1 /*
   2  * Copyright (c) 2007, 2017, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  */
  23 
  24 #include "precompiled.hpp"
  25 #include "compiler/compileLog.hpp"
  26 #include "libadt/vectset.hpp"
  27 #include "memory/allocation.inline.hpp"
  28 #include "memory/resourceArea.hpp"
  29 #include "opto/addnode.hpp"
  30 #include "opto/callnode.hpp"
  31 #include "opto/castnode.hpp"
  32 #include "opto/convertnode.hpp"
  33 #include "opto/divnode.hpp"
  34 #include "opto/matcher.hpp"
  35 #include "opto/memnode.hpp"
  36 #include "opto/mulnode.hpp"
  37 #include "opto/opcodes.hpp"
  38 #include "opto/opaquenode.hpp"
  39 #include "opto/superword.hpp"
  40 #include "opto/vectornode.hpp"
  41 #include "opto/movenode.hpp"
  42 
  43 //
  44 //                  S U P E R W O R D   T R A N S F O R M
  45 //=============================================================================
  46 
  47 //------------------------------SuperWord---------------------------
  48 SuperWord::SuperWord(PhaseIdealLoop* phase) :
  49   _phase(phase),
  50   _igvn(phase->_igvn),
  51   _arena(phase->C->comp_arena()),
  52   _packset(arena(), 8,  0, NULL),         // packs for the current block
  53   _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
  54   _block(arena(), 8,  0, NULL),           // nodes in current block
  55   _post_block(arena(), 8, 0, NULL),       // nodes common to current block which are marked as post loop vectorizable
  56   _data_entry(arena(), 8,  0, NULL),      // nodes with all inputs from outside
  57   _mem_slice_head(arena(), 8,  0, NULL),  // memory slice heads
  58   _mem_slice_tail(arena(), 8,  0, NULL),  // memory slice tails
  59   _node_info(arena(), 8,  0, SWNodeInfo::initial), // info needed per node
  60   _clone_map(phase->C->clone_map()),      // map of nodes created in cloning
  61   _cmovev_kit(_arena, this),              // map to facilitate CMoveVD creation
  62   _align_to_ref(NULL),                    // memory reference to align vectors to
  63   _disjoint_ptrs(arena(), 8,  0, OrderedPair::initial), // runtime disambiguated pointer pairs
  64   _dg(_arena),                            // dependence graph
  65   _visited(arena()),                      // visited node set
  66   _post_visited(arena()),                 // post visited node set
  67   _n_idx_list(arena(), 8),                // scratch list of (node,index) pairs
  68   _stk(arena(), 8, 0, NULL),              // scratch stack of nodes
  69   _nlist(arena(), 8, 0, NULL),            // scratch list of nodes
  70   _lpt(NULL),                             // loop tree node
  71   _lp(NULL),                              // LoopNode
  72   _bb(NULL),                              // basic block
  73   _iv(NULL),                              // induction var
  74   _race_possible(false),                  // cases where SDMU is true
  75   _early_return(true),                    // analysis evaluations routine
  76   _num_work_vecs(0),                      // amount of vector work we have
  77   _num_reductions(0),                     // amount of reduction work we have
  78   _do_vector_loop(phase->C->do_vector_loop()),  // whether to do vectorization/simd style
  79   _do_reserve_copy(DoReserveCopyInSuperWord),
  80   _ii_first(-1),                          // first loop generation index - only if do_vector_loop()
  81   _ii_last(-1),                           // last loop generation index - only if do_vector_loop()
  82   _ii_order(arena(), 8, 0, 0)
  83 {
  84 #ifndef PRODUCT
  85   _vector_loop_debug = 0;
  86   if (_phase->C->method() != NULL) {
  87     _vector_loop_debug = phase->C->directive()->VectorizeDebugOption;
  88   }
  89 
  90 #endif
  91 }
  92 
  93 //------------------------------transform_loop---------------------------
  94 void SuperWord::transform_loop(IdealLoopTree* lpt, bool do_optimization) {
  95   assert(UseSuperWord, "should be");
  96   // Do vectors exist on this architecture?
  97   if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return;
  98 
  99   assert(lpt->_head->is_CountedLoop(), "must be");
 100   CountedLoopNode *cl = lpt->_head->as_CountedLoop();
 101 
 102   if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
 103 
 104   bool post_loop_allowed = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
 105   if (post_loop_allowed) {
 106     if (cl->is_reduction_loop()) return; // no predication mapping
 107     Node *limit = cl->limit();
 108     if (limit->is_Con()) return; // non constant limits only
 109     // Now check the limit for expressions we do not handle
 110     if (limit->is_Add()) {
 111       Node *in2 = limit->in(2);
 112       if (in2->is_Con()) {
 113         int val = in2->get_int();
 114         // should not try to program these cases
 115         if (val < 0) return;
 116       }
 117     }
 118   }
 119 
 120   // skip any loop that has not been assigned max unroll by analysis
 121   if (do_optimization) {
 122     if (SuperWordLoopUnrollAnalysis && cl->slp_max_unroll() == 0) return;
 123   }
 124 
 125   // Check for no control flow in body (other than exit)
 126   Node *cl_exit = cl->loopexit();
 127   if (cl->is_main_loop() && (cl_exit->in(0) != lpt->_head)) {
 128     #ifndef PRODUCT
 129       if (TraceSuperWord) {
 130         tty->print_cr("SuperWord::transform_loop: loop too complicated, cl_exit->in(0) != lpt->_head");
 131         tty->print("cl_exit %d", cl_exit->_idx); cl_exit->dump();
 132         tty->print("cl_exit->in(0) %d", cl_exit->in(0)->_idx); cl_exit->in(0)->dump();
 133         tty->print("lpt->_head %d", lpt->_head->_idx); lpt->_head->dump();
 134         lpt->dump_head();
 135       }
 136     #endif
 137     return;
 138   }
 139 
 140   // Make sure the are no extra control users of the loop backedge
 141   if (cl->back_control()->outcnt() != 1) {
 142     return;
 143   }
 144 
 145   // Skip any loops already optimized by slp
 146   if (cl->is_vectorized_loop()) return;
 147 
 148   if (cl->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     bool small_basic_type = false;
 302     bool flag_small_bt = false;
 303     for (uint i = 0; i < lpt()->_body.size(); i++) {
 304       if (ignored_loop_nodes[i] != -1) continue;
 305 
 306       BasicType bt;
 307       Node* n = lpt()->_body.at(i);
 308       if (n->is_Mem()) {
 309         bt = n->as_Mem()->memory_type();
 310       } else {
 311         bt = n->bottom_type()->basic_type();
 312       }
 313 
 314       if (post_loop_allowed) {
 315         if (!small_basic_type) {
 316           switch (bt) {
 317           case T_CHAR:
 318           case T_BYTE:
 319           case T_SHORT:
 320             small_basic_type = true;
 321             break;
 322 
 323           case T_LONG:
 324             // TODO: Remove when support completed for mask context with LONG.
 325             //       Support needs to be augmented for logical qword operations, currently we map to dword
 326             //       buckets for vectors on logicals as these were legacy.
 327             small_basic_type = true;
 328             break;
 329 
 330           default:
 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 
 340       // If a max vector exists which is not larger than _local_loop_unroll_factor
 341       // stop looking, we already have the max vector to map to.
 342       if (cur_max_vector < local_loop_unroll_factor) {
 343         is_slp = false;
 344         if (TraceSuperWordLoopUnrollAnalysis) {
 345           tty->print_cr("slp analysis fails: unroll limit greater than max vector\n");
 346         }
 347         break;
 348       }
 349 
 350       // Map the maximal common vector
 351       if (VectorNode::implemented(n->Opcode(), cur_max_vector, bt)) {
 352         if (cur_max_vector < max_vector && !flag_small_bt) {
 353           max_vector = cur_max_vector;
 354         } else if (cur_max_vector > max_vector && UseSubwordForMaxVector) {
 355           // Analyse subword in the loop to set maximum vector size to take advantage of full vector width for subword types.
 356           // Here we analyze if narrowing is likely to happen and if it is we set vector size more aggressively.
 357           // We check for possibility of narrowing by looking through chain operations using subword types.
 358           if (is_subword_type(bt)) {
 359             uint start, end;
 360             VectorNode::vector_operands(n, &start, &end);
 361 
 362             for (uint j = start; j < end; j++) {
 363               Node* in = n->in(j);
 364               // Don't propagate through a memory
 365               if (!in->is_Mem() && in_bb(in) && in->bottom_type()->basic_type() == T_INT) {
 366                 bool same_type = true;
 367                 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
 368                   Node *use = in->fast_out(k);
 369                   if (!in_bb(use) && use->bottom_type()->basic_type() != bt) {
 370                     same_type = false;
 371                     break;
 372                   }
 373                 }
 374                 if (same_type) {
 375                   max_vector = cur_max_vector;
 376                   flag_small_bt = true;
 377                 }
 378               }
 379             }
 380           }
 381         }
 382         // We only process post loops on predicated targets where we want to
 383         // mask map the loop to a single iteration
 384         if (post_loop_allowed) {
 385           _post_block.at_put_grow(rpo_idx++, n);
 386         }
 387       }
 388     }
 389     if (is_slp) {
 390       local_loop_unroll_factor = max_vector;
 391       cl->mark_passed_slp();
 392     }
 393     cl->mark_was_slp();
 394     if (cl->is_main_loop()) {
 395       cl->set_slp_max_unroll(local_loop_unroll_factor);
 396     } else if (post_loop_allowed) {
 397       if (!small_basic_type) {
 398         // avoid replication context for small basic types in programmable masked loops
 399         cl->set_slp_max_unroll(local_loop_unroll_factor);
 400       }
 401     }
 402   }
 403 }
 404 
 405 //------------------------------SLP_extract---------------------------
 406 // Extract the superword level parallelism
 407 //
 408 // 1) A reverse post-order of nodes in the block is constructed.  By scanning
 409 //    this list from first to last, all definitions are visited before their uses.
 410 //
 411 // 2) A point-to-point dependence graph is constructed between memory references.
 412 //    This simplies the upcoming "independence" checker.
 413 //
 414 // 3) The maximum depth in the node graph from the beginning of the block
 415 //    to each node is computed.  This is used to prune the graph search
 416 //    in the independence checker.
 417 //
 418 // 4) For integer types, the necessary bit width is propagated backwards
 419 //    from stores to allow packed operations on byte, char, and short
 420 //    integers.  This reverses the promotion to type "int" that javac
 421 //    did for operations like: char c1,c2,c3;  c1 = c2 + c3.
 422 //
 423 // 5) One of the memory references is picked to be an aligned vector reference.
 424 //    The pre-loop trip count is adjusted to align this reference in the
 425 //    unrolled body.
 426 //
 427 // 6) The initial set of pack pairs is seeded with memory references.
 428 //
 429 // 7) The set of pack pairs is extended by following use->def and def->use links.
 430 //
 431 // 8) The pairs are combined into vector sized packs.
 432 //
 433 // 9) Reorder the memory slices to co-locate members of the memory packs.
 434 //
 435 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
 436 //    inserting scalar promotion, vector creation from multiple scalars, and
 437 //    extraction of scalar values from vectors.
 438 //
 439 void SuperWord::SLP_extract() {
 440 
 441 #ifndef PRODUCT
 442   if (_do_vector_loop && TraceSuperWord) {
 443     tty->print("SuperWord::SLP_extract\n");
 444     tty->print("input loop\n");
 445     _lpt->dump_head();
 446     _lpt->dump();
 447     for (uint i = 0; i < _lpt->_body.size(); i++) {
 448       _lpt->_body.at(i)->dump();
 449     }
 450   }
 451 #endif
 452   // Ready the block
 453   if (!construct_bb()) {
 454     return; // Exit if no interesting nodes or complex graph.
 455   }
 456 
 457   // build    _dg, _disjoint_ptrs
 458   dependence_graph();
 459 
 460   // compute function depth(Node*)
 461   compute_max_depth();
 462 
 463   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
 464   bool post_loop_allowed = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
 465   if (cl->is_main_loop()) {
 466     if (_do_vector_loop) {
 467       if (mark_generations() != -1) {
 468         hoist_loads_in_graph(); // this only rebuild the graph; all basic structs need rebuild explicitly
 469 
 470         if (!construct_bb()) {
 471           return; // Exit if no interesting nodes or complex graph.
 472         }
 473         dependence_graph();
 474         compute_max_depth();
 475       }
 476 
 477 #ifndef PRODUCT
 478       if (TraceSuperWord) {
 479         tty->print_cr("\nSuperWord::_do_vector_loop: graph after hoist_loads_in_graph");
 480         _lpt->dump_head();
 481         for (int j = 0; j < _block.length(); j++) {
 482           Node* n = _block.at(j);
 483           int d = depth(n);
 484           for (int i = 0; i < d; i++) tty->print("%s", "  ");
 485           tty->print("%d :", d);
 486           n->dump();
 487         }
 488       }
 489 #endif
 490     }
 491 
 492     compute_vector_element_type();
 493 
 494     // Attempt vectorization
 495 
 496     find_adjacent_refs();
 497 
 498     extend_packlist();
 499 
 500     if (_do_vector_loop) {
 501       if (_packset.length() == 0) {
 502         if (TraceSuperWord) {
 503           tty->print_cr("\nSuperWord::_do_vector_loop DFA could not build packset, now trying to build anyway");
 504         }
 505         pack_parallel();
 506       }
 507     }
 508 
 509     combine_packs();
 510 
 511     construct_my_pack_map();
 512 
 513     if (_do_vector_loop) {
 514       merge_packs_to_cmovd();
 515     }
 516 
 517     filter_packs();
 518 
 519     schedule();
 520   } else if (post_loop_allowed) {
 521     int saved_mapped_unroll_factor = cl->slp_max_unroll();
 522     if (saved_mapped_unroll_factor) {
 523       int vector_mapped_unroll_factor = saved_mapped_unroll_factor;
 524 
 525       // now reset the slp_unroll_factor so that we can check the analysis mapped
 526       // what the vector loop was mapped to
 527       cl->set_slp_max_unroll(0);
 528 
 529       // do the analysis on the post loop
 530       unrolling_analysis(vector_mapped_unroll_factor);
 531 
 532       // if our analyzed loop is a canonical fit, start processing it
 533       if (vector_mapped_unroll_factor == saved_mapped_unroll_factor) {
 534         // now add the vector nodes to packsets
 535         for (int i = 0; i < _post_block.length(); i++) {
 536           Node* n = _post_block.at(i);
 537           Node_List* singleton = new Node_List();
 538           singleton->push(n);
 539           _packset.append(singleton);
 540           set_my_pack(n, singleton);
 541         }
 542 
 543         // map base types for vector usage
 544         compute_vector_element_type();
 545       } else {
 546         return;
 547       }
 548     } else {
 549       // for some reason we could not map the slp analysis state of the vectorized loop
 550       return;
 551     }
 552   }
 553 
 554   output();
 555 }
 556 
 557 //------------------------------find_adjacent_refs---------------------------
 558 // Find the adjacent memory references and create pack pairs for them.
 559 // This is the initial set of packs that will then be extended by
 560 // following use->def and def->use links.  The align positions are
 561 // assigned relative to the reference "align_to_ref"
 562 void SuperWord::find_adjacent_refs() {
 563   // Get list of memory operations
 564   Node_List memops;
 565   for (int i = 0; i < _block.length(); i++) {
 566     Node* n = _block.at(i);
 567     if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
 568         is_java_primitive(n->as_Mem()->memory_type())) {
 569       int align = memory_alignment(n->as_Mem(), 0);
 570       if (align != bottom_align) {
 571         memops.push(n);
 572       }
 573     }
 574   }
 575 
 576   Node_List align_to_refs;
 577   int best_iv_adjustment = 0;
 578   MemNode* best_align_to_mem_ref = NULL;
 579 
 580   while (memops.size() != 0) {
 581     // Find a memory reference to align to.
 582     MemNode* mem_ref = find_align_to_ref(memops);
 583     if (mem_ref == NULL) break;
 584     align_to_refs.push(mem_ref);
 585     int iv_adjustment = get_iv_adjustment(mem_ref);
 586 
 587     if (best_align_to_mem_ref == NULL) {
 588       // Set memory reference which is the best from all memory operations
 589       // to be used for alignment. The pre-loop trip count is modified to align
 590       // this reference to a vector-aligned address.
 591       best_align_to_mem_ref = mem_ref;
 592       best_iv_adjustment = iv_adjustment;
 593       NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
 594     }
 595 
 596     SWPointer align_to_ref_p(mem_ref, this, NULL, false);
 597     // Set alignment relative to "align_to_ref" for all related memory operations.
 598     for (int i = memops.size() - 1; i >= 0; i--) {
 599       MemNode* s = memops.at(i)->as_Mem();
 600       if (isomorphic(s, mem_ref) &&
 601            (!_do_vector_loop || same_origin_idx(s, mem_ref))) {
 602         SWPointer p2(s, this, NULL, false);
 603         if (p2.comparable(align_to_ref_p)) {
 604           int align = memory_alignment(s, iv_adjustment);
 605           set_alignment(s, align);
 606         }
 607       }
 608     }
 609 
 610     // Create initial pack pairs of memory operations for which
 611     // alignment is set and vectors will be aligned.
 612     bool create_pack = true;
 613     if (memory_alignment(mem_ref, best_iv_adjustment) == 0 || _do_vector_loop) {
 614       if (!Matcher::misaligned_vectors_ok()) {
 615         int vw = vector_width(mem_ref);
 616         int vw_best = vector_width(best_align_to_mem_ref);
 617         if (vw > vw_best) {
 618           // Do not vectorize a memory access with more elements per vector
 619           // if unaligned memory access is not allowed because number of
 620           // iterations in pre-loop will be not enough to align it.
 621           create_pack = false;
 622         } else {
 623           SWPointer p2(best_align_to_mem_ref, this, NULL, false);
 624           if (align_to_ref_p.invar() != p2.invar()) {
 625             // Do not vectorize memory accesses with different invariants
 626             // if unaligned memory accesses are not allowed.
 627             create_pack = false;
 628           }
 629         }
 630       }
 631     } else {
 632       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 633         // Can't allow vectorization of unaligned memory accesses with the
 634         // same type since it could be overlapped accesses to the same array.
 635         create_pack = false;
 636       } else {
 637         // Allow independent (different type) unaligned memory operations
 638         // if HW supports them.
 639         if (!Matcher::misaligned_vectors_ok()) {
 640           create_pack = false;
 641         } else {
 642           // Check if packs of the same memory type but
 643           // with a different alignment were created before.
 644           for (uint i = 0; i < align_to_refs.size(); i++) {
 645             MemNode* mr = align_to_refs.at(i)->as_Mem();
 646             if (same_velt_type(mr, mem_ref) &&
 647                 memory_alignment(mr, iv_adjustment) != 0)
 648               create_pack = false;
 649           }
 650         }
 651       }
 652     }
 653     if (create_pack) {
 654       for (uint i = 0; i < memops.size(); i++) {
 655         Node* s1 = memops.at(i);
 656         int align = alignment(s1);
 657         if (align == top_align) continue;
 658         for (uint j = 0; j < memops.size(); j++) {
 659           Node* s2 = memops.at(j);
 660           if (alignment(s2) == top_align) continue;
 661           if (s1 != s2 && are_adjacent_refs(s1, s2)) {
 662             if (stmts_can_pack(s1, s2, align)) {
 663               Node_List* pair = new Node_List();
 664               pair->push(s1);
 665               pair->push(s2);
 666               if (!_do_vector_loop || same_origin_idx(s1, s2)) {
 667                 _packset.append(pair);
 668               }
 669             }
 670           }
 671         }
 672       }
 673     } else { // Don't create unaligned pack
 674       // First, remove remaining memory ops of the same type from the list.
 675       for (int i = memops.size() - 1; i >= 0; i--) {
 676         MemNode* s = memops.at(i)->as_Mem();
 677         if (same_velt_type(s, mem_ref)) {
 678           memops.remove(i);
 679         }
 680       }
 681 
 682       // Second, remove already constructed packs of the same type.
 683       for (int i = _packset.length() - 1; i >= 0; i--) {
 684         Node_List* p = _packset.at(i);
 685         MemNode* s = p->at(0)->as_Mem();
 686         if (same_velt_type(s, mem_ref)) {
 687           remove_pack_at(i);
 688         }
 689       }
 690 
 691       // If needed find the best memory reference for loop alignment again.
 692       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 693         // Put memory ops from remaining packs back on memops list for
 694         // the best alignment search.
 695         uint orig_msize = memops.size();
 696         for (int i = 0; i < _packset.length(); i++) {
 697           Node_List* p = _packset.at(i);
 698           MemNode* s = p->at(0)->as_Mem();
 699           assert(!same_velt_type(s, mem_ref), "sanity");
 700           memops.push(s);
 701         }
 702         best_align_to_mem_ref = find_align_to_ref(memops);
 703         if (best_align_to_mem_ref == NULL) {
 704           if (TraceSuperWord) {
 705             tty->print_cr("SuperWord::find_adjacent_refs(): best_align_to_mem_ref == NULL");
 706           }
 707           break;
 708         }
 709         best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
 710         NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
 711         // Restore list.
 712         while (memops.size() > orig_msize)
 713           (void)memops.pop();
 714       }
 715     } // unaligned memory accesses
 716 
 717     // Remove used mem nodes.
 718     for (int i = memops.size() - 1; i >= 0; i--) {
 719       MemNode* m = memops.at(i)->as_Mem();
 720       if (alignment(m) != top_align) {
 721         memops.remove(i);
 722       }
 723     }
 724 
 725   } // while (memops.size() != 0
 726   set_align_to_ref(best_align_to_mem_ref);
 727 
 728   if (TraceSuperWord) {
 729     tty->print_cr("\nAfter find_adjacent_refs");
 730     print_packset();
 731   }
 732 }
 733 
 734 #ifndef PRODUCT
 735 void SuperWord::find_adjacent_refs_trace_1(Node* best_align_to_mem_ref, int best_iv_adjustment) {
 736   if (is_trace_adjacent()) {
 737     tty->print("SuperWord::find_adjacent_refs best_align_to_mem_ref = %d, best_iv_adjustment = %d",
 738        best_align_to_mem_ref->_idx, best_iv_adjustment);
 739        best_align_to_mem_ref->dump();
 740   }
 741 }
 742 #endif
 743 
 744 //------------------------------find_align_to_ref---------------------------
 745 // Find a memory reference to align the loop induction variable to.
 746 // Looks first at stores then at loads, looking for a memory reference
 747 // with the largest number of references similar to it.
 748 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
 749   GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
 750 
 751   // Count number of comparable memory ops
 752   for (uint i = 0; i < memops.size(); i++) {
 753     MemNode* s1 = memops.at(i)->as_Mem();
 754     SWPointer p1(s1, this, NULL, false);
 755     // Discard if pre loop can't align this reference
 756     if (!ref_is_alignable(p1)) {
 757       *cmp_ct.adr_at(i) = 0;
 758       continue;
 759     }
 760     for (uint j = i+1; j < memops.size(); j++) {
 761       MemNode* s2 = memops.at(j)->as_Mem();
 762       if (isomorphic(s1, s2)) {
 763         SWPointer p2(s2, this, NULL, false);
 764         if (p1.comparable(p2)) {
 765           (*cmp_ct.adr_at(i))++;
 766           (*cmp_ct.adr_at(j))++;
 767         }
 768       }
 769     }
 770   }
 771 
 772   // Find Store (or Load) with the greatest number of "comparable" references,
 773   // biggest vector size, smallest data size and smallest iv offset.
 774   int max_ct        = 0;
 775   int max_vw        = 0;
 776   int max_idx       = -1;
 777   int min_size      = max_jint;
 778   int min_iv_offset = max_jint;
 779   for (uint j = 0; j < memops.size(); j++) {
 780     MemNode* s = memops.at(j)->as_Mem();
 781     if (s->is_Store()) {
 782       int vw = vector_width_in_bytes(s);
 783       assert(vw > 1, "sanity");
 784       SWPointer p(s, this, NULL, false);
 785       if ( cmp_ct.at(j) >  max_ct ||
 786           (cmp_ct.at(j) == max_ct &&
 787             ( vw >  max_vw ||
 788              (vw == max_vw &&
 789               ( data_size(s) <  min_size ||
 790                (data_size(s) == min_size &&
 791                 p.offset_in_bytes() < min_iv_offset)))))) {
 792         max_ct = cmp_ct.at(j);
 793         max_vw = vw;
 794         max_idx = j;
 795         min_size = data_size(s);
 796         min_iv_offset = p.offset_in_bytes();
 797       }
 798     }
 799   }
 800   // If no stores, look at loads
 801   if (max_ct == 0) {
 802     for (uint j = 0; j < memops.size(); j++) {
 803       MemNode* s = memops.at(j)->as_Mem();
 804       if (s->is_Load()) {
 805         int vw = vector_width_in_bytes(s);
 806         assert(vw > 1, "sanity");
 807         SWPointer p(s, this, NULL, false);
 808         if ( cmp_ct.at(j) >  max_ct ||
 809             (cmp_ct.at(j) == max_ct &&
 810               ( vw >  max_vw ||
 811                (vw == max_vw &&
 812                 ( data_size(s) <  min_size ||
 813                  (data_size(s) == min_size &&
 814                   p.offset_in_bytes() < min_iv_offset)))))) {
 815           max_ct = cmp_ct.at(j);
 816           max_vw = vw;
 817           max_idx = j;
 818           min_size = data_size(s);
 819           min_iv_offset = p.offset_in_bytes();
 820         }
 821       }
 822     }
 823   }
 824 
 825 #ifdef ASSERT
 826   if (TraceSuperWord && Verbose) {
 827     tty->print_cr("\nVector memops after find_align_to_ref");
 828     for (uint i = 0; i < memops.size(); i++) {
 829       MemNode* s = memops.at(i)->as_Mem();
 830       s->dump();
 831     }
 832   }
 833 #endif
 834 
 835   if (max_ct > 0) {
 836 #ifdef ASSERT
 837     if (TraceSuperWord) {
 838       tty->print("\nVector align to node: ");
 839       memops.at(max_idx)->as_Mem()->dump();
 840     }
 841 #endif
 842     return memops.at(max_idx)->as_Mem();
 843   }
 844   return NULL;
 845 }
 846 
 847 //------------------------------ref_is_alignable---------------------------
 848 // Can the preloop align the reference to position zero in the vector?
 849 bool SuperWord::ref_is_alignable(SWPointer& p) {
 850   if (!p.has_iv()) {
 851     return true;   // no induction variable
 852   }
 853   CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
 854   assert(pre_end != NULL, "we must have a correct pre-loop");
 855   assert(pre_end->stride_is_con(), "pre loop stride is constant");
 856   int preloop_stride = pre_end->stride_con();
 857 
 858   int span = preloop_stride * p.scale_in_bytes();
 859   int mem_size = p.memory_size();
 860   int offset   = p.offset_in_bytes();
 861   // Stride one accesses are alignable if offset is aligned to memory operation size.
 862   // Offset can be unaligned when UseUnalignedAccesses is used.
 863   if (ABS(span) == mem_size && (ABS(offset) % mem_size) == 0) {
 864     return true;
 865   }
 866   // If the initial offset from start of the object is computable,
 867   // check if the pre-loop can align the final offset accordingly.
 868   //
 869   // In other words: Can we find an i such that the offset
 870   // after i pre-loop iterations is aligned to vw?
 871   //   (init_offset + pre_loop) % vw == 0              (1)
 872   // where
 873   //   pre_loop = i * span
 874   // is the number of bytes added to the offset by i pre-loop iterations.
 875   //
 876   // For this to hold we need pre_loop to increase init_offset by
 877   //   pre_loop = vw - (init_offset % vw)
 878   //
 879   // This is only possible if pre_loop is divisible by span because each
 880   // pre-loop iteration increases the initial offset by 'span' bytes:
 881   //   (vw - (init_offset % vw)) % span == 0
 882   //
 883   int vw = vector_width_in_bytes(p.mem());
 884   assert(vw > 1, "sanity");
 885   Node* init_nd = pre_end->init_trip();
 886   if (init_nd->is_Con() && p.invar() == NULL) {
 887     int init = init_nd->bottom_type()->is_int()->get_con();
 888     int init_offset = init * p.scale_in_bytes() + offset;
 889     assert(init_offset >= 0, "positive offset from object start");
 890     if (vw % span == 0) {
 891       // If vm is a multiple of span, we use formula (1).
 892       if (span > 0) {
 893         return (vw - (init_offset % vw)) % span == 0;
 894       } else {
 895         assert(span < 0, "nonzero stride * scale");
 896         return (init_offset % vw) % -span == 0;
 897       }
 898     } else if (span % vw == 0) {
 899       // If span is a multiple of vw, we can simplify formula (1) to:
 900       //   (init_offset + i * span) % vw == 0
 901       //     =>
 902       //   (init_offset % vw) + ((i * span) % vw) == 0
 903       //     =>
 904       //   init_offset % vw == 0
 905       //
 906       // Because we add a multiple of vw to the initial offset, the final
 907       // offset is a multiple of vw if and only if init_offset is a multiple.
 908       //
 909       return (init_offset % vw) == 0;
 910     }
 911   }
 912   return false;
 913 }
 914 
 915 //---------------------------get_iv_adjustment---------------------------
 916 // Calculate loop's iv adjustment for this memory ops.
 917 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
 918   SWPointer align_to_ref_p(mem_ref, this, NULL, false);
 919   int offset = align_to_ref_p.offset_in_bytes();
 920   int scale  = align_to_ref_p.scale_in_bytes();
 921   int elt_size = align_to_ref_p.memory_size();
 922   int vw       = vector_width_in_bytes(mem_ref);
 923   assert(vw > 1, "sanity");
 924   int iv_adjustment;
 925   if (scale != 0) {
 926     int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
 927     // At least one iteration is executed in pre-loop by default. As result
 928     // several iterations are needed to align memory operations in main-loop even
 929     // if offset is 0.
 930     int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
 931     assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
 932            "(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size);
 933     iv_adjustment = iv_adjustment_in_bytes/elt_size;
 934   } else {
 935     // This memory op is not dependent on iv (scale == 0)
 936     iv_adjustment = 0;
 937   }
 938 
 939 #ifndef PRODUCT
 940   if (TraceSuperWord) {
 941     tty->print("SuperWord::get_iv_adjustment: n = %d, noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d: ",
 942       mem_ref->_idx, offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
 943     mem_ref->dump();
 944   }
 945 #endif
 946   return iv_adjustment;
 947 }
 948 
 949 //---------------------------dependence_graph---------------------------
 950 // Construct dependency graph.
 951 // Add dependence edges to load/store nodes for memory dependence
 952 //    A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
 953 void SuperWord::dependence_graph() {
 954   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
 955   // First, assign a dependence node to each memory node
 956   for (int i = 0; i < _block.length(); i++ ) {
 957     Node *n = _block.at(i);
 958     if (n->is_Mem() || (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
 959       _dg.make_node(n);
 960     }
 961   }
 962 
 963   // For each memory slice, create the dependences
 964   for (int i = 0; i < _mem_slice_head.length(); i++) {
 965     Node* n      = _mem_slice_head.at(i);
 966     Node* n_tail = _mem_slice_tail.at(i);
 967 
 968     // Get slice in predecessor order (last is first)
 969     if (cl->is_main_loop()) {
 970       mem_slice_preds(n_tail, n, _nlist);
 971     }
 972 
 973 #ifndef PRODUCT
 974     if(TraceSuperWord && Verbose) {
 975       tty->print_cr("SuperWord::dependence_graph: built a new mem slice");
 976       for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 977         _nlist.at(j)->dump();
 978       }
 979     }
 980 #endif
 981     // Make the slice dependent on the root
 982     DepMem* slice = _dg.dep(n);
 983     _dg.make_edge(_dg.root(), slice);
 984 
 985     // Create a sink for the slice
 986     DepMem* slice_sink = _dg.make_node(NULL);
 987     _dg.make_edge(slice_sink, _dg.tail());
 988 
 989     // Now visit each pair of memory ops, creating the edges
 990     for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 991       Node* s1 = _nlist.at(j);
 992 
 993       // If no dependency yet, use slice
 994       if (_dg.dep(s1)->in_cnt() == 0) {
 995         _dg.make_edge(slice, s1);
 996       }
 997       SWPointer p1(s1->as_Mem(), this, NULL, false);
 998       bool sink_dependent = true;
 999       for (int k = j - 1; k >= 0; k--) {
1000         Node* s2 = _nlist.at(k);
1001         if (s1->is_Load() && s2->is_Load())
1002           continue;
1003         SWPointer p2(s2->as_Mem(), this, NULL, false);
1004 
1005         int cmp = p1.cmp(p2);
1006         if (SuperWordRTDepCheck &&
1007             p1.base() != p2.base() && p1.valid() && p2.valid()) {
1008           // Create a runtime check to disambiguate
1009           OrderedPair pp(p1.base(), p2.base());
1010           _disjoint_ptrs.append_if_missing(pp);
1011         } else if (!SWPointer::not_equal(cmp)) {
1012           // Possibly same address
1013           _dg.make_edge(s1, s2);
1014           sink_dependent = false;
1015         }
1016       }
1017       if (sink_dependent) {
1018         _dg.make_edge(s1, slice_sink);
1019       }
1020     }
1021 
1022     if (TraceSuperWord) {
1023       tty->print_cr("\nDependence graph for slice: %d", n->_idx);
1024       for (int q = 0; q < _nlist.length(); q++) {
1025         _dg.print(_nlist.at(q));
1026       }
1027       tty->cr();
1028     }
1029 
1030     _nlist.clear();
1031   }
1032 
1033   if (TraceSuperWord) {
1034     tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
1035     for (int r = 0; r < _disjoint_ptrs.length(); r++) {
1036       _disjoint_ptrs.at(r).print();
1037       tty->cr();
1038     }
1039     tty->cr();
1040   }
1041 
1042 }
1043 
1044 //---------------------------mem_slice_preds---------------------------
1045 // Return a memory slice (node list) in predecessor order starting at "start"
1046 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
1047   assert(preds.length() == 0, "start empty");
1048   Node* n = start;
1049   Node* prev = NULL;
1050   while (true) {
1051     NOT_PRODUCT( if(is_trace_mem_slice()) tty->print_cr("SuperWord::mem_slice_preds: n %d", n->_idx);)
1052     assert(in_bb(n), "must be in block");
1053     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1054       Node* out = n->fast_out(i);
1055       if (out->is_Load()) {
1056         if (in_bb(out)) {
1057           preds.push(out);
1058           if (TraceSuperWord && Verbose) {
1059             tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", out->_idx);
1060           }
1061         }
1062       } else {
1063         // FIXME
1064         if (out->is_MergeMem() && !in_bb(out)) {
1065           // Either unrolling is causing a memory edge not to disappear,
1066           // or need to run igvn.optimize() again before SLP
1067         } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
1068           // Ditto.  Not sure what else to check further.
1069         } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
1070           // StoreCM has an input edge used as a precedence edge.
1071           // Maybe an issue when oop stores are vectorized.
1072         } else {
1073           assert(out == prev || prev == NULL, "no branches off of store slice");
1074         }
1075       }//else
1076     }//for
1077     if (n == stop) break;
1078     preds.push(n);
1079     if (TraceSuperWord && Verbose) {
1080       tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", n->_idx);
1081     }
1082     prev = n;
1083     assert(n->is_Mem(), "unexpected node %s", n->Name());
1084     n = n->in(MemNode::Memory);
1085   }
1086 }
1087 
1088 //------------------------------stmts_can_pack---------------------------
1089 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
1090 // s1 aligned at "align"
1091 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
1092 
1093   // Do not use superword for non-primitives
1094   BasicType bt1 = velt_basic_type(s1);
1095   BasicType bt2 = velt_basic_type(s2);
1096   if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
1097     return false;
1098   if (Matcher::max_vector_size(bt1) < 2) {
1099     return false; // No vectors for this type
1100   }
1101 
1102   if (isomorphic(s1, s2)) {
1103     if (independent(s1, s2) || reduction(s1, s2)) {
1104       if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
1105         if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
1106           int s1_align = alignment(s1);
1107           int s2_align = alignment(s2);
1108           if (s1_align == top_align || s1_align == align) {
1109             if (s2_align == top_align || s2_align == align + data_size(s1)) {
1110               return true;
1111             }
1112           }
1113         }
1114       }
1115     }
1116   }
1117   return false;
1118 }
1119 
1120 //------------------------------exists_at---------------------------
1121 // Does s exist in a pack at position pos?
1122 bool SuperWord::exists_at(Node* s, uint pos) {
1123   for (int i = 0; i < _packset.length(); i++) {
1124     Node_List* p = _packset.at(i);
1125     if (p->at(pos) == s) {
1126       return true;
1127     }
1128   }
1129   return false;
1130 }
1131 
1132 //------------------------------are_adjacent_refs---------------------------
1133 // Is s1 immediately before s2 in memory?
1134 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
1135   if (!s1->is_Mem() || !s2->is_Mem()) return false;
1136   if (!in_bb(s1)    || !in_bb(s2))    return false;
1137 
1138   // Do not use superword for non-primitives
1139   if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
1140       !is_java_primitive(s2->as_Mem()->memory_type())) {
1141     return false;
1142   }
1143 
1144   // FIXME - co_locate_pack fails on Stores in different mem-slices, so
1145   // only pack memops that are in the same alias set until that's fixed.
1146   if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
1147       _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
1148     return false;
1149   SWPointer p1(s1->as_Mem(), this, NULL, false);
1150   SWPointer p2(s2->as_Mem(), this, NULL, false);
1151   if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
1152   int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
1153   return diff == data_size(s1);
1154 }
1155 
1156 //------------------------------isomorphic---------------------------
1157 // Are s1 and s2 similar?
1158 bool SuperWord::isomorphic(Node* s1, Node* s2) {
1159   if (s1->Opcode() != s2->Opcode()) return false;
1160   if (s1->req() != s2->req()) return false;
1161   if (s1->in(0) != s2->in(0)) return false;
1162   if (!same_velt_type(s1, s2)) return false;
1163   return true;
1164 }
1165 
1166 //------------------------------independent---------------------------
1167 // Is there no data path from s1 to s2 or s2 to s1?
1168 bool SuperWord::independent(Node* s1, Node* s2) {
1169   //  assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
1170   int d1 = depth(s1);
1171   int d2 = depth(s2);
1172   if (d1 == d2) return s1 != s2;
1173   Node* deep    = d1 > d2 ? s1 : s2;
1174   Node* shallow = d1 > d2 ? s2 : s1;
1175 
1176   visited_clear();
1177 
1178   return independent_path(shallow, deep);
1179 }
1180 
1181 //------------------------------reduction---------------------------
1182 // Is there a data path between s1 and s2 and the nodes reductions?
1183 bool SuperWord::reduction(Node* s1, Node* s2) {
1184   bool retValue = false;
1185   int d1 = depth(s1);
1186   int d2 = depth(s2);
1187   if (d1 + 1 == d2) {
1188     if (s1->is_reduction() && s2->is_reduction()) {
1189       // This is an ordered set, so s1 should define s2
1190       for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1191         Node* t1 = s1->fast_out(i);
1192         if (t1 == s2) {
1193           // both nodes are reductions and connected
1194           retValue = true;
1195         }
1196       }
1197     }
1198   }
1199 
1200   return retValue;
1201 }
1202 
1203 //------------------------------independent_path------------------------------
1204 // Helper for independent
1205 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
1206   if (dp >= 1000) return false; // stop deep recursion
1207   visited_set(deep);
1208   int shal_depth = depth(shallow);
1209   assert(shal_depth <= depth(deep), "must be");
1210   for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
1211     Node* pred = preds.current();
1212     if (in_bb(pred) && !visited_test(pred)) {
1213       if (shallow == pred) {
1214         return false;
1215       }
1216       if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
1217         return false;
1218       }
1219     }
1220   }
1221   return true;
1222 }
1223 
1224 //------------------------------set_alignment---------------------------
1225 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
1226   set_alignment(s1, align);
1227   if (align == top_align || align == bottom_align) {
1228     set_alignment(s2, align);
1229   } else {
1230     set_alignment(s2, align + data_size(s1));
1231   }
1232 }
1233 
1234 //------------------------------data_size---------------------------
1235 int SuperWord::data_size(Node* s) {
1236   Node* use = NULL; //test if the node is a candidate for CMoveVD optimization, then return the size of CMov
1237   if (_do_vector_loop) {
1238     use = _cmovev_kit.is_Bool_candidate(s);
1239     if (use != NULL) {
1240       return data_size(use);
1241     }
1242     use = _cmovev_kit.is_CmpD_candidate(s);
1243     if (use != NULL) {
1244       return data_size(use);
1245     }
1246   }
1247   int bsize = type2aelembytes(velt_basic_type(s));
1248   assert(bsize != 0, "valid size");
1249   return bsize;
1250 }
1251 
1252 //------------------------------extend_packlist---------------------------
1253 // Extend packset by following use->def and def->use links from pack members.
1254 void SuperWord::extend_packlist() {
1255   bool changed;
1256   do {
1257     packset_sort(_packset.length());
1258     changed = false;
1259     for (int i = 0; i < _packset.length(); i++) {
1260       Node_List* p = _packset.at(i);
1261       changed |= follow_use_defs(p);
1262       changed |= follow_def_uses(p);
1263     }
1264   } while (changed);
1265 
1266   if (_race_possible) {
1267     for (int i = 0; i < _packset.length(); i++) {
1268       Node_List* p = _packset.at(i);
1269       order_def_uses(p);
1270     }
1271   }
1272 
1273   if (TraceSuperWord) {
1274     tty->print_cr("\nAfter extend_packlist");
1275     print_packset();
1276   }
1277 }
1278 
1279 //------------------------------follow_use_defs---------------------------
1280 // Extend the packset by visiting operand definitions of nodes in pack p
1281 bool SuperWord::follow_use_defs(Node_List* p) {
1282   assert(p->size() == 2, "just checking");
1283   Node* s1 = p->at(0);
1284   Node* s2 = p->at(1);
1285   assert(s1->req() == s2->req(), "just checking");
1286   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1287 
1288   if (s1->is_Load()) return false;
1289 
1290   int align = alignment(s1);
1291   NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: s1 %d, align %d", s1->_idx, align);)
1292   bool changed = false;
1293   int start = s1->is_Store() ? MemNode::ValueIn   : 1;
1294   int end   = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
1295   for (int j = start; j < end; j++) {
1296     Node* t1 = s1->in(j);
1297     Node* t2 = s2->in(j);
1298     if (!in_bb(t1) || !in_bb(t2))
1299       continue;
1300     if (stmts_can_pack(t1, t2, align)) {
1301       if (est_savings(t1, t2) >= 0) {
1302         Node_List* pair = new Node_List();
1303         pair->push(t1);
1304         pair->push(t2);
1305         _packset.append(pair);
1306         NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: set_alignment(%d, %d, %d)", t1->_idx, t2->_idx, align);)
1307         set_alignment(t1, t2, align);
1308         changed = true;
1309       }
1310     }
1311   }
1312   return changed;
1313 }
1314 
1315 //------------------------------follow_def_uses---------------------------
1316 // Extend the packset by visiting uses of nodes in pack p
1317 bool SuperWord::follow_def_uses(Node_List* p) {
1318   bool changed = false;
1319   Node* s1 = p->at(0);
1320   Node* s2 = p->at(1);
1321   assert(p->size() == 2, "just checking");
1322   assert(s1->req() == s2->req(), "just checking");
1323   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1324 
1325   if (s1->is_Store()) return false;
1326 
1327   int align = alignment(s1);
1328   NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: s1 %d, align %d", s1->_idx, align);)
1329   int savings = -1;
1330   int num_s1_uses = 0;
1331   Node* u1 = NULL;
1332   Node* u2 = NULL;
1333   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1334     Node* t1 = s1->fast_out(i);
1335     num_s1_uses++;
1336     if (!in_bb(t1)) continue;
1337     for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1338       Node* t2 = s2->fast_out(j);
1339       if (!in_bb(t2)) continue;
1340       if (!opnd_positions_match(s1, t1, s2, t2))
1341         continue;
1342       if (stmts_can_pack(t1, t2, align)) {
1343         int my_savings = est_savings(t1, t2);
1344         if (my_savings > savings) {
1345           savings = my_savings;
1346           u1 = t1;
1347           u2 = t2;
1348         }
1349       }
1350     }
1351   }
1352   if (num_s1_uses > 1) {
1353     _race_possible = true;
1354   }
1355   if (savings >= 0) {
1356     Node_List* pair = new Node_List();
1357     pair->push(u1);
1358     pair->push(u2);
1359     _packset.append(pair);
1360     NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: set_alignment(%d, %d, %d)", u1->_idx, u2->_idx, align);)
1361     set_alignment(u1, u2, align);
1362     changed = true;
1363   }
1364   return changed;
1365 }
1366 
1367 //------------------------------order_def_uses---------------------------
1368 // For extended packsets, ordinally arrange uses packset by major component
1369 void SuperWord::order_def_uses(Node_List* p) {
1370   Node* s1 = p->at(0);
1371 
1372   if (s1->is_Store()) return;
1373 
1374   // reductions are always managed beforehand
1375   if (s1->is_reduction()) return;
1376 
1377   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1378     Node* t1 = s1->fast_out(i);
1379 
1380     // Only allow operand swap on commuting operations
1381     if (!t1->is_Add() && !t1->is_Mul()) {
1382       break;
1383     }
1384 
1385     // Now find t1's packset
1386     Node_List* p2 = NULL;
1387     for (int j = 0; j < _packset.length(); j++) {
1388       p2 = _packset.at(j);
1389       Node* first = p2->at(0);
1390       if (t1 == first) {
1391         break;
1392       }
1393       p2 = NULL;
1394     }
1395     // Arrange all sub components by the major component
1396     if (p2 != NULL) {
1397       for (uint j = 1; j < p->size(); j++) {
1398         Node* d1 = p->at(j);
1399         Node* u1 = p2->at(j);
1400         opnd_positions_match(s1, t1, d1, u1);
1401       }
1402     }
1403   }
1404 }
1405 
1406 //---------------------------opnd_positions_match-------------------------
1407 // Is the use of d1 in u1 at the same operand position as d2 in u2?
1408 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
1409   // check reductions to see if they are marshalled to represent the reduction
1410   // operator in a specified opnd
1411   if (u1->is_reduction() && u2->is_reduction()) {
1412     // ensure reductions have phis and reduction definitions feeding the 1st operand
1413     Node* first = u1->in(2);
1414     if (first->is_Phi() || first->is_reduction()) {
1415       u1->swap_edges(1, 2);
1416     }
1417     // ensure reductions have phis and reduction definitions feeding the 1st operand
1418     first = u2->in(2);
1419     if (first->is_Phi() || first->is_reduction()) {
1420       u2->swap_edges(1, 2);
1421     }
1422     return true;
1423   }
1424 
1425   uint ct = u1->req();
1426   if (ct != u2->req()) return false;
1427   uint i1 = 0;
1428   uint i2 = 0;
1429   do {
1430     for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
1431     for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
1432     if (i1 != i2) {
1433       if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
1434         // Further analysis relies on operands position matching.
1435         u2->swap_edges(i1, i2);
1436       } else {
1437         return false;
1438       }
1439     }
1440   } while (i1 < ct);
1441   return true;
1442 }
1443 
1444 //------------------------------est_savings---------------------------
1445 // Estimate the savings from executing s1 and s2 as a pack
1446 int SuperWord::est_savings(Node* s1, Node* s2) {
1447   int save_in = 2 - 1; // 2 operations per instruction in packed form
1448 
1449   // inputs
1450   for (uint i = 1; i < s1->req(); i++) {
1451     Node* x1 = s1->in(i);
1452     Node* x2 = s2->in(i);
1453     if (x1 != x2) {
1454       if (are_adjacent_refs(x1, x2)) {
1455         save_in += adjacent_profit(x1, x2);
1456       } else if (!in_packset(x1, x2)) {
1457         save_in -= pack_cost(2);
1458       } else {
1459         save_in += unpack_cost(2);
1460       }
1461     }
1462   }
1463 
1464   // uses of result
1465   uint ct = 0;
1466   int save_use = 0;
1467   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1468     Node* s1_use = s1->fast_out(i);
1469     for (int j = 0; j < _packset.length(); j++) {
1470       Node_List* p = _packset.at(j);
1471       if (p->at(0) == s1_use) {
1472         for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
1473           Node* s2_use = s2->fast_out(k);
1474           if (p->at(p->size()-1) == s2_use) {
1475             ct++;
1476             if (are_adjacent_refs(s1_use, s2_use)) {
1477               save_use += adjacent_profit(s1_use, s2_use);
1478             }
1479           }
1480         }
1481       }
1482     }
1483   }
1484 
1485   if (ct < s1->outcnt()) save_use += unpack_cost(1);
1486   if (ct < s2->outcnt()) save_use += unpack_cost(1);
1487 
1488   return MAX2(save_in, save_use);
1489 }
1490 
1491 //------------------------------costs---------------------------
1492 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
1493 int SuperWord::pack_cost(int ct)   { return ct; }
1494 int SuperWord::unpack_cost(int ct) { return ct; }
1495 
1496 //------------------------------combine_packs---------------------------
1497 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
1498 void SuperWord::combine_packs() {
1499   bool changed = true;
1500   // Combine packs regardless max vector size.
1501   while (changed) {
1502     changed = false;
1503     for (int i = 0; i < _packset.length(); i++) {
1504       Node_List* p1 = _packset.at(i);
1505       if (p1 == NULL) continue;
1506       // Because of sorting we can start at i + 1
1507       for (int j = i + 1; j < _packset.length(); j++) {
1508         Node_List* p2 = _packset.at(j);
1509         if (p2 == NULL) continue;
1510         if (i == j) continue;
1511         if (p1->at(p1->size()-1) == p2->at(0)) {
1512           for (uint k = 1; k < p2->size(); k++) {
1513             p1->push(p2->at(k));
1514           }
1515           _packset.at_put(j, NULL);
1516           changed = true;
1517         }
1518       }
1519     }
1520   }
1521 
1522   // Split packs which have size greater then max vector size.
1523   for (int i = 0; i < _packset.length(); i++) {
1524     Node_List* p1 = _packset.at(i);
1525     if (p1 != NULL) {
1526       BasicType bt = velt_basic_type(p1->at(0));
1527       uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1528       assert(is_power_of_2(max_vlen), "sanity");
1529       uint psize = p1->size();
1530       if (!is_power_of_2(psize)) {
1531         // Skip pack which can't be vector.
1532         // case1: for(...) { a[i] = i; }    elements values are different (i+x)
1533         // case2: for(...) { a[i] = b[i+1]; }  can't align both, load and store
1534         _packset.at_put(i, NULL);
1535         continue;
1536       }
1537       if (psize > max_vlen) {
1538         Node_List* pack = new Node_List();
1539         for (uint j = 0; j < psize; j++) {
1540           pack->push(p1->at(j));
1541           if (pack->size() >= max_vlen) {
1542             assert(is_power_of_2(pack->size()), "sanity");
1543             _packset.append(pack);
1544             pack = new Node_List();
1545           }
1546         }
1547         _packset.at_put(i, NULL);
1548       }
1549     }
1550   }
1551 
1552   // Compress list.
1553   for (int i = _packset.length() - 1; i >= 0; i--) {
1554     Node_List* p1 = _packset.at(i);
1555     if (p1 == NULL) {
1556       _packset.remove_at(i);
1557     }
1558   }
1559 
1560   if (TraceSuperWord) {
1561     tty->print_cr("\nAfter combine_packs");
1562     print_packset();
1563   }
1564 }
1565 
1566 //-----------------------------construct_my_pack_map--------------------------
1567 // Construct the map from nodes to packs.  Only valid after the
1568 // point where a node is only in one pack (after combine_packs).
1569 void SuperWord::construct_my_pack_map() {
1570   Node_List* rslt = NULL;
1571   for (int i = 0; i < _packset.length(); i++) {
1572     Node_List* p = _packset.at(i);
1573     for (uint j = 0; j < p->size(); j++) {
1574       Node* s = p->at(j);
1575       assert(my_pack(s) == NULL, "only in one pack");
1576       set_my_pack(s, p);
1577     }
1578   }
1579 }
1580 
1581 //------------------------------filter_packs---------------------------
1582 // Remove packs that are not implemented or not profitable.
1583 void SuperWord::filter_packs() {
1584   // Remove packs that are not implemented
1585   for (int i = _packset.length() - 1; i >= 0; i--) {
1586     Node_List* pk = _packset.at(i);
1587     bool impl = implemented(pk);
1588     if (!impl) {
1589 #ifndef PRODUCT
1590       if (TraceSuperWord && Verbose) {
1591         tty->print_cr("Unimplemented");
1592         pk->at(0)->dump();
1593       }
1594 #endif
1595       remove_pack_at(i);
1596     }
1597     Node *n = pk->at(0);
1598     if (n->is_reduction()) {
1599       _num_reductions++;
1600     } else {
1601       _num_work_vecs++;
1602     }
1603   }
1604 
1605   // Remove packs that are not profitable
1606   bool changed;
1607   do {
1608     changed = false;
1609     for (int i = _packset.length() - 1; i >= 0; i--) {
1610       Node_List* pk = _packset.at(i);
1611       bool prof = profitable(pk);
1612       if (!prof) {
1613 #ifndef PRODUCT
1614         if (TraceSuperWord && Verbose) {
1615           tty->print_cr("Unprofitable");
1616           pk->at(0)->dump();
1617         }
1618 #endif
1619         remove_pack_at(i);
1620         changed = true;
1621       }
1622     }
1623   } while (changed);
1624 
1625 #ifndef PRODUCT
1626   if (TraceSuperWord) {
1627     tty->print_cr("\nAfter filter_packs");
1628     print_packset();
1629     tty->cr();
1630   }
1631 #endif
1632 }
1633 
1634 //------------------------------merge_packs_to_cmovd---------------------------
1635 // Merge CMoveD into new vector-nodes
1636 // We want to catch this pattern and subsume CmpD and Bool into CMoveD
1637 //
1638 //                   SubD             ConD
1639 //                  /  |               /
1640 //                 /   |           /   /
1641 //                /    |       /      /
1642 //               /     |   /         /
1643 //              /      /            /
1644 //             /    /  |           /
1645 //            v /      |          /
1646 //         CmpD        |         /
1647 //          |          |        /
1648 //          v          |       /
1649 //         Bool        |      /
1650 //           \         |     /
1651 //             \       |    /
1652 //               \     |   /
1653 //                 \   |  /
1654 //                   \ v /
1655 //                   CMoveD
1656 //
1657 
1658 void SuperWord::merge_packs_to_cmovd() {
1659   for (int i = _packset.length() - 1; i >= 0; i--) {
1660     _cmovev_kit.make_cmovevd_pack(_packset.at(i));
1661   }
1662   #ifndef PRODUCT
1663     if (TraceSuperWord) {
1664       tty->print_cr("\nSuperWord::merge_packs_to_cmovd(): After merge");
1665       print_packset();
1666       tty->cr();
1667     }
1668   #endif
1669 }
1670 
1671 Node* CMoveKit::is_Bool_candidate(Node* def) const {
1672   Node* use = NULL;
1673   if (!def->is_Bool() || def->in(0) != NULL || def->outcnt() != 1) {
1674     return NULL;
1675   }
1676   for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1677     use = def->fast_out(j);
1678     if (!_sw->same_generation(def, use) || !use->is_CMove()) {
1679       return NULL;
1680     }
1681   }
1682   return use;
1683 }
1684 
1685 Node* CMoveKit::is_CmpD_candidate(Node* def) const {
1686   Node* use = NULL;
1687   if (!def->is_Cmp() || def->in(0) != NULL || def->outcnt() != 1) {
1688     return NULL;
1689   }
1690   for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1691     use = def->fast_out(j);
1692     if (!_sw->same_generation(def, use) || (use = is_Bool_candidate(use)) == NULL || !_sw->same_generation(def, use)) {
1693       return NULL;
1694     }
1695   }
1696   return use;
1697 }
1698 
1699 Node_List* CMoveKit::make_cmovevd_pack(Node_List* cmovd_pk) {
1700   Node *cmovd = cmovd_pk->at(0);
1701   if (!cmovd->is_CMove()) {
1702     return NULL;
1703   }
1704   if (pack(cmovd) != NULL) { // already in the cmov pack
1705     return NULL;
1706   }
1707   if (cmovd->in(0) != NULL) {
1708     NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: CMoveD %d has control flow, escaping...", cmovd->_idx); cmovd->dump();})
1709     return NULL;
1710   }
1711 
1712   Node* bol = cmovd->as_CMove()->in(CMoveNode::Condition);
1713   if (!bol->is_Bool()
1714       || bol->outcnt() != 1
1715       || !_sw->same_generation(bol, cmovd)
1716       || bol->in(0) != NULL  // BoolNode has control flow!!
1717       || _sw->my_pack(bol) == NULL) {
1718       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();})
1719       return NULL;
1720   }
1721   Node_List* bool_pk = _sw->my_pack(bol);
1722   if (bool_pk->size() != cmovd_pk->size() ) {
1723     return NULL;
1724   }
1725 
1726   Node* cmpd = bol->in(1);
1727   if (!cmpd->is_Cmp()
1728       || cmpd->outcnt() != 1
1729       || !_sw->same_generation(cmpd, cmovd)
1730       || cmpd->in(0) != NULL  // CmpDNode has control flow!!
1731       || _sw->my_pack(cmpd) == NULL) {
1732       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();})
1733       return NULL;
1734   }
1735   Node_List* cmpd_pk = _sw->my_pack(cmpd);
1736   if (cmpd_pk->size() != cmovd_pk->size() ) {
1737     return NULL;
1738   }
1739 
1740   if (!test_cmpd_pack(cmpd_pk, cmovd_pk)) {
1741     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();})
1742     return NULL;
1743   }
1744 
1745   Node_List* new_cmpd_pk = new Node_List();
1746   uint sz = cmovd_pk->size() - 1;
1747   for (uint i = 0; i <= sz; ++i) {
1748     Node* cmov = cmovd_pk->at(i);
1749     Node* bol  = bool_pk->at(i);
1750     Node* cmp  = cmpd_pk->at(i);
1751 
1752     new_cmpd_pk->insert(i, cmov);
1753 
1754     map(cmov, new_cmpd_pk);
1755     map(bol, new_cmpd_pk);
1756     map(cmp, new_cmpd_pk);
1757 
1758     _sw->set_my_pack(cmov, new_cmpd_pk); // and keep old packs for cmp and bool
1759   }
1760   _sw->_packset.remove(cmovd_pk);
1761   _sw->_packset.remove(bool_pk);
1762   _sw->_packset.remove(cmpd_pk);
1763   _sw->_packset.append(new_cmpd_pk);
1764   NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print_cr("CMoveKit::make_cmovevd_pack: added syntactic CMoveD pack"); _sw->print_pack(new_cmpd_pk);})
1765   return new_cmpd_pk;
1766 }
1767 
1768 bool CMoveKit::test_cmpd_pack(Node_List* cmpd_pk, Node_List* cmovd_pk) {
1769   Node* cmpd0 = cmpd_pk->at(0);
1770   assert(cmpd0->is_Cmp(), "CMoveKit::test_cmpd_pack: should be CmpDNode");
1771   assert(cmovd_pk->at(0)->is_CMove(), "CMoveKit::test_cmpd_pack: should be CMoveD");
1772   assert(cmpd_pk->size() == cmovd_pk->size(), "CMoveKit::test_cmpd_pack: should be same size");
1773   Node* in1 = cmpd0->in(1);
1774   Node* in2 = cmpd0->in(2);
1775   Node_List* in1_pk = _sw->my_pack(in1);
1776   Node_List* in2_pk = _sw->my_pack(in2);
1777 
1778   if (  (in1_pk != NULL && in1_pk->size() != cmpd_pk->size())
1779      || (in2_pk != NULL && in2_pk->size() != cmpd_pk->size()) ) {
1780     return false;
1781   }
1782 
1783   // test if "all" in1 are in the same pack or the same node
1784   if (in1_pk == NULL) {
1785     for (uint j = 1; j < cmpd_pk->size(); j++) {
1786       if (cmpd_pk->at(j)->in(1) != in1) {
1787         return false;
1788       }
1789     }//for: in1_pk is not pack but all CmpD nodes in the pack have the same in(1)
1790   }
1791   // test if "all" in2 are in the same pack or the same node
1792   if (in2_pk == NULL) {
1793     for (uint j = 1; j < cmpd_pk->size(); j++) {
1794       if (cmpd_pk->at(j)->in(2) != in2) {
1795         return false;
1796       }
1797     }//for: in2_pk is not pack but all CmpD nodes in the pack have the same in(2)
1798   }
1799   //now check if cmpd_pk may be subsumed in vector built for cmovd_pk
1800   int cmovd_ind1, cmovd_ind2;
1801   if (cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1802    && cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1803       cmovd_ind1 = CMoveNode::IfFalse;
1804       cmovd_ind2 = CMoveNode::IfTrue;
1805   } else if (cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1806           && cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1807       cmovd_ind2 = CMoveNode::IfFalse;
1808       cmovd_ind1 = CMoveNode::IfTrue;
1809   }
1810   else {
1811     return false;
1812   }
1813 
1814   for (uint j = 1; j < cmpd_pk->size(); j++) {
1815     if (cmpd_pk->at(j)->in(1) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind1)
1816         || cmpd_pk->at(j)->in(2) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind2)) {
1817         return false;
1818     }//if
1819   }
1820   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(); })
1821   return true;
1822 }
1823 
1824 //------------------------------implemented---------------------------
1825 // Can code be generated for pack p?
1826 bool SuperWord::implemented(Node_List* p) {
1827   bool retValue = false;
1828   Node* p0 = p->at(0);
1829   if (p0 != NULL) {
1830     int opc = p0->Opcode();
1831     uint size = p->size();
1832     if (p0->is_reduction()) {
1833       const Type *arith_type = p0->bottom_type();
1834       // Length 2 reductions of INT/LONG do not offer performance benefits
1835       if (((arith_type->basic_type() == T_INT) || (arith_type->basic_type() == T_LONG)) && (size == 2)) {
1836         retValue = false;
1837       } else {
1838         retValue = ReductionNode::implemented(opc, size, arith_type->basic_type());
1839       }
1840     } else {
1841       retValue = VectorNode::implemented(opc, size, velt_basic_type(p0));
1842     }
1843     if (!retValue) {
1844       if (is_cmov_pack(p)) {
1845         NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::implemented: found cmpd pack"); print_pack(p);})
1846         return true;
1847       }
1848     }
1849   }
1850   return retValue;
1851 }
1852 
1853 bool SuperWord::is_cmov_pack(Node_List* p) {
1854   return _cmovev_kit.pack(p->at(0)) != NULL;
1855 }
1856 //------------------------------same_inputs--------------------------
1857 // For pack p, are all idx operands the same?
1858 bool SuperWord::same_inputs(Node_List* p, int idx) {
1859   Node* p0 = p->at(0);
1860   uint vlen = p->size();
1861   Node* p0_def = p0->in(idx);
1862   for (uint i = 1; i < vlen; i++) {
1863     Node* pi = p->at(i);
1864     Node* pi_def = pi->in(idx);
1865     if (p0_def != pi_def) {
1866       return false;
1867     }
1868   }
1869   return true;
1870 }
1871 
1872 //------------------------------profitable---------------------------
1873 // For pack p, are all operands and all uses (with in the block) vector?
1874 bool SuperWord::profitable(Node_List* p) {
1875   Node* p0 = p->at(0);
1876   uint start, end;
1877   VectorNode::vector_operands(p0, &start, &end);
1878 
1879   // Return false if some inputs are not vectors or vectors with different
1880   // size or alignment.
1881   // Also, for now, return false if not scalar promotion case when inputs are
1882   // the same. Later, implement PackNode and allow differing, non-vector inputs
1883   // (maybe just the ones from outside the block.)
1884   for (uint i = start; i < end; i++) {
1885     if (!is_vector_use(p0, i)) {
1886       return false;
1887     }
1888   }
1889   // Check if reductions are connected
1890   if (p0->is_reduction()) {
1891     Node* second_in = p0->in(2);
1892     Node_List* second_pk = my_pack(second_in);
1893     if ((second_pk == NULL) || (_num_work_vecs == _num_reductions)) {
1894       // Remove reduction flag if no parent pack or if not enough work
1895       // to cover reduction expansion overhead
1896       p0->remove_flag(Node::Flag_is_reduction);
1897       return false;
1898     } else if (second_pk->size() != p->size()) {
1899       return false;
1900     }
1901   }
1902   if (VectorNode::is_shift(p0)) {
1903     // For now, return false if shift count is vector or not scalar promotion
1904     // case (different shift counts) because it is not supported yet.
1905     Node* cnt = p0->in(2);
1906     Node_List* cnt_pk = my_pack(cnt);
1907     if (cnt_pk != NULL)
1908       return false;
1909     if (!same_inputs(p, 2))
1910       return false;
1911   }
1912   if (!p0->is_Store()) {
1913     // For now, return false if not all uses are vector.
1914     // Later, implement ExtractNode and allow non-vector uses (maybe
1915     // just the ones outside the block.)
1916     for (uint i = 0; i < p->size(); i++) {
1917       Node* def = p->at(i);
1918       if (is_cmov_pack_internal_node(p, def)) {
1919         continue;
1920       }
1921       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1922         Node* use = def->fast_out(j);
1923         for (uint k = 0; k < use->req(); k++) {
1924           Node* n = use->in(k);
1925           if (def == n) {
1926             // reductions can be loop carried dependences
1927             if (def->is_reduction() && use->is_Phi())
1928               continue;
1929             if (!is_vector_use(use, k)) {
1930               return false;
1931             }
1932           }
1933         }
1934       }
1935     }
1936   }
1937   return true;
1938 }
1939 
1940 //------------------------------schedule---------------------------
1941 // Adjust the memory graph for the packed operations
1942 void SuperWord::schedule() {
1943 
1944   // Co-locate in the memory graph the members of each memory pack
1945   for (int i = 0; i < _packset.length(); i++) {
1946     co_locate_pack(_packset.at(i));
1947   }
1948 }
1949 
1950 //-------------------------------remove_and_insert-------------------
1951 // Remove "current" from its current position in the memory graph and insert
1952 // it after the appropriate insertion point (lip or uip).
1953 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1954                                   Node *uip, Unique_Node_List &sched_before) {
1955   Node* my_mem = current->in(MemNode::Memory);
1956   bool sched_up = sched_before.member(current);
1957 
1958   // remove current_store from its current position in the memmory graph
1959   for (DUIterator i = current->outs(); current->has_out(i); i++) {
1960     Node* use = current->out(i);
1961     if (use->is_Mem()) {
1962       assert(use->in(MemNode::Memory) == current, "must be");
1963       if (use == prev) { // connect prev to my_mem
1964           _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1965           --i; //deleted this edge; rescan position
1966       } else if (sched_before.member(use)) {
1967         if (!sched_up) { // Will be moved together with current
1968           _igvn.replace_input_of(use, MemNode::Memory, uip);
1969           --i; //deleted this edge; rescan position
1970         }
1971       } else {
1972         if (sched_up) { // Will be moved together with current
1973           _igvn.replace_input_of(use, MemNode::Memory, lip);
1974           --i; //deleted this edge; rescan position
1975         }
1976       }
1977     }
1978   }
1979 
1980   Node *insert_pt =  sched_up ?  uip : lip;
1981 
1982   // all uses of insert_pt's memory state should use current's instead
1983   for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1984     Node* use = insert_pt->out(i);
1985     if (use->is_Mem()) {
1986       assert(use->in(MemNode::Memory) == insert_pt, "must be");
1987       _igvn.replace_input_of(use, MemNode::Memory, current);
1988       --i; //deleted this edge; rescan position
1989     } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1990       uint pos; //lip (lower insert point) must be the last one in the memory slice
1991       for (pos=1; pos < use->req(); pos++) {
1992         if (use->in(pos) == insert_pt) break;
1993       }
1994       _igvn.replace_input_of(use, pos, current);
1995       --i;
1996     }
1997   }
1998 
1999   //connect current to insert_pt
2000   _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
2001 }
2002 
2003 //------------------------------co_locate_pack----------------------------------
2004 // To schedule a store pack, we need to move any sandwiched memory ops either before
2005 // or after the pack, based upon dependence information:
2006 // (1) If any store in the pack depends on the sandwiched memory op, the
2007 //     sandwiched memory op must be scheduled BEFORE the pack;
2008 // (2) If a sandwiched memory op depends on any store in the pack, the
2009 //     sandwiched memory op must be scheduled AFTER the pack;
2010 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
2011 //     memory op (say memB), memB must be scheduled before memA. So, if memA is
2012 //     scheduled before the pack, memB must also be scheduled before the pack;
2013 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
2014 //     schedule this store AFTER the pack
2015 // (5) We know there is no dependence cycle, so there in no other case;
2016 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
2017 //
2018 // To schedule a load pack, we use the memory state of either the first or the last load in
2019 // the pack, based on the dependence constraint.
2020 void SuperWord::co_locate_pack(Node_List* pk) {
2021   if (pk->at(0)->is_Store()) {
2022     MemNode* first     = executed_first(pk)->as_Mem();
2023     MemNode* last      = executed_last(pk)->as_Mem();
2024     Unique_Node_List schedule_before_pack;
2025     Unique_Node_List memops;
2026 
2027     MemNode* current   = last->in(MemNode::Memory)->as_Mem();
2028     MemNode* previous  = last;
2029     while (true) {
2030       assert(in_bb(current), "stay in block");
2031       memops.push(previous);
2032       for (DUIterator i = current->outs(); current->has_out(i); i++) {
2033         Node* use = current->out(i);
2034         if (use->is_Mem() && use != previous)
2035           memops.push(use);
2036       }
2037       if (current == first) break;
2038       previous = current;
2039       current  = current->in(MemNode::Memory)->as_Mem();
2040     }
2041 
2042     // determine which memory operations should be scheduled before the pack
2043     for (uint i = 1; i < memops.size(); i++) {
2044       Node *s1 = memops.at(i);
2045       if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
2046         for (uint j = 0; j< i; j++) {
2047           Node *s2 = memops.at(j);
2048           if (!independent(s1, s2)) {
2049             if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
2050               schedule_before_pack.push(s1); // s1 must be scheduled before
2051               Node_List* mem_pk = my_pack(s1);
2052               if (mem_pk != NULL) {
2053                 for (uint ii = 0; ii < mem_pk->size(); ii++) {
2054                   Node* s = mem_pk->at(ii);  // follow partner
2055                   if (memops.member(s) && !schedule_before_pack.member(s))
2056                     schedule_before_pack.push(s);
2057                 }
2058               }
2059               break;
2060             }
2061           }
2062         }
2063       }
2064     }
2065 
2066     Node*    upper_insert_pt = first->in(MemNode::Memory);
2067     // Following code moves loads connected to upper_insert_pt below aliased stores.
2068     // Collect such loads here and reconnect them back to upper_insert_pt later.
2069     memops.clear();
2070     for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
2071       Node* use = upper_insert_pt->out(i);
2072       if (use->is_Mem() && !use->is_Store()) {
2073         memops.push(use);
2074       }
2075     }
2076 
2077     MemNode* lower_insert_pt = last;
2078     previous                 = last; //previous store in pk
2079     current                  = last->in(MemNode::Memory)->as_Mem();
2080 
2081     // start scheduling from "last" to "first"
2082     while (true) {
2083       assert(in_bb(current), "stay in block");
2084       assert(in_pack(previous, pk), "previous stays in pack");
2085       Node* my_mem = current->in(MemNode::Memory);
2086 
2087       if (in_pack(current, pk)) {
2088         // Forward users of my memory state (except "previous) to my input memory state
2089         for (DUIterator i = current->outs(); current->has_out(i); i++) {
2090           Node* use = current->out(i);
2091           if (use->is_Mem() && use != previous) {
2092             assert(use->in(MemNode::Memory) == current, "must be");
2093             if (schedule_before_pack.member(use)) {
2094               _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
2095             } else {
2096               _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
2097             }
2098             --i; // deleted this edge; rescan position
2099           }
2100         }
2101         previous = current;
2102       } else { // !in_pack(current, pk) ==> a sandwiched store
2103         remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
2104       }
2105 
2106       if (current == first) break;
2107       current = my_mem->as_Mem();
2108     } // end while
2109 
2110     // Reconnect loads back to upper_insert_pt.
2111     for (uint i = 0; i < memops.size(); i++) {
2112       Node *ld = memops.at(i);
2113       if (ld->in(MemNode::Memory) != upper_insert_pt) {
2114         _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
2115       }
2116     }
2117   } else if (pk->at(0)->is_Load()) { //load
2118     // all loads in the pack should have the same memory state. By default,
2119     // we use the memory state of the last load. However, if any load could
2120     // not be moved down due to the dependence constraint, we use the memory
2121     // state of the first load.
2122     Node* last_mem  = executed_last(pk)->in(MemNode::Memory);
2123     Node* first_mem = executed_first(pk)->in(MemNode::Memory);
2124     bool schedule_last = true;
2125     for (uint i = 0; i < pk->size(); i++) {
2126       Node* ld = pk->at(i);
2127       for (Node* current = last_mem; current != ld->in(MemNode::Memory);
2128            current=current->in(MemNode::Memory)) {
2129         assert(current != first_mem, "corrupted memory graph");
2130         if(current->is_Mem() && !independent(current, ld)){
2131           schedule_last = false; // a later store depends on this load
2132           break;
2133         }
2134       }
2135     }
2136 
2137     Node* mem_input = schedule_last ? last_mem : first_mem;
2138     _igvn.hash_delete(mem_input);
2139     // Give each load the same memory state
2140     for (uint i = 0; i < pk->size(); i++) {
2141       LoadNode* ld = pk->at(i)->as_Load();
2142       _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
2143     }
2144   }
2145 }
2146 
2147 #ifndef PRODUCT
2148 void SuperWord::print_loop(bool whole) {
2149   Node_Stack stack(_arena, _phase->C->unique() >> 2);
2150   Node_List rpo_list;
2151   VectorSet visited(_arena);
2152   visited.set(lpt()->_head->_idx);
2153   _phase->rpo(lpt()->_head, stack, visited, rpo_list);
2154   _phase->dump(lpt(), rpo_list.size(), rpo_list );
2155   if(whole) {
2156     tty->print_cr("\n Whole loop tree");
2157     _phase->dump();
2158     tty->print_cr(" End of whole loop tree\n");
2159   }
2160 }
2161 #endif
2162 
2163 //------------------------------output---------------------------
2164 // Convert packs into vector node operations
2165 void SuperWord::output() {
2166   if (_packset.length() == 0) return;
2167 
2168 #ifndef PRODUCT
2169   if (TraceLoopOpts) {
2170     tty->print("SuperWord::output    ");
2171     lpt()->dump_head();
2172   }
2173 #endif
2174 
2175   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2176   if (cl->is_main_loop()) {
2177     // MUST ENSURE main loop's initial value is properly aligned:
2178     //  (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
2179 
2180     align_initial_loop_index(align_to_ref());
2181 
2182     // Insert extract (unpack) operations for scalar uses
2183     for (int i = 0; i < _packset.length(); i++) {
2184       insert_extracts(_packset.at(i));
2185     }
2186   }
2187 
2188   Compile* C = _phase->C;
2189   uint max_vlen_in_bytes = 0;
2190   uint max_vlen = 0;
2191   bool can_process_post_loop = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
2192 
2193   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop before create_reserve_version_of_loop"); print_loop(true);})
2194 
2195   CountedLoopReserveKit make_reversable(_phase, _lpt, do_reserve_copy());
2196 
2197   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop after create_reserve_version_of_loop"); print_loop(true);})
2198 
2199   if (do_reserve_copy() && !make_reversable.has_reserved()) {
2200     NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: loop was not reserved correctly, exiting SuperWord");})
2201     return;
2202   }
2203 
2204   for (int i = 0; i < _block.length(); i++) {
2205     Node* n = _block.at(i);
2206     Node_List* p = my_pack(n);
2207     if (p && n == executed_last(p)) {
2208       uint vlen = p->size();
2209       uint vlen_in_bytes = 0;
2210       Node* vn = NULL;
2211       Node* low_adr = p->at(0);
2212       Node* first   = executed_first(p);
2213       if (can_process_post_loop) {
2214         // override vlen with the main loops vector length
2215         vlen = cl->slp_max_unroll();
2216       }
2217       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);})
2218       int   opc = n->Opcode();
2219       if (n->is_Load()) {
2220         Node* ctl = n->in(MemNode::Control);
2221         Node* mem = first->in(MemNode::Memory);
2222         SWPointer p1(n->as_Mem(), this, NULL, false);
2223         // Identify the memory dependency for the new loadVector node by
2224         // walking up through memory chain.
2225         // This is done to give flexibility to the new loadVector node so that
2226         // it can move above independent storeVector nodes.
2227         while (mem->is_StoreVector()) {
2228           SWPointer p2(mem->as_Mem(), this, NULL, false);
2229           int cmp = p1.cmp(p2);
2230           if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
2231             mem = mem->in(MemNode::Memory);
2232           } else {
2233             break; // dependent memory
2234           }
2235         }
2236         Node* adr = low_adr->in(MemNode::Address);
2237         const TypePtr* atyp = n->adr_type();
2238         vn = LoadVectorNode::make(opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n), control_dependency(p));
2239         vlen_in_bytes = vn->as_LoadVector()->memory_size();
2240       } else if (n->is_Store()) {
2241         // Promote value to be stored to vector
2242         Node* val = vector_opd(p, MemNode::ValueIn);
2243         if (val == NULL) {
2244           if (do_reserve_copy()) {
2245             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: val should not be NULL, exiting SuperWord");})
2246             return; //and reverse to backup IG
2247           }
2248           ShouldNotReachHere();
2249         }
2250 
2251         Node* ctl = n->in(MemNode::Control);
2252         Node* mem = first->in(MemNode::Memory);
2253         Node* adr = low_adr->in(MemNode::Address);
2254         const TypePtr* atyp = n->adr_type();
2255         vn = StoreVectorNode::make(opc, ctl, mem, adr, atyp, val, vlen);
2256         vlen_in_bytes = vn->as_StoreVector()->memory_size();
2257       } else if (n->req() == 3 && !is_cmov_pack(p)) {
2258         // Promote operands to vector
2259         Node* in1 = NULL;
2260         bool node_isa_reduction = n->is_reduction();
2261         if (node_isa_reduction) {
2262           // the input to the first reduction operation is retained
2263           in1 = low_adr->in(1);
2264         } else {
2265           in1 = vector_opd(p, 1);
2266           if (in1 == NULL) {
2267             if (do_reserve_copy()) {
2268               NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in1 should not be NULL, exiting SuperWord");})
2269               return; //and reverse to backup IG
2270             }
2271             ShouldNotReachHere();
2272           }
2273         }
2274         Node* in2 = vector_opd(p, 2);
2275         if (in2 == NULL) {
2276           if (do_reserve_copy()) {
2277             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in2 should not be NULL, exiting SuperWord");})
2278             return; //and reverse to backup IG
2279           }
2280           ShouldNotReachHere();
2281         }
2282         if (VectorNode::is_invariant_vector(in1) && (node_isa_reduction == false) && (n->is_Add() || n->is_Mul())) {
2283           // Move invariant vector input into second position to avoid register spilling.
2284           Node* tmp = in1;
2285           in1 = in2;
2286           in2 = tmp;
2287         }
2288         if (node_isa_reduction) {
2289           const Type *arith_type = n->bottom_type();
2290           vn = ReductionNode::make(opc, NULL, in1, in2, arith_type->basic_type());
2291           if (in2->is_Load()) {
2292             vlen_in_bytes = in2->as_LoadVector()->memory_size();
2293           } else {
2294             vlen_in_bytes = in2->as_Vector()->length_in_bytes();
2295           }
2296         } else {
2297           vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2298           vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2299         }
2300       } else if (opc == Op_SqrtD || opc == Op_AbsF || opc == Op_AbsD || opc == Op_NegF || opc == Op_NegD) {
2301         // Promote operand to vector (Sqrt/Abs/Neg are 2 address instructions)
2302         Node* in = vector_opd(p, 1);
2303         vn = VectorNode::make(opc, in, NULL, vlen, velt_basic_type(n));
2304         vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2305       } else if (is_cmov_pack(p)) {
2306         if (can_process_post_loop) {
2307           // do not refactor of flow in post loop context
2308           return;
2309         }
2310         if (!n->is_CMove()) {
2311           continue;
2312         }
2313         // place here CMoveVDNode
2314         NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::output: print before CMove vectorization"); print_loop(false);})
2315         Node* bol = n->in(CMoveNode::Condition);
2316         if (!bol->is_Bool() && bol->Opcode() == Op_ExtractI && bol->req() > 1 ) {
2317           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();})
2318           bol = bol->in(1); //may be ExtractNode
2319         }
2320 
2321         assert(bol->is_Bool(), "should be BoolNode - too late to bail out!");
2322         if (!bol->is_Bool()) {
2323           if (do_reserve_copy()) {
2324             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: expected %d bool node, exiting SuperWord", bol->_idx); bol->dump();})
2325             return; //and reverse to backup IG
2326           }
2327           ShouldNotReachHere();
2328         }
2329 
2330         int cond = (int)bol->as_Bool()->_test._test;
2331         Node* in_cc  = _igvn.intcon(cond);
2332         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created intcon in_cc node %d", in_cc->_idx); in_cc->dump();})
2333         Node* cc = bol->clone();
2334         cc->set_req(1, in_cc);
2335         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created bool cc node %d", cc->_idx); cc->dump();})
2336 
2337         Node* src1 = vector_opd(p, 2); //2=CMoveNode::IfFalse
2338         if (src1 == NULL) {
2339           if (do_reserve_copy()) {
2340             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src1 should not be NULL, exiting SuperWord");})
2341             return; //and reverse to backup IG
2342           }
2343           ShouldNotReachHere();
2344         }
2345         Node* src2 = vector_opd(p, 3); //3=CMoveNode::IfTrue
2346         if (src2 == NULL) {
2347           if (do_reserve_copy()) {
2348             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src2 should not be NULL, exiting SuperWord");})
2349             return; //and reverse to backup IG
2350           }
2351           ShouldNotReachHere();
2352         }
2353         BasicType bt = velt_basic_type(n);
2354         const TypeVect* vt = TypeVect::make(bt, vlen);
2355         vn = new CMoveVDNode(cc, src1, src2, vt);
2356         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created new CMove node %d: ", vn->_idx); vn->dump();})
2357       } else if (opc == Op_FmaD || opc == Op_FmaF) {
2358         // Promote operands to vector
2359         Node* in1 = vector_opd(p, 1);
2360         Node* in2 = vector_opd(p, 2);
2361         Node* in3 = vector_opd(p, 3);
2362         vn = VectorNode::make(opc, in1, in2, in3, vlen, velt_basic_type(n));
2363         vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2364       } else {
2365         if (do_reserve_copy()) {
2366           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: ShouldNotReachHere, exiting SuperWord");})
2367           return; //and reverse to backup IG
2368         }
2369         ShouldNotReachHere();
2370       }
2371 
2372       assert(vn != NULL, "sanity");
2373       if (vn == NULL) {
2374         if (do_reserve_copy()){
2375           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: got NULL node, cannot proceed, exiting SuperWord");})
2376           return; //and reverse to backup IG
2377         }
2378         ShouldNotReachHere();
2379       }
2380 
2381       _block.at_put(i, vn);
2382       _igvn.register_new_node_with_optimizer(vn);
2383       _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
2384       for (uint j = 0; j < p->size(); j++) {
2385         Node* pm = p->at(j);
2386         _igvn.replace_node(pm, vn);
2387       }
2388       _igvn._worklist.push(vn);
2389 
2390       if (can_process_post_loop) {
2391         // first check if the vector size if the maximum vector which we can use on the machine,
2392         // other vector size have reduced values for predicated data mapping.
2393         if (vlen_in_bytes != (uint)MaxVectorSize) {
2394           return;
2395         }
2396       }
2397 
2398       if (vlen_in_bytes >= max_vlen_in_bytes && vlen > max_vlen) {
2399         max_vlen = vlen;
2400         max_vlen_in_bytes = vlen_in_bytes;
2401       }
2402 #ifdef ASSERT
2403       if (TraceNewVectors) {
2404         tty->print("new Vector node: ");
2405         vn->dump();
2406       }
2407 #endif
2408     }
2409   }//for (int i = 0; i < _block.length(); i++)
2410 
2411   C->set_max_vector_size(max_vlen_in_bytes);
2412 
2413   if (SuperWordLoopUnrollAnalysis) {
2414     if (cl->has_passed_slp()) {
2415       uint slp_max_unroll_factor = cl->slp_max_unroll();
2416       if (slp_max_unroll_factor == max_vlen) {
2417         if (TraceSuperWordLoopUnrollAnalysis) {
2418           tty->print_cr("vector loop(unroll=%d, len=%d)\n", max_vlen, max_vlen_in_bytes*BitsPerByte);
2419         }
2420 
2421         // For atomic unrolled loops which are vector mapped, instigate more unrolling
2422         cl->set_notpassed_slp();
2423         if (cl->is_main_loop()) {
2424           // if vector resources are limited, do not allow additional unrolling, also
2425           // do not unroll more on pure vector loops which were not reduced so that we can
2426           // program the post loop to single iteration execution.
2427           if (FLOATPRESSURE > 8) {
2428             C->set_major_progress();
2429             cl->mark_do_unroll_only();
2430           }
2431         }
2432 
2433         if (do_reserve_copy()) {
2434           cl->mark_loop_vectorized();
2435           if (can_process_post_loop) {
2436             // Now create the difference of trip and limit and use it as our mask index.
2437             // Note: We limited the unroll of the vectorized loop so that
2438             //       only vlen-1 size iterations can remain to be mask programmed.
2439             Node *incr = cl->incr();
2440             SubINode *index = new SubINode(cl->limit(), cl->init_trip());
2441             _igvn.register_new_node_with_optimizer(index);
2442             SetVectMaskINode  *mask = new SetVectMaskINode(_phase->get_ctrl(cl->init_trip()), index);
2443             _igvn.register_new_node_with_optimizer(mask);
2444             // make this a single iteration loop
2445             AddINode *new_incr = new AddINode(incr->in(1), mask);
2446             _igvn.register_new_node_with_optimizer(new_incr);
2447             _phase->set_ctrl(new_incr, _phase->get_ctrl(incr));
2448             _igvn.replace_node(incr, new_incr);
2449             cl->mark_is_multiversioned();
2450             cl->loopexit()->add_flag(Node::Flag_has_vector_mask_set);
2451           }
2452         }
2453       }
2454     }
2455   }
2456 
2457   if (do_reserve_copy()) {
2458     make_reversable.use_new();
2459   }
2460   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("\n Final loop after SuperWord"); print_loop(true);})
2461   return;
2462 }
2463 
2464 //------------------------------vector_opd---------------------------
2465 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
2466 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
2467   Node* p0 = p->at(0);
2468   uint vlen = p->size();
2469   Node* opd = p0->in(opd_idx);
2470   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2471 
2472   if (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop()) {
2473     // override vlen with the main loops vector length
2474     vlen = cl->slp_max_unroll();
2475   }
2476 
2477   if (same_inputs(p, opd_idx)) {
2478     if (opd->is_Vector() || opd->is_LoadVector()) {
2479       assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
2480       if (opd_idx == 2 && VectorNode::is_shift(p0)) {
2481         NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("shift's count can't be vector");})
2482         return NULL;
2483       }
2484       return opd; // input is matching vector
2485     }
2486     if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
2487       Compile* C = _phase->C;
2488       Node* cnt = opd;
2489       // Vector instructions do not mask shift count, do it here.
2490       juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2491       const TypeInt* t = opd->find_int_type();
2492       if (t != NULL && t->is_con()) {
2493         juint shift = t->get_con();
2494         if (shift > mask) { // Unsigned cmp
2495           cnt = ConNode::make(TypeInt::make(shift & mask));
2496         }
2497       } else {
2498         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2499           cnt = ConNode::make(TypeInt::make(mask));
2500           _igvn.register_new_node_with_optimizer(cnt);
2501           cnt = new AndINode(opd, cnt);
2502           _igvn.register_new_node_with_optimizer(cnt);
2503           _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2504         }
2505         assert(opd->bottom_type()->isa_int(), "int type only");
2506         if (!opd->bottom_type()->isa_int()) {
2507           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should be int type only");})
2508           return NULL;
2509         }
2510         // Move non constant shift count into vector register.
2511         cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0));
2512       }
2513       if (cnt != opd) {
2514         _igvn.register_new_node_with_optimizer(cnt);
2515         _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2516       }
2517       return cnt;
2518     }
2519     assert(!opd->is_StoreVector(), "such vector is not expected here");
2520     if (opd->is_StoreVector()) {
2521       NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("StoreVector is not expected here");})
2522       return NULL;
2523     }
2524     // Convert scalar input to vector with the same number of elements as
2525     // p0's vector. Use p0's type because size of operand's container in
2526     // vector should match p0's size regardless operand's size.
2527     const Type* p0_t = velt_type(p0);
2528     VectorNode* vn = VectorNode::scalar2vector(opd, vlen, p0_t);
2529 
2530     _igvn.register_new_node_with_optimizer(vn);
2531     _phase->set_ctrl(vn, _phase->get_ctrl(opd));
2532 #ifdef ASSERT
2533     if (TraceNewVectors) {
2534       tty->print("new Vector node: ");
2535       vn->dump();
2536     }
2537 #endif
2538     return vn;
2539   }
2540 
2541   // Insert pack operation
2542   BasicType bt = velt_basic_type(p0);
2543   PackNode* pk = PackNode::make(opd, vlen, bt);
2544   DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
2545 
2546   for (uint i = 1; i < vlen; i++) {
2547     Node* pi = p->at(i);
2548     Node* in = pi->in(opd_idx);
2549     assert(my_pack(in) == NULL, "Should already have been unpacked");
2550     if (my_pack(in) != NULL) {
2551       NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should already have been unpacked");})
2552       return NULL;
2553     }
2554     assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
2555     pk->add_opd(in);
2556   }
2557   _igvn.register_new_node_with_optimizer(pk);
2558   _phase->set_ctrl(pk, _phase->get_ctrl(opd));
2559 #ifdef ASSERT
2560   if (TraceNewVectors) {
2561     tty->print("new Vector node: ");
2562     pk->dump();
2563   }
2564 #endif
2565   return pk;
2566 }
2567 
2568 //------------------------------insert_extracts---------------------------
2569 // If a use of pack p is not a vector use, then replace the
2570 // use with an extract operation.
2571 void SuperWord::insert_extracts(Node_List* p) {
2572   if (p->at(0)->is_Store()) return;
2573   assert(_n_idx_list.is_empty(), "empty (node,index) list");
2574 
2575   // Inspect each use of each pack member.  For each use that is
2576   // not a vector use, replace the use with an extract operation.
2577 
2578   for (uint i = 0; i < p->size(); i++) {
2579     Node* def = p->at(i);
2580     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2581       Node* use = def->fast_out(j);
2582       for (uint k = 0; k < use->req(); k++) {
2583         Node* n = use->in(k);
2584         if (def == n) {
2585           Node_List* u_pk = my_pack(use);
2586           if ((u_pk == NULL || !is_cmov_pack(u_pk) || use->is_CMove()) && !is_vector_use(use, k)) {
2587               _n_idx_list.push(use, k);
2588           }
2589         }
2590       }
2591     }
2592   }
2593 
2594   while (_n_idx_list.is_nonempty()) {
2595     Node* use = _n_idx_list.node();
2596     int   idx = _n_idx_list.index();
2597     _n_idx_list.pop();
2598     Node* def = use->in(idx);
2599 
2600     if (def->is_reduction()) continue;
2601 
2602     // Insert extract operation
2603     _igvn.hash_delete(def);
2604     int def_pos = alignment(def) / data_size(def);
2605 
2606     Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def));
2607     _igvn.register_new_node_with_optimizer(ex);
2608     _phase->set_ctrl(ex, _phase->get_ctrl(def));
2609     _igvn.replace_input_of(use, idx, ex);
2610     _igvn._worklist.push(def);
2611 
2612     bb_insert_after(ex, bb_idx(def));
2613     set_velt_type(ex, velt_type(def));
2614   }
2615 }
2616 
2617 //------------------------------is_vector_use---------------------------
2618 // Is use->in(u_idx) a vector use?
2619 bool SuperWord::is_vector_use(Node* use, int u_idx) {
2620   Node_List* u_pk = my_pack(use);
2621   if (u_pk == NULL) return false;
2622   if (use->is_reduction()) return true;
2623   Node* def = use->in(u_idx);
2624   Node_List* d_pk = my_pack(def);
2625   if (d_pk == NULL) {
2626     // check for scalar promotion
2627     Node* n = u_pk->at(0)->in(u_idx);
2628     for (uint i = 1; i < u_pk->size(); i++) {
2629       if (u_pk->at(i)->in(u_idx) != n) return false;
2630     }
2631     return true;
2632   }
2633   if (u_pk->size() != d_pk->size())
2634     return false;
2635   for (uint i = 0; i < u_pk->size(); i++) {
2636     Node* ui = u_pk->at(i);
2637     Node* di = d_pk->at(i);
2638     if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
2639       return false;
2640   }
2641   return true;
2642 }
2643 
2644 //------------------------------construct_bb---------------------------
2645 // Construct reverse postorder list of block members
2646 bool SuperWord::construct_bb() {
2647   Node* entry = bb();
2648 
2649   assert(_stk.length() == 0,            "stk is empty");
2650   assert(_block.length() == 0,          "block is empty");
2651   assert(_data_entry.length() == 0,     "data_entry is empty");
2652   assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
2653   assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
2654 
2655   // Find non-control nodes with no inputs from within block,
2656   // create a temporary map from node _idx to bb_idx for use
2657   // by the visited and post_visited sets,
2658   // and count number of nodes in block.
2659   int bb_ct = 0;
2660   for (uint i = 0; i < lpt()->_body.size(); i++) {
2661     Node *n = lpt()->_body.at(i);
2662     set_bb_idx(n, i); // Create a temporary map
2663     if (in_bb(n)) {
2664       if (n->is_LoadStore() || n->is_MergeMem() ||
2665           (n->is_Proj() && !n->as_Proj()->is_CFG())) {
2666         // Bailout if the loop has LoadStore, MergeMem or data Proj
2667         // nodes. Superword optimization does not work with them.
2668         return false;
2669       }
2670       bb_ct++;
2671       if (!n->is_CFG()) {
2672         bool found = false;
2673         for (uint j = 0; j < n->req(); j++) {
2674           Node* def = n->in(j);
2675           if (def && in_bb(def)) {
2676             found = true;
2677             break;
2678           }
2679         }
2680         if (!found) {
2681           assert(n != entry, "can't be entry");
2682           _data_entry.push(n);
2683         }
2684       }
2685     }
2686   }
2687 
2688   // Find memory slices (head and tail)
2689   for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
2690     Node *n = lp()->fast_out(i);
2691     if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
2692       Node* n_tail  = n->in(LoopNode::LoopBackControl);
2693       if (n_tail != n->in(LoopNode::EntryControl)) {
2694         if (!n_tail->is_Mem()) {
2695           assert(n_tail->is_Mem(), "unexpected node for memory slice: %s", n_tail->Name());
2696           return false; // Bailout
2697         }
2698         _mem_slice_head.push(n);
2699         _mem_slice_tail.push(n_tail);
2700       }
2701     }
2702   }
2703 
2704   // Create an RPO list of nodes in block
2705 
2706   visited_clear();
2707   post_visited_clear();
2708 
2709   // Push all non-control nodes with no inputs from within block, then control entry
2710   for (int j = 0; j < _data_entry.length(); j++) {
2711     Node* n = _data_entry.at(j);
2712     visited_set(n);
2713     _stk.push(n);
2714   }
2715   visited_set(entry);
2716   _stk.push(entry);
2717 
2718   // Do a depth first walk over out edges
2719   int rpo_idx = bb_ct - 1;
2720   int size;
2721   int reduction_uses = 0;
2722   while ((size = _stk.length()) > 0) {
2723     Node* n = _stk.top(); // Leave node on stack
2724     if (!visited_test_set(n)) {
2725       // forward arc in graph
2726     } else if (!post_visited_test(n)) {
2727       // cross or back arc
2728       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2729         Node *use = n->fast_out(i);
2730         if (in_bb(use) && !visited_test(use) &&
2731             // Don't go around backedge
2732             (!use->is_Phi() || n == entry)) {
2733           if (use->is_reduction()) {
2734             // First see if we can map the reduction on the given system we are on, then
2735             // make a data entry operation for each reduction we see.
2736             BasicType bt = use->bottom_type()->basic_type();
2737             if (ReductionNode::implemented(use->Opcode(), Matcher::min_vector_size(bt), bt)) {
2738               reduction_uses++;
2739             }
2740           }
2741           _stk.push(use);
2742         }
2743       }
2744       if (_stk.length() == size) {
2745         // There were no additional uses, post visit node now
2746         _stk.pop(); // Remove node from stack
2747         assert(rpo_idx >= 0, "");
2748         _block.at_put_grow(rpo_idx, n);
2749         rpo_idx--;
2750         post_visited_set(n);
2751         assert(rpo_idx >= 0 || _stk.is_empty(), "");
2752       }
2753     } else {
2754       _stk.pop(); // Remove post-visited node from stack
2755     }
2756   }//while
2757 
2758   int ii_current = -1;
2759   unsigned int load_idx = (unsigned int)-1;
2760   _ii_order.clear();
2761   // Create real map of block indices for nodes
2762   for (int j = 0; j < _block.length(); j++) {
2763     Node* n = _block.at(j);
2764     set_bb_idx(n, j);
2765     if (_do_vector_loop && n->is_Load()) {
2766       if (ii_current == -1) {
2767         ii_current = _clone_map.gen(n->_idx);
2768         _ii_order.push(ii_current);
2769         load_idx = _clone_map.idx(n->_idx);
2770       } else if (_clone_map.idx(n->_idx) == load_idx && _clone_map.gen(n->_idx) != ii_current) {
2771         ii_current = _clone_map.gen(n->_idx);
2772         _ii_order.push(ii_current);
2773       }
2774     }
2775   }//for
2776 
2777   // Ensure extra info is allocated.
2778   initialize_bb();
2779 
2780 #ifndef PRODUCT
2781   if (_vector_loop_debug && _ii_order.length() > 0) {
2782     tty->print("SuperWord::construct_bb: List of generations: ");
2783     for (int jj = 0; jj < _ii_order.length(); ++jj) {
2784       tty->print("  %d:%d", jj, _ii_order.at(jj));
2785     }
2786     tty->print_cr(" ");
2787   }
2788   if (TraceSuperWord) {
2789     print_bb();
2790     tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
2791     for (int m = 0; m < _data_entry.length(); m++) {
2792       tty->print("%3d ", m);
2793       _data_entry.at(m)->dump();
2794     }
2795     tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
2796     for (int m = 0; m < _mem_slice_head.length(); m++) {
2797       tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
2798       tty->print("    ");    _mem_slice_tail.at(m)->dump();
2799     }
2800   }
2801 #endif
2802   assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
2803   return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0);
2804 }
2805 
2806 //------------------------------initialize_bb---------------------------
2807 // Initialize per node info
2808 void SuperWord::initialize_bb() {
2809   Node* last = _block.at(_block.length() - 1);
2810   grow_node_info(bb_idx(last));
2811 }
2812 
2813 //------------------------------bb_insert_after---------------------------
2814 // Insert n into block after pos
2815 void SuperWord::bb_insert_after(Node* n, int pos) {
2816   int n_pos = pos + 1;
2817   // Make room
2818   for (int i = _block.length() - 1; i >= n_pos; i--) {
2819     _block.at_put_grow(i+1, _block.at(i));
2820   }
2821   for (int j = _node_info.length() - 1; j >= n_pos; j--) {
2822     _node_info.at_put_grow(j+1, _node_info.at(j));
2823   }
2824   // Set value
2825   _block.at_put_grow(n_pos, n);
2826   _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
2827   // Adjust map from node->_idx to _block index
2828   for (int i = n_pos; i < _block.length(); i++) {
2829     set_bb_idx(_block.at(i), i);
2830   }
2831 }
2832 
2833 //------------------------------compute_max_depth---------------------------
2834 // Compute max depth for expressions from beginning of block
2835 // Use to prune search paths during test for independence.
2836 void SuperWord::compute_max_depth() {
2837   int ct = 0;
2838   bool again;
2839   do {
2840     again = false;
2841     for (int i = 0; i < _block.length(); i++) {
2842       Node* n = _block.at(i);
2843       if (!n->is_Phi()) {
2844         int d_orig = depth(n);
2845         int d_in   = 0;
2846         for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
2847           Node* pred = preds.current();
2848           if (in_bb(pred)) {
2849             d_in = MAX2(d_in, depth(pred));
2850           }
2851         }
2852         if (d_in + 1 != d_orig) {
2853           set_depth(n, d_in + 1);
2854           again = true;
2855         }
2856       }
2857     }
2858     ct++;
2859   } while (again);
2860 
2861   if (TraceSuperWord && Verbose) {
2862     tty->print_cr("compute_max_depth iterated: %d times", ct);
2863   }
2864 }
2865 
2866 //-------------------------compute_vector_element_type-----------------------
2867 // Compute necessary vector element type for expressions
2868 // This propagates backwards a narrower integer type when the
2869 // upper bits of the value are not needed.
2870 // Example:  char a,b,c;  a = b + c;
2871 // Normally the type of the add is integer, but for packed character
2872 // operations the type of the add needs to be char.
2873 void SuperWord::compute_vector_element_type() {
2874   if (TraceSuperWord && Verbose) {
2875     tty->print_cr("\ncompute_velt_type:");
2876   }
2877 
2878   // Initial type
2879   for (int i = 0; i < _block.length(); i++) {
2880     Node* n = _block.at(i);
2881     set_velt_type(n, container_type(n));
2882   }
2883 
2884   // Propagate integer narrowed type backwards through operations
2885   // that don't depend on higher order bits
2886   for (int i = _block.length() - 1; i >= 0; i--) {
2887     Node* n = _block.at(i);
2888     // Only integer types need be examined
2889     const Type* vtn = velt_type(n);
2890     if (vtn->basic_type() == T_INT) {
2891       uint start, end;
2892       VectorNode::vector_operands(n, &start, &end);
2893 
2894       for (uint j = start; j < end; j++) {
2895         Node* in  = n->in(j);
2896         // Don't propagate through a memory
2897         if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
2898             data_size(n) < data_size(in)) {
2899           bool same_type = true;
2900           for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
2901             Node *use = in->fast_out(k);
2902             if (!in_bb(use) || !same_velt_type(use, n)) {
2903               same_type = false;
2904               break;
2905             }
2906           }
2907           if (same_type) {
2908             // For right shifts of small integer types (bool, byte, char, short)
2909             // we need precise information about sign-ness. Only Load nodes have
2910             // this information because Store nodes are the same for signed and
2911             // unsigned values. And any arithmetic operation after a load may
2912             // expand a value to signed Int so such right shifts can't be used
2913             // because vector elements do not have upper bits of Int.
2914             const Type* vt = vtn;
2915             if (VectorNode::is_shift(in)) {
2916               Node* load = in->in(1);
2917               if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
2918                 vt = velt_type(load);
2919               } else if (in->Opcode() != Op_LShiftI) {
2920                 // Widen type to Int to avoid creation of right shift vector
2921                 // (align + data_size(s1) check in stmts_can_pack() will fail).
2922                 // Note, left shifts work regardless type.
2923                 vt = TypeInt::INT;
2924               }
2925             }
2926             set_velt_type(in, vt);
2927           }
2928         }
2929       }
2930     }
2931   }
2932 #ifndef PRODUCT
2933   if (TraceSuperWord && Verbose) {
2934     for (int i = 0; i < _block.length(); i++) {
2935       Node* n = _block.at(i);
2936       velt_type(n)->dump();
2937       tty->print("\t");
2938       n->dump();
2939     }
2940   }
2941 #endif
2942 }
2943 
2944 //------------------------------memory_alignment---------------------------
2945 // Alignment within a vector memory reference
2946 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
2947   #ifndef PRODUCT
2948     if(TraceSuperWord && Verbose) {
2949       tty->print("SuperWord::memory_alignment within a vector memory reference for %d:  ", s->_idx); s->dump();
2950     }
2951   #endif
2952   NOT_PRODUCT(SWPointer::Tracer::Depth ddd(0);)
2953   SWPointer p(s, this, NULL, false);
2954   if (!p.valid()) {
2955     NOT_PRODUCT(if(is_trace_alignment()) tty->print("SWPointer::memory_alignment: SWPointer p invalid, return bottom_align");)
2956     return bottom_align;
2957   }
2958   int vw = vector_width_in_bytes(s);
2959   if (vw < 2) {
2960     NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SWPointer::memory_alignment: vector_width_in_bytes < 2, return bottom_align");)
2961     return bottom_align; // No vectors for this type
2962   }
2963   int offset  = p.offset_in_bytes();
2964   offset     += iv_adjust*p.memory_size();
2965   int off_rem = offset % vw;
2966   int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
2967   if (TraceSuperWord && Verbose) {
2968     tty->print_cr("SWPointer::memory_alignment: off_rem = %d, off_mod = %d", off_rem, off_mod);
2969   }
2970   return off_mod;
2971 }
2972 
2973 //---------------------------container_type---------------------------
2974 // Smallest type containing range of values
2975 const Type* SuperWord::container_type(Node* n) {
2976   if (n->is_Mem()) {
2977     BasicType bt = n->as_Mem()->memory_type();
2978     if (n->is_Store() && (bt == T_CHAR)) {
2979       // Use T_SHORT type instead of T_CHAR for stored values because any
2980       // preceding arithmetic operation extends values to signed Int.
2981       bt = T_SHORT;
2982     }
2983     if (n->Opcode() == Op_LoadUB) {
2984       // Adjust type for unsigned byte loads, it is important for right shifts.
2985       // T_BOOLEAN is used because there is no basic type representing type
2986       // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
2987       // size (one byte) and sign is important.
2988       bt = T_BOOLEAN;
2989     }
2990     return Type::get_const_basic_type(bt);
2991   }
2992   const Type* t = _igvn.type(n);
2993   if (t->basic_type() == T_INT) {
2994     // A narrow type of arithmetic operations will be determined by
2995     // propagating the type of memory operations.
2996     return TypeInt::INT;
2997   }
2998   return t;
2999 }
3000 
3001 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
3002   const Type* vt1 = velt_type(n1);
3003   const Type* vt2 = velt_type(n2);
3004   if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
3005     // Compare vectors element sizes for integer types.
3006     return data_size(n1) == data_size(n2);
3007   }
3008   return vt1 == vt2;
3009 }
3010 
3011 //------------------------------in_packset---------------------------
3012 // Are s1 and s2 in a pack pair and ordered as s1,s2?
3013 bool SuperWord::in_packset(Node* s1, Node* s2) {
3014   for (int i = 0; i < _packset.length(); i++) {
3015     Node_List* p = _packset.at(i);
3016     assert(p->size() == 2, "must be");
3017     if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
3018       return true;
3019     }
3020   }
3021   return false;
3022 }
3023 
3024 //------------------------------in_pack---------------------------
3025 // Is s in pack p?
3026 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
3027   for (uint i = 0; i < p->size(); i++) {
3028     if (p->at(i) == s) {
3029       return p;
3030     }
3031   }
3032   return NULL;
3033 }
3034 
3035 //------------------------------remove_pack_at---------------------------
3036 // Remove the pack at position pos in the packset
3037 void SuperWord::remove_pack_at(int pos) {
3038   Node_List* p = _packset.at(pos);
3039   for (uint i = 0; i < p->size(); i++) {
3040     Node* s = p->at(i);
3041     set_my_pack(s, NULL);
3042   }
3043   _packset.remove_at(pos);
3044 }
3045 
3046 void SuperWord::packset_sort(int n) {
3047   // simple bubble sort so that we capitalize with O(n) when its already sorted
3048   while (n != 0) {
3049     bool swapped = false;
3050     for (int i = 1; i < n; i++) {
3051       Node_List* q_low = _packset.at(i-1);
3052       Node_List* q_i = _packset.at(i);
3053 
3054       // only swap when we find something to swap
3055       if (alignment(q_low->at(0)) > alignment(q_i->at(0))) {
3056         Node_List* t = q_i;
3057         *(_packset.adr_at(i)) = q_low;
3058         *(_packset.adr_at(i-1)) = q_i;
3059         swapped = true;
3060       }
3061     }
3062     if (swapped == false) break;
3063     n--;
3064   }
3065 }
3066 
3067 //------------------------------executed_first---------------------------
3068 // Return the node executed first in pack p.  Uses the RPO block list
3069 // to determine order.
3070 Node* SuperWord::executed_first(Node_List* p) {
3071   Node* n = p->at(0);
3072   int n_rpo = bb_idx(n);
3073   for (uint i = 1; i < p->size(); i++) {
3074     Node* s = p->at(i);
3075     int s_rpo = bb_idx(s);
3076     if (s_rpo < n_rpo) {
3077       n = s;
3078       n_rpo = s_rpo;
3079     }
3080   }
3081   return n;
3082 }
3083 
3084 //------------------------------executed_last---------------------------
3085 // Return the node executed last in pack p.
3086 Node* SuperWord::executed_last(Node_List* p) {
3087   Node* n = p->at(0);
3088   int n_rpo = bb_idx(n);
3089   for (uint i = 1; i < p->size(); i++) {
3090     Node* s = p->at(i);
3091     int s_rpo = bb_idx(s);
3092     if (s_rpo > n_rpo) {
3093       n = s;
3094       n_rpo = s_rpo;
3095     }
3096   }
3097   return n;
3098 }
3099 
3100 LoadNode::ControlDependency SuperWord::control_dependency(Node_List* p) {
3101   LoadNode::ControlDependency dep = LoadNode::DependsOnlyOnTest;
3102   for (uint i = 0; i < p->size(); i++) {
3103     Node* n = p->at(i);
3104     assert(n->is_Load(), "only meaningful for loads");
3105     if (!n->depends_only_on_test()) {
3106       dep = LoadNode::Pinned;
3107     }
3108   }
3109   return dep;
3110 }
3111 
3112 
3113 //----------------------------align_initial_loop_index---------------------------
3114 // Adjust pre-loop limit so that in main loop, a load/store reference
3115 // to align_to_ref will be a position zero in the vector.
3116 //   (iv + k) mod vector_align == 0
3117 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
3118   CountedLoopNode *main_head = lp()->as_CountedLoop();
3119   assert(main_head->is_main_loop(), "");
3120   CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
3121   assert(pre_end != NULL, "we must have a correct pre-loop");
3122   Node *pre_opaq1 = pre_end->limit();
3123   assert(pre_opaq1->Opcode() == Op_Opaque1, "");
3124   Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
3125   Node *lim0 = pre_opaq->in(1);
3126 
3127   // Where we put new limit calculations
3128   Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
3129 
3130   // Ensure the original loop limit is available from the
3131   // pre-loop Opaque1 node.
3132   Node *orig_limit = pre_opaq->original_loop_limit();
3133   assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
3134 
3135   SWPointer align_to_ref_p(align_to_ref, this, NULL, false);
3136   assert(align_to_ref_p.valid(), "sanity");
3137 
3138   // Given:
3139   //     lim0 == original pre loop limit
3140   //     V == v_align (power of 2)
3141   //     invar == extra invariant piece of the address expression
3142   //     e == offset [ +/- invar ]
3143   //
3144   // When reassociating expressions involving '%' the basic rules are:
3145   //     (a - b) % k == 0   =>  a % k == b % k
3146   // and:
3147   //     (a + b) % k == 0   =>  a % k == (k - b) % k
3148   //
3149   // For stride > 0 && scale > 0,
3150   //   Derive the new pre-loop limit "lim" such that the two constraints:
3151   //     (1) lim = lim0 + N           (where N is some positive integer < V)
3152   //     (2) (e + lim) % V == 0
3153   //   are true.
3154   //
3155   //   Substituting (1) into (2),
3156   //     (e + lim0 + N) % V == 0
3157   //   solve for N:
3158   //     N = (V - (e + lim0)) % V
3159   //   substitute back into (1), so that new limit
3160   //     lim = lim0 + (V - (e + lim0)) % V
3161   //
3162   // For stride > 0 && scale < 0
3163   //   Constraints:
3164   //     lim = lim0 + N
3165   //     (e - lim) % V == 0
3166   //   Solving for lim:
3167   //     (e - lim0 - N) % V == 0
3168   //     N = (e - lim0) % V
3169   //     lim = lim0 + (e - lim0) % V
3170   //
3171   // For stride < 0 && scale > 0
3172   //   Constraints:
3173   //     lim = lim0 - N
3174   //     (e + lim) % V == 0
3175   //   Solving for lim:
3176   //     (e + lim0 - N) % V == 0
3177   //     N = (e + lim0) % V
3178   //     lim = lim0 - (e + lim0) % V
3179   //
3180   // For stride < 0 && scale < 0
3181   //   Constraints:
3182   //     lim = lim0 - N
3183   //     (e - lim) % V == 0
3184   //   Solving for lim:
3185   //     (e - lim0 + N) % V == 0
3186   //     N = (V - (e - lim0)) % V
3187   //     lim = lim0 - (V - (e - lim0)) % V
3188 
3189   int vw = vector_width_in_bytes(align_to_ref);
3190   int stride   = iv_stride();
3191   int scale    = align_to_ref_p.scale_in_bytes();
3192   int elt_size = align_to_ref_p.memory_size();
3193   int v_align  = vw / elt_size;
3194   assert(v_align > 1, "sanity");
3195   int offset   = align_to_ref_p.offset_in_bytes() / elt_size;
3196   Node *offsn  = _igvn.intcon(offset);
3197 
3198   Node *e = offsn;
3199   if (align_to_ref_p.invar() != NULL) {
3200     // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
3201     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3202     Node* invar = align_to_ref_p.invar();
3203     if (_igvn.type(invar)->isa_long()) {
3204       // Computations are done % (vector width/element size) so it's
3205       // safe to simply convert invar to an int and loose the upper 32
3206       // bit half.
3207       invar = new ConvL2INode(invar);
3208       _igvn.register_new_node_with_optimizer(invar);
3209     }
3210     Node* aref = new URShiftINode(invar, log2_elt);
3211     _igvn.register_new_node_with_optimizer(aref);
3212     _phase->set_ctrl(aref, pre_ctrl);
3213     if (align_to_ref_p.negate_invar()) {
3214       e = new SubINode(e, aref);
3215     } else {
3216       e = new AddINode(e, aref);
3217     }
3218     _igvn.register_new_node_with_optimizer(e);
3219     _phase->set_ctrl(e, pre_ctrl);
3220   }
3221   if (vw > ObjectAlignmentInBytes) {
3222     // incorporate base e +/- base && Mask >>> log2(elt)
3223     Node* xbase = new CastP2XNode(NULL, align_to_ref_p.base());
3224     _igvn.register_new_node_with_optimizer(xbase);
3225 #ifdef _LP64
3226     xbase  = new ConvL2INode(xbase);
3227     _igvn.register_new_node_with_optimizer(xbase);
3228 #endif
3229     Node* mask = _igvn.intcon(vw-1);
3230     Node* masked_xbase  = new AndINode(xbase, mask);
3231     _igvn.register_new_node_with_optimizer(masked_xbase);
3232     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3233     Node* bref     = new URShiftINode(masked_xbase, log2_elt);
3234     _igvn.register_new_node_with_optimizer(bref);
3235     _phase->set_ctrl(bref, pre_ctrl);
3236     e = new AddINode(e, bref);
3237     _igvn.register_new_node_with_optimizer(e);
3238     _phase->set_ctrl(e, pre_ctrl);
3239   }
3240 
3241   // compute e +/- lim0
3242   if (scale < 0) {
3243     e = new SubINode(e, lim0);
3244   } else {
3245     e = new AddINode(e, lim0);
3246   }
3247   _igvn.register_new_node_with_optimizer(e);
3248   _phase->set_ctrl(e, pre_ctrl);
3249 
3250   if (stride * scale > 0) {
3251     // compute V - (e +/- lim0)
3252     Node* va  = _igvn.intcon(v_align);
3253     e = new SubINode(va, e);
3254     _igvn.register_new_node_with_optimizer(e);
3255     _phase->set_ctrl(e, pre_ctrl);
3256   }
3257   // compute N = (exp) % V
3258   Node* va_msk = _igvn.intcon(v_align - 1);
3259   Node* N = new AndINode(e, va_msk);
3260   _igvn.register_new_node_with_optimizer(N);
3261   _phase->set_ctrl(N, pre_ctrl);
3262 
3263   //   substitute back into (1), so that new limit
3264   //     lim = lim0 + N
3265   Node* lim;
3266   if (stride < 0) {
3267     lim = new SubINode(lim0, N);
3268   } else {
3269     lim = new AddINode(lim0, N);
3270   }
3271   _igvn.register_new_node_with_optimizer(lim);
3272   _phase->set_ctrl(lim, pre_ctrl);
3273   Node* constrained =
3274     (stride > 0) ? (Node*) new MinINode(lim, orig_limit)
3275                  : (Node*) new MaxINode(lim, orig_limit);
3276   _igvn.register_new_node_with_optimizer(constrained);
3277   _phase->set_ctrl(constrained, pre_ctrl);
3278   _igvn.replace_input_of(pre_opaq, 1, constrained);
3279 }
3280 
3281 //----------------------------get_pre_loop_end---------------------------
3282 // Find pre loop end from main loop.  Returns null if none.
3283 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode* cl) {
3284   // The loop cannot be optimized if the graph shape at
3285   // the loop entry is inappropriate.
3286   if (!PhaseIdealLoop::is_canonical_loop_entry(cl)) {
3287     return NULL;
3288   }
3289 
3290   Node* p_f = cl->in(LoopNode::EntryControl)->in(0)->in(0);
3291   if (!p_f->is_IfFalse()) return NULL;
3292   if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
3293   CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
3294   CountedLoopNode* loop_node = pre_end->loopnode();
3295   if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
3296   return pre_end;
3297 }
3298 
3299 //------------------------------init---------------------------
3300 void SuperWord::init() {
3301   _dg.init();
3302   _packset.clear();
3303   _disjoint_ptrs.clear();
3304   _block.clear();
3305   _post_block.clear();
3306   _data_entry.clear();
3307   _mem_slice_head.clear();
3308   _mem_slice_tail.clear();
3309   _iteration_first.clear();
3310   _iteration_last.clear();
3311   _node_info.clear();
3312   _align_to_ref = NULL;
3313   _lpt = NULL;
3314   _lp = NULL;
3315   _bb = NULL;
3316   _iv = NULL;
3317   _race_possible = 0;
3318   _early_return = false;
3319   _num_work_vecs = 0;
3320   _num_reductions = 0;
3321 }
3322 
3323 //------------------------------restart---------------------------
3324 void SuperWord::restart() {
3325   _dg.init();
3326   _packset.clear();
3327   _disjoint_ptrs.clear();
3328   _block.clear();
3329   _post_block.clear();
3330   _data_entry.clear();
3331   _mem_slice_head.clear();
3332   _mem_slice_tail.clear();
3333   _node_info.clear();
3334 }
3335 
3336 //------------------------------print_packset---------------------------
3337 void SuperWord::print_packset() {
3338 #ifndef PRODUCT
3339   tty->print_cr("packset");
3340   for (int i = 0; i < _packset.length(); i++) {
3341     tty->print_cr("Pack: %d", i);
3342     Node_List* p = _packset.at(i);
3343     print_pack(p);
3344   }
3345 #endif
3346 }
3347 
3348 //------------------------------print_pack---------------------------
3349 void SuperWord::print_pack(Node_List* p) {
3350   for (uint i = 0; i < p->size(); i++) {
3351     print_stmt(p->at(i));
3352   }
3353 }
3354 
3355 //------------------------------print_bb---------------------------
3356 void SuperWord::print_bb() {
3357 #ifndef PRODUCT
3358   tty->print_cr("\nBlock");
3359   for (int i = 0; i < _block.length(); i++) {
3360     Node* n = _block.at(i);
3361     tty->print("%d ", i);
3362     if (n) {
3363       n->dump();
3364     }
3365   }
3366 #endif
3367 }
3368 
3369 //------------------------------print_stmt---------------------------
3370 void SuperWord::print_stmt(Node* s) {
3371 #ifndef PRODUCT
3372   tty->print(" align: %d \t", alignment(s));
3373   s->dump();
3374 #endif
3375 }
3376 
3377 //------------------------------blank---------------------------
3378 char* SuperWord::blank(uint depth) {
3379   static char blanks[101];
3380   assert(depth < 101, "too deep");
3381   for (uint i = 0; i < depth; i++) blanks[i] = ' ';
3382   blanks[depth] = '\0';
3383   return blanks;
3384 }
3385 
3386 
3387 //==============================SWPointer===========================
3388 #ifndef PRODUCT
3389 int SWPointer::Tracer::_depth = 0;
3390 #endif
3391 //----------------------------SWPointer------------------------
3392 SWPointer::SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only) :
3393   _mem(mem), _slp(slp),  _base(NULL),  _adr(NULL),
3394   _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3395   _nstack(nstack), _analyze_only(analyze_only),
3396   _stack_idx(0)
3397 #ifndef PRODUCT
3398   , _tracer(slp)
3399 #endif
3400 {
3401   NOT_PRODUCT(_tracer.ctor_1(mem);)
3402 
3403   Node* adr = mem->in(MemNode::Address);
3404   if (!adr->is_AddP()) {
3405     assert(!valid(), "too complex");
3406     return;
3407   }
3408   // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
3409   Node* base = adr->in(AddPNode::Base);
3410   // The base address should be loop invariant
3411   if (!invariant(base)) {
3412     assert(!valid(), "base address is loop variant");
3413     return;
3414   }
3415   //unsafe reference could not be aligned appropriately without runtime checking
3416   if (base == NULL || base->bottom_type() == Type::TOP) {
3417     assert(!valid(), "unsafe access");
3418     return;
3419   }
3420 
3421   NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.store_depth();)
3422   NOT_PRODUCT(_tracer.ctor_2(adr);)
3423 
3424   int i;
3425   for (i = 0; i < 3; i++) {
3426     NOT_PRODUCT(_tracer.ctor_3(adr, i);)
3427 
3428     if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
3429       assert(!valid(), "too complex");
3430       return;
3431     }
3432     adr = adr->in(AddPNode::Address);
3433     NOT_PRODUCT(_tracer.ctor_4(adr, i);)
3434 
3435     if (base == adr || !adr->is_AddP()) {
3436       NOT_PRODUCT(_tracer.ctor_5(adr, base, i);)
3437       break; // stop looking at addp's
3438     }
3439   }
3440   NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.restore_depth();)
3441   NOT_PRODUCT(_tracer.ctor_6(mem);)
3442 
3443   _base = base;
3444   _adr  = adr;
3445   assert(valid(), "Usable");
3446 }
3447 
3448 // Following is used to create a temporary object during
3449 // the pattern match of an address expression.
3450 SWPointer::SWPointer(SWPointer* p) :
3451   _mem(p->_mem), _slp(p->_slp),  _base(NULL),  _adr(NULL),
3452   _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3453   _nstack(p->_nstack), _analyze_only(p->_analyze_only),
3454   _stack_idx(p->_stack_idx)
3455   #ifndef PRODUCT
3456   , _tracer(p->_slp)
3457   #endif
3458 {}
3459 
3460 
3461 bool SWPointer::invariant(Node* n) {
3462   NOT_PRODUCT(Tracer::Depth dd;)
3463   Node *n_c = phase()->get_ctrl(n);
3464   NOT_PRODUCT(_tracer.invariant_1(n, n_c);)
3465   return !lpt()->is_member(phase()->get_loop(n_c));
3466 }
3467 //------------------------scaled_iv_plus_offset--------------------
3468 // Match: k*iv + offset
3469 // where: k is a constant that maybe zero, and
3470 //        offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
3471 bool SWPointer::scaled_iv_plus_offset(Node* n) {
3472   NOT_PRODUCT(Tracer::Depth ddd;)
3473   NOT_PRODUCT(_tracer.scaled_iv_plus_offset_1(n);)
3474 
3475   if (scaled_iv(n)) {
3476     NOT_PRODUCT(_tracer.scaled_iv_plus_offset_2(n);)
3477     return true;
3478   }
3479 
3480   if (offset_plus_k(n)) {
3481     NOT_PRODUCT(_tracer.scaled_iv_plus_offset_3(n);)
3482     return true;
3483   }
3484 
3485   int opc = n->Opcode();
3486   if (opc == Op_AddI) {
3487     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
3488       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_4(n);)
3489       return true;
3490     }
3491     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3492       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_5(n);)
3493       return true;
3494     }
3495   } else if (opc == Op_SubI) {
3496     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
3497       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_6(n);)
3498       return true;
3499     }
3500     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3501       _scale *= -1;
3502       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_7(n);)
3503       return true;
3504     }
3505   }
3506 
3507   NOT_PRODUCT(_tracer.scaled_iv_plus_offset_8(n);)
3508   return false;
3509 }
3510 
3511 //----------------------------scaled_iv------------------------
3512 // Match: k*iv where k is a constant that's not zero
3513 bool SWPointer::scaled_iv(Node* n) {
3514   NOT_PRODUCT(Tracer::Depth ddd;)
3515   NOT_PRODUCT(_tracer.scaled_iv_1(n);)
3516 
3517   if (_scale != 0) { // already found a scale
3518     NOT_PRODUCT(_tracer.scaled_iv_2(n, _scale);)
3519     return false;
3520   }
3521 
3522   if (n == iv()) {
3523     _scale = 1;
3524     NOT_PRODUCT(_tracer.scaled_iv_3(n, _scale);)
3525     return true;
3526   }
3527   if (_analyze_only && (invariant(n) == false)) {
3528     _nstack->push(n, _stack_idx++);
3529   }
3530 
3531   int opc = n->Opcode();
3532   if (opc == Op_MulI) {
3533     if (n->in(1) == iv() && n->in(2)->is_Con()) {
3534       _scale = n->in(2)->get_int();
3535       NOT_PRODUCT(_tracer.scaled_iv_4(n, _scale);)
3536       return true;
3537     } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
3538       _scale = n->in(1)->get_int();
3539       NOT_PRODUCT(_tracer.scaled_iv_5(n, _scale);)
3540       return true;
3541     }
3542   } else if (opc == Op_LShiftI) {
3543     if (n->in(1) == iv() && n->in(2)->is_Con()) {
3544       _scale = 1 << n->in(2)->get_int();
3545       NOT_PRODUCT(_tracer.scaled_iv_6(n, _scale);)
3546       return true;
3547     }
3548   } else if (opc == Op_ConvI2L) {
3549     if (n->in(1)->Opcode() == Op_CastII &&
3550         n->in(1)->as_CastII()->has_range_check()) {
3551       // Skip range check dependent CastII nodes
3552       n = n->in(1);
3553     }
3554     if (scaled_iv_plus_offset(n->in(1))) {
3555       NOT_PRODUCT(_tracer.scaled_iv_7(n);)
3556       return true;
3557     }
3558   } else if (opc == Op_LShiftL) {
3559     if (!has_iv() && _invar == NULL) {
3560       // Need to preserve the current _offset value, so
3561       // create a temporary object for this expression subtree.
3562       // Hacky, so should re-engineer the address pattern match.
3563       NOT_PRODUCT(Tracer::Depth dddd;)
3564       SWPointer tmp(this);
3565       NOT_PRODUCT(_tracer.scaled_iv_8(n, &tmp);)
3566 
3567       if (tmp.scaled_iv_plus_offset(n->in(1))) {
3568         if (tmp._invar == NULL || _slp->do_vector_loop()) {
3569           int mult = 1 << n->in(2)->get_int();
3570           _scale   = tmp._scale  * mult;
3571           _offset += tmp._offset * mult;
3572           NOT_PRODUCT(_tracer.scaled_iv_9(n, _scale, _offset, mult);)
3573           return true;
3574         }
3575       }
3576     }
3577   }
3578   NOT_PRODUCT(_tracer.scaled_iv_10(n);)
3579   return false;
3580 }
3581 
3582 //----------------------------offset_plus_k------------------------
3583 // Match: offset is (k [+/- invariant])
3584 // where k maybe zero and invariant is optional, but not both.
3585 bool SWPointer::offset_plus_k(Node* n, bool negate) {
3586   NOT_PRODUCT(Tracer::Depth ddd;)
3587   NOT_PRODUCT(_tracer.offset_plus_k_1(n);)
3588 
3589   int opc = n->Opcode();
3590   if (opc == Op_ConI) {
3591     _offset += negate ? -(n->get_int()) : n->get_int();
3592     NOT_PRODUCT(_tracer.offset_plus_k_2(n, _offset);)
3593     return true;
3594   } else if (opc == Op_ConL) {
3595     // Okay if value fits into an int
3596     const TypeLong* t = n->find_long_type();
3597     if (t->higher_equal(TypeLong::INT)) {
3598       jlong loff = n->get_long();
3599       jint  off  = (jint)loff;
3600       _offset += negate ? -off : loff;
3601       NOT_PRODUCT(_tracer.offset_plus_k_3(n, _offset);)
3602       return true;
3603     }
3604     NOT_PRODUCT(_tracer.offset_plus_k_4(n);)
3605     return false;
3606   }
3607   if (_invar != NULL) { // already has an invariant
3608     NOT_PRODUCT(_tracer.offset_plus_k_5(n, _invar);)
3609     return false;
3610   }
3611 
3612   if (_analyze_only && (invariant(n) == false)) {
3613     _nstack->push(n, _stack_idx++);
3614   }
3615   if (opc == Op_AddI) {
3616     if (n->in(2)->is_Con() && invariant(n->in(1))) {
3617       _negate_invar = negate;
3618       _invar = n->in(1);
3619       _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3620       NOT_PRODUCT(_tracer.offset_plus_k_6(n, _invar, _negate_invar, _offset);)
3621       return true;
3622     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3623       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3624       _negate_invar = negate;
3625       _invar = n->in(2);
3626       NOT_PRODUCT(_tracer.offset_plus_k_7(n, _invar, _negate_invar, _offset);)
3627       return true;
3628     }
3629   }
3630   if (opc == Op_SubI) {
3631     if (n->in(2)->is_Con() && invariant(n->in(1))) {
3632       _negate_invar = negate;
3633       _invar = n->in(1);
3634       _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3635       NOT_PRODUCT(_tracer.offset_plus_k_8(n, _invar, _negate_invar, _offset);)
3636       return true;
3637     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3638       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3639       _negate_invar = !negate;
3640       _invar = n->in(2);
3641       NOT_PRODUCT(_tracer.offset_plus_k_9(n, _invar, _negate_invar, _offset);)
3642       return true;
3643     }
3644   }
3645   if (invariant(n)) {
3646     if (opc == Op_ConvI2L) {
3647       n = n->in(1);
3648       if (n->Opcode() == Op_CastII &&
3649           n->as_CastII()->has_range_check()) {
3650         // Skip range check dependent CastII nodes
3651         assert(invariant(n), "sanity");
3652         n = n->in(1);
3653       }
3654     }
3655     _negate_invar = negate;
3656     _invar = n;
3657     NOT_PRODUCT(_tracer.offset_plus_k_10(n, _invar, _negate_invar, _offset);)
3658     return true;
3659   }
3660 
3661   NOT_PRODUCT(_tracer.offset_plus_k_11(n);)
3662   return false;
3663 }
3664 
3665 //----------------------------print------------------------
3666 void SWPointer::print() {
3667 #ifndef PRODUCT
3668   tty->print("base: %d  adr: %d  scale: %d  offset: %d  invar: %c%d\n",
3669              _base != NULL ? _base->_idx : 0,
3670              _adr  != NULL ? _adr->_idx  : 0,
3671              _scale, _offset,
3672              _negate_invar?'-':'+',
3673              _invar != NULL ? _invar->_idx : 0);
3674 #endif
3675 }
3676 
3677 //----------------------------tracing------------------------
3678 #ifndef PRODUCT
3679 void SWPointer::Tracer::print_depth() {
3680   for (int ii = 0; ii<_depth; ++ii) tty->print("  ");
3681 }
3682 
3683 void SWPointer::Tracer::ctor_1 (Node* mem) {
3684   if(_slp->is_trace_alignment()) {
3685     print_depth(); tty->print(" %d SWPointer::SWPointer: start alignment analysis", mem->_idx); mem->dump();
3686   }
3687 }
3688 
3689 void SWPointer::Tracer::ctor_2(Node* adr) {
3690   if(_slp->is_trace_alignment()) {
3691     //store_depth();
3692     inc_depth();
3693     print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: ", adr->_idx); adr->dump();
3694     inc_depth();
3695     print_depth(); tty->print(" %d (base) SWPointer::SWPointer: ", adr->in(AddPNode::Base)->_idx); adr->in(AddPNode::Base)->dump();
3696   }
3697 }
3698 
3699 void SWPointer::Tracer::ctor_3(Node* adr, int i) {
3700   if(_slp->is_trace_alignment()) {
3701     inc_depth();
3702     Node* offset = adr->in(AddPNode::Offset);
3703     print_depth(); tty->print(" %d (offset) SWPointer::SWPointer: i = %d: ", offset->_idx, i); offset->dump();
3704   }
3705 }
3706 
3707 void SWPointer::Tracer::ctor_4(Node* adr, int i) {
3708   if(_slp->is_trace_alignment()) {
3709     inc_depth();
3710     print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: i = %d: ", adr->_idx, i); adr->dump();
3711   }
3712 }
3713 
3714 void SWPointer::Tracer::ctor_5(Node* adr, Node* base, int i) {
3715   if(_slp->is_trace_alignment()) {
3716     inc_depth();
3717     if (base == adr) {
3718       print_depth(); tty->print_cr("  \\ %d (adr) == %d (base) SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, base->_idx, i);
3719     } else if (!adr->is_AddP()) {
3720       print_depth(); tty->print_cr("  \\ %d (adr) is NOT Addp SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, i);
3721     }
3722   }
3723 }
3724 
3725 void SWPointer::Tracer::ctor_6(Node* mem) {
3726   if(_slp->is_trace_alignment()) {
3727     //restore_depth();
3728     print_depth(); tty->print_cr(" %d (adr) SWPointer::SWPointer: stop analysis", mem->_idx);
3729   }
3730 }
3731 
3732 void SWPointer::Tracer::invariant_1(Node *n, Node *n_c) {
3733   if (_slp->do_vector_loop() && _slp->is_debug() && _slp->_lpt->is_member(_slp->_phase->get_loop(n_c)) != (int)_slp->in_bb(n)) {
3734     int is_member =  _slp->_lpt->is_member(_slp->_phase->get_loop(n_c));
3735     int in_bb     =  _slp->in_bb(n);
3736     print_depth(); tty->print("  \\ ");  tty->print_cr(" %d SWPointer::invariant  conditions differ: n_c %d", n->_idx, n_c->_idx);
3737     print_depth(); tty->print("  \\ ");  tty->print_cr("is_member %d, in_bb %d", is_member, in_bb);
3738     print_depth(); tty->print("  \\ ");  n->dump();
3739     print_depth(); tty->print("  \\ ");  n_c->dump();
3740   }
3741 }
3742 
3743 void SWPointer::Tracer::scaled_iv_plus_offset_1(Node* n) {
3744   if(_slp->is_trace_alignment()) {
3745     print_depth(); tty->print(" %d SWPointer::scaled_iv_plus_offset testing node: ", n->_idx);
3746     n->dump();
3747   }
3748 }
3749 
3750 void SWPointer::Tracer::scaled_iv_plus_offset_2(Node* n) {
3751   if(_slp->is_trace_alignment()) {
3752     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
3753   }
3754 }
3755 
3756 void SWPointer::Tracer::scaled_iv_plus_offset_3(Node* n) {
3757   if(_slp->is_trace_alignment()) {
3758     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
3759   }
3760 }
3761 
3762 void SWPointer::Tracer::scaled_iv_plus_offset_4(Node* n) {
3763   if(_slp->is_trace_alignment()) {
3764     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
3765     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
3766     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
3767   }
3768 }
3769 
3770 void SWPointer::Tracer::scaled_iv_plus_offset_5(Node* n) {
3771   if(_slp->is_trace_alignment()) {
3772     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
3773     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
3774     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
3775   }
3776 }
3777 
3778 void SWPointer::Tracer::scaled_iv_plus_offset_6(Node* n) {
3779   if(_slp->is_trace_alignment()) {
3780     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
3781     print_depth(); tty->print("  \\  %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
3782     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
3783   }
3784 }
3785 
3786 void SWPointer::Tracer::scaled_iv_plus_offset_7(Node* n) {
3787   if(_slp->is_trace_alignment()) {
3788     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
3789     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
3790     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
3791   }
3792 }
3793 
3794 void SWPointer::Tracer::scaled_iv_plus_offset_8(Node* n) {
3795   if(_slp->is_trace_alignment()) {
3796     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: FAILED", n->_idx);
3797   }
3798 }
3799 
3800 void SWPointer::Tracer::scaled_iv_1(Node* n) {
3801   if(_slp->is_trace_alignment()) {
3802     print_depth(); tty->print(" %d SWPointer::scaled_iv: testing node: ", n->_idx); n->dump();
3803   }
3804 }
3805 
3806 void SWPointer::Tracer::scaled_iv_2(Node* n, int scale) {
3807   if(_slp->is_trace_alignment()) {
3808     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED since another _scale has been detected before", n->_idx);
3809     print_depth(); tty->print_cr("  \\ SWPointer::scaled_iv: _scale (%d) != 0", scale);
3810   }
3811 }
3812 
3813 void SWPointer::Tracer::scaled_iv_3(Node* n, int scale) {
3814   if(_slp->is_trace_alignment()) {
3815     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: is iv, setting _scale = %d", n->_idx, scale);
3816   }
3817 }
3818 
3819 void SWPointer::Tracer::scaled_iv_4(Node* n, int scale) {
3820   if(_slp->is_trace_alignment()) {
3821     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
3822     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
3823     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3824   }
3825 }
3826 
3827 void SWPointer::Tracer::scaled_iv_5(Node* n, int scale) {
3828   if(_slp->is_trace_alignment()) {
3829     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
3830     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is iv: ", n->in(2)->_idx); n->in(2)->dump();
3831     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3832   }
3833 }
3834 
3835 void SWPointer::Tracer::scaled_iv_6(Node* n, int scale) {
3836   if(_slp->is_trace_alignment()) {
3837     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftI PASSED, setting _scale = %d", n->_idx, scale);
3838     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
3839     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3840   }
3841 }
3842 
3843 void SWPointer::Tracer::scaled_iv_7(Node* n) {
3844   if(_slp->is_trace_alignment()) {
3845     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_ConvI2L PASSED", n->_idx);
3846     print_depth(); tty->print_cr("  \\ SWPointer::scaled_iv: in(1) %d is scaled_iv_plus_offset: ", n->in(1)->_idx);
3847     inc_depth(); inc_depth();
3848     print_depth(); n->in(1)->dump();
3849     dec_depth(); dec_depth();
3850   }
3851 }
3852 
3853 void SWPointer::Tracer::scaled_iv_8(Node* n, SWPointer* tmp) {
3854   if(_slp->is_trace_alignment()) {
3855     print_depth(); tty->print(" %d SWPointer::scaled_iv: Op_LShiftL, creating tmp SWPointer: ", n->_idx); tmp->print();
3856   }
3857 }
3858 
3859 void SWPointer::Tracer::scaled_iv_9(Node* n, int scale, int _offset, int mult) {
3860   if(_slp->is_trace_alignment()) {
3861     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftL PASSED, setting _scale = %d, _offset = %d", n->_idx, scale, _offset);
3862     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",
3863     n->in(1)->_idx, n->in(2)->_idx, mult, scale, _offset);
3864     inc_depth(); inc_depth();
3865     print_depth(); n->in(1)->dump();
3866     print_depth(); n->in(2)->dump();
3867     dec_depth(); dec_depth();
3868   }
3869 }
3870 
3871 void SWPointer::Tracer::scaled_iv_10(Node* n) {
3872   if(_slp->is_trace_alignment()) {
3873     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED", n->_idx);
3874   }
3875 }
3876 
3877 void SWPointer::Tracer::offset_plus_k_1(Node* n) {
3878   if(_slp->is_trace_alignment()) {
3879     print_depth(); tty->print(" %d SWPointer::offset_plus_k: testing node: ", n->_idx); n->dump();
3880   }
3881 }
3882 
3883 void SWPointer::Tracer::offset_plus_k_2(Node* n, int _offset) {
3884   if(_slp->is_trace_alignment()) {
3885     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConI PASSED, setting _offset = %d", n->_idx, _offset);
3886   }
3887 }
3888 
3889 void SWPointer::Tracer::offset_plus_k_3(Node* n, int _offset) {
3890   if(_slp->is_trace_alignment()) {
3891     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConL PASSED, setting _offset = %d", n->_idx, _offset);
3892   }
3893 }
3894 
3895 void SWPointer::Tracer::offset_plus_k_4(Node* n) {
3896   if(_slp->is_trace_alignment()) {
3897     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
3898     print_depth(); tty->print_cr("  \\ " JLONG_FORMAT " SWPointer::offset_plus_k: Op_ConL FAILED, k is too big", n->get_long());
3899   }
3900 }
3901 
3902 void SWPointer::Tracer::offset_plus_k_5(Node* n, Node* _invar) {
3903   if(_slp->is_trace_alignment()) {
3904     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED since another invariant has been detected before", n->_idx);
3905     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: _invar != NULL: ", _invar->_idx); _invar->dump();
3906   }
3907 }
3908 
3909 void SWPointer::Tracer::offset_plus_k_6(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3910   if(_slp->is_trace_alignment()) {
3911     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
3912     n->_idx, _negate_invar, _invar->_idx, _offset);
3913     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3914     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
3915   }
3916 }
3917 
3918 void SWPointer::Tracer::offset_plus_k_7(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3919   if(_slp->is_trace_alignment()) {
3920     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
3921     n->_idx, _negate_invar, _invar->_idx, _offset);
3922     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3923     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
3924   }
3925 }
3926 
3927 void SWPointer::Tracer::offset_plus_k_8(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3928   if(_slp->is_trace_alignment()) {
3929     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_SubI is PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
3930     n->_idx, _negate_invar, _invar->_idx, _offset);
3931     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3932     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
3933   }
3934 }
3935 
3936 void SWPointer::Tracer::offset_plus_k_9(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3937   if(_slp->is_trace_alignment()) {
3938     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);
3939     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3940     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
3941   }
3942 }
3943 
3944 void SWPointer::Tracer::offset_plus_k_10(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3945   if(_slp->is_trace_alignment()) {
3946     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d", n->_idx, _negate_invar, _invar->_idx, _offset);
3947     print_depth(); tty->print_cr("  \\ %d SWPointer::offset_plus_k: is invariant", n->_idx);
3948   }
3949 }
3950 
3951 void SWPointer::Tracer::offset_plus_k_11(Node* n) {
3952   if(_slp->is_trace_alignment()) {
3953     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
3954   }
3955 }
3956 
3957 #endif
3958 // ========================= OrderedPair =====================
3959 
3960 const OrderedPair OrderedPair::initial;
3961 
3962 // ========================= SWNodeInfo =====================
3963 
3964 const SWNodeInfo SWNodeInfo::initial;
3965 
3966 
3967 // ============================ DepGraph ===========================
3968 
3969 //------------------------------make_node---------------------------
3970 // Make a new dependence graph node for an ideal node.
3971 DepMem* DepGraph::make_node(Node* node) {
3972   DepMem* m = new (_arena) DepMem(node);
3973   if (node != NULL) {
3974     assert(_map.at_grow(node->_idx) == NULL, "one init only");
3975     _map.at_put_grow(node->_idx, m);
3976   }
3977   return m;
3978 }
3979 
3980 //------------------------------make_edge---------------------------
3981 // Make a new dependence graph edge from dpred -> dsucc
3982 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
3983   DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
3984   dpred->set_out_head(e);
3985   dsucc->set_in_head(e);
3986   return e;
3987 }
3988 
3989 // ========================== DepMem ========================
3990 
3991 //------------------------------in_cnt---------------------------
3992 int DepMem::in_cnt() {
3993   int ct = 0;
3994   for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
3995   return ct;
3996 }
3997 
3998 //------------------------------out_cnt---------------------------
3999 int DepMem::out_cnt() {
4000   int ct = 0;
4001   for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
4002   return ct;
4003 }
4004 
4005 //------------------------------print-----------------------------
4006 void DepMem::print() {
4007 #ifndef PRODUCT
4008   tty->print("  DepNode %d (", _node->_idx);
4009   for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
4010     Node* pred = p->pred()->node();
4011     tty->print(" %d", pred != NULL ? pred->_idx : 0);
4012   }
4013   tty->print(") [");
4014   for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
4015     Node* succ = s->succ()->node();
4016     tty->print(" %d", succ != NULL ? succ->_idx : 0);
4017   }
4018   tty->print_cr(" ]");
4019 #endif
4020 }
4021 
4022 // =========================== DepEdge =========================
4023 
4024 //------------------------------DepPreds---------------------------
4025 void DepEdge::print() {
4026 #ifndef PRODUCT
4027   tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
4028 #endif
4029 }
4030 
4031 // =========================== DepPreds =========================
4032 // Iterator over predecessor edges in the dependence graph.
4033 
4034 //------------------------------DepPreds---------------------------
4035 DepPreds::DepPreds(Node* n, DepGraph& dg) {
4036   _n = n;
4037   _done = false;
4038   if (_n->is_Store() || _n->is_Load()) {
4039     _next_idx = MemNode::Address;
4040     _end_idx  = n->req();
4041     _dep_next = dg.dep(_n)->in_head();
4042   } else if (_n->is_Mem()) {
4043     _next_idx = 0;
4044     _end_idx  = 0;
4045     _dep_next = dg.dep(_n)->in_head();
4046   } else {
4047     _next_idx = 1;
4048     _end_idx  = _n->req();
4049     _dep_next = NULL;
4050   }
4051   next();
4052 }
4053 
4054 //------------------------------next---------------------------
4055 void DepPreds::next() {
4056   if (_dep_next != NULL) {
4057     _current  = _dep_next->pred()->node();
4058     _dep_next = _dep_next->next_in();
4059   } else if (_next_idx < _end_idx) {
4060     _current  = _n->in(_next_idx++);
4061   } else {
4062     _done = true;
4063   }
4064 }
4065 
4066 // =========================== DepSuccs =========================
4067 // Iterator over successor edges in the dependence graph.
4068 
4069 //------------------------------DepSuccs---------------------------
4070 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
4071   _n = n;
4072   _done = false;
4073   if (_n->is_Load()) {
4074     _next_idx = 0;
4075     _end_idx  = _n->outcnt();
4076     _dep_next = dg.dep(_n)->out_head();
4077   } else if (_n->is_Mem() || (_n->is_Phi() && _n->bottom_type() == Type::MEMORY)) {
4078     _next_idx = 0;
4079     _end_idx  = 0;
4080     _dep_next = dg.dep(_n)->out_head();
4081   } else {
4082     _next_idx = 0;
4083     _end_idx  = _n->outcnt();
4084     _dep_next = NULL;
4085   }
4086   next();
4087 }
4088 
4089 //-------------------------------next---------------------------
4090 void DepSuccs::next() {
4091   if (_dep_next != NULL) {
4092     _current  = _dep_next->succ()->node();
4093     _dep_next = _dep_next->next_out();
4094   } else if (_next_idx < _end_idx) {
4095     _current  = _n->raw_out(_next_idx++);
4096   } else {
4097     _done = true;
4098   }
4099 }
4100 
4101 //
4102 // --------------------------------- vectorization/simd -----------------------------------
4103 //
4104 bool SuperWord::same_origin_idx(Node* a, Node* b) const {
4105   return a != NULL && b != NULL && _clone_map.same_idx(a->_idx, b->_idx);
4106 }
4107 bool SuperWord::same_generation(Node* a, Node* b) const {
4108   return a != NULL && b != NULL && _clone_map.same_gen(a->_idx, b->_idx);
4109 }
4110 
4111 Node*  SuperWord::find_phi_for_mem_dep(LoadNode* ld) {
4112   assert(in_bb(ld), "must be in block");
4113   if (_clone_map.gen(ld->_idx) == _ii_first) {
4114 #ifndef PRODUCT
4115     if (_vector_loop_debug) {
4116       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(ld->_idx)=%d",
4117         _clone_map.gen(ld->_idx));
4118     }
4119 #endif
4120     return NULL; //we think that any ld in the first gen being vectorizable
4121   }
4122 
4123   Node* mem = ld->in(MemNode::Memory);
4124   if (mem->outcnt() <= 1) {
4125     // we don't want to remove the only edge from mem node to load
4126 #ifndef PRODUCT
4127     if (_vector_loop_debug) {
4128       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",
4129         mem->_idx, ld->_idx);
4130       ld->dump();
4131       mem->dump();
4132     }
4133 #endif
4134     return NULL;
4135   }
4136   if (!in_bb(mem) || same_generation(mem, ld)) {
4137 #ifndef PRODUCT
4138     if (_vector_loop_debug) {
4139       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(mem->_idx)=%d",
4140         _clone_map.gen(mem->_idx));
4141     }
4142 #endif
4143     return NULL; // does not depend on loop volatile node or depends on the same generation
4144   }
4145 
4146   //otherwise first node should depend on mem-phi
4147   Node* first = first_node(ld);
4148   assert(first->is_Load(), "must be Load");
4149   Node* phi = first->as_Load()->in(MemNode::Memory);
4150   if (!phi->is_Phi() || phi->bottom_type() != Type::MEMORY) {
4151 #ifndef PRODUCT
4152     if (_vector_loop_debug) {
4153       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");
4154       ld->dump();
4155       first->dump();
4156     }
4157 #endif
4158     return NULL;
4159   }
4160 
4161   Node* tail = 0;
4162   for (int m = 0; m < _mem_slice_head.length(); m++) {
4163     if (_mem_slice_head.at(m) == phi) {
4164       tail = _mem_slice_tail.at(m);
4165     }
4166   }
4167   if (tail == 0) { //test that found phi is in the list  _mem_slice_head
4168 #ifndef PRODUCT
4169     if (_vector_loop_debug) {
4170       tty->print_cr("SuperWord::find_phi_for_mem_dep load %d is not vectorizable node, its phi %d is not _mem_slice_head",
4171         ld->_idx, phi->_idx);
4172       ld->dump();
4173       phi->dump();
4174     }
4175 #endif
4176     return NULL;
4177   }
4178 
4179   // now all conditions are met
4180   return phi;
4181 }
4182 
4183 Node* SuperWord::first_node(Node* nd) {
4184   for (int ii = 0; ii < _iteration_first.length(); ii++) {
4185     Node* nnn = _iteration_first.at(ii);
4186     if (same_origin_idx(nnn, nd)) {
4187 #ifndef PRODUCT
4188       if (_vector_loop_debug) {
4189         tty->print_cr("SuperWord::first_node: %d is the first iteration node for %d (_clone_map.idx(nnn->_idx) = %d)",
4190           nnn->_idx, nd->_idx, _clone_map.idx(nnn->_idx));
4191       }
4192 #endif
4193       return nnn;
4194     }
4195   }
4196 
4197 #ifndef PRODUCT
4198   if (_vector_loop_debug) {
4199     tty->print_cr("SuperWord::first_node: did not find first iteration node for %d (_clone_map.idx(nd->_idx)=%d)",
4200       nd->_idx, _clone_map.idx(nd->_idx));
4201   }
4202 #endif
4203   return 0;
4204 }
4205 
4206 Node* SuperWord::last_node(Node* nd) {
4207   for (int ii = 0; ii < _iteration_last.length(); ii++) {
4208     Node* nnn = _iteration_last.at(ii);
4209     if (same_origin_idx(nnn, nd)) {
4210 #ifndef PRODUCT
4211       if (_vector_loop_debug) {
4212         tty->print_cr("SuperWord::last_node _clone_map.idx(nnn->_idx)=%d, _clone_map.idx(nd->_idx)=%d",
4213           _clone_map.idx(nnn->_idx), _clone_map.idx(nd->_idx));
4214       }
4215 #endif
4216       return nnn;
4217     }
4218   }
4219   return 0;
4220 }
4221 
4222 int SuperWord::mark_generations() {
4223   Node *ii_err = NULL, *tail_err = NULL;
4224   for (int i = 0; i < _mem_slice_head.length(); i++) {
4225     Node* phi  = _mem_slice_head.at(i);
4226     assert(phi->is_Phi(), "must be phi");
4227 
4228     Node* tail = _mem_slice_tail.at(i);
4229     if (_ii_last == -1) {
4230       tail_err = tail;
4231       _ii_last = _clone_map.gen(tail->_idx);
4232     }
4233     else if (_ii_last != _clone_map.gen(tail->_idx)) {
4234 #ifndef PRODUCT
4235       if (TraceSuperWord && Verbose) {
4236         tty->print_cr("SuperWord::mark_generations _ii_last error - found different generations in two tail nodes ");
4237         tail->dump();
4238         tail_err->dump();
4239       }
4240 #endif
4241       return -1;
4242     }
4243 
4244     // find first iteration in the loop
4245     for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
4246       Node* ii = phi->fast_out(i);
4247       if (in_bb(ii) && ii->is_Store()) { // we speculate that normally Stores of one and one only generation have deps from mem phi
4248         if (_ii_first == -1) {
4249           ii_err = ii;
4250           _ii_first = _clone_map.gen(ii->_idx);
4251         } else if (_ii_first != _clone_map.gen(ii->_idx)) {
4252 #ifndef PRODUCT
4253           if (TraceSuperWord && Verbose) {
4254             tty->print_cr("SuperWord::mark_generations: _ii_first was found before and not equal to one in this node (%d)", _ii_first);
4255             ii->dump();
4256             if (ii_err!= 0) {
4257               ii_err->dump();
4258             }
4259           }
4260 #endif
4261           return -1; // this phi has Stores from different generations of unroll and cannot be simd/vectorized
4262         }
4263       }
4264     }//for (DUIterator_Fast imax,
4265   }//for (int i...
4266 
4267   if (_ii_first == -1 || _ii_last == -1) {
4268     if (TraceSuperWord && Verbose) {
4269       tty->print_cr("SuperWord::mark_generations unknown error, something vent wrong");
4270     }
4271     return -1; // something vent wrong
4272   }
4273   // collect nodes in the first and last generations
4274   assert(_iteration_first.length() == 0, "_iteration_first must be empty");
4275   assert(_iteration_last.length() == 0, "_iteration_last must be empty");
4276   for (int j = 0; j < _block.length(); j++) {
4277     Node* n = _block.at(j);
4278     node_idx_t gen = _clone_map.gen(n->_idx);
4279     if ((signed)gen == _ii_first) {
4280       _iteration_first.push(n);
4281     } else if ((signed)gen == _ii_last) {
4282       _iteration_last.push(n);
4283     }
4284   }
4285 
4286   // building order of iterations
4287   if (_ii_order.length() == 0 && ii_err != 0) {
4288     assert(in_bb(ii_err) && ii_err->is_Store(), "should be Store in bb");
4289     Node* nd = ii_err;
4290     while(_clone_map.gen(nd->_idx) != _ii_last) {
4291       _ii_order.push(_clone_map.gen(nd->_idx));
4292       bool found = false;
4293       for (DUIterator_Fast imax, i = nd->fast_outs(imax); i < imax; i++) {
4294         Node* use = nd->fast_out(i);
4295         if (same_origin_idx(use, nd) && use->as_Store()->in(MemNode::Memory) == nd) {
4296           found = true;
4297           nd = use;
4298           break;
4299         }
4300       }//for
4301 
4302       if (found == false) {
4303         if (TraceSuperWord && Verbose) {
4304           tty->print_cr("SuperWord::mark_generations: Cannot build order of iterations - no dependent Store for %d", nd->_idx);
4305         }
4306         _ii_order.clear();
4307         return -1;
4308       }
4309     } //while
4310     _ii_order.push(_clone_map.gen(nd->_idx));
4311   }
4312 
4313 #ifndef PRODUCT
4314   if (_vector_loop_debug) {
4315     tty->print_cr("SuperWord::mark_generations");
4316     tty->print_cr("First generation (%d) nodes:", _ii_first);
4317     for (int ii = 0; ii < _iteration_first.length(); ii++)  _iteration_first.at(ii)->dump();
4318     tty->print_cr("Last generation (%d) nodes:", _ii_last);
4319     for (int ii = 0; ii < _iteration_last.length(); ii++)  _iteration_last.at(ii)->dump();
4320     tty->print_cr(" ");
4321 
4322     tty->print("SuperWord::List of generations: ");
4323     for (int jj = 0; jj < _ii_order.length(); ++jj) {
4324       tty->print("%d:%d ", jj, _ii_order.at(jj));
4325     }
4326     tty->print_cr(" ");
4327   }
4328 #endif
4329 
4330   return _ii_first;
4331 }
4332 
4333 bool SuperWord::fix_commutative_inputs(Node* gold, Node* fix) {
4334   assert(gold->is_Add() && fix->is_Add() || gold->is_Mul() && fix->is_Mul(), "should be only Add or Mul nodes");
4335   assert(same_origin_idx(gold, fix), "should be clones of the same node");
4336   Node* gin1 = gold->in(1);
4337   Node* gin2 = gold->in(2);
4338   Node* fin1 = fix->in(1);
4339   Node* fin2 = fix->in(2);
4340   bool swapped = false;
4341 
4342   if (in_bb(gin1) && in_bb(gin2) && in_bb(fin1) && in_bb(fin1)) {
4343     if (same_origin_idx(gin1, fin1) &&
4344         same_origin_idx(gin2, fin2)) {
4345       return true; // nothing to fix
4346     }
4347     if (same_origin_idx(gin1, fin2) &&
4348         same_origin_idx(gin2, fin1)) {
4349       fix->swap_edges(1, 2);
4350       swapped = true;
4351     }
4352   }
4353   // at least one input comes from outside of bb
4354   if (gin1->_idx == fin1->_idx)  {
4355     return true; // nothing to fix
4356   }
4357   if (!swapped && (gin1->_idx == fin2->_idx || gin2->_idx == fin1->_idx))  { //swapping is expensive, check condition first
4358     fix->swap_edges(1, 2);
4359     swapped = true;
4360   }
4361 
4362   if (swapped) {
4363 #ifndef PRODUCT
4364     if (_vector_loop_debug) {
4365       tty->print_cr("SuperWord::fix_commutative_inputs: fixed node %d", fix->_idx);
4366     }
4367 #endif
4368     return true;
4369   }
4370 
4371   if (TraceSuperWord && Verbose) {
4372     tty->print_cr("SuperWord::fix_commutative_inputs: cannot fix node %d", fix->_idx);
4373   }
4374 
4375   return false;
4376 }
4377 
4378 bool SuperWord::pack_parallel() {
4379 #ifndef PRODUCT
4380   if (_vector_loop_debug) {
4381     tty->print_cr("SuperWord::pack_parallel: START");
4382   }
4383 #endif
4384 
4385   _packset.clear();
4386 
4387   for (int ii = 0; ii < _iteration_first.length(); ii++) {
4388     Node* nd = _iteration_first.at(ii);
4389     if (in_bb(nd) && (nd->is_Load() || nd->is_Store() || nd->is_Add() || nd->is_Mul())) {
4390       Node_List* pk = new Node_List();
4391       pk->push(nd);
4392       for (int gen = 1; gen < _ii_order.length(); ++gen) {
4393         for (int kk = 0; kk < _block.length(); kk++) {
4394           Node* clone = _block.at(kk);
4395           if (same_origin_idx(clone, nd) &&
4396               _clone_map.gen(clone->_idx) == _ii_order.at(gen)) {
4397             if (nd->is_Add() || nd->is_Mul()) {
4398               fix_commutative_inputs(nd, clone);
4399             }
4400             pk->push(clone);
4401             if (pk->size() == 4) {
4402               _packset.append(pk);
4403 #ifndef PRODUCT
4404               if (_vector_loop_debug) {
4405                 tty->print_cr("SuperWord::pack_parallel: added pack ");
4406                 pk->dump();
4407               }
4408 #endif
4409               if (_clone_map.gen(clone->_idx) != _ii_last) {
4410                 pk = new Node_List();
4411               }
4412             }
4413             break;
4414           }
4415         }
4416       }//for
4417     }//if
4418   }//for
4419 
4420 #ifndef PRODUCT
4421   if (_vector_loop_debug) {
4422     tty->print_cr("SuperWord::pack_parallel: END");
4423   }
4424 #endif
4425 
4426   return true;
4427 }
4428 
4429 bool SuperWord::hoist_loads_in_graph() {
4430   GrowableArray<Node*> loads;
4431 
4432 #ifndef PRODUCT
4433   if (_vector_loop_debug) {
4434     tty->print_cr("SuperWord::hoist_loads_in_graph: total number _mem_slice_head.length() = %d", _mem_slice_head.length());
4435   }
4436 #endif
4437 
4438   for (int i = 0; i < _mem_slice_head.length(); i++) {
4439     Node* n = _mem_slice_head.at(i);
4440     if ( !in_bb(n) || !n->is_Phi() || n->bottom_type() != Type::MEMORY) {
4441       if (TraceSuperWord && Verbose) {
4442         tty->print_cr("SuperWord::hoist_loads_in_graph: skipping unexpected node n=%d", n->_idx);
4443       }
4444       continue;
4445     }
4446 
4447 #ifndef PRODUCT
4448     if (_vector_loop_debug) {
4449       tty->print_cr("SuperWord::hoist_loads_in_graph: processing phi %d  = _mem_slice_head.at(%d);", n->_idx, i);
4450     }
4451 #endif
4452 
4453     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4454       Node* ld = n->fast_out(i);
4455       if (ld->is_Load() && ld->as_Load()->in(MemNode::Memory) == n && in_bb(ld)) {
4456         for (int i = 0; i < _block.length(); i++) {
4457           Node* ld2 = _block.at(i);
4458           if (ld2->is_Load() && same_origin_idx(ld, ld2) &&
4459               !same_generation(ld, ld2)) { // <= do not collect the first generation ld
4460 #ifndef PRODUCT
4461             if (_vector_loop_debug) {
4462               tty->print_cr("SuperWord::hoist_loads_in_graph: will try to hoist load ld2->_idx=%d, cloned from %d (ld->_idx=%d)",
4463                 ld2->_idx, _clone_map.idx(ld->_idx), ld->_idx);
4464             }
4465 #endif
4466             // could not do on-the-fly, since iterator is immutable
4467             loads.push(ld2);
4468           }
4469         }// for
4470       }//if
4471     }//for (DUIterator_Fast imax,
4472   }//for (int i = 0; i
4473 
4474   for (int i = 0; i < loads.length(); i++) {
4475     LoadNode* ld = loads.at(i)->as_Load();
4476     Node* phi = find_phi_for_mem_dep(ld);
4477     if (phi != NULL) {
4478 #ifndef PRODUCT
4479       if (_vector_loop_debug) {
4480         tty->print_cr("SuperWord::hoist_loads_in_graph replacing MemNode::Memory(%d) edge in %d with one from %d",
4481           MemNode::Memory, ld->_idx, phi->_idx);
4482       }
4483 #endif
4484       _igvn.replace_input_of(ld, MemNode::Memory, phi);
4485     }
4486   }//for
4487 
4488   restart(); // invalidate all basic structures, since we rebuilt the graph
4489 
4490   if (TraceSuperWord && Verbose) {
4491     tty->print_cr("\nSuperWord::hoist_loads_in_graph() the graph was rebuilt, all structures invalidated and need rebuild");
4492   }
4493 
4494   return true;
4495 }
4496