1 /*
   2  * Copyright (c) 2007, 2015, 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 "opto/addnode.hpp"
  29 #include "opto/callnode.hpp"
  30 #include "opto/castnode.hpp"
  31 #include "opto/convertnode.hpp"
  32 #include "opto/divnode.hpp"
  33 #include "opto/matcher.hpp"
  34 #include "opto/memnode.hpp"
  35 #include "opto/mulnode.hpp"
  36 #include "opto/opcodes.hpp"
  37 #include "opto/opaquenode.hpp"
  38 #include "opto/superword.hpp"
  39 #include "opto/vectornode.hpp"
  40 #include "opto/movenode.hpp"
  41 
  42 //
  43 //                  S U P E R W O R D   T R A N S F O R M
  44 //=============================================================================
  45 
  46 //------------------------------SuperWord---------------------------
  47 SuperWord::SuperWord(PhaseIdealLoop* phase) :
  48   _phase(phase),
  49   _igvn(phase->_igvn),
  50   _arena(phase->C->comp_arena()),
  51   _packset(arena(), 8,  0, NULL),         // packs for the current block
  52   _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
  53   _block(arena(), 8,  0, NULL),           // nodes in current block
  54   _data_entry(arena(), 8,  0, NULL),      // nodes with all inputs from outside
  55   _mem_slice_head(arena(), 8,  0, NULL),  // memory slice heads
  56   _mem_slice_tail(arena(), 8,  0, NULL),  // memory slice tails
  57   _node_info(arena(), 8,  0, SWNodeInfo::initial), // info needed per node
  58   _clone_map(phase->C->clone_map()),      // map of nodes created in cloning
  59   _cmovev_kit(_arena, this),              // map to facilitate CMoveVD creation
  60   _align_to_ref(NULL),                    // memory reference to align vectors to
  61   _disjoint_ptrs(arena(), 8,  0, OrderedPair::initial), // runtime disambiguated pointer pairs
  62   _dg(_arena),                            // dependence graph
  63   _visited(arena()),                      // visited node set
  64   _post_visited(arena()),                 // post visited node set
  65   _n_idx_list(arena(), 8),                // scratch list of (node,index) pairs
  66   _stk(arena(), 8, 0, NULL),              // scratch stack of nodes
  67   _nlist(arena(), 8, 0, NULL),            // scratch list of nodes
  68   _lpt(NULL),                             // loop tree node
  69   _lp(NULL),                              // LoopNode
  70   _bb(NULL),                              // basic block
  71   _iv(NULL),                              // induction var
  72   _race_possible(false),                  // cases where SDMU is true
  73   _early_return(true),                    // analysis evaluations routine
  74   _num_work_vecs(0),                      // amount of vector work we have
  75   _num_reductions(0),                     // amount of reduction work we have
  76   _do_vector_loop(phase->C->do_vector_loop()),  // whether to do vectorization/simd style
  77   _do_reserve_copy(DoReserveCopyInSuperWord),
  78   _ii_first(-1),                          // first loop generation index - only if do_vector_loop()
  79   _ii_last(-1),                           // last loop generation index - only if do_vector_loop()
  80   _ii_order(arena(), 8, 0, 0)
  81 {
  82 #ifndef PRODUCT
  83   _vector_loop_debug = 0;
  84   if (_phase->C->method() != NULL) {
  85     _vector_loop_debug = phase->C->directive()->VectorizeDebugOption;
  86   }
  87 
  88 #endif
  89 }
  90 
  91 //------------------------------transform_loop---------------------------
  92 void SuperWord::transform_loop(IdealLoopTree* lpt, bool do_optimization) {
  93   assert(UseSuperWord, "should be");
  94   // Do vectors exist on this architecture?
  95   if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return;
  96 
  97   assert(lpt->_head->is_CountedLoop(), "must be");
  98   CountedLoopNode *cl = lpt->_head->as_CountedLoop();
  99 
 100   if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
 101 
 102   if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
 103   // Check for no control flow in body (other than exit)
 104   Node *cl_exit = cl->loopexit();
 105   if (cl_exit->in(0) != lpt->_head) {
 106     #ifndef PRODUCT
 107       if (TraceSuperWord) {
 108         tty->print_cr("SuperWord::transform_loop: loop too complicated, cl_exit->in(0) != lpt->_head");
 109         tty->print("cl_exit %d", cl_exit->_idx); cl_exit->dump();
 110         tty->print("cl_exit->in(0) %d", cl_exit->in(0)->_idx); cl_exit->in(0)->dump();
 111         tty->print("lpt->_head %d", lpt->_head->_idx); lpt->_head->dump();
 112         lpt->dump_head();
 113       }
 114     #endif
 115     return;
 116   }
 117 
 118   // Make sure the are no extra control users of the loop backedge
 119   if (cl->back_control()->outcnt() != 1) {
 120     return;
 121   }
 122 
 123   // We only re-enter slp when we vector mapped a queried loop and we want to
 124   // continue unrolling, in this case, slp is not subsequently done.
 125   if (cl->do_unroll_only()) return;
 126 
 127   // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
 128   CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
 129   if (pre_end == NULL) return;
 130   Node *pre_opaq1 = pre_end->limit();
 131   if (pre_opaq1->Opcode() != Op_Opaque1) return;
 132 
 133   init(); // initialize data structures
 134 
 135   set_lpt(lpt);
 136   set_lp(cl);
 137 
 138   // For now, define one block which is the entire loop body
 139   set_bb(cl);
 140 
 141   if (do_optimization) {
 142     assert(_packset.length() == 0, "packset must be empty");
 143     SLP_extract();
 144   }
 145 }
 146 
 147 //------------------------------early unrolling analysis------------------------------
 148 void SuperWord::unrolling_analysis(int &local_loop_unroll_factor) {
 149   bool is_slp = true;
 150   ResourceMark rm;
 151   size_t ignored_size = lpt()->_body.size();
 152   int *ignored_loop_nodes = NEW_RESOURCE_ARRAY(int, ignored_size);
 153   Node_Stack nstack((int)ignored_size);
 154   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
 155   Node *cl_exit = cl->loopexit();
 156 
 157   // First clear the entries
 158   for (uint i = 0; i < lpt()->_body.size(); i++) {
 159     ignored_loop_nodes[i] = -1;
 160   }
 161 
 162   int max_vector = Matcher::max_vector_size(T_INT);
 163 
 164   // Process the loop, some/all of the stack entries will not be in order, ergo
 165   // need to preprocess the ignored initial state before we process the loop
 166   for (uint i = 0; i < lpt()->_body.size(); i++) {
 167     Node* n = lpt()->_body.at(i);
 168     if (n == cl->incr() ||
 169       n->is_reduction() ||
 170       n->is_AddP() ||
 171       n->is_Cmp() ||
 172       n->is_IfTrue() ||
 173       n->is_CountedLoop() ||
 174       (n == cl_exit)) {
 175       ignored_loop_nodes[i] = n->_idx;
 176       continue;
 177     }
 178 
 179     if (n->is_If()) {
 180       IfNode *iff = n->as_If();
 181       if (iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN) {
 182         if (lpt()->is_loop_exit(iff)) {
 183           ignored_loop_nodes[i] = n->_idx;
 184           continue;
 185         }
 186       }
 187     }
 188 
 189     if (n->is_Phi() && (n->bottom_type() == Type::MEMORY)) {
 190       Node* n_tail = n->in(LoopNode::LoopBackControl);
 191       if (n_tail != n->in(LoopNode::EntryControl)) {
 192         if (!n_tail->is_Mem()) {
 193           is_slp = false;
 194           break;
 195         }
 196       }
 197     }
 198 
 199     // This must happen after check of phi/if
 200     if (n->is_Phi() || n->is_If()) {
 201       ignored_loop_nodes[i] = n->_idx;
 202       continue;
 203     }
 204 
 205     if (n->is_LoadStore() || n->is_MergeMem() ||
 206       (n->is_Proj() && !n->as_Proj()->is_CFG())) {
 207       is_slp = false;
 208       break;
 209     }
 210 
 211     // Ignore nodes with non-primitive type.
 212     BasicType bt;
 213     if (n->is_Mem()) {
 214       bt = n->as_Mem()->memory_type();
 215     } else {
 216       bt = n->bottom_type()->basic_type();
 217     }
 218     if (is_java_primitive(bt) == false) {
 219       ignored_loop_nodes[i] = n->_idx;
 220       continue;
 221     }
 222 
 223     if (n->is_Mem()) {
 224       MemNode* current = n->as_Mem();
 225       Node* adr = n->in(MemNode::Address);
 226       Node* n_ctrl = _phase->get_ctrl(adr);
 227 
 228       // save a queue of post process nodes
 229       if (n_ctrl != NULL && lpt()->is_member(_phase->get_loop(n_ctrl))) {
 230         // Process the memory expression
 231         int stack_idx = 0;
 232         bool have_side_effects = true;
 233         if (adr->is_AddP() == false) {
 234           nstack.push(adr, stack_idx++);
 235         } else {
 236           // Mark the components of the memory operation in nstack
 237           SWPointer p1(current, this, &nstack, true);
 238           have_side_effects = p1.node_stack()->is_nonempty();
 239         }
 240 
 241         // Process the pointer stack
 242         while (have_side_effects) {
 243           Node* pointer_node = nstack.node();
 244           for (uint j = 0; j < lpt()->_body.size(); j++) {
 245             Node* cur_node = lpt()->_body.at(j);
 246             if (cur_node == pointer_node) {
 247               ignored_loop_nodes[j] = cur_node->_idx;
 248               break;
 249             }
 250           }
 251           nstack.pop();
 252           have_side_effects = nstack.is_nonempty();
 253         }
 254       }
 255     }
 256   }
 257 
 258   if (is_slp) {
 259     // Now we try to find the maximum supported consistent vector which the machine
 260     // description can use
 261     for (uint i = 0; i < lpt()->_body.size(); i++) {
 262       if (ignored_loop_nodes[i] != -1) continue;
 263 
 264       BasicType bt;
 265       Node* n = lpt()->_body.at(i);
 266       if (n->is_Mem()) {
 267         bt = n->as_Mem()->memory_type();
 268       } else {
 269         bt = n->bottom_type()->basic_type();
 270       }
 271       if (is_java_primitive(bt) == false) continue;
 272 
 273       int cur_max_vector = Matcher::max_vector_size(bt);
 274 
 275       // If a max vector exists which is not larger than _local_loop_unroll_factor
 276       // stop looking, we already have the max vector to map to.
 277       if (cur_max_vector < local_loop_unroll_factor) {
 278         is_slp = false;
 279         if (TraceSuperWordLoopUnrollAnalysis) {
 280           tty->print_cr("slp analysis fails: unroll limit greater than max vector\n");
 281         }
 282         break;
 283       }
 284 
 285       // Map the maximal common vector
 286       if (VectorNode::implemented(n->Opcode(), cur_max_vector, bt)) {
 287         if (cur_max_vector < max_vector) {
 288           max_vector = cur_max_vector;
 289         }
 290       }
 291     }
 292     if (is_slp) {
 293       local_loop_unroll_factor = max_vector;
 294       cl->mark_passed_slp();
 295     }
 296     cl->mark_was_slp();
 297     cl->set_slp_max_unroll(local_loop_unroll_factor);
 298   }
 299 }
 300 
 301 //------------------------------SLP_extract---------------------------
 302 // Extract the superword level parallelism
 303 //
 304 // 1) A reverse post-order of nodes in the block is constructed.  By scanning
 305 //    this list from first to last, all definitions are visited before their uses.
 306 //
 307 // 2) A point-to-point dependence graph is constructed between memory references.
 308 //    This simplies the upcoming "independence" checker.
 309 //
 310 // 3) The maximum depth in the node graph from the beginning of the block
 311 //    to each node is computed.  This is used to prune the graph search
 312 //    in the independence checker.
 313 //
 314 // 4) For integer types, the necessary bit width is propagated backwards
 315 //    from stores to allow packed operations on byte, char, and short
 316 //    integers.  This reverses the promotion to type "int" that javac
 317 //    did for operations like: char c1,c2,c3;  c1 = c2 + c3.
 318 //
 319 // 5) One of the memory references is picked to be an aligned vector reference.
 320 //    The pre-loop trip count is adjusted to align this reference in the
 321 //    unrolled body.
 322 //
 323 // 6) The initial set of pack pairs is seeded with memory references.
 324 //
 325 // 7) The set of pack pairs is extended by following use->def and def->use links.
 326 //
 327 // 8) The pairs are combined into vector sized packs.
 328 //
 329 // 9) Reorder the memory slices to co-locate members of the memory packs.
 330 //
 331 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
 332 //    inserting scalar promotion, vector creation from multiple scalars, and
 333 //    extraction of scalar values from vectors.
 334 //
 335 void SuperWord::SLP_extract() {
 336 
 337 #ifndef PRODUCT
 338   if (_do_vector_loop && TraceSuperWord) {
 339     tty->print("SuperWord::SLP_extract\n");
 340     tty->print("input loop\n");
 341     _lpt->dump_head();
 342     _lpt->dump();
 343     for (uint i = 0; i < _lpt->_body.size(); i++) {
 344       _lpt->_body.at(i)->dump();
 345     }
 346   }
 347 #endif
 348   // Ready the block
 349   if (!construct_bb()) {
 350     return; // Exit if no interesting nodes or complex graph.
 351   }
 352   // build    _dg, _disjoint_ptrs
 353   dependence_graph();
 354 
 355   // compute function depth(Node*)
 356   compute_max_depth();
 357 
 358   if (_do_vector_loop) {
 359     if (mark_generations() != -1) {
 360       hoist_loads_in_graph(); // this only rebuild the graph; all basic structs need rebuild explicitly
 361 
 362       if (!construct_bb()) {
 363         return; // Exit if no interesting nodes or complex graph.
 364       }
 365       dependence_graph();
 366       compute_max_depth();
 367     }
 368 
 369 #ifndef PRODUCT
 370     if (TraceSuperWord) {
 371       tty->print_cr("\nSuperWord::_do_vector_loop: graph after hoist_loads_in_graph");
 372       _lpt->dump_head();
 373       for (int j = 0; j < _block.length(); j++) {
 374         Node* n = _block.at(j);
 375         int d = depth(n);
 376         for (int i = 0;  i < d; i++) tty->print("%s", "  ");
 377         tty->print("%d :", d);
 378         n->dump();
 379       }
 380     }
 381 #endif
 382   }
 383 
 384   compute_vector_element_type();
 385 
 386   // Attempt vectorization
 387 
 388   find_adjacent_refs();
 389 
 390   extend_packlist();
 391 
 392   if (_do_vector_loop) {
 393     if (_packset.length() == 0) {
 394       if (TraceSuperWord) {
 395         tty->print_cr("\nSuperWord::_do_vector_loop DFA could not build packset, now trying to build anyway");
 396       }
 397       pack_parallel();
 398     }
 399   }
 400 
 401   combine_packs();
 402 
 403   construct_my_pack_map();
 404 
 405   if (_do_vector_loop) {
 406     merge_packs_to_cmovd();
 407   }
 408 
 409   filter_packs();
 410 
 411   schedule();
 412 
 413   output();
 414 }
 415 
 416 //------------------------------find_adjacent_refs---------------------------
 417 // Find the adjacent memory references and create pack pairs for them.
 418 // This is the initial set of packs that will then be extended by
 419 // following use->def and def->use links.  The align positions are
 420 // assigned relative to the reference "align_to_ref"
 421 void SuperWord::find_adjacent_refs() {
 422   // Get list of memory operations
 423   Node_List memops;
 424   for (int i = 0; i < _block.length(); i++) {
 425     Node* n = _block.at(i);
 426     if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
 427         is_java_primitive(n->as_Mem()->memory_type())) {
 428       int align = memory_alignment(n->as_Mem(), 0);
 429       if (align != bottom_align) {
 430         memops.push(n);
 431       }
 432     }
 433   }
 434 
 435   Node_List align_to_refs;
 436   int best_iv_adjustment = 0;
 437   MemNode* best_align_to_mem_ref = NULL;
 438 
 439   while (memops.size() != 0) {
 440     // Find a memory reference to align to.
 441     MemNode* mem_ref = find_align_to_ref(memops);
 442     if (mem_ref == NULL) break;
 443     align_to_refs.push(mem_ref);
 444     int iv_adjustment = get_iv_adjustment(mem_ref);
 445 
 446     if (best_align_to_mem_ref == NULL) {
 447       // Set memory reference which is the best from all memory operations
 448       // to be used for alignment. The pre-loop trip count is modified to align
 449       // this reference to a vector-aligned address.
 450       best_align_to_mem_ref = mem_ref;
 451       best_iv_adjustment = iv_adjustment;
 452       NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
 453     }
 454 
 455     SWPointer align_to_ref_p(mem_ref, this, NULL, false);
 456     // Set alignment relative to "align_to_ref" for all related memory operations.
 457     for (int i = memops.size() - 1; i >= 0; i--) {
 458       MemNode* s = memops.at(i)->as_Mem();
 459       if (isomorphic(s, mem_ref) &&
 460            (!_do_vector_loop || same_origin_idx(s, mem_ref))) {
 461         SWPointer p2(s, this, NULL, false);
 462         if (p2.comparable(align_to_ref_p)) {
 463           int align = memory_alignment(s, iv_adjustment);
 464           set_alignment(s, align);
 465         }
 466       }
 467     }
 468 
 469     // Create initial pack pairs of memory operations for which
 470     // alignment is set and vectors will be aligned.
 471     bool create_pack = true;
 472     if (memory_alignment(mem_ref, best_iv_adjustment) == 0 || _do_vector_loop) {
 473       if (!Matcher::misaligned_vectors_ok()) {
 474         int vw = vector_width(mem_ref);
 475         int vw_best = vector_width(best_align_to_mem_ref);
 476         if (vw > vw_best) {
 477           // Do not vectorize a memory access with more elements per vector
 478           // if unaligned memory access is not allowed because number of
 479           // iterations in pre-loop will be not enough to align it.
 480           create_pack = false;
 481         } else {
 482           SWPointer p2(best_align_to_mem_ref, this, NULL, false);
 483           if (align_to_ref_p.invar() != p2.invar()) {
 484             // Do not vectorize memory accesses with different invariants
 485             // if unaligned memory accesses are not allowed.
 486             create_pack = false;
 487           }
 488         }
 489       }
 490     } else {
 491       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 492         // Can't allow vectorization of unaligned memory accesses with the
 493         // same type since it could be overlapped accesses to the same array.
 494         create_pack = false;
 495       } else {
 496         // Allow independent (different type) unaligned memory operations
 497         // if HW supports them.
 498         if (!Matcher::misaligned_vectors_ok()) {
 499           create_pack = false;
 500         } else {
 501           // Check if packs of the same memory type but
 502           // with a different alignment were created before.
 503           for (uint i = 0; i < align_to_refs.size(); i++) {
 504             MemNode* mr = align_to_refs.at(i)->as_Mem();
 505             if (same_velt_type(mr, mem_ref) &&
 506                 memory_alignment(mr, iv_adjustment) != 0)
 507               create_pack = false;
 508           }
 509         }
 510       }
 511     }
 512     if (create_pack) {
 513       for (uint i = 0; i < memops.size(); i++) {
 514         Node* s1 = memops.at(i);
 515         int align = alignment(s1);
 516         if (align == top_align) continue;
 517         for (uint j = 0; j < memops.size(); j++) {
 518           Node* s2 = memops.at(j);
 519           if (alignment(s2) == top_align) continue;
 520           if (s1 != s2 && are_adjacent_refs(s1, s2)) {
 521             if (stmts_can_pack(s1, s2, align)) {
 522               Node_List* pair = new Node_List();
 523               pair->push(s1);
 524               pair->push(s2);
 525               if (!_do_vector_loop || same_origin_idx(s1, s2)) {
 526                 _packset.append(pair);
 527               }
 528             }
 529           }
 530         }
 531       }
 532     } else { // Don't create unaligned pack
 533       // First, remove remaining memory ops of the same type from the list.
 534       for (int i = memops.size() - 1; i >= 0; i--) {
 535         MemNode* s = memops.at(i)->as_Mem();
 536         if (same_velt_type(s, mem_ref)) {
 537           memops.remove(i);
 538         }
 539       }
 540 
 541       // Second, remove already constructed packs of the same type.
 542       for (int i = _packset.length() - 1; i >= 0; i--) {
 543         Node_List* p = _packset.at(i);
 544         MemNode* s = p->at(0)->as_Mem();
 545         if (same_velt_type(s, mem_ref)) {
 546           remove_pack_at(i);
 547         }
 548       }
 549 
 550       // If needed find the best memory reference for loop alignment again.
 551       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 552         // Put memory ops from remaining packs back on memops list for
 553         // the best alignment search.
 554         uint orig_msize = memops.size();
 555         for (int i = 0; i < _packset.length(); i++) {
 556           Node_List* p = _packset.at(i);
 557           MemNode* s = p->at(0)->as_Mem();
 558           assert(!same_velt_type(s, mem_ref), "sanity");
 559           memops.push(s);
 560         }
 561         best_align_to_mem_ref = find_align_to_ref(memops);
 562         if (best_align_to_mem_ref == NULL) {
 563           if (TraceSuperWord) {
 564             tty->print_cr("SuperWord::find_adjacent_refs(): best_align_to_mem_ref == NULL");
 565           }
 566           break;
 567         }
 568         best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
 569         NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
 570         // Restore list.
 571         while (memops.size() > orig_msize)
 572           (void)memops.pop();
 573       }
 574     } // unaligned memory accesses
 575 
 576     // Remove used mem nodes.
 577     for (int i = memops.size() - 1; i >= 0; i--) {
 578       MemNode* m = memops.at(i)->as_Mem();
 579       if (alignment(m) != top_align) {
 580         memops.remove(i);
 581       }
 582     }
 583 
 584   } // while (memops.size() != 0
 585   set_align_to_ref(best_align_to_mem_ref);
 586 
 587   if (TraceSuperWord) {
 588     tty->print_cr("\nAfter find_adjacent_refs");
 589     print_packset();
 590   }
 591 }
 592 
 593 #ifndef PRODUCT
 594 void SuperWord::find_adjacent_refs_trace_1(Node* best_align_to_mem_ref, int best_iv_adjustment) {
 595   if (is_trace_adjacent()) {
 596     tty->print("SuperWord::find_adjacent_refs best_align_to_mem_ref = %d, best_iv_adjustment = %d",
 597        best_align_to_mem_ref->_idx, best_iv_adjustment);
 598        best_align_to_mem_ref->dump();
 599   }
 600 }
 601 #endif
 602 
 603 //------------------------------find_align_to_ref---------------------------
 604 // Find a memory reference to align the loop induction variable to.
 605 // Looks first at stores then at loads, looking for a memory reference
 606 // with the largest number of references similar to it.
 607 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
 608   GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
 609 
 610   // Count number of comparable memory ops
 611   for (uint i = 0; i < memops.size(); i++) {
 612     MemNode* s1 = memops.at(i)->as_Mem();
 613     SWPointer p1(s1, this, NULL, false);
 614     // Discard if pre loop can't align this reference
 615     if (!ref_is_alignable(p1)) {
 616       *cmp_ct.adr_at(i) = 0;
 617       continue;
 618     }
 619     for (uint j = i+1; j < memops.size(); j++) {
 620       MemNode* s2 = memops.at(j)->as_Mem();
 621       if (isomorphic(s1, s2)) {
 622         SWPointer p2(s2, this, NULL, false);
 623         if (p1.comparable(p2)) {
 624           (*cmp_ct.adr_at(i))++;
 625           (*cmp_ct.adr_at(j))++;
 626         }
 627       }
 628     }
 629   }
 630 
 631   // Find Store (or Load) with the greatest number of "comparable" references,
 632   // biggest vector size, smallest data size and smallest iv offset.
 633   int max_ct        = 0;
 634   int max_vw        = 0;
 635   int max_idx       = -1;
 636   int min_size      = max_jint;
 637   int min_iv_offset = max_jint;
 638   for (uint j = 0; j < memops.size(); j++) {
 639     MemNode* s = memops.at(j)->as_Mem();
 640     if (s->is_Store()) {
 641       int vw = vector_width_in_bytes(s);
 642       assert(vw > 1, "sanity");
 643       SWPointer p(s, this, NULL, false);
 644       if (cmp_ct.at(j) >  max_ct ||
 645           cmp_ct.at(j) == max_ct &&
 646             (vw >  max_vw ||
 647              vw == max_vw &&
 648               (data_size(s) <  min_size ||
 649                data_size(s) == min_size &&
 650                  (p.offset_in_bytes() < min_iv_offset)))) {
 651         max_ct = cmp_ct.at(j);
 652         max_vw = vw;
 653         max_idx = j;
 654         min_size = data_size(s);
 655         min_iv_offset = p.offset_in_bytes();
 656       }
 657     }
 658   }
 659   // If no stores, look at loads
 660   if (max_ct == 0) {
 661     for (uint j = 0; j < memops.size(); j++) {
 662       MemNode* s = memops.at(j)->as_Mem();
 663       if (s->is_Load()) {
 664         int vw = vector_width_in_bytes(s);
 665         assert(vw > 1, "sanity");
 666         SWPointer p(s, this, NULL, false);
 667         if (cmp_ct.at(j) >  max_ct ||
 668             cmp_ct.at(j) == max_ct &&
 669               (vw >  max_vw ||
 670                vw == max_vw &&
 671                 (data_size(s) <  min_size ||
 672                  data_size(s) == min_size &&
 673                    (p.offset_in_bytes() < min_iv_offset)))) {
 674           max_ct = cmp_ct.at(j);
 675           max_vw = vw;
 676           max_idx = j;
 677           min_size = data_size(s);
 678           min_iv_offset = p.offset_in_bytes();
 679         }
 680       }
 681     }
 682   }
 683 
 684 #ifdef ASSERT
 685   if (TraceSuperWord && Verbose) {
 686     tty->print_cr("\nVector memops after find_align_to_ref");
 687     for (uint i = 0; i < memops.size(); i++) {
 688       MemNode* s = memops.at(i)->as_Mem();
 689       s->dump();
 690     }
 691   }
 692 #endif
 693 
 694   if (max_ct > 0) {
 695 #ifdef ASSERT
 696     if (TraceSuperWord) {
 697       tty->print("\nVector align to node: ");
 698       memops.at(max_idx)->as_Mem()->dump();
 699     }
 700 #endif
 701     return memops.at(max_idx)->as_Mem();
 702   }
 703   return NULL;
 704 }
 705 
 706 //------------------------------ref_is_alignable---------------------------
 707 // Can the preloop align the reference to position zero in the vector?
 708 bool SuperWord::ref_is_alignable(SWPointer& p) {
 709   if (!p.has_iv()) {
 710     return true;   // no induction variable
 711   }
 712   CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
 713   assert(pre_end != NULL, "we must have a correct pre-loop");
 714   assert(pre_end->stride_is_con(), "pre loop stride is constant");
 715   int preloop_stride = pre_end->stride_con();
 716 
 717   int span = preloop_stride * p.scale_in_bytes();
 718   int mem_size = p.memory_size();
 719   int offset   = p.offset_in_bytes();
 720   // Stride one accesses are alignable if offset is aligned to memory operation size.
 721   // Offset can be unaligned when UseUnalignedAccesses is used.
 722   if (ABS(span) == mem_size && (ABS(offset) % mem_size) == 0) {
 723     return true;
 724   }
 725   // If the initial offset from start of the object is computable,
 726   // check if the pre-loop can align the final offset accordingly.
 727   //
 728   // In other words: Can we find an i such that the offset
 729   // after i pre-loop iterations is aligned to vw?
 730   //   (init_offset + pre_loop) % vw == 0              (1)
 731   // where
 732   //   pre_loop = i * span
 733   // is the number of bytes added to the offset by i pre-loop iterations.
 734   //
 735   // For this to hold we need pre_loop to increase init_offset by
 736   //   pre_loop = vw - (init_offset % vw)
 737   //
 738   // This is only possible if pre_loop is divisible by span because each
 739   // pre-loop iteration increases the initial offset by 'span' bytes:
 740   //   (vw - (init_offset % vw)) % span == 0
 741   //
 742   int vw = vector_width_in_bytes(p.mem());
 743   assert(vw > 1, "sanity");
 744   Node* init_nd = pre_end->init_trip();
 745   if (init_nd->is_Con() && p.invar() == NULL) {
 746     int init = init_nd->bottom_type()->is_int()->get_con();
 747     int init_offset = init * p.scale_in_bytes() + offset;
 748     assert(init_offset >= 0, "positive offset from object start");
 749     if (vw % span == 0) {
 750       // If vm is a multiple of span, we use formula (1).
 751       if (span > 0) {
 752         return (vw - (init_offset % vw)) % span == 0;
 753       } else {
 754         assert(span < 0, "nonzero stride * scale");
 755         return (init_offset % vw) % -span == 0;
 756       }
 757     } else if (span % vw == 0) {
 758       // If span is a multiple of vw, we can simplify formula (1) to:
 759       //   (init_offset + i * span) % vw == 0
 760       //     =>
 761       //   (init_offset % vw) + ((i * span) % vw) == 0
 762       //     =>
 763       //   init_offset % vw == 0
 764       //
 765       // Because we add a multiple of vw to the initial offset, the final
 766       // offset is a multiple of vw if and only if init_offset is a multiple.
 767       //
 768       return (init_offset % vw) == 0;
 769     }
 770   }
 771   return false;
 772 }
 773 
 774 //---------------------------get_iv_adjustment---------------------------
 775 // Calculate loop's iv adjustment for this memory ops.
 776 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
 777   SWPointer align_to_ref_p(mem_ref, this, NULL, false);
 778   int offset = align_to_ref_p.offset_in_bytes();
 779   int scale  = align_to_ref_p.scale_in_bytes();
 780   int elt_size = align_to_ref_p.memory_size();
 781   int vw       = vector_width_in_bytes(mem_ref);
 782   assert(vw > 1, "sanity");
 783   int iv_adjustment;
 784   if (scale != 0) {
 785     int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
 786     // At least one iteration is executed in pre-loop by default. As result
 787     // several iterations are needed to align memory operations in main-loop even
 788     // if offset is 0.
 789     int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
 790     assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
 791            "(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size);
 792     iv_adjustment = iv_adjustment_in_bytes/elt_size;
 793   } else {
 794     // This memory op is not dependent on iv (scale == 0)
 795     iv_adjustment = 0;
 796   }
 797 
 798 #ifndef PRODUCT
 799   if (TraceSuperWord) {
 800     tty->print("SuperWord::get_iv_adjustment: n = %d, noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d: ",
 801       mem_ref->_idx, offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
 802     mem_ref->dump();
 803   }
 804 #endif
 805   return iv_adjustment;
 806 }
 807 
 808 //---------------------------dependence_graph---------------------------
 809 // Construct dependency graph.
 810 // Add dependence edges to load/store nodes for memory dependence
 811 //    A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
 812 void SuperWord::dependence_graph() {
 813   // First, assign a dependence node to each memory node
 814   for (int i = 0; i < _block.length(); i++ ) {
 815     Node *n = _block.at(i);
 816     if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
 817       _dg.make_node(n);
 818     }
 819   }
 820 
 821   // For each memory slice, create the dependences
 822   for (int i = 0; i < _mem_slice_head.length(); i++) {
 823     Node* n      = _mem_slice_head.at(i);
 824     Node* n_tail = _mem_slice_tail.at(i);
 825 
 826     // Get slice in predecessor order (last is first)
 827     mem_slice_preds(n_tail, n, _nlist);
 828 
 829 #ifndef PRODUCT
 830     if(TraceSuperWord && Verbose) {
 831       tty->print_cr("SuperWord::dependence_graph: built a new mem slice");
 832       for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 833         _nlist.at(j)->dump();
 834       }
 835     }
 836 #endif
 837     // Make the slice dependent on the root
 838     DepMem* slice = _dg.dep(n);
 839     _dg.make_edge(_dg.root(), slice);
 840 
 841     // Create a sink for the slice
 842     DepMem* slice_sink = _dg.make_node(NULL);
 843     _dg.make_edge(slice_sink, _dg.tail());
 844 
 845     // Now visit each pair of memory ops, creating the edges
 846     for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 847       Node* s1 = _nlist.at(j);
 848 
 849       // If no dependency yet, use slice
 850       if (_dg.dep(s1)->in_cnt() == 0) {
 851         _dg.make_edge(slice, s1);
 852       }
 853       SWPointer p1(s1->as_Mem(), this, NULL, false);
 854       bool sink_dependent = true;
 855       for (int k = j - 1; k >= 0; k--) {
 856         Node* s2 = _nlist.at(k);
 857         if (s1->is_Load() && s2->is_Load())
 858           continue;
 859         SWPointer p2(s2->as_Mem(), this, NULL, false);
 860 
 861         int cmp = p1.cmp(p2);
 862         if (SuperWordRTDepCheck &&
 863             p1.base() != p2.base() && p1.valid() && p2.valid()) {
 864           // Create a runtime check to disambiguate
 865           OrderedPair pp(p1.base(), p2.base());
 866           _disjoint_ptrs.append_if_missing(pp);
 867         } else if (!SWPointer::not_equal(cmp)) {
 868           // Possibly same address
 869           _dg.make_edge(s1, s2);
 870           sink_dependent = false;
 871         }
 872       }
 873       if (sink_dependent) {
 874         _dg.make_edge(s1, slice_sink);
 875       }
 876     }
 877 
 878     if (TraceSuperWord) {
 879       tty->print_cr("\nDependence graph for slice: %d", n->_idx);
 880       for (int q = 0; q < _nlist.length(); q++) {
 881         _dg.print(_nlist.at(q));
 882       }
 883       tty->cr();
 884     }
 885 
 886     _nlist.clear();
 887   }
 888 
 889   if (TraceSuperWord) {
 890     tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
 891     for (int r = 0; r < _disjoint_ptrs.length(); r++) {
 892       _disjoint_ptrs.at(r).print();
 893       tty->cr();
 894     }
 895     tty->cr();
 896   }
 897 
 898 }
 899 
 900 //---------------------------mem_slice_preds---------------------------
 901 // Return a memory slice (node list) in predecessor order starting at "start"
 902 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
 903   assert(preds.length() == 0, "start empty");
 904   Node* n = start;
 905   Node* prev = NULL;
 906   while (true) {
 907     NOT_PRODUCT( if(is_trace_mem_slice()) tty->print_cr("SuperWord::mem_slice_preds: n %d", n->_idx);)
 908     assert(in_bb(n), "must be in block");
 909     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 910       Node* out = n->fast_out(i);
 911       if (out->is_Load()) {
 912         if (in_bb(out)) {
 913           preds.push(out);
 914           if (TraceSuperWord && Verbose) {
 915             tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", out->_idx);
 916           }
 917         }
 918       } else {
 919         // FIXME
 920         if (out->is_MergeMem() && !in_bb(out)) {
 921           // Either unrolling is causing a memory edge not to disappear,
 922           // or need to run igvn.optimize() again before SLP
 923         } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
 924           // Ditto.  Not sure what else to check further.
 925         } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
 926           // StoreCM has an input edge used as a precedence edge.
 927           // Maybe an issue when oop stores are vectorized.
 928         } else {
 929           assert(out == prev || prev == NULL, "no branches off of store slice");
 930         }
 931       }//else
 932     }//for
 933     if (n == stop) break;
 934     preds.push(n);
 935     if (TraceSuperWord && Verbose) {
 936       tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", n->_idx);
 937     }
 938     prev = n;
 939     assert(n->is_Mem(), "unexpected node %s", n->Name());
 940     n = n->in(MemNode::Memory);
 941   }
 942 }
 943 
 944 //------------------------------stmts_can_pack---------------------------
 945 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
 946 // s1 aligned at "align"
 947 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
 948 
 949   // Do not use superword for non-primitives
 950   BasicType bt1 = velt_basic_type(s1);
 951   BasicType bt2 = velt_basic_type(s2);
 952   if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
 953     return false;
 954   if (Matcher::max_vector_size(bt1) < 2) {
 955     return false; // No vectors for this type
 956   }
 957 
 958   if (isomorphic(s1, s2)) {
 959     if (independent(s1, s2) || reduction(s1, s2)) {
 960       if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
 961         if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
 962           int s1_align = alignment(s1);
 963           int s2_align = alignment(s2);
 964           if (s1_align == top_align || s1_align == align) {
 965             if (s2_align == top_align || s2_align == align + data_size(s1)) {
 966               return true;
 967             }
 968           }
 969         }
 970       }
 971     }
 972   }
 973   return false;
 974 }
 975 
 976 //------------------------------exists_at---------------------------
 977 // Does s exist in a pack at position pos?
 978 bool SuperWord::exists_at(Node* s, uint pos) {
 979   for (int i = 0; i < _packset.length(); i++) {
 980     Node_List* p = _packset.at(i);
 981     if (p->at(pos) == s) {
 982       return true;
 983     }
 984   }
 985   return false;
 986 }
 987 
 988 //------------------------------are_adjacent_refs---------------------------
 989 // Is s1 immediately before s2 in memory?
 990 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
 991   if (!s1->is_Mem() || !s2->is_Mem()) return false;
 992   if (!in_bb(s1)    || !in_bb(s2))    return false;
 993 
 994   // Do not use superword for non-primitives
 995   if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
 996       !is_java_primitive(s2->as_Mem()->memory_type())) {
 997     return false;
 998   }
 999 
1000   // FIXME - co_locate_pack fails on Stores in different mem-slices, so
1001   // only pack memops that are in the same alias set until that's fixed.
1002   if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
1003       _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
1004     return false;
1005   SWPointer p1(s1->as_Mem(), this, NULL, false);
1006   SWPointer p2(s2->as_Mem(), this, NULL, false);
1007   if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
1008   int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
1009   return diff == data_size(s1);
1010 }
1011 
1012 //------------------------------isomorphic---------------------------
1013 // Are s1 and s2 similar?
1014 bool SuperWord::isomorphic(Node* s1, Node* s2) {
1015   if (s1->Opcode() != s2->Opcode()) return false;
1016   if (s1->req() != s2->req()) return false;
1017   if (s1->in(0) != s2->in(0)) return false;
1018   if (!same_velt_type(s1, s2)) return false;
1019   return true;
1020 }
1021 
1022 //------------------------------independent---------------------------
1023 // Is there no data path from s1 to s2 or s2 to s1?
1024 bool SuperWord::independent(Node* s1, Node* s2) {
1025   //  assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
1026   int d1 = depth(s1);
1027   int d2 = depth(s2);
1028   if (d1 == d2) return s1 != s2;
1029   Node* deep    = d1 > d2 ? s1 : s2;
1030   Node* shallow = d1 > d2 ? s2 : s1;
1031 
1032   visited_clear();
1033 
1034   return independent_path(shallow, deep);
1035 }
1036 
1037 //------------------------------reduction---------------------------
1038 // Is there a data path between s1 and s2 and the nodes reductions?
1039 bool SuperWord::reduction(Node* s1, Node* s2) {
1040   bool retValue = false;
1041   int d1 = depth(s1);
1042   int d2 = depth(s2);
1043   if (d1 + 1 == d2) {
1044     if (s1->is_reduction() && s2->is_reduction()) {
1045       // This is an ordered set, so s1 should define s2
1046       for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1047         Node* t1 = s1->fast_out(i);
1048         if (t1 == s2) {
1049           // both nodes are reductions and connected
1050           retValue = true;
1051         }
1052       }
1053     }
1054   }
1055 
1056   return retValue;
1057 }
1058 
1059 //------------------------------independent_path------------------------------
1060 // Helper for independent
1061 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
1062   if (dp >= 1000) return false; // stop deep recursion
1063   visited_set(deep);
1064   int shal_depth = depth(shallow);
1065   assert(shal_depth <= depth(deep), "must be");
1066   for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
1067     Node* pred = preds.current();
1068     if (in_bb(pred) && !visited_test(pred)) {
1069       if (shallow == pred) {
1070         return false;
1071       }
1072       if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
1073         return false;
1074       }
1075     }
1076   }
1077   return true;
1078 }
1079 
1080 //------------------------------set_alignment---------------------------
1081 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
1082   set_alignment(s1, align);
1083   if (align == top_align || align == bottom_align) {
1084     set_alignment(s2, align);
1085   } else {
1086     set_alignment(s2, align + data_size(s1));
1087   }
1088 }
1089 
1090 //------------------------------data_size---------------------------
1091 int SuperWord::data_size(Node* s) {
1092   Node* use = NULL; //test if the node is a candidate for CMoveVD optimization, then return the size of CMov
1093   if (_do_vector_loop) {
1094     use = _cmovev_kit.is_Bool_candidate(s);
1095     if (use != NULL) {
1096       return data_size(use);
1097     }
1098     use = _cmovev_kit.is_CmpD_candidate(s);
1099     if (use != NULL) {
1100       return data_size(use);
1101     }
1102   }
1103   int bsize = type2aelembytes(velt_basic_type(s));
1104   assert(bsize != 0, "valid size");
1105   return bsize;
1106 }
1107 
1108 //------------------------------extend_packlist---------------------------
1109 // Extend packset by following use->def and def->use links from pack members.
1110 void SuperWord::extend_packlist() {
1111   bool changed;
1112   do {
1113     packset_sort(_packset.length());
1114     changed = false;
1115     for (int i = 0; i < _packset.length(); i++) {
1116       Node_List* p = _packset.at(i);
1117       changed |= follow_use_defs(p);
1118       changed |= follow_def_uses(p);
1119     }
1120   } while (changed);
1121 
1122   if (_race_possible) {
1123     for (int i = 0; i < _packset.length(); i++) {
1124       Node_List* p = _packset.at(i);
1125       order_def_uses(p);
1126     }
1127   }
1128 
1129   if (TraceSuperWord) {
1130     tty->print_cr("\nAfter extend_packlist");
1131     print_packset();
1132   }
1133 }
1134 
1135 //------------------------------follow_use_defs---------------------------
1136 // Extend the packset by visiting operand definitions of nodes in pack p
1137 bool SuperWord::follow_use_defs(Node_List* p) {
1138   assert(p->size() == 2, "just checking");
1139   Node* s1 = p->at(0);
1140   Node* s2 = p->at(1);
1141   assert(s1->req() == s2->req(), "just checking");
1142   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1143 
1144   if (s1->is_Load()) return false;
1145 
1146   int align = alignment(s1);
1147   NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: s1 %d, align %d", s1->_idx, align);)
1148   bool changed = false;
1149   int start = s1->is_Store() ? MemNode::ValueIn   : 1;
1150   int end   = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
1151   for (int j = start; j < end; j++) {
1152     Node* t1 = s1->in(j);
1153     Node* t2 = s2->in(j);
1154     if (!in_bb(t1) || !in_bb(t2))
1155       continue;
1156     if (stmts_can_pack(t1, t2, align)) {
1157       if (est_savings(t1, t2) >= 0) {
1158         Node_List* pair = new Node_List();
1159         pair->push(t1);
1160         pair->push(t2);
1161         _packset.append(pair);
1162         NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: set_alignment(%d, %d, %d)", t1->_idx, t2->_idx, align);)
1163         set_alignment(t1, t2, align);
1164         changed = true;
1165       }
1166     }
1167   }
1168   return changed;
1169 }
1170 
1171 //------------------------------follow_def_uses---------------------------
1172 // Extend the packset by visiting uses of nodes in pack p
1173 bool SuperWord::follow_def_uses(Node_List* p) {
1174   bool changed = false;
1175   Node* s1 = p->at(0);
1176   Node* s2 = p->at(1);
1177   assert(p->size() == 2, "just checking");
1178   assert(s1->req() == s2->req(), "just checking");
1179   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1180 
1181   if (s1->is_Store()) return false;
1182 
1183   int align = alignment(s1);
1184   NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: s1 %d, align %d", s1->_idx, align);)
1185   int savings = -1;
1186   int num_s1_uses = 0;
1187   Node* u1 = NULL;
1188   Node* u2 = NULL;
1189   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1190     Node* t1 = s1->fast_out(i);
1191     num_s1_uses++;
1192     if (!in_bb(t1)) continue;
1193     for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1194       Node* t2 = s2->fast_out(j);
1195       if (!in_bb(t2)) continue;
1196       if (!opnd_positions_match(s1, t1, s2, t2))
1197         continue;
1198       if (stmts_can_pack(t1, t2, align)) {
1199         int my_savings = est_savings(t1, t2);
1200         if (my_savings > savings) {
1201           savings = my_savings;
1202           u1 = t1;
1203           u2 = t2;
1204         }
1205       }
1206     }
1207   }
1208   if (num_s1_uses > 1) {
1209     _race_possible = true;
1210   }
1211   if (savings >= 0) {
1212     Node_List* pair = new Node_List();
1213     pair->push(u1);
1214     pair->push(u2);
1215     _packset.append(pair);
1216     NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: set_alignment(%d, %d, %d)", u1->_idx, u2->_idx, align);)
1217     set_alignment(u1, u2, align);
1218     changed = true;
1219   }
1220   return changed;
1221 }
1222 
1223 //------------------------------order_def_uses---------------------------
1224 // For extended packsets, ordinally arrange uses packset by major component
1225 void SuperWord::order_def_uses(Node_List* p) {
1226   Node* s1 = p->at(0);
1227 
1228   if (s1->is_Store()) return;
1229 
1230   // reductions are always managed beforehand
1231   if (s1->is_reduction()) return;
1232 
1233   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1234     Node* t1 = s1->fast_out(i);
1235 
1236     // Only allow operand swap on commuting operations
1237     if (!t1->is_Add() && !t1->is_Mul()) {
1238       break;
1239     }
1240 
1241     // Now find t1's packset
1242     Node_List* p2 = NULL;
1243     for (int j = 0; j < _packset.length(); j++) {
1244       p2 = _packset.at(j);
1245       Node* first = p2->at(0);
1246       if (t1 == first) {
1247         break;
1248       }
1249       p2 = NULL;
1250     }
1251     // Arrange all sub components by the major component
1252     if (p2 != NULL) {
1253       for (uint j = 1; j < p->size(); j++) {
1254         Node* d1 = p->at(j);
1255         Node* u1 = p2->at(j);
1256         opnd_positions_match(s1, t1, d1, u1);
1257       }
1258     }
1259   }
1260 }
1261 
1262 //---------------------------opnd_positions_match-------------------------
1263 // Is the use of d1 in u1 at the same operand position as d2 in u2?
1264 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
1265   // check reductions to see if they are marshalled to represent the reduction
1266   // operator in a specified opnd
1267   if (u1->is_reduction() && u2->is_reduction()) {
1268     // ensure reductions have phis and reduction definitions feeding the 1st operand
1269     Node* first = u1->in(2);
1270     if (first->is_Phi() || first->is_reduction()) {
1271       u1->swap_edges(1, 2);
1272     }
1273     // ensure reductions have phis and reduction definitions feeding the 1st operand
1274     first = u2->in(2);
1275     if (first->is_Phi() || first->is_reduction()) {
1276       u2->swap_edges(1, 2);
1277     }
1278     return true;
1279   }
1280 
1281   uint ct = u1->req();
1282   if (ct != u2->req()) return false;
1283   uint i1 = 0;
1284   uint i2 = 0;
1285   do {
1286     for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
1287     for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
1288     if (i1 != i2) {
1289       if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
1290         // Further analysis relies on operands position matching.
1291         u2->swap_edges(i1, i2);
1292       } else {
1293         return false;
1294       }
1295     }
1296   } while (i1 < ct);
1297   return true;
1298 }
1299 
1300 //------------------------------est_savings---------------------------
1301 // Estimate the savings from executing s1 and s2 as a pack
1302 int SuperWord::est_savings(Node* s1, Node* s2) {
1303   int save_in = 2 - 1; // 2 operations per instruction in packed form
1304 
1305   // inputs
1306   for (uint i = 1; i < s1->req(); i++) {
1307     Node* x1 = s1->in(i);
1308     Node* x2 = s2->in(i);
1309     if (x1 != x2) {
1310       if (are_adjacent_refs(x1, x2)) {
1311         save_in += adjacent_profit(x1, x2);
1312       } else if (!in_packset(x1, x2)) {
1313         save_in -= pack_cost(2);
1314       } else {
1315         save_in += unpack_cost(2);
1316       }
1317     }
1318   }
1319 
1320   // uses of result
1321   uint ct = 0;
1322   int save_use = 0;
1323   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1324     Node* s1_use = s1->fast_out(i);
1325     for (int j = 0; j < _packset.length(); j++) {
1326       Node_List* p = _packset.at(j);
1327       if (p->at(0) == s1_use) {
1328         for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
1329           Node* s2_use = s2->fast_out(k);
1330           if (p->at(p->size()-1) == s2_use) {
1331             ct++;
1332             if (are_adjacent_refs(s1_use, s2_use)) {
1333               save_use += adjacent_profit(s1_use, s2_use);
1334             }
1335           }
1336         }
1337       }
1338     }
1339   }
1340 
1341   if (ct < s1->outcnt()) save_use += unpack_cost(1);
1342   if (ct < s2->outcnt()) save_use += unpack_cost(1);
1343 
1344   return MAX2(save_in, save_use);
1345 }
1346 
1347 //------------------------------costs---------------------------
1348 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
1349 int SuperWord::pack_cost(int ct)   { return ct; }
1350 int SuperWord::unpack_cost(int ct) { return ct; }
1351 
1352 //------------------------------combine_packs---------------------------
1353 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
1354 void SuperWord::combine_packs() {
1355   bool changed = true;
1356   // Combine packs regardless max vector size.
1357   while (changed) {
1358     changed = false;
1359     for (int i = 0; i < _packset.length(); i++) {
1360       Node_List* p1 = _packset.at(i);
1361       if (p1 == NULL) continue;
1362       // Because of sorting we can start at i + 1
1363       for (int j = i + 1; j < _packset.length(); j++) {
1364         Node_List* p2 = _packset.at(j);
1365         if (p2 == NULL) continue;
1366         if (i == j) continue;
1367         if (p1->at(p1->size()-1) == p2->at(0)) {
1368           for (uint k = 1; k < p2->size(); k++) {
1369             p1->push(p2->at(k));
1370           }
1371           _packset.at_put(j, NULL);
1372           changed = true;
1373         }
1374       }
1375     }
1376   }
1377 
1378   // Split packs which have size greater then max vector size.
1379   for (int i = 0; i < _packset.length(); i++) {
1380     Node_List* p1 = _packset.at(i);
1381     if (p1 != NULL) {
1382       BasicType bt = velt_basic_type(p1->at(0));
1383       uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1384       assert(is_power_of_2(max_vlen), "sanity");
1385       uint psize = p1->size();
1386       if (!is_power_of_2(psize)) {
1387         // Skip pack which can't be vector.
1388         // case1: for(...) { a[i] = i; }    elements values are different (i+x)
1389         // case2: for(...) { a[i] = b[i+1]; }  can't align both, load and store
1390         _packset.at_put(i, NULL);
1391         continue;
1392       }
1393       if (psize > max_vlen) {
1394         Node_List* pack = new Node_List();
1395         for (uint j = 0; j < psize; j++) {
1396           pack->push(p1->at(j));
1397           if (pack->size() >= max_vlen) {
1398             assert(is_power_of_2(pack->size()), "sanity");
1399             _packset.append(pack);
1400             pack = new Node_List();
1401           }
1402         }
1403         _packset.at_put(i, NULL);
1404       }
1405     }
1406   }
1407 
1408   // Compress list.
1409   for (int i = _packset.length() - 1; i >= 0; i--) {
1410     Node_List* p1 = _packset.at(i);
1411     if (p1 == NULL) {
1412       _packset.remove_at(i);
1413     }
1414   }
1415 
1416   if (TraceSuperWord) {
1417     tty->print_cr("\nAfter combine_packs");
1418     print_packset();
1419   }
1420 }
1421 
1422 //-----------------------------construct_my_pack_map--------------------------
1423 // Construct the map from nodes to packs.  Only valid after the
1424 // point where a node is only in one pack (after combine_packs).
1425 void SuperWord::construct_my_pack_map() {
1426   Node_List* rslt = NULL;
1427   for (int i = 0; i < _packset.length(); i++) {
1428     Node_List* p = _packset.at(i);
1429     for (uint j = 0; j < p->size(); j++) {
1430       Node* s = p->at(j);
1431       assert(my_pack(s) == NULL, "only in one pack");
1432       set_my_pack(s, p);
1433     }
1434   }
1435 }
1436 
1437 //------------------------------filter_packs---------------------------
1438 // Remove packs that are not implemented or not profitable.
1439 void SuperWord::filter_packs() {
1440   // Remove packs that are not implemented
1441   for (int i = _packset.length() - 1; i >= 0; i--) {
1442     Node_List* pk = _packset.at(i);
1443     bool impl = implemented(pk);
1444     if (!impl) {
1445 #ifndef PRODUCT
1446       if (TraceSuperWord && Verbose) {
1447         tty->print_cr("Unimplemented");
1448         pk->at(0)->dump();
1449       }
1450 #endif
1451       remove_pack_at(i);
1452     }
1453     Node *n = pk->at(0);
1454     if (n->is_reduction()) {
1455       _num_reductions++;
1456     } else {
1457       _num_work_vecs++;
1458     }
1459   }
1460 
1461   // Remove packs that are not profitable
1462   bool changed;
1463   do {
1464     changed = false;
1465     for (int i = _packset.length() - 1; i >= 0; i--) {
1466       Node_List* pk = _packset.at(i);
1467       bool prof = profitable(pk);
1468       if (!prof) {
1469 #ifndef PRODUCT
1470         if (TraceSuperWord && Verbose) {
1471           tty->print_cr("Unprofitable");
1472           pk->at(0)->dump();
1473         }
1474 #endif
1475         remove_pack_at(i);
1476         changed = true;
1477       }
1478     }
1479   } while (changed);
1480 
1481 #ifndef PRODUCT
1482   if (TraceSuperWord) {
1483     tty->print_cr("\nAfter filter_packs");
1484     print_packset();
1485     tty->cr();
1486   }
1487 #endif
1488 }
1489 
1490 //------------------------------merge_packs_to_cmovd---------------------------
1491 // Merge CMoveD into new vector-nodes
1492 // We want to catch this pattern and subsume CmpD and Bool into CMoveD
1493 //
1494 //                   SubD             ConD
1495 //                  /  |               /
1496 //                 /   |           /   /
1497 //                /    |       /      /
1498 //               /     |   /         /
1499 //              /      /            /
1500 //             /    /  |           /
1501 //            v /      |          /
1502 //         CmpD        |         /
1503 //          |          |        /
1504 //          v          |       /
1505 //         Bool        |      /
1506 //           \         |     /
1507 //             \       |    /
1508 //               \     |   /
1509 //                 \   |  /
1510 //                   \ v /
1511 //                   CMoveD
1512 //
1513 
1514 void SuperWord::merge_packs_to_cmovd() {
1515   for (int i = _packset.length() - 1; i >= 0; i--) {
1516     _cmovev_kit.make_cmovevd_pack(_packset.at(i));
1517   }
1518   #ifndef PRODUCT
1519     if (TraceSuperWord) {
1520       tty->print_cr("\nSuperWord::merge_packs_to_cmovd(): After merge");
1521       print_packset();
1522       tty->cr();
1523     }
1524   #endif
1525 }
1526 
1527 Node* CMoveKit::is_Bool_candidate(Node* def) const {
1528   Node* use = NULL;
1529   if (!def->is_Bool() || def->in(0) != NULL || def->outcnt() != 1) {
1530     return NULL;
1531   }
1532   for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1533     use = def->fast_out(j);
1534     if (!_sw->same_generation(def, use) || !use->is_CMove()) {
1535       return NULL;
1536     }
1537   }
1538   return use;
1539 }
1540 
1541 Node* CMoveKit::is_CmpD_candidate(Node* def) const {
1542   Node* use = NULL;
1543   if (!def->is_Cmp() || def->in(0) != NULL || def->outcnt() != 1) {
1544     return NULL;
1545   }
1546   for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1547     use = def->fast_out(j);
1548     if (!_sw->same_generation(def, use) || (use = is_Bool_candidate(use)) == NULL || !_sw->same_generation(def, use)) {
1549       return NULL;
1550     }
1551   }
1552   return use;
1553 }
1554 
1555 Node_List* CMoveKit::make_cmovevd_pack(Node_List* cmovd_pk) {
1556   Node *cmovd = cmovd_pk->at(0);
1557   if (!cmovd->is_CMove()) {
1558     return NULL;
1559   }
1560   if (pack(cmovd) != NULL) { // already in the cmov pack
1561     return NULL;
1562   }
1563   if (cmovd->in(0) != NULL) {
1564     NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: CMoveD %d has control flow, escaping...", cmovd->_idx); cmovd->dump();})
1565     return NULL;
1566   }
1567 
1568   Node* bol = cmovd->as_CMove()->in(CMoveNode::Condition);
1569   if (!bol->is_Bool()
1570       || bol->outcnt() != 1
1571       || !_sw->same_generation(bol, cmovd)
1572       || bol->in(0) != NULL  // BoolNode has control flow!!
1573       || _sw->my_pack(bol) == NULL) {
1574       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();})
1575       return NULL;
1576   }
1577   Node_List* bool_pk = _sw->my_pack(bol);
1578   if (bool_pk->size() != cmovd_pk->size() ) {
1579     return NULL;
1580   }
1581 
1582   Node* cmpd = bol->in(1);
1583   if (!cmpd->is_Cmp()
1584       || cmpd->outcnt() != 1
1585       || !_sw->same_generation(cmpd, cmovd)
1586       || cmpd->in(0) != NULL  // CmpDNode has control flow!!
1587       || _sw->my_pack(cmpd) == NULL) {
1588       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();})
1589       return NULL;
1590   }
1591   Node_List* cmpd_pk = _sw->my_pack(cmpd);
1592   if (cmpd_pk->size() != cmovd_pk->size() ) {
1593     return NULL;
1594   }
1595 
1596   if (!test_cmpd_pack(cmpd_pk, cmovd_pk)) {
1597     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();})
1598     return NULL;
1599   }
1600 
1601   Node_List* new_cmpd_pk = new Node_List();
1602   uint sz = cmovd_pk->size() - 1;
1603   for (uint i = 0; i <= sz; ++i) {
1604     Node* cmov = cmovd_pk->at(i);
1605     Node* bol  = bool_pk->at(i);
1606     Node* cmp  = cmpd_pk->at(i);
1607 
1608     new_cmpd_pk->insert(i, cmov);
1609 
1610     map(cmov, new_cmpd_pk);
1611     map(bol, new_cmpd_pk);
1612     map(cmp, new_cmpd_pk);
1613 
1614     _sw->set_my_pack(cmov, new_cmpd_pk); // and keep old packs for cmp and bool
1615   }
1616   _sw->_packset.remove(cmovd_pk);
1617   _sw->_packset.remove(bool_pk);
1618   _sw->_packset.remove(cmpd_pk);
1619   _sw->_packset.append(new_cmpd_pk);
1620   NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print_cr("CMoveKit::make_cmovevd_pack: added syntactic CMoveD pack"); _sw->print_pack(new_cmpd_pk);})
1621   return new_cmpd_pk;
1622 }
1623 
1624 bool CMoveKit::test_cmpd_pack(Node_List* cmpd_pk, Node_List* cmovd_pk) {
1625   Node* cmpd0 = cmpd_pk->at(0);
1626   assert(cmpd0->is_Cmp(), "CMoveKit::test_cmpd_pack: should be CmpDNode");
1627   assert(cmovd_pk->at(0)->is_CMove(), "CMoveKit::test_cmpd_pack: should be CMoveD");
1628   assert(cmpd_pk->size() == cmovd_pk->size(), "CMoveKit::test_cmpd_pack: should be same size");
1629   Node* in1 = cmpd0->in(1);
1630   Node* in2 = cmpd0->in(2);
1631   Node_List* in1_pk = _sw->my_pack(in1);
1632   Node_List* in2_pk = _sw->my_pack(in2);
1633 
1634   if (in1_pk != NULL && in1_pk->size() != cmpd_pk->size()
1635     || in2_pk != NULL && in2_pk->size() != cmpd_pk->size() ) {
1636     return false;
1637   }
1638 
1639   // test if "all" in1 are in the same pack or the same node
1640   if (in1_pk == NULL) {
1641     for (uint j = 1; j < cmpd_pk->size(); j++) {
1642       if (cmpd_pk->at(j)->in(1) != in1) {
1643         return false;
1644       }
1645     }//for: in1_pk is not pack but all CmpD nodes in the pack have the same in(1)
1646   }
1647   // test if "all" in2 are in the same pack or the same node
1648   if (in2_pk == NULL) {
1649     for (uint j = 1; j < cmpd_pk->size(); j++) {
1650       if (cmpd_pk->at(j)->in(2) != in2) {
1651         return false;
1652       }
1653     }//for: in2_pk is not pack but all CmpD nodes in the pack have the same in(2)
1654   }
1655   //now check if cmpd_pk may be subsumed in vector built for cmovd_pk
1656   int cmovd_ind1, cmovd_ind2;
1657   if (cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1658    && cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1659       cmovd_ind1 = CMoveNode::IfFalse;
1660       cmovd_ind2 = CMoveNode::IfTrue;
1661   } else if (cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1662           && cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1663       cmovd_ind2 = CMoveNode::IfFalse;
1664       cmovd_ind1 = CMoveNode::IfTrue;
1665   }
1666   else {
1667     return false;
1668   }
1669 
1670   for (uint j = 1; j < cmpd_pk->size(); j++) {
1671     if (cmpd_pk->at(j)->in(1) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind1)
1672         || cmpd_pk->at(j)->in(2) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind2)) {
1673         return false;
1674     }//if
1675   }
1676   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(); })
1677   return true;
1678 }
1679 
1680 //------------------------------implemented---------------------------
1681 // Can code be generated for pack p?
1682 bool SuperWord::implemented(Node_List* p) {
1683   bool retValue = false;
1684   Node* p0 = p->at(0);
1685   if (p0 != NULL) {
1686     int opc = p0->Opcode();
1687     uint size = p->size();
1688     if (p0->is_reduction()) {
1689       const Type *arith_type = p0->bottom_type();
1690       // Length 2 reductions of INT/LONG do not offer performance benefits
1691       if (((arith_type->basic_type() == T_INT) || (arith_type->basic_type() == T_LONG)) && (size == 2)) {
1692         retValue = false;
1693       } else {
1694         retValue = ReductionNode::implemented(opc, size, arith_type->basic_type());
1695       }
1696     } else {
1697       retValue = VectorNode::implemented(opc, size, velt_basic_type(p0));
1698     }
1699     if (!retValue) {
1700       if (is_cmov_pack(p)) {
1701         NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::implemented: found cmpd pack"); print_pack(p);})
1702         return true;
1703       }
1704     }
1705   }
1706   return retValue;
1707 }
1708 
1709 bool SuperWord::is_cmov_pack(Node_List* p) {
1710   return _cmovev_kit.pack(p->at(0)) != NULL;
1711 }
1712 //------------------------------same_inputs--------------------------
1713 // For pack p, are all idx operands the same?
1714 bool SuperWord::same_inputs(Node_List* p, int idx) {
1715   Node* p0 = p->at(0);
1716   uint vlen = p->size();
1717   Node* p0_def = p0->in(idx);
1718   for (uint i = 1; i < vlen; i++) {
1719     Node* pi = p->at(i);
1720     Node* pi_def = pi->in(idx);
1721     if (p0_def != pi_def) {
1722       return false;
1723     }
1724   }
1725   return true;
1726 }
1727 
1728 //------------------------------profitable---------------------------
1729 // For pack p, are all operands and all uses (with in the block) vector?
1730 bool SuperWord::profitable(Node_List* p) {
1731   Node* p0 = p->at(0);
1732   uint start, end;
1733   VectorNode::vector_operands(p0, &start, &end);
1734 
1735   // Return false if some inputs are not vectors or vectors with different
1736   // size or alignment.
1737   // Also, for now, return false if not scalar promotion case when inputs are
1738   // the same. Later, implement PackNode and allow differing, non-vector inputs
1739   // (maybe just the ones from outside the block.)
1740   for (uint i = start; i < end; i++) {
1741     if (!is_vector_use(p0, i)) {
1742       return false;
1743     }
1744   }
1745   // Check if reductions are connected
1746   if (p0->is_reduction()) {
1747     Node* second_in = p0->in(2);
1748     Node_List* second_pk = my_pack(second_in);
1749     if ((second_pk == NULL) || (_num_work_vecs == _num_reductions)) {
1750       // Remove reduction flag if no parent pack or if not enough work
1751       // to cover reduction expansion overhead
1752       p0->remove_flag(Node::Flag_is_reduction);
1753       return false;
1754     } else if (second_pk->size() != p->size()) {
1755       return false;
1756     }
1757   }
1758   if (VectorNode::is_shift(p0)) {
1759     // For now, return false if shift count is vector or not scalar promotion
1760     // case (different shift counts) because it is not supported yet.
1761     Node* cnt = p0->in(2);
1762     Node_List* cnt_pk = my_pack(cnt);
1763     if (cnt_pk != NULL)
1764       return false;
1765     if (!same_inputs(p, 2))
1766       return false;
1767   }
1768   if (!p0->is_Store()) {
1769     // For now, return false if not all uses are vector.
1770     // Later, implement ExtractNode and allow non-vector uses (maybe
1771     // just the ones outside the block.)
1772     for (uint i = 0; i < p->size(); i++) {
1773       Node* def = p->at(i);
1774       if (is_cmov_pack_internal_node(p, def)) {
1775         continue;
1776       }
1777       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1778         Node* use = def->fast_out(j);
1779         for (uint k = 0; k < use->req(); k++) {
1780           Node* n = use->in(k);
1781           if (def == n) {
1782             // reductions can be loop carried dependences
1783             if (def->is_reduction() && use->is_Phi())
1784               continue;
1785             if (!is_vector_use(use, k)) {
1786               return false;
1787             }
1788           }
1789         }
1790       }
1791     }
1792   }
1793   return true;
1794 }
1795 
1796 //------------------------------schedule---------------------------
1797 // Adjust the memory graph for the packed operations
1798 void SuperWord::schedule() {
1799 
1800   // Co-locate in the memory graph the members of each memory pack
1801   for (int i = 0; i < _packset.length(); i++) {
1802     co_locate_pack(_packset.at(i));
1803   }
1804 }
1805 
1806 //-------------------------------remove_and_insert-------------------
1807 // Remove "current" from its current position in the memory graph and insert
1808 // it after the appropriate insertion point (lip or uip).
1809 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1810                                   Node *uip, Unique_Node_List &sched_before) {
1811   Node* my_mem = current->in(MemNode::Memory);
1812   bool sched_up = sched_before.member(current);
1813 
1814   // remove current_store from its current position in the memmory graph
1815   for (DUIterator i = current->outs(); current->has_out(i); i++) {
1816     Node* use = current->out(i);
1817     if (use->is_Mem()) {
1818       assert(use->in(MemNode::Memory) == current, "must be");
1819       if (use == prev) { // connect prev to my_mem
1820           _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1821           --i; //deleted this edge; rescan position
1822       } else if (sched_before.member(use)) {
1823         if (!sched_up) { // Will be moved together with current
1824           _igvn.replace_input_of(use, MemNode::Memory, uip);
1825           --i; //deleted this edge; rescan position
1826         }
1827       } else {
1828         if (sched_up) { // Will be moved together with current
1829           _igvn.replace_input_of(use, MemNode::Memory, lip);
1830           --i; //deleted this edge; rescan position
1831         }
1832       }
1833     }
1834   }
1835 
1836   Node *insert_pt =  sched_up ?  uip : lip;
1837 
1838   // all uses of insert_pt's memory state should use current's instead
1839   for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1840     Node* use = insert_pt->out(i);
1841     if (use->is_Mem()) {
1842       assert(use->in(MemNode::Memory) == insert_pt, "must be");
1843       _igvn.replace_input_of(use, MemNode::Memory, current);
1844       --i; //deleted this edge; rescan position
1845     } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1846       uint pos; //lip (lower insert point) must be the last one in the memory slice
1847       for (pos=1; pos < use->req(); pos++) {
1848         if (use->in(pos) == insert_pt) break;
1849       }
1850       _igvn.replace_input_of(use, pos, current);
1851       --i;
1852     }
1853   }
1854 
1855   //connect current to insert_pt
1856   _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
1857 }
1858 
1859 //------------------------------co_locate_pack----------------------------------
1860 // To schedule a store pack, we need to move any sandwiched memory ops either before
1861 // or after the pack, based upon dependence information:
1862 // (1) If any store in the pack depends on the sandwiched memory op, the
1863 //     sandwiched memory op must be scheduled BEFORE the pack;
1864 // (2) If a sandwiched memory op depends on any store in the pack, the
1865 //     sandwiched memory op must be scheduled AFTER the pack;
1866 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1867 //     memory op (say memB), memB must be scheduled before memA. So, if memA is
1868 //     scheduled before the pack, memB must also be scheduled before the pack;
1869 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1870 //     schedule this store AFTER the pack
1871 // (5) We know there is no dependence cycle, so there in no other case;
1872 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1873 //
1874 // To schedule a load pack, we use the memory state of either the first or the last load in
1875 // the pack, based on the dependence constraint.
1876 void SuperWord::co_locate_pack(Node_List* pk) {
1877   if (pk->at(0)->is_Store()) {
1878     MemNode* first     = executed_first(pk)->as_Mem();
1879     MemNode* last      = executed_last(pk)->as_Mem();
1880     Unique_Node_List schedule_before_pack;
1881     Unique_Node_List memops;
1882 
1883     MemNode* current   = last->in(MemNode::Memory)->as_Mem();
1884     MemNode* previous  = last;
1885     while (true) {
1886       assert(in_bb(current), "stay in block");
1887       memops.push(previous);
1888       for (DUIterator i = current->outs(); current->has_out(i); i++) {
1889         Node* use = current->out(i);
1890         if (use->is_Mem() && use != previous)
1891           memops.push(use);
1892       }
1893       if (current == first) break;
1894       previous = current;
1895       current  = current->in(MemNode::Memory)->as_Mem();
1896     }
1897 
1898     // determine which memory operations should be scheduled before the pack
1899     for (uint i = 1; i < memops.size(); i++) {
1900       Node *s1 = memops.at(i);
1901       if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1902         for (uint j = 0; j< i; j++) {
1903           Node *s2 = memops.at(j);
1904           if (!independent(s1, s2)) {
1905             if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1906               schedule_before_pack.push(s1); // s1 must be scheduled before
1907               Node_List* mem_pk = my_pack(s1);
1908               if (mem_pk != NULL) {
1909                 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1910                   Node* s = mem_pk->at(ii);  // follow partner
1911                   if (memops.member(s) && !schedule_before_pack.member(s))
1912                     schedule_before_pack.push(s);
1913                 }
1914               }
1915               break;
1916             }
1917           }
1918         }
1919       }
1920     }
1921 
1922     Node*    upper_insert_pt = first->in(MemNode::Memory);
1923     // Following code moves loads connected to upper_insert_pt below aliased stores.
1924     // Collect such loads here and reconnect them back to upper_insert_pt later.
1925     memops.clear();
1926     for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
1927       Node* use = upper_insert_pt->out(i);
1928       if (use->is_Mem() && !use->is_Store()) {
1929         memops.push(use);
1930       }
1931     }
1932 
1933     MemNode* lower_insert_pt = last;
1934     previous                 = last; //previous store in pk
1935     current                  = last->in(MemNode::Memory)->as_Mem();
1936 
1937     // start scheduling from "last" to "first"
1938     while (true) {
1939       assert(in_bb(current), "stay in block");
1940       assert(in_pack(previous, pk), "previous stays in pack");
1941       Node* my_mem = current->in(MemNode::Memory);
1942 
1943       if (in_pack(current, pk)) {
1944         // Forward users of my memory state (except "previous) to my input memory state
1945         for (DUIterator i = current->outs(); current->has_out(i); i++) {
1946           Node* use = current->out(i);
1947           if (use->is_Mem() && use != previous) {
1948             assert(use->in(MemNode::Memory) == current, "must be");
1949             if (schedule_before_pack.member(use)) {
1950               _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
1951             } else {
1952               _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
1953             }
1954             --i; // deleted this edge; rescan position
1955           }
1956         }
1957         previous = current;
1958       } else { // !in_pack(current, pk) ==> a sandwiched store
1959         remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1960       }
1961 
1962       if (current == first) break;
1963       current = my_mem->as_Mem();
1964     } // end while
1965 
1966     // Reconnect loads back to upper_insert_pt.
1967     for (uint i = 0; i < memops.size(); i++) {
1968       Node *ld = memops.at(i);
1969       if (ld->in(MemNode::Memory) != upper_insert_pt) {
1970         _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
1971       }
1972     }
1973   } else if (pk->at(0)->is_Load()) { //load
1974     // all loads in the pack should have the same memory state. By default,
1975     // we use the memory state of the last load. However, if any load could
1976     // not be moved down due to the dependence constraint, we use the memory
1977     // state of the first load.
1978     Node* last_mem  = executed_last(pk)->in(MemNode::Memory);
1979     Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1980     bool schedule_last = true;
1981     for (uint i = 0; i < pk->size(); i++) {
1982       Node* ld = pk->at(i);
1983       for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1984            current=current->in(MemNode::Memory)) {
1985         assert(current != first_mem, "corrupted memory graph");
1986         if(current->is_Mem() && !independent(current, ld)){
1987           schedule_last = false; // a later store depends on this load
1988           break;
1989         }
1990       }
1991     }
1992 
1993     Node* mem_input = schedule_last ? last_mem : first_mem;
1994     _igvn.hash_delete(mem_input);
1995     // Give each load the same memory state
1996     for (uint i = 0; i < pk->size(); i++) {
1997       LoadNode* ld = pk->at(i)->as_Load();
1998       _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
1999     }
2000   }
2001 }
2002 
2003 #ifndef PRODUCT
2004 void SuperWord::print_loop(bool whole) {
2005   Node_Stack stack(_arena, _phase->C->unique() >> 2);
2006   Node_List rpo_list;
2007   VectorSet visited(_arena);
2008   visited.set(lpt()->_head->_idx);
2009   _phase->rpo(lpt()->_head, stack, visited, rpo_list);
2010   _phase->dump(lpt(), rpo_list.size(), rpo_list );
2011   if(whole) {
2012     tty->print_cr("\n Whole loop tree");
2013     _phase->dump();
2014     tty->print_cr(" End of whole loop tree\n");
2015   }
2016 }
2017 #endif
2018 
2019 //------------------------------output---------------------------
2020 // Convert packs into vector node operations
2021 void SuperWord::output() {
2022   if (_packset.length() == 0) return;
2023 
2024 #ifndef PRODUCT
2025   if (TraceLoopOpts) {
2026     tty->print("SuperWord::output    ");
2027     lpt()->dump_head();
2028   }
2029 #endif
2030 
2031   // MUST ENSURE main loop's initial value is properly aligned:
2032   //  (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
2033 
2034   align_initial_loop_index(align_to_ref());
2035 
2036   // Insert extract (unpack) operations for scalar uses
2037   for (int i = 0; i < _packset.length(); i++) {
2038     insert_extracts(_packset.at(i));
2039   }
2040 
2041   Compile* C = _phase->C;
2042   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2043   uint max_vlen_in_bytes = 0;
2044   uint max_vlen = 0;
2045 
2046   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop before create_reserve_version_of_loop"); print_loop(true);})
2047 
2048   CountedLoopReserveKit make_reversable(_phase, _lpt, do_reserve_copy());
2049 
2050   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop after create_reserve_version_of_loop"); print_loop(true);})
2051 
2052   if (do_reserve_copy() && !make_reversable.has_reserved()) {
2053     NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: loop was not reserved correctly, exiting SuperWord");})
2054     return;
2055   }
2056 
2057   for (int i = 0; i < _block.length(); i++) {
2058     Node* n = _block.at(i);
2059     Node_List* p = my_pack(n);
2060     if (p && n == executed_last(p)) {
2061       uint vlen = p->size();
2062       uint vlen_in_bytes = 0;
2063       Node* vn = NULL;
2064       Node* low_adr = p->at(0);
2065       Node* first   = executed_first(p);
2066       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);})
2067       int   opc = n->Opcode();
2068       if (n->is_Load()) {
2069         Node* ctl = n->in(MemNode::Control);
2070         Node* mem = first->in(MemNode::Memory);
2071         SWPointer p1(n->as_Mem(), this, NULL, false);
2072         // Identify the memory dependency for the new loadVector node by
2073         // walking up through memory chain.
2074         // This is done to give flexibility to the new loadVector node so that
2075         // it can move above independent storeVector nodes.
2076         while (mem->is_StoreVector()) {
2077           SWPointer p2(mem->as_Mem(), this, NULL, false);
2078           int cmp = p1.cmp(p2);
2079           if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
2080             mem = mem->in(MemNode::Memory);
2081           } else {
2082             break; // dependent memory
2083           }
2084         }
2085         Node* adr = low_adr->in(MemNode::Address);
2086         const TypePtr* atyp = n->adr_type();
2087         vn = LoadVectorNode::make(opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n), control_dependency(p));
2088         vlen_in_bytes = vn->as_LoadVector()->memory_size();
2089       } else if (n->is_Store()) {
2090         // Promote value to be stored to vector
2091         Node* val = vector_opd(p, MemNode::ValueIn);
2092         if (val == NULL) {
2093           if (do_reserve_copy()) {
2094             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: val should not be NULL, exiting SuperWord");})
2095             return; //and reverse to backup IG
2096           }
2097           ShouldNotReachHere();
2098         }
2099 
2100         Node* ctl = n->in(MemNode::Control);
2101         Node* mem = first->in(MemNode::Memory);
2102         Node* adr = low_adr->in(MemNode::Address);
2103         const TypePtr* atyp = n->adr_type();
2104         vn = StoreVectorNode::make(opc, ctl, mem, adr, atyp, val, vlen);
2105         vlen_in_bytes = vn->as_StoreVector()->memory_size();
2106       } else if (n->req() == 3 && !is_cmov_pack(p)) {
2107         // Promote operands to vector
2108         Node* in1 = NULL;
2109         bool node_isa_reduction = n->is_reduction();
2110         if (node_isa_reduction) {
2111           // the input to the first reduction operation is retained
2112           in1 = low_adr->in(1);
2113         } else {
2114           in1 = vector_opd(p, 1);
2115           if (in1 == NULL) {
2116             if (do_reserve_copy()) {
2117               NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in1 should not be NULL, exiting SuperWord");})
2118               return; //and reverse to backup IG
2119             }
2120             ShouldNotReachHere();
2121           }
2122         }
2123         Node* in2 = vector_opd(p, 2);
2124         if (in2 == NULL) {
2125           if (do_reserve_copy()) {
2126             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in2 should not be NULL, exiting SuperWord");})
2127             return; //and reverse to backup IG
2128           }
2129           ShouldNotReachHere();
2130         }
2131         if (VectorNode::is_invariant_vector(in1) && (node_isa_reduction == false) && (n->is_Add() || n->is_Mul())) {
2132           // Move invariant vector input into second position to avoid register spilling.
2133           Node* tmp = in1;
2134           in1 = in2;
2135           in2 = tmp;
2136         }
2137         if (node_isa_reduction) {
2138           const Type *arith_type = n->bottom_type();
2139           vn = ReductionNode::make(opc, NULL, in1, in2, arith_type->basic_type());
2140           if (in2->is_Load()) {
2141             vlen_in_bytes = in2->as_LoadVector()->memory_size();
2142           } else {
2143             vlen_in_bytes = in2->as_Vector()->length_in_bytes();
2144           }
2145         } else {
2146           vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2147           vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2148         }
2149       } else if (opc == Op_SqrtD || opc == Op_AbsF || opc == Op_AbsD || opc == Op_NegF || opc == Op_NegD) {
2150         // Promote operand to vector (Sqrt/Abs/Neg are 2 address instructions)
2151         Node* in = vector_opd(p, 1);
2152         vn = VectorNode::make(opc, in, NULL, vlen, velt_basic_type(n));
2153         vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2154       } else if (is_cmov_pack(p)) {
2155         if (!n->is_CMove()) {
2156           continue;
2157         }
2158         // place here CMoveVDNode
2159         NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::output: print before CMove vectorization"); print_loop(false);})
2160         Node* bol = n->in(CMoveNode::Condition);
2161         if (!bol->is_Bool() && bol->Opcode() == Op_ExtractI && bol->req() > 1 ) {
2162           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();})
2163           bol = bol->in(1); //may be ExtractNode
2164         }
2165 
2166         assert(bol->is_Bool(), "should be BoolNode - too late to bail out!");
2167         if (!bol->is_Bool()) {
2168           if (do_reserve_copy()) {
2169             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: expected %d bool node, exiting SuperWord", bol->_idx); bol->dump();})
2170             return; //and reverse to backup IG
2171           }
2172           ShouldNotReachHere();
2173         }
2174 
2175         int cond = (int)bol->as_Bool()->_test._test;
2176         Node* in_cc  = _igvn.intcon(cond);
2177         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created intcon in_cc node %d", in_cc->_idx); in_cc->dump();})
2178         Node* cc = bol->clone();
2179         cc->set_req(1, in_cc);
2180         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created bool cc node %d", cc->_idx); cc->dump();})
2181 
2182         Node* src1 = vector_opd(p, 2); //2=CMoveNode::IfFalse
2183         if (src1 == NULL) {
2184           if (do_reserve_copy()) {
2185             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src1 should not be NULL, exiting SuperWord");})
2186             return; //and reverse to backup IG
2187           }
2188           ShouldNotReachHere();
2189         }
2190         Node* src2 = vector_opd(p, 3); //3=CMoveNode::IfTrue
2191         if (src2 == NULL) {
2192           if (do_reserve_copy()) {
2193             NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src2 should not be NULL, exiting SuperWord");})
2194             return; //and reverse to backup IG
2195           }
2196           ShouldNotReachHere();
2197         }
2198         BasicType bt = velt_basic_type(n);
2199         const TypeVect* vt = TypeVect::make(bt, vlen);
2200         vn = new CMoveVDNode(cc, src1, src2, vt);
2201         NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created new CMove node %d: ", vn->_idx); vn->dump();})
2202       } else {
2203         if (do_reserve_copy()) {
2204           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: ShouldNotReachHere, exiting SuperWord");})
2205           return; //and reverse to backup IG
2206         }
2207         ShouldNotReachHere();
2208       }
2209 
2210       assert(vn != NULL, "sanity");
2211       if (vn == NULL) {
2212         if (do_reserve_copy()){
2213           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: got NULL node, cannot proceed, exiting SuperWord");})
2214           return; //and reverse to backup IG
2215         }
2216         ShouldNotReachHere();
2217       }
2218 
2219       _igvn.register_new_node_with_optimizer(vn);
2220       _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
2221       for (uint j = 0; j < p->size(); j++) {
2222         Node* pm = p->at(j);
2223         _igvn.replace_node(pm, vn);
2224       }
2225       _igvn._worklist.push(vn);
2226 
2227       if (vlen_in_bytes > max_vlen_in_bytes) {
2228         max_vlen = vlen;
2229         max_vlen_in_bytes = vlen_in_bytes;
2230       }
2231 #ifdef ASSERT
2232       if (TraceNewVectors) {
2233         tty->print("new Vector node: ");
2234         vn->dump();
2235       }
2236 #endif
2237     }
2238   }//for (int i = 0; i < _block.length(); i++)
2239 
2240   C->set_max_vector_size(max_vlen_in_bytes);
2241 
2242   if (SuperWordLoopUnrollAnalysis) {
2243     if (cl->has_passed_slp()) {
2244       uint slp_max_unroll_factor = cl->slp_max_unroll();
2245       if (slp_max_unroll_factor == max_vlen) {
2246         if (TraceSuperWordLoopUnrollAnalysis) {
2247           tty->print_cr("vector loop(unroll=%d, len=%d)\n", max_vlen, max_vlen_in_bytes*BitsPerByte);
2248         }
2249         // For atomic unrolled loops which are vector mapped, instigate more unrolling.
2250         cl->set_notpassed_slp();
2251         // if vector resources are limited, do not allow additional unrolling
2252         if (FLOATPRESSURE > 8) {
2253           C->set_major_progress();
2254         }
2255         cl->mark_do_unroll_only();
2256         if (do_reserve_copy()) {
2257           cl->mark_loop_vectorized();
2258         }
2259       }
2260     }
2261   }
2262 
2263   if (do_reserve_copy()) {
2264     make_reversable.use_new();
2265   }
2266   NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("\n Final loop after SuperWord"); print_loop(true);})
2267   return;
2268 }
2269 
2270 //------------------------------vector_opd---------------------------
2271 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
2272 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
2273   Node* p0 = p->at(0);
2274   uint vlen = p->size();
2275   Node* opd = p0->in(opd_idx);
2276 
2277   if (same_inputs(p, opd_idx)) {
2278     if (opd->is_Vector() || opd->is_LoadVector()) {
2279       assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
2280       if (opd_idx == 2 && VectorNode::is_shift(p0)) {
2281         NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("shift's count can't be vector");})
2282         return NULL;
2283       }
2284       return opd; // input is matching vector
2285     }
2286     if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
2287       Compile* C = _phase->C;
2288       Node* cnt = opd;
2289       // Vector instructions do not mask shift count, do it here.
2290       juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2291       const TypeInt* t = opd->find_int_type();
2292       if (t != NULL && t->is_con()) {
2293         juint shift = t->get_con();
2294         if (shift > mask) { // Unsigned cmp
2295           cnt = ConNode::make(TypeInt::make(shift & mask));
2296         }
2297       } else {
2298         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2299           cnt = ConNode::make(TypeInt::make(mask));
2300           _igvn.register_new_node_with_optimizer(cnt);
2301           cnt = new AndINode(opd, cnt);
2302           _igvn.register_new_node_with_optimizer(cnt);
2303           _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2304         }
2305         assert(opd->bottom_type()->isa_int(), "int type only");
2306         if (!opd->bottom_type()->isa_int()) {
2307           NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should be int type only");})
2308           return NULL;
2309         }
2310         // Move non constant shift count into vector register.
2311         cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0));
2312       }
2313       if (cnt != opd) {
2314         _igvn.register_new_node_with_optimizer(cnt);
2315         _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2316       }
2317       return cnt;
2318     }
2319     assert(!opd->is_StoreVector(), "such vector is not expected here");
2320     if (opd->is_StoreVector()) {
2321       NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("StoreVector is not expected here");})
2322       return NULL;
2323     }
2324     // Convert scalar input to vector with the same number of elements as
2325     // p0's vector. Use p0's type because size of operand's container in
2326     // vector should match p0's size regardless operand's size.
2327     const Type* p0_t = velt_type(p0);
2328     VectorNode* vn = VectorNode::scalar2vector(opd, vlen, p0_t);
2329 
2330     _igvn.register_new_node_with_optimizer(vn);
2331     _phase->set_ctrl(vn, _phase->get_ctrl(opd));
2332 #ifdef ASSERT
2333     if (TraceNewVectors) {
2334       tty->print("new Vector node: ");
2335       vn->dump();
2336     }
2337 #endif
2338     return vn;
2339   }
2340 
2341   // Insert pack operation
2342   BasicType bt = velt_basic_type(p0);
2343   PackNode* pk = PackNode::make(opd, vlen, bt);
2344   DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
2345 
2346   for (uint i = 1; i < vlen; i++) {
2347     Node* pi = p->at(i);
2348     Node* in = pi->in(opd_idx);
2349     assert(my_pack(in) == NULL, "Should already have been unpacked");
2350     if (my_pack(in) != NULL) {
2351       NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should already have been unpacked");})
2352       return NULL;
2353     }
2354     assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
2355     pk->add_opd(in);
2356   }
2357   _igvn.register_new_node_with_optimizer(pk);
2358   _phase->set_ctrl(pk, _phase->get_ctrl(opd));
2359 #ifdef ASSERT
2360   if (TraceNewVectors) {
2361     tty->print("new Vector node: ");
2362     pk->dump();
2363   }
2364 #endif
2365   return pk;
2366 }
2367 
2368 //------------------------------insert_extracts---------------------------
2369 // If a use of pack p is not a vector use, then replace the
2370 // use with an extract operation.
2371 void SuperWord::insert_extracts(Node_List* p) {
2372   if (p->at(0)->is_Store()) return;
2373   assert(_n_idx_list.is_empty(), "empty (node,index) list");
2374 
2375   // Inspect each use of each pack member.  For each use that is
2376   // not a vector use, replace the use with an extract operation.
2377 
2378   for (uint i = 0; i < p->size(); i++) {
2379     Node* def = p->at(i);
2380     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2381       Node* use = def->fast_out(j);
2382       for (uint k = 0; k < use->req(); k++) {
2383         Node* n = use->in(k);
2384         if (def == n) {
2385           Node_List* u_pk = my_pack(use);
2386           if ((u_pk == NULL || !is_cmov_pack(u_pk) || use->is_CMove()) && !is_vector_use(use, k)) {
2387               _n_idx_list.push(use, k);
2388           }
2389         }
2390       }
2391     }
2392   }
2393 
2394   while (_n_idx_list.is_nonempty()) {
2395     Node* use = _n_idx_list.node();
2396     int   idx = _n_idx_list.index();
2397     _n_idx_list.pop();
2398     Node* def = use->in(idx);
2399 
2400     if (def->is_reduction()) continue;
2401 
2402     // Insert extract operation
2403     _igvn.hash_delete(def);
2404     int def_pos = alignment(def) / data_size(def);
2405 
2406     Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def));
2407     _igvn.register_new_node_with_optimizer(ex);
2408     _phase->set_ctrl(ex, _phase->get_ctrl(def));
2409     _igvn.replace_input_of(use, idx, ex);
2410     _igvn._worklist.push(def);
2411 
2412     bb_insert_after(ex, bb_idx(def));
2413     set_velt_type(ex, velt_type(def));
2414   }
2415 }
2416 
2417 //------------------------------is_vector_use---------------------------
2418 // Is use->in(u_idx) a vector use?
2419 bool SuperWord::is_vector_use(Node* use, int u_idx) {
2420   Node_List* u_pk = my_pack(use);
2421   if (u_pk == NULL) return false;
2422   if (use->is_reduction()) return true;
2423   Node* def = use->in(u_idx);
2424   Node_List* d_pk = my_pack(def);
2425   if (d_pk == NULL) {
2426     // check for scalar promotion
2427     Node* n = u_pk->at(0)->in(u_idx);
2428     for (uint i = 1; i < u_pk->size(); i++) {
2429       if (u_pk->at(i)->in(u_idx) != n) return false;
2430     }
2431     return true;
2432   }
2433   if (u_pk->size() != d_pk->size())
2434     return false;
2435   for (uint i = 0; i < u_pk->size(); i++) {
2436     Node* ui = u_pk->at(i);
2437     Node* di = d_pk->at(i);
2438     if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
2439       return false;
2440   }
2441   return true;
2442 }
2443 
2444 //------------------------------construct_bb---------------------------
2445 // Construct reverse postorder list of block members
2446 bool SuperWord::construct_bb() {
2447   Node* entry = bb();
2448 
2449   assert(_stk.length() == 0,            "stk is empty");
2450   assert(_block.length() == 0,          "block is empty");
2451   assert(_data_entry.length() == 0,     "data_entry is empty");
2452   assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
2453   assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
2454 
2455   // Find non-control nodes with no inputs from within block,
2456   // create a temporary map from node _idx to bb_idx for use
2457   // by the visited and post_visited sets,
2458   // and count number of nodes in block.
2459   int bb_ct = 0;
2460   for (uint i = 0; i < lpt()->_body.size(); i++) {
2461     Node *n = lpt()->_body.at(i);
2462     set_bb_idx(n, i); // Create a temporary map
2463     if (in_bb(n)) {
2464       if (n->is_LoadStore() || n->is_MergeMem() ||
2465           (n->is_Proj() && !n->as_Proj()->is_CFG())) {
2466         // Bailout if the loop has LoadStore, MergeMem or data Proj
2467         // nodes. Superword optimization does not work with them.
2468         return false;
2469       }
2470       bb_ct++;
2471       if (!n->is_CFG()) {
2472         bool found = false;
2473         for (uint j = 0; j < n->req(); j++) {
2474           Node* def = n->in(j);
2475           if (def && in_bb(def)) {
2476             found = true;
2477             break;
2478           }
2479         }
2480         if (!found) {
2481           assert(n != entry, "can't be entry");
2482           _data_entry.push(n);
2483         }
2484       }
2485     }
2486   }
2487 
2488   // Find memory slices (head and tail)
2489   for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
2490     Node *n = lp()->fast_out(i);
2491     if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
2492       Node* n_tail  = n->in(LoopNode::LoopBackControl);
2493       if (n_tail != n->in(LoopNode::EntryControl)) {
2494         if (!n_tail->is_Mem()) {
2495           assert(n_tail->is_Mem(), "unexpected node for memory slice: %s", n_tail->Name());
2496           return false; // Bailout
2497         }
2498         _mem_slice_head.push(n);
2499         _mem_slice_tail.push(n_tail);
2500       }
2501     }
2502   }
2503 
2504   // Create an RPO list of nodes in block
2505 
2506   visited_clear();
2507   post_visited_clear();
2508 
2509   // Push all non-control nodes with no inputs from within block, then control entry
2510   for (int j = 0; j < _data_entry.length(); j++) {
2511     Node* n = _data_entry.at(j);
2512     visited_set(n);
2513     _stk.push(n);
2514   }
2515   visited_set(entry);
2516   _stk.push(entry);
2517 
2518   // Do a depth first walk over out edges
2519   int rpo_idx = bb_ct - 1;
2520   int size;
2521   int reduction_uses = 0;
2522   while ((size = _stk.length()) > 0) {
2523     Node* n = _stk.top(); // Leave node on stack
2524     if (!visited_test_set(n)) {
2525       // forward arc in graph
2526     } else if (!post_visited_test(n)) {
2527       // cross or back arc
2528       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2529         Node *use = n->fast_out(i);
2530         if (in_bb(use) && !visited_test(use) &&
2531             // Don't go around backedge
2532             (!use->is_Phi() || n == entry)) {
2533           if (use->is_reduction()) {
2534             // First see if we can map the reduction on the given system we are on, then
2535             // make a data entry operation for each reduction we see.
2536             BasicType bt = use->bottom_type()->basic_type();
2537             if (ReductionNode::implemented(use->Opcode(), Matcher::min_vector_size(bt), bt)) {
2538               reduction_uses++;
2539             }
2540           }
2541           _stk.push(use);
2542         }
2543       }
2544       if (_stk.length() == size) {
2545         // There were no additional uses, post visit node now
2546         _stk.pop(); // Remove node from stack
2547         assert(rpo_idx >= 0, "");
2548         _block.at_put_grow(rpo_idx, n);
2549         rpo_idx--;
2550         post_visited_set(n);
2551         assert(rpo_idx >= 0 || _stk.is_empty(), "");
2552       }
2553     } else {
2554       _stk.pop(); // Remove post-visited node from stack
2555     }
2556   }//while
2557 
2558   int ii_current = -1;
2559   unsigned int load_idx = (unsigned int)-1;
2560   _ii_order.clear();
2561   // Create real map of block indices for nodes
2562   for (int j = 0; j < _block.length(); j++) {
2563     Node* n = _block.at(j);
2564     set_bb_idx(n, j);
2565     if (_do_vector_loop && n->is_Load()) {
2566       if (ii_current == -1) {
2567         ii_current = _clone_map.gen(n->_idx);
2568         _ii_order.push(ii_current);
2569         load_idx = _clone_map.idx(n->_idx);
2570       } else if (_clone_map.idx(n->_idx) == load_idx && _clone_map.gen(n->_idx) != ii_current) {
2571         ii_current = _clone_map.gen(n->_idx);
2572         _ii_order.push(ii_current);
2573       }
2574     }
2575   }//for
2576 
2577   // Ensure extra info is allocated.
2578   initialize_bb();
2579 
2580 #ifndef PRODUCT
2581   if (_vector_loop_debug && _ii_order.length() > 0) {
2582     tty->print("SuperWord::construct_bb: List of generations: ");
2583     for (int jj = 0; jj < _ii_order.length(); ++jj) {
2584       tty->print("  %d:%d", jj, _ii_order.at(jj));
2585     }
2586     tty->print_cr(" ");
2587   }
2588   if (TraceSuperWord) {
2589     print_bb();
2590     tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
2591     for (int m = 0; m < _data_entry.length(); m++) {
2592       tty->print("%3d ", m);
2593       _data_entry.at(m)->dump();
2594     }
2595     tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
2596     for (int m = 0; m < _mem_slice_head.length(); m++) {
2597       tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
2598       tty->print("    ");    _mem_slice_tail.at(m)->dump();
2599     }
2600   }
2601 #endif
2602   assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
2603   return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0);
2604 }
2605 
2606 //------------------------------initialize_bb---------------------------
2607 // Initialize per node info
2608 void SuperWord::initialize_bb() {
2609   Node* last = _block.at(_block.length() - 1);
2610   grow_node_info(bb_idx(last));
2611 }
2612 
2613 //------------------------------bb_insert_after---------------------------
2614 // Insert n into block after pos
2615 void SuperWord::bb_insert_after(Node* n, int pos) {
2616   int n_pos = pos + 1;
2617   // Make room
2618   for (int i = _block.length() - 1; i >= n_pos; i--) {
2619     _block.at_put_grow(i+1, _block.at(i));
2620   }
2621   for (int j = _node_info.length() - 1; j >= n_pos; j--) {
2622     _node_info.at_put_grow(j+1, _node_info.at(j));
2623   }
2624   // Set value
2625   _block.at_put_grow(n_pos, n);
2626   _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
2627   // Adjust map from node->_idx to _block index
2628   for (int i = n_pos; i < _block.length(); i++) {
2629     set_bb_idx(_block.at(i), i);
2630   }
2631 }
2632 
2633 //------------------------------compute_max_depth---------------------------
2634 // Compute max depth for expressions from beginning of block
2635 // Use to prune search paths during test for independence.
2636 void SuperWord::compute_max_depth() {
2637   int ct = 0;
2638   bool again;
2639   do {
2640     again = false;
2641     for (int i = 0; i < _block.length(); i++) {
2642       Node* n = _block.at(i);
2643       if (!n->is_Phi()) {
2644         int d_orig = depth(n);
2645         int d_in   = 0;
2646         for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
2647           Node* pred = preds.current();
2648           if (in_bb(pred)) {
2649             d_in = MAX2(d_in, depth(pred));
2650           }
2651         }
2652         if (d_in + 1 != d_orig) {
2653           set_depth(n, d_in + 1);
2654           again = true;
2655         }
2656       }
2657     }
2658     ct++;
2659   } while (again);
2660 
2661   if (TraceSuperWord && Verbose) {
2662     tty->print_cr("compute_max_depth iterated: %d times", ct);
2663   }
2664 }
2665 
2666 //-------------------------compute_vector_element_type-----------------------
2667 // Compute necessary vector element type for expressions
2668 // This propagates backwards a narrower integer type when the
2669 // upper bits of the value are not needed.
2670 // Example:  char a,b,c;  a = b + c;
2671 // Normally the type of the add is integer, but for packed character
2672 // operations the type of the add needs to be char.
2673 void SuperWord::compute_vector_element_type() {
2674   if (TraceSuperWord && Verbose) {
2675     tty->print_cr("\ncompute_velt_type:");
2676   }
2677 
2678   // Initial type
2679   for (int i = 0; i < _block.length(); i++) {
2680     Node* n = _block.at(i);
2681     set_velt_type(n, container_type(n));
2682   }
2683 
2684   // Propagate integer narrowed type backwards through operations
2685   // that don't depend on higher order bits
2686   for (int i = _block.length() - 1; i >= 0; i--) {
2687     Node* n = _block.at(i);
2688     // Only integer types need be examined
2689     const Type* vtn = velt_type(n);
2690     if (vtn->basic_type() == T_INT) {
2691       uint start, end;
2692       VectorNode::vector_operands(n, &start, &end);
2693 
2694       for (uint j = start; j < end; j++) {
2695         Node* in  = n->in(j);
2696         // Don't propagate through a memory
2697         if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
2698             data_size(n) < data_size(in)) {
2699           bool same_type = true;
2700           for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
2701             Node *use = in->fast_out(k);
2702             if (!in_bb(use) || !same_velt_type(use, n)) {
2703               same_type = false;
2704               break;
2705             }
2706           }
2707           if (same_type) {
2708             // For right shifts of small integer types (bool, byte, char, short)
2709             // we need precise information about sign-ness. Only Load nodes have
2710             // this information because Store nodes are the same for signed and
2711             // unsigned values. And any arithmetic operation after a load may
2712             // expand a value to signed Int so such right shifts can't be used
2713             // because vector elements do not have upper bits of Int.
2714             const Type* vt = vtn;
2715             if (VectorNode::is_shift(in)) {
2716               Node* load = in->in(1);
2717               if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
2718                 vt = velt_type(load);
2719               } else if (in->Opcode() != Op_LShiftI) {
2720                 // Widen type to Int to avoid creation of right shift vector
2721                 // (align + data_size(s1) check in stmts_can_pack() will fail).
2722                 // Note, left shifts work regardless type.
2723                 vt = TypeInt::INT;
2724               }
2725             }
2726             set_velt_type(in, vt);
2727           }
2728         }
2729       }
2730     }
2731   }
2732 #ifndef PRODUCT
2733   if (TraceSuperWord && Verbose) {
2734     for (int i = 0; i < _block.length(); i++) {
2735       Node* n = _block.at(i);
2736       velt_type(n)->dump();
2737       tty->print("\t");
2738       n->dump();
2739     }
2740   }
2741 #endif
2742 }
2743 
2744 //------------------------------memory_alignment---------------------------
2745 // Alignment within a vector memory reference
2746 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
2747   #ifndef PRODUCT
2748     if(TraceSuperWord && Verbose) {
2749       tty->print("SuperWord::memory_alignment within a vector memory reference for %d:  ", s->_idx); s->dump();
2750     }
2751   #endif
2752   NOT_PRODUCT(SWPointer::Tracer::Depth ddd(0);)
2753   SWPointer p(s, this, NULL, false);
2754   if (!p.valid()) {
2755     NOT_PRODUCT(if(is_trace_alignment()) tty->print("SWPointer::memory_alignment: SWPointer p invalid, return bottom_align");)
2756     return bottom_align;
2757   }
2758   int vw = vector_width_in_bytes(s);
2759   if (vw < 2) {
2760     NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SWPointer::memory_alignment: vector_width_in_bytes < 2, return bottom_align");)
2761     return bottom_align; // No vectors for this type
2762   }
2763   int offset  = p.offset_in_bytes();
2764   offset     += iv_adjust*p.memory_size();
2765   int off_rem = offset % vw;
2766   int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
2767   if (TraceSuperWord && Verbose) {
2768     tty->print_cr("SWPointer::memory_alignment: off_rem = %d, off_mod = %d", off_rem, off_mod);
2769   }
2770   return off_mod;
2771 }
2772 
2773 //---------------------------container_type---------------------------
2774 // Smallest type containing range of values
2775 const Type* SuperWord::container_type(Node* n) {
2776   if (n->is_Mem()) {
2777     BasicType bt = n->as_Mem()->memory_type();
2778     if (n->is_Store() && (bt == T_CHAR)) {
2779       // Use T_SHORT type instead of T_CHAR for stored values because any
2780       // preceding arithmetic operation extends values to signed Int.
2781       bt = T_SHORT;
2782     }
2783     if (n->Opcode() == Op_LoadUB) {
2784       // Adjust type for unsigned byte loads, it is important for right shifts.
2785       // T_BOOLEAN is used because there is no basic type representing type
2786       // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
2787       // size (one byte) and sign is important.
2788       bt = T_BOOLEAN;
2789     }
2790     return Type::get_const_basic_type(bt);
2791   }
2792   const Type* t = _igvn.type(n);
2793   if (t->basic_type() == T_INT) {
2794     // A narrow type of arithmetic operations will be determined by
2795     // propagating the type of memory operations.
2796     return TypeInt::INT;
2797   }
2798   return t;
2799 }
2800 
2801 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
2802   const Type* vt1 = velt_type(n1);
2803   const Type* vt2 = velt_type(n2);
2804   if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
2805     // Compare vectors element sizes for integer types.
2806     return data_size(n1) == data_size(n2);
2807   }
2808   return vt1 == vt2;
2809 }
2810 
2811 //------------------------------in_packset---------------------------
2812 // Are s1 and s2 in a pack pair and ordered as s1,s2?
2813 bool SuperWord::in_packset(Node* s1, Node* s2) {
2814   for (int i = 0; i < _packset.length(); i++) {
2815     Node_List* p = _packset.at(i);
2816     assert(p->size() == 2, "must be");
2817     if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
2818       return true;
2819     }
2820   }
2821   return false;
2822 }
2823 
2824 //------------------------------in_pack---------------------------
2825 // Is s in pack p?
2826 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
2827   for (uint i = 0; i < p->size(); i++) {
2828     if (p->at(i) == s) {
2829       return p;
2830     }
2831   }
2832   return NULL;
2833 }
2834 
2835 //------------------------------remove_pack_at---------------------------
2836 // Remove the pack at position pos in the packset
2837 void SuperWord::remove_pack_at(int pos) {
2838   Node_List* p = _packset.at(pos);
2839   for (uint i = 0; i < p->size(); i++) {
2840     Node* s = p->at(i);
2841     set_my_pack(s, NULL);
2842   }
2843   _packset.remove_at(pos);
2844 }
2845 
2846 void SuperWord::packset_sort(int n) {
2847   // simple bubble sort so that we capitalize with O(n) when its already sorted
2848   while (n != 0) {
2849     bool swapped = false;
2850     for (int i = 1; i < n; i++) {
2851       Node_List* q_low = _packset.at(i-1);
2852       Node_List* q_i = _packset.at(i);
2853 
2854       // only swap when we find something to swap
2855       if (alignment(q_low->at(0)) > alignment(q_i->at(0))) {
2856         Node_List* t = q_i;
2857         *(_packset.adr_at(i)) = q_low;
2858         *(_packset.adr_at(i-1)) = q_i;
2859         swapped = true;
2860       }
2861     }
2862     if (swapped == false) break;
2863     n--;
2864   }
2865 }
2866 
2867 //------------------------------executed_first---------------------------
2868 // Return the node executed first in pack p.  Uses the RPO block list
2869 // to determine order.
2870 Node* SuperWord::executed_first(Node_List* p) {
2871   Node* n = p->at(0);
2872   int n_rpo = bb_idx(n);
2873   for (uint i = 1; i < p->size(); i++) {
2874     Node* s = p->at(i);
2875     int s_rpo = bb_idx(s);
2876     if (s_rpo < n_rpo) {
2877       n = s;
2878       n_rpo = s_rpo;
2879     }
2880   }
2881   return n;
2882 }
2883 
2884 //------------------------------executed_last---------------------------
2885 // Return the node executed last in pack p.
2886 Node* SuperWord::executed_last(Node_List* p) {
2887   Node* n = p->at(0);
2888   int n_rpo = bb_idx(n);
2889   for (uint i = 1; i < p->size(); i++) {
2890     Node* s = p->at(i);
2891     int s_rpo = bb_idx(s);
2892     if (s_rpo > n_rpo) {
2893       n = s;
2894       n_rpo = s_rpo;
2895     }
2896   }
2897   return n;
2898 }
2899 
2900 LoadNode::ControlDependency SuperWord::control_dependency(Node_List* p) {
2901   LoadNode::ControlDependency dep = LoadNode::DependsOnlyOnTest;
2902   for (uint i = 0; i < p->size(); i++) {
2903     Node* n = p->at(i);
2904     assert(n->is_Load(), "only meaningful for loads");
2905     if (!n->depends_only_on_test()) {
2906       dep = LoadNode::Pinned;
2907     }
2908   }
2909   return dep;
2910 }
2911 
2912 
2913 //----------------------------align_initial_loop_index---------------------------
2914 // Adjust pre-loop limit so that in main loop, a load/store reference
2915 // to align_to_ref will be a position zero in the vector.
2916 //   (iv + k) mod vector_align == 0
2917 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
2918   CountedLoopNode *main_head = lp()->as_CountedLoop();
2919   assert(main_head->is_main_loop(), "");
2920   CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
2921   assert(pre_end != NULL, "we must have a correct pre-loop");
2922   Node *pre_opaq1 = pre_end->limit();
2923   assert(pre_opaq1->Opcode() == Op_Opaque1, "");
2924   Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2925   Node *lim0 = pre_opaq->in(1);
2926 
2927   // Where we put new limit calculations
2928   Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2929 
2930   // Ensure the original loop limit is available from the
2931   // pre-loop Opaque1 node.
2932   Node *orig_limit = pre_opaq->original_loop_limit();
2933   assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
2934 
2935   SWPointer align_to_ref_p(align_to_ref, this, NULL, false);
2936   assert(align_to_ref_p.valid(), "sanity");
2937 
2938   // Given:
2939   //     lim0 == original pre loop limit
2940   //     V == v_align (power of 2)
2941   //     invar == extra invariant piece of the address expression
2942   //     e == offset [ +/- invar ]
2943   //
2944   // When reassociating expressions involving '%' the basic rules are:
2945   //     (a - b) % k == 0   =>  a % k == b % k
2946   // and:
2947   //     (a + b) % k == 0   =>  a % k == (k - b) % k
2948   //
2949   // For stride > 0 && scale > 0,
2950   //   Derive the new pre-loop limit "lim" such that the two constraints:
2951   //     (1) lim = lim0 + N           (where N is some positive integer < V)
2952   //     (2) (e + lim) % V == 0
2953   //   are true.
2954   //
2955   //   Substituting (1) into (2),
2956   //     (e + lim0 + N) % V == 0
2957   //   solve for N:
2958   //     N = (V - (e + lim0)) % V
2959   //   substitute back into (1), so that new limit
2960   //     lim = lim0 + (V - (e + lim0)) % V
2961   //
2962   // For stride > 0 && scale < 0
2963   //   Constraints:
2964   //     lim = lim0 + N
2965   //     (e - lim) % V == 0
2966   //   Solving for lim:
2967   //     (e - lim0 - N) % V == 0
2968   //     N = (e - lim0) % V
2969   //     lim = lim0 + (e - lim0) % V
2970   //
2971   // For stride < 0 && scale > 0
2972   //   Constraints:
2973   //     lim = lim0 - N
2974   //     (e + lim) % V == 0
2975   //   Solving for lim:
2976   //     (e + lim0 - N) % V == 0
2977   //     N = (e + lim0) % V
2978   //     lim = lim0 - (e + lim0) % V
2979   //
2980   // For stride < 0 && scale < 0
2981   //   Constraints:
2982   //     lim = lim0 - N
2983   //     (e - lim) % V == 0
2984   //   Solving for lim:
2985   //     (e - lim0 + N) % V == 0
2986   //     N = (V - (e - lim0)) % V
2987   //     lim = lim0 - (V - (e - lim0)) % V
2988 
2989   int vw = vector_width_in_bytes(align_to_ref);
2990   int stride   = iv_stride();
2991   int scale    = align_to_ref_p.scale_in_bytes();
2992   int elt_size = align_to_ref_p.memory_size();
2993   int v_align  = vw / elt_size;
2994   assert(v_align > 1, "sanity");
2995   int offset   = align_to_ref_p.offset_in_bytes() / elt_size;
2996   Node *offsn  = _igvn.intcon(offset);
2997 
2998   Node *e = offsn;
2999   if (align_to_ref_p.invar() != NULL) {
3000     // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
3001     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3002     Node* aref     = new URShiftINode(align_to_ref_p.invar(), log2_elt);
3003     _igvn.register_new_node_with_optimizer(aref);
3004     _phase->set_ctrl(aref, pre_ctrl);
3005     if (align_to_ref_p.negate_invar()) {
3006       e = new SubINode(e, aref);
3007     } else {
3008       e = new AddINode(e, aref);
3009     }
3010     _igvn.register_new_node_with_optimizer(e);
3011     _phase->set_ctrl(e, pre_ctrl);
3012   }
3013   if (vw > ObjectAlignmentInBytes) {
3014     // incorporate base e +/- base && Mask >>> log2(elt)
3015     Node* xbase = new CastP2XNode(NULL, align_to_ref_p.base());
3016     _igvn.register_new_node_with_optimizer(xbase);
3017 #ifdef _LP64
3018     xbase  = new ConvL2INode(xbase);
3019     _igvn.register_new_node_with_optimizer(xbase);
3020 #endif
3021     Node* mask = _igvn.intcon(vw-1);
3022     Node* masked_xbase  = new AndINode(xbase, mask);
3023     _igvn.register_new_node_with_optimizer(masked_xbase);
3024     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3025     Node* bref     = new URShiftINode(masked_xbase, log2_elt);
3026     _igvn.register_new_node_with_optimizer(bref);
3027     _phase->set_ctrl(bref, pre_ctrl);
3028     e = new AddINode(e, bref);
3029     _igvn.register_new_node_with_optimizer(e);
3030     _phase->set_ctrl(e, pre_ctrl);
3031   }
3032 
3033   // compute e +/- lim0
3034   if (scale < 0) {
3035     e = new SubINode(e, lim0);
3036   } else {
3037     e = new AddINode(e, lim0);
3038   }
3039   _igvn.register_new_node_with_optimizer(e);
3040   _phase->set_ctrl(e, pre_ctrl);
3041 
3042   if (stride * scale > 0) {
3043     // compute V - (e +/- lim0)
3044     Node* va  = _igvn.intcon(v_align);
3045     e = new SubINode(va, e);
3046     _igvn.register_new_node_with_optimizer(e);
3047     _phase->set_ctrl(e, pre_ctrl);
3048   }
3049   // compute N = (exp) % V
3050   Node* va_msk = _igvn.intcon(v_align - 1);
3051   Node* N = new AndINode(e, va_msk);
3052   _igvn.register_new_node_with_optimizer(N);
3053   _phase->set_ctrl(N, pre_ctrl);
3054 
3055   //   substitute back into (1), so that new limit
3056   //     lim = lim0 + N
3057   Node* lim;
3058   if (stride < 0) {
3059     lim = new SubINode(lim0, N);
3060   } else {
3061     lim = new AddINode(lim0, N);
3062   }
3063   _igvn.register_new_node_with_optimizer(lim);
3064   _phase->set_ctrl(lim, pre_ctrl);
3065   Node* constrained =
3066     (stride > 0) ? (Node*) new MinINode(lim, orig_limit)
3067                  : (Node*) new MaxINode(lim, orig_limit);
3068   _igvn.register_new_node_with_optimizer(constrained);
3069   _phase->set_ctrl(constrained, pre_ctrl);
3070   _igvn.hash_delete(pre_opaq);
3071   pre_opaq->set_req(1, constrained);
3072 }
3073 
3074 //----------------------------get_pre_loop_end---------------------------
3075 // Find pre loop end from main loop.  Returns null if none.
3076 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode* cl) {
3077   Node* ctrl = cl->in(LoopNode::EntryControl);
3078   if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
3079   Node* iffm = ctrl->in(0);
3080   if (!iffm->is_If()) return NULL;
3081   Node* bolzm = iffm->in(1);
3082   if (!bolzm->is_Bool()) return NULL;
3083   Node* cmpzm = bolzm->in(1);
3084   if (!cmpzm->is_Cmp()) return NULL;
3085   Node* opqzm = cmpzm->in(2);
3086   // Can not optimize a loop if zero-trip Opaque1 node is optimized
3087   // away and then another round of loop opts attempted.
3088   if (opqzm->Opcode() != Op_Opaque1) {
3089     return NULL;
3090   }
3091   Node* p_f = iffm->in(0);
3092   if (!p_f->is_IfFalse()) return NULL;
3093   if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
3094   CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
3095   CountedLoopNode* loop_node = pre_end->loopnode();
3096   if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
3097   return pre_end;
3098 }
3099 
3100 
3101 //------------------------------init---------------------------
3102 void SuperWord::init() {
3103   _dg.init();
3104   _packset.clear();
3105   _disjoint_ptrs.clear();
3106   _block.clear();
3107   _data_entry.clear();
3108   _mem_slice_head.clear();
3109   _mem_slice_tail.clear();
3110   _iteration_first.clear();
3111   _iteration_last.clear();
3112   _node_info.clear();
3113   _align_to_ref = NULL;
3114   _lpt = NULL;
3115   _lp = NULL;
3116   _bb = NULL;
3117   _iv = NULL;
3118   _race_possible = 0;
3119   _early_return = false;
3120   _num_work_vecs = 0;
3121   _num_reductions = 0;
3122 }
3123 
3124 //------------------------------restart---------------------------
3125 void SuperWord::restart() {
3126   _dg.init();
3127   _packset.clear();
3128   _disjoint_ptrs.clear();
3129   _block.clear();
3130   _data_entry.clear();
3131   _mem_slice_head.clear();
3132   _mem_slice_tail.clear();
3133   _node_info.clear();
3134 }
3135 
3136 //------------------------------print_packset---------------------------
3137 void SuperWord::print_packset() {
3138 #ifndef PRODUCT
3139   tty->print_cr("packset");
3140   for (int i = 0; i < _packset.length(); i++) {
3141     tty->print_cr("Pack: %d", i);
3142     Node_List* p = _packset.at(i);
3143     print_pack(p);
3144   }
3145 #endif
3146 }
3147 
3148 //------------------------------print_pack---------------------------
3149 void SuperWord::print_pack(Node_List* p) {
3150   for (uint i = 0; i < p->size(); i++) {
3151     print_stmt(p->at(i));
3152   }
3153 }
3154 
3155 //------------------------------print_bb---------------------------
3156 void SuperWord::print_bb() {
3157 #ifndef PRODUCT
3158   tty->print_cr("\nBlock");
3159   for (int i = 0; i < _block.length(); i++) {
3160     Node* n = _block.at(i);
3161     tty->print("%d ", i);
3162     if (n) {
3163       n->dump();
3164     }
3165   }
3166 #endif
3167 }
3168 
3169 //------------------------------print_stmt---------------------------
3170 void SuperWord::print_stmt(Node* s) {
3171 #ifndef PRODUCT
3172   tty->print(" align: %d \t", alignment(s));
3173   s->dump();
3174 #endif
3175 }
3176 
3177 //------------------------------blank---------------------------
3178 char* SuperWord::blank(uint depth) {
3179   static char blanks[101];
3180   assert(depth < 101, "too deep");
3181   for (uint i = 0; i < depth; i++) blanks[i] = ' ';
3182   blanks[depth] = '\0';
3183   return blanks;
3184 }
3185 
3186 
3187 //==============================SWPointer===========================
3188 #ifndef PRODUCT
3189 int SWPointer::Tracer::_depth = 0;
3190 #endif
3191 //----------------------------SWPointer------------------------
3192 SWPointer::SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only) :
3193   _mem(mem), _slp(slp),  _base(NULL),  _adr(NULL),
3194   _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3195   _nstack(nstack), _analyze_only(analyze_only),
3196   _stack_idx(0)
3197 #ifndef PRODUCT
3198   , _tracer(slp)
3199 #endif
3200 {
3201   NOT_PRODUCT(_tracer.ctor_1(mem);)
3202 
3203   Node* adr = mem->in(MemNode::Address);
3204   if (!adr->is_AddP()) {
3205     assert(!valid(), "too complex");
3206     return;
3207   }
3208   // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
3209   Node* base = adr->in(AddPNode::Base);
3210   // The base address should be loop invariant
3211   if (!invariant(base)) {
3212     assert(!valid(), "base address is loop variant");
3213     return;
3214   }
3215   //unsafe reference could not be aligned appropriately without runtime checking
3216   if (base == NULL || base->bottom_type() == Type::TOP) {
3217     assert(!valid(), "unsafe access");
3218     return;
3219   }
3220 
3221   NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.store_depth();)
3222   NOT_PRODUCT(_tracer.ctor_2(adr);)
3223 
3224   int i;
3225   for (i = 0; i < 3; i++) {
3226     NOT_PRODUCT(_tracer.ctor_3(adr, i);)
3227 
3228     if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
3229       assert(!valid(), "too complex");
3230       return;
3231     }
3232     adr = adr->in(AddPNode::Address);
3233     NOT_PRODUCT(_tracer.ctor_4(adr, i);)
3234 
3235     if (base == adr || !adr->is_AddP()) {
3236       NOT_PRODUCT(_tracer.ctor_5(adr, base, i);)
3237       break; // stop looking at addp's
3238     }
3239   }
3240   NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.restore_depth();)
3241   NOT_PRODUCT(_tracer.ctor_6(mem);)
3242 
3243   _base = base;
3244   _adr  = adr;
3245   assert(valid(), "Usable");
3246 }
3247 
3248 // Following is used to create a temporary object during
3249 // the pattern match of an address expression.
3250 SWPointer::SWPointer(SWPointer* p) :
3251   _mem(p->_mem), _slp(p->_slp),  _base(NULL),  _adr(NULL),
3252   _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3253   _nstack(p->_nstack), _analyze_only(p->_analyze_only),
3254   _stack_idx(p->_stack_idx)
3255   #ifndef PRODUCT
3256   , _tracer(p->_slp)
3257   #endif
3258 {}
3259 
3260 
3261 bool SWPointer::invariant(Node* n) {
3262   NOT_PRODUCT(Tracer::Depth dd;)
3263   Node *n_c = phase()->get_ctrl(n);
3264   NOT_PRODUCT(_tracer.invariant_1(n, n_c);)
3265   return !lpt()->is_member(phase()->get_loop(n_c));
3266 }
3267 //------------------------scaled_iv_plus_offset--------------------
3268 // Match: k*iv + offset
3269 // where: k is a constant that maybe zero, and
3270 //        offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
3271 bool SWPointer::scaled_iv_plus_offset(Node* n) {
3272   NOT_PRODUCT(Tracer::Depth ddd;)
3273   NOT_PRODUCT(_tracer.scaled_iv_plus_offset_1(n);)
3274 
3275   if (scaled_iv(n)) {
3276     NOT_PRODUCT(_tracer.scaled_iv_plus_offset_2(n);)
3277     return true;
3278   }
3279 
3280   if (offset_plus_k(n)) {
3281     NOT_PRODUCT(_tracer.scaled_iv_plus_offset_3(n);)
3282     return true;
3283   }
3284 
3285   int opc = n->Opcode();
3286   if (opc == Op_AddI) {
3287     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
3288       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_4(n);)
3289       return true;
3290     }
3291     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3292       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_5(n);)
3293       return true;
3294     }
3295   } else if (opc == Op_SubI) {
3296     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
3297       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_6(n);)
3298       return true;
3299     }
3300     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3301       _scale *= -1;
3302       NOT_PRODUCT(_tracer.scaled_iv_plus_offset_7(n);)
3303       return true;
3304     }
3305   }
3306 
3307   NOT_PRODUCT(_tracer.scaled_iv_plus_offset_8(n);)
3308   return false;
3309 }
3310 
3311 //----------------------------scaled_iv------------------------
3312 // Match: k*iv where k is a constant that's not zero
3313 bool SWPointer::scaled_iv(Node* n) {
3314   NOT_PRODUCT(Tracer::Depth ddd;)
3315   NOT_PRODUCT(_tracer.scaled_iv_1(n);)
3316 
3317   if (_scale != 0) { // already found a scale
3318     NOT_PRODUCT(_tracer.scaled_iv_2(n, _scale);)
3319     return false;
3320   }
3321 
3322   if (n == iv()) {
3323     _scale = 1;
3324     NOT_PRODUCT(_tracer.scaled_iv_3(n, _scale);)
3325     return true;
3326   }
3327   if (_analyze_only && (invariant(n) == false)) {
3328     _nstack->push(n, _stack_idx++);
3329   }
3330 
3331   int opc = n->Opcode();
3332   if (opc == Op_MulI) {
3333     if (n->in(1) == iv() && n->in(2)->is_Con()) {
3334       _scale = n->in(2)->get_int();
3335       NOT_PRODUCT(_tracer.scaled_iv_4(n, _scale);)
3336       return true;
3337     } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
3338       _scale = n->in(1)->get_int();
3339       NOT_PRODUCT(_tracer.scaled_iv_5(n, _scale);)
3340       return true;
3341     }
3342   } else if (opc == Op_LShiftI) {
3343     if (n->in(1) == iv() && n->in(2)->is_Con()) {
3344       _scale = 1 << n->in(2)->get_int();
3345       NOT_PRODUCT(_tracer.scaled_iv_6(n, _scale);)
3346       return true;
3347     }
3348   } else if (opc == Op_ConvI2L) {
3349     if (n->in(1)->Opcode() == Op_CastII &&
3350         n->in(1)->as_CastII()->has_range_check()) {
3351       // Skip range check dependent CastII nodes
3352       n = n->in(1);
3353     }
3354     if (scaled_iv_plus_offset(n->in(1))) {
3355       NOT_PRODUCT(_tracer.scaled_iv_7(n);)
3356       return true;
3357     }
3358   } else if (opc == Op_LShiftL) {
3359     if (!has_iv() && _invar == NULL) {
3360       // Need to preserve the current _offset value, so
3361       // create a temporary object for this expression subtree.
3362       // Hacky, so should re-engineer the address pattern match.
3363       NOT_PRODUCT(Tracer::Depth dddd;)
3364       SWPointer tmp(this);
3365       NOT_PRODUCT(_tracer.scaled_iv_8(n, &tmp);)
3366 
3367       if (tmp.scaled_iv_plus_offset(n->in(1))) {
3368         if (tmp._invar == NULL || _slp->do_vector_loop()) {
3369           int mult = 1 << n->in(2)->get_int();
3370           _scale   = tmp._scale  * mult;
3371           _offset += tmp._offset * mult;
3372           NOT_PRODUCT(_tracer.scaled_iv_9(n, _scale, _offset, mult);)
3373           return true;
3374         }
3375       }
3376     }
3377   }
3378   NOT_PRODUCT(_tracer.scaled_iv_10(n);)
3379   return false;
3380 }
3381 
3382 //----------------------------offset_plus_k------------------------
3383 // Match: offset is (k [+/- invariant])
3384 // where k maybe zero and invariant is optional, but not both.
3385 bool SWPointer::offset_plus_k(Node* n, bool negate) {
3386   NOT_PRODUCT(Tracer::Depth ddd;)
3387   NOT_PRODUCT(_tracer.offset_plus_k_1(n);)
3388 
3389   int opc = n->Opcode();
3390   if (opc == Op_ConI) {
3391     _offset += negate ? -(n->get_int()) : n->get_int();
3392     NOT_PRODUCT(_tracer.offset_plus_k_2(n, _offset);)
3393     return true;
3394   } else if (opc == Op_ConL) {
3395     // Okay if value fits into an int
3396     const TypeLong* t = n->find_long_type();
3397     if (t->higher_equal(TypeLong::INT)) {
3398       jlong loff = n->get_long();
3399       jint  off  = (jint)loff;
3400       _offset += negate ? -off : loff;
3401       NOT_PRODUCT(_tracer.offset_plus_k_3(n, _offset);)
3402       return true;
3403     }
3404     NOT_PRODUCT(_tracer.offset_plus_k_4(n);)
3405     return false;
3406   }
3407   if (_invar != NULL) { // already has an invariant
3408     NOT_PRODUCT(_tracer.offset_plus_k_5(n, _invar);)
3409     return false;
3410   }
3411 
3412   if (_analyze_only && (invariant(n) == false)) {
3413     _nstack->push(n, _stack_idx++);
3414   }
3415   if (opc == Op_AddI) {
3416     if (n->in(2)->is_Con() && invariant(n->in(1))) {
3417       _negate_invar = negate;
3418       _invar = n->in(1);
3419       _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3420       NOT_PRODUCT(_tracer.offset_plus_k_6(n, _invar, _negate_invar, _offset);)
3421       return true;
3422     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3423       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3424       _negate_invar = negate;
3425       _invar = n->in(2);
3426       NOT_PRODUCT(_tracer.offset_plus_k_7(n, _invar, _negate_invar, _offset);)
3427       return true;
3428     }
3429   }
3430   if (opc == Op_SubI) {
3431     if (n->in(2)->is_Con() && invariant(n->in(1))) {
3432       _negate_invar = negate;
3433       _invar = n->in(1);
3434       _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3435       NOT_PRODUCT(_tracer.offset_plus_k_8(n, _invar, _negate_invar, _offset);)
3436       return true;
3437     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3438       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3439       _negate_invar = !negate;
3440       _invar = n->in(2);
3441       NOT_PRODUCT(_tracer.offset_plus_k_9(n, _invar, _negate_invar, _offset);)
3442       return true;
3443     }
3444   }
3445   if (invariant(n)) {
3446     if (opc == Op_ConvI2L) {
3447       n = n->in(1);
3448       if (n->Opcode() == Op_CastII &&
3449           n->as_CastII()->has_range_check()) {
3450         // Skip range check dependent CastII nodes
3451         assert(invariant(n), "sanity");
3452         n = n->in(1);
3453       }
3454     }
3455     if (n->bottom_type()->isa_int()) {
3456       _negate_invar = negate;
3457       _invar = n;
3458       NOT_PRODUCT(_tracer.offset_plus_k_10(n, _invar, _negate_invar, _offset);)
3459       return true;
3460     }
3461   }
3462 
3463   NOT_PRODUCT(_tracer.offset_plus_k_11(n);)
3464   return false;
3465 }
3466 
3467 //----------------------------print------------------------
3468 void SWPointer::print() {
3469 #ifndef PRODUCT
3470   tty->print("base: %d  adr: %d  scale: %d  offset: %d  invar: %c%d\n",
3471              _base != NULL ? _base->_idx : 0,
3472              _adr  != NULL ? _adr->_idx  : 0,
3473              _scale, _offset,
3474              _negate_invar?'-':'+',
3475              _invar != NULL ? _invar->_idx : 0);
3476 #endif
3477 }
3478 
3479 //----------------------------tracing------------------------
3480 #ifndef PRODUCT
3481 void SWPointer::Tracer::print_depth() {
3482   for (int ii = 0; ii<_depth; ++ii) tty->print("  ");
3483 }
3484 
3485 void SWPointer::Tracer::ctor_1 (Node* mem) {
3486   if(_slp->is_trace_alignment()) {
3487     print_depth(); tty->print(" %d SWPointer::SWPointer: start alignment analysis", mem->_idx); mem->dump();
3488   }
3489 }
3490 
3491 void SWPointer::Tracer::ctor_2(Node* adr) {
3492   if(_slp->is_trace_alignment()) {
3493     //store_depth();
3494     inc_depth();
3495     print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: ", adr->_idx); adr->dump();
3496     inc_depth();
3497     print_depth(); tty->print(" %d (base) SWPointer::SWPointer: ", adr->in(AddPNode::Base)->_idx); adr->in(AddPNode::Base)->dump();
3498   }
3499 }
3500 
3501 void SWPointer::Tracer::ctor_3(Node* adr, int i) {
3502   if(_slp->is_trace_alignment()) {
3503     inc_depth();
3504     Node* offset = adr->in(AddPNode::Offset);
3505     print_depth(); tty->print(" %d (offset) SWPointer::SWPointer: i = %d: ", offset->_idx, i); offset->dump();
3506   }
3507 }
3508 
3509 void SWPointer::Tracer::ctor_4(Node* adr, int i) {
3510   if(_slp->is_trace_alignment()) {
3511     inc_depth();
3512     print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: i = %d: ", adr->_idx, i); adr->dump();
3513   }
3514 }
3515 
3516 void SWPointer::Tracer::ctor_5(Node* adr, Node* base, int i) {
3517   if(_slp->is_trace_alignment()) {
3518     inc_depth();
3519     if (base == adr) {
3520       print_depth(); tty->print_cr("  \\ %d (adr) == %d (base) SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, base->_idx, i);
3521     } else if (!adr->is_AddP()) {
3522       print_depth(); tty->print_cr("  \\ %d (adr) is NOT Addp SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, i);
3523     }
3524   }
3525 }
3526 
3527 void SWPointer::Tracer::ctor_6(Node* mem) {
3528   if(_slp->is_trace_alignment()) {
3529     //restore_depth();
3530     print_depth(); tty->print_cr(" %d (adr) SWPointer::SWPointer: stop analysis", mem->_idx);
3531   }
3532 }
3533 
3534 void SWPointer::Tracer::invariant_1(Node *n, Node *n_c) {
3535   if (_slp->do_vector_loop() && _slp->is_debug() && _slp->_lpt->is_member(_slp->_phase->get_loop(n_c)) != (int)_slp->in_bb(n)) {
3536     int is_member =  _slp->_lpt->is_member(_slp->_phase->get_loop(n_c));
3537     int in_bb     =  _slp->in_bb(n);
3538     print_depth(); tty->print("  \\ ");  tty->print_cr(" %d SWPointer::invariant  conditions differ: n_c %d", n->_idx, n_c->_idx);
3539     print_depth(); tty->print("  \\ ");  tty->print_cr("is_member %d, in_bb %d", is_member, in_bb);
3540     print_depth(); tty->print("  \\ ");  n->dump();
3541     print_depth(); tty->print("  \\ ");  n_c->dump();
3542   }
3543 }
3544 
3545 void SWPointer::Tracer::scaled_iv_plus_offset_1(Node* n) {
3546   if(_slp->is_trace_alignment()) {
3547     print_depth(); tty->print(" %d SWPointer::scaled_iv_plus_offset testing node: ", n->_idx);
3548     n->dump();
3549   }
3550 }
3551 
3552 void SWPointer::Tracer::scaled_iv_plus_offset_2(Node* n) {
3553   if(_slp->is_trace_alignment()) {
3554     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
3555   }
3556 }
3557 
3558 void SWPointer::Tracer::scaled_iv_plus_offset_3(Node* n) {
3559   if(_slp->is_trace_alignment()) {
3560     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
3561   }
3562 }
3563 
3564 void SWPointer::Tracer::scaled_iv_plus_offset_4(Node* n) {
3565   if(_slp->is_trace_alignment()) {
3566     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
3567     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
3568     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
3569   }
3570 }
3571 
3572 void SWPointer::Tracer::scaled_iv_plus_offset_5(Node* n) {
3573   if(_slp->is_trace_alignment()) {
3574     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
3575     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
3576     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
3577   }
3578 }
3579 
3580 void SWPointer::Tracer::scaled_iv_plus_offset_6(Node* n) {
3581   if(_slp->is_trace_alignment()) {
3582     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
3583     print_depth(); tty->print("  \\  %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
3584     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
3585   }
3586 }
3587 
3588 void SWPointer::Tracer::scaled_iv_plus_offset_7(Node* n) {
3589   if(_slp->is_trace_alignment()) {
3590     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
3591     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
3592     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
3593   }
3594 }
3595 
3596 void SWPointer::Tracer::scaled_iv_plus_offset_8(Node* n) {
3597   if(_slp->is_trace_alignment()) {
3598     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: FAILED", n->_idx);
3599   }
3600 }
3601 
3602 void SWPointer::Tracer::scaled_iv_1(Node* n) {
3603   if(_slp->is_trace_alignment()) {
3604     print_depth(); tty->print(" %d SWPointer::scaled_iv: testing node: ", n->_idx); n->dump();
3605   }
3606 }
3607 
3608 void SWPointer::Tracer::scaled_iv_2(Node* n, int scale) {
3609   if(_slp->is_trace_alignment()) {
3610     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED since another _scale has been detected before", n->_idx);
3611     print_depth(); tty->print_cr("  \\ SWPointer::scaled_iv: _scale (%d) != 0", scale);
3612   }
3613 }
3614 
3615 void SWPointer::Tracer::scaled_iv_3(Node* n, int scale) {
3616   if(_slp->is_trace_alignment()) {
3617     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: is iv, setting _scale = %d", n->_idx, scale);
3618   }
3619 }
3620 
3621 void SWPointer::Tracer::scaled_iv_4(Node* n, int scale) {
3622   if(_slp->is_trace_alignment()) {
3623     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
3624     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
3625     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3626   }
3627 }
3628 
3629 void SWPointer::Tracer::scaled_iv_5(Node* n, int scale) {
3630   if(_slp->is_trace_alignment()) {
3631     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
3632     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is iv: ", n->in(2)->_idx); n->in(2)->dump();
3633     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3634   }
3635 }
3636 
3637 void SWPointer::Tracer::scaled_iv_6(Node* n, int scale) {
3638   if(_slp->is_trace_alignment()) {
3639     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftI PASSED, setting _scale = %d", n->_idx, scale);
3640     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
3641     print_depth(); tty->print("  \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3642   }
3643 }
3644 
3645 void SWPointer::Tracer::scaled_iv_7(Node* n) {
3646   if(_slp->is_trace_alignment()) {
3647     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_ConvI2L PASSED", n->_idx);
3648     print_depth(); tty->print_cr("  \\ SWPointer::scaled_iv: in(1) %d is scaled_iv_plus_offset: ", n->in(1)->_idx);
3649     inc_depth(); inc_depth();
3650     print_depth(); n->in(1)->dump();
3651     dec_depth(); dec_depth();
3652   }
3653 }
3654 
3655 void SWPointer::Tracer::scaled_iv_8(Node* n, SWPointer* tmp) {
3656   if(_slp->is_trace_alignment()) {
3657     print_depth(); tty->print(" %d SWPointer::scaled_iv: Op_LShiftL, creating tmp SWPointer: ", n->_idx); tmp->print();
3658   }
3659 }
3660 
3661 void SWPointer::Tracer::scaled_iv_9(Node* n, int scale, int _offset, int mult) {
3662   if(_slp->is_trace_alignment()) {
3663     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftL PASSED, setting _scale = %d, _offset = %d", n->_idx, scale, _offset);
3664     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",
3665     n->in(1)->_idx, n->in(2)->_idx, mult, scale, _offset);
3666     inc_depth(); inc_depth();
3667     print_depth(); n->in(1)->dump();
3668     print_depth(); n->in(2)->dump();
3669     dec_depth(); dec_depth();
3670   }
3671 }
3672 
3673 void SWPointer::Tracer::scaled_iv_10(Node* n) {
3674   if(_slp->is_trace_alignment()) {
3675     print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED", n->_idx);
3676   }
3677 }
3678 
3679 void SWPointer::Tracer::offset_plus_k_1(Node* n) {
3680   if(_slp->is_trace_alignment()) {
3681     print_depth(); tty->print(" %d SWPointer::offset_plus_k: testing node: ", n->_idx); n->dump();
3682   }
3683 }
3684 
3685 void SWPointer::Tracer::offset_plus_k_2(Node* n, int _offset) {
3686   if(_slp->is_trace_alignment()) {
3687     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConI PASSED, setting _offset = %d", n->_idx, _offset);
3688   }
3689 }
3690 
3691 void SWPointer::Tracer::offset_plus_k_3(Node* n, int _offset) {
3692   if(_slp->is_trace_alignment()) {
3693     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConL PASSED, setting _offset = %d", n->_idx, _offset);
3694   }
3695 }
3696 
3697 void SWPointer::Tracer::offset_plus_k_4(Node* n) {
3698   if(_slp->is_trace_alignment()) {
3699     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
3700     print_depth(); tty->print_cr("  \\ " JLONG_FORMAT " SWPointer::offset_plus_k: Op_ConL FAILED, k is too big", n->get_long());
3701   }
3702 }
3703 
3704 void SWPointer::Tracer::offset_plus_k_5(Node* n, Node* _invar) {
3705   if(_slp->is_trace_alignment()) {
3706     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED since another invariant has been detected before", n->_idx);
3707     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: _invar != NULL: ", _invar->_idx); _invar->dump();
3708   }
3709 }
3710 
3711 void SWPointer::Tracer::offset_plus_k_6(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3712   if(_slp->is_trace_alignment()) {
3713     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
3714     n->_idx, _negate_invar, _invar->_idx, _offset);
3715     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3716     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
3717   }
3718 }
3719 
3720 void SWPointer::Tracer::offset_plus_k_7(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3721   if(_slp->is_trace_alignment()) {
3722     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
3723     n->_idx, _negate_invar, _invar->_idx, _offset);
3724     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3725     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
3726   }
3727 }
3728 
3729 void SWPointer::Tracer::offset_plus_k_8(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3730   if(_slp->is_trace_alignment()) {
3731     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_SubI is PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
3732     n->_idx, _negate_invar, _invar->_idx, _offset);
3733     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3734     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
3735   }
3736 }
3737 
3738 void SWPointer::Tracer::offset_plus_k_9(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3739   if(_slp->is_trace_alignment()) {
3740     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);
3741     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3742     print_depth(); tty->print("  \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
3743   }
3744 }
3745 
3746 void SWPointer::Tracer::offset_plus_k_10(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3747   if(_slp->is_trace_alignment()) {
3748     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);
3749     print_depth(); tty->print_cr("  \\ %d SWPointer::offset_plus_k: is invariant", n->_idx);
3750   }
3751 }
3752 
3753 void SWPointer::Tracer::offset_plus_k_11(Node* n) {
3754   if(_slp->is_trace_alignment()) {
3755     print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
3756   }
3757 }
3758 
3759 #endif
3760 // ========================= OrderedPair =====================
3761 
3762 const OrderedPair OrderedPair::initial;
3763 
3764 // ========================= SWNodeInfo =====================
3765 
3766 const SWNodeInfo SWNodeInfo::initial;
3767 
3768 
3769 // ============================ DepGraph ===========================
3770 
3771 //------------------------------make_node---------------------------
3772 // Make a new dependence graph node for an ideal node.
3773 DepMem* DepGraph::make_node(Node* node) {
3774   DepMem* m = new (_arena) DepMem(node);
3775   if (node != NULL) {
3776     assert(_map.at_grow(node->_idx) == NULL, "one init only");
3777     _map.at_put_grow(node->_idx, m);
3778   }
3779   return m;
3780 }
3781 
3782 //------------------------------make_edge---------------------------
3783 // Make a new dependence graph edge from dpred -> dsucc
3784 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
3785   DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
3786   dpred->set_out_head(e);
3787   dsucc->set_in_head(e);
3788   return e;
3789 }
3790 
3791 // ========================== DepMem ========================
3792 
3793 //------------------------------in_cnt---------------------------
3794 int DepMem::in_cnt() {
3795   int ct = 0;
3796   for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
3797   return ct;
3798 }
3799 
3800 //------------------------------out_cnt---------------------------
3801 int DepMem::out_cnt() {
3802   int ct = 0;
3803   for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
3804   return ct;
3805 }
3806 
3807 //------------------------------print-----------------------------
3808 void DepMem::print() {
3809 #ifndef PRODUCT
3810   tty->print("  DepNode %d (", _node->_idx);
3811   for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
3812     Node* pred = p->pred()->node();
3813     tty->print(" %d", pred != NULL ? pred->_idx : 0);
3814   }
3815   tty->print(") [");
3816   for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
3817     Node* succ = s->succ()->node();
3818     tty->print(" %d", succ != NULL ? succ->_idx : 0);
3819   }
3820   tty->print_cr(" ]");
3821 #endif
3822 }
3823 
3824 // =========================== DepEdge =========================
3825 
3826 //------------------------------DepPreds---------------------------
3827 void DepEdge::print() {
3828 #ifndef PRODUCT
3829   tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
3830 #endif
3831 }
3832 
3833 // =========================== DepPreds =========================
3834 // Iterator over predecessor edges in the dependence graph.
3835 
3836 //------------------------------DepPreds---------------------------
3837 DepPreds::DepPreds(Node* n, DepGraph& dg) {
3838   _n = n;
3839   _done = false;
3840   if (_n->is_Store() || _n->is_Load()) {
3841     _next_idx = MemNode::Address;
3842     _end_idx  = n->req();
3843     _dep_next = dg.dep(_n)->in_head();
3844   } else if (_n->is_Mem()) {
3845     _next_idx = 0;
3846     _end_idx  = 0;
3847     _dep_next = dg.dep(_n)->in_head();
3848   } else {
3849     _next_idx = 1;
3850     _end_idx  = _n->req();
3851     _dep_next = NULL;
3852   }
3853   next();
3854 }
3855 
3856 //------------------------------next---------------------------
3857 void DepPreds::next() {
3858   if (_dep_next != NULL) {
3859     _current  = _dep_next->pred()->node();
3860     _dep_next = _dep_next->next_in();
3861   } else if (_next_idx < _end_idx) {
3862     _current  = _n->in(_next_idx++);
3863   } else {
3864     _done = true;
3865   }
3866 }
3867 
3868 // =========================== DepSuccs =========================
3869 // Iterator over successor edges in the dependence graph.
3870 
3871 //------------------------------DepSuccs---------------------------
3872 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
3873   _n = n;
3874   _done = false;
3875   if (_n->is_Load()) {
3876     _next_idx = 0;
3877     _end_idx  = _n->outcnt();
3878     _dep_next = dg.dep(_n)->out_head();
3879   } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
3880     _next_idx = 0;
3881     _end_idx  = 0;
3882     _dep_next = dg.dep(_n)->out_head();
3883   } else {
3884     _next_idx = 0;
3885     _end_idx  = _n->outcnt();
3886     _dep_next = NULL;
3887   }
3888   next();
3889 }
3890 
3891 //-------------------------------next---------------------------
3892 void DepSuccs::next() {
3893   if (_dep_next != NULL) {
3894     _current  = _dep_next->succ()->node();
3895     _dep_next = _dep_next->next_out();
3896   } else if (_next_idx < _end_idx) {
3897     _current  = _n->raw_out(_next_idx++);
3898   } else {
3899     _done = true;
3900   }
3901 }
3902 
3903 //
3904 // --------------------------------- vectorization/simd -----------------------------------
3905 //
3906 bool SuperWord::same_origin_idx(Node* a, Node* b) const {
3907   return a != NULL && b != NULL && _clone_map.same_idx(a->_idx, b->_idx);
3908 }
3909 bool SuperWord::same_generation(Node* a, Node* b) const {
3910   return a != NULL && b != NULL && _clone_map.same_gen(a->_idx, b->_idx);
3911 }
3912 
3913 Node*  SuperWord::find_phi_for_mem_dep(LoadNode* ld) {
3914   assert(in_bb(ld), "must be in block");
3915   if (_clone_map.gen(ld->_idx) == _ii_first) {
3916 #ifndef PRODUCT
3917     if (_vector_loop_debug) {
3918       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(ld->_idx)=%d",
3919         _clone_map.gen(ld->_idx));
3920     }
3921 #endif
3922     return NULL; //we think that any ld in the first gen being vectorizable
3923   }
3924 
3925   Node* mem = ld->in(MemNode::Memory);
3926   if (mem->outcnt() <= 1) {
3927     // we don't want to remove the only edge from mem node to load
3928 #ifndef PRODUCT
3929     if (_vector_loop_debug) {
3930       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",
3931         mem->_idx, ld->_idx);
3932       ld->dump();
3933       mem->dump();
3934     }
3935 #endif
3936     return NULL;
3937   }
3938   if (!in_bb(mem) || same_generation(mem, ld)) {
3939 #ifndef PRODUCT
3940     if (_vector_loop_debug) {
3941       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(mem->_idx)=%d",
3942         _clone_map.gen(mem->_idx));
3943     }
3944 #endif
3945     return NULL; // does not depend on loop volatile node or depends on the same generation
3946   }
3947 
3948   //otherwise first node should depend on mem-phi
3949   Node* first = first_node(ld);
3950   assert(first->is_Load(), "must be Load");
3951   Node* phi = first->as_Load()->in(MemNode::Memory);
3952   if (!phi->is_Phi() || phi->bottom_type() != Type::MEMORY) {
3953 #ifndef PRODUCT
3954     if (_vector_loop_debug) {
3955       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");
3956       ld->dump();
3957       first->dump();
3958     }
3959 #endif
3960     return NULL;
3961   }
3962 
3963   Node* tail = 0;
3964   for (int m = 0; m < _mem_slice_head.length(); m++) {
3965     if (_mem_slice_head.at(m) == phi) {
3966       tail = _mem_slice_tail.at(m);
3967     }
3968   }
3969   if (tail == 0) { //test that found phi is in the list  _mem_slice_head
3970 #ifndef PRODUCT
3971     if (_vector_loop_debug) {
3972       tty->print_cr("SuperWord::find_phi_for_mem_dep load %d is not vectorizable node, its phi %d is not _mem_slice_head",
3973         ld->_idx, phi->_idx);
3974       ld->dump();
3975       phi->dump();
3976     }
3977 #endif
3978     return NULL;
3979   }
3980 
3981   // now all conditions are met
3982   return phi;
3983 }
3984 
3985 Node* SuperWord::first_node(Node* nd) {
3986   for (int ii = 0; ii < _iteration_first.length(); ii++) {
3987     Node* nnn = _iteration_first.at(ii);
3988     if (same_origin_idx(nnn, nd)) {
3989 #ifndef PRODUCT
3990       if (_vector_loop_debug) {
3991         tty->print_cr("SuperWord::first_node: %d is the first iteration node for %d (_clone_map.idx(nnn->_idx) = %d)",
3992           nnn->_idx, nd->_idx, _clone_map.idx(nnn->_idx));
3993       }
3994 #endif
3995       return nnn;
3996     }
3997   }
3998 
3999 #ifndef PRODUCT
4000   if (_vector_loop_debug) {
4001     tty->print_cr("SuperWord::first_node: did not find first iteration node for %d (_clone_map.idx(nd->_idx)=%d)",
4002       nd->_idx, _clone_map.idx(nd->_idx));
4003   }
4004 #endif
4005   return 0;
4006 }
4007 
4008 Node* SuperWord::last_node(Node* nd) {
4009   for (int ii = 0; ii < _iteration_last.length(); ii++) {
4010     Node* nnn = _iteration_last.at(ii);
4011     if (same_origin_idx(nnn, nd)) {
4012 #ifndef PRODUCT
4013       if (_vector_loop_debug) {
4014         tty->print_cr("SuperWord::last_node _clone_map.idx(nnn->_idx)=%d, _clone_map.idx(nd->_idx)=%d",
4015           _clone_map.idx(nnn->_idx), _clone_map.idx(nd->_idx));
4016       }
4017 #endif
4018       return nnn;
4019     }
4020   }
4021   return 0;
4022 }
4023 
4024 int SuperWord::mark_generations() {
4025   Node *ii_err = NULL, *tail_err = NULL;
4026   for (int i = 0; i < _mem_slice_head.length(); i++) {
4027     Node* phi  = _mem_slice_head.at(i);
4028     assert(phi->is_Phi(), "must be phi");
4029 
4030     Node* tail = _mem_slice_tail.at(i);
4031     if (_ii_last == -1) {
4032       tail_err = tail;
4033       _ii_last = _clone_map.gen(tail->_idx);
4034     }
4035     else if (_ii_last != _clone_map.gen(tail->_idx)) {
4036 #ifndef PRODUCT
4037       if (TraceSuperWord && Verbose) {
4038         tty->print_cr("SuperWord::mark_generations _ii_last error - found different generations in two tail nodes ");
4039         tail->dump();
4040         tail_err->dump();
4041       }
4042 #endif
4043       return -1;
4044     }
4045 
4046     // find first iteration in the loop
4047     for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
4048       Node* ii = phi->fast_out(i);
4049       if (in_bb(ii) && ii->is_Store()) { // we speculate that normally Stores of one and one only generation have deps from mem phi
4050         if (_ii_first == -1) {
4051           ii_err = ii;
4052           _ii_first = _clone_map.gen(ii->_idx);
4053         } else if (_ii_first != _clone_map.gen(ii->_idx)) {
4054 #ifndef PRODUCT
4055           if (TraceSuperWord && Verbose) {
4056             tty->print_cr("SuperWord::mark_generations: _ii_first was found before and not equal to one in this node (%d)", _ii_first);
4057             ii->dump();
4058             if (ii_err!= 0) {
4059               ii_err->dump();
4060             }
4061           }
4062 #endif
4063           return -1; // this phi has Stores from different generations of unroll and cannot be simd/vectorized
4064         }
4065       }
4066     }//for (DUIterator_Fast imax,
4067   }//for (int i...
4068 
4069   if (_ii_first == -1 || _ii_last == -1) {
4070     if (TraceSuperWord && Verbose) {
4071       tty->print_cr("SuperWord::mark_generations unknown error, something vent wrong");
4072     }
4073     return -1; // something vent wrong
4074   }
4075   // collect nodes in the first and last generations
4076   assert(_iteration_first.length() == 0, "_iteration_first must be empty");
4077   assert(_iteration_last.length() == 0, "_iteration_last must be empty");
4078   for (int j = 0; j < _block.length(); j++) {
4079     Node* n = _block.at(j);
4080     node_idx_t gen = _clone_map.gen(n->_idx);
4081     if ((signed)gen == _ii_first) {
4082       _iteration_first.push(n);
4083     } else if ((signed)gen == _ii_last) {
4084       _iteration_last.push(n);
4085     }
4086   }
4087 
4088   // building order of iterations
4089   if (_ii_order.length() == 0 && ii_err != 0) {
4090     assert(in_bb(ii_err) && ii_err->is_Store(), "should be Store in bb");
4091     Node* nd = ii_err;
4092     while(_clone_map.gen(nd->_idx) != _ii_last) {
4093       _ii_order.push(_clone_map.gen(nd->_idx));
4094       bool found = false;
4095       for (DUIterator_Fast imax, i = nd->fast_outs(imax); i < imax; i++) {
4096         Node* use = nd->fast_out(i);
4097         if (same_origin_idx(use, nd) && use->as_Store()->in(MemNode::Memory) == nd) {
4098           found = true;
4099           nd = use;
4100           break;
4101         }
4102       }//for
4103 
4104       if (found == false) {
4105         if (TraceSuperWord && Verbose) {
4106           tty->print_cr("SuperWord::mark_generations: Cannot build order of iterations - no dependent Store for %d", nd->_idx);
4107         }
4108         _ii_order.clear();
4109         return -1;
4110       }
4111     } //while
4112     _ii_order.push(_clone_map.gen(nd->_idx));
4113   }
4114 
4115 #ifndef PRODUCT
4116   if (_vector_loop_debug) {
4117     tty->print_cr("SuperWord::mark_generations");
4118     tty->print_cr("First generation (%d) nodes:", _ii_first);
4119     for (int ii = 0; ii < _iteration_first.length(); ii++)  _iteration_first.at(ii)->dump();
4120     tty->print_cr("Last generation (%d) nodes:", _ii_last);
4121     for (int ii = 0; ii < _iteration_last.length(); ii++)  _iteration_last.at(ii)->dump();
4122     tty->print_cr(" ");
4123 
4124     tty->print("SuperWord::List of generations: ");
4125     for (int jj = 0; jj < _ii_order.length(); ++jj) {
4126       tty->print("%d:%d ", jj, _ii_order.at(jj));
4127     }
4128     tty->print_cr(" ");
4129   }
4130 #endif
4131 
4132   return _ii_first;
4133 }
4134 
4135 bool SuperWord::fix_commutative_inputs(Node* gold, Node* fix) {
4136   assert(gold->is_Add() && fix->is_Add() || gold->is_Mul() && fix->is_Mul(), "should be only Add or Mul nodes");
4137   assert(same_origin_idx(gold, fix), "should be clones of the same node");
4138   Node* gin1 = gold->in(1);
4139   Node* gin2 = gold->in(2);
4140   Node* fin1 = fix->in(1);
4141   Node* fin2 = fix->in(2);
4142   bool swapped = false;
4143 
4144   if (in_bb(gin1) && in_bb(gin2) && in_bb(fin1) && in_bb(fin1)) {
4145     if (same_origin_idx(gin1, fin1) &&
4146         same_origin_idx(gin2, fin2)) {
4147       return true; // nothing to fix
4148     }
4149     if (same_origin_idx(gin1, fin2) &&
4150         same_origin_idx(gin2, fin1)) {
4151       fix->swap_edges(1, 2);
4152       swapped = true;
4153     }
4154   }
4155   // at least one input comes from outside of bb
4156   if (gin1->_idx == fin1->_idx)  {
4157     return true; // nothing to fix
4158   }
4159   if (!swapped && (gin1->_idx == fin2->_idx || gin2->_idx == fin1->_idx))  { //swapping is expensive, check condition first
4160     fix->swap_edges(1, 2);
4161     swapped = true;
4162   }
4163 
4164   if (swapped) {
4165 #ifndef PRODUCT
4166     if (_vector_loop_debug) {
4167       tty->print_cr("SuperWord::fix_commutative_inputs: fixed node %d", fix->_idx);
4168     }
4169 #endif
4170     return true;
4171   }
4172 
4173   if (TraceSuperWord && Verbose) {
4174     tty->print_cr("SuperWord::fix_commutative_inputs: cannot fix node %d", fix->_idx);
4175   }
4176 
4177   return false;
4178 }
4179 
4180 bool SuperWord::pack_parallel() {
4181 #ifndef PRODUCT
4182   if (_vector_loop_debug) {
4183     tty->print_cr("SuperWord::pack_parallel: START");
4184   }
4185 #endif
4186 
4187   _packset.clear();
4188 
4189   for (int ii = 0; ii < _iteration_first.length(); ii++) {
4190     Node* nd = _iteration_first.at(ii);
4191     if (in_bb(nd) && (nd->is_Load() || nd->is_Store() || nd->is_Add() || nd->is_Mul())) {
4192       Node_List* pk = new Node_List();
4193       pk->push(nd);
4194       for (int gen = 1; gen < _ii_order.length(); ++gen) {
4195         for (int kk = 0; kk < _block.length(); kk++) {
4196           Node* clone = _block.at(kk);
4197           if (same_origin_idx(clone, nd) &&
4198               _clone_map.gen(clone->_idx) == _ii_order.at(gen)) {
4199             if (nd->is_Add() || nd->is_Mul()) {
4200               fix_commutative_inputs(nd, clone);
4201             }
4202             pk->push(clone);
4203             if (pk->size() == 4) {
4204               _packset.append(pk);
4205 #ifndef PRODUCT
4206               if (_vector_loop_debug) {
4207                 tty->print_cr("SuperWord::pack_parallel: added pack ");
4208                 pk->dump();
4209               }
4210 #endif
4211               if (_clone_map.gen(clone->_idx) != _ii_last) {
4212                 pk = new Node_List();
4213               }
4214             }
4215             break;
4216           }
4217         }
4218       }//for
4219     }//if
4220   }//for
4221 
4222 #ifndef PRODUCT
4223   if (_vector_loop_debug) {
4224     tty->print_cr("SuperWord::pack_parallel: END");
4225   }
4226 #endif
4227 
4228   return true;
4229 }
4230 
4231 bool SuperWord::hoist_loads_in_graph() {
4232   GrowableArray<Node*> loads;
4233 
4234 #ifndef PRODUCT
4235   if (_vector_loop_debug) {
4236     tty->print_cr("SuperWord::hoist_loads_in_graph: total number _mem_slice_head.length() = %d", _mem_slice_head.length());
4237   }
4238 #endif
4239 
4240   for (int i = 0; i < _mem_slice_head.length(); i++) {
4241     Node* n = _mem_slice_head.at(i);
4242     if ( !in_bb(n) || !n->is_Phi() || n->bottom_type() != Type::MEMORY) {
4243       if (TraceSuperWord && Verbose) {
4244         tty->print_cr("SuperWord::hoist_loads_in_graph: skipping unexpected node n=%d", n->_idx);
4245       }
4246       continue;
4247     }
4248 
4249 #ifndef PRODUCT
4250     if (_vector_loop_debug) {
4251       tty->print_cr("SuperWord::hoist_loads_in_graph: processing phi %d  = _mem_slice_head.at(%d);", n->_idx, i);
4252     }
4253 #endif
4254 
4255     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4256       Node* ld = n->fast_out(i);
4257       if (ld->is_Load() && ld->as_Load()->in(MemNode::Memory) == n && in_bb(ld)) {
4258         for (int i = 0; i < _block.length(); i++) {
4259           Node* ld2 = _block.at(i);
4260           if (ld2->is_Load() && same_origin_idx(ld, ld2) &&
4261               !same_generation(ld, ld2)) { // <= do not collect the first generation ld
4262 #ifndef PRODUCT
4263             if (_vector_loop_debug) {
4264               tty->print_cr("SuperWord::hoist_loads_in_graph: will try to hoist load ld2->_idx=%d, cloned from %d (ld->_idx=%d)",
4265                 ld2->_idx, _clone_map.idx(ld->_idx), ld->_idx);
4266             }
4267 #endif
4268             // could not do on-the-fly, since iterator is immutable
4269             loads.push(ld2);
4270           }
4271         }// for
4272       }//if
4273     }//for (DUIterator_Fast imax,
4274   }//for (int i = 0; i
4275 
4276   for (int i = 0; i < loads.length(); i++) {
4277     LoadNode* ld = loads.at(i)->as_Load();
4278     Node* phi = find_phi_for_mem_dep(ld);
4279     if (phi != NULL) {
4280 #ifndef PRODUCT
4281       if (_vector_loop_debug) {
4282         tty->print_cr("SuperWord::hoist_loads_in_graph replacing MemNode::Memory(%d) edge in %d with one from %d",
4283           MemNode::Memory, ld->_idx, phi->_idx);
4284       }
4285 #endif
4286       _igvn.replace_input_of(ld, MemNode::Memory, phi);
4287     }
4288   }//for
4289 
4290   restart(); // invalidate all basic structures, since we rebuilt the graph
4291 
4292   if (TraceSuperWord && Verbose) {
4293     tty->print_cr("\nSuperWord::hoist_loads_in_graph() the graph was rebuilt, all structures invalidated and need rebuild");
4294   }
4295 
4296   return true;
4297 }
4298 
--- EOF ---