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