1 /*
   2  * Copyright 2007-2009 Sun Microsystems, Inc.  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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  20  * CA 95054 USA or visit www.sun.com if you need additional information or
  21  * have any questions.
  22  */
  23 
  24 #include "incls/_precompiled.incl"
  25 #include "incls/_superword.cpp.incl"
  26 
  27 //
  28 //                  S U P E R W O R D   T R A N S F O R M
  29 //=============================================================================
  30 
  31 //------------------------------SuperWord---------------------------
  32 SuperWord::SuperWord(PhaseIdealLoop* phase) :
  33   _phase(phase),
  34   _igvn(phase->_igvn),
  35   _arena(phase->C->comp_arena()),
  36   _packset(arena(), 8,  0, NULL),         // packs for the current block
  37   _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
  38   _block(arena(), 8,  0, NULL),           // nodes in current block
  39   _data_entry(arena(), 8,  0, NULL),      // nodes with all inputs from outside
  40   _mem_slice_head(arena(), 8,  0, NULL),  // memory slice heads
  41   _mem_slice_tail(arena(), 8,  0, NULL),  // memory slice tails
  42   _node_info(arena(), 8,  0, SWNodeInfo::initial), // info needed per node
  43   _align_to_ref(NULL),                    // memory reference to align vectors to
  44   _disjoint_ptrs(arena(), 8,  0, OrderedPair::initial), // runtime disambiguated pointer pairs
  45   _dg(_arena),                            // dependence graph
  46   _visited(arena()),                      // visited node set
  47   _post_visited(arena()),                 // post visited node set
  48   _n_idx_list(arena(), 8),                // scratch list of (node,index) pairs
  49   _stk(arena(), 8, 0, NULL),              // scratch stack of nodes
  50   _nlist(arena(), 8, 0, NULL),            // scratch list of nodes
  51   _lpt(NULL),                             // loop tree node
  52   _lp(NULL),                              // LoopNode
  53   _bb(NULL),                              // basic block
  54   _iv(NULL)                               // induction var
  55 {}
  56 
  57 //------------------------------transform_loop---------------------------
  58 void SuperWord::transform_loop(IdealLoopTree* lpt) {
  59   assert(lpt->_head->is_CountedLoop(), "must be");
  60   CountedLoopNode *cl = lpt->_head->as_CountedLoop();
  61 
  62   if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
  63 
  64   // Check for no control flow in body (other than exit)
  65   Node *cl_exit = cl->loopexit();
  66   if (cl_exit->in(0) != lpt->_head) return;
  67 
  68   // Make sure the are no extra control users of the loop backedge
  69   if (cl->back_control()->outcnt() != 1) {
  70     return;
  71   }
  72 
  73   // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
  74   CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
  75   if (pre_end == NULL) return;
  76   Node *pre_opaq1 = pre_end->limit();
  77   if (pre_opaq1->Opcode() != Op_Opaque1) return;
  78 
  79   // Do vectors exist on this architecture?
  80   if (vector_width_in_bytes() == 0) return;
  81 
  82   init(); // initialize data structures
  83 
  84   set_lpt(lpt);
  85   set_lp(cl);
  86 
  87  // For now, define one block which is the entire loop body
  88   set_bb(cl);
  89 
  90   assert(_packset.length() == 0, "packset must be empty");
  91   SLP_extract();
  92 }
  93 
  94 //------------------------------SLP_extract---------------------------
  95 // Extract the superword level parallelism
  96 //
  97 // 1) A reverse post-order of nodes in the block is constructed.  By scanning
  98 //    this list from first to last, all definitions are visited before their uses.
  99 //
 100 // 2) A point-to-point dependence graph is constructed between memory references.
 101 //    This simplies the upcoming "independence" checker.
 102 //
 103 // 3) The maximum depth in the node graph from the beginning of the block
 104 //    to each node is computed.  This is used to prune the graph search
 105 //    in the independence checker.
 106 //
 107 // 4) For integer types, the necessary bit width is propagated backwards
 108 //    from stores to allow packed operations on byte, char, and short
 109 //    integers.  This reverses the promotion to type "int" that javac
 110 //    did for operations like: char c1,c2,c3;  c1 = c2 + c3.
 111 //
 112 // 5) One of the memory references is picked to be an aligned vector reference.
 113 //    The pre-loop trip count is adjusted to align this reference in the
 114 //    unrolled body.
 115 //
 116 // 6) The initial set of pack pairs is seeded with memory references.
 117 //
 118 // 7) The set of pack pairs is extended by following use->def and def->use links.
 119 //
 120 // 8) The pairs are combined into vector sized packs.
 121 //
 122 // 9) Reorder the memory slices to co-locate members of the memory packs.
 123 //
 124 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
 125 //    inserting scalar promotion, vector creation from multiple scalars, and
 126 //    extraction of scalar values from vectors.
 127 //
 128 void SuperWord::SLP_extract() {
 129 
 130   // Ready the block
 131 
 132   construct_bb();
 133 
 134   dependence_graph();
 135 
 136   compute_max_depth();
 137 
 138   compute_vector_element_type();
 139 
 140   // Attempt vectorization
 141 
 142   find_adjacent_refs();
 143 
 144   extend_packlist();
 145 
 146   combine_packs();
 147 
 148   construct_my_pack_map();
 149 
 150   filter_packs();
 151 
 152   schedule();
 153 
 154   output();
 155 }
 156 
 157 //------------------------------find_adjacent_refs---------------------------
 158 // Find the adjacent memory references and create pack pairs for them.
 159 // This is the initial set of packs that will then be extended by
 160 // following use->def and def->use links.  The align positions are
 161 // assigned relative to the reference "align_to_ref"
 162 void SuperWord::find_adjacent_refs() {
 163   // Get list of memory operations
 164   Node_List memops;
 165   for (int i = 0; i < _block.length(); i++) {
 166     Node* n = _block.at(i);
 167     if (n->is_Mem() && in_bb(n) &&
 168         is_java_primitive(n->as_Mem()->memory_type())) {
 169       int align = memory_alignment(n->as_Mem(), 0);
 170       if (align != bottom_align) {
 171         memops.push(n);
 172       }
 173     }
 174   }
 175   if (memops.size() == 0) return;
 176 
 177   // Find a memory reference to align to.  The pre-loop trip count
 178   // is modified to align this reference to a vector-aligned address
 179   find_align_to_ref(memops);
 180   if (align_to_ref() == NULL) return;
 181 
 182   SWPointer align_to_ref_p(align_to_ref(), this);
 183   int offset = align_to_ref_p.offset_in_bytes();
 184   int scale  = align_to_ref_p.scale_in_bytes();
 185   int vw              = vector_width_in_bytes();
 186   int stride_sign     = (scale * iv_stride()) > 0 ? 1 : -1;
 187   int iv_adjustment   = (stride_sign * vw - (offset % vw)) % vw;
 188 
 189 #ifndef PRODUCT
 190   if (TraceSuperWord)
 191     tty->print_cr("\noffset = %d iv_adjustment = %d  elt_align = %d scale = %d iv_stride = %d",
 192                   offset, iv_adjustment, align_to_ref_p.memory_size(), align_to_ref_p.scale_in_bytes(), iv_stride());
 193 #endif
 194 
 195   // Set alignment relative to "align_to_ref"
 196   for (int i = memops.size() - 1; i >= 0; i--) {
 197     MemNode* s = memops.at(i)->as_Mem();
 198     SWPointer p2(s, this);
 199     if (p2.comparable(align_to_ref_p)) {
 200       int align = memory_alignment(s, iv_adjustment);
 201       set_alignment(s, align);
 202     } else {
 203       memops.remove(i);
 204     }
 205   }
 206 
 207   // Create initial pack pairs of memory operations
 208   for (uint i = 0; i < memops.size(); i++) {
 209     Node* s1 = memops.at(i);
 210     for (uint j = 0; j < memops.size(); j++) {
 211       Node* s2 = memops.at(j);
 212       if (s1 != s2 && are_adjacent_refs(s1, s2)) {
 213         int align = alignment(s1);
 214         if (stmts_can_pack(s1, s2, align)) {
 215           Node_List* pair = new Node_List();
 216           pair->push(s1);
 217           pair->push(s2);
 218           _packset.append(pair);
 219         }
 220       }
 221     }
 222   }
 223 
 224 #ifndef PRODUCT
 225   if (TraceSuperWord) {
 226     tty->print_cr("\nAfter find_adjacent_refs");
 227     print_packset();
 228   }
 229 #endif
 230 }
 231 
 232 //------------------------------find_align_to_ref---------------------------
 233 // Find a memory reference to align the loop induction variable to.
 234 // Looks first at stores then at loads, looking for a memory reference
 235 // with the largest number of references similar to it.
 236 void SuperWord::find_align_to_ref(Node_List &memops) {
 237   GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
 238 
 239   // Count number of comparable memory ops
 240   for (uint i = 0; i < memops.size(); i++) {
 241     MemNode* s1 = memops.at(i)->as_Mem();
 242     SWPointer p1(s1, this);
 243     // Discard if pre loop can't align this reference
 244     if (!ref_is_alignable(p1)) {
 245       *cmp_ct.adr_at(i) = 0;
 246       continue;
 247     }
 248     for (uint j = i+1; j < memops.size(); j++) {
 249       MemNode* s2 = memops.at(j)->as_Mem();
 250       if (isomorphic(s1, s2)) {
 251         SWPointer p2(s2, this);
 252         if (p1.comparable(p2)) {
 253           (*cmp_ct.adr_at(i))++;
 254           (*cmp_ct.adr_at(j))++;
 255         }
 256       }
 257     }
 258   }
 259 
 260   // Find Store (or Load) with the greatest number of "comparable" references
 261   int max_ct        = 0;
 262   int max_idx       = -1;
 263   int min_size      = max_jint;
 264   int min_iv_offset = max_jint;
 265   for (uint j = 0; j < memops.size(); j++) {
 266     MemNode* s = memops.at(j)->as_Mem();
 267     if (s->is_Store()) {
 268       SWPointer p(s, this);
 269       if (cmp_ct.at(j) > max_ct ||
 270           cmp_ct.at(j) == max_ct && (data_size(s) < min_size ||
 271                                      data_size(s) == min_size &&
 272                                         p.offset_in_bytes() < min_iv_offset)) {
 273         max_ct = cmp_ct.at(j);
 274         max_idx = j;
 275         min_size = data_size(s);
 276         min_iv_offset = p.offset_in_bytes();
 277       }
 278     }
 279   }
 280   // If no stores, look at loads
 281   if (max_ct == 0) {
 282     for (uint j = 0; j < memops.size(); j++) {
 283       MemNode* s = memops.at(j)->as_Mem();
 284       if (s->is_Load()) {
 285         SWPointer p(s, this);
 286         if (cmp_ct.at(j) > max_ct ||
 287             cmp_ct.at(j) == max_ct && (data_size(s) < min_size ||
 288                                        data_size(s) == min_size &&
 289                                           p.offset_in_bytes() < min_iv_offset)) {
 290           max_ct = cmp_ct.at(j);
 291           max_idx = j;
 292           min_size = data_size(s);
 293           min_iv_offset = p.offset_in_bytes();
 294         }
 295       }
 296     }
 297   }
 298 
 299   if (max_ct > 0)
 300     set_align_to_ref(memops.at(max_idx)->as_Mem());
 301 
 302 #ifndef PRODUCT
 303   if (TraceSuperWord && Verbose) {
 304     tty->print_cr("\nVector memops after find_align_to_refs");
 305     for (uint i = 0; i < memops.size(); i++) {
 306       MemNode* s = memops.at(i)->as_Mem();
 307       s->dump();
 308     }
 309   }
 310 #endif
 311 }
 312 
 313 //------------------------------ref_is_alignable---------------------------
 314 // Can the preloop align the reference to position zero in the vector?
 315 bool SuperWord::ref_is_alignable(SWPointer& p) {
 316   if (!p.has_iv()) {
 317     return true;   // no induction variable
 318   }
 319   CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
 320   assert(pre_end->stride_is_con(), "pre loop stride is constant");
 321   int preloop_stride = pre_end->stride_con();
 322 
 323   int span = preloop_stride * p.scale_in_bytes();
 324 
 325   // Stride one accesses are alignable.
 326   if (ABS(span) == p.memory_size())
 327     return true;
 328 
 329   // If initial offset from start of object is computable,
 330   // compute alignment within the vector.
 331   int vw = vector_width_in_bytes();
 332   if (vw % span == 0) {
 333     Node* init_nd = pre_end->init_trip();
 334     if (init_nd->is_Con() && p.invar() == NULL) {
 335       int init = init_nd->bottom_type()->is_int()->get_con();
 336 
 337       int init_offset = init * p.scale_in_bytes() + p.offset_in_bytes();
 338       assert(init_offset >= 0, "positive offset from object start");
 339 
 340       if (span > 0) {
 341         return (vw - (init_offset % vw)) % span == 0;
 342       } else {
 343         assert(span < 0, "nonzero stride * scale");
 344         return (init_offset % vw) % -span == 0;
 345       }
 346     }
 347   }
 348   return false;
 349 }
 350 
 351 //---------------------------dependence_graph---------------------------
 352 // Construct dependency graph.
 353 // Add dependence edges to load/store nodes for memory dependence
 354 //    A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
 355 void SuperWord::dependence_graph() {
 356   // First, assign a dependence node to each memory node
 357   for (int i = 0; i < _block.length(); i++ ) {
 358     Node *n = _block.at(i);
 359     if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
 360       _dg.make_node(n);
 361     }
 362   }
 363 
 364   // For each memory slice, create the dependences
 365   for (int i = 0; i < _mem_slice_head.length(); i++) {
 366     Node* n      = _mem_slice_head.at(i);
 367     Node* n_tail = _mem_slice_tail.at(i);
 368 
 369     // Get slice in predecessor order (last is first)
 370     mem_slice_preds(n_tail, n, _nlist);
 371 
 372     // Make the slice dependent on the root
 373     DepMem* slice = _dg.dep(n);
 374     _dg.make_edge(_dg.root(), slice);
 375 
 376     // Create a sink for the slice
 377     DepMem* slice_sink = _dg.make_node(NULL);
 378     _dg.make_edge(slice_sink, _dg.tail());
 379 
 380     // Now visit each pair of memory ops, creating the edges
 381     for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 382       Node* s1 = _nlist.at(j);
 383 
 384       // If no dependency yet, use slice
 385       if (_dg.dep(s1)->in_cnt() == 0) {
 386         _dg.make_edge(slice, s1);
 387       }
 388       SWPointer p1(s1->as_Mem(), this);
 389       bool sink_dependent = true;
 390       for (int k = j - 1; k >= 0; k--) {
 391         Node* s2 = _nlist.at(k);
 392         if (s1->is_Load() && s2->is_Load())
 393           continue;
 394         SWPointer p2(s2->as_Mem(), this);
 395 
 396         int cmp = p1.cmp(p2);
 397         if (SuperWordRTDepCheck &&
 398             p1.base() != p2.base() && p1.valid() && p2.valid()) {
 399           // Create a runtime check to disambiguate
 400           OrderedPair pp(p1.base(), p2.base());
 401           _disjoint_ptrs.append_if_missing(pp);
 402         } else if (!SWPointer::not_equal(cmp)) {
 403           // Possibly same address
 404           _dg.make_edge(s1, s2);
 405           sink_dependent = false;
 406         }
 407       }
 408       if (sink_dependent) {
 409         _dg.make_edge(s1, slice_sink);
 410       }
 411     }
 412 #ifndef PRODUCT
 413     if (TraceSuperWord) {
 414       tty->print_cr("\nDependence graph for slice: %d", n->_idx);
 415       for (int q = 0; q < _nlist.length(); q++) {
 416         _dg.print(_nlist.at(q));
 417       }
 418       tty->cr();
 419     }
 420 #endif
 421     _nlist.clear();
 422   }
 423 
 424 #ifndef PRODUCT
 425   if (TraceSuperWord) {
 426     tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
 427     for (int r = 0; r < _disjoint_ptrs.length(); r++) {
 428       _disjoint_ptrs.at(r).print();
 429       tty->cr();
 430     }
 431     tty->cr();
 432   }
 433 #endif
 434 }
 435 
 436 //---------------------------mem_slice_preds---------------------------
 437 // Return a memory slice (node list) in predecessor order starting at "start"
 438 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
 439   assert(preds.length() == 0, "start empty");
 440   Node* n = start;
 441   Node* prev = NULL;
 442   while (true) {
 443     assert(in_bb(n), "must be in block");
 444     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 445       Node* out = n->fast_out(i);
 446       if (out->is_Load()) {
 447         if (in_bb(out)) {
 448           preds.push(out);
 449         }
 450       } else {
 451         // FIXME
 452         if (out->is_MergeMem() && !in_bb(out)) {
 453           // Either unrolling is causing a memory edge not to disappear,
 454           // or need to run igvn.optimize() again before SLP
 455         } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
 456           // Ditto.  Not sure what else to check further.
 457         } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
 458           // StoreCM has an input edge used as a precedence edge.
 459           // Maybe an issue when oop stores are vectorized.
 460         } else if( out->is_MergeMem() && prev &&
 461                    prev->Opcode() == Op_StoreCM && out == prev->in(MemNode::OopStore)) {
 462           // Oop store is a MergeMem! This should not happen. Temporarily remove the assertion
 463           // for this case because it could not be superwordized anyway.
 464         } else {
 465           assert(out == prev || prev == NULL, "no branches off of store slice");
 466         }
 467       }
 468     }
 469     if (n == stop) break;
 470     preds.push(n);
 471     prev = n;
 472     n = n->in(MemNode::Memory);
 473   }
 474 }
 475 
 476 //------------------------------stmts_can_pack---------------------------
 477 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
 478 // s1 aligned at "align"
 479 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
 480   if (isomorphic(s1, s2)) {
 481     if (independent(s1, s2)) {
 482       if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
 483         if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
 484           int s1_align = alignment(s1);
 485           int s2_align = alignment(s2);
 486           if (s1_align == top_align || s1_align == align) {
 487             if (s2_align == top_align || s2_align == align + data_size(s1)) {
 488               return true;
 489             }
 490           }
 491         }
 492       }
 493     }
 494   }
 495   return false;
 496 }
 497 
 498 //------------------------------exists_at---------------------------
 499 // Does s exist in a pack at position pos?
 500 bool SuperWord::exists_at(Node* s, uint pos) {
 501   for (int i = 0; i < _packset.length(); i++) {
 502     Node_List* p = _packset.at(i);
 503     if (p->at(pos) == s) {
 504       return true;
 505     }
 506   }
 507   return false;
 508 }
 509 
 510 //------------------------------are_adjacent_refs---------------------------
 511 // Is s1 immediately before s2 in memory?
 512 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
 513   if (!s1->is_Mem() || !s2->is_Mem()) return false;
 514   if (!in_bb(s1)    || !in_bb(s2))    return false;
 515   // FIXME - co_locate_pack fails on Stores in different mem-slices, so
 516   // only pack memops that are in the same alias set until that's fixed.
 517   if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
 518       _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
 519     return false;
 520   SWPointer p1(s1->as_Mem(), this);
 521   SWPointer p2(s2->as_Mem(), this);
 522   if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
 523   int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
 524   return diff == data_size(s1);
 525 }
 526 
 527 //------------------------------isomorphic---------------------------
 528 // Are s1 and s2 similar?
 529 bool SuperWord::isomorphic(Node* s1, Node* s2) {
 530   if (s1->Opcode() != s2->Opcode()) return false;
 531   if (s1->req() != s2->req()) return false;
 532   if (s1->in(0) != s2->in(0)) return false;
 533   if (velt_type(s1) != velt_type(s2)) return false;
 534   return true;
 535 }
 536 
 537 //------------------------------independent---------------------------
 538 // Is there no data path from s1 to s2 or s2 to s1?
 539 bool SuperWord::independent(Node* s1, Node* s2) {
 540   //  assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
 541   int d1 = depth(s1);
 542   int d2 = depth(s2);
 543   if (d1 == d2) return s1 != s2;
 544   Node* deep    = d1 > d2 ? s1 : s2;
 545   Node* shallow = d1 > d2 ? s2 : s1;
 546 
 547   visited_clear();
 548 
 549   return independent_path(shallow, deep);
 550 }
 551 
 552 //------------------------------independent_path------------------------------
 553 // Helper for independent
 554 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
 555   if (dp >= 1000) return false; // stop deep recursion
 556   visited_set(deep);
 557   int shal_depth = depth(shallow);
 558   assert(shal_depth <= depth(deep), "must be");
 559   for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
 560     Node* pred = preds.current();
 561     if (in_bb(pred) && !visited_test(pred)) {
 562       if (shallow == pred) {
 563         return false;
 564       }
 565       if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
 566         return false;
 567       }
 568     }
 569   }
 570   return true;
 571 }
 572 
 573 //------------------------------set_alignment---------------------------
 574 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
 575   set_alignment(s1, align);
 576   set_alignment(s2, align + data_size(s1));
 577 }
 578 
 579 //------------------------------data_size---------------------------
 580 int SuperWord::data_size(Node* s) {
 581   const Type* t = velt_type(s);
 582   BasicType  bt = t->array_element_basic_type();
 583   int bsize = type2aelembytes(bt);
 584   assert(bsize != 0, "valid size");
 585   return bsize;
 586 }
 587 
 588 //------------------------------extend_packlist---------------------------
 589 // Extend packset by following use->def and def->use links from pack members.
 590 void SuperWord::extend_packlist() {
 591   bool changed;
 592   do {
 593     changed = false;
 594     for (int i = 0; i < _packset.length(); i++) {
 595       Node_List* p = _packset.at(i);
 596       changed |= follow_use_defs(p);
 597       changed |= follow_def_uses(p);
 598     }
 599   } while (changed);
 600 
 601 #ifndef PRODUCT
 602   if (TraceSuperWord) {
 603     tty->print_cr("\nAfter extend_packlist");
 604     print_packset();
 605   }
 606 #endif
 607 }
 608 
 609 //------------------------------follow_use_defs---------------------------
 610 // Extend the packset by visiting operand definitions of nodes in pack p
 611 bool SuperWord::follow_use_defs(Node_List* p) {
 612   Node* s1 = p->at(0);
 613   Node* s2 = p->at(1);
 614   assert(p->size() == 2, "just checking");
 615   assert(s1->req() == s2->req(), "just checking");
 616   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
 617 
 618   if (s1->is_Load()) return false;
 619 
 620   int align = alignment(s1);
 621   bool changed = false;
 622   int start = s1->is_Store() ? MemNode::ValueIn   : 1;
 623   int end   = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
 624   for (int j = start; j < end; j++) {
 625     Node* t1 = s1->in(j);
 626     Node* t2 = s2->in(j);
 627     if (!in_bb(t1) || !in_bb(t2))
 628       continue;
 629     if (stmts_can_pack(t1, t2, align)) {
 630       if (est_savings(t1, t2) >= 0) {
 631         Node_List* pair = new Node_List();
 632         pair->push(t1);
 633         pair->push(t2);
 634         _packset.append(pair);
 635         set_alignment(t1, t2, align);
 636         changed = true;
 637       }
 638     }
 639   }
 640   return changed;
 641 }
 642 
 643 //------------------------------follow_def_uses---------------------------
 644 // Extend the packset by visiting uses of nodes in pack p
 645 bool SuperWord::follow_def_uses(Node_List* p) {
 646   bool changed = false;
 647   Node* s1 = p->at(0);
 648   Node* s2 = p->at(1);
 649   assert(p->size() == 2, "just checking");
 650   assert(s1->req() == s2->req(), "just checking");
 651   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
 652 
 653   if (s1->is_Store()) return false;
 654 
 655   int align = alignment(s1);
 656   int savings = -1;
 657   Node* u1 = NULL;
 658   Node* u2 = NULL;
 659   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
 660     Node* t1 = s1->fast_out(i);
 661     if (!in_bb(t1)) continue;
 662     for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
 663       Node* t2 = s2->fast_out(j);
 664       if (!in_bb(t2)) continue;
 665       if (!opnd_positions_match(s1, t1, s2, t2))
 666         continue;
 667       if (stmts_can_pack(t1, t2, align)) {
 668         int my_savings = est_savings(t1, t2);
 669         if (my_savings > savings) {
 670           savings = my_savings;
 671           u1 = t1;
 672           u2 = t2;
 673         }
 674       }
 675     }
 676   }
 677   if (savings >= 0) {
 678     Node_List* pair = new Node_List();
 679     pair->push(u1);
 680     pair->push(u2);
 681     _packset.append(pair);
 682     set_alignment(u1, u2, align);
 683     changed = true;
 684   }
 685   return changed;
 686 }
 687 
 688 //---------------------------opnd_positions_match-------------------------
 689 // Is the use of d1 in u1 at the same operand position as d2 in u2?
 690 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
 691   uint ct = u1->req();
 692   if (ct != u2->req()) return false;
 693   uint i1 = 0;
 694   uint i2 = 0;
 695   do {
 696     for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
 697     for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
 698     if (i1 != i2) {
 699       return false;
 700     }
 701   } while (i1 < ct);
 702   return true;
 703 }
 704 
 705 //------------------------------est_savings---------------------------
 706 // Estimate the savings from executing s1 and s2 as a pack
 707 int SuperWord::est_savings(Node* s1, Node* s2) {
 708   int save = 2 - 1; // 2 operations per instruction in packed form
 709 
 710   // inputs
 711   for (uint i = 1; i < s1->req(); i++) {
 712     Node* x1 = s1->in(i);
 713     Node* x2 = s2->in(i);
 714     if (x1 != x2) {
 715       if (are_adjacent_refs(x1, x2)) {
 716         save += adjacent_profit(x1, x2);
 717       } else if (!in_packset(x1, x2)) {
 718         save -= pack_cost(2);
 719       } else {
 720         save += unpack_cost(2);
 721       }
 722     }
 723   }
 724 
 725   // uses of result
 726   uint ct = 0;
 727   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
 728     Node* s1_use = s1->fast_out(i);
 729     for (int j = 0; j < _packset.length(); j++) {
 730       Node_List* p = _packset.at(j);
 731       if (p->at(0) == s1_use) {
 732         for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
 733           Node* s2_use = s2->fast_out(k);
 734           if (p->at(p->size()-1) == s2_use) {
 735             ct++;
 736             if (are_adjacent_refs(s1_use, s2_use)) {
 737               save += adjacent_profit(s1_use, s2_use);
 738             }
 739           }
 740         }
 741       }
 742     }
 743   }
 744 
 745   if (ct < s1->outcnt()) save += unpack_cost(1);
 746   if (ct < s2->outcnt()) save += unpack_cost(1);
 747 
 748   return save;
 749 }
 750 
 751 //------------------------------costs---------------------------
 752 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
 753 int SuperWord::pack_cost(int ct)   { return ct; }
 754 int SuperWord::unpack_cost(int ct) { return ct; }
 755 
 756 //------------------------------combine_packs---------------------------
 757 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
 758 void SuperWord::combine_packs() {
 759   bool changed;
 760   do {
 761     changed = false;
 762     for (int i = 0; i < _packset.length(); i++) {
 763       Node_List* p1 = _packset.at(i);
 764       if (p1 == NULL) continue;
 765       for (int j = 0; j < _packset.length(); j++) {
 766         Node_List* p2 = _packset.at(j);
 767         if (p2 == NULL) continue;
 768         if (p1->at(p1->size()-1) == p2->at(0)) {
 769           for (uint k = 1; k < p2->size(); k++) {
 770             p1->push(p2->at(k));
 771           }
 772           _packset.at_put(j, NULL);
 773           changed = true;
 774         }
 775       }
 776     }
 777   } while (changed);
 778 
 779   for (int i = _packset.length() - 1; i >= 0; i--) {
 780     Node_List* p1 = _packset.at(i);
 781     if (p1 == NULL) {
 782       _packset.remove_at(i);
 783     }
 784   }
 785 
 786 #ifndef PRODUCT
 787   if (TraceSuperWord) {
 788     tty->print_cr("\nAfter combine_packs");
 789     print_packset();
 790   }
 791 #endif
 792 }
 793 
 794 //-----------------------------construct_my_pack_map--------------------------
 795 // Construct the map from nodes to packs.  Only valid after the
 796 // point where a node is only in one pack (after combine_packs).
 797 void SuperWord::construct_my_pack_map() {
 798   Node_List* rslt = NULL;
 799   for (int i = 0; i < _packset.length(); i++) {
 800     Node_List* p = _packset.at(i);
 801     for (uint j = 0; j < p->size(); j++) {
 802       Node* s = p->at(j);
 803       assert(my_pack(s) == NULL, "only in one pack");
 804       set_my_pack(s, p);
 805     }
 806   }
 807 }
 808 
 809 //------------------------------filter_packs---------------------------
 810 // Remove packs that are not implemented or not profitable.
 811 void SuperWord::filter_packs() {
 812 
 813   // Remove packs that are not implemented
 814   for (int i = _packset.length() - 1; i >= 0; i--) {
 815     Node_List* pk = _packset.at(i);
 816     bool impl = implemented(pk);
 817     if (!impl) {
 818 #ifndef PRODUCT
 819       if (TraceSuperWord && Verbose) {
 820         tty->print_cr("Unimplemented");
 821         pk->at(0)->dump();
 822       }
 823 #endif
 824       remove_pack_at(i);
 825     }
 826   }
 827 
 828   // Remove packs that are not profitable
 829   bool changed;
 830   do {
 831     changed = false;
 832     for (int i = _packset.length() - 1; i >= 0; i--) {
 833       Node_List* pk = _packset.at(i);
 834       bool prof = profitable(pk);
 835       if (!prof) {
 836 #ifndef PRODUCT
 837         if (TraceSuperWord && Verbose) {
 838           tty->print_cr("Unprofitable");
 839           pk->at(0)->dump();
 840         }
 841 #endif
 842         remove_pack_at(i);
 843         changed = true;
 844       }
 845     }
 846   } while (changed);
 847 
 848 #ifndef PRODUCT
 849   if (TraceSuperWord) {
 850     tty->print_cr("\nAfter filter_packs");
 851     print_packset();
 852     tty->cr();
 853   }
 854 #endif
 855 }
 856 
 857 //------------------------------implemented---------------------------
 858 // Can code be generated for pack p?
 859 bool SuperWord::implemented(Node_List* p) {
 860   Node* p0 = p->at(0);
 861   int vopc = VectorNode::opcode(p0->Opcode(), p->size(), velt_type(p0));
 862   return vopc > 0 && Matcher::has_match_rule(vopc);
 863 }
 864 
 865 //------------------------------profitable---------------------------
 866 // For pack p, are all operands and all uses (with in the block) vector?
 867 bool SuperWord::profitable(Node_List* p) {
 868   Node* p0 = p->at(0);
 869   uint start, end;
 870   vector_opd_range(p0, &start, &end);
 871 
 872   // Return false if some input is not vector and inside block
 873   for (uint i = start; i < end; i++) {
 874     if (!is_vector_use(p0, i)) {
 875       // For now, return false if not scalar promotion case (inputs are the same.)
 876       // Later, implement PackNode and allow differing, non-vector inputs
 877       // (maybe just the ones from outside the block.)
 878       Node* p0_def = p0->in(i);
 879       for (uint j = 1; j < p->size(); j++) {
 880         Node* use = p->at(j);
 881         Node* def = use->in(i);
 882         if (p0_def != def)
 883           return false;
 884       }
 885     }
 886   }
 887   if (!p0->is_Store()) {
 888     // For now, return false if not all uses are vector.
 889     // Later, implement ExtractNode and allow non-vector uses (maybe
 890     // just the ones outside the block.)
 891     for (uint i = 0; i < p->size(); i++) {
 892       Node* def = p->at(i);
 893       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
 894         Node* use = def->fast_out(j);
 895         for (uint k = 0; k < use->req(); k++) {
 896           Node* n = use->in(k);
 897           if (def == n) {
 898             if (!is_vector_use(use, k)) {
 899               return false;
 900             }
 901           }
 902         }
 903       }
 904     }
 905   }
 906   return true;
 907 }
 908 
 909 //------------------------------schedule---------------------------
 910 // Adjust the memory graph for the packed operations
 911 void SuperWord::schedule() {
 912 
 913   // Co-locate in the memory graph the members of each memory pack
 914   for (int i = 0; i < _packset.length(); i++) {
 915     co_locate_pack(_packset.at(i));
 916   }
 917 }
 918 
 919 //-------------------------------remove_and_insert-------------------
 920 //remove "current" from its current position in the memory graph and insert
 921 //it after the appropriate insertion point (lip or uip)
 922 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
 923                                   Node *uip, Unique_Node_List &sched_before) {
 924   Node* my_mem = current->in(MemNode::Memory);
 925   _igvn.hash_delete(current);
 926   _igvn.hash_delete(my_mem);
 927 
 928   //remove current_store from its current position in the memmory graph
 929   for (DUIterator i = current->outs(); current->has_out(i); i++) {
 930     Node* use = current->out(i);
 931     if (use->is_Mem()) {
 932       assert(use->in(MemNode::Memory) == current, "must be");
 933       _igvn.hash_delete(use);
 934       if (use == prev) { // connect prev to my_mem
 935         use->set_req(MemNode::Memory, my_mem);
 936       } else if (sched_before.member(use)) {
 937         _igvn.hash_delete(uip);
 938         use->set_req(MemNode::Memory, uip);
 939       } else {
 940         _igvn.hash_delete(lip);
 941         use->set_req(MemNode::Memory, lip);
 942       }
 943       _igvn._worklist.push(use);
 944       --i; //deleted this edge; rescan position
 945     }
 946   }
 947 
 948   bool sched_up = sched_before.member(current);
 949   Node *insert_pt =  sched_up ?  uip : lip;
 950   _igvn.hash_delete(insert_pt);
 951 
 952   // all uses of insert_pt's memory state should use current's instead
 953   for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
 954     Node* use = insert_pt->out(i);
 955     if (use->is_Mem()) {
 956       assert(use->in(MemNode::Memory) == insert_pt, "must be");
 957       _igvn.hash_delete(use);
 958       use->set_req(MemNode::Memory, current);
 959       _igvn._worklist.push(use);
 960       --i; //deleted this edge; rescan position
 961     } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
 962       uint pos; //lip (lower insert point) must be the last one in the memory slice
 963       _igvn.hash_delete(use);
 964       for (pos=1; pos < use->req(); pos++) {
 965         if (use->in(pos) == insert_pt) break;
 966       }
 967       use->set_req(pos, current);
 968       _igvn._worklist.push(use);
 969       --i;
 970     }
 971   }
 972 
 973   //connect current to insert_pt
 974   current->set_req(MemNode::Memory, insert_pt);
 975   _igvn._worklist.push(current);
 976 }
 977 
 978 //------------------------------co_locate_pack----------------------------------
 979 // To schedule a store pack, we need to move any sandwiched memory ops either before
 980 // or after the pack, based upon dependence information:
 981 // (1) If any store in the pack depends on the sandwiched memory op, the
 982 //     sandwiched memory op must be scheduled BEFORE the pack;
 983 // (2) If a sandwiched memory op depends on any store in the pack, the
 984 //     sandwiched memory op must be scheduled AFTER the pack;
 985 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
 986 //     memory op (say memB), memB must be scheduled before memA. So, if memA is
 987 //     scheduled before the pack, memB must also be scheduled before the pack;
 988 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
 989 //     schedule this store AFTER the pack
 990 // (5) We know there is no dependence cycle, so there in no other case;
 991 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
 992 //
 993 // To schedule a load pack, we use the memory state of either the first or the last load in
 994 // the pack, based on the dependence constraint.
 995 void SuperWord::co_locate_pack(Node_List* pk) {
 996   if (pk->at(0)->is_Store()) {
 997     MemNode* first     = executed_first(pk)->as_Mem();
 998     MemNode* last      = executed_last(pk)->as_Mem();
 999     Unique_Node_List schedule_before_pack;
1000     Unique_Node_List memops;
1001 
1002     MemNode* current   = last->in(MemNode::Memory)->as_Mem();
1003     MemNode* previous  = last;
1004     while (true) {
1005       assert(in_bb(current), "stay in block");
1006       memops.push(previous);
1007       for (DUIterator i = current->outs(); current->has_out(i); i++) {
1008         Node* use = current->out(i);
1009         if (use->is_Mem() && use != previous)
1010           memops.push(use);
1011       }
1012       if(current == first) break;
1013       previous = current;
1014       current  = current->in(MemNode::Memory)->as_Mem();
1015     }
1016 
1017     // determine which memory operations should be scheduled before the pack
1018     for (uint i = 1; i < memops.size(); i++) {
1019       Node *s1 = memops.at(i);
1020       if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1021         for (uint j = 0; j< i; j++) {
1022           Node *s2 = memops.at(j);
1023           if (!independent(s1, s2)) {
1024             if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1025               schedule_before_pack.push(s1); //s1 must be scheduled before
1026               Node_List* mem_pk = my_pack(s1);
1027               if (mem_pk != NULL) {
1028                 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1029                   Node* s = mem_pk->at(ii); // follow partner
1030                   if (memops.member(s) && !schedule_before_pack.member(s))
1031                     schedule_before_pack.push(s);
1032                 }
1033               }
1034             }
1035           }
1036         }
1037       }
1038     }
1039 
1040     MemNode* lower_insert_pt = last;
1041     Node*    upper_insert_pt = first->in(MemNode::Memory);
1042     previous                 = last; //previous store in pk
1043     current                  = last->in(MemNode::Memory)->as_Mem();
1044 
1045     //start scheduling from "last" to "first"
1046     while (true) {
1047       assert(in_bb(current), "stay in block");
1048       assert(in_pack(previous, pk), "previous stays in pack");
1049       Node* my_mem = current->in(MemNode::Memory);
1050 
1051       if (in_pack(current, pk)) {
1052         // Forward users of my memory state (except "previous) to my input memory state
1053         _igvn.hash_delete(current);
1054         for (DUIterator i = current->outs(); current->has_out(i); i++) {
1055           Node* use = current->out(i);
1056           if (use->is_Mem() && use != previous) {
1057             assert(use->in(MemNode::Memory) == current, "must be");
1058             _igvn.hash_delete(use);
1059             if (schedule_before_pack.member(use)) {
1060               _igvn.hash_delete(upper_insert_pt);
1061               use->set_req(MemNode::Memory, upper_insert_pt);
1062             } else {
1063               _igvn.hash_delete(lower_insert_pt);
1064               use->set_req(MemNode::Memory, lower_insert_pt);
1065             }
1066             _igvn._worklist.push(use);
1067             --i; // deleted this edge; rescan position
1068           }
1069         }
1070         previous = current;
1071       } else { // !in_pack(current, pk) ==> a sandwiched store
1072         remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1073       }
1074 
1075       if (current == first) break;
1076       current = my_mem->as_Mem();
1077     } // end while
1078   } else if (pk->at(0)->is_Load()) { //load
1079     // all loads in the pack should have the same memory state. By default,
1080     // we use the memory state of the last load. However, if any load could
1081     // not be moved down due to the dependence constraint, we use the memory
1082     // state of the first load.
1083     Node* last_mem  = executed_last(pk)->in(MemNode::Memory);
1084     Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1085     bool schedule_last = true;
1086     for (uint i = 0; i < pk->size(); i++) {
1087       Node* ld = pk->at(i);
1088       for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1089            current=current->in(MemNode::Memory)) {
1090         assert(current != first_mem, "corrupted memory graph");
1091         if(current->is_Mem() && !independent(current, ld)){
1092           schedule_last = false; // a later store depends on this load
1093           break;
1094         }
1095       }
1096     }
1097 
1098     Node* mem_input = schedule_last ? last_mem : first_mem;
1099     _igvn.hash_delete(mem_input);
1100     // Give each load the same memory state
1101     for (uint i = 0; i < pk->size(); i++) {
1102       LoadNode* ld = pk->at(i)->as_Load();
1103       _igvn.hash_delete(ld);
1104       ld->set_req(MemNode::Memory, mem_input);
1105       _igvn._worklist.push(ld);
1106     }
1107   }
1108 }
1109 
1110 //------------------------------output---------------------------
1111 // Convert packs into vector node operations
1112 void SuperWord::output() {
1113   if (_packset.length() == 0) return;
1114 
1115   // MUST ENSURE main loop's initial value is properly aligned:
1116   //  (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1117 
1118   align_initial_loop_index(align_to_ref());
1119 
1120   // Insert extract (unpack) operations for scalar uses
1121   for (int i = 0; i < _packset.length(); i++) {
1122     insert_extracts(_packset.at(i));
1123   }
1124 
1125   for (int i = 0; i < _block.length(); i++) {
1126     Node* n = _block.at(i);
1127     Node_List* p = my_pack(n);
1128     if (p && n == executed_last(p)) {
1129       uint vlen = p->size();
1130       Node* vn = NULL;
1131       Node* low_adr = p->at(0);
1132       Node* first   = executed_first(p);
1133       if (n->is_Load()) {
1134         int   opc = n->Opcode();
1135         Node* ctl = n->in(MemNode::Control);
1136         Node* mem = first->in(MemNode::Memory);
1137         Node* adr = low_adr->in(MemNode::Address);
1138         const TypePtr* atyp = n->adr_type();
1139         vn = VectorLoadNode::make(_phase->C, opc, ctl, mem, adr, atyp, vlen);
1140 
1141       } else if (n->is_Store()) {
1142         // Promote value to be stored to vector
1143         VectorNode* val = vector_opd(p, MemNode::ValueIn);
1144 
1145         int   opc = n->Opcode();
1146         Node* ctl = n->in(MemNode::Control);
1147         Node* mem = first->in(MemNode::Memory);
1148         Node* adr = low_adr->in(MemNode::Address);
1149         const TypePtr* atyp = n->adr_type();
1150         vn = VectorStoreNode::make(_phase->C, opc, ctl, mem, adr, atyp, val, vlen);
1151 
1152       } else if (n->req() == 3) {
1153         // Promote operands to vector
1154         Node* in1 = vector_opd(p, 1);
1155         Node* in2 = vector_opd(p, 2);
1156         vn = VectorNode::make(_phase->C, n->Opcode(), in1, in2, vlen, velt_type(n));
1157 
1158       } else {
1159         ShouldNotReachHere();
1160       }
1161 
1162       _phase->_igvn.register_new_node_with_optimizer(vn);
1163       _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1164       for (uint j = 0; j < p->size(); j++) {
1165         Node* pm = p->at(j);
1166         _igvn.hash_delete(pm);
1167         _igvn.subsume_node(pm, vn);
1168       }
1169       _igvn._worklist.push(vn);
1170     }
1171   }
1172 }
1173 
1174 //------------------------------vector_opd---------------------------
1175 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1176 VectorNode* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1177   Node* p0 = p->at(0);
1178   uint vlen = p->size();
1179   Node* opd = p0->in(opd_idx);
1180 
1181   bool same_opd = true;
1182   for (uint i = 1; i < vlen; i++) {
1183     Node* pi = p->at(i);
1184     Node* in = pi->in(opd_idx);
1185     if (opd != in) {
1186       same_opd = false;
1187       break;
1188     }
1189   }
1190 
1191   if (same_opd) {
1192     if (opd->is_Vector()) {
1193       return (VectorNode*)opd; // input is matching vector
1194     }
1195     // Convert scalar input to vector. Use p0's type because it's container
1196     // maybe smaller than the operand's container.
1197     const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd);
1198     const Type* p0_t  = velt_type(p0);
1199     if (p0_t->higher_equal(opd_t)) opd_t = p0_t;
1200     VectorNode* vn    = VectorNode::scalar2vector(_phase->C, opd, vlen, opd_t);
1201 
1202     _phase->_igvn.register_new_node_with_optimizer(vn);
1203     _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1204     return vn;
1205   }
1206 
1207   // Insert pack operation
1208   const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd);
1209   PackNode* pk = PackNode::make(_phase->C, opd, opd_t);
1210 
1211   for (uint i = 1; i < vlen; i++) {
1212     Node* pi = p->at(i);
1213     Node* in = pi->in(opd_idx);
1214     assert(my_pack(in) == NULL, "Should already have been unpacked");
1215     assert(opd_t == velt_type(!in_bb(in) ? pi : in), "all same type");
1216     pk->add_opd(in);
1217   }
1218   _phase->_igvn.register_new_node_with_optimizer(pk);
1219   _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1220   return pk;
1221 }
1222 
1223 //------------------------------insert_extracts---------------------------
1224 // If a use of pack p is not a vector use, then replace the
1225 // use with an extract operation.
1226 void SuperWord::insert_extracts(Node_List* p) {
1227   if (p->at(0)->is_Store()) return;
1228   assert(_n_idx_list.is_empty(), "empty (node,index) list");
1229 
1230   // Inspect each use of each pack member.  For each use that is
1231   // not a vector use, replace the use with an extract operation.
1232 
1233   for (uint i = 0; i < p->size(); i++) {
1234     Node* def = p->at(i);
1235     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1236       Node* use = def->fast_out(j);
1237       for (uint k = 0; k < use->req(); k++) {
1238         Node* n = use->in(k);
1239         if (def == n) {
1240           if (!is_vector_use(use, k)) {
1241             _n_idx_list.push(use, k);
1242           }
1243         }
1244       }
1245     }
1246   }
1247 
1248   while (_n_idx_list.is_nonempty()) {
1249     Node* use = _n_idx_list.node();
1250     int   idx = _n_idx_list.index();
1251     _n_idx_list.pop();
1252     Node* def = use->in(idx);
1253 
1254     // Insert extract operation
1255     _igvn.hash_delete(def);
1256     _igvn.hash_delete(use);
1257     int def_pos = alignment(def) / data_size(def);
1258     const Type* def_t = velt_type(def);
1259 
1260     Node* ex = ExtractNode::make(_phase->C, def, def_pos, def_t);
1261     _phase->_igvn.register_new_node_with_optimizer(ex);
1262     _phase->set_ctrl(ex, _phase->get_ctrl(def));
1263     use->set_req(idx, ex);
1264     _igvn._worklist.push(def);
1265     _igvn._worklist.push(use);
1266 
1267     bb_insert_after(ex, bb_idx(def));
1268     set_velt_type(ex, def_t);
1269   }
1270 }
1271 
1272 //------------------------------is_vector_use---------------------------
1273 // Is use->in(u_idx) a vector use?
1274 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1275   Node_List* u_pk = my_pack(use);
1276   if (u_pk == NULL) return false;
1277   Node* def = use->in(u_idx);
1278   Node_List* d_pk = my_pack(def);
1279   if (d_pk == NULL) {
1280     // check for scalar promotion
1281     Node* n = u_pk->at(0)->in(u_idx);
1282     for (uint i = 1; i < u_pk->size(); i++) {
1283       if (u_pk->at(i)->in(u_idx) != n) return false;
1284     }
1285     return true;
1286   }
1287   if (u_pk->size() != d_pk->size())
1288     return false;
1289   for (uint i = 0; i < u_pk->size(); i++) {
1290     Node* ui = u_pk->at(i);
1291     Node* di = d_pk->at(i);
1292     if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1293       return false;
1294   }
1295   return true;
1296 }
1297 
1298 //------------------------------construct_bb---------------------------
1299 // Construct reverse postorder list of block members
1300 void SuperWord::construct_bb() {
1301   Node* entry = bb();
1302 
1303   assert(_stk.length() == 0,            "stk is empty");
1304   assert(_block.length() == 0,          "block is empty");
1305   assert(_data_entry.length() == 0,     "data_entry is empty");
1306   assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1307   assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1308 
1309   // Find non-control nodes with no inputs from within block,
1310   // create a temporary map from node _idx to bb_idx for use
1311   // by the visited and post_visited sets,
1312   // and count number of nodes in block.
1313   int bb_ct = 0;
1314   for (uint i = 0; i < lpt()->_body.size(); i++ ) {
1315     Node *n = lpt()->_body.at(i);
1316     set_bb_idx(n, i); // Create a temporary map
1317     if (in_bb(n)) {
1318       bb_ct++;
1319       if (!n->is_CFG()) {
1320         bool found = false;
1321         for (uint j = 0; j < n->req(); j++) {
1322           Node* def = n->in(j);
1323           if (def && in_bb(def)) {
1324             found = true;
1325             break;
1326           }
1327         }
1328         if (!found) {
1329           assert(n != entry, "can't be entry");
1330           _data_entry.push(n);
1331         }
1332       }
1333     }
1334   }
1335 
1336   // Find memory slices (head and tail)
1337   for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1338     Node *n = lp()->fast_out(i);
1339     if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1340       Node* n_tail  = n->in(LoopNode::LoopBackControl);
1341       if (n_tail != n->in(LoopNode::EntryControl)) {
1342         _mem_slice_head.push(n);
1343         _mem_slice_tail.push(n_tail);
1344       }
1345     }
1346   }
1347 
1348   // Create an RPO list of nodes in block
1349 
1350   visited_clear();
1351   post_visited_clear();
1352 
1353   // Push all non-control nodes with no inputs from within block, then control entry
1354   for (int j = 0; j < _data_entry.length(); j++) {
1355     Node* n = _data_entry.at(j);
1356     visited_set(n);
1357     _stk.push(n);
1358   }
1359   visited_set(entry);
1360   _stk.push(entry);
1361 
1362   // Do a depth first walk over out edges
1363   int rpo_idx = bb_ct - 1;
1364   int size;
1365   while ((size = _stk.length()) > 0) {
1366     Node* n = _stk.top(); // Leave node on stack
1367     if (!visited_test_set(n)) {
1368       // forward arc in graph
1369     } else if (!post_visited_test(n)) {
1370       // cross or back arc
1371       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1372         Node *use = n->fast_out(i);
1373         if (in_bb(use) && !visited_test(use) &&
1374             // Don't go around backedge
1375             (!use->is_Phi() || n == entry)) {
1376           _stk.push(use);
1377         }
1378       }
1379       if (_stk.length() == size) {
1380         // There were no additional uses, post visit node now
1381         _stk.pop(); // Remove node from stack
1382         assert(rpo_idx >= 0, "");
1383         _block.at_put_grow(rpo_idx, n);
1384         rpo_idx--;
1385         post_visited_set(n);
1386         assert(rpo_idx >= 0 || _stk.is_empty(), "");
1387       }
1388     } else {
1389       _stk.pop(); // Remove post-visited node from stack
1390     }
1391   }
1392 
1393   // Create real map of block indices for nodes
1394   for (int j = 0; j < _block.length(); j++) {
1395     Node* n = _block.at(j);
1396     set_bb_idx(n, j);
1397   }
1398 
1399   initialize_bb(); // Ensure extra info is allocated.
1400 
1401 #ifndef PRODUCT
1402   if (TraceSuperWord) {
1403     print_bb();
1404     tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1405     for (int m = 0; m < _data_entry.length(); m++) {
1406       tty->print("%3d ", m);
1407       _data_entry.at(m)->dump();
1408     }
1409     tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1410     for (int m = 0; m < _mem_slice_head.length(); m++) {
1411       tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1412       tty->print("    ");    _mem_slice_tail.at(m)->dump();
1413     }
1414   }
1415 #endif
1416   assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1417 }
1418 
1419 //------------------------------initialize_bb---------------------------
1420 // Initialize per node info
1421 void SuperWord::initialize_bb() {
1422   Node* last = _block.at(_block.length() - 1);
1423   grow_node_info(bb_idx(last));
1424 }
1425 
1426 //------------------------------bb_insert_after---------------------------
1427 // Insert n into block after pos
1428 void SuperWord::bb_insert_after(Node* n, int pos) {
1429   int n_pos = pos + 1;
1430   // Make room
1431   for (int i = _block.length() - 1; i >= n_pos; i--) {
1432     _block.at_put_grow(i+1, _block.at(i));
1433   }
1434   for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1435     _node_info.at_put_grow(j+1, _node_info.at(j));
1436   }
1437   // Set value
1438   _block.at_put_grow(n_pos, n);
1439   _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1440   // Adjust map from node->_idx to _block index
1441   for (int i = n_pos; i < _block.length(); i++) {
1442     set_bb_idx(_block.at(i), i);
1443   }
1444 }
1445 
1446 //------------------------------compute_max_depth---------------------------
1447 // Compute max depth for expressions from beginning of block
1448 // Use to prune search paths during test for independence.
1449 void SuperWord::compute_max_depth() {
1450   int ct = 0;
1451   bool again;
1452   do {
1453     again = false;
1454     for (int i = 0; i < _block.length(); i++) {
1455       Node* n = _block.at(i);
1456       if (!n->is_Phi()) {
1457         int d_orig = depth(n);
1458         int d_in   = 0;
1459         for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1460           Node* pred = preds.current();
1461           if (in_bb(pred)) {
1462             d_in = MAX2(d_in, depth(pred));
1463           }
1464         }
1465         if (d_in + 1 != d_orig) {
1466           set_depth(n, d_in + 1);
1467           again = true;
1468         }
1469       }
1470     }
1471     ct++;
1472   } while (again);
1473 #ifndef PRODUCT
1474   if (TraceSuperWord && Verbose)
1475     tty->print_cr("compute_max_depth iterated: %d times", ct);
1476 #endif
1477 }
1478 
1479 //-------------------------compute_vector_element_type-----------------------
1480 // Compute necessary vector element type for expressions
1481 // This propagates backwards a narrower integer type when the
1482 // upper bits of the value are not needed.
1483 // Example:  char a,b,c;  a = b + c;
1484 // Normally the type of the add is integer, but for packed character
1485 // operations the type of the add needs to be char.
1486 void SuperWord::compute_vector_element_type() {
1487 #ifndef PRODUCT
1488   if (TraceSuperWord && Verbose)
1489     tty->print_cr("\ncompute_velt_type:");
1490 #endif
1491 
1492   // Initial type
1493   for (int i = 0; i < _block.length(); i++) {
1494     Node* n = _block.at(i);
1495     const Type* t  = n->is_Mem() ? Type::get_const_basic_type(n->as_Mem()->memory_type())
1496                                  : _igvn.type(n);
1497     const Type* vt = container_type(t);
1498     set_velt_type(n, vt);
1499   }
1500 
1501   // Propagate narrowed type backwards through operations
1502   // that don't depend on higher order bits
1503   for (int i = _block.length() - 1; i >= 0; i--) {
1504     Node* n = _block.at(i);
1505     // Only integer types need be examined
1506     if (n->bottom_type()->isa_int()) {
1507       uint start, end;
1508       vector_opd_range(n, &start, &end);
1509       const Type* vt = velt_type(n);
1510 
1511       for (uint j = start; j < end; j++) {
1512         Node* in  = n->in(j);
1513         // Don't propagate through a type conversion
1514         if (n->bottom_type() != in->bottom_type())
1515           continue;
1516         switch(in->Opcode()) {
1517         case Op_AddI:    case Op_AddL:
1518         case Op_SubI:    case Op_SubL:
1519         case Op_MulI:    case Op_MulL:
1520         case Op_AndI:    case Op_AndL:
1521         case Op_OrI:     case Op_OrL:
1522         case Op_XorI:    case Op_XorL:
1523         case Op_LShiftI: case Op_LShiftL:
1524         case Op_CMoveI:  case Op_CMoveL:
1525           if (in_bb(in)) {
1526             bool same_type = true;
1527             for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1528               Node *use = in->fast_out(k);
1529               if (!in_bb(use) || velt_type(use) != vt) {
1530                 same_type = false;
1531                 break;
1532               }
1533             }
1534             if (same_type) {
1535               set_velt_type(in, vt);
1536             }
1537           }
1538         }
1539       }
1540     }
1541   }
1542 #ifndef PRODUCT
1543   if (TraceSuperWord && Verbose) {
1544     for (int i = 0; i < _block.length(); i++) {
1545       Node* n = _block.at(i);
1546       velt_type(n)->dump();
1547       tty->print("\t");
1548       n->dump();
1549     }
1550   }
1551 #endif
1552 }
1553 
1554 //------------------------------memory_alignment---------------------------
1555 // Alignment within a vector memory reference
1556 int SuperWord::memory_alignment(MemNode* s, int iv_adjust_in_bytes) {
1557   SWPointer p(s, this);
1558   if (!p.valid()) {
1559     return bottom_align;
1560   }
1561   int offset  = p.offset_in_bytes();
1562   offset     += iv_adjust_in_bytes;
1563   int off_rem = offset % vector_width_in_bytes();
1564   int off_mod = off_rem >= 0 ? off_rem : off_rem + vector_width_in_bytes();
1565   return off_mod;
1566 }
1567 
1568 //---------------------------container_type---------------------------
1569 // Smallest type containing range of values
1570 const Type* SuperWord::container_type(const Type* t) {
1571   const Type* tp = t->make_ptr();
1572   if (tp && tp->isa_aryptr()) {
1573     t = tp->is_aryptr()->elem();
1574   }
1575   if (t->basic_type() == T_INT) {
1576     if (t->higher_equal(TypeInt::BOOL))  return TypeInt::BOOL;
1577     if (t->higher_equal(TypeInt::BYTE))  return TypeInt::BYTE;
1578     if (t->higher_equal(TypeInt::CHAR))  return TypeInt::CHAR;
1579     if (t->higher_equal(TypeInt::SHORT)) return TypeInt::SHORT;
1580     return TypeInt::INT;
1581   }
1582   return t;
1583 }
1584 
1585 //-------------------------vector_opd_range-----------------------
1586 // (Start, end] half-open range defining which operands are vector
1587 void SuperWord::vector_opd_range(Node* n, uint* start, uint* end) {
1588   switch (n->Opcode()) {
1589   case Op_LoadB:   case Op_LoadUS:
1590   case Op_LoadI:   case Op_LoadL:
1591   case Op_LoadF:   case Op_LoadD:
1592   case Op_LoadP:
1593     *start = 0;
1594     *end   = 0;
1595     return;
1596   case Op_StoreB:  case Op_StoreC:
1597   case Op_StoreI:  case Op_StoreL:
1598   case Op_StoreF:  case Op_StoreD:
1599   case Op_StoreP:
1600     *start = MemNode::ValueIn;
1601     *end   = *start + 1;
1602     return;
1603   case Op_LShiftI: case Op_LShiftL:
1604     *start = 1;
1605     *end   = 2;
1606     return;
1607   case Op_CMoveI:  case Op_CMoveL:  case Op_CMoveF:  case Op_CMoveD:
1608     *start = 2;
1609     *end   = n->req();
1610     return;
1611   }
1612   *start = 1;
1613   *end   = n->req(); // default is all operands
1614 }
1615 
1616 //------------------------------in_packset---------------------------
1617 // Are s1 and s2 in a pack pair and ordered as s1,s2?
1618 bool SuperWord::in_packset(Node* s1, Node* s2) {
1619   for (int i = 0; i < _packset.length(); i++) {
1620     Node_List* p = _packset.at(i);
1621     assert(p->size() == 2, "must be");
1622     if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
1623       return true;
1624     }
1625   }
1626   return false;
1627 }
1628 
1629 //------------------------------in_pack---------------------------
1630 // Is s in pack p?
1631 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
1632   for (uint i = 0; i < p->size(); i++) {
1633     if (p->at(i) == s) {
1634       return p;
1635     }
1636   }
1637   return NULL;
1638 }
1639 
1640 //------------------------------remove_pack_at---------------------------
1641 // Remove the pack at position pos in the packset
1642 void SuperWord::remove_pack_at(int pos) {
1643   Node_List* p = _packset.at(pos);
1644   for (uint i = 0; i < p->size(); i++) {
1645     Node* s = p->at(i);
1646     set_my_pack(s, NULL);
1647   }
1648   _packset.remove_at(pos);
1649 }
1650 
1651 //------------------------------executed_first---------------------------
1652 // Return the node executed first in pack p.  Uses the RPO block list
1653 // to determine order.
1654 Node* SuperWord::executed_first(Node_List* p) {
1655   Node* n = p->at(0);
1656   int n_rpo = bb_idx(n);
1657   for (uint i = 1; i < p->size(); i++) {
1658     Node* s = p->at(i);
1659     int s_rpo = bb_idx(s);
1660     if (s_rpo < n_rpo) {
1661       n = s;
1662       n_rpo = s_rpo;
1663     }
1664   }
1665   return n;
1666 }
1667 
1668 //------------------------------executed_last---------------------------
1669 // Return the node executed last in pack p.
1670 Node* SuperWord::executed_last(Node_List* p) {
1671   Node* n = p->at(0);
1672   int n_rpo = bb_idx(n);
1673   for (uint i = 1; i < p->size(); i++) {
1674     Node* s = p->at(i);
1675     int s_rpo = bb_idx(s);
1676     if (s_rpo > n_rpo) {
1677       n = s;
1678       n_rpo = s_rpo;
1679     }
1680   }
1681   return n;
1682 }
1683 
1684 //----------------------------align_initial_loop_index---------------------------
1685 // Adjust pre-loop limit so that in main loop, a load/store reference
1686 // to align_to_ref will be a position zero in the vector.
1687 //   (iv + k) mod vector_align == 0
1688 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
1689   CountedLoopNode *main_head = lp()->as_CountedLoop();
1690   assert(main_head->is_main_loop(), "");
1691   CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
1692   assert(pre_end != NULL, "");
1693   Node *pre_opaq1 = pre_end->limit();
1694   assert(pre_opaq1->Opcode() == Op_Opaque1, "");
1695   Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1696   Node *lim0 = pre_opaq->in(1);
1697 
1698   // Where we put new limit calculations
1699   Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1700 
1701   // Ensure the original loop limit is available from the
1702   // pre-loop Opaque1 node.
1703   Node *orig_limit = pre_opaq->original_loop_limit();
1704   assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
1705 
1706   SWPointer align_to_ref_p(align_to_ref, this);
1707 
1708   // Given:
1709   //     lim0 == original pre loop limit
1710   //     V == v_align (power of 2)
1711   //     invar == extra invariant piece of the address expression
1712   //     e == k [ +/- invar ]
1713   //
1714   // When reassociating expressions involving '%' the basic rules are:
1715   //     (a - b) % k == 0   =>  a % k == b % k
1716   // and:
1717   //     (a + b) % k == 0   =>  a % k == (k - b) % k
1718   //
1719   // For stride > 0 && scale > 0,
1720   //   Derive the new pre-loop limit "lim" such that the two constraints:
1721   //     (1) lim = lim0 + N           (where N is some positive integer < V)
1722   //     (2) (e + lim) % V == 0
1723   //   are true.
1724   //
1725   //   Substituting (1) into (2),
1726   //     (e + lim0 + N) % V == 0
1727   //   solve for N:
1728   //     N = (V - (e + lim0)) % V
1729   //   substitute back into (1), so that new limit
1730   //     lim = lim0 + (V - (e + lim0)) % V
1731   //
1732   // For stride > 0 && scale < 0
1733   //   Constraints:
1734   //     lim = lim0 + N
1735   //     (e - lim) % V == 0
1736   //   Solving for lim:
1737   //     (e - lim0 - N) % V == 0
1738   //     N = (e - lim0) % V
1739   //     lim = lim0 + (e - lim0) % V
1740   //
1741   // For stride < 0 && scale > 0
1742   //   Constraints:
1743   //     lim = lim0 - N
1744   //     (e + lim) % V == 0
1745   //   Solving for lim:
1746   //     (e + lim0 - N) % V == 0
1747   //     N = (e + lim0) % V
1748   //     lim = lim0 - (e + lim0) % V
1749   //
1750   // For stride < 0 && scale < 0
1751   //   Constraints:
1752   //     lim = lim0 - N
1753   //     (e - lim) % V == 0
1754   //   Solving for lim:
1755   //     (e - lim0 + N) % V == 0
1756   //     N = (V - (e - lim0)) % V
1757   //     lim = lim0 - (V - (e - lim0)) % V
1758 
1759   int stride   = iv_stride();
1760   int scale    = align_to_ref_p.scale_in_bytes();
1761   int elt_size = align_to_ref_p.memory_size();
1762   int v_align  = vector_width_in_bytes() / elt_size;
1763   int k        = align_to_ref_p.offset_in_bytes() / elt_size;
1764 
1765   Node *kn   = _igvn.intcon(k);
1766 
1767   Node *e = kn;
1768   if (align_to_ref_p.invar() != NULL) {
1769     // incorporate any extra invariant piece producing k +/- invar >>> log2(elt)
1770     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
1771     Node* aref     = new (_phase->C, 3) URShiftINode(align_to_ref_p.invar(), log2_elt);
1772     _phase->_igvn.register_new_node_with_optimizer(aref);
1773     _phase->set_ctrl(aref, pre_ctrl);
1774     if (align_to_ref_p.negate_invar()) {
1775       e = new (_phase->C, 3) SubINode(e, aref);
1776     } else {
1777       e = new (_phase->C, 3) AddINode(e, aref);
1778     }
1779     _phase->_igvn.register_new_node_with_optimizer(e);
1780     _phase->set_ctrl(e, pre_ctrl);
1781   }
1782 
1783   // compute e +/- lim0
1784   if (scale < 0) {
1785     e = new (_phase->C, 3) SubINode(e, lim0);
1786   } else {
1787     e = new (_phase->C, 3) AddINode(e, lim0);
1788   }
1789   _phase->_igvn.register_new_node_with_optimizer(e);
1790   _phase->set_ctrl(e, pre_ctrl);
1791 
1792   if (stride * scale > 0) {
1793     // compute V - (e +/- lim0)
1794     Node* va  = _igvn.intcon(v_align);
1795     e = new (_phase->C, 3) SubINode(va, e);
1796     _phase->_igvn.register_new_node_with_optimizer(e);
1797     _phase->set_ctrl(e, pre_ctrl);
1798   }
1799   // compute N = (exp) % V
1800   Node* va_msk = _igvn.intcon(v_align - 1);
1801   Node* N = new (_phase->C, 3) AndINode(e, va_msk);
1802   _phase->_igvn.register_new_node_with_optimizer(N);
1803   _phase->set_ctrl(N, pre_ctrl);
1804 
1805   //   substitute back into (1), so that new limit
1806   //     lim = lim0 + N
1807   Node* lim;
1808   if (stride < 0) {
1809     lim = new (_phase->C, 3) SubINode(lim0, N);
1810   } else {
1811     lim = new (_phase->C, 3) AddINode(lim0, N);
1812   }
1813   _phase->_igvn.register_new_node_with_optimizer(lim);
1814   _phase->set_ctrl(lim, pre_ctrl);
1815   Node* constrained =
1816     (stride > 0) ? (Node*) new (_phase->C,3) MinINode(lim, orig_limit)
1817                  : (Node*) new (_phase->C,3) MaxINode(lim, orig_limit);
1818   _phase->_igvn.register_new_node_with_optimizer(constrained);
1819   _phase->set_ctrl(constrained, pre_ctrl);
1820   _igvn.hash_delete(pre_opaq);
1821   pre_opaq->set_req(1, constrained);
1822 }
1823 
1824 //----------------------------get_pre_loop_end---------------------------
1825 // Find pre loop end from main loop.  Returns null if none.
1826 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
1827   Node *ctrl = cl->in(LoopNode::EntryControl);
1828   if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
1829   Node *iffm = ctrl->in(0);
1830   if (!iffm->is_If()) return NULL;
1831   Node *p_f = iffm->in(0);
1832   if (!p_f->is_IfFalse()) return NULL;
1833   if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
1834   CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1835   if (!pre_end->loopnode()->is_pre_loop()) return NULL;
1836   return pre_end;
1837 }
1838 
1839 
1840 //------------------------------init---------------------------
1841 void SuperWord::init() {
1842   _dg.init();
1843   _packset.clear();
1844   _disjoint_ptrs.clear();
1845   _block.clear();
1846   _data_entry.clear();
1847   _mem_slice_head.clear();
1848   _mem_slice_tail.clear();
1849   _node_info.clear();
1850   _align_to_ref = NULL;
1851   _lpt = NULL;
1852   _lp = NULL;
1853   _bb = NULL;
1854   _iv = NULL;
1855 }
1856 
1857 //------------------------------print_packset---------------------------
1858 void SuperWord::print_packset() {
1859 #ifndef PRODUCT
1860   tty->print_cr("packset");
1861   for (int i = 0; i < _packset.length(); i++) {
1862     tty->print_cr("Pack: %d", i);
1863     Node_List* p = _packset.at(i);
1864     print_pack(p);
1865   }
1866 #endif
1867 }
1868 
1869 //------------------------------print_pack---------------------------
1870 void SuperWord::print_pack(Node_List* p) {
1871   for (uint i = 0; i < p->size(); i++) {
1872     print_stmt(p->at(i));
1873   }
1874 }
1875 
1876 //------------------------------print_bb---------------------------
1877 void SuperWord::print_bb() {
1878 #ifndef PRODUCT
1879   tty->print_cr("\nBlock");
1880   for (int i = 0; i < _block.length(); i++) {
1881     Node* n = _block.at(i);
1882     tty->print("%d ", i);
1883     if (n) {
1884       n->dump();
1885     }
1886   }
1887 #endif
1888 }
1889 
1890 //------------------------------print_stmt---------------------------
1891 void SuperWord::print_stmt(Node* s) {
1892 #ifndef PRODUCT
1893   tty->print(" align: %d \t", alignment(s));
1894   s->dump();
1895 #endif
1896 }
1897 
1898 //------------------------------blank---------------------------
1899 char* SuperWord::blank(uint depth) {
1900   static char blanks[101];
1901   assert(depth < 101, "too deep");
1902   for (uint i = 0; i < depth; i++) blanks[i] = ' ';
1903   blanks[depth] = '\0';
1904   return blanks;
1905 }
1906 
1907 
1908 //==============================SWPointer===========================
1909 
1910 //----------------------------SWPointer------------------------
1911 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
1912   _mem(mem), _slp(slp),  _base(NULL),  _adr(NULL),
1913   _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
1914 
1915   Node* adr = mem->in(MemNode::Address);
1916   if (!adr->is_AddP()) {
1917     assert(!valid(), "too complex");
1918     return;
1919   }
1920   // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
1921   Node* base = adr->in(AddPNode::Base);
1922   for (int i = 0; i < 3; i++) {
1923     if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
1924       assert(!valid(), "too complex");
1925       return;
1926     }
1927     adr = adr->in(AddPNode::Address);
1928     if (base == adr || !adr->is_AddP()) {
1929       break; // stop looking at addp's
1930     }
1931   }
1932   _base = base;
1933   _adr  = adr;
1934   assert(valid(), "Usable");
1935 }
1936 
1937 // Following is used to create a temporary object during
1938 // the pattern match of an address expression.
1939 SWPointer::SWPointer(SWPointer* p) :
1940   _mem(p->_mem), _slp(p->_slp),  _base(NULL),  _adr(NULL),
1941   _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
1942 
1943 //------------------------scaled_iv_plus_offset--------------------
1944 // Match: k*iv + offset
1945 // where: k is a constant that maybe zero, and
1946 //        offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
1947 bool SWPointer::scaled_iv_plus_offset(Node* n) {
1948   if (scaled_iv(n)) {
1949     return true;
1950   }
1951   if (offset_plus_k(n)) {
1952     return true;
1953   }
1954   int opc = n->Opcode();
1955   if (opc == Op_AddI) {
1956     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
1957       return true;
1958     }
1959     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
1960       return true;
1961     }
1962   } else if (opc == Op_SubI) {
1963     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
1964       return true;
1965     }
1966     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
1967       _scale *= -1;
1968       return true;
1969     }
1970   }
1971   return false;
1972 }
1973 
1974 //----------------------------scaled_iv------------------------
1975 // Match: k*iv where k is a constant that's not zero
1976 bool SWPointer::scaled_iv(Node* n) {
1977   if (_scale != 0) {
1978     return false;  // already found a scale
1979   }
1980   if (n == iv()) {
1981     _scale = 1;
1982     return true;
1983   }
1984   int opc = n->Opcode();
1985   if (opc == Op_MulI) {
1986     if (n->in(1) == iv() && n->in(2)->is_Con()) {
1987       _scale = n->in(2)->get_int();
1988       return true;
1989     } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
1990       _scale = n->in(1)->get_int();
1991       return true;
1992     }
1993   } else if (opc == Op_LShiftI) {
1994     if (n->in(1) == iv() && n->in(2)->is_Con()) {
1995       _scale = 1 << n->in(2)->get_int();
1996       return true;
1997     }
1998   } else if (opc == Op_ConvI2L) {
1999     if (scaled_iv_plus_offset(n->in(1))) {
2000       return true;
2001     }
2002   } else if (opc == Op_LShiftL) {
2003     if (!has_iv() && _invar == NULL) {
2004       // Need to preserve the current _offset value, so
2005       // create a temporary object for this expression subtree.
2006       // Hacky, so should re-engineer the address pattern match.
2007       SWPointer tmp(this);
2008       if (tmp.scaled_iv_plus_offset(n->in(1))) {
2009         if (tmp._invar == NULL) {
2010           int mult = 1 << n->in(2)->get_int();
2011           _scale   = tmp._scale  * mult;
2012           _offset += tmp._offset * mult;
2013           return true;
2014         }
2015       }
2016     }
2017   }
2018   return false;
2019 }
2020 
2021 //----------------------------offset_plus_k------------------------
2022 // Match: offset is (k [+/- invariant])
2023 // where k maybe zero and invariant is optional, but not both.
2024 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2025   int opc = n->Opcode();
2026   if (opc == Op_ConI) {
2027     _offset += negate ? -(n->get_int()) : n->get_int();
2028     return true;
2029   } else if (opc == Op_ConL) {
2030     // Okay if value fits into an int
2031     const TypeLong* t = n->find_long_type();
2032     if (t->higher_equal(TypeLong::INT)) {
2033       jlong loff = n->get_long();
2034       jint  off  = (jint)loff;
2035       _offset += negate ? -off : loff;
2036       return true;
2037     }
2038     return false;
2039   }
2040   if (_invar != NULL) return false; // already have an invariant
2041   if (opc == Op_AddI) {
2042     if (n->in(2)->is_Con() && invariant(n->in(1))) {
2043       _negate_invar = negate;
2044       _invar = n->in(1);
2045       _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2046       return true;
2047     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2048       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2049       _negate_invar = negate;
2050       _invar = n->in(2);
2051       return true;
2052     }
2053   }
2054   if (opc == Op_SubI) {
2055     if (n->in(2)->is_Con() && invariant(n->in(1))) {
2056       _negate_invar = negate;
2057       _invar = n->in(1);
2058       _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2059       return true;
2060     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2061       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2062       _negate_invar = !negate;
2063       _invar = n->in(2);
2064       return true;
2065     }
2066   }
2067   if (invariant(n)) {
2068     _negate_invar = negate;
2069     _invar = n;
2070     return true;
2071   }
2072   return false;
2073 }
2074 
2075 //----------------------------print------------------------
2076 void SWPointer::print() {
2077 #ifndef PRODUCT
2078   tty->print("base: %d  adr: %d  scale: %d  offset: %d  invar: %c%d\n",
2079              _base != NULL ? _base->_idx : 0,
2080              _adr  != NULL ? _adr->_idx  : 0,
2081              _scale, _offset,
2082              _negate_invar?'-':'+',
2083              _invar != NULL ? _invar->_idx : 0);
2084 #endif
2085 }
2086 
2087 // ========================= OrderedPair =====================
2088 
2089 const OrderedPair OrderedPair::initial;
2090 
2091 // ========================= SWNodeInfo =====================
2092 
2093 const SWNodeInfo SWNodeInfo::initial;
2094 
2095 
2096 // ============================ DepGraph ===========================
2097 
2098 //------------------------------make_node---------------------------
2099 // Make a new dependence graph node for an ideal node.
2100 DepMem* DepGraph::make_node(Node* node) {
2101   DepMem* m = new (_arena) DepMem(node);
2102   if (node != NULL) {
2103     assert(_map.at_grow(node->_idx) == NULL, "one init only");
2104     _map.at_put_grow(node->_idx, m);
2105   }
2106   return m;
2107 }
2108 
2109 //------------------------------make_edge---------------------------
2110 // Make a new dependence graph edge from dpred -> dsucc
2111 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2112   DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2113   dpred->set_out_head(e);
2114   dsucc->set_in_head(e);
2115   return e;
2116 }
2117 
2118 // ========================== DepMem ========================
2119 
2120 //------------------------------in_cnt---------------------------
2121 int DepMem::in_cnt() {
2122   int ct = 0;
2123   for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2124   return ct;
2125 }
2126 
2127 //------------------------------out_cnt---------------------------
2128 int DepMem::out_cnt() {
2129   int ct = 0;
2130   for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2131   return ct;
2132 }
2133 
2134 //------------------------------print-----------------------------
2135 void DepMem::print() {
2136 #ifndef PRODUCT
2137   tty->print("  DepNode %d (", _node->_idx);
2138   for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2139     Node* pred = p->pred()->node();
2140     tty->print(" %d", pred != NULL ? pred->_idx : 0);
2141   }
2142   tty->print(") [");
2143   for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2144     Node* succ = s->succ()->node();
2145     tty->print(" %d", succ != NULL ? succ->_idx : 0);
2146   }
2147   tty->print_cr(" ]");
2148 #endif
2149 }
2150 
2151 // =========================== DepEdge =========================
2152 
2153 //------------------------------DepPreds---------------------------
2154 void DepEdge::print() {
2155 #ifndef PRODUCT
2156   tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2157 #endif
2158 }
2159 
2160 // =========================== DepPreds =========================
2161 // Iterator over predecessor edges in the dependence graph.
2162 
2163 //------------------------------DepPreds---------------------------
2164 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2165   _n = n;
2166   _done = false;
2167   if (_n->is_Store() || _n->is_Load()) {
2168     _next_idx = MemNode::Address;
2169     _end_idx  = n->req();
2170     _dep_next = dg.dep(_n)->in_head();
2171   } else if (_n->is_Mem()) {
2172     _next_idx = 0;
2173     _end_idx  = 0;
2174     _dep_next = dg.dep(_n)->in_head();
2175   } else {
2176     _next_idx = 1;
2177     _end_idx  = _n->req();
2178     _dep_next = NULL;
2179   }
2180   next();
2181 }
2182 
2183 //------------------------------next---------------------------
2184 void DepPreds::next() {
2185   if (_dep_next != NULL) {
2186     _current  = _dep_next->pred()->node();
2187     _dep_next = _dep_next->next_in();
2188   } else if (_next_idx < _end_idx) {
2189     _current  = _n->in(_next_idx++);
2190   } else {
2191     _done = true;
2192   }
2193 }
2194 
2195 // =========================== DepSuccs =========================
2196 // Iterator over successor edges in the dependence graph.
2197 
2198 //------------------------------DepSuccs---------------------------
2199 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2200   _n = n;
2201   _done = false;
2202   if (_n->is_Load()) {
2203     _next_idx = 0;
2204     _end_idx  = _n->outcnt();
2205     _dep_next = dg.dep(_n)->out_head();
2206   } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2207     _next_idx = 0;
2208     _end_idx  = 0;
2209     _dep_next = dg.dep(_n)->out_head();
2210   } else {
2211     _next_idx = 0;
2212     _end_idx  = _n->outcnt();
2213     _dep_next = NULL;
2214   }
2215   next();
2216 }
2217 
2218 //-------------------------------next---------------------------
2219 void DepSuccs::next() {
2220   if (_dep_next != NULL) {
2221     _current  = _dep_next->succ()->node();
2222     _dep_next = _dep_next->next_out();
2223   } else if (_next_idx < _end_idx) {
2224     _current  = _n->raw_out(_next_idx++);
2225   } else {
2226     _done = true;
2227   }
2228 }