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 {
 461           assert(out == prev || prev == NULL, "no branches off of store slice");
 462         }
 463       }
 464     }
 465     if (n == stop) break;
 466     preds.push(n);
 467     prev = n;
 468     n = n->in(MemNode::Memory);
 469   }
 470 }
 471 
 472 //------------------------------stmts_can_pack---------------------------
 473 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
 474 // s1 aligned at "align"
 475 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
 476 
 477   // Do not use superword for non-primitives
 478   if((s1->is_Mem() && !is_java_primitive(s1->as_Mem()->memory_type())) ||
 479      (s2->is_Mem() && !is_java_primitive(s2->as_Mem()->memory_type())))
 480     return false;
 481 
 482   if (isomorphic(s1, s2)) {
 483     if (independent(s1, s2)) {
 484       if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
 485         if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
 486           int s1_align = alignment(s1);
 487           int s2_align = alignment(s2);
 488           if (s1_align == top_align || s1_align == align) {
 489             if (s2_align == top_align || s2_align == align + data_size(s1)) {
 490               return true;
 491             }
 492           }
 493         }
 494       }
 495     }
 496   }
 497   return false;
 498 }
 499 
 500 //------------------------------exists_at---------------------------
 501 // Does s exist in a pack at position pos?
 502 bool SuperWord::exists_at(Node* s, uint pos) {
 503   for (int i = 0; i < _packset.length(); i++) {
 504     Node_List* p = _packset.at(i);
 505     if (p->at(pos) == s) {
 506       return true;
 507     }
 508   }
 509   return false;
 510 }
 511 
 512 //------------------------------are_adjacent_refs---------------------------
 513 // Is s1 immediately before s2 in memory?
 514 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
 515   if (!s1->is_Mem() || !s2->is_Mem()) return false;
 516   if (!in_bb(s1)    || !in_bb(s2))    return false;
 517   // FIXME - co_locate_pack fails on Stores in different mem-slices, so
 518   // only pack memops that are in the same alias set until that's fixed.
 519   if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
 520       _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
 521     return false;
 522   SWPointer p1(s1->as_Mem(), this);
 523   SWPointer p2(s2->as_Mem(), this);
 524   if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
 525   int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
 526   return diff == data_size(s1);
 527 }
 528 
 529 //------------------------------isomorphic---------------------------
 530 // Are s1 and s2 similar?
 531 bool SuperWord::isomorphic(Node* s1, Node* s2) {
 532   if (s1->Opcode() != s2->Opcode()) return false;
 533   if (s1->req() != s2->req()) return false;
 534   if (s1->in(0) != s2->in(0)) return false;
 535   if (velt_type(s1) != velt_type(s2)) return false;
 536   return true;
 537 }
 538 
 539 //------------------------------independent---------------------------
 540 // Is there no data path from s1 to s2 or s2 to s1?
 541 bool SuperWord::independent(Node* s1, Node* s2) {
 542   //  assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
 543   int d1 = depth(s1);
 544   int d2 = depth(s2);
 545   if (d1 == d2) return s1 != s2;
 546   Node* deep    = d1 > d2 ? s1 : s2;
 547   Node* shallow = d1 > d2 ? s2 : s1;
 548 
 549   visited_clear();
 550 
 551   return independent_path(shallow, deep);
 552 }
 553 
 554 //------------------------------independent_path------------------------------
 555 // Helper for independent
 556 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
 557   if (dp >= 1000) return false; // stop deep recursion
 558   visited_set(deep);
 559   int shal_depth = depth(shallow);
 560   assert(shal_depth <= depth(deep), "must be");
 561   for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
 562     Node* pred = preds.current();
 563     if (in_bb(pred) && !visited_test(pred)) {
 564       if (shallow == pred) {
 565         return false;
 566       }
 567       if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
 568         return false;
 569       }
 570     }
 571   }
 572   return true;
 573 }
 574 
 575 //------------------------------set_alignment---------------------------
 576 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
 577   set_alignment(s1, align);
 578   set_alignment(s2, align + data_size(s1));
 579 }
 580 
 581 //------------------------------data_size---------------------------
 582 int SuperWord::data_size(Node* s) {
 583   const Type* t = velt_type(s);
 584   BasicType  bt = t->array_element_basic_type();
 585   int bsize = type2aelembytes(bt);
 586   assert(bsize != 0, "valid size");
 587   return bsize;
 588 }
 589 
 590 //------------------------------extend_packlist---------------------------
 591 // Extend packset by following use->def and def->use links from pack members.
 592 void SuperWord::extend_packlist() {
 593   bool changed;
 594   do {
 595     changed = false;
 596     for (int i = 0; i < _packset.length(); i++) {
 597       Node_List* p = _packset.at(i);
 598       changed |= follow_use_defs(p);
 599       changed |= follow_def_uses(p);
 600     }
 601   } while (changed);
 602 
 603 #ifndef PRODUCT
 604   if (TraceSuperWord) {
 605     tty->print_cr("\nAfter extend_packlist");
 606     print_packset();
 607   }
 608 #endif
 609 }
 610 
 611 //------------------------------follow_use_defs---------------------------
 612 // Extend the packset by visiting operand definitions of nodes in pack p
 613 bool SuperWord::follow_use_defs(Node_List* p) {
 614   Node* s1 = p->at(0);
 615   Node* s2 = p->at(1);
 616   assert(p->size() == 2, "just checking");
 617   assert(s1->req() == s2->req(), "just checking");
 618   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
 619 
 620   if (s1->is_Load()) return false;
 621 
 622   int align = alignment(s1);
 623   bool changed = false;
 624   int start = s1->is_Store() ? MemNode::ValueIn   : 1;
 625   int end   = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
 626   for (int j = start; j < end; j++) {
 627     Node* t1 = s1->in(j);
 628     Node* t2 = s2->in(j);
 629     if (!in_bb(t1) || !in_bb(t2))
 630       continue;
 631     if (stmts_can_pack(t1, t2, align)) {
 632       if (est_savings(t1, t2) >= 0) {
 633         Node_List* pair = new Node_List();
 634         pair->push(t1);
 635         pair->push(t2);
 636         _packset.append(pair);
 637         set_alignment(t1, t2, align);
 638         changed = true;
 639       }
 640     }
 641   }
 642   return changed;
 643 }
 644 
 645 //------------------------------follow_def_uses---------------------------
 646 // Extend the packset by visiting uses of nodes in pack p
 647 bool SuperWord::follow_def_uses(Node_List* p) {
 648   bool changed = false;
 649   Node* s1 = p->at(0);
 650   Node* s2 = p->at(1);
 651   assert(p->size() == 2, "just checking");
 652   assert(s1->req() == s2->req(), "just checking");
 653   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
 654 
 655   if (s1->is_Store()) return false;
 656 
 657   int align = alignment(s1);
 658   int savings = -1;
 659   Node* u1 = NULL;
 660   Node* u2 = NULL;
 661   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
 662     Node* t1 = s1->fast_out(i);
 663     if (!in_bb(t1)) continue;
 664     for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
 665       Node* t2 = s2->fast_out(j);
 666       if (!in_bb(t2)) continue;
 667       if (!opnd_positions_match(s1, t1, s2, t2))
 668         continue;
 669       if (stmts_can_pack(t1, t2, align)) {
 670         int my_savings = est_savings(t1, t2);
 671         if (my_savings > savings) {
 672           savings = my_savings;
 673           u1 = t1;
 674           u2 = t2;
 675         }
 676       }
 677     }
 678   }
 679   if (savings >= 0) {
 680     Node_List* pair = new Node_List();
 681     pair->push(u1);
 682     pair->push(u2);
 683     _packset.append(pair);
 684     set_alignment(u1, u2, align);
 685     changed = true;
 686   }
 687   return changed;
 688 }
 689 
 690 //---------------------------opnd_positions_match-------------------------
 691 // Is the use of d1 in u1 at the same operand position as d2 in u2?
 692 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
 693   uint ct = u1->req();
 694   if (ct != u2->req()) return false;
 695   uint i1 = 0;
 696   uint i2 = 0;
 697   do {
 698     for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
 699     for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
 700     if (i1 != i2) {
 701       return false;
 702     }
 703   } while (i1 < ct);
 704   return true;
 705 }
 706 
 707 //------------------------------est_savings---------------------------
 708 // Estimate the savings from executing s1 and s2 as a pack
 709 int SuperWord::est_savings(Node* s1, Node* s2) {
 710   int save = 2 - 1; // 2 operations per instruction in packed form
 711 
 712   // inputs
 713   for (uint i = 1; i < s1->req(); i++) {
 714     Node* x1 = s1->in(i);
 715     Node* x2 = s2->in(i);
 716     if (x1 != x2) {
 717       if (are_adjacent_refs(x1, x2)) {
 718         save += adjacent_profit(x1, x2);
 719       } else if (!in_packset(x1, x2)) {
 720         save -= pack_cost(2);
 721       } else {
 722         save += unpack_cost(2);
 723       }
 724     }
 725   }
 726 
 727   // uses of result
 728   uint ct = 0;
 729   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
 730     Node* s1_use = s1->fast_out(i);
 731     for (int j = 0; j < _packset.length(); j++) {
 732       Node_List* p = _packset.at(j);
 733       if (p->at(0) == s1_use) {
 734         for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
 735           Node* s2_use = s2->fast_out(k);
 736           if (p->at(p->size()-1) == s2_use) {
 737             ct++;
 738             if (are_adjacent_refs(s1_use, s2_use)) {
 739               save += adjacent_profit(s1_use, s2_use);
 740             }
 741           }
 742         }
 743       }
 744     }
 745   }
 746 
 747   if (ct < s1->outcnt()) save += unpack_cost(1);
 748   if (ct < s2->outcnt()) save += unpack_cost(1);
 749 
 750   return save;
 751 }
 752 
 753 //------------------------------costs---------------------------
 754 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
 755 int SuperWord::pack_cost(int ct)   { return ct; }
 756 int SuperWord::unpack_cost(int ct) { return ct; }
 757 
 758 //------------------------------combine_packs---------------------------
 759 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
 760 void SuperWord::combine_packs() {
 761   bool changed;
 762   do {
 763     changed = false;
 764     for (int i = 0; i < _packset.length(); i++) {
 765       Node_List* p1 = _packset.at(i);
 766       if (p1 == NULL) continue;
 767       for (int j = 0; j < _packset.length(); j++) {
 768         Node_List* p2 = _packset.at(j);
 769         if (p2 == NULL) continue;
 770         if (p1->at(p1->size()-1) == p2->at(0)) {
 771           for (uint k = 1; k < p2->size(); k++) {
 772             p1->push(p2->at(k));
 773           }
 774           _packset.at_put(j, NULL);
 775           changed = true;
 776         }
 777       }
 778     }
 779   } while (changed);
 780 
 781   for (int i = _packset.length() - 1; i >= 0; i--) {
 782     Node_List* p1 = _packset.at(i);
 783     if (p1 == NULL) {
 784       _packset.remove_at(i);
 785     }
 786   }
 787 
 788 #ifndef PRODUCT
 789   if (TraceSuperWord) {
 790     tty->print_cr("\nAfter combine_packs");
 791     print_packset();
 792   }
 793 #endif
 794 }
 795 
 796 //-----------------------------construct_my_pack_map--------------------------
 797 // Construct the map from nodes to packs.  Only valid after the
 798 // point where a node is only in one pack (after combine_packs).
 799 void SuperWord::construct_my_pack_map() {
 800   Node_List* rslt = NULL;
 801   for (int i = 0; i < _packset.length(); i++) {
 802     Node_List* p = _packset.at(i);
 803     for (uint j = 0; j < p->size(); j++) {
 804       Node* s = p->at(j);
 805       assert(my_pack(s) == NULL, "only in one pack");
 806       set_my_pack(s, p);
 807     }
 808   }
 809 }
 810 
 811 //------------------------------filter_packs---------------------------
 812 // Remove packs that are not implemented or not profitable.
 813 void SuperWord::filter_packs() {
 814 
 815   // Remove packs that are not implemented
 816   for (int i = _packset.length() - 1; i >= 0; i--) {
 817     Node_List* pk = _packset.at(i);
 818     bool impl = implemented(pk);
 819     if (!impl) {
 820 #ifndef PRODUCT
 821       if (TraceSuperWord && Verbose) {
 822         tty->print_cr("Unimplemented");
 823         pk->at(0)->dump();
 824       }
 825 #endif
 826       remove_pack_at(i);
 827     }
 828   }
 829 
 830   // Remove packs that are not profitable
 831   bool changed;
 832   do {
 833     changed = false;
 834     for (int i = _packset.length() - 1; i >= 0; i--) {
 835       Node_List* pk = _packset.at(i);
 836       bool prof = profitable(pk);
 837       if (!prof) {
 838 #ifndef PRODUCT
 839         if (TraceSuperWord && Verbose) {
 840           tty->print_cr("Unprofitable");
 841           pk->at(0)->dump();
 842         }
 843 #endif
 844         remove_pack_at(i);
 845         changed = true;
 846       }
 847     }
 848   } while (changed);
 849 
 850 #ifndef PRODUCT
 851   if (TraceSuperWord) {
 852     tty->print_cr("\nAfter filter_packs");
 853     print_packset();
 854     tty->cr();
 855   }
 856 #endif
 857 }
 858 
 859 //------------------------------implemented---------------------------
 860 // Can code be generated for pack p?
 861 bool SuperWord::implemented(Node_List* p) {
 862   Node* p0 = p->at(0);
 863   int vopc = VectorNode::opcode(p0->Opcode(), p->size(), velt_type(p0));
 864   return vopc > 0 && Matcher::has_match_rule(vopc);
 865 }
 866 
 867 //------------------------------profitable---------------------------
 868 // For pack p, are all operands and all uses (with in the block) vector?
 869 bool SuperWord::profitable(Node_List* p) {
 870   Node* p0 = p->at(0);
 871   uint start, end;
 872   vector_opd_range(p0, &start, &end);
 873 
 874   // Return false if some input is not vector and inside block
 875   for (uint i = start; i < end; i++) {
 876     if (!is_vector_use(p0, i)) {
 877       // For now, return false if not scalar promotion case (inputs are the same.)
 878       // Later, implement PackNode and allow differing, non-vector inputs
 879       // (maybe just the ones from outside the block.)
 880       Node* p0_def = p0->in(i);
 881       for (uint j = 1; j < p->size(); j++) {
 882         Node* use = p->at(j);
 883         Node* def = use->in(i);
 884         if (p0_def != def)
 885           return false;
 886       }
 887     }
 888   }
 889   if (!p0->is_Store()) {
 890     // For now, return false if not all uses are vector.
 891     // Later, implement ExtractNode and allow non-vector uses (maybe
 892     // just the ones outside the block.)
 893     for (uint i = 0; i < p->size(); i++) {
 894       Node* def = p->at(i);
 895       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
 896         Node* use = def->fast_out(j);
 897         for (uint k = 0; k < use->req(); k++) {
 898           Node* n = use->in(k);
 899           if (def == n) {
 900             if (!is_vector_use(use, k)) {
 901               return false;
 902             }
 903           }
 904         }
 905       }
 906     }
 907   }
 908   return true;
 909 }
 910 
 911 //------------------------------schedule---------------------------
 912 // Adjust the memory graph for the packed operations
 913 void SuperWord::schedule() {
 914 
 915   // Co-locate in the memory graph the members of each memory pack
 916   for (int i = 0; i < _packset.length(); i++) {
 917     co_locate_pack(_packset.at(i));
 918   }
 919 }
 920 
 921 //-------------------------------remove_and_insert-------------------
 922 //remove "current" from its current position in the memory graph and insert
 923 //it after the appropriate insertion point (lip or uip)
 924 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
 925                                   Node *uip, Unique_Node_List &sched_before) {
 926   Node* my_mem = current->in(MemNode::Memory);
 927   _igvn.hash_delete(current);
 928   _igvn.hash_delete(my_mem);
 929 
 930   //remove current_store from its current position in the memmory graph
 931   for (DUIterator i = current->outs(); current->has_out(i); i++) {
 932     Node* use = current->out(i);
 933     if (use->is_Mem()) {
 934       assert(use->in(MemNode::Memory) == current, "must be");
 935       _igvn.hash_delete(use);
 936       if (use == prev) { // connect prev to my_mem
 937         use->set_req(MemNode::Memory, my_mem);
 938       } else if (sched_before.member(use)) {
 939         _igvn.hash_delete(uip);
 940         use->set_req(MemNode::Memory, uip);
 941       } else {
 942         _igvn.hash_delete(lip);
 943         use->set_req(MemNode::Memory, lip);
 944       }
 945       _igvn._worklist.push(use);
 946       --i; //deleted this edge; rescan position
 947     }
 948   }
 949 
 950   bool sched_up = sched_before.member(current);
 951   Node *insert_pt =  sched_up ?  uip : lip;
 952   _igvn.hash_delete(insert_pt);
 953 
 954   // all uses of insert_pt's memory state should use current's instead
 955   for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
 956     Node* use = insert_pt->out(i);
 957     if (use->is_Mem()) {
 958       assert(use->in(MemNode::Memory) == insert_pt, "must be");
 959       _igvn.hash_delete(use);
 960       use->set_req(MemNode::Memory, current);
 961       _igvn._worklist.push(use);
 962       --i; //deleted this edge; rescan position
 963     } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
 964       uint pos; //lip (lower insert point) must be the last one in the memory slice
 965       _igvn.hash_delete(use);
 966       for (pos=1; pos < use->req(); pos++) {
 967         if (use->in(pos) == insert_pt) break;
 968       }
 969       use->set_req(pos, current);
 970       _igvn._worklist.push(use);
 971       --i;
 972     }
 973   }
 974 
 975   //connect current to insert_pt
 976   current->set_req(MemNode::Memory, insert_pt);
 977   _igvn._worklist.push(current);
 978 }
 979 
 980 //------------------------------co_locate_pack----------------------------------
 981 // To schedule a store pack, we need to move any sandwiched memory ops either before
 982 // or after the pack, based upon dependence information:
 983 // (1) If any store in the pack depends on the sandwiched memory op, the
 984 //     sandwiched memory op must be scheduled BEFORE the pack;
 985 // (2) If a sandwiched memory op depends on any store in the pack, the
 986 //     sandwiched memory op must be scheduled AFTER the pack;
 987 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
 988 //     memory op (say memB), memB must be scheduled before memA. So, if memA is
 989 //     scheduled before the pack, memB must also be scheduled before the pack;
 990 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
 991 //     schedule this store AFTER the pack
 992 // (5) We know there is no dependence cycle, so there in no other case;
 993 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
 994 //
 995 // To schedule a load pack, we use the memory state of either the first or the last load in
 996 // the pack, based on the dependence constraint.
 997 void SuperWord::co_locate_pack(Node_List* pk) {
 998   if (pk->at(0)->is_Store()) {
 999     MemNode* first     = executed_first(pk)->as_Mem();
1000     MemNode* last      = executed_last(pk)->as_Mem();
1001     Unique_Node_List schedule_before_pack;
1002     Unique_Node_List memops;
1003 
1004     MemNode* current   = last->in(MemNode::Memory)->as_Mem();
1005     MemNode* previous  = last;
1006     while (true) {
1007       assert(in_bb(current), "stay in block");
1008       memops.push(previous);
1009       for (DUIterator i = current->outs(); current->has_out(i); i++) {
1010         Node* use = current->out(i);
1011         if (use->is_Mem() && use != previous)
1012           memops.push(use);
1013       }
1014       if(current == first) break;
1015       previous = current;
1016       current  = current->in(MemNode::Memory)->as_Mem();
1017     }
1018 
1019     // determine which memory operations should be scheduled before the pack
1020     for (uint i = 1; i < memops.size(); i++) {
1021       Node *s1 = memops.at(i);
1022       if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1023         for (uint j = 0; j< i; j++) {
1024           Node *s2 = memops.at(j);
1025           if (!independent(s1, s2)) {
1026             if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1027               schedule_before_pack.push(s1); //s1 must be scheduled before
1028               Node_List* mem_pk = my_pack(s1);
1029               if (mem_pk != NULL) {
1030                 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1031                   Node* s = mem_pk->at(ii); // follow partner
1032                   if (memops.member(s) && !schedule_before_pack.member(s))
1033                     schedule_before_pack.push(s);
1034                 }
1035               }
1036             }
1037           }
1038         }
1039       }
1040     }
1041 
1042     MemNode* lower_insert_pt = last;
1043     Node*    upper_insert_pt = first->in(MemNode::Memory);
1044     previous                 = last; //previous store in pk
1045     current                  = last->in(MemNode::Memory)->as_Mem();
1046 
1047     //start scheduling from "last" to "first"
1048     while (true) {
1049       assert(in_bb(current), "stay in block");
1050       assert(in_pack(previous, pk), "previous stays in pack");
1051       Node* my_mem = current->in(MemNode::Memory);
1052 
1053       if (in_pack(current, pk)) {
1054         // Forward users of my memory state (except "previous) to my input memory state
1055         _igvn.hash_delete(current);
1056         for (DUIterator i = current->outs(); current->has_out(i); i++) {
1057           Node* use = current->out(i);
1058           if (use->is_Mem() && use != previous) {
1059             assert(use->in(MemNode::Memory) == current, "must be");
1060             _igvn.hash_delete(use);
1061             if (schedule_before_pack.member(use)) {
1062               _igvn.hash_delete(upper_insert_pt);
1063               use->set_req(MemNode::Memory, upper_insert_pt);
1064             } else {
1065               _igvn.hash_delete(lower_insert_pt);
1066               use->set_req(MemNode::Memory, lower_insert_pt);
1067             }
1068             _igvn._worklist.push(use);
1069             --i; // deleted this edge; rescan position
1070           }
1071         }
1072         previous = current;
1073       } else { // !in_pack(current, pk) ==> a sandwiched store
1074         remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1075       }
1076 
1077       if (current == first) break;
1078       current = my_mem->as_Mem();
1079     } // end while
1080   } else if (pk->at(0)->is_Load()) { //load
1081     // all loads in the pack should have the same memory state. By default,
1082     // we use the memory state of the last load. However, if any load could
1083     // not be moved down due to the dependence constraint, we use the memory
1084     // state of the first load.
1085     Node* last_mem  = executed_last(pk)->in(MemNode::Memory);
1086     Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1087     bool schedule_last = true;
1088     for (uint i = 0; i < pk->size(); i++) {
1089       Node* ld = pk->at(i);
1090       for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1091            current=current->in(MemNode::Memory)) {
1092         assert(current != first_mem, "corrupted memory graph");
1093         if(current->is_Mem() && !independent(current, ld)){
1094           schedule_last = false; // a later store depends on this load
1095           break;
1096         }
1097       }
1098     }
1099 
1100     Node* mem_input = schedule_last ? last_mem : first_mem;
1101     _igvn.hash_delete(mem_input);
1102     // Give each load the same memory state
1103     for (uint i = 0; i < pk->size(); i++) {
1104       LoadNode* ld = pk->at(i)->as_Load();
1105       _igvn.hash_delete(ld);
1106       ld->set_req(MemNode::Memory, mem_input);
1107       _igvn._worklist.push(ld);
1108     }
1109   }
1110 }
1111 
1112 //------------------------------output---------------------------
1113 // Convert packs into vector node operations
1114 void SuperWord::output() {
1115   if (_packset.length() == 0) return;
1116 
1117   // MUST ENSURE main loop's initial value is properly aligned:
1118   //  (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1119 
1120   align_initial_loop_index(align_to_ref());
1121 
1122   // Insert extract (unpack) operations for scalar uses
1123   for (int i = 0; i < _packset.length(); i++) {
1124     insert_extracts(_packset.at(i));
1125   }
1126 
1127   for (int i = 0; i < _block.length(); i++) {
1128     Node* n = _block.at(i);
1129     Node_List* p = my_pack(n);
1130     if (p && n == executed_last(p)) {
1131       uint vlen = p->size();
1132       Node* vn = NULL;
1133       Node* low_adr = p->at(0);
1134       Node* first   = executed_first(p);
1135       if (n->is_Load()) {
1136         int   opc = n->Opcode();
1137         Node* ctl = n->in(MemNode::Control);
1138         Node* mem = first->in(MemNode::Memory);
1139         Node* adr = low_adr->in(MemNode::Address);
1140         const TypePtr* atyp = n->adr_type();
1141         vn = VectorLoadNode::make(_phase->C, opc, ctl, mem, adr, atyp, vlen);
1142 
1143       } else if (n->is_Store()) {
1144         // Promote value to be stored to vector
1145         VectorNode* val = vector_opd(p, MemNode::ValueIn);
1146 
1147         int   opc = n->Opcode();
1148         Node* ctl = n->in(MemNode::Control);
1149         Node* mem = first->in(MemNode::Memory);
1150         Node* adr = low_adr->in(MemNode::Address);
1151         const TypePtr* atyp = n->adr_type();
1152         vn = VectorStoreNode::make(_phase->C, opc, ctl, mem, adr, atyp, val, vlen);
1153 
1154       } else if (n->req() == 3) {
1155         // Promote operands to vector
1156         Node* in1 = vector_opd(p, 1);
1157         Node* in2 = vector_opd(p, 2);
1158         vn = VectorNode::make(_phase->C, n->Opcode(), in1, in2, vlen, velt_type(n));
1159 
1160       } else {
1161         ShouldNotReachHere();
1162       }
1163 
1164       _phase->_igvn.register_new_node_with_optimizer(vn);
1165       _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1166       for (uint j = 0; j < p->size(); j++) {
1167         Node* pm = p->at(j);
1168         _igvn.hash_delete(pm);
1169         _igvn.subsume_node(pm, vn);
1170       }
1171       _igvn._worklist.push(vn);
1172     }
1173   }
1174 }
1175 
1176 //------------------------------vector_opd---------------------------
1177 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1178 VectorNode* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1179   Node* p0 = p->at(0);
1180   uint vlen = p->size();
1181   Node* opd = p0->in(opd_idx);
1182 
1183   bool same_opd = true;
1184   for (uint i = 1; i < vlen; i++) {
1185     Node* pi = p->at(i);
1186     Node* in = pi->in(opd_idx);
1187     if (opd != in) {
1188       same_opd = false;
1189       break;
1190     }
1191   }
1192 
1193   if (same_opd) {
1194     if (opd->is_Vector()) {
1195       return (VectorNode*)opd; // input is matching vector
1196     }
1197     // Convert scalar input to vector. Use p0's type because it's container
1198     // maybe smaller than the operand's container.
1199     const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd);
1200     const Type* p0_t  = velt_type(p0);
1201     if (p0_t->higher_equal(opd_t)) opd_t = p0_t;
1202     VectorNode* vn    = VectorNode::scalar2vector(_phase->C, opd, vlen, opd_t);
1203 
1204     _phase->_igvn.register_new_node_with_optimizer(vn);
1205     _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1206     return vn;
1207   }
1208 
1209   // Insert pack operation
1210   const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd);
1211   PackNode* pk = PackNode::make(_phase->C, opd, opd_t);
1212 
1213   for (uint i = 1; i < vlen; i++) {
1214     Node* pi = p->at(i);
1215     Node* in = pi->in(opd_idx);
1216     assert(my_pack(in) == NULL, "Should already have been unpacked");
1217     assert(opd_t == velt_type(!in_bb(in) ? pi : in), "all same type");
1218     pk->add_opd(in);
1219   }
1220   _phase->_igvn.register_new_node_with_optimizer(pk);
1221   _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1222   return pk;
1223 }
1224 
1225 //------------------------------insert_extracts---------------------------
1226 // If a use of pack p is not a vector use, then replace the
1227 // use with an extract operation.
1228 void SuperWord::insert_extracts(Node_List* p) {
1229   if (p->at(0)->is_Store()) return;
1230   assert(_n_idx_list.is_empty(), "empty (node,index) list");
1231 
1232   // Inspect each use of each pack member.  For each use that is
1233   // not a vector use, replace the use with an extract operation.
1234 
1235   for (uint i = 0; i < p->size(); i++) {
1236     Node* def = p->at(i);
1237     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1238       Node* use = def->fast_out(j);
1239       for (uint k = 0; k < use->req(); k++) {
1240         Node* n = use->in(k);
1241         if (def == n) {
1242           if (!is_vector_use(use, k)) {
1243             _n_idx_list.push(use, k);
1244           }
1245         }
1246       }
1247     }
1248   }
1249 
1250   while (_n_idx_list.is_nonempty()) {
1251     Node* use = _n_idx_list.node();
1252     int   idx = _n_idx_list.index();
1253     _n_idx_list.pop();
1254     Node* def = use->in(idx);
1255 
1256     // Insert extract operation
1257     _igvn.hash_delete(def);
1258     _igvn.hash_delete(use);
1259     int def_pos = alignment(def) / data_size(def);
1260     const Type* def_t = velt_type(def);
1261 
1262     Node* ex = ExtractNode::make(_phase->C, def, def_pos, def_t);
1263     _phase->_igvn.register_new_node_with_optimizer(ex);
1264     _phase->set_ctrl(ex, _phase->get_ctrl(def));
1265     use->set_req(idx, ex);
1266     _igvn._worklist.push(def);
1267     _igvn._worklist.push(use);
1268 
1269     bb_insert_after(ex, bb_idx(def));
1270     set_velt_type(ex, def_t);
1271   }
1272 }
1273 
1274 //------------------------------is_vector_use---------------------------
1275 // Is use->in(u_idx) a vector use?
1276 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1277   Node_List* u_pk = my_pack(use);
1278   if (u_pk == NULL) return false;
1279   Node* def = use->in(u_idx);
1280   Node_List* d_pk = my_pack(def);
1281   if (d_pk == NULL) {
1282     // check for scalar promotion
1283     Node* n = u_pk->at(0)->in(u_idx);
1284     for (uint i = 1; i < u_pk->size(); i++) {
1285       if (u_pk->at(i)->in(u_idx) != n) return false;
1286     }
1287     return true;
1288   }
1289   if (u_pk->size() != d_pk->size())
1290     return false;
1291   for (uint i = 0; i < u_pk->size(); i++) {
1292     Node* ui = u_pk->at(i);
1293     Node* di = d_pk->at(i);
1294     if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1295       return false;
1296   }
1297   return true;
1298 }
1299 
1300 //------------------------------construct_bb---------------------------
1301 // Construct reverse postorder list of block members
1302 void SuperWord::construct_bb() {
1303   Node* entry = bb();
1304 
1305   assert(_stk.length() == 0,            "stk is empty");
1306   assert(_block.length() == 0,          "block is empty");
1307   assert(_data_entry.length() == 0,     "data_entry is empty");
1308   assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1309   assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1310 
1311   // Find non-control nodes with no inputs from within block,
1312   // create a temporary map from node _idx to bb_idx for use
1313   // by the visited and post_visited sets,
1314   // and count number of nodes in block.
1315   int bb_ct = 0;
1316   for (uint i = 0; i < lpt()->_body.size(); i++ ) {
1317     Node *n = lpt()->_body.at(i);
1318     set_bb_idx(n, i); // Create a temporary map
1319     if (in_bb(n)) {
1320       bb_ct++;
1321       if (!n->is_CFG()) {
1322         bool found = false;
1323         for (uint j = 0; j < n->req(); j++) {
1324           Node* def = n->in(j);
1325           if (def && in_bb(def)) {
1326             found = true;
1327             break;
1328           }
1329         }
1330         if (!found) {
1331           assert(n != entry, "can't be entry");
1332           _data_entry.push(n);
1333         }
1334       }
1335     }
1336   }
1337 
1338   // Find memory slices (head and tail)
1339   for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1340     Node *n = lp()->fast_out(i);
1341     if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1342       Node* n_tail  = n->in(LoopNode::LoopBackControl);
1343       if (n_tail != n->in(LoopNode::EntryControl)) {
1344         _mem_slice_head.push(n);
1345         _mem_slice_tail.push(n_tail);
1346       }
1347     }
1348   }
1349 
1350   // Create an RPO list of nodes in block
1351 
1352   visited_clear();
1353   post_visited_clear();
1354 
1355   // Push all non-control nodes with no inputs from within block, then control entry
1356   for (int j = 0; j < _data_entry.length(); j++) {
1357     Node* n = _data_entry.at(j);
1358     visited_set(n);
1359     _stk.push(n);
1360   }
1361   visited_set(entry);
1362   _stk.push(entry);
1363 
1364   // Do a depth first walk over out edges
1365   int rpo_idx = bb_ct - 1;
1366   int size;
1367   while ((size = _stk.length()) > 0) {
1368     Node* n = _stk.top(); // Leave node on stack
1369     if (!visited_test_set(n)) {
1370       // forward arc in graph
1371     } else if (!post_visited_test(n)) {
1372       // cross or back arc
1373       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1374         Node *use = n->fast_out(i);
1375         if (in_bb(use) && !visited_test(use) &&
1376             // Don't go around backedge
1377             (!use->is_Phi() || n == entry)) {
1378           _stk.push(use);
1379         }
1380       }
1381       if (_stk.length() == size) {
1382         // There were no additional uses, post visit node now
1383         _stk.pop(); // Remove node from stack
1384         assert(rpo_idx >= 0, "");
1385         _block.at_put_grow(rpo_idx, n);
1386         rpo_idx--;
1387         post_visited_set(n);
1388         assert(rpo_idx >= 0 || _stk.is_empty(), "");
1389       }
1390     } else {
1391       _stk.pop(); // Remove post-visited node from stack
1392     }
1393   }
1394 
1395   // Create real map of block indices for nodes
1396   for (int j = 0; j < _block.length(); j++) {
1397     Node* n = _block.at(j);
1398     set_bb_idx(n, j);
1399   }
1400 
1401   initialize_bb(); // Ensure extra info is allocated.
1402 
1403 #ifndef PRODUCT
1404   if (TraceSuperWord) {
1405     print_bb();
1406     tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1407     for (int m = 0; m < _data_entry.length(); m++) {
1408       tty->print("%3d ", m);
1409       _data_entry.at(m)->dump();
1410     }
1411     tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1412     for (int m = 0; m < _mem_slice_head.length(); m++) {
1413       tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1414       tty->print("    ");    _mem_slice_tail.at(m)->dump();
1415     }
1416   }
1417 #endif
1418   assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1419 }
1420 
1421 //------------------------------initialize_bb---------------------------
1422 // Initialize per node info
1423 void SuperWord::initialize_bb() {
1424   Node* last = _block.at(_block.length() - 1);
1425   grow_node_info(bb_idx(last));
1426 }
1427 
1428 //------------------------------bb_insert_after---------------------------
1429 // Insert n into block after pos
1430 void SuperWord::bb_insert_after(Node* n, int pos) {
1431   int n_pos = pos + 1;
1432   // Make room
1433   for (int i = _block.length() - 1; i >= n_pos; i--) {
1434     _block.at_put_grow(i+1, _block.at(i));
1435   }
1436   for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1437     _node_info.at_put_grow(j+1, _node_info.at(j));
1438   }
1439   // Set value
1440   _block.at_put_grow(n_pos, n);
1441   _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1442   // Adjust map from node->_idx to _block index
1443   for (int i = n_pos; i < _block.length(); i++) {
1444     set_bb_idx(_block.at(i), i);
1445   }
1446 }
1447 
1448 //------------------------------compute_max_depth---------------------------
1449 // Compute max depth for expressions from beginning of block
1450 // Use to prune search paths during test for independence.
1451 void SuperWord::compute_max_depth() {
1452   int ct = 0;
1453   bool again;
1454   do {
1455     again = false;
1456     for (int i = 0; i < _block.length(); i++) {
1457       Node* n = _block.at(i);
1458       if (!n->is_Phi()) {
1459         int d_orig = depth(n);
1460         int d_in   = 0;
1461         for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1462           Node* pred = preds.current();
1463           if (in_bb(pred)) {
1464             d_in = MAX2(d_in, depth(pred));
1465           }
1466         }
1467         if (d_in + 1 != d_orig) {
1468           set_depth(n, d_in + 1);
1469           again = true;
1470         }
1471       }
1472     }
1473     ct++;
1474   } while (again);
1475 #ifndef PRODUCT
1476   if (TraceSuperWord && Verbose)
1477     tty->print_cr("compute_max_depth iterated: %d times", ct);
1478 #endif
1479 }
1480 
1481 //-------------------------compute_vector_element_type-----------------------
1482 // Compute necessary vector element type for expressions
1483 // This propagates backwards a narrower integer type when the
1484 // upper bits of the value are not needed.
1485 // Example:  char a,b,c;  a = b + c;
1486 // Normally the type of the add is integer, but for packed character
1487 // operations the type of the add needs to be char.
1488 void SuperWord::compute_vector_element_type() {
1489 #ifndef PRODUCT
1490   if (TraceSuperWord && Verbose)
1491     tty->print_cr("\ncompute_velt_type:");
1492 #endif
1493 
1494   // Initial type
1495   for (int i = 0; i < _block.length(); i++) {
1496     Node* n = _block.at(i);
1497     const Type* t  = n->is_Mem() ? Type::get_const_basic_type(n->as_Mem()->memory_type())
1498                                  : _igvn.type(n);
1499     const Type* vt = container_type(t);
1500     set_velt_type(n, vt);
1501   }
1502 
1503   // Propagate narrowed type backwards through operations
1504   // that don't depend on higher order bits
1505   for (int i = _block.length() - 1; i >= 0; i--) {
1506     Node* n = _block.at(i);
1507     // Only integer types need be examined
1508     if (n->bottom_type()->isa_int()) {
1509       uint start, end;
1510       vector_opd_range(n, &start, &end);
1511       const Type* vt = velt_type(n);
1512 
1513       for (uint j = start; j < end; j++) {
1514         Node* in  = n->in(j);
1515         // Don't propagate through a type conversion
1516         if (n->bottom_type() != in->bottom_type())
1517           continue;
1518         switch(in->Opcode()) {
1519         case Op_AddI:    case Op_AddL:
1520         case Op_SubI:    case Op_SubL:
1521         case Op_MulI:    case Op_MulL:
1522         case Op_AndI:    case Op_AndL:
1523         case Op_OrI:     case Op_OrL:
1524         case Op_XorI:    case Op_XorL:
1525         case Op_LShiftI: case Op_LShiftL:
1526         case Op_CMoveI:  case Op_CMoveL:
1527           if (in_bb(in)) {
1528             bool same_type = true;
1529             for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1530               Node *use = in->fast_out(k);
1531               if (!in_bb(use) || velt_type(use) != vt) {
1532                 same_type = false;
1533                 break;
1534               }
1535             }
1536             if (same_type) {
1537               set_velt_type(in, vt);
1538             }
1539           }
1540         }
1541       }
1542     }
1543   }
1544 #ifndef PRODUCT
1545   if (TraceSuperWord && Verbose) {
1546     for (int i = 0; i < _block.length(); i++) {
1547       Node* n = _block.at(i);
1548       velt_type(n)->dump();
1549       tty->print("\t");
1550       n->dump();
1551     }
1552   }
1553 #endif
1554 }
1555 
1556 //------------------------------memory_alignment---------------------------
1557 // Alignment within a vector memory reference
1558 int SuperWord::memory_alignment(MemNode* s, int iv_adjust_in_bytes) {
1559   SWPointer p(s, this);
1560   if (!p.valid()) {
1561     return bottom_align;
1562   }
1563   int offset  = p.offset_in_bytes();
1564   offset     += iv_adjust_in_bytes;
1565   int off_rem = offset % vector_width_in_bytes();
1566   int off_mod = off_rem >= 0 ? off_rem : off_rem + vector_width_in_bytes();
1567   return off_mod;
1568 }
1569 
1570 //---------------------------container_type---------------------------
1571 // Smallest type containing range of values
1572 const Type* SuperWord::container_type(const Type* t) {
1573   const Type* tp = t->make_ptr();
1574   if (tp && tp->isa_aryptr()) {
1575     t = tp->is_aryptr()->elem();
1576   }
1577   if (t->basic_type() == T_INT) {
1578     if (t->higher_equal(TypeInt::BOOL))  return TypeInt::BOOL;
1579     if (t->higher_equal(TypeInt::BYTE))  return TypeInt::BYTE;
1580     if (t->higher_equal(TypeInt::CHAR))  return TypeInt::CHAR;
1581     if (t->higher_equal(TypeInt::SHORT)) return TypeInt::SHORT;
1582     return TypeInt::INT;
1583   }
1584   return t;
1585 }
1586 
1587 //-------------------------vector_opd_range-----------------------
1588 // (Start, end] half-open range defining which operands are vector
1589 void SuperWord::vector_opd_range(Node* n, uint* start, uint* end) {
1590   switch (n->Opcode()) {
1591   case Op_LoadB:   case Op_LoadUS:
1592   case Op_LoadI:   case Op_LoadL:
1593   case Op_LoadF:   case Op_LoadD:
1594   case Op_LoadP:
1595     *start = 0;
1596     *end   = 0;
1597     return;
1598   case Op_StoreB:  case Op_StoreC:
1599   case Op_StoreI:  case Op_StoreL:
1600   case Op_StoreF:  case Op_StoreD:
1601   case Op_StoreP:
1602     *start = MemNode::ValueIn;
1603     *end   = *start + 1;
1604     return;
1605   case Op_LShiftI: case Op_LShiftL:
1606     *start = 1;
1607     *end   = 2;
1608     return;
1609   case Op_CMoveI:  case Op_CMoveL:  case Op_CMoveF:  case Op_CMoveD:
1610     *start = 2;
1611     *end   = n->req();
1612     return;
1613   }
1614   *start = 1;
1615   *end   = n->req(); // default is all operands
1616 }
1617 
1618 //------------------------------in_packset---------------------------
1619 // Are s1 and s2 in a pack pair and ordered as s1,s2?
1620 bool SuperWord::in_packset(Node* s1, Node* s2) {
1621   for (int i = 0; i < _packset.length(); i++) {
1622     Node_List* p = _packset.at(i);
1623     assert(p->size() == 2, "must be");
1624     if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
1625       return true;
1626     }
1627   }
1628   return false;
1629 }
1630 
1631 //------------------------------in_pack---------------------------
1632 // Is s in pack p?
1633 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
1634   for (uint i = 0; i < p->size(); i++) {
1635     if (p->at(i) == s) {
1636       return p;
1637     }
1638   }
1639   return NULL;
1640 }
1641 
1642 //------------------------------remove_pack_at---------------------------
1643 // Remove the pack at position pos in the packset
1644 void SuperWord::remove_pack_at(int pos) {
1645   Node_List* p = _packset.at(pos);
1646   for (uint i = 0; i < p->size(); i++) {
1647     Node* s = p->at(i);
1648     set_my_pack(s, NULL);
1649   }
1650   _packset.remove_at(pos);
1651 }
1652 
1653 //------------------------------executed_first---------------------------
1654 // Return the node executed first in pack p.  Uses the RPO block list
1655 // to determine order.
1656 Node* SuperWord::executed_first(Node_List* p) {
1657   Node* n = p->at(0);
1658   int n_rpo = bb_idx(n);
1659   for (uint i = 1; i < p->size(); i++) {
1660     Node* s = p->at(i);
1661     int s_rpo = bb_idx(s);
1662     if (s_rpo < n_rpo) {
1663       n = s;
1664       n_rpo = s_rpo;
1665     }
1666   }
1667   return n;
1668 }
1669 
1670 //------------------------------executed_last---------------------------
1671 // Return the node executed last in pack p.
1672 Node* SuperWord::executed_last(Node_List* p) {
1673   Node* n = p->at(0);
1674   int n_rpo = bb_idx(n);
1675   for (uint i = 1; i < p->size(); i++) {
1676     Node* s = p->at(i);
1677     int s_rpo = bb_idx(s);
1678     if (s_rpo > n_rpo) {
1679       n = s;
1680       n_rpo = s_rpo;
1681     }
1682   }
1683   return n;
1684 }
1685 
1686 //----------------------------align_initial_loop_index---------------------------
1687 // Adjust pre-loop limit so that in main loop, a load/store reference
1688 // to align_to_ref will be a position zero in the vector.
1689 //   (iv + k) mod vector_align == 0
1690 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
1691   CountedLoopNode *main_head = lp()->as_CountedLoop();
1692   assert(main_head->is_main_loop(), "");
1693   CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
1694   assert(pre_end != NULL, "");
1695   Node *pre_opaq1 = pre_end->limit();
1696   assert(pre_opaq1->Opcode() == Op_Opaque1, "");
1697   Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1698   Node *lim0 = pre_opaq->in(1);
1699 
1700   // Where we put new limit calculations
1701   Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1702 
1703   // Ensure the original loop limit is available from the
1704   // pre-loop Opaque1 node.
1705   Node *orig_limit = pre_opaq->original_loop_limit();
1706   assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
1707 
1708   SWPointer align_to_ref_p(align_to_ref, this);
1709 
1710   // Given:
1711   //     lim0 == original pre loop limit
1712   //     V == v_align (power of 2)
1713   //     invar == extra invariant piece of the address expression
1714   //     e == k [ +/- invar ]
1715   //
1716   // When reassociating expressions involving '%' the basic rules are:
1717   //     (a - b) % k == 0   =>  a % k == b % k
1718   // and:
1719   //     (a + b) % k == 0   =>  a % k == (k - b) % k
1720   //
1721   // For stride > 0 && scale > 0,
1722   //   Derive the new pre-loop limit "lim" such that the two constraints:
1723   //     (1) lim = lim0 + N           (where N is some positive integer < V)
1724   //     (2) (e + lim) % V == 0
1725   //   are true.
1726   //
1727   //   Substituting (1) into (2),
1728   //     (e + lim0 + N) % V == 0
1729   //   solve for N:
1730   //     N = (V - (e + lim0)) % V
1731   //   substitute back into (1), so that new limit
1732   //     lim = lim0 + (V - (e + lim0)) % V
1733   //
1734   // For stride > 0 && scale < 0
1735   //   Constraints:
1736   //     lim = lim0 + N
1737   //     (e - lim) % V == 0
1738   //   Solving for lim:
1739   //     (e - lim0 - N) % V == 0
1740   //     N = (e - lim0) % V
1741   //     lim = lim0 + (e - lim0) % V
1742   //
1743   // For stride < 0 && scale > 0
1744   //   Constraints:
1745   //     lim = lim0 - N
1746   //     (e + lim) % V == 0
1747   //   Solving for lim:
1748   //     (e + lim0 - N) % V == 0
1749   //     N = (e + lim0) % V
1750   //     lim = lim0 - (e + lim0) % V
1751   //
1752   // For stride < 0 && scale < 0
1753   //   Constraints:
1754   //     lim = lim0 - N
1755   //     (e - lim) % V == 0
1756   //   Solving for lim:
1757   //     (e - lim0 + N) % V == 0
1758   //     N = (V - (e - lim0)) % V
1759   //     lim = lim0 - (V - (e - lim0)) % V
1760 
1761   int stride   = iv_stride();
1762   int scale    = align_to_ref_p.scale_in_bytes();
1763   int elt_size = align_to_ref_p.memory_size();
1764   int v_align  = vector_width_in_bytes() / elt_size;
1765   int k        = align_to_ref_p.offset_in_bytes() / elt_size;
1766 
1767   Node *kn   = _igvn.intcon(k);
1768 
1769   Node *e = kn;
1770   if (align_to_ref_p.invar() != NULL) {
1771     // incorporate any extra invariant piece producing k +/- invar >>> log2(elt)
1772     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
1773     Node* aref     = new (_phase->C, 3) URShiftINode(align_to_ref_p.invar(), log2_elt);
1774     _phase->_igvn.register_new_node_with_optimizer(aref);
1775     _phase->set_ctrl(aref, pre_ctrl);
1776     if (align_to_ref_p.negate_invar()) {
1777       e = new (_phase->C, 3) SubINode(e, aref);
1778     } else {
1779       e = new (_phase->C, 3) AddINode(e, aref);
1780     }
1781     _phase->_igvn.register_new_node_with_optimizer(e);
1782     _phase->set_ctrl(e, pre_ctrl);
1783   }
1784 
1785   // compute e +/- lim0
1786   if (scale < 0) {
1787     e = new (_phase->C, 3) SubINode(e, lim0);
1788   } else {
1789     e = new (_phase->C, 3) AddINode(e, lim0);
1790   }
1791   _phase->_igvn.register_new_node_with_optimizer(e);
1792   _phase->set_ctrl(e, pre_ctrl);
1793 
1794   if (stride * scale > 0) {
1795     // compute V - (e +/- lim0)
1796     Node* va  = _igvn.intcon(v_align);
1797     e = new (_phase->C, 3) SubINode(va, e);
1798     _phase->_igvn.register_new_node_with_optimizer(e);
1799     _phase->set_ctrl(e, pre_ctrl);
1800   }
1801   // compute N = (exp) % V
1802   Node* va_msk = _igvn.intcon(v_align - 1);
1803   Node* N = new (_phase->C, 3) AndINode(e, va_msk);
1804   _phase->_igvn.register_new_node_with_optimizer(N);
1805   _phase->set_ctrl(N, pre_ctrl);
1806 
1807   //   substitute back into (1), so that new limit
1808   //     lim = lim0 + N
1809   Node* lim;
1810   if (stride < 0) {
1811     lim = new (_phase->C, 3) SubINode(lim0, N);
1812   } else {
1813     lim = new (_phase->C, 3) AddINode(lim0, N);
1814   }
1815   _phase->_igvn.register_new_node_with_optimizer(lim);
1816   _phase->set_ctrl(lim, pre_ctrl);
1817   Node* constrained =
1818     (stride > 0) ? (Node*) new (_phase->C,3) MinINode(lim, orig_limit)
1819                  : (Node*) new (_phase->C,3) MaxINode(lim, orig_limit);
1820   _phase->_igvn.register_new_node_with_optimizer(constrained);
1821   _phase->set_ctrl(constrained, pre_ctrl);
1822   _igvn.hash_delete(pre_opaq);
1823   pre_opaq->set_req(1, constrained);
1824 }
1825 
1826 //----------------------------get_pre_loop_end---------------------------
1827 // Find pre loop end from main loop.  Returns null if none.
1828 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
1829   Node *ctrl = cl->in(LoopNode::EntryControl);
1830   if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
1831   Node *iffm = ctrl->in(0);
1832   if (!iffm->is_If()) return NULL;
1833   Node *p_f = iffm->in(0);
1834   if (!p_f->is_IfFalse()) return NULL;
1835   if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
1836   CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1837   if (!pre_end->loopnode()->is_pre_loop()) return NULL;
1838   return pre_end;
1839 }
1840 
1841 
1842 //------------------------------init---------------------------
1843 void SuperWord::init() {
1844   _dg.init();
1845   _packset.clear();
1846   _disjoint_ptrs.clear();
1847   _block.clear();
1848   _data_entry.clear();
1849   _mem_slice_head.clear();
1850   _mem_slice_tail.clear();
1851   _node_info.clear();
1852   _align_to_ref = NULL;
1853   _lpt = NULL;
1854   _lp = NULL;
1855   _bb = NULL;
1856   _iv = NULL;
1857 }
1858 
1859 //------------------------------print_packset---------------------------
1860 void SuperWord::print_packset() {
1861 #ifndef PRODUCT
1862   tty->print_cr("packset");
1863   for (int i = 0; i < _packset.length(); i++) {
1864     tty->print_cr("Pack: %d", i);
1865     Node_List* p = _packset.at(i);
1866     print_pack(p);
1867   }
1868 #endif
1869 }
1870 
1871 //------------------------------print_pack---------------------------
1872 void SuperWord::print_pack(Node_List* p) {
1873   for (uint i = 0; i < p->size(); i++) {
1874     print_stmt(p->at(i));
1875   }
1876 }
1877 
1878 //------------------------------print_bb---------------------------
1879 void SuperWord::print_bb() {
1880 #ifndef PRODUCT
1881   tty->print_cr("\nBlock");
1882   for (int i = 0; i < _block.length(); i++) {
1883     Node* n = _block.at(i);
1884     tty->print("%d ", i);
1885     if (n) {
1886       n->dump();
1887     }
1888   }
1889 #endif
1890 }
1891 
1892 //------------------------------print_stmt---------------------------
1893 void SuperWord::print_stmt(Node* s) {
1894 #ifndef PRODUCT
1895   tty->print(" align: %d \t", alignment(s));
1896   s->dump();
1897 #endif
1898 }
1899 
1900 //------------------------------blank---------------------------
1901 char* SuperWord::blank(uint depth) {
1902   static char blanks[101];
1903   assert(depth < 101, "too deep");
1904   for (uint i = 0; i < depth; i++) blanks[i] = ' ';
1905   blanks[depth] = '\0';
1906   return blanks;
1907 }
1908 
1909 
1910 //==============================SWPointer===========================
1911 
1912 //----------------------------SWPointer------------------------
1913 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
1914   _mem(mem), _slp(slp),  _base(NULL),  _adr(NULL),
1915   _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
1916 
1917   Node* adr = mem->in(MemNode::Address);
1918   if (!adr->is_AddP()) {
1919     assert(!valid(), "too complex");
1920     return;
1921   }
1922   // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
1923   Node* base = adr->in(AddPNode::Base);
1924   for (int i = 0; i < 3; i++) {
1925     if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
1926       assert(!valid(), "too complex");
1927       return;
1928     }
1929     adr = adr->in(AddPNode::Address);
1930     if (base == adr || !adr->is_AddP()) {
1931       break; // stop looking at addp's
1932     }
1933   }
1934   _base = base;
1935   _adr  = adr;
1936   assert(valid(), "Usable");
1937 }
1938 
1939 // Following is used to create a temporary object during
1940 // the pattern match of an address expression.
1941 SWPointer::SWPointer(SWPointer* p) :
1942   _mem(p->_mem), _slp(p->_slp),  _base(NULL),  _adr(NULL),
1943   _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
1944 
1945 //------------------------scaled_iv_plus_offset--------------------
1946 // Match: k*iv + offset
1947 // where: k is a constant that maybe zero, and
1948 //        offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
1949 bool SWPointer::scaled_iv_plus_offset(Node* n) {
1950   if (scaled_iv(n)) {
1951     return true;
1952   }
1953   if (offset_plus_k(n)) {
1954     return true;
1955   }
1956   int opc = n->Opcode();
1957   if (opc == Op_AddI) {
1958     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
1959       return true;
1960     }
1961     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
1962       return true;
1963     }
1964   } else if (opc == Op_SubI) {
1965     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
1966       return true;
1967     }
1968     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
1969       _scale *= -1;
1970       return true;
1971     }
1972   }
1973   return false;
1974 }
1975 
1976 //----------------------------scaled_iv------------------------
1977 // Match: k*iv where k is a constant that's not zero
1978 bool SWPointer::scaled_iv(Node* n) {
1979   if (_scale != 0) {
1980     return false;  // already found a scale
1981   }
1982   if (n == iv()) {
1983     _scale = 1;
1984     return true;
1985   }
1986   int opc = n->Opcode();
1987   if (opc == Op_MulI) {
1988     if (n->in(1) == iv() && n->in(2)->is_Con()) {
1989       _scale = n->in(2)->get_int();
1990       return true;
1991     } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
1992       _scale = n->in(1)->get_int();
1993       return true;
1994     }
1995   } else if (opc == Op_LShiftI) {
1996     if (n->in(1) == iv() && n->in(2)->is_Con()) {
1997       _scale = 1 << n->in(2)->get_int();
1998       return true;
1999     }
2000   } else if (opc == Op_ConvI2L) {
2001     if (scaled_iv_plus_offset(n->in(1))) {
2002       return true;
2003     }
2004   } else if (opc == Op_LShiftL) {
2005     if (!has_iv() && _invar == NULL) {
2006       // Need to preserve the current _offset value, so
2007       // create a temporary object for this expression subtree.
2008       // Hacky, so should re-engineer the address pattern match.
2009       SWPointer tmp(this);
2010       if (tmp.scaled_iv_plus_offset(n->in(1))) {
2011         if (tmp._invar == NULL) {
2012           int mult = 1 << n->in(2)->get_int();
2013           _scale   = tmp._scale  * mult;
2014           _offset += tmp._offset * mult;
2015           return true;
2016         }
2017       }
2018     }
2019   }
2020   return false;
2021 }
2022 
2023 //----------------------------offset_plus_k------------------------
2024 // Match: offset is (k [+/- invariant])
2025 // where k maybe zero and invariant is optional, but not both.
2026 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2027   int opc = n->Opcode();
2028   if (opc == Op_ConI) {
2029     _offset += negate ? -(n->get_int()) : n->get_int();
2030     return true;
2031   } else if (opc == Op_ConL) {
2032     // Okay if value fits into an int
2033     const TypeLong* t = n->find_long_type();
2034     if (t->higher_equal(TypeLong::INT)) {
2035       jlong loff = n->get_long();
2036       jint  off  = (jint)loff;
2037       _offset += negate ? -off : loff;
2038       return true;
2039     }
2040     return false;
2041   }
2042   if (_invar != NULL) return false; // already have an invariant
2043   if (opc == Op_AddI) {
2044     if (n->in(2)->is_Con() && invariant(n->in(1))) {
2045       _negate_invar = negate;
2046       _invar = n->in(1);
2047       _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2048       return true;
2049     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2050       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2051       _negate_invar = negate;
2052       _invar = n->in(2);
2053       return true;
2054     }
2055   }
2056   if (opc == Op_SubI) {
2057     if (n->in(2)->is_Con() && invariant(n->in(1))) {
2058       _negate_invar = negate;
2059       _invar = n->in(1);
2060       _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2061       return true;
2062     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2063       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2064       _negate_invar = !negate;
2065       _invar = n->in(2);
2066       return true;
2067     }
2068   }
2069   if (invariant(n)) {
2070     _negate_invar = negate;
2071     _invar = n;
2072     return true;
2073   }
2074   return false;
2075 }
2076 
2077 //----------------------------print------------------------
2078 void SWPointer::print() {
2079 #ifndef PRODUCT
2080   tty->print("base: %d  adr: %d  scale: %d  offset: %d  invar: %c%d\n",
2081              _base != NULL ? _base->_idx : 0,
2082              _adr  != NULL ? _adr->_idx  : 0,
2083              _scale, _offset,
2084              _negate_invar?'-':'+',
2085              _invar != NULL ? _invar->_idx : 0);
2086 #endif
2087 }
2088 
2089 // ========================= OrderedPair =====================
2090 
2091 const OrderedPair OrderedPair::initial;
2092 
2093 // ========================= SWNodeInfo =====================
2094 
2095 const SWNodeInfo SWNodeInfo::initial;
2096 
2097 
2098 // ============================ DepGraph ===========================
2099 
2100 //------------------------------make_node---------------------------
2101 // Make a new dependence graph node for an ideal node.
2102 DepMem* DepGraph::make_node(Node* node) {
2103   DepMem* m = new (_arena) DepMem(node);
2104   if (node != NULL) {
2105     assert(_map.at_grow(node->_idx) == NULL, "one init only");
2106     _map.at_put_grow(node->_idx, m);
2107   }
2108   return m;
2109 }
2110 
2111 //------------------------------make_edge---------------------------
2112 // Make a new dependence graph edge from dpred -> dsucc
2113 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2114   DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2115   dpred->set_out_head(e);
2116   dsucc->set_in_head(e);
2117   return e;
2118 }
2119 
2120 // ========================== DepMem ========================
2121 
2122 //------------------------------in_cnt---------------------------
2123 int DepMem::in_cnt() {
2124   int ct = 0;
2125   for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2126   return ct;
2127 }
2128 
2129 //------------------------------out_cnt---------------------------
2130 int DepMem::out_cnt() {
2131   int ct = 0;
2132   for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2133   return ct;
2134 }
2135 
2136 //------------------------------print-----------------------------
2137 void DepMem::print() {
2138 #ifndef PRODUCT
2139   tty->print("  DepNode %d (", _node->_idx);
2140   for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2141     Node* pred = p->pred()->node();
2142     tty->print(" %d", pred != NULL ? pred->_idx : 0);
2143   }
2144   tty->print(") [");
2145   for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2146     Node* succ = s->succ()->node();
2147     tty->print(" %d", succ != NULL ? succ->_idx : 0);
2148   }
2149   tty->print_cr(" ]");
2150 #endif
2151 }
2152 
2153 // =========================== DepEdge =========================
2154 
2155 //------------------------------DepPreds---------------------------
2156 void DepEdge::print() {
2157 #ifndef PRODUCT
2158   tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2159 #endif
2160 }
2161 
2162 // =========================== DepPreds =========================
2163 // Iterator over predecessor edges in the dependence graph.
2164 
2165 //------------------------------DepPreds---------------------------
2166 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2167   _n = n;
2168   _done = false;
2169   if (_n->is_Store() || _n->is_Load()) {
2170     _next_idx = MemNode::Address;
2171     _end_idx  = n->req();
2172     _dep_next = dg.dep(_n)->in_head();
2173   } else if (_n->is_Mem()) {
2174     _next_idx = 0;
2175     _end_idx  = 0;
2176     _dep_next = dg.dep(_n)->in_head();
2177   } else {
2178     _next_idx = 1;
2179     _end_idx  = _n->req();
2180     _dep_next = NULL;
2181   }
2182   next();
2183 }
2184 
2185 //------------------------------next---------------------------
2186 void DepPreds::next() {
2187   if (_dep_next != NULL) {
2188     _current  = _dep_next->pred()->node();
2189     _dep_next = _dep_next->next_in();
2190   } else if (_next_idx < _end_idx) {
2191     _current  = _n->in(_next_idx++);
2192   } else {
2193     _done = true;
2194   }
2195 }
2196 
2197 // =========================== DepSuccs =========================
2198 // Iterator over successor edges in the dependence graph.
2199 
2200 //------------------------------DepSuccs---------------------------
2201 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2202   _n = n;
2203   _done = false;
2204   if (_n->is_Load()) {
2205     _next_idx = 0;
2206     _end_idx  = _n->outcnt();
2207     _dep_next = dg.dep(_n)->out_head();
2208   } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2209     _next_idx = 0;
2210     _end_idx  = 0;
2211     _dep_next = dg.dep(_n)->out_head();
2212   } else {
2213     _next_idx = 0;
2214     _end_idx  = _n->outcnt();
2215     _dep_next = NULL;
2216   }
2217   next();
2218 }
2219 
2220 //-------------------------------next---------------------------
2221 void DepSuccs::next() {
2222   if (_dep_next != NULL) {
2223     _current  = _dep_next->succ()->node();
2224     _dep_next = _dep_next->next_out();
2225   } else if (_next_idx < _end_idx) {
2226     _current  = _n->raw_out(_next_idx++);
2227   } else {
2228     _done = true;
2229   }
2230 }