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() || AlignVector) {
 617         int vw = vector_width(mem_ref);
 618         int vw_best = vector_width(best_align_to_mem_ref);
 619         if (vw > vw_best) {
 620           // Do not vectorize a memory access with more elements per vector
 621           // if unaligned memory access is not allowed because number of
 622           // iterations in pre-loop will be not enough to align it.
 623           create_pack = false;
 624         } else {
 625           SWPointer p2(best_align_to_mem_ref, this, NULL, false);
 626           if (align_to_ref_p.invar() != p2.invar()) {
 627             // Do not vectorize memory accesses with different invariants
 628             // if unaligned memory accesses are not allowed.
 629             create_pack = false;
 630           }
 631         }
 632       }
 633     } else {
 634       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 635         // Can't allow vectorization of unaligned memory accesses with the
 636         // same type since it could be overlapped accesses to the same array.
 637         create_pack = false;
 638       } else {
 639         // Allow independent (different type) unaligned memory operations
 640         // if HW supports them.
 641         if (!Matcher::misaligned_vectors_ok() || AlignVector) {
 642           create_pack = false;
 643         } else {
 644           // Check if packs of the same memory type but
 645           // with a different alignment were created before.
 646           for (uint i = 0; i < align_to_refs.size(); i++) {
 647             MemNode* mr = align_to_refs.at(i)->as_Mem();
 648             if (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 (!same_velt_type(s1, s2)) return false;
1213   Node* s1_ctrl = s1->in(0);
1214   Node* s2_ctrl = s2->in(0);
1215   // If the control nodes are equivalent, no further checks are required to test for isomorphism.
1216   if (s1_ctrl == s2_ctrl) {
1217     return true;
1218   } else {
1219     bool s1_ctrl_inv = ((s1_ctrl == NULL) ? true : lpt()->is_invariant(s1_ctrl));
1220     bool s2_ctrl_inv = ((s2_ctrl == NULL) ? true : lpt()->is_invariant(s2_ctrl));
1221     // If the control nodes are not invariant for the loop, fail isomorphism test.
1222     if (!s1_ctrl_inv || !s2_ctrl_inv) {
1223       return false;
1224     }
1225     if(s1_ctrl != NULL && s2_ctrl != NULL) {
1226       if (s1_ctrl->is_Proj()) {
1227         s1_ctrl = s1_ctrl->in(0);
1228         assert(lpt()->is_invariant(s1_ctrl), "must be invariant");
1229       }
1230       if (s2_ctrl->is_Proj()) {
1231         s2_ctrl = s2_ctrl->in(0);
1232         assert(lpt()->is_invariant(s2_ctrl), "must be invariant");
1233       }
1234       if (!s1_ctrl->is_RangeCheck() || !s2_ctrl->is_RangeCheck()) {
1235         return false;
1236       }
1237     }
1238     // Control nodes are invariant. However, we have no way of checking whether they resolve
1239     // in an equivalent manner. But, we know that invariant range checks are guaranteed to
1240     // throw before the loop (if they would have thrown). Thus, the loop would not have been reached.
1241     // Therefore, if the control nodes for both are range checks, we accept them to be isomorphic.
1242     for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1243       Node* t1 = s1->fast_out(i);
1244       for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1245         Node* t2 = s2->fast_out(j);
1246         if (VectorNode::is_muladds2i(t1) && VectorNode::is_muladds2i(t2)) {
1247           return true;
1248         }
1249       }
1250     }
1251   }
1252   return false;
1253 }
1254 
1255 //------------------------------independent---------------------------
1256 // Is there no data path from s1 to s2 or s2 to s1?
1257 bool SuperWord::independent(Node* s1, Node* s2) {
1258   //  assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
1259   int d1 = depth(s1);
1260   int d2 = depth(s2);
1261   if (d1 == d2) return s1 != s2;
1262   Node* deep    = d1 > d2 ? s1 : s2;
1263   Node* shallow = d1 > d2 ? s2 : s1;
1264 
1265   visited_clear();
1266 
1267   return independent_path(shallow, deep);
1268 }
1269 
1270 //--------------------------have_similar_inputs-----------------------
1271 // For a node pair (s1, s2) which is isomorphic and independent,
1272 // do s1 and s2 have similar input edges?
1273 bool SuperWord::have_similar_inputs(Node* s1, Node* s2) {
1274   // assert(isomorphic(s1, s2) == true, "check isomorphic");
1275   // assert(independent(s1, s2) == true, "check independent");
1276   if (s1->req() > 1 && !s1->is_Store() && !s1->is_Load()) {
1277     for (uint i = 1; i < s1->req(); i++) {
1278       if (s1->in(i)->Opcode() != s2->in(i)->Opcode()) return false;
1279     }
1280   }
1281   return true;
1282 }
1283 
1284 //------------------------------reduction---------------------------
1285 // Is there a data path between s1 and s2 and the nodes reductions?
1286 bool SuperWord::reduction(Node* s1, Node* s2) {
1287   bool retValue = false;
1288   int d1 = depth(s1);
1289   int d2 = depth(s2);
1290   if (d1 + 1 == d2) {
1291     if (s1->is_reduction() && s2->is_reduction()) {
1292       // This is an ordered set, so s1 should define s2
1293       for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1294         Node* t1 = s1->fast_out(i);
1295         if (t1 == s2) {
1296           // both nodes are reductions and connected
1297           retValue = true;
1298         }
1299       }
1300     }
1301   }
1302 
1303   return retValue;
1304 }
1305 
1306 //------------------------------independent_path------------------------------
1307 // Helper for independent
1308 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
1309   if (dp >= 1000) return false; // stop deep recursion
1310   visited_set(deep);
1311   int shal_depth = depth(shallow);
1312   assert(shal_depth <= depth(deep), "must be");
1313   for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
1314     Node* pred = preds.current();
1315     if (in_bb(pred) && !visited_test(pred)) {
1316       if (shallow == pred) {
1317         return false;
1318       }
1319       if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
1320         return false;
1321       }
1322     }
1323   }
1324   return true;
1325 }
1326 
1327 //------------------------------set_alignment---------------------------
1328 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
1329   set_alignment(s1, align);
1330   if (align == top_align || align == bottom_align) {
1331     set_alignment(s2, align);
1332   } else {
1333     set_alignment(s2, align + data_size(s1));
1334   }
1335 }
1336 
1337 //------------------------------data_size---------------------------
1338 int SuperWord::data_size(Node* s) {
1339   Node* use = NULL; //test if the node is a candidate for CMoveV optimization, then return the size of CMov
1340   if (UseVectorCmov) {
1341     use = _cmovev_kit.is_Bool_candidate(s);
1342     if (use != NULL) {
1343       return data_size(use);
1344     }
1345     use = _cmovev_kit.is_CmpD_candidate(s);
1346     if (use != NULL) {
1347       return data_size(use);
1348     }
1349   }
1350 
1351   int bsize = type2aelembytes(velt_basic_type(s));
1352   assert(bsize != 0, "valid size");
1353   return bsize;
1354 }
1355 
1356 //------------------------------extend_packlist---------------------------
1357 // Extend packset by following use->def and def->use links from pack members.
1358 void SuperWord::extend_packlist() {
1359   bool changed;
1360   do {
1361     packset_sort(_packset.length());
1362     changed = false;
1363     for (int i = 0; i < _packset.length(); i++) {
1364       Node_List* p = _packset.at(i);
1365       changed |= follow_use_defs(p);
1366       changed |= follow_def_uses(p);
1367     }
1368   } while (changed);
1369 
1370   if (_race_possible) {
1371     for (int i = 0; i < _packset.length(); i++) {
1372       Node_List* p = _packset.at(i);
1373       order_def_uses(p);
1374     }
1375   }
1376 
1377   if (TraceSuperWord) {
1378     tty->print_cr("\nAfter extend_packlist");
1379     print_packset();
1380   }
1381 }
1382 
1383 //------------------------------follow_use_defs---------------------------
1384 // Extend the packset by visiting operand definitions of nodes in pack p
1385 bool SuperWord::follow_use_defs(Node_List* p) {
1386   assert(p->size() == 2, "just checking");
1387   Node* s1 = p->at(0);
1388   Node* s2 = p->at(1);
1389   assert(s1->req() == s2->req(), "just checking");
1390   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1391 
1392   if (s1->is_Load()) return false;
1393 
1394   int align = alignment(s1);
1395   NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: s1 %d, align %d", s1->_idx, align);)
1396   bool changed = false;
1397   int start = s1->is_Store() ? MemNode::ValueIn   : 1;
1398   int end   = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
1399   for (int j = start; j < end; j++) {
1400     Node* t1 = s1->in(j);
1401     Node* t2 = s2->in(j);
1402     if (!in_bb(t1) || !in_bb(t2))
1403       continue;
1404     if (stmts_can_pack(t1, t2, align)) {
1405       if (est_savings(t1, t2) >= 0) {
1406         Node_List* pair = new Node_List();
1407         pair->push(t1);
1408         pair->push(t2);
1409         _packset.append(pair);
1410         NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: set_alignment(%d, %d, %d)", t1->_idx, t2->_idx, align);)
1411         set_alignment(t1, t2, align);
1412         changed = true;
1413       }
1414     }
1415   }
1416   return changed;
1417 }
1418 
1419 //------------------------------follow_def_uses---------------------------
1420 // Extend the packset by visiting uses of nodes in pack p
1421 bool SuperWord::follow_def_uses(Node_List* p) {
1422   bool changed = false;
1423   Node* s1 = p->at(0);
1424   Node* s2 = p->at(1);
1425   assert(p->size() == 2, "just checking");
1426   assert(s1->req() == s2->req(), "just checking");
1427   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1428 
1429   if (s1->is_Store()) return false;
1430 
1431   int align = alignment(s1);
1432   NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: s1 %d, align %d", s1->_idx, align);)
1433   int savings = -1;
1434   int num_s1_uses = 0;
1435   Node* u1 = NULL;
1436   Node* u2 = NULL;
1437   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1438     Node* t1 = s1->fast_out(i);
1439     num_s1_uses++;
1440     if (!in_bb(t1)) continue;
1441     for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1442       Node* t2 = s2->fast_out(j);
1443       if (!in_bb(t2)) continue;
1444       if (t2->Opcode() == Op_AddI && t2 == _lp->as_CountedLoop()->incr()) continue; // don't mess with the iv
1445       if (!opnd_positions_match(s1, t1, s2, t2))
1446         continue;
1447       if (stmts_can_pack(t1, t2, align)) {
1448         int my_savings = est_savings(t1, t2);
1449         if (my_savings > savings) {
1450           savings = my_savings;
1451           u1 = t1;
1452           u2 = t2;
1453         }
1454       }
1455     }
1456   }
1457   if (num_s1_uses > 1) {
1458     _race_possible = true;
1459   }
1460   if (savings >= 0) {
1461     Node_List* pair = new Node_List();
1462     pair->push(u1);
1463     pair->push(u2);
1464     _packset.append(pair);
1465     NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: set_alignment(%d, %d, %d)", u1->_idx, u2->_idx, align);)
1466     set_alignment(u1, u2, align);
1467     changed = true;
1468   }
1469   return changed;
1470 }
1471 
1472 //------------------------------order_def_uses---------------------------
1473 // For extended packsets, ordinally arrange uses packset by major component
1474 void SuperWord::order_def_uses(Node_List* p) {
1475   Node* s1 = p->at(0);
1476 
1477   if (s1->is_Store()) return;
1478 
1479   // reductions are always managed beforehand
1480   if (s1->is_reduction()) return;
1481 
1482   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1483     Node* t1 = s1->fast_out(i);
1484 
1485     // Only allow operand swap on commuting operations
1486     if (!t1->is_Add() && !t1->is_Mul() && !VectorNode::is_muladds2i(t1)) {
1487       break;
1488     }
1489 
1490     // Now find t1's packset
1491     Node_List* p2 = NULL;
1492     for (int j = 0; j < _packset.length(); j++) {
1493       p2 = _packset.at(j);
1494       Node* first = p2->at(0);
1495       if (t1 == first) {
1496         break;
1497       }
1498       p2 = NULL;
1499     }
1500     // Arrange all sub components by the major component
1501     if (p2 != NULL) {
1502       for (uint j = 1; j < p->size(); j++) {
1503         Node* d1 = p->at(j);
1504         Node* u1 = p2->at(j);
1505         opnd_positions_match(s1, t1, d1, u1);
1506       }
1507     }
1508   }
1509 }
1510 
1511 //---------------------------opnd_positions_match-------------------------
1512 // Is the use of d1 in u1 at the same operand position as d2 in u2?
1513 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
1514   // check reductions to see if they are marshalled to represent the reduction
1515   // operator in a specified opnd
1516   if (u1->is_reduction() && u2->is_reduction()) {
1517     // ensure reductions have phis and reduction definitions feeding the 1st operand
1518     Node* first = u1->in(2);
1519     if (first->is_Phi() || first->is_reduction()) {
1520       u1->swap_edges(1, 2);
1521     }
1522     // ensure reductions have phis and reduction definitions feeding the 1st operand
1523     first = u2->in(2);
1524     if (first->is_Phi() || first->is_reduction()) {
1525       u2->swap_edges(1, 2);
1526     }
1527     return true;
1528   }
1529 
1530   uint ct = u1->req();
1531   if (ct != u2->req()) return false;
1532   uint i1 = 0;
1533   uint i2 = 0;
1534   do {
1535     for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
1536     for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
1537     if (i1 != i2) {
1538       if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
1539         // Further analysis relies on operands position matching.
1540         u2->swap_edges(i1, i2);
1541       } else if (VectorNode::is_muladds2i(u2) && u1 != u2) {
1542         if (i1 == 5 - i2) { // ((i1 == 3 && i2 == 2) || (i1 == 2 && i2 == 3) || (i1 == 1 && i2 == 4) || (i1 == 4 && i2 == 1))
1543           u2->swap_edges(1, 2);
1544           u2->swap_edges(3, 4);
1545         }
1546         if (i1 == 3 - i2 || i1 == 7 - i2) { // ((i1 == 1 && i2 == 2) || (i1 == 2 && i2 == 1) || (i1 == 3 && i2 == 4) || (i1 == 4 && i2 == 3))
1547           u2->swap_edges(2, 3);
1548           u2->swap_edges(1, 4);
1549         }
1550         return false; // Just swap the edges, the muladds2i nodes get packed in follow_use_defs
1551       } else {
1552         return false;
1553       }
1554     } else if (i1 == i2 && VectorNode::is_muladds2i(u2) && u1 != u2) {
1555       u2->swap_edges(1, 3);
1556       u2->swap_edges(2, 4);
1557       return false; // Just swap the edges, the muladds2i nodes get packed in follow_use_defs
1558     }
1559   } while (i1 < ct);
1560   return true;
1561 }
1562 
1563 //------------------------------est_savings---------------------------
1564 // Estimate the savings from executing s1 and s2 as a pack
1565 int SuperWord::est_savings(Node* s1, Node* s2) {
1566   int save_in = 2 - 1; // 2 operations per instruction in packed form
1567 
1568   // inputs
1569   for (uint i = 1; i < s1->req(); i++) {
1570     Node* x1 = s1->in(i);
1571     Node* x2 = s2->in(i);
1572     if (x1 != x2) {
1573       if (are_adjacent_refs(x1, x2)) {
1574         save_in += adjacent_profit(x1, x2);
1575       } else if (!in_packset(x1, x2)) {
1576         save_in -= pack_cost(2);
1577       } else {
1578         save_in += unpack_cost(2);
1579       }
1580     }
1581   }
1582 
1583   // uses of result
1584   uint ct = 0;
1585   int save_use = 0;
1586   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1587     Node* s1_use = s1->fast_out(i);
1588     for (int j = 0; j < _packset.length(); j++) {
1589       Node_List* p = _packset.at(j);
1590       if (p->at(0) == s1_use) {
1591         for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
1592           Node* s2_use = s2->fast_out(k);
1593           if (p->at(p->size()-1) == s2_use) {
1594             ct++;
1595             if (are_adjacent_refs(s1_use, s2_use)) {
1596               save_use += adjacent_profit(s1_use, s2_use);
1597             }
1598           }
1599         }
1600       }
1601     }
1602   }
1603 
1604   if (ct < s1->outcnt()) save_use += unpack_cost(1);
1605   if (ct < s2->outcnt()) save_use += unpack_cost(1);
1606 
1607   return MAX2(save_in, save_use);
1608 }
1609 
1610 //------------------------------costs---------------------------
1611 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
1612 int SuperWord::pack_cost(int ct)   { return ct; }
1613 int SuperWord::unpack_cost(int ct) { return ct; }
1614 
1615 //------------------------------combine_packs---------------------------
1616 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
1617 void SuperWord::combine_packs() {
1618   bool changed = true;
1619   // Combine packs regardless max vector size.
1620   while (changed) {
1621     changed = false;
1622     for (int i = 0; i < _packset.length(); i++) {
1623       Node_List* p1 = _packset.at(i);
1624       if (p1 == NULL) continue;
1625       // Because of sorting we can start at i + 1
1626       for (int j = i + 1; j < _packset.length(); j++) {
1627         Node_List* p2 = _packset.at(j);
1628         if (p2 == NULL) continue;
1629         if (i == j) continue;
1630         if (p1->at(p1->size()-1) == p2->at(0)) {
1631           for (uint k = 1; k < p2->size(); k++) {
1632             p1->push(p2->at(k));
1633           }
1634           _packset.at_put(j, NULL);
1635           changed = true;
1636         }
1637       }
1638     }
1639   }
1640 
1641   // Split packs which have size greater then max vector size.
1642   for (int i = 0; i < _packset.length(); i++) {
1643     Node_List* p1 = _packset.at(i);
1644     if (p1 != NULL) {
1645       BasicType bt = velt_basic_type(p1->at(0));
1646       uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1647       assert(is_power_of_2(max_vlen), "sanity");
1648       uint psize = p1->size();
1649       if (!is_power_of_2(psize)) {
1650         // Skip pack which can't be vector.
1651         // case1: for(...) { a[i] = i; }    elements values are different (i+x)
1652         // case2: for(...) { a[i] = b[i+1]; }  can't align both, load and store
1653         _packset.at_put(i, NULL);
1654         continue;
1655       }
1656       if (psize > max_vlen) {
1657         Node_List* pack = new Node_List();
1658         for (uint j = 0; j < psize; j++) {
1659           pack->push(p1->at(j));
1660           if (pack->size() >= max_vlen) {
1661             assert(is_power_of_2(pack->size()), "sanity");
1662             _packset.append(pack);
1663             pack = new Node_List();
1664           }
1665         }
1666         _packset.at_put(i, NULL);
1667       }
1668     }
1669   }
1670 
1671   // Compress list.
1672   for (int i = _packset.length() - 1; i >= 0; i--) {
1673     Node_List* p1 = _packset.at(i);
1674     if (p1 == NULL) {
1675       _packset.remove_at(i);
1676     }
1677   }
1678 
1679   if (TraceSuperWord) {
1680     tty->print_cr("\nAfter combine_packs");
1681     print_packset();
1682   }
1683 }
1684 
1685 //-----------------------------construct_my_pack_map--------------------------
1686 // Construct the map from nodes to packs.  Only valid after the
1687 // point where a node is only in one pack (after combine_packs).
1688 void SuperWord::construct_my_pack_map() {
1689   Node_List* rslt = NULL;
1690   for (int i = 0; i < _packset.length(); i++) {
1691     Node_List* p = _packset.at(i);
1692     for (uint j = 0; j < p->size(); j++) {
1693       Node* s = p->at(j);
1694       assert(my_pack(s) == NULL, "only in one pack");
1695       set_my_pack(s, p);
1696     }
1697   }
1698 }
1699 
1700 //------------------------------filter_packs---------------------------
1701 // Remove packs that are not implemented or not profitable.
1702 void SuperWord::filter_packs() {
1703   // Remove packs that are not implemented
1704   for (int i = _packset.length() - 1; i >= 0; i--) {
1705     Node_List* pk = _packset.at(i);
1706     bool impl = implemented(pk);
1707     if (!impl) {
1708 #ifndef PRODUCT
1709       if (TraceSuperWord && Verbose) {
1710         tty->print_cr("Unimplemented");
1711         pk->at(0)->dump();
1712       }
1713 #endif
1714       remove_pack_at(i);
1715     }
1716     Node *n = pk->at(0);
1717     if (n->is_reduction()) {
1718       _num_reductions++;
1719     } else {
1720       _num_work_vecs++;
1721     }
1722   }
1723 
1724   // Remove packs that are not profitable
1725   bool changed;
1726   do {
1727     changed = false;
1728     for (int i = _packset.length() - 1; i >= 0; i--) {
1729       Node_List* pk = _packset.at(i);
1730       bool prof = profitable(pk);
1731       if (!prof) {
1732 #ifndef PRODUCT
1733         if (TraceSuperWord && Verbose) {
1734           tty->print_cr("Unprofitable");
1735           pk->at(0)->dump();
1736         }
1737 #endif
1738         remove_pack_at(i);
1739         changed = true;
1740       }
1741     }
1742   } while (changed);
1743 
1744 #ifndef PRODUCT
1745   if (TraceSuperWord) {
1746     tty->print_cr("\nAfter filter_packs");
1747     print_packset();
1748     tty->cr();
1749   }
1750 #endif
1751 }
1752 
1753 //------------------------------merge_packs_to_cmovd---------------------------
1754 // Merge CMoveD into new vector-nodes
1755 // We want to catch this pattern and subsume CmpD and Bool into CMoveD
1756 //
1757 //                   SubD             ConD
1758 //                  /  |               /
1759 //                 /   |           /   /
1760 //                /    |       /      /
1761 //               /     |   /         /
1762 //              /      /            /
1763 //             /    /  |           /
1764 //            v /      |          /
1765 //         CmpD        |         /
1766 //          |          |        /
1767 //          v          |       /
1768 //         Bool        |      /
1769 //           \         |     /
1770 //             \       |    /
1771 //               \     |   /
1772 //                 \   |  /
1773 //                   \ v /
1774 //                   CMoveD
1775 //
1776 
1777 void SuperWord::merge_packs_to_cmovd() {
1778   for (int i = _packset.length() - 1; i >= 0; i--) {
1779     _cmovev_kit.make_cmovevd_pack(_packset.at(i));
1780   }
1781   #ifndef PRODUCT
1782     if (TraceSuperWord) {
1783       tty->print_cr("\nSuperWord::merge_packs_to_cmovd(): After merge");
1784       print_packset();
1785       tty->cr();
1786     }
1787   #endif
1788 }
1789 
1790 Node* CMoveKit::is_Bool_candidate(Node* def) const {
1791   Node* use = NULL;
1792   if (!def->is_Bool() || def->in(0) != NULL || def->outcnt() != 1) {
1793     return NULL;
1794   }
1795   for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1796     use = def->fast_out(j);
1797     if (!_sw->same_generation(def, use) || !use->is_CMove()) {
1798       return NULL;
1799     }
1800   }
1801   return use;
1802 }
1803 
1804 Node* CMoveKit::is_CmpD_candidate(Node* def) const {
1805   Node* use = NULL;
1806   if (!def->is_Cmp() || def->in(0) != NULL || def->outcnt() != 1) {
1807     return NULL;
1808   }
1809   for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1810     use = def->fast_out(j);
1811     if (!_sw->same_generation(def, use) || (use = is_Bool_candidate(use)) == NULL || !_sw->same_generation(def, use)) {
1812       return NULL;
1813     }
1814   }
1815   return use;
1816 }
1817 
1818 Node_List* CMoveKit::make_cmovevd_pack(Node_List* cmovd_pk) {
1819   Node *cmovd = cmovd_pk->at(0);
1820   if (!cmovd->is_CMove()) {
1821     return NULL;
1822   }
1823   if (cmovd->Opcode() != Op_CMoveF && cmovd->Opcode() != Op_CMoveD) {
1824     return NULL;
1825   }
1826   if (pack(cmovd) != NULL) { // already in the cmov pack
1827     return NULL;
1828   }
1829   if (cmovd->in(0) != NULL) {
1830     NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: CMoveD %d has control flow, escaping...", cmovd->_idx); cmovd->dump();})
1831     return NULL;
1832   }
1833 
1834   Node* bol = cmovd->as_CMove()->in(CMoveNode::Condition);
1835   if (!bol->is_Bool()
1836       || bol->outcnt() != 1
1837       || !_sw->same_generation(bol, cmovd)
1838       || bol->in(0) != NULL  // BoolNode has control flow!!
1839       || _sw->my_pack(bol) == NULL) {
1840       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();})
1841       return NULL;
1842   }
1843   Node_List* bool_pk = _sw->my_pack(bol);
1844   if (bool_pk->size() != cmovd_pk->size() ) {
1845     return NULL;
1846   }
1847 
1848   Node* cmpd = bol->in(1);
1849   if (!cmpd->is_Cmp()
1850       || cmpd->outcnt() != 1
1851       || !_sw->same_generation(cmpd, cmovd)
1852       || cmpd->in(0) != NULL  // CmpDNode has control flow!!
1853       || _sw->my_pack(cmpd) == NULL) {
1854       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();})
1855       return NULL;
1856   }
1857   Node_List* cmpd_pk = _sw->my_pack(cmpd);
1858   if (cmpd_pk->size() != cmovd_pk->size() ) {
1859     return NULL;
1860   }
1861 
1862   if (!test_cmpd_pack(cmpd_pk, cmovd_pk)) {
1863     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();})
1864     return NULL;
1865   }
1866 
1867   Node_List* new_cmpd_pk = new Node_List();
1868   uint sz = cmovd_pk->size() - 1;
1869   for (uint i = 0; i <= sz; ++i) {
1870     Node* cmov = cmovd_pk->at(i);
1871     Node* bol  = bool_pk->at(i);
1872     Node* cmp  = cmpd_pk->at(i);
1873 
1874     new_cmpd_pk->insert(i, cmov);
1875 
1876     map(cmov, new_cmpd_pk);
1877     map(bol, new_cmpd_pk);
1878     map(cmp, new_cmpd_pk);
1879 
1880     _sw->set_my_pack(cmov, new_cmpd_pk); // and keep old packs for cmp and bool
1881   }
1882   _sw->_packset.remove(cmovd_pk);
1883   _sw->_packset.remove(bool_pk);
1884   _sw->_packset.remove(cmpd_pk);
1885   _sw->_packset.append(new_cmpd_pk);
1886   NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print_cr("CMoveKit::make_cmovevd_pack: added syntactic CMoveD pack"); _sw->print_pack(new_cmpd_pk);})
1887   return new_cmpd_pk;
1888 }
1889 
1890 bool CMoveKit::test_cmpd_pack(Node_List* cmpd_pk, Node_List* cmovd_pk) {
1891   Node* cmpd0 = cmpd_pk->at(0);
1892   assert(cmpd0->is_Cmp(), "CMoveKit::test_cmpd_pack: should be CmpDNode");
1893   assert(cmovd_pk->at(0)->is_CMove(), "CMoveKit::test_cmpd_pack: should be CMoveD");
1894   assert(cmpd_pk->size() == cmovd_pk->size(), "CMoveKit::test_cmpd_pack: should be same size");
1895   Node* in1 = cmpd0->in(1);
1896   Node* in2 = cmpd0->in(2);
1897   Node_List* in1_pk = _sw->my_pack(in1);
1898   Node_List* in2_pk = _sw->my_pack(in2);
1899 
1900   if (  (in1_pk != NULL && in1_pk->size() != cmpd_pk->size())
1901      || (in2_pk != NULL && in2_pk->size() != cmpd_pk->size()) ) {
1902     return false;
1903   }
1904 
1905   // test if "all" in1 are in the same pack or the same node
1906   if (in1_pk == NULL) {
1907     for (uint j = 1; j < cmpd_pk->size(); j++) {
1908       if (cmpd_pk->at(j)->in(1) != in1) {
1909         return false;
1910       }
1911     }//for: in1_pk is not pack but all CmpD nodes in the pack have the same in(1)
1912   }
1913   // test if "all" in2 are in the same pack or the same node
1914   if (in2_pk == NULL) {
1915     for (uint j = 1; j < cmpd_pk->size(); j++) {
1916       if (cmpd_pk->at(j)->in(2) != in2) {
1917         return false;
1918       }
1919     }//for: in2_pk is not pack but all CmpD nodes in the pack have the same in(2)
1920   }
1921   //now check if cmpd_pk may be subsumed in vector built for cmovd_pk
1922   int cmovd_ind1, cmovd_ind2;
1923   if (cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1924    && cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1925       cmovd_ind1 = CMoveNode::IfFalse;
1926       cmovd_ind2 = CMoveNode::IfTrue;
1927   } else if (cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1928           && cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1929       cmovd_ind2 = CMoveNode::IfFalse;
1930       cmovd_ind1 = CMoveNode::IfTrue;
1931   }
1932   else {
1933     return false;
1934   }
1935 
1936   for (uint j = 1; j < cmpd_pk->size(); j++) {
1937     if (cmpd_pk->at(j)->in(1) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind1)
1938         || cmpd_pk->at(j)->in(2) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind2)) {
1939         return false;
1940     }//if
1941   }
1942   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(); })
1943   return true;
1944 }
1945 
1946 //------------------------------implemented---------------------------
1947 // Can code be generated for pack p?
1948 bool SuperWord::implemented(Node_List* p) {
1949   bool retValue = false;
1950   Node* p0 = p->at(0);
1951   if (p0 != NULL) {
1952     int opc = p0->Opcode();
1953     uint size = p->size();
1954     if (p0->is_reduction()) {
1955       const Type *arith_type = p0->bottom_type();
1956       // Length 2 reductions of INT/LONG do not offer performance benefits
1957       if (((arith_type->basic_type() == T_INT) || (arith_type->basic_type() == T_LONG)) && (size == 2)) {
1958         retValue = false;
1959       } else {
1960         retValue = ReductionNode::implemented(opc, size, arith_type->basic_type());
1961       }
1962     } else {
1963       retValue = VectorNode::implemented(opc, size, velt_basic_type(p0));
1964     }
1965     if (!retValue) {
1966       if (is_cmov_pack(p)) {
1967         NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::implemented: found cmpd pack"); print_pack(p);})
1968         return true;
1969       }
1970     }
1971   }
1972   return retValue;
1973 }
1974 
1975 bool SuperWord::is_cmov_pack(Node_List* p) {
1976   return _cmovev_kit.pack(p->at(0)) != NULL;
1977 }
1978 //------------------------------same_inputs--------------------------
1979 // For pack p, are all idx operands the same?
1980 bool SuperWord::same_inputs(Node_List* p, int idx) {
1981   Node* p0 = p->at(0);
1982   uint vlen = p->size();
1983   Node* p0_def = p0->in(idx);
1984   for (uint i = 1; i < vlen; i++) {
1985     Node* pi = p->at(i);
1986     Node* pi_def = pi->in(idx);
1987     if (p0_def != pi_def) {
1988       return false;
1989     }
1990   }
1991   return true;
1992 }
1993 
1994 //------------------------------profitable---------------------------
1995 // For pack p, are all operands and all uses (with in the block) vector?
1996 bool SuperWord::profitable(Node_List* p) {
1997   Node* p0 = p->at(0);
1998   uint start, end;
1999   VectorNode::vector_operands(p0, &start, &end);
2000 
2001   // Return false if some inputs are not vectors or vectors with different
2002   // size or alignment.
2003   // Also, for now, return false if not scalar promotion case when inputs are
2004   // the same. Later, implement PackNode and allow differing, non-vector inputs
2005   // (maybe just the ones from outside the block.)
2006   for (uint i = start; i < end; i++) {
2007     if (!is_vector_use(p0, i)) {
2008       return false;
2009     }
2010   }
2011   // Check if reductions are connected
2012   if (p0->is_reduction()) {
2013     Node* second_in = p0->in(2);
2014     Node_List* second_pk = my_pack(second_in);
2015     if ((second_pk == NULL) || (_num_work_vecs == _num_reductions)) {
2016       // Remove reduction flag if no parent pack or if not enough work
2017       // to cover reduction expansion overhead
2018       p0->remove_flag(Node::Flag_is_reduction);
2019       return false;
2020     } else if (second_pk->size() != p->size()) {
2021       return false;
2022     }
2023   }
2024   if (VectorNode::is_shift(p0)) {
2025     // For now, return false if shift count is vector or not scalar promotion
2026     // case (different shift counts) because it is not supported yet.
2027     Node* cnt = p0->in(2);
2028     Node_List* cnt_pk = my_pack(cnt);
2029     if (cnt_pk != NULL)
2030       return false;
2031     if (!same_inputs(p, 2))
2032       return false;
2033   }
2034   if (!p0->is_Store()) {
2035     // For now, return false if not all uses are vector.
2036     // Later, implement ExtractNode and allow non-vector uses (maybe
2037     // just the ones outside the block.)
2038     for (uint i = 0; i < p->size(); i++) {
2039       Node* def = p->at(i);
2040       if (is_cmov_pack_internal_node(p, def)) {
2041         continue;
2042       }
2043       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2044         Node* use = def->fast_out(j);
2045         for (uint k = 0; k < use->req(); k++) {
2046           Node* n = use->in(k);
2047           if (def == n) {
2048             // reductions should only have a Phi use at the the loop
2049             // head and out of loop uses
2050             if (def->is_reduction() &&
2051                 ((use->is_Phi() && use->in(0) == _lpt->_head) ||
2052                  !_lpt->is_member(_phase->get_loop(_phase->ctrl_or_self(use))))) {
2053               assert(i == p->size()-1, "must be last element of the pack");
2054               continue;
2055             }
2056             if (!is_vector_use(use, k)) {
2057               return false;
2058             }
2059           }
2060         }
2061       }
2062     }
2063   }
2064   return true;
2065 }
2066 
2067 //------------------------------schedule---------------------------
2068 // Adjust the memory graph for the packed operations
2069 void SuperWord::schedule() {
2070 
2071   // Co-locate in the memory graph the members of each memory pack
2072   for (int i = 0; i < _packset.length(); i++) {
2073     co_locate_pack(_packset.at(i));
2074   }
2075 }
2076 
2077 //-------------------------------remove_and_insert-------------------
2078 // Remove "current" from its current position in the memory graph and insert
2079 // it after the appropriate insertion point (lip or uip).
2080 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
2081                                   Node *uip, Unique_Node_List &sched_before) {
2082   Node* my_mem = current->in(MemNode::Memory);
2083   bool sched_up = sched_before.member(current);
2084 
2085   // remove current_store from its current position in the memmory graph
2086   for (DUIterator i = current->outs(); current->has_out(i); i++) {
2087     Node* use = current->out(i);
2088     if (use->is_Mem()) {
2089       assert(use->in(MemNode::Memory) == current, "must be");
2090       if (use == prev) { // connect prev to my_mem
2091           _igvn.replace_input_of(use, MemNode::Memory, my_mem);
2092           --i; //deleted this edge; rescan position
2093       } else if (sched_before.member(use)) {
2094         if (!sched_up) { // Will be moved together with current
2095           _igvn.replace_input_of(use, MemNode::Memory, uip);
2096           --i; //deleted this edge; rescan position
2097         }
2098       } else {
2099         if (sched_up) { // Will be moved together with current
2100           _igvn.replace_input_of(use, MemNode::Memory, lip);
2101           --i; //deleted this edge; rescan position
2102         }
2103       }
2104     }
2105   }
2106 
2107   Node *insert_pt =  sched_up ?  uip : lip;
2108 
2109   // all uses of insert_pt's memory state should use current's instead
2110   for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
2111     Node* use = insert_pt->out(i);
2112     if (use->is_Mem()) {
2113       assert(use->in(MemNode::Memory) == insert_pt, "must be");
2114       _igvn.replace_input_of(use, MemNode::Memory, current);
2115       --i; //deleted this edge; rescan position
2116     } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
2117       uint pos; //lip (lower insert point) must be the last one in the memory slice
2118       for (pos=1; pos < use->req(); pos++) {
2119         if (use->in(pos) == insert_pt) break;
2120       }
2121       _igvn.replace_input_of(use, pos, current);
2122       --i;
2123     }
2124   }
2125 
2126   //connect current to insert_pt
2127   _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
2128 }
2129 
2130 //------------------------------co_locate_pack----------------------------------
2131 // To schedule a store pack, we need to move any sandwiched memory ops either before
2132 // or after the pack, based upon dependence information:
2133 // (1) If any store in the pack depends on the sandwiched memory op, the
2134 //     sandwiched memory op must be scheduled BEFORE the pack;
2135 // (2) If a sandwiched memory op depends on any store in the pack, the
2136 //     sandwiched memory op must be scheduled AFTER the pack;
2137 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
2138 //     memory op (say memB), memB must be scheduled before memA. So, if memA is
2139 //     scheduled before the pack, memB must also be scheduled before the pack;
2140 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
2141 //     schedule this store AFTER the pack
2142 // (5) We know there is no dependence cycle, so there in no other case;
2143 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
2144 //
2145 // To schedule a load pack, we use the memory state of either the first or the last load in
2146 // the pack, based on the dependence constraint.
2147 void SuperWord::co_locate_pack(Node_List* pk) {
2148   if (pk->at(0)->is_Store()) {
2149     MemNode* first     = executed_first(pk)->as_Mem();
2150     MemNode* last      = executed_last(pk)->as_Mem();
2151     Unique_Node_List schedule_before_pack;
2152     Unique_Node_List memops;
2153 
2154     MemNode* current   = last->in(MemNode::Memory)->as_Mem();
2155     MemNode* previous  = last;
2156     while (true) {
2157       assert(in_bb(current), "stay in block");
2158       memops.push(previous);
2159       for (DUIterator i = current->outs(); current->has_out(i); i++) {
2160         Node* use = current->out(i);
2161         if (use->is_Mem() && use != previous)
2162           memops.push(use);
2163       }
2164       if (current == first) break;
2165       previous = current;
2166       current  = current->in(MemNode::Memory)->as_Mem();
2167     }
2168 
2169     // determine which memory operations should be scheduled before the pack
2170     for (uint i = 1; i < memops.size(); i++) {
2171       Node *s1 = memops.at(i);
2172       if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
2173         for (uint j = 0; j< i; j++) {
2174           Node *s2 = memops.at(j);
2175           if (!independent(s1, s2)) {
2176             if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
2177               schedule_before_pack.push(s1); // s1 must be scheduled before
2178               Node_List* mem_pk = my_pack(s1);
2179               if (mem_pk != NULL) {
2180                 for (uint ii = 0; ii < mem_pk->size(); ii++) {
2181                   Node* s = mem_pk->at(ii);  // follow partner
2182                   if (memops.member(s) && !schedule_before_pack.member(s))
2183                     schedule_before_pack.push(s);
2184                 }
2185               }
2186               break;
2187             }
2188           }
2189         }
2190       }
2191     }
2192 
2193     Node*    upper_insert_pt = first->in(MemNode::Memory);
2194     // Following code moves loads connected to upper_insert_pt below aliased stores.
2195     // Collect such loads here and reconnect them back to upper_insert_pt later.
2196     memops.clear();
2197     for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
2198       Node* use = upper_insert_pt->out(i);
2199       if (use->is_Mem() && !use->is_Store()) {
2200         memops.push(use);
2201       }
2202     }
2203 
2204     MemNode* lower_insert_pt = last;
2205     previous                 = last; //previous store in pk
2206     current                  = last->in(MemNode::Memory)->as_Mem();
2207 
2208     // start scheduling from "last" to "first"
2209     while (true) {
2210       assert(in_bb(current), "stay in block");
2211       assert(in_pack(previous, pk), "previous stays in pack");
2212       Node* my_mem = current->in(MemNode::Memory);
2213 
2214       if (in_pack(current, pk)) {
2215         // Forward users of my memory state (except "previous) to my input memory state
2216         for (DUIterator i = current->outs(); current->has_out(i); i++) {
2217           Node* use = current->out(i);
2218           if (use->is_Mem() && use != previous) {
2219             assert(use->in(MemNode::Memory) == current, "must be");
2220             if (schedule_before_pack.member(use)) {
2221               _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
2222             } else {
2223               _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
2224             }
2225             --i; // deleted this edge; rescan position
2226           }
2227         }
2228         previous = current;
2229       } else { // !in_pack(current, pk) ==> a sandwiched store
2230         remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
2231       }
2232 
2233       if (current == first) break;
2234       current = my_mem->as_Mem();
2235     } // end while
2236 
2237     // Reconnect loads back to upper_insert_pt.
2238     for (uint i = 0; i < memops.size(); i++) {
2239       Node *ld = memops.at(i);
2240       if (ld->in(MemNode::Memory) != upper_insert_pt) {
2241         _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
2242       }
2243     }
2244   } else if (pk->at(0)->is_Load()) { //load
2245     // all loads in the pack should have the same memory state. By default,
2246     // we use the memory state of the last load. However, if any load could
2247     // not be moved down due to the dependence constraint, we use the memory
2248     // state of the first load.
2249     Node* first_mem = pk->at(0)->in(MemNode::Memory);
2250     Node* last_mem = first_mem;
2251     for (uint i = 1; i < pk->size(); i++) {
2252       Node* ld = pk->at(i);
2253       Node* mem = ld->in(MemNode::Memory);
2254       assert(in_bb(first_mem) || in_bb(mem) || mem == first_mem, "2 different memory state from outside the loop?");
2255       if (in_bb(mem)) {
2256         if (in_bb(first_mem) && bb_idx(mem) < bb_idx(first_mem)) {
2257           first_mem = mem;
2258         }
2259         if (!in_bb(last_mem) || bb_idx(mem) > bb_idx(last_mem)) {
2260           last_mem = mem;
2261         }
2262       }
2263     }
2264     bool schedule_last = true;
2265     for (uint i = 0; i < pk->size(); i++) {
2266       Node* ld = pk->at(i);
2267       for (Node* current = last_mem; current != ld->in(MemNode::Memory);
2268            current=current->in(MemNode::Memory)) {
2269         assert(current != first_mem, "corrupted memory graph");
2270         if(current->is_Mem() && !independent(current, ld)){
2271           schedule_last = false; // a later store depends on this load
2272           break;
2273         }
2274       }
2275     }
2276 
2277     Node* mem_input = schedule_last ? last_mem : first_mem;
2278     _igvn.hash_delete(mem_input);
2279     // Give each load the same memory state
2280     for (uint i = 0; i < pk->size(); i++) {
2281       LoadNode* ld = pk->at(i)->as_Load();
2282       _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
2283     }
2284   }
2285 }
2286 
2287 #ifndef PRODUCT
2288 void SuperWord::print_loop(bool whole) {
2289   Node_Stack stack(_arena, _phase->C->unique() >> 2);
2290   Node_List rpo_list;
2291   VectorSet visited(_arena);
2292   visited.set(lpt()->_head->_idx);
2293   _phase->rpo(lpt()->_head, stack, visited, rpo_list);
2294   _phase->dump(lpt(), rpo_list.size(), rpo_list );
2295   if(whole) {
2296     tty->print_cr("\n Whole loop tree");
2297     _phase->dump();
2298     tty->print_cr(" End of whole loop tree\n");
2299   }
2300 }
2301 #endif
2302 
2303 //------------------------------output---------------------------
2304 // Convert packs into vector node operations
2305 void SuperWord::output() {
2306   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2307   Compile* C = _phase->C;
2308   if (_packset.length() == 0) {
2309     if (cl->is_main_loop()) {
2310       // Instigate more unrolling for optimization when vectorization fails.
2311       C->set_major_progress();
2312       cl->set_notpassed_slp();
2313       cl->mark_do_unroll_only();
2314     }
2315     return;
2316   }
2317 
2318 #ifndef PRODUCT
2319   if (TraceLoopOpts) {
2320     tty->print("SuperWord::output    ");
2321     lpt()->dump_head();
2322   }
2323 #endif
2324 
2325   if (cl->is_main_loop()) {
2326     // MUST ENSURE main loop's initial value is properly aligned:
2327     //  (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
2328 
2329     align_initial_loop_index(align_to_ref());
2330 
2331     // Insert extract (unpack) operations for scalar uses
2332     for (int i = 0; i < _packset.length(); i++) {
2333       insert_extracts(_packset.at(i));
2334     }
2335   }
2336 
2337   uint max_vlen_in_bytes = 0;
2338   uint max_vlen = 0;
2339   bool can_process_post_loop = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
2340 
2341   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop before create_reserve_version_of_loop"); print_loop(true);})
2342 
2343   CountedLoopReserveKit make_reversable(_phase, _lpt, do_reserve_copy());
2344 
2345   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop after create_reserve_version_of_loop"); print_loop(true);})
2346 
2347   if (do_reserve_copy() && !make_reversable.has_reserved()) {
2348     NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: loop was not reserved correctly, exiting SuperWord");})
2349     return;
2350   }
2351 
2352   for (int i = 0; i < _block.length(); i++) {
2353     Node* n = _block.at(i);
2354     Node_List* p = my_pack(n);
2355     if (p && n == executed_last(p)) {
2356       uint vlen = p->size();
2357       uint vlen_in_bytes = 0;
2358       Node* vn = NULL;
2359       Node* low_adr = p->at(0);
2360       Node* first   = executed_first(p);
2361       if (can_process_post_loop) {
2362         // override vlen with the main loops vector length
2363         vlen = cl->slp_max_unroll();
2364       }
2365       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);})
2366       int   opc = n->Opcode();
2367       if (n->is_Load()) {
2368         Node* ctl = n->in(MemNode::Control);
2369         Node* mem = first->in(MemNode::Memory);
2370         SWPointer p1(n->as_Mem(), this, NULL, false);
2371         // Identify the memory dependency for the new loadVector node by
2372         // walking up through memory chain.
2373         // This is done to give flexibility to the new loadVector node so that
2374         // it can move above independent storeVector nodes.
2375         while (mem->is_StoreVector()) {
2376           SWPointer p2(mem->as_Mem(), this, NULL, false);
2377           int cmp = p1.cmp(p2);
2378           if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
2379             mem = mem->in(MemNode::Memory);
2380           } else {
2381             break; // dependent memory
2382           }
2383         }
2384         Node* adr = low_adr->in(MemNode::Address);
2385         const TypePtr* atyp = n->adr_type();
2386         vn = LoadVectorNode::make(opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n), control_dependency(p));
2387         vlen_in_bytes = vn->as_LoadVector()->memory_size();
2388       } else if (n->is_Store()) {
2389         // Promote value to be stored to vector
2390         Node* val = vector_opd(p, MemNode::ValueIn);
2391         if (val == NULL) {
2392           if (do_reserve_copy()) {
2393             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: val should not be NULL, exiting SuperWord");})
2394             return; //and reverse to backup IG
2395           }
2396           ShouldNotReachHere();
2397         }
2398 
2399         Node* ctl = n->in(MemNode::Control);
2400         Node* mem = first->in(MemNode::Memory);
2401         Node* adr = low_adr->in(MemNode::Address);
2402         const TypePtr* atyp = n->adr_type();
2403         vn = StoreVectorNode::make(opc, ctl, mem, adr, atyp, val, vlen);
2404         vlen_in_bytes = vn->as_StoreVector()->memory_size();
2405       } else if (VectorNode::is_muladds2i(n)) {
2406         assert(n->req() == 5u, "MulAddS2I should have 4 operands.");
2407         Node* in1 = vector_opd(p, 1);
2408         Node* in2 = vector_opd(p, 2);
2409         vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2410         vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2411       } else if (n->req() == 3 && !is_cmov_pack(p)) {
2412         // Promote operands to vector
2413         Node* in1 = NULL;
2414         bool node_isa_reduction = n->is_reduction();
2415         if (node_isa_reduction) {
2416           // the input to the first reduction operation is retained
2417           in1 = low_adr->in(1);
2418         } else {
2419           in1 = vector_opd(p, 1);
2420           if (in1 == NULL) {
2421             if (do_reserve_copy()) {
2422               NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in1 should not be NULL, exiting SuperWord");})
2423               return; //and reverse to backup IG
2424             }
2425             ShouldNotReachHere();
2426           }
2427         }
2428         Node* in2 = vector_opd(p, 2);
2429         if (in2 == NULL) {
2430           if (do_reserve_copy()) {
2431             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in2 should not be NULL, exiting SuperWord");})
2432             return; //and reverse to backup IG
2433           }
2434           ShouldNotReachHere();
2435         }
2436         if (VectorNode::is_invariant_vector(in1) && (node_isa_reduction == false) && (n->is_Add() || n->is_Mul())) {
2437           // Move invariant vector input into second position to avoid register spilling.
2438           Node* tmp = in1;
2439           in1 = in2;
2440           in2 = tmp;
2441         }
2442         if (node_isa_reduction) {
2443           const Type *arith_type = n->bottom_type();
2444           vn = ReductionNode::make(opc, NULL, in1, in2, arith_type->basic_type());
2445           if (in2->is_Load()) {
2446             vlen_in_bytes = in2->as_LoadVector()->memory_size();
2447           } else {
2448             vlen_in_bytes = in2->as_Vector()->length_in_bytes();
2449           }
2450         } else {
2451           vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2452           vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2453         }
2454       } else if (opc == Op_SqrtF || opc == Op_SqrtD ||
2455                  opc == Op_AbsF || opc == Op_AbsD ||
2456                  opc == Op_AbsI || opc == Op_AbsL ||
2457                  opc == Op_NegF || opc == Op_NegD ||
2458                  opc == Op_PopCountI) {
2459         assert(n->req() == 2, "only one input expected");
2460         Node* in = vector_opd(p, 1);
2461         vn = VectorNode::make(opc, in, NULL, vlen, velt_basic_type(n));
2462         vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2463       } else if (is_cmov_pack(p)) {
2464         if (can_process_post_loop) {
2465           // do not refactor of flow in post loop context
2466           return;
2467         }
2468         if (!n->is_CMove()) {
2469           continue;
2470         }
2471         // place here CMoveVDNode
2472         NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::output: print before CMove vectorization"); print_loop(false);})
2473         Node* bol = n->in(CMoveNode::Condition);
2474         if (!bol->is_Bool() && bol->Opcode() == Op_ExtractI && bol->req() > 1 ) {
2475           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();})
2476           bol = bol->in(1); //may be ExtractNode
2477         }
2478 
2479         assert(bol->is_Bool(), "should be BoolNode - too late to bail out!");
2480         if (!bol->is_Bool()) {
2481           if (do_reserve_copy()) {
2482             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: expected %d bool node, exiting SuperWord", bol->_idx); bol->dump();})
2483             return; //and reverse to backup IG
2484           }
2485           ShouldNotReachHere();
2486         }
2487 
2488         int cond = (int)bol->as_Bool()->_test._test;
2489         Node* in_cc  = _igvn.intcon(cond);
2490         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created intcon in_cc node %d", in_cc->_idx); in_cc->dump();})
2491         Node* cc = bol->clone();
2492         cc->set_req(1, in_cc);
2493         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created bool cc node %d", cc->_idx); cc->dump();})
2494 
2495         Node* src1 = vector_opd(p, 2); //2=CMoveNode::IfFalse
2496         if (src1 == NULL) {
2497           if (do_reserve_copy()) {
2498             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src1 should not be NULL, exiting SuperWord");})
2499             return; //and reverse to backup IG
2500           }
2501           ShouldNotReachHere();
2502         }
2503         Node* src2 = vector_opd(p, 3); //3=CMoveNode::IfTrue
2504         if (src2 == NULL) {
2505           if (do_reserve_copy()) {
2506             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src2 should not be NULL, exiting SuperWord");})
2507             return; //and reverse to backup IG
2508           }
2509           ShouldNotReachHere();
2510         }
2511         BasicType bt = velt_basic_type(n);
2512         const TypeVect* vt = TypeVect::make(bt, vlen);
2513         assert(bt == T_FLOAT || bt == T_DOUBLE, "Only vectorization for FP cmovs is supported");
2514         if (bt == T_FLOAT) {
2515           vn = new CMoveVFNode(cc, src1, src2, vt);
2516         } else {
2517           assert(bt == T_DOUBLE, "Expected double");
2518           vn = new CMoveVDNode(cc, src1, src2, vt);
2519         }
2520         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created new CMove node %d: ", vn->_idx); vn->dump();})
2521       } else if (opc == Op_FmaD || opc == Op_FmaF) {
2522         // Promote operands to vector
2523         Node* in1 = vector_opd(p, 1);
2524         Node* in2 = vector_opd(p, 2);
2525         Node* in3 = vector_opd(p, 3);
2526         vn = VectorNode::make(opc, in1, in2, in3, vlen, velt_basic_type(n));
2527         vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2528       } else {
2529         if (do_reserve_copy()) {
2530           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: ShouldNotReachHere, exiting SuperWord");})
2531           return; //and reverse to backup IG
2532         }
2533         ShouldNotReachHere();
2534       }
2535 
2536       assert(vn != NULL, "sanity");
2537       if (vn == NULL) {
2538         if (do_reserve_copy()){
2539           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: got NULL node, cannot proceed, exiting SuperWord");})
2540           return; //and reverse to backup IG
2541         }
2542         ShouldNotReachHere();
2543       }
2544 
2545       _block.at_put(i, vn);
2546       _igvn.register_new_node_with_optimizer(vn);
2547       _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
2548       for (uint j = 0; j < p->size(); j++) {
2549         Node* pm = p->at(j);
2550         _igvn.replace_node(pm, vn);
2551       }
2552       _igvn._worklist.push(vn);
2553 
2554       if (can_process_post_loop) {
2555         // first check if the vector size if the maximum vector which we can use on the machine,
2556         // other vector size have reduced values for predicated data mapping.
2557         if (vlen_in_bytes != (uint)MaxVectorSize) {
2558           return;
2559         }
2560       }
2561 
2562       if (vlen_in_bytes >= max_vlen_in_bytes && vlen > max_vlen) {
2563         max_vlen = vlen;
2564         max_vlen_in_bytes = vlen_in_bytes;
2565       }
2566 #ifdef ASSERT
2567       if (TraceNewVectors) {
2568         tty->print("new Vector node: ");
2569         vn->dump();
2570       }
2571 #endif
2572     }
2573   }//for (int i = 0; i < _block.length(); i++)
2574 
2575   if (max_vlen_in_bytes > C->max_vector_size()) {
2576     C->set_max_vector_size(max_vlen_in_bytes);
2577   }
2578   if (max_vlen_in_bytes > 0) {
2579     cl->mark_loop_vectorized();
2580   }
2581 
2582   if (SuperWordLoopUnrollAnalysis) {
2583     if (cl->has_passed_slp()) {
2584       uint slp_max_unroll_factor = cl->slp_max_unroll();
2585       if (slp_max_unroll_factor == max_vlen) {
2586         if (TraceSuperWordLoopUnrollAnalysis) {
2587           tty->print_cr("vector loop(unroll=%d, len=%d)\n", max_vlen, max_vlen_in_bytes*BitsPerByte);
2588         }
2589 
2590         // For atomic unrolled loops which are vector mapped, instigate more unrolling
2591         cl->set_notpassed_slp();
2592         if (cl->is_main_loop()) {
2593           // if vector resources are limited, do not allow additional unrolling, also
2594           // do not unroll more on pure vector loops which were not reduced so that we can
2595           // program the post loop to single iteration execution.
2596           if (FLOATPRESSURE > 8) {
2597             C->set_major_progress();
2598             cl->mark_do_unroll_only();
2599           }
2600         }
2601 
2602         if (do_reserve_copy()) {
2603           if (can_process_post_loop) {
2604             // Now create the difference of trip and limit and use it as our mask index.
2605             // Note: We limited the unroll of the vectorized loop so that
2606             //       only vlen-1 size iterations can remain to be mask programmed.
2607             Node *incr = cl->incr();
2608             SubINode *index = new SubINode(cl->limit(), cl->init_trip());
2609             _igvn.register_new_node_with_optimizer(index);
2610             SetVectMaskINode  *mask = new SetVectMaskINode(_phase->get_ctrl(cl->init_trip()), index);
2611             _igvn.register_new_node_with_optimizer(mask);
2612             // make this a single iteration loop
2613             AddINode *new_incr = new AddINode(incr->in(1), mask);
2614             _igvn.register_new_node_with_optimizer(new_incr);
2615             _phase->set_ctrl(new_incr, _phase->get_ctrl(incr));
2616             _igvn.replace_node(incr, new_incr);
2617             cl->mark_is_multiversioned();
2618             cl->loopexit()->add_flag(Node::Flag_has_vector_mask_set);
2619           }
2620         }
2621       }
2622     }
2623   }
2624 
2625   if (do_reserve_copy()) {
2626     make_reversable.use_new();
2627   }
2628   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("\n Final loop after SuperWord"); print_loop(true);})
2629   return;
2630 }
2631 
2632 //------------------------------vector_opd---------------------------
2633 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
2634 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
2635   Node* p0 = p->at(0);
2636   uint vlen = p->size();
2637   Node* opd = p0->in(opd_idx);
2638   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2639 
2640   if (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop()) {
2641     // override vlen with the main loops vector length
2642     vlen = cl->slp_max_unroll();
2643   }
2644 
2645   if (same_inputs(p, opd_idx)) {
2646     if (opd->is_Vector() || opd->is_LoadVector()) {
2647       assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
2648       if (opd_idx == 2 && VectorNode::is_shift(p0)) {
2649         NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("shift's count can't be vector");})
2650         return NULL;
2651       }
2652       return opd; // input is matching vector
2653     }
2654     if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
2655       Compile* C = _phase->C;
2656       Node* cnt = opd;
2657       // Vector instructions do not mask shift count, do it here.
2658       juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2659       const TypeInt* t = opd->find_int_type();
2660       if (t != NULL && t->is_con()) {
2661         juint shift = t->get_con();
2662         if (shift > mask) { // Unsigned cmp
2663           cnt = ConNode::make(TypeInt::make(shift & mask));
2664         }
2665       } else {
2666         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2667           cnt = ConNode::make(TypeInt::make(mask));
2668           _igvn.register_new_node_with_optimizer(cnt);
2669           cnt = new AndINode(opd, cnt);
2670           _igvn.register_new_node_with_optimizer(cnt);
2671           _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2672         }
2673         assert(opd->bottom_type()->isa_int(), "int type only");
2674         if (!opd->bottom_type()->isa_int()) {
2675           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should be int type only");})
2676           return NULL;
2677         }
2678         // Move non constant shift count into vector register.
2679         cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0));
2680       }
2681       if (cnt != opd) {
2682         _igvn.register_new_node_with_optimizer(cnt);
2683         _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2684       }
2685       return cnt;
2686     }
2687     assert(!opd->is_StoreVector(), "such vector is not expected here");
2688     if (opd->is_StoreVector()) {
2689       NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("StoreVector is not expected here");})
2690       return NULL;
2691     }
2692     // Convert scalar input to vector with the same number of elements as
2693     // p0's vector. Use p0's type because size of operand's container in
2694     // vector should match p0's size regardless operand's size.
2695     const Type* p0_t = velt_type(p0);
2696     VectorNode* vn = VectorNode::scalar2vector(opd, vlen, p0_t);
2697 
2698     _igvn.register_new_node_with_optimizer(vn);
2699     _phase->set_ctrl(vn, _phase->get_ctrl(opd));
2700 #ifdef ASSERT
2701     if (TraceNewVectors) {
2702       tty->print("new Vector node: ");
2703       vn->dump();
2704     }
2705 #endif
2706     return vn;
2707   }
2708 
2709   // Insert pack operation
2710   BasicType bt = velt_basic_type(p0);
2711   PackNode* pk = PackNode::make(opd, vlen, bt);
2712   DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
2713 
2714   for (uint i = 1; i < vlen; i++) {
2715     Node* pi = p->at(i);
2716     Node* in = pi->in(opd_idx);
2717     assert(my_pack(in) == NULL, "Should already have been unpacked");
2718     if (my_pack(in) != NULL) {
2719       NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should already have been unpacked");})
2720       return NULL;
2721     }
2722     assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
2723     pk->add_opd(in);
2724     if (VectorNode::is_muladds2i(pi)) {
2725       Node* in2 = pi->in(opd_idx + 2);
2726       assert(my_pack(in2) == NULL, "Should already have been unpacked");
2727       if (my_pack(in2) != NULL) {
2728         NOT_PRODUCT(if (is_trace_loop_reverse() || TraceLoopOpts) { tty->print_cr("Should already have been unpacked"); })
2729           return NULL;
2730       }
2731       assert(opd_bt == in2->bottom_type()->basic_type(), "all same type");
2732       pk->add_opd(in2);
2733     }
2734   }
2735   _igvn.register_new_node_with_optimizer(pk);
2736   _phase->set_ctrl(pk, _phase->get_ctrl(opd));
2737 #ifdef ASSERT
2738   if (TraceNewVectors) {
2739     tty->print("new Vector node: ");
2740     pk->dump();
2741   }
2742 #endif
2743   return pk;
2744 }
2745 
2746 //------------------------------insert_extracts---------------------------
2747 // If a use of pack p is not a vector use, then replace the
2748 // use with an extract operation.
2749 void SuperWord::insert_extracts(Node_List* p) {
2750   if (p->at(0)->is_Store()) return;
2751   assert(_n_idx_list.is_empty(), "empty (node,index) list");
2752 
2753   // Inspect each use of each pack member.  For each use that is
2754   // not a vector use, replace the use with an extract operation.
2755 
2756   for (uint i = 0; i < p->size(); i++) {
2757     Node* def = p->at(i);
2758     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2759       Node* use = def->fast_out(j);
2760       for (uint k = 0; k < use->req(); k++) {
2761         Node* n = use->in(k);
2762         if (def == n) {
2763           Node_List* u_pk = my_pack(use);
2764           if ((u_pk == NULL || !is_cmov_pack(u_pk) || use->is_CMove()) && !is_vector_use(use, k)) {
2765               _n_idx_list.push(use, k);
2766           }
2767         }
2768       }
2769     }
2770   }
2771 
2772   while (_n_idx_list.is_nonempty()) {
2773     Node* use = _n_idx_list.node();
2774     int   idx = _n_idx_list.index();
2775     _n_idx_list.pop();
2776     Node* def = use->in(idx);
2777 
2778     if (def->is_reduction()) continue;
2779 
2780     // Insert extract operation
2781     _igvn.hash_delete(def);
2782     int def_pos = alignment(def) / data_size(def);
2783 
2784     Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def));
2785     _igvn.register_new_node_with_optimizer(ex);
2786     _phase->set_ctrl(ex, _phase->get_ctrl(def));
2787     _igvn.replace_input_of(use, idx, ex);
2788     _igvn._worklist.push(def);
2789 
2790     bb_insert_after(ex, bb_idx(def));
2791     set_velt_type(ex, velt_type(def));
2792   }
2793 }
2794 
2795 //------------------------------is_vector_use---------------------------
2796 // Is use->in(u_idx) a vector use?
2797 bool SuperWord::is_vector_use(Node* use, int u_idx) {
2798   Node_List* u_pk = my_pack(use);
2799   if (u_pk == NULL) return false;
2800   if (use->is_reduction()) return true;
2801   Node* def = use->in(u_idx);
2802   Node_List* d_pk = my_pack(def);
2803   if (d_pk == NULL) {
2804     // check for scalar promotion
2805     Node* n = u_pk->at(0)->in(u_idx);
2806     for (uint i = 1; i < u_pk->size(); i++) {
2807       if (u_pk->at(i)->in(u_idx) != n) return false;
2808     }
2809     return true;
2810   }
2811   if (VectorNode::is_muladds2i(use)) {
2812     // MulAddS2I takes shorts and produces ints - hence the special checks
2813     // on alignment and size.
2814     if (u_pk->size() * 2 != d_pk->size()) {
2815       return false;
2816     }
2817     for (uint i = 0; i < MIN2(d_pk->size(), u_pk->size()); i++) {
2818       Node* ui = u_pk->at(i);
2819       Node* di = d_pk->at(i);
2820       if (alignment(ui) != alignment(di) * 2) {
2821         return false;
2822       }
2823     }
2824     return true;
2825   }
2826   if (u_pk->size() != d_pk->size())
2827     return false;
2828   for (uint i = 0; i < u_pk->size(); i++) {
2829     Node* ui = u_pk->at(i);
2830     Node* di = d_pk->at(i);
2831     if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
2832       return false;
2833   }
2834   return true;
2835 }
2836 
2837 //------------------------------construct_bb---------------------------
2838 // Construct reverse postorder list of block members
2839 bool SuperWord::construct_bb() {
2840   Node* entry = bb();
2841 
2842   assert(_stk.length() == 0,            "stk is empty");
2843   assert(_block.length() == 0,          "block is empty");
2844   assert(_data_entry.length() == 0,     "data_entry is empty");
2845   assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
2846   assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
2847 
2848   // Find non-control nodes with no inputs from within block,
2849   // create a temporary map from node _idx to bb_idx for use
2850   // by the visited and post_visited sets,
2851   // and count number of nodes in block.
2852   int bb_ct = 0;
2853   for (uint i = 0; i < lpt()->_body.size(); i++) {
2854     Node *n = lpt()->_body.at(i);
2855     set_bb_idx(n, i); // Create a temporary map
2856     if (in_bb(n)) {
2857       if (n->is_LoadStore() || n->is_MergeMem() ||
2858           (n->is_Proj() && !n->as_Proj()->is_CFG())) {
2859         // Bailout if the loop has LoadStore, MergeMem or data Proj
2860         // nodes. Superword optimization does not work with them.
2861         return false;
2862       }
2863       bb_ct++;
2864       if (!n->is_CFG()) {
2865         bool found = false;
2866         for (uint j = 0; j < n->req(); j++) {
2867           Node* def = n->in(j);
2868           if (def && in_bb(def)) {
2869             found = true;
2870             break;
2871           }
2872         }
2873         if (!found) {
2874           assert(n != entry, "can't be entry");
2875           _data_entry.push(n);
2876         }
2877       }
2878     }
2879   }
2880 
2881   // Find memory slices (head and tail)
2882   for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
2883     Node *n = lp()->fast_out(i);
2884     if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
2885       Node* n_tail  = n->in(LoopNode::LoopBackControl);
2886       if (n_tail != n->in(LoopNode::EntryControl)) {
2887         if (!n_tail->is_Mem()) {
2888           assert(n_tail->is_Mem(), "unexpected node for memory slice: %s", n_tail->Name());
2889           return false; // Bailout
2890         }
2891         _mem_slice_head.push(n);
2892         _mem_slice_tail.push(n_tail);
2893       }
2894     }
2895   }
2896 
2897   // Create an RPO list of nodes in block
2898 
2899   visited_clear();
2900   post_visited_clear();
2901 
2902   // Push all non-control nodes with no inputs from within block, then control entry
2903   for (int j = 0; j < _data_entry.length(); j++) {
2904     Node* n = _data_entry.at(j);
2905     visited_set(n);
2906     _stk.push(n);
2907   }
2908   visited_set(entry);
2909   _stk.push(entry);
2910 
2911   // Do a depth first walk over out edges
2912   int rpo_idx = bb_ct - 1;
2913   int size;
2914   int reduction_uses = 0;
2915   while ((size = _stk.length()) > 0) {
2916     Node* n = _stk.top(); // Leave node on stack
2917     if (!visited_test_set(n)) {
2918       // forward arc in graph
2919     } else if (!post_visited_test(n)) {
2920       // cross or back arc
2921       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2922         Node *use = n->fast_out(i);
2923         if (in_bb(use) && !visited_test(use) &&
2924             // Don't go around backedge
2925             (!use->is_Phi() || n == entry)) {
2926           if (use->is_reduction()) {
2927             // First see if we can map the reduction on the given system we are on, then
2928             // make a data entry operation for each reduction we see.
2929             BasicType bt = use->bottom_type()->basic_type();
2930             if (ReductionNode::implemented(use->Opcode(), Matcher::min_vector_size(bt), bt)) {
2931               reduction_uses++;
2932             }
2933           }
2934           _stk.push(use);
2935         }
2936       }
2937       if (_stk.length() == size) {
2938         // There were no additional uses, post visit node now
2939         _stk.pop(); // Remove node from stack
2940         assert(rpo_idx >= 0, "");
2941         _block.at_put_grow(rpo_idx, n);
2942         rpo_idx--;
2943         post_visited_set(n);
2944         assert(rpo_idx >= 0 || _stk.is_empty(), "");
2945       }
2946     } else {
2947       _stk.pop(); // Remove post-visited node from stack
2948     }
2949   }//while
2950 
2951   int ii_current = -1;
2952   unsigned int load_idx = (unsigned int)-1;
2953   _ii_order.clear();
2954   // Create real map of block indices for nodes
2955   for (int j = 0; j < _block.length(); j++) {
2956     Node* n = _block.at(j);
2957     set_bb_idx(n, j);
2958     if (_do_vector_loop && n->is_Load()) {
2959       if (ii_current == -1) {
2960         ii_current = _clone_map.gen(n->_idx);
2961         _ii_order.push(ii_current);
2962         load_idx = _clone_map.idx(n->_idx);
2963       } else if (_clone_map.idx(n->_idx) == load_idx && _clone_map.gen(n->_idx) != ii_current) {
2964         ii_current = _clone_map.gen(n->_idx);
2965         _ii_order.push(ii_current);
2966       }
2967     }
2968   }//for
2969 
2970   // Ensure extra info is allocated.
2971   initialize_bb();
2972 
2973 #ifndef PRODUCT
2974   if (_vector_loop_debug && _ii_order.length() > 0) {
2975     tty->print("SuperWord::construct_bb: List of generations: ");
2976     for (int jj = 0; jj < _ii_order.length(); ++jj) {
2977       tty->print("  %d:%d", jj, _ii_order.at(jj));
2978     }
2979     tty->print_cr(" ");
2980   }
2981   if (TraceSuperWord) {
2982     print_bb();
2983     tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
2984     for (int m = 0; m < _data_entry.length(); m++) {
2985       tty->print("%3d ", m);
2986       _data_entry.at(m)->dump();
2987     }
2988     tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
2989     for (int m = 0; m < _mem_slice_head.length(); m++) {
2990       tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
2991       tty->print("    ");    _mem_slice_tail.at(m)->dump();
2992     }
2993   }
2994 #endif
2995   assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
2996   return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0);
2997 }
2998 
2999 //------------------------------initialize_bb---------------------------
3000 // Initialize per node info
3001 void SuperWord::initialize_bb() {
3002   Node* last = _block.at(_block.length() - 1);
3003   grow_node_info(bb_idx(last));
3004 }
3005 
3006 //------------------------------bb_insert_after---------------------------
3007 // Insert n into block after pos
3008 void SuperWord::bb_insert_after(Node* n, int pos) {
3009   int n_pos = pos + 1;
3010   // Make room
3011   for (int i = _block.length() - 1; i >= n_pos; i--) {
3012     _block.at_put_grow(i+1, _block.at(i));
3013   }
3014   for (int j = _node_info.length() - 1; j >= n_pos; j--) {
3015     _node_info.at_put_grow(j+1, _node_info.at(j));
3016   }
3017   // Set value
3018   _block.at_put_grow(n_pos, n);
3019   _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
3020   // Adjust map from node->_idx to _block index
3021   for (int i = n_pos; i < _block.length(); i++) {
3022     set_bb_idx(_block.at(i), i);
3023   }
3024 }
3025 
3026 //------------------------------compute_max_depth---------------------------
3027 // Compute max depth for expressions from beginning of block
3028 // Use to prune search paths during test for independence.
3029 void SuperWord::compute_max_depth() {
3030   int ct = 0;
3031   bool again;
3032   do {
3033     again = false;
3034     for (int i = 0; i < _block.length(); i++) {
3035       Node* n = _block.at(i);
3036       if (!n->is_Phi()) {
3037         int d_orig = depth(n);
3038         int d_in   = 0;
3039         for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
3040           Node* pred = preds.current();
3041           if (in_bb(pred)) {
3042             d_in = MAX2(d_in, depth(pred));
3043           }
3044         }
3045         if (d_in + 1 != d_orig) {
3046           set_depth(n, d_in + 1);
3047           again = true;
3048         }
3049       }
3050     }
3051     ct++;
3052   } while (again);
3053 
3054   if (TraceSuperWord && Verbose) {
3055     tty->print_cr("compute_max_depth iterated: %d times", ct);
3056   }
3057 }
3058 
3059 //-------------------------compute_vector_element_type-----------------------
3060 // Compute necessary vector element type for expressions
3061 // This propagates backwards a narrower integer type when the
3062 // upper bits of the value are not needed.
3063 // Example:  char a,b,c;  a = b + c;
3064 // Normally the type of the add is integer, but for packed character
3065 // operations the type of the add needs to be char.
3066 void SuperWord::compute_vector_element_type() {
3067   if (TraceSuperWord && Verbose) {
3068     tty->print_cr("\ncompute_velt_type:");
3069   }
3070 
3071   // Initial type
3072   for (int i = 0; i < _block.length(); i++) {
3073     Node* n = _block.at(i);
3074     set_velt_type(n, container_type(n));
3075   }
3076 
3077   // Propagate integer narrowed type backwards through operations
3078   // that don't depend on higher order bits
3079   for (int i = _block.length() - 1; i >= 0; i--) {
3080     Node* n = _block.at(i);
3081     // Only integer types need be examined
3082     const Type* vtn = velt_type(n);
3083     if (vtn->basic_type() == T_INT) {
3084       uint start, end;
3085       VectorNode::vector_operands(n, &start, &end);
3086 
3087       for (uint j = start; j < end; j++) {
3088         Node* in  = n->in(j);
3089         // Don't propagate through a memory
3090         if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
3091             data_size(n) < data_size(in)) {
3092           bool same_type = true;
3093           for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
3094             Node *use = in->fast_out(k);
3095             if (!in_bb(use) || !same_velt_type(use, n)) {
3096               same_type = false;
3097               break;
3098             }
3099           }
3100           if (same_type) {
3101             // For right shifts of small integer types (bool, byte, char, short)
3102             // we need precise information about sign-ness. Only Load nodes have
3103             // this information because Store nodes are the same for signed and
3104             // unsigned values. And any arithmetic operation after a load may
3105             // expand a value to signed Int so such right shifts can't be used
3106             // because vector elements do not have upper bits of Int.
3107             const Type* vt = vtn;
3108             if (VectorNode::is_shift(in)) {
3109               Node* load = in->in(1);
3110               if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
3111                 vt = velt_type(load);
3112               } else if (in->Opcode() != Op_LShiftI) {
3113                 // Widen type to Int to avoid creation of right shift vector
3114                 // (align + data_size(s1) check in stmts_can_pack() will fail).
3115                 // Note, left shifts work regardless type.
3116                 vt = TypeInt::INT;
3117               }
3118             }
3119             set_velt_type(in, vt);
3120           }
3121         }
3122       }
3123     }
3124   }
3125 #ifndef PRODUCT
3126   if (TraceSuperWord && Verbose) {
3127     for (int i = 0; i < _block.length(); i++) {
3128       Node* n = _block.at(i);
3129       velt_type(n)->dump();
3130       tty->print("\t");
3131       n->dump();
3132     }
3133   }
3134 #endif
3135 }
3136 
3137 //------------------------------memory_alignment---------------------------
3138 // Alignment within a vector memory reference
3139 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
3140   #ifndef PRODUCT
3141     if(TraceSuperWord && Verbose) {
3142       tty->print("SuperWord::memory_alignment within a vector memory reference for %d:  ", s->_idx); s->dump();
3143     }
3144   #endif
3145   NOT_PRODUCT(SWPointer::Tracer::Depth ddd(0);)
3146   SWPointer p(s, this, NULL, false);
3147   if (!p.valid()) {
3148     NOT_PRODUCT(if(is_trace_alignment()) tty->print("SWPointer::memory_alignment: SWPointer p invalid, return bottom_align");)
3149     return bottom_align;
3150   }
3151   int vw = get_vw_bytes_special(s);
3152   if (vw < 2) {
3153     NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SWPointer::memory_alignment: vector_width_in_bytes < 2, return bottom_align");)
3154     return bottom_align; // No vectors for this type
3155   }
3156   int offset  = p.offset_in_bytes();
3157   offset     += iv_adjust*p.memory_size();
3158   int off_rem = offset % vw;
3159   int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
3160   if (TraceSuperWord && Verbose) {
3161     tty->print_cr("SWPointer::memory_alignment: off_rem = %d, off_mod = %d", off_rem, off_mod);
3162   }
3163   return off_mod;
3164 }
3165 
3166 //---------------------------container_type---------------------------
3167 // Smallest type containing range of values
3168 const Type* SuperWord::container_type(Node* n) {
3169   if (n->is_Mem()) {
3170     BasicType bt = n->as_Mem()->memory_type();
3171     if (n->is_Store() && (bt == T_CHAR)) {
3172       // Use T_SHORT type instead of T_CHAR for stored values because any
3173       // preceding arithmetic operation extends values to signed Int.
3174       bt = T_SHORT;
3175     }
3176     if (n->Opcode() == Op_LoadUB) {
3177       // Adjust type for unsigned byte loads, it is important for right shifts.
3178       // T_BOOLEAN is used because there is no basic type representing type
3179       // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
3180       // size (one byte) and sign is important.
3181       bt = T_BOOLEAN;
3182     }
3183     return Type::get_const_basic_type(bt);
3184   }
3185   const Type* t = _igvn.type(n);
3186   if (t->basic_type() == T_INT) {
3187     // A narrow type of arithmetic operations will be determined by
3188     // propagating the type of memory operations.
3189     return TypeInt::INT;
3190   }
3191   return t;
3192 }
3193 
3194 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
3195   const Type* vt1 = velt_type(n1);
3196   const Type* vt2 = velt_type(n2);
3197   if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
3198     // Compare vectors element sizes for integer types.
3199     return data_size(n1) == data_size(n2);
3200   }
3201   return vt1 == vt2;
3202 }
3203 
3204 //------------------------------in_packset---------------------------
3205 // Are s1 and s2 in a pack pair and ordered as s1,s2?
3206 bool SuperWord::in_packset(Node* s1, Node* s2) {
3207   for (int i = 0; i < _packset.length(); i++) {
3208     Node_List* p = _packset.at(i);
3209     assert(p->size() == 2, "must be");
3210     if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
3211       return true;
3212     }
3213   }
3214   return false;
3215 }
3216 
3217 //------------------------------in_pack---------------------------
3218 // Is s in pack p?
3219 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
3220   for (uint i = 0; i < p->size(); i++) {
3221     if (p->at(i) == s) {
3222       return p;
3223     }
3224   }
3225   return NULL;
3226 }
3227 
3228 //------------------------------remove_pack_at---------------------------
3229 // Remove the pack at position pos in the packset
3230 void SuperWord::remove_pack_at(int pos) {
3231   Node_List* p = _packset.at(pos);
3232   for (uint i = 0; i < p->size(); i++) {
3233     Node* s = p->at(i);
3234     set_my_pack(s, NULL);
3235   }
3236   _packset.remove_at(pos);
3237 }
3238 
3239 void SuperWord::packset_sort(int n) {
3240   // simple bubble sort so that we capitalize with O(n) when its already sorted
3241   while (n != 0) {
3242     bool swapped = false;
3243     for (int i = 1; i < n; i++) {
3244       Node_List* q_low = _packset.at(i-1);
3245       Node_List* q_i = _packset.at(i);
3246 
3247       // only swap when we find something to swap
3248       if (alignment(q_low->at(0)) > alignment(q_i->at(0))) {
3249         Node_List* t = q_i;
3250         *(_packset.adr_at(i)) = q_low;
3251         *(_packset.adr_at(i-1)) = q_i;
3252         swapped = true;
3253       }
3254     }
3255     if (swapped == false) break;
3256     n--;
3257   }
3258 }
3259 
3260 //------------------------------executed_first---------------------------
3261 // Return the node executed first in pack p.  Uses the RPO block list
3262 // to determine order.
3263 Node* SuperWord::executed_first(Node_List* p) {
3264   Node* n = p->at(0);
3265   int n_rpo = bb_idx(n);
3266   for (uint i = 1; i < p->size(); i++) {
3267     Node* s = p->at(i);
3268     int s_rpo = bb_idx(s);
3269     if (s_rpo < n_rpo) {
3270       n = s;
3271       n_rpo = s_rpo;
3272     }
3273   }
3274   return n;
3275 }
3276 
3277 //------------------------------executed_last---------------------------
3278 // Return the node executed last in pack p.
3279 Node* SuperWord::executed_last(Node_List* p) {
3280   Node* n = p->at(0);
3281   int n_rpo = bb_idx(n);
3282   for (uint i = 1; i < p->size(); i++) {
3283     Node* s = p->at(i);
3284     int s_rpo = bb_idx(s);
3285     if (s_rpo > n_rpo) {
3286       n = s;
3287       n_rpo = s_rpo;
3288     }
3289   }
3290   return n;
3291 }
3292 
3293 LoadNode::ControlDependency SuperWord::control_dependency(Node_List* p) {
3294   LoadNode::ControlDependency dep = LoadNode::DependsOnlyOnTest;
3295   for (uint i = 0; i < p->size(); i++) {
3296     Node* n = p->at(i);
3297     assert(n->is_Load(), "only meaningful for loads");
3298     if (!n->depends_only_on_test()) {
3299       dep = LoadNode::Pinned;
3300     }
3301   }
3302   return dep;
3303 }
3304 
3305 
3306 //----------------------------align_initial_loop_index---------------------------
3307 // Adjust pre-loop limit so that in main loop, a load/store reference
3308 // to align_to_ref will be a position zero in the vector.
3309 //   (iv + k) mod vector_align == 0
3310 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
3311   CountedLoopNode *main_head = lp()->as_CountedLoop();
3312   assert(main_head->is_main_loop(), "");
3313   CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
3314   assert(pre_end != NULL, "we must have a correct pre-loop");
3315   Node *pre_opaq1 = pre_end->limit();
3316   assert(pre_opaq1->Opcode() == Op_Opaque1, "");
3317   Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
3318   Node *lim0 = pre_opaq->in(1);
3319 
3320   // Where we put new limit calculations
3321   Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
3322 
3323   // Ensure the original loop limit is available from the
3324   // pre-loop Opaque1 node.
3325   Node *orig_limit = pre_opaq->original_loop_limit();
3326   assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
3327 
3328   SWPointer align_to_ref_p(align_to_ref, this, NULL, false);
3329   assert(align_to_ref_p.valid(), "sanity");
3330 
3331   // Given:
3332   //     lim0 == original pre loop limit
3333   //     V == v_align (power of 2)
3334   //     invar == extra invariant piece of the address expression
3335   //     e == offset [ +/- invar ]
3336   //
3337   // When reassociating expressions involving '%' the basic rules are:
3338   //     (a - b) % k == 0   =>  a % k == b % k
3339   // and:
3340   //     (a + b) % k == 0   =>  a % k == (k - b) % k
3341   //
3342   // For stride > 0 && scale > 0,
3343   //   Derive the new pre-loop limit "lim" such that the two constraints:
3344   //     (1) lim = lim0 + N           (where N is some positive integer < V)
3345   //     (2) (e + lim) % V == 0
3346   //   are true.
3347   //
3348   //   Substituting (1) into (2),
3349   //     (e + lim0 + N) % V == 0
3350   //   solve for N:
3351   //     N = (V - (e + lim0)) % V
3352   //   substitute back into (1), so that new limit
3353   //     lim = lim0 + (V - (e + lim0)) % V
3354   //
3355   // For stride > 0 && scale < 0
3356   //   Constraints:
3357   //     lim = lim0 + N
3358   //     (e - lim) % V == 0
3359   //   Solving for lim:
3360   //     (e - lim0 - N) % V == 0
3361   //     N = (e - lim0) % V
3362   //     lim = lim0 + (e - lim0) % V
3363   //
3364   // For stride < 0 && scale > 0
3365   //   Constraints:
3366   //     lim = lim0 - N
3367   //     (e + lim) % V == 0
3368   //   Solving for lim:
3369   //     (e + lim0 - N) % V == 0
3370   //     N = (e + lim0) % V
3371   //     lim = lim0 - (e + lim0) % V
3372   //
3373   // For stride < 0 && scale < 0
3374   //   Constraints:
3375   //     lim = lim0 - N
3376   //     (e - lim) % V == 0
3377   //   Solving for lim:
3378   //     (e - lim0 + N) % V == 0
3379   //     N = (V - (e - lim0)) % V
3380   //     lim = lim0 - (V - (e - lim0)) % V
3381 
3382   int vw = vector_width_in_bytes(align_to_ref);
3383   int stride   = iv_stride();
3384   int scale    = align_to_ref_p.scale_in_bytes();
3385   int elt_size = align_to_ref_p.memory_size();
3386   int v_align  = vw / elt_size;
3387   assert(v_align > 1, "sanity");
3388   int offset   = align_to_ref_p.offset_in_bytes() / elt_size;
3389   Node *offsn  = _igvn.intcon(offset);
3390 
3391   Node *e = offsn;
3392   if (align_to_ref_p.invar() != NULL) {
3393     // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
3394     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3395     Node* invar = align_to_ref_p.invar();
3396     if (_igvn.type(invar)->isa_long()) {
3397       // Computations are done % (vector width/element size) so it's
3398       // safe to simply convert invar to an int and loose the upper 32
3399       // bit half.
3400       invar = new ConvL2INode(invar);
3401       _igvn.register_new_node_with_optimizer(invar);
3402     }
3403     Node* aref = new URShiftINode(invar, log2_elt);
3404     _igvn.register_new_node_with_optimizer(aref);
3405     _phase->set_ctrl(aref, pre_ctrl);
3406     if (align_to_ref_p.negate_invar()) {
3407       e = new SubINode(e, aref);
3408     } else {
3409       e = new AddINode(e, aref);
3410     }
3411     _igvn.register_new_node_with_optimizer(e);
3412     _phase->set_ctrl(e, pre_ctrl);
3413   }
3414   if (vw > ObjectAlignmentInBytes || align_to_ref_p.base()->is_top()) {
3415     // incorporate base e +/- base && Mask >>> log2(elt)
3416     Node* xbase = new CastP2XNode(NULL, align_to_ref_p.adr());
3417     _igvn.register_new_node_with_optimizer(xbase);
3418 #ifdef _LP64
3419     xbase  = new ConvL2INode(xbase);
3420     _igvn.register_new_node_with_optimizer(xbase);
3421 #endif
3422     Node* mask = _igvn.intcon(vw-1);
3423     Node* masked_xbase  = new AndINode(xbase, mask);
3424     _igvn.register_new_node_with_optimizer(masked_xbase);
3425     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3426     Node* bref     = new URShiftINode(masked_xbase, log2_elt);
3427     _igvn.register_new_node_with_optimizer(bref);
3428     _phase->set_ctrl(bref, pre_ctrl);
3429     e = new AddINode(e, bref);
3430     _igvn.register_new_node_with_optimizer(e);
3431     _phase->set_ctrl(e, pre_ctrl);
3432   }
3433 
3434   // compute e +/- lim0
3435   if (scale < 0) {
3436     e = new SubINode(e, lim0);
3437   } else {
3438     e = new AddINode(e, lim0);
3439   }
3440   _igvn.register_new_node_with_optimizer(e);
3441   _phase->set_ctrl(e, pre_ctrl);
3442 
3443   if (stride * scale > 0) {
3444     // compute V - (e +/- lim0)
3445     Node* va  = _igvn.intcon(v_align);
3446     e = new SubINode(va, e);
3447     _igvn.register_new_node_with_optimizer(e);
3448     _phase->set_ctrl(e, pre_ctrl);
3449   }
3450   // compute N = (exp) % V
3451   Node* va_msk = _igvn.intcon(v_align - 1);
3452   Node* N = new AndINode(e, va_msk);
3453   _igvn.register_new_node_with_optimizer(N);
3454   _phase->set_ctrl(N, pre_ctrl);
3455 
3456   //   substitute back into (1), so that new limit
3457   //     lim = lim0 + N
3458   Node* lim;
3459   if (stride < 0) {
3460     lim = new SubINode(lim0, N);
3461   } else {
3462     lim = new AddINode(lim0, N);
3463   }
3464   _igvn.register_new_node_with_optimizer(lim);
3465   _phase->set_ctrl(lim, pre_ctrl);
3466   Node* constrained =
3467     (stride > 0) ? (Node*) new MinINode(lim, orig_limit)
3468                  : (Node*) new MaxINode(lim, orig_limit);
3469   _igvn.register_new_node_with_optimizer(constrained);
3470   _phase->set_ctrl(constrained, pre_ctrl);
3471   _igvn.replace_input_of(pre_opaq, 1, constrained);
3472 }
3473 
3474 //----------------------------get_pre_loop_end---------------------------
3475 // Find pre loop end from main loop.  Returns null if none.
3476 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode* cl) {
3477   // The loop cannot be optimized if the graph shape at
3478   // the loop entry is inappropriate.
3479   if (!PhaseIdealLoop::is_canonical_loop_entry(cl)) {
3480     return NULL;
3481   }
3482 
3483   Node* p_f = cl->skip_predicates()->in(0)->in(0);
3484   if (!p_f->is_IfFalse()) return NULL;
3485   if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
3486   CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
3487   CountedLoopNode* loop_node = pre_end->loopnode();
3488   if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
3489   return pre_end;
3490 }
3491 
3492 //------------------------------init---------------------------
3493 void SuperWord::init() {
3494   _dg.init();
3495   _packset.clear();
3496   _disjoint_ptrs.clear();
3497   _block.clear();
3498   _post_block.clear();
3499   _data_entry.clear();
3500   _mem_slice_head.clear();
3501   _mem_slice_tail.clear();
3502   _iteration_first.clear();
3503   _iteration_last.clear();
3504   _node_info.clear();
3505   _align_to_ref = NULL;
3506   _lpt = NULL;
3507   _lp = NULL;
3508   _bb = NULL;
3509   _iv = NULL;
3510   _race_possible = 0;
3511   _early_return = false;
3512   _num_work_vecs = 0;
3513   _num_reductions = 0;
3514 }
3515 
3516 //------------------------------restart---------------------------
3517 void SuperWord::restart() {
3518   _dg.init();
3519   _packset.clear();
3520   _disjoint_ptrs.clear();
3521   _block.clear();
3522   _post_block.clear();
3523   _data_entry.clear();
3524   _mem_slice_head.clear();
3525   _mem_slice_tail.clear();
3526   _node_info.clear();
3527 }
3528 
3529 //------------------------------print_packset---------------------------
3530 void SuperWord::print_packset() {
3531 #ifndef PRODUCT
3532   tty->print_cr("packset");
3533   for (int i = 0; i < _packset.length(); i++) {
3534     tty->print_cr("Pack: %d", i);
3535     Node_List* p = _packset.at(i);
3536     print_pack(p);
3537   }
3538 #endif
3539 }
3540 
3541 //------------------------------print_pack---------------------------
3542 void SuperWord::print_pack(Node_List* p) {
3543   for (uint i = 0; i < p->size(); i++) {
3544     print_stmt(p->at(i));
3545   }
3546 }
3547 
3548 //------------------------------print_bb---------------------------
3549 void SuperWord::print_bb() {
3550 #ifndef PRODUCT
3551   tty->print_cr("\nBlock");
3552   for (int i = 0; i < _block.length(); i++) {
3553     Node* n = _block.at(i);
3554     tty->print("%d ", i);
3555     if (n) {
3556       n->dump();
3557     }
3558   }
3559 #endif
3560 }
3561 
3562 //------------------------------print_stmt---------------------------
3563 void SuperWord::print_stmt(Node* s) {
3564 #ifndef PRODUCT
3565   tty->print(" align: %d \t", alignment(s));
3566   s->dump();
3567 #endif
3568 }
3569 
3570 //------------------------------blank---------------------------
3571 char* SuperWord::blank(uint depth) {
3572   static char blanks[101];
3573   assert(depth < 101, "too deep");
3574   for (uint i = 0; i < depth; i++) blanks[i] = ' ';
3575   blanks[depth] = '\0';
3576   return blanks;
3577 }
3578 
3579 
3580 //==============================SWPointer===========================
3581 #ifndef PRODUCT
3582 int SWPointer::Tracer::_depth = 0;
3583 #endif
3584 //----------------------------SWPointer------------------------
3585 SWPointer::SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only) :
3586   _mem(mem), _slp(slp),  _base(NULL),  _adr(NULL),
3587   _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3588   _nstack(nstack), _analyze_only(analyze_only),
3589   _stack_idx(0)
3590 #ifndef PRODUCT
3591   , _tracer(slp)
3592 #endif
3593 {
3594   NOT_PRODUCT(_tracer.ctor_1(mem);)
3595 
3596   Node* adr = mem->in(MemNode::Address);
3597   if (!adr->is_AddP()) {
3598     assert(!valid(), "too complex");
3599     return;
3600   }
3601   // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
3602   Node* base = adr->in(AddPNode::Base);
3603   // The base address should be loop invariant
3604   if (!invariant(base)) {
3605     assert(!valid(), "base address is loop variant");
3606     return;
3607   }
3608   // unsafe references require misaligned vector access support
3609   if (base->is_top() && !Matcher::misaligned_vectors_ok()) {
3610     assert(!valid(), "unsafe access");
3611     return;
3612   }
3613 
3614   NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.store_depth();)
3615   NOT_PRODUCT(_tracer.ctor_2(adr);)
3616 
3617   int i;
3618   for (i = 0; i < 3; i++) {
3619     NOT_PRODUCT(_tracer.ctor_3(adr, i);)
3620 
3621     if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
3622       assert(!valid(), "too complex");
3623       return;
3624     }
3625     adr = adr->in(AddPNode::Address);
3626     NOT_PRODUCT(_tracer.ctor_4(adr, i);)
3627 
3628     if (base == adr || !adr->is_AddP()) {
3629       NOT_PRODUCT(_tracer.ctor_5(adr, base, i);)
3630       break; // stop looking at addp's
3631     }
3632   }
3633   if (!invariant(adr)) {
3634     assert(!valid(), "adr is loop variant");
3635     return;
3636   }
3637 
3638   if (!base->is_top() && adr != base) {
3639     assert(!valid(), "adr and base differ");
3640     return;
3641   }
3642 
3643   NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.restore_depth();)
3644   NOT_PRODUCT(_tracer.ctor_6(mem);)
3645 
3646   _base = base;
3647   _adr  = adr;
3648   assert(valid(), "Usable");
3649 }
3650 
3651 // Following is used to create a temporary object during
3652 // the pattern match of an address expression.
3653 SWPointer::SWPointer(SWPointer* p) :
3654   _mem(p->_mem), _slp(p->_slp),  _base(NULL),  _adr(NULL),
3655   _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3656   _nstack(p->_nstack), _analyze_only(p->_analyze_only),
3657   _stack_idx(p->_stack_idx)
3658   #ifndef PRODUCT
3659   , _tracer(p->_slp)
3660   #endif
3661 {}
3662 
3663 
3664 bool SWPointer::invariant(Node* n) {
3665   NOT_PRODUCT(Tracer::Depth dd;)
3666   Node *n_c = phase()->get_ctrl(n);
3667   NOT_PRODUCT(_tracer.invariant_1(n, n_c);)
3668   return !lpt()->is_member(phase()->get_loop(n_c));
3669 }
3670 //------------------------scaled_iv_plus_offset--------------------
3671 // Match: k*iv + offset
3672 // where: k is a constant that maybe zero, and
3673 //        offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
3674 bool SWPointer::scaled_iv_plus_offset(Node* n) {
3675   NOT_PRODUCT(Tracer::Depth ddd;)
3676   NOT_PRODUCT(_tracer.scaled_iv_plus_offset_1(n);)
3677 
3678   if (scaled_iv(n)) {
3679     NOT_PRODUCT(_tracer.scaled_iv_plus_offset_2(n);)
3680     return true;
3681   }
3682 
3683   if (offset_plus_k(n)) {
3684     NOT_PRODUCT(_tracer.scaled_iv_plus_offset_3(n);)
3685     return true;
3686   }
3687 
3688   int opc = n->Opcode();
3689   if (opc == Op_AddI) {
3690     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
3691       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_4(n);)
3692       return true;
3693     }
3694     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3695       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_5(n);)
3696       return true;
3697     }
3698   } else if (opc == Op_SubI) {
3699     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
3700       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_6(n);)
3701       return true;
3702     }
3703     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3704       _scale *= -1;
3705       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_7(n);)
3706       return true;
3707     }
3708   }
3709 
3710   NOT_PRODUCT(_tracer.scaled_iv_plus_offset_8(n);)
3711   return false;
3712 }
3713 
3714 //----------------------------scaled_iv------------------------
3715 // Match: k*iv where k is a constant that's not zero
3716 bool SWPointer::scaled_iv(Node* n) {
3717   NOT_PRODUCT(Tracer::Depth ddd;)
3718   NOT_PRODUCT(_tracer.scaled_iv_1(n);)
3719 
3720   if (_scale != 0) { // already found a scale
3721     NOT_PRODUCT(_tracer.scaled_iv_2(n, _scale);)
3722     return false;
3723   }
3724 
3725   if (n == iv()) {
3726     _scale = 1;
3727     NOT_PRODUCT(_tracer.scaled_iv_3(n, _scale);)
3728     return true;
3729   }
3730   if (_analyze_only && (invariant(n) == false)) {
3731     _nstack->push(n, _stack_idx++);
3732   }
3733 
3734   int opc = n->Opcode();
3735   if (opc == Op_MulI) {
3736     if (n->in(1) == iv() && n->in(2)->is_Con()) {
3737       _scale = n->in(2)->get_int();
3738       NOT_PRODUCT(_tracer.scaled_iv_4(n, _scale);)
3739       return true;
3740     } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
3741       _scale = n->in(1)->get_int();
3742       NOT_PRODUCT(_tracer.scaled_iv_5(n, _scale);)
3743       return true;
3744     }
3745   } else if (opc == Op_LShiftI) {
3746     if (n->in(1) == iv() && n->in(2)->is_Con()) {
3747       _scale = 1 << n->in(2)->get_int();
3748       NOT_PRODUCT(_tracer.scaled_iv_6(n, _scale);)
3749       return true;
3750     }
3751   } else if (opc == Op_ConvI2L) {
3752     if (n->in(1)->Opcode() == Op_CastII &&
3753         n->in(1)->as_CastII()->has_range_check()) {
3754       // Skip range check dependent CastII nodes
3755       n = n->in(1);
3756     }
3757     if (scaled_iv_plus_offset(n->in(1))) {
3758       NOT_PRODUCT(_tracer.scaled_iv_7(n);)
3759       return true;
3760     }
3761   } else if (opc == Op_LShiftL) {
3762     if (!has_iv() && _invar == NULL) {
3763       // Need to preserve the current _offset value, so
3764       // create a temporary object for this expression subtree.
3765       // Hacky, so should re-engineer the address pattern match.
3766       NOT_PRODUCT(Tracer::Depth dddd;)
3767       SWPointer tmp(this);
3768       NOT_PRODUCT(_tracer.scaled_iv_8(n, &tmp);)
3769 
3770       if (tmp.scaled_iv_plus_offset(n->in(1))) {
3771         if (tmp._invar == NULL || _slp->do_vector_loop()) {
3772           int mult = 1 << n->in(2)->get_int();
3773           _scale   = tmp._scale  * mult;
3774           _offset += tmp._offset * mult;
3775           NOT_PRODUCT(_tracer.scaled_iv_9(n, _scale, _offset, mult);)
3776           return true;
3777         }
3778       }
3779     }
3780   }
3781   NOT_PRODUCT(_tracer.scaled_iv_10(n);)
3782   return false;
3783 }
3784 
3785 //----------------------------offset_plus_k------------------------
3786 // Match: offset is (k [+/- invariant])
3787 // where k maybe zero and invariant is optional, but not both.
3788 bool SWPointer::offset_plus_k(Node* n, bool negate) {
3789   NOT_PRODUCT(Tracer::Depth ddd;)
3790   NOT_PRODUCT(_tracer.offset_plus_k_1(n);)
3791 
3792   int opc = n->Opcode();
3793   if (opc == Op_ConI) {
3794     _offset += negate ? -(n->get_int()) : n->get_int();
3795     NOT_PRODUCT(_tracer.offset_plus_k_2(n, _offset);)
3796     return true;
3797   } else if (opc == Op_ConL) {
3798     // Okay if value fits into an int
3799     const TypeLong* t = n->find_long_type();
3800     if (t->higher_equal(TypeLong::INT)) {
3801       jlong loff = n->get_long();
3802       jint  off  = (jint)loff;
3803       _offset += negate ? -off : loff;
3804       NOT_PRODUCT(_tracer.offset_plus_k_3(n, _offset);)
3805       return true;
3806     }
3807     NOT_PRODUCT(_tracer.offset_plus_k_4(n);)
3808     return false;
3809   }
3810   if (_invar != NULL) { // already has an invariant
3811     NOT_PRODUCT(_tracer.offset_plus_k_5(n, _invar);)
3812     return false;
3813   }
3814 
3815   if (_analyze_only && (invariant(n) == false)) {
3816     _nstack->push(n, _stack_idx++);
3817   }
3818   if (opc == Op_AddI) {
3819     if (n->in(2)->is_Con() && invariant(n->in(1))) {
3820       _negate_invar = negate;
3821       _invar = n->in(1);
3822       _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3823       NOT_PRODUCT(_tracer.offset_plus_k_6(n, _invar, _negate_invar, _offset);)
3824       return true;
3825     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3826       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3827       _negate_invar = negate;
3828       _invar = n->in(2);
3829       NOT_PRODUCT(_tracer.offset_plus_k_7(n, _invar, _negate_invar, _offset);)
3830       return true;
3831     }
3832   }
3833   if (opc == Op_SubI) {
3834     if (n->in(2)->is_Con() && invariant(n->in(1))) {
3835       _negate_invar = negate;
3836       _invar = n->in(1);
3837       _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3838       NOT_PRODUCT(_tracer.offset_plus_k_8(n, _invar, _negate_invar, _offset);)
3839       return true;
3840     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3841       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3842       _negate_invar = !negate;
3843       _invar = n->in(2);
3844       NOT_PRODUCT(_tracer.offset_plus_k_9(n, _invar, _negate_invar, _offset);)
3845       return true;
3846     }
3847   }
3848   if (invariant(n)) {
3849     if (opc == Op_ConvI2L) {
3850       n = n->in(1);
3851       if (n->Opcode() == Op_CastII &&
3852           n->as_CastII()->has_range_check()) {
3853         // Skip range check dependent CastII nodes
3854         assert(invariant(n), "sanity");
3855         n = n->in(1);
3856       }
3857     }
3858     _negate_invar = negate;
3859     _invar = n;
3860     NOT_PRODUCT(_tracer.offset_plus_k_10(n, _invar, _negate_invar, _offset);)
3861     return true;
3862   }
3863 
3864   NOT_PRODUCT(_tracer.offset_plus_k_11(n);)
3865   return false;
3866 }
3867 
3868 //----------------------------print------------------------
3869 void SWPointer::print() {
3870 #ifndef PRODUCT
3871   tty->print("base: %d  adr: %d  scale: %d  offset: %d  invar: %c%d\n",
3872              _base != NULL ? _base->_idx : 0,
3873              _adr  != NULL ? _adr->_idx  : 0,
3874              _scale, _offset,
3875              _negate_invar?'-':'+',
3876              _invar != NULL ? _invar->_idx : 0);
3877 #endif
3878 }
3879 
3880 //----------------------------tracing------------------------
3881 #ifndef PRODUCT
3882 void SWPointer::Tracer::print_depth() {
3883   for (int ii = 0; ii<_depth; ++ii) tty->print("  ");
3884 }
3885 
3886 void SWPointer::Tracer::ctor_1 (Node* mem) {
3887   if(_slp->is_trace_alignment()) {
3888     print_depth(); tty->print(" %d SWPointer::SWPointer: start alignment analysis", mem->_idx); mem->dump();
3889   }
3890 }
3891 
3892 void SWPointer::Tracer::ctor_2(Node* adr) {
3893   if(_slp->is_trace_alignment()) {
3894     //store_depth();
3895     inc_depth();
3896     print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: ", adr->_idx); adr->dump();
3897     inc_depth();
3898     print_depth(); tty->print(" %d (base) SWPointer::SWPointer: ", adr->in(AddPNode::Base)->_idx); adr->in(AddPNode::Base)->dump();
3899   }
3900 }
3901 
3902 void SWPointer::Tracer::ctor_3(Node* adr, int i) {
3903   if(_slp->is_trace_alignment()) {
3904     inc_depth();
3905     Node* offset = adr->in(AddPNode::Offset);
3906     print_depth(); tty->print(" %d (offset) SWPointer::SWPointer: i = %d: ", offset->_idx, i); offset->dump();
3907   }
3908 }
3909 
3910 void SWPointer::Tracer::ctor_4(Node* adr, int i) {
3911   if(_slp->is_trace_alignment()) {
3912     inc_depth();
3913     print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: i = %d: ", adr->_idx, i); adr->dump();
3914   }
3915 }
3916 
3917 void SWPointer::Tracer::ctor_5(Node* adr, Node* base, int i) {
3918   if(_slp->is_trace_alignment()) {
3919     inc_depth();
3920     if (base == adr) {
3921       print_depth(); tty->print_cr("  \\ %d (adr) == %d (base) SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, base->_idx, i);
3922     } else if (!adr->is_AddP()) {
3923       print_depth(); tty->print_cr("  \\ %d (adr) is NOT Addp SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, i);
3924     }
3925   }
3926 }
3927 
3928 void SWPointer::Tracer::ctor_6(Node* mem) {
3929   if(_slp->is_trace_alignment()) {
3930     //restore_depth();
3931     print_depth(); tty->print_cr(" %d (adr) SWPointer::SWPointer: stop analysis", mem->_idx);
3932   }
3933 }
3934 
3935 void SWPointer::Tracer::invariant_1(Node *n, Node *n_c) {
3936   if (_slp->do_vector_loop() && _slp->is_debug() && _slp->_lpt->is_member(_slp->_phase->get_loop(n_c)) != (int)_slp->in_bb(n)) {
3937     int is_member =  _slp->_lpt->is_member(_slp->_phase->get_loop(n_c));
3938     int in_bb     =  _slp->in_bb(n);
3939     print_depth(); tty->print("  \\ ");  tty->print_cr(" %d SWPointer::invariant  conditions differ: n_c %d", n->_idx, n_c->_idx);
3940     print_depth(); tty->print("  \\ ");  tty->print_cr("is_member %d, in_bb %d", is_member, in_bb);
3941     print_depth(); tty->print("  \\ ");  n->dump();
3942     print_depth(); tty->print("  \\ ");  n_c->dump();
3943   }
3944 }
3945 
3946 void SWPointer::Tracer::scaled_iv_plus_offset_1(Node* n) {
3947   if(_slp->is_trace_alignment()) {
3948     print_depth(); tty->print(" %d SWPointer::scaled_iv_plus_offset testing node: ", n->_idx);
3949     n->dump();
3950   }
3951 }
3952 
3953 void SWPointer::Tracer::scaled_iv_plus_offset_2(Node* n) {
3954   if(_slp->is_trace_alignment()) {
3955     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
3956   }
3957 }
3958 
3959 void SWPointer::Tracer::scaled_iv_plus_offset_3(Node* n) {
3960   if(_slp->is_trace_alignment()) {
3961     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
3962   }
3963 }
3964 
3965 void SWPointer::Tracer::scaled_iv_plus_offset_4(Node* n) {
3966   if(_slp->is_trace_alignment()) {
3967     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
3968     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
3969     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
3970   }
3971 }
3972 
3973 void SWPointer::Tracer::scaled_iv_plus_offset_5(Node* n) {
3974   if(_slp->is_trace_alignment()) {
3975     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
3976     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
3977     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
3978   }
3979 }
3980 
3981 void SWPointer::Tracer::scaled_iv_plus_offset_6(Node* n) {
3982   if(_slp->is_trace_alignment()) {
3983     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
3984     print_depth(); tty->print("  \\  %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
3985     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
3986   }
3987 }
3988 
3989 void SWPointer::Tracer::scaled_iv_plus_offset_7(Node* n) {
3990   if(_slp->is_trace_alignment()) {
3991     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
3992     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
3993     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
3994   }
3995 }
3996 
3997 void SWPointer::Tracer::scaled_iv_plus_offset_8(Node* n) {
3998   if(_slp->is_trace_alignment()) {
3999     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: FAILED", n->_idx);
4000   }
4001 }
4002 
4003 void SWPointer::Tracer::scaled_iv_1(Node* n) {
4004   if(_slp->is_trace_alignment()) {
4005     print_depth(); tty->print(" %d SWPointer::scaled_iv: testing node: ", n->_idx); n->dump();
4006   }
4007 }
4008 
4009 void SWPointer::Tracer::scaled_iv_2(Node* n, int scale) {
4010   if(_slp->is_trace_alignment()) {
4011     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED since another _scale has been detected before", n->_idx);
4012     print_depth(); tty->print_cr("  \\ SWPointer::scaled_iv: _scale (%d) != 0", scale);
4013   }
4014 }
4015 
4016 void SWPointer::Tracer::scaled_iv_3(Node* n, int scale) {
4017   if(_slp->is_trace_alignment()) {
4018     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: is iv, setting _scale = %d", n->_idx, scale);
4019   }
4020 }
4021 
4022 void SWPointer::Tracer::scaled_iv_4(Node* n, int scale) {
4023   if(_slp->is_trace_alignment()) {
4024     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
4025     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
4026     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4027   }
4028 }
4029 
4030 void SWPointer::Tracer::scaled_iv_5(Node* n, int scale) {
4031   if(_slp->is_trace_alignment()) {
4032     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
4033     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is iv: ", n->in(2)->_idx); n->in(2)->dump();
4034     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
4035   }
4036 }
4037 
4038 void SWPointer::Tracer::scaled_iv_6(Node* n, int scale) {
4039   if(_slp->is_trace_alignment()) {
4040     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftI PASSED, setting _scale = %d", n->_idx, scale);
4041     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
4042     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4043   }
4044 }
4045 
4046 void SWPointer::Tracer::scaled_iv_7(Node* n) {
4047   if(_slp->is_trace_alignment()) {
4048     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_ConvI2L PASSED", n->_idx);
4049     print_depth(); tty->print_cr("  \\ SWPointer::scaled_iv: in(1) %d is scaled_iv_plus_offset: ", n->in(1)->_idx);
4050     inc_depth(); inc_depth();
4051     print_depth(); n->in(1)->dump();
4052     dec_depth(); dec_depth();
4053   }
4054 }
4055 
4056 void SWPointer::Tracer::scaled_iv_8(Node* n, SWPointer* tmp) {
4057   if(_slp->is_trace_alignment()) {
4058     print_depth(); tty->print(" %d SWPointer::scaled_iv: Op_LShiftL, creating tmp SWPointer: ", n->_idx); tmp->print();
4059   }
4060 }
4061 
4062 void SWPointer::Tracer::scaled_iv_9(Node* n, int scale, int _offset, int mult) {
4063   if(_slp->is_trace_alignment()) {
4064     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftL PASSED, setting _scale = %d, _offset = %d", n->_idx, scale, _offset);
4065     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",
4066     n->in(1)->_idx, n->in(2)->_idx, mult, scale, _offset);
4067     inc_depth(); inc_depth();
4068     print_depth(); n->in(1)->dump();
4069     print_depth(); n->in(2)->dump();
4070     dec_depth(); dec_depth();
4071   }
4072 }
4073 
4074 void SWPointer::Tracer::scaled_iv_10(Node* n) {
4075   if(_slp->is_trace_alignment()) {
4076     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED", n->_idx);
4077   }
4078 }
4079 
4080 void SWPointer::Tracer::offset_plus_k_1(Node* n) {
4081   if(_slp->is_trace_alignment()) {
4082     print_depth(); tty->print(" %d SWPointer::offset_plus_k: testing node: ", n->_idx); n->dump();
4083   }
4084 }
4085 
4086 void SWPointer::Tracer::offset_plus_k_2(Node* n, int _offset) {
4087   if(_slp->is_trace_alignment()) {
4088     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConI PASSED, setting _offset = %d", n->_idx, _offset);
4089   }
4090 }
4091 
4092 void SWPointer::Tracer::offset_plus_k_3(Node* n, int _offset) {
4093   if(_slp->is_trace_alignment()) {
4094     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConL PASSED, setting _offset = %d", n->_idx, _offset);
4095   }
4096 }
4097 
4098 void SWPointer::Tracer::offset_plus_k_4(Node* n) {
4099   if(_slp->is_trace_alignment()) {
4100     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
4101     print_depth(); tty->print_cr("  \\ " JLONG_FORMAT " SWPointer::offset_plus_k: Op_ConL FAILED, k is too big", n->get_long());
4102   }
4103 }
4104 
4105 void SWPointer::Tracer::offset_plus_k_5(Node* n, Node* _invar) {
4106   if(_slp->is_trace_alignment()) {
4107     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED since another invariant has been detected before", n->_idx);
4108     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: _invar != NULL: ", _invar->_idx); _invar->dump();
4109   }
4110 }
4111 
4112 void SWPointer::Tracer::offset_plus_k_6(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4113   if(_slp->is_trace_alignment()) {
4114     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
4115     n->_idx, _negate_invar, _invar->_idx, _offset);
4116     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4117     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
4118   }
4119 }
4120 
4121 void SWPointer::Tracer::offset_plus_k_7(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4122   if(_slp->is_trace_alignment()) {
4123     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
4124     n->_idx, _negate_invar, _invar->_idx, _offset);
4125     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
4126     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
4127   }
4128 }
4129 
4130 void SWPointer::Tracer::offset_plus_k_8(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4131   if(_slp->is_trace_alignment()) {
4132     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_SubI is PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
4133     n->_idx, _negate_invar, _invar->_idx, _offset);
4134     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4135     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
4136   }
4137 }
4138 
4139 void SWPointer::Tracer::offset_plus_k_9(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4140   if(_slp->is_trace_alignment()) {
4141     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);
4142     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
4143     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
4144   }
4145 }
4146 
4147 void SWPointer::Tracer::offset_plus_k_10(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4148   if(_slp->is_trace_alignment()) {
4149     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);
4150     print_depth(); tty->print_cr("  \\ %d SWPointer::offset_plus_k: is invariant", n->_idx);
4151   }
4152 }
4153 
4154 void SWPointer::Tracer::offset_plus_k_11(Node* n) {
4155   if(_slp->is_trace_alignment()) {
4156     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
4157   }
4158 }
4159 
4160 #endif
4161 // ========================= OrderedPair =====================
4162 
4163 const OrderedPair OrderedPair::initial;
4164 
4165 // ========================= SWNodeInfo =====================
4166 
4167 const SWNodeInfo SWNodeInfo::initial;
4168 
4169 
4170 // ============================ DepGraph ===========================
4171 
4172 //------------------------------make_node---------------------------
4173 // Make a new dependence graph node for an ideal node.
4174 DepMem* DepGraph::make_node(Node* node) {
4175   DepMem* m = new (_arena) DepMem(node);
4176   if (node != NULL) {
4177     assert(_map.at_grow(node->_idx) == NULL, "one init only");
4178     _map.at_put_grow(node->_idx, m);
4179   }
4180   return m;
4181 }
4182 
4183 //------------------------------make_edge---------------------------
4184 // Make a new dependence graph edge from dpred -> dsucc
4185 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
4186   DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
4187   dpred->set_out_head(e);
4188   dsucc->set_in_head(e);
4189   return e;
4190 }
4191 
4192 // ========================== DepMem ========================
4193 
4194 //------------------------------in_cnt---------------------------
4195 int DepMem::in_cnt() {
4196   int ct = 0;
4197   for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
4198   return ct;
4199 }
4200 
4201 //------------------------------out_cnt---------------------------
4202 int DepMem::out_cnt() {
4203   int ct = 0;
4204   for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
4205   return ct;
4206 }
4207 
4208 //------------------------------print-----------------------------
4209 void DepMem::print() {
4210 #ifndef PRODUCT
4211   tty->print("  DepNode %d (", _node->_idx);
4212   for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
4213     Node* pred = p->pred()->node();
4214     tty->print(" %d", pred != NULL ? pred->_idx : 0);
4215   }
4216   tty->print(") [");
4217   for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
4218     Node* succ = s->succ()->node();
4219     tty->print(" %d", succ != NULL ? succ->_idx : 0);
4220   }
4221   tty->print_cr(" ]");
4222 #endif
4223 }
4224 
4225 // =========================== DepEdge =========================
4226 
4227 //------------------------------DepPreds---------------------------
4228 void DepEdge::print() {
4229 #ifndef PRODUCT
4230   tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
4231 #endif
4232 }
4233 
4234 // =========================== DepPreds =========================
4235 // Iterator over predecessor edges in the dependence graph.
4236 
4237 //------------------------------DepPreds---------------------------
4238 DepPreds::DepPreds(Node* n, DepGraph& dg) {
4239   _n = n;
4240   _done = false;
4241   if (_n->is_Store() || _n->is_Load()) {
4242     _next_idx = MemNode::Address;
4243     _end_idx  = n->req();
4244     _dep_next = dg.dep(_n)->in_head();
4245   } else if (_n->is_Mem()) {
4246     _next_idx = 0;
4247     _end_idx  = 0;
4248     _dep_next = dg.dep(_n)->in_head();
4249   } else {
4250     _next_idx = 1;
4251     _end_idx  = _n->req();
4252     _dep_next = NULL;
4253   }
4254   next();
4255 }
4256 
4257 //------------------------------next---------------------------
4258 void DepPreds::next() {
4259   if (_dep_next != NULL) {
4260     _current  = _dep_next->pred()->node();
4261     _dep_next = _dep_next->next_in();
4262   } else if (_next_idx < _end_idx) {
4263     _current  = _n->in(_next_idx++);
4264   } else {
4265     _done = true;
4266   }
4267 }
4268 
4269 // =========================== DepSuccs =========================
4270 // Iterator over successor edges in the dependence graph.
4271 
4272 //------------------------------DepSuccs---------------------------
4273 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
4274   _n = n;
4275   _done = false;
4276   if (_n->is_Load()) {
4277     _next_idx = 0;
4278     _end_idx  = _n->outcnt();
4279     _dep_next = dg.dep(_n)->out_head();
4280   } else if (_n->is_Mem() || (_n->is_Phi() && _n->bottom_type() == Type::MEMORY)) {
4281     _next_idx = 0;
4282     _end_idx  = 0;
4283     _dep_next = dg.dep(_n)->out_head();
4284   } else {
4285     _next_idx = 0;
4286     _end_idx  = _n->outcnt();
4287     _dep_next = NULL;
4288   }
4289   next();
4290 }
4291 
4292 //-------------------------------next---------------------------
4293 void DepSuccs::next() {
4294   if (_dep_next != NULL) {
4295     _current  = _dep_next->succ()->node();
4296     _dep_next = _dep_next->next_out();
4297   } else if (_next_idx < _end_idx) {
4298     _current  = _n->raw_out(_next_idx++);
4299   } else {
4300     _done = true;
4301   }
4302 }
4303 
4304 //
4305 // --------------------------------- vectorization/simd -----------------------------------
4306 //
4307 bool SuperWord::same_origin_idx(Node* a, Node* b) const {
4308   return a != NULL && b != NULL && _clone_map.same_idx(a->_idx, b->_idx);
4309 }
4310 bool SuperWord::same_generation(Node* a, Node* b) const {
4311   return a != NULL && b != NULL && _clone_map.same_gen(a->_idx, b->_idx);
4312 }
4313 
4314 Node*  SuperWord::find_phi_for_mem_dep(LoadNode* ld) {
4315   assert(in_bb(ld), "must be in block");
4316   if (_clone_map.gen(ld->_idx) == _ii_first) {
4317 #ifndef PRODUCT
4318     if (_vector_loop_debug) {
4319       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(ld->_idx)=%d",
4320         _clone_map.gen(ld->_idx));
4321     }
4322 #endif
4323     return NULL; //we think that any ld in the first gen being vectorizable
4324   }
4325 
4326   Node* mem = ld->in(MemNode::Memory);
4327   if (mem->outcnt() <= 1) {
4328     // we don't want to remove the only edge from mem node to load
4329 #ifndef PRODUCT
4330     if (_vector_loop_debug) {
4331       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",
4332         mem->_idx, ld->_idx);
4333       ld->dump();
4334       mem->dump();
4335     }
4336 #endif
4337     return NULL;
4338   }
4339   if (!in_bb(mem) || same_generation(mem, ld)) {
4340 #ifndef PRODUCT
4341     if (_vector_loop_debug) {
4342       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(mem->_idx)=%d",
4343         _clone_map.gen(mem->_idx));
4344     }
4345 #endif
4346     return NULL; // does not depend on loop volatile node or depends on the same generation
4347   }
4348 
4349   //otherwise first node should depend on mem-phi
4350   Node* first = first_node(ld);
4351   assert(first->is_Load(), "must be Load");
4352   Node* phi = first->as_Load()->in(MemNode::Memory);
4353   if (!phi->is_Phi() || phi->bottom_type() != Type::MEMORY) {
4354 #ifndef PRODUCT
4355     if (_vector_loop_debug) {
4356       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");
4357       ld->dump();
4358       first->dump();
4359     }
4360 #endif
4361     return NULL;
4362   }
4363 
4364   Node* tail = 0;
4365   for (int m = 0; m < _mem_slice_head.length(); m++) {
4366     if (_mem_slice_head.at(m) == phi) {
4367       tail = _mem_slice_tail.at(m);
4368     }
4369   }
4370   if (tail == 0) { //test that found phi is in the list  _mem_slice_head
4371 #ifndef PRODUCT
4372     if (_vector_loop_debug) {
4373       tty->print_cr("SuperWord::find_phi_for_mem_dep load %d is not vectorizable node, its phi %d is not _mem_slice_head",
4374         ld->_idx, phi->_idx);
4375       ld->dump();
4376       phi->dump();
4377     }
4378 #endif
4379     return NULL;
4380   }
4381 
4382   // now all conditions are met
4383   return phi;
4384 }
4385 
4386 Node* SuperWord::first_node(Node* nd) {
4387   for (int ii = 0; ii < _iteration_first.length(); ii++) {
4388     Node* nnn = _iteration_first.at(ii);
4389     if (same_origin_idx(nnn, nd)) {
4390 #ifndef PRODUCT
4391       if (_vector_loop_debug) {
4392         tty->print_cr("SuperWord::first_node: %d is the first iteration node for %d (_clone_map.idx(nnn->_idx) = %d)",
4393           nnn->_idx, nd->_idx, _clone_map.idx(nnn->_idx));
4394       }
4395 #endif
4396       return nnn;
4397     }
4398   }
4399 
4400 #ifndef PRODUCT
4401   if (_vector_loop_debug) {
4402     tty->print_cr("SuperWord::first_node: did not find first iteration node for %d (_clone_map.idx(nd->_idx)=%d)",
4403       nd->_idx, _clone_map.idx(nd->_idx));
4404   }
4405 #endif
4406   return 0;
4407 }
4408 
4409 Node* SuperWord::last_node(Node* nd) {
4410   for (int ii = 0; ii < _iteration_last.length(); ii++) {
4411     Node* nnn = _iteration_last.at(ii);
4412     if (same_origin_idx(nnn, nd)) {
4413 #ifndef PRODUCT
4414       if (_vector_loop_debug) {
4415         tty->print_cr("SuperWord::last_node _clone_map.idx(nnn->_idx)=%d, _clone_map.idx(nd->_idx)=%d",
4416           _clone_map.idx(nnn->_idx), _clone_map.idx(nd->_idx));
4417       }
4418 #endif
4419       return nnn;
4420     }
4421   }
4422   return 0;
4423 }
4424 
4425 int SuperWord::mark_generations() {
4426   Node *ii_err = NULL, *tail_err = NULL;
4427   for (int i = 0; i < _mem_slice_head.length(); i++) {
4428     Node* phi  = _mem_slice_head.at(i);
4429     assert(phi->is_Phi(), "must be phi");
4430 
4431     Node* tail = _mem_slice_tail.at(i);
4432     if (_ii_last == -1) {
4433       tail_err = tail;
4434       _ii_last = _clone_map.gen(tail->_idx);
4435     }
4436     else if (_ii_last != _clone_map.gen(tail->_idx)) {
4437 #ifndef PRODUCT
4438       if (TraceSuperWord && Verbose) {
4439         tty->print_cr("SuperWord::mark_generations _ii_last error - found different generations in two tail nodes ");
4440         tail->dump();
4441         tail_err->dump();
4442       }
4443 #endif
4444       return -1;
4445     }
4446 
4447     // find first iteration in the loop
4448     for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
4449       Node* ii = phi->fast_out(i);
4450       if (in_bb(ii) && ii->is_Store()) { // we speculate that normally Stores of one and one only generation have deps from mem phi
4451         if (_ii_first == -1) {
4452           ii_err = ii;
4453           _ii_first = _clone_map.gen(ii->_idx);
4454         } else if (_ii_first != _clone_map.gen(ii->_idx)) {
4455 #ifndef PRODUCT
4456           if (TraceSuperWord && Verbose) {
4457             tty->print_cr("SuperWord::mark_generations: _ii_first was found before and not equal to one in this node (%d)", _ii_first);
4458             ii->dump();
4459             if (ii_err!= 0) {
4460               ii_err->dump();
4461             }
4462           }
4463 #endif
4464           return -1; // this phi has Stores from different generations of unroll and cannot be simd/vectorized
4465         }
4466       }
4467     }//for (DUIterator_Fast imax,
4468   }//for (int i...
4469 
4470   if (_ii_first == -1 || _ii_last == -1) {
4471     if (TraceSuperWord && Verbose) {
4472       tty->print_cr("SuperWord::mark_generations unknown error, something vent wrong");
4473     }
4474     return -1; // something vent wrong
4475   }
4476   // collect nodes in the first and last generations
4477   assert(_iteration_first.length() == 0, "_iteration_first must be empty");
4478   assert(_iteration_last.length() == 0, "_iteration_last must be empty");
4479   for (int j = 0; j < _block.length(); j++) {
4480     Node* n = _block.at(j);
4481     node_idx_t gen = _clone_map.gen(n->_idx);
4482     if ((signed)gen == _ii_first) {
4483       _iteration_first.push(n);
4484     } else if ((signed)gen == _ii_last) {
4485       _iteration_last.push(n);
4486     }
4487   }
4488 
4489   // building order of iterations
4490   if (_ii_order.length() == 0 && ii_err != 0) {
4491     assert(in_bb(ii_err) && ii_err->is_Store(), "should be Store in bb");
4492     Node* nd = ii_err;
4493     while(_clone_map.gen(nd->_idx) != _ii_last) {
4494       _ii_order.push(_clone_map.gen(nd->_idx));
4495       bool found = false;
4496       for (DUIterator_Fast imax, i = nd->fast_outs(imax); i < imax; i++) {
4497         Node* use = nd->fast_out(i);
4498         if (same_origin_idx(use, nd) && use->as_Store()->in(MemNode::Memory) == nd) {
4499           found = true;
4500           nd = use;
4501           break;
4502         }
4503       }//for
4504 
4505       if (found == false) {
4506         if (TraceSuperWord && Verbose) {
4507           tty->print_cr("SuperWord::mark_generations: Cannot build order of iterations - no dependent Store for %d", nd->_idx);
4508         }
4509         _ii_order.clear();
4510         return -1;
4511       }
4512     } //while
4513     _ii_order.push(_clone_map.gen(nd->_idx));
4514   }
4515 
4516 #ifndef PRODUCT
4517   if (_vector_loop_debug) {
4518     tty->print_cr("SuperWord::mark_generations");
4519     tty->print_cr("First generation (%d) nodes:", _ii_first);
4520     for (int ii = 0; ii < _iteration_first.length(); ii++)  _iteration_first.at(ii)->dump();
4521     tty->print_cr("Last generation (%d) nodes:", _ii_last);
4522     for (int ii = 0; ii < _iteration_last.length(); ii++)  _iteration_last.at(ii)->dump();
4523     tty->print_cr(" ");
4524 
4525     tty->print("SuperWord::List of generations: ");
4526     for (int jj = 0; jj < _ii_order.length(); ++jj) {
4527       tty->print("%d:%d ", jj, _ii_order.at(jj));
4528     }
4529     tty->print_cr(" ");
4530   }
4531 #endif
4532 
4533   return _ii_first;
4534 }
4535 
4536 bool SuperWord::fix_commutative_inputs(Node* gold, Node* fix) {
4537   assert(gold->is_Add() && fix->is_Add() || gold->is_Mul() && fix->is_Mul(), "should be only Add or Mul nodes");
4538   assert(same_origin_idx(gold, fix), "should be clones of the same node");
4539   Node* gin1 = gold->in(1);
4540   Node* gin2 = gold->in(2);
4541   Node* fin1 = fix->in(1);
4542   Node* fin2 = fix->in(2);
4543   bool swapped = false;
4544 
4545   if (in_bb(gin1) && in_bb(gin2) && in_bb(fin1) && in_bb(fin1)) {
4546     if (same_origin_idx(gin1, fin1) &&
4547         same_origin_idx(gin2, fin2)) {
4548       return true; // nothing to fix
4549     }
4550     if (same_origin_idx(gin1, fin2) &&
4551         same_origin_idx(gin2, fin1)) {
4552       fix->swap_edges(1, 2);
4553       swapped = true;
4554     }
4555   }
4556   // at least one input comes from outside of bb
4557   if (gin1->_idx == fin1->_idx)  {
4558     return true; // nothing to fix
4559   }
4560   if (!swapped && (gin1->_idx == fin2->_idx || gin2->_idx == fin1->_idx))  { //swapping is expensive, check condition first
4561     fix->swap_edges(1, 2);
4562     swapped = true;
4563   }
4564 
4565   if (swapped) {
4566 #ifndef PRODUCT
4567     if (_vector_loop_debug) {
4568       tty->print_cr("SuperWord::fix_commutative_inputs: fixed node %d", fix->_idx);
4569     }
4570 #endif
4571     return true;
4572   }
4573 
4574   if (TraceSuperWord && Verbose) {
4575     tty->print_cr("SuperWord::fix_commutative_inputs: cannot fix node %d", fix->_idx);
4576   }
4577 
4578   return false;
4579 }
4580 
4581 bool SuperWord::pack_parallel() {
4582 #ifndef PRODUCT
4583   if (_vector_loop_debug) {
4584     tty->print_cr("SuperWord::pack_parallel: START");
4585   }
4586 #endif
4587 
4588   _packset.clear();
4589 
4590   for (int ii = 0; ii < _iteration_first.length(); ii++) {
4591     Node* nd = _iteration_first.at(ii);
4592     if (in_bb(nd) && (nd->is_Load() || nd->is_Store() || nd->is_Add() || nd->is_Mul())) {
4593       Node_List* pk = new Node_List();
4594       pk->push(nd);
4595       for (int gen = 1; gen < _ii_order.length(); ++gen) {
4596         for (int kk = 0; kk < _block.length(); kk++) {
4597           Node* clone = _block.at(kk);
4598           if (same_origin_idx(clone, nd) &&
4599               _clone_map.gen(clone->_idx) == _ii_order.at(gen)) {
4600             if (nd->is_Add() || nd->is_Mul()) {
4601               fix_commutative_inputs(nd, clone);
4602             }
4603             pk->push(clone);
4604             if (pk->size() == 4) {
4605               _packset.append(pk);
4606 #ifndef PRODUCT
4607               if (_vector_loop_debug) {
4608                 tty->print_cr("SuperWord::pack_parallel: added pack ");
4609                 pk->dump();
4610               }
4611 #endif
4612               if (_clone_map.gen(clone->_idx) != _ii_last) {
4613                 pk = new Node_List();
4614               }
4615             }
4616             break;
4617           }
4618         }
4619       }//for
4620     }//if
4621   }//for
4622 
4623 #ifndef PRODUCT
4624   if (_vector_loop_debug) {
4625     tty->print_cr("SuperWord::pack_parallel: END");
4626   }
4627 #endif
4628 
4629   return true;
4630 }
4631 
4632 bool SuperWord::hoist_loads_in_graph() {
4633   GrowableArray<Node*> loads;
4634 
4635 #ifndef PRODUCT
4636   if (_vector_loop_debug) {
4637     tty->print_cr("SuperWord::hoist_loads_in_graph: total number _mem_slice_head.length() = %d", _mem_slice_head.length());
4638   }
4639 #endif
4640 
4641   for (int i = 0; i < _mem_slice_head.length(); i++) {
4642     Node* n = _mem_slice_head.at(i);
4643     if ( !in_bb(n) || !n->is_Phi() || n->bottom_type() != Type::MEMORY) {
4644       if (TraceSuperWord && Verbose) {
4645         tty->print_cr("SuperWord::hoist_loads_in_graph: skipping unexpected node n=%d", n->_idx);
4646       }
4647       continue;
4648     }
4649 
4650 #ifndef PRODUCT
4651     if (_vector_loop_debug) {
4652       tty->print_cr("SuperWord::hoist_loads_in_graph: processing phi %d  = _mem_slice_head.at(%d);", n->_idx, i);
4653     }
4654 #endif
4655 
4656     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4657       Node* ld = n->fast_out(i);
4658       if (ld->is_Load() && ld->as_Load()->in(MemNode::Memory) == n && in_bb(ld)) {
4659         for (int i = 0; i < _block.length(); i++) {
4660           Node* ld2 = _block.at(i);
4661           if (ld2->is_Load() && same_origin_idx(ld, ld2) &&
4662               !same_generation(ld, ld2)) { // <= do not collect the first generation ld
4663 #ifndef PRODUCT
4664             if (_vector_loop_debug) {
4665               tty->print_cr("SuperWord::hoist_loads_in_graph: will try to hoist load ld2->_idx=%d, cloned from %d (ld->_idx=%d)",
4666                 ld2->_idx, _clone_map.idx(ld->_idx), ld->_idx);
4667             }
4668 #endif
4669             // could not do on-the-fly, since iterator is immutable
4670             loads.push(ld2);
4671           }
4672         }// for
4673       }//if
4674     }//for (DUIterator_Fast imax,
4675   }//for (int i = 0; i
4676 
4677   for (int i = 0; i < loads.length(); i++) {
4678     LoadNode* ld = loads.at(i)->as_Load();
4679     Node* phi = find_phi_for_mem_dep(ld);
4680     if (phi != NULL) {
4681 #ifndef PRODUCT
4682       if (_vector_loop_debug) {
4683         tty->print_cr("SuperWord::hoist_loads_in_graph replacing MemNode::Memory(%d) edge in %d with one from %d",
4684           MemNode::Memory, ld->_idx, phi->_idx);
4685       }
4686 #endif
4687       _igvn.replace_input_of(ld, MemNode::Memory, phi);
4688     }
4689   }//for
4690 
4691   restart(); // invalidate all basic structures, since we rebuilt the graph
4692 
4693   if (TraceSuperWord && Verbose) {
4694     tty->print_cr("\nSuperWord::hoist_loads_in_graph() the graph was rebuilt, all structures invalidated and need rebuild");
4695   }
4696 
4697   return true;
4698 }