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/divnode.hpp"
  31 #include "opto/matcher.hpp"
  32 #include "opto/memnode.hpp"
  33 #include "opto/mulnode.hpp"
  34 #include "opto/opcodes.hpp"
  35 #include "opto/superword.hpp"
  36 #include "opto/vectornode.hpp"
  37 
  38 //
  39 //                  S U P E R W O R D   T R A N S F O R M
  40 //=============================================================================
  41 
  42 //------------------------------SuperWord---------------------------
  43 SuperWord::SuperWord(PhaseIdealLoop* phase) :
  44   _phase(phase),
  45   _igvn(phase->_igvn),
  46   _arena(phase->C->comp_arena()),
  47   _packset(arena(), 8,  0, NULL),         // packs for the current block
  48   _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
  49   _block(arena(), 8,  0, NULL),           // nodes in current block
  50   _data_entry(arena(), 8,  0, NULL),      // nodes with all inputs from outside
  51   _mem_slice_head(arena(), 8,  0, NULL),  // memory slice heads
  52   _mem_slice_tail(arena(), 8,  0, NULL),  // memory slice tails
  53   _node_info(arena(), 8,  0, SWNodeInfo::initial), // info needed per node
  54   _align_to_ref(NULL),                    // memory reference to align vectors to
  55   _disjoint_ptrs(arena(), 8,  0, OrderedPair::initial), // runtime disambiguated pointer pairs
  56   _dg(_arena),                            // dependence graph
  57   _visited(arena()),                      // visited node set
  58   _post_visited(arena()),                 // post visited node set
  59   _n_idx_list(arena(), 8),                // scratch list of (node,index) pairs
  60   _stk(arena(), 8, 0, NULL),              // scratch stack of nodes
  61   _nlist(arena(), 8, 0, NULL),            // scratch list of nodes
  62   _lpt(NULL),                             // loop tree node
  63   _lp(NULL),                              // LoopNode
  64   _bb(NULL),                              // basic block
  65   _iv(NULL)                               // induction var
  66 {}
  67 
  68 //------------------------------transform_loop---------------------------
  69 void SuperWord::transform_loop(IdealLoopTree* lpt) {
  70   assert(UseSuperWord, "should be");
  71   // Do vectors exist on this architecture?
  72   if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return;
  73 
  74   assert(lpt->_head->is_CountedLoop(), "must be");
  75   CountedLoopNode *cl = lpt->_head->as_CountedLoop();
  76 
  77   if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
  78 
  79   if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
  80 
  81   // Check for no control flow in body (other than exit)
  82   Node *cl_exit = cl->loopexit();
  83   if (cl_exit->in(0) != lpt->_head) return;
  84 
  85   // Make sure the are no extra control users of the loop backedge
  86   if (cl->back_control()->outcnt() != 1) {
  87     return;
  88   }
  89 
  90   // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
  91   CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
  92   if (pre_end == NULL) return;
  93   Node *pre_opaq1 = pre_end->limit();
  94   if (pre_opaq1->Opcode() != Op_Opaque1) return;
  95 
  96   init(); // initialize data structures
  97 
  98   set_lpt(lpt);
  99   set_lp(cl);
 100 
 101   // For now, define one block which is the entire loop body
 102   set_bb(cl);
 103 
 104   assert(_packset.length() == 0, "packset must be empty");
 105   SLP_extract();
 106 }
 107 
 108 //------------------------------SLP_extract---------------------------
 109 // Extract the superword level parallelism
 110 //
 111 // 1) A reverse post-order of nodes in the block is constructed.  By scanning
 112 //    this list from first to last, all definitions are visited before their uses.
 113 //
 114 // 2) A point-to-point dependence graph is constructed between memory references.
 115 //    This simplies the upcoming "independence" checker.
 116 //
 117 // 3) The maximum depth in the node graph from the beginning of the block
 118 //    to each node is computed.  This is used to prune the graph search
 119 //    in the independence checker.
 120 //
 121 // 4) For integer types, the necessary bit width is propagated backwards
 122 //    from stores to allow packed operations on byte, char, and short
 123 //    integers.  This reverses the promotion to type "int" that javac
 124 //    did for operations like: char c1,c2,c3;  c1 = c2 + c3.
 125 //
 126 // 5) One of the memory references is picked to be an aligned vector reference.
 127 //    The pre-loop trip count is adjusted to align this reference in the
 128 //    unrolled body.
 129 //
 130 // 6) The initial set of pack pairs is seeded with memory references.
 131 //
 132 // 7) The set of pack pairs is extended by following use->def and def->use links.
 133 //
 134 // 8) The pairs are combined into vector sized packs.
 135 //
 136 // 9) Reorder the memory slices to co-locate members of the memory packs.
 137 //
 138 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
 139 //    inserting scalar promotion, vector creation from multiple scalars, and
 140 //    extraction of scalar values from vectors.
 141 //
 142 void SuperWord::SLP_extract() {
 143 
 144   // Ready the block
 145 
 146   if (!construct_bb())
 147     return; // Exit if no interesting nodes or complex graph.
 148 
 149   dependence_graph();
 150 
 151   compute_max_depth();
 152 
 153   compute_vector_element_type();
 154 
 155   // Attempt vectorization
 156 
 157   find_adjacent_refs();
 158 
 159   extend_packlist();
 160 
 161   combine_packs();
 162 
 163   construct_my_pack_map();
 164 
 165   filter_packs();
 166 
 167   schedule();
 168 
 169   output();
 170 }
 171 
 172 //------------------------------find_adjacent_refs---------------------------
 173 // Find the adjacent memory references and create pack pairs for them.
 174 // This is the initial set of packs that will then be extended by
 175 // following use->def and def->use links.  The align positions are
 176 // assigned relative to the reference "align_to_ref"
 177 void SuperWord::find_adjacent_refs() {
 178   // Get list of memory operations
 179   Node_List memops;
 180   for (int i = 0; i < _block.length(); i++) {
 181     Node* n = _block.at(i);
 182     if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
 183         is_java_primitive(n->as_Mem()->memory_type())) {
 184       int align = memory_alignment(n->as_Mem(), 0);
 185       if (align != bottom_align) {
 186         memops.push(n);
 187       }
 188     }
 189   }
 190 
 191   Node_List align_to_refs;
 192   int best_iv_adjustment = 0;
 193   MemNode* best_align_to_mem_ref = NULL;
 194 
 195   while (memops.size() != 0) {
 196     // Find a memory reference to align to.
 197     MemNode* mem_ref = find_align_to_ref(memops);
 198     if (mem_ref == NULL) break;
 199     align_to_refs.push(mem_ref);
 200     int iv_adjustment = get_iv_adjustment(mem_ref);
 201 
 202     if (best_align_to_mem_ref == NULL) {
 203       // Set memory reference which is the best from all memory operations
 204       // to be used for alignment. The pre-loop trip count is modified to align
 205       // this reference to a vector-aligned address.
 206       best_align_to_mem_ref = mem_ref;
 207       best_iv_adjustment = iv_adjustment;
 208     }
 209 
 210     SWPointer align_to_ref_p(mem_ref, this);
 211     // Set alignment relative to "align_to_ref" for all related memory operations.
 212     for (int i = memops.size() - 1; i >= 0; i--) {
 213       MemNode* s = memops.at(i)->as_Mem();
 214       if (isomorphic(s, mem_ref)) {
 215         SWPointer p2(s, this);
 216         if (p2.comparable(align_to_ref_p)) {
 217           int align = memory_alignment(s, iv_adjustment);
 218           set_alignment(s, align);
 219         }
 220       }
 221     }
 222 
 223     // Create initial pack pairs of memory operations for which
 224     // alignment is set and vectors will be aligned.
 225     bool create_pack = true;
 226     if (memory_alignment(mem_ref, best_iv_adjustment) == 0) {
 227       if (!Matcher::misaligned_vectors_ok()) {
 228         int vw = vector_width(mem_ref);
 229         int vw_best = vector_width(best_align_to_mem_ref);
 230         if (vw > vw_best) {
 231           // Do not vectorize a memory access with more elements per vector
 232           // if unaligned memory access is not allowed because number of
 233           // iterations in pre-loop will be not enough to align it.
 234           create_pack = false;
 235         }
 236       }
 237     } else {
 238       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 239         // Can't allow vectorization of unaligned memory accesses with the
 240         // same type since it could be overlapped accesses to the same array.
 241         create_pack = false;
 242       } else {
 243         // Allow independent (different type) unaligned memory operations
 244         // if HW supports them.
 245         if (!Matcher::misaligned_vectors_ok()) {
 246           create_pack = false;
 247         } else {
 248           // Check if packs of the same memory type but
 249           // with a different alignment were created before.
 250           for (uint i = 0; i < align_to_refs.size(); i++) {
 251             MemNode* mr = align_to_refs.at(i)->as_Mem();
 252             if (same_velt_type(mr, mem_ref) &&
 253                 memory_alignment(mr, iv_adjustment) != 0)
 254               create_pack = false;
 255           }
 256         }
 257       }
 258     }
 259     if (create_pack) {
 260       for (uint i = 0; i < memops.size(); i++) {
 261         Node* s1 = memops.at(i);
 262         int align = alignment(s1);
 263         if (align == top_align) continue;
 264         for (uint j = 0; j < memops.size(); j++) {
 265           Node* s2 = memops.at(j);
 266           if (alignment(s2) == top_align) continue;
 267           if (s1 != s2 && are_adjacent_refs(s1, s2)) {
 268             if (stmts_can_pack(s1, s2, align)) {
 269               Node_List* pair = new Node_List();
 270               pair->push(s1);
 271               pair->push(s2);
 272               _packset.append(pair);
 273             }
 274           }
 275         }
 276       }
 277     } else { // Don't create unaligned pack
 278       // First, remove remaining memory ops of the same type from the list.
 279       for (int i = memops.size() - 1; i >= 0; i--) {
 280         MemNode* s = memops.at(i)->as_Mem();
 281         if (same_velt_type(s, mem_ref)) {
 282           memops.remove(i);
 283         }
 284       }
 285 
 286       // Second, remove already constructed packs of the same type.
 287       for (int i = _packset.length() - 1; i >= 0; i--) {
 288         Node_List* p = _packset.at(i);
 289         MemNode* s = p->at(0)->as_Mem();
 290         if (same_velt_type(s, mem_ref)) {
 291           remove_pack_at(i);
 292         }
 293       }
 294 
 295       // If needed find the best memory reference for loop alignment again.
 296       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 297         // Put memory ops from remaining packs back on memops list for
 298         // the best alignment search.
 299         uint orig_msize = memops.size();
 300         for (int i = 0; i < _packset.length(); i++) {
 301           Node_List* p = _packset.at(i);
 302           MemNode* s = p->at(0)->as_Mem();
 303           assert(!same_velt_type(s, mem_ref), "sanity");
 304           memops.push(s);
 305         }
 306         MemNode* best_align_to_mem_ref = find_align_to_ref(memops);
 307         if (best_align_to_mem_ref == NULL) break;
 308         best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
 309         // Restore list.
 310         while (memops.size() > orig_msize)
 311           (void)memops.pop();
 312       }
 313     } // unaligned memory accesses
 314 
 315     // Remove used mem nodes.
 316     for (int i = memops.size() - 1; i >= 0; i--) {
 317       MemNode* m = memops.at(i)->as_Mem();
 318       if (alignment(m) != top_align) {
 319         memops.remove(i);
 320       }
 321     }
 322 
 323   } // while (memops.size() != 0
 324   set_align_to_ref(best_align_to_mem_ref);
 325 
 326 #ifndef PRODUCT
 327   if (TraceSuperWord) {
 328     tty->print_cr("\nAfter find_adjacent_refs");
 329     print_packset();
 330   }
 331 #endif
 332 }
 333 
 334 //------------------------------find_align_to_ref---------------------------
 335 // Find a memory reference to align the loop induction variable to.
 336 // Looks first at stores then at loads, looking for a memory reference
 337 // with the largest number of references similar to it.
 338 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
 339   GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
 340 
 341   // Count number of comparable memory ops
 342   for (uint i = 0; i < memops.size(); i++) {
 343     MemNode* s1 = memops.at(i)->as_Mem();
 344     SWPointer p1(s1, this);
 345     // Discard if pre loop can't align this reference
 346     if (!ref_is_alignable(p1)) {
 347       *cmp_ct.adr_at(i) = 0;
 348       continue;
 349     }
 350     for (uint j = i+1; j < memops.size(); j++) {
 351       MemNode* s2 = memops.at(j)->as_Mem();
 352       if (isomorphic(s1, s2)) {
 353         SWPointer p2(s2, this);
 354         if (p1.comparable(p2)) {
 355           (*cmp_ct.adr_at(i))++;
 356           (*cmp_ct.adr_at(j))++;
 357         }
 358       }
 359     }
 360   }
 361 
 362   // Find Store (or Load) with the greatest number of "comparable" references,
 363   // biggest vector size, smallest data size and smallest iv offset.
 364   int max_ct        = 0;
 365   int max_vw        = 0;
 366   int max_idx       = -1;
 367   int min_size      = max_jint;
 368   int min_iv_offset = max_jint;
 369   for (uint j = 0; j < memops.size(); j++) {
 370     MemNode* s = memops.at(j)->as_Mem();
 371     if (s->is_Store()) {
 372       int vw = vector_width_in_bytes(s);
 373       assert(vw > 1, "sanity");
 374       SWPointer p(s, this);
 375       if (cmp_ct.at(j) >  max_ct ||
 376           cmp_ct.at(j) == max_ct &&
 377             (vw >  max_vw ||
 378              vw == max_vw &&
 379               (data_size(s) <  min_size ||
 380                data_size(s) == min_size &&
 381                  (p.offset_in_bytes() < min_iv_offset)))) {
 382         max_ct = cmp_ct.at(j);
 383         max_vw = vw;
 384         max_idx = j;
 385         min_size = data_size(s);
 386         min_iv_offset = p.offset_in_bytes();
 387       }
 388     }
 389   }
 390   // If no stores, look at loads
 391   if (max_ct == 0) {
 392     for (uint j = 0; j < memops.size(); j++) {
 393       MemNode* s = memops.at(j)->as_Mem();
 394       if (s->is_Load()) {
 395         int vw = vector_width_in_bytes(s);
 396         assert(vw > 1, "sanity");
 397         SWPointer p(s, this);
 398         if (cmp_ct.at(j) >  max_ct ||
 399             cmp_ct.at(j) == max_ct &&
 400               (vw >  max_vw ||
 401                vw == max_vw &&
 402                 (data_size(s) <  min_size ||
 403                  data_size(s) == min_size &&
 404                    (p.offset_in_bytes() < min_iv_offset)))) {
 405           max_ct = cmp_ct.at(j);
 406           max_vw = vw;
 407           max_idx = j;
 408           min_size = data_size(s);
 409           min_iv_offset = p.offset_in_bytes();
 410         }
 411       }
 412     }
 413   }
 414 
 415 #ifdef ASSERT
 416   if (TraceSuperWord && Verbose) {
 417     tty->print_cr("\nVector memops after find_align_to_refs");
 418     for (uint i = 0; i < memops.size(); i++) {
 419       MemNode* s = memops.at(i)->as_Mem();
 420       s->dump();
 421     }
 422   }
 423 #endif
 424 
 425   if (max_ct > 0) {
 426 #ifdef ASSERT
 427     if (TraceSuperWord) {
 428       tty->print("\nVector align to node: ");
 429       memops.at(max_idx)->as_Mem()->dump();
 430     }
 431 #endif
 432     return memops.at(max_idx)->as_Mem();
 433   }
 434   return NULL;
 435 }
 436 
 437 //------------------------------ref_is_alignable---------------------------
 438 // Can the preloop align the reference to position zero in the vector?
 439 bool SuperWord::ref_is_alignable(SWPointer& p) {
 440   if (!p.has_iv()) {
 441     return true;   // no induction variable
 442   }
 443   CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
 444   assert(pre_end->stride_is_con(), "pre loop stride is constant");
 445   int preloop_stride = pre_end->stride_con();
 446 
 447   int span = preloop_stride * p.scale_in_bytes();
 448 
 449   // Stride one accesses are alignable.
 450   if (ABS(span) == p.memory_size())
 451     return true;
 452 
 453   // If initial offset from start of object is computable,
 454   // compute alignment within the vector.
 455   int vw = vector_width_in_bytes(p.mem());
 456   assert(vw > 1, "sanity");
 457   if (vw % span == 0) {
 458     Node* init_nd = pre_end->init_trip();
 459     if (init_nd->is_Con() && p.invar() == NULL) {
 460       int init = init_nd->bottom_type()->is_int()->get_con();
 461 
 462       int init_offset = init * p.scale_in_bytes() + p.offset_in_bytes();
 463       assert(init_offset >= 0, "positive offset from object start");
 464 
 465       if (span > 0) {
 466         return (vw - (init_offset % vw)) % span == 0;
 467       } else {
 468         assert(span < 0, "nonzero stride * scale");
 469         return (init_offset % vw) % -span == 0;
 470       }
 471     }
 472   }
 473   return false;
 474 }
 475 
 476 //---------------------------get_iv_adjustment---------------------------
 477 // Calculate loop's iv adjustment for this memory ops.
 478 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
 479   SWPointer align_to_ref_p(mem_ref, this);
 480   int offset = align_to_ref_p.offset_in_bytes();
 481   int scale  = align_to_ref_p.scale_in_bytes();
 482   int vw       = vector_width_in_bytes(mem_ref);
 483   assert(vw > 1, "sanity");
 484   int stride_sign   = (scale * iv_stride()) > 0 ? 1 : -1;
 485   // At least one iteration is executed in pre-loop by default. As result
 486   // several iterations are needed to align memory operations in main-loop even
 487   // if offset is 0.
 488   int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
 489   int elt_size = align_to_ref_p.memory_size();
 490   assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
 491          err_msg_res("(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size));
 492   int iv_adjustment = iv_adjustment_in_bytes/elt_size;
 493 
 494 #ifndef PRODUCT
 495   if (TraceSuperWord)
 496     tty->print_cr("\noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d",
 497                   offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
 498 #endif
 499   return iv_adjustment;
 500 }
 501 
 502 //---------------------------dependence_graph---------------------------
 503 // Construct dependency graph.
 504 // Add dependence edges to load/store nodes for memory dependence
 505 //    A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
 506 void SuperWord::dependence_graph() {
 507   // First, assign a dependence node to each memory node
 508   for (int i = 0; i < _block.length(); i++ ) {
 509     Node *n = _block.at(i);
 510     if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
 511       _dg.make_node(n);
 512     }
 513   }
 514 
 515   // For each memory slice, create the dependences
 516   for (int i = 0; i < _mem_slice_head.length(); i++) {
 517     Node* n      = _mem_slice_head.at(i);
 518     Node* n_tail = _mem_slice_tail.at(i);
 519 
 520     // Get slice in predecessor order (last is first)
 521     mem_slice_preds(n_tail, n, _nlist);
 522 
 523     // Make the slice dependent on the root
 524     DepMem* slice = _dg.dep(n);
 525     _dg.make_edge(_dg.root(), slice);
 526 
 527     // Create a sink for the slice
 528     DepMem* slice_sink = _dg.make_node(NULL);
 529     _dg.make_edge(slice_sink, _dg.tail());
 530 
 531     // Now visit each pair of memory ops, creating the edges
 532     for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 533       Node* s1 = _nlist.at(j);
 534 
 535       // If no dependency yet, use slice
 536       if (_dg.dep(s1)->in_cnt() == 0) {
 537         _dg.make_edge(slice, s1);
 538       }
 539       SWPointer p1(s1->as_Mem(), this);
 540       bool sink_dependent = true;
 541       for (int k = j - 1; k >= 0; k--) {
 542         Node* s2 = _nlist.at(k);
 543         if (s1->is_Load() && s2->is_Load())
 544           continue;
 545         SWPointer p2(s2->as_Mem(), this);
 546 
 547         int cmp = p1.cmp(p2);
 548         if (SuperWordRTDepCheck &&
 549             p1.base() != p2.base() && p1.valid() && p2.valid()) {
 550           // Create a runtime check to disambiguate
 551           OrderedPair pp(p1.base(), p2.base());
 552           _disjoint_ptrs.append_if_missing(pp);
 553         } else if (!SWPointer::not_equal(cmp)) {
 554           // Possibly same address
 555           _dg.make_edge(s1, s2);
 556           sink_dependent = false;
 557         }
 558       }
 559       if (sink_dependent) {
 560         _dg.make_edge(s1, slice_sink);
 561       }
 562     }
 563 #ifndef PRODUCT
 564     if (TraceSuperWord) {
 565       tty->print_cr("\nDependence graph for slice: %d", n->_idx);
 566       for (int q = 0; q < _nlist.length(); q++) {
 567         _dg.print(_nlist.at(q));
 568       }
 569       tty->cr();
 570     }
 571 #endif
 572     _nlist.clear();
 573   }
 574 
 575 #ifndef PRODUCT
 576   if (TraceSuperWord) {
 577     tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
 578     for (int r = 0; r < _disjoint_ptrs.length(); r++) {
 579       _disjoint_ptrs.at(r).print();
 580       tty->cr();
 581     }
 582     tty->cr();
 583   }
 584 #endif
 585 }
 586 
 587 //---------------------------mem_slice_preds---------------------------
 588 // Return a memory slice (node list) in predecessor order starting at "start"
 589 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
 590   assert(preds.length() == 0, "start empty");
 591   Node* n = start;
 592   Node* prev = NULL;
 593   while (true) {
 594     assert(in_bb(n), "must be in block");
 595     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 596       Node* out = n->fast_out(i);
 597       if (out->is_Load()) {
 598         if (in_bb(out)) {
 599           preds.push(out);
 600         }
 601       } else {
 602         // FIXME
 603         if (out->is_MergeMem() && !in_bb(out)) {
 604           // Either unrolling is causing a memory edge not to disappear,
 605           // or need to run igvn.optimize() again before SLP
 606         } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
 607           // Ditto.  Not sure what else to check further.
 608         } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
 609           // StoreCM has an input edge used as a precedence edge.
 610           // Maybe an issue when oop stores are vectorized.
 611         } else {
 612           assert(out == prev || prev == NULL, "no branches off of store slice");
 613         }
 614       }
 615     }
 616     if (n == stop) break;
 617     preds.push(n);
 618     prev = n;
 619     assert(n->is_Mem(), err_msg_res("unexpected node %s", n->Name()));
 620     n = n->in(MemNode::Memory);
 621   }
 622 }
 623 
 624 //------------------------------stmts_can_pack---------------------------
 625 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
 626 // s1 aligned at "align"
 627 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
 628 
 629   // Do not use superword for non-primitives
 630   BasicType bt1 = velt_basic_type(s1);
 631   BasicType bt2 = velt_basic_type(s2);
 632   if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
 633     return false;
 634   if (Matcher::max_vector_size(bt1) < 2) {
 635     return false; // No vectors for this type
 636   }
 637 
 638   if (isomorphic(s1, s2)) {
 639     if (independent(s1, s2)) {
 640       if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
 641         if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
 642           int s1_align = alignment(s1);
 643           int s2_align = alignment(s2);
 644           if (s1_align == top_align || s1_align == align) {
 645             if (s2_align == top_align || s2_align == align + data_size(s1)) {
 646               return true;
 647             }
 648           }
 649         }
 650       }
 651     }
 652   }
 653   return false;
 654 }
 655 
 656 //------------------------------exists_at---------------------------
 657 // Does s exist in a pack at position pos?
 658 bool SuperWord::exists_at(Node* s, uint pos) {
 659   for (int i = 0; i < _packset.length(); i++) {
 660     Node_List* p = _packset.at(i);
 661     if (p->at(pos) == s) {
 662       return true;
 663     }
 664   }
 665   return false;
 666 }
 667 
 668 //------------------------------are_adjacent_refs---------------------------
 669 // Is s1 immediately before s2 in memory?
 670 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
 671   if (!s1->is_Mem() || !s2->is_Mem()) return false;
 672   if (!in_bb(s1)    || !in_bb(s2))    return false;
 673 
 674   // Do not use superword for non-primitives
 675   if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
 676       !is_java_primitive(s2->as_Mem()->memory_type())) {
 677     return false;
 678   }
 679 
 680   // FIXME - co_locate_pack fails on Stores in different mem-slices, so
 681   // only pack memops that are in the same alias set until that's fixed.
 682   if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
 683       _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
 684     return false;
 685   SWPointer p1(s1->as_Mem(), this);
 686   SWPointer p2(s2->as_Mem(), this);
 687   if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
 688   int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
 689   return diff == data_size(s1);
 690 }
 691 
 692 //------------------------------isomorphic---------------------------
 693 // Are s1 and s2 similar?
 694 bool SuperWord::isomorphic(Node* s1, Node* s2) {
 695   if (s1->Opcode() != s2->Opcode()) return false;
 696   if (s1->req() != s2->req()) return false;
 697   if (s1->in(0) != s2->in(0)) return false;
 698   if (!same_velt_type(s1, s2)) return false;
 699   return true;
 700 }
 701 
 702 //------------------------------independent---------------------------
 703 // Is there no data path from s1 to s2 or s2 to s1?
 704 bool SuperWord::independent(Node* s1, Node* s2) {
 705   //  assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
 706   int d1 = depth(s1);
 707   int d2 = depth(s2);
 708   if (d1 == d2) return s1 != s2;
 709   Node* deep    = d1 > d2 ? s1 : s2;
 710   Node* shallow = d1 > d2 ? s2 : s1;
 711 
 712   visited_clear();
 713 
 714   return independent_path(shallow, deep);
 715 }
 716 
 717 //------------------------------independent_path------------------------------
 718 // Helper for independent
 719 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
 720   if (dp >= 1000) return false; // stop deep recursion
 721   visited_set(deep);
 722   int shal_depth = depth(shallow);
 723   assert(shal_depth <= depth(deep), "must be");
 724   for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
 725     Node* pred = preds.current();
 726     if (in_bb(pred) && !visited_test(pred)) {
 727       if (shallow == pred) {
 728         return false;
 729       }
 730       if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
 731         return false;
 732       }
 733     }
 734   }
 735   return true;
 736 }
 737 
 738 //------------------------------set_alignment---------------------------
 739 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
 740   set_alignment(s1, align);
 741   if (align == top_align || align == bottom_align) {
 742     set_alignment(s2, align);
 743   } else {
 744     set_alignment(s2, align + data_size(s1));
 745   }
 746 }
 747 
 748 //------------------------------data_size---------------------------
 749 int SuperWord::data_size(Node* s) {
 750   int bsize = type2aelembytes(velt_basic_type(s));
 751   assert(bsize != 0, "valid size");
 752   return bsize;
 753 }
 754 
 755 //------------------------------extend_packlist---------------------------
 756 // Extend packset by following use->def and def->use links from pack members.
 757 void SuperWord::extend_packlist() {
 758   bool changed;
 759   do {
 760     changed = false;
 761     for (int i = 0; i < _packset.length(); i++) {
 762       Node_List* p = _packset.at(i);
 763       changed |= follow_use_defs(p);
 764       changed |= follow_def_uses(p);
 765     }
 766   } while (changed);
 767 
 768 #ifndef PRODUCT
 769   if (TraceSuperWord) {
 770     tty->print_cr("\nAfter extend_packlist");
 771     print_packset();
 772   }
 773 #endif
 774 }
 775 
 776 //------------------------------follow_use_defs---------------------------
 777 // Extend the packset by visiting operand definitions of nodes in pack p
 778 bool SuperWord::follow_use_defs(Node_List* p) {
 779   assert(p->size() == 2, "just checking");
 780   Node* s1 = p->at(0);
 781   Node* s2 = p->at(1);
 782   assert(s1->req() == s2->req(), "just checking");
 783   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
 784 
 785   if (s1->is_Load()) return false;
 786 
 787   int align = alignment(s1);
 788   bool changed = false;
 789   int start = s1->is_Store() ? MemNode::ValueIn   : 1;
 790   int end   = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
 791   for (int j = start; j < end; j++) {
 792     Node* t1 = s1->in(j);
 793     Node* t2 = s2->in(j);
 794     if (!in_bb(t1) || !in_bb(t2))
 795       continue;
 796     if (stmts_can_pack(t1, t2, align)) {
 797       if (est_savings(t1, t2) >= 0) {
 798         Node_List* pair = new Node_List();
 799         pair->push(t1);
 800         pair->push(t2);
 801         _packset.append(pair);
 802         set_alignment(t1, t2, align);
 803         changed = true;
 804       }
 805     }
 806   }
 807   return changed;
 808 }
 809 
 810 //------------------------------follow_def_uses---------------------------
 811 // Extend the packset by visiting uses of nodes in pack p
 812 bool SuperWord::follow_def_uses(Node_List* p) {
 813   bool changed = false;
 814   Node* s1 = p->at(0);
 815   Node* s2 = p->at(1);
 816   assert(p->size() == 2, "just checking");
 817   assert(s1->req() == s2->req(), "just checking");
 818   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
 819 
 820   if (s1->is_Store()) return false;
 821 
 822   int align = alignment(s1);
 823   int savings = -1;
 824   Node* u1 = NULL;
 825   Node* u2 = NULL;
 826   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
 827     Node* t1 = s1->fast_out(i);
 828     if (!in_bb(t1)) continue;
 829     for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
 830       Node* t2 = s2->fast_out(j);
 831       if (!in_bb(t2)) continue;
 832       if (!opnd_positions_match(s1, t1, s2, t2))
 833         continue;
 834       if (stmts_can_pack(t1, t2, align)) {
 835         int my_savings = est_savings(t1, t2);
 836         if (my_savings > savings) {
 837           savings = my_savings;
 838           u1 = t1;
 839           u2 = t2;
 840         }
 841       }
 842     }
 843   }
 844   if (savings >= 0) {
 845     Node_List* pair = new Node_List();
 846     pair->push(u1);
 847     pair->push(u2);
 848     _packset.append(pair);
 849     set_alignment(u1, u2, align);
 850     changed = true;
 851   }
 852   return changed;
 853 }
 854 
 855 //---------------------------opnd_positions_match-------------------------
 856 // Is the use of d1 in u1 at the same operand position as d2 in u2?
 857 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
 858   uint ct = u1->req();
 859   if (ct != u2->req()) return false;
 860   uint i1 = 0;
 861   uint i2 = 0;
 862   do {
 863     for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
 864     for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
 865     if (i1 != i2) {
 866       if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
 867         // Further analysis relies on operands position matching.
 868         u2->swap_edges(i1, i2);
 869       } else {
 870         return false;
 871       }
 872     }
 873   } while (i1 < ct);
 874   return true;
 875 }
 876 
 877 //------------------------------est_savings---------------------------
 878 // Estimate the savings from executing s1 and s2 as a pack
 879 int SuperWord::est_savings(Node* s1, Node* s2) {
 880   int save_in = 2 - 1; // 2 operations per instruction in packed form
 881 
 882   // inputs
 883   for (uint i = 1; i < s1->req(); i++) {
 884     Node* x1 = s1->in(i);
 885     Node* x2 = s2->in(i);
 886     if (x1 != x2) {
 887       if (are_adjacent_refs(x1, x2)) {
 888         save_in += adjacent_profit(x1, x2);
 889       } else if (!in_packset(x1, x2)) {
 890         save_in -= pack_cost(2);
 891       } else {
 892         save_in += unpack_cost(2);
 893       }
 894     }
 895   }
 896 
 897   // uses of result
 898   uint ct = 0;
 899   int save_use = 0;
 900   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
 901     Node* s1_use = s1->fast_out(i);
 902     for (int j = 0; j < _packset.length(); j++) {
 903       Node_List* p = _packset.at(j);
 904       if (p->at(0) == s1_use) {
 905         for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
 906           Node* s2_use = s2->fast_out(k);
 907           if (p->at(p->size()-1) == s2_use) {
 908             ct++;
 909             if (are_adjacent_refs(s1_use, s2_use)) {
 910               save_use += adjacent_profit(s1_use, s2_use);
 911             }
 912           }
 913         }
 914       }
 915     }
 916   }
 917 
 918   if (ct < s1->outcnt()) save_use += unpack_cost(1);
 919   if (ct < s2->outcnt()) save_use += unpack_cost(1);
 920 
 921   return MAX2(save_in, save_use);
 922 }
 923 
 924 //------------------------------costs---------------------------
 925 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
 926 int SuperWord::pack_cost(int ct)   { return ct; }
 927 int SuperWord::unpack_cost(int ct) { return ct; }
 928 
 929 //------------------------------combine_packs---------------------------
 930 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
 931 void SuperWord::combine_packs() {
 932   bool changed = true;
 933   // Combine packs regardless max vector size.
 934   while (changed) {
 935     changed = false;
 936     for (int i = 0; i < _packset.length(); i++) {
 937       Node_List* p1 = _packset.at(i);
 938       if (p1 == NULL) continue;
 939       for (int j = 0; j < _packset.length(); j++) {
 940         Node_List* p2 = _packset.at(j);
 941         if (p2 == NULL) continue;
 942         if (i == j) continue;
 943         if (p1->at(p1->size()-1) == p2->at(0)) {
 944           for (uint k = 1; k < p2->size(); k++) {
 945             p1->push(p2->at(k));
 946           }
 947           _packset.at_put(j, NULL);
 948           changed = true;
 949         }
 950       }
 951     }
 952   }
 953 
 954   // Split packs which have size greater then max vector size.
 955   for (int i = 0; i < _packset.length(); i++) {
 956     Node_List* p1 = _packset.at(i);
 957     if (p1 != NULL) {
 958       BasicType bt = velt_basic_type(p1->at(0));
 959       uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
 960       assert(is_power_of_2(max_vlen), "sanity");
 961       uint psize = p1->size();
 962       if (!is_power_of_2(psize)) {
 963         // Skip pack which can't be vector.
 964         // case1: for(...) { a[i] = i; }    elements values are different (i+x)
 965         // case2: for(...) { a[i] = b[i+1]; }  can't align both, load and store
 966         _packset.at_put(i, NULL);
 967         continue;
 968       }
 969       if (psize > max_vlen) {
 970         Node_List* pack = new Node_List();
 971         for (uint j = 0; j < psize; j++) {
 972           pack->push(p1->at(j));
 973           if (pack->size() >= max_vlen) {
 974             assert(is_power_of_2(pack->size()), "sanity");
 975             _packset.append(pack);
 976             pack = new Node_List();
 977           }
 978         }
 979         _packset.at_put(i, NULL);
 980       }
 981     }
 982   }
 983 
 984   // Compress list.
 985   for (int i = _packset.length() - 1; i >= 0; i--) {
 986     Node_List* p1 = _packset.at(i);
 987     if (p1 == NULL) {
 988       _packset.remove_at(i);
 989     }
 990   }
 991 
 992 #ifndef PRODUCT
 993   if (TraceSuperWord) {
 994     tty->print_cr("\nAfter combine_packs");
 995     print_packset();
 996   }
 997 #endif
 998 }
 999 
1000 //-----------------------------construct_my_pack_map--------------------------
1001 // Construct the map from nodes to packs.  Only valid after the
1002 // point where a node is only in one pack (after combine_packs).
1003 void SuperWord::construct_my_pack_map() {
1004   Node_List* rslt = NULL;
1005   for (int i = 0; i < _packset.length(); i++) {
1006     Node_List* p = _packset.at(i);
1007     for (uint j = 0; j < p->size(); j++) {
1008       Node* s = p->at(j);
1009       assert(my_pack(s) == NULL, "only in one pack");
1010       set_my_pack(s, p);
1011     }
1012   }
1013 }
1014 
1015 //------------------------------filter_packs---------------------------
1016 // Remove packs that are not implemented or not profitable.
1017 void SuperWord::filter_packs() {
1018 
1019   // Remove packs that are not implemented
1020   for (int i = _packset.length() - 1; i >= 0; i--) {
1021     Node_List* pk = _packset.at(i);
1022     bool impl = implemented(pk);
1023     if (!impl) {
1024 #ifndef PRODUCT
1025       if (TraceSuperWord && Verbose) {
1026         tty->print_cr("Unimplemented");
1027         pk->at(0)->dump();
1028       }
1029 #endif
1030       remove_pack_at(i);
1031     }
1032   }
1033 
1034   // Remove packs that are not profitable
1035   bool changed;
1036   do {
1037     changed = false;
1038     for (int i = _packset.length() - 1; i >= 0; i--) {
1039       Node_List* pk = _packset.at(i);
1040       bool prof = profitable(pk);
1041       if (!prof) {
1042 #ifndef PRODUCT
1043         if (TraceSuperWord && Verbose) {
1044           tty->print_cr("Unprofitable");
1045           pk->at(0)->dump();
1046         }
1047 #endif
1048         remove_pack_at(i);
1049         changed = true;
1050       }
1051     }
1052   } while (changed);
1053 
1054 #ifndef PRODUCT
1055   if (TraceSuperWord) {
1056     tty->print_cr("\nAfter filter_packs");
1057     print_packset();
1058     tty->cr();
1059   }
1060 #endif
1061 }
1062 
1063 //------------------------------implemented---------------------------
1064 // Can code be generated for pack p?
1065 bool SuperWord::implemented(Node_List* p) {
1066   Node* p0 = p->at(0);
1067   return VectorNode::implemented(p0->Opcode(), p->size(), velt_basic_type(p0));
1068 }
1069 
1070 //------------------------------same_inputs--------------------------
1071 // For pack p, are all idx operands the same?
1072 static bool same_inputs(Node_List* p, int idx) {
1073   Node* p0 = p->at(0);
1074   uint vlen = p->size();
1075   Node* p0_def = p0->in(idx);
1076   for (uint i = 1; i < vlen; i++) {
1077     Node* pi = p->at(i);
1078     Node* pi_def = pi->in(idx);
1079     if (p0_def != pi_def)
1080       return false;
1081   }
1082   return true;
1083 }
1084 
1085 //------------------------------profitable---------------------------
1086 // For pack p, are all operands and all uses (with in the block) vector?
1087 bool SuperWord::profitable(Node_List* p) {
1088   Node* p0 = p->at(0);
1089   uint start, end;
1090   VectorNode::vector_operands(p0, &start, &end);
1091 
1092   // Return false if some inputs are not vectors or vectors with different
1093   // size or alignment.
1094   // Also, for now, return false if not scalar promotion case when inputs are
1095   // the same. Later, implement PackNode and allow differing, non-vector inputs
1096   // (maybe just the ones from outside the block.)
1097   for (uint i = start; i < end; i++) {
1098     if (!is_vector_use(p0, i))
1099       return false;
1100   }
1101   if (VectorNode::is_shift(p0)) {
1102     // For now, return false if shift count is vector or not scalar promotion
1103     // case (different shift counts) because it is not supported yet.
1104     Node* cnt = p0->in(2);
1105     Node_List* cnt_pk = my_pack(cnt);
1106     if (cnt_pk != NULL)
1107       return false;
1108     if (!same_inputs(p, 2))
1109       return false;
1110   }
1111   if (!p0->is_Store()) {
1112     // For now, return false if not all uses are vector.
1113     // Later, implement ExtractNode and allow non-vector uses (maybe
1114     // just the ones outside the block.)
1115     for (uint i = 0; i < p->size(); i++) {
1116       Node* def = p->at(i);
1117       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1118         Node* use = def->fast_out(j);
1119         for (uint k = 0; k < use->req(); k++) {
1120           Node* n = use->in(k);
1121           if (def == n) {
1122             if (!is_vector_use(use, k)) {
1123               return false;
1124             }
1125           }
1126         }
1127       }
1128     }
1129   }
1130   return true;
1131 }
1132 
1133 //------------------------------schedule---------------------------
1134 // Adjust the memory graph for the packed operations
1135 void SuperWord::schedule() {
1136 
1137   // Co-locate in the memory graph the members of each memory pack
1138   for (int i = 0; i < _packset.length(); i++) {
1139     co_locate_pack(_packset.at(i));
1140   }
1141 }
1142 
1143 //-------------------------------remove_and_insert-------------------
1144 // Remove "current" from its current position in the memory graph and insert
1145 // it after the appropriate insertion point (lip or uip).
1146 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1147                                   Node *uip, Unique_Node_List &sched_before) {
1148   Node* my_mem = current->in(MemNode::Memory);
1149   bool sched_up = sched_before.member(current);
1150 
1151   // remove current_store from its current position in the memmory graph
1152   for (DUIterator i = current->outs(); current->has_out(i); i++) {
1153     Node* use = current->out(i);
1154     if (use->is_Mem()) {
1155       assert(use->in(MemNode::Memory) == current, "must be");
1156       if (use == prev) { // connect prev to my_mem
1157           _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1158           --i; //deleted this edge; rescan position
1159       } else if (sched_before.member(use)) {
1160         if (!sched_up) { // Will be moved together with current
1161           _igvn.replace_input_of(use, MemNode::Memory, uip);
1162           --i; //deleted this edge; rescan position
1163         }
1164       } else {
1165         if (sched_up) { // Will be moved together with current
1166           _igvn.replace_input_of(use, MemNode::Memory, lip);
1167           --i; //deleted this edge; rescan position
1168         }
1169       }
1170     }
1171   }
1172 
1173   Node *insert_pt =  sched_up ?  uip : lip;
1174 
1175   // all uses of insert_pt's memory state should use current's instead
1176   for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1177     Node* use = insert_pt->out(i);
1178     if (use->is_Mem()) {
1179       assert(use->in(MemNode::Memory) == insert_pt, "must be");
1180       _igvn.replace_input_of(use, MemNode::Memory, current);
1181       --i; //deleted this edge; rescan position
1182     } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1183       uint pos; //lip (lower insert point) must be the last one in the memory slice
1184       for (pos=1; pos < use->req(); pos++) {
1185         if (use->in(pos) == insert_pt) break;
1186       }
1187       _igvn.replace_input_of(use, pos, current);
1188       --i;
1189     }
1190   }
1191 
1192   //connect current to insert_pt
1193   _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
1194 }
1195 
1196 //------------------------------co_locate_pack----------------------------------
1197 // To schedule a store pack, we need to move any sandwiched memory ops either before
1198 // or after the pack, based upon dependence information:
1199 // (1) If any store in the pack depends on the sandwiched memory op, the
1200 //     sandwiched memory op must be scheduled BEFORE the pack;
1201 // (2) If a sandwiched memory op depends on any store in the pack, the
1202 //     sandwiched memory op must be scheduled AFTER the pack;
1203 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1204 //     memory op (say memB), memB must be scheduled before memA. So, if memA is
1205 //     scheduled before the pack, memB must also be scheduled before the pack;
1206 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1207 //     schedule this store AFTER the pack
1208 // (5) We know there is no dependence cycle, so there in no other case;
1209 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1210 //
1211 // To schedule a load pack, we use the memory state of either the first or the last load in
1212 // the pack, based on the dependence constraint.
1213 void SuperWord::co_locate_pack(Node_List* pk) {
1214   if (pk->at(0)->is_Store()) {
1215     MemNode* first     = executed_first(pk)->as_Mem();
1216     MemNode* last      = executed_last(pk)->as_Mem();
1217     Unique_Node_List schedule_before_pack;
1218     Unique_Node_List memops;
1219 
1220     MemNode* current   = last->in(MemNode::Memory)->as_Mem();
1221     MemNode* previous  = last;
1222     while (true) {
1223       assert(in_bb(current), "stay in block");
1224       memops.push(previous);
1225       for (DUIterator i = current->outs(); current->has_out(i); i++) {
1226         Node* use = current->out(i);
1227         if (use->is_Mem() && use != previous)
1228           memops.push(use);
1229       }
1230       if (current == first) break;
1231       previous = current;
1232       current  = current->in(MemNode::Memory)->as_Mem();
1233     }
1234 
1235     // determine which memory operations should be scheduled before the pack
1236     for (uint i = 1; i < memops.size(); i++) {
1237       Node *s1 = memops.at(i);
1238       if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1239         for (uint j = 0; j< i; j++) {
1240           Node *s2 = memops.at(j);
1241           if (!independent(s1, s2)) {
1242             if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1243               schedule_before_pack.push(s1); // s1 must be scheduled before
1244               Node_List* mem_pk = my_pack(s1);
1245               if (mem_pk != NULL) {
1246                 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1247                   Node* s = mem_pk->at(ii);  // follow partner
1248                   if (memops.member(s) && !schedule_before_pack.member(s))
1249                     schedule_before_pack.push(s);
1250                 }
1251               }
1252               break;
1253             }
1254           }
1255         }
1256       }
1257     }
1258 
1259     Node*    upper_insert_pt = first->in(MemNode::Memory);
1260     // Following code moves loads connected to upper_insert_pt below aliased stores.
1261     // Collect such loads here and reconnect them back to upper_insert_pt later.
1262     memops.clear();
1263     for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
1264       Node* use = upper_insert_pt->out(i);
1265       if (use->is_Mem() && !use->is_Store()) {
1266         memops.push(use);
1267       }
1268     }
1269 
1270     MemNode* lower_insert_pt = last;
1271     previous                 = last; //previous store in pk
1272     current                  = last->in(MemNode::Memory)->as_Mem();
1273 
1274     // start scheduling from "last" to "first"
1275     while (true) {
1276       assert(in_bb(current), "stay in block");
1277       assert(in_pack(previous, pk), "previous stays in pack");
1278       Node* my_mem = current->in(MemNode::Memory);
1279 
1280       if (in_pack(current, pk)) {
1281         // Forward users of my memory state (except "previous) to my input memory state
1282         for (DUIterator i = current->outs(); current->has_out(i); i++) {
1283           Node* use = current->out(i);
1284           if (use->is_Mem() && use != previous) {
1285             assert(use->in(MemNode::Memory) == current, "must be");
1286             if (schedule_before_pack.member(use)) {
1287               _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
1288             } else {
1289               _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
1290             }
1291             --i; // deleted this edge; rescan position
1292           }
1293         }
1294         previous = current;
1295       } else { // !in_pack(current, pk) ==> a sandwiched store
1296         remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1297       }
1298 
1299       if (current == first) break;
1300       current = my_mem->as_Mem();
1301     } // end while
1302 
1303     // Reconnect loads back to upper_insert_pt.
1304     for (uint i = 0; i < memops.size(); i++) {
1305       Node *ld = memops.at(i);
1306       if (ld->in(MemNode::Memory) != upper_insert_pt) {
1307         _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
1308       }
1309     }
1310   } else if (pk->at(0)->is_Load()) { //load
1311     // all loads in the pack should have the same memory state. By default,
1312     // we use the memory state of the last load. However, if any load could
1313     // not be moved down due to the dependence constraint, we use the memory
1314     // state of the first load.
1315     Node* last_mem  = executed_last(pk)->in(MemNode::Memory);
1316     Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1317     bool schedule_last = true;
1318     for (uint i = 0; i < pk->size(); i++) {
1319       Node* ld = pk->at(i);
1320       for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1321            current=current->in(MemNode::Memory)) {
1322         assert(current != first_mem, "corrupted memory graph");
1323         if(current->is_Mem() && !independent(current, ld)){
1324           schedule_last = false; // a later store depends on this load
1325           break;
1326         }
1327       }
1328     }
1329 
1330     Node* mem_input = schedule_last ? last_mem : first_mem;
1331     _igvn.hash_delete(mem_input);
1332     // Give each load the same memory state
1333     for (uint i = 0; i < pk->size(); i++) {
1334       LoadNode* ld = pk->at(i)->as_Load();
1335       _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
1336     }
1337   }
1338 }
1339 
1340 //------------------------------output---------------------------
1341 // Convert packs into vector node operations
1342 void SuperWord::output() {
1343   if (_packset.length() == 0) return;
1344 
1345 #ifndef PRODUCT
1346   if (TraceLoopOpts) {
1347     tty->print("SuperWord    ");
1348     lpt()->dump_head();
1349   }
1350 #endif
1351 
1352   // MUST ENSURE main loop's initial value is properly aligned:
1353   //  (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1354 
1355   align_initial_loop_index(align_to_ref());
1356 
1357   // Insert extract (unpack) operations for scalar uses
1358   for (int i = 0; i < _packset.length(); i++) {
1359     insert_extracts(_packset.at(i));
1360   }
1361 
1362   Compile* C = _phase->C;
1363   uint max_vlen_in_bytes = 0;
1364   for (int i = 0; i < _block.length(); i++) {
1365     Node* n = _block.at(i);
1366     Node_List* p = my_pack(n);
1367     if (p && n == executed_last(p)) {
1368       uint vlen = p->size();
1369       uint vlen_in_bytes = 0;
1370       Node* vn = NULL;
1371       Node* low_adr = p->at(0);
1372       Node* first   = executed_first(p);
1373       int   opc = n->Opcode();
1374       if (n->is_Load()) {
1375         Node* ctl = n->in(MemNode::Control);
1376         Node* mem = first->in(MemNode::Memory);
1377         SWPointer p1(n->as_Mem(), this);
1378         // Identify the memory dependency for the new loadVector node by
1379         // walking up through memory chain.
1380         // This is done to give flexibility to the new loadVector node so that
1381         // it can move above independent storeVector nodes.
1382         while (mem->is_StoreVector()) {
1383           SWPointer p2(mem->as_Mem(), this);
1384           int cmp = p1.cmp(p2);
1385           if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
1386             mem = mem->in(MemNode::Memory);
1387           } else {
1388             break; // dependent memory
1389           }
1390         }
1391         Node* adr = low_adr->in(MemNode::Address);
1392         const TypePtr* atyp = n->adr_type();
1393         vn = LoadVectorNode::make(C, opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n));
1394         vlen_in_bytes = vn->as_LoadVector()->memory_size();
1395       } else if (n->is_Store()) {
1396         // Promote value to be stored to vector
1397         Node* val = vector_opd(p, MemNode::ValueIn);
1398         Node* ctl = n->in(MemNode::Control);
1399         Node* mem = first->in(MemNode::Memory);
1400         Node* adr = low_adr->in(MemNode::Address);
1401         const TypePtr* atyp = n->adr_type();
1402         vn = StoreVectorNode::make(C, opc, ctl, mem, adr, atyp, val, vlen);
1403         vlen_in_bytes = vn->as_StoreVector()->memory_size();
1404       } else if (n->req() == 3) {
1405         // Promote operands to vector
1406         Node* in1 = vector_opd(p, 1);
1407         Node* in2 = vector_opd(p, 2);
1408         if (VectorNode::is_invariant_vector(in1) && (n->is_Add() || n->is_Mul())) {
1409           // Move invariant vector input into second position to avoid register spilling.
1410           Node* tmp = in1;
1411           in1 = in2;
1412           in2 = tmp;
1413         }
1414         vn = VectorNode::make(C, opc, in1, in2, vlen, velt_basic_type(n));
1415         vlen_in_bytes = vn->as_Vector()->length_in_bytes();
1416       } else {
1417         ShouldNotReachHere();
1418       }
1419       assert(vn != NULL, "sanity");
1420       _igvn.register_new_node_with_optimizer(vn);
1421       _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1422       for (uint j = 0; j < p->size(); j++) {
1423         Node* pm = p->at(j);
1424         _igvn.replace_node(pm, vn);
1425       }
1426       _igvn._worklist.push(vn);
1427 
1428       if (vlen_in_bytes > max_vlen_in_bytes) {
1429         max_vlen_in_bytes = vlen_in_bytes;
1430       }
1431 #ifdef ASSERT
1432       if (TraceNewVectors) {
1433         tty->print("new Vector node: ");
1434         vn->dump();
1435       }
1436 #endif
1437     }
1438   }
1439   C->set_max_vector_size(max_vlen_in_bytes);
1440 }
1441 
1442 //------------------------------vector_opd---------------------------
1443 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1444 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1445   Node* p0 = p->at(0);
1446   uint vlen = p->size();
1447   Node* opd = p0->in(opd_idx);
1448 
1449   if (same_inputs(p, opd_idx)) {
1450     if (opd->is_Vector() || opd->is_LoadVector()) {
1451       assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
1452       return opd; // input is matching vector
1453     }
1454     if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
1455       Compile* C = _phase->C;
1456       Node* cnt = opd;
1457       // Vector instructions do not mask shift count, do it here.
1458       juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
1459       const TypeInt* t = opd->find_int_type();
1460       if (t != NULL && t->is_con()) {
1461         juint shift = t->get_con();
1462         if (shift > mask) { // Unsigned cmp
1463           cnt = ConNode::make(C, TypeInt::make(shift & mask));
1464         }
1465       } else {
1466         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
1467           cnt = ConNode::make(C, TypeInt::make(mask));
1468           _igvn.register_new_node_with_optimizer(cnt);
1469           cnt = new (C) AndINode(opd, cnt);
1470           _igvn.register_new_node_with_optimizer(cnt);
1471           _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1472         }
1473         assert(opd->bottom_type()->isa_int(), "int type only");
1474         // Move non constant shift count into vector register.
1475         cnt = VectorNode::shift_count(C, p0, cnt, vlen, velt_basic_type(p0));
1476       }
1477       if (cnt != opd) {
1478         _igvn.register_new_node_with_optimizer(cnt);
1479         _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1480       }
1481       return cnt;
1482     }
1483     assert(!opd->is_StoreVector(), "such vector is not expected here");
1484     // Convert scalar input to vector with the same number of elements as
1485     // p0's vector. Use p0's type because size of operand's container in
1486     // vector should match p0's size regardless operand's size.
1487     const Type* p0_t = velt_type(p0);
1488     VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, p0_t);
1489 
1490     _igvn.register_new_node_with_optimizer(vn);
1491     _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1492 #ifdef ASSERT
1493     if (TraceNewVectors) {
1494       tty->print("new Vector node: ");
1495       vn->dump();
1496     }
1497 #endif
1498     return vn;
1499   }
1500 
1501   // Insert pack operation
1502   BasicType bt = velt_basic_type(p0);
1503   PackNode* pk = PackNode::make(_phase->C, opd, vlen, bt);
1504   DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
1505 
1506   for (uint i = 1; i < vlen; i++) {
1507     Node* pi = p->at(i);
1508     Node* in = pi->in(opd_idx);
1509     assert(my_pack(in) == NULL, "Should already have been unpacked");
1510     assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
1511     pk->add_opd(in);
1512   }
1513   _igvn.register_new_node_with_optimizer(pk);
1514   _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1515 #ifdef ASSERT
1516   if (TraceNewVectors) {
1517     tty->print("new Vector node: ");
1518     pk->dump();
1519   }
1520 #endif
1521   return pk;
1522 }
1523 
1524 //------------------------------insert_extracts---------------------------
1525 // If a use of pack p is not a vector use, then replace the
1526 // use with an extract operation.
1527 void SuperWord::insert_extracts(Node_List* p) {
1528   if (p->at(0)->is_Store()) return;
1529   assert(_n_idx_list.is_empty(), "empty (node,index) list");
1530 
1531   // Inspect each use of each pack member.  For each use that is
1532   // not a vector use, replace the use with an extract operation.
1533 
1534   for (uint i = 0; i < p->size(); i++) {
1535     Node* def = p->at(i);
1536     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1537       Node* use = def->fast_out(j);
1538       for (uint k = 0; k < use->req(); k++) {
1539         Node* n = use->in(k);
1540         if (def == n) {
1541           if (!is_vector_use(use, k)) {
1542             _n_idx_list.push(use, k);
1543           }
1544         }
1545       }
1546     }
1547   }
1548 
1549   while (_n_idx_list.is_nonempty()) {
1550     Node* use = _n_idx_list.node();
1551     int   idx = _n_idx_list.index();
1552     _n_idx_list.pop();
1553     Node* def = use->in(idx);
1554 
1555     // Insert extract operation
1556     _igvn.hash_delete(def);
1557     int def_pos = alignment(def) / data_size(def);
1558 
1559     Node* ex = ExtractNode::make(_phase->C, def, def_pos, velt_basic_type(def));
1560     _igvn.register_new_node_with_optimizer(ex);
1561     _phase->set_ctrl(ex, _phase->get_ctrl(def));
1562     _igvn.replace_input_of(use, idx, ex);
1563     _igvn._worklist.push(def);
1564 
1565     bb_insert_after(ex, bb_idx(def));
1566     set_velt_type(ex, velt_type(def));
1567   }
1568 }
1569 
1570 //------------------------------is_vector_use---------------------------
1571 // Is use->in(u_idx) a vector use?
1572 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1573   Node_List* u_pk = my_pack(use);
1574   if (u_pk == NULL) return false;
1575   Node* def = use->in(u_idx);
1576   Node_List* d_pk = my_pack(def);
1577   if (d_pk == NULL) {
1578     // check for scalar promotion
1579     Node* n = u_pk->at(0)->in(u_idx);
1580     for (uint i = 1; i < u_pk->size(); i++) {
1581       if (u_pk->at(i)->in(u_idx) != n) return false;
1582     }
1583     return true;
1584   }
1585   if (u_pk->size() != d_pk->size())
1586     return false;
1587   for (uint i = 0; i < u_pk->size(); i++) {
1588     Node* ui = u_pk->at(i);
1589     Node* di = d_pk->at(i);
1590     if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1591       return false;
1592   }
1593   return true;
1594 }
1595 
1596 //------------------------------construct_bb---------------------------
1597 // Construct reverse postorder list of block members
1598 bool SuperWord::construct_bb() {
1599   Node* entry = bb();
1600 
1601   assert(_stk.length() == 0,            "stk is empty");
1602   assert(_block.length() == 0,          "block is empty");
1603   assert(_data_entry.length() == 0,     "data_entry is empty");
1604   assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1605   assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1606 
1607   // Find non-control nodes with no inputs from within block,
1608   // create a temporary map from node _idx to bb_idx for use
1609   // by the visited and post_visited sets,
1610   // and count number of nodes in block.
1611   int bb_ct = 0;
1612   for (uint i = 0; i < lpt()->_body.size(); i++ ) {
1613     Node *n = lpt()->_body.at(i);
1614     set_bb_idx(n, i); // Create a temporary map
1615     if (in_bb(n)) {
1616       if (n->is_LoadStore() || n->is_MergeMem() ||
1617           (n->is_Proj() && !n->as_Proj()->is_CFG())) {
1618         // Bailout if the loop has LoadStore, MergeMem or data Proj
1619         // nodes. Superword optimization does not work with them.
1620         return false;
1621       }
1622       bb_ct++;
1623       if (!n->is_CFG()) {
1624         bool found = false;
1625         for (uint j = 0; j < n->req(); j++) {
1626           Node* def = n->in(j);
1627           if (def && in_bb(def)) {
1628             found = true;
1629             break;
1630           }
1631         }
1632         if (!found) {
1633           assert(n != entry, "can't be entry");
1634           _data_entry.push(n);
1635         }
1636       }
1637     }
1638   }
1639 
1640   // Find memory slices (head and tail)
1641   for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1642     Node *n = lp()->fast_out(i);
1643     if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1644       Node* n_tail  = n->in(LoopNode::LoopBackControl);
1645       if (n_tail != n->in(LoopNode::EntryControl)) {
1646         if (!n_tail->is_Mem()) {
1647           assert(n_tail->is_Mem(), err_msg_res("unexpected node for memory slice: %s", n_tail->Name()));
1648           return false; // Bailout
1649         }
1650         _mem_slice_head.push(n);
1651         _mem_slice_tail.push(n_tail);
1652       }
1653     }
1654   }
1655 
1656   // Create an RPO list of nodes in block
1657 
1658   visited_clear();
1659   post_visited_clear();
1660 
1661   // Push all non-control nodes with no inputs from within block, then control entry
1662   for (int j = 0; j < _data_entry.length(); j++) {
1663     Node* n = _data_entry.at(j);
1664     visited_set(n);
1665     _stk.push(n);
1666   }
1667   visited_set(entry);
1668   _stk.push(entry);
1669 
1670   // Do a depth first walk over out edges
1671   int rpo_idx = bb_ct - 1;
1672   int size;
1673   while ((size = _stk.length()) > 0) {
1674     Node* n = _stk.top(); // Leave node on stack
1675     if (!visited_test_set(n)) {
1676       // forward arc in graph
1677     } else if (!post_visited_test(n)) {
1678       // cross or back arc
1679       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1680         Node *use = n->fast_out(i);
1681         if (in_bb(use) && !visited_test(use) &&
1682             // Don't go around backedge
1683             (!use->is_Phi() || n == entry)) {
1684           _stk.push(use);
1685         }
1686       }
1687       if (_stk.length() == size) {
1688         // There were no additional uses, post visit node now
1689         _stk.pop(); // Remove node from stack
1690         assert(rpo_idx >= 0, "");
1691         _block.at_put_grow(rpo_idx, n);
1692         rpo_idx--;
1693         post_visited_set(n);
1694         assert(rpo_idx >= 0 || _stk.is_empty(), "");
1695       }
1696     } else {
1697       _stk.pop(); // Remove post-visited node from stack
1698     }
1699   }
1700 
1701   // Create real map of block indices for nodes
1702   for (int j = 0; j < _block.length(); j++) {
1703     Node* n = _block.at(j);
1704     set_bb_idx(n, j);
1705   }
1706 
1707   initialize_bb(); // Ensure extra info is allocated.
1708 
1709 #ifndef PRODUCT
1710   if (TraceSuperWord) {
1711     print_bb();
1712     tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1713     for (int m = 0; m < _data_entry.length(); m++) {
1714       tty->print("%3d ", m);
1715       _data_entry.at(m)->dump();
1716     }
1717     tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1718     for (int m = 0; m < _mem_slice_head.length(); m++) {
1719       tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1720       tty->print("    ");    _mem_slice_tail.at(m)->dump();
1721     }
1722   }
1723 #endif
1724   assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1725   return (_mem_slice_head.length() > 0) || (_data_entry.length() > 0);
1726 }
1727 
1728 //------------------------------initialize_bb---------------------------
1729 // Initialize per node info
1730 void SuperWord::initialize_bb() {
1731   Node* last = _block.at(_block.length() - 1);
1732   grow_node_info(bb_idx(last));
1733 }
1734 
1735 //------------------------------bb_insert_after---------------------------
1736 // Insert n into block after pos
1737 void SuperWord::bb_insert_after(Node* n, int pos) {
1738   int n_pos = pos + 1;
1739   // Make room
1740   for (int i = _block.length() - 1; i >= n_pos; i--) {
1741     _block.at_put_grow(i+1, _block.at(i));
1742   }
1743   for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1744     _node_info.at_put_grow(j+1, _node_info.at(j));
1745   }
1746   // Set value
1747   _block.at_put_grow(n_pos, n);
1748   _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1749   // Adjust map from node->_idx to _block index
1750   for (int i = n_pos; i < _block.length(); i++) {
1751     set_bb_idx(_block.at(i), i);
1752   }
1753 }
1754 
1755 //------------------------------compute_max_depth---------------------------
1756 // Compute max depth for expressions from beginning of block
1757 // Use to prune search paths during test for independence.
1758 void SuperWord::compute_max_depth() {
1759   int ct = 0;
1760   bool again;
1761   do {
1762     again = false;
1763     for (int i = 0; i < _block.length(); i++) {
1764       Node* n = _block.at(i);
1765       if (!n->is_Phi()) {
1766         int d_orig = depth(n);
1767         int d_in   = 0;
1768         for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1769           Node* pred = preds.current();
1770           if (in_bb(pred)) {
1771             d_in = MAX2(d_in, depth(pred));
1772           }
1773         }
1774         if (d_in + 1 != d_orig) {
1775           set_depth(n, d_in + 1);
1776           again = true;
1777         }
1778       }
1779     }
1780     ct++;
1781   } while (again);
1782 #ifndef PRODUCT
1783   if (TraceSuperWord && Verbose)
1784     tty->print_cr("compute_max_depth iterated: %d times", ct);
1785 #endif
1786 }
1787 
1788 //-------------------------compute_vector_element_type-----------------------
1789 // Compute necessary vector element type for expressions
1790 // This propagates backwards a narrower integer type when the
1791 // upper bits of the value are not needed.
1792 // Example:  char a,b,c;  a = b + c;
1793 // Normally the type of the add is integer, but for packed character
1794 // operations the type of the add needs to be char.
1795 void SuperWord::compute_vector_element_type() {
1796 #ifndef PRODUCT
1797   if (TraceSuperWord && Verbose)
1798     tty->print_cr("\ncompute_velt_type:");
1799 #endif
1800 
1801   // Initial type
1802   for (int i = 0; i < _block.length(); i++) {
1803     Node* n = _block.at(i);
1804     set_velt_type(n, container_type(n));
1805   }
1806 
1807   // Propagate integer narrowed type backwards through operations
1808   // that don't depend on higher order bits
1809   for (int i = _block.length() - 1; i >= 0; i--) {
1810     Node* n = _block.at(i);
1811     // Only integer types need be examined
1812     const Type* vtn = velt_type(n);
1813     if (vtn->basic_type() == T_INT) {
1814       uint start, end;
1815       VectorNode::vector_operands(n, &start, &end);
1816 
1817       for (uint j = start; j < end; j++) {
1818         Node* in  = n->in(j);
1819         // Don't propagate through a memory
1820         if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
1821             data_size(n) < data_size(in)) {
1822           bool same_type = true;
1823           for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1824             Node *use = in->fast_out(k);
1825             if (!in_bb(use) || !same_velt_type(use, n)) {
1826               same_type = false;
1827               break;
1828             }
1829           }
1830           if (same_type) {
1831             // For right shifts of small integer types (bool, byte, char, short)
1832             // we need precise information about sign-ness. Only Load nodes have
1833             // this information because Store nodes are the same for signed and
1834             // unsigned values. And any arithmetic operation after a load may
1835             // expand a value to signed Int so such right shifts can't be used
1836             // because vector elements do not have upper bits of Int.
1837             const Type* vt = vtn;
1838             if (VectorNode::is_shift(in)) {
1839               Node* load = in->in(1);
1840               if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
1841                 vt = velt_type(load);
1842               } else if (in->Opcode() != Op_LShiftI) {
1843                 // Widen type to Int to avoid creation of right shift vector
1844                 // (align + data_size(s1) check in stmts_can_pack() will fail).
1845                 // Note, left shifts work regardless type.
1846                 vt = TypeInt::INT;
1847               }
1848             }
1849             set_velt_type(in, vt);
1850           }
1851         }
1852       }
1853     }
1854   }
1855 #ifndef PRODUCT
1856   if (TraceSuperWord && Verbose) {
1857     for (int i = 0; i < _block.length(); i++) {
1858       Node* n = _block.at(i);
1859       velt_type(n)->dump();
1860       tty->print("\t");
1861       n->dump();
1862     }
1863   }
1864 #endif
1865 }
1866 
1867 //------------------------------memory_alignment---------------------------
1868 // Alignment within a vector memory reference
1869 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
1870   SWPointer p(s, this);
1871   if (!p.valid()) {
1872     return bottom_align;
1873   }
1874   int vw = vector_width_in_bytes(s);
1875   if (vw < 2) {
1876     return bottom_align; // No vectors for this type
1877   }
1878   int offset  = p.offset_in_bytes();
1879   offset     += iv_adjust*p.memory_size();
1880   int off_rem = offset % vw;
1881   int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
1882   return off_mod;
1883 }
1884 
1885 //---------------------------container_type---------------------------
1886 // Smallest type containing range of values
1887 const Type* SuperWord::container_type(Node* n) {
1888   if (n->is_Mem()) {
1889     BasicType bt = n->as_Mem()->memory_type();
1890     if (n->is_Store() && (bt == T_CHAR)) {
1891       // Use T_SHORT type instead of T_CHAR for stored values because any
1892       // preceding arithmetic operation extends values to signed Int.
1893       bt = T_SHORT;
1894     }
1895     if (n->Opcode() == Op_LoadUB) {
1896       // Adjust type for unsigned byte loads, it is important for right shifts.
1897       // T_BOOLEAN is used because there is no basic type representing type
1898       // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
1899       // size (one byte) and sign is important.
1900       bt = T_BOOLEAN;
1901     }
1902     return Type::get_const_basic_type(bt);
1903   }
1904   const Type* t = _igvn.type(n);
1905   if (t->basic_type() == T_INT) {
1906     // A narrow type of arithmetic operations will be determined by
1907     // propagating the type of memory operations.
1908     return TypeInt::INT;
1909   }
1910   return t;
1911 }
1912 
1913 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
1914   const Type* vt1 = velt_type(n1);
1915   const Type* vt2 = velt_type(n2);
1916   if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
1917     // Compare vectors element sizes for integer types.
1918     return data_size(n1) == data_size(n2);
1919   }
1920   return vt1 == vt2;
1921 }
1922 
1923 //------------------------------in_packset---------------------------
1924 // Are s1 and s2 in a pack pair and ordered as s1,s2?
1925 bool SuperWord::in_packset(Node* s1, Node* s2) {
1926   for (int i = 0; i < _packset.length(); i++) {
1927     Node_List* p = _packset.at(i);
1928     assert(p->size() == 2, "must be");
1929     if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
1930       return true;
1931     }
1932   }
1933   return false;
1934 }
1935 
1936 //------------------------------in_pack---------------------------
1937 // Is s in pack p?
1938 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
1939   for (uint i = 0; i < p->size(); i++) {
1940     if (p->at(i) == s) {
1941       return p;
1942     }
1943   }
1944   return NULL;
1945 }
1946 
1947 //------------------------------remove_pack_at---------------------------
1948 // Remove the pack at position pos in the packset
1949 void SuperWord::remove_pack_at(int pos) {
1950   Node_List* p = _packset.at(pos);
1951   for (uint i = 0; i < p->size(); i++) {
1952     Node* s = p->at(i);
1953     set_my_pack(s, NULL);
1954   }
1955   _packset.remove_at(pos);
1956 }
1957 
1958 //------------------------------executed_first---------------------------
1959 // Return the node executed first in pack p.  Uses the RPO block list
1960 // to determine order.
1961 Node* SuperWord::executed_first(Node_List* p) {
1962   Node* n = p->at(0);
1963   int n_rpo = bb_idx(n);
1964   for (uint i = 1; i < p->size(); i++) {
1965     Node* s = p->at(i);
1966     int s_rpo = bb_idx(s);
1967     if (s_rpo < n_rpo) {
1968       n = s;
1969       n_rpo = s_rpo;
1970     }
1971   }
1972   return n;
1973 }
1974 
1975 //------------------------------executed_last---------------------------
1976 // Return the node executed last in pack p.
1977 Node* SuperWord::executed_last(Node_List* p) {
1978   Node* n = p->at(0);
1979   int n_rpo = bb_idx(n);
1980   for (uint i = 1; i < p->size(); i++) {
1981     Node* s = p->at(i);
1982     int s_rpo = bb_idx(s);
1983     if (s_rpo > n_rpo) {
1984       n = s;
1985       n_rpo = s_rpo;
1986     }
1987   }
1988   return n;
1989 }
1990 
1991 //----------------------------align_initial_loop_index---------------------------
1992 // Adjust pre-loop limit so that in main loop, a load/store reference
1993 // to align_to_ref will be a position zero in the vector.
1994 //   (iv + k) mod vector_align == 0
1995 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
1996   CountedLoopNode *main_head = lp()->as_CountedLoop();
1997   assert(main_head->is_main_loop(), "");
1998   CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
1999   assert(pre_end != NULL, "");
2000   Node *pre_opaq1 = pre_end->limit();
2001   assert(pre_opaq1->Opcode() == Op_Opaque1, "");
2002   Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2003   Node *lim0 = pre_opaq->in(1);
2004 
2005   // Where we put new limit calculations
2006   Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2007 
2008   // Ensure the original loop limit is available from the
2009   // pre-loop Opaque1 node.
2010   Node *orig_limit = pre_opaq->original_loop_limit();
2011   assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
2012 
2013   SWPointer align_to_ref_p(align_to_ref, this);
2014   assert(align_to_ref_p.valid(), "sanity");
2015 
2016   // Given:
2017   //     lim0 == original pre loop limit
2018   //     V == v_align (power of 2)
2019   //     invar == extra invariant piece of the address expression
2020   //     e == offset [ +/- invar ]
2021   //
2022   // When reassociating expressions involving '%' the basic rules are:
2023   //     (a - b) % k == 0   =>  a % k == b % k
2024   // and:
2025   //     (a + b) % k == 0   =>  a % k == (k - b) % k
2026   //
2027   // For stride > 0 && scale > 0,
2028   //   Derive the new pre-loop limit "lim" such that the two constraints:
2029   //     (1) lim = lim0 + N           (where N is some positive integer < V)
2030   //     (2) (e + lim) % V == 0
2031   //   are true.
2032   //
2033   //   Substituting (1) into (2),
2034   //     (e + lim0 + N) % V == 0
2035   //   solve for N:
2036   //     N = (V - (e + lim0)) % V
2037   //   substitute back into (1), so that new limit
2038   //     lim = lim0 + (V - (e + lim0)) % V
2039   //
2040   // For stride > 0 && scale < 0
2041   //   Constraints:
2042   //     lim = lim0 + N
2043   //     (e - lim) % V == 0
2044   //   Solving for lim:
2045   //     (e - lim0 - N) % V == 0
2046   //     N = (e - lim0) % V
2047   //     lim = lim0 + (e - lim0) % V
2048   //
2049   // For stride < 0 && scale > 0
2050   //   Constraints:
2051   //     lim = lim0 - N
2052   //     (e + lim) % V == 0
2053   //   Solving for lim:
2054   //     (e + lim0 - N) % V == 0
2055   //     N = (e + lim0) % V
2056   //     lim = lim0 - (e + lim0) % V
2057   //
2058   // For stride < 0 && scale < 0
2059   //   Constraints:
2060   //     lim = lim0 - N
2061   //     (e - lim) % V == 0
2062   //   Solving for lim:
2063   //     (e - lim0 + N) % V == 0
2064   //     N = (V - (e - lim0)) % V
2065   //     lim = lim0 - (V - (e - lim0)) % V
2066 
2067   int vw = vector_width_in_bytes(align_to_ref);
2068   int stride   = iv_stride();
2069   int scale    = align_to_ref_p.scale_in_bytes();
2070   int elt_size = align_to_ref_p.memory_size();
2071   int v_align  = vw / elt_size;
2072   assert(v_align > 1, "sanity");
2073   int offset   = align_to_ref_p.offset_in_bytes() / elt_size;
2074   Node *offsn  = _igvn.intcon(offset);
2075 
2076   Node *e = offsn;
2077   if (align_to_ref_p.invar() != NULL) {
2078     // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
2079     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2080     Node* aref     = new (_phase->C) URShiftINode(align_to_ref_p.invar(), log2_elt);
2081     _igvn.register_new_node_with_optimizer(aref);
2082     _phase->set_ctrl(aref, pre_ctrl);
2083     if (align_to_ref_p.negate_invar()) {
2084       e = new (_phase->C) SubINode(e, aref);
2085     } else {
2086       e = new (_phase->C) AddINode(e, aref);
2087     }
2088     _igvn.register_new_node_with_optimizer(e);
2089     _phase->set_ctrl(e, pre_ctrl);
2090   }
2091   if (vw > ObjectAlignmentInBytes) {
2092     // incorporate base e +/- base && Mask >>> log2(elt)
2093     Node* xbase = new(_phase->C) CastP2XNode(NULL, align_to_ref_p.base());
2094     _igvn.register_new_node_with_optimizer(xbase);
2095 #ifdef _LP64
2096     xbase  = new (_phase->C) ConvL2INode(xbase);
2097     _igvn.register_new_node_with_optimizer(xbase);
2098 #endif
2099     Node* mask = _igvn.intcon(vw-1);
2100     Node* masked_xbase  = new (_phase->C) AndINode(xbase, mask);
2101     _igvn.register_new_node_with_optimizer(masked_xbase);
2102     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2103     Node* bref     = new (_phase->C) URShiftINode(masked_xbase, log2_elt);
2104     _igvn.register_new_node_with_optimizer(bref);
2105     _phase->set_ctrl(bref, pre_ctrl);
2106     e = new (_phase->C) AddINode(e, bref);
2107     _igvn.register_new_node_with_optimizer(e);
2108     _phase->set_ctrl(e, pre_ctrl);
2109   }
2110 
2111   // compute e +/- lim0
2112   if (scale < 0) {
2113     e = new (_phase->C) SubINode(e, lim0);
2114   } else {
2115     e = new (_phase->C) AddINode(e, lim0);
2116   }
2117   _igvn.register_new_node_with_optimizer(e);
2118   _phase->set_ctrl(e, pre_ctrl);
2119 
2120   if (stride * scale > 0) {
2121     // compute V - (e +/- lim0)
2122     Node* va  = _igvn.intcon(v_align);
2123     e = new (_phase->C) SubINode(va, e);
2124     _igvn.register_new_node_with_optimizer(e);
2125     _phase->set_ctrl(e, pre_ctrl);
2126   }
2127   // compute N = (exp) % V
2128   Node* va_msk = _igvn.intcon(v_align - 1);
2129   Node* N = new (_phase->C) AndINode(e, va_msk);
2130   _igvn.register_new_node_with_optimizer(N);
2131   _phase->set_ctrl(N, pre_ctrl);
2132 
2133   //   substitute back into (1), so that new limit
2134   //     lim = lim0 + N
2135   Node* lim;
2136   if (stride < 0) {
2137     lim = new (_phase->C) SubINode(lim0, N);
2138   } else {
2139     lim = new (_phase->C) AddINode(lim0, N);
2140   }
2141   _igvn.register_new_node_with_optimizer(lim);
2142   _phase->set_ctrl(lim, pre_ctrl);
2143   Node* constrained =
2144     (stride > 0) ? (Node*) new (_phase->C) MinINode(lim, orig_limit)
2145                  : (Node*) new (_phase->C) MaxINode(lim, orig_limit);
2146   _igvn.register_new_node_with_optimizer(constrained);
2147   _phase->set_ctrl(constrained, pre_ctrl);
2148   _igvn.hash_delete(pre_opaq);
2149   pre_opaq->set_req(1, constrained);
2150 }
2151 
2152 //----------------------------get_pre_loop_end---------------------------
2153 // Find pre loop end from main loop.  Returns null if none.
2154 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
2155   Node *ctrl = cl->in(LoopNode::EntryControl);
2156   if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
2157   Node *iffm = ctrl->in(0);
2158   if (!iffm->is_If()) return NULL;
2159   Node *p_f = iffm->in(0);
2160   if (!p_f->is_IfFalse()) return NULL;
2161   if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
2162   CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2163   if (!pre_end->loopnode()->is_pre_loop()) return NULL;
2164   return pre_end;
2165 }
2166 
2167 
2168 //------------------------------init---------------------------
2169 void SuperWord::init() {
2170   _dg.init();
2171   _packset.clear();
2172   _disjoint_ptrs.clear();
2173   _block.clear();
2174   _data_entry.clear();
2175   _mem_slice_head.clear();
2176   _mem_slice_tail.clear();
2177   _node_info.clear();
2178   _align_to_ref = NULL;
2179   _lpt = NULL;
2180   _lp = NULL;
2181   _bb = NULL;
2182   _iv = NULL;
2183 }
2184 
2185 //------------------------------print_packset---------------------------
2186 void SuperWord::print_packset() {
2187 #ifndef PRODUCT
2188   tty->print_cr("packset");
2189   for (int i = 0; i < _packset.length(); i++) {
2190     tty->print_cr("Pack: %d", i);
2191     Node_List* p = _packset.at(i);
2192     print_pack(p);
2193   }
2194 #endif
2195 }
2196 
2197 //------------------------------print_pack---------------------------
2198 void SuperWord::print_pack(Node_List* p) {
2199   for (uint i = 0; i < p->size(); i++) {
2200     print_stmt(p->at(i));
2201   }
2202 }
2203 
2204 //------------------------------print_bb---------------------------
2205 void SuperWord::print_bb() {
2206 #ifndef PRODUCT
2207   tty->print_cr("\nBlock");
2208   for (int i = 0; i < _block.length(); i++) {
2209     Node* n = _block.at(i);
2210     tty->print("%d ", i);
2211     if (n) {
2212       n->dump();
2213     }
2214   }
2215 #endif
2216 }
2217 
2218 //------------------------------print_stmt---------------------------
2219 void SuperWord::print_stmt(Node* s) {
2220 #ifndef PRODUCT
2221   tty->print(" align: %d \t", alignment(s));
2222   s->dump();
2223 #endif
2224 }
2225 
2226 //------------------------------blank---------------------------
2227 char* SuperWord::blank(uint depth) {
2228   static char blanks[101];
2229   assert(depth < 101, "too deep");
2230   for (uint i = 0; i < depth; i++) blanks[i] = ' ';
2231   blanks[depth] = '\0';
2232   return blanks;
2233 }
2234 
2235 
2236 //==============================SWPointer===========================
2237 
2238 //----------------------------SWPointer------------------------
2239 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
2240   _mem(mem), _slp(slp),  _base(NULL),  _adr(NULL),
2241   _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
2242 
2243   Node* adr = mem->in(MemNode::Address);
2244   if (!adr->is_AddP()) {
2245     assert(!valid(), "too complex");
2246     return;
2247   }
2248   // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
2249   Node* base = adr->in(AddPNode::Base);
2250   //unsafe reference could not be aligned appropriately without runtime checking
2251   if (base == NULL || base->bottom_type() == Type::TOP) {
2252     assert(!valid(), "unsafe access");
2253     return;
2254   }
2255   for (int i = 0; i < 3; i++) {
2256     if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
2257       assert(!valid(), "too complex");
2258       return;
2259     }
2260     adr = adr->in(AddPNode::Address);
2261     if (base == adr || !adr->is_AddP()) {
2262       break; // stop looking at addp's
2263     }
2264   }
2265   _base = base;
2266   _adr  = adr;
2267   assert(valid(), "Usable");
2268 }
2269 
2270 // Following is used to create a temporary object during
2271 // the pattern match of an address expression.
2272 SWPointer::SWPointer(SWPointer* p) :
2273   _mem(p->_mem), _slp(p->_slp),  _base(NULL),  _adr(NULL),
2274   _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
2275 
2276 //------------------------scaled_iv_plus_offset--------------------
2277 // Match: k*iv + offset
2278 // where: k is a constant that maybe zero, and
2279 //        offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
2280 bool SWPointer::scaled_iv_plus_offset(Node* n) {
2281   if (scaled_iv(n)) {
2282     return true;
2283   }
2284   if (offset_plus_k(n)) {
2285     return true;
2286   }
2287   int opc = n->Opcode();
2288   if (opc == Op_AddI) {
2289     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
2290       return true;
2291     }
2292     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2293       return true;
2294     }
2295   } else if (opc == Op_SubI) {
2296     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
2297       return true;
2298     }
2299     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2300       _scale *= -1;
2301       return true;
2302     }
2303   }
2304   return false;
2305 }
2306 
2307 //----------------------------scaled_iv------------------------
2308 // Match: k*iv where k is a constant that's not zero
2309 bool SWPointer::scaled_iv(Node* n) {
2310   if (_scale != 0) {
2311     return false;  // already found a scale
2312   }
2313   if (n == iv()) {
2314     _scale = 1;
2315     return true;
2316   }
2317   int opc = n->Opcode();
2318   if (opc == Op_MulI) {
2319     if (n->in(1) == iv() && n->in(2)->is_Con()) {
2320       _scale = n->in(2)->get_int();
2321       return true;
2322     } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2323       _scale = n->in(1)->get_int();
2324       return true;
2325     }
2326   } else if (opc == Op_LShiftI) {
2327     if (n->in(1) == iv() && n->in(2)->is_Con()) {
2328       _scale = 1 << n->in(2)->get_int();
2329       return true;
2330     }
2331   } else if (opc == Op_ConvI2L) {
2332     if (scaled_iv_plus_offset(n->in(1))) {
2333       return true;
2334     }
2335   } else if (opc == Op_LShiftL) {
2336     if (!has_iv() && _invar == NULL) {
2337       // Need to preserve the current _offset value, so
2338       // create a temporary object for this expression subtree.
2339       // Hacky, so should re-engineer the address pattern match.
2340       SWPointer tmp(this);
2341       if (tmp.scaled_iv_plus_offset(n->in(1))) {
2342         if (tmp._invar == NULL) {
2343           int mult = 1 << n->in(2)->get_int();
2344           _scale   = tmp._scale  * mult;
2345           _offset += tmp._offset * mult;
2346           return true;
2347         }
2348       }
2349     }
2350   }
2351   return false;
2352 }
2353 
2354 //----------------------------offset_plus_k------------------------
2355 // Match: offset is (k [+/- invariant])
2356 // where k maybe zero and invariant is optional, but not both.
2357 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2358   int opc = n->Opcode();
2359   if (opc == Op_ConI) {
2360     _offset += negate ? -(n->get_int()) : n->get_int();
2361     return true;
2362   } else if (opc == Op_ConL) {
2363     // Okay if value fits into an int
2364     const TypeLong* t = n->find_long_type();
2365     if (t->higher_equal(TypeLong::INT)) {
2366       jlong loff = n->get_long();
2367       jint  off  = (jint)loff;
2368       _offset += negate ? -off : loff;
2369       return true;
2370     }
2371     return false;
2372   }
2373   if (_invar != NULL) return false; // already have an invariant
2374   if (opc == Op_AddI) {
2375     if (n->in(2)->is_Con() && invariant(n->in(1))) {
2376       _negate_invar = negate;
2377       _invar = n->in(1);
2378       _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2379       return true;
2380     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2381       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2382       _negate_invar = negate;
2383       _invar = n->in(2);
2384       return true;
2385     }
2386   }
2387   if (opc == Op_SubI) {
2388     if (n->in(2)->is_Con() && invariant(n->in(1))) {
2389       _negate_invar = negate;
2390       _invar = n->in(1);
2391       _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2392       return true;
2393     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2394       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2395       _negate_invar = !negate;
2396       _invar = n->in(2);
2397       return true;
2398     }
2399   }
2400   if (invariant(n)) {
2401     _negate_invar = negate;
2402     _invar = n;
2403     return true;
2404   }
2405   return false;
2406 }
2407 
2408 //----------------------------print------------------------
2409 void SWPointer::print() {
2410 #ifndef PRODUCT
2411   tty->print("base: %d  adr: %d  scale: %d  offset: %d  invar: %c%d\n",
2412              _base != NULL ? _base->_idx : 0,
2413              _adr  != NULL ? _adr->_idx  : 0,
2414              _scale, _offset,
2415              _negate_invar?'-':'+',
2416              _invar != NULL ? _invar->_idx : 0);
2417 #endif
2418 }
2419 
2420 // ========================= OrderedPair =====================
2421 
2422 const OrderedPair OrderedPair::initial;
2423 
2424 // ========================= SWNodeInfo =====================
2425 
2426 const SWNodeInfo SWNodeInfo::initial;
2427 
2428 
2429 // ============================ DepGraph ===========================
2430 
2431 //------------------------------make_node---------------------------
2432 // Make a new dependence graph node for an ideal node.
2433 DepMem* DepGraph::make_node(Node* node) {
2434   DepMem* m = new (_arena) DepMem(node);
2435   if (node != NULL) {
2436     assert(_map.at_grow(node->_idx) == NULL, "one init only");
2437     _map.at_put_grow(node->_idx, m);
2438   }
2439   return m;
2440 }
2441 
2442 //------------------------------make_edge---------------------------
2443 // Make a new dependence graph edge from dpred -> dsucc
2444 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2445   DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2446   dpred->set_out_head(e);
2447   dsucc->set_in_head(e);
2448   return e;
2449 }
2450 
2451 // ========================== DepMem ========================
2452 
2453 //------------------------------in_cnt---------------------------
2454 int DepMem::in_cnt() {
2455   int ct = 0;
2456   for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2457   return ct;
2458 }
2459 
2460 //------------------------------out_cnt---------------------------
2461 int DepMem::out_cnt() {
2462   int ct = 0;
2463   for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2464   return ct;
2465 }
2466 
2467 //------------------------------print-----------------------------
2468 void DepMem::print() {
2469 #ifndef PRODUCT
2470   tty->print("  DepNode %d (", _node->_idx);
2471   for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2472     Node* pred = p->pred()->node();
2473     tty->print(" %d", pred != NULL ? pred->_idx : 0);
2474   }
2475   tty->print(") [");
2476   for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2477     Node* succ = s->succ()->node();
2478     tty->print(" %d", succ != NULL ? succ->_idx : 0);
2479   }
2480   tty->print_cr(" ]");
2481 #endif
2482 }
2483 
2484 // =========================== DepEdge =========================
2485 
2486 //------------------------------DepPreds---------------------------
2487 void DepEdge::print() {
2488 #ifndef PRODUCT
2489   tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2490 #endif
2491 }
2492 
2493 // =========================== DepPreds =========================
2494 // Iterator over predecessor edges in the dependence graph.
2495 
2496 //------------------------------DepPreds---------------------------
2497 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2498   _n = n;
2499   _done = false;
2500   if (_n->is_Store() || _n->is_Load()) {
2501     _next_idx = MemNode::Address;
2502     _end_idx  = n->req();
2503     _dep_next = dg.dep(_n)->in_head();
2504   } else if (_n->is_Mem()) {
2505     _next_idx = 0;
2506     _end_idx  = 0;
2507     _dep_next = dg.dep(_n)->in_head();
2508   } else {
2509     _next_idx = 1;
2510     _end_idx  = _n->req();
2511     _dep_next = NULL;
2512   }
2513   next();
2514 }
2515 
2516 //------------------------------next---------------------------
2517 void DepPreds::next() {
2518   if (_dep_next != NULL) {
2519     _current  = _dep_next->pred()->node();
2520     _dep_next = _dep_next->next_in();
2521   } else if (_next_idx < _end_idx) {
2522     _current  = _n->in(_next_idx++);
2523   } else {
2524     _done = true;
2525   }
2526 }
2527 
2528 // =========================== DepSuccs =========================
2529 // Iterator over successor edges in the dependence graph.
2530 
2531 //------------------------------DepSuccs---------------------------
2532 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2533   _n = n;
2534   _done = false;
2535   if (_n->is_Load()) {
2536     _next_idx = 0;
2537     _end_idx  = _n->outcnt();
2538     _dep_next = dg.dep(_n)->out_head();
2539   } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2540     _next_idx = 0;
2541     _end_idx  = 0;
2542     _dep_next = dg.dep(_n)->out_head();
2543   } else {
2544     _next_idx = 0;
2545     _end_idx  = _n->outcnt();
2546     _dep_next = NULL;
2547   }
2548   next();
2549 }
2550 
2551 //-------------------------------next---------------------------
2552 void DepSuccs::next() {
2553   if (_dep_next != NULL) {
2554     _current  = _dep_next->succ()->node();
2555     _dep_next = _dep_next->next_out();
2556   } else if (_next_idx < _end_idx) {
2557     _current  = _n->raw_out(_next_idx++);
2558   } else {
2559     _done = true;
2560   }
2561 }