1 /*
   2  * Copyright (c) 2007, 2017, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  */
  23 
  24 #include "precompiled.hpp"
  25 #include "compiler/compileLog.hpp"
  26 #include "libadt/vectset.hpp"
  27 #include "memory/allocation.inline.hpp"
  28 #include "memory/resourceArea.hpp"
  29 #include "opto/addnode.hpp"
  30 #include "opto/callnode.hpp"
  31 #include "opto/castnode.hpp"
  32 #include "opto/convertnode.hpp"
  33 #include "opto/divnode.hpp"
  34 #include "opto/matcher.hpp"
  35 #include "opto/memnode.hpp"
  36 #include "opto/mulnode.hpp"
  37 #include "opto/opcodes.hpp"
  38 #include "opto/opaquenode.hpp"
  39 #include "opto/superword.hpp"
  40 #include "opto/vectornode.hpp"
  41 #include "opto/movenode.hpp"
  42 
  43 //
  44 //                  S U P E R W O R D   T R A N S F O R M
  45 //=============================================================================
  46 
  47 //------------------------------SuperWord---------------------------
  48 SuperWord::SuperWord(PhaseIdealLoop* phase) :
  49   _phase(phase),
  50   _igvn(phase->_igvn),
  51   _arena(phase->C->comp_arena()),
  52   _packset(arena(), 8,  0, NULL),         // packs for the current block
  53   _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
  54   _block(arena(), 8,  0, NULL),           // nodes in current block
  55   _post_block(arena(), 8, 0, NULL),       // nodes common to current block which are marked as post loop vectorizable
  56   _data_entry(arena(), 8,  0, NULL),      // nodes with all inputs from outside
  57   _mem_slice_head(arena(), 8,  0, NULL),  // memory slice heads
  58   _mem_slice_tail(arena(), 8,  0, NULL),  // memory slice tails
  59   _node_info(arena(), 8,  0, SWNodeInfo::initial), // info needed per node
  60   _clone_map(phase->C->clone_map()),      // map of nodes created in cloning
  61   _cmovev_kit(_arena, this),              // map to facilitate CMoveVD creation
  62   _align_to_ref(NULL),                    // memory reference to align vectors to
  63   _disjoint_ptrs(arena(), 8,  0, OrderedPair::initial), // runtime disambiguated pointer pairs
  64   _dg(_arena),                            // dependence graph
  65   _visited(arena()),                      // visited node set
  66   _post_visited(arena()),                 // post visited node set
  67   _n_idx_list(arena(), 8),                // scratch list of (node,index) pairs
  68   _stk(arena(), 8, 0, NULL),              // scratch stack of nodes
  69   _nlist(arena(), 8, 0, NULL),            // scratch list of nodes
  70   _lpt(NULL),                             // loop tree node
  71   _lp(NULL),                              // LoopNode
  72   _bb(NULL),                              // basic block
  73   _iv(NULL),                              // induction var
  74   _race_possible(false),                  // cases where SDMU is true
  75   _early_return(true),                    // analysis evaluations routine
  76   _num_work_vecs(0),                      // amount of vector work we have
  77   _num_reductions(0),                     // amount of reduction work we have
  78   _do_vector_loop(phase->C->do_vector_loop()),  // whether to do vectorization/simd style
  79   _do_reserve_copy(DoReserveCopyInSuperWord),
  80   _ii_first(-1),                          // first loop generation index - only if do_vector_loop()
  81   _ii_last(-1),                           // last loop generation index - only if do_vector_loop()
  82   _ii_order(arena(), 8, 0, 0)
  83 {
  84 #ifndef PRODUCT
  85   _vector_loop_debug = 0;
  86   if (_phase->C->method() != NULL) {
  87     _vector_loop_debug = phase->C->directive()->VectorizeDebugOption;
  88   }
  89 
  90 #endif
  91 }
  92 
  93 //------------------------------transform_loop---------------------------
  94 void SuperWord::transform_loop(IdealLoopTree* lpt, bool do_optimization) {
  95   assert(UseSuperWord, "should be");
  96   // Do vectors exist on this architecture?
  97   if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return;
  98 
  99   assert(lpt->_head->is_CountedLoop(), "must be");
 100   CountedLoopNode *cl = lpt->_head->as_CountedLoop();
 101 
 102   if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
 103 
 104   bool post_loop_allowed = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
 105   if (post_loop_allowed) {
 106     if (cl->is_reduction_loop()) return; // no predication mapping
 107     Node *limit = cl->limit();
 108     if (limit->is_Con()) return; // non constant limits only
 109     // Now check the limit for expressions we do not handle
 110     if (limit->is_Add()) {
 111       Node *in2 = limit->in(2);
 112       if (in2->is_Con()) {
 113         int val = in2->get_int();
 114         // should not try to program these cases
 115         if (val < 0) return;
 116       }
 117     }
 118   }
 119 
 120   // skip any loop that has not been assigned max unroll by analysis
 121   if (do_optimization) {
 122     if (SuperWordLoopUnrollAnalysis && cl->slp_max_unroll() == 0) return;
 123   }
 124 
 125   // Check for no control flow in body (other than exit)
 126   Node *cl_exit = cl->loopexit();
 127   if (cl->is_main_loop() && (cl_exit->in(0) != lpt->_head)) {
 128     #ifndef PRODUCT
 129       if (TraceSuperWord) {
 130         tty->print_cr("SuperWord::transform_loop: loop too complicated, cl_exit->in(0) != lpt->_head");
 131         tty->print("cl_exit %d", cl_exit->_idx); cl_exit->dump();
 132         tty->print("cl_exit->in(0) %d", cl_exit->in(0)->_idx); cl_exit->in(0)->dump();
 133         tty->print("lpt->_head %d", lpt->_head->_idx); lpt->_head->dump();
 134         lpt->dump_head();
 135       }
 136     #endif
 137     return;
 138   }
 139 
 140   // Make sure the are no extra control users of the loop backedge
 141   if (cl->back_control()->outcnt() != 1) {
 142     return;
 143   }
 144 
 145   // Skip any loops already optimized by slp
 146   if (cl->is_vectorized_loop()) return;
 147 
 148   if (cl->do_unroll_only()) return;
 149 
 150   if (cl->is_main_loop()) {
 151     // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
 152     CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
 153     if (pre_end == NULL) return;
 154     Node *pre_opaq1 = pre_end->limit();
 155     if (pre_opaq1->Opcode() != Op_Opaque1) return;
 156   }
 157 
 158   init(); // initialize data structures
 159 
 160   set_lpt(lpt);
 161   set_lp(cl);
 162 
 163   // For now, define one block which is the entire loop body
 164   set_bb(cl);
 165 
 166   if (do_optimization) {
 167     assert(_packset.length() == 0, "packset must be empty");
 168     SLP_extract();
 169     if (PostLoopMultiversioning && Matcher::has_predicated_vectors()) {
 170       if (cl->is_vectorized_loop() && cl->is_main_loop() && !cl->is_reduction_loop()) {
 171         IdealLoopTree *lpt_next = lpt->_next;
 172         CountedLoopNode *cl_next = lpt_next->_head->as_CountedLoop();
 173         _phase->has_range_checks(lpt_next);
 174         if (cl_next->is_post_loop() && !cl_next->range_checks_present()) {
 175           if (!cl_next->is_vectorized_loop()) {
 176             int slp_max_unroll_factor = cl->slp_max_unroll();
 177             cl_next->set_slp_max_unroll(slp_max_unroll_factor);
 178           }
 179         }
 180       }
 181     }
 182   }
 183 }
 184 
 185 //------------------------------early unrolling analysis------------------------------
 186 void SuperWord::unrolling_analysis(int &local_loop_unroll_factor) {
 187   bool is_slp = true;
 188   ResourceMark rm;
 189   size_t ignored_size = lpt()->_body.size();
 190   int *ignored_loop_nodes = NEW_RESOURCE_ARRAY(int, ignored_size);
 191   Node_Stack nstack((int)ignored_size);
 192   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
 193   Node *cl_exit = cl->loopexit();
 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                 }
 380               }
 381             }
 382           }
 383         }
 384         // We only process post loops on predicated targets where we want to
 385         // mask map the loop to a single iteration
 386         if (post_loop_allowed) {
 387           _post_block.at_put_grow(rpo_idx++, n);
 388         }
 389       }
 390     }
 391     if (is_slp) {
 392       local_loop_unroll_factor = max_vector;
 393       cl->mark_passed_slp();
 394     }
 395     cl->mark_was_slp();
 396     if (cl->is_main_loop()) {
 397       cl->set_slp_max_unroll(local_loop_unroll_factor);
 398     } else if (post_loop_allowed) {
 399       if (!small_basic_type) {
 400         // avoid replication context for small basic types in programmable masked loops
 401         cl->set_slp_max_unroll(local_loop_unroll_factor);
 402       }
 403     }
 404   }
 405 }
 406 
 407 //------------------------------SLP_extract---------------------------
 408 // Extract the superword level parallelism
 409 //
 410 // 1) A reverse post-order of nodes in the block is constructed.  By scanning
 411 //    this list from first to last, all definitions are visited before their uses.
 412 //
 413 // 2) A point-to-point dependence graph is constructed between memory references.
 414 //    This simplies the upcoming "independence" checker.
 415 //
 416 // 3) The maximum depth in the node graph from the beginning of the block
 417 //    to each node is computed.  This is used to prune the graph search
 418 //    in the independence checker.
 419 //
 420 // 4) For integer types, the necessary bit width is propagated backwards
 421 //    from stores to allow packed operations on byte, char, and short
 422 //    integers.  This reverses the promotion to type "int" that javac
 423 //    did for operations like: char c1,c2,c3;  c1 = c2 + c3.
 424 //
 425 // 5) One of the memory references is picked to be an aligned vector reference.
 426 //    The pre-loop trip count is adjusted to align this reference in the
 427 //    unrolled body.
 428 //
 429 // 6) The initial set of pack pairs is seeded with memory references.
 430 //
 431 // 7) The set of pack pairs is extended by following use->def and def->use links.
 432 //
 433 // 8) The pairs are combined into vector sized packs.
 434 //
 435 // 9) Reorder the memory slices to co-locate members of the memory packs.
 436 //
 437 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
 438 //    inserting scalar promotion, vector creation from multiple scalars, and
 439 //    extraction of scalar values from vectors.
 440 //
 441 void SuperWord::SLP_extract() {
 442 
 443 #ifndef PRODUCT
 444   if (_do_vector_loop && TraceSuperWord) {
 445     tty->print("SuperWord::SLP_extract\n");
 446     tty->print("input loop\n");
 447     _lpt->dump_head();
 448     _lpt->dump();
 449     for (uint i = 0; i < _lpt->_body.size(); i++) {
 450       _lpt->_body.at(i)->dump();
 451     }
 452   }
 453 #endif
 454   // Ready the block
 455   if (!construct_bb()) {
 456     return; // Exit if no interesting nodes or complex graph.
 457   }
 458 
 459   // build    _dg, _disjoint_ptrs
 460   dependence_graph();
 461 
 462   // compute function depth(Node*)
 463   compute_max_depth();
 464 
 465   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
 466   bool post_loop_allowed = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
 467   if (cl->is_main_loop()) {
 468     if (_do_vector_loop) {
 469       if (mark_generations() != -1) {
 470         hoist_loads_in_graph(); // this only rebuild the graph; all basic structs need rebuild explicitly
 471 
 472         if (!construct_bb()) {
 473           return; // Exit if no interesting nodes or complex graph.
 474         }
 475         dependence_graph();
 476         compute_max_depth();
 477       }
 478 
 479 #ifndef PRODUCT
 480       if (TraceSuperWord) {
 481         tty->print_cr("\nSuperWord::_do_vector_loop: graph after hoist_loads_in_graph");
 482         _lpt->dump_head();
 483         for (int j = 0; j < _block.length(); j++) {
 484           Node* n = _block.at(j);
 485           int d = depth(n);
 486           for (int i = 0; i < d; i++) tty->print("%s", "  ");
 487           tty->print("%d :", d);
 488           n->dump();
 489         }
 490       }
 491 #endif
 492     }
 493 
 494     compute_vector_element_type();
 495 
 496     // Attempt vectorization
 497 
 498     find_adjacent_refs();
 499 
 500     extend_packlist();
 501 
 502     if (_do_vector_loop) {
 503       if (_packset.length() == 0) {
 504         if (TraceSuperWord) {
 505           tty->print_cr("\nSuperWord::_do_vector_loop DFA could not build packset, now trying to build anyway");
 506         }
 507         pack_parallel();
 508       }
 509     }
 510 
 511     combine_packs();
 512 
 513     construct_my_pack_map();
 514 
 515     if (_do_vector_loop) {
 516       merge_packs_to_cmovd();
 517     }
 518 
 519     filter_packs();
 520 
 521     schedule();
 522   } else if (post_loop_allowed) {
 523     int saved_mapped_unroll_factor = cl->slp_max_unroll();
 524     if (saved_mapped_unroll_factor) {
 525       int vector_mapped_unroll_factor = saved_mapped_unroll_factor;
 526 
 527       // now reset the slp_unroll_factor so that we can check the analysis mapped
 528       // what the vector loop was mapped to
 529       cl->set_slp_max_unroll(0);
 530 
 531       // do the analysis on the post loop
 532       unrolling_analysis(vector_mapped_unroll_factor);
 533 
 534       // if our analyzed loop is a canonical fit, start processing it
 535       if (vector_mapped_unroll_factor == saved_mapped_unroll_factor) {
 536         // now add the vector nodes to packsets
 537         for (int i = 0; i < _post_block.length(); i++) {
 538           Node* n = _post_block.at(i);
 539           Node_List* singleton = new Node_List();
 540           singleton->push(n);
 541           _packset.append(singleton);
 542           set_my_pack(n, singleton);
 543         }
 544 
 545         // map base types for vector usage
 546         compute_vector_element_type();
 547       } else {
 548         return;
 549       }
 550     } else {
 551       // for some reason we could not map the slp analysis state of the vectorized loop
 552       return;
 553     }
 554   }
 555 
 556   output();
 557 }
 558 
 559 //------------------------------find_adjacent_refs---------------------------
 560 // Find the adjacent memory references and create pack pairs for them.
 561 // This is the initial set of packs that will then be extended by
 562 // following use->def and def->use links.  The align positions are
 563 // assigned relative to the reference "align_to_ref"
 564 void SuperWord::find_adjacent_refs() {
 565   // Get list of memory operations
 566   Node_List memops;
 567   for (int i = 0; i < _block.length(); i++) {
 568     Node* n = _block.at(i);
 569     if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
 570         is_java_primitive(n->as_Mem()->memory_type())) {
 571       int align = memory_alignment(n->as_Mem(), 0);
 572       if (align != bottom_align) {
 573         memops.push(n);
 574       }
 575     }
 576   }
 577 
 578   Node_List align_to_refs;
 579   int best_iv_adjustment = 0;
 580   MemNode* best_align_to_mem_ref = NULL;
 581 
 582   while (memops.size() != 0) {
 583     // Find a memory reference to align to.
 584     MemNode* mem_ref = find_align_to_ref(memops);
 585     if (mem_ref == NULL) break;
 586     align_to_refs.push(mem_ref);
 587     int iv_adjustment = get_iv_adjustment(mem_ref);
 588 
 589     if (best_align_to_mem_ref == NULL) {
 590       // Set memory reference which is the best from all memory operations
 591       // to be used for alignment. The pre-loop trip count is modified to align
 592       // this reference to a vector-aligned address.
 593       best_align_to_mem_ref = mem_ref;
 594       best_iv_adjustment = iv_adjustment;
 595       NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
 596     }
 597 
 598     SWPointer align_to_ref_p(mem_ref, this, NULL, false);
 599     // Set alignment relative to "align_to_ref" for all related memory operations.
 600     for (int i = memops.size() - 1; i >= 0; i--) {
 601       MemNode* s = memops.at(i)->as_Mem();
 602       if (isomorphic(s, mem_ref) &&
 603            (!_do_vector_loop || same_origin_idx(s, mem_ref))) {
 604         SWPointer p2(s, this, NULL, false);
 605         if (p2.comparable(align_to_ref_p)) {
 606           int align = memory_alignment(s, iv_adjustment);
 607           set_alignment(s, align);
 608         }
 609       }
 610     }
 611 
 612     // Create initial pack pairs of memory operations for which
 613     // alignment is set and vectors will be aligned.
 614     bool create_pack = true;
 615     if (memory_alignment(mem_ref, best_iv_adjustment) == 0 || _do_vector_loop) {
 616       if (!Matcher::misaligned_vectors_ok()) {
 617         int vw = vector_width(mem_ref);
 618         int vw_best = vector_width(best_align_to_mem_ref);
 619         if (vw > vw_best) {
 620           // Do not vectorize a memory access with more elements per vector
 621           // if unaligned memory access is not allowed because number of
 622           // iterations in pre-loop will be not enough to align it.
 623           create_pack = false;
 624         } else {
 625           SWPointer p2(best_align_to_mem_ref, this, NULL, false);
 626           if (align_to_ref_p.invar() != p2.invar()) {
 627             // Do not vectorize memory accesses with different invariants
 628             // if unaligned memory accesses are not allowed.
 629             create_pack = false;
 630           }
 631         }
 632       }
 633     } else {
 634       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 635         // Can't allow vectorization of unaligned memory accesses with the
 636         // same type since it could be overlapped accesses to the same array.
 637         create_pack = false;
 638       } else {
 639         // Allow independent (different type) unaligned memory operations
 640         // if HW supports them.
 641         if (!Matcher::misaligned_vectors_ok()) {
 642           create_pack = false;
 643         } else {
 644           // Check if packs of the same memory type but
 645           // with a different alignment were created before.
 646           for (uint i = 0; i < align_to_refs.size(); i++) {
 647             MemNode* mr = align_to_refs.at(i)->as_Mem();
 648             if (same_velt_type(mr, mem_ref) &&
 649                 memory_alignment(mr, iv_adjustment) != 0)
 650               create_pack = false;
 651           }
 652         }
 653       }
 654     }
 655     if (create_pack) {
 656       for (uint i = 0; i < memops.size(); i++) {
 657         Node* s1 = memops.at(i);
 658         int align = alignment(s1);
 659         if (align == top_align) continue;
 660         for (uint j = 0; j < memops.size(); j++) {
 661           Node* s2 = memops.at(j);
 662           if (alignment(s2) == top_align) continue;
 663           if (s1 != s2 && are_adjacent_refs(s1, s2)) {
 664             if (stmts_can_pack(s1, s2, align)) {
 665               Node_List* pair = new Node_List();
 666               pair->push(s1);
 667               pair->push(s2);
 668               if (!_do_vector_loop || same_origin_idx(s1, s2)) {
 669                 _packset.append(pair);
 670               }
 671             }
 672           }
 673         }
 674       }
 675     } else { // Don't create unaligned pack
 676       // First, remove remaining memory ops of the same type from the list.
 677       for (int i = memops.size() - 1; i >= 0; i--) {
 678         MemNode* s = memops.at(i)->as_Mem();
 679         if (same_velt_type(s, mem_ref)) {
 680           memops.remove(i);
 681         }
 682       }
 683 
 684       // Second, remove already constructed packs of the same type.
 685       for (int i = _packset.length() - 1; i >= 0; i--) {
 686         Node_List* p = _packset.at(i);
 687         MemNode* s = p->at(0)->as_Mem();
 688         if (same_velt_type(s, mem_ref)) {
 689           remove_pack_at(i);
 690         }
 691       }
 692 
 693       // If needed find the best memory reference for loop alignment again.
 694       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 695         // Put memory ops from remaining packs back on memops list for
 696         // the best alignment search.
 697         uint orig_msize = memops.size();
 698         for (int i = 0; i < _packset.length(); i++) {
 699           Node_List* p = _packset.at(i);
 700           MemNode* s = p->at(0)->as_Mem();
 701           assert(!same_velt_type(s, mem_ref), "sanity");
 702           memops.push(s);
 703         }
 704         best_align_to_mem_ref = find_align_to_ref(memops);
 705         if (best_align_to_mem_ref == NULL) {
 706           if (TraceSuperWord) {
 707             tty->print_cr("SuperWord::find_adjacent_refs(): best_align_to_mem_ref == NULL");
 708           }
 709           break;
 710         }
 711         best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
 712         NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
 713         // Restore list.
 714         while (memops.size() > orig_msize)
 715           (void)memops.pop();
 716       }
 717     } // unaligned memory accesses
 718 
 719     // Remove used mem nodes.
 720     for (int i = memops.size() - 1; i >= 0; i--) {
 721       MemNode* m = memops.at(i)->as_Mem();
 722       if (alignment(m) != top_align) {
 723         memops.remove(i);
 724       }
 725     }
 726 
 727   } // while (memops.size() != 0
 728   set_align_to_ref(best_align_to_mem_ref);
 729 
 730   if (TraceSuperWord) {
 731     tty->print_cr("\nAfter find_adjacent_refs");
 732     print_packset();
 733   }
 734 }
 735 
 736 #ifndef PRODUCT
 737 void SuperWord::find_adjacent_refs_trace_1(Node* best_align_to_mem_ref, int best_iv_adjustment) {
 738   if (is_trace_adjacent()) {
 739     tty->print("SuperWord::find_adjacent_refs best_align_to_mem_ref = %d, best_iv_adjustment = %d",
 740        best_align_to_mem_ref->_idx, best_iv_adjustment);
 741        best_align_to_mem_ref->dump();
 742   }
 743 }
 744 #endif
 745 
 746 //------------------------------find_align_to_ref---------------------------
 747 // Find a memory reference to align the loop induction variable to.
 748 // Looks first at stores then at loads, looking for a memory reference
 749 // with the largest number of references similar to it.
 750 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
 751   GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
 752 
 753   // Count number of comparable memory ops
 754   for (uint i = 0; i < memops.size(); i++) {
 755     MemNode* s1 = memops.at(i)->as_Mem();
 756     SWPointer p1(s1, this, NULL, false);
 757     // Discard if pre loop can't align this reference
 758     if (!ref_is_alignable(p1)) {
 759       *cmp_ct.adr_at(i) = 0;
 760       continue;
 761     }
 762     for (uint j = i+1; j < memops.size(); j++) {
 763       MemNode* s2 = memops.at(j)->as_Mem();
 764       if (isomorphic(s1, s2)) {
 765         SWPointer p2(s2, this, NULL, false);
 766         if (p1.comparable(p2)) {
 767           (*cmp_ct.adr_at(i))++;
 768           (*cmp_ct.adr_at(j))++;
 769         }
 770       }
 771     }
 772   }
 773 
 774   // Find Store (or Load) with the greatest number of "comparable" references,
 775   // biggest vector size, smallest data size and smallest iv offset.
 776   int max_ct        = 0;
 777   int max_vw        = 0;
 778   int max_idx       = -1;
 779   int min_size      = max_jint;
 780   int min_iv_offset = max_jint;
 781   for (uint j = 0; j < memops.size(); j++) {
 782     MemNode* s = memops.at(j)->as_Mem();
 783     if (s->is_Store()) {
 784       int vw = vector_width_in_bytes(s);
 785       assert(vw > 1, "sanity");
 786       SWPointer p(s, this, NULL, false);
 787       if ( cmp_ct.at(j) >  max_ct ||
 788           (cmp_ct.at(j) == max_ct &&
 789             ( vw >  max_vw ||
 790              (vw == max_vw &&
 791               ( data_size(s) <  min_size ||
 792                (data_size(s) == min_size &&
 793                 p.offset_in_bytes() < min_iv_offset)))))) {
 794         max_ct = cmp_ct.at(j);
 795         max_vw = vw;
 796         max_idx = j;
 797         min_size = data_size(s);
 798         min_iv_offset = p.offset_in_bytes();
 799       }
 800     }
 801   }
 802   // If no stores, look at loads
 803   if (max_ct == 0) {
 804     for (uint j = 0; j < memops.size(); j++) {
 805       MemNode* s = memops.at(j)->as_Mem();
 806       if (s->is_Load()) {
 807         int vw = vector_width_in_bytes(s);
 808         assert(vw > 1, "sanity");
 809         SWPointer p(s, this, NULL, false);
 810         if ( cmp_ct.at(j) >  max_ct ||
 811             (cmp_ct.at(j) == max_ct &&
 812               ( vw >  max_vw ||
 813                (vw == max_vw &&
 814                 ( data_size(s) <  min_size ||
 815                  (data_size(s) == min_size &&
 816                   p.offset_in_bytes() < min_iv_offset)))))) {
 817           max_ct = cmp_ct.at(j);
 818           max_vw = vw;
 819           max_idx = j;
 820           min_size = data_size(s);
 821           min_iv_offset = p.offset_in_bytes();
 822         }
 823       }
 824     }
 825   }
 826 
 827 #ifdef ASSERT
 828   if (TraceSuperWord && Verbose) {
 829     tty->print_cr("\nVector memops after find_align_to_ref");
 830     for (uint i = 0; i < memops.size(); i++) {
 831       MemNode* s = memops.at(i)->as_Mem();
 832       s->dump();
 833     }
 834   }
 835 #endif
 836 
 837   if (max_ct > 0) {
 838 #ifdef ASSERT
 839     if (TraceSuperWord) {
 840       tty->print("\nVector align to node: ");
 841       memops.at(max_idx)->as_Mem()->dump();
 842     }
 843 #endif
 844     return memops.at(max_idx)->as_Mem();
 845   }
 846   return NULL;
 847 }
 848 
 849 //------------------------------ref_is_alignable---------------------------
 850 // Can the preloop align the reference to position zero in the vector?
 851 bool SuperWord::ref_is_alignable(SWPointer& p) {
 852   if (!p.has_iv()) {
 853     return true;   // no induction variable
 854   }
 855   CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
 856   assert(pre_end != NULL, "we must have a correct pre-loop");
 857   assert(pre_end->stride_is_con(), "pre loop stride is constant");
 858   int preloop_stride = pre_end->stride_con();
 859 
 860   int span = preloop_stride * p.scale_in_bytes();
 861   int mem_size = p.memory_size();
 862   int offset   = p.offset_in_bytes();
 863   // Stride one accesses are alignable if offset is aligned to memory operation size.
 864   // Offset can be unaligned when UseUnalignedAccesses is used.
 865   if (ABS(span) == mem_size && (ABS(offset) % mem_size) == 0) {
 866     return true;
 867   }
 868   // If the initial offset from start of the object is computable,
 869   // check if the pre-loop can align the final offset accordingly.
 870   //
 871   // In other words: Can we find an i such that the offset
 872   // after i pre-loop iterations is aligned to vw?
 873   //   (init_offset + pre_loop) % vw == 0              (1)
 874   // where
 875   //   pre_loop = i * span
 876   // is the number of bytes added to the offset by i pre-loop iterations.
 877   //
 878   // For this to hold we need pre_loop to increase init_offset by
 879   //   pre_loop = vw - (init_offset % vw)
 880   //
 881   // This is only possible if pre_loop is divisible by span because each
 882   // pre-loop iteration increases the initial offset by 'span' bytes:
 883   //   (vw - (init_offset % vw)) % span == 0
 884   //
 885   int vw = vector_width_in_bytes(p.mem());
 886   assert(vw > 1, "sanity");
 887   Node* init_nd = pre_end->init_trip();
 888   if (init_nd->is_Con() && p.invar() == NULL) {
 889     int init = init_nd->bottom_type()->is_int()->get_con();
 890     int init_offset = init * p.scale_in_bytes() + offset;
 891     assert(init_offset >= 0, "positive offset from object start");
 892     if (vw % span == 0) {
 893       // If vm is a multiple of span, we use formula (1).
 894       if (span > 0) {
 895         return (vw - (init_offset % vw)) % span == 0;
 896       } else {
 897         assert(span < 0, "nonzero stride * scale");
 898         return (init_offset % vw) % -span == 0;
 899       }
 900     } else if (span % vw == 0) {
 901       // If span is a multiple of vw, we can simplify formula (1) to:
 902       //   (init_offset + i * span) % vw == 0
 903       //     =>
 904       //   (init_offset % vw) + ((i * span) % vw) == 0
 905       //     =>
 906       //   init_offset % vw == 0
 907       //
 908       // Because we add a multiple of vw to the initial offset, the final
 909       // offset is a multiple of vw if and only if init_offset is a multiple.
 910       //
 911       return (init_offset % vw) == 0;
 912     }
 913   }
 914   return false;
 915 }
 916 
 917 //---------------------------get_iv_adjustment---------------------------
 918 // Calculate loop's iv adjustment for this memory ops.
 919 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
 920   SWPointer align_to_ref_p(mem_ref, this, NULL, false);
 921   int offset = align_to_ref_p.offset_in_bytes();
 922   int scale  = align_to_ref_p.scale_in_bytes();
 923   int elt_size = align_to_ref_p.memory_size();
 924   int vw       = vector_width_in_bytes(mem_ref);
 925   assert(vw > 1, "sanity");
 926   int iv_adjustment;
 927   if (scale != 0) {
 928     int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
 929     // At least one iteration is executed in pre-loop by default. As result
 930     // several iterations are needed to align memory operations in main-loop even
 931     // if offset is 0.
 932     int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
 933     assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
 934            "(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size);
 935     iv_adjustment = iv_adjustment_in_bytes/elt_size;
 936   } else {
 937     // This memory op is not dependent on iv (scale == 0)
 938     iv_adjustment = 0;
 939   }
 940 
 941 #ifndef PRODUCT
 942   if (TraceSuperWord) {
 943     tty->print("SuperWord::get_iv_adjustment: n = %d, noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d: ",
 944       mem_ref->_idx, offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
 945     mem_ref->dump();
 946   }
 947 #endif
 948   return iv_adjustment;
 949 }
 950 
 951 //---------------------------dependence_graph---------------------------
 952 // Construct dependency graph.
 953 // Add dependence edges to load/store nodes for memory dependence
 954 //    A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
 955 void SuperWord::dependence_graph() {
 956   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
 957   // First, assign a dependence node to each memory node
 958   for (int i = 0; i < _block.length(); i++ ) {
 959     Node *n = _block.at(i);
 960     if (n->is_Mem() || (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
 961       _dg.make_node(n);
 962     }
 963   }
 964 
 965   // For each memory slice, create the dependences
 966   for (int i = 0; i < _mem_slice_head.length(); i++) {
 967     Node* n      = _mem_slice_head.at(i);
 968     Node* n_tail = _mem_slice_tail.at(i);
 969 
 970     // Get slice in predecessor order (last is first)
 971     if (cl->is_main_loop()) {
 972       mem_slice_preds(n_tail, n, _nlist);
 973     }
 974 
 975 #ifndef PRODUCT
 976     if(TraceSuperWord && Verbose) {
 977       tty->print_cr("SuperWord::dependence_graph: built a new mem slice");
 978       for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 979         _nlist.at(j)->dump();
 980       }
 981     }
 982 #endif
 983     // Make the slice dependent on the root
 984     DepMem* slice = _dg.dep(n);
 985     _dg.make_edge(_dg.root(), slice);
 986 
 987     // Create a sink for the slice
 988     DepMem* slice_sink = _dg.make_node(NULL);
 989     _dg.make_edge(slice_sink, _dg.tail());
 990 
 991     // Now visit each pair of memory ops, creating the edges
 992     for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 993       Node* s1 = _nlist.at(j);
 994 
 995       // If no dependency yet, use slice
 996       if (_dg.dep(s1)->in_cnt() == 0) {
 997         _dg.make_edge(slice, s1);
 998       }
 999       SWPointer p1(s1->as_Mem(), this, NULL, false);
1000       bool sink_dependent = true;
1001       for (int k = j - 1; k >= 0; k--) {
1002         Node* s2 = _nlist.at(k);
1003         if (s1->is_Load() && s2->is_Load())
1004           continue;
1005         SWPointer p2(s2->as_Mem(), this, NULL, false);
1006 
1007         int cmp = p1.cmp(p2);
1008         if (SuperWordRTDepCheck &&
1009             p1.base() != p2.base() && p1.valid() && p2.valid()) {
1010           // Create a runtime check to disambiguate
1011           OrderedPair pp(p1.base(), p2.base());
1012           _disjoint_ptrs.append_if_missing(pp);
1013         } else if (!SWPointer::not_equal(cmp)) {
1014           // Possibly same address
1015           _dg.make_edge(s1, s2);
1016           sink_dependent = false;
1017         }
1018       }
1019       if (sink_dependent) {
1020         _dg.make_edge(s1, slice_sink);
1021       }
1022     }
1023 
1024     if (TraceSuperWord) {
1025       tty->print_cr("\nDependence graph for slice: %d", n->_idx);
1026       for (int q = 0; q < _nlist.length(); q++) {
1027         _dg.print(_nlist.at(q));
1028       }
1029       tty->cr();
1030     }
1031 
1032     _nlist.clear();
1033   }
1034 
1035   if (TraceSuperWord) {
1036     tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
1037     for (int r = 0; r < _disjoint_ptrs.length(); r++) {
1038       _disjoint_ptrs.at(r).print();
1039       tty->cr();
1040     }
1041     tty->cr();
1042   }
1043 
1044 }
1045 
1046 //---------------------------mem_slice_preds---------------------------
1047 // Return a memory slice (node list) in predecessor order starting at "start"
1048 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
1049   assert(preds.length() == 0, "start empty");
1050   Node* n = start;
1051   Node* prev = NULL;
1052   while (true) {
1053     NOT_PRODUCT( if(is_trace_mem_slice()) tty->print_cr("SuperWord::mem_slice_preds: n %d", n->_idx);)
1054     assert(in_bb(n), "must be in block");
1055     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1056       Node* out = n->fast_out(i);
1057       if (out->is_Load()) {
1058         if (in_bb(out)) {
1059           preds.push(out);
1060           if (TraceSuperWord && Verbose) {
1061             tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", out->_idx);
1062           }
1063         }
1064       } else {
1065         // FIXME
1066         if (out->is_MergeMem() && !in_bb(out)) {
1067           // Either unrolling is causing a memory edge not to disappear,
1068           // or need to run igvn.optimize() again before SLP
1069         } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
1070           // Ditto.  Not sure what else to check further.
1071         } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
1072           // StoreCM has an input edge used as a precedence edge.
1073           // Maybe an issue when oop stores are vectorized.
1074         } else {
1075           assert(out == prev || prev == NULL, "no branches off of store slice");
1076         }
1077       }//else
1078     }//for
1079     if (n == stop) break;
1080     preds.push(n);
1081     if (TraceSuperWord && Verbose) {
1082       tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", n->_idx);
1083     }
1084     prev = n;
1085     assert(n->is_Mem(), "unexpected node %s", n->Name());
1086     n = n->in(MemNode::Memory);
1087   }
1088 }
1089 
1090 //------------------------------stmts_can_pack---------------------------
1091 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
1092 // s1 aligned at "align"
1093 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
1094 
1095   // Do not use superword for non-primitives
1096   BasicType bt1 = velt_basic_type(s1);
1097   BasicType bt2 = velt_basic_type(s2);
1098   if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
1099     return false;
1100   if (Matcher::max_vector_size(bt1) < 2) {
1101     return false; // No vectors for this type
1102   }
1103 
1104   if (isomorphic(s1, s2)) {
1105     if (independent(s1, s2) || reduction(s1, s2)) {
1106       if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
1107         if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
1108           int s1_align = alignment(s1);
1109           int s2_align = alignment(s2);
1110           if (s1_align == top_align || s1_align == align) {
1111             if (s2_align == top_align || s2_align == align + data_size(s1)) {
1112               return true;
1113             }
1114           }
1115         }
1116       }
1117     }
1118   }
1119   return false;
1120 }
1121 
1122 //------------------------------exists_at---------------------------
1123 // Does s exist in a pack at position pos?
1124 bool SuperWord::exists_at(Node* s, uint pos) {
1125   for (int i = 0; i < _packset.length(); i++) {
1126     Node_List* p = _packset.at(i);
1127     if (p->at(pos) == s) {
1128       return true;
1129     }
1130   }
1131   return false;
1132 }
1133 
1134 //------------------------------are_adjacent_refs---------------------------
1135 // Is s1 immediately before s2 in memory?
1136 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
1137   if (!s1->is_Mem() || !s2->is_Mem()) return false;
1138   if (!in_bb(s1)    || !in_bb(s2))    return false;
1139 
1140   // Do not use superword for non-primitives
1141   if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
1142       !is_java_primitive(s2->as_Mem()->memory_type())) {
1143     return false;
1144   }
1145 
1146   // FIXME - co_locate_pack fails on Stores in different mem-slices, so
1147   // only pack memops that are in the same alias set until that's fixed.
1148   if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
1149       _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
1150     return false;
1151   SWPointer p1(s1->as_Mem(), this, NULL, false);
1152   SWPointer p2(s2->as_Mem(), this, NULL, false);
1153   if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
1154   int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
1155   return diff == data_size(s1);
1156 }
1157 
1158 //------------------------------isomorphic---------------------------
1159 // Are s1 and s2 similar?
1160 bool SuperWord::isomorphic(Node* s1, Node* s2) {
1161   if (s1->Opcode() != s2->Opcode()) return false;
1162   if (s1->req() != s2->req()) return false;
1163   if (s1->in(0) != s2->in(0)) return false;
1164   if (!same_velt_type(s1, s2)) return false;
1165   return true;
1166 }
1167 
1168 //------------------------------independent---------------------------
1169 // Is there no data path from s1 to s2 or s2 to s1?
1170 bool SuperWord::independent(Node* s1, Node* s2) {
1171   //  assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
1172   int d1 = depth(s1);
1173   int d2 = depth(s2);
1174   if (d1 == d2) return s1 != s2;
1175   Node* deep    = d1 > d2 ? s1 : s2;
1176   Node* shallow = d1 > d2 ? s2 : s1;
1177 
1178   visited_clear();
1179 
1180   return independent_path(shallow, deep);
1181 }
1182 
1183 //------------------------------reduction---------------------------
1184 // Is there a data path between s1 and s2 and the nodes reductions?
1185 bool SuperWord::reduction(Node* s1, Node* s2) {
1186   bool retValue = false;
1187   int d1 = depth(s1);
1188   int d2 = depth(s2);
1189   if (d1 + 1 == d2) {
1190     if (s1->is_reduction() && s2->is_reduction()) {
1191       // This is an ordered set, so s1 should define s2
1192       for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1193         Node* t1 = s1->fast_out(i);
1194         if (t1 == s2) {
1195           // both nodes are reductions and connected
1196           retValue = true;
1197         }
1198       }
1199     }
1200   }
1201 
1202   return retValue;
1203 }
1204 
1205 //------------------------------independent_path------------------------------
1206 // Helper for independent
1207 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
1208   if (dp >= 1000) return false; // stop deep recursion
1209   visited_set(deep);
1210   int shal_depth = depth(shallow);
1211   assert(shal_depth <= depth(deep), "must be");
1212   for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
1213     Node* pred = preds.current();
1214     if (in_bb(pred) && !visited_test(pred)) {
1215       if (shallow == pred) {
1216         return false;
1217       }
1218       if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
1219         return false;
1220       }
1221     }
1222   }
1223   return true;
1224 }
1225 
1226 //------------------------------set_alignment---------------------------
1227 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
1228   set_alignment(s1, align);
1229   if (align == top_align || align == bottom_align) {
1230     set_alignment(s2, align);
1231   } else {
1232     set_alignment(s2, align + data_size(s1));
1233   }
1234 }
1235 
1236 //------------------------------data_size---------------------------
1237 int SuperWord::data_size(Node* s) {
1238   Node* use = NULL; //test if the node is a candidate for CMoveVD optimization, then return the size of CMov
1239   if (_do_vector_loop) {
1240     use = _cmovev_kit.is_Bool_candidate(s);
1241     if (use != NULL) {
1242       return data_size(use);
1243     }
1244     use = _cmovev_kit.is_CmpD_candidate(s);
1245     if (use != NULL) {
1246       return data_size(use);
1247     }
1248   }
1249   int bsize = type2aelembytes(velt_basic_type(s));
1250   assert(bsize != 0, "valid size");
1251   return bsize;
1252 }
1253 
1254 //------------------------------extend_packlist---------------------------
1255 // Extend packset by following use->def and def->use links from pack members.
1256 void SuperWord::extend_packlist() {
1257   bool changed;
1258   do {
1259     packset_sort(_packset.length());
1260     changed = false;
1261     for (int i = 0; i < _packset.length(); i++) {
1262       Node_List* p = _packset.at(i);
1263       changed |= follow_use_defs(p);
1264       changed |= follow_def_uses(p);
1265     }
1266   } while (changed);
1267 
1268   if (_race_possible) {
1269     for (int i = 0; i < _packset.length(); i++) {
1270       Node_List* p = _packset.at(i);
1271       order_def_uses(p);
1272     }
1273   }
1274 
1275   if (TraceSuperWord) {
1276     tty->print_cr("\nAfter extend_packlist");
1277     print_packset();
1278   }
1279 }
1280 
1281 //------------------------------follow_use_defs---------------------------
1282 // Extend the packset by visiting operand definitions of nodes in pack p
1283 bool SuperWord::follow_use_defs(Node_List* p) {
1284   assert(p->size() == 2, "just checking");
1285   Node* s1 = p->at(0);
1286   Node* s2 = p->at(1);
1287   assert(s1->req() == s2->req(), "just checking");
1288   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1289 
1290   if (s1->is_Load()) return false;
1291 
1292   int align = alignment(s1);
1293   NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: s1 %d, align %d", s1->_idx, align);)
1294   bool changed = false;
1295   int start = s1->is_Store() ? MemNode::ValueIn   : 1;
1296   int end   = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
1297   for (int j = start; j < end; j++) {
1298     Node* t1 = s1->in(j);
1299     Node* t2 = s2->in(j);
1300     if (!in_bb(t1) || !in_bb(t2))
1301       continue;
1302     if (stmts_can_pack(t1, t2, align)) {
1303       if (est_savings(t1, t2) >= 0) {
1304         Node_List* pair = new Node_List();
1305         pair->push(t1);
1306         pair->push(t2);
1307         _packset.append(pair);
1308         NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: set_alignment(%d, %d, %d)", t1->_idx, t2->_idx, align);)
1309         set_alignment(t1, t2, align);
1310         changed = true;
1311       }
1312     }
1313   }
1314   return changed;
1315 }
1316 
1317 //------------------------------follow_def_uses---------------------------
1318 // Extend the packset by visiting uses of nodes in pack p
1319 bool SuperWord::follow_def_uses(Node_List* p) {
1320   bool changed = false;
1321   Node* s1 = p->at(0);
1322   Node* s2 = p->at(1);
1323   assert(p->size() == 2, "just checking");
1324   assert(s1->req() == s2->req(), "just checking");
1325   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1326 
1327   if (s1->is_Store()) return false;
1328 
1329   int align = alignment(s1);
1330   NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: s1 %d, align %d", s1->_idx, align);)
1331   int savings = -1;
1332   int num_s1_uses = 0;
1333   Node* u1 = NULL;
1334   Node* u2 = NULL;
1335   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1336     Node* t1 = s1->fast_out(i);
1337     num_s1_uses++;
1338     if (!in_bb(t1)) continue;
1339     for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1340       Node* t2 = s2->fast_out(j);
1341       if (!in_bb(t2)) continue;
1342       if (!opnd_positions_match(s1, t1, s2, t2))
1343         continue;
1344       if (stmts_can_pack(t1, t2, align)) {
1345         int my_savings = est_savings(t1, t2);
1346         if (my_savings > savings) {
1347           savings = my_savings;
1348           u1 = t1;
1349           u2 = t2;
1350         }
1351       }
1352     }
1353   }
1354   if (num_s1_uses > 1) {
1355     _race_possible = true;
1356   }
1357   if (savings >= 0) {
1358     Node_List* pair = new Node_List();
1359     pair->push(u1);
1360     pair->push(u2);
1361     _packset.append(pair);
1362     NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: set_alignment(%d, %d, %d)", u1->_idx, u2->_idx, align);)
1363     set_alignment(u1, u2, align);
1364     changed = true;
1365   }
1366   return changed;
1367 }
1368 
1369 //------------------------------order_def_uses---------------------------
1370 // For extended packsets, ordinally arrange uses packset by major component
1371 void SuperWord::order_def_uses(Node_List* p) {
1372   Node* s1 = p->at(0);
1373 
1374   if (s1->is_Store()) return;
1375 
1376   // reductions are always managed beforehand
1377   if (s1->is_reduction()) return;
1378 
1379   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1380     Node* t1 = s1->fast_out(i);
1381 
1382     // Only allow operand swap on commuting operations
1383     if (!t1->is_Add() && !t1->is_Mul()) {
1384       break;
1385     }
1386 
1387     // Now find t1's packset
1388     Node_List* p2 = NULL;
1389     for (int j = 0; j < _packset.length(); j++) {
1390       p2 = _packset.at(j);
1391       Node* first = p2->at(0);
1392       if (t1 == first) {
1393         break;
1394       }
1395       p2 = NULL;
1396     }
1397     // Arrange all sub components by the major component
1398     if (p2 != NULL) {
1399       for (uint j = 1; j < p->size(); j++) {
1400         Node* d1 = p->at(j);
1401         Node* u1 = p2->at(j);
1402         opnd_positions_match(s1, t1, d1, u1);
1403       }
1404     }
1405   }
1406 }
1407 
1408 //---------------------------opnd_positions_match-------------------------
1409 // Is the use of d1 in u1 at the same operand position as d2 in u2?
1410 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
1411   // check reductions to see if they are marshalled to represent the reduction
1412   // operator in a specified opnd
1413   if (u1->is_reduction() && u2->is_reduction()) {
1414     // ensure reductions have phis and reduction definitions feeding the 1st operand
1415     Node* first = u1->in(2);
1416     if (first->is_Phi() || first->is_reduction()) {
1417       u1->swap_edges(1, 2);
1418     }
1419     // ensure reductions have phis and reduction definitions feeding the 1st operand
1420     first = u2->in(2);
1421     if (first->is_Phi() || first->is_reduction()) {
1422       u2->swap_edges(1, 2);
1423     }
1424     return true;
1425   }
1426 
1427   uint ct = u1->req();
1428   if (ct != u2->req()) return false;
1429   uint i1 = 0;
1430   uint i2 = 0;
1431   do {
1432     for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
1433     for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
1434     if (i1 != i2) {
1435       if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
1436         // Further analysis relies on operands position matching.
1437         u2->swap_edges(i1, i2);
1438       } else {
1439         return false;
1440       }
1441     }
1442   } while (i1 < ct);
1443   return true;
1444 }
1445 
1446 //------------------------------est_savings---------------------------
1447 // Estimate the savings from executing s1 and s2 as a pack
1448 int SuperWord::est_savings(Node* s1, Node* s2) {
1449   int save_in = 2 - 1; // 2 operations per instruction in packed form
1450 
1451   // inputs
1452   for (uint i = 1; i < s1->req(); i++) {
1453     Node* x1 = s1->in(i);
1454     Node* x2 = s2->in(i);
1455     if (x1 != x2) {
1456       if (are_adjacent_refs(x1, x2)) {
1457         save_in += adjacent_profit(x1, x2);
1458       } else if (!in_packset(x1, x2)) {
1459         save_in -= pack_cost(2);
1460       } else {
1461         save_in += unpack_cost(2);
1462       }
1463     }
1464   }
1465 
1466   // uses of result
1467   uint ct = 0;
1468   int save_use = 0;
1469   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1470     Node* s1_use = s1->fast_out(i);
1471     for (int j = 0; j < _packset.length(); j++) {
1472       Node_List* p = _packset.at(j);
1473       if (p->at(0) == s1_use) {
1474         for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
1475           Node* s2_use = s2->fast_out(k);
1476           if (p->at(p->size()-1) == s2_use) {
1477             ct++;
1478             if (are_adjacent_refs(s1_use, s2_use)) {
1479               save_use += adjacent_profit(s1_use, s2_use);
1480             }
1481           }
1482         }
1483       }
1484     }
1485   }
1486 
1487   if (ct < s1->outcnt()) save_use += unpack_cost(1);
1488   if (ct < s2->outcnt()) save_use += unpack_cost(1);
1489 
1490   return MAX2(save_in, save_use);
1491 }
1492 
1493 //------------------------------costs---------------------------
1494 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
1495 int SuperWord::pack_cost(int ct)   { return ct; }
1496 int SuperWord::unpack_cost(int ct) { return ct; }
1497 
1498 //------------------------------combine_packs---------------------------
1499 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
1500 void SuperWord::combine_packs() {
1501   bool changed = true;
1502   // Combine packs regardless max vector size.
1503   while (changed) {
1504     changed = false;
1505     for (int i = 0; i < _packset.length(); i++) {
1506       Node_List* p1 = _packset.at(i);
1507       if (p1 == NULL) continue;
1508       // Because of sorting we can start at i + 1
1509       for (int j = i + 1; j < _packset.length(); j++) {
1510         Node_List* p2 = _packset.at(j);
1511         if (p2 == NULL) continue;
1512         if (i == j) continue;
1513         if (p1->at(p1->size()-1) == p2->at(0)) {
1514           for (uint k = 1; k < p2->size(); k++) {
1515             p1->push(p2->at(k));
1516           }
1517           _packset.at_put(j, NULL);
1518           changed = true;
1519         }
1520       }
1521     }
1522   }
1523 
1524   // Split packs which have size greater then max vector size.
1525   for (int i = 0; i < _packset.length(); i++) {
1526     Node_List* p1 = _packset.at(i);
1527     if (p1 != NULL) {
1528       BasicType bt = velt_basic_type(p1->at(0));
1529       uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1530       assert(is_power_of_2(max_vlen), "sanity");
1531       uint psize = p1->size();
1532       if (!is_power_of_2(psize)) {
1533         // Skip pack which can't be vector.
1534         // case1: for(...) { a[i] = i; }    elements values are different (i+x)
1535         // case2: for(...) { a[i] = b[i+1]; }  can't align both, load and store
1536         _packset.at_put(i, NULL);
1537         continue;
1538       }
1539       if (psize > max_vlen) {
1540         Node_List* pack = new Node_List();
1541         for (uint j = 0; j < psize; j++) {
1542           pack->push(p1->at(j));
1543           if (pack->size() >= max_vlen) {
1544             assert(is_power_of_2(pack->size()), "sanity");
1545             _packset.append(pack);
1546             pack = new Node_List();
1547           }
1548         }
1549         _packset.at_put(i, NULL);
1550       }
1551     }
1552   }
1553 
1554   // Compress list.
1555   for (int i = _packset.length() - 1; i >= 0; i--) {
1556     Node_List* p1 = _packset.at(i);
1557     if (p1 == NULL) {
1558       _packset.remove_at(i);
1559     }
1560   }
1561 
1562   if (TraceSuperWord) {
1563     tty->print_cr("\nAfter combine_packs");
1564     print_packset();
1565   }
1566 }
1567 
1568 //-----------------------------construct_my_pack_map--------------------------
1569 // Construct the map from nodes to packs.  Only valid after the
1570 // point where a node is only in one pack (after combine_packs).
1571 void SuperWord::construct_my_pack_map() {
1572   Node_List* rslt = NULL;
1573   for (int i = 0; i < _packset.length(); i++) {
1574     Node_List* p = _packset.at(i);
1575     for (uint j = 0; j < p->size(); j++) {
1576       Node* s = p->at(j);
1577       assert(my_pack(s) == NULL, "only in one pack");
1578       set_my_pack(s, p);
1579     }
1580   }
1581 }
1582 
1583 //------------------------------filter_packs---------------------------
1584 // Remove packs that are not implemented or not profitable.
1585 void SuperWord::filter_packs() {
1586   // Remove packs that are not implemented
1587   for (int i = _packset.length() - 1; i >= 0; i--) {
1588     Node_List* pk = _packset.at(i);
1589     bool impl = implemented(pk);
1590     if (!impl) {
1591 #ifndef PRODUCT
1592       if (TraceSuperWord && Verbose) {
1593         tty->print_cr("Unimplemented");
1594         pk->at(0)->dump();
1595       }
1596 #endif
1597       remove_pack_at(i);
1598     }
1599     Node *n = pk->at(0);
1600     if (n->is_reduction()) {
1601       _num_reductions++;
1602     } else {
1603       _num_work_vecs++;
1604     }
1605   }
1606 
1607   // Remove packs that are not profitable
1608   bool changed;
1609   do {
1610     changed = false;
1611     for (int i = _packset.length() - 1; i >= 0; i--) {
1612       Node_List* pk = _packset.at(i);
1613       bool prof = profitable(pk);
1614       if (!prof) {
1615 #ifndef PRODUCT
1616         if (TraceSuperWord && Verbose) {
1617           tty->print_cr("Unprofitable");
1618           pk->at(0)->dump();
1619         }
1620 #endif
1621         remove_pack_at(i);
1622         changed = true;
1623       }
1624     }
1625   } while (changed);
1626 
1627 #ifndef PRODUCT
1628   if (TraceSuperWord) {
1629     tty->print_cr("\nAfter filter_packs");
1630     print_packset();
1631     tty->cr();
1632   }
1633 #endif
1634 }
1635 
1636 //------------------------------merge_packs_to_cmovd---------------------------
1637 // Merge CMoveD into new vector-nodes
1638 // We want to catch this pattern and subsume CmpD and Bool into CMoveD
1639 //
1640 //                   SubD             ConD
1641 //                  /  |               /
1642 //                 /   |           /   /
1643 //                /    |       /      /
1644 //               /     |   /         /
1645 //              /      /            /
1646 //             /    /  |           /
1647 //            v /      |          /
1648 //         CmpD        |         /
1649 //          |          |        /
1650 //          v          |       /
1651 //         Bool        |      /
1652 //           \         |     /
1653 //             \       |    /
1654 //               \     |   /
1655 //                 \   |  /
1656 //                   \ v /
1657 //                   CMoveD
1658 //
1659 
1660 void SuperWord::merge_packs_to_cmovd() {
1661   for (int i = _packset.length() - 1; i >= 0; i--) {
1662     _cmovev_kit.make_cmovevd_pack(_packset.at(i));
1663   }
1664   #ifndef PRODUCT
1665     if (TraceSuperWord) {
1666       tty->print_cr("\nSuperWord::merge_packs_to_cmovd(): After merge");
1667       print_packset();
1668       tty->cr();
1669     }
1670   #endif
1671 }
1672 
1673 Node* CMoveKit::is_Bool_candidate(Node* def) const {
1674   Node* use = NULL;
1675   if (!def->is_Bool() || def->in(0) != NULL || def->outcnt() != 1) {
1676     return NULL;
1677   }
1678   for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1679     use = def->fast_out(j);
1680     if (!_sw->same_generation(def, use) || !use->is_CMove()) {
1681       return NULL;
1682     }
1683   }
1684   return use;
1685 }
1686 
1687 Node* CMoveKit::is_CmpD_candidate(Node* def) const {
1688   Node* use = NULL;
1689   if (!def->is_Cmp() || def->in(0) != NULL || def->outcnt() != 1) {
1690     return NULL;
1691   }
1692   for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1693     use = def->fast_out(j);
1694     if (!_sw->same_generation(def, use) || (use = is_Bool_candidate(use)) == NULL || !_sw->same_generation(def, use)) {
1695       return NULL;
1696     }
1697   }
1698   return use;
1699 }
1700 
1701 Node_List* CMoveKit::make_cmovevd_pack(Node_List* cmovd_pk) {
1702   Node *cmovd = cmovd_pk->at(0);
1703   if (!cmovd->is_CMove()) {
1704     return NULL;
1705   }
1706   if (pack(cmovd) != NULL) { // already in the cmov pack
1707     return NULL;
1708   }
1709   if (cmovd->in(0) != NULL) {
1710     NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: CMoveD %d has control flow, escaping...", cmovd->_idx); cmovd->dump();})
1711     return NULL;
1712   }
1713 
1714   Node* bol = cmovd->as_CMove()->in(CMoveNode::Condition);
1715   if (!bol->is_Bool()
1716       || bol->outcnt() != 1
1717       || !_sw->same_generation(bol, cmovd)
1718       || bol->in(0) != NULL  // BoolNode has control flow!!
1719       || _sw->my_pack(bol) == NULL) {
1720       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();})
1721       return NULL;
1722   }
1723   Node_List* bool_pk = _sw->my_pack(bol);
1724   if (bool_pk->size() != cmovd_pk->size() ) {
1725     return NULL;
1726   }
1727 
1728   Node* cmpd = bol->in(1);
1729   if (!cmpd->is_Cmp()
1730       || cmpd->outcnt() != 1
1731       || !_sw->same_generation(cmpd, cmovd)
1732       || cmpd->in(0) != NULL  // CmpDNode has control flow!!
1733       || _sw->my_pack(cmpd) == NULL) {
1734       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();})
1735       return NULL;
1736   }
1737   Node_List* cmpd_pk = _sw->my_pack(cmpd);
1738   if (cmpd_pk->size() != cmovd_pk->size() ) {
1739     return NULL;
1740   }
1741 
1742   if (!test_cmpd_pack(cmpd_pk, cmovd_pk)) {
1743     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();})
1744     return NULL;
1745   }
1746 
1747   Node_List* new_cmpd_pk = new Node_List();
1748   uint sz = cmovd_pk->size() - 1;
1749   for (uint i = 0; i <= sz; ++i) {
1750     Node* cmov = cmovd_pk->at(i);
1751     Node* bol  = bool_pk->at(i);
1752     Node* cmp  = cmpd_pk->at(i);
1753 
1754     new_cmpd_pk->insert(i, cmov);
1755 
1756     map(cmov, new_cmpd_pk);
1757     map(bol, new_cmpd_pk);
1758     map(cmp, new_cmpd_pk);
1759 
1760     _sw->set_my_pack(cmov, new_cmpd_pk); // and keep old packs for cmp and bool
1761   }
1762   _sw->_packset.remove(cmovd_pk);
1763   _sw->_packset.remove(bool_pk);
1764   _sw->_packset.remove(cmpd_pk);
1765   _sw->_packset.append(new_cmpd_pk);
1766   NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print_cr("CMoveKit::make_cmovevd_pack: added syntactic CMoveD pack"); _sw->print_pack(new_cmpd_pk);})
1767   return new_cmpd_pk;
1768 }
1769 
1770 bool CMoveKit::test_cmpd_pack(Node_List* cmpd_pk, Node_List* cmovd_pk) {
1771   Node* cmpd0 = cmpd_pk->at(0);
1772   assert(cmpd0->is_Cmp(), "CMoveKit::test_cmpd_pack: should be CmpDNode");
1773   assert(cmovd_pk->at(0)->is_CMove(), "CMoveKit::test_cmpd_pack: should be CMoveD");
1774   assert(cmpd_pk->size() == cmovd_pk->size(), "CMoveKit::test_cmpd_pack: should be same size");
1775   Node* in1 = cmpd0->in(1);
1776   Node* in2 = cmpd0->in(2);
1777   Node_List* in1_pk = _sw->my_pack(in1);
1778   Node_List* in2_pk = _sw->my_pack(in2);
1779 
1780   if (  (in1_pk != NULL && in1_pk->size() != cmpd_pk->size())
1781      || (in2_pk != NULL && in2_pk->size() != cmpd_pk->size()) ) {
1782     return false;
1783   }
1784 
1785   // test if "all" in1 are in the same pack or the same node
1786   if (in1_pk == NULL) {
1787     for (uint j = 1; j < cmpd_pk->size(); j++) {
1788       if (cmpd_pk->at(j)->in(1) != in1) {
1789         return false;
1790       }
1791     }//for: in1_pk is not pack but all CmpD nodes in the pack have the same in(1)
1792   }
1793   // test if "all" in2 are in the same pack or the same node
1794   if (in2_pk == NULL) {
1795     for (uint j = 1; j < cmpd_pk->size(); j++) {
1796       if (cmpd_pk->at(j)->in(2) != in2) {
1797         return false;
1798       }
1799     }//for: in2_pk is not pack but all CmpD nodes in the pack have the same in(2)
1800   }
1801   //now check if cmpd_pk may be subsumed in vector built for cmovd_pk
1802   int cmovd_ind1, cmovd_ind2;
1803   if (cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1804    && cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1805       cmovd_ind1 = CMoveNode::IfFalse;
1806       cmovd_ind2 = CMoveNode::IfTrue;
1807   } else if (cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1808           && cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1809       cmovd_ind2 = CMoveNode::IfFalse;
1810       cmovd_ind1 = CMoveNode::IfTrue;
1811   }
1812   else {
1813     return false;
1814   }
1815 
1816   for (uint j = 1; j < cmpd_pk->size(); j++) {
1817     if (cmpd_pk->at(j)->in(1) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind1)
1818         || cmpd_pk->at(j)->in(2) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind2)) {
1819         return false;
1820     }//if
1821   }
1822   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(); })
1823   return true;
1824 }
1825 
1826 //------------------------------implemented---------------------------
1827 // Can code be generated for pack p?
1828 bool SuperWord::implemented(Node_List* p) {
1829   bool retValue = false;
1830   Node* p0 = p->at(0);
1831   if (p0 != NULL) {
1832     int opc = p0->Opcode();
1833     uint size = p->size();
1834     if (p0->is_reduction()) {
1835       const Type *arith_type = p0->bottom_type();
1836       // Length 2 reductions of INT/LONG do not offer performance benefits
1837       if (((arith_type->basic_type() == T_INT) || (arith_type->basic_type() == T_LONG)) && (size == 2)) {
1838         retValue = false;
1839       } else {
1840         retValue = ReductionNode::implemented(opc, size, arith_type->basic_type());
1841       }
1842     } else {
1843       retValue = VectorNode::implemented(opc, size, velt_basic_type(p0));
1844     }
1845     if (!retValue) {
1846       if (is_cmov_pack(p)) {
1847         NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::implemented: found cmpd pack"); print_pack(p);})
1848         return true;
1849       }
1850     }
1851   }
1852   return retValue;
1853 }
1854 
1855 bool SuperWord::is_cmov_pack(Node_List* p) {
1856   return _cmovev_kit.pack(p->at(0)) != NULL;
1857 }
1858 //------------------------------same_inputs--------------------------
1859 // For pack p, are all idx operands the same?
1860 bool SuperWord::same_inputs(Node_List* p, int idx) {
1861   Node* p0 = p->at(0);
1862   uint vlen = p->size();
1863   Node* p0_def = p0->in(idx);
1864   for (uint i = 1; i < vlen; i++) {
1865     Node* pi = p->at(i);
1866     Node* pi_def = pi->in(idx);
1867     if (p0_def != pi_def) {
1868       return false;
1869     }
1870   }
1871   return true;
1872 }
1873 
1874 //------------------------------profitable---------------------------
1875 // For pack p, are all operands and all uses (with in the block) vector?
1876 bool SuperWord::profitable(Node_List* p) {
1877   Node* p0 = p->at(0);
1878   uint start, end;
1879   VectorNode::vector_operands(p0, &start, &end);
1880 
1881   // Return false if some inputs are not vectors or vectors with different
1882   // size or alignment.
1883   // Also, for now, return false if not scalar promotion case when inputs are
1884   // the same. Later, implement PackNode and allow differing, non-vector inputs
1885   // (maybe just the ones from outside the block.)
1886   for (uint i = start; i < end; i++) {
1887     if (!is_vector_use(p0, i)) {
1888       return false;
1889     }
1890   }
1891   // Check if reductions are connected
1892   if (p0->is_reduction()) {
1893     Node* second_in = p0->in(2);
1894     Node_List* second_pk = my_pack(second_in);
1895     if ((second_pk == NULL) || (_num_work_vecs == _num_reductions)) {
1896       // Remove reduction flag if no parent pack or if not enough work
1897       // to cover reduction expansion overhead
1898       p0->remove_flag(Node::Flag_is_reduction);
1899       return false;
1900     } else if (second_pk->size() != p->size()) {
1901       return false;
1902     }
1903   }
1904   if (VectorNode::is_shift(p0)) {
1905     // For now, return false if shift count is vector or not scalar promotion
1906     // case (different shift counts) because it is not supported yet.
1907     Node* cnt = p0->in(2);
1908     Node_List* cnt_pk = my_pack(cnt);
1909     if (cnt_pk != NULL)
1910       return false;
1911     if (!same_inputs(p, 2))
1912       return false;
1913   }
1914   if (!p0->is_Store()) {
1915     // For now, return false if not all uses are vector.
1916     // Later, implement ExtractNode and allow non-vector uses (maybe
1917     // just the ones outside the block.)
1918     for (uint i = 0; i < p->size(); i++) {
1919       Node* def = p->at(i);
1920       if (is_cmov_pack_internal_node(p, def)) {
1921         continue;
1922       }
1923       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1924         Node* use = def->fast_out(j);
1925         for (uint k = 0; k < use->req(); k++) {
1926           Node* n = use->in(k);
1927           if (def == n) {
1928             // reductions can be loop carried dependences
1929             if (def->is_reduction() && use->is_Phi())
1930               continue;
1931             if (!is_vector_use(use, k)) {
1932               return false;
1933             }
1934           }
1935         }
1936       }
1937     }
1938   }
1939   return true;
1940 }
1941 
1942 //------------------------------schedule---------------------------
1943 // Adjust the memory graph for the packed operations
1944 void SuperWord::schedule() {
1945 
1946   // Co-locate in the memory graph the members of each memory pack
1947   for (int i = 0; i < _packset.length(); i++) {
1948     co_locate_pack(_packset.at(i));
1949   }
1950 }
1951 
1952 //-------------------------------remove_and_insert-------------------
1953 // Remove "current" from its current position in the memory graph and insert
1954 // it after the appropriate insertion point (lip or uip).
1955 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1956                                   Node *uip, Unique_Node_List &sched_before) {
1957   Node* my_mem = current->in(MemNode::Memory);
1958   bool sched_up = sched_before.member(current);
1959 
1960   // remove current_store from its current position in the memmory graph
1961   for (DUIterator i = current->outs(); current->has_out(i); i++) {
1962     Node* use = current->out(i);
1963     if (use->is_Mem()) {
1964       assert(use->in(MemNode::Memory) == current, "must be");
1965       if (use == prev) { // connect prev to my_mem
1966           _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1967           --i; //deleted this edge; rescan position
1968       } else if (sched_before.member(use)) {
1969         if (!sched_up) { // Will be moved together with current
1970           _igvn.replace_input_of(use, MemNode::Memory, uip);
1971           --i; //deleted this edge; rescan position
1972         }
1973       } else {
1974         if (sched_up) { // Will be moved together with current
1975           _igvn.replace_input_of(use, MemNode::Memory, lip);
1976           --i; //deleted this edge; rescan position
1977         }
1978       }
1979     }
1980   }
1981 
1982   Node *insert_pt =  sched_up ?  uip : lip;
1983 
1984   // all uses of insert_pt's memory state should use current's instead
1985   for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1986     Node* use = insert_pt->out(i);
1987     if (use->is_Mem()) {
1988       assert(use->in(MemNode::Memory) == insert_pt, "must be");
1989       _igvn.replace_input_of(use, MemNode::Memory, current);
1990       --i; //deleted this edge; rescan position
1991     } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1992       uint pos; //lip (lower insert point) must be the last one in the memory slice
1993       for (pos=1; pos < use->req(); pos++) {
1994         if (use->in(pos) == insert_pt) break;
1995       }
1996       _igvn.replace_input_of(use, pos, current);
1997       --i;
1998     }
1999   }
2000 
2001   //connect current to insert_pt
2002   _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
2003 }
2004 
2005 //------------------------------co_locate_pack----------------------------------
2006 // To schedule a store pack, we need to move any sandwiched memory ops either before
2007 // or after the pack, based upon dependence information:
2008 // (1) If any store in the pack depends on the sandwiched memory op, the
2009 //     sandwiched memory op must be scheduled BEFORE the pack;
2010 // (2) If a sandwiched memory op depends on any store in the pack, the
2011 //     sandwiched memory op must be scheduled AFTER the pack;
2012 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
2013 //     memory op (say memB), memB must be scheduled before memA. So, if memA is
2014 //     scheduled before the pack, memB must also be scheduled before the pack;
2015 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
2016 //     schedule this store AFTER the pack
2017 // (5) We know there is no dependence cycle, so there in no other case;
2018 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
2019 //
2020 // To schedule a load pack, we use the memory state of either the first or the last load in
2021 // the pack, based on the dependence constraint.
2022 void SuperWord::co_locate_pack(Node_List* pk) {
2023   if (pk->at(0)->is_Store()) {
2024     MemNode* first     = executed_first(pk)->as_Mem();
2025     MemNode* last      = executed_last(pk)->as_Mem();
2026     Unique_Node_List schedule_before_pack;
2027     Unique_Node_List memops;
2028 
2029     MemNode* current   = last->in(MemNode::Memory)->as_Mem();
2030     MemNode* previous  = last;
2031     while (true) {
2032       assert(in_bb(current), "stay in block");
2033       memops.push(previous);
2034       for (DUIterator i = current->outs(); current->has_out(i); i++) {
2035         Node* use = current->out(i);
2036         if (use->is_Mem() && use != previous)
2037           memops.push(use);
2038       }
2039       if (current == first) break;
2040       previous = current;
2041       current  = current->in(MemNode::Memory)->as_Mem();
2042     }
2043 
2044     // determine which memory operations should be scheduled before the pack
2045     for (uint i = 1; i < memops.size(); i++) {
2046       Node *s1 = memops.at(i);
2047       if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
2048         for (uint j = 0; j< i; j++) {
2049           Node *s2 = memops.at(j);
2050           if (!independent(s1, s2)) {
2051             if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
2052               schedule_before_pack.push(s1); // s1 must be scheduled before
2053               Node_List* mem_pk = my_pack(s1);
2054               if (mem_pk != NULL) {
2055                 for (uint ii = 0; ii < mem_pk->size(); ii++) {
2056                   Node* s = mem_pk->at(ii);  // follow partner
2057                   if (memops.member(s) && !schedule_before_pack.member(s))
2058                     schedule_before_pack.push(s);
2059                 }
2060               }
2061               break;
2062             }
2063           }
2064         }
2065       }
2066     }
2067 
2068     Node*    upper_insert_pt = first->in(MemNode::Memory);
2069     // Following code moves loads connected to upper_insert_pt below aliased stores.
2070     // Collect such loads here and reconnect them back to upper_insert_pt later.
2071     memops.clear();
2072     for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
2073       Node* use = upper_insert_pt->out(i);
2074       if (use->is_Mem() && !use->is_Store()) {
2075         memops.push(use);
2076       }
2077     }
2078 
2079     MemNode* lower_insert_pt = last;
2080     previous                 = last; //previous store in pk
2081     current                  = last->in(MemNode::Memory)->as_Mem();
2082 
2083     // start scheduling from "last" to "first"
2084     while (true) {
2085       assert(in_bb(current), "stay in block");
2086       assert(in_pack(previous, pk), "previous stays in pack");
2087       Node* my_mem = current->in(MemNode::Memory);
2088 
2089       if (in_pack(current, pk)) {
2090         // Forward users of my memory state (except "previous) to my input memory state
2091         for (DUIterator i = current->outs(); current->has_out(i); i++) {
2092           Node* use = current->out(i);
2093           if (use->is_Mem() && use != previous) {
2094             assert(use->in(MemNode::Memory) == current, "must be");
2095             if (schedule_before_pack.member(use)) {
2096               _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
2097             } else {
2098               _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
2099             }
2100             --i; // deleted this edge; rescan position
2101           }
2102         }
2103         previous = current;
2104       } else { // !in_pack(current, pk) ==> a sandwiched store
2105         remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
2106       }
2107 
2108       if (current == first) break;
2109       current = my_mem->as_Mem();
2110     } // end while
2111 
2112     // Reconnect loads back to upper_insert_pt.
2113     for (uint i = 0; i < memops.size(); i++) {
2114       Node *ld = memops.at(i);
2115       if (ld->in(MemNode::Memory) != upper_insert_pt) {
2116         _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
2117       }
2118     }
2119   } else if (pk->at(0)->is_Load()) { //load
2120     // all loads in the pack should have the same memory state. By default,
2121     // we use the memory state of the last load. However, if any load could
2122     // not be moved down due to the dependence constraint, we use the memory
2123     // state of the first load.
2124     Node* last_mem  = executed_last(pk)->in(MemNode::Memory);
2125     Node* first_mem = executed_first(pk)->in(MemNode::Memory);
2126     bool schedule_last = true;
2127     for (uint i = 0; i < pk->size(); i++) {
2128       Node* ld = pk->at(i);
2129       for (Node* current = last_mem; current != ld->in(MemNode::Memory);
2130            current=current->in(MemNode::Memory)) {
2131         assert(current != first_mem, "corrupted memory graph");
2132         if(current->is_Mem() && !independent(current, ld)){
2133           schedule_last = false; // a later store depends on this load
2134           break;
2135         }
2136       }
2137     }
2138 
2139     Node* mem_input = schedule_last ? last_mem : first_mem;
2140     _igvn.hash_delete(mem_input);
2141     // Give each load the same memory state
2142     for (uint i = 0; i < pk->size(); i++) {
2143       LoadNode* ld = pk->at(i)->as_Load();
2144       _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
2145     }
2146   }
2147 }
2148 
2149 #ifndef PRODUCT
2150 void SuperWord::print_loop(bool whole) {
2151   Node_Stack stack(_arena, _phase->C->unique() >> 2);
2152   Node_List rpo_list;
2153   VectorSet visited(_arena);
2154   visited.set(lpt()->_head->_idx);
2155   _phase->rpo(lpt()->_head, stack, visited, rpo_list);
2156   _phase->dump(lpt(), rpo_list.size(), rpo_list );
2157   if(whole) {
2158     tty->print_cr("\n Whole loop tree");
2159     _phase->dump();
2160     tty->print_cr(" End of whole loop tree\n");
2161   }
2162 }
2163 #endif
2164 
2165 //------------------------------output---------------------------
2166 // Convert packs into vector node operations
2167 void SuperWord::output() {
2168   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2169   Compile* C = _phase->C;
2170   if (_packset.length() == 0) {
2171     // Instigate more unrolling for optimization when vectorization fails.
2172     C->set_major_progress();
2173     cl->set_notpassed_slp();
2174     cl->mark_do_unroll_only();
2175     return;
2176   }
2177 
2178 #ifndef PRODUCT
2179   if (TraceLoopOpts) {
2180     tty->print("SuperWord::output    ");
2181     lpt()->dump_head();
2182   }
2183 #endif
2184 
2185   if (cl->is_main_loop()) {
2186     // MUST ENSURE main loop's initial value is properly aligned:
2187     //  (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
2188 
2189     align_initial_loop_index(align_to_ref());
2190 
2191     // Insert extract (unpack) operations for scalar uses
2192     for (int i = 0; i < _packset.length(); i++) {
2193       insert_extracts(_packset.at(i));
2194     }
2195   }
2196 
2197   uint max_vlen_in_bytes = 0;
2198   uint max_vlen = 0;
2199   bool can_process_post_loop = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
2200 
2201   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop before create_reserve_version_of_loop"); print_loop(true);})
2202 
2203   CountedLoopReserveKit make_reversable(_phase, _lpt, do_reserve_copy());
2204 
2205   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop after create_reserve_version_of_loop"); print_loop(true);})
2206 
2207   if (do_reserve_copy() && !make_reversable.has_reserved()) {
2208     NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: loop was not reserved correctly, exiting SuperWord");})
2209     return;
2210   }
2211 
2212   for (int i = 0; i < _block.length(); i++) {
2213     Node* n = _block.at(i);
2214     Node_List* p = my_pack(n);
2215     if (p && n == executed_last(p)) {
2216       uint vlen = p->size();
2217       uint vlen_in_bytes = 0;
2218       Node* vn = NULL;
2219       Node* low_adr = p->at(0);
2220       Node* first   = executed_first(p);
2221       if (can_process_post_loop) {
2222         // override vlen with the main loops vector length
2223         vlen = cl->slp_max_unroll();
2224       }
2225       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);})
2226       int   opc = n->Opcode();
2227       if (n->is_Load()) {
2228         Node* ctl = n->in(MemNode::Control);
2229         Node* mem = first->in(MemNode::Memory);
2230         SWPointer p1(n->as_Mem(), this, NULL, false);
2231         // Identify the memory dependency for the new loadVector node by
2232         // walking up through memory chain.
2233         // This is done to give flexibility to the new loadVector node so that
2234         // it can move above independent storeVector nodes.
2235         while (mem->is_StoreVector()) {
2236           SWPointer p2(mem->as_Mem(), this, NULL, false);
2237           int cmp = p1.cmp(p2);
2238           if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
2239             mem = mem->in(MemNode::Memory);
2240           } else {
2241             break; // dependent memory
2242           }
2243         }
2244         Node* adr = low_adr->in(MemNode::Address);
2245         const TypePtr* atyp = n->adr_type();
2246         vn = LoadVectorNode::make(opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n), control_dependency(p));
2247         vlen_in_bytes = vn->as_LoadVector()->memory_size();
2248       } else if (n->is_Store()) {
2249         // Promote value to be stored to vector
2250         Node* val = vector_opd(p, MemNode::ValueIn);
2251         if (val == NULL) {
2252           if (do_reserve_copy()) {
2253             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: val should not be NULL, exiting SuperWord");})
2254             return; //and reverse to backup IG
2255           }
2256           ShouldNotReachHere();
2257         }
2258 
2259         Node* ctl = n->in(MemNode::Control);
2260         Node* mem = first->in(MemNode::Memory);
2261         Node* adr = low_adr->in(MemNode::Address);
2262         const TypePtr* atyp = n->adr_type();
2263         vn = StoreVectorNode::make(opc, ctl, mem, adr, atyp, val, vlen);
2264         vlen_in_bytes = vn->as_StoreVector()->memory_size();
2265       } else if (n->req() == 3 && !is_cmov_pack(p)) {
2266         // Promote operands to vector
2267         Node* in1 = NULL;
2268         bool node_isa_reduction = n->is_reduction();
2269         if (node_isa_reduction) {
2270           // the input to the first reduction operation is retained
2271           in1 = low_adr->in(1);
2272         } else {
2273           in1 = vector_opd(p, 1);
2274           if (in1 == NULL) {
2275             if (do_reserve_copy()) {
2276               NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in1 should not be NULL, exiting SuperWord");})
2277               return; //and reverse to backup IG
2278             }
2279             ShouldNotReachHere();
2280           }
2281         }
2282         Node* in2 = vector_opd(p, 2);
2283         if (in2 == NULL) {
2284           if (do_reserve_copy()) {
2285             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in2 should not be NULL, exiting SuperWord");})
2286             return; //and reverse to backup IG
2287           }
2288           ShouldNotReachHere();
2289         }
2290         if (VectorNode::is_invariant_vector(in1) && (node_isa_reduction == false) && (n->is_Add() || n->is_Mul())) {
2291           // Move invariant vector input into second position to avoid register spilling.
2292           Node* tmp = in1;
2293           in1 = in2;
2294           in2 = tmp;
2295         }
2296         if (node_isa_reduction) {
2297           const Type *arith_type = n->bottom_type();
2298           vn = ReductionNode::make(opc, NULL, in1, in2, arith_type->basic_type());
2299           if (in2->is_Load()) {
2300             vlen_in_bytes = in2->as_LoadVector()->memory_size();
2301           } else {
2302             vlen_in_bytes = in2->as_Vector()->length_in_bytes();
2303           }
2304         } else {
2305           vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2306           vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2307         }
2308       } else if (opc == Op_SqrtD || opc == Op_AbsF || opc == Op_AbsD || opc == Op_NegF || opc == Op_NegD) {
2309         // Promote operand to vector (Sqrt/Abs/Neg are 2 address instructions)
2310         Node* in = vector_opd(p, 1);
2311         vn = VectorNode::make(opc, in, NULL, vlen, velt_basic_type(n));
2312         vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2313       } else if (is_cmov_pack(p)) {
2314         if (can_process_post_loop) {
2315           // do not refactor of flow in post loop context
2316           return;
2317         }
2318         if (!n->is_CMove()) {
2319           continue;
2320         }
2321         // place here CMoveVDNode
2322         NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::output: print before CMove vectorization"); print_loop(false);})
2323         Node* bol = n->in(CMoveNode::Condition);
2324         if (!bol->is_Bool() && bol->Opcode() == Op_ExtractI && bol->req() > 1 ) {
2325           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();})
2326           bol = bol->in(1); //may be ExtractNode
2327         }
2328 
2329         assert(bol->is_Bool(), "should be BoolNode - too late to bail out!");
2330         if (!bol->is_Bool()) {
2331           if (do_reserve_copy()) {
2332             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: expected %d bool node, exiting SuperWord", bol->_idx); bol->dump();})
2333             return; //and reverse to backup IG
2334           }
2335           ShouldNotReachHere();
2336         }
2337 
2338         int cond = (int)bol->as_Bool()->_test._test;
2339         Node* in_cc  = _igvn.intcon(cond);
2340         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created intcon in_cc node %d", in_cc->_idx); in_cc->dump();})
2341         Node* cc = bol->clone();
2342         cc->set_req(1, in_cc);
2343         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created bool cc node %d", cc->_idx); cc->dump();})
2344 
2345         Node* src1 = vector_opd(p, 2); //2=CMoveNode::IfFalse
2346         if (src1 == NULL) {
2347           if (do_reserve_copy()) {
2348             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src1 should not be NULL, exiting SuperWord");})
2349             return; //and reverse to backup IG
2350           }
2351           ShouldNotReachHere();
2352         }
2353         Node* src2 = vector_opd(p, 3); //3=CMoveNode::IfTrue
2354         if (src2 == NULL) {
2355           if (do_reserve_copy()) {
2356             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src2 should not be NULL, exiting SuperWord");})
2357             return; //and reverse to backup IG
2358           }
2359           ShouldNotReachHere();
2360         }
2361         BasicType bt = velt_basic_type(n);
2362         const TypeVect* vt = TypeVect::make(bt, vlen);
2363         vn = new CMoveVDNode(cc, src1, src2, vt);
2364         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created new CMove node %d: ", vn->_idx); vn->dump();})
2365       } else if (opc == Op_FmaD || opc == Op_FmaF) {
2366         // Promote operands to vector
2367         Node* in1 = vector_opd(p, 1);
2368         Node* in2 = vector_opd(p, 2);
2369         Node* in3 = vector_opd(p, 3);
2370         vn = VectorNode::make(opc, in1, in2, in3, vlen, velt_basic_type(n));
2371         vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2372       } else {
2373         if (do_reserve_copy()) {
2374           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: ShouldNotReachHere, exiting SuperWord");})
2375           return; //and reverse to backup IG
2376         }
2377         ShouldNotReachHere();
2378       }
2379 
2380       assert(vn != NULL, "sanity");
2381       if (vn == NULL) {
2382         if (do_reserve_copy()){
2383           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: got NULL node, cannot proceed, exiting SuperWord");})
2384           return; //and reverse to backup IG
2385         }
2386         ShouldNotReachHere();
2387       }
2388 
2389       _block.at_put(i, vn);
2390       _igvn.register_new_node_with_optimizer(vn);
2391       _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
2392       for (uint j = 0; j < p->size(); j++) {
2393         Node* pm = p->at(j);
2394         _igvn.replace_node(pm, vn);
2395       }
2396       _igvn._worklist.push(vn);
2397 
2398       if (can_process_post_loop) {
2399         // first check if the vector size if the maximum vector which we can use on the machine,
2400         // other vector size have reduced values for predicated data mapping.
2401         if (vlen_in_bytes != (uint)MaxVectorSize) {
2402           return;
2403         }
2404       }
2405 
2406       if (vlen_in_bytes >= max_vlen_in_bytes && vlen > max_vlen) {
2407         max_vlen = vlen;
2408         max_vlen_in_bytes = vlen_in_bytes;
2409       }
2410 #ifdef ASSERT
2411       if (TraceNewVectors) {
2412         tty->print("new Vector node: ");
2413         vn->dump();
2414       }
2415 #endif
2416     }
2417   }//for (int i = 0; i < _block.length(); i++)
2418 
2419   C->set_max_vector_size(max_vlen_in_bytes);
2420 
2421   if (SuperWordLoopUnrollAnalysis) {
2422     if (cl->has_passed_slp()) {
2423       uint slp_max_unroll_factor = cl->slp_max_unroll();
2424       if (slp_max_unroll_factor == max_vlen) {
2425         if (TraceSuperWordLoopUnrollAnalysis) {
2426           tty->print_cr("vector loop(unroll=%d, len=%d)\n", max_vlen, max_vlen_in_bytes*BitsPerByte);
2427         }
2428 
2429         // For atomic unrolled loops which are vector mapped, instigate more unrolling
2430         cl->set_notpassed_slp();
2431         if (cl->is_main_loop()) {
2432           // if vector resources are limited, do not allow additional unrolling, also
2433           // do not unroll more on pure vector loops which were not reduced so that we can
2434           // program the post loop to single iteration execution.
2435           if (FLOATPRESSURE > 8) {
2436             C->set_major_progress();
2437             cl->mark_do_unroll_only();
2438           }
2439         }
2440 
2441         if (do_reserve_copy()) {
2442           cl->mark_loop_vectorized();
2443           if (can_process_post_loop) {
2444             // Now create the difference of trip and limit and use it as our mask index.
2445             // Note: We limited the unroll of the vectorized loop so that
2446             //       only vlen-1 size iterations can remain to be mask programmed.
2447             Node *incr = cl->incr();
2448             SubINode *index = new SubINode(cl->limit(), cl->init_trip());
2449             _igvn.register_new_node_with_optimizer(index);
2450             SetVectMaskINode  *mask = new SetVectMaskINode(_phase->get_ctrl(cl->init_trip()), index);
2451             _igvn.register_new_node_with_optimizer(mask);
2452             // make this a single iteration loop
2453             AddINode *new_incr = new AddINode(incr->in(1), mask);
2454             _igvn.register_new_node_with_optimizer(new_incr);
2455             _phase->set_ctrl(new_incr, _phase->get_ctrl(incr));
2456             _igvn.replace_node(incr, new_incr);
2457             cl->mark_is_multiversioned();
2458             cl->loopexit()->add_flag(Node::Flag_has_vector_mask_set);
2459           }
2460         }
2461       }
2462     }
2463   }
2464 
2465   if (do_reserve_copy()) {
2466     make_reversable.use_new();
2467   }
2468   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("\n Final loop after SuperWord"); print_loop(true);})
2469   return;
2470 }
2471 
2472 //------------------------------vector_opd---------------------------
2473 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
2474 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
2475   Node* p0 = p->at(0);
2476   uint vlen = p->size();
2477   Node* opd = p0->in(opd_idx);
2478   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2479 
2480   if (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop()) {
2481     // override vlen with the main loops vector length
2482     vlen = cl->slp_max_unroll();
2483   }
2484 
2485   if (same_inputs(p, opd_idx)) {
2486     if (opd->is_Vector() || opd->is_LoadVector()) {
2487       assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
2488       if (opd_idx == 2 && VectorNode::is_shift(p0)) {
2489         NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("shift's count can't be vector");})
2490         return NULL;
2491       }
2492       return opd; // input is matching vector
2493     }
2494     if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
2495       Compile* C = _phase->C;
2496       Node* cnt = opd;
2497       // Vector instructions do not mask shift count, do it here.
2498       juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2499       const TypeInt* t = opd->find_int_type();
2500       if (t != NULL && t->is_con()) {
2501         juint shift = t->get_con();
2502         if (shift > mask) { // Unsigned cmp
2503           cnt = ConNode::make(TypeInt::make(shift & mask));
2504         }
2505       } else {
2506         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2507           cnt = ConNode::make(TypeInt::make(mask));
2508           _igvn.register_new_node_with_optimizer(cnt);
2509           cnt = new AndINode(opd, cnt);
2510           _igvn.register_new_node_with_optimizer(cnt);
2511           _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2512         }
2513         assert(opd->bottom_type()->isa_int(), "int type only");
2514         if (!opd->bottom_type()->isa_int()) {
2515           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should be int type only");})
2516           return NULL;
2517         }
2518         // Move non constant shift count into vector register.
2519         cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0));
2520       }
2521       if (cnt != opd) {
2522         _igvn.register_new_node_with_optimizer(cnt);
2523         _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2524       }
2525       return cnt;
2526     }
2527     assert(!opd->is_StoreVector(), "such vector is not expected here");
2528     if (opd->is_StoreVector()) {
2529       NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("StoreVector is not expected here");})
2530       return NULL;
2531     }
2532     // Convert scalar input to vector with the same number of elements as
2533     // p0's vector. Use p0's type because size of operand's container in
2534     // vector should match p0's size regardless operand's size.
2535     const Type* p0_t = velt_type(p0);
2536     VectorNode* vn = VectorNode::scalar2vector(opd, vlen, p0_t);
2537 
2538     _igvn.register_new_node_with_optimizer(vn);
2539     _phase->set_ctrl(vn, _phase->get_ctrl(opd));
2540 #ifdef ASSERT
2541     if (TraceNewVectors) {
2542       tty->print("new Vector node: ");
2543       vn->dump();
2544     }
2545 #endif
2546     return vn;
2547   }
2548 
2549   // Insert pack operation
2550   BasicType bt = velt_basic_type(p0);
2551   PackNode* pk = PackNode::make(opd, vlen, bt);
2552   DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
2553 
2554   for (uint i = 1; i < vlen; i++) {
2555     Node* pi = p->at(i);
2556     Node* in = pi->in(opd_idx);
2557     assert(my_pack(in) == NULL, "Should already have been unpacked");
2558     if (my_pack(in) != NULL) {
2559       NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should already have been unpacked");})
2560       return NULL;
2561     }
2562     assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
2563     pk->add_opd(in);
2564   }
2565   _igvn.register_new_node_with_optimizer(pk);
2566   _phase->set_ctrl(pk, _phase->get_ctrl(opd));
2567 #ifdef ASSERT
2568   if (TraceNewVectors) {
2569     tty->print("new Vector node: ");
2570     pk->dump();
2571   }
2572 #endif
2573   return pk;
2574 }
2575 
2576 //------------------------------insert_extracts---------------------------
2577 // If a use of pack p is not a vector use, then replace the
2578 // use with an extract operation.
2579 void SuperWord::insert_extracts(Node_List* p) {
2580   if (p->at(0)->is_Store()) return;
2581   assert(_n_idx_list.is_empty(), "empty (node,index) list");
2582 
2583   // Inspect each use of each pack member.  For each use that is
2584   // not a vector use, replace the use with an extract operation.
2585 
2586   for (uint i = 0; i < p->size(); i++) {
2587     Node* def = p->at(i);
2588     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2589       Node* use = def->fast_out(j);
2590       for (uint k = 0; k < use->req(); k++) {
2591         Node* n = use->in(k);
2592         if (def == n) {
2593           Node_List* u_pk = my_pack(use);
2594           if ((u_pk == NULL || !is_cmov_pack(u_pk) || use->is_CMove()) && !is_vector_use(use, k)) {
2595               _n_idx_list.push(use, k);
2596           }
2597         }
2598       }
2599     }
2600   }
2601 
2602   while (_n_idx_list.is_nonempty()) {
2603     Node* use = _n_idx_list.node();
2604     int   idx = _n_idx_list.index();
2605     _n_idx_list.pop();
2606     Node* def = use->in(idx);
2607 
2608     if (def->is_reduction()) continue;
2609 
2610     // Insert extract operation
2611     _igvn.hash_delete(def);
2612     int def_pos = alignment(def) / data_size(def);
2613 
2614     Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def));
2615     _igvn.register_new_node_with_optimizer(ex);
2616     _phase->set_ctrl(ex, _phase->get_ctrl(def));
2617     _igvn.replace_input_of(use, idx, ex);
2618     _igvn._worklist.push(def);
2619 
2620     bb_insert_after(ex, bb_idx(def));
2621     set_velt_type(ex, velt_type(def));
2622   }
2623 }
2624 
2625 //------------------------------is_vector_use---------------------------
2626 // Is use->in(u_idx) a vector use?
2627 bool SuperWord::is_vector_use(Node* use, int u_idx) {
2628   Node_List* u_pk = my_pack(use);
2629   if (u_pk == NULL) return false;
2630   if (use->is_reduction()) return true;
2631   Node* def = use->in(u_idx);
2632   Node_List* d_pk = my_pack(def);
2633   if (d_pk == NULL) {
2634     // check for scalar promotion
2635     Node* n = u_pk->at(0)->in(u_idx);
2636     for (uint i = 1; i < u_pk->size(); i++) {
2637       if (u_pk->at(i)->in(u_idx) != n) return false;
2638     }
2639     return true;
2640   }
2641   if (u_pk->size() != d_pk->size())
2642     return false;
2643   for (uint i = 0; i < u_pk->size(); i++) {
2644     Node* ui = u_pk->at(i);
2645     Node* di = d_pk->at(i);
2646     if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
2647       return false;
2648   }
2649   return true;
2650 }
2651 
2652 //------------------------------construct_bb---------------------------
2653 // Construct reverse postorder list of block members
2654 bool SuperWord::construct_bb() {
2655   Node* entry = bb();
2656 
2657   assert(_stk.length() == 0,            "stk is empty");
2658   assert(_block.length() == 0,          "block is empty");
2659   assert(_data_entry.length() == 0,     "data_entry is empty");
2660   assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
2661   assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
2662 
2663   // Find non-control nodes with no inputs from within block,
2664   // create a temporary map from node _idx to bb_idx for use
2665   // by the visited and post_visited sets,
2666   // and count number of nodes in block.
2667   int bb_ct = 0;
2668   for (uint i = 0; i < lpt()->_body.size(); i++) {
2669     Node *n = lpt()->_body.at(i);
2670     set_bb_idx(n, i); // Create a temporary map
2671     if (in_bb(n)) {
2672       if (n->is_LoadStore() || n->is_MergeMem() ||
2673           (n->is_Proj() && !n->as_Proj()->is_CFG())) {
2674         // Bailout if the loop has LoadStore, MergeMem or data Proj
2675         // nodes. Superword optimization does not work with them.
2676         return false;
2677       }
2678       bb_ct++;
2679       if (!n->is_CFG()) {
2680         bool found = false;
2681         for (uint j = 0; j < n->req(); j++) {
2682           Node* def = n->in(j);
2683           if (def && in_bb(def)) {
2684             found = true;
2685             break;
2686           }
2687         }
2688         if (!found) {
2689           assert(n != entry, "can't be entry");
2690           _data_entry.push(n);
2691         }
2692       }
2693     }
2694   }
2695 
2696   // Find memory slices (head and tail)
2697   for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
2698     Node *n = lp()->fast_out(i);
2699     if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
2700       Node* n_tail  = n->in(LoopNode::LoopBackControl);
2701       if (n_tail != n->in(LoopNode::EntryControl)) {
2702         if (!n_tail->is_Mem()) {
2703           assert(n_tail->is_Mem(), "unexpected node for memory slice: %s", n_tail->Name());
2704           return false; // Bailout
2705         }
2706         _mem_slice_head.push(n);
2707         _mem_slice_tail.push(n_tail);
2708       }
2709     }
2710   }
2711 
2712   // Create an RPO list of nodes in block
2713 
2714   visited_clear();
2715   post_visited_clear();
2716 
2717   // Push all non-control nodes with no inputs from within block, then control entry
2718   for (int j = 0; j < _data_entry.length(); j++) {
2719     Node* n = _data_entry.at(j);
2720     visited_set(n);
2721     _stk.push(n);
2722   }
2723   visited_set(entry);
2724   _stk.push(entry);
2725 
2726   // Do a depth first walk over out edges
2727   int rpo_idx = bb_ct - 1;
2728   int size;
2729   int reduction_uses = 0;
2730   while ((size = _stk.length()) > 0) {
2731     Node* n = _stk.top(); // Leave node on stack
2732     if (!visited_test_set(n)) {
2733       // forward arc in graph
2734     } else if (!post_visited_test(n)) {
2735       // cross or back arc
2736       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2737         Node *use = n->fast_out(i);
2738         if (in_bb(use) && !visited_test(use) &&
2739             // Don't go around backedge
2740             (!use->is_Phi() || n == entry)) {
2741           if (use->is_reduction()) {
2742             // First see if we can map the reduction on the given system we are on, then
2743             // make a data entry operation for each reduction we see.
2744             BasicType bt = use->bottom_type()->basic_type();
2745             if (ReductionNode::implemented(use->Opcode(), Matcher::min_vector_size(bt), bt)) {
2746               reduction_uses++;
2747             }
2748           }
2749           _stk.push(use);
2750         }
2751       }
2752       if (_stk.length() == size) {
2753         // There were no additional uses, post visit node now
2754         _stk.pop(); // Remove node from stack
2755         assert(rpo_idx >= 0, "");
2756         _block.at_put_grow(rpo_idx, n);
2757         rpo_idx--;
2758         post_visited_set(n);
2759         assert(rpo_idx >= 0 || _stk.is_empty(), "");
2760       }
2761     } else {
2762       _stk.pop(); // Remove post-visited node from stack
2763     }
2764   }//while
2765 
2766   int ii_current = -1;
2767   unsigned int load_idx = (unsigned int)-1;
2768   _ii_order.clear();
2769   // Create real map of block indices for nodes
2770   for (int j = 0; j < _block.length(); j++) {
2771     Node* n = _block.at(j);
2772     set_bb_idx(n, j);
2773     if (_do_vector_loop && n->is_Load()) {
2774       if (ii_current == -1) {
2775         ii_current = _clone_map.gen(n->_idx);
2776         _ii_order.push(ii_current);
2777         load_idx = _clone_map.idx(n->_idx);
2778       } else if (_clone_map.idx(n->_idx) == load_idx && _clone_map.gen(n->_idx) != ii_current) {
2779         ii_current = _clone_map.gen(n->_idx);
2780         _ii_order.push(ii_current);
2781       }
2782     }
2783   }//for
2784 
2785   // Ensure extra info is allocated.
2786   initialize_bb();
2787 
2788 #ifndef PRODUCT
2789   if (_vector_loop_debug && _ii_order.length() > 0) {
2790     tty->print("SuperWord::construct_bb: List of generations: ");
2791     for (int jj = 0; jj < _ii_order.length(); ++jj) {
2792       tty->print("  %d:%d", jj, _ii_order.at(jj));
2793     }
2794     tty->print_cr(" ");
2795   }
2796   if (TraceSuperWord) {
2797     print_bb();
2798     tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
2799     for (int m = 0; m < _data_entry.length(); m++) {
2800       tty->print("%3d ", m);
2801       _data_entry.at(m)->dump();
2802     }
2803     tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
2804     for (int m = 0; m < _mem_slice_head.length(); m++) {
2805       tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
2806       tty->print("    ");    _mem_slice_tail.at(m)->dump();
2807     }
2808   }
2809 #endif
2810   assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
2811   return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0);
2812 }
2813 
2814 //------------------------------initialize_bb---------------------------
2815 // Initialize per node info
2816 void SuperWord::initialize_bb() {
2817   Node* last = _block.at(_block.length() - 1);
2818   grow_node_info(bb_idx(last));
2819 }
2820 
2821 //------------------------------bb_insert_after---------------------------
2822 // Insert n into block after pos
2823 void SuperWord::bb_insert_after(Node* n, int pos) {
2824   int n_pos = pos + 1;
2825   // Make room
2826   for (int i = _block.length() - 1; i >= n_pos; i--) {
2827     _block.at_put_grow(i+1, _block.at(i));
2828   }
2829   for (int j = _node_info.length() - 1; j >= n_pos; j--) {
2830     _node_info.at_put_grow(j+1, _node_info.at(j));
2831   }
2832   // Set value
2833   _block.at_put_grow(n_pos, n);
2834   _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
2835   // Adjust map from node->_idx to _block index
2836   for (int i = n_pos; i < _block.length(); i++) {
2837     set_bb_idx(_block.at(i), i);
2838   }
2839 }
2840 
2841 //------------------------------compute_max_depth---------------------------
2842 // Compute max depth for expressions from beginning of block
2843 // Use to prune search paths during test for independence.
2844 void SuperWord::compute_max_depth() {
2845   int ct = 0;
2846   bool again;
2847   do {
2848     again = false;
2849     for (int i = 0; i < _block.length(); i++) {
2850       Node* n = _block.at(i);
2851       if (!n->is_Phi()) {
2852         int d_orig = depth(n);
2853         int d_in   = 0;
2854         for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
2855           Node* pred = preds.current();
2856           if (in_bb(pred)) {
2857             d_in = MAX2(d_in, depth(pred));
2858           }
2859         }
2860         if (d_in + 1 != d_orig) {
2861           set_depth(n, d_in + 1);
2862           again = true;
2863         }
2864       }
2865     }
2866     ct++;
2867   } while (again);
2868 
2869   if (TraceSuperWord && Verbose) {
2870     tty->print_cr("compute_max_depth iterated: %d times", ct);
2871   }
2872 }
2873 
2874 //-------------------------compute_vector_element_type-----------------------
2875 // Compute necessary vector element type for expressions
2876 // This propagates backwards a narrower integer type when the
2877 // upper bits of the value are not needed.
2878 // Example:  char a,b,c;  a = b + c;
2879 // Normally the type of the add is integer, but for packed character
2880 // operations the type of the add needs to be char.
2881 void SuperWord::compute_vector_element_type() {
2882   if (TraceSuperWord && Verbose) {
2883     tty->print_cr("\ncompute_velt_type:");
2884   }
2885 
2886   // Initial type
2887   for (int i = 0; i < _block.length(); i++) {
2888     Node* n = _block.at(i);
2889     set_velt_type(n, container_type(n));
2890   }
2891 
2892   // Propagate integer narrowed type backwards through operations
2893   // that don't depend on higher order bits
2894   for (int i = _block.length() - 1; i >= 0; i--) {
2895     Node* n = _block.at(i);
2896     // Only integer types need be examined
2897     const Type* vtn = velt_type(n);
2898     if (vtn->basic_type() == T_INT) {
2899       uint start, end;
2900       VectorNode::vector_operands(n, &start, &end);
2901 
2902       for (uint j = start; j < end; j++) {
2903         Node* in  = n->in(j);
2904         // Don't propagate through a memory
2905         if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
2906             data_size(n) < data_size(in)) {
2907           bool same_type = true;
2908           for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
2909             Node *use = in->fast_out(k);
2910             if (!in_bb(use) || !same_velt_type(use, n)) {
2911               same_type = false;
2912               break;
2913             }
2914           }
2915           if (same_type) {
2916             // For right shifts of small integer types (bool, byte, char, short)
2917             // we need precise information about sign-ness. Only Load nodes have
2918             // this information because Store nodes are the same for signed and
2919             // unsigned values. And any arithmetic operation after a load may
2920             // expand a value to signed Int so such right shifts can't be used
2921             // because vector elements do not have upper bits of Int.
2922             const Type* vt = vtn;
2923             if (VectorNode::is_shift(in)) {
2924               Node* load = in->in(1);
2925               if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
2926                 vt = velt_type(load);
2927               } else if (in->Opcode() != Op_LShiftI) {
2928                 // Widen type to Int to avoid creation of right shift vector
2929                 // (align + data_size(s1) check in stmts_can_pack() will fail).
2930                 // Note, left shifts work regardless type.
2931                 vt = TypeInt::INT;
2932               }
2933             }
2934             set_velt_type(in, vt);
2935           }
2936         }
2937       }
2938     }
2939   }
2940 #ifndef PRODUCT
2941   if (TraceSuperWord && Verbose) {
2942     for (int i = 0; i < _block.length(); i++) {
2943       Node* n = _block.at(i);
2944       velt_type(n)->dump();
2945       tty->print("\t");
2946       n->dump();
2947     }
2948   }
2949 #endif
2950 }
2951 
2952 //------------------------------memory_alignment---------------------------
2953 // Alignment within a vector memory reference
2954 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
2955   #ifndef PRODUCT
2956     if(TraceSuperWord && Verbose) {
2957       tty->print("SuperWord::memory_alignment within a vector memory reference for %d:  ", s->_idx); s->dump();
2958     }
2959   #endif
2960   NOT_PRODUCT(SWPointer::Tracer::Depth ddd(0);)
2961   SWPointer p(s, this, NULL, false);
2962   if (!p.valid()) {
2963     NOT_PRODUCT(if(is_trace_alignment()) tty->print("SWPointer::memory_alignment: SWPointer p invalid, return bottom_align");)
2964     return bottom_align;
2965   }
2966   int vw = vector_width_in_bytes(s);
2967   if (vw < 2) {
2968     NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SWPointer::memory_alignment: vector_width_in_bytes < 2, return bottom_align");)
2969     return bottom_align; // No vectors for this type
2970   }
2971   int offset  = p.offset_in_bytes();
2972   offset     += iv_adjust*p.memory_size();
2973   int off_rem = offset % vw;
2974   int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
2975   if (TraceSuperWord && Verbose) {
2976     tty->print_cr("SWPointer::memory_alignment: off_rem = %d, off_mod = %d", off_rem, off_mod);
2977   }
2978   return off_mod;
2979 }
2980 
2981 //---------------------------container_type---------------------------
2982 // Smallest type containing range of values
2983 const Type* SuperWord::container_type(Node* n) {
2984   if (n->is_Mem()) {
2985     BasicType bt = n->as_Mem()->memory_type();
2986     if (n->is_Store() && (bt == T_CHAR)) {
2987       // Use T_SHORT type instead of T_CHAR for stored values because any
2988       // preceding arithmetic operation extends values to signed Int.
2989       bt = T_SHORT;
2990     }
2991     if (n->Opcode() == Op_LoadUB) {
2992       // Adjust type for unsigned byte loads, it is important for right shifts.
2993       // T_BOOLEAN is used because there is no basic type representing type
2994       // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
2995       // size (one byte) and sign is important.
2996       bt = T_BOOLEAN;
2997     }
2998     return Type::get_const_basic_type(bt);
2999   }
3000   const Type* t = _igvn.type(n);
3001   if (t->basic_type() == T_INT) {
3002     // A narrow type of arithmetic operations will be determined by
3003     // propagating the type of memory operations.
3004     return TypeInt::INT;
3005   }
3006   return t;
3007 }
3008 
3009 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
3010   const Type* vt1 = velt_type(n1);
3011   const Type* vt2 = velt_type(n2);
3012   if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
3013     // Compare vectors element sizes for integer types.
3014     return data_size(n1) == data_size(n2);
3015   }
3016   return vt1 == vt2;
3017 }
3018 
3019 //------------------------------in_packset---------------------------
3020 // Are s1 and s2 in a pack pair and ordered as s1,s2?
3021 bool SuperWord::in_packset(Node* s1, Node* s2) {
3022   for (int i = 0; i < _packset.length(); i++) {
3023     Node_List* p = _packset.at(i);
3024     assert(p->size() == 2, "must be");
3025     if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
3026       return true;
3027     }
3028   }
3029   return false;
3030 }
3031 
3032 //------------------------------in_pack---------------------------
3033 // Is s in pack p?
3034 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
3035   for (uint i = 0; i < p->size(); i++) {
3036     if (p->at(i) == s) {
3037       return p;
3038     }
3039   }
3040   return NULL;
3041 }
3042 
3043 //------------------------------remove_pack_at---------------------------
3044 // Remove the pack at position pos in the packset
3045 void SuperWord::remove_pack_at(int pos) {
3046   Node_List* p = _packset.at(pos);
3047   for (uint i = 0; i < p->size(); i++) {
3048     Node* s = p->at(i);
3049     set_my_pack(s, NULL);
3050   }
3051   _packset.remove_at(pos);
3052 }
3053 
3054 void SuperWord::packset_sort(int n) {
3055   // simple bubble sort so that we capitalize with O(n) when its already sorted
3056   while (n != 0) {
3057     bool swapped = false;
3058     for (int i = 1; i < n; i++) {
3059       Node_List* q_low = _packset.at(i-1);
3060       Node_List* q_i = _packset.at(i);
3061 
3062       // only swap when we find something to swap
3063       if (alignment(q_low->at(0)) > alignment(q_i->at(0))) {
3064         Node_List* t = q_i;
3065         *(_packset.adr_at(i)) = q_low;
3066         *(_packset.adr_at(i-1)) = q_i;
3067         swapped = true;
3068       }
3069     }
3070     if (swapped == false) break;
3071     n--;
3072   }
3073 }
3074 
3075 //------------------------------executed_first---------------------------
3076 // Return the node executed first in pack p.  Uses the RPO block list
3077 // to determine order.
3078 Node* SuperWord::executed_first(Node_List* p) {
3079   Node* n = p->at(0);
3080   int n_rpo = bb_idx(n);
3081   for (uint i = 1; i < p->size(); i++) {
3082     Node* s = p->at(i);
3083     int s_rpo = bb_idx(s);
3084     if (s_rpo < n_rpo) {
3085       n = s;
3086       n_rpo = s_rpo;
3087     }
3088   }
3089   return n;
3090 }
3091 
3092 //------------------------------executed_last---------------------------
3093 // Return the node executed last in pack p.
3094 Node* SuperWord::executed_last(Node_List* p) {
3095   Node* n = p->at(0);
3096   int n_rpo = bb_idx(n);
3097   for (uint i = 1; i < p->size(); i++) {
3098     Node* s = p->at(i);
3099     int s_rpo = bb_idx(s);
3100     if (s_rpo > n_rpo) {
3101       n = s;
3102       n_rpo = s_rpo;
3103     }
3104   }
3105   return n;
3106 }
3107 
3108 LoadNode::ControlDependency SuperWord::control_dependency(Node_List* p) {
3109   LoadNode::ControlDependency dep = LoadNode::DependsOnlyOnTest;
3110   for (uint i = 0; i < p->size(); i++) {
3111     Node* n = p->at(i);
3112     assert(n->is_Load(), "only meaningful for loads");
3113     if (!n->depends_only_on_test()) {
3114       dep = LoadNode::Pinned;
3115     }
3116   }
3117   return dep;
3118 }
3119 
3120 
3121 //----------------------------align_initial_loop_index---------------------------
3122 // Adjust pre-loop limit so that in main loop, a load/store reference
3123 // to align_to_ref will be a position zero in the vector.
3124 //   (iv + k) mod vector_align == 0
3125 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
3126   CountedLoopNode *main_head = lp()->as_CountedLoop();
3127   assert(main_head->is_main_loop(), "");
3128   CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
3129   assert(pre_end != NULL, "we must have a correct pre-loop");
3130   Node *pre_opaq1 = pre_end->limit();
3131   assert(pre_opaq1->Opcode() == Op_Opaque1, "");
3132   Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
3133   Node *lim0 = pre_opaq->in(1);
3134 
3135   // Where we put new limit calculations
3136   Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
3137 
3138   // Ensure the original loop limit is available from the
3139   // pre-loop Opaque1 node.
3140   Node *orig_limit = pre_opaq->original_loop_limit();
3141   assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
3142 
3143   SWPointer align_to_ref_p(align_to_ref, this, NULL, false);
3144   assert(align_to_ref_p.valid(), "sanity");
3145 
3146   // Given:
3147   //     lim0 == original pre loop limit
3148   //     V == v_align (power of 2)
3149   //     invar == extra invariant piece of the address expression
3150   //     e == offset [ +/- invar ]
3151   //
3152   // When reassociating expressions involving '%' the basic rules are:
3153   //     (a - b) % k == 0   =>  a % k == b % k
3154   // and:
3155   //     (a + b) % k == 0   =>  a % k == (k - b) % k
3156   //
3157   // For stride > 0 && scale > 0,
3158   //   Derive the new pre-loop limit "lim" such that the two constraints:
3159   //     (1) lim = lim0 + N           (where N is some positive integer < V)
3160   //     (2) (e + lim) % V == 0
3161   //   are true.
3162   //
3163   //   Substituting (1) into (2),
3164   //     (e + lim0 + N) % V == 0
3165   //   solve for N:
3166   //     N = (V - (e + lim0)) % V
3167   //   substitute back into (1), so that new limit
3168   //     lim = lim0 + (V - (e + lim0)) % V
3169   //
3170   // For stride > 0 && scale < 0
3171   //   Constraints:
3172   //     lim = lim0 + N
3173   //     (e - lim) % V == 0
3174   //   Solving for lim:
3175   //     (e - lim0 - N) % V == 0
3176   //     N = (e - lim0) % V
3177   //     lim = lim0 + (e - lim0) % V
3178   //
3179   // For stride < 0 && scale > 0
3180   //   Constraints:
3181   //     lim = lim0 - N
3182   //     (e + lim) % V == 0
3183   //   Solving for lim:
3184   //     (e + lim0 - N) % V == 0
3185   //     N = (e + lim0) % V
3186   //     lim = lim0 - (e + lim0) % V
3187   //
3188   // For stride < 0 && scale < 0
3189   //   Constraints:
3190   //     lim = lim0 - N
3191   //     (e - lim) % V == 0
3192   //   Solving for lim:
3193   //     (e - lim0 + N) % V == 0
3194   //     N = (V - (e - lim0)) % V
3195   //     lim = lim0 - (V - (e - lim0)) % V
3196 
3197   int vw = vector_width_in_bytes(align_to_ref);
3198   int stride   = iv_stride();
3199   int scale    = align_to_ref_p.scale_in_bytes();
3200   int elt_size = align_to_ref_p.memory_size();
3201   int v_align  = vw / elt_size;
3202   assert(v_align > 1, "sanity");
3203   int offset   = align_to_ref_p.offset_in_bytes() / elt_size;
3204   Node *offsn  = _igvn.intcon(offset);
3205 
3206   Node *e = offsn;
3207   if (align_to_ref_p.invar() != NULL) {
3208     // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
3209     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3210     Node* invar = align_to_ref_p.invar();
3211     if (_igvn.type(invar)->isa_long()) {
3212       // Computations are done % (vector width/element size) so it's
3213       // safe to simply convert invar to an int and loose the upper 32
3214       // bit half.
3215       invar = new ConvL2INode(invar);
3216       _igvn.register_new_node_with_optimizer(invar);
3217     }
3218     Node* aref = new URShiftINode(invar, log2_elt);
3219     _igvn.register_new_node_with_optimizer(aref);
3220     _phase->set_ctrl(aref, pre_ctrl);
3221     if (align_to_ref_p.negate_invar()) {
3222       e = new SubINode(e, aref);
3223     } else {
3224       e = new AddINode(e, aref);
3225     }
3226     _igvn.register_new_node_with_optimizer(e);
3227     _phase->set_ctrl(e, pre_ctrl);
3228   }
3229   if (vw > ObjectAlignmentInBytes) {
3230     // incorporate base e +/- base && Mask >>> log2(elt)
3231     Node* xbase = new CastP2XNode(NULL, align_to_ref_p.base());
3232     _igvn.register_new_node_with_optimizer(xbase);
3233 #ifdef _LP64
3234     xbase  = new ConvL2INode(xbase);
3235     _igvn.register_new_node_with_optimizer(xbase);
3236 #endif
3237     Node* mask = _igvn.intcon(vw-1);
3238     Node* masked_xbase  = new AndINode(xbase, mask);
3239     _igvn.register_new_node_with_optimizer(masked_xbase);
3240     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3241     Node* bref     = new URShiftINode(masked_xbase, log2_elt);
3242     _igvn.register_new_node_with_optimizer(bref);
3243     _phase->set_ctrl(bref, pre_ctrl);
3244     e = new AddINode(e, bref);
3245     _igvn.register_new_node_with_optimizer(e);
3246     _phase->set_ctrl(e, pre_ctrl);
3247   }
3248 
3249   // compute e +/- lim0
3250   if (scale < 0) {
3251     e = new SubINode(e, lim0);
3252   } else {
3253     e = new AddINode(e, lim0);
3254   }
3255   _igvn.register_new_node_with_optimizer(e);
3256   _phase->set_ctrl(e, pre_ctrl);
3257 
3258   if (stride * scale > 0) {
3259     // compute V - (e +/- lim0)
3260     Node* va  = _igvn.intcon(v_align);
3261     e = new SubINode(va, e);
3262     _igvn.register_new_node_with_optimizer(e);
3263     _phase->set_ctrl(e, pre_ctrl);
3264   }
3265   // compute N = (exp) % V
3266   Node* va_msk = _igvn.intcon(v_align - 1);
3267   Node* N = new AndINode(e, va_msk);
3268   _igvn.register_new_node_with_optimizer(N);
3269   _phase->set_ctrl(N, pre_ctrl);
3270 
3271   //   substitute back into (1), so that new limit
3272   //     lim = lim0 + N
3273   Node* lim;
3274   if (stride < 0) {
3275     lim = new SubINode(lim0, N);
3276   } else {
3277     lim = new AddINode(lim0, N);
3278   }
3279   _igvn.register_new_node_with_optimizer(lim);
3280   _phase->set_ctrl(lim, pre_ctrl);
3281   Node* constrained =
3282     (stride > 0) ? (Node*) new MinINode(lim, orig_limit)
3283                  : (Node*) new MaxINode(lim, orig_limit);
3284   _igvn.register_new_node_with_optimizer(constrained);
3285   _phase->set_ctrl(constrained, pre_ctrl);
3286   _igvn.replace_input_of(pre_opaq, 1, constrained);
3287 }
3288 
3289 //----------------------------get_pre_loop_end---------------------------
3290 // Find pre loop end from main loop.  Returns null if none.
3291 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode* cl) {
3292   // The loop cannot be optimized if the graph shape at
3293   // the loop entry is inappropriate.
3294   if (!PhaseIdealLoop::is_canonical_loop_entry(cl)) {
3295     return NULL;
3296   }
3297 
3298   Node* p_f = cl->in(LoopNode::EntryControl)->in(0)->in(0);
3299   if (!p_f->is_IfFalse()) return NULL;
3300   if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
3301   CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
3302   CountedLoopNode* loop_node = pre_end->loopnode();
3303   if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
3304   return pre_end;
3305 }
3306 
3307 //------------------------------init---------------------------
3308 void SuperWord::init() {
3309   _dg.init();
3310   _packset.clear();
3311   _disjoint_ptrs.clear();
3312   _block.clear();
3313   _post_block.clear();
3314   _data_entry.clear();
3315   _mem_slice_head.clear();
3316   _mem_slice_tail.clear();
3317   _iteration_first.clear();
3318   _iteration_last.clear();
3319   _node_info.clear();
3320   _align_to_ref = NULL;
3321   _lpt = NULL;
3322   _lp = NULL;
3323   _bb = NULL;
3324   _iv = NULL;
3325   _race_possible = 0;
3326   _early_return = false;
3327   _num_work_vecs = 0;
3328   _num_reductions = 0;
3329 }
3330 
3331 //------------------------------restart---------------------------
3332 void SuperWord::restart() {
3333   _dg.init();
3334   _packset.clear();
3335   _disjoint_ptrs.clear();
3336   _block.clear();
3337   _post_block.clear();
3338   _data_entry.clear();
3339   _mem_slice_head.clear();
3340   _mem_slice_tail.clear();
3341   _node_info.clear();
3342 }
3343 
3344 //------------------------------print_packset---------------------------
3345 void SuperWord::print_packset() {
3346 #ifndef PRODUCT
3347   tty->print_cr("packset");
3348   for (int i = 0; i < _packset.length(); i++) {
3349     tty->print_cr("Pack: %d", i);
3350     Node_List* p = _packset.at(i);
3351     print_pack(p);
3352   }
3353 #endif
3354 }
3355 
3356 //------------------------------print_pack---------------------------
3357 void SuperWord::print_pack(Node_List* p) {
3358   for (uint i = 0; i < p->size(); i++) {
3359     print_stmt(p->at(i));
3360   }
3361 }
3362 
3363 //------------------------------print_bb---------------------------
3364 void SuperWord::print_bb() {
3365 #ifndef PRODUCT
3366   tty->print_cr("\nBlock");
3367   for (int i = 0; i < _block.length(); i++) {
3368     Node* n = _block.at(i);
3369     tty->print("%d ", i);
3370     if (n) {
3371       n->dump();
3372     }
3373   }
3374 #endif
3375 }
3376 
3377 //------------------------------print_stmt---------------------------
3378 void SuperWord::print_stmt(Node* s) {
3379 #ifndef PRODUCT
3380   tty->print(" align: %d \t", alignment(s));
3381   s->dump();
3382 #endif
3383 }
3384 
3385 //------------------------------blank---------------------------
3386 char* SuperWord::blank(uint depth) {
3387   static char blanks[101];
3388   assert(depth < 101, "too deep");
3389   for (uint i = 0; i < depth; i++) blanks[i] = ' ';
3390   blanks[depth] = '\0';
3391   return blanks;
3392 }
3393 
3394 
3395 //==============================SWPointer===========================
3396 #ifndef PRODUCT
3397 int SWPointer::Tracer::_depth = 0;
3398 #endif
3399 //----------------------------SWPointer------------------------
3400 SWPointer::SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only) :
3401   _mem(mem), _slp(slp),  _base(NULL),  _adr(NULL),
3402   _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3403   _nstack(nstack), _analyze_only(analyze_only),
3404   _stack_idx(0)
3405 #ifndef PRODUCT
3406   , _tracer(slp)
3407 #endif
3408 {
3409   NOT_PRODUCT(_tracer.ctor_1(mem);)
3410 
3411   Node* adr = mem->in(MemNode::Address);
3412   if (!adr->is_AddP()) {
3413     assert(!valid(), "too complex");
3414     return;
3415   }
3416   // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
3417   Node* base = adr->in(AddPNode::Base);
3418   // The base address should be loop invariant
3419   if (!invariant(base)) {
3420     assert(!valid(), "base address is loop variant");
3421     return;
3422   }
3423   //unsafe reference could not be aligned appropriately without runtime checking
3424   if (base == NULL || base->bottom_type() == Type::TOP) {
3425     assert(!valid(), "unsafe access");
3426     return;
3427   }
3428 
3429   NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.store_depth();)
3430   NOT_PRODUCT(_tracer.ctor_2(adr);)
3431 
3432   int i;
3433   for (i = 0; i < 3; i++) {
3434     NOT_PRODUCT(_tracer.ctor_3(adr, i);)
3435 
3436     if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
3437       assert(!valid(), "too complex");
3438       return;
3439     }
3440     adr = adr->in(AddPNode::Address);
3441     NOT_PRODUCT(_tracer.ctor_4(adr, i);)
3442 
3443     if (base == adr || !adr->is_AddP()) {
3444       NOT_PRODUCT(_tracer.ctor_5(adr, base, i);)
3445       break; // stop looking at addp's
3446     }
3447   }
3448   NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.restore_depth();)
3449   NOT_PRODUCT(_tracer.ctor_6(mem);)
3450 
3451   _base = base;
3452   _adr  = adr;
3453   assert(valid(), "Usable");
3454 }
3455 
3456 // Following is used to create a temporary object during
3457 // the pattern match of an address expression.
3458 SWPointer::SWPointer(SWPointer* p) :
3459   _mem(p->_mem), _slp(p->_slp),  _base(NULL),  _adr(NULL),
3460   _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3461   _nstack(p->_nstack), _analyze_only(p->_analyze_only),
3462   _stack_idx(p->_stack_idx)
3463   #ifndef PRODUCT
3464   , _tracer(p->_slp)
3465   #endif
3466 {}
3467 
3468 
3469 bool SWPointer::invariant(Node* n) {
3470   NOT_PRODUCT(Tracer::Depth dd;)
3471   Node *n_c = phase()->get_ctrl(n);
3472   NOT_PRODUCT(_tracer.invariant_1(n, n_c);)
3473   return !lpt()->is_member(phase()->get_loop(n_c));
3474 }
3475 //------------------------scaled_iv_plus_offset--------------------
3476 // Match: k*iv + offset
3477 // where: k is a constant that maybe zero, and
3478 //        offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
3479 bool SWPointer::scaled_iv_plus_offset(Node* n) {
3480   NOT_PRODUCT(Tracer::Depth ddd;)
3481   NOT_PRODUCT(_tracer.scaled_iv_plus_offset_1(n);)
3482 
3483   if (scaled_iv(n)) {
3484     NOT_PRODUCT(_tracer.scaled_iv_plus_offset_2(n);)
3485     return true;
3486   }
3487 
3488   if (offset_plus_k(n)) {
3489     NOT_PRODUCT(_tracer.scaled_iv_plus_offset_3(n);)
3490     return true;
3491   }
3492 
3493   int opc = n->Opcode();
3494   if (opc == Op_AddI) {
3495     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
3496       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_4(n);)
3497       return true;
3498     }
3499     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3500       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_5(n);)
3501       return true;
3502     }
3503   } else if (opc == Op_SubI) {
3504     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
3505       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_6(n);)
3506       return true;
3507     }
3508     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3509       _scale *= -1;
3510       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_7(n);)
3511       return true;
3512     }
3513   }
3514 
3515   NOT_PRODUCT(_tracer.scaled_iv_plus_offset_8(n);)
3516   return false;
3517 }
3518 
3519 //----------------------------scaled_iv------------------------
3520 // Match: k*iv where k is a constant that's not zero
3521 bool SWPointer::scaled_iv(Node* n) {
3522   NOT_PRODUCT(Tracer::Depth ddd;)
3523   NOT_PRODUCT(_tracer.scaled_iv_1(n);)
3524 
3525   if (_scale != 0) { // already found a scale
3526     NOT_PRODUCT(_tracer.scaled_iv_2(n, _scale);)
3527     return false;
3528   }
3529 
3530   if (n == iv()) {
3531     _scale = 1;
3532     NOT_PRODUCT(_tracer.scaled_iv_3(n, _scale);)
3533     return true;
3534   }
3535   if (_analyze_only && (invariant(n) == false)) {
3536     _nstack->push(n, _stack_idx++);
3537   }
3538 
3539   int opc = n->Opcode();
3540   if (opc == Op_MulI) {
3541     if (n->in(1) == iv() && n->in(2)->is_Con()) {
3542       _scale = n->in(2)->get_int();
3543       NOT_PRODUCT(_tracer.scaled_iv_4(n, _scale);)
3544       return true;
3545     } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
3546       _scale = n->in(1)->get_int();
3547       NOT_PRODUCT(_tracer.scaled_iv_5(n, _scale);)
3548       return true;
3549     }
3550   } else if (opc == Op_LShiftI) {
3551     if (n->in(1) == iv() && n->in(2)->is_Con()) {
3552       _scale = 1 << n->in(2)->get_int();
3553       NOT_PRODUCT(_tracer.scaled_iv_6(n, _scale);)
3554       return true;
3555     }
3556   } else if (opc == Op_ConvI2L) {
3557     if (n->in(1)->Opcode() == Op_CastII &&
3558         n->in(1)->as_CastII()->has_range_check()) {
3559       // Skip range check dependent CastII nodes
3560       n = n->in(1);
3561     }
3562     if (scaled_iv_plus_offset(n->in(1))) {
3563       NOT_PRODUCT(_tracer.scaled_iv_7(n);)
3564       return true;
3565     }
3566   } else if (opc == Op_LShiftL) {
3567     if (!has_iv() && _invar == NULL) {
3568       // Need to preserve the current _offset value, so
3569       // create a temporary object for this expression subtree.
3570       // Hacky, so should re-engineer the address pattern match.
3571       NOT_PRODUCT(Tracer::Depth dddd;)
3572       SWPointer tmp(this);
3573       NOT_PRODUCT(_tracer.scaled_iv_8(n, &tmp);)
3574 
3575       if (tmp.scaled_iv_plus_offset(n->in(1))) {
3576         if (tmp._invar == NULL || _slp->do_vector_loop()) {
3577           int mult = 1 << n->in(2)->get_int();
3578           _scale   = tmp._scale  * mult;
3579           _offset += tmp._offset * mult;
3580           NOT_PRODUCT(_tracer.scaled_iv_9(n, _scale, _offset, mult);)
3581           return true;
3582         }
3583       }
3584     }
3585   }
3586   NOT_PRODUCT(_tracer.scaled_iv_10(n);)
3587   return false;
3588 }
3589 
3590 //----------------------------offset_plus_k------------------------
3591 // Match: offset is (k [+/- invariant])
3592 // where k maybe zero and invariant is optional, but not both.
3593 bool SWPointer::offset_plus_k(Node* n, bool negate) {
3594   NOT_PRODUCT(Tracer::Depth ddd;)
3595   NOT_PRODUCT(_tracer.offset_plus_k_1(n);)
3596 
3597   int opc = n->Opcode();
3598   if (opc == Op_ConI) {
3599     _offset += negate ? -(n->get_int()) : n->get_int();
3600     NOT_PRODUCT(_tracer.offset_plus_k_2(n, _offset);)
3601     return true;
3602   } else if (opc == Op_ConL) {
3603     // Okay if value fits into an int
3604     const TypeLong* t = n->find_long_type();
3605     if (t->higher_equal(TypeLong::INT)) {
3606       jlong loff = n->get_long();
3607       jint  off  = (jint)loff;
3608       _offset += negate ? -off : loff;
3609       NOT_PRODUCT(_tracer.offset_plus_k_3(n, _offset);)
3610       return true;
3611     }
3612     NOT_PRODUCT(_tracer.offset_plus_k_4(n);)
3613     return false;
3614   }
3615   if (_invar != NULL) { // already has an invariant
3616     NOT_PRODUCT(_tracer.offset_plus_k_5(n, _invar);)
3617     return false;
3618   }
3619 
3620   if (_analyze_only && (invariant(n) == false)) {
3621     _nstack->push(n, _stack_idx++);
3622   }
3623   if (opc == Op_AddI) {
3624     if (n->in(2)->is_Con() && invariant(n->in(1))) {
3625       _negate_invar = negate;
3626       _invar = n->in(1);
3627       _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3628       NOT_PRODUCT(_tracer.offset_plus_k_6(n, _invar, _negate_invar, _offset);)
3629       return true;
3630     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3631       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3632       _negate_invar = negate;
3633       _invar = n->in(2);
3634       NOT_PRODUCT(_tracer.offset_plus_k_7(n, _invar, _negate_invar, _offset);)
3635       return true;
3636     }
3637   }
3638   if (opc == Op_SubI) {
3639     if (n->in(2)->is_Con() && invariant(n->in(1))) {
3640       _negate_invar = negate;
3641       _invar = n->in(1);
3642       _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3643       NOT_PRODUCT(_tracer.offset_plus_k_8(n, _invar, _negate_invar, _offset);)
3644       return true;
3645     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3646       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3647       _negate_invar = !negate;
3648       _invar = n->in(2);
3649       NOT_PRODUCT(_tracer.offset_plus_k_9(n, _invar, _negate_invar, _offset);)
3650       return true;
3651     }
3652   }
3653   if (invariant(n)) {
3654     if (opc == Op_ConvI2L) {
3655       n = n->in(1);
3656       if (n->Opcode() == Op_CastII &&
3657           n->as_CastII()->has_range_check()) {
3658         // Skip range check dependent CastII nodes
3659         assert(invariant(n), "sanity");
3660         n = n->in(1);
3661       }
3662     }
3663     _negate_invar = negate;
3664     _invar = n;
3665     NOT_PRODUCT(_tracer.offset_plus_k_10(n, _invar, _negate_invar, _offset);)
3666     return true;
3667   }
3668 
3669   NOT_PRODUCT(_tracer.offset_plus_k_11(n);)
3670   return false;
3671 }
3672 
3673 //----------------------------print------------------------
3674 void SWPointer::print() {
3675 #ifndef PRODUCT
3676   tty->print("base: %d  adr: %d  scale: %d  offset: %d  invar: %c%d\n",
3677              _base != NULL ? _base->_idx : 0,
3678              _adr  != NULL ? _adr->_idx  : 0,
3679              _scale, _offset,
3680              _negate_invar?'-':'+',
3681              _invar != NULL ? _invar->_idx : 0);
3682 #endif
3683 }
3684 
3685 //----------------------------tracing------------------------
3686 #ifndef PRODUCT
3687 void SWPointer::Tracer::print_depth() {
3688   for (int ii = 0; ii<_depth; ++ii) tty->print("  ");
3689 }
3690 
3691 void SWPointer::Tracer::ctor_1 (Node* mem) {
3692   if(_slp->is_trace_alignment()) {
3693     print_depth(); tty->print(" %d SWPointer::SWPointer: start alignment analysis", mem->_idx); mem->dump();
3694   }
3695 }
3696 
3697 void SWPointer::Tracer::ctor_2(Node* adr) {
3698   if(_slp->is_trace_alignment()) {
3699     //store_depth();
3700     inc_depth();
3701     print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: ", adr->_idx); adr->dump();
3702     inc_depth();
3703     print_depth(); tty->print(" %d (base) SWPointer::SWPointer: ", adr->in(AddPNode::Base)->_idx); adr->in(AddPNode::Base)->dump();
3704   }
3705 }
3706 
3707 void SWPointer::Tracer::ctor_3(Node* adr, int i) {
3708   if(_slp->is_trace_alignment()) {
3709     inc_depth();
3710     Node* offset = adr->in(AddPNode::Offset);
3711     print_depth(); tty->print(" %d (offset) SWPointer::SWPointer: i = %d: ", offset->_idx, i); offset->dump();
3712   }
3713 }
3714 
3715 void SWPointer::Tracer::ctor_4(Node* adr, int i) {
3716   if(_slp->is_trace_alignment()) {
3717     inc_depth();
3718     print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: i = %d: ", adr->_idx, i); adr->dump();
3719   }
3720 }
3721 
3722 void SWPointer::Tracer::ctor_5(Node* adr, Node* base, int i) {
3723   if(_slp->is_trace_alignment()) {
3724     inc_depth();
3725     if (base == adr) {
3726       print_depth(); tty->print_cr("  \\ %d (adr) == %d (base) SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, base->_idx, i);
3727     } else if (!adr->is_AddP()) {
3728       print_depth(); tty->print_cr("  \\ %d (adr) is NOT Addp SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, i);
3729     }
3730   }
3731 }
3732 
3733 void SWPointer::Tracer::ctor_6(Node* mem) {
3734   if(_slp->is_trace_alignment()) {
3735     //restore_depth();
3736     print_depth(); tty->print_cr(" %d (adr) SWPointer::SWPointer: stop analysis", mem->_idx);
3737   }
3738 }
3739 
3740 void SWPointer::Tracer::invariant_1(Node *n, Node *n_c) {
3741   if (_slp->do_vector_loop() && _slp->is_debug() && _slp->_lpt->is_member(_slp->_phase->get_loop(n_c)) != (int)_slp->in_bb(n)) {
3742     int is_member =  _slp->_lpt->is_member(_slp->_phase->get_loop(n_c));
3743     int in_bb     =  _slp->in_bb(n);
3744     print_depth(); tty->print("  \\ ");  tty->print_cr(" %d SWPointer::invariant  conditions differ: n_c %d", n->_idx, n_c->_idx);
3745     print_depth(); tty->print("  \\ ");  tty->print_cr("is_member %d, in_bb %d", is_member, in_bb);
3746     print_depth(); tty->print("  \\ ");  n->dump();
3747     print_depth(); tty->print("  \\ ");  n_c->dump();
3748   }
3749 }
3750 
3751 void SWPointer::Tracer::scaled_iv_plus_offset_1(Node* n) {
3752   if(_slp->is_trace_alignment()) {
3753     print_depth(); tty->print(" %d SWPointer::scaled_iv_plus_offset testing node: ", n->_idx);
3754     n->dump();
3755   }
3756 }
3757 
3758 void SWPointer::Tracer::scaled_iv_plus_offset_2(Node* n) {
3759   if(_slp->is_trace_alignment()) {
3760     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
3761   }
3762 }
3763 
3764 void SWPointer::Tracer::scaled_iv_plus_offset_3(Node* n) {
3765   if(_slp->is_trace_alignment()) {
3766     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
3767   }
3768 }
3769 
3770 void SWPointer::Tracer::scaled_iv_plus_offset_4(Node* n) {
3771   if(_slp->is_trace_alignment()) {
3772     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
3773     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
3774     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
3775   }
3776 }
3777 
3778 void SWPointer::Tracer::scaled_iv_plus_offset_5(Node* n) {
3779   if(_slp->is_trace_alignment()) {
3780     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
3781     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
3782     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
3783   }
3784 }
3785 
3786 void SWPointer::Tracer::scaled_iv_plus_offset_6(Node* n) {
3787   if(_slp->is_trace_alignment()) {
3788     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
3789     print_depth(); tty->print("  \\  %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
3790     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
3791   }
3792 }
3793 
3794 void SWPointer::Tracer::scaled_iv_plus_offset_7(Node* n) {
3795   if(_slp->is_trace_alignment()) {
3796     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
3797     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
3798     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
3799   }
3800 }
3801 
3802 void SWPointer::Tracer::scaled_iv_plus_offset_8(Node* n) {
3803   if(_slp->is_trace_alignment()) {
3804     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: FAILED", n->_idx);
3805   }
3806 }
3807 
3808 void SWPointer::Tracer::scaled_iv_1(Node* n) {
3809   if(_slp->is_trace_alignment()) {
3810     print_depth(); tty->print(" %d SWPointer::scaled_iv: testing node: ", n->_idx); n->dump();
3811   }
3812 }
3813 
3814 void SWPointer::Tracer::scaled_iv_2(Node* n, int scale) {
3815   if(_slp->is_trace_alignment()) {
3816     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED since another _scale has been detected before", n->_idx);
3817     print_depth(); tty->print_cr("  \\ SWPointer::scaled_iv: _scale (%d) != 0", scale);
3818   }
3819 }
3820 
3821 void SWPointer::Tracer::scaled_iv_3(Node* n, int scale) {
3822   if(_slp->is_trace_alignment()) {
3823     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: is iv, setting _scale = %d", n->_idx, scale);
3824   }
3825 }
3826 
3827 void SWPointer::Tracer::scaled_iv_4(Node* n, int scale) {
3828   if(_slp->is_trace_alignment()) {
3829     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
3830     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
3831     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3832   }
3833 }
3834 
3835 void SWPointer::Tracer::scaled_iv_5(Node* n, int scale) {
3836   if(_slp->is_trace_alignment()) {
3837     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
3838     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is iv: ", n->in(2)->_idx); n->in(2)->dump();
3839     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3840   }
3841 }
3842 
3843 void SWPointer::Tracer::scaled_iv_6(Node* n, int scale) {
3844   if(_slp->is_trace_alignment()) {
3845     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftI PASSED, setting _scale = %d", n->_idx, scale);
3846     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
3847     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3848   }
3849 }
3850 
3851 void SWPointer::Tracer::scaled_iv_7(Node* n) {
3852   if(_slp->is_trace_alignment()) {
3853     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_ConvI2L PASSED", n->_idx);
3854     print_depth(); tty->print_cr("  \\ SWPointer::scaled_iv: in(1) %d is scaled_iv_plus_offset: ", n->in(1)->_idx);
3855     inc_depth(); inc_depth();
3856     print_depth(); n->in(1)->dump();
3857     dec_depth(); dec_depth();
3858   }
3859 }
3860 
3861 void SWPointer::Tracer::scaled_iv_8(Node* n, SWPointer* tmp) {
3862   if(_slp->is_trace_alignment()) {
3863     print_depth(); tty->print(" %d SWPointer::scaled_iv: Op_LShiftL, creating tmp SWPointer: ", n->_idx); tmp->print();
3864   }
3865 }
3866 
3867 void SWPointer::Tracer::scaled_iv_9(Node* n, int scale, int _offset, int mult) {
3868   if(_slp->is_trace_alignment()) {
3869     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftL PASSED, setting _scale = %d, _offset = %d", n->_idx, scale, _offset);
3870     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",
3871     n->in(1)->_idx, n->in(2)->_idx, mult, scale, _offset);
3872     inc_depth(); inc_depth();
3873     print_depth(); n->in(1)->dump();
3874     print_depth(); n->in(2)->dump();
3875     dec_depth(); dec_depth();
3876   }
3877 }
3878 
3879 void SWPointer::Tracer::scaled_iv_10(Node* n) {
3880   if(_slp->is_trace_alignment()) {
3881     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED", n->_idx);
3882   }
3883 }
3884 
3885 void SWPointer::Tracer::offset_plus_k_1(Node* n) {
3886   if(_slp->is_trace_alignment()) {
3887     print_depth(); tty->print(" %d SWPointer::offset_plus_k: testing node: ", n->_idx); n->dump();
3888   }
3889 }
3890 
3891 void SWPointer::Tracer::offset_plus_k_2(Node* n, int _offset) {
3892   if(_slp->is_trace_alignment()) {
3893     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConI PASSED, setting _offset = %d", n->_idx, _offset);
3894   }
3895 }
3896 
3897 void SWPointer::Tracer::offset_plus_k_3(Node* n, int _offset) {
3898   if(_slp->is_trace_alignment()) {
3899     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConL PASSED, setting _offset = %d", n->_idx, _offset);
3900   }
3901 }
3902 
3903 void SWPointer::Tracer::offset_plus_k_4(Node* n) {
3904   if(_slp->is_trace_alignment()) {
3905     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
3906     print_depth(); tty->print_cr("  \\ " JLONG_FORMAT " SWPointer::offset_plus_k: Op_ConL FAILED, k is too big", n->get_long());
3907   }
3908 }
3909 
3910 void SWPointer::Tracer::offset_plus_k_5(Node* n, Node* _invar) {
3911   if(_slp->is_trace_alignment()) {
3912     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED since another invariant has been detected before", n->_idx);
3913     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: _invar != NULL: ", _invar->_idx); _invar->dump();
3914   }
3915 }
3916 
3917 void SWPointer::Tracer::offset_plus_k_6(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3918   if(_slp->is_trace_alignment()) {
3919     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
3920     n->_idx, _negate_invar, _invar->_idx, _offset);
3921     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3922     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
3923   }
3924 }
3925 
3926 void SWPointer::Tracer::offset_plus_k_7(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3927   if(_slp->is_trace_alignment()) {
3928     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
3929     n->_idx, _negate_invar, _invar->_idx, _offset);
3930     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3931     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
3932   }
3933 }
3934 
3935 void SWPointer::Tracer::offset_plus_k_8(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3936   if(_slp->is_trace_alignment()) {
3937     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_SubI is PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
3938     n->_idx, _negate_invar, _invar->_idx, _offset);
3939     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3940     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
3941   }
3942 }
3943 
3944 void SWPointer::Tracer::offset_plus_k_9(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3945   if(_slp->is_trace_alignment()) {
3946     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);
3947     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3948     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
3949   }
3950 }
3951 
3952 void SWPointer::Tracer::offset_plus_k_10(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3953   if(_slp->is_trace_alignment()) {
3954     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);
3955     print_depth(); tty->print_cr("  \\ %d SWPointer::offset_plus_k: is invariant", n->_idx);
3956   }
3957 }
3958 
3959 void SWPointer::Tracer::offset_plus_k_11(Node* n) {
3960   if(_slp->is_trace_alignment()) {
3961     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
3962   }
3963 }
3964 
3965 #endif
3966 // ========================= OrderedPair =====================
3967 
3968 const OrderedPair OrderedPair::initial;
3969 
3970 // ========================= SWNodeInfo =====================
3971 
3972 const SWNodeInfo SWNodeInfo::initial;
3973 
3974 
3975 // ============================ DepGraph ===========================
3976 
3977 //------------------------------make_node---------------------------
3978 // Make a new dependence graph node for an ideal node.
3979 DepMem* DepGraph::make_node(Node* node) {
3980   DepMem* m = new (_arena) DepMem(node);
3981   if (node != NULL) {
3982     assert(_map.at_grow(node->_idx) == NULL, "one init only");
3983     _map.at_put_grow(node->_idx, m);
3984   }
3985   return m;
3986 }
3987 
3988 //------------------------------make_edge---------------------------
3989 // Make a new dependence graph edge from dpred -> dsucc
3990 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
3991   DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
3992   dpred->set_out_head(e);
3993   dsucc->set_in_head(e);
3994   return e;
3995 }
3996 
3997 // ========================== DepMem ========================
3998 
3999 //------------------------------in_cnt---------------------------
4000 int DepMem::in_cnt() {
4001   int ct = 0;
4002   for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
4003   return ct;
4004 }
4005 
4006 //------------------------------out_cnt---------------------------
4007 int DepMem::out_cnt() {
4008   int ct = 0;
4009   for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
4010   return ct;
4011 }
4012 
4013 //------------------------------print-----------------------------
4014 void DepMem::print() {
4015 #ifndef PRODUCT
4016   tty->print("  DepNode %d (", _node->_idx);
4017   for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
4018     Node* pred = p->pred()->node();
4019     tty->print(" %d", pred != NULL ? pred->_idx : 0);
4020   }
4021   tty->print(") [");
4022   for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
4023     Node* succ = s->succ()->node();
4024     tty->print(" %d", succ != NULL ? succ->_idx : 0);
4025   }
4026   tty->print_cr(" ]");
4027 #endif
4028 }
4029 
4030 // =========================== DepEdge =========================
4031 
4032 //------------------------------DepPreds---------------------------
4033 void DepEdge::print() {
4034 #ifndef PRODUCT
4035   tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
4036 #endif
4037 }
4038 
4039 // =========================== DepPreds =========================
4040 // Iterator over predecessor edges in the dependence graph.
4041 
4042 //------------------------------DepPreds---------------------------
4043 DepPreds::DepPreds(Node* n, DepGraph& dg) {
4044   _n = n;
4045   _done = false;
4046   if (_n->is_Store() || _n->is_Load()) {
4047     _next_idx = MemNode::Address;
4048     _end_idx  = n->req();
4049     _dep_next = dg.dep(_n)->in_head();
4050   } else if (_n->is_Mem()) {
4051     _next_idx = 0;
4052     _end_idx  = 0;
4053     _dep_next = dg.dep(_n)->in_head();
4054   } else {
4055     _next_idx = 1;
4056     _end_idx  = _n->req();
4057     _dep_next = NULL;
4058   }
4059   next();
4060 }
4061 
4062 //------------------------------next---------------------------
4063 void DepPreds::next() {
4064   if (_dep_next != NULL) {
4065     _current  = _dep_next->pred()->node();
4066     _dep_next = _dep_next->next_in();
4067   } else if (_next_idx < _end_idx) {
4068     _current  = _n->in(_next_idx++);
4069   } else {
4070     _done = true;
4071   }
4072 }
4073 
4074 // =========================== DepSuccs =========================
4075 // Iterator over successor edges in the dependence graph.
4076 
4077 //------------------------------DepSuccs---------------------------
4078 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
4079   _n = n;
4080   _done = false;
4081   if (_n->is_Load()) {
4082     _next_idx = 0;
4083     _end_idx  = _n->outcnt();
4084     _dep_next = dg.dep(_n)->out_head();
4085   } else if (_n->is_Mem() || (_n->is_Phi() && _n->bottom_type() == Type::MEMORY)) {
4086     _next_idx = 0;
4087     _end_idx  = 0;
4088     _dep_next = dg.dep(_n)->out_head();
4089   } else {
4090     _next_idx = 0;
4091     _end_idx  = _n->outcnt();
4092     _dep_next = NULL;
4093   }
4094   next();
4095 }
4096 
4097 //-------------------------------next---------------------------
4098 void DepSuccs::next() {
4099   if (_dep_next != NULL) {
4100     _current  = _dep_next->succ()->node();
4101     _dep_next = _dep_next->next_out();
4102   } else if (_next_idx < _end_idx) {
4103     _current  = _n->raw_out(_next_idx++);
4104   } else {
4105     _done = true;
4106   }
4107 }
4108 
4109 //
4110 // --------------------------------- vectorization/simd -----------------------------------
4111 //
4112 bool SuperWord::same_origin_idx(Node* a, Node* b) const {
4113   return a != NULL && b != NULL && _clone_map.same_idx(a->_idx, b->_idx);
4114 }
4115 bool SuperWord::same_generation(Node* a, Node* b) const {
4116   return a != NULL && b != NULL && _clone_map.same_gen(a->_idx, b->_idx);
4117 }
4118 
4119 Node*  SuperWord::find_phi_for_mem_dep(LoadNode* ld) {
4120   assert(in_bb(ld), "must be in block");
4121   if (_clone_map.gen(ld->_idx) == _ii_first) {
4122 #ifndef PRODUCT
4123     if (_vector_loop_debug) {
4124       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(ld->_idx)=%d",
4125         _clone_map.gen(ld->_idx));
4126     }
4127 #endif
4128     return NULL; //we think that any ld in the first gen being vectorizable
4129   }
4130 
4131   Node* mem = ld->in(MemNode::Memory);
4132   if (mem->outcnt() <= 1) {
4133     // we don't want to remove the only edge from mem node to load
4134 #ifndef PRODUCT
4135     if (_vector_loop_debug) {
4136       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",
4137         mem->_idx, ld->_idx);
4138       ld->dump();
4139       mem->dump();
4140     }
4141 #endif
4142     return NULL;
4143   }
4144   if (!in_bb(mem) || same_generation(mem, ld)) {
4145 #ifndef PRODUCT
4146     if (_vector_loop_debug) {
4147       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(mem->_idx)=%d",
4148         _clone_map.gen(mem->_idx));
4149     }
4150 #endif
4151     return NULL; // does not depend on loop volatile node or depends on the same generation
4152   }
4153 
4154   //otherwise first node should depend on mem-phi
4155   Node* first = first_node(ld);
4156   assert(first->is_Load(), "must be Load");
4157   Node* phi = first->as_Load()->in(MemNode::Memory);
4158   if (!phi->is_Phi() || phi->bottom_type() != Type::MEMORY) {
4159 #ifndef PRODUCT
4160     if (_vector_loop_debug) {
4161       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");
4162       ld->dump();
4163       first->dump();
4164     }
4165 #endif
4166     return NULL;
4167   }
4168 
4169   Node* tail = 0;
4170   for (int m = 0; m < _mem_slice_head.length(); m++) {
4171     if (_mem_slice_head.at(m) == phi) {
4172       tail = _mem_slice_tail.at(m);
4173     }
4174   }
4175   if (tail == 0) { //test that found phi is in the list  _mem_slice_head
4176 #ifndef PRODUCT
4177     if (_vector_loop_debug) {
4178       tty->print_cr("SuperWord::find_phi_for_mem_dep load %d is not vectorizable node, its phi %d is not _mem_slice_head",
4179         ld->_idx, phi->_idx);
4180       ld->dump();
4181       phi->dump();
4182     }
4183 #endif
4184     return NULL;
4185   }
4186 
4187   // now all conditions are met
4188   return phi;
4189 }
4190 
4191 Node* SuperWord::first_node(Node* nd) {
4192   for (int ii = 0; ii < _iteration_first.length(); ii++) {
4193     Node* nnn = _iteration_first.at(ii);
4194     if (same_origin_idx(nnn, nd)) {
4195 #ifndef PRODUCT
4196       if (_vector_loop_debug) {
4197         tty->print_cr("SuperWord::first_node: %d is the first iteration node for %d (_clone_map.idx(nnn->_idx) = %d)",
4198           nnn->_idx, nd->_idx, _clone_map.idx(nnn->_idx));
4199       }
4200 #endif
4201       return nnn;
4202     }
4203   }
4204 
4205 #ifndef PRODUCT
4206   if (_vector_loop_debug) {
4207     tty->print_cr("SuperWord::first_node: did not find first iteration node for %d (_clone_map.idx(nd->_idx)=%d)",
4208       nd->_idx, _clone_map.idx(nd->_idx));
4209   }
4210 #endif
4211   return 0;
4212 }
4213 
4214 Node* SuperWord::last_node(Node* nd) {
4215   for (int ii = 0; ii < _iteration_last.length(); ii++) {
4216     Node* nnn = _iteration_last.at(ii);
4217     if (same_origin_idx(nnn, nd)) {
4218 #ifndef PRODUCT
4219       if (_vector_loop_debug) {
4220         tty->print_cr("SuperWord::last_node _clone_map.idx(nnn->_idx)=%d, _clone_map.idx(nd->_idx)=%d",
4221           _clone_map.idx(nnn->_idx), _clone_map.idx(nd->_idx));
4222       }
4223 #endif
4224       return nnn;
4225     }
4226   }
4227   return 0;
4228 }
4229 
4230 int SuperWord::mark_generations() {
4231   Node *ii_err = NULL, *tail_err = NULL;
4232   for (int i = 0; i < _mem_slice_head.length(); i++) {
4233     Node* phi  = _mem_slice_head.at(i);
4234     assert(phi->is_Phi(), "must be phi");
4235 
4236     Node* tail = _mem_slice_tail.at(i);
4237     if (_ii_last == -1) {
4238       tail_err = tail;
4239       _ii_last = _clone_map.gen(tail->_idx);
4240     }
4241     else if (_ii_last != _clone_map.gen(tail->_idx)) {
4242 #ifndef PRODUCT
4243       if (TraceSuperWord && Verbose) {
4244         tty->print_cr("SuperWord::mark_generations _ii_last error - found different generations in two tail nodes ");
4245         tail->dump();
4246         tail_err->dump();
4247       }
4248 #endif
4249       return -1;
4250     }
4251 
4252     // find first iteration in the loop
4253     for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
4254       Node* ii = phi->fast_out(i);
4255       if (in_bb(ii) && ii->is_Store()) { // we speculate that normally Stores of one and one only generation have deps from mem phi
4256         if (_ii_first == -1) {
4257           ii_err = ii;
4258           _ii_first = _clone_map.gen(ii->_idx);
4259         } else if (_ii_first != _clone_map.gen(ii->_idx)) {
4260 #ifndef PRODUCT
4261           if (TraceSuperWord && Verbose) {
4262             tty->print_cr("SuperWord::mark_generations: _ii_first was found before and not equal to one in this node (%d)", _ii_first);
4263             ii->dump();
4264             if (ii_err!= 0) {
4265               ii_err->dump();
4266             }
4267           }
4268 #endif
4269           return -1; // this phi has Stores from different generations of unroll and cannot be simd/vectorized
4270         }
4271       }
4272     }//for (DUIterator_Fast imax,
4273   }//for (int i...
4274 
4275   if (_ii_first == -1 || _ii_last == -1) {
4276     if (TraceSuperWord && Verbose) {
4277       tty->print_cr("SuperWord::mark_generations unknown error, something vent wrong");
4278     }
4279     return -1; // something vent wrong
4280   }
4281   // collect nodes in the first and last generations
4282   assert(_iteration_first.length() == 0, "_iteration_first must be empty");
4283   assert(_iteration_last.length() == 0, "_iteration_last must be empty");
4284   for (int j = 0; j < _block.length(); j++) {
4285     Node* n = _block.at(j);
4286     node_idx_t gen = _clone_map.gen(n->_idx);
4287     if ((signed)gen == _ii_first) {
4288       _iteration_first.push(n);
4289     } else if ((signed)gen == _ii_last) {
4290       _iteration_last.push(n);
4291     }
4292   }
4293 
4294   // building order of iterations
4295   if (_ii_order.length() == 0 && ii_err != 0) {
4296     assert(in_bb(ii_err) && ii_err->is_Store(), "should be Store in bb");
4297     Node* nd = ii_err;
4298     while(_clone_map.gen(nd->_idx) != _ii_last) {
4299       _ii_order.push(_clone_map.gen(nd->_idx));
4300       bool found = false;
4301       for (DUIterator_Fast imax, i = nd->fast_outs(imax); i < imax; i++) {
4302         Node* use = nd->fast_out(i);
4303         if (same_origin_idx(use, nd) && use->as_Store()->in(MemNode::Memory) == nd) {
4304           found = true;
4305           nd = use;
4306           break;
4307         }
4308       }//for
4309 
4310       if (found == false) {
4311         if (TraceSuperWord && Verbose) {
4312           tty->print_cr("SuperWord::mark_generations: Cannot build order of iterations - no dependent Store for %d", nd->_idx);
4313         }
4314         _ii_order.clear();
4315         return -1;
4316       }
4317     } //while
4318     _ii_order.push(_clone_map.gen(nd->_idx));
4319   }
4320 
4321 #ifndef PRODUCT
4322   if (_vector_loop_debug) {
4323     tty->print_cr("SuperWord::mark_generations");
4324     tty->print_cr("First generation (%d) nodes:", _ii_first);
4325     for (int ii = 0; ii < _iteration_first.length(); ii++)  _iteration_first.at(ii)->dump();
4326     tty->print_cr("Last generation (%d) nodes:", _ii_last);
4327     for (int ii = 0; ii < _iteration_last.length(); ii++)  _iteration_last.at(ii)->dump();
4328     tty->print_cr(" ");
4329 
4330     tty->print("SuperWord::List of generations: ");
4331     for (int jj = 0; jj < _ii_order.length(); ++jj) {
4332       tty->print("%d:%d ", jj, _ii_order.at(jj));
4333     }
4334     tty->print_cr(" ");
4335   }
4336 #endif
4337 
4338   return _ii_first;
4339 }
4340 
4341 bool SuperWord::fix_commutative_inputs(Node* gold, Node* fix) {
4342   assert(gold->is_Add() && fix->is_Add() || gold->is_Mul() && fix->is_Mul(), "should be only Add or Mul nodes");
4343   assert(same_origin_idx(gold, fix), "should be clones of the same node");
4344   Node* gin1 = gold->in(1);
4345   Node* gin2 = gold->in(2);
4346   Node* fin1 = fix->in(1);
4347   Node* fin2 = fix->in(2);
4348   bool swapped = false;
4349 
4350   if (in_bb(gin1) && in_bb(gin2) && in_bb(fin1) && in_bb(fin1)) {
4351     if (same_origin_idx(gin1, fin1) &&
4352         same_origin_idx(gin2, fin2)) {
4353       return true; // nothing to fix
4354     }
4355     if (same_origin_idx(gin1, fin2) &&
4356         same_origin_idx(gin2, fin1)) {
4357       fix->swap_edges(1, 2);
4358       swapped = true;
4359     }
4360   }
4361   // at least one input comes from outside of bb
4362   if (gin1->_idx == fin1->_idx)  {
4363     return true; // nothing to fix
4364   }
4365   if (!swapped && (gin1->_idx == fin2->_idx || gin2->_idx == fin1->_idx))  { //swapping is expensive, check condition first
4366     fix->swap_edges(1, 2);
4367     swapped = true;
4368   }
4369 
4370   if (swapped) {
4371 #ifndef PRODUCT
4372     if (_vector_loop_debug) {
4373       tty->print_cr("SuperWord::fix_commutative_inputs: fixed node %d", fix->_idx);
4374     }
4375 #endif
4376     return true;
4377   }
4378 
4379   if (TraceSuperWord && Verbose) {
4380     tty->print_cr("SuperWord::fix_commutative_inputs: cannot fix node %d", fix->_idx);
4381   }
4382 
4383   return false;
4384 }
4385 
4386 bool SuperWord::pack_parallel() {
4387 #ifndef PRODUCT
4388   if (_vector_loop_debug) {
4389     tty->print_cr("SuperWord::pack_parallel: START");
4390   }
4391 #endif
4392 
4393   _packset.clear();
4394 
4395   for (int ii = 0; ii < _iteration_first.length(); ii++) {
4396     Node* nd = _iteration_first.at(ii);
4397     if (in_bb(nd) && (nd->is_Load() || nd->is_Store() || nd->is_Add() || nd->is_Mul())) {
4398       Node_List* pk = new Node_List();
4399       pk->push(nd);
4400       for (int gen = 1; gen < _ii_order.length(); ++gen) {
4401         for (int kk = 0; kk < _block.length(); kk++) {
4402           Node* clone = _block.at(kk);
4403           if (same_origin_idx(clone, nd) &&
4404               _clone_map.gen(clone->_idx) == _ii_order.at(gen)) {
4405             if (nd->is_Add() || nd->is_Mul()) {
4406               fix_commutative_inputs(nd, clone);
4407             }
4408             pk->push(clone);
4409             if (pk->size() == 4) {
4410               _packset.append(pk);
4411 #ifndef PRODUCT
4412               if (_vector_loop_debug) {
4413                 tty->print_cr("SuperWord::pack_parallel: added pack ");
4414                 pk->dump();
4415               }
4416 #endif
4417               if (_clone_map.gen(clone->_idx) != _ii_last) {
4418                 pk = new Node_List();
4419               }
4420             }
4421             break;
4422           }
4423         }
4424       }//for
4425     }//if
4426   }//for
4427 
4428 #ifndef PRODUCT
4429   if (_vector_loop_debug) {
4430     tty->print_cr("SuperWord::pack_parallel: END");
4431   }
4432 #endif
4433 
4434   return true;
4435 }
4436 
4437 bool SuperWord::hoist_loads_in_graph() {
4438   GrowableArray<Node*> loads;
4439 
4440 #ifndef PRODUCT
4441   if (_vector_loop_debug) {
4442     tty->print_cr("SuperWord::hoist_loads_in_graph: total number _mem_slice_head.length() = %d", _mem_slice_head.length());
4443   }
4444 #endif
4445 
4446   for (int i = 0; i < _mem_slice_head.length(); i++) {
4447     Node* n = _mem_slice_head.at(i);
4448     if ( !in_bb(n) || !n->is_Phi() || n->bottom_type() != Type::MEMORY) {
4449       if (TraceSuperWord && Verbose) {
4450         tty->print_cr("SuperWord::hoist_loads_in_graph: skipping unexpected node n=%d", n->_idx);
4451       }
4452       continue;
4453     }
4454 
4455 #ifndef PRODUCT
4456     if (_vector_loop_debug) {
4457       tty->print_cr("SuperWord::hoist_loads_in_graph: processing phi %d  = _mem_slice_head.at(%d);", n->_idx, i);
4458     }
4459 #endif
4460 
4461     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4462       Node* ld = n->fast_out(i);
4463       if (ld->is_Load() && ld->as_Load()->in(MemNode::Memory) == n && in_bb(ld)) {
4464         for (int i = 0; i < _block.length(); i++) {
4465           Node* ld2 = _block.at(i);
4466           if (ld2->is_Load() && same_origin_idx(ld, ld2) &&
4467               !same_generation(ld, ld2)) { // <= do not collect the first generation ld
4468 #ifndef PRODUCT
4469             if (_vector_loop_debug) {
4470               tty->print_cr("SuperWord::hoist_loads_in_graph: will try to hoist load ld2->_idx=%d, cloned from %d (ld->_idx=%d)",
4471                 ld2->_idx, _clone_map.idx(ld->_idx), ld->_idx);
4472             }
4473 #endif
4474             // could not do on-the-fly, since iterator is immutable
4475             loads.push(ld2);
4476           }
4477         }// for
4478       }//if
4479     }//for (DUIterator_Fast imax,
4480   }//for (int i = 0; i
4481 
4482   for (int i = 0; i < loads.length(); i++) {
4483     LoadNode* ld = loads.at(i)->as_Load();
4484     Node* phi = find_phi_for_mem_dep(ld);
4485     if (phi != NULL) {
4486 #ifndef PRODUCT
4487       if (_vector_loop_debug) {
4488         tty->print_cr("SuperWord::hoist_loads_in_graph replacing MemNode::Memory(%d) edge in %d with one from %d",
4489           MemNode::Memory, ld->_idx, phi->_idx);
4490       }
4491 #endif
4492       _igvn.replace_input_of(ld, MemNode::Memory, phi);
4493     }
4494   }//for
4495 
4496   restart(); // invalidate all basic structures, since we rebuilt the graph
4497 
4498   if (TraceSuperWord && Verbose) {
4499     tty->print_cr("\nSuperWord::hoist_loads_in_graph() the graph was rebuilt, all structures invalidated and need rebuild");
4500   }
4501 
4502   return true;
4503 }