outcnt() != 1
1752       || !_sw->same_generation(cmpd, cmovd)
1753       || cmpd->in(0) != NULL  // CmpDNode has control flow!!
1754       || _sw->my_pack(cmpd) == NULL) {
1755       NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: CmpD %d does not fit CMoveD %d for building vector, escaping...", cmpd->_idx, cmovd->_idx); cmpd->dump();})
1756       return NULL;
1757   }
1758   Node_List* cmpd_pk = _sw->my_pack(cmpd);
1759   if (cmpd_pk->size() != cmovd_pk->size() ) {
1760     return NULL;
1761   }
1762 
1763   if (!test_cmpd_pack(cmpd_pk, cmovd_pk)) {
1764     NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: cmpd pack for CmpD %d failed vectorization test", cmpd->_idx); cmpd->dump();})
1765     return NULL;
1766   }
1767 
1768   Node_List* new_cmpd_pk = new Node_List();
1769   uint sz = cmovd_pk->size() - 1;
1770   for (uint i = 0; i <= sz; ++i) {
1771     Node* cmov = cmovd_pk->at(i);
1772     Node* bol  = bool_pk->at(i);
1773     Node* cmp  = cmpd_pk->at(i);
1774 
1775     new_cmpd_pk->insert(i, cmov);
1776 
1777     map(cmov, new_cmpd_pk);
1778     map(bol, new_cmpd_pk);
1779     map(cmp, new_cmpd_pk);
1780 
1781     _sw->set_my_pack(cmov, new_cmpd_pk); // and keep old packs for cmp and bool
1782   }
1783   _sw->_packset.remove(cmovd_pk);
1784   _sw->_packset.remove(bool_pk);
1785   _sw->_packset.remove(cmpd_pk);
1786   _sw->_packset.append(new_cmpd_pk);
1787   NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print_cr("CMoveKit::make_cmovevd_pack: added syntactic CMoveD pack"); _sw->print_pack(new_cmpd_pk);})
1788   return new_cmpd_pk;
1789 }
1790 
1791 bool CMoveKit::test_cmpd_pack(Node_List* cmpd_pk, Node_List* cmovd_pk) {
1792   Node* cmpd0 = cmpd_pk->at(0);
1793   assert(cmpd0->is_Cmp(), "CMoveKit::test_cmpd_pack: should be CmpDNode");
1794   assert(cmovd_pk->at(0)->is_CMove(), "CMoveKit::test_cmpd_pack: should be CMoveD");
1795   assert(cmpd_pk->size() == cmovd_pk->size(), "CMoveKit::test_cmpd_pack: should be same size");
1796   Node* in1 = cmpd0->in(1);
1797   Node* in2 = cmpd0->in(2);
1798   Node_List* in1_pk = _sw->my_pack(in1);
1799   Node_List* in2_pk = _sw->my_pack(in2);
1800 
1801   if (  (in1_pk != NULL && in1_pk->size() != cmpd_pk->size())
1802      || (in2_pk != NULL && in2_pk->size() != cmpd_pk->size()) ) {
1803     return false;
1804   }
1805 
1806   // test if "all" in1 are in the same pack or the same node
1807   if (in1_pk == NULL) {
1808     for (uint j = 1; j < cmpd_pk->size(); j++) {
1809       if (cmpd_pk->at(j)->in(1) != in1) {
1810         return false;
1811       }
1812     }//for: in1_pk is not pack but all CmpD nodes in the pack have the same in(1)
1813   }
1814   // test if "all" in2 are in the same pack or the same node
1815   if (in2_pk == NULL) {
1816     for (uint j = 1; j < cmpd_pk->size(); j++) {
1817       if (cmpd_pk->at(j)->in(2) != in2) {
1818         return false;
1819       }
1820     }//for: in2_pk is not pack but all CmpD nodes in the pack have the same in(2)
1821   }
1822   //now check if cmpd_pk may be subsumed in vector built for cmovd_pk
1823   int cmovd_ind1, cmovd_ind2;
1824   if (cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1825    && cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1826       cmovd_ind1 = CMoveNode::IfFalse;
1827       cmovd_ind2 = CMoveNode::IfTrue;
1828   } else if (cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1829           && cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1830       cmovd_ind2 = CMoveNode::IfFalse;
1831       cmovd_ind1 = CMoveNode::IfTrue;
1832   }
1833   else {
1834     return false;
1835   }
1836 
1837   for (uint j = 1; j < cmpd_pk->size(); j++) {
1838     if (cmpd_pk->at(j)->in(1) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind1)
1839         || cmpd_pk->at(j)->in(2) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind2)) {
1840         return false;
1841     }//if
1842   }
1843   NOT_PRODUCT(if(_sw->is_trace_cmov()) { tty->print("CMoveKit::test_cmpd_pack: cmpd pack for 1st CmpD %d is OK for vectorization: ", cmpd0->_idx); cmpd0->dump(); })
1844   return true;
1845 }
1846 
1847 //------------------------------implemented---------------------------
1848 // Can code be generated for pack p?
1849 bool SuperWord::implemented(Node_List* p) {
1850   bool retValue = false;
1851   Node* p0 = p->at(0);
1852   if (p0 != NULL) {
1853     int opc = p0->Opcode();
1854     uint size = p->size();
1855     if (p0->is_reduction()) {
1856       const Type *arith_type = p0->bottom_type();
1857       // Length 2 reductions of INT/LONG do not offer performance benefits
1858       if (((arith_type->basic_type() == T_INT) || (arith_type->basic_type() == T_LONG)) && (size == 2)) {
1859         retValue = false;
1860       } else {
1861         retValue = ReductionNode::implemented(opc, size, arith_type->basic_type());
1862       }
1863     } else {
1864       retValue = VectorNode::implemented(opc, size, velt_basic_type(p0));
1865     }
1866     if (!retValue) {
1867       if (is_cmov_pack(p)) {
1868         NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::implemented: found cmpd pack"); print_pack(p);})
1869         return true;
1870       }
1871     }
1872   }
1873   return retValue;
1874 }
1875 
1876 bool SuperWord::is_cmov_pack(Node_List* p) {
1877   return _cmovev_kit.pack(p->at(0)) != NULL;
1878 }
1879 //------------------------------same_inputs--------------------------
1880 // For pack p, are all idx operands the same?
1881 bool SuperWord::same_inputs(Node_List* p, int idx) {
1882   Node* p0 = p->at(0);
1883   uint vlen = p->size();
1884   Node* p0_def = p0->in(idx);
1885   for (uint i = 1; i < vlen; i++) {
1886     Node* pi = p->at(i);
1887     Node* pi_def = pi->in(idx);
1888     if (p0_def != pi_def) {
1889       return false;
1890     }
1891   }
1892   return true;
1893 }
1894 
1895 //------------------------------profitable---------------------------
1896 // For pack p, are all operands and all uses (with in the block) vector?
1897 bool SuperWord::profitable(Node_List* p) {
1898   Node* p0 = p->at(0);
1899   uint start, end;
1900   VectorNode::vector_operands(p0, &start, &end);
1901 
1902   // Return false if some inputs are not vectors or vectors with different
1903   // size or alignment.
1904   // Also, for now, return false if not scalar promotion case when inputs are
1905   // the same. Later, implement PackNode and allow differing, non-vector inputs
1906   // (maybe just the ones from outside the block.)
1907   for (uint i = start; i < end; i++) {
1908     if (!is_vector_use(p0, i)) {
1909       return false;
1910     }
1911   }
1912   // Check if reductions are connected
1913   if (p0->is_reduction()) {
1914     Node* second_in = p0->in(2);
1915     Node_List* second_pk = my_pack(second_in);
1916     if ((second_pk == NULL) || (_num_work_vecs == _num_reductions)) {
1917       // Remove reduction flag if no parent pack or if not enough work
1918       // to cover reduction expansion overhead
1919       p0->remove_flag(Node::Flag_is_reduction);
1920       return false;
1921     } else if (second_pk->size() != p->size()) {
1922       return false;
1923     }
1924   }
1925   if (VectorNode::is_shift(p0)) {
1926     // For now, return false if shift count is vector or not scalar promotion
1927     // case (different shift counts) because it is not supported yet.
1928     Node* cnt = p0->in(2);
1929     Node_List* cnt_pk = my_pack(cnt);
1930     if (cnt_pk != NULL)
1931       return false;
1932     if (!same_inputs(p, 2))
1933       return false;
1934   }
1935   if (!p0->is_Store()) {
1936     // For now, return false if not all uses are vector.
1937     // Later, implement ExtractNode and allow non-vector uses (maybe
1938     // just the ones outside the block.)
1939     for (uint i = 0; i < p->size(); i++) {
1940       Node* def = p->at(i);
1941       if (is_cmov_pack_internal_node(p, def)) {
1942         continue;
1943       }
1944       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1945         Node* use = def->fast_out(j);
1946         for (uint k = 0; k < use->req(); k++) {
1947           Node* n = use->in(k);
1948           if (def == n) {
1949             // Reductions should only have a Phi use at the loop head or a non-phi use
1950             // outside of the loop if it is the last element of the pack (e.g. SafePoint).
1951             if (def->is_reduction() &&
1952                 ((use->is_Phi() && use->in(0) == _lpt->_head) ||
1953                  (!_lpt->is_member(_phase->get_loop(_phase->ctrl_or_self(use))) && i == p->size()-1))) {
1954               continue;
1955             }
1956             if (!is_vector_use(use, k)) {
1957               return false;
1958             }
1959           }
1960         }
1961       }
1962     }
1963   }
1964   return true;
1965 }
1966 
1967 //------------------------------schedule---------------------------
1968 // Adjust the memory graph for the packed operations
1969 void SuperWord::schedule() {
1970 
1971   // Co-locate in the memory graph the members of each memory pack
1972   for (int i = 0; i < _packset.length(); i++) {
1973     co_locate_pack(_packset.at(i));
1974   }
1975 }
1976 
1977 //-------------------------------remove_and_insert-------------------
1978 // Remove "current" from its current position in the memory graph and insert
1979 // it after the appropriate insertion point (lip or uip).
1980 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1981                                   Node *uip, Unique_Node_List &sched_before) {
1982   Node* my_mem = current->in(MemNode::Memory);
1983   bool sched_up = sched_before.member(current);
1984 
1985   // remove current_store from its current position in the memmory graph
1986   for (DUIterator i = current->outs(); current->has_out(i); i++) {
1987     Node* use = current->out(i);
1988     if (use->is_Mem()) {
1989       assert(use->in(MemNode::Memory) == current, "must be");
1990       if (use == prev) { // connect prev to my_mem
1991           _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1992           --i; //deleted this edge; rescan position
1993       } else if (sched_before.member(use)) {
1994         if (!sched_up) { // Will be moved together with current
1995           _igvn.replace_input_of(use, MemNode::Memory, uip);
1996           --i; //deleted this edge; rescan position
1997         }
1998       } else {
1999         if (sched_up) { // Will be moved together with current
2000           _igvn.replace_input_of(use, MemNode::Memory, lip);
2001           --i; //deleted this edge; rescan position
2002         }
2003       }
2004     }
2005   }
2006 
2007   Node *insert_pt =  sched_up ?  uip : lip;
2008 
2009   // all uses of insert_pt's memory state should use current's instead
2010   for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
2011     Node* use = insert_pt->out(i);
2012     if (use->is_Mem()) {
2013       assert(use->in(MemNode::Memory) == insert_pt, "must be");
2014       _igvn.replace_input_of(use, MemNode::Memory, current);
2015       --i; //deleted this edge; rescan position
2016     } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
2017       uint pos; //lip (lower insert point) must be the last one in the memory slice
2018       for (pos=1; pos < use->req(); pos++) {
2019         if (use->in(pos) == insert_pt) break;
2020       }
2021       _igvn.replace_input_of(use, pos, current);
2022       --i;
2023     }
2024   }
2025 
2026   //connect current to insert_pt
2027   _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
2028 }
2029 
2030 //------------------------------co_locate_pack----------------------------------
2031 // To schedule a store pack, we need to move any sandwiched memory ops either before
2032 // or after the pack, based upon dependence information:
2033 // (1) If any store in the pack depends on the sandwiched memory op, the
2034 //     sandwiched memory op must be scheduled BEFORE the pack;
2035 // (2) If a sandwiched memory op depends on any store in the pack, the
2036 //     sandwiched memory op must be scheduled AFTER the pack;
2037 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
2038 //     memory op (say memB), memB must be scheduled before memA. So, if memA is
2039 //     scheduled before the pack, memB must also be scheduled before the pack;
2040 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
2041 //     schedule this store AFTER the pack
2042 // (5) We know there is no dependence cycle, so there in no other case;
2043 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
2044 //
2045 // To schedule a load pack, we use the memory state of either the first or the last load in
2046 // the pack, based on the dependence constraint.
2047 void SuperWord::co_locate_pack(Node_List* pk) {
2048   if (pk->at(0)->is_Store()) {
2049     MemNode* first     = executed_first(pk)->as_Mem();
2050     MemNode* last      = executed_last(pk)->as_Mem();
2051     Unique_Node_List schedule_before_pack;
2052     Unique_Node_List memops;
2053 
2054     MemNode* current   = last->in(MemNode::Memory)->as_Mem();
2055     MemNode* previous  = last;
2056     while (true) {
2057       assert(in_bb(current), "stay in block");
2058       memops.push(previous);
2059       for (DUIterator i = current->outs(); current->has_out(i); i++) {
2060         Node* use = current->out(i);
2061         if (use->is_Mem() && use != previous)
2062           memops.push(use);
2063       }
2064       if (current == first) break;
2065       previous = current;
2066       current  = current->in(MemNode::Memory)->as_Mem();
2067     }
2068 
2069     // determine which memory operations should be scheduled before the pack
2070     for (uint i = 1; i < memops.size(); i++) {
2071       Node *s1 = memops.at(i);
2072       if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
2073         for (uint j = 0; j< i; j++) {
2074           Node *s2 = memops.at(j);
2075           if (!independent(s1, s2)) {
2076             if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
2077               schedule_before_pack.push(s1); // s1 must be scheduled before
2078               Node_List* mem_pk = my_pack(s1);
2079               if (mem_pk != NULL) {
2080                 for (uint ii = 0; ii < mem_pk->size(); ii++) {
2081                   Node* s = mem_pk->at(ii);  // follow partner
2082                   if (memops.member(s) && !schedule_before_pack.member(s))
2083                     schedule_before_pack.push(s);
2084                 }
2085               }
2086               break;
2087             }
2088           }
2089         }
2090       }
2091     }
2092 
2093     Node*    upper_insert_pt = first->in(MemNode::Memory);
2094     // Following code moves loads connected to upper_insert_pt below aliased stores.
2095     // Collect such loads here and reconnect them back to upper_insert_pt later.
2096     memops.clear();
2097     for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
2098       Node* use = upper_insert_pt->out(i);
2099       if (use->is_Mem() && !use->is_Store()) {
2100         memops.push(use);
2101       }
2102     }
2103 
2104     MemNode* lower_insert_pt = last;
2105     previous                 = last; //previous store in pk
2106     current                  = last->in(MemNode::Memory)->as_Mem();
2107 
2108     // start scheduling from "last" to "first"
2109     while (true) {
2110       assert(in_bb(current), "stay in block");
2111       assert(in_pack(previous, pk), "previous stays in pack");
2112       Node* my_mem = current->in(MemNode::Memory);
2113 
2114       if (in_pack(current, pk)) {
2115         // Forward users of my memory state (except "previous) to my input memory state
2116         for (DUIterator i = current->outs(); current->has_out(i); i++) {
2117           Node* use = current->out(i);
2118           if (use->is_Mem() && use != previous) {
2119             assert(use->in(MemNode::Memory) == current, "must be");
2120             if (schedule_before_pack.member(use)) {
2121               _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
2122             } else {
2123               _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
2124             }
2125             --i; // deleted this edge; rescan position
2126           }
2127         }
2128         previous = current;
2129       } else { // !in_pack(current, pk) ==> a sandwiched store
2130         remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
2131       }
2132 
2133       if (current == first) break;
2134       current = my_mem->as_Mem();
2135     } // end while
2136 
2137     // Reconnect loads back to upper_insert_pt.
2138     for (uint i = 0; i < memops.size(); i++) {
2139       Node *ld = memops.at(i);
2140       if (ld->in(MemNode::Memory) != upper_insert_pt) {
2141         _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
2142       }
2143     }
2144   } else if (pk->at(0)->is_Load()) { //load
2145     // all loads in the pack should have the same memory state. By default,
2146     // we use the memory state of the last load. However, if any load could
2147     // not be moved down due to the dependence constraint, we use the memory
2148     // state of the first load.
2149     Node* first_mem = pk->at(0)->in(MemNode::Memory);
2150     Node* last_mem = first_mem;
2151     for (uint i = 1; i < pk->size(); i++) {
2152       Node* ld = pk->at(i);
2153       Node* mem = ld->in(MemNode::Memory);
2154       assert(in_bb(first_mem) || in_bb(mem) || mem == first_mem, "2 different memory state from outside the loop?");
2155       if (in_bb(mem)) {
2156         if (in_bb(first_mem) && bb_idx(mem) < bb_idx(first_mem)) {
2157           first_mem = mem;
2158         }
2159         if (!in_bb(last_mem) || bb_idx(mem) > bb_idx(last_mem)) {
2160           last_mem = mem;
2161         }
2162       }
2163     }
2164     bool schedule_last = true;
2165     for (uint i = 0; i < pk->size(); i++) {
2166       Node* ld = pk->at(i);
2167       for (Node* current = last_mem; current != ld->in(MemNode::Memory);
2168            current=current->in(MemNode::Memory)) {
2169         assert(current != first_mem, "corrupted memory graph");
2170         if(current->is_Mem() && !independent(current, ld)){
2171           schedule_last = false; // a later store depends on this load
2172           break;
2173         }
2174       }
2175     }
2176 
2177     Node* mem_input = schedule_last ? last_mem : first_mem;
2178     _igvn.hash_delete(mem_input);
2179     // Give each load the same memory state
2180     for (uint i = 0; i < pk->size(); i++) {
2181       LoadNode* ld = pk->at(i)->as_Load();
2182       _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
2183     }
2184   }
2185 }
2186 
2187 #ifndef PRODUCT
2188 void SuperWord::print_loop(bool whole) {
2189   Node_Stack stack(_arena, _phase->C->unique() >> 2);
2190   Node_List rpo_list;
2191   VectorSet visited(_arena);
2192   visited.set(lpt()->_head->_idx);
2193   _phase->rpo(lpt()->_head, stack, visited, rpo_list);
2194   _phase->dump(lpt(), rpo_list.size(), rpo_list );
2195   if(whole) {
2196     tty->print_cr("\n Whole loop tree");
2197     _phase->dump();
2198     tty->print_cr(" End of whole loop tree\n");
2199   }
2200 }
2201 #endif
2202 
2203 //------------------------------output---------------------------
2204 // Convert packs into vector node operations
2205 void SuperWord::output() {
2206   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2207   Compile* C = _phase->C;
2208   if (_packset.length() == 0) {
2209     if (cl->is_main_loop()) {
2210       // Instigate more unrolling for optimization when vectorization fails.
2211       C->set_major_progress();
2212       cl->set_notpassed_slp();
2213       cl->mark_do_unroll_only();
2214     }
2215     return;
2216   }
2217 
2218 #ifndef PRODUCT
2219   if (TraceLoopOpts) {
2220     tty->print("SuperWord::output    ");
2221     lpt()->dump_head();
2222   }
2223 #endif
2224 
2225   if (cl->is_main_loop()) {
2226     // MUST ENSURE main loop's initial value is properly aligned:
2227     //  (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
2228 
2229     align_initial_loop_index(align_to_ref());
2230 
2231     // Insert extract (unpack) operations for scalar uses
2232     for (int i = 0; i < _packset.length(); i++) {
2233       insert_extracts(_packset.at(i));
2234     }
2235   }
2236 
2237   uint max_vlen_in_bytes = 0;
2238   uint max_vlen = 0;
2239   bool can_process_post_loop = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
2240 
2241   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop before create_reserve_version_of_loop"); print_loop(true);})
2242 
2243   CountedLoopReserveKit make_reversable(_phase, _lpt, do_reserve_copy());
2244 
2245   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop after create_reserve_version_of_loop"); print_loop(true);})
2246 
2247   if (do_reserve_copy() && !make_reversable.has_reserved()) {
2248     NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: loop was not reserved correctly, exiting SuperWord");})
2249     return;
2250   }
2251 
2252   for (int i = 0; i < _block.length(); i++) {
2253     Node* n = _block.at(i);
2254     Node_List* p = my_pack(n);
2255     if (p && n == executed_last(p)) {
2256       uint vlen = p->size();
2257       uint vlen_in_bytes = 0;
2258       Node* vn = NULL;
2259       Node* low_adr = p->at(0);
2260       Node* first   = executed_first(p);
2261       if (can_process_post_loop) {
2262         // override vlen with the main loops vector length
2263         vlen = cl->slp_max_unroll();
2264       }
2265       NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::output: %d executed first, %d executed last in pack", first->_idx, n->_idx); print_pack(p);})
2266       int   opc = n->Opcode();
2267       if (n->is_Load()) {
2268         Node* ctl = n->in(MemNode::Control);
2269         Node* mem = first->in(MemNode::Memory);
2270         SWPointer p1(n->as_Mem(), this, NULL, false);
2271         // Identify the memory dependency for the new loadVector node by
2272         // walking up through memory chain.
2273         // This is done to give flexibility to the new loadVector node so that
2274         // it can move above independent storeVector nodes.
2275         while (mem->is_StoreVector()) {
2276           SWPointer p2(mem->as_Mem(), this, NULL, false);
2277           int cmp = p1.cmp(p2);
2278           if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
2279             mem = mem->in(MemNode::Memory);
2280           } else {
2281             break; // dependent memory
2282           }
2283         }
2284         Node* adr = low_adr->in(MemNode::Address);
2285         const TypePtr* atyp = n->adr_type();
2286         vn = LoadVectorNode::make(opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n), control_dependency(p));
2287         vlen_in_bytes = vn->as_LoadVector()->memory_size();
2288       } else if (n->is_Store()) {
2289         // Promote value to be stored to vector
2290         Node* val = vector_opd(p, MemNode::ValueIn);
2291         if (val == NULL) {
2292           if (do_reserve_copy()) {
2293             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: val should not be NULL, exiting SuperWord");})
2294             return; //and reverse to backup IG
2295           }
2296           ShouldNotReachHere();
2297         }
2298 
2299         Node* ctl = n->in(MemNode::Control);
2300         Node* mem = first->in(MemNode::Memory);
2301         Node* adr = low_adr->in(MemNode::Address);
2302         const TypePtr* atyp = n->adr_type();
2303         vn = StoreVectorNode::make(opc, ctl, mem, adr, atyp, val, vlen);
2304         vlen_in_bytes = vn->as_StoreVector()->memory_size();
2305       } else if (n->req() == 3 && !is_cmov_pack(p)) {
2306         // Promote operands to vector
2307         Node* in1 = NULL;
2308         bool node_isa_reduction = n->is_reduction();
2309         if (node_isa_reduction) {
2310           // the input to the first reduction operation is retained
2311           in1 = low_adr->in(1);
2312         } else {
2313           in1 = vector_opd(p, 1);
2314           if (in1 == NULL) {
2315             if (do_reserve_copy()) {
2316               NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in1 should not be NULL, exiting SuperWord");})
2317               return; //and reverse to backup IG
2318             }
2319             ShouldNotReachHere();
2320           }
2321         }
2322         Node* in2 = vector_opd(p, 2);
2323         if (in2 == NULL) {
2324           if (do_reserve_copy()) {
2325             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in2 should not be NULL, exiting SuperWord");})
2326             return; //and reverse to backup IG
2327           }
2328           ShouldNotReachHere();
2329         }
2330         if (VectorNode::is_invariant_vector(in1) && (node_isa_reduction == false) && (n->is_Add() || n->is_Mul())) {
2331           // Move invariant vector input into second position to avoid register spilling.
2332           Node* tmp = in1;
2333           in1 = in2;
2334           in2 = tmp;
2335         }
2336         if (node_isa_reduction) {
2337           const Type *arith_type = n->bottom_type();
2338           vn = ReductionNode::make(opc, NULL, in1, in2, arith_type->basic_type());
2339           if (in2->is_Load()) {
2340             vlen_in_bytes = in2->as_LoadVector()->memory_size();
2341           } else {
2342             vlen_in_bytes = in2->as_Vector()->length_in_bytes();
2343           }
2344         } else {
2345           vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2346           vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2347         }
2348       } else if (opc == Op_SqrtF || opc == Op_SqrtD ||
2349                  opc == Op_AbsF || opc == Op_AbsD ||
2350                  opc == Op_NegF || opc == Op_NegD ||
2351                  opc == Op_PopCountI) {
2352         assert(n->req() == 2, "only one input expected");
2353         Node* in = vector_opd(p, 1);
2354         vn = VectorNode::make(opc, in, NULL, vlen, velt_basic_type(n));
2355         vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2356       } else if (is_cmov_pack(p)) {
2357         if (can_process_post_loop) {
2358           // do not refactor of flow in post loop context
2359           return;
2360         }
2361         if (!n->is_CMove()) {
2362           continue;
2363         }
2364         // place here CMoveVDNode
2365         NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::output: print before CMove vectorization"); print_loop(false);})
2366         Node* bol = n->in(CMoveNode::Condition);
2367         if (!bol->is_Bool() && bol->Opcode() == Op_ExtractI && bol->req() > 1 ) {
2368           NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::output: %d is not Bool node, trying its in(1) node %d", bol->_idx, bol->in(1)->_idx); bol->dump(); bol->in(1)->dump();})
2369           bol = bol->in(1); //may be ExtractNode
2370         }
2371 
2372         assert(bol->is_Bool(), "should be BoolNode - too late to bail out!");
2373         if (!bol->is_Bool()) {
2374           if (do_reserve_copy()) {
2375             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: expected %d bool node, exiting SuperWord", bol->_idx); bol->dump();})
2376             return; //and reverse to backup IG
2377           }
2378           ShouldNotReachHere();
2379         }
2380 
2381         int cond = (int)bol->as_Bool()->_test._test;
2382         Node* in_cc  = _igvn.intcon(cond);
2383         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created intcon in_cc node %d", in_cc->_idx); in_cc->dump();})
2384         Node* cc = bol->clone();
2385         cc->set_req(1, in_cc);
2386         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created bool cc node %d", cc->_idx); cc->dump();})
2387 
2388         Node* src1 = vector_opd(p, 2); //2=CMoveNode::IfFalse
2389         if (src1 == NULL) {
2390           if (do_reserve_copy()) {
2391             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src1 should not be NULL, exiting SuperWord");})
2392             return; //and reverse to backup IG
2393           }
2394           ShouldNotReachHere();
2395         }
2396         Node* src2 = vector_opd(p, 3); //3=CMoveNode::IfTrue
2397         if (src2 == NULL) {
2398           if (do_reserve_copy()) {
2399             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src2 should not be NULL, exiting SuperWord");})
2400             return; //and reverse to backup IG
2401           }
2402           ShouldNotReachHere();
2403         }
2404         BasicType bt = velt_basic_type(n);
2405         const TypeVect* vt = TypeVect::make(bt, vlen);
2406         assert(bt == T_FLOAT || bt == T_DOUBLE, "Only vectorization for FP cmovs is supported");
2407         if (bt == T_FLOAT) {
2408           vn = new CMoveVFNode(cc, src1, src2, vt);
2409         } else {
2410           assert(bt == T_DOUBLE, "Expected double");
2411           vn = new CMoveVDNode(cc, src1, src2, vt);
2412         }
2413         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created new CMove node %d: ", vn->_idx); vn->dump();})
2414       } else if (opc == Op_FmaD || opc == Op_FmaF) {
2415         // Promote operands to vector
2416         Node* in1 = vector_opd(p, 1);
2417         Node* in2 = vector_opd(p, 2);
2418         Node* in3 = vector_opd(p, 3);
2419         vn = VectorNode::make(opc, in1, in2, in3, vlen, velt_basic_type(n));
2420         vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2421       } else {
2422         if (do_reserve_copy()) {
2423           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: ShouldNotReachHere, exiting SuperWord");})
2424           return; //and reverse to backup IG
2425         }
2426         ShouldNotReachHere();
2427       }
2428 
2429       assert(vn != NULL, "sanity");
2430       if (vn == NULL) {
2431         if (do_reserve_copy()){
2432           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: got NULL node, cannot proceed, exiting SuperWord");})
2433           return; //and reverse to backup IG
2434         }
2435         ShouldNotReachHere();
2436       }
2437 
2438       _block.at_put(i, vn);
2439       _igvn.register_new_node_with_optimizer(vn);
2440       _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
2441       for (uint j = 0; j < p->size(); j++) {
2442         Node* pm = p->at(j);
2443         _igvn.replace_node(pm, vn);
2444       }
2445       _igvn._worklist.push(vn);
2446 
2447       if (can_process_post_loop) {
2448         // first check if the vector size if the maximum vector which we can use on the machine,
2449         // other vector size have reduced values for predicated data mapping.
2450         if (vlen_in_bytes != (uint)MaxVectorSize) {
2451           return;
2452         }
2453       }
2454 
2455       if (vlen_in_bytes >= max_vlen_in_bytes && vlen > max_vlen) {
2456         max_vlen = vlen;
2457         max_vlen_in_bytes = vlen_in_bytes;
2458       }
2459 #ifdef ASSERT
2460       if (TraceNewVectors) {
2461         tty->print("new Vector node: ");
2462         vn->dump();
2463       }
2464 #endif
2465     }
2466   }//for (int i = 0; i < _block.length(); i++)
2467 
2468   if (max_vlen_in_bytes > C->max_vector_size()) {
2469     C->set_max_vector_size(max_vlen_in_bytes);
2470   }
2471   if (max_vlen_in_bytes > 0) {
2472     cl->mark_loop_vectorized();
2473   }
2474 
2475   if (SuperWordLoopUnrollAnalysis) {
2476     if (cl->has_passed_slp()) {
2477       uint slp_max_unroll_factor = cl->slp_max_unroll();
2478       if (slp_max_unroll_factor == max_vlen) {
2479         if (TraceSuperWordLoopUnrollAnalysis) {
2480           tty->print_cr("vector loop(unroll=%d, len=%d)\n", max_vlen, max_vlen_in_bytes*BitsPerByte);
2481         }
2482 
2483         // For atomic unrolled loops which are vector mapped, instigate more unrolling
2484         cl->set_notpassed_slp();
2485         if (cl->is_main_loop()) {
2486           // if vector resources are limited, do not allow additional unrolling, also
2487           // do not unroll more on pure vector loops which were not reduced so that we can
2488           // program the post loop to single iteration execution.
2489           if (FLOATPRESSURE > 8) {
2490             C->set_major_progress();
2491             cl->mark_do_unroll_only();
2492           }
2493         }
2494 
2495         if (do_reserve_copy()) {
2496           if (can_process_post_loop) {
2497             // Now create the difference of trip and limit and use it as our mask index.
2498             // Note: We limited the unroll of the vectorized loop so that
2499             //       only vlen-1 size iterations can remain to be mask programmed.
2500             Node *incr = cl->incr();
2501             SubINode *index = new SubINode(cl->limit(), cl->init_trip());
2502             _igvn.register_new_node_with_optimizer(index);
2503             SetVectMaskINode  *mask = new SetVectMaskINode(_phase->get_ctrl(cl->init_trip()), index);
2504             _igvn.register_new_node_with_optimizer(mask);
2505             // make this a single iteration loop
2506             AddINode *new_incr = new AddINode(incr->in(1), mask);
2507             _igvn.register_new_node_with_optimizer(new_incr);
2508             _phase->set_ctrl(new_incr, _phase->get_ctrl(incr));
2509             _igvn.replace_node(incr, new_incr);
2510             cl->mark_is_multiversioned();
2511             cl->loopexit()->add_flag(Node::Flag_has_vector_mask_set);
2512           }
2513         }
2514       }
2515     }
2516   }
2517 
2518   if (do_reserve_copy()) {
2519     make_reversable.use_new();
2520   }
2521   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("\n Final loop after SuperWord"); print_loop(true);})
2522   return;
2523 }
2524 
2525 //------------------------------vector_opd---------------------------
2526 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
2527 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
2528   Node* p0 = p->at(0);
2529   uint vlen = p->size();
2530   Node* opd = p0->in(opd_idx);
2531   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2532 
2533   if (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop()) {
2534     // override vlen with the main loops vector length
2535     vlen = cl->slp_max_unroll();
2536   }
2537 
2538   if (same_inputs(p, opd_idx)) {
2539     if (opd->is_Vector() || opd->is_LoadVector()) {
2540       assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
2541       if (opd_idx == 2 && VectorNode::is_shift(p0)) {
2542         NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("shift's count can't be vector");})
2543         return NULL;
2544       }
2545       return opd; // input is matching vector
2546     }
2547     if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
2548       Compile* C = _phase->C;
2549       Node* cnt = opd;
2550       // Vector instructions do not mask shift count, do it here.
2551       juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2552       const TypeInt* t = opd->find_int_type();
2553       if (t != NULL && t->is_con()) {
2554         juint shift = t->get_con();
2555         if (shift > mask) { // Unsigned cmp
2556           cnt = ConNode::make(TypeInt::make(shift & mask));
2557         }
2558       } else {
2559         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2560           cnt = ConNode::make(TypeInt::make(mask));
2561           _igvn.register_new_node_with_optimizer(cnt);
2562           cnt = new AndINode(opd, cnt);
2563           _igvn.register_new_node_with_optimizer(cnt);
2564           _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2565         }
2566         assert(opd->bottom_type()->isa_int(), "int type only");
2567         if (!opd->bottom_type()->isa_int()) {
2568           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should be int type only");})
2569           return NULL;
2570         }
2571         // Move non constant shift count into vector register.
2572         cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0));
2573       }
2574       if (cnt != opd) {
2575         _igvn.register_new_node_with_optimizer(cnt);
2576         _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2577       }
2578       return cnt;
2579     }
2580     assert(!opd->is_StoreVector(), "such vector is not expected here");
2581     if (opd->is_StoreVector()) {
2582       NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("StoreVector is not expected here");})
2583       return NULL;
2584     }
2585     // Convert scalar input to vector with the same number of elements as
2586     // p0's vector. Use p0's type because size of operand's container in
2587     // vector should match p0's size regardless operand's size.
2588     const Type* p0_t = velt_type(p0);
2589     VectorNode* vn = VectorNode::scalar2vector(opd, vlen, p0_t);
2590 
2591     _igvn.register_new_node_with_optimizer(vn);
2592     _phase->set_ctrl(vn, _phase->get_ctrl(opd));
2593 #ifdef ASSERT
2594     if (TraceNewVectors) {
2595       tty->print("new Vector node: ");
2596       vn->dump();
2597     }
2598 #endif
2599     return vn;
2600   }
2601 
2602   // Insert pack operation
2603   BasicType bt = velt_basic_type(p0);
2604   PackNode* pk = PackNode::make(opd, vlen, bt);
2605   DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
2606 
2607   for (uint i = 1; i < vlen; i++) {
2608     Node* pi = p->at(i);
2609     Node* in = pi->in(opd_idx);
2610     assert(my_pack(in) == NULL, "Should already have been unpacked");
2611     if (my_pack(in) != NULL) {
2612       NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should already have been unpacked");})
2613       return NULL;
2614     }
2615     assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
2616     pk->add_opd(in);
2617   }
2618   _igvn.register_new_node_with_optimizer(pk);
2619   _phase->set_ctrl(pk, _phase->get_ctrl(opd));
2620 #ifdef ASSERT
2621   if (TraceNewVectors) {
2622     tty->print("new Vector node: ");
2623     pk->dump();
2624   }
2625 #endif
2626   return pk;
2627 }
2628 
2629 //------------------------------insert_extracts---------------------------
2630 // If a use of pack p is not a vector use, then replace the
2631 // use with an extract operation.
2632 void SuperWord::insert_extracts(Node_List* p) {
2633   if (p->at(0)->is_Store()) return;
2634   assert(_n_idx_list.is_empty(), "empty (node,index) list");
2635 
2636   // Inspect each use of each pack member.  For each use that is
2637   // not a vector use, replace the use with an extract operation.
2638 
2639   for (uint i = 0; i < p->size(); i++) {
2640     Node* def = p->at(i);
2641     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2642       Node* use = def->fast_out(j);
2643       for (uint k = 0; k < use->req(); k++) {
2644         Node* n = use->in(k);
2645         if (def == n) {
2646           Node_List* u_pk = my_pack(use);
2647           if ((u_pk == NULL || !is_cmov_pack(u_pk) || use->is_CMove()) && !is_vector_use(use, k)) {
2648               _n_idx_list.push(use, k);
2649           }
2650         }
2651       }
2652     }
2653   }
2654 
2655   while (_n_idx_list.is_nonempty()) {
2656     Node* use = _n_idx_list.node();
2657     int   idx = _n_idx_list.index();
2658     _n_idx_list.pop();
2659     Node* def = use->in(idx);
2660 
2661     if (def->is_reduction()) continue;
2662 
2663     // Insert extract operation
2664     _igvn.hash_delete(def);
2665     int def_pos = alignment(def) / data_size(def);
2666 
2667     Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def));
2668     _igvn.register_new_node_with_optimizer(ex);
2669     _phase->set_ctrl(ex, _phase->get_ctrl(def));
2670     _igvn.replace_input_of(use, idx, ex);
2671     _igvn._worklist.push(def);
2672 
2673     bb_insert_after(ex, bb_idx(def));
2674     set_velt_type(ex, velt_type(def));
2675   }
2676 }
2677 
2678 //------------------------------is_vector_use---------------------------
2679 // Is use->in(u_idx) a vector use?
2680 bool SuperWord::is_vector_use(Node* use, int u_idx) {
2681   Node_List* u_pk = my_pack(use);
2682   if (u_pk == NULL) return false;
2683   if (use->is_reduction()) return true;
2684   Node* def = use->in(u_idx);
2685   Node_List* d_pk = my_pack(def);
2686   if (d_pk == NULL) {
2687     // check for scalar promotion
2688     Node* n = u_pk->at(0)->in(u_idx);
2689     for (uint i = 1; i < u_pk->size(); i++) {
2690       if (u_pk->at(i)->in(u_idx) != n) return false;
2691     }
2692     return true;
2693   }
2694   if (u_pk->size() != d_pk->size())
2695     return false;
2696   for (uint i = 0; i < u_pk->size(); i++) {
2697     Node* ui = u_pk->at(i);
2698     Node* di = d_pk->at(i);
2699     if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
2700       return false;
2701   }
2702   return true;
2703 }
2704 
2705 //------------------------------construct_bb---------------------------
2706 // Construct reverse postorder list of block members
2707 bool SuperWord::construct_bb() {
2708   Node* entry = bb();
2709 
2710   assert(_stk.length() == 0,            "stk is empty");
2711   assert(_block.length() == 0,          "block is empty");
2712   assert(_data_entry.length() == 0,     "data_entry is empty");
2713   assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
2714   assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
2715 
2716   // Find non-control nodes with no inputs from within block,
2717   // create a temporary map from node _idx to bb_idx for use
2718   // by the visited and post_visited sets,
2719   // and count number of nodes in block.
2720   int bb_ct = 0;
2721   for (uint i = 0; i < lpt()->_body.size(); i++) {
2722     Node *n = lpt()->_body.at(i);
2723     set_bb_idx(n, i); // Create a temporary map
2724     if (in_bb(n)) {
2725       if (n->is_LoadStore() || n->is_MergeMem() ||
2726           (n->is_Proj() && !n->as_Proj()->is_CFG())) {
2727         // Bailout if the loop has LoadStore, MergeMem or data Proj
2728         // nodes. Superword optimization does not work with them.
2729         return false;
2730       }
2731       bb_ct++;
2732       if (!n->is_CFG()) {
2733         bool found = false;
2734         for (uint j = 0; j < n->req(); j++) {
2735           Node* def = n->in(j);
2736           if (def && in_bb(def)) {
2737             found = true;
2738             break;
2739           }
2740         }
2741         if (!found) {
2742           assert(n != entry, "can't be entry");
2743           _data_entry.push(n);
2744         }
2745       }
2746     }
2747   }
2748 
2749   // Find memory slices (head and tail)
2750   for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
2751     Node *n = lp()->fast_out(i);
2752     if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
2753       Node* n_tail  = n->in(LoopNode::LoopBackControl);
2754       if (n_tail != n->in(LoopNode::EntryControl)) {
2755         if (!n_tail->is_Mem()) {
2756           assert(n_tail->is_Mem(), "unexpected node for memory slice: %s", n_tail->Name());
2757           return false; // Bailout
2758         }
2759         _mem_slice_head.push(n);
2760         _mem_slice_tail.push(n_tail);
2761       }
2762     }
2763   }
2764 
2765   // Create an RPO list of nodes in block
2766 
2767   visited_clear();
2768   post_visited_clear();
2769 
2770   // Push all non-control nodes with no inputs from within block, then control entry
2771   for (int j = 0; j < _data_entry.length(); j++) {
2772     Node* n = _data_entry.at(j);
2773     visited_set(n);
2774     _stk.push(n);
2775   }
2776   visited_set(entry);
2777   _stk.push(entry);
2778 
2779   // Do a depth first walk over out edges
2780   int rpo_idx = bb_ct - 1;
2781   int size;
2782   int reduction_uses = 0;
2783   while ((size = _stk.length()) > 0) {
2784     Node* n = _stk.top(); // Leave node on stack
2785     if (!visited_test_set(n)) {
2786       // forward arc in graph
2787     } else if (!post_visited_test(n)) {
2788       // cross or back arc
2789       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2790         Node *use = n->fast_out(i);
2791         if (in_bb(use) && !visited_test(use) &&
2792             // Don't go around backedge
2793             (!use->is_Phi() || n == entry)) {
2794           if (use->is_reduction()) {
2795             // First see if we can map the reduction on the given system we are on, then
2796             // make a data entry operation for each reduction we see.
2797             BasicType bt = use->bottom_type()->basic_type();
2798             if (ReductionNode::implemented(use->Opcode(), Matcher::min_vector_size(bt), bt)) {
2799               reduction_uses++;
2800             }
2801           }
2802           _stk.push(use);
2803         }
2804       }
2805       if (_stk.length() == size) {
2806         // There were no additional uses, post visit node now
2807         _stk.pop(); // Remove node from stack
2808         assert(rpo_idx >= 0, "");
2809         _block.at_put_grow(rpo_idx, n);
2810         rpo_idx--;
2811         post_visited_set(n);
2812         assert(rpo_idx >= 0 || _stk.is_empty(), "");
2813       }
2814     } else {
2815       _stk.pop(); // Remove post-visited node from stack
2816     }
2817   }//while
2818 
2819   int ii_current = -1;
2820   unsigned int load_idx = (unsigned int)-1;
2821   _ii_order.clear();
2822   // Create real map of block indices for nodes
2823   for (int j = 0; j < _block.length(); j++) {
2824     Node* n = _block.at(j);
2825     set_bb_idx(n, j);
2826     if (_do_vector_loop && n->is_Load()) {
2827       if (ii_current == -1) {
2828         ii_current = _clone_map.gen(n->_idx);
2829         _ii_order.push(ii_current);
2830         load_idx = _clone_map.idx(n->_idx);
2831       } else if (_clone_map.idx(n->_idx) == load_idx && _clone_map.gen(n->_idx) != ii_current) {
2832         ii_current = _clone_map.gen(n->_idx);
2833         _ii_order.push(ii_current);
2834       }
2835     }
2836   }//for
2837 
2838   // Ensure extra info is allocated.
2839   initialize_bb();
2840 
2841 #ifndef PRODUCT
2842   if (_vector_loop_debug && _ii_order.length() > 0) {
2843     tty->print("SuperWord::construct_bb: List of generations: ");
2844     for (int jj = 0; jj < _ii_order.length(); ++jj) {
2845       tty->print("  %d:%d", jj, _ii_order.at(jj));
2846     }
2847     tty->print_cr(" ");
2848   }
2849   if (TraceSuperWord) {
2850     print_bb();
2851     tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
2852     for (int m = 0; m < _data_entry.length(); m++) {
2853       tty->print("%3d ", m);
2854       _data_entry.at(m)->dump();
2855     }
2856     tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
2857     for (int m = 0; m < _mem_slice_head.length(); m++) {
2858       tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
2859       tty->print("    ");    _mem_slice_tail.at(m)->dump();
2860     }
2861   }
2862 #endif
2863   assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
2864   return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0);
2865 }
2866 
2867 //------------------------------initialize_bb---------------------------
2868 // Initialize per node info
2869 void SuperWord::initialize_bb() {
2870   Node* last = _block.at(_block.length() - 1);
2871   grow_node_info(bb_idx(last));
2872 }
2873 
2874 //------------------------------bb_insert_after---------------------------
2875 // Insert n into block after pos
2876 void SuperWord::bb_insert_after(Node* n, int pos) {
2877   int n_pos = pos + 1;
2878   // Make room
2879   for (int i = _block.length() - 1; i >= n_pos; i--) {
2880     _block.at_put_grow(i+1, _block.at(i));
2881   }
2882   for (int j = _node_info.length() - 1; j >= n_pos; j--) {
2883     _node_info.at_put_grow(j+1, _node_info.at(j));
2884   }
2885   // Set value
2886   _block.at_put_grow(n_pos, n);
2887   _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
2888   // Adjust map from node->_idx to _block index
2889   for (int i = n_pos; i < _block.length(); i++) {
2890     set_bb_idx(_block.at(i), i);
2891   }
2892 }
2893 
2894 //------------------------------compute_max_depth---------------------------
2895 // Compute max depth for expressions from beginning of block
2896 // Use to prune search paths during test for independence.
2897 void SuperWord::compute_max_depth() {
2898   int ct = 0;
2899   bool again;
2900   do {
2901     again = false;
2902     for (int i = 0; i < _block.length(); i++) {
2903       Node* n = _block.at(i);
2904       if (!n->is_Phi()) {
2905         int d_orig = depth(n);
2906         int d_in   = 0;
2907         for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
2908           Node* pred = preds.current();
2909           if (in_bb(pred)) {
2910             d_in = MAX2(d_in, depth(pred));
2911           }
2912         }
2913         if (d_in + 1 != d_orig) {
2914           set_depth(n, d_in + 1);
2915           again = true;
2916         }
2917       }
2918     }
2919     ct++;
2920   } while (again);
2921 
2922   if (TraceSuperWord && Verbose) {
2923     tty->print_cr("compute_max_depth iterated: %d times", ct);
2924   }
2925 }
2926 
2927 //-------------------------compute_vector_element_type-----------------------
2928 // Compute necessary vector element type for expressions
2929 // This propagates backwards a narrower integer type when the
2930 // upper bits of the value are not needed.
2931 // Example:  char a,b,c;  a = b + c;
2932 // Normally the type of the add is integer, but for packed character
2933 // operations the type of the add needs to be char.
2934 void SuperWord::compute_vector_element_type() {
2935   if (TraceSuperWord && Verbose) {
2936     tty->print_cr("\ncompute_velt_type:");
2937   }
2938 
2939   // Initial type
2940   for (int i = 0; i < _block.length(); i++) {
2941     Node* n = _block.at(i);
2942     set_velt_type(n, container_type(n));
2943   }
2944 
2945   // Propagate integer narrowed type backwards through operations
2946   // that don't depend on higher order bits
2947   for (int i = _block.length() - 1; i >= 0; i--) {
2948     Node* n = _block.at(i);
2949     // Only integer types need be examined
2950     const Type* vtn = velt_type(n);
2951     if (vtn->basic_type() == T_INT) {
2952       uint start, end;
2953       VectorNode::vector_operands(n, &start, &end);
2954 
2955       for (uint j = start; j < end; j++) {
2956         Node* in  = n->in(j);
2957         // Don't propagate through a memory
2958         if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
2959             data_size(n) < data_size(in)) {
2960           bool same_type = true;
2961           for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
2962             Node *use = in->fast_out(k);
2963             if (!in_bb(use) || !same_velt_type(use, n)) {
2964               same_type = false;
2965               break;
2966             }
2967           }
2968           if (same_type) {
2969             // For right shifts of small integer types (bool, byte, char, short)
2970             // we need precise information about sign-ness. Only Load nodes have
2971             // this information because Store nodes are the same for signed and
2972             // unsigned values. And any arithmetic operation after a load may
2973             // expand a value to signed Int so such right shifts can't be used
2974             // because vector elements do not have upper bits of Int.
2975             const Type* vt = vtn;
2976             if (VectorNode::is_shift(in)) {
2977               Node* load = in->in(1);
2978               if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
2979                 vt = velt_type(load);
2980               } else if (in->Opcode() != Op_LShiftI) {
2981                 // Widen type to Int to avoid creation of right shift vector
2982                 // (align + data_size(s1) check in stmts_can_pack() will fail).
2983                 // Note, left shifts work regardless type.
2984                 vt = TypeInt::INT;
2985               }
2986             }
2987             set_velt_type(in, vt);
2988           }
2989         }
2990       }
2991     }
2992   }
2993 #ifndef PRODUCT
2994   if (TraceSuperWord && Verbose) {
2995     for (int i = 0; i < _block.length(); i++) {
2996       Node* n = _block.at(i);
2997       velt_type(n)->dump();
2998       tty->print("\t");
2999       n->dump();
3000     }
3001   }
3002 #endif
3003 }
3004 
3005 //------------------------------memory_alignment---------------------------
3006 // Alignment within a vector memory reference
3007 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
3008   #ifndef PRODUCT
3009     if(TraceSuperWord && Verbose) {
3010       tty->print("SuperWord::memory_alignment within a vector memory reference for %d:  ", s->_idx); s->dump();
3011     }
3012   #endif
3013   NOT_PRODUCT(SWPointer::Tracer::Depth ddd(0);)
3014   SWPointer p(s, this, NULL, false);
3015   if (!p.valid()) {
3016     NOT_PRODUCT(if(is_trace_alignment()) tty->print("SWPointer::memory_alignment: SWPointer p invalid, return bottom_align");)
3017     return bottom_align;
3018   }
3019   int vw = vector_width_in_bytes(s);
3020   if (vw < 2) {
3021     NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SWPointer::memory_alignment: vector_width_in_bytes < 2, return bottom_align");)
3022     return bottom_align; // No vectors for this type
3023   }
3024   int offset  = p.offset_in_bytes();
3025   offset     += iv_adjust*p.memory_size();
3026   int off_rem = offset % vw;
3027   int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
3028   if (TraceSuperWord && Verbose) {
3029     tty->print_cr("SWPointer::memory_alignment: off_rem = %d, off_mod = %d", off_rem, off_mod);
3030   }
3031   return off_mod;
3032 }
3033 
3034 //---------------------------container_type---------------------------
3035 // Smallest type containing range of values
3036 const Type* SuperWord::container_type(Node* n) {
3037   if (n->is_Mem()) {
3038     BasicType bt = n->as_Mem()->memory_type();
3039     if (n->is_Store() && (bt == T_CHAR)) {
3040       // Use T_SHORT type instead of T_CHAR for stored values because any
3041       // preceding arithmetic operation extends values to signed Int.
3042       bt = T_SHORT;
3043     }
3044     if (n->Opcode() == Op_LoadUB) {
3045       // Adjust type for unsigned byte loads, it is important for right shifts.
3046       // T_BOOLEAN is used because there is no basic type representing type
3047       // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
3048       // size (one byte) and sign is important.
3049       bt = T_BOOLEAN;
3050     }
3051     return Type::get_const_basic_type(bt);
3052   }
3053   const Type* t = _igvn.type(n);
3054   if (t->basic_type() == T_INT) {
3055     // A narrow type of arithmetic operations will be determined by
3056     // propagating the type of memory operations.
3057     return TypeInt::INT;
3058   }
3059   return t;
3060 }
3061 
3062 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
3063   const Type* vt1 = velt_type(n1);
3064   const Type* vt2 = velt_type(n2);
3065   if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
3066     // Compare vectors element sizes for integer types.
3067     return data_size(n1) == data_size(n2);
3068   }
3069   return vt1 == vt2;
3070 }
3071 
3072 //------------------------------in_packset---------------------------
3073 // Are s1 and s2 in a pack pair and ordered as s1,s2?
3074 bool SuperWord::in_packset(Node* s1, Node* s2) {
3075   for (int i = 0; i < _packset.length(); i++) {
3076     Node_List* p = _packset.at(i);
3077     assert(p->size() == 2, "must be");
3078     if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
3079       return true;
3080     }
3081   }
3082   return false;
3083 }
3084 
3085 //------------------------------in_pack---------------------------
3086 // Is s in pack p?
3087 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
3088   for (uint i = 0; i < p->size(); i++) {
3089     if (p->at(i) == s) {
3090       return p;
3091     }
3092   }
3093   return NULL;
3094 }
3095 
3096 //------------------------------remove_pack_at---------------------------
3097 // Remove the pack at position pos in the packset
3098 void SuperWord::remove_pack_at(int pos) {
3099   Node_List* p = _packset.at(pos);
3100   for (uint i = 0; i < p->size(); i++) {
3101     Node* s = p->at(i);
3102     set_my_pack(s, NULL);
3103   }
3104   _packset.remove_at(pos);
3105 }
3106 
3107 void SuperWord::packset_sort(int n) {
3108   // simple bubble sort so that we capitalize with O(n) when its already sorted
3109   while (n != 0) {
3110     bool swapped = false;
3111     for (int i = 1; i < n; i++) {
3112       Node_List* q_low = _packset.at(i-1);
3113       Node_List* q_i = _packset.at(i);
3114 
3115       // only swap when we find something to swap
3116       if (alignment(q_low->at(0)) > alignment(q_i->at(0))) {
3117         Node_List* t = q_i;
3118         *(_packset.adr_at(i)) = q_low;
3119         *(_packset.adr_at(i-1)) = q_i;
3120         swapped = true;
3121       }
3122     }
3123     if (swapped == false) break;
3124     n--;
3125   }
3126 }
3127 
3128 //------------------------------executed_first---------------------------
3129 // Return the node executed first in pack p.  Uses the RPO block list
3130 // to determine order.
3131 Node* SuperWord::executed_first(Node_List* p) {
3132   Node* n = p->at(0);
3133   int n_rpo = bb_idx(n);
3134   for (uint i = 1; i < p->size(); i++) {
3135     Node* s = p->at(i);
3136     int s_rpo = bb_idx(s);
3137     if (s_rpo < n_rpo) {
3138       n = s;
3139       n_rpo = s_rpo;
3140     }
3141   }
3142   return n;
3143 }
3144 
3145 //------------------------------executed_last---------------------------
3146 // Return the node executed last in pack p.
3147 Node* SuperWord::executed_last(Node_List* p) {
3148   Node* n = p->at(0);
3149   int n_rpo = bb_idx(n);
3150   for (uint i = 1; i < p->size(); i++) {
3151     Node* s = p->at(i);
3152     int s_rpo = bb_idx(s);
3153     if (s_rpo > n_rpo) {
3154       n = s;
3155       n_rpo = s_rpo;
3156     }
3157   }
3158   return n;
3159 }
3160 
3161 LoadNode::ControlDependency SuperWord::control_dependency(Node_List* p) {
3162   LoadNode::ControlDependency dep = LoadNode::DependsOnlyOnTest;
3163   for (uint i = 0; i < p->size(); i++) {
3164     Node* n = p->at(i);
3165     assert(n->is_Load(), "only meaningful for loads");
3166     if (!n->depends_only_on_test()) {
3167       dep = LoadNode::Pinned;
3168     }
3169   }
3170   return dep;
3171 }
3172 
3173 
3174 //----------------------------align_initial_loop_index---------------------------
3175 // Adjust pre-loop limit so that in main loop, a load/store reference
3176 // to align_to_ref will be a position zero in the vector.
3177 //   (iv + k) mod vector_align == 0
3178 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
3179   CountedLoopNode *main_head = lp()->as_CountedLoop();
3180   assert(main_head->is_main_loop(), "");
3181   CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
3182   assert(pre_end != NULL, "we must have a correct pre-loop");
3183   Node *pre_opaq1 = pre_end->limit();
3184   assert(pre_opaq1->Opcode() == Op_Opaque1, "");
3185   Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
3186   Node *lim0 = pre_opaq->in(1);
3187 
3188   // Where we put new limit calculations
3189   Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
3190 
3191   // Ensure the original loop limit is available from the
3192   // pre-loop Opaque1 node.
3193   Node *orig_limit = pre_opaq->original_loop_limit();
3194   assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
3195 
3196   SWPointer align_to_ref_p(align_to_ref, this, NULL, false);
3197   assert(align_to_ref_p.valid(), "sanity");
3198 
3199   // Given:
3200   //     lim0 == original pre loop limit
3201   //     V == v_align (power of 2)
3202   //     invar == extra invariant piece of the address expression
3203   //     e == offset [ +/- invar ]
3204   //
3205   // When reassociating expressions involving '%' the basic rules are:
3206   //     (a - b) % k == 0   =>  a % k == b % k
3207   // and:
3208   //     (a + b) % k == 0   =>  a % k == (k - b) % k
3209   //
3210   // For stride > 0 && scale > 0,
3211   //   Derive the new pre-loop limit "lim" such that the two constraints:
3212   //     (1) lim = lim0 + N           (where N is some positive integer < V)
3213   //     (2) (e + lim) % V == 0
3214   //   are true.
3215   //
3216   //   Substituting (1) into (2),
3217   //     (e + lim0 + N) % V == 0
3218   //   solve for N:
3219   //     N = (V - (e + lim0)) % V
3220   //   substitute back into (1), so that new limit
3221   //     lim = lim0 + (V - (e + lim0)) % V
3222   //
3223   // For stride > 0 && scale < 0
3224   //   Constraints:
3225   //     lim = lim0 + N
3226   //     (e - lim) % V == 0
3227   //   Solving for lim:
3228   //     (e - lim0 - N) % V == 0
3229   //     N = (e - lim0) % V
3230   //     lim = lim0 + (e - lim0) % V
3231   //
3232   // For stride < 0 && scale > 0
3233   //   Constraints:
3234   //     lim = lim0 - N
3235   //     (e + lim) % V == 0
3236   //   Solving for lim:
3237   //     (e + lim0 - N) % V == 0
3238   //     N = (e + lim0) % V
3239   //     lim = lim0 - (e + lim0) % V
3240   //
3241   // For stride < 0 && scale < 0
3242   //   Constraints:
3243   //     lim = lim0 - N
3244   //     (e - lim) % V == 0
3245   //   Solving for lim:
3246   //     (e - lim0 + N) % V == 0
3247   //     N = (V - (e - lim0)) % V
3248   //     lim = lim0 - (V - (e - lim0)) % V
3249 
3250   int vw = vector_width_in_bytes(align_to_ref);
3251   int stride   = iv_stride();
3252   int scale    = align_to_ref_p.scale_in_bytes();
3253   int elt_size = align_to_ref_p.memory_size();
3254   int v_align  = vw / elt_size;
3255   assert(v_align > 1, "sanity");
3256   int offset   = align_to_ref_p.offset_in_bytes() / elt_size;
3257   Node *offsn  = _igvn.intcon(offset);
3258 
3259   Node *e = offsn;
3260   if (align_to_ref_p.invar() != NULL) {
3261     // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
3262     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3263     Node* invar = align_to_ref_p.invar();
3264     if (_igvn.type(invar)->isa_long()) {
3265       // Computations are done % (vector width/element size) so it's
3266       // safe to simply convert invar to an int and loose the upper 32
3267       // bit half.
3268       invar = new ConvL2INode(invar);
3269       _igvn.register_new_node_with_optimizer(invar);
3270     }
3271     Node* aref = new URShiftINode(invar, log2_elt);
3272     _igvn.register_new_node_with_optimizer(aref);
3273     _phase->set_ctrl(aref, pre_ctrl);
3274     if (align_to_ref_p.negate_invar()) {
3275       e = new SubINode(e, aref);
3276     } else {
3277       e = new AddINode(e, aref);
3278     }
3279     _igvn.register_new_node_with_optimizer(e);
3280     _phase->set_ctrl(e, pre_ctrl);
3281   }
3282   if (vw > ObjectAlignmentInBytes || align_to_ref_p.base()->is_top()) {
3283     // incorporate base e +/- base && Mask >>> log2(elt)
3284     Node* xbase = new CastP2XNode(NULL, align_to_ref_p.adr());
3285     _igvn.register_new_node_with_optimizer(xbase);
3286 #ifdef _LP64
3287     xbase  = new ConvL2INode(xbase);
3288     _igvn.register_new_node_with_optimizer(xbase);
3289 #endif
3290     Node* mask = _igvn.intcon(vw-1);
3291     Node* masked_xbase  = new AndINode(xbase, mask);
3292     _igvn.register_new_node_with_optimizer(masked_xbase);
3293     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3294     Node* bref     = new URShiftINode(masked_xbase, log2_elt);
3295     _igvn.register_new_node_with_optimizer(bref);
3296     _phase->set_ctrl(bref, pre_ctrl);
3297     e = new AddINode(e, bref);
3298     _igvn.register_new_node_with_optimizer(e);
3299     _phase->set_ctrl(e, pre_ctrl);
3300   }
3301 
3302   // compute e +/- lim0
3303   if (scale < 0) {
3304     e = new SubINode(e, lim0);
3305   } else {
3306     e = new AddINode(e, lim0);
3307   }
3308   _igvn.register_new_node_with_optimizer(e);
3309   _phase->set_ctrl(e, pre_ctrl);
3310 
3311   if (stride * scale > 0) {
3312     // compute V - (e +/- lim0)
3313     Node* va  = _igvn.intcon(v_align);
3314     e = new SubINode(va, e);
3315     _igvn.register_new_node_with_optimizer(e);
3316     _phase->set_ctrl(e, pre_ctrl);
3317   }
3318   // compute N = (exp) % V
3319   Node* va_msk = _igvn.intcon(v_align - 1);
3320   Node* N = new AndINode(e, va_msk);
3321   _igvn.register_new_node_with_optimizer(N);
3322   _phase->set_ctrl(N, pre_ctrl);
3323 
3324   //   substitute back into (1), so that new limit
3325   //     lim = lim0 + N
3326   Node* lim;
3327   if (stride < 0) {
3328     lim = new SubINode(lim0, N);
3329   } else {
3330     lim = new AddINode(lim0, N);
3331   }
3332   _igvn.register_new_node_with_optimizer(lim);
3333   _phase->set_ctrl(lim, pre_ctrl);
3334   Node* constrained =
3335     (stride > 0) ? (Node*) new MinINode(lim, orig_limit)
3336                  : (Node*) new MaxINode(lim, orig_limit);
3337   _igvn.register_new_node_with_optimizer(constrained);
3338   _phase->set_ctrl(constrained, pre_ctrl);
3339   _igvn.replace_input_of(pre_opaq, 1, constrained);
3340 }
3341 
3342 //----------------------------get_pre_loop_end---------------------------
3343 // Find pre loop end from main loop.  Returns null if none.
3344 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode* cl) {
3345   // The loop cannot be optimized if the graph shape at
3346   // the loop entry is inappropriate.
3347   if (!PhaseIdealLoop::is_canonical_loop_entry(cl)) {
3348     return NULL;
3349   }
3350 
3351   Node* p_f = cl->skip_predicates()->in(0)->in(0);
3352   if (!p_f->is_IfFalse()) return NULL;
3353   if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
3354   CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
3355   CountedLoopNode* loop_node = pre_end->loopnode();
3356   if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
3357   return pre_end;
3358 }
3359 
3360 //------------------------------init---------------------------
3361 void SuperWord::init() {
3362   _dg.init();
3363   _packset.clear();
3364   _disjoint_ptrs.clear();
3365   _block.clear();
3366   _post_block.clear();
3367   _data_entry.clear();
3368   _mem_slice_head.clear();
3369   _mem_slice_tail.clear();
3370   _iteration_first.clear();
3371   _iteration_last.clear();
3372   _node_info.clear();
3373   _align_to_ref = NULL;
3374   _lpt = NULL;
3375   _lp = NULL;
3376   _bb = NULL;
3377   _iv = NULL;
3378   _race_possible = 0;
3379   _early_return = false;
3380   _num_work_vecs = 0;
3381   _num_reductions = 0;
3382 }
3383 
3384 //------------------------------restart---------------------------
3385 void SuperWord::restart() {
3386   _dg.init();
3387   _packset.clear();
3388   _disjoint_ptrs.clear();
3389   _block.clear();
3390   _post_block.clear();
3391   _data_entry.clear();
3392   _mem_slice_head.clear();
3393   _mem_slice_tail.clear();
3394   _node_info.clear();
3395 }
3396 
3397 //------------------------------print_packset---------------------------
3398 void SuperWord::print_packset() {
3399 #ifndef PRODUCT
3400   tty->print_cr("packset");
3401   for (int i = 0; i < _packset.length(); i++) {
3402     tty->print_cr("Pack: %d", i);
3403     Node_List* p = _packset.at(i);
3404     print_pack(p);
3405   }
3406 #endif
3407 }
3408 
3409 //------------------------------print_pack---------------------------
3410 void SuperWord::print_pack(Node_List* p) {
3411   for (uint i = 0; i < p->size(); i++) {
3412     print_stmt(p->at(i));
3413   }
3414 }
3415 
3416 //------------------------------print_bb---------------------------
3417 void SuperWord::print_bb() {
3418 #ifndef PRODUCT
3419   tty->print_cr("\nBlock");
3420   for (int i = 0; i < _block.length(); i++) {
3421     Node* n = _block.at(i);
3422     tty->print("%d ", i);
3423     if (n) {
3424       n->dump();
3425     }
3426   }
3427 #endif
3428 }
3429 
3430 //------------------------------print_stmt---------------------------
3431 void SuperWord::print_stmt(Node* s) {
3432 #ifndef PRODUCT
3433   tty->print(" align: %d \t", alignment(s));
3434   s->dump();
3435 #endif
3436 }
3437 
3438 //------------------------------blank---------------------------
3439 char* SuperWord::blank(uint depth) {
3440   static char blanks[101];
3441   assert(depth < 101, "too deep");
3442   for (uint i = 0; i < depth; i++) blanks[i] = ' ';
3443   blanks[depth] = '\0';
3444   return blanks;
3445 }
3446 
3447 
3448 //==============================SWPointer===========================
3449 #ifndef PRODUCT
3450 int SWPointer::Tracer::_depth = 0;
3451 #endif
3452 //----------------------------SWPointer------------------------
3453 SWPointer::SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only) :
3454   _mem(mem), _slp(slp),  _base(NULL),  _adr(NULL),
3455   _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3456   _nstack(nstack), _analyze_only(analyze_only),
3457   _stack_idx(0)
3458 #ifndef PRODUCT
3459   , _tracer(slp)
3460 #endif
3461 {
3462   NOT_PRODUCT(_tracer.ctor_1(mem);)
3463 
3464   Node* adr = mem->in(MemNode::Address);
3465   if (!adr->is_AddP()) {
3466     assert(!valid(), "too complex");
3467     return;
3468   }
3469   // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
3470   Node* base = adr->in(AddPNode::Base);
3471   // The base address should be loop invariant
3472   if (!invariant(base)) {
3473     assert(!valid(), "base address is loop variant");
3474     return;
3475   }
3476   // unsafe references require misaligned vector access support
3477   if (base->is_top() && !Matcher::misaligned_vectors_ok()) {
3478     assert(!valid(), "unsafe access");
3479     return;
3480   }
3481 
3482   // Detect a Shenandoah write barrier between the pre and main loop
3483   // (which could break loop alignment code)
3484   if (UseShenandoahGC) {
3485     CountedLoopNode *main_head = slp->lp()->as_CountedLoop();
3486     if (main_head->is_main_loop()) {
3487       Node* c = main_head->skip_predicates()->in(0)->in(0)->in(0);
3488       if (!c->is_CountedLoopEnd()) {
3489         // in case of a reserve copy
3490         c = main_head->skip_strip_mined()->in(LoopNode::EntryControl)->in(0)->in(0);
3491         c = CountedLoopNode::skip_predicates_from_entry(c);
3492         c = c->in(0)->in(0)->in(0);
3493         assert(c->is_CountedLoopEnd(), "where's the pre loop?");
3494       }
3495       CountedLoopEndNode* pre_end = c->as_CountedLoopEnd();
3496       CountedLoopNode* pre_loop = pre_end->loopnode();
3497       assert(pre_loop->is_pre_loop(), "where's the pre loop?");
3498 
3499       Node* base_c = phase()->get_ctrl(base);
3500       if (!phase()->is_dominator(base_c, pre_loop)) {
3501         return;
3502       }
3503     }
3504   }
3505 
3506   NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.store_depth();)
3507   NOT_PRODUCT(_tracer.ctor_2(adr);)
3508 
3509   int i;
3510   for (i = 0; i < 3; i++) {
3511     NOT_PRODUCT(_tracer.ctor_3(adr, i);)
3512 
3513     if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
3514       assert(!valid(), "too complex");
3515       return;
3516     }
3517     adr = adr->in(AddPNode::Address);
3518     NOT_PRODUCT(_tracer.ctor_4(adr, i);)
3519 
3520     if (base == adr || !adr->is_AddP()) {
3521       NOT_PRODUCT(_tracer.ctor_5(adr, base, i);)
3522       break; // stop looking at addp's
3523     }
3524   }
3525   if (!invariant(adr)) {
3526     assert(!valid(), "adr is loop variant");
3527     return;
3528   }
3529 
3530   if (!base->is_top() && adr != base) {
3531     assert(!valid(), "adr and base differ");
3532     return;
3533   }
3534 
3535   NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.restore_depth();)
3536   NOT_PRODUCT(_tracer.ctor_6(mem);)
3537 
3538   _base = base;
3539   _adr  = adr;
3540   assert(valid(), "Usable");
3541 }
3542 
3543 // Following is used to create a temporary object during
3544 // the pattern match of an address expression.
3545 SWPointer::SWPointer(SWPointer* p) :
3546   _mem(p->_mem), _slp(p->_slp),  _base(NULL),  _adr(NULL),
3547   _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3548   _nstack(p->_nstack), _analyze_only(p->_analyze_only),
3549   _stack_idx(p->_stack_idx)
3550   #ifndef PRODUCT
3551   , _tracer(p->_slp)
3552   #endif
3553 {}
3554 
3555 
3556 bool SWPointer::invariant(Node* n) {
3557   NOT_PRODUCT(Tracer::Depth dd;)
3558   Node *n_c = phase()->get_ctrl(n);
3559   NOT_PRODUCT(_tracer.invariant_1(n, n_c);)
3560   return !lpt()->is_member(phase()->get_loop(n_c));
3561 }
3562 //------------------------scaled_iv_plus_offset--------------------
3563 // Match: k*iv + offset
3564 // where: k is a constant that maybe zero, and
3565 //        offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
3566 bool SWPointer::scaled_iv_plus_offset(Node* n) {
3567   NOT_PRODUCT(Tracer::Depth ddd;)
3568   NOT_PRODUCT(_tracer.scaled_iv_plus_offset_1(n);)
3569 
3570   if (scaled_iv(n)) {
3571     NOT_PRODUCT(_tracer.scaled_iv_plus_offset_2(n);)
3572     return true;
3573   }
3574 
3575   if (offset_plus_k(n)) {
3576     NOT_PRODUCT(_tracer.scaled_iv_plus_offset_3(n);)
3577     return true;
3578   }
3579 
3580   int opc = n->Opcode();
3581   if (opc == Op_AddI) {
3582     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
3583       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_4(n);)
3584       return true;
3585     }
3586     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3587       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_5(n);)
3588       return true;
3589     }
3590   } else if (opc == Op_SubI) {
3591     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
3592       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_6(n);)
3593       return true;
3594     }
3595     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3596       _scale *= -1;
3597       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_7(n);)
3598       return true;
3599     }
3600   }
3601 
3602   NOT_PRODUCT(_tracer.scaled_iv_plus_offset_8(n);)
3603   return false;
3604 }
3605 
3606 //----------------------------scaled_iv------------------------
3607 // Match: k*iv where k is a constant that's not zero
3608 bool SWPointer::scaled_iv(Node* n) {
3609   NOT_PRODUCT(Tracer::Depth ddd;)
3610   NOT_PRODUCT(_tracer.scaled_iv_1(n);)
3611 
3612   if (_scale != 0) { // already found a scale
3613     NOT_PRODUCT(_tracer.scaled_iv_2(n, _scale);)
3614     return false;
3615   }
3616 
3617   if (n == iv()) {
3618     _scale = 1;
3619     NOT_PRODUCT(_tracer.scaled_iv_3(n, _scale);)
3620     return true;
3621   }
3622   if (_analyze_only && (invariant(n) == false)) {
3623     _nstack->push(n, _stack_idx++);
3624   }
3625 
3626   int opc = n->Opcode();
3627   if (opc == Op_MulI) {
3628     if (n->in(1) == iv() && n->in(2)->is_Con()) {
3629       _scale = n->in(2)->get_int();
3630       NOT_PRODUCT(_tracer.scaled_iv_4(n, _scale);)
3631       return true;
3632     } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
3633       _scale = n->in(1)->get_int();
3634       NOT_PRODUCT(_tracer.scaled_iv_5(n, _scale);)
3635       return true;
3636     }
3637   } else if (opc == Op_LShiftI) {
3638     if (n->in(1) == iv() && n->in(2)->is_Con()) {
3639       _scale = 1 << n->in(2)->get_int();
3640       NOT_PRODUCT(_tracer.scaled_iv_6(n, _scale);)
3641       return true;
3642     }
3643   } else if (opc == Op_ConvI2L) {
3644     if (n->in(1)->Opcode() == Op_CastII &&
3645         n->in(1)->as_CastII()->has_range_check()) {
3646       // Skip range check dependent CastII nodes
3647       n = n->in(1);
3648     }
3649     if (scaled_iv_plus_offset(n->in(1))) {
3650       NOT_PRODUCT(_tracer.scaled_iv_7(n);)
3651       return true;
3652     }
3653   } else if (opc == Op_LShiftL) {
3654     if (!has_iv() && _invar == NULL) {
3655       // Need to preserve the current _offset value, so
3656       // create a temporary object for this expression subtree.
3657       // Hacky, so should re-engineer the address pattern match.
3658       NOT_PRODUCT(Tracer::Depth dddd;)
3659       SWPointer tmp(this);
3660       NOT_PRODUCT(_tracer.scaled_iv_8(n, &tmp);)
3661 
3662       if (tmp.scaled_iv_plus_offset(n->in(1))) {
3663         if (tmp._invar == NULL || _slp->do_vector_loop()) {
3664           int mult = 1 << n->in(2)->get_int();
3665           _scale   = tmp._scale  * mult;
3666           _offset += tmp._offset * mult;
3667           NOT_PRODUCT(_tracer.scaled_iv_9(n, _scale, _offset, mult);)
3668           return true;
3669         }
3670       }
3671     }
3672   }
3673   NOT_PRODUCT(_tracer.scaled_iv_10(n);)
3674   return false;
3675 }
3676 
3677 //----------------------------offset_plus_k------------------------
3678 // Match: offset is (k [+/- invariant])
3679 // where k maybe zero and invariant is optional, but not both.
3680 bool SWPointer::offset_plus_k(Node* n, bool negate) {
3681   NOT_PRODUCT(Tracer::Depth ddd;)
3682   NOT_PRODUCT(_tracer.offset_plus_k_1(n);)
3683 
3684   int opc = n->Opcode();
3685   if (opc == Op_ConI) {
3686     _offset += negate ? -(n->get_int()) : n->get_int();
3687     NOT_PRODUCT(_tracer.offset_plus_k_2(n, _offset);)
3688     return true;
3689   } else if (opc == Op_ConL) {
3690     // Okay if value fits into an int
3691     const TypeLong* t = n->find_long_type();
3692     if (t->higher_equal(TypeLong::INT)) {
3693       jlong loff = n->get_long();
3694       jint  off  = (jint)loff;
3695       _offset += negate ? -off : loff;
3696       NOT_PRODUCT(_tracer.offset_plus_k_3(n, _offset);)
3697       return true;
3698     }
3699     NOT_PRODUCT(_tracer.offset_plus_k_4(n);)
3700     return false;
3701   }
3702   if (_invar != NULL) { // already has an invariant
3703     NOT_PRODUCT(_tracer.offset_plus_k_5(n, _invar);)
3704     return false;
3705   }
3706 
3707   if (_analyze_only && (invariant(n) == false)) {
3708     _nstack->push(n, _stack_idx++);
3709   }
3710   if (opc == Op_AddI) {
3711     if (n->in(2)->is_Con() && invariant(n->in(1))) {
3712       _negate_invar = negate;
3713       _invar = n->in(1);
3714       _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3715       NOT_PRODUCT(_tracer.offset_plus_k_6(n, _invar, _negate_invar, _offset);)
3716       return true;
3717     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3718       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3719       _negate_invar = negate;
3720       _invar = n->in(2);
3721       NOT_PRODUCT(_tracer.offset_plus_k_7(n, _invar, _negate_invar, _offset);)
3722       return true;
3723     }
3724   }
3725   if (opc == Op_SubI) {
3726     if (n->in(2)->is_Con() && invariant(n->in(1))) {
3727       _negate_invar = negate;
3728       _invar = n->in(1);
3729       _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3730       NOT_PRODUCT(_tracer.offset_plus_k_8(n, _invar, _negate_invar, _offset);)
3731       return true;
3732     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3733       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3734       _negate_invar = !negate;
3735       _invar = n->in(2);
3736       NOT_PRODUCT(_tracer.offset_plus_k_9(n, _invar, _negate_invar, _offset);)
3737       return true;
3738     }
3739   }
3740   if (invariant(n)) {
3741     if (opc == Op_ConvI2L) {
3742       n = n->in(1);
3743       if (n->Opcode() == Op_CastII &&
3744           n->as_CastII()->has_range_check()) {
3745         // Skip range check dependent CastII nodes
3746         assert(invariant(n), "sanity");
3747         n = n->in(1);
3748       }
3749     }
3750     _negate_invar = negate;
3751     _invar = n;
3752     NOT_PRODUCT(_tracer.offset_plus_k_10(n, _invar, _negate_invar, _offset);)
3753     return true;
3754   }
3755 
3756   NOT_PRODUCT(_tracer.offset_plus_k_11(n);)
3757   return false;
3758 }
3759 
3760 //----------------------------print------------------------
3761 void SWPointer::print() {
3762 #ifndef PRODUCT
3763   tty->print("base: %d  adr: %d  scale: %d  offset: %d  invar: %c%d\n",
3764              _base != NULL ? _base->_idx : 0,
3765              _adr  != NULL ? _adr->_idx  : 0,
3766              _scale, _offset,
3767              _negate_invar?'-':'+',
3768              _invar != NULL ? _invar->_idx : 0);
3769 #endif
3770 }
3771 
3772 //----------------------------tracing------------------------
3773 #ifndef PRODUCT
3774 void SWPointer::Tracer::print_depth() {
3775   for (int ii = 0; ii<_depth; ++ii) tty->print("  ");
3776 }
3777 
3778 void SWPointer::Tracer::ctor_1 (Node* mem) {
3779   if(_slp->is_trace_alignment()) {
3780     print_depth(); tty->print(" %d SWPointer::SWPointer: start alignment analysis", mem->_idx); mem->dump();
3781   }
3782 }
3783 
3784 void SWPointer::Tracer::ctor_2(Node* adr) {
3785   if(_slp->is_trace_alignment()) {
3786     //store_depth();
3787     inc_depth();
3788     print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: ", adr->_idx); adr->dump();
3789     inc_depth();
3790     print_depth(); tty->print(" %d (base) SWPointer::SWPointer: ", adr->in(AddPNode::Base)->_idx); adr->in(AddPNode::Base)->dump();
3791   }
3792 }
3793 
3794 void SWPointer::Tracer::ctor_3(Node* adr, int i) {
3795   if(_slp->is_trace_alignment()) {
3796     inc_depth();
3797     Node* offset = adr->in(AddPNode::Offset);
3798     print_depth(); tty->print(" %d (offset) SWPointer::SWPointer: i = %d: ", offset->_idx, i); offset->dump();
3799   }
3800 }
3801 
3802 void SWPointer::Tracer::ctor_4(Node* adr, int i) {
3803   if(_slp->is_trace_alignment()) {
3804     inc_depth();
3805     print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: i = %d: ", adr->_idx, i); adr->dump();
3806   }
3807 }
3808 
3809 void SWPointer::Tracer::ctor_5(Node* adr, Node* base, int i) {
3810   if(_slp->is_trace_alignment()) {
3811     inc_depth();
3812     if (base == adr) {
3813       print_depth(); tty->print_cr("  \\ %d (adr) == %d (base) SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, base->_idx, i);
3814     } else if (!adr->is_AddP()) {
3815       print_depth(); tty->print_cr("  \\ %d (adr) is NOT Addp SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, i);
3816     }
3817   }
3818 }
3819 
3820 void SWPointer::Tracer::ctor_6(Node* mem) {
3821   if(_slp->is_trace_alignment()) {
3822     //restore_depth();
3823     print_depth(); tty->print_cr(" %d (adr) SWPointer::SWPointer: stop analysis", mem->_idx);
3824   }
3825 }
3826 
3827 void SWPointer::Tracer::invariant_1(Node *n, Node *n_c) {
3828   if (_slp->do_vector_loop() && _slp->is_debug() && _slp->_lpt->is_member(_slp->_phase->get_loop(n_c)) != (int)_slp->in_bb(n)) {
3829     int is_member =  _slp->_lpt->is_member(_slp->_phase->get_loop(n_c));
3830     int in_bb     =  _slp->in_bb(n);
3831     print_depth(); tty->print("  \\ ");  tty->print_cr(" %d SWPointer::invariant  conditions differ: n_c %d", n->_idx, n_c->_idx);
3832     print_depth(); tty->print("  \\ ");  tty->print_cr("is_member %d, in_bb %d", is_member, in_bb);
3833     print_depth(); tty->print("  \\ ");  n->dump();
3834     print_depth(); tty->print("  \\ ");  n_c->dump();
3835   }
3836 }
3837 
3838 void SWPointer::Tracer::scaled_iv_plus_offset_1(Node* n) {
3839   if(_slp->is_trace_alignment()) {
3840     print_depth(); tty->print(" %d SWPointer::scaled_iv_plus_offset testing node: ", n->_idx);
3841     n->dump();
3842   }
3843 }
3844 
3845 void SWPointer::Tracer::scaled_iv_plus_offset_2(Node* n) {
3846   if(_slp->is_trace_alignment()) {
3847     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
3848   }
3849 }
3850 
3851 void SWPointer::Tracer::scaled_iv_plus_offset_3(Node* n) {
3852   if(_slp->is_trace_alignment()) {
3853     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
3854   }
3855 }
3856 
3857 void SWPointer::Tracer::scaled_iv_plus_offset_4(Node* n) {
3858   if(_slp->is_trace_alignment()) {
3859     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
3860     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
3861     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
3862   }
3863 }
3864 
3865 void SWPointer::Tracer::scaled_iv_plus_offset_5(Node* n) {
3866   if(_slp->is_trace_alignment()) {
3867     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
3868     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
3869     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
3870   }
3871 }
3872 
3873 void SWPointer::Tracer::scaled_iv_plus_offset_6(Node* n) {
3874   if(_slp->is_trace_alignment()) {
3875     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
3876     print_depth(); tty->print("  \\  %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
3877     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
3878   }
3879 }
3880 
3881 void SWPointer::Tracer::scaled_iv_plus_offset_7(Node* n) {
3882   if(_slp->is_trace_alignment()) {
3883     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
3884     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
3885     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
3886   }
3887 }
3888 
3889 void SWPointer::Tracer::scaled_iv_plus_offset_8(Node* n) {
3890   if(_slp->is_trace_alignment()) {
3891     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: FAILED", n->_idx);
3892   }
3893 }
3894 
3895 void SWPointer::Tracer::scaled_iv_1(Node* n) {
3896   if(_slp->is_trace_alignment()) {
3897     print_depth(); tty->print(" %d SWPointer::scaled_iv: testing node: ", n->_idx); n->dump();
3898   }
3899 }
3900 
3901 void SWPointer::Tracer::scaled_iv_2(Node* n, int scale) {
3902   if(_slp->is_trace_alignment()) {
3903     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED since another _scale has been detected before", n->_idx);
3904     print_depth(); tty->print_cr("  \\ SWPointer::scaled_iv: _scale (%d) != 0", scale);
3905   }
3906 }
3907 
3908 void SWPointer::Tracer::scaled_iv_3(Node* n, int scale) {
3909   if(_slp->is_trace_alignment()) {
3910     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: is iv, setting _scale = %d", n->_idx, scale);
3911   }
3912 }
3913 
3914 void SWPointer::Tracer::scaled_iv_4(Node* n, int scale) {
3915   if(_slp->is_trace_alignment()) {
3916     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
3917     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
3918     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3919   }
3920 }
3921 
3922 void SWPointer::Tracer::scaled_iv_5(Node* n, int scale) {
3923   if(_slp->is_trace_alignment()) {
3924     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
3925     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is iv: ", n->in(2)->_idx); n->in(2)->dump();
3926     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3927   }
3928 }
3929 
3930 void SWPointer::Tracer::scaled_iv_6(Node* n, int scale) {
3931   if(_slp->is_trace_alignment()) {
3932     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftI PASSED, setting _scale = %d", n->_idx, scale);
3933     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
3934     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3935   }
3936 }
3937 
3938 void SWPointer::Tracer::scaled_iv_7(Node* n) {
3939   if(_slp->is_trace_alignment()) {
3940     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_ConvI2L PASSED", n->_idx);
3941     print_depth(); tty->print_cr("  \\ SWPointer::scaled_iv: in(1) %d is scaled_iv_plus_offset: ", n->in(1)->_idx);
3942     inc_depth(); inc_depth();
3943     print_depth(); n->in(1)->dump();
3944     dec_depth(); dec_depth();
3945   }
3946 }
3947 
3948 void SWPointer::Tracer::scaled_iv_8(Node* n, SWPointer* tmp) {
3949   if(_slp->is_trace_alignment()) {
3950     print_depth(); tty->print(" %d SWPointer::scaled_iv: Op_LShiftL, creating tmp SWPointer: ", n->_idx); tmp->print();
3951   }
3952 }
3953 
3954 void SWPointer::Tracer::scaled_iv_9(Node* n, int scale, int _offset, int mult) {
3955   if(_slp->is_trace_alignment()) {
3956     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftL PASSED, setting _scale = %d, _offset = %d", n->_idx, scale, _offset);
3957     print_depth(); tty->print_cr("  \\ SWPointer::scaled_iv: in(1) %d is scaled_iv_plus_offset, in(2) %d used to get mult = %d: _scale = %d, _offset = %d",
3958     n->in(1)->_idx, n->in(2)->_idx, mult, scale, _offset);
3959     inc_depth(); inc_depth();
3960     print_depth(); n->in(1)->dump();
3961     print_depth(); n->in(2)->dump();
3962     dec_depth(); dec_depth();
3963   }
3964 }
3965 
3966 void SWPointer::Tracer::scaled_iv_10(Node* n) {
3967   if(_slp->is_trace_alignment()) {
3968     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED", n->_idx);
3969   }
3970 }
3971 
3972 void SWPointer::Tracer::offset_plus_k_1(Node* n) {
3973   if(_slp->is_trace_alignment()) {
3974     print_depth(); tty->print(" %d SWPointer::offset_plus_k: testing node: ", n->_idx); n->dump();
3975   }
3976 }
3977 
3978 void SWPointer::Tracer::offset_plus_k_2(Node* n, int _offset) {
3979   if(_slp->is_trace_alignment()) {
3980     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConI PASSED, setting _offset = %d", n->_idx, _offset);
3981   }
3982 }
3983 
3984 void SWPointer::Tracer::offset_plus_k_3(Node* n, int _offset) {
3985   if(_slp->is_trace_alignment()) {
3986     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConL PASSED, setting _offset = %d", n->_idx, _offset);
3987   }
3988 }
3989 
3990 void SWPointer::Tracer::offset_plus_k_4(Node* n) {
3991   if(_slp->is_trace_alignment()) {
3992     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
3993     print_depth(); tty->print_cr("  \\ " JLONG_FORMAT " SWPointer::offset_plus_k: Op_ConL FAILED, k is too big", n->get_long());
3994   }
3995 }
3996 
3997 void SWPointer::Tracer::offset_plus_k_5(Node* n, Node* _invar) {
3998   if(_slp->is_trace_alignment()) {
3999     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED since another invariant has been detected before", n->_idx);
4000     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: _invar != NULL: ", _invar->_idx); _invar->dump();
4001   }
4002 }
4003 
4004 void SWPointer::Tracer::offset_plus_k_6(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4005   if(_slp->is_trace_alignment()) {
4006     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
4007     n->_idx, _negate_invar, _invar->_idx, _offset);
4008     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4009     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
4010   }
4011 }
4012 
4013 void SWPointer::Tracer::offset_plus_k_7(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4014   if(_slp->is_trace_alignment()) {
4015     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
4016     n->_idx, _negate_invar, _invar->_idx, _offset);
4017     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
4018     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
4019   }
4020 }
4021 
4022 void SWPointer::Tracer::offset_plus_k_8(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4023   if(_slp->is_trace_alignment()) {
4024     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_SubI is PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
4025     n->_idx, _negate_invar, _invar->_idx, _offset);
4026     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4027     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
4028   }
4029 }
4030 
4031 void SWPointer::Tracer::offset_plus_k_9(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4032   if(_slp->is_trace_alignment()) {
4033     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_SubI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d", n->_idx, _negate_invar, _invar->_idx, _offset);
4034     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
4035     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
4036   }
4037 }
4038 
4039 void SWPointer::Tracer::offset_plus_k_10(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4040   if(_slp->is_trace_alignment()) {
4041     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d", n->_idx, _negate_invar, _invar->_idx, _offset);
4042     print_depth(); tty->print_cr("  \\ %d SWPointer::offset_plus_k: is invariant", n->_idx);
4043   }
4044 }
4045 
4046 void SWPointer::Tracer::offset_plus_k_11(Node* n) {
4047   if(_slp->is_trace_alignment()) {
4048     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
4049   }
4050 }
4051 
4052 #endif
4053 // ========================= OrderedPair =====================
4054 
4055 const OrderedPair OrderedPair::initial;
4056 
4057 // ========================= SWNodeInfo =====================
4058 
4059 const SWNodeInfo SWNodeInfo::initial;
4060 
4061 
4062 // ============================ DepGraph ===========================
4063 
4064 //------------------------------make_node---------------------------
4065 // Make a new dependence graph node for an ideal node.
4066 DepMem* DepGraph::make_node(Node* node) {
4067   DepMem* m = new (_arena) DepMem(node);
4068   if (node != NULL) {
4069     assert(_map.at_grow(node->_idx) == NULL, "one init only");
4070     _map.at_put_grow(node->_idx, m);
4071   }
4072   return m;
4073 }
4074 
4075 //------------------------------make_edge---------------------------
4076 // Make a new dependence graph edge from dpred -> dsucc
4077 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
4078   DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
4079   dpred->set_out_head(e);
4080   dsucc->set_in_head(e);
4081   return e;
4082 }
4083 
4084 // ========================== DepMem ========================
4085 
4086 //------------------------------in_cnt---------------------------
4087 int DepMem::in_cnt() {
4088   int ct = 0;
4089   for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
4090   return ct;
4091 }
4092 
4093 //------------------------------out_cnt---------------------------
4094 int DepMem::out_cnt() {
4095   int ct = 0;
4096   for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
4097   return ct;
4098 }
4099 
4100 //------------------------------print-----------------------------
4101 void DepMem::print() {
4102 #ifndef PRODUCT
4103   tty->print("  DepNode %d (", _node->_idx);
4104   for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
4105     Node* pred = p->pred()->node();
4106     tty->print(" %d", pred != NULL ? pred->_idx : 0);
4107   }
4108   tty->print(") [");
4109   for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
4110     Node* succ = s->succ()->node();
4111     tty->print(" %d", succ != NULL ? succ->_idx : 0);
4112   }
4113   tty->print_cr(" ]");
4114 #endif
4115 }
4116 
4117 // =========================== DepEdge =========================
4118 
4119 //------------------------------DepPreds---------------------------
4120 void DepEdge::print() {
4121 #ifndef PRODUCT
4122   tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
4123 #endif
4124 }
4125 
4126 // =========================== DepPreds =========================
4127 // Iterator over predecessor edges in the dependence graph.
4128 
4129 //------------------------------DepPreds---------------------------
4130 DepPreds::DepPreds(Node* n, DepGraph& dg) {
4131   _n = n;
4132   _done = false;
4133   if (_n->is_Store() || _n->is_Load()) {
4134     _next_idx = MemNode::Address;
4135     _end_idx  = n->req();
4136     _dep_next = dg.dep(_n)->in_head();
4137   } else if (_n->is_Mem()) {
4138     _next_idx = 0;
4139     _end_idx  = 0;
4140     _dep_next = dg.dep(_n)->in_head();
4141   } else {
4142     _next_idx = 1;
4143     _end_idx  = _n->req();
4144     _dep_next = NULL;
4145   }
4146   next();
4147 }
4148 
4149 //------------------------------next---------------------------
4150 void DepPreds::next() {
4151   if (_dep_next != NULL) {
4152     _current  = _dep_next->pred()->node();
4153     _dep_next = _dep_next->next_in();
4154   } else if (_next_idx < _end_idx) {
4155     _current  = _n->in(_next_idx++);
4156   } else {
4157     _done = true;
4158   }
4159 }
4160 
4161 // =========================== DepSuccs =========================
4162 // Iterator over successor edges in the dependence graph.
4163 
4164 //------------------------------DepSuccs---------------------------
4165 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
4166   _n = n;
4167   _done = false;
4168   if (_n->is_Load()) {
4169     _next_idx = 0;
4170     _end_idx  = _n->outcnt();
4171     _dep_next = dg.dep(_n)->out_head();
4172   } else if (_n->is_Mem() || (_n->is_Phi() && _n->bottom_type() == Type::MEMORY)) {
4173     _next_idx = 0;
4174     _end_idx  = 0;
4175     _dep_next = dg.dep(_n)->out_head();
4176   } else {
4177     _next_idx = 0;
4178     _end_idx  = _n->outcnt();
4179     _dep_next = NULL;
4180   }
4181   next();
4182 }
4183 
4184 //-------------------------------next---------------------------
4185 void DepSuccs::next() {
4186   if (_dep_next != NULL) {
4187     _current  = _dep_next->succ()->node();
4188     _dep_next = _dep_next->next_out();
4189   } else if (_next_idx < _end_idx) {
4190     _current  = _n->raw_out(_next_idx++);
4191   } else {
4192     _done = true;
4193   }
4194 }
4195 
4196 //
4197 // --------------------------------- vectorization/simd -----------------------------------
4198 //
4199 bool SuperWord::same_origin_idx(Node* a, Node* b) const {
4200   return a != NULL && b != NULL && _clone_map.same_idx(a->_idx, b->_idx);
4201 }
4202 bool SuperWord::same_generation(Node* a, Node* b) const {
4203   return a != NULL && b != NULL && _clone_map.same_gen(a->_idx, b->_idx);
4204 }
4205 
4206 Node*  SuperWord::find_phi_for_mem_dep(LoadNode* ld) {
4207   assert(in_bb(ld), "must be in block");
4208   if (_clone_map.gen(ld->_idx) == _ii_first) {
4209 #ifndef PRODUCT
4210     if (_vector_loop_debug) {
4211       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(ld->_idx)=%d",
4212         _clone_map.gen(ld->_idx));
4213     }
4214 #endif
4215     return NULL; //we think that any ld in the first gen being vectorizable
4216   }
4217 
4218   Node* mem = ld->in(MemNode::Memory);
4219   if (mem->outcnt() <= 1) {
4220     // we don't want to remove the only edge from mem node to load
4221 #ifndef PRODUCT
4222     if (_vector_loop_debug) {
4223       tty->print_cr("SuperWord::find_phi_for_mem_dep input node %d to load %d has no other outputs and edge mem->load cannot be removed",
4224         mem->_idx, ld->_idx);
4225       ld->dump();
4226       mem->dump();
4227     }
4228 #endif
4229     return NULL;
4230   }
4231   if (!in_bb(mem) || same_generation(mem, ld)) {
4232 #ifndef PRODUCT
4233     if (_vector_loop_debug) {
4234       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(mem->_idx)=%d",
4235         _clone_map.gen(mem->_idx));
4236     }
4237 #endif
4238     return NULL; // does not depend on loop volatile node or depends on the same generation
4239   }
4240 
4241   //otherwise first node should depend on mem-phi
4242   Node* first = first_node(ld);
4243   assert(first->is_Load(), "must be Load");
4244   Node* phi = first->as_Load()->in(MemNode::Memory);
4245   if (!phi->is_Phi() || phi->bottom_type() != Type::MEMORY) {
4246 #ifndef PRODUCT
4247     if (_vector_loop_debug) {
4248       tty->print_cr("SuperWord::find_phi_for_mem_dep load is not vectorizable node, since it's `first` does not take input from mem phi");
4249       ld->dump();
4250       first->dump();
4251     }
4252 #endif
4253     return NULL;
4254   }
4255 
4256   Node* tail = 0;
4257   for (int m = 0; m < _mem_slice_head.length(); m++) {
4258     if (_mem_slice_head.at(m) == phi) {
4259       tail = _mem_slice_tail.at(m);
4260     }
4261   }
4262   if (tail == 0) { //test that found phi is in the list  _mem_slice_head
4263 #ifndef PRODUCT
4264     if (_vector_loop_debug) {
4265       tty->print_cr("SuperWord::find_phi_for_mem_dep load %d is not vectorizable node, its phi %d is not _mem_slice_head",
4266         ld->_idx, phi->_idx);
4267       ld->dump();
4268       phi->dump();
4269     }
4270 #endif
4271     return NULL;
4272   }
4273 
4274   // now all conditions are met
4275   return phi;
4276 }
4277 
4278 Node* SuperWord::first_node(Node* nd) {
4279   for (int ii = 0; ii < _iteration_first.length(); ii++) {
4280     Node* nnn = _iteration_first.at(ii);
4281     if (same_origin_idx(nnn, nd)) {
4282 #ifndef PRODUCT
4283       if (_vector_loop_debug) {
4284         tty->print_cr("SuperWord::first_node: %d is the first iteration node for %d (_clone_map.idx(nnn->_idx) = %d)",
4285           nnn->_idx, nd->_idx, _clone_map.idx(nnn->_idx));
4286       }
4287 #endif
4288       return nnn;
4289     }
4290   }
4291 
4292 #ifndef PRODUCT
4293   if (_vector_loop_debug) {
4294     tty->print_cr("SuperWord::first_node: did not find first iteration node for %d (_clone_map.idx(nd->_idx)=%d)",
4295       nd->_idx, _clone_map.idx(nd->_idx));
4296   }
4297 #endif
4298   return 0;
4299 }
4300 
4301 Node* SuperWord::last_node(Node* nd) {
4302   for (int ii = 0; ii < _iteration_last.length(); ii++) {
4303     Node* nnn = _iteration_last.at(ii);
4304     if (same_origin_idx(nnn, nd)) {
4305 #ifndef PRODUCT
4306       if (_vector_loop_debug) {
4307         tty->print_cr("SuperWord::last_node _clone_map.idx(nnn->_idx)=%d, _clone_map.idx(nd->_idx)=%d",
4308           _clone_map.idx(nnn->_idx), _clone_map.idx(nd->_idx));
4309       }
4310 #endif
4311       return nnn;
4312     }
4313   }
4314   return 0;
4315 }
4316 
4317 int SuperWord::mark_generations() {
4318   Node *ii_err = NULL, *tail_err = NULL;
4319   for (int i = 0; i < _mem_slice_head.length(); i++) {
4320     Node* phi  = _mem_slice_head.at(i);
4321     assert(phi->is_Phi(), "must be phi");
4322 
4323     Node* tail = _mem_slice_tail.at(i);
4324     if (_ii_last == -1) {
4325       tail_err = tail;
4326       _ii_last = _clone_map.gen(tail->_idx);
4327     }
4328     else if (_ii_last != _clone_map.gen(tail->_idx)) {
4329 #ifndef PRODUCT
4330       if (TraceSuperWord && Verbose) {
4331         tty->print_cr("SuperWord::mark_generations _ii_last error - found different generations in two tail nodes ");
4332         tail->dump();
4333         tail_err->dump();
4334       }
4335 #endif
4336       return -1;
4337     }
4338 
4339     // find first iteration in the loop
4340     for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
4341       Node* ii = phi->fast_out(i);
4342       if (in_bb(ii) && ii->is_Store()) { // we speculate that normally Stores of one and one only generation have deps from mem phi
4343         if (_ii_first == -1) {
4344           ii_err = ii;
4345           _ii_first = _clone_map.gen(ii->_idx);
4346         } else if (_ii_first != _clone_map.gen(ii->_idx)) {
4347 #ifndef PRODUCT
4348           if (TraceSuperWord && Verbose) {
4349             tty->print_cr("SuperWord::mark_generations: _ii_first was found before and not equal to one in this node (%d)", _ii_first);
4350             ii->dump();
4351             if (ii_err!= 0) {
4352               ii_err->dump();
4353             }
4354           }
4355 #endif
4356           return -1; // this phi has Stores from different generations of unroll and cannot be simd/vectorized
4357         }
4358       }
4359     }//for (DUIterator_Fast imax,
4360   }//for (int i...
4361 
4362   if (_ii_first == -1 || _ii_last == -1) {
4363     if (TraceSuperWord && Verbose) {
4364       tty->print_cr("SuperWord::mark_generations unknown error, something vent wrong");
4365     }
4366     return -1; // something vent wrong
4367   }
4368   // collect nodes in the first and last generations
4369   assert(_iteration_first.length() == 0, "_iteration_first must be empty");
4370   assert(_iteration_last.length() == 0, "_iteration_last must be empty");
4371   for (int j = 0; j < _block.length(); j++) {
4372     Node* n = _block.at(j);
4373     node_idx_t gen = _clone_map.gen(n->_idx);
4374     if ((signed)gen == _ii_first) {
4375       _iteration_first.push(n);
4376     } else if ((signed)gen == _ii_last) {
4377       _iteration_last.push(n);
4378     }
4379   }
4380 
4381   // building order of iterations
4382   if (_ii_order.length() == 0 && ii_err != 0) {
4383     assert(in_bb(ii_err) && ii_err->is_Store(), "should be Store in bb");
4384     Node* nd = ii_err;
4385     while(_clone_map.gen(nd->_idx) != _ii_last) {
4386       _ii_order.push(_clone_map.gen(nd->_idx));
4387       bool found = false;
4388       for (DUIterator_Fast imax, i = nd->fast_outs(imax); i < imax; i++) {
4389         Node* use = nd->fast_out(i);
4390         if (same_origin_idx(use, nd) && use->as_Store()->in(MemNode::Memory) == nd) {
4391           found = true;
4392           nd = use;
4393           break;
4394         }
4395       }//for
4396 
4397       if (found == false) {
4398         if (TraceSuperWord && Verbose) {
4399           tty->print_cr("SuperWord::mark_generations: Cannot build order of iterations - no dependent Store for %d", nd->_idx);
4400         }
4401         _ii_order.clear();
4402         return -1;
4403       }
4404     } //while
4405     _ii_order.push(_clone_map.gen(nd->_idx));
4406   }
4407 
4408 #ifndef PRODUCT
4409   if (_vector_loop_debug) {
4410     tty->print_cr("SuperWord::mark_generations");
4411     tty->print_cr("First generation (%d) nodes:", _ii_first);
4412     for (int ii = 0; ii < _iteration_first.length(); ii++)  _iteration_first.at(ii)->dump();
4413     tty->print_cr("Last generation (%d) nodes:", _ii_last);
4414     for (int ii = 0; ii < _iteration_last.length(); ii++)  _iteration_last.at(ii)->dump();
4415     tty->print_cr(" ");
4416 
4417     tty->print("SuperWord::List of generations: ");
4418     for (int jj = 0; jj < _ii_order.length(); ++jj) {
4419       tty->print("%d:%d ", jj, _ii_order.at(jj));
4420     }
4421     tty->print_cr(" ");
4422   }
4423 #endif
4424 
4425   return _ii_first;
4426 }
4427 
4428 bool SuperWord::fix_commutative_inputs(Node* gold, Node* fix) {
4429   assert(gold->is_Add() && fix->is_Add() || gold->is_Mul() && fix->is_Mul(), "should be only Add or Mul nodes");
4430   assert(same_origin_idx(gold, fix), "should be clones of the same node");
4431   Node* gin1 = gold->in(1);
4432   Node* gin2 = gold->in(2);
4433   Node* fin1 = fix->in(1);
4434   Node* fin2 = fix->in(2);
4435   bool swapped = false;
4436 
4437   if (in_bb(gin1) && in_bb(gin2) && in_bb(fin1) && in_bb(fin1)) {
4438     if (same_origin_idx(gin1, fin1) &&
4439         same_origin_idx(gin2, fin2)) {
4440       return true; // nothing to fix
4441     }
4442     if (same_origin_idx(gin1, fin2) &&
4443         same_origin_idx(gin2, fin1)) {
4444       fix->swap_edges(1, 2);
4445       swapped = true;
4446     }
4447   }
4448   // at least one input comes from outside of bb
4449   if (gin1->_idx == fin1->_idx)  {
4450     return true; // nothing to fix
4451   }
4452   if (!swapped && (gin1->_idx == fin2->_idx || gin2->_idx == fin1->_idx))  { //swapping is expensive, check condition first
4453     fix->swap_edges(1, 2);
4454     swapped = true;
4455   }
4456 
4457   if (swapped) {
4458 #ifndef PRODUCT
4459     if (_vector_loop_debug) {
4460       tty->print_cr("SuperWord::fix_commutative_inputs: fixed node %d", fix->_idx);
4461     }
4462 #endif
4463     return true;
4464   }
4465 
4466   if (TraceSuperWord && Verbose) {
4467     tty->print_cr("SuperWord::fix_commutative_inputs: cannot fix node %d", fix->_idx);
4468   }
4469 
4470   return false;
4471 }
4472 
4473 bool SuperWord::pack_parallel() {
4474 #ifndef PRODUCT
4475   if (_vector_loop_debug) {
4476     tty->print_cr("SuperWord::pack_parallel: START");
4477   }
4478 #endif
4479 
4480   _packset.clear();
4481 
4482   for (int ii = 0; ii < _iteration_first.length(); ii++) {
4483     Node* nd = _iteration_first.at(ii);
4484     if (in_bb(nd) && (nd->is_Load() || nd->is_Store() || nd->is_Add() || nd->is_Mul())) {
4485       Node_List* pk = new Node_List();
4486       pk->push(nd);
4487       for (int gen = 1; gen < _ii_order.length(); ++gen) {
4488         for (int kk = 0; kk < _block.length(); kk++) {
4489           Node* clone = _block.at(kk);
4490           if (same_origin_idx(clone, nd) &&
4491               _clone_map.gen(clone->_idx) == _ii_order.at(gen)) {
4492             if (nd->is_Add() || nd->is_Mul()) {
4493               fix_commutative_inputs(nd, clone);
4494             }
4495             pk->push(clone);
4496             if (pk->size() == 4) {
4497               _packset.append(pk);
4498 #ifndef PRODUCT
4499               if (_vector_loop_debug) {
4500                 tty->print_cr("SuperWord::pack_parallel: added pack ");
4501                 pk->dump();
4502               }
4503 #endif
4504               if (_clone_map.gen(clone->_idx) != _ii_last) {
4505                 pk = new Node_List();
4506               }
4507             }
4508             break;
4509           }
4510         }
4511       }//for
4512     }//if
4513   }//for
4514 
4515 #ifndef PRODUCT
4516   if (_vector_loop_debug) {
4517     tty->print_cr("SuperWord::pack_parallel: END");
4518   }
4519 #endif
4520 
4521   return true;
4522 }
4523 
4524 bool SuperWord::hoist_loads_in_graph() {
4525   GrowableArray<Node*> loads;
4526 
4527 #ifndef PRODUCT
4528   if (_vector_loop_debug) {
4529     tty->print_cr("SuperWord::hoist_loads_in_graph: total number _mem_slice_head.length() = %d", _mem_slice_head.length());
4530   }
4531 #endif
4532 
4533   for (int i = 0; i < _mem_slice_head.length(); i++) {
4534     Node* n = _mem_slice_head.at(i);
4535     if ( !in_bb(n) || !n->is_Phi() || n->bottom_type() != Type::MEMORY) {
4536       if (TraceSuperWord && Verbose) {
4537         tty->print_cr("SuperWord::hoist_loads_in_graph: skipping unexpected node n=%d", n->_idx);
4538       }
4539       continue;
4540     }
4541 
4542 #ifndef PRODUCT
4543     if (_vector_loop_debug) {
4544       tty->print_cr("SuperWord::hoist_loads_in_graph: processing phi %d  = _mem_slice_head.at(%d);", n->_idx, i);
4545     }
4546 #endif
4547 
4548     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4549       Node* ld = n->fast_out(i);
4550       if (ld->is_Load() && ld->as_Load()->in(MemNode::Memory) == n && in_bb(ld)) {
4551         for (int i = 0; i < _block.length(); i++) {
4552           Node* ld2 = _block.at(i);
4553           if (ld2->is_Load() && same_origin_idx(ld, ld2) &&
4554               !same_generation(ld, ld2)) { // <= do not collect the first generation ld
4555 #ifndef PRODUCT
4556             if (_vector_loop_debug) {
4557               tty->print_cr("SuperWord::hoist_loads_in_graph: will try to hoist load ld2->_idx=%d, cloned from %d (ld->_idx=%d)",
4558                 ld2->_idx, _clone_map.idx(ld->_idx), ld->_idx);
4559             }
4560 #endif
4561             // could not do on-the-fly, since iterator is immutable
4562             loads.push(ld2);
4563           }
4564         }// for
4565       }//if
4566     }//for (DUIterator_Fast imax,
4567   }//for (int i = 0; i
4568 
4569   for (int i = 0; i < loads.length(); i++) {
4570     LoadNode* ld = loads.at(i)->as_Load();
4571     Node* phi = find_phi_for_mem_dep(ld);
4572     if (phi != NULL) {
4573 #ifndef PRODUCT
4574       if (_vector_loop_debug) {
4575         tty->print_cr("SuperWord::hoist_loads_in_graph replacing MemNode::Memory(%d) edge in %d with one from %d",
4576           MemNode::Memory, ld->_idx, phi->_idx);
4577       }
4578 #endif
4579       _igvn.replace_input_of(ld, MemNode::Memory, phi);
4580     }
4581   }//for
4582 
4583   restart(); // invalidate all basic structures, since we rebuilt the graph
4584 
4585   if (TraceSuperWord && Verbose) {
4586     tty->print_cr("\nSuperWord::hoist_loads_in_graph() the graph was rebuilt, all structures invalidated and need rebuild");
4587   }
4588 
4589   return true;
4590 }