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     for (uint i = 0; i < lpt()->_body.size(); i++) {
 303       if (ignored_loop_nodes[i] != -1) continue;
 304 
 305       BasicType bt;
 306       Node* n = lpt()->_body.at(i);
 307       if (n->is_Mem()) {
 308         bt = n->as_Mem()->memory_type();
 309       } else {
 310         bt = n->bottom_type()->basic_type();
 311       }
 312 
 313       if (post_loop_allowed) {
 314         if (!small_basic_type) {
 315           switch (bt) {
 316           case T_CHAR:
 317           case T_BYTE:
 318           case T_SHORT:
 319             small_basic_type = true;
 320             break;
 321 
 322           case T_LONG:
 323             // TODO: Remove when support completed for mask context with LONG.
 324             //       Support needs to be augmented for logical qword operations, currently we map to dword
 325             //       buckets for vectors on logicals as these were legacy.
 326             small_basic_type = true;
 327             break;
 328 
 329           default:
 330             break;
 331           }
 332         }
 333       }
 334 
 335       if (is_java_primitive(bt) == false) continue;
 336 
 337       int cur_max_vector = Matcher::max_vector_size(bt);
 338 
 339       // If a max vector exists which is not larger than _local_loop_unroll_factor
 340       // stop looking, we already have the max vector to map to.
 341       if (cur_max_vector < local_loop_unroll_factor) {
 342         is_slp = false;
 343         if (TraceSuperWordLoopUnrollAnalysis) {
 344           tty->print_cr("slp analysis fails: unroll limit greater than max vector\n");
 345         }
 346         break;
 347       }
 348 
 349       // Map the maximal common vector
 350       if (VectorNode::implemented(n->Opcode(), cur_max_vector, bt)) {
 351         if (cur_max_vector < max_vector) {
 352           max_vector = cur_max_vector;
 353         }
 354 
 355         // We only process post loops on predicated targets where we want to
 356         // mask map the loop to a single iteration
 357         if (post_loop_allowed) {
 358           _post_block.at_put_grow(rpo_idx++, n);
 359         }
 360       }
 361     }
 362     if (is_slp) {
 363       local_loop_unroll_factor = max_vector;
 364       cl->mark_passed_slp();
 365     }
 366     cl->mark_was_slp();
 367     if (cl->is_main_loop()) {
 368       cl->set_slp_max_unroll(local_loop_unroll_factor);
 369     } else if (post_loop_allowed) {
 370       if (!small_basic_type) {
 371         // avoid replication context for small basic types in programmable masked loops
 372         cl->set_slp_max_unroll(local_loop_unroll_factor);
 373       }
 374     }
 375   }
 376 }
 377 
 378 //------------------------------SLP_extract---------------------------
 379 // Extract the superword level parallelism
 380 //
 381 // 1) A reverse post-order of nodes in the block is constructed.  By scanning
 382 //    this list from first to last, all definitions are visited before their uses.
 383 //
 384 // 2) A point-to-point dependence graph is constructed between memory references.
 385 //    This simplies the upcoming "independence" checker.
 386 //
 387 // 3) The maximum depth in the node graph from the beginning of the block
 388 //    to each node is computed.  This is used to prune the graph search
 389 //    in the independence checker.
 390 //
 391 // 4) For integer types, the necessary bit width is propagated backwards
 392 //    from stores to allow packed operations on byte, char, and short
 393 //    integers.  This reverses the promotion to type "int" that javac
 394 //    did for operations like: char c1,c2,c3;  c1 = c2 + c3.
 395 //
 396 // 5) One of the memory references is picked to be an aligned vector reference.
 397 //    The pre-loop trip count is adjusted to align this reference in the
 398 //    unrolled body.
 399 //
 400 // 6) The initial set of pack pairs is seeded with memory references.
 401 //
 402 // 7) The set of pack pairs is extended by following use->def and def->use links.
 403 //
 404 // 8) The pairs are combined into vector sized packs.
 405 //
 406 // 9) Reorder the memory slices to co-locate members of the memory packs.
 407 //
 408 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
 409 //    inserting scalar promotion, vector creation from multiple scalars, and
 410 //    extraction of scalar values from vectors.
 411 //
 412 void SuperWord::SLP_extract() {
 413 
 414 #ifndef PRODUCT
 415   if (_do_vector_loop && TraceSuperWord) {
 416     tty->print("SuperWord::SLP_extract\n");
 417     tty->print("input loop\n");
 418     _lpt->dump_head();
 419     _lpt->dump();
 420     for (uint i = 0; i < _lpt->_body.size(); i++) {
 421       _lpt->_body.at(i)->dump();
 422     }
 423   }
 424 #endif
 425   // Ready the block
 426   if (!construct_bb()) {
 427     return; // Exit if no interesting nodes or complex graph.
 428   }
 429 
 430   // build    _dg, _disjoint_ptrs
 431   dependence_graph();
 432 
 433   // compute function depth(Node*)
 434   compute_max_depth();
 435 
 436   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
 437   bool post_loop_allowed = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
 438   if (cl->is_main_loop()) {
 439     if (_do_vector_loop) {
 440       if (mark_generations() != -1) {
 441         hoist_loads_in_graph(); // this only rebuild the graph; all basic structs need rebuild explicitly
 442 
 443         if (!construct_bb()) {
 444           return; // Exit if no interesting nodes or complex graph.
 445         }
 446         dependence_graph();
 447         compute_max_depth();
 448       }
 449 
 450 #ifndef PRODUCT
 451       if (TraceSuperWord) {
 452         tty->print_cr("\nSuperWord::_do_vector_loop: graph after hoist_loads_in_graph");
 453         _lpt->dump_head();
 454         for (int j = 0; j < _block.length(); j++) {
 455           Node* n = _block.at(j);
 456           int d = depth(n);
 457           for (int i = 0; i < d; i++) tty->print("%s", "  ");
 458           tty->print("%d :", d);
 459           n->dump();
 460         }
 461       }
 462 #endif
 463     }
 464 
 465     compute_vector_element_type();
 466 
 467     // Attempt vectorization
 468 
 469     find_adjacent_refs();
 470 
 471     extend_packlist();
 472 
 473     if (_do_vector_loop) {
 474       if (_packset.length() == 0) {
 475         if (TraceSuperWord) {
 476           tty->print_cr("\nSuperWord::_do_vector_loop DFA could not build packset, now trying to build anyway");
 477         }
 478         pack_parallel();
 479       }
 480     }
 481 
 482     combine_packs();
 483 
 484     construct_my_pack_map();
 485 
 486     if (_do_vector_loop) {
 487       merge_packs_to_cmovd();
 488     }
 489 
 490     filter_packs();
 491 
 492     schedule();
 493   } else if (post_loop_allowed) {
 494     int saved_mapped_unroll_factor = cl->slp_max_unroll();
 495     if (saved_mapped_unroll_factor) {
 496       int vector_mapped_unroll_factor = saved_mapped_unroll_factor;
 497 
 498       // now reset the slp_unroll_factor so that we can check the analysis mapped
 499       // what the vector loop was mapped to
 500       cl->set_slp_max_unroll(0);
 501 
 502       // do the analysis on the post loop
 503       unrolling_analysis(vector_mapped_unroll_factor);
 504 
 505       // if our analyzed loop is a canonical fit, start processing it
 506       if (vector_mapped_unroll_factor == saved_mapped_unroll_factor) {
 507         // now add the vector nodes to packsets
 508         for (int i = 0; i < _post_block.length(); i++) {
 509           Node* n = _post_block.at(i);
 510           Node_List* singleton = new Node_List();
 511           singleton->push(n);
 512           _packset.append(singleton);
 513           set_my_pack(n, singleton);
 514         }
 515 
 516         // map base types for vector usage
 517         compute_vector_element_type();
 518       } else {
 519         return;
 520       }
 521     } else {
 522       // for some reason we could not map the slp analysis state of the vectorized loop
 523       return;
 524     }
 525   }
 526 
 527   output();
 528 }
 529 
 530 //------------------------------find_adjacent_refs---------------------------
 531 // Find the adjacent memory references and create pack pairs for them.
 532 // This is the initial set of packs that will then be extended by
 533 // following use->def and def->use links.  The align positions are
 534 // assigned relative to the reference "align_to_ref"
 535 void SuperWord::find_adjacent_refs() {
 536   // Get list of memory operations
 537   Node_List memops;
 538   for (int i = 0; i < _block.length(); i++) {
 539     Node* n = _block.at(i);
 540     if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
 541         is_java_primitive(n->as_Mem()->memory_type())) {
 542       int align = memory_alignment(n->as_Mem(), 0);
 543       if (align != bottom_align) {
 544         memops.push(n);
 545       }
 546     }
 547   }
 548 
 549   Node_List align_to_refs;
 550   int best_iv_adjustment = 0;
 551   MemNode* best_align_to_mem_ref = NULL;
 552 
 553   while (memops.size() != 0) {
 554     // Find a memory reference to align to.
 555     MemNode* mem_ref = find_align_to_ref(memops);
 556     if (mem_ref == NULL) break;
 557     align_to_refs.push(mem_ref);
 558     int iv_adjustment = get_iv_adjustment(mem_ref);
 559 
 560     if (best_align_to_mem_ref == NULL) {
 561       // Set memory reference which is the best from all memory operations
 562       // to be used for alignment. The pre-loop trip count is modified to align
 563       // this reference to a vector-aligned address.
 564       best_align_to_mem_ref = mem_ref;
 565       best_iv_adjustment = iv_adjustment;
 566       NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
 567     }
 568 
 569     SWPointer align_to_ref_p(mem_ref, this, NULL, false);
 570     // Set alignment relative to "align_to_ref" for all related memory operations.
 571     for (int i = memops.size() - 1; i >= 0; i--) {
 572       MemNode* s = memops.at(i)->as_Mem();
 573       if (isomorphic(s, mem_ref) &&
 574            (!_do_vector_loop || same_origin_idx(s, mem_ref))) {
 575         SWPointer p2(s, this, NULL, false);
 576         if (p2.comparable(align_to_ref_p)) {
 577           int align = memory_alignment(s, iv_adjustment);
 578           set_alignment(s, align);
 579         }
 580       }
 581     }
 582 
 583     // Create initial pack pairs of memory operations for which
 584     // alignment is set and vectors will be aligned.
 585     bool create_pack = true;
 586     if (memory_alignment(mem_ref, best_iv_adjustment) == 0 || _do_vector_loop) {
 587       if (!Matcher::misaligned_vectors_ok()) {
 588         int vw = vector_width(mem_ref);
 589         int vw_best = vector_width(best_align_to_mem_ref);
 590         if (vw > vw_best) {
 591           // Do not vectorize a memory access with more elements per vector
 592           // if unaligned memory access is not allowed because number of
 593           // iterations in pre-loop will be not enough to align it.
 594           create_pack = false;
 595         } else {
 596           SWPointer p2(best_align_to_mem_ref, this, NULL, false);
 597           if (align_to_ref_p.invar() != p2.invar()) {
 598             // Do not vectorize memory accesses with different invariants
 599             // if unaligned memory accesses are not allowed.
 600             create_pack = false;
 601           }
 602         }
 603       }
 604     } else {
 605       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 606         // Can't allow vectorization of unaligned memory accesses with the
 607         // same type since it could be overlapped accesses to the same array.
 608         create_pack = false;
 609       } else {
 610         // Allow independent (different type) unaligned memory operations
 611         // if HW supports them.
 612         if (!Matcher::misaligned_vectors_ok()) {
 613           create_pack = false;
 614         } else {
 615           // Check if packs of the same memory type but
 616           // with a different alignment were created before.
 617           for (uint i = 0; i < align_to_refs.size(); i++) {
 618             MemNode* mr = align_to_refs.at(i)->as_Mem();
 619             if (same_velt_type(mr, mem_ref) &&
 620                 memory_alignment(mr, iv_adjustment) != 0)
 621               create_pack = false;
 622           }
 623         }
 624       }
 625     }
 626     if (create_pack) {
 627       for (uint i = 0; i < memops.size(); i++) {
 628         Node* s1 = memops.at(i);
 629         int align = alignment(s1);
 630         if (align == top_align) continue;
 631         for (uint j = 0; j < memops.size(); j++) {
 632           Node* s2 = memops.at(j);
 633           if (alignment(s2) == top_align) continue;
 634           if (s1 != s2 && are_adjacent_refs(s1, s2)) {
 635             if (stmts_can_pack(s1, s2, align)) {
 636               Node_List* pair = new Node_List();
 637               pair->push(s1);
 638               pair->push(s2);
 639               if (!_do_vector_loop || same_origin_idx(s1, s2)) {
 640                 _packset.append(pair);
 641               }
 642             }
 643           }
 644         }
 645       }
 646     } else { // Don't create unaligned pack
 647       // First, remove remaining memory ops of the same type from the list.
 648       for (int i = memops.size() - 1; i >= 0; i--) {
 649         MemNode* s = memops.at(i)->as_Mem();
 650         if (same_velt_type(s, mem_ref)) {
 651           memops.remove(i);
 652         }
 653       }
 654 
 655       // Second, remove already constructed packs of the same type.
 656       for (int i = _packset.length() - 1; i >= 0; i--) {
 657         Node_List* p = _packset.at(i);
 658         MemNode* s = p->at(0)->as_Mem();
 659         if (same_velt_type(s, mem_ref)) {
 660           remove_pack_at(i);
 661         }
 662       }
 663 
 664       // If needed find the best memory reference for loop alignment again.
 665       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 666         // Put memory ops from remaining packs back on memops list for
 667         // the best alignment search.
 668         uint orig_msize = memops.size();
 669         for (int i = 0; i < _packset.length(); i++) {
 670           Node_List* p = _packset.at(i);
 671           MemNode* s = p->at(0)->as_Mem();
 672           assert(!same_velt_type(s, mem_ref), "sanity");
 673           memops.push(s);
 674         }
 675         best_align_to_mem_ref = find_align_to_ref(memops);
 676         if (best_align_to_mem_ref == NULL) {
 677           if (TraceSuperWord) {
 678             tty->print_cr("SuperWord::find_adjacent_refs(): best_align_to_mem_ref == NULL");
 679           }
 680           break;
 681         }
 682         best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
 683         NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
 684         // Restore list.
 685         while (memops.size() > orig_msize)
 686           (void)memops.pop();
 687       }
 688     } // unaligned memory accesses
 689 
 690     // Remove used mem nodes.
 691     for (int i = memops.size() - 1; i >= 0; i--) {
 692       MemNode* m = memops.at(i)->as_Mem();
 693       if (alignment(m) != top_align) {
 694         memops.remove(i);
 695       }
 696     }
 697 
 698   } // while (memops.size() != 0
 699   set_align_to_ref(best_align_to_mem_ref);
 700 
 701   if (TraceSuperWord) {
 702     tty->print_cr("\nAfter find_adjacent_refs");
 703     print_packset();
 704   }
 705 }
 706 
 707 #ifndef PRODUCT
 708 void SuperWord::find_adjacent_refs_trace_1(Node* best_align_to_mem_ref, int best_iv_adjustment) {
 709   if (is_trace_adjacent()) {
 710     tty->print("SuperWord::find_adjacent_refs best_align_to_mem_ref = %d, best_iv_adjustment = %d",
 711        best_align_to_mem_ref->_idx, best_iv_adjustment);
 712        best_align_to_mem_ref->dump();
 713   }
 714 }
 715 #endif
 716 
 717 //------------------------------find_align_to_ref---------------------------
 718 // Find a memory reference to align the loop induction variable to.
 719 // Looks first at stores then at loads, looking for a memory reference
 720 // with the largest number of references similar to it.
 721 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
 722   GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
 723 
 724   // Count number of comparable memory ops
 725   for (uint i = 0; i < memops.size(); i++) {
 726     MemNode* s1 = memops.at(i)->as_Mem();
 727     SWPointer p1(s1, this, NULL, false);
 728     // Discard if pre loop can't align this reference
 729     if (!ref_is_alignable(p1)) {
 730       *cmp_ct.adr_at(i) = 0;
 731       continue;
 732     }
 733     for (uint j = i+1; j < memops.size(); j++) {
 734       MemNode* s2 = memops.at(j)->as_Mem();
 735       if (isomorphic(s1, s2)) {
 736         SWPointer p2(s2, this, NULL, false);
 737         if (p1.comparable(p2)) {
 738           (*cmp_ct.adr_at(i))++;
 739           (*cmp_ct.adr_at(j))++;
 740         }
 741       }
 742     }
 743   }
 744 
 745   // Find Store (or Load) with the greatest number of "comparable" references,
 746   // biggest vector size, smallest data size and smallest iv offset.
 747   int max_ct        = 0;
 748   int max_vw        = 0;
 749   int max_idx       = -1;
 750   int min_size      = max_jint;
 751   int min_iv_offset = max_jint;
 752   for (uint j = 0; j < memops.size(); j++) {
 753     MemNode* s = memops.at(j)->as_Mem();
 754     if (s->is_Store()) {
 755       int vw = vector_width_in_bytes(s);
 756       assert(vw > 1, "sanity");
 757       SWPointer p(s, this, NULL, false);
 758       if ( cmp_ct.at(j) >  max_ct ||
 759           (cmp_ct.at(j) == max_ct &&
 760             ( vw >  max_vw ||
 761              (vw == max_vw &&
 762               ( data_size(s) <  min_size ||
 763                (data_size(s) == min_size &&
 764                 p.offset_in_bytes() < min_iv_offset)))))) {
 765         max_ct = cmp_ct.at(j);
 766         max_vw = vw;
 767         max_idx = j;
 768         min_size = data_size(s);
 769         min_iv_offset = p.offset_in_bytes();
 770       }
 771     }
 772   }
 773   // If no stores, look at loads
 774   if (max_ct == 0) {
 775     for (uint j = 0; j < memops.size(); j++) {
 776       MemNode* s = memops.at(j)->as_Mem();
 777       if (s->is_Load()) {
 778         int vw = vector_width_in_bytes(s);
 779         assert(vw > 1, "sanity");
 780         SWPointer p(s, this, NULL, false);
 781         if ( cmp_ct.at(j) >  max_ct ||
 782             (cmp_ct.at(j) == max_ct &&
 783               ( vw >  max_vw ||
 784                (vw == max_vw &&
 785                 ( data_size(s) <  min_size ||
 786                  (data_size(s) == min_size &&
 787                   p.offset_in_bytes() < min_iv_offset)))))) {
 788           max_ct = cmp_ct.at(j);
 789           max_vw = vw;
 790           max_idx = j;
 791           min_size = data_size(s);
 792           min_iv_offset = p.offset_in_bytes();
 793         }
 794       }
 795     }
 796   }
 797 
 798 #ifdef ASSERT
 799   if (TraceSuperWord && Verbose) {
 800     tty->print_cr("\nVector memops after find_align_to_ref");
 801     for (uint i = 0; i < memops.size(); i++) {
 802       MemNode* s = memops.at(i)->as_Mem();
 803       s->dump();
 804     }
 805   }
 806 #endif
 807 
 808   if (max_ct > 0) {
 809 #ifdef ASSERT
 810     if (TraceSuperWord) {
 811       tty->print("\nVector align to node: ");
 812       memops.at(max_idx)->as_Mem()->dump();
 813     }
 814 #endif
 815     return memops.at(max_idx)->as_Mem();
 816   }
 817   return NULL;
 818 }
 819 
 820 //------------------------------ref_is_alignable---------------------------
 821 // Can the preloop align the reference to position zero in the vector?
 822 bool SuperWord::ref_is_alignable(SWPointer& p) {
 823   if (!p.has_iv()) {
 824     return true;   // no induction variable
 825   }
 826   CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
 827   assert(pre_end != NULL, "we must have a correct pre-loop");
 828   assert(pre_end->stride_is_con(), "pre loop stride is constant");
 829   int preloop_stride = pre_end->stride_con();
 830 
 831   int span = preloop_stride * p.scale_in_bytes();
 832   int mem_size = p.memory_size();
 833   int offset   = p.offset_in_bytes();
 834   // Stride one accesses are alignable if offset is aligned to memory operation size.
 835   // Offset can be unaligned when UseUnalignedAccesses is used.
 836   if (ABS(span) == mem_size && (ABS(offset) % mem_size) == 0) {
 837     return true;
 838   }
 839   // If the initial offset from start of the object is computable,
 840   // check if the pre-loop can align the final offset accordingly.
 841   //
 842   // In other words: Can we find an i such that the offset
 843   // after i pre-loop iterations is aligned to vw?
 844   //   (init_offset + pre_loop) % vw == 0              (1)
 845   // where
 846   //   pre_loop = i * span
 847   // is the number of bytes added to the offset by i pre-loop iterations.
 848   //
 849   // For this to hold we need pre_loop to increase init_offset by
 850   //   pre_loop = vw - (init_offset % vw)
 851   //
 852   // This is only possible if pre_loop is divisible by span because each
 853   // pre-loop iteration increases the initial offset by 'span' bytes:
 854   //   (vw - (init_offset % vw)) % span == 0
 855   //
 856   int vw = vector_width_in_bytes(p.mem());
 857   assert(vw > 1, "sanity");
 858   Node* init_nd = pre_end->init_trip();
 859   if (init_nd->is_Con() && p.invar() == NULL) {
 860     int init = init_nd->bottom_type()->is_int()->get_con();
 861     int init_offset = init * p.scale_in_bytes() + offset;
 862     assert(init_offset >= 0, "positive offset from object start");
 863     if (vw % span == 0) {
 864       // If vm is a multiple of span, we use formula (1).
 865       if (span > 0) {
 866         return (vw - (init_offset % vw)) % span == 0;
 867       } else {
 868         assert(span < 0, "nonzero stride * scale");
 869         return (init_offset % vw) % -span == 0;
 870       }
 871     } else if (span % vw == 0) {
 872       // If span is a multiple of vw, we can simplify formula (1) to:
 873       //   (init_offset + i * span) % vw == 0
 874       //     =>
 875       //   (init_offset % vw) + ((i * span) % vw) == 0
 876       //     =>
 877       //   init_offset % vw == 0
 878       //
 879       // Because we add a multiple of vw to the initial offset, the final
 880       // offset is a multiple of vw if and only if init_offset is a multiple.
 881       //
 882       return (init_offset % vw) == 0;
 883     }
 884   }
 885   return false;
 886 }
 887 
 888 //---------------------------get_iv_adjustment---------------------------
 889 // Calculate loop's iv adjustment for this memory ops.
 890 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
 891   SWPointer align_to_ref_p(mem_ref, this, NULL, false);
 892   int offset = align_to_ref_p.offset_in_bytes();
 893   int scale  = align_to_ref_p.scale_in_bytes();
 894   int elt_size = align_to_ref_p.memory_size();
 895   int vw       = vector_width_in_bytes(mem_ref);
 896   assert(vw > 1, "sanity");
 897   int iv_adjustment;
 898   if (scale != 0) {
 899     int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
 900     // At least one iteration is executed in pre-loop by default. As result
 901     // several iterations are needed to align memory operations in main-loop even
 902     // if offset is 0.
 903     int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
 904     assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
 905            "(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size);
 906     iv_adjustment = iv_adjustment_in_bytes/elt_size;
 907   } else {
 908     // This memory op is not dependent on iv (scale == 0)
 909     iv_adjustment = 0;
 910   }
 911 
 912 #ifndef PRODUCT
 913   if (TraceSuperWord) {
 914     tty->print("SuperWord::get_iv_adjustment: n = %d, noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d: ",
 915       mem_ref->_idx, offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
 916     mem_ref->dump();
 917   }
 918 #endif
 919   return iv_adjustment;
 920 }
 921 
 922 //---------------------------dependence_graph---------------------------
 923 // Construct dependency graph.
 924 // Add dependence edges to load/store nodes for memory dependence
 925 //    A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
 926 void SuperWord::dependence_graph() {
 927   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
 928   // First, assign a dependence node to each memory node
 929   for (int i = 0; i < _block.length(); i++ ) {
 930     Node *n = _block.at(i);
 931     if (n->is_Mem() || (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
 932       _dg.make_node(n);
 933     }
 934   }
 935 
 936   // For each memory slice, create the dependences
 937   for (int i = 0; i < _mem_slice_head.length(); i++) {
 938     Node* n      = _mem_slice_head.at(i);
 939     Node* n_tail = _mem_slice_tail.at(i);
 940 
 941     // Get slice in predecessor order (last is first)
 942     if (cl->is_main_loop()) {
 943       mem_slice_preds(n_tail, n, _nlist);
 944     }
 945 
 946 #ifndef PRODUCT
 947     if(TraceSuperWord && Verbose) {
 948       tty->print_cr("SuperWord::dependence_graph: built a new mem slice");
 949       for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 950         _nlist.at(j)->dump();
 951       }
 952     }
 953 #endif
 954     // Make the slice dependent on the root
 955     DepMem* slice = _dg.dep(n);
 956     _dg.make_edge(_dg.root(), slice);
 957 
 958     // Create a sink for the slice
 959     DepMem* slice_sink = _dg.make_node(NULL);
 960     _dg.make_edge(slice_sink, _dg.tail());
 961 
 962     // Now visit each pair of memory ops, creating the edges
 963     for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 964       Node* s1 = _nlist.at(j);
 965 
 966       // If no dependency yet, use slice
 967       if (_dg.dep(s1)->in_cnt() == 0) {
 968         _dg.make_edge(slice, s1);
 969       }
 970       SWPointer p1(s1->as_Mem(), this, NULL, false);
 971       bool sink_dependent = true;
 972       for (int k = j - 1; k >= 0; k--) {
 973         Node* s2 = _nlist.at(k);
 974         if (s1->is_Load() && s2->is_Load())
 975           continue;
 976         SWPointer p2(s2->as_Mem(), this, NULL, false);
 977 
 978         int cmp = p1.cmp(p2);
 979         if (SuperWordRTDepCheck &&
 980             p1.base() != p2.base() && p1.valid() && p2.valid()) {
 981           // Create a runtime check to disambiguate
 982           OrderedPair pp(p1.base(), p2.base());
 983           _disjoint_ptrs.append_if_missing(pp);
 984         } else if (!SWPointer::not_equal(cmp)) {
 985           // Possibly same address
 986           _dg.make_edge(s1, s2);
 987           sink_dependent = false;
 988         }
 989       }
 990       if (sink_dependent) {
 991         _dg.make_edge(s1, slice_sink);
 992       }
 993     }
 994 
 995     if (TraceSuperWord) {
 996       tty->print_cr("\nDependence graph for slice: %d", n->_idx);
 997       for (int q = 0; q < _nlist.length(); q++) {
 998         _dg.print(_nlist.at(q));
 999       }
1000       tty->cr();
1001     }
1002 
1003     _nlist.clear();
1004   }
1005 
1006   if (TraceSuperWord) {
1007     tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
1008     for (int r = 0; r < _disjoint_ptrs.length(); r++) {
1009       _disjoint_ptrs.at(r).print();
1010       tty->cr();
1011     }
1012     tty->cr();
1013   }
1014 
1015 }
1016 
1017 //---------------------------mem_slice_preds---------------------------
1018 // Return a memory slice (node list) in predecessor order starting at "start"
1019 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
1020   assert(preds.length() == 0, "start empty");
1021   Node* n = start;
1022   Node* prev = NULL;
1023   while (true) {
1024     NOT_PRODUCT( if(is_trace_mem_slice()) tty->print_cr("SuperWord::mem_slice_preds: n %d", n->_idx);)
1025     assert(in_bb(n), "must be in block");
1026     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1027       Node* out = n->fast_out(i);
1028       if (out->is_Load()) {
1029         if (in_bb(out)) {
1030           preds.push(out);
1031           if (TraceSuperWord && Verbose) {
1032             tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", out->_idx);
1033           }
1034         }
1035       } else {
1036         // FIXME
1037         if (out->is_MergeMem() && !in_bb(out)) {
1038           // Either unrolling is causing a memory edge not to disappear,
1039           // or need to run igvn.optimize() again before SLP
1040         } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
1041           // Ditto.  Not sure what else to check further.
1042         } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
1043           // StoreCM has an input edge used as a precedence edge.
1044           // Maybe an issue when oop stores are vectorized.
1045         } else {
1046           assert(out == prev || prev == NULL, "no branches off of store slice");
1047         }
1048       }//else
1049     }//for
1050     if (n == stop) break;
1051     preds.push(n);
1052     if (TraceSuperWord && Verbose) {
1053       tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", n->_idx);
1054     }
1055     prev = n;
1056     assert(n->is_Mem(), "unexpected node %s", n->Name());
1057     n = n->in(MemNode::Memory);
1058   }
1059 }
1060 
1061 //------------------------------stmts_can_pack---------------------------
1062 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
1063 // s1 aligned at "align"
1064 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
1065 
1066   // Do not use superword for non-primitives
1067   BasicType bt1 = velt_basic_type(s1);
1068   BasicType bt2 = velt_basic_type(s2);
1069   if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
1070     return false;
1071   if (Matcher::max_vector_size(bt1) < 2) {
1072     return false; // No vectors for this type
1073   }
1074 
1075   if (isomorphic(s1, s2)) {
1076     if (independent(s1, s2) || reduction(s1, s2)) {
1077       if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
1078         if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
1079           int s1_align = alignment(s1);
1080           int s2_align = alignment(s2);
1081           if (s1_align == top_align || s1_align == align) {
1082             if (s2_align == top_align || s2_align == align + data_size(s1)) {
1083               return true;
1084             }
1085           }
1086         }
1087       }
1088     }
1089   }
1090   return false;
1091 }
1092 
1093 //------------------------------exists_at---------------------------
1094 // Does s exist in a pack at position pos?
1095 bool SuperWord::exists_at(Node* s, uint pos) {
1096   for (int i = 0; i < _packset.length(); i++) {
1097     Node_List* p = _packset.at(i);
1098     if (p->at(pos) == s) {
1099       return true;
1100     }
1101   }
1102   return false;
1103 }
1104 
1105 //------------------------------are_adjacent_refs---------------------------
1106 // Is s1 immediately before s2 in memory?
1107 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
1108   if (!s1->is_Mem() || !s2->is_Mem()) return false;
1109   if (!in_bb(s1)    || !in_bb(s2))    return false;
1110 
1111   // Do not use superword for non-primitives
1112   if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
1113       !is_java_primitive(s2->as_Mem()->memory_type())) {
1114     return false;
1115   }
1116 
1117   // FIXME - co_locate_pack fails on Stores in different mem-slices, so
1118   // only pack memops that are in the same alias set until that's fixed.
1119   if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
1120       _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
1121     return false;
1122   SWPointer p1(s1->as_Mem(), this, NULL, false);
1123   SWPointer p2(s2->as_Mem(), this, NULL, false);
1124   if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
1125   int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
1126   return diff == data_size(s1);
1127 }
1128 
1129 //------------------------------isomorphic---------------------------
1130 // Are s1 and s2 similar?
1131 bool SuperWord::isomorphic(Node* s1, Node* s2) {
1132   if (s1->Opcode() != s2->Opcode()) return false;
1133   if (s1->req() != s2->req()) return false;
1134   if (s1->in(0) != s2->in(0)) return false;
1135   if (!same_velt_type(s1, s2)) return false;
1136   return true;
1137 }
1138 
1139 //------------------------------independent---------------------------
1140 // Is there no data path from s1 to s2 or s2 to s1?
1141 bool SuperWord::independent(Node* s1, Node* s2) {
1142   //  assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
1143   int d1 = depth(s1);
1144   int d2 = depth(s2);
1145   if (d1 == d2) return s1 != s2;
1146   Node* deep    = d1 > d2 ? s1 : s2;
1147   Node* shallow = d1 > d2 ? s2 : s1;
1148 
1149   visited_clear();
1150 
1151   return independent_path(shallow, deep);
1152 }
1153 
1154 //------------------------------reduction---------------------------
1155 // Is there a data path between s1 and s2 and the nodes reductions?
1156 bool SuperWord::reduction(Node* s1, Node* s2) {
1157   bool retValue = false;
1158   int d1 = depth(s1);
1159   int d2 = depth(s2);
1160   if (d1 + 1 == d2) {
1161     if (s1->is_reduction() && s2->is_reduction()) {
1162       // This is an ordered set, so s1 should define s2
1163       for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1164         Node* t1 = s1->fast_out(i);
1165         if (t1 == s2) {
1166           // both nodes are reductions and connected
1167           retValue = true;
1168         }
1169       }
1170     }
1171   }
1172 
1173   return retValue;
1174 }
1175 
1176 //------------------------------independent_path------------------------------
1177 // Helper for independent
1178 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
1179   if (dp >= 1000) return false; // stop deep recursion
1180   visited_set(deep);
1181   int shal_depth = depth(shallow);
1182   assert(shal_depth <= depth(deep), "must be");
1183   for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
1184     Node* pred = preds.current();
1185     if (in_bb(pred) && !visited_test(pred)) {
1186       if (shallow == pred) {
1187         return false;
1188       }
1189       if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
1190         return false;
1191       }
1192     }
1193   }
1194   return true;
1195 }
1196 
1197 //------------------------------set_alignment---------------------------
1198 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
1199   set_alignment(s1, align);
1200   if (align == top_align || align == bottom_align) {
1201     set_alignment(s2, align);
1202   } else {
1203     set_alignment(s2, align + data_size(s1));
1204   }
1205 }
1206 
1207 //------------------------------data_size---------------------------
1208 int SuperWord::data_size(Node* s) {
1209   Node* use = NULL; //test if the node is a candidate for CMoveVD optimization, then return the size of CMov
1210   if (_do_vector_loop) {
1211     use = _cmovev_kit.is_Bool_candidate(s);
1212     if (use != NULL) {
1213       return data_size(use);
1214     }
1215     use = _cmovev_kit.is_CmpD_candidate(s);
1216     if (use != NULL) {
1217       return data_size(use);
1218     }
1219   }
1220   int bsize = type2aelembytes(velt_basic_type(s));
1221   assert(bsize != 0, "valid size");
1222   return bsize;
1223 }
1224 
1225 //------------------------------extend_packlist---------------------------
1226 // Extend packset by following use->def and def->use links from pack members.
1227 void SuperWord::extend_packlist() {
1228   bool changed;
1229   do {
1230     packset_sort(_packset.length());
1231     changed = false;
1232     for (int i = 0; i < _packset.length(); i++) {
1233       Node_List* p = _packset.at(i);
1234       changed |= follow_use_defs(p);
1235       changed |= follow_def_uses(p);
1236     }
1237   } while (changed);
1238 
1239   if (_race_possible) {
1240     for (int i = 0; i < _packset.length(); i++) {
1241       Node_List* p = _packset.at(i);
1242       order_def_uses(p);
1243     }
1244   }
1245 
1246   if (TraceSuperWord) {
1247     tty->print_cr("\nAfter extend_packlist");
1248     print_packset();
1249   }
1250 }
1251 
1252 //------------------------------follow_use_defs---------------------------
1253 // Extend the packset by visiting operand definitions of nodes in pack p
1254 bool SuperWord::follow_use_defs(Node_List* p) {
1255   assert(p->size() == 2, "just checking");
1256   Node* s1 = p->at(0);
1257   Node* s2 = p->at(1);
1258   assert(s1->req() == s2->req(), "just checking");
1259   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1260 
1261   if (s1->is_Load()) return false;
1262 
1263   int align = alignment(s1);
1264   NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: s1 %d, align %d", s1->_idx, align);)
1265   bool changed = false;
1266   int start = s1->is_Store() ? MemNode::ValueIn   : 1;
1267   int end   = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
1268   for (int j = start; j < end; j++) {
1269     Node* t1 = s1->in(j);
1270     Node* t2 = s2->in(j);
1271     if (!in_bb(t1) || !in_bb(t2))
1272       continue;
1273     if (stmts_can_pack(t1, t2, align)) {
1274       if (est_savings(t1, t2) >= 0) {
1275         Node_List* pair = new Node_List();
1276         pair->push(t1);
1277         pair->push(t2);
1278         _packset.append(pair);
1279         NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: set_alignment(%d, %d, %d)", t1->_idx, t2->_idx, align);)
1280         set_alignment(t1, t2, align);
1281         changed = true;
1282       }
1283     }
1284   }
1285   return changed;
1286 }
1287 
1288 //------------------------------follow_def_uses---------------------------
1289 // Extend the packset by visiting uses of nodes in pack p
1290 bool SuperWord::follow_def_uses(Node_List* p) {
1291   bool changed = false;
1292   Node* s1 = p->at(0);
1293   Node* s2 = p->at(1);
1294   assert(p->size() == 2, "just checking");
1295   assert(s1->req() == s2->req(), "just checking");
1296   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1297 
1298   if (s1->is_Store()) return false;
1299 
1300   int align = alignment(s1);
1301   NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: s1 %d, align %d", s1->_idx, align);)
1302   int savings = -1;
1303   int num_s1_uses = 0;
1304   Node* u1 = NULL;
1305   Node* u2 = NULL;
1306   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1307     Node* t1 = s1->fast_out(i);
1308     num_s1_uses++;
1309     if (!in_bb(t1)) continue;
1310     for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1311       Node* t2 = s2->fast_out(j);
1312       if (!in_bb(t2)) continue;
1313       if (!opnd_positions_match(s1, t1, s2, t2))
1314         continue;
1315       if (stmts_can_pack(t1, t2, align)) {
1316         int my_savings = est_savings(t1, t2);
1317         if (my_savings > savings) {
1318           savings = my_savings;
1319           u1 = t1;
1320           u2 = t2;
1321         }
1322       }
1323     }
1324   }
1325   if (num_s1_uses > 1) {
1326     _race_possible = true;
1327   }
1328   if (savings >= 0) {
1329     Node_List* pair = new Node_List();
1330     pair->push(u1);
1331     pair->push(u2);
1332     _packset.append(pair);
1333     NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: set_alignment(%d, %d, %d)", u1->_idx, u2->_idx, align);)
1334     set_alignment(u1, u2, align);
1335     changed = true;
1336   }
1337   return changed;
1338 }
1339 
1340 //------------------------------order_def_uses---------------------------
1341 // For extended packsets, ordinally arrange uses packset by major component
1342 void SuperWord::order_def_uses(Node_List* p) {
1343   Node* s1 = p->at(0);
1344 
1345   if (s1->is_Store()) return;
1346 
1347   // reductions are always managed beforehand
1348   if (s1->is_reduction()) return;
1349 
1350   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1351     Node* t1 = s1->fast_out(i);
1352 
1353     // Only allow operand swap on commuting operations
1354     if (!t1->is_Add() && !t1->is_Mul()) {
1355       break;
1356     }
1357 
1358     // Now find t1's packset
1359     Node_List* p2 = NULL;
1360     for (int j = 0; j < _packset.length(); j++) {
1361       p2 = _packset.at(j);
1362       Node* first = p2->at(0);
1363       if (t1 == first) {
1364         break;
1365       }
1366       p2 = NULL;
1367     }
1368     // Arrange all sub components by the major component
1369     if (p2 != NULL) {
1370       for (uint j = 1; j < p->size(); j++) {
1371         Node* d1 = p->at(j);
1372         Node* u1 = p2->at(j);
1373         opnd_positions_match(s1, t1, d1, u1);
1374       }
1375     }
1376   }
1377 }
1378 
1379 //---------------------------opnd_positions_match-------------------------
1380 // Is the use of d1 in u1 at the same operand position as d2 in u2?
1381 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
1382   // check reductions to see if they are marshalled to represent the reduction
1383   // operator in a specified opnd
1384   if (u1->is_reduction() && u2->is_reduction()) {
1385     // ensure reductions have phis and reduction definitions feeding the 1st operand
1386     Node* first = u1->in(2);
1387     if (first->is_Phi() || first->is_reduction()) {
1388       u1->swap_edges(1, 2);
1389     }
1390     // ensure reductions have phis and reduction definitions feeding the 1st operand
1391     first = u2->in(2);
1392     if (first->is_Phi() || first->is_reduction()) {
1393       u2->swap_edges(1, 2);
1394     }
1395     return true;
1396   }
1397 
1398   uint ct = u1->req();
1399   if (ct != u2->req()) return false;
1400   uint i1 = 0;
1401   uint i2 = 0;
1402   do {
1403     for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
1404     for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
1405     if (i1 != i2) {
1406       if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
1407         // Further analysis relies on operands position matching.
1408         u2->swap_edges(i1, i2);
1409       } else {
1410         return false;
1411       }
1412     }
1413   } while (i1 < ct);
1414   return true;
1415 }
1416 
1417 //------------------------------est_savings---------------------------
1418 // Estimate the savings from executing s1 and s2 as a pack
1419 int SuperWord::est_savings(Node* s1, Node* s2) {
1420   int save_in = 2 - 1; // 2 operations per instruction in packed form
1421 
1422   // inputs
1423   for (uint i = 1; i < s1->req(); i++) {
1424     Node* x1 = s1->in(i);
1425     Node* x2 = s2->in(i);
1426     if (x1 != x2) {
1427       if (are_adjacent_refs(x1, x2)) {
1428         save_in += adjacent_profit(x1, x2);
1429       } else if (!in_packset(x1, x2)) {
1430         save_in -= pack_cost(2);
1431       } else {
1432         save_in += unpack_cost(2);
1433       }
1434     }
1435   }
1436 
1437   // uses of result
1438   uint ct = 0;
1439   int save_use = 0;
1440   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1441     Node* s1_use = s1->fast_out(i);
1442     for (int j = 0; j < _packset.length(); j++) {
1443       Node_List* p = _packset.at(j);
1444       if (p->at(0) == s1_use) {
1445         for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
1446           Node* s2_use = s2->fast_out(k);
1447           if (p->at(p->size()-1) == s2_use) {
1448             ct++;
1449             if (are_adjacent_refs(s1_use, s2_use)) {
1450               save_use += adjacent_profit(s1_use, s2_use);
1451             }
1452           }
1453         }
1454       }
1455     }
1456   }
1457 
1458   if (ct < s1->outcnt()) save_use += unpack_cost(1);
1459   if (ct < s2->outcnt()) save_use += unpack_cost(1);
1460 
1461   return MAX2(save_in, save_use);
1462 }
1463 
1464 //------------------------------costs---------------------------
1465 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
1466 int SuperWord::pack_cost(int ct)   { return ct; }
1467 int SuperWord::unpack_cost(int ct) { return ct; }
1468 
1469 //------------------------------combine_packs---------------------------
1470 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
1471 void SuperWord::combine_packs() {
1472   bool changed = true;
1473   // Combine packs regardless max vector size.
1474   while (changed) {
1475     changed = false;
1476     for (int i = 0; i < _packset.length(); i++) {
1477       Node_List* p1 = _packset.at(i);
1478       if (p1 == NULL) continue;
1479       // Because of sorting we can start at i + 1
1480       for (int j = i + 1; j < _packset.length(); j++) {
1481         Node_List* p2 = _packset.at(j);
1482         if (p2 == NULL) continue;
1483         if (i == j) continue;
1484         if (p1->at(p1->size()-1) == p2->at(0)) {
1485           for (uint k = 1; k < p2->size(); k++) {
1486             p1->push(p2->at(k));
1487           }
1488           _packset.at_put(j, NULL);
1489           changed = true;
1490         }
1491       }
1492     }
1493   }
1494 
1495   // Split packs which have size greater then max vector size.
1496   for (int i = 0; i < _packset.length(); i++) {
1497     Node_List* p1 = _packset.at(i);
1498     if (p1 != NULL) {
1499       BasicType bt = velt_basic_type(p1->at(0));
1500       uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1501       assert(is_power_of_2(max_vlen), "sanity");
1502       uint psize = p1->size();
1503       if (!is_power_of_2(psize)) {
1504         // Skip pack which can't be vector.
1505         // case1: for(...) { a[i] = i; }    elements values are different (i+x)
1506         // case2: for(...) { a[i] = b[i+1]; }  can't align both, load and store
1507         _packset.at_put(i, NULL);
1508         continue;
1509       }
1510       if (psize > max_vlen) {
1511         Node_List* pack = new Node_List();
1512         for (uint j = 0; j < psize; j++) {
1513           pack->push(p1->at(j));
1514           if (pack->size() >= max_vlen) {
1515             assert(is_power_of_2(pack->size()), "sanity");
1516             _packset.append(pack);
1517             pack = new Node_List();
1518           }
1519         }
1520         _packset.at_put(i, NULL);
1521       }
1522     }
1523   }
1524 
1525   // Compress list.
1526   for (int i = _packset.length() - 1; i >= 0; i--) {
1527     Node_List* p1 = _packset.at(i);
1528     if (p1 == NULL) {
1529       _packset.remove_at(i);
1530     }
1531   }
1532 
1533   if (TraceSuperWord) {
1534     tty->print_cr("\nAfter combine_packs");
1535     print_packset();
1536   }
1537 }
1538 
1539 //-----------------------------construct_my_pack_map--------------------------
1540 // Construct the map from nodes to packs.  Only valid after the
1541 // point where a node is only in one pack (after combine_packs).
1542 void SuperWord::construct_my_pack_map() {
1543   Node_List* rslt = NULL;
1544   for (int i = 0; i < _packset.length(); i++) {
1545     Node_List* p = _packset.at(i);
1546     for (uint j = 0; j < p->size(); j++) {
1547       Node* s = p->at(j);
1548       assert(my_pack(s) == NULL, "only in one pack");
1549       set_my_pack(s, p);
1550     }
1551   }
1552 }
1553 
1554 //------------------------------filter_packs---------------------------
1555 // Remove packs that are not implemented or not profitable.
1556 void SuperWord::filter_packs() {
1557   // Remove packs that are not implemented
1558   for (int i = _packset.length() - 1; i >= 0; i--) {
1559     Node_List* pk = _packset.at(i);
1560     bool impl = implemented(pk);
1561     if (!impl) {
1562 #ifndef PRODUCT
1563       if (TraceSuperWord && Verbose) {
1564         tty->print_cr("Unimplemented");
1565         pk->at(0)->dump();
1566       }
1567 #endif
1568       remove_pack_at(i);
1569     }
1570     Node *n = pk->at(0);
1571     if (n->is_reduction()) {
1572       _num_reductions++;
1573     } else {
1574       _num_work_vecs++;
1575     }
1576   }
1577 
1578   // Remove packs that are not profitable
1579   bool changed;
1580   do {
1581     changed = false;
1582     for (int i = _packset.length() - 1; i >= 0; i--) {
1583       Node_List* pk = _packset.at(i);
1584       bool prof = profitable(pk);
1585       if (!prof) {
1586 #ifndef PRODUCT
1587         if (TraceSuperWord && Verbose) {
1588           tty->print_cr("Unprofitable");
1589           pk->at(0)->dump();
1590         }
1591 #endif
1592         remove_pack_at(i);
1593         changed = true;
1594       }
1595     }
1596   } while (changed);
1597 
1598 #ifndef PRODUCT
1599   if (TraceSuperWord) {
1600     tty->print_cr("\nAfter filter_packs");
1601     print_packset();
1602     tty->cr();
1603   }
1604 #endif
1605 }
1606 
1607 //------------------------------merge_packs_to_cmovd---------------------------
1608 // Merge CMoveD into new vector-nodes
1609 // We want to catch this pattern and subsume CmpD and Bool into CMoveD
1610 //
1611 //                   SubD             ConD
1612 //                  /  |               /
1613 //                 /   |           /   /
1614 //                /    |       /      /
1615 //               /     |   /         /
1616 //              /      /            /
1617 //             /    /  |           /
1618 //            v /      |          /
1619 //         CmpD        |         /
1620 //          |          |        /
1621 //          v          |       /
1622 //         Bool        |      /
1623 //           \         |     /
1624 //             \       |    /
1625 //               \     |   /
1626 //                 \   |  /
1627 //                   \ v /
1628 //                   CMoveD
1629 //
1630 
1631 void SuperWord::merge_packs_to_cmovd() {
1632   for (int i = _packset.length() - 1; i >= 0; i--) {
1633     _cmovev_kit.make_cmovevd_pack(_packset.at(i));
1634   }
1635   #ifndef PRODUCT
1636     if (TraceSuperWord) {
1637       tty->print_cr("\nSuperWord::merge_packs_to_cmovd(): After merge");
1638       print_packset();
1639       tty->cr();
1640     }
1641   #endif
1642 }
1643 
1644 Node* CMoveKit::is_Bool_candidate(Node* def) const {
1645   Node* use = NULL;
1646   if (!def->is_Bool() || def->in(0) != NULL || def->outcnt() != 1) {
1647     return NULL;
1648   }
1649   for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1650     use = def->fast_out(j);
1651     if (!_sw->same_generation(def, use) || !use->is_CMove()) {
1652       return NULL;
1653     }
1654   }
1655   return use;
1656 }
1657 
1658 Node* CMoveKit::is_CmpD_candidate(Node* def) const {
1659   Node* use = NULL;
1660   if (!def->is_Cmp() || def->in(0) != NULL || def->outcnt() != 1) {
1661     return NULL;
1662   }
1663   for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1664     use = def->fast_out(j);
1665     if (!_sw->same_generation(def, use) || (use = is_Bool_candidate(use)) == NULL || !_sw->same_generation(def, use)) {
1666       return NULL;
1667     }
1668   }
1669   return use;
1670 }
1671 
1672 Node_List* CMoveKit::make_cmovevd_pack(Node_List* cmovd_pk) {
1673   Node *cmovd = cmovd_pk->at(0);
1674   if (!cmovd->is_CMove()) {
1675     return NULL;
1676   }
1677   if (pack(cmovd) != NULL) { // already in the cmov pack
1678     return NULL;
1679   }
1680   if (cmovd->in(0) != NULL) {
1681     NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: CMoveD %d has control flow, escaping...", cmovd->_idx); cmovd->dump();})
1682     return NULL;
1683   }
1684 
1685   Node* bol = cmovd->as_CMove()->in(CMoveNode::Condition);
1686   if (!bol->is_Bool()
1687       || bol->outcnt() != 1
1688       || !_sw->same_generation(bol, cmovd)
1689       || bol->in(0) != NULL  // BoolNode has control flow!!
1690       || _sw->my_pack(bol) == NULL) {
1691       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();})
1692       return NULL;
1693   }
1694   Node_List* bool_pk = _sw->my_pack(bol);
1695   if (bool_pk->size() != cmovd_pk->size() ) {
1696     return NULL;
1697   }
1698 
1699   Node* cmpd = bol->in(1);
1700   if (!cmpd->is_Cmp()
1701       || cmpd->outcnt() != 1
1702       || !_sw->same_generation(cmpd, cmovd)
1703       || cmpd->in(0) != NULL  // CmpDNode has control flow!!
1704       || _sw->my_pack(cmpd) == NULL) {
1705       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();})
1706       return NULL;
1707   }
1708   Node_List* cmpd_pk = _sw->my_pack(cmpd);
1709   if (cmpd_pk->size() != cmovd_pk->size() ) {
1710     return NULL;
1711   }
1712 
1713   if (!test_cmpd_pack(cmpd_pk, cmovd_pk)) {
1714     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();})
1715     return NULL;
1716   }
1717 
1718   Node_List* new_cmpd_pk = new Node_List();
1719   uint sz = cmovd_pk->size() - 1;
1720   for (uint i = 0; i <= sz; ++i) {
1721     Node* cmov = cmovd_pk->at(i);
1722     Node* bol  = bool_pk->at(i);
1723     Node* cmp  = cmpd_pk->at(i);
1724 
1725     new_cmpd_pk->insert(i, cmov);
1726 
1727     map(cmov, new_cmpd_pk);
1728     map(bol, new_cmpd_pk);
1729     map(cmp, new_cmpd_pk);
1730 
1731     _sw->set_my_pack(cmov, new_cmpd_pk); // and keep old packs for cmp and bool
1732   }
1733   _sw->_packset.remove(cmovd_pk);
1734   _sw->_packset.remove(bool_pk);
1735   _sw->_packset.remove(cmpd_pk);
1736   _sw->_packset.append(new_cmpd_pk);
1737   NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print_cr("CMoveKit::make_cmovevd_pack: added syntactic CMoveD pack"); _sw->print_pack(new_cmpd_pk);})
1738   return new_cmpd_pk;
1739 }
1740 
1741 bool CMoveKit::test_cmpd_pack(Node_List* cmpd_pk, Node_List* cmovd_pk) {
1742   Node* cmpd0 = cmpd_pk->at(0);
1743   assert(cmpd0->is_Cmp(), "CMoveKit::test_cmpd_pack: should be CmpDNode");
1744   assert(cmovd_pk->at(0)->is_CMove(), "CMoveKit::test_cmpd_pack: should be CMoveD");
1745   assert(cmpd_pk->size() == cmovd_pk->size(), "CMoveKit::test_cmpd_pack: should be same size");
1746   Node* in1 = cmpd0->in(1);
1747   Node* in2 = cmpd0->in(2);
1748   Node_List* in1_pk = _sw->my_pack(in1);
1749   Node_List* in2_pk = _sw->my_pack(in2);
1750 
1751   if (  (in1_pk != NULL && in1_pk->size() != cmpd_pk->size())
1752      || (in2_pk != NULL && in2_pk->size() != cmpd_pk->size()) ) {
1753     return false;
1754   }
1755 
1756   // test if "all" in1 are in the same pack or the same node
1757   if (in1_pk == NULL) {
1758     for (uint j = 1; j < cmpd_pk->size(); j++) {
1759       if (cmpd_pk->at(j)->in(1) != in1) {
1760         return false;
1761       }
1762     }//for: in1_pk is not pack but all CmpD nodes in the pack have the same in(1)
1763   }
1764   // test if "all" in2 are in the same pack or the same node
1765   if (in2_pk == NULL) {
1766     for (uint j = 1; j < cmpd_pk->size(); j++) {
1767       if (cmpd_pk->at(j)->in(2) != in2) {
1768         return false;
1769       }
1770     }//for: in2_pk is not pack but all CmpD nodes in the pack have the same in(2)
1771   }
1772   //now check if cmpd_pk may be subsumed in vector built for cmovd_pk
1773   int cmovd_ind1, cmovd_ind2;
1774   if (cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1775    && cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1776       cmovd_ind1 = CMoveNode::IfFalse;
1777       cmovd_ind2 = CMoveNode::IfTrue;
1778   } else if (cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1779           && cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1780       cmovd_ind2 = CMoveNode::IfFalse;
1781       cmovd_ind1 = CMoveNode::IfTrue;
1782   }
1783   else {
1784     return false;
1785   }
1786 
1787   for (uint j = 1; j < cmpd_pk->size(); j++) {
1788     if (cmpd_pk->at(j)->in(1) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind1)
1789         || cmpd_pk->at(j)->in(2) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind2)) {
1790         return false;
1791     }//if
1792   }
1793   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(); })
1794   return true;
1795 }
1796 
1797 //------------------------------implemented---------------------------
1798 // Can code be generated for pack p?
1799 bool SuperWord::implemented(Node_List* p) {
1800   bool retValue = false;
1801   Node* p0 = p->at(0);
1802   if (p0 != NULL) {
1803     int opc = p0->Opcode();
1804     uint size = p->size();
1805     if (p0->is_reduction()) {
1806       const Type *arith_type = p0->bottom_type();
1807       // Length 2 reductions of INT/LONG do not offer performance benefits
1808       if (((arith_type->basic_type() == T_INT) || (arith_type->basic_type() == T_LONG)) && (size == 2)) {
1809         retValue = false;
1810       } else {
1811         retValue = ReductionNode::implemented(opc, size, arith_type->basic_type());
1812       }
1813     } else {
1814       retValue = VectorNode::implemented(opc, size, velt_basic_type(p0));
1815     }
1816     if (!retValue) {
1817       if (is_cmov_pack(p)) {
1818         NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::implemented: found cmpd pack"); print_pack(p);})
1819         return true;
1820       }
1821     }
1822   }
1823   return retValue;
1824 }
1825 
1826 bool SuperWord::is_cmov_pack(Node_List* p) {
1827   return _cmovev_kit.pack(p->at(0)) != NULL;
1828 }
1829 //------------------------------same_inputs--------------------------
1830 // For pack p, are all idx operands the same?
1831 bool SuperWord::same_inputs(Node_List* p, int idx) {
1832   Node* p0 = p->at(0);
1833   uint vlen = p->size();
1834   Node* p0_def = p0->in(idx);
1835   for (uint i = 1; i < vlen; i++) {
1836     Node* pi = p->at(i);
1837     Node* pi_def = pi->in(idx);
1838     if (p0_def != pi_def) {
1839       return false;
1840     }
1841   }
1842   return true;
1843 }
1844 
1845 //------------------------------profitable---------------------------
1846 // For pack p, are all operands and all uses (with in the block) vector?
1847 bool SuperWord::profitable(Node_List* p) {
1848   Node* p0 = p->at(0);
1849   uint start, end;
1850   VectorNode::vector_operands(p0, &start, &end);
1851 
1852   // Return false if some inputs are not vectors or vectors with different
1853   // size or alignment.
1854   // Also, for now, return false if not scalar promotion case when inputs are
1855   // the same. Later, implement PackNode and allow differing, non-vector inputs
1856   // (maybe just the ones from outside the block.)
1857   for (uint i = start; i < end; i++) {
1858     if (!is_vector_use(p0, i)) {
1859       return false;
1860     }
1861   }
1862   // Check if reductions are connected
1863   if (p0->is_reduction()) {
1864     Node* second_in = p0->in(2);
1865     Node_List* second_pk = my_pack(second_in);
1866     if ((second_pk == NULL) || (_num_work_vecs == _num_reductions)) {
1867       // Remove reduction flag if no parent pack or if not enough work
1868       // to cover reduction expansion overhead
1869       p0->remove_flag(Node::Flag_is_reduction);
1870       return false;
1871     } else if (second_pk->size() != p->size()) {
1872       return false;
1873     }
1874   }
1875   if (VectorNode::is_shift(p0)) {
1876     // For now, return false if shift count is vector or not scalar promotion
1877     // case (different shift counts) because it is not supported yet.
1878     Node* cnt = p0->in(2);
1879     Node_List* cnt_pk = my_pack(cnt);
1880     if (cnt_pk != NULL)
1881       return false;
1882     if (!same_inputs(p, 2))
1883       return false;
1884   }
1885   if (!p0->is_Store()) {
1886     // For now, return false if not all uses are vector.
1887     // Later, implement ExtractNode and allow non-vector uses (maybe
1888     // just the ones outside the block.)
1889     for (uint i = 0; i < p->size(); i++) {
1890       Node* def = p->at(i);
1891       if (is_cmov_pack_internal_node(p, def)) {
1892         continue;
1893       }
1894       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1895         Node* use = def->fast_out(j);
1896         for (uint k = 0; k < use->req(); k++) {
1897           Node* n = use->in(k);
1898           if (def == n) {
1899             // reductions can be loop carried dependences
1900             if (def->is_reduction() && use->is_Phi())
1901               continue;
1902             if (!is_vector_use(use, k)) {
1903               return false;
1904             }
1905           }
1906         }
1907       }
1908     }
1909   }
1910   return true;
1911 }
1912 
1913 //------------------------------schedule---------------------------
1914 // Adjust the memory graph for the packed operations
1915 void SuperWord::schedule() {
1916 
1917   // Co-locate in the memory graph the members of each memory pack
1918   for (int i = 0; i < _packset.length(); i++) {
1919     co_locate_pack(_packset.at(i));
1920   }
1921 }
1922 
1923 //-------------------------------remove_and_insert-------------------
1924 // Remove "current" from its current position in the memory graph and insert
1925 // it after the appropriate insertion point (lip or uip).
1926 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1927                                   Node *uip, Unique_Node_List &sched_before) {
1928   Node* my_mem = current->in(MemNode::Memory);
1929   bool sched_up = sched_before.member(current);
1930 
1931   // remove current_store from its current position in the memmory graph
1932   for (DUIterator i = current->outs(); current->has_out(i); i++) {
1933     Node* use = current->out(i);
1934     if (use->is_Mem()) {
1935       assert(use->in(MemNode::Memory) == current, "must be");
1936       if (use == prev) { // connect prev to my_mem
1937           _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1938           --i; //deleted this edge; rescan position
1939       } else if (sched_before.member(use)) {
1940         if (!sched_up) { // Will be moved together with current
1941           _igvn.replace_input_of(use, MemNode::Memory, uip);
1942           --i; //deleted this edge; rescan position
1943         }
1944       } else {
1945         if (sched_up) { // Will be moved together with current
1946           _igvn.replace_input_of(use, MemNode::Memory, lip);
1947           --i; //deleted this edge; rescan position
1948         }
1949       }
1950     }
1951   }
1952 
1953   Node *insert_pt =  sched_up ?  uip : lip;
1954 
1955   // all uses of insert_pt's memory state should use current's instead
1956   for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1957     Node* use = insert_pt->out(i);
1958     if (use->is_Mem()) {
1959       assert(use->in(MemNode::Memory) == insert_pt, "must be");
1960       _igvn.replace_input_of(use, MemNode::Memory, current);
1961       --i; //deleted this edge; rescan position
1962     } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1963       uint pos; //lip (lower insert point) must be the last one in the memory slice
1964       for (pos=1; pos < use->req(); pos++) {
1965         if (use->in(pos) == insert_pt) break;
1966       }
1967       _igvn.replace_input_of(use, pos, current);
1968       --i;
1969     }
1970   }
1971 
1972   //connect current to insert_pt
1973   _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
1974 }
1975 
1976 //------------------------------co_locate_pack----------------------------------
1977 // To schedule a store pack, we need to move any sandwiched memory ops either before
1978 // or after the pack, based upon dependence information:
1979 // (1) If any store in the pack depends on the sandwiched memory op, the
1980 //     sandwiched memory op must be scheduled BEFORE the pack;
1981 // (2) If a sandwiched memory op depends on any store in the pack, the
1982 //     sandwiched memory op must be scheduled AFTER the pack;
1983 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1984 //     memory op (say memB), memB must be scheduled before memA. So, if memA is
1985 //     scheduled before the pack, memB must also be scheduled before the pack;
1986 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1987 //     schedule this store AFTER the pack
1988 // (5) We know there is no dependence cycle, so there in no other case;
1989 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1990 //
1991 // To schedule a load pack, we use the memory state of either the first or the last load in
1992 // the pack, based on the dependence constraint.
1993 void SuperWord::co_locate_pack(Node_List* pk) {
1994   if (pk->at(0)->is_Store()) {
1995     MemNode* first     = executed_first(pk)->as_Mem();
1996     MemNode* last      = executed_last(pk)->as_Mem();
1997     Unique_Node_List schedule_before_pack;
1998     Unique_Node_List memops;
1999 
2000     MemNode* current   = last->in(MemNode::Memory)->as_Mem();
2001     MemNode* previous  = last;
2002     while (true) {
2003       assert(in_bb(current), "stay in block");
2004       memops.push(previous);
2005       for (DUIterator i = current->outs(); current->has_out(i); i++) {
2006         Node* use = current->out(i);
2007         if (use->is_Mem() && use != previous)
2008           memops.push(use);
2009       }
2010       if (current == first) break;
2011       previous = current;
2012       current  = current->in(MemNode::Memory)->as_Mem();
2013     }
2014 
2015     // determine which memory operations should be scheduled before the pack
2016     for (uint i = 1; i < memops.size(); i++) {
2017       Node *s1 = memops.at(i);
2018       if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
2019         for (uint j = 0; j< i; j++) {
2020           Node *s2 = memops.at(j);
2021           if (!independent(s1, s2)) {
2022             if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
2023               schedule_before_pack.push(s1); // s1 must be scheduled before
2024               Node_List* mem_pk = my_pack(s1);
2025               if (mem_pk != NULL) {
2026                 for (uint ii = 0; ii < mem_pk->size(); ii++) {
2027                   Node* s = mem_pk->at(ii);  // follow partner
2028                   if (memops.member(s) && !schedule_before_pack.member(s))
2029                     schedule_before_pack.push(s);
2030                 }
2031               }
2032               break;
2033             }
2034           }
2035         }
2036       }
2037     }
2038 
2039     Node*    upper_insert_pt = first->in(MemNode::Memory);
2040     // Following code moves loads connected to upper_insert_pt below aliased stores.
2041     // Collect such loads here and reconnect them back to upper_insert_pt later.
2042     memops.clear();
2043     for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
2044       Node* use = upper_insert_pt->out(i);
2045       if (use->is_Mem() && !use->is_Store()) {
2046         memops.push(use);
2047       }
2048     }
2049 
2050     MemNode* lower_insert_pt = last;
2051     previous                 = last; //previous store in pk
2052     current                  = last->in(MemNode::Memory)->as_Mem();
2053 
2054     // start scheduling from "last" to "first"
2055     while (true) {
2056       assert(in_bb(current), "stay in block");
2057       assert(in_pack(previous, pk), "previous stays in pack");
2058       Node* my_mem = current->in(MemNode::Memory);
2059 
2060       if (in_pack(current, pk)) {
2061         // Forward users of my memory state (except "previous) to my input memory state
2062         for (DUIterator i = current->outs(); current->has_out(i); i++) {
2063           Node* use = current->out(i);
2064           if (use->is_Mem() && use != previous) {
2065             assert(use->in(MemNode::Memory) == current, "must be");
2066             if (schedule_before_pack.member(use)) {
2067               _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
2068             } else {
2069               _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
2070             }
2071             --i; // deleted this edge; rescan position
2072           }
2073         }
2074         previous = current;
2075       } else { // !in_pack(current, pk) ==> a sandwiched store
2076         remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
2077       }
2078 
2079       if (current == first) break;
2080       current = my_mem->as_Mem();
2081     } // end while
2082 
2083     // Reconnect loads back to upper_insert_pt.
2084     for (uint i = 0; i < memops.size(); i++) {
2085       Node *ld = memops.at(i);
2086       if (ld->in(MemNode::Memory) != upper_insert_pt) {
2087         _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
2088       }
2089     }
2090   } else if (pk->at(0)->is_Load()) { //load
2091     // all loads in the pack should have the same memory state. By default,
2092     // we use the memory state of the last load. However, if any load could
2093     // not be moved down due to the dependence constraint, we use the memory
2094     // state of the first load.
2095     Node* last_mem  = executed_last(pk)->in(MemNode::Memory);
2096     Node* first_mem = executed_first(pk)->in(MemNode::Memory);
2097     bool schedule_last = true;
2098     for (uint i = 0; i < pk->size(); i++) {
2099       Node* ld = pk->at(i);
2100       for (Node* current = last_mem; current != ld->in(MemNode::Memory);
2101            current=current->in(MemNode::Memory)) {
2102         assert(current != first_mem, "corrupted memory graph");
2103         if(current->is_Mem() && !independent(current, ld)){
2104           schedule_last = false; // a later store depends on this load
2105           break;
2106         }
2107       }
2108     }
2109 
2110     Node* mem_input = schedule_last ? last_mem : first_mem;
2111     _igvn.hash_delete(mem_input);
2112     // Give each load the same memory state
2113     for (uint i = 0; i < pk->size(); i++) {
2114       LoadNode* ld = pk->at(i)->as_Load();
2115       _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
2116     }
2117   }
2118 }
2119 
2120 #ifndef PRODUCT
2121 void SuperWord::print_loop(bool whole) {
2122   Node_Stack stack(_arena, _phase->C->unique() >> 2);
2123   Node_List rpo_list;
2124   VectorSet visited(_arena);
2125   visited.set(lpt()->_head->_idx);
2126   _phase->rpo(lpt()->_head, stack, visited, rpo_list);
2127   _phase->dump(lpt(), rpo_list.size(), rpo_list );
2128   if(whole) {
2129     tty->print_cr("\n Whole loop tree");
2130     _phase->dump();
2131     tty->print_cr(" End of whole loop tree\n");
2132   }
2133 }
2134 #endif
2135 
2136 //------------------------------output---------------------------
2137 // Convert packs into vector node operations
2138 void SuperWord::output() {
2139   if (_packset.length() == 0) return;
2140 
2141 #ifndef PRODUCT
2142   if (TraceLoopOpts) {
2143     tty->print("SuperWord::output    ");
2144     lpt()->dump_head();
2145   }
2146 #endif
2147 
2148   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2149   if (cl->is_main_loop()) {
2150     // MUST ENSURE main loop's initial value is properly aligned:
2151     //  (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
2152 
2153     align_initial_loop_index(align_to_ref());
2154 
2155     // Insert extract (unpack) operations for scalar uses
2156     for (int i = 0; i < _packset.length(); i++) {
2157       insert_extracts(_packset.at(i));
2158     }
2159   }
2160 
2161   Compile* C = _phase->C;
2162   uint max_vlen_in_bytes = 0;
2163   uint max_vlen = 0;
2164   bool can_process_post_loop = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
2165 
2166   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop before create_reserve_version_of_loop"); print_loop(true);})
2167 
2168   CountedLoopReserveKit make_reversable(_phase, _lpt, do_reserve_copy());
2169 
2170   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop after create_reserve_version_of_loop"); print_loop(true);})
2171 
2172   if (do_reserve_copy() && !make_reversable.has_reserved()) {
2173     NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: loop was not reserved correctly, exiting SuperWord");})
2174     return;
2175   }
2176 
2177   for (int i = 0; i < _block.length(); i++) {
2178     Node* n = _block.at(i);
2179     Node_List* p = my_pack(n);
2180     if (p && n == executed_last(p)) {
2181       uint vlen = p->size();
2182       uint vlen_in_bytes = 0;
2183       Node* vn = NULL;
2184       Node* low_adr = p->at(0);
2185       Node* first   = executed_first(p);
2186       if (can_process_post_loop) {
2187         // override vlen with the main loops vector length
2188         vlen = cl->slp_max_unroll();
2189       }
2190       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);})
2191       int   opc = n->Opcode();
2192       if (n->is_Load()) {
2193         Node* ctl = n->in(MemNode::Control);
2194         Node* mem = first->in(MemNode::Memory);
2195         SWPointer p1(n->as_Mem(), this, NULL, false);
2196         // Identify the memory dependency for the new loadVector node by
2197         // walking up through memory chain.
2198         // This is done to give flexibility to the new loadVector node so that
2199         // it can move above independent storeVector nodes.
2200         while (mem->is_StoreVector()) {
2201           SWPointer p2(mem->as_Mem(), this, NULL, false);
2202           int cmp = p1.cmp(p2);
2203           if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
2204             mem = mem->in(MemNode::Memory);
2205           } else {
2206             break; // dependent memory
2207           }
2208         }
2209         Node* adr = low_adr->in(MemNode::Address);
2210         const TypePtr* atyp = n->adr_type();
2211         vn = LoadVectorNode::make(opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n), control_dependency(p));
2212         vlen_in_bytes = vn->as_LoadVector()->memory_size();
2213       } else if (n->is_Store()) {
2214         // Promote value to be stored to vector
2215         Node* val = vector_opd(p, MemNode::ValueIn);
2216         if (val == NULL) {
2217           if (do_reserve_copy()) {
2218             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: val should not be NULL, exiting SuperWord");})
2219             return; //and reverse to backup IG
2220           }
2221           ShouldNotReachHere();
2222         }
2223 
2224         Node* ctl = n->in(MemNode::Control);
2225         Node* mem = first->in(MemNode::Memory);
2226         Node* adr = low_adr->in(MemNode::Address);
2227         const TypePtr* atyp = n->adr_type();
2228         vn = StoreVectorNode::make(opc, ctl, mem, adr, atyp, val, vlen);
2229         vlen_in_bytes = vn->as_StoreVector()->memory_size();
2230       } else if (n->req() == 3 && !is_cmov_pack(p)) {
2231         // Promote operands to vector
2232         Node* in1 = NULL;
2233         bool node_isa_reduction = n->is_reduction();
2234         if (node_isa_reduction) {
2235           // the input to the first reduction operation is retained
2236           in1 = low_adr->in(1);
2237         } else {
2238           in1 = vector_opd(p, 1);
2239           if (in1 == NULL) {
2240             if (do_reserve_copy()) {
2241               NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in1 should not be NULL, exiting SuperWord");})
2242               return; //and reverse to backup IG
2243             }
2244             ShouldNotReachHere();
2245           }
2246         }
2247         Node* in2 = vector_opd(p, 2);
2248         if (in2 == NULL) {
2249           if (do_reserve_copy()) {
2250             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in2 should not be NULL, exiting SuperWord");})
2251             return; //and reverse to backup IG
2252           }
2253           ShouldNotReachHere();
2254         }
2255         if (VectorNode::is_invariant_vector(in1) && (node_isa_reduction == false) && (n->is_Add() || n->is_Mul())) {
2256           // Move invariant vector input into second position to avoid register spilling.
2257           Node* tmp = in1;
2258           in1 = in2;
2259           in2 = tmp;
2260         }
2261         if (node_isa_reduction) {
2262           const Type *arith_type = n->bottom_type();
2263           vn = ReductionNode::make(opc, NULL, in1, in2, arith_type->basic_type());
2264           if (in2->is_Load()) {
2265             vlen_in_bytes = in2->as_LoadVector()->memory_size();
2266           } else {
2267             vlen_in_bytes = in2->as_Vector()->length_in_bytes();
2268           }
2269         } else {
2270           vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2271           vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2272         }
2273       } else if (opc == Op_SqrtD || opc == Op_AbsF || opc == Op_AbsD || opc == Op_NegF || opc == Op_NegD) {
2274         // Promote operand to vector (Sqrt/Abs/Neg are 2 address instructions)
2275         Node* in = vector_opd(p, 1);
2276         vn = VectorNode::make(opc, in, NULL, vlen, velt_basic_type(n));
2277         vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2278       } else if (is_cmov_pack(p)) {
2279         if (can_process_post_loop) {
2280           // do not refactor of flow in post loop context
2281           return;
2282         }
2283         if (!n->is_CMove()) {
2284           continue;
2285         }
2286         // place here CMoveVDNode
2287         NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::output: print before CMove vectorization"); print_loop(false);})
2288         Node* bol = n->in(CMoveNode::Condition);
2289         if (!bol->is_Bool() && bol->Opcode() == Op_ExtractI && bol->req() > 1 ) {
2290           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();})
2291           bol = bol->in(1); //may be ExtractNode
2292         }
2293 
2294         assert(bol->is_Bool(), "should be BoolNode - too late to bail out!");
2295         if (!bol->is_Bool()) {
2296           if (do_reserve_copy()) {
2297             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: expected %d bool node, exiting SuperWord", bol->_idx); bol->dump();})
2298             return; //and reverse to backup IG
2299           }
2300           ShouldNotReachHere();
2301         }
2302 
2303         int cond = (int)bol->as_Bool()->_test._test;
2304         Node* in_cc  = _igvn.intcon(cond);
2305         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created intcon in_cc node %d", in_cc->_idx); in_cc->dump();})
2306         Node* cc = bol->clone();
2307         cc->set_req(1, in_cc);
2308         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created bool cc node %d", cc->_idx); cc->dump();})
2309 
2310         Node* src1 = vector_opd(p, 2); //2=CMoveNode::IfFalse
2311         if (src1 == NULL) {
2312           if (do_reserve_copy()) {
2313             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src1 should not be NULL, exiting SuperWord");})
2314             return; //and reverse to backup IG
2315           }
2316           ShouldNotReachHere();
2317         }
2318         Node* src2 = vector_opd(p, 3); //3=CMoveNode::IfTrue
2319         if (src2 == NULL) {
2320           if (do_reserve_copy()) {
2321             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src2 should not be NULL, exiting SuperWord");})
2322             return; //and reverse to backup IG
2323           }
2324           ShouldNotReachHere();
2325         }
2326         BasicType bt = velt_basic_type(n);
2327         const TypeVect* vt = TypeVect::make(bt, vlen);
2328         vn = new CMoveVDNode(cc, src1, src2, vt);
2329         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created new CMove node %d: ", vn->_idx); vn->dump();})
2330       } else if (opc == Op_FmaD || opc == Op_FmaF) {
2331         // Promote operands to vector
2332         Node* in1 = vector_opd(p, 1);
2333         Node* in2 = vector_opd(p, 2);
2334         Node* in3 = vector_opd(p, 3);
2335         vn = VectorNode::make(opc, in1, in2, in3, vlen, velt_basic_type(n));
2336         vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2337       } else {
2338         if (do_reserve_copy()) {
2339           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: ShouldNotReachHere, exiting SuperWord");})
2340           return; //and reverse to backup IG
2341         }
2342         ShouldNotReachHere();
2343       }
2344 
2345       assert(vn != NULL, "sanity");
2346       if (vn == NULL) {
2347         if (do_reserve_copy()){
2348           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: got NULL node, cannot proceed, exiting SuperWord");})
2349           return; //and reverse to backup IG
2350         }
2351         ShouldNotReachHere();
2352       }
2353 
2354       _block.at_put(i, vn);
2355       _igvn.register_new_node_with_optimizer(vn);
2356       _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
2357       for (uint j = 0; j < p->size(); j++) {
2358         Node* pm = p->at(j);
2359         _igvn.replace_node(pm, vn);
2360       }
2361       _igvn._worklist.push(vn);
2362 
2363       if (can_process_post_loop) {
2364         // first check if the vector size if the maximum vector which we can use on the machine,
2365         // other vector size have reduced values for predicated data mapping.
2366         if (vlen_in_bytes != (uint)MaxVectorSize) {
2367           return;
2368         }
2369       }
2370 
2371       if (vlen_in_bytes > max_vlen_in_bytes) {
2372         max_vlen = vlen;
2373         max_vlen_in_bytes = vlen_in_bytes;
2374       }
2375 #ifdef ASSERT
2376       if (TraceNewVectors) {
2377         tty->print("new Vector node: ");
2378         vn->dump();
2379       }
2380 #endif
2381     }
2382   }//for (int i = 0; i < _block.length(); i++)
2383 
2384   C->set_max_vector_size(max_vlen_in_bytes);
2385 
2386   if (SuperWordLoopUnrollAnalysis) {
2387     if (cl->has_passed_slp()) {
2388       uint slp_max_unroll_factor = cl->slp_max_unroll();
2389       if (slp_max_unroll_factor == max_vlen) {
2390         if (TraceSuperWordLoopUnrollAnalysis) {
2391           tty->print_cr("vector loop(unroll=%d, len=%d)\n", max_vlen, max_vlen_in_bytes*BitsPerByte);
2392         }
2393 
2394         // For atomic unrolled loops which are vector mapped, instigate more unrolling
2395         cl->set_notpassed_slp();
2396         if (cl->is_main_loop()) {
2397           // if vector resources are limited, do not allow additional unrolling, also
2398           // do not unroll more on pure vector loops which were not reduced so that we can
2399           // program the post loop to single iteration execution.
2400           if (FLOATPRESSURE > 8) {
2401             C->set_major_progress();
2402             cl->mark_do_unroll_only();
2403           }
2404         }
2405 
2406         if (do_reserve_copy()) {
2407           cl->mark_loop_vectorized();
2408           if (can_process_post_loop) {
2409             // Now create the difference of trip and limit and use it as our mask index.
2410             // Note: We limited the unroll of the vectorized loop so that
2411             //       only vlen-1 size iterations can remain to be mask programmed.
2412             Node *incr = cl->incr();
2413             SubINode *index = new SubINode(cl->limit(), cl->init_trip());
2414             _igvn.register_new_node_with_optimizer(index);
2415             SetVectMaskINode  *mask = new SetVectMaskINode(_phase->get_ctrl(cl->init_trip()), index);
2416             _igvn.register_new_node_with_optimizer(mask);
2417             // make this a single iteration loop
2418             AddINode *new_incr = new AddINode(incr->in(1), mask);
2419             _igvn.register_new_node_with_optimizer(new_incr);
2420             _phase->set_ctrl(new_incr, _phase->get_ctrl(incr));
2421             _igvn.replace_node(incr, new_incr);
2422             cl->mark_is_multiversioned();
2423             cl->loopexit()->add_flag(Node::Flag_has_vector_mask_set);
2424           }
2425         }
2426       }
2427     }
2428   }
2429 
2430   if (do_reserve_copy()) {
2431     make_reversable.use_new();
2432   }
2433   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("\n Final loop after SuperWord"); print_loop(true);})
2434   return;
2435 }
2436 
2437 //------------------------------vector_opd---------------------------
2438 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
2439 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
2440   Node* p0 = p->at(0);
2441   uint vlen = p->size();
2442   Node* opd = p0->in(opd_idx);
2443   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2444 
2445   if (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop()) {
2446     // override vlen with the main loops vector length
2447     vlen = cl->slp_max_unroll();
2448   }
2449 
2450   if (same_inputs(p, opd_idx)) {
2451     if (opd->is_Vector() || opd->is_LoadVector()) {
2452       assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
2453       if (opd_idx == 2 && VectorNode::is_shift(p0)) {
2454         NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("shift's count can't be vector");})
2455         return NULL;
2456       }
2457       return opd; // input is matching vector
2458     }
2459     if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
2460       Compile* C = _phase->C;
2461       Node* cnt = opd;
2462       // Vector instructions do not mask shift count, do it here.
2463       juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2464       const TypeInt* t = opd->find_int_type();
2465       if (t != NULL && t->is_con()) {
2466         juint shift = t->get_con();
2467         if (shift > mask) { // Unsigned cmp
2468           cnt = ConNode::make(TypeInt::make(shift & mask));
2469         }
2470       } else {
2471         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2472           cnt = ConNode::make(TypeInt::make(mask));
2473           _igvn.register_new_node_with_optimizer(cnt);
2474           cnt = new AndINode(opd, cnt);
2475           _igvn.register_new_node_with_optimizer(cnt);
2476           _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2477         }
2478         assert(opd->bottom_type()->isa_int(), "int type only");
2479         if (!opd->bottom_type()->isa_int()) {
2480           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should be int type only");})
2481           return NULL;
2482         }
2483         // Move non constant shift count into vector register.
2484         cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0));
2485       }
2486       if (cnt != opd) {
2487         _igvn.register_new_node_with_optimizer(cnt);
2488         _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2489       }
2490       return cnt;
2491     }
2492     assert(!opd->is_StoreVector(), "such vector is not expected here");
2493     if (opd->is_StoreVector()) {
2494       NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("StoreVector is not expected here");})
2495       return NULL;
2496     }
2497     // Convert scalar input to vector with the same number of elements as
2498     // p0's vector. Use p0's type because size of operand's container in
2499     // vector should match p0's size regardless operand's size.
2500     const Type* p0_t = velt_type(p0);
2501     VectorNode* vn = VectorNode::scalar2vector(opd, vlen, p0_t);
2502 
2503     _igvn.register_new_node_with_optimizer(vn);
2504     _phase->set_ctrl(vn, _phase->get_ctrl(opd));
2505 #ifdef ASSERT
2506     if (TraceNewVectors) {
2507       tty->print("new Vector node: ");
2508       vn->dump();
2509     }
2510 #endif
2511     return vn;
2512   }
2513 
2514   // Insert pack operation
2515   BasicType bt = velt_basic_type(p0);
2516   PackNode* pk = PackNode::make(opd, vlen, bt);
2517   DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
2518 
2519   for (uint i = 1; i < vlen; i++) {
2520     Node* pi = p->at(i);
2521     Node* in = pi->in(opd_idx);
2522     assert(my_pack(in) == NULL, "Should already have been unpacked");
2523     if (my_pack(in) != NULL) {
2524       NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should already have been unpacked");})
2525       return NULL;
2526     }
2527     assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
2528     pk->add_opd(in);
2529   }
2530   _igvn.register_new_node_with_optimizer(pk);
2531   _phase->set_ctrl(pk, _phase->get_ctrl(opd));
2532 #ifdef ASSERT
2533   if (TraceNewVectors) {
2534     tty->print("new Vector node: ");
2535     pk->dump();
2536   }
2537 #endif
2538   return pk;
2539 }
2540 
2541 //------------------------------insert_extracts---------------------------
2542 // If a use of pack p is not a vector use, then replace the
2543 // use with an extract operation.
2544 void SuperWord::insert_extracts(Node_List* p) {
2545   if (p->at(0)->is_Store()) return;
2546   assert(_n_idx_list.is_empty(), "empty (node,index) list");
2547 
2548   // Inspect each use of each pack member.  For each use that is
2549   // not a vector use, replace the use with an extract operation.
2550 
2551   for (uint i = 0; i < p->size(); i++) {
2552     Node* def = p->at(i);
2553     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2554       Node* use = def->fast_out(j);
2555       for (uint k = 0; k < use->req(); k++) {
2556         Node* n = use->in(k);
2557         if (def == n) {
2558           Node_List* u_pk = my_pack(use);
2559           if ((u_pk == NULL || !is_cmov_pack(u_pk) || use->is_CMove()) && !is_vector_use(use, k)) {
2560               _n_idx_list.push(use, k);
2561           }
2562         }
2563       }
2564     }
2565   }
2566 
2567   while (_n_idx_list.is_nonempty()) {
2568     Node* use = _n_idx_list.node();
2569     int   idx = _n_idx_list.index();
2570     _n_idx_list.pop();
2571     Node* def = use->in(idx);
2572 
2573     if (def->is_reduction()) continue;
2574 
2575     // Insert extract operation
2576     _igvn.hash_delete(def);
2577     int def_pos = alignment(def) / data_size(def);
2578 
2579     Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def));
2580     _igvn.register_new_node_with_optimizer(ex);
2581     _phase->set_ctrl(ex, _phase->get_ctrl(def));
2582     _igvn.replace_input_of(use, idx, ex);
2583     _igvn._worklist.push(def);
2584 
2585     bb_insert_after(ex, bb_idx(def));
2586     set_velt_type(ex, velt_type(def));
2587   }
2588 }
2589 
2590 //------------------------------is_vector_use---------------------------
2591 // Is use->in(u_idx) a vector use?
2592 bool SuperWord::is_vector_use(Node* use, int u_idx) {
2593   Node_List* u_pk = my_pack(use);
2594   if (u_pk == NULL) return false;
2595   if (use->is_reduction()) return true;
2596   Node* def = use->in(u_idx);
2597   Node_List* d_pk = my_pack(def);
2598   if (d_pk == NULL) {
2599     // check for scalar promotion
2600     Node* n = u_pk->at(0)->in(u_idx);
2601     for (uint i = 1; i < u_pk->size(); i++) {
2602       if (u_pk->at(i)->in(u_idx) != n) return false;
2603     }
2604     return true;
2605   }
2606   if (u_pk->size() != d_pk->size())
2607     return false;
2608   for (uint i = 0; i < u_pk->size(); i++) {
2609     Node* ui = u_pk->at(i);
2610     Node* di = d_pk->at(i);
2611     if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
2612       return false;
2613   }
2614   return true;
2615 }
2616 
2617 //------------------------------construct_bb---------------------------
2618 // Construct reverse postorder list of block members
2619 bool SuperWord::construct_bb() {
2620   Node* entry = bb();
2621 
2622   assert(_stk.length() == 0,            "stk is empty");
2623   assert(_block.length() == 0,          "block is empty");
2624   assert(_data_entry.length() == 0,     "data_entry is empty");
2625   assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
2626   assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
2627 
2628   // Find non-control nodes with no inputs from within block,
2629   // create a temporary map from node _idx to bb_idx for use
2630   // by the visited and post_visited sets,
2631   // and count number of nodes in block.
2632   int bb_ct = 0;
2633   for (uint i = 0; i < lpt()->_body.size(); i++) {
2634     Node *n = lpt()->_body.at(i);
2635     set_bb_idx(n, i); // Create a temporary map
2636     if (in_bb(n)) {
2637       if (n->is_LoadStore() || n->is_MergeMem() ||
2638           (n->is_Proj() && !n->as_Proj()->is_CFG())) {
2639         // Bailout if the loop has LoadStore, MergeMem or data Proj
2640         // nodes. Superword optimization does not work with them.
2641         return false;
2642       }
2643       bb_ct++;
2644       if (!n->is_CFG()) {
2645         bool found = false;
2646         for (uint j = 0; j < n->req(); j++) {
2647           Node* def = n->in(j);
2648           if (def && in_bb(def)) {
2649             found = true;
2650             break;
2651           }
2652         }
2653         if (!found) {
2654           assert(n != entry, "can't be entry");
2655           _data_entry.push(n);
2656         }
2657       }
2658     }
2659   }
2660 
2661   // Find memory slices (head and tail)
2662   for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
2663     Node *n = lp()->fast_out(i);
2664     if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
2665       Node* n_tail  = n->in(LoopNode::LoopBackControl);
2666       if (n_tail != n->in(LoopNode::EntryControl)) {
2667         if (!n_tail->is_Mem()) {
2668           assert(n_tail->is_Mem(), "unexpected node for memory slice: %s", n_tail->Name());
2669           return false; // Bailout
2670         }
2671         _mem_slice_head.push(n);
2672         _mem_slice_tail.push(n_tail);
2673       }
2674     }
2675   }
2676 
2677   // Create an RPO list of nodes in block
2678 
2679   visited_clear();
2680   post_visited_clear();
2681 
2682   // Push all non-control nodes with no inputs from within block, then control entry
2683   for (int j = 0; j < _data_entry.length(); j++) {
2684     Node* n = _data_entry.at(j);
2685     visited_set(n);
2686     _stk.push(n);
2687   }
2688   visited_set(entry);
2689   _stk.push(entry);
2690 
2691   // Do a depth first walk over out edges
2692   int rpo_idx = bb_ct - 1;
2693   int size;
2694   int reduction_uses = 0;
2695   while ((size = _stk.length()) > 0) {
2696     Node* n = _stk.top(); // Leave node on stack
2697     if (!visited_test_set(n)) {
2698       // forward arc in graph
2699     } else if (!post_visited_test(n)) {
2700       // cross or back arc
2701       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2702         Node *use = n->fast_out(i);
2703         if (in_bb(use) && !visited_test(use) &&
2704             // Don't go around backedge
2705             (!use->is_Phi() || n == entry)) {
2706           if (use->is_reduction()) {
2707             // First see if we can map the reduction on the given system we are on, then
2708             // make a data entry operation for each reduction we see.
2709             BasicType bt = use->bottom_type()->basic_type();
2710             if (ReductionNode::implemented(use->Opcode(), Matcher::min_vector_size(bt), bt)) {
2711               reduction_uses++;
2712             }
2713           }
2714           _stk.push(use);
2715         }
2716       }
2717       if (_stk.length() == size) {
2718         // There were no additional uses, post visit node now
2719         _stk.pop(); // Remove node from stack
2720         assert(rpo_idx >= 0, "");
2721         _block.at_put_grow(rpo_idx, n);
2722         rpo_idx--;
2723         post_visited_set(n);
2724         assert(rpo_idx >= 0 || _stk.is_empty(), "");
2725       }
2726     } else {
2727       _stk.pop(); // Remove post-visited node from stack
2728     }
2729   }//while
2730 
2731   int ii_current = -1;
2732   unsigned int load_idx = (unsigned int)-1;
2733   _ii_order.clear();
2734   // Create real map of block indices for nodes
2735   for (int j = 0; j < _block.length(); j++) {
2736     Node* n = _block.at(j);
2737     set_bb_idx(n, j);
2738     if (_do_vector_loop && n->is_Load()) {
2739       if (ii_current == -1) {
2740         ii_current = _clone_map.gen(n->_idx);
2741         _ii_order.push(ii_current);
2742         load_idx = _clone_map.idx(n->_idx);
2743       } else if (_clone_map.idx(n->_idx) == load_idx && _clone_map.gen(n->_idx) != ii_current) {
2744         ii_current = _clone_map.gen(n->_idx);
2745         _ii_order.push(ii_current);
2746       }
2747     }
2748   }//for
2749 
2750   // Ensure extra info is allocated.
2751   initialize_bb();
2752 
2753 #ifndef PRODUCT
2754   if (_vector_loop_debug && _ii_order.length() > 0) {
2755     tty->print("SuperWord::construct_bb: List of generations: ");
2756     for (int jj = 0; jj < _ii_order.length(); ++jj) {
2757       tty->print("  %d:%d", jj, _ii_order.at(jj));
2758     }
2759     tty->print_cr(" ");
2760   }
2761   if (TraceSuperWord) {
2762     print_bb();
2763     tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
2764     for (int m = 0; m < _data_entry.length(); m++) {
2765       tty->print("%3d ", m);
2766       _data_entry.at(m)->dump();
2767     }
2768     tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
2769     for (int m = 0; m < _mem_slice_head.length(); m++) {
2770       tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
2771       tty->print("    ");    _mem_slice_tail.at(m)->dump();
2772     }
2773   }
2774 #endif
2775   assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
2776   return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0);
2777 }
2778 
2779 //------------------------------initialize_bb---------------------------
2780 // Initialize per node info
2781 void SuperWord::initialize_bb() {
2782   Node* last = _block.at(_block.length() - 1);
2783   grow_node_info(bb_idx(last));
2784 }
2785 
2786 //------------------------------bb_insert_after---------------------------
2787 // Insert n into block after pos
2788 void SuperWord::bb_insert_after(Node* n, int pos) {
2789   int n_pos = pos + 1;
2790   // Make room
2791   for (int i = _block.length() - 1; i >= n_pos; i--) {
2792     _block.at_put_grow(i+1, _block.at(i));
2793   }
2794   for (int j = _node_info.length() - 1; j >= n_pos; j--) {
2795     _node_info.at_put_grow(j+1, _node_info.at(j));
2796   }
2797   // Set value
2798   _block.at_put_grow(n_pos, n);
2799   _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
2800   // Adjust map from node->_idx to _block index
2801   for (int i = n_pos; i < _block.length(); i++) {
2802     set_bb_idx(_block.at(i), i);
2803   }
2804 }
2805 
2806 //------------------------------compute_max_depth---------------------------
2807 // Compute max depth for expressions from beginning of block
2808 // Use to prune search paths during test for independence.
2809 void SuperWord::compute_max_depth() {
2810   int ct = 0;
2811   bool again;
2812   do {
2813     again = false;
2814     for (int i = 0; i < _block.length(); i++) {
2815       Node* n = _block.at(i);
2816       if (!n->is_Phi()) {
2817         int d_orig = depth(n);
2818         int d_in   = 0;
2819         for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
2820           Node* pred = preds.current();
2821           if (in_bb(pred)) {
2822             d_in = MAX2(d_in, depth(pred));
2823           }
2824         }
2825         if (d_in + 1 != d_orig) {
2826           set_depth(n, d_in + 1);
2827           again = true;
2828         }
2829       }
2830     }
2831     ct++;
2832   } while (again);
2833 
2834   if (TraceSuperWord && Verbose) {
2835     tty->print_cr("compute_max_depth iterated: %d times", ct);
2836   }
2837 }
2838 
2839 //-------------------------compute_vector_element_type-----------------------
2840 // Compute necessary vector element type for expressions
2841 // This propagates backwards a narrower integer type when the
2842 // upper bits of the value are not needed.
2843 // Example:  char a,b,c;  a = b + c;
2844 // Normally the type of the add is integer, but for packed character
2845 // operations the type of the add needs to be char.
2846 void SuperWord::compute_vector_element_type() {
2847   if (TraceSuperWord && Verbose) {
2848     tty->print_cr("\ncompute_velt_type:");
2849   }
2850 
2851   // Initial type
2852   for (int i = 0; i < _block.length(); i++) {
2853     Node* n = _block.at(i);
2854     set_velt_type(n, container_type(n));
2855   }
2856 
2857   // Propagate integer narrowed type backwards through operations
2858   // that don't depend on higher order bits
2859   for (int i = _block.length() - 1; i >= 0; i--) {
2860     Node* n = _block.at(i);
2861     // Only integer types need be examined
2862     const Type* vtn = velt_type(n);
2863     if (vtn->basic_type() == T_INT) {
2864       uint start, end;
2865       VectorNode::vector_operands(n, &start, &end);
2866 
2867       for (uint j = start; j < end; j++) {
2868         Node* in  = n->in(j);
2869         // Don't propagate through a memory
2870         if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
2871             data_size(n) < data_size(in)) {
2872           bool same_type = true;
2873           for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
2874             Node *use = in->fast_out(k);
2875             if (!in_bb(use) || !same_velt_type(use, n)) {
2876               same_type = false;
2877               break;
2878             }
2879           }
2880           if (same_type) {
2881             // For right shifts of small integer types (bool, byte, char, short)
2882             // we need precise information about sign-ness. Only Load nodes have
2883             // this information because Store nodes are the same for signed and
2884             // unsigned values. And any arithmetic operation after a load may
2885             // expand a value to signed Int so such right shifts can't be used
2886             // because vector elements do not have upper bits of Int.
2887             const Type* vt = vtn;
2888             if (VectorNode::is_shift(in)) {
2889               Node* load = in->in(1);
2890               if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
2891                 vt = velt_type(load);
2892               } else if (in->Opcode() != Op_LShiftI) {
2893                 // Widen type to Int to avoid creation of right shift vector
2894                 // (align + data_size(s1) check in stmts_can_pack() will fail).
2895                 // Note, left shifts work regardless type.
2896                 vt = TypeInt::INT;
2897               }
2898             }
2899             set_velt_type(in, vt);
2900           }
2901         }
2902       }
2903     }
2904   }
2905 #ifndef PRODUCT
2906   if (TraceSuperWord && Verbose) {
2907     for (int i = 0; i < _block.length(); i++) {
2908       Node* n = _block.at(i);
2909       velt_type(n)->dump();
2910       tty->print("\t");
2911       n->dump();
2912     }
2913   }
2914 #endif
2915 }
2916 
2917 //------------------------------memory_alignment---------------------------
2918 // Alignment within a vector memory reference
2919 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
2920   #ifndef PRODUCT
2921     if(TraceSuperWord && Verbose) {
2922       tty->print("SuperWord::memory_alignment within a vector memory reference for %d:  ", s->_idx); s->dump();
2923     }
2924   #endif
2925   NOT_PRODUCT(SWPointer::Tracer::Depth ddd(0);)
2926   SWPointer p(s, this, NULL, false);
2927   if (!p.valid()) {
2928     NOT_PRODUCT(if(is_trace_alignment()) tty->print("SWPointer::memory_alignment: SWPointer p invalid, return bottom_align");)
2929     return bottom_align;
2930   }
2931   int vw = vector_width_in_bytes(s);
2932   if (vw < 2) {
2933     NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SWPointer::memory_alignment: vector_width_in_bytes < 2, return bottom_align");)
2934     return bottom_align; // No vectors for this type
2935   }
2936   int offset  = p.offset_in_bytes();
2937   offset     += iv_adjust*p.memory_size();
2938   int off_rem = offset % vw;
2939   int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
2940   if (TraceSuperWord && Verbose) {
2941     tty->print_cr("SWPointer::memory_alignment: off_rem = %d, off_mod = %d", off_rem, off_mod);
2942   }
2943   return off_mod;
2944 }
2945 
2946 //---------------------------container_type---------------------------
2947 // Smallest type containing range of values
2948 const Type* SuperWord::container_type(Node* n) {
2949   if (n->is_Mem()) {
2950     BasicType bt = n->as_Mem()->memory_type();
2951     if (n->is_Store() && (bt == T_CHAR)) {
2952       // Use T_SHORT type instead of T_CHAR for stored values because any
2953       // preceding arithmetic operation extends values to signed Int.
2954       bt = T_SHORT;
2955     }
2956     if (n->Opcode() == Op_LoadUB) {
2957       // Adjust type for unsigned byte loads, it is important for right shifts.
2958       // T_BOOLEAN is used because there is no basic type representing type
2959       // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
2960       // size (one byte) and sign is important.
2961       bt = T_BOOLEAN;
2962     }
2963     return Type::get_const_basic_type(bt);
2964   }
2965   const Type* t = _igvn.type(n);
2966   if (t->basic_type() == T_INT) {
2967     // A narrow type of arithmetic operations will be determined by
2968     // propagating the type of memory operations.
2969     return TypeInt::INT;
2970   }
2971   return t;
2972 }
2973 
2974 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
2975   const Type* vt1 = velt_type(n1);
2976   const Type* vt2 = velt_type(n2);
2977   if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
2978     // Compare vectors element sizes for integer types.
2979     return data_size(n1) == data_size(n2);
2980   }
2981   return vt1 == vt2;
2982 }
2983 
2984 //------------------------------in_packset---------------------------
2985 // Are s1 and s2 in a pack pair and ordered as s1,s2?
2986 bool SuperWord::in_packset(Node* s1, Node* s2) {
2987   for (int i = 0; i < _packset.length(); i++) {
2988     Node_List* p = _packset.at(i);
2989     assert(p->size() == 2, "must be");
2990     if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
2991       return true;
2992     }
2993   }
2994   return false;
2995 }
2996 
2997 //------------------------------in_pack---------------------------
2998 // Is s in pack p?
2999 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
3000   for (uint i = 0; i < p->size(); i++) {
3001     if (p->at(i) == s) {
3002       return p;
3003     }
3004   }
3005   return NULL;
3006 }
3007 
3008 //------------------------------remove_pack_at---------------------------
3009 // Remove the pack at position pos in the packset
3010 void SuperWord::remove_pack_at(int pos) {
3011   Node_List* p = _packset.at(pos);
3012   for (uint i = 0; i < p->size(); i++) {
3013     Node* s = p->at(i);
3014     set_my_pack(s, NULL);
3015   }
3016   _packset.remove_at(pos);
3017 }
3018 
3019 void SuperWord::packset_sort(int n) {
3020   // simple bubble sort so that we capitalize with O(n) when its already sorted
3021   while (n != 0) {
3022     bool swapped = false;
3023     for (int i = 1; i < n; i++) {
3024       Node_List* q_low = _packset.at(i-1);
3025       Node_List* q_i = _packset.at(i);
3026 
3027       // only swap when we find something to swap
3028       if (alignment(q_low->at(0)) > alignment(q_i->at(0))) {
3029         Node_List* t = q_i;
3030         *(_packset.adr_at(i)) = q_low;
3031         *(_packset.adr_at(i-1)) = q_i;
3032         swapped = true;
3033       }
3034     }
3035     if (swapped == false) break;
3036     n--;
3037   }
3038 }
3039 
3040 //------------------------------executed_first---------------------------
3041 // Return the node executed first in pack p.  Uses the RPO block list
3042 // to determine order.
3043 Node* SuperWord::executed_first(Node_List* p) {
3044   Node* n = p->at(0);
3045   int n_rpo = bb_idx(n);
3046   for (uint i = 1; i < p->size(); i++) {
3047     Node* s = p->at(i);
3048     int s_rpo = bb_idx(s);
3049     if (s_rpo < n_rpo) {
3050       n = s;
3051       n_rpo = s_rpo;
3052     }
3053   }
3054   return n;
3055 }
3056 
3057 //------------------------------executed_last---------------------------
3058 // Return the node executed last in pack p.
3059 Node* SuperWord::executed_last(Node_List* p) {
3060   Node* n = p->at(0);
3061   int n_rpo = bb_idx(n);
3062   for (uint i = 1; i < p->size(); i++) {
3063     Node* s = p->at(i);
3064     int s_rpo = bb_idx(s);
3065     if (s_rpo > n_rpo) {
3066       n = s;
3067       n_rpo = s_rpo;
3068     }
3069   }
3070   return n;
3071 }
3072 
3073 LoadNode::ControlDependency SuperWord::control_dependency(Node_List* p) {
3074   LoadNode::ControlDependency dep = LoadNode::DependsOnlyOnTest;
3075   for (uint i = 0; i < p->size(); i++) {
3076     Node* n = p->at(i);
3077     assert(n->is_Load(), "only meaningful for loads");
3078     if (!n->depends_only_on_test()) {
3079       dep = LoadNode::Pinned;
3080     }
3081   }
3082   return dep;
3083 }
3084 
3085 
3086 //----------------------------align_initial_loop_index---------------------------
3087 // Adjust pre-loop limit so that in main loop, a load/store reference
3088 // to align_to_ref will be a position zero in the vector.
3089 //   (iv + k) mod vector_align == 0
3090 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
3091   CountedLoopNode *main_head = lp()->as_CountedLoop();
3092   assert(main_head->is_main_loop(), "");
3093   CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
3094   assert(pre_end != NULL, "we must have a correct pre-loop");
3095   Node *pre_opaq1 = pre_end->limit();
3096   assert(pre_opaq1->Opcode() == Op_Opaque1, "");
3097   Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
3098   Node *lim0 = pre_opaq->in(1);
3099 
3100   // Where we put new limit calculations
3101   Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
3102 
3103   // Ensure the original loop limit is available from the
3104   // pre-loop Opaque1 node.
3105   Node *orig_limit = pre_opaq->original_loop_limit();
3106   assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
3107 
3108   SWPointer align_to_ref_p(align_to_ref, this, NULL, false);
3109   assert(align_to_ref_p.valid(), "sanity");
3110 
3111   // Given:
3112   //     lim0 == original pre loop limit
3113   //     V == v_align (power of 2)
3114   //     invar == extra invariant piece of the address expression
3115   //     e == offset [ +/- invar ]
3116   //
3117   // When reassociating expressions involving '%' the basic rules are:
3118   //     (a - b) % k == 0   =>  a % k == b % k
3119   // and:
3120   //     (a + b) % k == 0   =>  a % k == (k - b) % k
3121   //
3122   // For stride > 0 && scale > 0,
3123   //   Derive the new pre-loop limit "lim" such that the two constraints:
3124   //     (1) lim = lim0 + N           (where N is some positive integer < V)
3125   //     (2) (e + lim) % V == 0
3126   //   are true.
3127   //
3128   //   Substituting (1) into (2),
3129   //     (e + lim0 + N) % V == 0
3130   //   solve for N:
3131   //     N = (V - (e + lim0)) % V
3132   //   substitute back into (1), so that new limit
3133   //     lim = lim0 + (V - (e + lim0)) % V
3134   //
3135   // For stride > 0 && scale < 0
3136   //   Constraints:
3137   //     lim = lim0 + N
3138   //     (e - lim) % V == 0
3139   //   Solving for lim:
3140   //     (e - lim0 - N) % V == 0
3141   //     N = (e - lim0) % V
3142   //     lim = lim0 + (e - lim0) % V
3143   //
3144   // For stride < 0 && scale > 0
3145   //   Constraints:
3146   //     lim = lim0 - N
3147   //     (e + lim) % V == 0
3148   //   Solving for lim:
3149   //     (e + lim0 - N) % V == 0
3150   //     N = (e + lim0) % V
3151   //     lim = lim0 - (e + lim0) % V
3152   //
3153   // For stride < 0 && scale < 0
3154   //   Constraints:
3155   //     lim = lim0 - N
3156   //     (e - lim) % V == 0
3157   //   Solving for lim:
3158   //     (e - lim0 + N) % V == 0
3159   //     N = (V - (e - lim0)) % V
3160   //     lim = lim0 - (V - (e - lim0)) % V
3161 
3162   int vw = vector_width_in_bytes(align_to_ref);
3163   int stride   = iv_stride();
3164   int scale    = align_to_ref_p.scale_in_bytes();
3165   int elt_size = align_to_ref_p.memory_size();
3166   int v_align  = vw / elt_size;
3167   assert(v_align > 1, "sanity");
3168   int offset   = align_to_ref_p.offset_in_bytes() / elt_size;
3169   Node *offsn  = _igvn.intcon(offset);
3170 
3171   Node *e = offsn;
3172   if (align_to_ref_p.invar() != NULL) {
3173     // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
3174     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3175     Node* aref     = new URShiftINode(align_to_ref_p.invar(), log2_elt);
3176     _igvn.register_new_node_with_optimizer(aref);
3177     _phase->set_ctrl(aref, pre_ctrl);
3178     if (align_to_ref_p.negate_invar()) {
3179       e = new SubINode(e, aref);
3180     } else {
3181       e = new AddINode(e, aref);
3182     }
3183     _igvn.register_new_node_with_optimizer(e);
3184     _phase->set_ctrl(e, pre_ctrl);
3185   }
3186   if (vw > ObjectAlignmentInBytes) {
3187     // incorporate base e +/- base && Mask >>> log2(elt)
3188     Node* xbase = new CastP2XNode(NULL, align_to_ref_p.base());
3189     _igvn.register_new_node_with_optimizer(xbase);
3190 #ifdef _LP64
3191     xbase  = new ConvL2INode(xbase);
3192     _igvn.register_new_node_with_optimizer(xbase);
3193 #endif
3194     Node* mask = _igvn.intcon(vw-1);
3195     Node* masked_xbase  = new AndINode(xbase, mask);
3196     _igvn.register_new_node_with_optimizer(masked_xbase);
3197     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3198     Node* bref     = new URShiftINode(masked_xbase, log2_elt);
3199     _igvn.register_new_node_with_optimizer(bref);
3200     _phase->set_ctrl(bref, pre_ctrl);
3201     e = new AddINode(e, bref);
3202     _igvn.register_new_node_with_optimizer(e);
3203     _phase->set_ctrl(e, pre_ctrl);
3204   }
3205 
3206   // compute e +/- lim0
3207   if (scale < 0) {
3208     e = new SubINode(e, lim0);
3209   } else {
3210     e = new AddINode(e, lim0);
3211   }
3212   _igvn.register_new_node_with_optimizer(e);
3213   _phase->set_ctrl(e, pre_ctrl);
3214 
3215   if (stride * scale > 0) {
3216     // compute V - (e +/- lim0)
3217     Node* va  = _igvn.intcon(v_align);
3218     e = new SubINode(va, e);
3219     _igvn.register_new_node_with_optimizer(e);
3220     _phase->set_ctrl(e, pre_ctrl);
3221   }
3222   // compute N = (exp) % V
3223   Node* va_msk = _igvn.intcon(v_align - 1);
3224   Node* N = new AndINode(e, va_msk);
3225   _igvn.register_new_node_with_optimizer(N);
3226   _phase->set_ctrl(N, pre_ctrl);
3227 
3228   //   substitute back into (1), so that new limit
3229   //     lim = lim0 + N
3230   Node* lim;
3231   if (stride < 0) {
3232     lim = new SubINode(lim0, N);
3233   } else {
3234     lim = new AddINode(lim0, N);
3235   }
3236   _igvn.register_new_node_with_optimizer(lim);
3237   _phase->set_ctrl(lim, pre_ctrl);
3238   Node* constrained =
3239     (stride > 0) ? (Node*) new MinINode(lim, orig_limit)
3240                  : (Node*) new MaxINode(lim, orig_limit);
3241   _igvn.register_new_node_with_optimizer(constrained);
3242   _phase->set_ctrl(constrained, pre_ctrl);
3243   _igvn.replace_input_of(pre_opaq, 1, constrained);
3244 }
3245 
3246 //----------------------------get_pre_loop_end---------------------------
3247 // Find pre loop end from main loop.  Returns null if none.
3248 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode* cl) {
3249   // The loop cannot be optimized if the graph shape at
3250   // the loop entry is inappropriate.
3251   if (!PhaseIdealLoop::is_canonical_loop_entry(cl)) {
3252     return NULL;
3253   }
3254 
3255   Node* p_f = cl->in(LoopNode::EntryControl)->in(0)->in(0);
3256   if (!p_f->is_IfFalse()) return NULL;
3257   if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
3258   CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
3259   CountedLoopNode* loop_node = pre_end->loopnode();
3260   if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
3261   return pre_end;
3262 }
3263 
3264 //------------------------------init---------------------------
3265 void SuperWord::init() {
3266   _dg.init();
3267   _packset.clear();
3268   _disjoint_ptrs.clear();
3269   _block.clear();
3270   _post_block.clear();
3271   _data_entry.clear();
3272   _mem_slice_head.clear();
3273   _mem_slice_tail.clear();
3274   _iteration_first.clear();
3275   _iteration_last.clear();
3276   _node_info.clear();
3277   _align_to_ref = NULL;
3278   _lpt = NULL;
3279   _lp = NULL;
3280   _bb = NULL;
3281   _iv = NULL;
3282   _race_possible = 0;
3283   _early_return = false;
3284   _num_work_vecs = 0;
3285   _num_reductions = 0;
3286 }
3287 
3288 //------------------------------restart---------------------------
3289 void SuperWord::restart() {
3290   _dg.init();
3291   _packset.clear();
3292   _disjoint_ptrs.clear();
3293   _block.clear();
3294   _post_block.clear();
3295   _data_entry.clear();
3296   _mem_slice_head.clear();
3297   _mem_slice_tail.clear();
3298   _node_info.clear();
3299 }
3300 
3301 //------------------------------print_packset---------------------------
3302 void SuperWord::print_packset() {
3303 #ifndef PRODUCT
3304   tty->print_cr("packset");
3305   for (int i = 0; i < _packset.length(); i++) {
3306     tty->print_cr("Pack: %d", i);
3307     Node_List* p = _packset.at(i);
3308     print_pack(p);
3309   }
3310 #endif
3311 }
3312 
3313 //------------------------------print_pack---------------------------
3314 void SuperWord::print_pack(Node_List* p) {
3315   for (uint i = 0; i < p->size(); i++) {
3316     print_stmt(p->at(i));
3317   }
3318 }
3319 
3320 //------------------------------print_bb---------------------------
3321 void SuperWord::print_bb() {
3322 #ifndef PRODUCT
3323   tty->print_cr("\nBlock");
3324   for (int i = 0; i < _block.length(); i++) {
3325     Node* n = _block.at(i);
3326     tty->print("%d ", i);
3327     if (n) {
3328       n->dump();
3329     }
3330   }
3331 #endif
3332 }
3333 
3334 //------------------------------print_stmt---------------------------
3335 void SuperWord::print_stmt(Node* s) {
3336 #ifndef PRODUCT
3337   tty->print(" align: %d \t", alignment(s));
3338   s->dump();
3339 #endif
3340 }
3341 
3342 //------------------------------blank---------------------------
3343 char* SuperWord::blank(uint depth) {
3344   static char blanks[101];
3345   assert(depth < 101, "too deep");
3346   for (uint i = 0; i < depth; i++) blanks[i] = ' ';
3347   blanks[depth] = '\0';
3348   return blanks;
3349 }
3350 
3351 
3352 //==============================SWPointer===========================
3353 #ifndef PRODUCT
3354 int SWPointer::Tracer::_depth = 0;
3355 #endif
3356 //----------------------------SWPointer------------------------
3357 SWPointer::SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only) :
3358   _mem(mem), _slp(slp),  _base(NULL),  _adr(NULL),
3359   _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3360   _nstack(nstack), _analyze_only(analyze_only),
3361   _stack_idx(0)
3362 #ifndef PRODUCT
3363   , _tracer(slp)
3364 #endif
3365 {
3366   NOT_PRODUCT(_tracer.ctor_1(mem);)
3367 
3368   Node* adr = mem->in(MemNode::Address);
3369   if (!adr->is_AddP()) {
3370     assert(!valid(), "too complex");
3371     return;
3372   }
3373   // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
3374   Node* base = adr->in(AddPNode::Base);
3375   // The base address should be loop invariant
3376   if (!invariant(base)) {
3377     assert(!valid(), "base address is loop variant");
3378     return;
3379   }
3380   //unsafe reference could not be aligned appropriately without runtime checking
3381   if (base == NULL || base->bottom_type() == Type::TOP) {
3382     assert(!valid(), "unsafe access");
3383     return;
3384   }
3385 
3386   NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.store_depth();)
3387   NOT_PRODUCT(_tracer.ctor_2(adr);)
3388 
3389   int i;
3390   for (i = 0; i < 3; i++) {
3391     NOT_PRODUCT(_tracer.ctor_3(adr, i);)
3392 
3393     if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
3394       assert(!valid(), "too complex");
3395       return;
3396     }
3397     adr = adr->in(AddPNode::Address);
3398     NOT_PRODUCT(_tracer.ctor_4(adr, i);)
3399 
3400     if (base == adr || !adr->is_AddP()) {
3401       NOT_PRODUCT(_tracer.ctor_5(adr, base, i);)
3402       break; // stop looking at addp's
3403     }
3404   }
3405   NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.restore_depth();)
3406   NOT_PRODUCT(_tracer.ctor_6(mem);)
3407 
3408   _base = base;
3409   _adr  = adr;
3410   assert(valid(), "Usable");
3411 }
3412 
3413 // Following is used to create a temporary object during
3414 // the pattern match of an address expression.
3415 SWPointer::SWPointer(SWPointer* p) :
3416   _mem(p->_mem), _slp(p->_slp),  _base(NULL),  _adr(NULL),
3417   _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3418   _nstack(p->_nstack), _analyze_only(p->_analyze_only),
3419   _stack_idx(p->_stack_idx)
3420   #ifndef PRODUCT
3421   , _tracer(p->_slp)
3422   #endif
3423 {}
3424 
3425 
3426 bool SWPointer::invariant(Node* n) {
3427   NOT_PRODUCT(Tracer::Depth dd;)
3428   Node *n_c = phase()->get_ctrl(n);
3429   NOT_PRODUCT(_tracer.invariant_1(n, n_c);)
3430   return !lpt()->is_member(phase()->get_loop(n_c));
3431 }
3432 //------------------------scaled_iv_plus_offset--------------------
3433 // Match: k*iv + offset
3434 // where: k is a constant that maybe zero, and
3435 //        offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
3436 bool SWPointer::scaled_iv_plus_offset(Node* n) {
3437   NOT_PRODUCT(Tracer::Depth ddd;)
3438   NOT_PRODUCT(_tracer.scaled_iv_plus_offset_1(n);)
3439 
3440   if (scaled_iv(n)) {
3441     NOT_PRODUCT(_tracer.scaled_iv_plus_offset_2(n);)
3442     return true;
3443   }
3444 
3445   if (offset_plus_k(n)) {
3446     NOT_PRODUCT(_tracer.scaled_iv_plus_offset_3(n);)
3447     return true;
3448   }
3449 
3450   int opc = n->Opcode();
3451   if (opc == Op_AddI) {
3452     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
3453       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_4(n);)
3454       return true;
3455     }
3456     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3457       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_5(n);)
3458       return true;
3459     }
3460   } else if (opc == Op_SubI) {
3461     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
3462       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_6(n);)
3463       return true;
3464     }
3465     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3466       _scale *= -1;
3467       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_7(n);)
3468       return true;
3469     }
3470   }
3471 
3472   NOT_PRODUCT(_tracer.scaled_iv_plus_offset_8(n);)
3473   return false;
3474 }
3475 
3476 //----------------------------scaled_iv------------------------
3477 // Match: k*iv where k is a constant that's not zero
3478 bool SWPointer::scaled_iv(Node* n) {
3479   NOT_PRODUCT(Tracer::Depth ddd;)
3480   NOT_PRODUCT(_tracer.scaled_iv_1(n);)
3481 
3482   if (_scale != 0) { // already found a scale
3483     NOT_PRODUCT(_tracer.scaled_iv_2(n, _scale);)
3484     return false;
3485   }
3486 
3487   if (n == iv()) {
3488     _scale = 1;
3489     NOT_PRODUCT(_tracer.scaled_iv_3(n, _scale);)
3490     return true;
3491   }
3492   if (_analyze_only && (invariant(n) == false)) {
3493     _nstack->push(n, _stack_idx++);
3494   }
3495 
3496   int opc = n->Opcode();
3497   if (opc == Op_MulI) {
3498     if (n->in(1) == iv() && n->in(2)->is_Con()) {
3499       _scale = n->in(2)->get_int();
3500       NOT_PRODUCT(_tracer.scaled_iv_4(n, _scale);)
3501       return true;
3502     } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
3503       _scale = n->in(1)->get_int();
3504       NOT_PRODUCT(_tracer.scaled_iv_5(n, _scale);)
3505       return true;
3506     }
3507   } else if (opc == Op_LShiftI) {
3508     if (n->in(1) == iv() && n->in(2)->is_Con()) {
3509       _scale = 1 << n->in(2)->get_int();
3510       NOT_PRODUCT(_tracer.scaled_iv_6(n, _scale);)
3511       return true;
3512     }
3513   } else if (opc == Op_ConvI2L) {
3514     if (n->in(1)->Opcode() == Op_CastII &&
3515         n->in(1)->as_CastII()->has_range_check()) {
3516       // Skip range check dependent CastII nodes
3517       n = n->in(1);
3518     }
3519     if (scaled_iv_plus_offset(n->in(1))) {
3520       NOT_PRODUCT(_tracer.scaled_iv_7(n);)
3521       return true;
3522     }
3523   } else if (opc == Op_LShiftL) {
3524     if (!has_iv() && _invar == NULL) {
3525       // Need to preserve the current _offset value, so
3526       // create a temporary object for this expression subtree.
3527       // Hacky, so should re-engineer the address pattern match.
3528       NOT_PRODUCT(Tracer::Depth dddd;)
3529       SWPointer tmp(this);
3530       NOT_PRODUCT(_tracer.scaled_iv_8(n, &tmp);)
3531 
3532       if (tmp.scaled_iv_plus_offset(n->in(1))) {
3533         if (tmp._invar == NULL || _slp->do_vector_loop()) {
3534           int mult = 1 << n->in(2)->get_int();
3535           _scale   = tmp._scale  * mult;
3536           _offset += tmp._offset * mult;
3537           NOT_PRODUCT(_tracer.scaled_iv_9(n, _scale, _offset, mult);)
3538           return true;
3539         }
3540       }
3541     }
3542   }
3543   NOT_PRODUCT(_tracer.scaled_iv_10(n);)
3544   return false;
3545 }
3546 
3547 //----------------------------offset_plus_k------------------------
3548 // Match: offset is (k [+/- invariant])
3549 // where k maybe zero and invariant is optional, but not both.
3550 bool SWPointer::offset_plus_k(Node* n, bool negate) {
3551   NOT_PRODUCT(Tracer::Depth ddd;)
3552   NOT_PRODUCT(_tracer.offset_plus_k_1(n);)
3553 
3554   int opc = n->Opcode();
3555   if (opc == Op_ConI) {
3556     _offset += negate ? -(n->get_int()) : n->get_int();
3557     NOT_PRODUCT(_tracer.offset_plus_k_2(n, _offset);)
3558     return true;
3559   } else if (opc == Op_ConL) {
3560     // Okay if value fits into an int
3561     const TypeLong* t = n->find_long_type();
3562     if (t->higher_equal(TypeLong::INT)) {
3563       jlong loff = n->get_long();
3564       jint  off  = (jint)loff;
3565       _offset += negate ? -off : loff;
3566       NOT_PRODUCT(_tracer.offset_plus_k_3(n, _offset);)
3567       return true;
3568     }
3569     NOT_PRODUCT(_tracer.offset_plus_k_4(n);)
3570     return false;
3571   }
3572   if (_invar != NULL) { // already has an invariant
3573     NOT_PRODUCT(_tracer.offset_plus_k_5(n, _invar);)
3574     return false;
3575   }
3576 
3577   if (_analyze_only && (invariant(n) == false)) {
3578     _nstack->push(n, _stack_idx++);
3579   }
3580   if (opc == Op_AddI) {
3581     if (n->in(2)->is_Con() && invariant(n->in(1))) {
3582       _negate_invar = negate;
3583       _invar = n->in(1);
3584       _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3585       NOT_PRODUCT(_tracer.offset_plus_k_6(n, _invar, _negate_invar, _offset);)
3586       return true;
3587     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3588       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3589       _negate_invar = negate;
3590       _invar = n->in(2);
3591       NOT_PRODUCT(_tracer.offset_plus_k_7(n, _invar, _negate_invar, _offset);)
3592       return true;
3593     }
3594   }
3595   if (opc == Op_SubI) {
3596     if (n->in(2)->is_Con() && invariant(n->in(1))) {
3597       _negate_invar = negate;
3598       _invar = n->in(1);
3599       _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3600       NOT_PRODUCT(_tracer.offset_plus_k_8(n, _invar, _negate_invar, _offset);)
3601       return true;
3602     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3603       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3604       _negate_invar = !negate;
3605       _invar = n->in(2);
3606       NOT_PRODUCT(_tracer.offset_plus_k_9(n, _invar, _negate_invar, _offset);)
3607       return true;
3608     }
3609   }
3610   if (invariant(n)) {
3611     if (opc == Op_ConvI2L) {
3612       n = n->in(1);
3613       if (n->Opcode() == Op_CastII &&
3614           n->as_CastII()->has_range_check()) {
3615         // Skip range check dependent CastII nodes
3616         assert(invariant(n), "sanity");
3617         n = n->in(1);
3618       }
3619     }
3620     if (n->bottom_type()->isa_int()) {
3621       _negate_invar = negate;
3622       _invar = n;
3623       NOT_PRODUCT(_tracer.offset_plus_k_10(n, _invar, _negate_invar, _offset);)
3624       return true;
3625     }
3626   }
3627 
3628   NOT_PRODUCT(_tracer.offset_plus_k_11(n);)
3629   return false;
3630 }
3631 
3632 //----------------------------print------------------------
3633 void SWPointer::print() {
3634 #ifndef PRODUCT
3635   tty->print("base: %d  adr: %d  scale: %d  offset: %d  invar: %c%d\n",
3636              _base != NULL ? _base->_idx : 0,
3637              _adr  != NULL ? _adr->_idx  : 0,
3638              _scale, _offset,
3639              _negate_invar?'-':'+',
3640              _invar != NULL ? _invar->_idx : 0);
3641 #endif
3642 }
3643 
3644 //----------------------------tracing------------------------
3645 #ifndef PRODUCT
3646 void SWPointer::Tracer::print_depth() {
3647   for (int ii = 0; ii<_depth; ++ii) tty->print("  ");
3648 }
3649 
3650 void SWPointer::Tracer::ctor_1 (Node* mem) {
3651   if(_slp->is_trace_alignment()) {
3652     print_depth(); tty->print(" %d SWPointer::SWPointer: start alignment analysis", mem->_idx); mem->dump();
3653   }
3654 }
3655 
3656 void SWPointer::Tracer::ctor_2(Node* adr) {
3657   if(_slp->is_trace_alignment()) {
3658     //store_depth();
3659     inc_depth();
3660     print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: ", adr->_idx); adr->dump();
3661     inc_depth();
3662     print_depth(); tty->print(" %d (base) SWPointer::SWPointer: ", adr->in(AddPNode::Base)->_idx); adr->in(AddPNode::Base)->dump();
3663   }
3664 }
3665 
3666 void SWPointer::Tracer::ctor_3(Node* adr, int i) {
3667   if(_slp->is_trace_alignment()) {
3668     inc_depth();
3669     Node* offset = adr->in(AddPNode::Offset);
3670     print_depth(); tty->print(" %d (offset) SWPointer::SWPointer: i = %d: ", offset->_idx, i); offset->dump();
3671   }
3672 }
3673 
3674 void SWPointer::Tracer::ctor_4(Node* adr, int i) {
3675   if(_slp->is_trace_alignment()) {
3676     inc_depth();
3677     print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: i = %d: ", adr->_idx, i); adr->dump();
3678   }
3679 }
3680 
3681 void SWPointer::Tracer::ctor_5(Node* adr, Node* base, int i) {
3682   if(_slp->is_trace_alignment()) {
3683     inc_depth();
3684     if (base == adr) {
3685       print_depth(); tty->print_cr("  \\ %d (adr) == %d (base) SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, base->_idx, i);
3686     } else if (!adr->is_AddP()) {
3687       print_depth(); tty->print_cr("  \\ %d (adr) is NOT Addp SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, i);
3688     }
3689   }
3690 }
3691 
3692 void SWPointer::Tracer::ctor_6(Node* mem) {
3693   if(_slp->is_trace_alignment()) {
3694     //restore_depth();
3695     print_depth(); tty->print_cr(" %d (adr) SWPointer::SWPointer: stop analysis", mem->_idx);
3696   }
3697 }
3698 
3699 void SWPointer::Tracer::invariant_1(Node *n, Node *n_c) {
3700   if (_slp->do_vector_loop() && _slp->is_debug() && _slp->_lpt->is_member(_slp->_phase->get_loop(n_c)) != (int)_slp->in_bb(n)) {
3701     int is_member =  _slp->_lpt->is_member(_slp->_phase->get_loop(n_c));
3702     int in_bb     =  _slp->in_bb(n);
3703     print_depth(); tty->print("  \\ ");  tty->print_cr(" %d SWPointer::invariant  conditions differ: n_c %d", n->_idx, n_c->_idx);
3704     print_depth(); tty->print("  \\ ");  tty->print_cr("is_member %d, in_bb %d", is_member, in_bb);
3705     print_depth(); tty->print("  \\ ");  n->dump();
3706     print_depth(); tty->print("  \\ ");  n_c->dump();
3707   }
3708 }
3709 
3710 void SWPointer::Tracer::scaled_iv_plus_offset_1(Node* n) {
3711   if(_slp->is_trace_alignment()) {
3712     print_depth(); tty->print(" %d SWPointer::scaled_iv_plus_offset testing node: ", n->_idx);
3713     n->dump();
3714   }
3715 }
3716 
3717 void SWPointer::Tracer::scaled_iv_plus_offset_2(Node* n) {
3718   if(_slp->is_trace_alignment()) {
3719     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
3720   }
3721 }
3722 
3723 void SWPointer::Tracer::scaled_iv_plus_offset_3(Node* n) {
3724   if(_slp->is_trace_alignment()) {
3725     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
3726   }
3727 }
3728 
3729 void SWPointer::Tracer::scaled_iv_plus_offset_4(Node* n) {
3730   if(_slp->is_trace_alignment()) {
3731     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
3732     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
3733     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
3734   }
3735 }
3736 
3737 void SWPointer::Tracer::scaled_iv_plus_offset_5(Node* n) {
3738   if(_slp->is_trace_alignment()) {
3739     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
3740     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
3741     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
3742   }
3743 }
3744 
3745 void SWPointer::Tracer::scaled_iv_plus_offset_6(Node* n) {
3746   if(_slp->is_trace_alignment()) {
3747     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
3748     print_depth(); tty->print("  \\  %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
3749     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
3750   }
3751 }
3752 
3753 void SWPointer::Tracer::scaled_iv_plus_offset_7(Node* n) {
3754   if(_slp->is_trace_alignment()) {
3755     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
3756     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
3757     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
3758   }
3759 }
3760 
3761 void SWPointer::Tracer::scaled_iv_plus_offset_8(Node* n) {
3762   if(_slp->is_trace_alignment()) {
3763     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: FAILED", n->_idx);
3764   }
3765 }
3766 
3767 void SWPointer::Tracer::scaled_iv_1(Node* n) {
3768   if(_slp->is_trace_alignment()) {
3769     print_depth(); tty->print(" %d SWPointer::scaled_iv: testing node: ", n->_idx); n->dump();
3770   }
3771 }
3772 
3773 void SWPointer::Tracer::scaled_iv_2(Node* n, int scale) {
3774   if(_slp->is_trace_alignment()) {
3775     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED since another _scale has been detected before", n->_idx);
3776     print_depth(); tty->print_cr("  \\ SWPointer::scaled_iv: _scale (%d) != 0", scale);
3777   }
3778 }
3779 
3780 void SWPointer::Tracer::scaled_iv_3(Node* n, int scale) {
3781   if(_slp->is_trace_alignment()) {
3782     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: is iv, setting _scale = %d", n->_idx, scale);
3783   }
3784 }
3785 
3786 void SWPointer::Tracer::scaled_iv_4(Node* n, int scale) {
3787   if(_slp->is_trace_alignment()) {
3788     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
3789     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
3790     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3791   }
3792 }
3793 
3794 void SWPointer::Tracer::scaled_iv_5(Node* n, int scale) {
3795   if(_slp->is_trace_alignment()) {
3796     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
3797     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is iv: ", n->in(2)->_idx); n->in(2)->dump();
3798     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3799   }
3800 }
3801 
3802 void SWPointer::Tracer::scaled_iv_6(Node* n, int scale) {
3803   if(_slp->is_trace_alignment()) {
3804     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftI PASSED, setting _scale = %d", n->_idx, scale);
3805     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
3806     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3807   }
3808 }
3809 
3810 void SWPointer::Tracer::scaled_iv_7(Node* n) {
3811   if(_slp->is_trace_alignment()) {
3812     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_ConvI2L PASSED", n->_idx);
3813     print_depth(); tty->print_cr("  \\ SWPointer::scaled_iv: in(1) %d is scaled_iv_plus_offset: ", n->in(1)->_idx);
3814     inc_depth(); inc_depth();
3815     print_depth(); n->in(1)->dump();
3816     dec_depth(); dec_depth();
3817   }
3818 }
3819 
3820 void SWPointer::Tracer::scaled_iv_8(Node* n, SWPointer* tmp) {
3821   if(_slp->is_trace_alignment()) {
3822     print_depth(); tty->print(" %d SWPointer::scaled_iv: Op_LShiftL, creating tmp SWPointer: ", n->_idx); tmp->print();
3823   }
3824 }
3825 
3826 void SWPointer::Tracer::scaled_iv_9(Node* n, int scale, int _offset, int mult) {
3827   if(_slp->is_trace_alignment()) {
3828     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftL PASSED, setting _scale = %d, _offset = %d", n->_idx, scale, _offset);
3829     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",
3830     n->in(1)->_idx, n->in(2)->_idx, mult, scale, _offset);
3831     inc_depth(); inc_depth();
3832     print_depth(); n->in(1)->dump();
3833     print_depth(); n->in(2)->dump();
3834     dec_depth(); dec_depth();
3835   }
3836 }
3837 
3838 void SWPointer::Tracer::scaled_iv_10(Node* n) {
3839   if(_slp->is_trace_alignment()) {
3840     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED", n->_idx);
3841   }
3842 }
3843 
3844 void SWPointer::Tracer::offset_plus_k_1(Node* n) {
3845   if(_slp->is_trace_alignment()) {
3846     print_depth(); tty->print(" %d SWPointer::offset_plus_k: testing node: ", n->_idx); n->dump();
3847   }
3848 }
3849 
3850 void SWPointer::Tracer::offset_plus_k_2(Node* n, int _offset) {
3851   if(_slp->is_trace_alignment()) {
3852     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConI PASSED, setting _offset = %d", n->_idx, _offset);
3853   }
3854 }
3855 
3856 void SWPointer::Tracer::offset_plus_k_3(Node* n, int _offset) {
3857   if(_slp->is_trace_alignment()) {
3858     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConL PASSED, setting _offset = %d", n->_idx, _offset);
3859   }
3860 }
3861 
3862 void SWPointer::Tracer::offset_plus_k_4(Node* n) {
3863   if(_slp->is_trace_alignment()) {
3864     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
3865     print_depth(); tty->print_cr("  \\ " JLONG_FORMAT " SWPointer::offset_plus_k: Op_ConL FAILED, k is too big", n->get_long());
3866   }
3867 }
3868 
3869 void SWPointer::Tracer::offset_plus_k_5(Node* n, Node* _invar) {
3870   if(_slp->is_trace_alignment()) {
3871     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED since another invariant has been detected before", n->_idx);
3872     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: _invar != NULL: ", _invar->_idx); _invar->dump();
3873   }
3874 }
3875 
3876 void SWPointer::Tracer::offset_plus_k_6(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3877   if(_slp->is_trace_alignment()) {
3878     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
3879     n->_idx, _negate_invar, _invar->_idx, _offset);
3880     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3881     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
3882   }
3883 }
3884 
3885 void SWPointer::Tracer::offset_plus_k_7(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3886   if(_slp->is_trace_alignment()) {
3887     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
3888     n->_idx, _negate_invar, _invar->_idx, _offset);
3889     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3890     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
3891   }
3892 }
3893 
3894 void SWPointer::Tracer::offset_plus_k_8(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3895   if(_slp->is_trace_alignment()) {
3896     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_SubI is PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
3897     n->_idx, _negate_invar, _invar->_idx, _offset);
3898     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3899     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
3900   }
3901 }
3902 
3903 void SWPointer::Tracer::offset_plus_k_9(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3904   if(_slp->is_trace_alignment()) {
3905     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);
3906     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3907     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
3908   }
3909 }
3910 
3911 void SWPointer::Tracer::offset_plus_k_10(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3912   if(_slp->is_trace_alignment()) {
3913     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);
3914     print_depth(); tty->print_cr("  \\ %d SWPointer::offset_plus_k: is invariant", n->_idx);
3915   }
3916 }
3917 
3918 void SWPointer::Tracer::offset_plus_k_11(Node* n) {
3919   if(_slp->is_trace_alignment()) {
3920     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
3921   }
3922 }
3923 
3924 #endif
3925 // ========================= OrderedPair =====================
3926 
3927 const OrderedPair OrderedPair::initial;
3928 
3929 // ========================= SWNodeInfo =====================
3930 
3931 const SWNodeInfo SWNodeInfo::initial;
3932 
3933 
3934 // ============================ DepGraph ===========================
3935 
3936 //------------------------------make_node---------------------------
3937 // Make a new dependence graph node for an ideal node.
3938 DepMem* DepGraph::make_node(Node* node) {
3939   DepMem* m = new (_arena) DepMem(node);
3940   if (node != NULL) {
3941     assert(_map.at_grow(node->_idx) == NULL, "one init only");
3942     _map.at_put_grow(node->_idx, m);
3943   }
3944   return m;
3945 }
3946 
3947 //------------------------------make_edge---------------------------
3948 // Make a new dependence graph edge from dpred -> dsucc
3949 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
3950   DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
3951   dpred->set_out_head(e);
3952   dsucc->set_in_head(e);
3953   return e;
3954 }
3955 
3956 // ========================== DepMem ========================
3957 
3958 //------------------------------in_cnt---------------------------
3959 int DepMem::in_cnt() {
3960   int ct = 0;
3961   for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
3962   return ct;
3963 }
3964 
3965 //------------------------------out_cnt---------------------------
3966 int DepMem::out_cnt() {
3967   int ct = 0;
3968   for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
3969   return ct;
3970 }
3971 
3972 //------------------------------print-----------------------------
3973 void DepMem::print() {
3974 #ifndef PRODUCT
3975   tty->print("  DepNode %d (", _node->_idx);
3976   for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
3977     Node* pred = p->pred()->node();
3978     tty->print(" %d", pred != NULL ? pred->_idx : 0);
3979   }
3980   tty->print(") [");
3981   for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
3982     Node* succ = s->succ()->node();
3983     tty->print(" %d", succ != NULL ? succ->_idx : 0);
3984   }
3985   tty->print_cr(" ]");
3986 #endif
3987 }
3988 
3989 // =========================== DepEdge =========================
3990 
3991 //------------------------------DepPreds---------------------------
3992 void DepEdge::print() {
3993 #ifndef PRODUCT
3994   tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
3995 #endif
3996 }
3997 
3998 // =========================== DepPreds =========================
3999 // Iterator over predecessor edges in the dependence graph.
4000 
4001 //------------------------------DepPreds---------------------------
4002 DepPreds::DepPreds(Node* n, DepGraph& dg) {
4003   _n = n;
4004   _done = false;
4005   if (_n->is_Store() || _n->is_Load()) {
4006     _next_idx = MemNode::Address;
4007     _end_idx  = n->req();
4008     _dep_next = dg.dep(_n)->in_head();
4009   } else if (_n->is_Mem()) {
4010     _next_idx = 0;
4011     _end_idx  = 0;
4012     _dep_next = dg.dep(_n)->in_head();
4013   } else {
4014     _next_idx = 1;
4015     _end_idx  = _n->req();
4016     _dep_next = NULL;
4017   }
4018   next();
4019 }
4020 
4021 //------------------------------next---------------------------
4022 void DepPreds::next() {
4023   if (_dep_next != NULL) {
4024     _current  = _dep_next->pred()->node();
4025     _dep_next = _dep_next->next_in();
4026   } else if (_next_idx < _end_idx) {
4027     _current  = _n->in(_next_idx++);
4028   } else {
4029     _done = true;
4030   }
4031 }
4032 
4033 // =========================== DepSuccs =========================
4034 // Iterator over successor edges in the dependence graph.
4035 
4036 //------------------------------DepSuccs---------------------------
4037 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
4038   _n = n;
4039   _done = false;
4040   if (_n->is_Load()) {
4041     _next_idx = 0;
4042     _end_idx  = _n->outcnt();
4043     _dep_next = dg.dep(_n)->out_head();
4044   } else if (_n->is_Mem() || (_n->is_Phi() && _n->bottom_type() == Type::MEMORY)) {
4045     _next_idx = 0;
4046     _end_idx  = 0;
4047     _dep_next = dg.dep(_n)->out_head();
4048   } else {
4049     _next_idx = 0;
4050     _end_idx  = _n->outcnt();
4051     _dep_next = NULL;
4052   }
4053   next();
4054 }
4055 
4056 //-------------------------------next---------------------------
4057 void DepSuccs::next() {
4058   if (_dep_next != NULL) {
4059     _current  = _dep_next->succ()->node();
4060     _dep_next = _dep_next->next_out();
4061   } else if (_next_idx < _end_idx) {
4062     _current  = _n->raw_out(_next_idx++);
4063   } else {
4064     _done = true;
4065   }
4066 }
4067 
4068 //
4069 // --------------------------------- vectorization/simd -----------------------------------
4070 //
4071 bool SuperWord::same_origin_idx(Node* a, Node* b) const {
4072   return a != NULL && b != NULL && _clone_map.same_idx(a->_idx, b->_idx);
4073 }
4074 bool SuperWord::same_generation(Node* a, Node* b) const {
4075   return a != NULL && b != NULL && _clone_map.same_gen(a->_idx, b->_idx);
4076 }
4077 
4078 Node*  SuperWord::find_phi_for_mem_dep(LoadNode* ld) {
4079   assert(in_bb(ld), "must be in block");
4080   if (_clone_map.gen(ld->_idx) == _ii_first) {
4081 #ifndef PRODUCT
4082     if (_vector_loop_debug) {
4083       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(ld->_idx)=%d",
4084         _clone_map.gen(ld->_idx));
4085     }
4086 #endif
4087     return NULL; //we think that any ld in the first gen being vectorizable
4088   }
4089 
4090   Node* mem = ld->in(MemNode::Memory);
4091   if (mem->outcnt() <= 1) {
4092     // we don't want to remove the only edge from mem node to load
4093 #ifndef PRODUCT
4094     if (_vector_loop_debug) {
4095       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",
4096         mem->_idx, ld->_idx);
4097       ld->dump();
4098       mem->dump();
4099     }
4100 #endif
4101     return NULL;
4102   }
4103   if (!in_bb(mem) || same_generation(mem, ld)) {
4104 #ifndef PRODUCT
4105     if (_vector_loop_debug) {
4106       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(mem->_idx)=%d",
4107         _clone_map.gen(mem->_idx));
4108     }
4109 #endif
4110     return NULL; // does not depend on loop volatile node or depends on the same generation
4111   }
4112 
4113   //otherwise first node should depend on mem-phi
4114   Node* first = first_node(ld);
4115   assert(first->is_Load(), "must be Load");
4116   Node* phi = first->as_Load()->in(MemNode::Memory);
4117   if (!phi->is_Phi() || phi->bottom_type() != Type::MEMORY) {
4118 #ifndef PRODUCT
4119     if (_vector_loop_debug) {
4120       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");
4121       ld->dump();
4122       first->dump();
4123     }
4124 #endif
4125     return NULL;
4126   }
4127 
4128   Node* tail = 0;
4129   for (int m = 0; m < _mem_slice_head.length(); m++) {
4130     if (_mem_slice_head.at(m) == phi) {
4131       tail = _mem_slice_tail.at(m);
4132     }
4133   }
4134   if (tail == 0) { //test that found phi is in the list  _mem_slice_head
4135 #ifndef PRODUCT
4136     if (_vector_loop_debug) {
4137       tty->print_cr("SuperWord::find_phi_for_mem_dep load %d is not vectorizable node, its phi %d is not _mem_slice_head",
4138         ld->_idx, phi->_idx);
4139       ld->dump();
4140       phi->dump();
4141     }
4142 #endif
4143     return NULL;
4144   }
4145 
4146   // now all conditions are met
4147   return phi;
4148 }
4149 
4150 Node* SuperWord::first_node(Node* nd) {
4151   for (int ii = 0; ii < _iteration_first.length(); ii++) {
4152     Node* nnn = _iteration_first.at(ii);
4153     if (same_origin_idx(nnn, nd)) {
4154 #ifndef PRODUCT
4155       if (_vector_loop_debug) {
4156         tty->print_cr("SuperWord::first_node: %d is the first iteration node for %d (_clone_map.idx(nnn->_idx) = %d)",
4157           nnn->_idx, nd->_idx, _clone_map.idx(nnn->_idx));
4158       }
4159 #endif
4160       return nnn;
4161     }
4162   }
4163 
4164 #ifndef PRODUCT
4165   if (_vector_loop_debug) {
4166     tty->print_cr("SuperWord::first_node: did not find first iteration node for %d (_clone_map.idx(nd->_idx)=%d)",
4167       nd->_idx, _clone_map.idx(nd->_idx));
4168   }
4169 #endif
4170   return 0;
4171 }
4172 
4173 Node* SuperWord::last_node(Node* nd) {
4174   for (int ii = 0; ii < _iteration_last.length(); ii++) {
4175     Node* nnn = _iteration_last.at(ii);
4176     if (same_origin_idx(nnn, nd)) {
4177 #ifndef PRODUCT
4178       if (_vector_loop_debug) {
4179         tty->print_cr("SuperWord::last_node _clone_map.idx(nnn->_idx)=%d, _clone_map.idx(nd->_idx)=%d",
4180           _clone_map.idx(nnn->_idx), _clone_map.idx(nd->_idx));
4181       }
4182 #endif
4183       return nnn;
4184     }
4185   }
4186   return 0;
4187 }
4188 
4189 int SuperWord::mark_generations() {
4190   Node *ii_err = NULL, *tail_err = NULL;
4191   for (int i = 0; i < _mem_slice_head.length(); i++) {
4192     Node* phi  = _mem_slice_head.at(i);
4193     assert(phi->is_Phi(), "must be phi");
4194 
4195     Node* tail = _mem_slice_tail.at(i);
4196     if (_ii_last == -1) {
4197       tail_err = tail;
4198       _ii_last = _clone_map.gen(tail->_idx);
4199     }
4200     else if (_ii_last != _clone_map.gen(tail->_idx)) {
4201 #ifndef PRODUCT
4202       if (TraceSuperWord && Verbose) {
4203         tty->print_cr("SuperWord::mark_generations _ii_last error - found different generations in two tail nodes ");
4204         tail->dump();
4205         tail_err->dump();
4206       }
4207 #endif
4208       return -1;
4209     }
4210 
4211     // find first iteration in the loop
4212     for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
4213       Node* ii = phi->fast_out(i);
4214       if (in_bb(ii) && ii->is_Store()) { // we speculate that normally Stores of one and one only generation have deps from mem phi
4215         if (_ii_first == -1) {
4216           ii_err = ii;
4217           _ii_first = _clone_map.gen(ii->_idx);
4218         } else if (_ii_first != _clone_map.gen(ii->_idx)) {
4219 #ifndef PRODUCT
4220           if (TraceSuperWord && Verbose) {
4221             tty->print_cr("SuperWord::mark_generations: _ii_first was found before and not equal to one in this node (%d)", _ii_first);
4222             ii->dump();
4223             if (ii_err!= 0) {
4224               ii_err->dump();
4225             }
4226           }
4227 #endif
4228           return -1; // this phi has Stores from different generations of unroll and cannot be simd/vectorized
4229         }
4230       }
4231     }//for (DUIterator_Fast imax,
4232   }//for (int i...
4233 
4234   if (_ii_first == -1 || _ii_last == -1) {
4235     if (TraceSuperWord && Verbose) {
4236       tty->print_cr("SuperWord::mark_generations unknown error, something vent wrong");
4237     }
4238     return -1; // something vent wrong
4239   }
4240   // collect nodes in the first and last generations
4241   assert(_iteration_first.length() == 0, "_iteration_first must be empty");
4242   assert(_iteration_last.length() == 0, "_iteration_last must be empty");
4243   for (int j = 0; j < _block.length(); j++) {
4244     Node* n = _block.at(j);
4245     node_idx_t gen = _clone_map.gen(n->_idx);
4246     if ((signed)gen == _ii_first) {
4247       _iteration_first.push(n);
4248     } else if ((signed)gen == _ii_last) {
4249       _iteration_last.push(n);
4250     }
4251   }
4252 
4253   // building order of iterations
4254   if (_ii_order.length() == 0 && ii_err != 0) {
4255     assert(in_bb(ii_err) && ii_err->is_Store(), "should be Store in bb");
4256     Node* nd = ii_err;
4257     while(_clone_map.gen(nd->_idx) != _ii_last) {
4258       _ii_order.push(_clone_map.gen(nd->_idx));
4259       bool found = false;
4260       for (DUIterator_Fast imax, i = nd->fast_outs(imax); i < imax; i++) {
4261         Node* use = nd->fast_out(i);
4262         if (same_origin_idx(use, nd) && use->as_Store()->in(MemNode::Memory) == nd) {
4263           found = true;
4264           nd = use;
4265           break;
4266         }
4267       }//for
4268 
4269       if (found == false) {
4270         if (TraceSuperWord && Verbose) {
4271           tty->print_cr("SuperWord::mark_generations: Cannot build order of iterations - no dependent Store for %d", nd->_idx);
4272         }
4273         _ii_order.clear();
4274         return -1;
4275       }
4276     } //while
4277     _ii_order.push(_clone_map.gen(nd->_idx));
4278   }
4279 
4280 #ifndef PRODUCT
4281   if (_vector_loop_debug) {
4282     tty->print_cr("SuperWord::mark_generations");
4283     tty->print_cr("First generation (%d) nodes:", _ii_first);
4284     for (int ii = 0; ii < _iteration_first.length(); ii++)  _iteration_first.at(ii)->dump();
4285     tty->print_cr("Last generation (%d) nodes:", _ii_last);
4286     for (int ii = 0; ii < _iteration_last.length(); ii++)  _iteration_last.at(ii)->dump();
4287     tty->print_cr(" ");
4288 
4289     tty->print("SuperWord::List of generations: ");
4290     for (int jj = 0; jj < _ii_order.length(); ++jj) {
4291       tty->print("%d:%d ", jj, _ii_order.at(jj));
4292     }
4293     tty->print_cr(" ");
4294   }
4295 #endif
4296 
4297   return _ii_first;
4298 }
4299 
4300 bool SuperWord::fix_commutative_inputs(Node* gold, Node* fix) {
4301   assert(gold->is_Add() && fix->is_Add() || gold->is_Mul() && fix->is_Mul(), "should be only Add or Mul nodes");
4302   assert(same_origin_idx(gold, fix), "should be clones of the same node");
4303   Node* gin1 = gold->in(1);
4304   Node* gin2 = gold->in(2);
4305   Node* fin1 = fix->in(1);
4306   Node* fin2 = fix->in(2);
4307   bool swapped = false;
4308 
4309   if (in_bb(gin1) && in_bb(gin2) && in_bb(fin1) && in_bb(fin1)) {
4310     if (same_origin_idx(gin1, fin1) &&
4311         same_origin_idx(gin2, fin2)) {
4312       return true; // nothing to fix
4313     }
4314     if (same_origin_idx(gin1, fin2) &&
4315         same_origin_idx(gin2, fin1)) {
4316       fix->swap_edges(1, 2);
4317       swapped = true;
4318     }
4319   }
4320   // at least one input comes from outside of bb
4321   if (gin1->_idx == fin1->_idx)  {
4322     return true; // nothing to fix
4323   }
4324   if (!swapped && (gin1->_idx == fin2->_idx || gin2->_idx == fin1->_idx))  { //swapping is expensive, check condition first
4325     fix->swap_edges(1, 2);
4326     swapped = true;
4327   }
4328 
4329   if (swapped) {
4330 #ifndef PRODUCT
4331     if (_vector_loop_debug) {
4332       tty->print_cr("SuperWord::fix_commutative_inputs: fixed node %d", fix->_idx);
4333     }
4334 #endif
4335     return true;
4336   }
4337 
4338   if (TraceSuperWord && Verbose) {
4339     tty->print_cr("SuperWord::fix_commutative_inputs: cannot fix node %d", fix->_idx);
4340   }
4341 
4342   return false;
4343 }
4344 
4345 bool SuperWord::pack_parallel() {
4346 #ifndef PRODUCT
4347   if (_vector_loop_debug) {
4348     tty->print_cr("SuperWord::pack_parallel: START");
4349   }
4350 #endif
4351 
4352   _packset.clear();
4353 
4354   for (int ii = 0; ii < _iteration_first.length(); ii++) {
4355     Node* nd = _iteration_first.at(ii);
4356     if (in_bb(nd) && (nd->is_Load() || nd->is_Store() || nd->is_Add() || nd->is_Mul())) {
4357       Node_List* pk = new Node_List();
4358       pk->push(nd);
4359       for (int gen = 1; gen < _ii_order.length(); ++gen) {
4360         for (int kk = 0; kk < _block.length(); kk++) {
4361           Node* clone = _block.at(kk);
4362           if (same_origin_idx(clone, nd) &&
4363               _clone_map.gen(clone->_idx) == _ii_order.at(gen)) {
4364             if (nd->is_Add() || nd->is_Mul()) {
4365               fix_commutative_inputs(nd, clone);
4366             }
4367             pk->push(clone);
4368             if (pk->size() == 4) {
4369               _packset.append(pk);
4370 #ifndef PRODUCT
4371               if (_vector_loop_debug) {
4372                 tty->print_cr("SuperWord::pack_parallel: added pack ");
4373                 pk->dump();
4374               }
4375 #endif
4376               if (_clone_map.gen(clone->_idx) != _ii_last) {
4377                 pk = new Node_List();
4378               }
4379             }
4380             break;
4381           }
4382         }
4383       }//for
4384     }//if
4385   }//for
4386 
4387 #ifndef PRODUCT
4388   if (_vector_loop_debug) {
4389     tty->print_cr("SuperWord::pack_parallel: END");
4390   }
4391 #endif
4392 
4393   return true;
4394 }
4395 
4396 bool SuperWord::hoist_loads_in_graph() {
4397   GrowableArray<Node*> loads;
4398 
4399 #ifndef PRODUCT
4400   if (_vector_loop_debug) {
4401     tty->print_cr("SuperWord::hoist_loads_in_graph: total number _mem_slice_head.length() = %d", _mem_slice_head.length());
4402   }
4403 #endif
4404 
4405   for (int i = 0; i < _mem_slice_head.length(); i++) {
4406     Node* n = _mem_slice_head.at(i);
4407     if ( !in_bb(n) || !n->is_Phi() || n->bottom_type() != Type::MEMORY) {
4408       if (TraceSuperWord && Verbose) {
4409         tty->print_cr("SuperWord::hoist_loads_in_graph: skipping unexpected node n=%d", n->_idx);
4410       }
4411       continue;
4412     }
4413 
4414 #ifndef PRODUCT
4415     if (_vector_loop_debug) {
4416       tty->print_cr("SuperWord::hoist_loads_in_graph: processing phi %d  = _mem_slice_head.at(%d);", n->_idx, i);
4417     }
4418 #endif
4419 
4420     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4421       Node* ld = n->fast_out(i);
4422       if (ld->is_Load() && ld->as_Load()->in(MemNode::Memory) == n && in_bb(ld)) {
4423         for (int i = 0; i < _block.length(); i++) {
4424           Node* ld2 = _block.at(i);
4425           if (ld2->is_Load() && same_origin_idx(ld, ld2) &&
4426               !same_generation(ld, ld2)) { // <= do not collect the first generation ld
4427 #ifndef PRODUCT
4428             if (_vector_loop_debug) {
4429               tty->print_cr("SuperWord::hoist_loads_in_graph: will try to hoist load ld2->_idx=%d, cloned from %d (ld->_idx=%d)",
4430                 ld2->_idx, _clone_map.idx(ld->_idx), ld->_idx);
4431             }
4432 #endif
4433             // could not do on-the-fly, since iterator is immutable
4434             loads.push(ld2);
4435           }
4436         }// for
4437       }//if
4438     }//for (DUIterator_Fast imax,
4439   }//for (int i = 0; i
4440 
4441   for (int i = 0; i < loads.length(); i++) {
4442     LoadNode* ld = loads.at(i)->as_Load();
4443     Node* phi = find_phi_for_mem_dep(ld);
4444     if (phi != NULL) {
4445 #ifndef PRODUCT
4446       if (_vector_loop_debug) {
4447         tty->print_cr("SuperWord::hoist_loads_in_graph replacing MemNode::Memory(%d) edge in %d with one from %d",
4448           MemNode::Memory, ld->_idx, phi->_idx);
4449       }
4450 #endif
4451       _igvn.replace_input_of(ld, MemNode::Memory, phi);
4452     }
4453   }//for
4454 
4455   restart(); // invalidate all basic structures, since we rebuilt the graph
4456 
4457   if (TraceSuperWord && Verbose) {
4458     tty->print_cr("\nSuperWord::hoist_loads_in_graph() the graph was rebuilt, all structures invalidated and need rebuild");
4459   }
4460 
4461   return true;
4462 }
4463