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   _arena(phase->C->comp_arena()),
  51   _igvn(phase->_igvn),
  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   _nlist(arena(), 8, 0, NULL),            // scratch list of nodes
  69   _stk(arena(), 8, 0, NULL),              // scratch stack 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   _do_vector_loop(phase->C->do_vector_loop()),  // whether to do vectorization/simd style
  77   _do_reserve_copy(DoReserveCopyInSuperWord),
  78   _num_work_vecs(0),                      // amount of vector work we have
  79   _num_reductions(0),                     // amount of reduction work we have
  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()) {
 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()) {
 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 (mr == mem_ref) {
 649               // Skip when we are looking at same memory operation.
 650               continue;
 651             }
 652             if (same_velt_type(mr, mem_ref) &&
 653                 memory_alignment(mr, iv_adjustment) != 0)
 654               create_pack = false;
 655           }
 656         }
 657       }
 658     }
 659     if (create_pack) {
 660       for (uint i = 0; i < memops.size(); i++) {
 661         Node* s1 = memops.at(i);
 662         int align = alignment(s1);
 663         if (align == top_align) continue;
 664         for (uint j = 0; j < memops.size(); j++) {
 665           Node* s2 = memops.at(j);
 666           if (alignment(s2) == top_align) continue;
 667           if (s1 != s2 && are_adjacent_refs(s1, s2)) {
 668             if (stmts_can_pack(s1, s2, align)) {
 669               Node_List* pair = new Node_List();
 670               pair->push(s1);
 671               pair->push(s2);
 672               if (!_do_vector_loop || same_origin_idx(s1, s2)) {
 673                 _packset.append(pair);
 674               }
 675             }
 676           }
 677         }
 678       }
 679     } else { // Don't create unaligned pack
 680       // First, remove remaining memory ops of the same type from the list.
 681       for (int i = memops.size() - 1; i >= 0; i--) {
 682         MemNode* s = memops.at(i)->as_Mem();
 683         if (same_velt_type(s, mem_ref)) {
 684           memops.remove(i);
 685         }
 686       }
 687 
 688       // Second, remove already constructed packs of the same type.
 689       for (int i = _packset.length() - 1; i >= 0; i--) {
 690         Node_List* p = _packset.at(i);
 691         MemNode* s = p->at(0)->as_Mem();
 692         if (same_velt_type(s, mem_ref)) {
 693           remove_pack_at(i);
 694         }
 695       }
 696 
 697       // If needed find the best memory reference for loop alignment again.
 698       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 699         // Put memory ops from remaining packs back on memops list for
 700         // the best alignment search.
 701         uint orig_msize = memops.size();
 702         for (int i = 0; i < _packset.length(); i++) {
 703           Node_List* p = _packset.at(i);
 704           MemNode* s = p->at(0)->as_Mem();
 705           assert(!same_velt_type(s, mem_ref), "sanity");
 706           memops.push(s);
 707         }
 708         best_align_to_mem_ref = find_align_to_ref(memops);
 709         if (best_align_to_mem_ref == NULL) {
 710           if (TraceSuperWord) {
 711             tty->print_cr("SuperWord::find_adjacent_refs(): best_align_to_mem_ref == NULL");
 712           }
 713           break;
 714         }
 715         best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
 716         NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
 717         // Restore list.
 718         while (memops.size() > orig_msize)
 719           (void)memops.pop();
 720       }
 721     } // unaligned memory accesses
 722 
 723     // Remove used mem nodes.
 724     for (int i = memops.size() - 1; i >= 0; i--) {
 725       MemNode* m = memops.at(i)->as_Mem();
 726       if (alignment(m) != top_align) {
 727         memops.remove(i);
 728       }
 729     }
 730 
 731   } // while (memops.size() != 0
 732   set_align_to_ref(best_align_to_mem_ref);
 733 
 734   if (TraceSuperWord) {
 735     tty->print_cr("\nAfter find_adjacent_refs");
 736     print_packset();
 737   }
 738 }
 739 
 740 #ifndef PRODUCT
 741 void SuperWord::find_adjacent_refs_trace_1(Node* best_align_to_mem_ref, int best_iv_adjustment) {
 742   if (is_trace_adjacent()) {
 743     tty->print("SuperWord::find_adjacent_refs best_align_to_mem_ref = %d, best_iv_adjustment = %d",
 744        best_align_to_mem_ref->_idx, best_iv_adjustment);
 745        best_align_to_mem_ref->dump();
 746   }
 747 }
 748 #endif
 749 
 750 //------------------------------find_align_to_ref---------------------------
 751 // Find a memory reference to align the loop induction variable to.
 752 // Looks first at stores then at loads, looking for a memory reference
 753 // with the largest number of references similar to it.
 754 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
 755   GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
 756 
 757   // Count number of comparable memory ops
 758   for (uint i = 0; i < memops.size(); i++) {
 759     MemNode* s1 = memops.at(i)->as_Mem();
 760     SWPointer p1(s1, this, NULL, false);
 761     // Discard if pre loop can't align this reference
 762     if (!ref_is_alignable(p1)) {
 763       *cmp_ct.adr_at(i) = 0;
 764       continue;
 765     }
 766     for (uint j = i+1; j < memops.size(); j++) {
 767       MemNode* s2 = memops.at(j)->as_Mem();
 768       if (isomorphic(s1, s2)) {
 769         SWPointer p2(s2, this, NULL, false);
 770         if (p1.comparable(p2)) {
 771           (*cmp_ct.adr_at(i))++;
 772           (*cmp_ct.adr_at(j))++;
 773         }
 774       }
 775     }
 776   }
 777 
 778   // Find Store (or Load) with the greatest number of "comparable" references,
 779   // biggest vector size, smallest data size and smallest iv offset.
 780   int max_ct        = 0;
 781   int max_vw        = 0;
 782   int max_idx       = -1;
 783   int min_size      = max_jint;
 784   int min_iv_offset = max_jint;
 785   for (uint j = 0; j < memops.size(); j++) {
 786     MemNode* s = memops.at(j)->as_Mem();
 787     if (s->is_Store()) {
 788       int vw = vector_width_in_bytes(s);
 789       assert(vw > 1, "sanity");
 790       SWPointer p(s, this, NULL, false);
 791       if ( cmp_ct.at(j) >  max_ct ||
 792           (cmp_ct.at(j) == max_ct &&
 793             ( vw >  max_vw ||
 794              (vw == max_vw &&
 795               ( data_size(s) <  min_size ||
 796                (data_size(s) == min_size &&
 797                 p.offset_in_bytes() < min_iv_offset)))))) {
 798         max_ct = cmp_ct.at(j);
 799         max_vw = vw;
 800         max_idx = j;
 801         min_size = data_size(s);
 802         min_iv_offset = p.offset_in_bytes();
 803       }
 804     }
 805   }
 806   // If no stores, look at loads
 807   if (max_ct == 0) {
 808     for (uint j = 0; j < memops.size(); j++) {
 809       MemNode* s = memops.at(j)->as_Mem();
 810       if (s->is_Load()) {
 811         int vw = vector_width_in_bytes(s);
 812         assert(vw > 1, "sanity");
 813         SWPointer p(s, this, NULL, false);
 814         if ( cmp_ct.at(j) >  max_ct ||
 815             (cmp_ct.at(j) == max_ct &&
 816               ( vw >  max_vw ||
 817                (vw == max_vw &&
 818                 ( data_size(s) <  min_size ||
 819                  (data_size(s) == min_size &&
 820                   p.offset_in_bytes() < min_iv_offset)))))) {
 821           max_ct = cmp_ct.at(j);
 822           max_vw = vw;
 823           max_idx = j;
 824           min_size = data_size(s);
 825           min_iv_offset = p.offset_in_bytes();
 826         }
 827       }
 828     }
 829   }
 830 
 831 #ifdef ASSERT
 832   if (TraceSuperWord && Verbose) {
 833     tty->print_cr("\nVector memops after find_align_to_ref");
 834     for (uint i = 0; i < memops.size(); i++) {
 835       MemNode* s = memops.at(i)->as_Mem();
 836       s->dump();
 837     }
 838   }
 839 #endif
 840 
 841   if (max_ct > 0) {
 842 #ifdef ASSERT
 843     if (TraceSuperWord) {
 844       tty->print("\nVector align to node: ");
 845       memops.at(max_idx)->as_Mem()->dump();
 846     }
 847 #endif
 848     return memops.at(max_idx)->as_Mem();
 849   }
 850   return NULL;
 851 }
 852 
 853 //------------------span_works_for_memory_size-----------------------------
 854 static bool span_works_for_memory_size(MemNode* mem, int span, int mem_size, int offset) {
 855   bool span_matches_memory = false;
 856   if ((mem_size == type2aelembytes(T_BYTE) || mem_size == type2aelembytes(T_SHORT))
 857     && ABS(span) == type2aelembytes(T_INT)) {
 858     // There is a mismatch on span size compared to memory.
 859     for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
 860       Node* use = mem->fast_out(j);
 861       if (!VectorNode::is_type_transition_to_int(use)) {
 862         return false;
 863       }
 864     }
 865     // If all uses transition to integer, it means that we can successfully align even on mismatch.
 866     return true;
 867   }
 868   else {
 869     span_matches_memory = ABS(span) == mem_size;
 870   }
 871   return span_matches_memory && (ABS(offset) % mem_size) == 0;
 872 }
 873 
 874 //------------------------------ref_is_alignable---------------------------
 875 // Can the preloop align the reference to position zero in the vector?
 876 bool SuperWord::ref_is_alignable(SWPointer& p) {
 877   if (!p.has_iv()) {
 878     return true;   // no induction variable
 879   }
 880   CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
 881   assert(pre_end != NULL, "we must have a correct pre-loop");
 882   assert(pre_end->stride_is_con(), "pre loop stride is constant");
 883   int preloop_stride = pre_end->stride_con();
 884 
 885   int span = preloop_stride * p.scale_in_bytes();
 886   int mem_size = p.memory_size();
 887   int offset   = p.offset_in_bytes();
 888   // Stride one accesses are alignable if offset is aligned to memory operation size.
 889   // Offset can be unaligned when UseUnalignedAccesses is used.
 890   if (span_works_for_memory_size(p.mem(), span, mem_size, offset)) {
 891     return true;
 892   }
 893   // If the initial offset from start of the object is computable,
 894   // check if the pre-loop can align the final offset accordingly.
 895   //
 896   // In other words: Can we find an i such that the offset
 897   // after i pre-loop iterations is aligned to vw?
 898   //   (init_offset + pre_loop) % vw == 0              (1)
 899   // where
 900   //   pre_loop = i * span
 901   // is the number of bytes added to the offset by i pre-loop iterations.
 902   //
 903   // For this to hold we need pre_loop to increase init_offset by
 904   //   pre_loop = vw - (init_offset % vw)
 905   //
 906   // This is only possible if pre_loop is divisible by span because each
 907   // pre-loop iteration increases the initial offset by 'span' bytes:
 908   //   (vw - (init_offset % vw)) % span == 0
 909   //
 910   int vw = vector_width_in_bytes(p.mem());
 911   assert(vw > 1, "sanity");
 912   Node* init_nd = pre_end->init_trip();
 913   if (init_nd->is_Con() && p.invar() == NULL) {
 914     int init = init_nd->bottom_type()->is_int()->get_con();
 915     int init_offset = init * p.scale_in_bytes() + offset;
 916     if (init_offset < 0) { // negative offset from object start?
 917       return false;        // may happen in dead loop
 918     }
 919     if (vw % span == 0) {
 920       // If vm is a multiple of span, we use formula (1).
 921       if (span > 0) {
 922         return (vw - (init_offset % vw)) % span == 0;
 923       } else {
 924         assert(span < 0, "nonzero stride * scale");
 925         return (init_offset % vw) % -span == 0;
 926       }
 927     } else if (span % vw == 0) {
 928       // If span is a multiple of vw, we can simplify formula (1) to:
 929       //   (init_offset + i * span) % vw == 0
 930       //     =>
 931       //   (init_offset % vw) + ((i * span) % vw) == 0
 932       //     =>
 933       //   init_offset % vw == 0
 934       //
 935       // Because we add a multiple of vw to the initial offset, the final
 936       // offset is a multiple of vw if and only if init_offset is a multiple.
 937       //
 938       return (init_offset % vw) == 0;
 939     }
 940   }
 941   return false;
 942 }
 943 //---------------------------get_vw_bytes_special------------------------
 944 int SuperWord::get_vw_bytes_special(MemNode* s) {
 945   // Get the vector width in bytes.
 946   int vw = vector_width_in_bytes(s);
 947 
 948   // Check for special case where there is an MulAddS2I usage where short vectors are going to need combined.
 949   BasicType btype = velt_basic_type(s);
 950   if (type2aelembytes(btype) == 2) {
 951     bool should_combine_adjacent = true;
 952     for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
 953       Node* user = s->fast_out(i);
 954       if (!VectorNode::is_muladds2i(user)) {
 955         should_combine_adjacent = false;
 956       }
 957     }
 958     if (should_combine_adjacent) {
 959       vw = MIN2(Matcher::max_vector_size(btype)*type2aelembytes(btype), vw * 2);
 960     }
 961   }
 962 
 963   return vw;
 964 }
 965 
 966 //---------------------------get_iv_adjustment---------------------------
 967 // Calculate loop's iv adjustment for this memory ops.
 968 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
 969   SWPointer align_to_ref_p(mem_ref, this, NULL, false);
 970   int offset = align_to_ref_p.offset_in_bytes();
 971   int scale  = align_to_ref_p.scale_in_bytes();
 972   int elt_size = align_to_ref_p.memory_size();
 973   int vw       = get_vw_bytes_special(mem_ref);
 974   assert(vw > 1, "sanity");
 975   int iv_adjustment;
 976   if (scale != 0) {
 977     int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
 978     // At least one iteration is executed in pre-loop by default. As result
 979     // several iterations are needed to align memory operations in main-loop even
 980     // if offset is 0.
 981     int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
 982     assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
 983            "(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size);
 984     iv_adjustment = iv_adjustment_in_bytes/elt_size;
 985   } else {
 986     // This memory op is not dependent on iv (scale == 0)
 987     iv_adjustment = 0;
 988   }
 989 
 990 #ifndef PRODUCT
 991   if (TraceSuperWord) {
 992     tty->print("SuperWord::get_iv_adjustment: n = %d, noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d: ",
 993       mem_ref->_idx, offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
 994     mem_ref->dump();
 995   }
 996 #endif
 997   return iv_adjustment;
 998 }
 999 
1000 //---------------------------dependence_graph---------------------------
1001 // Construct dependency graph.
1002 // Add dependence edges to load/store nodes for memory dependence
1003 //    A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
1004 void SuperWord::dependence_graph() {
1005   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
1006   // First, assign a dependence node to each memory node
1007   for (int i = 0; i < _block.length(); i++ ) {
1008     Node *n = _block.at(i);
1009     if (n->is_Mem() || (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1010       _dg.make_node(n);
1011     }
1012   }
1013 
1014   // For each memory slice, create the dependences
1015   for (int i = 0; i < _mem_slice_head.length(); i++) {
1016     Node* n      = _mem_slice_head.at(i);
1017     Node* n_tail = _mem_slice_tail.at(i);
1018 
1019     // Get slice in predecessor order (last is first)
1020     if (cl->is_main_loop()) {
1021       mem_slice_preds(n_tail, n, _nlist);
1022     }
1023 
1024 #ifndef PRODUCT
1025     if(TraceSuperWord && Verbose) {
1026       tty->print_cr("SuperWord::dependence_graph: built a new mem slice");
1027       for (int j = _nlist.length() - 1; j >= 0 ; j--) {
1028         _nlist.at(j)->dump();
1029       }
1030     }
1031 #endif
1032     // Make the slice dependent on the root
1033     DepMem* slice = _dg.dep(n);
1034     _dg.make_edge(_dg.root(), slice);
1035 
1036     // Create a sink for the slice
1037     DepMem* slice_sink = _dg.make_node(NULL);
1038     _dg.make_edge(slice_sink, _dg.tail());
1039 
1040     // Now visit each pair of memory ops, creating the edges
1041     for (int j = _nlist.length() - 1; j >= 0 ; j--) {
1042       Node* s1 = _nlist.at(j);
1043 
1044       // If no dependency yet, use slice
1045       if (_dg.dep(s1)->in_cnt() == 0) {
1046         _dg.make_edge(slice, s1);
1047       }
1048       SWPointer p1(s1->as_Mem(), this, NULL, false);
1049       bool sink_dependent = true;
1050       for (int k = j - 1; k >= 0; k--) {
1051         Node* s2 = _nlist.at(k);
1052         if (s1->is_Load() && s2->is_Load())
1053           continue;
1054         SWPointer p2(s2->as_Mem(), this, NULL, false);
1055 
1056         int cmp = p1.cmp(p2);
1057         if (SuperWordRTDepCheck &&
1058             p1.base() != p2.base() && p1.valid() && p2.valid()) {
1059           // Create a runtime check to disambiguate
1060           OrderedPair pp(p1.base(), p2.base());
1061           _disjoint_ptrs.append_if_missing(pp);
1062         } else if (!SWPointer::not_equal(cmp)) {
1063           // Possibly same address
1064           _dg.make_edge(s1, s2);
1065           sink_dependent = false;
1066         }
1067       }
1068       if (sink_dependent) {
1069         _dg.make_edge(s1, slice_sink);
1070       }
1071     }
1072 
1073     if (TraceSuperWord) {
1074       tty->print_cr("\nDependence graph for slice: %d", n->_idx);
1075       for (int q = 0; q < _nlist.length(); q++) {
1076         _dg.print(_nlist.at(q));
1077       }
1078       tty->cr();
1079     }
1080 
1081     _nlist.clear();
1082   }
1083 
1084   if (TraceSuperWord) {
1085     tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
1086     for (int r = 0; r < _disjoint_ptrs.length(); r++) {
1087       _disjoint_ptrs.at(r).print();
1088       tty->cr();
1089     }
1090     tty->cr();
1091   }
1092 
1093 }
1094 
1095 //---------------------------mem_slice_preds---------------------------
1096 // Return a memory slice (node list) in predecessor order starting at "start"
1097 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
1098   assert(preds.length() == 0, "start empty");
1099   Node* n = start;
1100   Node* prev = NULL;
1101   while (true) {
1102     NOT_PRODUCT( if(is_trace_mem_slice()) tty->print_cr("SuperWord::mem_slice_preds: n %d", n->_idx);)
1103     assert(in_bb(n), "must be in block");
1104     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1105       Node* out = n->fast_out(i);
1106       if (out->is_Load()) {
1107         if (in_bb(out)) {
1108           preds.push(out);
1109           if (TraceSuperWord && Verbose) {
1110             tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", out->_idx);
1111           }
1112         }
1113       } else {
1114         // FIXME
1115         if (out->is_MergeMem() && !in_bb(out)) {
1116           // Either unrolling is causing a memory edge not to disappear,
1117           // or need to run igvn.optimize() again before SLP
1118         } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
1119           // Ditto.  Not sure what else to check further.
1120         } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
1121           // StoreCM has an input edge used as a precedence edge.
1122           // Maybe an issue when oop stores are vectorized.
1123         } else {
1124           assert(out == prev || prev == NULL, "no branches off of store slice");
1125         }
1126       }//else
1127     }//for
1128     if (n == stop) break;
1129     preds.push(n);
1130     if (TraceSuperWord && Verbose) {
1131       tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", n->_idx);
1132     }
1133     prev = n;
1134     assert(n->is_Mem(), "unexpected node %s", n->Name());
1135     n = n->in(MemNode::Memory);
1136   }
1137 }
1138 
1139 //------------------------------stmts_can_pack---------------------------
1140 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
1141 // s1 aligned at "align"
1142 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
1143 
1144   // Do not use superword for non-primitives
1145   BasicType bt1 = velt_basic_type(s1);
1146   BasicType bt2 = velt_basic_type(s2);
1147   if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
1148     return false;
1149   if (Matcher::max_vector_size(bt1) < 2) {
1150     return false; // No vectors for this type
1151   }
1152 
1153   if (isomorphic(s1, s2)) {
1154     if ((independent(s1, s2) && have_similar_inputs(s1, s2)) || reduction(s1, s2)) {
1155       if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
1156         if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
1157           int s1_align = alignment(s1);
1158           int s2_align = alignment(s2);
1159           if (s1_align == top_align || s1_align == align) {
1160             if (s2_align == top_align || s2_align == align + data_size(s1)) {
1161               return true;
1162             }
1163           }
1164         }
1165       }
1166     }
1167   }
1168   return false;
1169 }
1170 
1171 //------------------------------exists_at---------------------------
1172 // Does s exist in a pack at position pos?
1173 bool SuperWord::exists_at(Node* s, uint pos) {
1174   for (int i = 0; i < _packset.length(); i++) {
1175     Node_List* p = _packset.at(i);
1176     if (p->at(pos) == s) {
1177       return true;
1178     }
1179   }
1180   return false;
1181 }
1182 
1183 //------------------------------are_adjacent_refs---------------------------
1184 // Is s1 immediately before s2 in memory?
1185 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
1186   if (!s1->is_Mem() || !s2->is_Mem()) return false;
1187   if (!in_bb(s1)    || !in_bb(s2))    return false;
1188 
1189   // Do not use superword for non-primitives
1190   if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
1191       !is_java_primitive(s2->as_Mem()->memory_type())) {
1192     return false;
1193   }
1194 
1195   // FIXME - co_locate_pack fails on Stores in different mem-slices, so
1196   // only pack memops that are in the same alias set until that's fixed.
1197   if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
1198       _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
1199     return false;
1200   SWPointer p1(s1->as_Mem(), this, NULL, false);
1201   SWPointer p2(s2->as_Mem(), this, NULL, false);
1202   if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
1203   int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
1204   return diff == data_size(s1);
1205 }
1206 
1207 //------------------------------isomorphic---------------------------
1208 // Are s1 and s2 similar?
1209 bool SuperWord::isomorphic(Node* s1, Node* s2) {
1210   if (s1->Opcode() != s2->Opcode()) return false;
1211   if (s1->req() != s2->req()) return false;
1212   if (s1->in(0) != s2->in(0)) return false;
1213   if (!same_velt_type(s1, s2)) return false;
1214   return true;
1215 }
1216 
1217 //------------------------------independent---------------------------
1218 // Is there no data path from s1 to s2 or s2 to s1?
1219 bool SuperWord::independent(Node* s1, Node* s2) {
1220   //  assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
1221   int d1 = depth(s1);
1222   int d2 = depth(s2);
1223   if (d1 == d2) return s1 != s2;
1224   Node* deep    = d1 > d2 ? s1 : s2;
1225   Node* shallow = d1 > d2 ? s2 : s1;
1226 
1227   visited_clear();
1228 
1229   return independent_path(shallow, deep);
1230 }
1231 
1232 //--------------------------have_similar_inputs-----------------------
1233 // For a node pair (s1, s2) which is isomorphic and independent,
1234 // do s1 and s2 have similar input edges?
1235 bool SuperWord::have_similar_inputs(Node* s1, Node* s2) {
1236   // assert(isomorphic(s1, s2) == true, "check isomorphic");
1237   // assert(independent(s1, s2) == true, "check independent");
1238   if (s1->req() > 1 && !s1->is_Store() && !s1->is_Load()) {
1239     for (uint i = 1; i < s1->req(); i++) {
1240       if (s1->in(i)->Opcode() != s2->in(i)->Opcode()) return false;
1241     }
1242   }
1243   return true;
1244 }
1245 
1246 //------------------------------reduction---------------------------
1247 // Is there a data path between s1 and s2 and the nodes reductions?
1248 bool SuperWord::reduction(Node* s1, Node* s2) {
1249   bool retValue = false;
1250   int d1 = depth(s1);
1251   int d2 = depth(s2);
1252   if (d1 + 1 == d2) {
1253     if (s1->is_reduction() && s2->is_reduction()) {
1254       // This is an ordered set, so s1 should define s2
1255       for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1256         Node* t1 = s1->fast_out(i);
1257         if (t1 == s2) {
1258           // both nodes are reductions and connected
1259           retValue = true;
1260         }
1261       }
1262     }
1263   }
1264 
1265   return retValue;
1266 }
1267 
1268 //------------------------------independent_path------------------------------
1269 // Helper for independent
1270 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
1271   if (dp >= 1000) return false; // stop deep recursion
1272   visited_set(deep);
1273   int shal_depth = depth(shallow);
1274   assert(shal_depth <= depth(deep), "must be");
1275   for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
1276     Node* pred = preds.current();
1277     if (in_bb(pred) && !visited_test(pred)) {
1278       if (shallow == pred) {
1279         return false;
1280       }
1281       if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
1282         return false;
1283       }
1284     }
1285   }
1286   return true;
1287 }
1288 
1289 //------------------------------set_alignment---------------------------
1290 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
1291   set_alignment(s1, align);
1292   if (align == top_align || align == bottom_align) {
1293     set_alignment(s2, align);
1294   } else {
1295     set_alignment(s2, align + data_size(s1));
1296   }
1297 }
1298 
1299 //------------------------------data_size---------------------------
1300 int SuperWord::data_size(Node* s) {
1301   Node* use = NULL; //test if the node is a candidate for CMoveV optimization, then return the size of CMov
1302   if (UseVectorCmov) {
1303     use = _cmovev_kit.is_Bool_candidate(s);
1304     if (use != NULL) {
1305       return data_size(use);
1306     }
1307     use = _cmovev_kit.is_CmpD_candidate(s);
1308     if (use != NULL) {
1309       return data_size(use);
1310     }
1311   }
1312 
1313   int bsize = type2aelembytes(velt_basic_type(s));
1314   assert(bsize != 0, "valid size");
1315   return bsize;
1316 }
1317 
1318 //------------------------------extend_packlist---------------------------
1319 // Extend packset by following use->def and def->use links from pack members.
1320 void SuperWord::extend_packlist() {
1321   bool changed;
1322   do {
1323     packset_sort(_packset.length());
1324     changed = false;
1325     for (int i = 0; i < _packset.length(); i++) {
1326       Node_List* p = _packset.at(i);
1327       changed |= follow_use_defs(p);
1328       changed |= follow_def_uses(p);
1329     }
1330   } while (changed);
1331 
1332   if (_race_possible) {
1333     for (int i = 0; i < _packset.length(); i++) {
1334       Node_List* p = _packset.at(i);
1335       order_def_uses(p);
1336     }
1337   }
1338 
1339   if (TraceSuperWord) {
1340     tty->print_cr("\nAfter extend_packlist");
1341     print_packset();
1342   }
1343 }
1344 
1345 //------------------------------follow_use_defs---------------------------
1346 // Extend the packset by visiting operand definitions of nodes in pack p
1347 bool SuperWord::follow_use_defs(Node_List* p) {
1348   assert(p->size() == 2, "just checking");
1349   Node* s1 = p->at(0);
1350   Node* s2 = p->at(1);
1351   assert(s1->req() == s2->req(), "just checking");
1352   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1353 
1354   if (s1->is_Load()) return false;
1355 
1356   int align = alignment(s1);
1357   NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: s1 %d, align %d", s1->_idx, align);)
1358   bool changed = false;
1359   int start = s1->is_Store() ? MemNode::ValueIn   : 1;
1360   int end   = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
1361   for (int j = start; j < end; j++) {
1362     Node* t1 = s1->in(j);
1363     Node* t2 = s2->in(j);
1364     if (!in_bb(t1) || !in_bb(t2))
1365       continue;
1366     if (stmts_can_pack(t1, t2, align)) {
1367       if (est_savings(t1, t2) >= 0) {
1368         Node_List* pair = new Node_List();
1369         pair->push(t1);
1370         pair->push(t2);
1371         _packset.append(pair);
1372         NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: set_alignment(%d, %d, %d)", t1->_idx, t2->_idx, align);)
1373         set_alignment(t1, t2, align);
1374         changed = true;
1375       }
1376     }
1377   }
1378   return changed;
1379 }
1380 
1381 //------------------------------follow_def_uses---------------------------
1382 // Extend the packset by visiting uses of nodes in pack p
1383 bool SuperWord::follow_def_uses(Node_List* p) {
1384   bool changed = false;
1385   Node* s1 = p->at(0);
1386   Node* s2 = p->at(1);
1387   assert(p->size() == 2, "just checking");
1388   assert(s1->req() == s2->req(), "just checking");
1389   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1390 
1391   if (s1->is_Store()) return false;
1392 
1393   int align = alignment(s1);
1394   NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: s1 %d, align %d", s1->_idx, align);)
1395   int savings = -1;
1396   int num_s1_uses = 0;
1397   Node* u1 = NULL;
1398   Node* u2 = NULL;
1399   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1400     Node* t1 = s1->fast_out(i);
1401     num_s1_uses++;
1402     if (!in_bb(t1)) continue;
1403     for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1404       Node* t2 = s2->fast_out(j);
1405       if (!in_bb(t2)) continue;
1406       if (t2->Opcode() == Op_AddI && t2 == _lp->as_CountedLoop()->incr()) continue; // don't mess with the iv
1407       if (!opnd_positions_match(s1, t1, s2, t2))
1408         continue;
1409       if (stmts_can_pack(t1, t2, align)) {
1410         int my_savings = est_savings(t1, t2);
1411         if (my_savings > savings) {
1412           savings = my_savings;
1413           u1 = t1;
1414           u2 = t2;
1415         }
1416       }
1417     }
1418   }
1419   if (num_s1_uses > 1) {
1420     _race_possible = true;
1421   }
1422   if (savings >= 0) {
1423     Node_List* pair = new Node_List();
1424     pair->push(u1);
1425     pair->push(u2);
1426     _packset.append(pair);
1427     NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: set_alignment(%d, %d, %d)", u1->_idx, u2->_idx, align);)
1428     set_alignment(u1, u2, align);
1429     changed = true;
1430   }
1431   return changed;
1432 }
1433 
1434 //------------------------------order_def_uses---------------------------
1435 // For extended packsets, ordinally arrange uses packset by major component
1436 void SuperWord::order_def_uses(Node_List* p) {
1437   Node* s1 = p->at(0);
1438 
1439   if (s1->is_Store()) return;
1440 
1441   // reductions are always managed beforehand
1442   if (s1->is_reduction()) return;
1443 
1444   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1445     Node* t1 = s1->fast_out(i);
1446 
1447     // Only allow operand swap on commuting operations
1448     if (!t1->is_Add() && !t1->is_Mul() && !VectorNode::is_muladds2i(t1)) {
1449       break;
1450     }
1451 
1452     // Now find t1's packset
1453     Node_List* p2 = NULL;
1454     for (int j = 0; j < _packset.length(); j++) {
1455       p2 = _packset.at(j);
1456       Node* first = p2->at(0);
1457       if (t1 == first) {
1458         break;
1459       }
1460       p2 = NULL;
1461     }
1462     // Arrange all sub components by the major component
1463     if (p2 != NULL) {
1464       for (uint j = 1; j < p->size(); j++) {
1465         Node* d1 = p->at(j);
1466         Node* u1 = p2->at(j);
1467         opnd_positions_match(s1, t1, d1, u1);
1468       }
1469     }
1470   }
1471 }
1472 
1473 //---------------------------opnd_positions_match-------------------------
1474 // Is the use of d1 in u1 at the same operand position as d2 in u2?
1475 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
1476   // check reductions to see if they are marshalled to represent the reduction
1477   // operator in a specified opnd
1478   if (u1->is_reduction() && u2->is_reduction()) {
1479     // ensure reductions have phis and reduction definitions feeding the 1st operand
1480     Node* first = u1->in(2);
1481     if (first->is_Phi() || first->is_reduction()) {
1482       u1->swap_edges(1, 2);
1483     }
1484     // ensure reductions have phis and reduction definitions feeding the 1st operand
1485     first = u2->in(2);
1486     if (first->is_Phi() || first->is_reduction()) {
1487       u2->swap_edges(1, 2);
1488     }
1489     return true;
1490   }
1491 
1492   uint ct = u1->req();
1493   if (ct != u2->req()) return false;
1494   uint i1 = 0;
1495   uint i2 = 0;
1496   do {
1497     for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
1498     for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
1499     if (i1 != i2) {
1500       if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
1501         // Further analysis relies on operands position matching.
1502         u2->swap_edges(i1, i2);
1503       } else if (VectorNode::is_muladds2i(u2) && u1 != u2) {
1504         if (i1 == 5 - i2) { // ((i1 == 3 && i2 == 2) || (i1 == 2 && i2 == 3) || (i1 == 1 && i2 == 4) || (i1 == 4 && i2 == 1))
1505           u2->swap_edges(1, 2);
1506           u2->swap_edges(3, 4);
1507         }
1508         if (i1 == 3 - i2 || i1 == 7 - i2) { // ((i1 == 1 && i2 == 2) || (i1 == 2 && i2 == 1) || (i1 == 3 && i2 == 4) || (i1 == 4 && i2 == 3))
1509           u2->swap_edges(2, 3);
1510           u2->swap_edges(1, 4);
1511         }
1512         return false; // Just swap the edges, the muladds2i nodes get packed in follow_use_defs
1513       } else {
1514         return false;
1515       }
1516     } else if (i1 == i2 && VectorNode::is_muladds2i(u2) && u1 != u2) {
1517       u2->swap_edges(1, 3);
1518       u2->swap_edges(2, 4);
1519       return false; // Just swap the edges, the muladds2i nodes get packed in follow_use_defs
1520     }
1521   } while (i1 < ct);
1522   return true;
1523 }
1524 
1525 //------------------------------est_savings---------------------------
1526 // Estimate the savings from executing s1 and s2 as a pack
1527 int SuperWord::est_savings(Node* s1, Node* s2) {
1528   int save_in = 2 - 1; // 2 operations per instruction in packed form
1529 
1530   // inputs
1531   for (uint i = 1; i < s1->req(); i++) {
1532     Node* x1 = s1->in(i);
1533     Node* x2 = s2->in(i);
1534     if (x1 != x2) {
1535       if (are_adjacent_refs(x1, x2)) {
1536         save_in += adjacent_profit(x1, x2);
1537       } else if (!in_packset(x1, x2)) {
1538         save_in -= pack_cost(2);
1539       } else {
1540         save_in += unpack_cost(2);
1541       }
1542     }
1543   }
1544 
1545   // uses of result
1546   uint ct = 0;
1547   int save_use = 0;
1548   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1549     Node* s1_use = s1->fast_out(i);
1550     for (int j = 0; j < _packset.length(); j++) {
1551       Node_List* p = _packset.at(j);
1552       if (p->at(0) == s1_use) {
1553         for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
1554           Node* s2_use = s2->fast_out(k);
1555           if (p->at(p->size()-1) == s2_use) {
1556             ct++;
1557             if (are_adjacent_refs(s1_use, s2_use)) {
1558               save_use += adjacent_profit(s1_use, s2_use);
1559             }
1560           }
1561         }
1562       }
1563     }
1564   }
1565 
1566   if (ct < s1->outcnt()) save_use += unpack_cost(1);
1567   if (ct < s2->outcnt()) save_use += unpack_cost(1);
1568 
1569   return MAX2(save_in, save_use);
1570 }
1571 
1572 //------------------------------costs---------------------------
1573 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
1574 int SuperWord::pack_cost(int ct)   { return ct; }
1575 int SuperWord::unpack_cost(int ct) { return ct; }
1576 
1577 //------------------------------combine_packs---------------------------
1578 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
1579 void SuperWord::combine_packs() {
1580   bool changed = true;
1581   // Combine packs regardless max vector size.
1582   while (changed) {
1583     changed = false;
1584     for (int i = 0; i < _packset.length(); i++) {
1585       Node_List* p1 = _packset.at(i);
1586       if (p1 == NULL) continue;
1587       // Because of sorting we can start at i + 1
1588       for (int j = i + 1; j < _packset.length(); j++) {
1589         Node_List* p2 = _packset.at(j);
1590         if (p2 == NULL) continue;
1591         if (i == j) continue;
1592         if (p1->at(p1->size()-1) == p2->at(0)) {
1593           for (uint k = 1; k < p2->size(); k++) {
1594             p1->push(p2->at(k));
1595           }
1596           _packset.at_put(j, NULL);
1597           changed = true;
1598         }
1599       }
1600     }
1601   }
1602 
1603   // Split packs which have size greater then max vector size.
1604   for (int i = 0; i < _packset.length(); i++) {
1605     Node_List* p1 = _packset.at(i);
1606     if (p1 != NULL) {
1607       BasicType bt = velt_basic_type(p1->at(0));
1608       uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1609       assert(is_power_of_2(max_vlen), "sanity");
1610       uint psize = p1->size();
1611       if (!is_power_of_2(psize)) {
1612         // Skip pack which can't be vector.
1613         // case1: for(...) { a[i] = i; }    elements values are different (i+x)
1614         // case2: for(...) { a[i] = b[i+1]; }  can't align both, load and store
1615         _packset.at_put(i, NULL);
1616         continue;
1617       }
1618       if (psize > max_vlen) {
1619         Node_List* pack = new Node_List();
1620         for (uint j = 0; j < psize; j++) {
1621           pack->push(p1->at(j));
1622           if (pack->size() >= max_vlen) {
1623             assert(is_power_of_2(pack->size()), "sanity");
1624             _packset.append(pack);
1625             pack = new Node_List();
1626           }
1627         }
1628         _packset.at_put(i, NULL);
1629       }
1630     }
1631   }
1632 
1633   // Compress list.
1634   for (int i = _packset.length() - 1; i >= 0; i--) {
1635     Node_List* p1 = _packset.at(i);
1636     if (p1 == NULL) {
1637       _packset.remove_at(i);
1638     }
1639   }
1640 
1641   if (TraceSuperWord) {
1642     tty->print_cr("\nAfter combine_packs");
1643     print_packset();
1644   }
1645 }
1646 
1647 //-----------------------------construct_my_pack_map--------------------------
1648 // Construct the map from nodes to packs.  Only valid after the
1649 // point where a node is only in one pack (after combine_packs).
1650 void SuperWord::construct_my_pack_map() {
1651   Node_List* rslt = NULL;
1652   for (int i = 0; i < _packset.length(); i++) {
1653     Node_List* p = _packset.at(i);
1654     for (uint j = 0; j < p->size(); j++) {
1655       Node* s = p->at(j);
1656       assert(my_pack(s) == NULL, "only in one pack");
1657       set_my_pack(s, p);
1658     }
1659   }
1660 }
1661 
1662 //------------------------------filter_packs---------------------------
1663 // Remove packs that are not implemented or not profitable.
1664 void SuperWord::filter_packs() {
1665   // Remove packs that are not implemented
1666   for (int i = _packset.length() - 1; i >= 0; i--) {
1667     Node_List* pk = _packset.at(i);
1668     bool impl = implemented(pk);
1669     if (!impl) {
1670 #ifndef PRODUCT
1671       if (TraceSuperWord && Verbose) {
1672         tty->print_cr("Unimplemented");
1673         pk->at(0)->dump();
1674       }
1675 #endif
1676       remove_pack_at(i);
1677     }
1678     Node *n = pk->at(0);
1679     if (n->is_reduction()) {
1680       _num_reductions++;
1681     } else {
1682       _num_work_vecs++;
1683     }
1684   }
1685 
1686   // Remove packs that are not profitable
1687   bool changed;
1688   do {
1689     changed = false;
1690     for (int i = _packset.length() - 1; i >= 0; i--) {
1691       Node_List* pk = _packset.at(i);
1692       bool prof = profitable(pk);
1693       if (!prof) {
1694 #ifndef PRODUCT
1695         if (TraceSuperWord && Verbose) {
1696           tty->print_cr("Unprofitable");
1697           pk->at(0)->dump();
1698         }
1699 #endif
1700         remove_pack_at(i);
1701         changed = true;
1702       }
1703     }
1704   } while (changed);
1705 
1706 #ifndef PRODUCT
1707   if (TraceSuperWord) {
1708     tty->print_cr("\nAfter filter_packs");
1709     print_packset();
1710     tty->cr();
1711   }
1712 #endif
1713 }
1714 
1715 //------------------------------merge_packs_to_cmovd---------------------------
1716 // Merge CMoveD into new vector-nodes
1717 // We want to catch this pattern and subsume CmpD and Bool into CMoveD
1718 //
1719 //                   SubD             ConD
1720 //                  /  |               /
1721 //                 /   |           /   /
1722 //                /    |       /      /
1723 //               /     |   /         /
1724 //              /      /            /
1725 //             /    /  |           /
1726 //            v /      |          /
1727 //         CmpD        |         /
1728 //          |          |        /
1729 //          v          |       /
1730 //         Bool        |      /
1731 //           \         |     /
1732 //             \       |    /
1733 //               \     |   /
1734 //                 \   |  /
1735 //                   \ v /
1736 //                   CMoveD
1737 //
1738 
1739 void SuperWord::merge_packs_to_cmovd() {
1740   for (int i = _packset.length() - 1; i >= 0; i--) {
1741     _cmovev_kit.make_cmovevd_pack(_packset.at(i));
1742   }
1743   #ifndef PRODUCT
1744     if (TraceSuperWord) {
1745       tty->print_cr("\nSuperWord::merge_packs_to_cmovd(): After merge");
1746       print_packset();
1747       tty->cr();
1748     }
1749   #endif
1750 }
1751 
1752 Node* CMoveKit::is_Bool_candidate(Node* def) const {
1753   Node* use = NULL;
1754   if (!def->is_Bool() || def->in(0) != NULL || def->outcnt() != 1) {
1755     return NULL;
1756   }
1757   for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1758     use = def->fast_out(j);
1759     if (!_sw->same_generation(def, use) || !use->is_CMove()) {
1760       return NULL;
1761     }
1762   }
1763   return use;
1764 }
1765 
1766 Node* CMoveKit::is_CmpD_candidate(Node* def) const {
1767   Node* use = NULL;
1768   if (!def->is_Cmp() || def->in(0) != NULL || def->outcnt() != 1) {
1769     return NULL;
1770   }
1771   for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1772     use = def->fast_out(j);
1773     if (!_sw->same_generation(def, use) || (use = is_Bool_candidate(use)) == NULL || !_sw->same_generation(def, use)) {
1774       return NULL;
1775     }
1776   }
1777   return use;
1778 }
1779 
1780 Node_List* CMoveKit::make_cmovevd_pack(Node_List* cmovd_pk) {
1781   Node *cmovd = cmovd_pk->at(0);
1782   if (!cmovd->is_CMove()) {
1783     return NULL;
1784   }
1785   if (cmovd->Opcode() != Op_CMoveF && cmovd->Opcode() != Op_CMoveD) {
1786     return NULL;
1787   }
1788   if (pack(cmovd) != NULL) { // already in the cmov pack
1789     return NULL;
1790   }
1791   if (cmovd->in(0) != NULL) {
1792     NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: CMoveD %d has control flow, escaping...", cmovd->_idx); cmovd->dump();})
1793     return NULL;
1794   }
1795 
1796   Node* bol = cmovd->as_CMove()->in(CMoveNode::Condition);
1797   if (!bol->is_Bool()
1798       || bol->outcnt() != 1
1799       || !_sw->same_generation(bol, cmovd)
1800       || bol->in(0) != NULL  // BoolNode has control flow!!
1801       || _sw->my_pack(bol) == NULL) {
1802       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();})
1803       return NULL;
1804   }
1805   Node_List* bool_pk = _sw->my_pack(bol);
1806   if (bool_pk->size() != cmovd_pk->size() ) {
1807     return NULL;
1808   }
1809 
1810   Node* cmpd = bol->in(1);
1811   if (!cmpd->is_Cmp()
1812       || cmpd->outcnt() != 1
1813       || !_sw->same_generation(cmpd, cmovd)
1814       || cmpd->in(0) != NULL  // CmpDNode has control flow!!
1815       || _sw->my_pack(cmpd) == NULL) {
1816       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();})
1817       return NULL;
1818   }
1819   Node_List* cmpd_pk = _sw->my_pack(cmpd);
1820   if (cmpd_pk->size() != cmovd_pk->size() ) {
1821     return NULL;
1822   }
1823 
1824   if (!test_cmpd_pack(cmpd_pk, cmovd_pk)) {
1825     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();})
1826     return NULL;
1827   }
1828 
1829   Node_List* new_cmpd_pk = new Node_List();
1830   uint sz = cmovd_pk->size() - 1;
1831   for (uint i = 0; i <= sz; ++i) {
1832     Node* cmov = cmovd_pk->at(i);
1833     Node* bol  = bool_pk->at(i);
1834     Node* cmp  = cmpd_pk->at(i);
1835 
1836     new_cmpd_pk->insert(i, cmov);
1837 
1838     map(cmov, new_cmpd_pk);
1839     map(bol, new_cmpd_pk);
1840     map(cmp, new_cmpd_pk);
1841 
1842     _sw->set_my_pack(cmov, new_cmpd_pk); // and keep old packs for cmp and bool
1843   }
1844   _sw->_packset.remove(cmovd_pk);
1845   _sw->_packset.remove(bool_pk);
1846   _sw->_packset.remove(cmpd_pk);
1847   _sw->_packset.append(new_cmpd_pk);
1848   NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print_cr("CMoveKit::make_cmovevd_pack: added syntactic CMoveD pack"); _sw->print_pack(new_cmpd_pk);})
1849   return new_cmpd_pk;
1850 }
1851 
1852 bool CMoveKit::test_cmpd_pack(Node_List* cmpd_pk, Node_List* cmovd_pk) {
1853   Node* cmpd0 = cmpd_pk->at(0);
1854   assert(cmpd0->is_Cmp(), "CMoveKit::test_cmpd_pack: should be CmpDNode");
1855   assert(cmovd_pk->at(0)->is_CMove(), "CMoveKit::test_cmpd_pack: should be CMoveD");
1856   assert(cmpd_pk->size() == cmovd_pk->size(), "CMoveKit::test_cmpd_pack: should be same size");
1857   Node* in1 = cmpd0->in(1);
1858   Node* in2 = cmpd0->in(2);
1859   Node_List* in1_pk = _sw->my_pack(in1);
1860   Node_List* in2_pk = _sw->my_pack(in2);
1861 
1862   if (  (in1_pk != NULL && in1_pk->size() != cmpd_pk->size())
1863      || (in2_pk != NULL && in2_pk->size() != cmpd_pk->size()) ) {
1864     return false;
1865   }
1866 
1867   // test if "all" in1 are in the same pack or the same node
1868   if (in1_pk == NULL) {
1869     for (uint j = 1; j < cmpd_pk->size(); j++) {
1870       if (cmpd_pk->at(j)->in(1) != in1) {
1871         return false;
1872       }
1873     }//for: in1_pk is not pack but all CmpD nodes in the pack have the same in(1)
1874   }
1875   // test if "all" in2 are in the same pack or the same node
1876   if (in2_pk == NULL) {
1877     for (uint j = 1; j < cmpd_pk->size(); j++) {
1878       if (cmpd_pk->at(j)->in(2) != in2) {
1879         return false;
1880       }
1881     }//for: in2_pk is not pack but all CmpD nodes in the pack have the same in(2)
1882   }
1883   //now check if cmpd_pk may be subsumed in vector built for cmovd_pk
1884   int cmovd_ind1, cmovd_ind2;
1885   if (cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1886    && cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1887       cmovd_ind1 = CMoveNode::IfFalse;
1888       cmovd_ind2 = CMoveNode::IfTrue;
1889   } else if (cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1890           && cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1891       cmovd_ind2 = CMoveNode::IfFalse;
1892       cmovd_ind1 = CMoveNode::IfTrue;
1893   }
1894   else {
1895     return false;
1896   }
1897 
1898   for (uint j = 1; j < cmpd_pk->size(); j++) {
1899     if (cmpd_pk->at(j)->in(1) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind1)
1900         || cmpd_pk->at(j)->in(2) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind2)) {
1901         return false;
1902     }//if
1903   }
1904   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(); })
1905   return true;
1906 }
1907 
1908 //------------------------------implemented---------------------------
1909 // Can code be generated for pack p?
1910 bool SuperWord::implemented(Node_List* p) {
1911   bool retValue = false;
1912   Node* p0 = p->at(0);
1913   if (p0 != NULL) {
1914     int opc = p0->Opcode();
1915     uint size = p->size();
1916     if (p0->is_reduction()) {
1917       const Type *arith_type = p0->bottom_type();
1918       // Length 2 reductions of INT/LONG do not offer performance benefits
1919       if (((arith_type->basic_type() == T_INT) || (arith_type->basic_type() == T_LONG)) && (size == 2)) {
1920         retValue = false;
1921       } else {
1922         retValue = ReductionNode::implemented(opc, size, arith_type->basic_type());
1923       }
1924     } else {
1925       retValue = VectorNode::implemented(opc, size, velt_basic_type(p0));
1926     }
1927     if (!retValue) {
1928       if (is_cmov_pack(p)) {
1929         NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::implemented: found cmpd pack"); print_pack(p);})
1930         return true;
1931       }
1932     }
1933   }
1934   return retValue;
1935 }
1936 
1937 bool SuperWord::is_cmov_pack(Node_List* p) {
1938   return _cmovev_kit.pack(p->at(0)) != NULL;
1939 }
1940 //------------------------------same_inputs--------------------------
1941 // For pack p, are all idx operands the same?
1942 bool SuperWord::same_inputs(Node_List* p, int idx) {
1943   Node* p0 = p->at(0);
1944   uint vlen = p->size();
1945   Node* p0_def = p0->in(idx);
1946   for (uint i = 1; i < vlen; i++) {
1947     Node* pi = p->at(i);
1948     Node* pi_def = pi->in(idx);
1949     if (p0_def != pi_def) {
1950       return false;
1951     }
1952   }
1953   return true;
1954 }
1955 
1956 //------------------------------profitable---------------------------
1957 // For pack p, are all operands and all uses (with in the block) vector?
1958 bool SuperWord::profitable(Node_List* p) {
1959   Node* p0 = p->at(0);
1960   uint start, end;
1961   VectorNode::vector_operands(p0, &start, &end);
1962 
1963   // Return false if some inputs are not vectors or vectors with different
1964   // size or alignment.
1965   // Also, for now, return false if not scalar promotion case when inputs are
1966   // the same. Later, implement PackNode and allow differing, non-vector inputs
1967   // (maybe just the ones from outside the block.)
1968   for (uint i = start; i < end; i++) {
1969     if (!is_vector_use(p0, i)) {
1970       return false;
1971     }
1972   }
1973   // Check if reductions are connected
1974   if (p0->is_reduction()) {
1975     Node* second_in = p0->in(2);
1976     Node_List* second_pk = my_pack(second_in);
1977     if ((second_pk == NULL) || (_num_work_vecs == _num_reductions)) {
1978       // Remove reduction flag if no parent pack or if not enough work
1979       // to cover reduction expansion overhead
1980       p0->remove_flag(Node::Flag_is_reduction);
1981       return false;
1982     } else if (second_pk->size() != p->size()) {
1983       return false;
1984     }
1985   }
1986   if (VectorNode::is_shift(p0)) {
1987     // For now, return false if shift count is vector or not scalar promotion
1988     // case (different shift counts) because it is not supported yet.
1989     Node* cnt = p0->in(2);
1990     Node_List* cnt_pk = my_pack(cnt);
1991     if (cnt_pk != NULL)
1992       return false;
1993     if (!same_inputs(p, 2))
1994       return false;
1995   }
1996   if (!p0->is_Store()) {
1997     // For now, return false if not all uses are vector.
1998     // Later, implement ExtractNode and allow non-vector uses (maybe
1999     // just the ones outside the block.)
2000     for (uint i = 0; i < p->size(); i++) {
2001       Node* def = p->at(i);
2002       if (is_cmov_pack_internal_node(p, def)) {
2003         continue;
2004       }
2005       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2006         Node* use = def->fast_out(j);
2007         for (uint k = 0; k < use->req(); k++) {
2008           Node* n = use->in(k);
2009           if (def == n) {
2010             // reductions should only have a Phi use at the the loop
2011             // head and out of loop uses
2012             if (def->is_reduction() &&
2013                 ((use->is_Phi() && use->in(0) == _lpt->_head) ||
2014                  !_lpt->is_member(_phase->get_loop(_phase->ctrl_or_self(use))))) {
2015               assert(i == p->size()-1, "must be last element of the pack");
2016               continue;
2017             }
2018             if (!is_vector_use(use, k)) {
2019               return false;
2020             }
2021           }
2022         }
2023       }
2024     }
2025   }
2026   return true;
2027 }
2028 
2029 //------------------------------schedule---------------------------
2030 // Adjust the memory graph for the packed operations
2031 void SuperWord::schedule() {
2032 
2033   // Co-locate in the memory graph the members of each memory pack
2034   for (int i = 0; i < _packset.length(); i++) {
2035     co_locate_pack(_packset.at(i));
2036   }
2037 }
2038 
2039 //-------------------------------remove_and_insert-------------------
2040 // Remove "current" from its current position in the memory graph and insert
2041 // it after the appropriate insertion point (lip or uip).
2042 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
2043                                   Node *uip, Unique_Node_List &sched_before) {
2044   Node* my_mem = current->in(MemNode::Memory);
2045   bool sched_up = sched_before.member(current);
2046 
2047   // remove current_store from its current position in the memmory graph
2048   for (DUIterator i = current->outs(); current->has_out(i); i++) {
2049     Node* use = current->out(i);
2050     if (use->is_Mem()) {
2051       assert(use->in(MemNode::Memory) == current, "must be");
2052       if (use == prev) { // connect prev to my_mem
2053           _igvn.replace_input_of(use, MemNode::Memory, my_mem);
2054           --i; //deleted this edge; rescan position
2055       } else if (sched_before.member(use)) {
2056         if (!sched_up) { // Will be moved together with current
2057           _igvn.replace_input_of(use, MemNode::Memory, uip);
2058           --i; //deleted this edge; rescan position
2059         }
2060       } else {
2061         if (sched_up) { // Will be moved together with current
2062           _igvn.replace_input_of(use, MemNode::Memory, lip);
2063           --i; //deleted this edge; rescan position
2064         }
2065       }
2066     }
2067   }
2068 
2069   Node *insert_pt =  sched_up ?  uip : lip;
2070 
2071   // all uses of insert_pt's memory state should use current's instead
2072   for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
2073     Node* use = insert_pt->out(i);
2074     if (use->is_Mem()) {
2075       assert(use->in(MemNode::Memory) == insert_pt, "must be");
2076       _igvn.replace_input_of(use, MemNode::Memory, current);
2077       --i; //deleted this edge; rescan position
2078     } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
2079       uint pos; //lip (lower insert point) must be the last one in the memory slice
2080       for (pos=1; pos < use->req(); pos++) {
2081         if (use->in(pos) == insert_pt) break;
2082       }
2083       _igvn.replace_input_of(use, pos, current);
2084       --i;
2085     }
2086   }
2087 
2088   //connect current to insert_pt
2089   _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
2090 }
2091 
2092 //------------------------------co_locate_pack----------------------------------
2093 // To schedule a store pack, we need to move any sandwiched memory ops either before
2094 // or after the pack, based upon dependence information:
2095 // (1) If any store in the pack depends on the sandwiched memory op, the
2096 //     sandwiched memory op must be scheduled BEFORE the pack;
2097 // (2) If a sandwiched memory op depends on any store in the pack, the
2098 //     sandwiched memory op must be scheduled AFTER the pack;
2099 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
2100 //     memory op (say memB), memB must be scheduled before memA. So, if memA is
2101 //     scheduled before the pack, memB must also be scheduled before the pack;
2102 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
2103 //     schedule this store AFTER the pack
2104 // (5) We know there is no dependence cycle, so there in no other case;
2105 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
2106 //
2107 // To schedule a load pack, we use the memory state of either the first or the last load in
2108 // the pack, based on the dependence constraint.
2109 void SuperWord::co_locate_pack(Node_List* pk) {
2110   if (pk->at(0)->is_Store()) {
2111     MemNode* first     = executed_first(pk)->as_Mem();
2112     MemNode* last      = executed_last(pk)->as_Mem();
2113     Unique_Node_List schedule_before_pack;
2114     Unique_Node_List memops;
2115 
2116     MemNode* current   = last->in(MemNode::Memory)->as_Mem();
2117     MemNode* previous  = last;
2118     while (true) {
2119       assert(in_bb(current), "stay in block");
2120       memops.push(previous);
2121       for (DUIterator i = current->outs(); current->has_out(i); i++) {
2122         Node* use = current->out(i);
2123         if (use->is_Mem() && use != previous)
2124           memops.push(use);
2125       }
2126       if (current == first) break;
2127       previous = current;
2128       current  = current->in(MemNode::Memory)->as_Mem();
2129     }
2130 
2131     // determine which memory operations should be scheduled before the pack
2132     for (uint i = 1; i < memops.size(); i++) {
2133       Node *s1 = memops.at(i);
2134       if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
2135         for (uint j = 0; j< i; j++) {
2136           Node *s2 = memops.at(j);
2137           if (!independent(s1, s2)) {
2138             if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
2139               schedule_before_pack.push(s1); // s1 must be scheduled before
2140               Node_List* mem_pk = my_pack(s1);
2141               if (mem_pk != NULL) {
2142                 for (uint ii = 0; ii < mem_pk->size(); ii++) {
2143                   Node* s = mem_pk->at(ii);  // follow partner
2144                   if (memops.member(s) && !schedule_before_pack.member(s))
2145                     schedule_before_pack.push(s);
2146                 }
2147               }
2148               break;
2149             }
2150           }
2151         }
2152       }
2153     }
2154 
2155     Node*    upper_insert_pt = first->in(MemNode::Memory);
2156     // Following code moves loads connected to upper_insert_pt below aliased stores.
2157     // Collect such loads here and reconnect them back to upper_insert_pt later.
2158     memops.clear();
2159     for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
2160       Node* use = upper_insert_pt->out(i);
2161       if (use->is_Mem() && !use->is_Store()) {
2162         memops.push(use);
2163       }
2164     }
2165 
2166     MemNode* lower_insert_pt = last;
2167     previous                 = last; //previous store in pk
2168     current                  = last->in(MemNode::Memory)->as_Mem();
2169 
2170     // start scheduling from "last" to "first"
2171     while (true) {
2172       assert(in_bb(current), "stay in block");
2173       assert(in_pack(previous, pk), "previous stays in pack");
2174       Node* my_mem = current->in(MemNode::Memory);
2175 
2176       if (in_pack(current, pk)) {
2177         // Forward users of my memory state (except "previous) to my input memory state
2178         for (DUIterator i = current->outs(); current->has_out(i); i++) {
2179           Node* use = current->out(i);
2180           if (use->is_Mem() && use != previous) {
2181             assert(use->in(MemNode::Memory) == current, "must be");
2182             if (schedule_before_pack.member(use)) {
2183               _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
2184             } else {
2185               _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
2186             }
2187             --i; // deleted this edge; rescan position
2188           }
2189         }
2190         previous = current;
2191       } else { // !in_pack(current, pk) ==> a sandwiched store
2192         remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
2193       }
2194 
2195       if (current == first) break;
2196       current = my_mem->as_Mem();
2197     } // end while
2198 
2199     // Reconnect loads back to upper_insert_pt.
2200     for (uint i = 0; i < memops.size(); i++) {
2201       Node *ld = memops.at(i);
2202       if (ld->in(MemNode::Memory) != upper_insert_pt) {
2203         _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
2204       }
2205     }
2206   } else if (pk->at(0)->is_Load()) { //load
2207     // all loads in the pack should have the same memory state. By default,
2208     // we use the memory state of the last load. However, if any load could
2209     // not be moved down due to the dependence constraint, we use the memory
2210     // state of the first load.
2211     Node* first_mem = pk->at(0)->in(MemNode::Memory);
2212     Node* last_mem = first_mem;
2213     for (uint i = 1; i < pk->size(); i++) {
2214       Node* ld = pk->at(i);
2215       Node* mem = ld->in(MemNode::Memory);
2216       assert(in_bb(first_mem) || in_bb(mem) || mem == first_mem, "2 different memory state from outside the loop?");
2217       if (in_bb(mem)) {
2218         if (in_bb(first_mem) && bb_idx(mem) < bb_idx(first_mem)) {
2219           first_mem = mem;
2220         }
2221         if (!in_bb(last_mem) || bb_idx(mem) > bb_idx(last_mem)) {
2222           last_mem = mem;
2223         }
2224       }
2225     }
2226     bool schedule_last = true;
2227     for (uint i = 0; i < pk->size(); i++) {
2228       Node* ld = pk->at(i);
2229       for (Node* current = last_mem; current != ld->in(MemNode::Memory);
2230            current=current->in(MemNode::Memory)) {
2231         assert(current != first_mem, "corrupted memory graph");
2232         if(current->is_Mem() && !independent(current, ld)){
2233           schedule_last = false; // a later store depends on this load
2234           break;
2235         }
2236       }
2237     }
2238 
2239     Node* mem_input = schedule_last ? last_mem : first_mem;
2240     _igvn.hash_delete(mem_input);
2241     // Give each load the same memory state
2242     for (uint i = 0; i < pk->size(); i++) {
2243       LoadNode* ld = pk->at(i)->as_Load();
2244       _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
2245     }
2246   }
2247 }
2248 
2249 #ifndef PRODUCT
2250 void SuperWord::print_loop(bool whole) {
2251   Node_Stack stack(_arena, _phase->C->unique() >> 2);
2252   Node_List rpo_list;
2253   VectorSet visited(_arena);
2254   visited.set(lpt()->_head->_idx);
2255   _phase->rpo(lpt()->_head, stack, visited, rpo_list);
2256   _phase->dump(lpt(), rpo_list.size(), rpo_list );
2257   if(whole) {
2258     tty->print_cr("\n Whole loop tree");
2259     _phase->dump();
2260     tty->print_cr(" End of whole loop tree\n");
2261   }
2262 }
2263 #endif
2264 
2265 //------------------------------output---------------------------
2266 // Convert packs into vector node operations
2267 void SuperWord::output() {
2268   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2269   Compile* C = _phase->C;
2270   if (_packset.length() == 0) {
2271     if (cl->is_main_loop()) {
2272       // Instigate more unrolling for optimization when vectorization fails.
2273       C->set_major_progress();
2274       cl->set_notpassed_slp();
2275       cl->mark_do_unroll_only();
2276     }
2277     return;
2278   }
2279 
2280 #ifndef PRODUCT
2281   if (TraceLoopOpts) {
2282     tty->print("SuperWord::output    ");
2283     lpt()->dump_head();
2284   }
2285 #endif
2286 
2287   if (cl->is_main_loop()) {
2288     // MUST ENSURE main loop's initial value is properly aligned:
2289     //  (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
2290 
2291     align_initial_loop_index(align_to_ref());
2292 
2293     // Insert extract (unpack) operations for scalar uses
2294     for (int i = 0; i < _packset.length(); i++) {
2295       insert_extracts(_packset.at(i));
2296     }
2297   }
2298 
2299   uint max_vlen_in_bytes = 0;
2300   uint max_vlen = 0;
2301   bool can_process_post_loop = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
2302 
2303   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop before create_reserve_version_of_loop"); print_loop(true);})
2304 
2305   CountedLoopReserveKit make_reversable(_phase, _lpt, do_reserve_copy());
2306 
2307   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop after create_reserve_version_of_loop"); print_loop(true);})
2308 
2309   if (do_reserve_copy() && !make_reversable.has_reserved()) {
2310     NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: loop was not reserved correctly, exiting SuperWord");})
2311     return;
2312   }
2313 
2314   for (int i = 0; i < _block.length(); i++) {
2315     Node* n = _block.at(i);
2316     Node_List* p = my_pack(n);
2317     if (p && n == executed_last(p)) {
2318       uint vlen = p->size();
2319       uint vlen_in_bytes = 0;
2320       Node* vn = NULL;
2321       Node* low_adr = p->at(0);
2322       Node* first   = executed_first(p);
2323       if (can_process_post_loop) {
2324         // override vlen with the main loops vector length
2325         vlen = cl->slp_max_unroll();
2326       }
2327       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);})
2328       int   opc = n->Opcode();
2329       if (n->is_Load()) {
2330         Node* ctl = n->in(MemNode::Control);
2331         Node* mem = first->in(MemNode::Memory);
2332         SWPointer p1(n->as_Mem(), this, NULL, false);
2333         // Identify the memory dependency for the new loadVector node by
2334         // walking up through memory chain.
2335         // This is done to give flexibility to the new loadVector node so that
2336         // it can move above independent storeVector nodes.
2337         while (mem->is_StoreVector()) {
2338           SWPointer p2(mem->as_Mem(), this, NULL, false);
2339           int cmp = p1.cmp(p2);
2340           if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
2341             mem = mem->in(MemNode::Memory);
2342           } else {
2343             break; // dependent memory
2344           }
2345         }
2346         Node* adr = low_adr->in(MemNode::Address);
2347         const TypePtr* atyp = n->adr_type();
2348         vn = LoadVectorNode::make(opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n), control_dependency(p));
2349         vlen_in_bytes = vn->as_LoadVector()->memory_size();
2350       } else if (n->is_Store()) {
2351         // Promote value to be stored to vector
2352         Node* val = vector_opd(p, MemNode::ValueIn);
2353         if (val == NULL) {
2354           if (do_reserve_copy()) {
2355             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: val should not be NULL, exiting SuperWord");})
2356             return; //and reverse to backup IG
2357           }
2358           ShouldNotReachHere();
2359         }
2360 
2361         Node* ctl = n->in(MemNode::Control);
2362         Node* mem = first->in(MemNode::Memory);
2363         Node* adr = low_adr->in(MemNode::Address);
2364         const TypePtr* atyp = n->adr_type();
2365         vn = StoreVectorNode::make(opc, ctl, mem, adr, atyp, val, vlen);
2366         vlen_in_bytes = vn->as_StoreVector()->memory_size();
2367       } else if (VectorNode::is_muladds2i(n)) {
2368         assert(n->req() == 5u, "MulAddS2I should have 4 operands.");
2369         Node* in1 = vector_opd(p, 1);
2370         Node* in2 = vector_opd(p, 2);
2371         vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2372         vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2373       } else if (n->req() == 3 && !is_cmov_pack(p)) {
2374         // Promote operands to vector
2375         Node* in1 = NULL;
2376         bool node_isa_reduction = n->is_reduction();
2377         if (node_isa_reduction) {
2378           // the input to the first reduction operation is retained
2379           in1 = low_adr->in(1);
2380         } else {
2381           in1 = vector_opd(p, 1);
2382           if (in1 == NULL) {
2383             if (do_reserve_copy()) {
2384               NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in1 should not be NULL, exiting SuperWord");})
2385               return; //and reverse to backup IG
2386             }
2387             ShouldNotReachHere();
2388           }
2389         }
2390         Node* in2 = vector_opd(p, 2);
2391         if (in2 == NULL) {
2392           if (do_reserve_copy()) {
2393             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in2 should not be NULL, exiting SuperWord");})
2394             return; //and reverse to backup IG
2395           }
2396           ShouldNotReachHere();
2397         }
2398         if (VectorNode::is_invariant_vector(in1) && (node_isa_reduction == false) && (n->is_Add() || n->is_Mul())) {
2399           // Move invariant vector input into second position to avoid register spilling.
2400           Node* tmp = in1;
2401           in1 = in2;
2402           in2 = tmp;
2403         }
2404         if (node_isa_reduction) {
2405           const Type *arith_type = n->bottom_type();
2406           vn = ReductionNode::make(opc, NULL, in1, in2, arith_type->basic_type());
2407           if (in2->is_Load()) {
2408             vlen_in_bytes = in2->as_LoadVector()->memory_size();
2409           } else {
2410             vlen_in_bytes = in2->as_Vector()->length_in_bytes();
2411           }
2412         } else {
2413           vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2414           vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2415         }
2416       } else if (opc == Op_SqrtF || opc == Op_SqrtD ||
2417                  opc == Op_AbsF || opc == Op_AbsD ||
2418                  opc == Op_NegF || opc == Op_NegD ||
2419                  opc == Op_PopCountI) {
2420         assert(n->req() == 2, "only one input expected");
2421         Node* in = vector_opd(p, 1);
2422         vn = VectorNode::make(opc, in, NULL, vlen, velt_basic_type(n));
2423         vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2424       } else if (is_cmov_pack(p)) {
2425         if (can_process_post_loop) {
2426           // do not refactor of flow in post loop context
2427           return;
2428         }
2429         if (!n->is_CMove()) {
2430           continue;
2431         }
2432         // place here CMoveVDNode
2433         NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::output: print before CMove vectorization"); print_loop(false);})
2434         Node* bol = n->in(CMoveNode::Condition);
2435         if (!bol->is_Bool() && bol->Opcode() == Op_ExtractI && bol->req() > 1 ) {
2436           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();})
2437           bol = bol->in(1); //may be ExtractNode
2438         }
2439 
2440         assert(bol->is_Bool(), "should be BoolNode - too late to bail out!");
2441         if (!bol->is_Bool()) {
2442           if (do_reserve_copy()) {
2443             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: expected %d bool node, exiting SuperWord", bol->_idx); bol->dump();})
2444             return; //and reverse to backup IG
2445           }
2446           ShouldNotReachHere();
2447         }
2448 
2449         int cond = (int)bol->as_Bool()->_test._test;
2450         Node* in_cc  = _igvn.intcon(cond);
2451         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created intcon in_cc node %d", in_cc->_idx); in_cc->dump();})
2452         Node* cc = bol->clone();
2453         cc->set_req(1, in_cc);
2454         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created bool cc node %d", cc->_idx); cc->dump();})
2455 
2456         Node* src1 = vector_opd(p, 2); //2=CMoveNode::IfFalse
2457         if (src1 == NULL) {
2458           if (do_reserve_copy()) {
2459             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src1 should not be NULL, exiting SuperWord");})
2460             return; //and reverse to backup IG
2461           }
2462           ShouldNotReachHere();
2463         }
2464         Node* src2 = vector_opd(p, 3); //3=CMoveNode::IfTrue
2465         if (src2 == NULL) {
2466           if (do_reserve_copy()) {
2467             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src2 should not be NULL, exiting SuperWord");})
2468             return; //and reverse to backup IG
2469           }
2470           ShouldNotReachHere();
2471         }
2472         BasicType bt = velt_basic_type(n);
2473         const TypeVect* vt = TypeVect::make(bt, vlen);
2474         assert(bt == T_FLOAT || bt == T_DOUBLE, "Only vectorization for FP cmovs is supported");
2475         if (bt == T_FLOAT) {
2476           vn = new CMoveVFNode(cc, src1, src2, vt);
2477         } else {
2478           assert(bt == T_DOUBLE, "Expected double");
2479           vn = new CMoveVDNode(cc, src1, src2, vt);
2480         }
2481         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created new CMove node %d: ", vn->_idx); vn->dump();})
2482       } else if (opc == Op_FmaD || opc == Op_FmaF) {
2483         // Promote operands to vector
2484         Node* in1 = vector_opd(p, 1);
2485         Node* in2 = vector_opd(p, 2);
2486         Node* in3 = vector_opd(p, 3);
2487         vn = VectorNode::make(opc, in1, in2, in3, vlen, velt_basic_type(n));
2488         vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2489       } else {
2490         if (do_reserve_copy()) {
2491           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: ShouldNotReachHere, exiting SuperWord");})
2492           return; //and reverse to backup IG
2493         }
2494         ShouldNotReachHere();
2495       }
2496 
2497       assert(vn != NULL, "sanity");
2498       if (vn == NULL) {
2499         if (do_reserve_copy()){
2500           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: got NULL node, cannot proceed, exiting SuperWord");})
2501           return; //and reverse to backup IG
2502         }
2503         ShouldNotReachHere();
2504       }
2505 
2506       _block.at_put(i, vn);
2507       _igvn.register_new_node_with_optimizer(vn);
2508       _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
2509       for (uint j = 0; j < p->size(); j++) {
2510         Node* pm = p->at(j);
2511         _igvn.replace_node(pm, vn);
2512       }
2513       _igvn._worklist.push(vn);
2514 
2515       if (can_process_post_loop) {
2516         // first check if the vector size if the maximum vector which we can use on the machine,
2517         // other vector size have reduced values for predicated data mapping.
2518         if (vlen_in_bytes != (uint)MaxVectorSize) {
2519           return;
2520         }
2521       }
2522 
2523       if (vlen_in_bytes >= max_vlen_in_bytes && vlen > max_vlen) {
2524         max_vlen = vlen;
2525         max_vlen_in_bytes = vlen_in_bytes;
2526       }
2527 #ifdef ASSERT
2528       if (TraceNewVectors) {
2529         tty->print("new Vector node: ");
2530         vn->dump();
2531       }
2532 #endif
2533     }
2534   }//for (int i = 0; i < _block.length(); i++)
2535 
2536   if (max_vlen_in_bytes > C->max_vector_size()) {
2537     C->set_max_vector_size(max_vlen_in_bytes);
2538   }
2539   if (max_vlen_in_bytes > 0) {
2540     cl->mark_loop_vectorized();
2541   }
2542 
2543   if (SuperWordLoopUnrollAnalysis) {
2544     if (cl->has_passed_slp()) {
2545       uint slp_max_unroll_factor = cl->slp_max_unroll();
2546       if (slp_max_unroll_factor == max_vlen) {
2547         if (TraceSuperWordLoopUnrollAnalysis) {
2548           tty->print_cr("vector loop(unroll=%d, len=%d)\n", max_vlen, max_vlen_in_bytes*BitsPerByte);
2549         }
2550 
2551         // For atomic unrolled loops which are vector mapped, instigate more unrolling
2552         cl->set_notpassed_slp();
2553         if (cl->is_main_loop()) {
2554           // if vector resources are limited, do not allow additional unrolling, also
2555           // do not unroll more on pure vector loops which were not reduced so that we can
2556           // program the post loop to single iteration execution.
2557           if (FLOATPRESSURE > 8) {
2558             C->set_major_progress();
2559             cl->mark_do_unroll_only();
2560           }
2561         }
2562 
2563         if (do_reserve_copy()) {
2564           if (can_process_post_loop) {
2565             // Now create the difference of trip and limit and use it as our mask index.
2566             // Note: We limited the unroll of the vectorized loop so that
2567             //       only vlen-1 size iterations can remain to be mask programmed.
2568             Node *incr = cl->incr();
2569             SubINode *index = new SubINode(cl->limit(), cl->init_trip());
2570             _igvn.register_new_node_with_optimizer(index);
2571             SetVectMaskINode  *mask = new SetVectMaskINode(_phase->get_ctrl(cl->init_trip()), index);
2572             _igvn.register_new_node_with_optimizer(mask);
2573             // make this a single iteration loop
2574             AddINode *new_incr = new AddINode(incr->in(1), mask);
2575             _igvn.register_new_node_with_optimizer(new_incr);
2576             _phase->set_ctrl(new_incr, _phase->get_ctrl(incr));
2577             _igvn.replace_node(incr, new_incr);
2578             cl->mark_is_multiversioned();
2579             cl->loopexit()->add_flag(Node::Flag_has_vector_mask_set);
2580           }
2581         }
2582       }
2583     }
2584   }
2585 
2586   if (do_reserve_copy()) {
2587     make_reversable.use_new();
2588   }
2589   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("\n Final loop after SuperWord"); print_loop(true);})
2590   return;
2591 }
2592 
2593 //------------------------------vector_opd---------------------------
2594 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
2595 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
2596   Node* p0 = p->at(0);
2597   uint vlen = p->size();
2598   Node* opd = p0->in(opd_idx);
2599   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2600 
2601   if (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop()) {
2602     // override vlen with the main loops vector length
2603     vlen = cl->slp_max_unroll();
2604   }
2605 
2606   if (same_inputs(p, opd_idx)) {
2607     if (opd->is_Vector() || opd->is_LoadVector()) {
2608       assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
2609       if (opd_idx == 2 && VectorNode::is_shift(p0)) {
2610         NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("shift's count can't be vector");})
2611         return NULL;
2612       }
2613       return opd; // input is matching vector
2614     }
2615     if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
2616       Compile* C = _phase->C;
2617       Node* cnt = opd;
2618       // Vector instructions do not mask shift count, do it here.
2619       juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2620       const TypeInt* t = opd->find_int_type();
2621       if (t != NULL && t->is_con()) {
2622         juint shift = t->get_con();
2623         if (shift > mask) { // Unsigned cmp
2624           cnt = ConNode::make(TypeInt::make(shift & mask));
2625         }
2626       } else {
2627         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2628           cnt = ConNode::make(TypeInt::make(mask));
2629           _igvn.register_new_node_with_optimizer(cnt);
2630           cnt = new AndINode(opd, cnt);
2631           _igvn.register_new_node_with_optimizer(cnt);
2632           _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2633         }
2634         assert(opd->bottom_type()->isa_int(), "int type only");
2635         if (!opd->bottom_type()->isa_int()) {
2636           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should be int type only");})
2637           return NULL;
2638         }
2639         // Move non constant shift count into vector register.
2640         cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0));
2641       }
2642       if (cnt != opd) {
2643         _igvn.register_new_node_with_optimizer(cnt);
2644         _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2645       }
2646       return cnt;
2647     }
2648     assert(!opd->is_StoreVector(), "such vector is not expected here");
2649     if (opd->is_StoreVector()) {
2650       NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("StoreVector is not expected here");})
2651       return NULL;
2652     }
2653     // Convert scalar input to vector with the same number of elements as
2654     // p0's vector. Use p0's type because size of operand's container in
2655     // vector should match p0's size regardless operand's size.
2656     const Type* p0_t = velt_type(p0);
2657     VectorNode* vn = VectorNode::scalar2vector(opd, vlen, p0_t);
2658 
2659     _igvn.register_new_node_with_optimizer(vn);
2660     _phase->set_ctrl(vn, _phase->get_ctrl(opd));
2661 #ifdef ASSERT
2662     if (TraceNewVectors) {
2663       tty->print("new Vector node: ");
2664       vn->dump();
2665     }
2666 #endif
2667     return vn;
2668   }
2669 
2670   // Insert pack operation
2671   BasicType bt = velt_basic_type(p0);
2672   PackNode* pk = PackNode::make(opd, vlen, bt);
2673   DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
2674 
2675   for (uint i = 1; i < vlen; i++) {
2676     Node* pi = p->at(i);
2677     Node* in = pi->in(opd_idx);
2678     assert(my_pack(in) == NULL, "Should already have been unpacked");
2679     if (my_pack(in) != NULL) {
2680       NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should already have been unpacked");})
2681       return NULL;
2682     }
2683     assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
2684     pk->add_opd(in);
2685     if (VectorNode::is_muladds2i(pi)) {
2686       Node* in2 = pi->in(opd_idx + 2);
2687       assert(my_pack(in2) == NULL, "Should already have been unpacked");
2688       if (my_pack(in2) != NULL) {
2689         NOT_PRODUCT(if (is_trace_loop_reverse() || TraceLoopOpts) { tty->print_cr("Should already have been unpacked"); })
2690           return NULL;
2691       }
2692       assert(opd_bt == in2->bottom_type()->basic_type(), "all same type");
2693       pk->add_opd(in2);
2694     }
2695   }
2696   _igvn.register_new_node_with_optimizer(pk);
2697   _phase->set_ctrl(pk, _phase->get_ctrl(opd));
2698 #ifdef ASSERT
2699   if (TraceNewVectors) {
2700     tty->print("new Vector node: ");
2701     pk->dump();
2702   }
2703 #endif
2704   return pk;
2705 }
2706 
2707 //------------------------------insert_extracts---------------------------
2708 // If a use of pack p is not a vector use, then replace the
2709 // use with an extract operation.
2710 void SuperWord::insert_extracts(Node_List* p) {
2711   if (p->at(0)->is_Store()) return;
2712   assert(_n_idx_list.is_empty(), "empty (node,index) list");
2713 
2714   // Inspect each use of each pack member.  For each use that is
2715   // not a vector use, replace the use with an extract operation.
2716 
2717   for (uint i = 0; i < p->size(); i++) {
2718     Node* def = p->at(i);
2719     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2720       Node* use = def->fast_out(j);
2721       for (uint k = 0; k < use->req(); k++) {
2722         Node* n = use->in(k);
2723         if (def == n) {
2724           Node_List* u_pk = my_pack(use);
2725           if ((u_pk == NULL || !is_cmov_pack(u_pk) || use->is_CMove()) && !is_vector_use(use, k)) {
2726               _n_idx_list.push(use, k);
2727           }
2728         }
2729       }
2730     }
2731   }
2732 
2733   while (_n_idx_list.is_nonempty()) {
2734     Node* use = _n_idx_list.node();
2735     int   idx = _n_idx_list.index();
2736     _n_idx_list.pop();
2737     Node* def = use->in(idx);
2738 
2739     if (def->is_reduction()) continue;
2740 
2741     // Insert extract operation
2742     _igvn.hash_delete(def);
2743     int def_pos = alignment(def) / data_size(def);
2744 
2745     Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def));
2746     _igvn.register_new_node_with_optimizer(ex);
2747     _phase->set_ctrl(ex, _phase->get_ctrl(def));
2748     _igvn.replace_input_of(use, idx, ex);
2749     _igvn._worklist.push(def);
2750 
2751     bb_insert_after(ex, bb_idx(def));
2752     set_velt_type(ex, velt_type(def));
2753   }
2754 }
2755 
2756 //------------------------------is_vector_use---------------------------
2757 // Is use->in(u_idx) a vector use?
2758 bool SuperWord::is_vector_use(Node* use, int u_idx) {
2759   Node_List* u_pk = my_pack(use);
2760   if (u_pk == NULL) return false;
2761   if (use->is_reduction()) return true;
2762   Node* def = use->in(u_idx);
2763   Node_List* d_pk = my_pack(def);
2764   if (d_pk == NULL) {
2765     // check for scalar promotion
2766     Node* n = u_pk->at(0)->in(u_idx);
2767     for (uint i = 1; i < u_pk->size(); i++) {
2768       if (u_pk->at(i)->in(u_idx) != n) return false;
2769     }
2770     return true;
2771   }
2772   if (VectorNode::is_muladds2i(use)) {
2773     // MulAddS2I takes shorts and produces ints - hence the special checks
2774     // on alignment and size.
2775     if (u_pk->size() * 2 != d_pk->size()) {
2776       return false;
2777     }
2778     for (uint i = 0; i < MIN2(d_pk->size(), u_pk->size()); i++) {
2779       Node* ui = u_pk->at(i);
2780       Node* di = d_pk->at(i);
2781       if (alignment(ui) != alignment(di) * 2) {
2782         return false;
2783       }
2784     }
2785     return true;
2786   }
2787   if (u_pk->size() != d_pk->size())
2788     return false;
2789   for (uint i = 0; i < u_pk->size(); i++) {
2790     Node* ui = u_pk->at(i);
2791     Node* di = d_pk->at(i);
2792     if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
2793       return false;
2794   }
2795   return true;
2796 }
2797 
2798 //------------------------------construct_bb---------------------------
2799 // Construct reverse postorder list of block members
2800 bool SuperWord::construct_bb() {
2801   Node* entry = bb();
2802 
2803   assert(_stk.length() == 0,            "stk is empty");
2804   assert(_block.length() == 0,          "block is empty");
2805   assert(_data_entry.length() == 0,     "data_entry is empty");
2806   assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
2807   assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
2808 
2809   // Find non-control nodes with no inputs from within block,
2810   // create a temporary map from node _idx to bb_idx for use
2811   // by the visited and post_visited sets,
2812   // and count number of nodes in block.
2813   int bb_ct = 0;
2814   for (uint i = 0; i < lpt()->_body.size(); i++) {
2815     Node *n = lpt()->_body.at(i);
2816     set_bb_idx(n, i); // Create a temporary map
2817     if (in_bb(n)) {
2818       if (n->is_LoadStore() || n->is_MergeMem() ||
2819           (n->is_Proj() && !n->as_Proj()->is_CFG())) {
2820         // Bailout if the loop has LoadStore, MergeMem or data Proj
2821         // nodes. Superword optimization does not work with them.
2822         return false;
2823       }
2824       bb_ct++;
2825       if (!n->is_CFG()) {
2826         bool found = false;
2827         for (uint j = 0; j < n->req(); j++) {
2828           Node* def = n->in(j);
2829           if (def && in_bb(def)) {
2830             found = true;
2831             break;
2832           }
2833         }
2834         if (!found) {
2835           assert(n != entry, "can't be entry");
2836           _data_entry.push(n);
2837         }
2838       }
2839     }
2840   }
2841 
2842   // Find memory slices (head and tail)
2843   for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
2844     Node *n = lp()->fast_out(i);
2845     if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
2846       Node* n_tail  = n->in(LoopNode::LoopBackControl);
2847       if (n_tail != n->in(LoopNode::EntryControl)) {
2848         if (!n_tail->is_Mem()) {
2849           assert(n_tail->is_Mem(), "unexpected node for memory slice: %s", n_tail->Name());
2850           return false; // Bailout
2851         }
2852         _mem_slice_head.push(n);
2853         _mem_slice_tail.push(n_tail);
2854       }
2855     }
2856   }
2857 
2858   // Create an RPO list of nodes in block
2859 
2860   visited_clear();
2861   post_visited_clear();
2862 
2863   // Push all non-control nodes with no inputs from within block, then control entry
2864   for (int j = 0; j < _data_entry.length(); j++) {
2865     Node* n = _data_entry.at(j);
2866     visited_set(n);
2867     _stk.push(n);
2868   }
2869   visited_set(entry);
2870   _stk.push(entry);
2871 
2872   // Do a depth first walk over out edges
2873   int rpo_idx = bb_ct - 1;
2874   int size;
2875   int reduction_uses = 0;
2876   while ((size = _stk.length()) > 0) {
2877     Node* n = _stk.top(); // Leave node on stack
2878     if (!visited_test_set(n)) {
2879       // forward arc in graph
2880     } else if (!post_visited_test(n)) {
2881       // cross or back arc
2882       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2883         Node *use = n->fast_out(i);
2884         if (in_bb(use) && !visited_test(use) &&
2885             // Don't go around backedge
2886             (!use->is_Phi() || n == entry)) {
2887           if (use->is_reduction()) {
2888             // First see if we can map the reduction on the given system we are on, then
2889             // make a data entry operation for each reduction we see.
2890             BasicType bt = use->bottom_type()->basic_type();
2891             if (ReductionNode::implemented(use->Opcode(), Matcher::min_vector_size(bt), bt)) {
2892               reduction_uses++;
2893             }
2894           }
2895           _stk.push(use);
2896         }
2897       }
2898       if (_stk.length() == size) {
2899         // There were no additional uses, post visit node now
2900         _stk.pop(); // Remove node from stack
2901         assert(rpo_idx >= 0, "");
2902         _block.at_put_grow(rpo_idx, n);
2903         rpo_idx--;
2904         post_visited_set(n);
2905         assert(rpo_idx >= 0 || _stk.is_empty(), "");
2906       }
2907     } else {
2908       _stk.pop(); // Remove post-visited node from stack
2909     }
2910   }//while
2911 
2912   int ii_current = -1;
2913   unsigned int load_idx = (unsigned int)-1;
2914   _ii_order.clear();
2915   // Create real map of block indices for nodes
2916   for (int j = 0; j < _block.length(); j++) {
2917     Node* n = _block.at(j);
2918     set_bb_idx(n, j);
2919     if (_do_vector_loop && n->is_Load()) {
2920       if (ii_current == -1) {
2921         ii_current = _clone_map.gen(n->_idx);
2922         _ii_order.push(ii_current);
2923         load_idx = _clone_map.idx(n->_idx);
2924       } else if (_clone_map.idx(n->_idx) == load_idx && _clone_map.gen(n->_idx) != ii_current) {
2925         ii_current = _clone_map.gen(n->_idx);
2926         _ii_order.push(ii_current);
2927       }
2928     }
2929   }//for
2930 
2931   // Ensure extra info is allocated.
2932   initialize_bb();
2933 
2934 #ifndef PRODUCT
2935   if (_vector_loop_debug && _ii_order.length() > 0) {
2936     tty->print("SuperWord::construct_bb: List of generations: ");
2937     for (int jj = 0; jj < _ii_order.length(); ++jj) {
2938       tty->print("  %d:%d", jj, _ii_order.at(jj));
2939     }
2940     tty->print_cr(" ");
2941   }
2942   if (TraceSuperWord) {
2943     print_bb();
2944     tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
2945     for (int m = 0; m < _data_entry.length(); m++) {
2946       tty->print("%3d ", m);
2947       _data_entry.at(m)->dump();
2948     }
2949     tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
2950     for (int m = 0; m < _mem_slice_head.length(); m++) {
2951       tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
2952       tty->print("    ");    _mem_slice_tail.at(m)->dump();
2953     }
2954   }
2955 #endif
2956   assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
2957   return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0);
2958 }
2959 
2960 //------------------------------initialize_bb---------------------------
2961 // Initialize per node info
2962 void SuperWord::initialize_bb() {
2963   Node* last = _block.at(_block.length() - 1);
2964   grow_node_info(bb_idx(last));
2965 }
2966 
2967 //------------------------------bb_insert_after---------------------------
2968 // Insert n into block after pos
2969 void SuperWord::bb_insert_after(Node* n, int pos) {
2970   int n_pos = pos + 1;
2971   // Make room
2972   for (int i = _block.length() - 1; i >= n_pos; i--) {
2973     _block.at_put_grow(i+1, _block.at(i));
2974   }
2975   for (int j = _node_info.length() - 1; j >= n_pos; j--) {
2976     _node_info.at_put_grow(j+1, _node_info.at(j));
2977   }
2978   // Set value
2979   _block.at_put_grow(n_pos, n);
2980   _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
2981   // Adjust map from node->_idx to _block index
2982   for (int i = n_pos; i < _block.length(); i++) {
2983     set_bb_idx(_block.at(i), i);
2984   }
2985 }
2986 
2987 //------------------------------compute_max_depth---------------------------
2988 // Compute max depth for expressions from beginning of block
2989 // Use to prune search paths during test for independence.
2990 void SuperWord::compute_max_depth() {
2991   int ct = 0;
2992   bool again;
2993   do {
2994     again = false;
2995     for (int i = 0; i < _block.length(); i++) {
2996       Node* n = _block.at(i);
2997       if (!n->is_Phi()) {
2998         int d_orig = depth(n);
2999         int d_in   = 0;
3000         for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
3001           Node* pred = preds.current();
3002           if (in_bb(pred)) {
3003             d_in = MAX2(d_in, depth(pred));
3004           }
3005         }
3006         if (d_in + 1 != d_orig) {
3007           set_depth(n, d_in + 1);
3008           again = true;
3009         }
3010       }
3011     }
3012     ct++;
3013   } while (again);
3014 
3015   if (TraceSuperWord && Verbose) {
3016     tty->print_cr("compute_max_depth iterated: %d times", ct);
3017   }
3018 }
3019 
3020 //-------------------------compute_vector_element_type-----------------------
3021 // Compute necessary vector element type for expressions
3022 // This propagates backwards a narrower integer type when the
3023 // upper bits of the value are not needed.
3024 // Example:  char a,b,c;  a = b + c;
3025 // Normally the type of the add is integer, but for packed character
3026 // operations the type of the add needs to be char.
3027 void SuperWord::compute_vector_element_type() {
3028   if (TraceSuperWord && Verbose) {
3029     tty->print_cr("\ncompute_velt_type:");
3030   }
3031 
3032   // Initial type
3033   for (int i = 0; i < _block.length(); i++) {
3034     Node* n = _block.at(i);
3035     set_velt_type(n, container_type(n));
3036   }
3037 
3038   // Propagate integer narrowed type backwards through operations
3039   // that don't depend on higher order bits
3040   for (int i = _block.length() - 1; i >= 0; i--) {
3041     Node* n = _block.at(i);
3042     // Only integer types need be examined
3043     const Type* vtn = velt_type(n);
3044     if (vtn->basic_type() == T_INT) {
3045       uint start, end;
3046       VectorNode::vector_operands(n, &start, &end);
3047 
3048       for (uint j = start; j < end; j++) {
3049         Node* in  = n->in(j);
3050         // Don't propagate through a memory
3051         if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
3052             data_size(n) < data_size(in)) {
3053           bool same_type = true;
3054           for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
3055             Node *use = in->fast_out(k);
3056             if (!in_bb(use) || !same_velt_type(use, n)) {
3057               same_type = false;
3058               break;
3059             }
3060           }
3061           if (same_type) {
3062             // For right shifts of small integer types (bool, byte, char, short)
3063             // we need precise information about sign-ness. Only Load nodes have
3064             // this information because Store nodes are the same for signed and
3065             // unsigned values. And any arithmetic operation after a load may
3066             // expand a value to signed Int so such right shifts can't be used
3067             // because vector elements do not have upper bits of Int.
3068             const Type* vt = vtn;
3069             if (VectorNode::is_shift(in)) {
3070               Node* load = in->in(1);
3071               if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
3072                 vt = velt_type(load);
3073               } else if (in->Opcode() != Op_LShiftI) {
3074                 // Widen type to Int to avoid creation of right shift vector
3075                 // (align + data_size(s1) check in stmts_can_pack() will fail).
3076                 // Note, left shifts work regardless type.
3077                 vt = TypeInt::INT;
3078               }
3079             }
3080             set_velt_type(in, vt);
3081           }
3082         }
3083       }
3084     }
3085   }
3086 #ifndef PRODUCT
3087   if (TraceSuperWord && Verbose) {
3088     for (int i = 0; i < _block.length(); i++) {
3089       Node* n = _block.at(i);
3090       velt_type(n)->dump();
3091       tty->print("\t");
3092       n->dump();
3093     }
3094   }
3095 #endif
3096 }
3097 
3098 //------------------------------memory_alignment---------------------------
3099 // Alignment within a vector memory reference
3100 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
3101   #ifndef PRODUCT
3102     if(TraceSuperWord && Verbose) {
3103       tty->print("SuperWord::memory_alignment within a vector memory reference for %d:  ", s->_idx); s->dump();
3104     }
3105   #endif
3106   NOT_PRODUCT(SWPointer::Tracer::Depth ddd(0);)
3107   SWPointer p(s, this, NULL, false);
3108   if (!p.valid()) {
3109     NOT_PRODUCT(if(is_trace_alignment()) tty->print("SWPointer::memory_alignment: SWPointer p invalid, return bottom_align");)
3110     return bottom_align;
3111   }
3112   int vw = get_vw_bytes_special(s);
3113   if (vw < 2) {
3114     NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SWPointer::memory_alignment: vector_width_in_bytes < 2, return bottom_align");)
3115     return bottom_align; // No vectors for this type
3116   }
3117   int offset  = p.offset_in_bytes();
3118   offset     += iv_adjust*p.memory_size();
3119   int off_rem = offset % vw;
3120   int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
3121   if (TraceSuperWord && Verbose) {
3122     tty->print_cr("SWPointer::memory_alignment: off_rem = %d, off_mod = %d", off_rem, off_mod);
3123   }
3124   return off_mod;
3125 }
3126 
3127 //---------------------------container_type---------------------------
3128 // Smallest type containing range of values
3129 const Type* SuperWord::container_type(Node* n) {
3130   if (n->is_Mem()) {
3131     BasicType bt = n->as_Mem()->memory_type();
3132     if (n->is_Store() && (bt == T_CHAR)) {
3133       // Use T_SHORT type instead of T_CHAR for stored values because any
3134       // preceding arithmetic operation extends values to signed Int.
3135       bt = T_SHORT;
3136     }
3137     if (n->Opcode() == Op_LoadUB) {
3138       // Adjust type for unsigned byte loads, it is important for right shifts.
3139       // T_BOOLEAN is used because there is no basic type representing type
3140       // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
3141       // size (one byte) and sign is important.
3142       bt = T_BOOLEAN;
3143     }
3144     return Type::get_const_basic_type(bt);
3145   }
3146   const Type* t = _igvn.type(n);
3147   if (t->basic_type() == T_INT) {
3148     // A narrow type of arithmetic operations will be determined by
3149     // propagating the type of memory operations.
3150     return TypeInt::INT;
3151   }
3152   return t;
3153 }
3154 
3155 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
3156   const Type* vt1 = velt_type(n1);
3157   const Type* vt2 = velt_type(n2);
3158   if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
3159     // Compare vectors element sizes for integer types.
3160     return data_size(n1) == data_size(n2);
3161   }
3162   return vt1 == vt2;
3163 }
3164 
3165 //------------------------------in_packset---------------------------
3166 // Are s1 and s2 in a pack pair and ordered as s1,s2?
3167 bool SuperWord::in_packset(Node* s1, Node* s2) {
3168   for (int i = 0; i < _packset.length(); i++) {
3169     Node_List* p = _packset.at(i);
3170     assert(p->size() == 2, "must be");
3171     if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
3172       return true;
3173     }
3174   }
3175   return false;
3176 }
3177 
3178 //------------------------------in_pack---------------------------
3179 // Is s in pack p?
3180 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
3181   for (uint i = 0; i < p->size(); i++) {
3182     if (p->at(i) == s) {
3183       return p;
3184     }
3185   }
3186   return NULL;
3187 }
3188 
3189 //------------------------------remove_pack_at---------------------------
3190 // Remove the pack at position pos in the packset
3191 void SuperWord::remove_pack_at(int pos) {
3192   Node_List* p = _packset.at(pos);
3193   for (uint i = 0; i < p->size(); i++) {
3194     Node* s = p->at(i);
3195     set_my_pack(s, NULL);
3196   }
3197   _packset.remove_at(pos);
3198 }
3199 
3200 void SuperWord::packset_sort(int n) {
3201   // simple bubble sort so that we capitalize with O(n) when its already sorted
3202   while (n != 0) {
3203     bool swapped = false;
3204     for (int i = 1; i < n; i++) {
3205       Node_List* q_low = _packset.at(i-1);
3206       Node_List* q_i = _packset.at(i);
3207 
3208       // only swap when we find something to swap
3209       if (alignment(q_low->at(0)) > alignment(q_i->at(0))) {
3210         Node_List* t = q_i;
3211         *(_packset.adr_at(i)) = q_low;
3212         *(_packset.adr_at(i-1)) = q_i;
3213         swapped = true;
3214       }
3215     }
3216     if (swapped == false) break;
3217     n--;
3218   }
3219 }
3220 
3221 //------------------------------executed_first---------------------------
3222 // Return the node executed first in pack p.  Uses the RPO block list
3223 // to determine order.
3224 Node* SuperWord::executed_first(Node_List* p) {
3225   Node* n = p->at(0);
3226   int n_rpo = bb_idx(n);
3227   for (uint i = 1; i < p->size(); i++) {
3228     Node* s = p->at(i);
3229     int s_rpo = bb_idx(s);
3230     if (s_rpo < n_rpo) {
3231       n = s;
3232       n_rpo = s_rpo;
3233     }
3234   }
3235   return n;
3236 }
3237 
3238 //------------------------------executed_last---------------------------
3239 // Return the node executed last in pack p.
3240 Node* SuperWord::executed_last(Node_List* p) {
3241   Node* n = p->at(0);
3242   int n_rpo = bb_idx(n);
3243   for (uint i = 1; i < p->size(); i++) {
3244     Node* s = p->at(i);
3245     int s_rpo = bb_idx(s);
3246     if (s_rpo > n_rpo) {
3247       n = s;
3248       n_rpo = s_rpo;
3249     }
3250   }
3251   return n;
3252 }
3253 
3254 LoadNode::ControlDependency SuperWord::control_dependency(Node_List* p) {
3255   LoadNode::ControlDependency dep = LoadNode::DependsOnlyOnTest;
3256   for (uint i = 0; i < p->size(); i++) {
3257     Node* n = p->at(i);
3258     assert(n->is_Load(), "only meaningful for loads");
3259     if (!n->depends_only_on_test()) {
3260       dep = LoadNode::Pinned;
3261     }
3262   }
3263   return dep;
3264 }
3265 
3266 
3267 //----------------------------align_initial_loop_index---------------------------
3268 // Adjust pre-loop limit so that in main loop, a load/store reference
3269 // to align_to_ref will be a position zero in the vector.
3270 //   (iv + k) mod vector_align == 0
3271 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
3272   CountedLoopNode *main_head = lp()->as_CountedLoop();
3273   assert(main_head->is_main_loop(), "");
3274   CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
3275   assert(pre_end != NULL, "we must have a correct pre-loop");
3276   Node *pre_opaq1 = pre_end->limit();
3277   assert(pre_opaq1->Opcode() == Op_Opaque1, "");
3278   Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
3279   Node *lim0 = pre_opaq->in(1);
3280 
3281   // Where we put new limit calculations
3282   Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
3283 
3284   // Ensure the original loop limit is available from the
3285   // pre-loop Opaque1 node.
3286   Node *orig_limit = pre_opaq->original_loop_limit();
3287   assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
3288 
3289   SWPointer align_to_ref_p(align_to_ref, this, NULL, false);
3290   assert(align_to_ref_p.valid(), "sanity");
3291 
3292   // Given:
3293   //     lim0 == original pre loop limit
3294   //     V == v_align (power of 2)
3295   //     invar == extra invariant piece of the address expression
3296   //     e == offset [ +/- invar ]
3297   //
3298   // When reassociating expressions involving '%' the basic rules are:
3299   //     (a - b) % k == 0   =>  a % k == b % k
3300   // and:
3301   //     (a + b) % k == 0   =>  a % k == (k - b) % k
3302   //
3303   // For stride > 0 && scale > 0,
3304   //   Derive the new pre-loop limit "lim" such that the two constraints:
3305   //     (1) lim = lim0 + N           (where N is some positive integer < V)
3306   //     (2) (e + lim) % V == 0
3307   //   are true.
3308   //
3309   //   Substituting (1) into (2),
3310   //     (e + lim0 + N) % V == 0
3311   //   solve for N:
3312   //     N = (V - (e + lim0)) % V
3313   //   substitute back into (1), so that new limit
3314   //     lim = lim0 + (V - (e + lim0)) % V
3315   //
3316   // For stride > 0 && scale < 0
3317   //   Constraints:
3318   //     lim = lim0 + N
3319   //     (e - lim) % V == 0
3320   //   Solving for lim:
3321   //     (e - lim0 - N) % V == 0
3322   //     N = (e - lim0) % V
3323   //     lim = lim0 + (e - lim0) % V
3324   //
3325   // For stride < 0 && scale > 0
3326   //   Constraints:
3327   //     lim = lim0 - N
3328   //     (e + lim) % V == 0
3329   //   Solving for lim:
3330   //     (e + lim0 - N) % V == 0
3331   //     N = (e + lim0) % V
3332   //     lim = lim0 - (e + lim0) % V
3333   //
3334   // For stride < 0 && scale < 0
3335   //   Constraints:
3336   //     lim = lim0 - N
3337   //     (e - lim) % V == 0
3338   //   Solving for lim:
3339   //     (e - lim0 + N) % V == 0
3340   //     N = (V - (e - lim0)) % V
3341   //     lim = lim0 - (V - (e - lim0)) % V
3342 
3343   int vw = vector_width_in_bytes(align_to_ref);
3344   int stride   = iv_stride();
3345   int scale    = align_to_ref_p.scale_in_bytes();
3346   int elt_size = align_to_ref_p.memory_size();
3347   int v_align  = vw / elt_size;
3348   assert(v_align > 1, "sanity");
3349   int offset   = align_to_ref_p.offset_in_bytes() / elt_size;
3350   Node *offsn  = _igvn.intcon(offset);
3351 
3352   Node *e = offsn;
3353   if (align_to_ref_p.invar() != NULL) {
3354     // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
3355     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3356     Node* invar = align_to_ref_p.invar();
3357     if (_igvn.type(invar)->isa_long()) {
3358       // Computations are done % (vector width/element size) so it's
3359       // safe to simply convert invar to an int and loose the upper 32
3360       // bit half.
3361       invar = new ConvL2INode(invar);
3362       _igvn.register_new_node_with_optimizer(invar);
3363     }
3364     Node* aref = new URShiftINode(invar, log2_elt);
3365     _igvn.register_new_node_with_optimizer(aref);
3366     _phase->set_ctrl(aref, pre_ctrl);
3367     if (align_to_ref_p.negate_invar()) {
3368       e = new SubINode(e, aref);
3369     } else {
3370       e = new AddINode(e, aref);
3371     }
3372     _igvn.register_new_node_with_optimizer(e);
3373     _phase->set_ctrl(e, pre_ctrl);
3374   }
3375   if (vw > ObjectAlignmentInBytes) {
3376     // incorporate base e +/- base && Mask >>> log2(elt)
3377     Node* xbase = new CastP2XNode(NULL, align_to_ref_p.base());
3378     _igvn.register_new_node_with_optimizer(xbase);
3379 #ifdef _LP64
3380     xbase  = new ConvL2INode(xbase);
3381     _igvn.register_new_node_with_optimizer(xbase);
3382 #endif
3383     Node* mask = _igvn.intcon(vw-1);
3384     Node* masked_xbase  = new AndINode(xbase, mask);
3385     _igvn.register_new_node_with_optimizer(masked_xbase);
3386     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3387     Node* bref     = new URShiftINode(masked_xbase, log2_elt);
3388     _igvn.register_new_node_with_optimizer(bref);
3389     _phase->set_ctrl(bref, pre_ctrl);
3390     e = new AddINode(e, bref);
3391     _igvn.register_new_node_with_optimizer(e);
3392     _phase->set_ctrl(e, pre_ctrl);
3393   }
3394 
3395   // compute e +/- lim0
3396   if (scale < 0) {
3397     e = new SubINode(e, lim0);
3398   } else {
3399     e = new AddINode(e, lim0);
3400   }
3401   _igvn.register_new_node_with_optimizer(e);
3402   _phase->set_ctrl(e, pre_ctrl);
3403 
3404   if (stride * scale > 0) {
3405     // compute V - (e +/- lim0)
3406     Node* va  = _igvn.intcon(v_align);
3407     e = new SubINode(va, e);
3408     _igvn.register_new_node_with_optimizer(e);
3409     _phase->set_ctrl(e, pre_ctrl);
3410   }
3411   // compute N = (exp) % V
3412   Node* va_msk = _igvn.intcon(v_align - 1);
3413   Node* N = new AndINode(e, va_msk);
3414   _igvn.register_new_node_with_optimizer(N);
3415   _phase->set_ctrl(N, pre_ctrl);
3416 
3417   //   substitute back into (1), so that new limit
3418   //     lim = lim0 + N
3419   Node* lim;
3420   if (stride < 0) {
3421     lim = new SubINode(lim0, N);
3422   } else {
3423     lim = new AddINode(lim0, N);
3424   }
3425   _igvn.register_new_node_with_optimizer(lim);
3426   _phase->set_ctrl(lim, pre_ctrl);
3427   Node* constrained =
3428     (stride > 0) ? (Node*) new MinINode(lim, orig_limit)
3429                  : (Node*) new MaxINode(lim, orig_limit);
3430   _igvn.register_new_node_with_optimizer(constrained);
3431   _phase->set_ctrl(constrained, pre_ctrl);
3432   _igvn.replace_input_of(pre_opaq, 1, constrained);
3433 }
3434 
3435 //----------------------------get_pre_loop_end---------------------------
3436 // Find pre loop end from main loop.  Returns null if none.
3437 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode* cl) {
3438   // The loop cannot be optimized if the graph shape at
3439   // the loop entry is inappropriate.
3440   if (!PhaseIdealLoop::is_canonical_loop_entry(cl)) {
3441     return NULL;
3442   }
3443 
3444   Node* p_f = cl->skip_predicates()->in(0)->in(0);
3445   if (!p_f->is_IfFalse()) return NULL;
3446   if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
3447   CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
3448   CountedLoopNode* loop_node = pre_end->loopnode();
3449   if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
3450   return pre_end;
3451 }
3452 
3453 //------------------------------init---------------------------
3454 void SuperWord::init() {
3455   _dg.init();
3456   _packset.clear();
3457   _disjoint_ptrs.clear();
3458   _block.clear();
3459   _post_block.clear();
3460   _data_entry.clear();
3461   _mem_slice_head.clear();
3462   _mem_slice_tail.clear();
3463   _iteration_first.clear();
3464   _iteration_last.clear();
3465   _node_info.clear();
3466   _align_to_ref = NULL;
3467   _lpt = NULL;
3468   _lp = NULL;
3469   _bb = NULL;
3470   _iv = NULL;
3471   _race_possible = 0;
3472   _early_return = false;
3473   _num_work_vecs = 0;
3474   _num_reductions = 0;
3475 }
3476 
3477 //------------------------------restart---------------------------
3478 void SuperWord::restart() {
3479   _dg.init();
3480   _packset.clear();
3481   _disjoint_ptrs.clear();
3482   _block.clear();
3483   _post_block.clear();
3484   _data_entry.clear();
3485   _mem_slice_head.clear();
3486   _mem_slice_tail.clear();
3487   _node_info.clear();
3488 }
3489 
3490 //------------------------------print_packset---------------------------
3491 void SuperWord::print_packset() {
3492 #ifndef PRODUCT
3493   tty->print_cr("packset");
3494   for (int i = 0; i < _packset.length(); i++) {
3495     tty->print_cr("Pack: %d", i);
3496     Node_List* p = _packset.at(i);
3497     print_pack(p);
3498   }
3499 #endif
3500 }
3501 
3502 //------------------------------print_pack---------------------------
3503 void SuperWord::print_pack(Node_List* p) {
3504   for (uint i = 0; i < p->size(); i++) {
3505     print_stmt(p->at(i));
3506   }
3507 }
3508 
3509 //------------------------------print_bb---------------------------
3510 void SuperWord::print_bb() {
3511 #ifndef PRODUCT
3512   tty->print_cr("\nBlock");
3513   for (int i = 0; i < _block.length(); i++) {
3514     Node* n = _block.at(i);
3515     tty->print("%d ", i);
3516     if (n) {
3517       n->dump();
3518     }
3519   }
3520 #endif
3521 }
3522 
3523 //------------------------------print_stmt---------------------------
3524 void SuperWord::print_stmt(Node* s) {
3525 #ifndef PRODUCT
3526   tty->print(" align: %d \t", alignment(s));
3527   s->dump();
3528 #endif
3529 }
3530 
3531 //------------------------------blank---------------------------
3532 char* SuperWord::blank(uint depth) {
3533   static char blanks[101];
3534   assert(depth < 101, "too deep");
3535   for (uint i = 0; i < depth; i++) blanks[i] = ' ';
3536   blanks[depth] = '\0';
3537   return blanks;
3538 }
3539 
3540 
3541 //==============================SWPointer===========================
3542 #ifndef PRODUCT
3543 int SWPointer::Tracer::_depth = 0;
3544 #endif
3545 //----------------------------SWPointer------------------------
3546 SWPointer::SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only) :
3547   _mem(mem), _slp(slp),  _base(NULL),  _adr(NULL),
3548   _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3549   _nstack(nstack), _analyze_only(analyze_only),
3550   _stack_idx(0)
3551 #ifndef PRODUCT
3552   , _tracer(slp)
3553 #endif
3554 {
3555   NOT_PRODUCT(_tracer.ctor_1(mem);)
3556 
3557   Node* adr = mem->in(MemNode::Address);
3558   if (!adr->is_AddP()) {
3559     assert(!valid(), "too complex");
3560     return;
3561   }
3562   // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
3563   Node* base = adr->in(AddPNode::Base);
3564   // The base address should be loop invariant
3565   if (!invariant(base)) {
3566     assert(!valid(), "base address is loop variant");
3567     return;
3568   }
3569   //unsafe reference could not be aligned appropriately without runtime checking
3570   if (base == NULL || base->bottom_type() == Type::TOP) {
3571     assert(!valid(), "unsafe access");
3572     return;
3573   }
3574 
3575   NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.store_depth();)
3576   NOT_PRODUCT(_tracer.ctor_2(adr);)
3577 
3578   int i;
3579   for (i = 0; i < 3; i++) {
3580     NOT_PRODUCT(_tracer.ctor_3(adr, i);)
3581 
3582     if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
3583       assert(!valid(), "too complex");
3584       return;
3585     }
3586     adr = adr->in(AddPNode::Address);
3587     NOT_PRODUCT(_tracer.ctor_4(adr, i);)
3588 
3589     if (base == adr || !adr->is_AddP()) {
3590       NOT_PRODUCT(_tracer.ctor_5(adr, base, i);)
3591       break; // stop looking at addp's
3592     }
3593   }
3594   NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.restore_depth();)
3595   NOT_PRODUCT(_tracer.ctor_6(mem);)
3596 
3597   _base = base;
3598   _adr  = adr;
3599   assert(valid(), "Usable");
3600 }
3601 
3602 // Following is used to create a temporary object during
3603 // the pattern match of an address expression.
3604 SWPointer::SWPointer(SWPointer* p) :
3605   _mem(p->_mem), _slp(p->_slp),  _base(NULL),  _adr(NULL),
3606   _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3607   _nstack(p->_nstack), _analyze_only(p->_analyze_only),
3608   _stack_idx(p->_stack_idx)
3609   #ifndef PRODUCT
3610   , _tracer(p->_slp)
3611   #endif
3612 {}
3613 
3614 
3615 bool SWPointer::invariant(Node* n) {
3616   NOT_PRODUCT(Tracer::Depth dd;)
3617   Node *n_c = phase()->get_ctrl(n);
3618   NOT_PRODUCT(_tracer.invariant_1(n, n_c);)
3619   return !lpt()->is_member(phase()->get_loop(n_c));
3620 }
3621 //------------------------scaled_iv_plus_offset--------------------
3622 // Match: k*iv + offset
3623 // where: k is a constant that maybe zero, and
3624 //        offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
3625 bool SWPointer::scaled_iv_plus_offset(Node* n) {
3626   NOT_PRODUCT(Tracer::Depth ddd;)
3627   NOT_PRODUCT(_tracer.scaled_iv_plus_offset_1(n);)
3628 
3629   if (scaled_iv(n)) {
3630     NOT_PRODUCT(_tracer.scaled_iv_plus_offset_2(n);)
3631     return true;
3632   }
3633 
3634   if (offset_plus_k(n)) {
3635     NOT_PRODUCT(_tracer.scaled_iv_plus_offset_3(n);)
3636     return true;
3637   }
3638 
3639   int opc = n->Opcode();
3640   if (opc == Op_AddI) {
3641     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
3642       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_4(n);)
3643       return true;
3644     }
3645     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3646       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_5(n);)
3647       return true;
3648     }
3649   } else if (opc == Op_SubI) {
3650     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
3651       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_6(n);)
3652       return true;
3653     }
3654     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3655       _scale *= -1;
3656       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_7(n);)
3657       return true;
3658     }
3659   }
3660 
3661   NOT_PRODUCT(_tracer.scaled_iv_plus_offset_8(n);)
3662   return false;
3663 }
3664 
3665 //----------------------------scaled_iv------------------------
3666 // Match: k*iv where k is a constant that's not zero
3667 bool SWPointer::scaled_iv(Node* n) {
3668   NOT_PRODUCT(Tracer::Depth ddd;)
3669   NOT_PRODUCT(_tracer.scaled_iv_1(n);)
3670 
3671   if (_scale != 0) { // already found a scale
3672     NOT_PRODUCT(_tracer.scaled_iv_2(n, _scale);)
3673     return false;
3674   }
3675 
3676   if (n == iv()) {
3677     _scale = 1;
3678     NOT_PRODUCT(_tracer.scaled_iv_3(n, _scale);)
3679     return true;
3680   }
3681   if (_analyze_only && (invariant(n) == false)) {
3682     _nstack->push(n, _stack_idx++);
3683   }
3684 
3685   int opc = n->Opcode();
3686   if (opc == Op_MulI) {
3687     if (n->in(1) == iv() && n->in(2)->is_Con()) {
3688       _scale = n->in(2)->get_int();
3689       NOT_PRODUCT(_tracer.scaled_iv_4(n, _scale);)
3690       return true;
3691     } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
3692       _scale = n->in(1)->get_int();
3693       NOT_PRODUCT(_tracer.scaled_iv_5(n, _scale);)
3694       return true;
3695     }
3696   } else if (opc == Op_LShiftI) {
3697     if (n->in(1) == iv() && n->in(2)->is_Con()) {
3698       _scale = 1 << n->in(2)->get_int();
3699       NOT_PRODUCT(_tracer.scaled_iv_6(n, _scale);)
3700       return true;
3701     }
3702   } else if (opc == Op_ConvI2L) {
3703     if (n->in(1)->Opcode() == Op_CastII &&
3704         n->in(1)->as_CastII()->has_range_check()) {
3705       // Skip range check dependent CastII nodes
3706       n = n->in(1);
3707     }
3708     if (scaled_iv_plus_offset(n->in(1))) {
3709       NOT_PRODUCT(_tracer.scaled_iv_7(n);)
3710       return true;
3711     }
3712   } else if (opc == Op_LShiftL) {
3713     if (!has_iv() && _invar == NULL) {
3714       // Need to preserve the current _offset value, so
3715       // create a temporary object for this expression subtree.
3716       // Hacky, so should re-engineer the address pattern match.
3717       NOT_PRODUCT(Tracer::Depth dddd;)
3718       SWPointer tmp(this);
3719       NOT_PRODUCT(_tracer.scaled_iv_8(n, &tmp);)
3720 
3721       if (tmp.scaled_iv_plus_offset(n->in(1))) {
3722         if (tmp._invar == NULL || _slp->do_vector_loop()) {
3723           int mult = 1 << n->in(2)->get_int();
3724           _scale   = tmp._scale  * mult;
3725           _offset += tmp._offset * mult;
3726           NOT_PRODUCT(_tracer.scaled_iv_9(n, _scale, _offset, mult);)
3727           return true;
3728         }
3729       }
3730     }
3731   }
3732   NOT_PRODUCT(_tracer.scaled_iv_10(n);)
3733   return false;
3734 }
3735 
3736 //----------------------------offset_plus_k------------------------
3737 // Match: offset is (k [+/- invariant])
3738 // where k maybe zero and invariant is optional, but not both.
3739 bool SWPointer::offset_plus_k(Node* n, bool negate) {
3740   NOT_PRODUCT(Tracer::Depth ddd;)
3741   NOT_PRODUCT(_tracer.offset_plus_k_1(n);)
3742 
3743   int opc = n->Opcode();
3744   if (opc == Op_ConI) {
3745     _offset += negate ? -(n->get_int()) : n->get_int();
3746     NOT_PRODUCT(_tracer.offset_plus_k_2(n, _offset);)
3747     return true;
3748   } else if (opc == Op_ConL) {
3749     // Okay if value fits into an int
3750     const TypeLong* t = n->find_long_type();
3751     if (t->higher_equal(TypeLong::INT)) {
3752       jlong loff = n->get_long();
3753       jint  off  = (jint)loff;
3754       _offset += negate ? -off : loff;
3755       NOT_PRODUCT(_tracer.offset_plus_k_3(n, _offset);)
3756       return true;
3757     }
3758     NOT_PRODUCT(_tracer.offset_plus_k_4(n);)
3759     return false;
3760   }
3761   if (_invar != NULL) { // already has an invariant
3762     NOT_PRODUCT(_tracer.offset_plus_k_5(n, _invar);)
3763     return false;
3764   }
3765 
3766   if (_analyze_only && (invariant(n) == false)) {
3767     _nstack->push(n, _stack_idx++);
3768   }
3769   if (opc == Op_AddI) {
3770     if (n->in(2)->is_Con() && invariant(n->in(1))) {
3771       _negate_invar = negate;
3772       _invar = n->in(1);
3773       _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3774       NOT_PRODUCT(_tracer.offset_plus_k_6(n, _invar, _negate_invar, _offset);)
3775       return true;
3776     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3777       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3778       _negate_invar = negate;
3779       _invar = n->in(2);
3780       NOT_PRODUCT(_tracer.offset_plus_k_7(n, _invar, _negate_invar, _offset);)
3781       return true;
3782     }
3783   }
3784   if (opc == Op_SubI) {
3785     if (n->in(2)->is_Con() && invariant(n->in(1))) {
3786       _negate_invar = negate;
3787       _invar = n->in(1);
3788       _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3789       NOT_PRODUCT(_tracer.offset_plus_k_8(n, _invar, _negate_invar, _offset);)
3790       return true;
3791     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3792       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3793       _negate_invar = !negate;
3794       _invar = n->in(2);
3795       NOT_PRODUCT(_tracer.offset_plus_k_9(n, _invar, _negate_invar, _offset);)
3796       return true;
3797     }
3798   }
3799   if (invariant(n)) {
3800     if (opc == Op_ConvI2L) {
3801       n = n->in(1);
3802       if (n->Opcode() == Op_CastII &&
3803           n->as_CastII()->has_range_check()) {
3804         // Skip range check dependent CastII nodes
3805         assert(invariant(n), "sanity");
3806         n = n->in(1);
3807       }
3808     }
3809     _negate_invar = negate;
3810     _invar = n;
3811     NOT_PRODUCT(_tracer.offset_plus_k_10(n, _invar, _negate_invar, _offset);)
3812     return true;
3813   }
3814 
3815   NOT_PRODUCT(_tracer.offset_plus_k_11(n);)
3816   return false;
3817 }
3818 
3819 //----------------------------print------------------------
3820 void SWPointer::print() {
3821 #ifndef PRODUCT
3822   tty->print("base: %d  adr: %d  scale: %d  offset: %d  invar: %c%d\n",
3823              _base != NULL ? _base->_idx : 0,
3824              _adr  != NULL ? _adr->_idx  : 0,
3825              _scale, _offset,
3826              _negate_invar?'-':'+',
3827              _invar != NULL ? _invar->_idx : 0);
3828 #endif
3829 }
3830 
3831 //----------------------------tracing------------------------
3832 #ifndef PRODUCT
3833 void SWPointer::Tracer::print_depth() {
3834   for (int ii = 0; ii<_depth; ++ii) tty->print("  ");
3835 }
3836 
3837 void SWPointer::Tracer::ctor_1 (Node* mem) {
3838   if(_slp->is_trace_alignment()) {
3839     print_depth(); tty->print(" %d SWPointer::SWPointer: start alignment analysis", mem->_idx); mem->dump();
3840   }
3841 }
3842 
3843 void SWPointer::Tracer::ctor_2(Node* adr) {
3844   if(_slp->is_trace_alignment()) {
3845     //store_depth();
3846     inc_depth();
3847     print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: ", adr->_idx); adr->dump();
3848     inc_depth();
3849     print_depth(); tty->print(" %d (base) SWPointer::SWPointer: ", adr->in(AddPNode::Base)->_idx); adr->in(AddPNode::Base)->dump();
3850   }
3851 }
3852 
3853 void SWPointer::Tracer::ctor_3(Node* adr, int i) {
3854   if(_slp->is_trace_alignment()) {
3855     inc_depth();
3856     Node* offset = adr->in(AddPNode::Offset);
3857     print_depth(); tty->print(" %d (offset) SWPointer::SWPointer: i = %d: ", offset->_idx, i); offset->dump();
3858   }
3859 }
3860 
3861 void SWPointer::Tracer::ctor_4(Node* adr, int i) {
3862   if(_slp->is_trace_alignment()) {
3863     inc_depth();
3864     print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: i = %d: ", adr->_idx, i); adr->dump();
3865   }
3866 }
3867 
3868 void SWPointer::Tracer::ctor_5(Node* adr, Node* base, int i) {
3869   if(_slp->is_trace_alignment()) {
3870     inc_depth();
3871     if (base == adr) {
3872       print_depth(); tty->print_cr("  \\ %d (adr) == %d (base) SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, base->_idx, i);
3873     } else if (!adr->is_AddP()) {
3874       print_depth(); tty->print_cr("  \\ %d (adr) is NOT Addp SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, i);
3875     }
3876   }
3877 }
3878 
3879 void SWPointer::Tracer::ctor_6(Node* mem) {
3880   if(_slp->is_trace_alignment()) {
3881     //restore_depth();
3882     print_depth(); tty->print_cr(" %d (adr) SWPointer::SWPointer: stop analysis", mem->_idx);
3883   }
3884 }
3885 
3886 void SWPointer::Tracer::invariant_1(Node *n, Node *n_c) {
3887   if (_slp->do_vector_loop() && _slp->is_debug() && _slp->_lpt->is_member(_slp->_phase->get_loop(n_c)) != (int)_slp->in_bb(n)) {
3888     int is_member =  _slp->_lpt->is_member(_slp->_phase->get_loop(n_c));
3889     int in_bb     =  _slp->in_bb(n);
3890     print_depth(); tty->print("  \\ ");  tty->print_cr(" %d SWPointer::invariant  conditions differ: n_c %d", n->_idx, n_c->_idx);
3891     print_depth(); tty->print("  \\ ");  tty->print_cr("is_member %d, in_bb %d", is_member, in_bb);
3892     print_depth(); tty->print("  \\ ");  n->dump();
3893     print_depth(); tty->print("  \\ ");  n_c->dump();
3894   }
3895 }
3896 
3897 void SWPointer::Tracer::scaled_iv_plus_offset_1(Node* n) {
3898   if(_slp->is_trace_alignment()) {
3899     print_depth(); tty->print(" %d SWPointer::scaled_iv_plus_offset testing node: ", n->_idx);
3900     n->dump();
3901   }
3902 }
3903 
3904 void SWPointer::Tracer::scaled_iv_plus_offset_2(Node* n) {
3905   if(_slp->is_trace_alignment()) {
3906     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
3907   }
3908 }
3909 
3910 void SWPointer::Tracer::scaled_iv_plus_offset_3(Node* n) {
3911   if(_slp->is_trace_alignment()) {
3912     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
3913   }
3914 }
3915 
3916 void SWPointer::Tracer::scaled_iv_plus_offset_4(Node* n) {
3917   if(_slp->is_trace_alignment()) {
3918     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
3919     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
3920     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
3921   }
3922 }
3923 
3924 void SWPointer::Tracer::scaled_iv_plus_offset_5(Node* n) {
3925   if(_slp->is_trace_alignment()) {
3926     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
3927     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
3928     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
3929   }
3930 }
3931 
3932 void SWPointer::Tracer::scaled_iv_plus_offset_6(Node* n) {
3933   if(_slp->is_trace_alignment()) {
3934     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
3935     print_depth(); tty->print("  \\  %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
3936     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
3937   }
3938 }
3939 
3940 void SWPointer::Tracer::scaled_iv_plus_offset_7(Node* n) {
3941   if(_slp->is_trace_alignment()) {
3942     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
3943     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
3944     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
3945   }
3946 }
3947 
3948 void SWPointer::Tracer::scaled_iv_plus_offset_8(Node* n) {
3949   if(_slp->is_trace_alignment()) {
3950     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: FAILED", n->_idx);
3951   }
3952 }
3953 
3954 void SWPointer::Tracer::scaled_iv_1(Node* n) {
3955   if(_slp->is_trace_alignment()) {
3956     print_depth(); tty->print(" %d SWPointer::scaled_iv: testing node: ", n->_idx); n->dump();
3957   }
3958 }
3959 
3960 void SWPointer::Tracer::scaled_iv_2(Node* n, int scale) {
3961   if(_slp->is_trace_alignment()) {
3962     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED since another _scale has been detected before", n->_idx);
3963     print_depth(); tty->print_cr("  \\ SWPointer::scaled_iv: _scale (%d) != 0", scale);
3964   }
3965 }
3966 
3967 void SWPointer::Tracer::scaled_iv_3(Node* n, int scale) {
3968   if(_slp->is_trace_alignment()) {
3969     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: is iv, setting _scale = %d", n->_idx, scale);
3970   }
3971 }
3972 
3973 void SWPointer::Tracer::scaled_iv_4(Node* n, int scale) {
3974   if(_slp->is_trace_alignment()) {
3975     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
3976     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
3977     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3978   }
3979 }
3980 
3981 void SWPointer::Tracer::scaled_iv_5(Node* n, int scale) {
3982   if(_slp->is_trace_alignment()) {
3983     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
3984     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is iv: ", n->in(2)->_idx); n->in(2)->dump();
3985     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3986   }
3987 }
3988 
3989 void SWPointer::Tracer::scaled_iv_6(Node* n, int scale) {
3990   if(_slp->is_trace_alignment()) {
3991     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftI PASSED, setting _scale = %d", n->_idx, scale);
3992     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
3993     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3994   }
3995 }
3996 
3997 void SWPointer::Tracer::scaled_iv_7(Node* n) {
3998   if(_slp->is_trace_alignment()) {
3999     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_ConvI2L PASSED", n->_idx);
4000     print_depth(); tty->print_cr("  \\ SWPointer::scaled_iv: in(1) %d is scaled_iv_plus_offset: ", n->in(1)->_idx);
4001     inc_depth(); inc_depth();
4002     print_depth(); n->in(1)->dump();
4003     dec_depth(); dec_depth();
4004   }
4005 }
4006 
4007 void SWPointer::Tracer::scaled_iv_8(Node* n, SWPointer* tmp) {
4008   if(_slp->is_trace_alignment()) {
4009     print_depth(); tty->print(" %d SWPointer::scaled_iv: Op_LShiftL, creating tmp SWPointer: ", n->_idx); tmp->print();
4010   }
4011 }
4012 
4013 void SWPointer::Tracer::scaled_iv_9(Node* n, int scale, int _offset, int mult) {
4014   if(_slp->is_trace_alignment()) {
4015     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftL PASSED, setting _scale = %d, _offset = %d", n->_idx, scale, _offset);
4016     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",
4017     n->in(1)->_idx, n->in(2)->_idx, mult, scale, _offset);
4018     inc_depth(); inc_depth();
4019     print_depth(); n->in(1)->dump();
4020     print_depth(); n->in(2)->dump();
4021     dec_depth(); dec_depth();
4022   }
4023 }
4024 
4025 void SWPointer::Tracer::scaled_iv_10(Node* n) {
4026   if(_slp->is_trace_alignment()) {
4027     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED", n->_idx);
4028   }
4029 }
4030 
4031 void SWPointer::Tracer::offset_plus_k_1(Node* n) {
4032   if(_slp->is_trace_alignment()) {
4033     print_depth(); tty->print(" %d SWPointer::offset_plus_k: testing node: ", n->_idx); n->dump();
4034   }
4035 }
4036 
4037 void SWPointer::Tracer::offset_plus_k_2(Node* n, int _offset) {
4038   if(_slp->is_trace_alignment()) {
4039     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConI PASSED, setting _offset = %d", n->_idx, _offset);
4040   }
4041 }
4042 
4043 void SWPointer::Tracer::offset_plus_k_3(Node* n, int _offset) {
4044   if(_slp->is_trace_alignment()) {
4045     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConL PASSED, setting _offset = %d", n->_idx, _offset);
4046   }
4047 }
4048 
4049 void SWPointer::Tracer::offset_plus_k_4(Node* n) {
4050   if(_slp->is_trace_alignment()) {
4051     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
4052     print_depth(); tty->print_cr("  \\ " JLONG_FORMAT " SWPointer::offset_plus_k: Op_ConL FAILED, k is too big", n->get_long());
4053   }
4054 }
4055 
4056 void SWPointer::Tracer::offset_plus_k_5(Node* n, Node* _invar) {
4057   if(_slp->is_trace_alignment()) {
4058     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED since another invariant has been detected before", n->_idx);
4059     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: _invar != NULL: ", _invar->_idx); _invar->dump();
4060   }
4061 }
4062 
4063 void SWPointer::Tracer::offset_plus_k_6(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4064   if(_slp->is_trace_alignment()) {
4065     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
4066     n->_idx, _negate_invar, _invar->_idx, _offset);
4067     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4068     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
4069   }
4070 }
4071 
4072 void SWPointer::Tracer::offset_plus_k_7(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4073   if(_slp->is_trace_alignment()) {
4074     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
4075     n->_idx, _negate_invar, _invar->_idx, _offset);
4076     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
4077     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
4078   }
4079 }
4080 
4081 void SWPointer::Tracer::offset_plus_k_8(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4082   if(_slp->is_trace_alignment()) {
4083     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_SubI is PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
4084     n->_idx, _negate_invar, _invar->_idx, _offset);
4085     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4086     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
4087   }
4088 }
4089 
4090 void SWPointer::Tracer::offset_plus_k_9(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4091   if(_slp->is_trace_alignment()) {
4092     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);
4093     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
4094     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
4095   }
4096 }
4097 
4098 void SWPointer::Tracer::offset_plus_k_10(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4099   if(_slp->is_trace_alignment()) {
4100     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);
4101     print_depth(); tty->print_cr("  \\ %d SWPointer::offset_plus_k: is invariant", n->_idx);
4102   }
4103 }
4104 
4105 void SWPointer::Tracer::offset_plus_k_11(Node* n) {
4106   if(_slp->is_trace_alignment()) {
4107     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
4108   }
4109 }
4110 
4111 #endif
4112 // ========================= OrderedPair =====================
4113 
4114 const OrderedPair OrderedPair::initial;
4115 
4116 // ========================= SWNodeInfo =====================
4117 
4118 const SWNodeInfo SWNodeInfo::initial;
4119 
4120 
4121 // ============================ DepGraph ===========================
4122 
4123 //------------------------------make_node---------------------------
4124 // Make a new dependence graph node for an ideal node.
4125 DepMem* DepGraph::make_node(Node* node) {
4126   DepMem* m = new (_arena) DepMem(node);
4127   if (node != NULL) {
4128     assert(_map.at_grow(node->_idx) == NULL, "one init only");
4129     _map.at_put_grow(node->_idx, m);
4130   }
4131   return m;
4132 }
4133 
4134 //------------------------------make_edge---------------------------
4135 // Make a new dependence graph edge from dpred -> dsucc
4136 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
4137   DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
4138   dpred->set_out_head(e);
4139   dsucc->set_in_head(e);
4140   return e;
4141 }
4142 
4143 // ========================== DepMem ========================
4144 
4145 //------------------------------in_cnt---------------------------
4146 int DepMem::in_cnt() {
4147   int ct = 0;
4148   for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
4149   return ct;
4150 }
4151 
4152 //------------------------------out_cnt---------------------------
4153 int DepMem::out_cnt() {
4154   int ct = 0;
4155   for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
4156   return ct;
4157 }
4158 
4159 //------------------------------print-----------------------------
4160 void DepMem::print() {
4161 #ifndef PRODUCT
4162   tty->print("  DepNode %d (", _node->_idx);
4163   for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
4164     Node* pred = p->pred()->node();
4165     tty->print(" %d", pred != NULL ? pred->_idx : 0);
4166   }
4167   tty->print(") [");
4168   for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
4169     Node* succ = s->succ()->node();
4170     tty->print(" %d", succ != NULL ? succ->_idx : 0);
4171   }
4172   tty->print_cr(" ]");
4173 #endif
4174 }
4175 
4176 // =========================== DepEdge =========================
4177 
4178 //------------------------------DepPreds---------------------------
4179 void DepEdge::print() {
4180 #ifndef PRODUCT
4181   tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
4182 #endif
4183 }
4184 
4185 // =========================== DepPreds =========================
4186 // Iterator over predecessor edges in the dependence graph.
4187 
4188 //------------------------------DepPreds---------------------------
4189 DepPreds::DepPreds(Node* n, DepGraph& dg) {
4190   _n = n;
4191   _done = false;
4192   if (_n->is_Store() || _n->is_Load()) {
4193     _next_idx = MemNode::Address;
4194     _end_idx  = n->req();
4195     _dep_next = dg.dep(_n)->in_head();
4196   } else if (_n->is_Mem()) {
4197     _next_idx = 0;
4198     _end_idx  = 0;
4199     _dep_next = dg.dep(_n)->in_head();
4200   } else {
4201     _next_idx = 1;
4202     _end_idx  = _n->req();
4203     _dep_next = NULL;
4204   }
4205   next();
4206 }
4207 
4208 //------------------------------next---------------------------
4209 void DepPreds::next() {
4210   if (_dep_next != NULL) {
4211     _current  = _dep_next->pred()->node();
4212     _dep_next = _dep_next->next_in();
4213   } else if (_next_idx < _end_idx) {
4214     _current  = _n->in(_next_idx++);
4215   } else {
4216     _done = true;
4217   }
4218 }
4219 
4220 // =========================== DepSuccs =========================
4221 // Iterator over successor edges in the dependence graph.
4222 
4223 //------------------------------DepSuccs---------------------------
4224 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
4225   _n = n;
4226   _done = false;
4227   if (_n->is_Load()) {
4228     _next_idx = 0;
4229     _end_idx  = _n->outcnt();
4230     _dep_next = dg.dep(_n)->out_head();
4231   } else if (_n->is_Mem() || (_n->is_Phi() && _n->bottom_type() == Type::MEMORY)) {
4232     _next_idx = 0;
4233     _end_idx  = 0;
4234     _dep_next = dg.dep(_n)->out_head();
4235   } else {
4236     _next_idx = 0;
4237     _end_idx  = _n->outcnt();
4238     _dep_next = NULL;
4239   }
4240   next();
4241 }
4242 
4243 //-------------------------------next---------------------------
4244 void DepSuccs::next() {
4245   if (_dep_next != NULL) {
4246     _current  = _dep_next->succ()->node();
4247     _dep_next = _dep_next->next_out();
4248   } else if (_next_idx < _end_idx) {
4249     _current  = _n->raw_out(_next_idx++);
4250   } else {
4251     _done = true;
4252   }
4253 }
4254 
4255 //
4256 // --------------------------------- vectorization/simd -----------------------------------
4257 //
4258 bool SuperWord::same_origin_idx(Node* a, Node* b) const {
4259   return a != NULL && b != NULL && _clone_map.same_idx(a->_idx, b->_idx);
4260 }
4261 bool SuperWord::same_generation(Node* a, Node* b) const {
4262   return a != NULL && b != NULL && _clone_map.same_gen(a->_idx, b->_idx);
4263 }
4264 
4265 Node*  SuperWord::find_phi_for_mem_dep(LoadNode* ld) {
4266   assert(in_bb(ld), "must be in block");
4267   if (_clone_map.gen(ld->_idx) == _ii_first) {
4268 #ifndef PRODUCT
4269     if (_vector_loop_debug) {
4270       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(ld->_idx)=%d",
4271         _clone_map.gen(ld->_idx));
4272     }
4273 #endif
4274     return NULL; //we think that any ld in the first gen being vectorizable
4275   }
4276 
4277   Node* mem = ld->in(MemNode::Memory);
4278   if (mem->outcnt() <= 1) {
4279     // we don't want to remove the only edge from mem node to load
4280 #ifndef PRODUCT
4281     if (_vector_loop_debug) {
4282       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",
4283         mem->_idx, ld->_idx);
4284       ld->dump();
4285       mem->dump();
4286     }
4287 #endif
4288     return NULL;
4289   }
4290   if (!in_bb(mem) || same_generation(mem, ld)) {
4291 #ifndef PRODUCT
4292     if (_vector_loop_debug) {
4293       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(mem->_idx)=%d",
4294         _clone_map.gen(mem->_idx));
4295     }
4296 #endif
4297     return NULL; // does not depend on loop volatile node or depends on the same generation
4298   }
4299 
4300   //otherwise first node should depend on mem-phi
4301   Node* first = first_node(ld);
4302   assert(first->is_Load(), "must be Load");
4303   Node* phi = first->as_Load()->in(MemNode::Memory);
4304   if (!phi->is_Phi() || phi->bottom_type() != Type::MEMORY) {
4305 #ifndef PRODUCT
4306     if (_vector_loop_debug) {
4307       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");
4308       ld->dump();
4309       first->dump();
4310     }
4311 #endif
4312     return NULL;
4313   }
4314 
4315   Node* tail = 0;
4316   for (int m = 0; m < _mem_slice_head.length(); m++) {
4317     if (_mem_slice_head.at(m) == phi) {
4318       tail = _mem_slice_tail.at(m);
4319     }
4320   }
4321   if (tail == 0) { //test that found phi is in the list  _mem_slice_head
4322 #ifndef PRODUCT
4323     if (_vector_loop_debug) {
4324       tty->print_cr("SuperWord::find_phi_for_mem_dep load %d is not vectorizable node, its phi %d is not _mem_slice_head",
4325         ld->_idx, phi->_idx);
4326       ld->dump();
4327       phi->dump();
4328     }
4329 #endif
4330     return NULL;
4331   }
4332 
4333   // now all conditions are met
4334   return phi;
4335 }
4336 
4337 Node* SuperWord::first_node(Node* nd) {
4338   for (int ii = 0; ii < _iteration_first.length(); ii++) {
4339     Node* nnn = _iteration_first.at(ii);
4340     if (same_origin_idx(nnn, nd)) {
4341 #ifndef PRODUCT
4342       if (_vector_loop_debug) {
4343         tty->print_cr("SuperWord::first_node: %d is the first iteration node for %d (_clone_map.idx(nnn->_idx) = %d)",
4344           nnn->_idx, nd->_idx, _clone_map.idx(nnn->_idx));
4345       }
4346 #endif
4347       return nnn;
4348     }
4349   }
4350 
4351 #ifndef PRODUCT
4352   if (_vector_loop_debug) {
4353     tty->print_cr("SuperWord::first_node: did not find first iteration node for %d (_clone_map.idx(nd->_idx)=%d)",
4354       nd->_idx, _clone_map.idx(nd->_idx));
4355   }
4356 #endif
4357   return 0;
4358 }
4359 
4360 Node* SuperWord::last_node(Node* nd) {
4361   for (int ii = 0; ii < _iteration_last.length(); ii++) {
4362     Node* nnn = _iteration_last.at(ii);
4363     if (same_origin_idx(nnn, nd)) {
4364 #ifndef PRODUCT
4365       if (_vector_loop_debug) {
4366         tty->print_cr("SuperWord::last_node _clone_map.idx(nnn->_idx)=%d, _clone_map.idx(nd->_idx)=%d",
4367           _clone_map.idx(nnn->_idx), _clone_map.idx(nd->_idx));
4368       }
4369 #endif
4370       return nnn;
4371     }
4372   }
4373   return 0;
4374 }
4375 
4376 int SuperWord::mark_generations() {
4377   Node *ii_err = NULL, *tail_err = NULL;
4378   for (int i = 0; i < _mem_slice_head.length(); i++) {
4379     Node* phi  = _mem_slice_head.at(i);
4380     assert(phi->is_Phi(), "must be phi");
4381 
4382     Node* tail = _mem_slice_tail.at(i);
4383     if (_ii_last == -1) {
4384       tail_err = tail;
4385       _ii_last = _clone_map.gen(tail->_idx);
4386     }
4387     else if (_ii_last != _clone_map.gen(tail->_idx)) {
4388 #ifndef PRODUCT
4389       if (TraceSuperWord && Verbose) {
4390         tty->print_cr("SuperWord::mark_generations _ii_last error - found different generations in two tail nodes ");
4391         tail->dump();
4392         tail_err->dump();
4393       }
4394 #endif
4395       return -1;
4396     }
4397 
4398     // find first iteration in the loop
4399     for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
4400       Node* ii = phi->fast_out(i);
4401       if (in_bb(ii) && ii->is_Store()) { // we speculate that normally Stores of one and one only generation have deps from mem phi
4402         if (_ii_first == -1) {
4403           ii_err = ii;
4404           _ii_first = _clone_map.gen(ii->_idx);
4405         } else if (_ii_first != _clone_map.gen(ii->_idx)) {
4406 #ifndef PRODUCT
4407           if (TraceSuperWord && Verbose) {
4408             tty->print_cr("SuperWord::mark_generations: _ii_first was found before and not equal to one in this node (%d)", _ii_first);
4409             ii->dump();
4410             if (ii_err!= 0) {
4411               ii_err->dump();
4412             }
4413           }
4414 #endif
4415           return -1; // this phi has Stores from different generations of unroll and cannot be simd/vectorized
4416         }
4417       }
4418     }//for (DUIterator_Fast imax,
4419   }//for (int i...
4420 
4421   if (_ii_first == -1 || _ii_last == -1) {
4422     if (TraceSuperWord && Verbose) {
4423       tty->print_cr("SuperWord::mark_generations unknown error, something vent wrong");
4424     }
4425     return -1; // something vent wrong
4426   }
4427   // collect nodes in the first and last generations
4428   assert(_iteration_first.length() == 0, "_iteration_first must be empty");
4429   assert(_iteration_last.length() == 0, "_iteration_last must be empty");
4430   for (int j = 0; j < _block.length(); j++) {
4431     Node* n = _block.at(j);
4432     node_idx_t gen = _clone_map.gen(n->_idx);
4433     if ((signed)gen == _ii_first) {
4434       _iteration_first.push(n);
4435     } else if ((signed)gen == _ii_last) {
4436       _iteration_last.push(n);
4437     }
4438   }
4439 
4440   // building order of iterations
4441   if (_ii_order.length() == 0 && ii_err != 0) {
4442     assert(in_bb(ii_err) && ii_err->is_Store(), "should be Store in bb");
4443     Node* nd = ii_err;
4444     while(_clone_map.gen(nd->_idx) != _ii_last) {
4445       _ii_order.push(_clone_map.gen(nd->_idx));
4446       bool found = false;
4447       for (DUIterator_Fast imax, i = nd->fast_outs(imax); i < imax; i++) {
4448         Node* use = nd->fast_out(i);
4449         if (same_origin_idx(use, nd) && use->as_Store()->in(MemNode::Memory) == nd) {
4450           found = true;
4451           nd = use;
4452           break;
4453         }
4454       }//for
4455 
4456       if (found == false) {
4457         if (TraceSuperWord && Verbose) {
4458           tty->print_cr("SuperWord::mark_generations: Cannot build order of iterations - no dependent Store for %d", nd->_idx);
4459         }
4460         _ii_order.clear();
4461         return -1;
4462       }
4463     } //while
4464     _ii_order.push(_clone_map.gen(nd->_idx));
4465   }
4466 
4467 #ifndef PRODUCT
4468   if (_vector_loop_debug) {
4469     tty->print_cr("SuperWord::mark_generations");
4470     tty->print_cr("First generation (%d) nodes:", _ii_first);
4471     for (int ii = 0; ii < _iteration_first.length(); ii++)  _iteration_first.at(ii)->dump();
4472     tty->print_cr("Last generation (%d) nodes:", _ii_last);
4473     for (int ii = 0; ii < _iteration_last.length(); ii++)  _iteration_last.at(ii)->dump();
4474     tty->print_cr(" ");
4475 
4476     tty->print("SuperWord::List of generations: ");
4477     for (int jj = 0; jj < _ii_order.length(); ++jj) {
4478       tty->print("%d:%d ", jj, _ii_order.at(jj));
4479     }
4480     tty->print_cr(" ");
4481   }
4482 #endif
4483 
4484   return _ii_first;
4485 }
4486 
4487 bool SuperWord::fix_commutative_inputs(Node* gold, Node* fix) {
4488   assert(gold->is_Add() && fix->is_Add() || gold->is_Mul() && fix->is_Mul(), "should be only Add or Mul nodes");
4489   assert(same_origin_idx(gold, fix), "should be clones of the same node");
4490   Node* gin1 = gold->in(1);
4491   Node* gin2 = gold->in(2);
4492   Node* fin1 = fix->in(1);
4493   Node* fin2 = fix->in(2);
4494   bool swapped = false;
4495 
4496   if (in_bb(gin1) && in_bb(gin2) && in_bb(fin1) && in_bb(fin1)) {
4497     if (same_origin_idx(gin1, fin1) &&
4498         same_origin_idx(gin2, fin2)) {
4499       return true; // nothing to fix
4500     }
4501     if (same_origin_idx(gin1, fin2) &&
4502         same_origin_idx(gin2, fin1)) {
4503       fix->swap_edges(1, 2);
4504       swapped = true;
4505     }
4506   }
4507   // at least one input comes from outside of bb
4508   if (gin1->_idx == fin1->_idx)  {
4509     return true; // nothing to fix
4510   }
4511   if (!swapped && (gin1->_idx == fin2->_idx || gin2->_idx == fin1->_idx))  { //swapping is expensive, check condition first
4512     fix->swap_edges(1, 2);
4513     swapped = true;
4514   }
4515 
4516   if (swapped) {
4517 #ifndef PRODUCT
4518     if (_vector_loop_debug) {
4519       tty->print_cr("SuperWord::fix_commutative_inputs: fixed node %d", fix->_idx);
4520     }
4521 #endif
4522     return true;
4523   }
4524 
4525   if (TraceSuperWord && Verbose) {
4526     tty->print_cr("SuperWord::fix_commutative_inputs: cannot fix node %d", fix->_idx);
4527   }
4528 
4529   return false;
4530 }
4531 
4532 bool SuperWord::pack_parallel() {
4533 #ifndef PRODUCT
4534   if (_vector_loop_debug) {
4535     tty->print_cr("SuperWord::pack_parallel: START");
4536   }
4537 #endif
4538 
4539   _packset.clear();
4540 
4541   for (int ii = 0; ii < _iteration_first.length(); ii++) {
4542     Node* nd = _iteration_first.at(ii);
4543     if (in_bb(nd) && (nd->is_Load() || nd->is_Store() || nd->is_Add() || nd->is_Mul())) {
4544       Node_List* pk = new Node_List();
4545       pk->push(nd);
4546       for (int gen = 1; gen < _ii_order.length(); ++gen) {
4547         for (int kk = 0; kk < _block.length(); kk++) {
4548           Node* clone = _block.at(kk);
4549           if (same_origin_idx(clone, nd) &&
4550               _clone_map.gen(clone->_idx) == _ii_order.at(gen)) {
4551             if (nd->is_Add() || nd->is_Mul()) {
4552               fix_commutative_inputs(nd, clone);
4553             }
4554             pk->push(clone);
4555             if (pk->size() == 4) {
4556               _packset.append(pk);
4557 #ifndef PRODUCT
4558               if (_vector_loop_debug) {
4559                 tty->print_cr("SuperWord::pack_parallel: added pack ");
4560                 pk->dump();
4561               }
4562 #endif
4563               if (_clone_map.gen(clone->_idx) != _ii_last) {
4564                 pk = new Node_List();
4565               }
4566             }
4567             break;
4568           }
4569         }
4570       }//for
4571     }//if
4572   }//for
4573 
4574 #ifndef PRODUCT
4575   if (_vector_loop_debug) {
4576     tty->print_cr("SuperWord::pack_parallel: END");
4577   }
4578 #endif
4579 
4580   return true;
4581 }
4582 
4583 bool SuperWord::hoist_loads_in_graph() {
4584   GrowableArray<Node*> loads;
4585 
4586 #ifndef PRODUCT
4587   if (_vector_loop_debug) {
4588     tty->print_cr("SuperWord::hoist_loads_in_graph: total number _mem_slice_head.length() = %d", _mem_slice_head.length());
4589   }
4590 #endif
4591 
4592   for (int i = 0; i < _mem_slice_head.length(); i++) {
4593     Node* n = _mem_slice_head.at(i);
4594     if ( !in_bb(n) || !n->is_Phi() || n->bottom_type() != Type::MEMORY) {
4595       if (TraceSuperWord && Verbose) {
4596         tty->print_cr("SuperWord::hoist_loads_in_graph: skipping unexpected node n=%d", n->_idx);
4597       }
4598       continue;
4599     }
4600 
4601 #ifndef PRODUCT
4602     if (_vector_loop_debug) {
4603       tty->print_cr("SuperWord::hoist_loads_in_graph: processing phi %d  = _mem_slice_head.at(%d);", n->_idx, i);
4604     }
4605 #endif
4606 
4607     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4608       Node* ld = n->fast_out(i);
4609       if (ld->is_Load() && ld->as_Load()->in(MemNode::Memory) == n && in_bb(ld)) {
4610         for (int i = 0; i < _block.length(); i++) {
4611           Node* ld2 = _block.at(i);
4612           if (ld2->is_Load() && same_origin_idx(ld, ld2) &&
4613               !same_generation(ld, ld2)) { // <= do not collect the first generation ld
4614 #ifndef PRODUCT
4615             if (_vector_loop_debug) {
4616               tty->print_cr("SuperWord::hoist_loads_in_graph: will try to hoist load ld2->_idx=%d, cloned from %d (ld->_idx=%d)",
4617                 ld2->_idx, _clone_map.idx(ld->_idx), ld->_idx);
4618             }
4619 #endif
4620             // could not do on-the-fly, since iterator is immutable
4621             loads.push(ld2);
4622           }
4623         }// for
4624       }//if
4625     }//for (DUIterator_Fast imax,
4626   }//for (int i = 0; i
4627 
4628   for (int i = 0; i < loads.length(); i++) {
4629     LoadNode* ld = loads.at(i)->as_Load();
4630     Node* phi = find_phi_for_mem_dep(ld);
4631     if (phi != NULL) {
4632 #ifndef PRODUCT
4633       if (_vector_loop_debug) {
4634         tty->print_cr("SuperWord::hoist_loads_in_graph replacing MemNode::Memory(%d) edge in %d with one from %d",
4635           MemNode::Memory, ld->_idx, phi->_idx);
4636       }
4637 #endif
4638       _igvn.replace_input_of(ld, MemNode::Memory, phi);
4639     }
4640   }//for
4641 
4642   restart(); // invalidate all basic structures, since we rebuilt the graph
4643 
4644   if (TraceSuperWord && Verbose) {
4645     tty->print_cr("\nSuperWord::hoist_loads_in_graph() the graph was rebuilt, all structures invalidated and need rebuild");
4646   }
4647 
4648   return true;
4649 }