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