1 /*
   2  * Copyright (c) 2007, 2013, 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/optonode.hpp"
  38 #include "opto/superword.hpp"
  39 #include "opto/vectornode.hpp"
  40 
  41 //
  42 //                  S U P E R W O R D   T R A N S F O R M
  43 //=============================================================================
  44 
  45 //------------------------------SuperWord---------------------------
  46 SuperWord::SuperWord(PhaseIdealLoop* phase) :
  47   _phase(phase),
  48   _igvn(phase->_igvn),
  49   _arena(phase->C->comp_arena()),
  50   _packset(arena(), 8,  0, NULL),         // packs for the current block
  51   _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
  52   _block(arena(), 8,  0, NULL),           // nodes in current block
  53   _data_entry(arena(), 8,  0, NULL),      // nodes with all inputs from outside
  54   _mem_slice_head(arena(), 8,  0, NULL),  // memory slice heads
  55   _mem_slice_tail(arena(), 8,  0, NULL),  // memory slice tails
  56   _node_info(arena(), 8,  0, SWNodeInfo::initial), // info needed per node
  57   _align_to_ref(NULL),                    // memory reference to align vectors to
  58   _disjoint_ptrs(arena(), 8,  0, OrderedPair::initial), // runtime disambiguated pointer pairs
  59   _dg(_arena),                            // dependence graph
  60   _visited(arena()),                      // visited node set
  61   _post_visited(arena()),                 // post visited node set
  62   _n_idx_list(arena(), 8),                // scratch list of (node,index) pairs
  63   _stk(arena(), 8, 0, NULL),              // scratch stack of nodes
  64   _nlist(arena(), 8, 0, NULL),            // scratch list of nodes
  65   _lpt(NULL),                             // loop tree node
  66   _lp(NULL),                              // LoopNode
  67   _bb(NULL),                              // basic block
  68   _iv(NULL)                               // induction var
  69 {}
  70 
  71 //------------------------------transform_loop---------------------------
  72 void SuperWord::transform_loop(IdealLoopTree* lpt) {
  73   assert(UseSuperWord, "should be");
  74   // Do vectors exist on this architecture?
  75   if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return;
  76 
  77   assert(lpt->_head->is_CountedLoop(), "must be");
  78   CountedLoopNode *cl = lpt->_head->as_CountedLoop();
  79 
  80   if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
  81 
  82   if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
  83 
  84   // Check for no control flow in body (other than exit)
  85   Node *cl_exit = cl->loopexit();
  86   if (cl_exit->in(0) != lpt->_head) return;
  87 
  88   // Make sure the are no extra control users of the loop backedge
  89   if (cl->back_control()->outcnt() != 1) {
  90     return;
  91   }
  92 
  93   // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
  94   CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
  95   if (pre_end == NULL) return;
  96   Node *pre_opaq1 = pre_end->limit();
  97   if (pre_opaq1->Opcode() != Op_Opaque1) return;
  98 
  99   init(); // initialize data structures
 100 
 101   set_lpt(lpt);
 102   set_lp(cl);
 103 
 104   // For now, define one block which is the entire loop body
 105   set_bb(cl);
 106 
 107   assert(_packset.length() == 0, "packset must be empty");
 108   SLP_extract();
 109 }
 110 
 111 //------------------------------SLP_extract---------------------------
 112 // Extract the superword level parallelism
 113 //
 114 // 1) A reverse post-order of nodes in the block is constructed.  By scanning
 115 //    this list from first to last, all definitions are visited before their uses.
 116 //
 117 // 2) A point-to-point dependence graph is constructed between memory references.
 118 //    This simplies the upcoming "independence" checker.
 119 //
 120 // 3) The maximum depth in the node graph from the beginning of the block
 121 //    to each node is computed.  This is used to prune the graph search
 122 //    in the independence checker.
 123 //
 124 // 4) For integer types, the necessary bit width is propagated backwards
 125 //    from stores to allow packed operations on byte, char, and short
 126 //    integers.  This reverses the promotion to type "int" that javac
 127 //    did for operations like: char c1,c2,c3;  c1 = c2 + c3.
 128 //
 129 // 5) One of the memory references is picked to be an aligned vector reference.
 130 //    The pre-loop trip count is adjusted to align this reference in the
 131 //    unrolled body.
 132 //
 133 // 6) The initial set of pack pairs is seeded with memory references.
 134 //
 135 // 7) The set of pack pairs is extended by following use->def and def->use links.
 136 //
 137 // 8) The pairs are combined into vector sized packs.
 138 //
 139 // 9) Reorder the memory slices to co-locate members of the memory packs.
 140 //
 141 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
 142 //    inserting scalar promotion, vector creation from multiple scalars, and
 143 //    extraction of scalar values from vectors.
 144 //
 145 void SuperWord::SLP_extract() {
 146 
 147   // Ready the block
 148 
 149   if (!construct_bb())
 150     return; // Exit if no interesting nodes or complex graph.
 151 
 152   dependence_graph();
 153 
 154   compute_max_depth();
 155 
 156   compute_vector_element_type();
 157 
 158   // Attempt vectorization
 159 
 160   find_adjacent_refs();
 161 
 162   extend_packlist();
 163 
 164   combine_packs();
 165 
 166   construct_my_pack_map();
 167 
 168   filter_packs();
 169 
 170   schedule();
 171 
 172   output();
 173 }
 174 
 175 //------------------------------find_adjacent_refs---------------------------
 176 // Find the adjacent memory references and create pack pairs for them.
 177 // This is the initial set of packs that will then be extended by
 178 // following use->def and def->use links.  The align positions are
 179 // assigned relative to the reference "align_to_ref"
 180 void SuperWord::find_adjacent_refs() {
 181   // Get list of memory operations
 182   Node_List memops;
 183   for (int i = 0; i < _block.length(); i++) {
 184     Node* n = _block.at(i);
 185     if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
 186         is_java_primitive(n->as_Mem()->memory_type())) {
 187       int align = memory_alignment(n->as_Mem(), 0);
 188       if (align != bottom_align) {
 189         memops.push(n);
 190       }
 191     }
 192   }
 193 
 194   Node_List align_to_refs;
 195   int best_iv_adjustment = 0;
 196   MemNode* best_align_to_mem_ref = NULL;
 197 
 198   while (memops.size() != 0) {
 199     // Find a memory reference to align to.
 200     MemNode* mem_ref = find_align_to_ref(memops);
 201     if (mem_ref == NULL) break;
 202     align_to_refs.push(mem_ref);
 203     int iv_adjustment = get_iv_adjustment(mem_ref);
 204 
 205     if (best_align_to_mem_ref == NULL) {
 206       // Set memory reference which is the best from all memory operations
 207       // to be used for alignment. The pre-loop trip count is modified to align
 208       // this reference to a vector-aligned address.
 209       best_align_to_mem_ref = mem_ref;
 210       best_iv_adjustment = iv_adjustment;
 211     }
 212 
 213     SWPointer align_to_ref_p(mem_ref, this);
 214     // Set alignment relative to "align_to_ref" for all related memory operations.
 215     for (int i = memops.size() - 1; i >= 0; i--) {
 216       MemNode* s = memops.at(i)->as_Mem();
 217       if (isomorphic(s, mem_ref)) {
 218         SWPointer p2(s, this);
 219         if (p2.comparable(align_to_ref_p)) {
 220           int align = memory_alignment(s, iv_adjustment);
 221           set_alignment(s, align);
 222         }
 223       }
 224     }
 225 
 226     // Create initial pack pairs of memory operations for which
 227     // alignment is set and vectors will be aligned.
 228     bool create_pack = true;
 229     if (memory_alignment(mem_ref, best_iv_adjustment) == 0) {
 230       if (!Matcher::misaligned_vectors_ok()) {
 231         int vw = vector_width(mem_ref);
 232         int vw_best = vector_width(best_align_to_mem_ref);
 233         if (vw > vw_best) {
 234           // Do not vectorize a memory access with more elements per vector
 235           // if unaligned memory access is not allowed because number of
 236           // iterations in pre-loop will be not enough to align it.
 237           create_pack = false;
 238         }
 239       }
 240     } else {
 241       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 242         // Can't allow vectorization of unaligned memory accesses with the
 243         // same type since it could be overlapped accesses to the same array.
 244         create_pack = false;
 245       } else {
 246         // Allow independent (different type) unaligned memory operations
 247         // if HW supports them.
 248         if (!Matcher::misaligned_vectors_ok()) {
 249           create_pack = false;
 250         } else {
 251           // Check if packs of the same memory type but
 252           // with a different alignment were created before.
 253           for (uint i = 0; i < align_to_refs.size(); i++) {
 254             MemNode* mr = align_to_refs.at(i)->as_Mem();
 255             if (same_velt_type(mr, mem_ref) &&
 256                 memory_alignment(mr, iv_adjustment) != 0)
 257               create_pack = false;
 258           }
 259         }
 260       }
 261     }
 262     if (create_pack) {
 263       for (uint i = 0; i < memops.size(); i++) {
 264         Node* s1 = memops.at(i);
 265         int align = alignment(s1);
 266         if (align == top_align) continue;
 267         for (uint j = 0; j < memops.size(); j++) {
 268           Node* s2 = memops.at(j);
 269           if (alignment(s2) == top_align) continue;
 270           if (s1 != s2 && are_adjacent_refs(s1, s2)) {
 271             if (stmts_can_pack(s1, s2, align)) {
 272               Node_List* pair = new Node_List();
 273               pair->push(s1);
 274               pair->push(s2);
 275               _packset.append(pair);
 276             }
 277           }
 278         }
 279       }
 280     } else { // Don't create unaligned pack
 281       // First, remove remaining memory ops of the same type from the list.
 282       for (int i = memops.size() - 1; i >= 0; i--) {
 283         MemNode* s = memops.at(i)->as_Mem();
 284         if (same_velt_type(s, mem_ref)) {
 285           memops.remove(i);
 286         }
 287       }
 288 
 289       // Second, remove already constructed packs of the same type.
 290       for (int i = _packset.length() - 1; i >= 0; i--) {
 291         Node_List* p = _packset.at(i);
 292         MemNode* s = p->at(0)->as_Mem();
 293         if (same_velt_type(s, mem_ref)) {
 294           remove_pack_at(i);
 295         }
 296       }
 297 
 298       // If needed find the best memory reference for loop alignment again.
 299       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 300         // Put memory ops from remaining packs back on memops list for
 301         // the best alignment search.
 302         uint orig_msize = memops.size();
 303         for (int i = 0; i < _packset.length(); i++) {
 304           Node_List* p = _packset.at(i);
 305           MemNode* s = p->at(0)->as_Mem();
 306           assert(!same_velt_type(s, mem_ref), "sanity");
 307           memops.push(s);
 308         }
 309         MemNode* best_align_to_mem_ref = find_align_to_ref(memops);
 310         if (best_align_to_mem_ref == NULL) break;
 311         best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
 312         // Restore list.
 313         while (memops.size() > orig_msize)
 314           (void)memops.pop();
 315       }
 316     } // unaligned memory accesses
 317 
 318     // Remove used mem nodes.
 319     for (int i = memops.size() - 1; i >= 0; i--) {
 320       MemNode* m = memops.at(i)->as_Mem();
 321       if (alignment(m) != top_align) {
 322         memops.remove(i);
 323       }
 324     }
 325 
 326   } // while (memops.size() != 0
 327   set_align_to_ref(best_align_to_mem_ref);
 328 
 329 #ifndef PRODUCT
 330   if (TraceSuperWord) {
 331     tty->print_cr("\nAfter find_adjacent_refs");
 332     print_packset();
 333   }
 334 #endif
 335 }
 336 
 337 //------------------------------find_align_to_ref---------------------------
 338 // Find a memory reference to align the loop induction variable to.
 339 // Looks first at stores then at loads, looking for a memory reference
 340 // with the largest number of references similar to it.
 341 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
 342   GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
 343 
 344   // Count number of comparable memory ops
 345   for (uint i = 0; i < memops.size(); i++) {
 346     MemNode* s1 = memops.at(i)->as_Mem();
 347     SWPointer p1(s1, this);
 348     // Discard if pre loop can't align this reference
 349     if (!ref_is_alignable(p1)) {
 350       *cmp_ct.adr_at(i) = 0;
 351       continue;
 352     }
 353     for (uint j = i+1; j < memops.size(); j++) {
 354       MemNode* s2 = memops.at(j)->as_Mem();
 355       if (isomorphic(s1, s2)) {
 356         SWPointer p2(s2, this);
 357         if (p1.comparable(p2)) {
 358           (*cmp_ct.adr_at(i))++;
 359           (*cmp_ct.adr_at(j))++;
 360         }
 361       }
 362     }
 363   }
 364 
 365   // Find Store (or Load) with the greatest number of "comparable" references,
 366   // biggest vector size, smallest data size and smallest iv offset.
 367   int max_ct        = 0;
 368   int max_vw        = 0;
 369   int max_idx       = -1;
 370   int min_size      = max_jint;
 371   int min_iv_offset = max_jint;
 372   for (uint j = 0; j < memops.size(); j++) {
 373     MemNode* s = memops.at(j)->as_Mem();
 374     if (s->is_Store()) {
 375       int vw = vector_width_in_bytes(s);
 376       assert(vw > 1, "sanity");
 377       SWPointer p(s, this);
 378       if (cmp_ct.at(j) >  max_ct ||
 379           cmp_ct.at(j) == max_ct &&
 380             (vw >  max_vw ||
 381              vw == max_vw &&
 382               (data_size(s) <  min_size ||
 383                data_size(s) == min_size &&
 384                  (p.offset_in_bytes() < min_iv_offset)))) {
 385         max_ct = cmp_ct.at(j);
 386         max_vw = vw;
 387         max_idx = j;
 388         min_size = data_size(s);
 389         min_iv_offset = p.offset_in_bytes();
 390       }
 391     }
 392   }
 393   // If no stores, look at loads
 394   if (max_ct == 0) {
 395     for (uint j = 0; j < memops.size(); j++) {
 396       MemNode* s = memops.at(j)->as_Mem();
 397       if (s->is_Load()) {
 398         int vw = vector_width_in_bytes(s);
 399         assert(vw > 1, "sanity");
 400         SWPointer p(s, this);
 401         if (cmp_ct.at(j) >  max_ct ||
 402             cmp_ct.at(j) == max_ct &&
 403               (vw >  max_vw ||
 404                vw == max_vw &&
 405                 (data_size(s) <  min_size ||
 406                  data_size(s) == min_size &&
 407                    (p.offset_in_bytes() < min_iv_offset)))) {
 408           max_ct = cmp_ct.at(j);
 409           max_vw = vw;
 410           max_idx = j;
 411           min_size = data_size(s);
 412           min_iv_offset = p.offset_in_bytes();
 413         }
 414       }
 415     }
 416   }
 417 
 418 #ifdef ASSERT
 419   if (TraceSuperWord && Verbose) {
 420     tty->print_cr("\nVector memops after find_align_to_refs");
 421     for (uint i = 0; i < memops.size(); i++) {
 422       MemNode* s = memops.at(i)->as_Mem();
 423       s->dump();
 424     }
 425   }
 426 #endif
 427 
 428   if (max_ct > 0) {
 429 #ifdef ASSERT
 430     if (TraceSuperWord) {
 431       tty->print("\nVector align to node: ");
 432       memops.at(max_idx)->as_Mem()->dump();
 433     }
 434 #endif
 435     return memops.at(max_idx)->as_Mem();
 436   }
 437   return NULL;
 438 }
 439 
 440 //------------------------------ref_is_alignable---------------------------
 441 // Can the preloop align the reference to position zero in the vector?
 442 bool SuperWord::ref_is_alignable(SWPointer& p) {
 443   if (!p.has_iv()) {
 444     return true;   // no induction variable
 445   }
 446   CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
 447   assert(pre_end != NULL, "we must have a correct pre-loop");
 448   assert(pre_end->stride_is_con(), "pre loop stride is constant");
 449   int preloop_stride = pre_end->stride_con();
 450 
 451   int span = preloop_stride * p.scale_in_bytes();
 452 
 453   // Stride one accesses are alignable.
 454   if (ABS(span) == p.memory_size())
 455     return true;
 456 
 457   // If initial offset from start of object is computable,
 458   // compute alignment within the vector.
 459   int vw = vector_width_in_bytes(p.mem());
 460   assert(vw > 1, "sanity");
 461   if (vw % span == 0) {
 462     Node* init_nd = pre_end->init_trip();
 463     if (init_nd->is_Con() && p.invar() == NULL) {
 464       int init = init_nd->bottom_type()->is_int()->get_con();
 465 
 466       int init_offset = init * p.scale_in_bytes() + p.offset_in_bytes();
 467       assert(init_offset >= 0, "positive offset from object start");
 468 
 469       if (span > 0) {
 470         return (vw - (init_offset % vw)) % span == 0;
 471       } else {
 472         assert(span < 0, "nonzero stride * scale");
 473         return (init_offset % vw) % -span == 0;
 474       }
 475     }
 476   }
 477   return false;
 478 }
 479 
 480 //---------------------------get_iv_adjustment---------------------------
 481 // Calculate loop's iv adjustment for this memory ops.
 482 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
 483   SWPointer align_to_ref_p(mem_ref, this);
 484   int offset = align_to_ref_p.offset_in_bytes();
 485   int scale  = align_to_ref_p.scale_in_bytes();
 486   int vw       = vector_width_in_bytes(mem_ref);
 487   assert(vw > 1, "sanity");
 488   int stride_sign   = (scale * iv_stride()) > 0 ? 1 : -1;
 489   // At least one iteration is executed in pre-loop by default. As result
 490   // several iterations are needed to align memory operations in main-loop even
 491   // if offset is 0.
 492   int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
 493   int elt_size = align_to_ref_p.memory_size();
 494   assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
 495          err_msg_res("(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size));
 496   int iv_adjustment = iv_adjustment_in_bytes/elt_size;
 497 
 498 #ifndef PRODUCT
 499   if (TraceSuperWord)
 500     tty->print_cr("\noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d",
 501                   offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
 502 #endif
 503   return iv_adjustment;
 504 }
 505 
 506 //---------------------------dependence_graph---------------------------
 507 // Construct dependency graph.
 508 // Add dependence edges to load/store nodes for memory dependence
 509 //    A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
 510 void SuperWord::dependence_graph() {
 511   // First, assign a dependence node to each memory node
 512   for (int i = 0; i < _block.length(); i++ ) {
 513     Node *n = _block.at(i);
 514     if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
 515       _dg.make_node(n);
 516     }
 517   }
 518 
 519   // For each memory slice, create the dependences
 520   for (int i = 0; i < _mem_slice_head.length(); i++) {
 521     Node* n      = _mem_slice_head.at(i);
 522     Node* n_tail = _mem_slice_tail.at(i);
 523 
 524     // Get slice in predecessor order (last is first)
 525     mem_slice_preds(n_tail, n, _nlist);
 526 
 527     // Make the slice dependent on the root
 528     DepMem* slice = _dg.dep(n);
 529     _dg.make_edge(_dg.root(), slice);
 530 
 531     // Create a sink for the slice
 532     DepMem* slice_sink = _dg.make_node(NULL);
 533     _dg.make_edge(slice_sink, _dg.tail());
 534 
 535     // Now visit each pair of memory ops, creating the edges
 536     for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 537       Node* s1 = _nlist.at(j);
 538 
 539       // If no dependency yet, use slice
 540       if (_dg.dep(s1)->in_cnt() == 0) {
 541         _dg.make_edge(slice, s1);
 542       }
 543       SWPointer p1(s1->as_Mem(), this);
 544       bool sink_dependent = true;
 545       for (int k = j - 1; k >= 0; k--) {
 546         Node* s2 = _nlist.at(k);
 547         if (s1->is_Load() && s2->is_Load())
 548           continue;
 549         SWPointer p2(s2->as_Mem(), this);
 550 
 551         int cmp = p1.cmp(p2);
 552         if (SuperWordRTDepCheck &&
 553             p1.base() != p2.base() && p1.valid() && p2.valid()) {
 554           // Create a runtime check to disambiguate
 555           OrderedPair pp(p1.base(), p2.base());
 556           _disjoint_ptrs.append_if_missing(pp);
 557         } else if (!SWPointer::not_equal(cmp)) {
 558           // Possibly same address
 559           _dg.make_edge(s1, s2);
 560           sink_dependent = false;
 561         }
 562       }
 563       if (sink_dependent) {
 564         _dg.make_edge(s1, slice_sink);
 565       }
 566     }
 567 #ifndef PRODUCT
 568     if (TraceSuperWord) {
 569       tty->print_cr("\nDependence graph for slice: %d", n->_idx);
 570       for (int q = 0; q < _nlist.length(); q++) {
 571         _dg.print(_nlist.at(q));
 572       }
 573       tty->cr();
 574     }
 575 #endif
 576     _nlist.clear();
 577   }
 578 
 579 #ifndef PRODUCT
 580   if (TraceSuperWord) {
 581     tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
 582     for (int r = 0; r < _disjoint_ptrs.length(); r++) {
 583       _disjoint_ptrs.at(r).print();
 584       tty->cr();
 585     }
 586     tty->cr();
 587   }
 588 #endif
 589 }
 590 
 591 //---------------------------mem_slice_preds---------------------------
 592 // Return a memory slice (node list) in predecessor order starting at "start"
 593 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
 594   assert(preds.length() == 0, "start empty");
 595   Node* n = start;
 596   Node* prev = NULL;
 597   while (true) {
 598     assert(in_bb(n), "must be in block");
 599     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 600       Node* out = n->fast_out(i);
 601       if (out->is_Load()) {
 602         if (in_bb(out)) {
 603           preds.push(out);
 604         }
 605       } else {
 606         // FIXME
 607         if (out->is_MergeMem() && !in_bb(out)) {
 608           // Either unrolling is causing a memory edge not to disappear,
 609           // or need to run igvn.optimize() again before SLP
 610         } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
 611           // Ditto.  Not sure what else to check further.
 612         } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
 613           // StoreCM has an input edge used as a precedence edge.
 614           // Maybe an issue when oop stores are vectorized.
 615         } else {
 616           assert(out == prev || prev == NULL, "no branches off of store slice");
 617         }
 618       }
 619     }
 620     if (n == stop) break;
 621     preds.push(n);
 622     prev = n;
 623     assert(n->is_Mem(), err_msg_res("unexpected node %s", n->Name()));
 624     n = n->in(MemNode::Memory);
 625   }
 626 }
 627 
 628 //------------------------------stmts_can_pack---------------------------
 629 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
 630 // s1 aligned at "align"
 631 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
 632 
 633   // Do not use superword for non-primitives
 634   BasicType bt1 = velt_basic_type(s1);
 635   BasicType bt2 = velt_basic_type(s2);
 636   if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
 637     return false;
 638   if (Matcher::max_vector_size(bt1) < 2) {
 639     return false; // No vectors for this type
 640   }
 641 
 642   if (isomorphic(s1, s2)) {
 643     if (independent(s1, s2)) {
 644       if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
 645         if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
 646           int s1_align = alignment(s1);
 647           int s2_align = alignment(s2);
 648           if (s1_align == top_align || s1_align == align) {
 649             if (s2_align == top_align || s2_align == align + data_size(s1)) {
 650               return true;
 651             }
 652           }
 653         }
 654       }
 655     }
 656   }
 657   return false;
 658 }
 659 
 660 //------------------------------exists_at---------------------------
 661 // Does s exist in a pack at position pos?
 662 bool SuperWord::exists_at(Node* s, uint pos) {
 663   for (int i = 0; i < _packset.length(); i++) {
 664     Node_List* p = _packset.at(i);
 665     if (p->at(pos) == s) {
 666       return true;
 667     }
 668   }
 669   return false;
 670 }
 671 
 672 //------------------------------are_adjacent_refs---------------------------
 673 // Is s1 immediately before s2 in memory?
 674 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
 675   if (!s1->is_Mem() || !s2->is_Mem()) return false;
 676   if (!in_bb(s1)    || !in_bb(s2))    return false;
 677 
 678   // Do not use superword for non-primitives
 679   if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
 680       !is_java_primitive(s2->as_Mem()->memory_type())) {
 681     return false;
 682   }
 683 
 684   // FIXME - co_locate_pack fails on Stores in different mem-slices, so
 685   // only pack memops that are in the same alias set until that's fixed.
 686   if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
 687       _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
 688     return false;
 689   SWPointer p1(s1->as_Mem(), this);
 690   SWPointer p2(s2->as_Mem(), this);
 691   if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
 692   int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
 693   return diff == data_size(s1);
 694 }
 695 
 696 //------------------------------isomorphic---------------------------
 697 // Are s1 and s2 similar?
 698 bool SuperWord::isomorphic(Node* s1, Node* s2) {
 699   if (s1->Opcode() != s2->Opcode()) return false;
 700   if (s1->req() != s2->req()) return false;
 701   if (s1->in(0) != s2->in(0)) return false;
 702   if (!same_velt_type(s1, s2)) return false;
 703   return true;
 704 }
 705 
 706 //------------------------------independent---------------------------
 707 // Is there no data path from s1 to s2 or s2 to s1?
 708 bool SuperWord::independent(Node* s1, Node* s2) {
 709   //  assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
 710   int d1 = depth(s1);
 711   int d2 = depth(s2);
 712   if (d1 == d2) return s1 != s2;
 713   Node* deep    = d1 > d2 ? s1 : s2;
 714   Node* shallow = d1 > d2 ? s2 : s1;
 715 
 716   visited_clear();
 717 
 718   return independent_path(shallow, deep);
 719 }
 720 
 721 //------------------------------independent_path------------------------------
 722 // Helper for independent
 723 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
 724   if (dp >= 1000) return false; // stop deep recursion
 725   visited_set(deep);
 726   int shal_depth = depth(shallow);
 727   assert(shal_depth <= depth(deep), "must be");
 728   for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
 729     Node* pred = preds.current();
 730     if (in_bb(pred) && !visited_test(pred)) {
 731       if (shallow == pred) {
 732         return false;
 733       }
 734       if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
 735         return false;
 736       }
 737     }
 738   }
 739   return true;
 740 }
 741 
 742 //------------------------------set_alignment---------------------------
 743 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
 744   set_alignment(s1, align);
 745   if (align == top_align || align == bottom_align) {
 746     set_alignment(s2, align);
 747   } else {
 748     set_alignment(s2, align + data_size(s1));
 749   }
 750 }
 751 
 752 //------------------------------data_size---------------------------
 753 int SuperWord::data_size(Node* s) {
 754   int bsize = type2aelembytes(velt_basic_type(s));
 755   assert(bsize != 0, "valid size");
 756   return bsize;
 757 }
 758 
 759 //------------------------------extend_packlist---------------------------
 760 // Extend packset by following use->def and def->use links from pack members.
 761 void SuperWord::extend_packlist() {
 762   bool changed;
 763   do {
 764     changed = false;
 765     for (int i = 0; i < _packset.length(); i++) {
 766       Node_List* p = _packset.at(i);
 767       changed |= follow_use_defs(p);
 768       changed |= follow_def_uses(p);
 769     }
 770   } while (changed);
 771 
 772 #ifndef PRODUCT
 773   if (TraceSuperWord) {
 774     tty->print_cr("\nAfter extend_packlist");
 775     print_packset();
 776   }
 777 #endif
 778 }
 779 
 780 //------------------------------follow_use_defs---------------------------
 781 // Extend the packset by visiting operand definitions of nodes in pack p
 782 bool SuperWord::follow_use_defs(Node_List* p) {
 783   assert(p->size() == 2, "just checking");
 784   Node* s1 = p->at(0);
 785   Node* s2 = p->at(1);
 786   assert(s1->req() == s2->req(), "just checking");
 787   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
 788 
 789   if (s1->is_Load()) return false;
 790 
 791   int align = alignment(s1);
 792   bool changed = false;
 793   int start = s1->is_Store() ? MemNode::ValueIn   : 1;
 794   int end   = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
 795   for (int j = start; j < end; j++) {
 796     Node* t1 = s1->in(j);
 797     Node* t2 = s2->in(j);
 798     if (!in_bb(t1) || !in_bb(t2))
 799       continue;
 800     if (stmts_can_pack(t1, t2, align)) {
 801       if (est_savings(t1, t2) >= 0) {
 802         Node_List* pair = new Node_List();
 803         pair->push(t1);
 804         pair->push(t2);
 805         _packset.append(pair);
 806         set_alignment(t1, t2, align);
 807         changed = true;
 808       }
 809     }
 810   }
 811   return changed;
 812 }
 813 
 814 //------------------------------follow_def_uses---------------------------
 815 // Extend the packset by visiting uses of nodes in pack p
 816 bool SuperWord::follow_def_uses(Node_List* p) {
 817   bool changed = false;
 818   Node* s1 = p->at(0);
 819   Node* s2 = p->at(1);
 820   assert(p->size() == 2, "just checking");
 821   assert(s1->req() == s2->req(), "just checking");
 822   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
 823 
 824   if (s1->is_Store()) return false;
 825 
 826   int align = alignment(s1);
 827   int savings = -1;
 828   Node* u1 = NULL;
 829   Node* u2 = NULL;
 830   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
 831     Node* t1 = s1->fast_out(i);
 832     if (!in_bb(t1)) continue;
 833     for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
 834       Node* t2 = s2->fast_out(j);
 835       if (!in_bb(t2)) continue;
 836       if (!opnd_positions_match(s1, t1, s2, t2))
 837         continue;
 838       if (stmts_can_pack(t1, t2, align)) {
 839         int my_savings = est_savings(t1, t2);
 840         if (my_savings > savings) {
 841           savings = my_savings;
 842           u1 = t1;
 843           u2 = t2;
 844         }
 845       }
 846     }
 847   }
 848   if (savings >= 0) {
 849     Node_List* pair = new Node_List();
 850     pair->push(u1);
 851     pair->push(u2);
 852     _packset.append(pair);
 853     set_alignment(u1, u2, align);
 854     changed = true;
 855   }
 856   return changed;
 857 }
 858 
 859 //---------------------------opnd_positions_match-------------------------
 860 // Is the use of d1 in u1 at the same operand position as d2 in u2?
 861 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
 862   uint ct = u1->req();
 863   if (ct != u2->req()) return false;
 864   uint i1 = 0;
 865   uint i2 = 0;
 866   do {
 867     for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
 868     for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
 869     if (i1 != i2) {
 870       if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
 871         // Further analysis relies on operands position matching.
 872         u2->swap_edges(i1, i2);
 873       } else {
 874         return false;
 875       }
 876     }
 877   } while (i1 < ct);
 878   return true;
 879 }
 880 
 881 //------------------------------est_savings---------------------------
 882 // Estimate the savings from executing s1 and s2 as a pack
 883 int SuperWord::est_savings(Node* s1, Node* s2) {
 884   int save_in = 2 - 1; // 2 operations per instruction in packed form
 885 
 886   // inputs
 887   for (uint i = 1; i < s1->req(); i++) {
 888     Node* x1 = s1->in(i);
 889     Node* x2 = s2->in(i);
 890     if (x1 != x2) {
 891       if (are_adjacent_refs(x1, x2)) {
 892         save_in += adjacent_profit(x1, x2);
 893       } else if (!in_packset(x1, x2)) {
 894         save_in -= pack_cost(2);
 895       } else {
 896         save_in += unpack_cost(2);
 897       }
 898     }
 899   }
 900 
 901   // uses of result
 902   uint ct = 0;
 903   int save_use = 0;
 904   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
 905     Node* s1_use = s1->fast_out(i);
 906     for (int j = 0; j < _packset.length(); j++) {
 907       Node_List* p = _packset.at(j);
 908       if (p->at(0) == s1_use) {
 909         for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
 910           Node* s2_use = s2->fast_out(k);
 911           if (p->at(p->size()-1) == s2_use) {
 912             ct++;
 913             if (are_adjacent_refs(s1_use, s2_use)) {
 914               save_use += adjacent_profit(s1_use, s2_use);
 915             }
 916           }
 917         }
 918       }
 919     }
 920   }
 921 
 922   if (ct < s1->outcnt()) save_use += unpack_cost(1);
 923   if (ct < s2->outcnt()) save_use += unpack_cost(1);
 924 
 925   return MAX2(save_in, save_use);
 926 }
 927 
 928 //------------------------------costs---------------------------
 929 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
 930 int SuperWord::pack_cost(int ct)   { return ct; }
 931 int SuperWord::unpack_cost(int ct) { return ct; }
 932 
 933 //------------------------------combine_packs---------------------------
 934 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
 935 void SuperWord::combine_packs() {
 936   bool changed = true;
 937   // Combine packs regardless max vector size.
 938   while (changed) {
 939     changed = false;
 940     for (int i = 0; i < _packset.length(); i++) {
 941       Node_List* p1 = _packset.at(i);
 942       if (p1 == NULL) continue;
 943       for (int j = 0; j < _packset.length(); j++) {
 944         Node_List* p2 = _packset.at(j);
 945         if (p2 == NULL) continue;
 946         if (i == j) continue;
 947         if (p1->at(p1->size()-1) == p2->at(0)) {
 948           for (uint k = 1; k < p2->size(); k++) {
 949             p1->push(p2->at(k));
 950           }
 951           _packset.at_put(j, NULL);
 952           changed = true;
 953         }
 954       }
 955     }
 956   }
 957 
 958   // Split packs which have size greater then max vector size.
 959   for (int i = 0; i < _packset.length(); i++) {
 960     Node_List* p1 = _packset.at(i);
 961     if (p1 != NULL) {
 962       BasicType bt = velt_basic_type(p1->at(0));
 963       uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
 964       assert(is_power_of_2(max_vlen), "sanity");
 965       uint psize = p1->size();
 966       if (!is_power_of_2(psize)) {
 967         // Skip pack which can't be vector.
 968         // case1: for(...) { a[i] = i; }    elements values are different (i+x)
 969         // case2: for(...) { a[i] = b[i+1]; }  can't align both, load and store
 970         _packset.at_put(i, NULL);
 971         continue;
 972       }
 973       if (psize > max_vlen) {
 974         Node_List* pack = new Node_List();
 975         for (uint j = 0; j < psize; j++) {
 976           pack->push(p1->at(j));
 977           if (pack->size() >= max_vlen) {
 978             assert(is_power_of_2(pack->size()), "sanity");
 979             _packset.append(pack);
 980             pack = new Node_List();
 981           }
 982         }
 983         _packset.at_put(i, NULL);
 984       }
 985     }
 986   }
 987 
 988   // Compress list.
 989   for (int i = _packset.length() - 1; i >= 0; i--) {
 990     Node_List* p1 = _packset.at(i);
 991     if (p1 == NULL) {
 992       _packset.remove_at(i);
 993     }
 994   }
 995 
 996 #ifndef PRODUCT
 997   if (TraceSuperWord) {
 998     tty->print_cr("\nAfter combine_packs");
 999     print_packset();
1000   }
1001 #endif
1002 }
1003 
1004 //-----------------------------construct_my_pack_map--------------------------
1005 // Construct the map from nodes to packs.  Only valid after the
1006 // point where a node is only in one pack (after combine_packs).
1007 void SuperWord::construct_my_pack_map() {
1008   Node_List* rslt = NULL;
1009   for (int i = 0; i < _packset.length(); i++) {
1010     Node_List* p = _packset.at(i);
1011     for (uint j = 0; j < p->size(); j++) {
1012       Node* s = p->at(j);
1013       assert(my_pack(s) == NULL, "only in one pack");
1014       set_my_pack(s, p);
1015     }
1016   }
1017 }
1018 
1019 //------------------------------filter_packs---------------------------
1020 // Remove packs that are not implemented or not profitable.
1021 void SuperWord::filter_packs() {
1022 
1023   // Remove packs that are not implemented
1024   for (int i = _packset.length() - 1; i >= 0; i--) {
1025     Node_List* pk = _packset.at(i);
1026     bool impl = implemented(pk);
1027     if (!impl) {
1028 #ifndef PRODUCT
1029       if (TraceSuperWord && Verbose) {
1030         tty->print_cr("Unimplemented");
1031         pk->at(0)->dump();
1032       }
1033 #endif
1034       remove_pack_at(i);
1035     }
1036   }
1037 
1038   // Remove packs that are not profitable
1039   bool changed;
1040   do {
1041     changed = false;
1042     for (int i = _packset.length() - 1; i >= 0; i--) {
1043       Node_List* pk = _packset.at(i);
1044       bool prof = profitable(pk);
1045       if (!prof) {
1046 #ifndef PRODUCT
1047         if (TraceSuperWord && Verbose) {
1048           tty->print_cr("Unprofitable");
1049           pk->at(0)->dump();
1050         }
1051 #endif
1052         remove_pack_at(i);
1053         changed = true;
1054       }
1055     }
1056   } while (changed);
1057 
1058 #ifndef PRODUCT
1059   if (TraceSuperWord) {
1060     tty->print_cr("\nAfter filter_packs");
1061     print_packset();
1062     tty->cr();
1063   }
1064 #endif
1065 }
1066 
1067 //------------------------------implemented---------------------------
1068 // Can code be generated for pack p?
1069 bool SuperWord::implemented(Node_List* p) {
1070   Node* p0 = p->at(0);
1071   return VectorNode::implemented(p0->Opcode(), p->size(), velt_basic_type(p0));
1072 }
1073 
1074 //------------------------------same_inputs--------------------------
1075 // For pack p, are all idx operands the same?
1076 static bool same_inputs(Node_List* p, int idx) {
1077   Node* p0 = p->at(0);
1078   uint vlen = p->size();
1079   Node* p0_def = p0->in(idx);
1080   for (uint i = 1; i < vlen; i++) {
1081     Node* pi = p->at(i);
1082     Node* pi_def = pi->in(idx);
1083     if (p0_def != pi_def)
1084       return false;
1085   }
1086   return true;
1087 }
1088 
1089 //------------------------------profitable---------------------------
1090 // For pack p, are all operands and all uses (with in the block) vector?
1091 bool SuperWord::profitable(Node_List* p) {
1092   Node* p0 = p->at(0);
1093   uint start, end;
1094   VectorNode::vector_operands(p0, &start, &end);
1095 
1096   // Return false if some inputs are not vectors or vectors with different
1097   // size or alignment.
1098   // Also, for now, return false if not scalar promotion case when inputs are
1099   // the same. Later, implement PackNode and allow differing, non-vector inputs
1100   // (maybe just the ones from outside the block.)
1101   for (uint i = start; i < end; i++) {
1102     if (!is_vector_use(p0, i))
1103       return false;
1104   }
1105   if (VectorNode::is_shift(p0)) {
1106     // For now, return false if shift count is vector or not scalar promotion
1107     // case (different shift counts) because it is not supported yet.
1108     Node* cnt = p0->in(2);
1109     Node_List* cnt_pk = my_pack(cnt);
1110     if (cnt_pk != NULL)
1111       return false;
1112     if (!same_inputs(p, 2))
1113       return false;
1114   }
1115   if (!p0->is_Store()) {
1116     // For now, return false if not all uses are vector.
1117     // Later, implement ExtractNode and allow non-vector uses (maybe
1118     // just the ones outside the block.)
1119     for (uint i = 0; i < p->size(); i++) {
1120       Node* def = p->at(i);
1121       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1122         Node* use = def->fast_out(j);
1123         for (uint k = 0; k < use->req(); k++) {
1124           Node* n = use->in(k);
1125           if (def == n) {
1126             if (!is_vector_use(use, k)) {
1127               return false;
1128             }
1129           }
1130         }
1131       }
1132     }
1133   }
1134   return true;
1135 }
1136 
1137 //------------------------------schedule---------------------------
1138 // Adjust the memory graph for the packed operations
1139 void SuperWord::schedule() {
1140 
1141   // Co-locate in the memory graph the members of each memory pack
1142   for (int i = 0; i < _packset.length(); i++) {
1143     co_locate_pack(_packset.at(i));
1144   }
1145 }
1146 
1147 //-------------------------------remove_and_insert-------------------
1148 // Remove "current" from its current position in the memory graph and insert
1149 // it after the appropriate insertion point (lip or uip).
1150 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1151                                   Node *uip, Unique_Node_List &sched_before) {
1152   Node* my_mem = current->in(MemNode::Memory);
1153   bool sched_up = sched_before.member(current);
1154 
1155   // remove current_store from its current position in the memmory graph
1156   for (DUIterator i = current->outs(); current->has_out(i); i++) {
1157     Node* use = current->out(i);
1158     if (use->is_Mem()) {
1159       assert(use->in(MemNode::Memory) == current, "must be");
1160       if (use == prev) { // connect prev to my_mem
1161           _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1162           --i; //deleted this edge; rescan position
1163       } else if (sched_before.member(use)) {
1164         if (!sched_up) { // Will be moved together with current
1165           _igvn.replace_input_of(use, MemNode::Memory, uip);
1166           --i; //deleted this edge; rescan position
1167         }
1168       } else {
1169         if (sched_up) { // Will be moved together with current
1170           _igvn.replace_input_of(use, MemNode::Memory, lip);
1171           --i; //deleted this edge; rescan position
1172         }
1173       }
1174     }
1175   }
1176 
1177   Node *insert_pt =  sched_up ?  uip : lip;
1178 
1179   // all uses of insert_pt's memory state should use current's instead
1180   for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1181     Node* use = insert_pt->out(i);
1182     if (use->is_Mem()) {
1183       assert(use->in(MemNode::Memory) == insert_pt, "must be");
1184       _igvn.replace_input_of(use, MemNode::Memory, current);
1185       --i; //deleted this edge; rescan position
1186     } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1187       uint pos; //lip (lower insert point) must be the last one in the memory slice
1188       for (pos=1; pos < use->req(); pos++) {
1189         if (use->in(pos) == insert_pt) break;
1190       }
1191       _igvn.replace_input_of(use, pos, current);
1192       --i;
1193     }
1194   }
1195 
1196   //connect current to insert_pt
1197   _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
1198 }
1199 
1200 //------------------------------co_locate_pack----------------------------------
1201 // To schedule a store pack, we need to move any sandwiched memory ops either before
1202 // or after the pack, based upon dependence information:
1203 // (1) If any store in the pack depends on the sandwiched memory op, the
1204 //     sandwiched memory op must be scheduled BEFORE the pack;
1205 // (2) If a sandwiched memory op depends on any store in the pack, the
1206 //     sandwiched memory op must be scheduled AFTER the pack;
1207 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1208 //     memory op (say memB), memB must be scheduled before memA. So, if memA is
1209 //     scheduled before the pack, memB must also be scheduled before the pack;
1210 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1211 //     schedule this store AFTER the pack
1212 // (5) We know there is no dependence cycle, so there in no other case;
1213 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1214 //
1215 // To schedule a load pack, we use the memory state of either the first or the last load in
1216 // the pack, based on the dependence constraint.
1217 void SuperWord::co_locate_pack(Node_List* pk) {
1218   if (pk->at(0)->is_Store()) {
1219     MemNode* first     = executed_first(pk)->as_Mem();
1220     MemNode* last      = executed_last(pk)->as_Mem();
1221     Unique_Node_List schedule_before_pack;
1222     Unique_Node_List memops;
1223 
1224     MemNode* current   = last->in(MemNode::Memory)->as_Mem();
1225     MemNode* previous  = last;
1226     while (true) {
1227       assert(in_bb(current), "stay in block");
1228       memops.push(previous);
1229       for (DUIterator i = current->outs(); current->has_out(i); i++) {
1230         Node* use = current->out(i);
1231         if (use->is_Mem() && use != previous)
1232           memops.push(use);
1233       }
1234       if (current == first) break;
1235       previous = current;
1236       current  = current->in(MemNode::Memory)->as_Mem();
1237     }
1238 
1239     // determine which memory operations should be scheduled before the pack
1240     for (uint i = 1; i < memops.size(); i++) {
1241       Node *s1 = memops.at(i);
1242       if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1243         for (uint j = 0; j< i; j++) {
1244           Node *s2 = memops.at(j);
1245           if (!independent(s1, s2)) {
1246             if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1247               schedule_before_pack.push(s1); // s1 must be scheduled before
1248               Node_List* mem_pk = my_pack(s1);
1249               if (mem_pk != NULL) {
1250                 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1251                   Node* s = mem_pk->at(ii);  // follow partner
1252                   if (memops.member(s) && !schedule_before_pack.member(s))
1253                     schedule_before_pack.push(s);
1254                 }
1255               }
1256               break;
1257             }
1258           }
1259         }
1260       }
1261     }
1262 
1263     Node*    upper_insert_pt = first->in(MemNode::Memory);
1264     // Following code moves loads connected to upper_insert_pt below aliased stores.
1265     // Collect such loads here and reconnect them back to upper_insert_pt later.
1266     memops.clear();
1267     for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
1268       Node* use = upper_insert_pt->out(i);
1269       if (!use->is_Store())
1270         memops.push(use);
1271     }
1272 
1273     MemNode* lower_insert_pt = last;
1274     previous                 = last; //previous store in pk
1275     current                  = last->in(MemNode::Memory)->as_Mem();
1276 
1277     // start scheduling from "last" to "first"
1278     while (true) {
1279       assert(in_bb(current), "stay in block");
1280       assert(in_pack(previous, pk), "previous stays in pack");
1281       Node* my_mem = current->in(MemNode::Memory);
1282 
1283       if (in_pack(current, pk)) {
1284         // Forward users of my memory state (except "previous) to my input memory state
1285         for (DUIterator i = current->outs(); current->has_out(i); i++) {
1286           Node* use = current->out(i);
1287           if (use->is_Mem() && use != previous) {
1288             assert(use->in(MemNode::Memory) == current, "must be");
1289             if (schedule_before_pack.member(use)) {
1290               _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
1291             } else {
1292               _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
1293             }
1294             --i; // deleted this edge; rescan position
1295           }
1296         }
1297         previous = current;
1298       } else { // !in_pack(current, pk) ==> a sandwiched store
1299         remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1300       }
1301 
1302       if (current == first) break;
1303       current = my_mem->as_Mem();
1304     } // end while
1305 
1306     // Reconnect loads back to upper_insert_pt.
1307     for (uint i = 0; i < memops.size(); i++) {
1308       Node *ld = memops.at(i);
1309       if (ld->in(MemNode::Memory) != upper_insert_pt) {
1310         _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
1311       }
1312     }
1313   } else if (pk->at(0)->is_Load()) { //load
1314     // all loads in the pack should have the same memory state. By default,
1315     // we use the memory state of the last load. However, if any load could
1316     // not be moved down due to the dependence constraint, we use the memory
1317     // state of the first load.
1318     Node* last_mem  = executed_last(pk)->in(MemNode::Memory);
1319     Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1320     bool schedule_last = true;
1321     for (uint i = 0; i < pk->size(); i++) {
1322       Node* ld = pk->at(i);
1323       for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1324            current=current->in(MemNode::Memory)) {
1325         assert(current != first_mem, "corrupted memory graph");
1326         if(current->is_Mem() && !independent(current, ld)){
1327           schedule_last = false; // a later store depends on this load
1328           break;
1329         }
1330       }
1331     }
1332 
1333     Node* mem_input = schedule_last ? last_mem : first_mem;
1334     _igvn.hash_delete(mem_input);
1335     // Give each load the same memory state
1336     for (uint i = 0; i < pk->size(); i++) {
1337       LoadNode* ld = pk->at(i)->as_Load();
1338       _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
1339     }
1340   }
1341 }
1342 
1343 //------------------------------output---------------------------
1344 // Convert packs into vector node operations
1345 void SuperWord::output() {
1346   if (_packset.length() == 0) return;
1347 
1348 #ifndef PRODUCT
1349   if (TraceLoopOpts) {
1350     tty->print("SuperWord    ");
1351     lpt()->dump_head();
1352   }
1353 #endif
1354 
1355   // MUST ENSURE main loop's initial value is properly aligned:
1356   //  (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1357 
1358   align_initial_loop_index(align_to_ref());
1359 
1360   // Insert extract (unpack) operations for scalar uses
1361   for (int i = 0; i < _packset.length(); i++) {
1362     insert_extracts(_packset.at(i));
1363   }
1364 
1365   Compile* C = _phase->C;
1366   uint max_vlen_in_bytes = 0;
1367   for (int i = 0; i < _block.length(); i++) {
1368     Node* n = _block.at(i);
1369     Node_List* p = my_pack(n);
1370     if (p && n == executed_last(p)) {
1371       uint vlen = p->size();
1372       uint vlen_in_bytes = 0;
1373       Node* vn = NULL;
1374       Node* low_adr = p->at(0);
1375       Node* first   = executed_first(p);
1376       int   opc = n->Opcode();
1377       if (n->is_Load()) {
1378         Node* ctl = n->in(MemNode::Control);
1379         Node* mem = first->in(MemNode::Memory);
1380         Node* adr = low_adr->in(MemNode::Address);
1381         const TypePtr* atyp = n->adr_type();
1382         vn = LoadVectorNode::make(C, opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n));
1383         vlen_in_bytes = vn->as_LoadVector()->memory_size();
1384       } else if (n->is_Store()) {
1385         // Promote value to be stored to vector
1386         Node* val = vector_opd(p, MemNode::ValueIn);
1387         Node* ctl = n->in(MemNode::Control);
1388         Node* mem = first->in(MemNode::Memory);
1389         Node* adr = low_adr->in(MemNode::Address);
1390         const TypePtr* atyp = n->adr_type();
1391         vn = StoreVectorNode::make(C, opc, ctl, mem, adr, atyp, val, vlen);
1392         vlen_in_bytes = vn->as_StoreVector()->memory_size();
1393       } else if (n->req() == 3) {
1394         // Promote operands to vector
1395         Node* in1 = vector_opd(p, 1);
1396         Node* in2 = vector_opd(p, 2);
1397         if (VectorNode::is_invariant_vector(in1) && (n->is_Add() || n->is_Mul())) {
1398           // Move invariant vector input into second position to avoid register spilling.
1399           Node* tmp = in1;
1400           in1 = in2;
1401           in2 = tmp;
1402         }
1403         vn = VectorNode::make(C, opc, in1, in2, vlen, velt_basic_type(n));
1404         vlen_in_bytes = vn->as_Vector()->length_in_bytes();
1405       } else {
1406         ShouldNotReachHere();
1407       }
1408       assert(vn != NULL, "sanity");
1409       _igvn.register_new_node_with_optimizer(vn);
1410       _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1411       for (uint j = 0; j < p->size(); j++) {
1412         Node* pm = p->at(j);
1413         _igvn.replace_node(pm, vn);
1414       }
1415       _igvn._worklist.push(vn);
1416 
1417       if (vlen_in_bytes > max_vlen_in_bytes) {
1418         max_vlen_in_bytes = vlen_in_bytes;
1419       }
1420 #ifdef ASSERT
1421       if (TraceNewVectors) {
1422         tty->print("new Vector node: ");
1423         vn->dump();
1424       }
1425 #endif
1426     }
1427   }
1428   C->set_max_vector_size(max_vlen_in_bytes);
1429 }
1430 
1431 //------------------------------vector_opd---------------------------
1432 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1433 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1434   Node* p0 = p->at(0);
1435   uint vlen = p->size();
1436   Node* opd = p0->in(opd_idx);
1437 
1438   if (same_inputs(p, opd_idx)) {
1439     if (opd->is_Vector() || opd->is_LoadVector()) {
1440       assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
1441       return opd; // input is matching vector
1442     }
1443     if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
1444       Compile* C = _phase->C;
1445       Node* cnt = opd;
1446       // Vector instructions do not mask shift count, do it here.
1447       juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
1448       const TypeInt* t = opd->find_int_type();
1449       if (t != NULL && t->is_con()) {
1450         juint shift = t->get_con();
1451         if (shift > mask) { // Unsigned cmp
1452           cnt = ConNode::make(C, TypeInt::make(shift & mask));
1453         }
1454       } else {
1455         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
1456           cnt = ConNode::make(C, TypeInt::make(mask));
1457           _igvn.register_new_node_with_optimizer(cnt);
1458           cnt = new (C) AndINode(opd, cnt);
1459           _igvn.register_new_node_with_optimizer(cnt);
1460           _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1461         }
1462         assert(opd->bottom_type()->isa_int(), "int type only");
1463         // Move non constant shift count into vector register.
1464         cnt = VectorNode::shift_count(C, p0, cnt, vlen, velt_basic_type(p0));
1465       }
1466       if (cnt != opd) {
1467         _igvn.register_new_node_with_optimizer(cnt);
1468         _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1469       }
1470       return cnt;
1471     }
1472     assert(!opd->is_StoreVector(), "such vector is not expected here");
1473     // Convert scalar input to vector with the same number of elements as
1474     // p0's vector. Use p0's type because size of operand's container in
1475     // vector should match p0's size regardless operand's size.
1476     const Type* p0_t = velt_type(p0);
1477     VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, p0_t);
1478 
1479     _igvn.register_new_node_with_optimizer(vn);
1480     _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1481 #ifdef ASSERT
1482     if (TraceNewVectors) {
1483       tty->print("new Vector node: ");
1484       vn->dump();
1485     }
1486 #endif
1487     return vn;
1488   }
1489 
1490   // Insert pack operation
1491   BasicType bt = velt_basic_type(p0);
1492   PackNode* pk = PackNode::make(_phase->C, opd, vlen, bt);
1493   DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
1494 
1495   for (uint i = 1; i < vlen; i++) {
1496     Node* pi = p->at(i);
1497     Node* in = pi->in(opd_idx);
1498     assert(my_pack(in) == NULL, "Should already have been unpacked");
1499     assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
1500     pk->add_opd(in);
1501   }
1502   _igvn.register_new_node_with_optimizer(pk);
1503   _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1504 #ifdef ASSERT
1505   if (TraceNewVectors) {
1506     tty->print("new Vector node: ");
1507     pk->dump();
1508   }
1509 #endif
1510   return pk;
1511 }
1512 
1513 //------------------------------insert_extracts---------------------------
1514 // If a use of pack p is not a vector use, then replace the
1515 // use with an extract operation.
1516 void SuperWord::insert_extracts(Node_List* p) {
1517   if (p->at(0)->is_Store()) return;
1518   assert(_n_idx_list.is_empty(), "empty (node,index) list");
1519 
1520   // Inspect each use of each pack member.  For each use that is
1521   // not a vector use, replace the use with an extract operation.
1522 
1523   for (uint i = 0; i < p->size(); i++) {
1524     Node* def = p->at(i);
1525     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1526       Node* use = def->fast_out(j);
1527       for (uint k = 0; k < use->req(); k++) {
1528         Node* n = use->in(k);
1529         if (def == n) {
1530           if (!is_vector_use(use, k)) {
1531             _n_idx_list.push(use, k);
1532           }
1533         }
1534       }
1535     }
1536   }
1537 
1538   while (_n_idx_list.is_nonempty()) {
1539     Node* use = _n_idx_list.node();
1540     int   idx = _n_idx_list.index();
1541     _n_idx_list.pop();
1542     Node* def = use->in(idx);
1543 
1544     // Insert extract operation
1545     _igvn.hash_delete(def);
1546     int def_pos = alignment(def) / data_size(def);
1547 
1548     Node* ex = ExtractNode::make(_phase->C, def, def_pos, velt_basic_type(def));
1549     _igvn.register_new_node_with_optimizer(ex);
1550     _phase->set_ctrl(ex, _phase->get_ctrl(def));
1551     _igvn.replace_input_of(use, idx, ex);
1552     _igvn._worklist.push(def);
1553 
1554     bb_insert_after(ex, bb_idx(def));
1555     set_velt_type(ex, velt_type(def));
1556   }
1557 }
1558 
1559 //------------------------------is_vector_use---------------------------
1560 // Is use->in(u_idx) a vector use?
1561 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1562   Node_List* u_pk = my_pack(use);
1563   if (u_pk == NULL) return false;
1564   Node* def = use->in(u_idx);
1565   Node_List* d_pk = my_pack(def);
1566   if (d_pk == NULL) {
1567     // check for scalar promotion
1568     Node* n = u_pk->at(0)->in(u_idx);
1569     for (uint i = 1; i < u_pk->size(); i++) {
1570       if (u_pk->at(i)->in(u_idx) != n) return false;
1571     }
1572     return true;
1573   }
1574   if (u_pk->size() != d_pk->size())
1575     return false;
1576   for (uint i = 0; i < u_pk->size(); i++) {
1577     Node* ui = u_pk->at(i);
1578     Node* di = d_pk->at(i);
1579     if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1580       return false;
1581   }
1582   return true;
1583 }
1584 
1585 //------------------------------construct_bb---------------------------
1586 // Construct reverse postorder list of block members
1587 bool SuperWord::construct_bb() {
1588   Node* entry = bb();
1589 
1590   assert(_stk.length() == 0,            "stk is empty");
1591   assert(_block.length() == 0,          "block is empty");
1592   assert(_data_entry.length() == 0,     "data_entry is empty");
1593   assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1594   assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1595 
1596   // Find non-control nodes with no inputs from within block,
1597   // create a temporary map from node _idx to bb_idx for use
1598   // by the visited and post_visited sets,
1599   // and count number of nodes in block.
1600   int bb_ct = 0;
1601   for (uint i = 0; i < lpt()->_body.size(); i++ ) {
1602     Node *n = lpt()->_body.at(i);
1603     set_bb_idx(n, i); // Create a temporary map
1604     if (in_bb(n)) {
1605       if (n->is_LoadStore() || n->is_MergeMem() ||
1606           (n->is_Proj() && !n->as_Proj()->is_CFG())) {
1607         // Bailout if the loop has LoadStore, MergeMem or data Proj
1608         // nodes. Superword optimization does not work with them.
1609         return false;
1610       }
1611       bb_ct++;
1612       if (!n->is_CFG()) {
1613         bool found = false;
1614         for (uint j = 0; j < n->req(); j++) {
1615           Node* def = n->in(j);
1616           if (def && in_bb(def)) {
1617             found = true;
1618             break;
1619           }
1620         }
1621         if (!found) {
1622           assert(n != entry, "can't be entry");
1623           _data_entry.push(n);
1624         }
1625       }
1626     }
1627   }
1628 
1629   // Find memory slices (head and tail)
1630   for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1631     Node *n = lp()->fast_out(i);
1632     if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1633       Node* n_tail  = n->in(LoopNode::LoopBackControl);
1634       if (n_tail != n->in(LoopNode::EntryControl)) {
1635         if (!n_tail->is_Mem()) {
1636           assert(n_tail->is_Mem(), err_msg_res("unexpected node for memory slice: %s", n_tail->Name()));
1637           return false; // Bailout
1638         }
1639         _mem_slice_head.push(n);
1640         _mem_slice_tail.push(n_tail);
1641       }
1642     }
1643   }
1644 
1645   // Create an RPO list of nodes in block
1646 
1647   visited_clear();
1648   post_visited_clear();
1649 
1650   // Push all non-control nodes with no inputs from within block, then control entry
1651   for (int j = 0; j < _data_entry.length(); j++) {
1652     Node* n = _data_entry.at(j);
1653     visited_set(n);
1654     _stk.push(n);
1655   }
1656   visited_set(entry);
1657   _stk.push(entry);
1658 
1659   // Do a depth first walk over out edges
1660   int rpo_idx = bb_ct - 1;
1661   int size;
1662   while ((size = _stk.length()) > 0) {
1663     Node* n = _stk.top(); // Leave node on stack
1664     if (!visited_test_set(n)) {
1665       // forward arc in graph
1666     } else if (!post_visited_test(n)) {
1667       // cross or back arc
1668       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1669         Node *use = n->fast_out(i);
1670         if (in_bb(use) && !visited_test(use) &&
1671             // Don't go around backedge
1672             (!use->is_Phi() || n == entry)) {
1673           _stk.push(use);
1674         }
1675       }
1676       if (_stk.length() == size) {
1677         // There were no additional uses, post visit node now
1678         _stk.pop(); // Remove node from stack
1679         assert(rpo_idx >= 0, "");
1680         _block.at_put_grow(rpo_idx, n);
1681         rpo_idx--;
1682         post_visited_set(n);
1683         assert(rpo_idx >= 0 || _stk.is_empty(), "");
1684       }
1685     } else {
1686       _stk.pop(); // Remove post-visited node from stack
1687     }
1688   }
1689 
1690   // Create real map of block indices for nodes
1691   for (int j = 0; j < _block.length(); j++) {
1692     Node* n = _block.at(j);
1693     set_bb_idx(n, j);
1694   }
1695 
1696   initialize_bb(); // Ensure extra info is allocated.
1697 
1698 #ifndef PRODUCT
1699   if (TraceSuperWord) {
1700     print_bb();
1701     tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1702     for (int m = 0; m < _data_entry.length(); m++) {
1703       tty->print("%3d ", m);
1704       _data_entry.at(m)->dump();
1705     }
1706     tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1707     for (int m = 0; m < _mem_slice_head.length(); m++) {
1708       tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1709       tty->print("    ");    _mem_slice_tail.at(m)->dump();
1710     }
1711   }
1712 #endif
1713   assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1714   return (_mem_slice_head.length() > 0) || (_data_entry.length() > 0);
1715 }
1716 
1717 //------------------------------initialize_bb---------------------------
1718 // Initialize per node info
1719 void SuperWord::initialize_bb() {
1720   Node* last = _block.at(_block.length() - 1);
1721   grow_node_info(bb_idx(last));
1722 }
1723 
1724 //------------------------------bb_insert_after---------------------------
1725 // Insert n into block after pos
1726 void SuperWord::bb_insert_after(Node* n, int pos) {
1727   int n_pos = pos + 1;
1728   // Make room
1729   for (int i = _block.length() - 1; i >= n_pos; i--) {
1730     _block.at_put_grow(i+1, _block.at(i));
1731   }
1732   for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1733     _node_info.at_put_grow(j+1, _node_info.at(j));
1734   }
1735   // Set value
1736   _block.at_put_grow(n_pos, n);
1737   _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1738   // Adjust map from node->_idx to _block index
1739   for (int i = n_pos; i < _block.length(); i++) {
1740     set_bb_idx(_block.at(i), i);
1741   }
1742 }
1743 
1744 //------------------------------compute_max_depth---------------------------
1745 // Compute max depth for expressions from beginning of block
1746 // Use to prune search paths during test for independence.
1747 void SuperWord::compute_max_depth() {
1748   int ct = 0;
1749   bool again;
1750   do {
1751     again = false;
1752     for (int i = 0; i < _block.length(); i++) {
1753       Node* n = _block.at(i);
1754       if (!n->is_Phi()) {
1755         int d_orig = depth(n);
1756         int d_in   = 0;
1757         for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1758           Node* pred = preds.current();
1759           if (in_bb(pred)) {
1760             d_in = MAX2(d_in, depth(pred));
1761           }
1762         }
1763         if (d_in + 1 != d_orig) {
1764           set_depth(n, d_in + 1);
1765           again = true;
1766         }
1767       }
1768     }
1769     ct++;
1770   } while (again);
1771 #ifndef PRODUCT
1772   if (TraceSuperWord && Verbose)
1773     tty->print_cr("compute_max_depth iterated: %d times", ct);
1774 #endif
1775 }
1776 
1777 //-------------------------compute_vector_element_type-----------------------
1778 // Compute necessary vector element type for expressions
1779 // This propagates backwards a narrower integer type when the
1780 // upper bits of the value are not needed.
1781 // Example:  char a,b,c;  a = b + c;
1782 // Normally the type of the add is integer, but for packed character
1783 // operations the type of the add needs to be char.
1784 void SuperWord::compute_vector_element_type() {
1785 #ifndef PRODUCT
1786   if (TraceSuperWord && Verbose)
1787     tty->print_cr("\ncompute_velt_type:");
1788 #endif
1789 
1790   // Initial type
1791   for (int i = 0; i < _block.length(); i++) {
1792     Node* n = _block.at(i);
1793     set_velt_type(n, container_type(n));
1794   }
1795 
1796   // Propagate integer narrowed type backwards through operations
1797   // that don't depend on higher order bits
1798   for (int i = _block.length() - 1; i >= 0; i--) {
1799     Node* n = _block.at(i);
1800     // Only integer types need be examined
1801     const Type* vtn = velt_type(n);
1802     if (vtn->basic_type() == T_INT) {
1803       uint start, end;
1804       VectorNode::vector_operands(n, &start, &end);
1805 
1806       for (uint j = start; j < end; j++) {
1807         Node* in  = n->in(j);
1808         // Don't propagate through a memory
1809         if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
1810             data_size(n) < data_size(in)) {
1811           bool same_type = true;
1812           for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1813             Node *use = in->fast_out(k);
1814             if (!in_bb(use) || !same_velt_type(use, n)) {
1815               same_type = false;
1816               break;
1817             }
1818           }
1819           if (same_type) {
1820             // For right shifts of small integer types (bool, byte, char, short)
1821             // we need precise information about sign-ness. Only Load nodes have
1822             // this information because Store nodes are the same for signed and
1823             // unsigned values. And any arithmetic operation after a load may
1824             // expand a value to signed Int so such right shifts can't be used
1825             // because vector elements do not have upper bits of Int.
1826             const Type* vt = vtn;
1827             if (VectorNode::is_shift(in)) {
1828               Node* load = in->in(1);
1829               if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
1830                 vt = velt_type(load);
1831               } else if (in->Opcode() != Op_LShiftI) {
1832                 // Widen type to Int to avoid creation of right shift vector
1833                 // (align + data_size(s1) check in stmts_can_pack() will fail).
1834                 // Note, left shifts work regardless type.
1835                 vt = TypeInt::INT;
1836               }
1837             }
1838             set_velt_type(in, vt);
1839           }
1840         }
1841       }
1842     }
1843   }
1844 #ifndef PRODUCT
1845   if (TraceSuperWord && Verbose) {
1846     for (int i = 0; i < _block.length(); i++) {
1847       Node* n = _block.at(i);
1848       velt_type(n)->dump();
1849       tty->print("\t");
1850       n->dump();
1851     }
1852   }
1853 #endif
1854 }
1855 
1856 //------------------------------memory_alignment---------------------------
1857 // Alignment within a vector memory reference
1858 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
1859   SWPointer p(s, this);
1860   if (!p.valid()) {
1861     return bottom_align;
1862   }
1863   int vw = vector_width_in_bytes(s);
1864   if (vw < 2) {
1865     return bottom_align; // No vectors for this type
1866   }
1867   int offset  = p.offset_in_bytes();
1868   offset     += iv_adjust*p.memory_size();
1869   int off_rem = offset % vw;
1870   int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
1871   return off_mod;
1872 }
1873 
1874 //---------------------------container_type---------------------------
1875 // Smallest type containing range of values
1876 const Type* SuperWord::container_type(Node* n) {
1877   if (n->is_Mem()) {
1878     BasicType bt = n->as_Mem()->memory_type();
1879     if (n->is_Store() && (bt == T_CHAR)) {
1880       // Use T_SHORT type instead of T_CHAR for stored values because any
1881       // preceding arithmetic operation extends values to signed Int.
1882       bt = T_SHORT;
1883     }
1884     if (n->Opcode() == Op_LoadUB) {
1885       // Adjust type for unsigned byte loads, it is important for right shifts.
1886       // T_BOOLEAN is used because there is no basic type representing type
1887       // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
1888       // size (one byte) and sign is important.
1889       bt = T_BOOLEAN;
1890     }
1891     return Type::get_const_basic_type(bt);
1892   }
1893   const Type* t = _igvn.type(n);
1894   if (t->basic_type() == T_INT) {
1895     // A narrow type of arithmetic operations will be determined by
1896     // propagating the type of memory operations.
1897     return TypeInt::INT;
1898   }
1899   return t;
1900 }
1901 
1902 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
1903   const Type* vt1 = velt_type(n1);
1904   const Type* vt2 = velt_type(n2);
1905   if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
1906     // Compare vectors element sizes for integer types.
1907     return data_size(n1) == data_size(n2);
1908   }
1909   return vt1 == vt2;
1910 }
1911 
1912 //------------------------------in_packset---------------------------
1913 // Are s1 and s2 in a pack pair and ordered as s1,s2?
1914 bool SuperWord::in_packset(Node* s1, Node* s2) {
1915   for (int i = 0; i < _packset.length(); i++) {
1916     Node_List* p = _packset.at(i);
1917     assert(p->size() == 2, "must be");
1918     if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
1919       return true;
1920     }
1921   }
1922   return false;
1923 }
1924 
1925 //------------------------------in_pack---------------------------
1926 // Is s in pack p?
1927 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
1928   for (uint i = 0; i < p->size(); i++) {
1929     if (p->at(i) == s) {
1930       return p;
1931     }
1932   }
1933   return NULL;
1934 }
1935 
1936 //------------------------------remove_pack_at---------------------------
1937 // Remove the pack at position pos in the packset
1938 void SuperWord::remove_pack_at(int pos) {
1939   Node_List* p = _packset.at(pos);
1940   for (uint i = 0; i < p->size(); i++) {
1941     Node* s = p->at(i);
1942     set_my_pack(s, NULL);
1943   }
1944   _packset.remove_at(pos);
1945 }
1946 
1947 //------------------------------executed_first---------------------------
1948 // Return the node executed first in pack p.  Uses the RPO block list
1949 // to determine order.
1950 Node* SuperWord::executed_first(Node_List* p) {
1951   Node* n = p->at(0);
1952   int n_rpo = bb_idx(n);
1953   for (uint i = 1; i < p->size(); i++) {
1954     Node* s = p->at(i);
1955     int s_rpo = bb_idx(s);
1956     if (s_rpo < n_rpo) {
1957       n = s;
1958       n_rpo = s_rpo;
1959     }
1960   }
1961   return n;
1962 }
1963 
1964 //------------------------------executed_last---------------------------
1965 // Return the node executed last in pack p.
1966 Node* SuperWord::executed_last(Node_List* p) {
1967   Node* n = p->at(0);
1968   int n_rpo = bb_idx(n);
1969   for (uint i = 1; i < p->size(); i++) {
1970     Node* s = p->at(i);
1971     int s_rpo = bb_idx(s);
1972     if (s_rpo > n_rpo) {
1973       n = s;
1974       n_rpo = s_rpo;
1975     }
1976   }
1977   return n;
1978 }
1979 
1980 //----------------------------align_initial_loop_index---------------------------
1981 // Adjust pre-loop limit so that in main loop, a load/store reference
1982 // to align_to_ref will be a position zero in the vector.
1983 //   (iv + k) mod vector_align == 0
1984 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
1985   CountedLoopNode *main_head = lp()->as_CountedLoop();
1986   assert(main_head->is_main_loop(), "");
1987   CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
1988   assert(pre_end != NULL, "we must have a correct pre-loop");
1989   Node *pre_opaq1 = pre_end->limit();
1990   assert(pre_opaq1->Opcode() == Op_Opaque1, "");
1991   Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1992   Node *lim0 = pre_opaq->in(1);
1993 
1994   // Where we put new limit calculations
1995   Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1996 
1997   // Ensure the original loop limit is available from the
1998   // pre-loop Opaque1 node.
1999   Node *orig_limit = pre_opaq->original_loop_limit();
2000   assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
2001 
2002   SWPointer align_to_ref_p(align_to_ref, this);
2003   assert(align_to_ref_p.valid(), "sanity");
2004 
2005   // Given:
2006   //     lim0 == original pre loop limit
2007   //     V == v_align (power of 2)
2008   //     invar == extra invariant piece of the address expression
2009   //     e == offset [ +/- invar ]
2010   //
2011   // When reassociating expressions involving '%' the basic rules are:
2012   //     (a - b) % k == 0   =>  a % k == b % k
2013   // and:
2014   //     (a + b) % k == 0   =>  a % k == (k - b) % k
2015   //
2016   // For stride > 0 && scale > 0,
2017   //   Derive the new pre-loop limit "lim" such that the two constraints:
2018   //     (1) lim = lim0 + N           (where N is some positive integer < V)
2019   //     (2) (e + lim) % V == 0
2020   //   are true.
2021   //
2022   //   Substituting (1) into (2),
2023   //     (e + lim0 + N) % V == 0
2024   //   solve for N:
2025   //     N = (V - (e + lim0)) % V
2026   //   substitute back into (1), so that new limit
2027   //     lim = lim0 + (V - (e + lim0)) % V
2028   //
2029   // For stride > 0 && scale < 0
2030   //   Constraints:
2031   //     lim = lim0 + N
2032   //     (e - lim) % V == 0
2033   //   Solving for lim:
2034   //     (e - lim0 - N) % V == 0
2035   //     N = (e - lim0) % V
2036   //     lim = lim0 + (e - lim0) % V
2037   //
2038   // For stride < 0 && scale > 0
2039   //   Constraints:
2040   //     lim = lim0 - N
2041   //     (e + lim) % V == 0
2042   //   Solving for lim:
2043   //     (e + lim0 - N) % V == 0
2044   //     N = (e + lim0) % V
2045   //     lim = lim0 - (e + lim0) % V
2046   //
2047   // For stride < 0 && scale < 0
2048   //   Constraints:
2049   //     lim = lim0 - N
2050   //     (e - lim) % V == 0
2051   //   Solving for lim:
2052   //     (e - lim0 + N) % V == 0
2053   //     N = (V - (e - lim0)) % V
2054   //     lim = lim0 - (V - (e - lim0)) % V
2055 
2056   int vw = vector_width_in_bytes(align_to_ref);
2057   int stride   = iv_stride();
2058   int scale    = align_to_ref_p.scale_in_bytes();
2059   int elt_size = align_to_ref_p.memory_size();
2060   int v_align  = vw / elt_size;
2061   assert(v_align > 1, "sanity");
2062   int offset   = align_to_ref_p.offset_in_bytes() / elt_size;
2063   Node *offsn  = _igvn.intcon(offset);
2064 
2065   Node *e = offsn;
2066   if (align_to_ref_p.invar() != NULL) {
2067     // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
2068     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2069     Node* aref     = new (_phase->C) URShiftINode(align_to_ref_p.invar(), log2_elt);
2070     _igvn.register_new_node_with_optimizer(aref);
2071     _phase->set_ctrl(aref, pre_ctrl);
2072     if (align_to_ref_p.negate_invar()) {
2073       e = new (_phase->C) SubINode(e, aref);
2074     } else {
2075       e = new (_phase->C) AddINode(e, aref);
2076     }
2077     _igvn.register_new_node_with_optimizer(e);
2078     _phase->set_ctrl(e, pre_ctrl);
2079   }
2080   if (vw > ObjectAlignmentInBytes) {
2081     // incorporate base e +/- base && Mask >>> log2(elt)
2082     Node* xbase = new(_phase->C) CastP2XNode(NULL, align_to_ref_p.base());
2083     _igvn.register_new_node_with_optimizer(xbase);
2084 #ifdef _LP64
2085     xbase  = new (_phase->C) ConvL2INode(xbase);
2086     _igvn.register_new_node_with_optimizer(xbase);
2087 #endif
2088     Node* mask = _igvn.intcon(vw-1);
2089     Node* masked_xbase  = new (_phase->C) AndINode(xbase, mask);
2090     _igvn.register_new_node_with_optimizer(masked_xbase);
2091     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2092     Node* bref     = new (_phase->C) URShiftINode(masked_xbase, log2_elt);
2093     _igvn.register_new_node_with_optimizer(bref);
2094     _phase->set_ctrl(bref, pre_ctrl);
2095     e = new (_phase->C) AddINode(e, bref);
2096     _igvn.register_new_node_with_optimizer(e);
2097     _phase->set_ctrl(e, pre_ctrl);
2098   }
2099 
2100   // compute e +/- lim0
2101   if (scale < 0) {
2102     e = new (_phase->C) SubINode(e, lim0);
2103   } else {
2104     e = new (_phase->C) AddINode(e, lim0);
2105   }
2106   _igvn.register_new_node_with_optimizer(e);
2107   _phase->set_ctrl(e, pre_ctrl);
2108 
2109   if (stride * scale > 0) {
2110     // compute V - (e +/- lim0)
2111     Node* va  = _igvn.intcon(v_align);
2112     e = new (_phase->C) SubINode(va, e);
2113     _igvn.register_new_node_with_optimizer(e);
2114     _phase->set_ctrl(e, pre_ctrl);
2115   }
2116   // compute N = (exp) % V
2117   Node* va_msk = _igvn.intcon(v_align - 1);
2118   Node* N = new (_phase->C) AndINode(e, va_msk);
2119   _igvn.register_new_node_with_optimizer(N);
2120   _phase->set_ctrl(N, pre_ctrl);
2121 
2122   //   substitute back into (1), so that new limit
2123   //     lim = lim0 + N
2124   Node* lim;
2125   if (stride < 0) {
2126     lim = new (_phase->C) SubINode(lim0, N);
2127   } else {
2128     lim = new (_phase->C) AddINode(lim0, N);
2129   }
2130   _igvn.register_new_node_with_optimizer(lim);
2131   _phase->set_ctrl(lim, pre_ctrl);
2132   Node* constrained =
2133     (stride > 0) ? (Node*) new (_phase->C) MinINode(lim, orig_limit)
2134                  : (Node*) new (_phase->C) MaxINode(lim, orig_limit);
2135   _igvn.register_new_node_with_optimizer(constrained);
2136   _phase->set_ctrl(constrained, pre_ctrl);
2137   _igvn.hash_delete(pre_opaq);
2138   pre_opaq->set_req(1, constrained);
2139 }
2140 
2141 //----------------------------get_pre_loop_end---------------------------
2142 // Find pre loop end from main loop.  Returns null if none.
2143 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
2144   Node *ctrl = cl->in(LoopNode::EntryControl);
2145   if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
2146   Node *iffm = ctrl->in(0);
2147   if (!iffm->is_If()) return NULL;
2148   Node *p_f = iffm->in(0);
2149   if (!p_f->is_IfFalse()) return NULL;
2150   if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
2151   CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2152   CountedLoopNode* loop_node = pre_end->loopnode();
2153   if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
2154   return pre_end;
2155 }
2156 
2157 
2158 //------------------------------init---------------------------
2159 void SuperWord::init() {
2160   _dg.init();
2161   _packset.clear();
2162   _disjoint_ptrs.clear();
2163   _block.clear();
2164   _data_entry.clear();
2165   _mem_slice_head.clear();
2166   _mem_slice_tail.clear();
2167   _node_info.clear();
2168   _align_to_ref = NULL;
2169   _lpt = NULL;
2170   _lp = NULL;
2171   _bb = NULL;
2172   _iv = NULL;
2173 }
2174 
2175 //------------------------------print_packset---------------------------
2176 void SuperWord::print_packset() {
2177 #ifndef PRODUCT
2178   tty->print_cr("packset");
2179   for (int i = 0; i < _packset.length(); i++) {
2180     tty->print_cr("Pack: %d", i);
2181     Node_List* p = _packset.at(i);
2182     print_pack(p);
2183   }
2184 #endif
2185 }
2186 
2187 //------------------------------print_pack---------------------------
2188 void SuperWord::print_pack(Node_List* p) {
2189   for (uint i = 0; i < p->size(); i++) {
2190     print_stmt(p->at(i));
2191   }
2192 }
2193 
2194 //------------------------------print_bb---------------------------
2195 void SuperWord::print_bb() {
2196 #ifndef PRODUCT
2197   tty->print_cr("\nBlock");
2198   for (int i = 0; i < _block.length(); i++) {
2199     Node* n = _block.at(i);
2200     tty->print("%d ", i);
2201     if (n) {
2202       n->dump();
2203     }
2204   }
2205 #endif
2206 }
2207 
2208 //------------------------------print_stmt---------------------------
2209 void SuperWord::print_stmt(Node* s) {
2210 #ifndef PRODUCT
2211   tty->print(" align: %d \t", alignment(s));
2212   s->dump();
2213 #endif
2214 }
2215 
2216 //------------------------------blank---------------------------
2217 char* SuperWord::blank(uint depth) {
2218   static char blanks[101];
2219   assert(depth < 101, "too deep");
2220   for (uint i = 0; i < depth; i++) blanks[i] = ' ';
2221   blanks[depth] = '\0';
2222   return blanks;
2223 }
2224 
2225 
2226 //==============================SWPointer===========================
2227 
2228 //----------------------------SWPointer------------------------
2229 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
2230   _mem(mem), _slp(slp),  _base(NULL),  _adr(NULL),
2231   _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
2232 
2233   Node* adr = mem->in(MemNode::Address);
2234   if (!adr->is_AddP()) {
2235     assert(!valid(), "too complex");
2236     return;
2237   }
2238   // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
2239   Node* base = adr->in(AddPNode::Base);
2240   //unsafe reference could not be aligned appropriately without runtime checking
2241   if (base == NULL || base->bottom_type() == Type::TOP) {
2242     assert(!valid(), "unsafe access");
2243     return;
2244   }
2245   for (int i = 0; i < 3; i++) {
2246     if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
2247       assert(!valid(), "too complex");
2248       return;
2249     }
2250     adr = adr->in(AddPNode::Address);
2251     if (base == adr || !adr->is_AddP()) {
2252       break; // stop looking at addp's
2253     }
2254   }
2255   _base = base;
2256   _adr  = adr;
2257   assert(valid(), "Usable");
2258 }
2259 
2260 // Following is used to create a temporary object during
2261 // the pattern match of an address expression.
2262 SWPointer::SWPointer(SWPointer* p) :
2263   _mem(p->_mem), _slp(p->_slp),  _base(NULL),  _adr(NULL),
2264   _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
2265 
2266 //------------------------scaled_iv_plus_offset--------------------
2267 // Match: k*iv + offset
2268 // where: k is a constant that maybe zero, and
2269 //        offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
2270 bool SWPointer::scaled_iv_plus_offset(Node* n) {
2271   if (scaled_iv(n)) {
2272     return true;
2273   }
2274   if (offset_plus_k(n)) {
2275     return true;
2276   }
2277   int opc = n->Opcode();
2278   if (opc == Op_AddI) {
2279     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
2280       return true;
2281     }
2282     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2283       return true;
2284     }
2285   } else if (opc == Op_SubI) {
2286     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
2287       return true;
2288     }
2289     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2290       _scale *= -1;
2291       return true;
2292     }
2293   }
2294   return false;
2295 }
2296 
2297 //----------------------------scaled_iv------------------------
2298 // Match: k*iv where k is a constant that's not zero
2299 bool SWPointer::scaled_iv(Node* n) {
2300   if (_scale != 0) {
2301     return false;  // already found a scale
2302   }
2303   if (n == iv()) {
2304     _scale = 1;
2305     return true;
2306   }
2307   int opc = n->Opcode();
2308   if (opc == Op_MulI) {
2309     if (n->in(1) == iv() && n->in(2)->is_Con()) {
2310       _scale = n->in(2)->get_int();
2311       return true;
2312     } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2313       _scale = n->in(1)->get_int();
2314       return true;
2315     }
2316   } else if (opc == Op_LShiftI) {
2317     if (n->in(1) == iv() && n->in(2)->is_Con()) {
2318       _scale = 1 << n->in(2)->get_int();
2319       return true;
2320     }
2321   } else if (opc == Op_ConvI2L) {
2322     if (scaled_iv_plus_offset(n->in(1))) {
2323       return true;
2324     }
2325   } else if (opc == Op_LShiftL) {
2326     if (!has_iv() && _invar == NULL) {
2327       // Need to preserve the current _offset value, so
2328       // create a temporary object for this expression subtree.
2329       // Hacky, so should re-engineer the address pattern match.
2330       SWPointer tmp(this);
2331       if (tmp.scaled_iv_plus_offset(n->in(1))) {
2332         if (tmp._invar == NULL) {
2333           int mult = 1 << n->in(2)->get_int();
2334           _scale   = tmp._scale  * mult;
2335           _offset += tmp._offset * mult;
2336           return true;
2337         }
2338       }
2339     }
2340   }
2341   return false;
2342 }
2343 
2344 //----------------------------offset_plus_k------------------------
2345 // Match: offset is (k [+/- invariant])
2346 // where k maybe zero and invariant is optional, but not both.
2347 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2348   int opc = n->Opcode();
2349   if (opc == Op_ConI) {
2350     _offset += negate ? -(n->get_int()) : n->get_int();
2351     return true;
2352   } else if (opc == Op_ConL) {
2353     // Okay if value fits into an int
2354     const TypeLong* t = n->find_long_type();
2355     if (t->higher_equal(TypeLong::INT)) {
2356       jlong loff = n->get_long();
2357       jint  off  = (jint)loff;
2358       _offset += negate ? -off : loff;
2359       return true;
2360     }
2361     return false;
2362   }
2363   if (_invar != NULL) return false; // already have an invariant
2364   if (opc == Op_AddI) {
2365     if (n->in(2)->is_Con() && invariant(n->in(1))) {
2366       _negate_invar = negate;
2367       _invar = n->in(1);
2368       _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2369       return true;
2370     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2371       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2372       _negate_invar = negate;
2373       _invar = n->in(2);
2374       return true;
2375     }
2376   }
2377   if (opc == Op_SubI) {
2378     if (n->in(2)->is_Con() && invariant(n->in(1))) {
2379       _negate_invar = negate;
2380       _invar = n->in(1);
2381       _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2382       return true;
2383     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2384       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2385       _negate_invar = !negate;
2386       _invar = n->in(2);
2387       return true;
2388     }
2389   }
2390   if (invariant(n)) {
2391     _negate_invar = negate;
2392     _invar = n;
2393     return true;
2394   }
2395   return false;
2396 }
2397 
2398 //----------------------------print------------------------
2399 void SWPointer::print() {
2400 #ifndef PRODUCT
2401   tty->print("base: %d  adr: %d  scale: %d  offset: %d  invar: %c%d\n",
2402              _base != NULL ? _base->_idx : 0,
2403              _adr  != NULL ? _adr->_idx  : 0,
2404              _scale, _offset,
2405              _negate_invar?'-':'+',
2406              _invar != NULL ? _invar->_idx : 0);
2407 #endif
2408 }
2409 
2410 // ========================= OrderedPair =====================
2411 
2412 const OrderedPair OrderedPair::initial;
2413 
2414 // ========================= SWNodeInfo =====================
2415 
2416 const SWNodeInfo SWNodeInfo::initial;
2417 
2418 
2419 // ============================ DepGraph ===========================
2420 
2421 //------------------------------make_node---------------------------
2422 // Make a new dependence graph node for an ideal node.
2423 DepMem* DepGraph::make_node(Node* node) {
2424   DepMem* m = new (_arena) DepMem(node);
2425   if (node != NULL) {
2426     assert(_map.at_grow(node->_idx) == NULL, "one init only");
2427     _map.at_put_grow(node->_idx, m);
2428   }
2429   return m;
2430 }
2431 
2432 //------------------------------make_edge---------------------------
2433 // Make a new dependence graph edge from dpred -> dsucc
2434 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2435   DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2436   dpred->set_out_head(e);
2437   dsucc->set_in_head(e);
2438   return e;
2439 }
2440 
2441 // ========================== DepMem ========================
2442 
2443 //------------------------------in_cnt---------------------------
2444 int DepMem::in_cnt() {
2445   int ct = 0;
2446   for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2447   return ct;
2448 }
2449 
2450 //------------------------------out_cnt---------------------------
2451 int DepMem::out_cnt() {
2452   int ct = 0;
2453   for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2454   return ct;
2455 }
2456 
2457 //------------------------------print-----------------------------
2458 void DepMem::print() {
2459 #ifndef PRODUCT
2460   tty->print("  DepNode %d (", _node->_idx);
2461   for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2462     Node* pred = p->pred()->node();
2463     tty->print(" %d", pred != NULL ? pred->_idx : 0);
2464   }
2465   tty->print(") [");
2466   for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2467     Node* succ = s->succ()->node();
2468     tty->print(" %d", succ != NULL ? succ->_idx : 0);
2469   }
2470   tty->print_cr(" ]");
2471 #endif
2472 }
2473 
2474 // =========================== DepEdge =========================
2475 
2476 //------------------------------DepPreds---------------------------
2477 void DepEdge::print() {
2478 #ifndef PRODUCT
2479   tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2480 #endif
2481 }
2482 
2483 // =========================== DepPreds =========================
2484 // Iterator over predecessor edges in the dependence graph.
2485 
2486 //------------------------------DepPreds---------------------------
2487 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2488   _n = n;
2489   _done = false;
2490   if (_n->is_Store() || _n->is_Load()) {
2491     _next_idx = MemNode::Address;
2492     _end_idx  = n->req();
2493     _dep_next = dg.dep(_n)->in_head();
2494   } else if (_n->is_Mem()) {
2495     _next_idx = 0;
2496     _end_idx  = 0;
2497     _dep_next = dg.dep(_n)->in_head();
2498   } else {
2499     _next_idx = 1;
2500     _end_idx  = _n->req();
2501     _dep_next = NULL;
2502   }
2503   next();
2504 }
2505 
2506 //------------------------------next---------------------------
2507 void DepPreds::next() {
2508   if (_dep_next != NULL) {
2509     _current  = _dep_next->pred()->node();
2510     _dep_next = _dep_next->next_in();
2511   } else if (_next_idx < _end_idx) {
2512     _current  = _n->in(_next_idx++);
2513   } else {
2514     _done = true;
2515   }
2516 }
2517 
2518 // =========================== DepSuccs =========================
2519 // Iterator over successor edges in the dependence graph.
2520 
2521 //------------------------------DepSuccs---------------------------
2522 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2523   _n = n;
2524   _done = false;
2525   if (_n->is_Load()) {
2526     _next_idx = 0;
2527     _end_idx  = _n->outcnt();
2528     _dep_next = dg.dep(_n)->out_head();
2529   } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2530     _next_idx = 0;
2531     _end_idx  = 0;
2532     _dep_next = dg.dep(_n)->out_head();
2533   } else {
2534     _next_idx = 0;
2535     _end_idx  = _n->outcnt();
2536     _dep_next = NULL;
2537   }
2538   next();
2539 }
2540 
2541 //-------------------------------next---------------------------
2542 void DepSuccs::next() {
2543   if (_dep_next != NULL) {
2544     _current  = _dep_next->succ()->node();
2545     _dep_next = _dep_next->next_out();
2546   } else if (_next_idx < _end_idx) {
2547     _current  = _n->raw_out(_next_idx++);
2548   } else {
2549     _done = true;
2550   }
2551 }