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