1 /*
   2  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  */
  23 
  24 #include "precompiled.hpp"
  25 #include "compiler/compileLog.hpp"
  26 #include "libadt/vectset.hpp"
  27 #include "memory/allocation.inline.hpp"
  28 #include "opto/addnode.hpp"
  29 #include "opto/callnode.hpp"
  30 #include "opto/divnode.hpp"
  31 #include "opto/matcher.hpp"
  32 #include "opto/memnode.hpp"
  33 #include "opto/mulnode.hpp"
  34 #include "opto/opcodes.hpp"
  35 #include "opto/superword.hpp"
  36 #include "opto/vectornode.hpp"
  37 
  38 //
  39 //                  S U P E R W O R D   T R A N S F O R M
  40 //=============================================================================
  41 
  42 //------------------------------SuperWord---------------------------
  43 SuperWord::SuperWord(PhaseIdealLoop* phase) :
  44   _phase(phase),
  45   _igvn(phase->_igvn),
  46   _arena(phase->C->comp_arena()),
  47   _packset(arena(), 8,  0, NULL),         // packs for the current block
  48   _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
  49   _block(arena(), 8,  0, NULL),           // nodes in current block
  50   _data_entry(arena(), 8,  0, NULL),      // nodes with all inputs from outside
  51   _mem_slice_head(arena(), 8,  0, NULL),  // memory slice heads
  52   _mem_slice_tail(arena(), 8,  0, NULL),  // memory slice tails
  53   _node_info(arena(), 8,  0, SWNodeInfo::initial), // info needed per node
  54   _align_to_ref(NULL),                    // memory reference to align vectors to
  55   _disjoint_ptrs(arena(), 8,  0, OrderedPair::initial), // runtime disambiguated pointer pairs
  56   _dg(_arena),                            // dependence graph
  57   _visited(arena()),                      // visited node set
  58   _post_visited(arena()),                 // post visited node set
  59   _n_idx_list(arena(), 8),                // scratch list of (node,index) pairs
  60   _stk(arena(), 8, 0, NULL),              // scratch stack of nodes
  61   _nlist(arena(), 8, 0, NULL),            // scratch list of nodes
  62   _lpt(NULL),                             // loop tree node
  63   _lp(NULL),                              // LoopNode
  64   _bb(NULL),                              // basic block
  65   _iv(NULL)                               // induction var
  66 {}
  67 
  68 //------------------------------transform_loop---------------------------
  69 void SuperWord::transform_loop(IdealLoopTree* lpt) {
  70   assert(lpt->_head->is_CountedLoop(), "must be");
  71   CountedLoopNode *cl = lpt->_head->as_CountedLoop();
  72 
  73   if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
  74 
  75   // Check for no control flow in body (other than exit)
  76   Node *cl_exit = cl->loopexit();
  77   if (cl_exit->in(0) != lpt->_head) return;
  78 
  79   // Make sure the are no extra control users of the loop backedge
  80   if (cl->back_control()->outcnt() != 1) {
  81     return;
  82   }
  83 
  84   // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
  85   CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
  86   if (pre_end == NULL) return;
  87   Node *pre_opaq1 = pre_end->limit();
  88   if (pre_opaq1->Opcode() != Op_Opaque1) return;
  89 
  90   // Do vectors exist on this architecture?
  91   if (vector_width_in_bytes() == 0) return;
  92 
  93   init(); // initialize data structures
  94 
  95   set_lpt(lpt);
  96   set_lp(cl);
  97 
  98  // For now, define one block which is the entire loop body
  99   set_bb(cl);
 100 
 101   assert(_packset.length() == 0, "packset must be empty");
 102   SLP_extract();
 103 }
 104 
 105 //------------------------------SLP_extract---------------------------
 106 // Extract the superword level parallelism
 107 //
 108 // 1) A reverse post-order of nodes in the block is constructed.  By scanning
 109 //    this list from first to last, all definitions are visited before their uses.
 110 //
 111 // 2) A point-to-point dependence graph is constructed between memory references.
 112 //    This simplies the upcoming "independence" checker.
 113 //
 114 // 3) The maximum depth in the node graph from the beginning of the block
 115 //    to each node is computed.  This is used to prune the graph search
 116 //    in the independence checker.
 117 //
 118 // 4) For integer types, the necessary bit width is propagated backwards
 119 //    from stores to allow packed operations on byte, char, and short
 120 //    integers.  This reverses the promotion to type "int" that javac
 121 //    did for operations like: char c1,c2,c3;  c1 = c2 + c3.
 122 //
 123 // 5) One of the memory references is picked to be an aligned vector reference.
 124 //    The pre-loop trip count is adjusted to align this reference in the
 125 //    unrolled body.
 126 //
 127 // 6) The initial set of pack pairs is seeded with memory references.
 128 //
 129 // 7) The set of pack pairs is extended by following use->def and def->use links.
 130 //
 131 // 8) The pairs are combined into vector sized packs.
 132 //
 133 // 9) Reorder the memory slices to co-locate members of the memory packs.
 134 //
 135 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
 136 //    inserting scalar promotion, vector creation from multiple scalars, and
 137 //    extraction of scalar values from vectors.
 138 //
 139 void SuperWord::SLP_extract() {
 140 
 141   // Ready the block
 142 
 143   construct_bb();
 144 
 145   dependence_graph();
 146 
 147   compute_max_depth();
 148 
 149   compute_vector_element_type();
 150 
 151   // Attempt vectorization
 152 
 153   find_adjacent_refs();
 154 
 155   extend_packlist();
 156 
 157   combine_packs();
 158 
 159   construct_my_pack_map();
 160 
 161   filter_packs();
 162 
 163   schedule();
 164 
 165   output();
 166 }
 167 
 168 //------------------------------find_adjacent_refs---------------------------
 169 // Find the adjacent memory references and create pack pairs for them.
 170 // This is the initial set of packs that will then be extended by
 171 // following use->def and def->use links.  The align positions are
 172 // assigned relative to the reference "align_to_ref"
 173 void SuperWord::find_adjacent_refs() {
 174   // Get list of memory operations
 175   Node_List memops;
 176   for (int i = 0; i < _block.length(); i++) {
 177     Node* n = _block.at(i);
 178     if (n->is_Mem() && in_bb(n) &&
 179         is_java_primitive(n->as_Mem()->memory_type())) {
 180       int align = memory_alignment(n->as_Mem(), 0);
 181       if (align != bottom_align) {
 182         memops.push(n);
 183       }
 184     }
 185   }
 186   if (memops.size() == 0) return;
 187 
 188   // Find a memory reference to align to.  The pre-loop trip count
 189   // is modified to align this reference to a vector-aligned address
 190   find_align_to_ref(memops);
 191   if (align_to_ref() == NULL) return;
 192 
 193   SWPointer align_to_ref_p(align_to_ref(), this);
 194   int offset = align_to_ref_p.offset_in_bytes();
 195   int scale  = align_to_ref_p.scale_in_bytes();
 196   int vw              = vector_width_in_bytes();
 197   int stride_sign     = (scale * iv_stride()) > 0 ? 1 : -1;
 198   int iv_adjustment   = (stride_sign * vw - (offset % vw)) % vw;
 199 
 200 #ifndef PRODUCT
 201   if (TraceSuperWord)
 202     tty->print_cr("\noffset = %d iv_adjustment = %d  elt_align = %d scale = %d iv_stride = %d",
 203                   offset, iv_adjustment, align_to_ref_p.memory_size(), align_to_ref_p.scale_in_bytes(), iv_stride());
 204 #endif
 205 
 206   // Set alignment relative to "align_to_ref"
 207   for (int i = memops.size() - 1; i >= 0; i--) {
 208     MemNode* s = memops.at(i)->as_Mem();
 209     SWPointer p2(s, this);
 210     if (p2.comparable(align_to_ref_p)) {
 211       int align = memory_alignment(s, iv_adjustment);
 212       set_alignment(s, align);
 213     } else {
 214       memops.remove(i);
 215     }
 216   }
 217 
 218   // Create initial pack pairs of memory operations
 219   for (uint i = 0; i < memops.size(); i++) {
 220     Node* s1 = memops.at(i);
 221     for (uint j = 0; j < memops.size(); j++) {
 222       Node* s2 = memops.at(j);
 223       if (s1 != s2 && are_adjacent_refs(s1, s2)) {
 224         int align = alignment(s1);
 225         if (stmts_can_pack(s1, s2, align)) {
 226           Node_List* pair = new Node_List();
 227           pair->push(s1);
 228           pair->push(s2);
 229           _packset.append(pair);
 230         }
 231       }
 232     }
 233   }
 234 
 235 #ifndef PRODUCT
 236   if (TraceSuperWord) {
 237     tty->print_cr("\nAfter find_adjacent_refs");
 238     print_packset();
 239   }
 240 #endif
 241 }
 242 
 243 //------------------------------find_align_to_ref---------------------------
 244 // Find a memory reference to align the loop induction variable to.
 245 // Looks first at stores then at loads, looking for a memory reference
 246 // with the largest number of references similar to it.
 247 void SuperWord::find_align_to_ref(Node_List &memops) {
 248   GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
 249 
 250   // Count number of comparable memory ops
 251   for (uint i = 0; i < memops.size(); i++) {
 252     MemNode* s1 = memops.at(i)->as_Mem();
 253     SWPointer p1(s1, this);
 254     // Discard if pre loop can't align this reference
 255     if (!ref_is_alignable(p1)) {
 256       *cmp_ct.adr_at(i) = 0;
 257       continue;
 258     }
 259     for (uint j = i+1; j < memops.size(); j++) {
 260       MemNode* s2 = memops.at(j)->as_Mem();
 261       if (isomorphic(s1, s2)) {
 262         SWPointer p2(s2, this);
 263         if (p1.comparable(p2)) {
 264           (*cmp_ct.adr_at(i))++;
 265           (*cmp_ct.adr_at(j))++;
 266         }
 267       }
 268     }
 269   }
 270 
 271   // Find Store (or Load) with the greatest number of "comparable" references
 272   int max_ct        = 0;
 273   int max_idx       = -1;
 274   int min_size      = max_jint;
 275   int min_iv_offset = max_jint;
 276   for (uint j = 0; j < memops.size(); j++) {
 277     MemNode* s = memops.at(j)->as_Mem();
 278     if (s->is_Store()) {
 279       SWPointer p(s, this);
 280       if (cmp_ct.at(j) > max_ct ||
 281           cmp_ct.at(j) == max_ct && (data_size(s) < min_size ||
 282                                      data_size(s) == min_size &&
 283                                         p.offset_in_bytes() < min_iv_offset)) {
 284         max_ct = cmp_ct.at(j);
 285         max_idx = j;
 286         min_size = data_size(s);
 287         min_iv_offset = p.offset_in_bytes();
 288       }
 289     }
 290   }
 291   // If no stores, look at loads
 292   if (max_ct == 0) {
 293     for (uint j = 0; j < memops.size(); j++) {
 294       MemNode* s = memops.at(j)->as_Mem();
 295       if (s->is_Load()) {
 296         SWPointer p(s, this);
 297         if (cmp_ct.at(j) > max_ct ||
 298             cmp_ct.at(j) == max_ct && (data_size(s) < min_size ||
 299                                        data_size(s) == min_size &&
 300                                           p.offset_in_bytes() < min_iv_offset)) {
 301           max_ct = cmp_ct.at(j);
 302           max_idx = j;
 303           min_size = data_size(s);
 304           min_iv_offset = p.offset_in_bytes();
 305         }
 306       }
 307     }
 308   }
 309 
 310   if (max_ct > 0)
 311     set_align_to_ref(memops.at(max_idx)->as_Mem());
 312 
 313 #ifndef PRODUCT
 314   if (TraceSuperWord && Verbose) {
 315     tty->print_cr("\nVector memops after find_align_to_refs");
 316     for (uint i = 0; i < memops.size(); i++) {
 317       MemNode* s = memops.at(i)->as_Mem();
 318       s->dump();
 319     }
 320   }
 321 #endif
 322 }
 323 
 324 //------------------------------ref_is_alignable---------------------------
 325 // Can the preloop align the reference to position zero in the vector?
 326 bool SuperWord::ref_is_alignable(SWPointer& p) {
 327   if (!p.has_iv()) {
 328     return true;   // no induction variable
 329   }
 330   CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
 331   assert(pre_end->stride_is_con(), "pre loop stride is constant");
 332   int preloop_stride = pre_end->stride_con();
 333 
 334   int span = preloop_stride * p.scale_in_bytes();
 335 
 336   // Stride one accesses are alignable.
 337   if (ABS(span) == p.memory_size())
 338     return true;
 339 
 340   // If initial offset from start of object is computable,
 341   // compute alignment within the vector.
 342   int vw = vector_width_in_bytes();
 343   if (vw % span == 0) {
 344     Node* init_nd = pre_end->init_trip();
 345     if (init_nd->is_Con() && p.invar() == NULL) {
 346       int init = init_nd->bottom_type()->is_int()->get_con();
 347 
 348       int init_offset = init * p.scale_in_bytes() + p.offset_in_bytes();
 349       assert(init_offset >= 0, "positive offset from object start");
 350 
 351       if (span > 0) {
 352         return (vw - (init_offset % vw)) % span == 0;
 353       } else {
 354         assert(span < 0, "nonzero stride * scale");
 355         return (init_offset % vw) % -span == 0;
 356       }
 357     }
 358   }
 359   return false;
 360 }
 361 
 362 //---------------------------dependence_graph---------------------------
 363 // Construct dependency graph.
 364 // Add dependence edges to load/store nodes for memory dependence
 365 //    A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
 366 void SuperWord::dependence_graph() {
 367   // First, assign a dependence node to each memory node
 368   for (int i = 0; i < _block.length(); i++ ) {
 369     Node *n = _block.at(i);
 370     if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
 371       _dg.make_node(n);
 372     }
 373   }
 374 
 375   // For each memory slice, create the dependences
 376   for (int i = 0; i < _mem_slice_head.length(); i++) {
 377     Node* n      = _mem_slice_head.at(i);
 378     Node* n_tail = _mem_slice_tail.at(i);
 379 
 380     // Get slice in predecessor order (last is first)
 381     mem_slice_preds(n_tail, n, _nlist);
 382 
 383     // Make the slice dependent on the root
 384     DepMem* slice = _dg.dep(n);
 385     _dg.make_edge(_dg.root(), slice);
 386 
 387     // Create a sink for the slice
 388     DepMem* slice_sink = _dg.make_node(NULL);
 389     _dg.make_edge(slice_sink, _dg.tail());
 390 
 391     // Now visit each pair of memory ops, creating the edges
 392     for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 393       Node* s1 = _nlist.at(j);
 394 
 395       // If no dependency yet, use slice
 396       if (_dg.dep(s1)->in_cnt() == 0) {
 397         _dg.make_edge(slice, s1);
 398       }
 399       SWPointer p1(s1->as_Mem(), this);
 400       bool sink_dependent = true;
 401       for (int k = j - 1; k >= 0; k--) {
 402         Node* s2 = _nlist.at(k);
 403         if (s1->is_Load() && s2->is_Load())
 404           continue;
 405         SWPointer p2(s2->as_Mem(), this);
 406 
 407         int cmp = p1.cmp(p2);
 408         if (SuperWordRTDepCheck &&
 409             p1.base() != p2.base() && p1.valid() && p2.valid()) {
 410           // Create a runtime check to disambiguate
 411           OrderedPair pp(p1.base(), p2.base());
 412           _disjoint_ptrs.append_if_missing(pp);
 413         } else if (!SWPointer::not_equal(cmp)) {
 414           // Possibly same address
 415           _dg.make_edge(s1, s2);
 416           sink_dependent = false;
 417         }
 418       }
 419       if (sink_dependent) {
 420         _dg.make_edge(s1, slice_sink);
 421       }
 422     }
 423 #ifndef PRODUCT
 424     if (TraceSuperWord) {
 425       tty->print_cr("\nDependence graph for slice: %d", n->_idx);
 426       for (int q = 0; q < _nlist.length(); q++) {
 427         _dg.print(_nlist.at(q));
 428       }
 429       tty->cr();
 430     }
 431 #endif
 432     _nlist.clear();
 433   }
 434 
 435 #ifndef PRODUCT
 436   if (TraceSuperWord) {
 437     tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
 438     for (int r = 0; r < _disjoint_ptrs.length(); r++) {
 439       _disjoint_ptrs.at(r).print();
 440       tty->cr();
 441     }
 442     tty->cr();
 443   }
 444 #endif
 445 }
 446 
 447 //---------------------------mem_slice_preds---------------------------
 448 // Return a memory slice (node list) in predecessor order starting at "start"
 449 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
 450   assert(preds.length() == 0, "start empty");
 451   Node* n = start;
 452   Node* prev = NULL;
 453   while (true) {
 454     assert(in_bb(n), "must be in block");
 455     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 456       Node* out = n->fast_out(i);
 457       if (out->is_Load()) {
 458         if (in_bb(out)) {
 459           preds.push(out);
 460         }
 461       } else {
 462         // FIXME
 463         if (out->is_MergeMem() && !in_bb(out)) {
 464           // Either unrolling is causing a memory edge not to disappear,
 465           // or need to run igvn.optimize() again before SLP
 466         } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
 467           // Ditto.  Not sure what else to check further.
 468         } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
 469           // StoreCM has an input edge used as a precedence edge.
 470           // Maybe an issue when oop stores are vectorized.
 471         } else {
 472           assert(out == prev || prev == NULL, "no branches off of store slice");
 473         }
 474       }
 475     }
 476     if (n == stop) break;
 477     preds.push(n);
 478     prev = n;
 479     n = n->in(MemNode::Memory);
 480   }
 481 }
 482 
 483 //------------------------------stmts_can_pack---------------------------
 484 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
 485 // s1 aligned at "align"
 486 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
 487 
 488   // Do not use superword for non-primitives
 489   if((s1->is_Mem() && !is_java_primitive(s1->as_Mem()->memory_type())) ||
 490      (s2->is_Mem() && !is_java_primitive(s2->as_Mem()->memory_type())))
 491     return false;
 492 
 493   if (isomorphic(s1, s2)) {
 494     if (independent(s1, s2)) {
 495       if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
 496         if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
 497           int s1_align = alignment(s1);
 498           int s2_align = alignment(s2);
 499           if (s1_align == top_align || s1_align == align) {
 500             if (s2_align == top_align || s2_align == align + data_size(s1)) {
 501               return true;
 502             }
 503           }
 504         }
 505       }
 506     }
 507   }
 508   return false;
 509 }
 510 
 511 //------------------------------exists_at---------------------------
 512 // Does s exist in a pack at position pos?
 513 bool SuperWord::exists_at(Node* s, uint pos) {
 514   for (int i = 0; i < _packset.length(); i++) {
 515     Node_List* p = _packset.at(i);
 516     if (p->at(pos) == s) {
 517       return true;
 518     }
 519   }
 520   return false;
 521 }
 522 
 523 //------------------------------are_adjacent_refs---------------------------
 524 // Is s1 immediately before s2 in memory?
 525 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
 526   if (!s1->is_Mem() || !s2->is_Mem()) return false;
 527   if (!in_bb(s1)    || !in_bb(s2))    return false;
 528 
 529   // Do not use superword for non-primitives
 530   if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
 531       !is_java_primitive(s2->as_Mem()->memory_type())) {
 532     return false;
 533   }
 534 
 535   // FIXME - co_locate_pack fails on Stores in different mem-slices, so
 536   // only pack memops that are in the same alias set until that's fixed.
 537   if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
 538       _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
 539     return false;
 540   SWPointer p1(s1->as_Mem(), this);
 541   SWPointer p2(s2->as_Mem(), this);
 542   if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
 543   int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
 544   return diff == data_size(s1);
 545 }
 546 
 547 //------------------------------isomorphic---------------------------
 548 // Are s1 and s2 similar?
 549 bool SuperWord::isomorphic(Node* s1, Node* s2) {
 550   if (s1->Opcode() != s2->Opcode()) return false;
 551   if (s1->req() != s2->req()) return false;
 552   if (s1->in(0) != s2->in(0)) return false;
 553   if (velt_type(s1) != velt_type(s2)) return false;
 554   return true;
 555 }
 556 
 557 //------------------------------independent---------------------------
 558 // Is there no data path from s1 to s2 or s2 to s1?
 559 bool SuperWord::independent(Node* s1, Node* s2) {
 560   //  assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
 561   int d1 = depth(s1);
 562   int d2 = depth(s2);
 563   if (d1 == d2) return s1 != s2;
 564   Node* deep    = d1 > d2 ? s1 : s2;
 565   Node* shallow = d1 > d2 ? s2 : s1;
 566 
 567   visited_clear();
 568 
 569   return independent_path(shallow, deep);
 570 }
 571 
 572 //------------------------------independent_path------------------------------
 573 // Helper for independent
 574 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
 575   if (dp >= 1000) return false; // stop deep recursion
 576   visited_set(deep);
 577   int shal_depth = depth(shallow);
 578   assert(shal_depth <= depth(deep), "must be");
 579   for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
 580     Node* pred = preds.current();
 581     if (in_bb(pred) && !visited_test(pred)) {
 582       if (shallow == pred) {
 583         return false;
 584       }
 585       if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
 586         return false;
 587       }
 588     }
 589   }
 590   return true;
 591 }
 592 
 593 //------------------------------set_alignment---------------------------
 594 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
 595   set_alignment(s1, align);
 596   set_alignment(s2, align + data_size(s1));
 597 }
 598 
 599 //------------------------------data_size---------------------------
 600 int SuperWord::data_size(Node* s) {
 601   const Type* t = velt_type(s);
 602   BasicType  bt = t->array_element_basic_type();
 603   int bsize = type2aelembytes(bt);
 604   assert(bsize != 0, "valid size");
 605   return bsize;
 606 }
 607 
 608 //------------------------------extend_packlist---------------------------
 609 // Extend packset by following use->def and def->use links from pack members.
 610 void SuperWord::extend_packlist() {
 611   bool changed;
 612   do {
 613     changed = false;
 614     for (int i = 0; i < _packset.length(); i++) {
 615       Node_List* p = _packset.at(i);
 616       changed |= follow_use_defs(p);
 617       changed |= follow_def_uses(p);
 618     }
 619   } while (changed);
 620 
 621 #ifndef PRODUCT
 622   if (TraceSuperWord) {
 623     tty->print_cr("\nAfter extend_packlist");
 624     print_packset();
 625   }
 626 #endif
 627 }
 628 
 629 //------------------------------follow_use_defs---------------------------
 630 // Extend the packset by visiting operand definitions of nodes in pack p
 631 bool SuperWord::follow_use_defs(Node_List* p) {
 632   Node* s1 = p->at(0);
 633   Node* s2 = p->at(1);
 634   assert(p->size() == 2, "just checking");
 635   assert(s1->req() == s2->req(), "just checking");
 636   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
 637 
 638   if (s1->is_Load()) return false;
 639 
 640   int align = alignment(s1);
 641   bool changed = false;
 642   int start = s1->is_Store() ? MemNode::ValueIn   : 1;
 643   int end   = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
 644   for (int j = start; j < end; j++) {
 645     Node* t1 = s1->in(j);
 646     Node* t2 = s2->in(j);
 647     if (!in_bb(t1) || !in_bb(t2))
 648       continue;
 649     if (stmts_can_pack(t1, t2, align)) {
 650       if (est_savings(t1, t2) >= 0) {
 651         Node_List* pair = new Node_List();
 652         pair->push(t1);
 653         pair->push(t2);
 654         _packset.append(pair);
 655         set_alignment(t1, t2, align);
 656         changed = true;
 657       }
 658     }
 659   }
 660   return changed;
 661 }
 662 
 663 //------------------------------follow_def_uses---------------------------
 664 // Extend the packset by visiting uses of nodes in pack p
 665 bool SuperWord::follow_def_uses(Node_List* p) {
 666   bool changed = false;
 667   Node* s1 = p->at(0);
 668   Node* s2 = p->at(1);
 669   assert(p->size() == 2, "just checking");
 670   assert(s1->req() == s2->req(), "just checking");
 671   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
 672 
 673   if (s1->is_Store()) return false;
 674 
 675   int align = alignment(s1);
 676   int savings = -1;
 677   Node* u1 = NULL;
 678   Node* u2 = NULL;
 679   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
 680     Node* t1 = s1->fast_out(i);
 681     if (!in_bb(t1)) continue;
 682     for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
 683       Node* t2 = s2->fast_out(j);
 684       if (!in_bb(t2)) continue;
 685       if (!opnd_positions_match(s1, t1, s2, t2))
 686         continue;
 687       if (stmts_can_pack(t1, t2, align)) {
 688         int my_savings = est_savings(t1, t2);
 689         if (my_savings > savings) {
 690           savings = my_savings;
 691           u1 = t1;
 692           u2 = t2;
 693         }
 694       }
 695     }
 696   }
 697   if (savings >= 0) {
 698     Node_List* pair = new Node_List();
 699     pair->push(u1);
 700     pair->push(u2);
 701     _packset.append(pair);
 702     set_alignment(u1, u2, align);
 703     changed = true;
 704   }
 705   return changed;
 706 }
 707 
 708 //---------------------------opnd_positions_match-------------------------
 709 // Is the use of d1 in u1 at the same operand position as d2 in u2?
 710 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
 711   uint ct = u1->req();
 712   if (ct != u2->req()) return false;
 713   uint i1 = 0;
 714   uint i2 = 0;
 715   do {
 716     for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
 717     for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
 718     if (i1 != i2) {
 719       return false;
 720     }
 721   } while (i1 < ct);
 722   return true;
 723 }
 724 
 725 //------------------------------est_savings---------------------------
 726 // Estimate the savings from executing s1 and s2 as a pack
 727 int SuperWord::est_savings(Node* s1, Node* s2) {
 728   int save = 2 - 1; // 2 operations per instruction in packed form
 729 
 730   // inputs
 731   for (uint i = 1; i < s1->req(); i++) {
 732     Node* x1 = s1->in(i);
 733     Node* x2 = s2->in(i);
 734     if (x1 != x2) {
 735       if (are_adjacent_refs(x1, x2)) {
 736         save += adjacent_profit(x1, x2);
 737       } else if (!in_packset(x1, x2)) {
 738         save -= pack_cost(2);
 739       } else {
 740         save += unpack_cost(2);
 741       }
 742     }
 743   }
 744 
 745   // uses of result
 746   uint ct = 0;
 747   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
 748     Node* s1_use = s1->fast_out(i);
 749     for (int j = 0; j < _packset.length(); j++) {
 750       Node_List* p = _packset.at(j);
 751       if (p->at(0) == s1_use) {
 752         for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
 753           Node* s2_use = s2->fast_out(k);
 754           if (p->at(p->size()-1) == s2_use) {
 755             ct++;
 756             if (are_adjacent_refs(s1_use, s2_use)) {
 757               save += adjacent_profit(s1_use, s2_use);
 758             }
 759           }
 760         }
 761       }
 762     }
 763   }
 764 
 765   if (ct < s1->outcnt()) save += unpack_cost(1);
 766   if (ct < s2->outcnt()) save += unpack_cost(1);
 767 
 768   return save;
 769 }
 770 
 771 //------------------------------costs---------------------------
 772 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
 773 int SuperWord::pack_cost(int ct)   { return ct; }
 774 int SuperWord::unpack_cost(int ct) { return ct; }
 775 
 776 //------------------------------combine_packs---------------------------
 777 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
 778 void SuperWord::combine_packs() {
 779   bool changed;
 780   do {
 781     changed = false;
 782     for (int i = 0; i < _packset.length(); i++) {
 783       Node_List* p1 = _packset.at(i);
 784       if (p1 == NULL) continue;
 785       for (int j = 0; j < _packset.length(); j++) {
 786         Node_List* p2 = _packset.at(j);
 787         if (p2 == NULL) continue;
 788         if (p1->at(p1->size()-1) == p2->at(0)) {
 789           for (uint k = 1; k < p2->size(); k++) {
 790             p1->push(p2->at(k));
 791           }
 792           _packset.at_put(j, NULL);
 793           changed = true;
 794         }
 795       }
 796     }
 797   } while (changed);
 798 
 799   for (int i = _packset.length() - 1; i >= 0; i--) {
 800     Node_List* p1 = _packset.at(i);
 801     if (p1 == NULL) {
 802       _packset.remove_at(i);
 803     }
 804   }
 805 
 806 #ifndef PRODUCT
 807   if (TraceSuperWord) {
 808     tty->print_cr("\nAfter combine_packs");
 809     print_packset();
 810   }
 811 #endif
 812 }
 813 
 814 //-----------------------------construct_my_pack_map--------------------------
 815 // Construct the map from nodes to packs.  Only valid after the
 816 // point where a node is only in one pack (after combine_packs).
 817 void SuperWord::construct_my_pack_map() {
 818   Node_List* rslt = NULL;
 819   for (int i = 0; i < _packset.length(); i++) {
 820     Node_List* p = _packset.at(i);
 821     for (uint j = 0; j < p->size(); j++) {
 822       Node* s = p->at(j);
 823       assert(my_pack(s) == NULL, "only in one pack");
 824       set_my_pack(s, p);
 825     }
 826   }
 827 }
 828 
 829 //------------------------------filter_packs---------------------------
 830 // Remove packs that are not implemented or not profitable.
 831 void SuperWord::filter_packs() {
 832 
 833   // Remove packs that are not implemented
 834   for (int i = _packset.length() - 1; i >= 0; i--) {
 835     Node_List* pk = _packset.at(i);
 836     bool impl = implemented(pk);
 837     if (!impl) {
 838 #ifndef PRODUCT
 839       if (TraceSuperWord && Verbose) {
 840         tty->print_cr("Unimplemented");
 841         pk->at(0)->dump();
 842       }
 843 #endif
 844       remove_pack_at(i);
 845     }
 846   }
 847 
 848   // Remove packs that are not profitable
 849   bool changed;
 850   do {
 851     changed = false;
 852     for (int i = _packset.length() - 1; i >= 0; i--) {
 853       Node_List* pk = _packset.at(i);
 854       bool prof = profitable(pk);
 855       if (!prof) {
 856 #ifndef PRODUCT
 857         if (TraceSuperWord && Verbose) {
 858           tty->print_cr("Unprofitable");
 859           pk->at(0)->dump();
 860         }
 861 #endif
 862         remove_pack_at(i);
 863         changed = true;
 864       }
 865     }
 866   } while (changed);
 867 
 868 #ifndef PRODUCT
 869   if (TraceSuperWord) {
 870     tty->print_cr("\nAfter filter_packs");
 871     print_packset();
 872     tty->cr();
 873   }
 874 #endif
 875 }
 876 
 877 //------------------------------implemented---------------------------
 878 // Can code be generated for pack p?
 879 bool SuperWord::implemented(Node_List* p) {
 880   Node* p0 = p->at(0);
 881   int vopc = VectorNode::opcode(p0->Opcode(), p->size(), velt_type(p0));
 882   return vopc > 0 && Matcher::has_match_rule(vopc);
 883 }
 884 
 885 //------------------------------profitable---------------------------
 886 // For pack p, are all operands and all uses (with in the block) vector?
 887 bool SuperWord::profitable(Node_List* p) {
 888   Node* p0 = p->at(0);
 889   uint start, end;
 890   vector_opd_range(p0, &start, &end);
 891 
 892   // Return false if some input is not vector and inside block
 893   for (uint i = start; i < end; i++) {
 894     if (!is_vector_use(p0, i)) {
 895       // For now, return false if not scalar promotion case (inputs are the same.)
 896       // Later, implement PackNode and allow differing, non-vector inputs
 897       // (maybe just the ones from outside the block.)
 898       Node* p0_def = p0->in(i);
 899       for (uint j = 1; j < p->size(); j++) {
 900         Node* use = p->at(j);
 901         Node* def = use->in(i);
 902         if (p0_def != def)
 903           return false;
 904       }
 905     }
 906   }
 907   if (!p0->is_Store()) {
 908     // For now, return false if not all uses are vector.
 909     // Later, implement ExtractNode and allow non-vector uses (maybe
 910     // just the ones outside the block.)
 911     for (uint i = 0; i < p->size(); i++) {
 912       Node* def = p->at(i);
 913       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
 914         Node* use = def->fast_out(j);
 915         for (uint k = 0; k < use->req(); k++) {
 916           Node* n = use->in(k);
 917           if (def == n) {
 918             if (!is_vector_use(use, k)) {
 919               return false;
 920             }
 921           }
 922         }
 923       }
 924     }
 925   }
 926   return true;
 927 }
 928 
 929 //------------------------------schedule---------------------------
 930 // Adjust the memory graph for the packed operations
 931 void SuperWord::schedule() {
 932 
 933   // Co-locate in the memory graph the members of each memory pack
 934   for (int i = 0; i < _packset.length(); i++) {
 935     co_locate_pack(_packset.at(i));
 936   }
 937 }
 938 
 939 //-------------------------------remove_and_insert-------------------
 940 //remove "current" from its current position in the memory graph and insert
 941 //it after the appropriate insertion point (lip or uip)
 942 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
 943                                   Node *uip, Unique_Node_List &sched_before) {
 944   Node* my_mem = current->in(MemNode::Memory);
 945   _igvn.hash_delete(current);
 946   _igvn.hash_delete(my_mem);
 947 
 948   //remove current_store from its current position in the memmory graph
 949   for (DUIterator i = current->outs(); current->has_out(i); i++) {
 950     Node* use = current->out(i);
 951     if (use->is_Mem()) {
 952       assert(use->in(MemNode::Memory) == current, "must be");
 953       _igvn.hash_delete(use);
 954       if (use == prev) { // connect prev to my_mem
 955         use->set_req(MemNode::Memory, my_mem);
 956       } else if (sched_before.member(use)) {
 957         _igvn.hash_delete(uip);
 958         use->set_req(MemNode::Memory, uip);
 959       } else {
 960         _igvn.hash_delete(lip);
 961         use->set_req(MemNode::Memory, lip);
 962       }
 963       _igvn._worklist.push(use);
 964       --i; //deleted this edge; rescan position
 965     }
 966   }
 967 
 968   bool sched_up = sched_before.member(current);
 969   Node *insert_pt =  sched_up ?  uip : lip;
 970   _igvn.hash_delete(insert_pt);
 971 
 972   // all uses of insert_pt's memory state should use current's instead
 973   for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
 974     Node* use = insert_pt->out(i);
 975     if (use->is_Mem()) {
 976       assert(use->in(MemNode::Memory) == insert_pt, "must be");
 977       _igvn.hash_delete(use);
 978       use->set_req(MemNode::Memory, current);
 979       _igvn._worklist.push(use);
 980       --i; //deleted this edge; rescan position
 981     } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
 982       uint pos; //lip (lower insert point) must be the last one in the memory slice
 983       _igvn.hash_delete(use);
 984       for (pos=1; pos < use->req(); pos++) {
 985         if (use->in(pos) == insert_pt) break;
 986       }
 987       use->set_req(pos, current);
 988       _igvn._worklist.push(use);
 989       --i;
 990     }
 991   }
 992 
 993   //connect current to insert_pt
 994   current->set_req(MemNode::Memory, insert_pt);
 995   _igvn._worklist.push(current);
 996 }
 997 
 998 //------------------------------co_locate_pack----------------------------------
 999 // To schedule a store pack, we need to move any sandwiched memory ops either before
1000 // or after the pack, based upon dependence information:
1001 // (1) If any store in the pack depends on the sandwiched memory op, the
1002 //     sandwiched memory op must be scheduled BEFORE the pack;
1003 // (2) If a sandwiched memory op depends on any store in the pack, the
1004 //     sandwiched memory op must be scheduled AFTER the pack;
1005 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1006 //     memory op (say memB), memB must be scheduled before memA. So, if memA is
1007 //     scheduled before the pack, memB must also be scheduled before the pack;
1008 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1009 //     schedule this store AFTER the pack
1010 // (5) We know there is no dependence cycle, so there in no other case;
1011 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1012 //
1013 // To schedule a load pack, we use the memory state of either the first or the last load in
1014 // the pack, based on the dependence constraint.
1015 void SuperWord::co_locate_pack(Node_List* pk) {
1016   if (pk->at(0)->is_Store()) {
1017     MemNode* first     = executed_first(pk)->as_Mem();
1018     MemNode* last      = executed_last(pk)->as_Mem();
1019     Unique_Node_List schedule_before_pack;
1020     Unique_Node_List memops;
1021 
1022     MemNode* current   = last->in(MemNode::Memory)->as_Mem();
1023     MemNode* previous  = last;
1024     while (true) {
1025       assert(in_bb(current), "stay in block");
1026       memops.push(previous);
1027       for (DUIterator i = current->outs(); current->has_out(i); i++) {
1028         Node* use = current->out(i);
1029         if (use->is_Mem() && use != previous)
1030           memops.push(use);
1031       }
1032       if(current == first) break;
1033       previous = current;
1034       current  = current->in(MemNode::Memory)->as_Mem();
1035     }
1036 
1037     // determine which memory operations should be scheduled before the pack
1038     for (uint i = 1; i < memops.size(); i++) {
1039       Node *s1 = memops.at(i);
1040       if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1041         for (uint j = 0; j< i; j++) {
1042           Node *s2 = memops.at(j);
1043           if (!independent(s1, s2)) {
1044             if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1045               schedule_before_pack.push(s1); //s1 must be scheduled before
1046               Node_List* mem_pk = my_pack(s1);
1047               if (mem_pk != NULL) {
1048                 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1049                   Node* s = mem_pk->at(ii); // follow partner
1050                   if (memops.member(s) && !schedule_before_pack.member(s))
1051                     schedule_before_pack.push(s);
1052                 }
1053               }
1054             }
1055           }
1056         }
1057       }
1058     }
1059 
1060     MemNode* lower_insert_pt = last;
1061     Node*    upper_insert_pt = first->in(MemNode::Memory);
1062     previous                 = last; //previous store in pk
1063     current                  = last->in(MemNode::Memory)->as_Mem();
1064 
1065     //start scheduling from "last" to "first"
1066     while (true) {
1067       assert(in_bb(current), "stay in block");
1068       assert(in_pack(previous, pk), "previous stays in pack");
1069       Node* my_mem = current->in(MemNode::Memory);
1070 
1071       if (in_pack(current, pk)) {
1072         // Forward users of my memory state (except "previous) to my input memory state
1073         _igvn.hash_delete(current);
1074         for (DUIterator i = current->outs(); current->has_out(i); i++) {
1075           Node* use = current->out(i);
1076           if (use->is_Mem() && use != previous) {
1077             assert(use->in(MemNode::Memory) == current, "must be");
1078             _igvn.hash_delete(use);
1079             if (schedule_before_pack.member(use)) {
1080               _igvn.hash_delete(upper_insert_pt);
1081               use->set_req(MemNode::Memory, upper_insert_pt);
1082             } else {
1083               _igvn.hash_delete(lower_insert_pt);
1084               use->set_req(MemNode::Memory, lower_insert_pt);
1085             }
1086             _igvn._worklist.push(use);
1087             --i; // deleted this edge; rescan position
1088           }
1089         }
1090         previous = current;
1091       } else { // !in_pack(current, pk) ==> a sandwiched store
1092         remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1093       }
1094 
1095       if (current == first) break;
1096       current = my_mem->as_Mem();
1097     } // end while
1098   } else if (pk->at(0)->is_Load()) { //load
1099     // all loads in the pack should have the same memory state. By default,
1100     // we use the memory state of the last load. However, if any load could
1101     // not be moved down due to the dependence constraint, we use the memory
1102     // state of the first load.
1103     Node* last_mem  = executed_last(pk)->in(MemNode::Memory);
1104     Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1105     bool schedule_last = true;
1106     for (uint i = 0; i < pk->size(); i++) {
1107       Node* ld = pk->at(i);
1108       for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1109            current=current->in(MemNode::Memory)) {
1110         assert(current != first_mem, "corrupted memory graph");
1111         if(current->is_Mem() && !independent(current, ld)){
1112           schedule_last = false; // a later store depends on this load
1113           break;
1114         }
1115       }
1116     }
1117 
1118     Node* mem_input = schedule_last ? last_mem : first_mem;
1119     _igvn.hash_delete(mem_input);
1120     // Give each load the same memory state
1121     for (uint i = 0; i < pk->size(); i++) {
1122       LoadNode* ld = pk->at(i)->as_Load();
1123       _igvn.hash_delete(ld);
1124       ld->set_req(MemNode::Memory, mem_input);
1125       _igvn._worklist.push(ld);
1126     }
1127   }
1128 }
1129 
1130 //------------------------------output---------------------------
1131 // Convert packs into vector node operations
1132 void SuperWord::output() {
1133   if (_packset.length() == 0) return;
1134 
1135 #ifndef PRODUCT
1136   if (TraceLoopOpts) {
1137     tty->print("SuperWord    ");
1138     lpt()->dump_head();
1139   }
1140 #endif
1141 
1142   // MUST ENSURE main loop's initial value is properly aligned:
1143   //  (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1144 
1145   align_initial_loop_index(align_to_ref());
1146 
1147   // Insert extract (unpack) operations for scalar uses
1148   for (int i = 0; i < _packset.length(); i++) {
1149     insert_extracts(_packset.at(i));
1150   }
1151 
1152   for (int i = 0; i < _block.length(); i++) {
1153     Node* n = _block.at(i);
1154     Node_List* p = my_pack(n);
1155     if (p && n == executed_last(p)) {
1156       uint vlen = p->size();
1157       Node* vn = NULL;
1158       Node* low_adr = p->at(0);
1159       Node* first   = executed_first(p);
1160       if (n->is_Load()) {
1161         int   opc = n->Opcode();
1162         Node* ctl = n->in(MemNode::Control);
1163         Node* mem = first->in(MemNode::Memory);
1164         Node* adr = low_adr->in(MemNode::Address);
1165         const TypePtr* atyp = n->adr_type();
1166         vn = VectorLoadNode::make(_phase->C, opc, ctl, mem, adr, atyp, vlen);
1167 
1168       } else if (n->is_Store()) {
1169         // Promote value to be stored to vector
1170         Node* val = vector_opd(p, MemNode::ValueIn);
1171 
1172         int   opc = n->Opcode();
1173         Node* ctl = n->in(MemNode::Control);
1174         Node* mem = first->in(MemNode::Memory);
1175         Node* adr = low_adr->in(MemNode::Address);
1176         const TypePtr* atyp = n->adr_type();
1177         vn = VectorStoreNode::make(_phase->C, opc, ctl, mem, adr, atyp, val, vlen);
1178 
1179       } else if (n->req() == 3) {
1180         // Promote operands to vector
1181         Node* in1 = vector_opd(p, 1);
1182         Node* in2 = vector_opd(p, 2);
1183         vn = VectorNode::make(_phase->C, n->Opcode(), in1, in2, vlen, velt_type(n));
1184 
1185       } else {
1186         ShouldNotReachHere();
1187       }
1188 
1189       _phase->_igvn.register_new_node_with_optimizer(vn);
1190       _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1191       for (uint j = 0; j < p->size(); j++) {
1192         Node* pm = p->at(j);
1193         _igvn.replace_node(pm, vn);
1194       }
1195       _igvn._worklist.push(vn);
1196     }
1197   }
1198 }
1199 
1200 //------------------------------vector_opd---------------------------
1201 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1202 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1203   Node* p0 = p->at(0);
1204   uint vlen = p->size();
1205   Node* opd = p0->in(opd_idx);
1206 
1207   bool same_opd = true;
1208   for (uint i = 1; i < vlen; i++) {
1209     Node* pi = p->at(i);
1210     Node* in = pi->in(opd_idx);
1211     if (opd != in) {
1212       same_opd = false;
1213       break;
1214     }
1215   }
1216 
1217   if (same_opd) {
1218     if (opd->is_Vector() || opd->is_VectorLoad()) {
1219       return opd; // input is matching vector
1220     }
1221     assert(!opd->is_VectorStore(), "such vector is not expected here");
1222     // Convert scalar input to vector. Use p0's type because it's container
1223     // maybe smaller than the operand's container.
1224     const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd);
1225     const Type* p0_t  = velt_type(p0);
1226     if (p0_t->higher_equal(opd_t)) opd_t = p0_t;
1227     VectorNode* vn    = VectorNode::scalar2vector(_phase->C, opd, vlen, opd_t);
1228 
1229     _phase->_igvn.register_new_node_with_optimizer(vn);
1230     _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1231     return vn;
1232   }
1233 
1234   // Insert pack operation
1235   const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd);
1236   PackNode* pk = PackNode::make(_phase->C, opd, opd_t);
1237 
1238   for (uint i = 1; i < vlen; i++) {
1239     Node* pi = p->at(i);
1240     Node* in = pi->in(opd_idx);
1241     assert(my_pack(in) == NULL, "Should already have been unpacked");
1242     assert(opd_t == velt_type(!in_bb(in) ? pi : in), "all same type");
1243     pk->add_opd(in);
1244   }
1245   _phase->_igvn.register_new_node_with_optimizer(pk);
1246   _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1247   return pk;
1248 }
1249 
1250 //------------------------------insert_extracts---------------------------
1251 // If a use of pack p is not a vector use, then replace the
1252 // use with an extract operation.
1253 void SuperWord::insert_extracts(Node_List* p) {
1254   if (p->at(0)->is_Store()) return;
1255   assert(_n_idx_list.is_empty(), "empty (node,index) list");
1256 
1257   // Inspect each use of each pack member.  For each use that is
1258   // not a vector use, replace the use with an extract operation.
1259 
1260   for (uint i = 0; i < p->size(); i++) {
1261     Node* def = p->at(i);
1262     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1263       Node* use = def->fast_out(j);
1264       for (uint k = 0; k < use->req(); k++) {
1265         Node* n = use->in(k);
1266         if (def == n) {
1267           if (!is_vector_use(use, k)) {
1268             _n_idx_list.push(use, k);
1269           }
1270         }
1271       }
1272     }
1273   }
1274 
1275   while (_n_idx_list.is_nonempty()) {
1276     Node* use = _n_idx_list.node();
1277     int   idx = _n_idx_list.index();
1278     _n_idx_list.pop();
1279     Node* def = use->in(idx);
1280 
1281     // Insert extract operation
1282     _igvn.hash_delete(def);
1283     _igvn.hash_delete(use);
1284     int def_pos = alignment(def) / data_size(def);
1285     const Type* def_t = velt_type(def);
1286 
1287     Node* ex = ExtractNode::make(_phase->C, def, def_pos, def_t);
1288     _phase->_igvn.register_new_node_with_optimizer(ex);
1289     _phase->set_ctrl(ex, _phase->get_ctrl(def));
1290     use->set_req(idx, ex);
1291     _igvn._worklist.push(def);
1292     _igvn._worklist.push(use);
1293 
1294     bb_insert_after(ex, bb_idx(def));
1295     set_velt_type(ex, def_t);
1296   }
1297 }
1298 
1299 //------------------------------is_vector_use---------------------------
1300 // Is use->in(u_idx) a vector use?
1301 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1302   Node_List* u_pk = my_pack(use);
1303   if (u_pk == NULL) return false;
1304   Node* def = use->in(u_idx);
1305   Node_List* d_pk = my_pack(def);
1306   if (d_pk == NULL) {
1307     // check for scalar promotion
1308     Node* n = u_pk->at(0)->in(u_idx);
1309     for (uint i = 1; i < u_pk->size(); i++) {
1310       if (u_pk->at(i)->in(u_idx) != n) return false;
1311     }
1312     return true;
1313   }
1314   if (u_pk->size() != d_pk->size())
1315     return false;
1316   for (uint i = 0; i < u_pk->size(); i++) {
1317     Node* ui = u_pk->at(i);
1318     Node* di = d_pk->at(i);
1319     if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1320       return false;
1321   }
1322   return true;
1323 }
1324 
1325 //------------------------------construct_bb---------------------------
1326 // Construct reverse postorder list of block members
1327 void SuperWord::construct_bb() {
1328   Node* entry = bb();
1329 
1330   assert(_stk.length() == 0,            "stk is empty");
1331   assert(_block.length() == 0,          "block is empty");
1332   assert(_data_entry.length() == 0,     "data_entry is empty");
1333   assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1334   assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1335 
1336   // Find non-control nodes with no inputs from within block,
1337   // create a temporary map from node _idx to bb_idx for use
1338   // by the visited and post_visited sets,
1339   // and count number of nodes in block.
1340   int bb_ct = 0;
1341   for (uint i = 0; i < lpt()->_body.size(); i++ ) {
1342     Node *n = lpt()->_body.at(i);
1343     set_bb_idx(n, i); // Create a temporary map
1344     if (in_bb(n)) {
1345       bb_ct++;
1346       if (!n->is_CFG()) {
1347         bool found = false;
1348         for (uint j = 0; j < n->req(); j++) {
1349           Node* def = n->in(j);
1350           if (def && in_bb(def)) {
1351             found = true;
1352             break;
1353           }
1354         }
1355         if (!found) {
1356           assert(n != entry, "can't be entry");
1357           _data_entry.push(n);
1358         }
1359       }
1360     }
1361   }
1362 
1363   // Find memory slices (head and tail)
1364   for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1365     Node *n = lp()->fast_out(i);
1366     if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1367       Node* n_tail  = n->in(LoopNode::LoopBackControl);
1368       if (n_tail != n->in(LoopNode::EntryControl)) {
1369         _mem_slice_head.push(n);
1370         _mem_slice_tail.push(n_tail);
1371       }
1372     }
1373   }
1374 
1375   // Create an RPO list of nodes in block
1376 
1377   visited_clear();
1378   post_visited_clear();
1379 
1380   // Push all non-control nodes with no inputs from within block, then control entry
1381   for (int j = 0; j < _data_entry.length(); j++) {
1382     Node* n = _data_entry.at(j);
1383     visited_set(n);
1384     _stk.push(n);
1385   }
1386   visited_set(entry);
1387   _stk.push(entry);
1388 
1389   // Do a depth first walk over out edges
1390   int rpo_idx = bb_ct - 1;
1391   int size;
1392   while ((size = _stk.length()) > 0) {
1393     Node* n = _stk.top(); // Leave node on stack
1394     if (!visited_test_set(n)) {
1395       // forward arc in graph
1396     } else if (!post_visited_test(n)) {
1397       // cross or back arc
1398       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1399         Node *use = n->fast_out(i);
1400         if (in_bb(use) && !visited_test(use) &&
1401             // Don't go around backedge
1402             (!use->is_Phi() || n == entry)) {
1403           _stk.push(use);
1404         }
1405       }
1406       if (_stk.length() == size) {
1407         // There were no additional uses, post visit node now
1408         _stk.pop(); // Remove node from stack
1409         assert(rpo_idx >= 0, "");
1410         _block.at_put_grow(rpo_idx, n);
1411         rpo_idx--;
1412         post_visited_set(n);
1413         assert(rpo_idx >= 0 || _stk.is_empty(), "");
1414       }
1415     } else {
1416       _stk.pop(); // Remove post-visited node from stack
1417     }
1418   }
1419 
1420   // Create real map of block indices for nodes
1421   for (int j = 0; j < _block.length(); j++) {
1422     Node* n = _block.at(j);
1423     set_bb_idx(n, j);
1424   }
1425 
1426   initialize_bb(); // Ensure extra info is allocated.
1427 
1428 #ifndef PRODUCT
1429   if (TraceSuperWord) {
1430     print_bb();
1431     tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1432     for (int m = 0; m < _data_entry.length(); m++) {
1433       tty->print("%3d ", m);
1434       _data_entry.at(m)->dump();
1435     }
1436     tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1437     for (int m = 0; m < _mem_slice_head.length(); m++) {
1438       tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1439       tty->print("    ");    _mem_slice_tail.at(m)->dump();
1440     }
1441   }
1442 #endif
1443   assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1444 }
1445 
1446 //------------------------------initialize_bb---------------------------
1447 // Initialize per node info
1448 void SuperWord::initialize_bb() {
1449   Node* last = _block.at(_block.length() - 1);
1450   grow_node_info(bb_idx(last));
1451 }
1452 
1453 //------------------------------bb_insert_after---------------------------
1454 // Insert n into block after pos
1455 void SuperWord::bb_insert_after(Node* n, int pos) {
1456   int n_pos = pos + 1;
1457   // Make room
1458   for (int i = _block.length() - 1; i >= n_pos; i--) {
1459     _block.at_put_grow(i+1, _block.at(i));
1460   }
1461   for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1462     _node_info.at_put_grow(j+1, _node_info.at(j));
1463   }
1464   // Set value
1465   _block.at_put_grow(n_pos, n);
1466   _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1467   // Adjust map from node->_idx to _block index
1468   for (int i = n_pos; i < _block.length(); i++) {
1469     set_bb_idx(_block.at(i), i);
1470   }
1471 }
1472 
1473 //------------------------------compute_max_depth---------------------------
1474 // Compute max depth for expressions from beginning of block
1475 // Use to prune search paths during test for independence.
1476 void SuperWord::compute_max_depth() {
1477   int ct = 0;
1478   bool again;
1479   do {
1480     again = false;
1481     for (int i = 0; i < _block.length(); i++) {
1482       Node* n = _block.at(i);
1483       if (!n->is_Phi()) {
1484         int d_orig = depth(n);
1485         int d_in   = 0;
1486         for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1487           Node* pred = preds.current();
1488           if (in_bb(pred)) {
1489             d_in = MAX2(d_in, depth(pred));
1490           }
1491         }
1492         if (d_in + 1 != d_orig) {
1493           set_depth(n, d_in + 1);
1494           again = true;
1495         }
1496       }
1497     }
1498     ct++;
1499   } while (again);
1500 #ifndef PRODUCT
1501   if (TraceSuperWord && Verbose)
1502     tty->print_cr("compute_max_depth iterated: %d times", ct);
1503 #endif
1504 }
1505 
1506 //-------------------------compute_vector_element_type-----------------------
1507 // Compute necessary vector element type for expressions
1508 // This propagates backwards a narrower integer type when the
1509 // upper bits of the value are not needed.
1510 // Example:  char a,b,c;  a = b + c;
1511 // Normally the type of the add is integer, but for packed character
1512 // operations the type of the add needs to be char.
1513 void SuperWord::compute_vector_element_type() {
1514 #ifndef PRODUCT
1515   if (TraceSuperWord && Verbose)
1516     tty->print_cr("\ncompute_velt_type:");
1517 #endif
1518 
1519   // Initial type
1520   for (int i = 0; i < _block.length(); i++) {
1521     Node* n = _block.at(i);
1522     const Type* t  = n->is_Mem() ? Type::get_const_basic_type(n->as_Mem()->memory_type())
1523                                  : _igvn.type(n);
1524     const Type* vt = container_type(t);
1525     set_velt_type(n, vt);
1526   }
1527 
1528   // Propagate narrowed type backwards through operations
1529   // that don't depend on higher order bits
1530   for (int i = _block.length() - 1; i >= 0; i--) {
1531     Node* n = _block.at(i);
1532     // Only integer types need be examined
1533     if (n->bottom_type()->isa_int()) {
1534       uint start, end;
1535       vector_opd_range(n, &start, &end);
1536       const Type* vt = velt_type(n);
1537 
1538       for (uint j = start; j < end; j++) {
1539         Node* in  = n->in(j);
1540         // Don't propagate through a type conversion
1541         if (n->bottom_type() != in->bottom_type())
1542           continue;
1543         switch(in->Opcode()) {
1544         case Op_AddI:    case Op_AddL:
1545         case Op_SubI:    case Op_SubL:
1546         case Op_MulI:    case Op_MulL:
1547         case Op_AndI:    case Op_AndL:
1548         case Op_OrI:     case Op_OrL:
1549         case Op_XorI:    case Op_XorL:
1550         case Op_LShiftI: case Op_LShiftL:
1551         case Op_CMoveI:  case Op_CMoveL:
1552           if (in_bb(in)) {
1553             bool same_type = true;
1554             for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1555               Node *use = in->fast_out(k);
1556               if (!in_bb(use) || velt_type(use) != vt) {
1557                 same_type = false;
1558                 break;
1559               }
1560             }
1561             if (same_type) {
1562               set_velt_type(in, vt);
1563             }
1564           }
1565         }
1566       }
1567     }
1568   }
1569 #ifndef PRODUCT
1570   if (TraceSuperWord && Verbose) {
1571     for (int i = 0; i < _block.length(); i++) {
1572       Node* n = _block.at(i);
1573       velt_type(n)->dump();
1574       tty->print("\t");
1575       n->dump();
1576     }
1577   }
1578 #endif
1579 }
1580 
1581 //------------------------------memory_alignment---------------------------
1582 // Alignment within a vector memory reference
1583 int SuperWord::memory_alignment(MemNode* s, int iv_adjust_in_bytes) {
1584   SWPointer p(s, this);
1585   if (!p.valid()) {
1586     return bottom_align;
1587   }
1588   int offset  = p.offset_in_bytes();
1589   offset     += iv_adjust_in_bytes;
1590   int off_rem = offset % vector_width_in_bytes();
1591   int off_mod = off_rem >= 0 ? off_rem : off_rem + vector_width_in_bytes();
1592   return off_mod;
1593 }
1594 
1595 //---------------------------container_type---------------------------
1596 // Smallest type containing range of values
1597 const Type* SuperWord::container_type(const Type* t) {
1598   const Type* tp = t->make_ptr();
1599   if (tp && tp->isa_aryptr()) {
1600     t = tp->is_aryptr()->elem();
1601   }
1602   if (t->basic_type() == T_INT) {
1603     if (t->higher_equal(TypeInt::BOOL))  return TypeInt::BOOL;
1604     if (t->higher_equal(TypeInt::BYTE))  return TypeInt::BYTE;
1605     if (t->higher_equal(TypeInt::CHAR))  return TypeInt::CHAR;
1606     if (t->higher_equal(TypeInt::SHORT)) return TypeInt::SHORT;
1607     return TypeInt::INT;
1608   }
1609   return t;
1610 }
1611 
1612 //-------------------------vector_opd_range-----------------------
1613 // (Start, end] half-open range defining which operands are vector
1614 void SuperWord::vector_opd_range(Node* n, uint* start, uint* end) {
1615   switch (n->Opcode()) {
1616   case Op_LoadB:   case Op_LoadUS:
1617   case Op_LoadI:   case Op_LoadL:
1618   case Op_LoadF:   case Op_LoadD:
1619   case Op_LoadP:
1620     *start = 0;
1621     *end   = 0;
1622     return;
1623   case Op_StoreB:  case Op_StoreC:
1624   case Op_StoreI:  case Op_StoreL:
1625   case Op_StoreF:  case Op_StoreD:
1626   case Op_StoreP:
1627     *start = MemNode::ValueIn;
1628     *end   = *start + 1;
1629     return;
1630   case Op_LShiftI: case Op_LShiftL:
1631     *start = 1;
1632     *end   = 2;
1633     return;
1634   case Op_CMoveI:  case Op_CMoveL:  case Op_CMoveF:  case Op_CMoveD:
1635     *start = 2;
1636     *end   = n->req();
1637     return;
1638   }
1639   *start = 1;
1640   *end   = n->req(); // default is all operands
1641 }
1642 
1643 //------------------------------in_packset---------------------------
1644 // Are s1 and s2 in a pack pair and ordered as s1,s2?
1645 bool SuperWord::in_packset(Node* s1, Node* s2) {
1646   for (int i = 0; i < _packset.length(); i++) {
1647     Node_List* p = _packset.at(i);
1648     assert(p->size() == 2, "must be");
1649     if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
1650       return true;
1651     }
1652   }
1653   return false;
1654 }
1655 
1656 //------------------------------in_pack---------------------------
1657 // Is s in pack p?
1658 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
1659   for (uint i = 0; i < p->size(); i++) {
1660     if (p->at(i) == s) {
1661       return p;
1662     }
1663   }
1664   return NULL;
1665 }
1666 
1667 //------------------------------remove_pack_at---------------------------
1668 // Remove the pack at position pos in the packset
1669 void SuperWord::remove_pack_at(int pos) {
1670   Node_List* p = _packset.at(pos);
1671   for (uint i = 0; i < p->size(); i++) {
1672     Node* s = p->at(i);
1673     set_my_pack(s, NULL);
1674   }
1675   _packset.remove_at(pos);
1676 }
1677 
1678 //------------------------------executed_first---------------------------
1679 // Return the node executed first in pack p.  Uses the RPO block list
1680 // to determine order.
1681 Node* SuperWord::executed_first(Node_List* p) {
1682   Node* n = p->at(0);
1683   int n_rpo = bb_idx(n);
1684   for (uint i = 1; i < p->size(); i++) {
1685     Node* s = p->at(i);
1686     int s_rpo = bb_idx(s);
1687     if (s_rpo < n_rpo) {
1688       n = s;
1689       n_rpo = s_rpo;
1690     }
1691   }
1692   return n;
1693 }
1694 
1695 //------------------------------executed_last---------------------------
1696 // Return the node executed last in pack p.
1697 Node* SuperWord::executed_last(Node_List* p) {
1698   Node* n = p->at(0);
1699   int n_rpo = bb_idx(n);
1700   for (uint i = 1; i < p->size(); i++) {
1701     Node* s = p->at(i);
1702     int s_rpo = bb_idx(s);
1703     if (s_rpo > n_rpo) {
1704       n = s;
1705       n_rpo = s_rpo;
1706     }
1707   }
1708   return n;
1709 }
1710 
1711 //----------------------------align_initial_loop_index---------------------------
1712 // Adjust pre-loop limit so that in main loop, a load/store reference
1713 // to align_to_ref will be a position zero in the vector.
1714 //   (iv + k) mod vector_align == 0
1715 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
1716   CountedLoopNode *main_head = lp()->as_CountedLoop();
1717   assert(main_head->is_main_loop(), "");
1718   CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
1719   assert(pre_end != NULL, "");
1720   Node *pre_opaq1 = pre_end->limit();
1721   assert(pre_opaq1->Opcode() == Op_Opaque1, "");
1722   Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1723   Node *lim0 = pre_opaq->in(1);
1724 
1725   // Where we put new limit calculations
1726   Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1727 
1728   // Ensure the original loop limit is available from the
1729   // pre-loop Opaque1 node.
1730   Node *orig_limit = pre_opaq->original_loop_limit();
1731   assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
1732 
1733   SWPointer align_to_ref_p(align_to_ref, this);
1734 
1735   // Given:
1736   //     lim0 == original pre loop limit
1737   //     V == v_align (power of 2)
1738   //     invar == extra invariant piece of the address expression
1739   //     e == k [ +/- invar ]
1740   //
1741   // When reassociating expressions involving '%' the basic rules are:
1742   //     (a - b) % k == 0   =>  a % k == b % k
1743   // and:
1744   //     (a + b) % k == 0   =>  a % k == (k - b) % k
1745   //
1746   // For stride > 0 && scale > 0,
1747   //   Derive the new pre-loop limit "lim" such that the two constraints:
1748   //     (1) lim = lim0 + N           (where N is some positive integer < V)
1749   //     (2) (e + lim) % V == 0
1750   //   are true.
1751   //
1752   //   Substituting (1) into (2),
1753   //     (e + lim0 + N) % V == 0
1754   //   solve for N:
1755   //     N = (V - (e + lim0)) % V
1756   //   substitute back into (1), so that new limit
1757   //     lim = lim0 + (V - (e + lim0)) % V
1758   //
1759   // For stride > 0 && scale < 0
1760   //   Constraints:
1761   //     lim = lim0 + N
1762   //     (e - lim) % V == 0
1763   //   Solving for lim:
1764   //     (e - lim0 - N) % V == 0
1765   //     N = (e - lim0) % V
1766   //     lim = lim0 + (e - lim0) % V
1767   //
1768   // For stride < 0 && scale > 0
1769   //   Constraints:
1770   //     lim = lim0 - N
1771   //     (e + lim) % V == 0
1772   //   Solving for lim:
1773   //     (e + lim0 - N) % V == 0
1774   //     N = (e + lim0) % V
1775   //     lim = lim0 - (e + lim0) % V
1776   //
1777   // For stride < 0 && scale < 0
1778   //   Constraints:
1779   //     lim = lim0 - N
1780   //     (e - lim) % V == 0
1781   //   Solving for lim:
1782   //     (e - lim0 + N) % V == 0
1783   //     N = (V - (e - lim0)) % V
1784   //     lim = lim0 - (V - (e - lim0)) % V
1785 
1786   int stride   = iv_stride();
1787   int scale    = align_to_ref_p.scale_in_bytes();
1788   int elt_size = align_to_ref_p.memory_size();
1789   int v_align  = vector_width_in_bytes() / elt_size;
1790   int k        = align_to_ref_p.offset_in_bytes() / elt_size;
1791 
1792   Node *kn   = _igvn.intcon(k);
1793 
1794   Node *e = kn;
1795   if (align_to_ref_p.invar() != NULL) {
1796     // incorporate any extra invariant piece producing k +/- invar >>> log2(elt)
1797     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
1798     Node* aref     = new (_phase->C, 3) URShiftINode(align_to_ref_p.invar(), log2_elt);
1799     _phase->_igvn.register_new_node_with_optimizer(aref);
1800     _phase->set_ctrl(aref, pre_ctrl);
1801     if (align_to_ref_p.negate_invar()) {
1802       e = new (_phase->C, 3) SubINode(e, aref);
1803     } else {
1804       e = new (_phase->C, 3) AddINode(e, aref);
1805     }
1806     _phase->_igvn.register_new_node_with_optimizer(e);
1807     _phase->set_ctrl(e, pre_ctrl);
1808   }
1809 
1810   // compute e +/- lim0
1811   if (scale < 0) {
1812     e = new (_phase->C, 3) SubINode(e, lim0);
1813   } else {
1814     e = new (_phase->C, 3) AddINode(e, lim0);
1815   }
1816   _phase->_igvn.register_new_node_with_optimizer(e);
1817   _phase->set_ctrl(e, pre_ctrl);
1818 
1819   if (stride * scale > 0) {
1820     // compute V - (e +/- lim0)
1821     Node* va  = _igvn.intcon(v_align);
1822     e = new (_phase->C, 3) SubINode(va, e);
1823     _phase->_igvn.register_new_node_with_optimizer(e);
1824     _phase->set_ctrl(e, pre_ctrl);
1825   }
1826   // compute N = (exp) % V
1827   Node* va_msk = _igvn.intcon(v_align - 1);
1828   Node* N = new (_phase->C, 3) AndINode(e, va_msk);
1829   _phase->_igvn.register_new_node_with_optimizer(N);
1830   _phase->set_ctrl(N, pre_ctrl);
1831 
1832   //   substitute back into (1), so that new limit
1833   //     lim = lim0 + N
1834   Node* lim;
1835   if (stride < 0) {
1836     lim = new (_phase->C, 3) SubINode(lim0, N);
1837   } else {
1838     lim = new (_phase->C, 3) AddINode(lim0, N);
1839   }
1840   _phase->_igvn.register_new_node_with_optimizer(lim);
1841   _phase->set_ctrl(lim, pre_ctrl);
1842   Node* constrained =
1843     (stride > 0) ? (Node*) new (_phase->C,3) MinINode(lim, orig_limit)
1844                  : (Node*) new (_phase->C,3) MaxINode(lim, orig_limit);
1845   _phase->_igvn.register_new_node_with_optimizer(constrained);
1846   _phase->set_ctrl(constrained, pre_ctrl);
1847   _igvn.hash_delete(pre_opaq);
1848   pre_opaq->set_req(1, constrained);
1849 }
1850 
1851 //----------------------------get_pre_loop_end---------------------------
1852 // Find pre loop end from main loop.  Returns null if none.
1853 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
1854   Node *ctrl = cl->in(LoopNode::EntryControl);
1855   if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
1856   Node *iffm = ctrl->in(0);
1857   if (!iffm->is_If()) return NULL;
1858   Node *p_f = iffm->in(0);
1859   if (!p_f->is_IfFalse()) return NULL;
1860   if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
1861   CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1862   if (!pre_end->loopnode()->is_pre_loop()) return NULL;
1863   return pre_end;
1864 }
1865 
1866 
1867 //------------------------------init---------------------------
1868 void SuperWord::init() {
1869   _dg.init();
1870   _packset.clear();
1871   _disjoint_ptrs.clear();
1872   _block.clear();
1873   _data_entry.clear();
1874   _mem_slice_head.clear();
1875   _mem_slice_tail.clear();
1876   _node_info.clear();
1877   _align_to_ref = NULL;
1878   _lpt = NULL;
1879   _lp = NULL;
1880   _bb = NULL;
1881   _iv = NULL;
1882 }
1883 
1884 //------------------------------print_packset---------------------------
1885 void SuperWord::print_packset() {
1886 #ifndef PRODUCT
1887   tty->print_cr("packset");
1888   for (int i = 0; i < _packset.length(); i++) {
1889     tty->print_cr("Pack: %d", i);
1890     Node_List* p = _packset.at(i);
1891     print_pack(p);
1892   }
1893 #endif
1894 }
1895 
1896 //------------------------------print_pack---------------------------
1897 void SuperWord::print_pack(Node_List* p) {
1898   for (uint i = 0; i < p->size(); i++) {
1899     print_stmt(p->at(i));
1900   }
1901 }
1902 
1903 //------------------------------print_bb---------------------------
1904 void SuperWord::print_bb() {
1905 #ifndef PRODUCT
1906   tty->print_cr("\nBlock");
1907   for (int i = 0; i < _block.length(); i++) {
1908     Node* n = _block.at(i);
1909     tty->print("%d ", i);
1910     if (n) {
1911       n->dump();
1912     }
1913   }
1914 #endif
1915 }
1916 
1917 //------------------------------print_stmt---------------------------
1918 void SuperWord::print_stmt(Node* s) {
1919 #ifndef PRODUCT
1920   tty->print(" align: %d \t", alignment(s));
1921   s->dump();
1922 #endif
1923 }
1924 
1925 //------------------------------blank---------------------------
1926 char* SuperWord::blank(uint depth) {
1927   static char blanks[101];
1928   assert(depth < 101, "too deep");
1929   for (uint i = 0; i < depth; i++) blanks[i] = ' ';
1930   blanks[depth] = '\0';
1931   return blanks;
1932 }
1933 
1934 
1935 //==============================SWPointer===========================
1936 
1937 //----------------------------SWPointer------------------------
1938 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
1939   _mem(mem), _slp(slp),  _base(NULL),  _adr(NULL),
1940   _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
1941 
1942   Node* adr = mem->in(MemNode::Address);
1943   if (!adr->is_AddP()) {
1944     assert(!valid(), "too complex");
1945     return;
1946   }
1947   // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
1948   Node* base = adr->in(AddPNode::Base);
1949   //unsafe reference could not be aligned appropriately without runtime checking
1950   if (base == NULL || base->bottom_type() == Type::TOP) {
1951     assert(!valid(), "unsafe access");
1952     return;
1953   }
1954   for (int i = 0; i < 3; i++) {
1955     if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
1956       assert(!valid(), "too complex");
1957       return;
1958     }
1959     adr = adr->in(AddPNode::Address);
1960     if (base == adr || !adr->is_AddP()) {
1961       break; // stop looking at addp's
1962     }
1963   }
1964   _base = base;
1965   _adr  = adr;
1966   assert(valid(), "Usable");
1967 }
1968 
1969 // Following is used to create a temporary object during
1970 // the pattern match of an address expression.
1971 SWPointer::SWPointer(SWPointer* p) :
1972   _mem(p->_mem), _slp(p->_slp),  _base(NULL),  _adr(NULL),
1973   _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
1974 
1975 //------------------------scaled_iv_plus_offset--------------------
1976 // Match: k*iv + offset
1977 // where: k is a constant that maybe zero, and
1978 //        offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
1979 bool SWPointer::scaled_iv_plus_offset(Node* n) {
1980   if (scaled_iv(n)) {
1981     return true;
1982   }
1983   if (offset_plus_k(n)) {
1984     return true;
1985   }
1986   int opc = n->Opcode();
1987   if (opc == Op_AddI) {
1988     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
1989       return true;
1990     }
1991     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
1992       return true;
1993     }
1994   } else if (opc == Op_SubI) {
1995     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
1996       return true;
1997     }
1998     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
1999       _scale *= -1;
2000       return true;
2001     }
2002   }
2003   return false;
2004 }
2005 
2006 //----------------------------scaled_iv------------------------
2007 // Match: k*iv where k is a constant that's not zero
2008 bool SWPointer::scaled_iv(Node* n) {
2009   if (_scale != 0) {
2010     return false;  // already found a scale
2011   }
2012   if (n == iv()) {
2013     _scale = 1;
2014     return true;
2015   }
2016   int opc = n->Opcode();
2017   if (opc == Op_MulI) {
2018     if (n->in(1) == iv() && n->in(2)->is_Con()) {
2019       _scale = n->in(2)->get_int();
2020       return true;
2021     } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2022       _scale = n->in(1)->get_int();
2023       return true;
2024     }
2025   } else if (opc == Op_LShiftI) {
2026     if (n->in(1) == iv() && n->in(2)->is_Con()) {
2027       _scale = 1 << n->in(2)->get_int();
2028       return true;
2029     }
2030   } else if (opc == Op_ConvI2L) {
2031     if (scaled_iv_plus_offset(n->in(1))) {
2032       return true;
2033     }
2034   } else if (opc == Op_LShiftL) {
2035     if (!has_iv() && _invar == NULL) {
2036       // Need to preserve the current _offset value, so
2037       // create a temporary object for this expression subtree.
2038       // Hacky, so should re-engineer the address pattern match.
2039       SWPointer tmp(this);
2040       if (tmp.scaled_iv_plus_offset(n->in(1))) {
2041         if (tmp._invar == NULL) {
2042           int mult = 1 << n->in(2)->get_int();
2043           _scale   = tmp._scale  * mult;
2044           _offset += tmp._offset * mult;
2045           return true;
2046         }
2047       }
2048     }
2049   }
2050   return false;
2051 }
2052 
2053 //----------------------------offset_plus_k------------------------
2054 // Match: offset is (k [+/- invariant])
2055 // where k maybe zero and invariant is optional, but not both.
2056 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2057   int opc = n->Opcode();
2058   if (opc == Op_ConI) {
2059     _offset += negate ? -(n->get_int()) : n->get_int();
2060     return true;
2061   } else if (opc == Op_ConL) {
2062     // Okay if value fits into an int
2063     const TypeLong* t = n->find_long_type();
2064     if (t->higher_equal(TypeLong::INT)) {
2065       jlong loff = n->get_long();
2066       jint  off  = (jint)loff;
2067       _offset += negate ? -off : loff;
2068       return true;
2069     }
2070     return false;
2071   }
2072   if (_invar != NULL) return false; // already have an invariant
2073   if (opc == Op_AddI) {
2074     if (n->in(2)->is_Con() && invariant(n->in(1))) {
2075       _negate_invar = negate;
2076       _invar = n->in(1);
2077       _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2078       return true;
2079     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2080       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2081       _negate_invar = negate;
2082       _invar = n->in(2);
2083       return true;
2084     }
2085   }
2086   if (opc == Op_SubI) {
2087     if (n->in(2)->is_Con() && invariant(n->in(1))) {
2088       _negate_invar = negate;
2089       _invar = n->in(1);
2090       _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2091       return true;
2092     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2093       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2094       _negate_invar = !negate;
2095       _invar = n->in(2);
2096       return true;
2097     }
2098   }
2099   if (invariant(n)) {
2100     _negate_invar = negate;
2101     _invar = n;
2102     return true;
2103   }
2104   return false;
2105 }
2106 
2107 //----------------------------print------------------------
2108 void SWPointer::print() {
2109 #ifndef PRODUCT
2110   tty->print("base: %d  adr: %d  scale: %d  offset: %d  invar: %c%d\n",
2111              _base != NULL ? _base->_idx : 0,
2112              _adr  != NULL ? _adr->_idx  : 0,
2113              _scale, _offset,
2114              _negate_invar?'-':'+',
2115              _invar != NULL ? _invar->_idx : 0);
2116 #endif
2117 }
2118 
2119 // ========================= OrderedPair =====================
2120 
2121 const OrderedPair OrderedPair::initial;
2122 
2123 // ========================= SWNodeInfo =====================
2124 
2125 const SWNodeInfo SWNodeInfo::initial;
2126 
2127 
2128 // ============================ DepGraph ===========================
2129 
2130 //------------------------------make_node---------------------------
2131 // Make a new dependence graph node for an ideal node.
2132 DepMem* DepGraph::make_node(Node* node) {
2133   DepMem* m = new (_arena) DepMem(node);
2134   if (node != NULL) {
2135     assert(_map.at_grow(node->_idx) == NULL, "one init only");
2136     _map.at_put_grow(node->_idx, m);
2137   }
2138   return m;
2139 }
2140 
2141 //------------------------------make_edge---------------------------
2142 // Make a new dependence graph edge from dpred -> dsucc
2143 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2144   DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2145   dpred->set_out_head(e);
2146   dsucc->set_in_head(e);
2147   return e;
2148 }
2149 
2150 // ========================== DepMem ========================
2151 
2152 //------------------------------in_cnt---------------------------
2153 int DepMem::in_cnt() {
2154   int ct = 0;
2155   for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2156   return ct;
2157 }
2158 
2159 //------------------------------out_cnt---------------------------
2160 int DepMem::out_cnt() {
2161   int ct = 0;
2162   for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2163   return ct;
2164 }
2165 
2166 //------------------------------print-----------------------------
2167 void DepMem::print() {
2168 #ifndef PRODUCT
2169   tty->print("  DepNode %d (", _node->_idx);
2170   for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2171     Node* pred = p->pred()->node();
2172     tty->print(" %d", pred != NULL ? pred->_idx : 0);
2173   }
2174   tty->print(") [");
2175   for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2176     Node* succ = s->succ()->node();
2177     tty->print(" %d", succ != NULL ? succ->_idx : 0);
2178   }
2179   tty->print_cr(" ]");
2180 #endif
2181 }
2182 
2183 // =========================== DepEdge =========================
2184 
2185 //------------------------------DepPreds---------------------------
2186 void DepEdge::print() {
2187 #ifndef PRODUCT
2188   tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2189 #endif
2190 }
2191 
2192 // =========================== DepPreds =========================
2193 // Iterator over predecessor edges in the dependence graph.
2194 
2195 //------------------------------DepPreds---------------------------
2196 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2197   _n = n;
2198   _done = false;
2199   if (_n->is_Store() || _n->is_Load()) {
2200     _next_idx = MemNode::Address;
2201     _end_idx  = n->req();
2202     _dep_next = dg.dep(_n)->in_head();
2203   } else if (_n->is_Mem()) {
2204     _next_idx = 0;
2205     _end_idx  = 0;
2206     _dep_next = dg.dep(_n)->in_head();
2207   } else {
2208     _next_idx = 1;
2209     _end_idx  = _n->req();
2210     _dep_next = NULL;
2211   }
2212   next();
2213 }
2214 
2215 //------------------------------next---------------------------
2216 void DepPreds::next() {
2217   if (_dep_next != NULL) {
2218     _current  = _dep_next->pred()->node();
2219     _dep_next = _dep_next->next_in();
2220   } else if (_next_idx < _end_idx) {
2221     _current  = _n->in(_next_idx++);
2222   } else {
2223     _done = true;
2224   }
2225 }
2226 
2227 // =========================== DepSuccs =========================
2228 // Iterator over successor edges in the dependence graph.
2229 
2230 //------------------------------DepSuccs---------------------------
2231 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2232   _n = n;
2233   _done = false;
2234   if (_n->is_Load()) {
2235     _next_idx = 0;
2236     _end_idx  = _n->outcnt();
2237     _dep_next = dg.dep(_n)->out_head();
2238   } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2239     _next_idx = 0;
2240     _end_idx  = 0;
2241     _dep_next = dg.dep(_n)->out_head();
2242   } else {
2243     _next_idx = 0;
2244     _end_idx  = _n->outcnt();
2245     _dep_next = NULL;
2246   }
2247   next();
2248 }
2249 
2250 //-------------------------------next---------------------------
2251 void DepSuccs::next() {
2252   if (_dep_next != NULL) {
2253     _current  = _dep_next->succ()->node();
2254     _dep_next = _dep_next->next_out();
2255   } else if (_next_idx < _end_idx) {
2256     _current  = _n->raw_out(_next_idx++);
2257   } else {
2258     _done = true;
2259   }
2260 }